Index: head/sys/amd64/sgx/sgx.c =================================================================== --- head/sys/amd64/sgx/sgx.c (revision 353538) +++ head/sys/amd64/sgx/sgx.c (revision 353539) @@ -1,1218 +1,1218 @@ /*- * Copyright (c) 2017 Ruslan Bukin * All rights reserved. * * This software was developed by BAE Systems, the University of Cambridge * Computer Laboratory, and Memorial University under DARPA/AFRL contract * FA8650-15-C-7558 ("CADETS"), as part of the DARPA Transparent Computing * (TC) research program. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ /* * Design overview. * * The driver provides character device for mmap(2) and ioctl(2) system calls * allowing user to manage isolated compartments ("enclaves") in user VA space. * * The driver duties is EPC pages management, enclave management, user data * validation. * * This driver requires Intel SGX support from hardware. * * /dev/sgx: * .mmap: * sgx_mmap_single() allocates VM object with following pager * operations: * a) sgx_pg_ctor(): * VM object constructor does nothing * b) sgx_pg_dtor(): * VM object destructor destroys the SGX enclave associated * with the object: it frees all the EPC pages allocated for * enclave and removes the enclave. * c) sgx_pg_fault(): * VM object fault handler does nothing * * .ioctl: * sgx_ioctl(): * a) SGX_IOC_ENCLAVE_CREATE * Adds Enclave SECS page: initial step of enclave creation. * b) SGX_IOC_ENCLAVE_ADD_PAGE * Adds TCS, REG pages to the enclave. * c) SGX_IOC_ENCLAVE_INIT * Finalizes enclave creation. * * Enclave lifecycle: * .-- ECREATE -- Add SECS page * Kernel | EADD -- Add TCS, REG pages * space | EEXTEND -- Measure the page (take unique hash) * ENCLS | EPA -- Allocate version array page * '-- EINIT -- Finalize enclave creation * User .-- EENTER -- Go to entry point of enclave * space | EEXIT -- Exit back to main application * ENCLU '-- ERESUME -- Resume enclave execution (e.g. after exception) * * Enclave lifecycle from driver point of view: * 1) User calls mmap() on /dev/sgx: we allocate a VM object * 2) User calls ioctl SGX_IOC_ENCLAVE_CREATE: we look for the VM object * associated with user process created on step 1, create SECS physical * page and store it in enclave's VM object queue by special index * SGX_SECS_VM_OBJECT_INDEX. * 3) User calls ioctl SGX_IOC_ENCLAVE_ADD_PAGE: we look for enclave created * on step 2, create TCS or REG physical page and map it to specified by * user address of enclave VM object. * 4) User finalizes enclave creation with ioctl SGX_IOC_ENCLAVE_INIT call. * 5) User can freely enter to and exit from enclave using ENCLU instructions * from userspace: the driver does nothing here. * 6) User proceed munmap(2) system call (or the process with enclave dies): * we destroy the enclave associated with the object. * * EPC page types and their indexes in VM object queue: * - PT_SECS index is special and equals SGX_SECS_VM_OBJECT_INDEX (-1); * - PT_TCS and PT_REG indexes are specified by user in addr field of ioctl * request data and determined as follows: * pidx = OFF_TO_IDX(addp->addr - vmh->base); * - PT_VA index is special, created for PT_REG, PT_TCS and PT_SECS pages * and determined by formula: * va_page_idx = - SGX_VA_PAGES_OFFS - (page_idx / SGX_VA_PAGE_SLOTS); * PT_VA page can hold versions of up to 512 pages, and slot for each * page in PT_VA page is determined as follows: * va_slot_idx = page_idx % SGX_VA_PAGE_SLOTS; * - PT_TRIM is unused. * * Locking: * SGX ENCLS set of instructions have limitations on concurrency: * some instructions can't be executed same time on different CPUs. * We use sc->mtx_encls lock around them to prevent concurrent execution. * sc->mtx lock is used to manage list of created enclaves and the state of * SGX driver. * * Eviction of EPC pages: * Eviction support is not implemented in this driver, however the driver * manages VA (version array) pages: it allocates a VA slot for each EPC * page. This will be required for eviction support in future. * VA pages and slots are currently unused. * * Intel® 64 and IA-32 Architectures Software Developer's Manual * https://software.intel.com/en-us/articles/intel-sdm */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #define SGX_DEBUG #undef SGX_DEBUG #ifdef SGX_DEBUG #define dprintf(fmt, ...) printf(fmt, ##__VA_ARGS__) #else #define dprintf(fmt, ...) #endif static struct cdev_pager_ops sgx_pg_ops; struct sgx_softc sgx_sc; static int sgx_get_epc_page(struct sgx_softc *sc, struct epc_page **epc) { vmem_addr_t addr; int i; if (vmem_alloc(sc->vmem_epc, PAGE_SIZE, M_FIRSTFIT | M_NOWAIT, &addr) == 0) { i = (addr - sc->epc_base) / PAGE_SIZE; *epc = &sc->epc_pages[i]; return (0); } return (ENOMEM); } static void sgx_put_epc_page(struct sgx_softc *sc, struct epc_page *epc) { vmem_addr_t addr; if (epc == NULL) return; addr = (epc->index * PAGE_SIZE) + sc->epc_base; vmem_free(sc->vmem_epc, addr, PAGE_SIZE); } static int sgx_va_slot_init_by_index(struct sgx_softc *sc, vm_object_t object, uint64_t idx) { struct epc_page *epc; vm_page_t page; vm_page_t p; int ret; VM_OBJECT_ASSERT_WLOCKED(object); p = vm_page_lookup(object, idx); if (p == NULL) { ret = sgx_get_epc_page(sc, &epc); if (ret) { dprintf("%s: No free EPC pages available.\n", __func__); return (ret); } mtx_lock(&sc->mtx_encls); sgx_epa((void *)epc->base); mtx_unlock(&sc->mtx_encls); page = PHYS_TO_VM_PAGE(epc->phys); - vm_page_insert(page, object, idx); page->valid = VM_PAGE_BITS_ALL; + vm_page_insert(page, object, idx); } return (0); } static int sgx_va_slot_init(struct sgx_softc *sc, struct sgx_enclave *enclave, uint64_t addr) { vm_pindex_t pidx; uint64_t va_page_idx; uint64_t idx; vm_object_t object; int va_slot; int ret; object = enclave->object; VM_OBJECT_ASSERT_WLOCKED(object); pidx = OFF_TO_IDX(addr); va_slot = pidx % SGX_VA_PAGE_SLOTS; va_page_idx = pidx / SGX_VA_PAGE_SLOTS; idx = - SGX_VA_PAGES_OFFS - va_page_idx; ret = sgx_va_slot_init_by_index(sc, object, idx); return (ret); } static int sgx_mem_find(struct sgx_softc *sc, uint64_t addr, vm_map_entry_t *entry0, vm_object_t *object0) { vm_map_t map; vm_map_entry_t entry; vm_object_t object; map = &curproc->p_vmspace->vm_map; vm_map_lock_read(map); if (!vm_map_lookup_entry(map, addr, &entry)) { vm_map_unlock_read(map); dprintf("%s: Can't find enclave.\n", __func__); return (EINVAL); } object = entry->object.vm_object; if (object == NULL || object->handle == NULL) { vm_map_unlock_read(map); return (EINVAL); } if (object->type != OBJT_MGTDEVICE || object->un_pager.devp.ops != &sgx_pg_ops) { vm_map_unlock_read(map); return (EINVAL); } vm_object_reference(object); *object0 = object; *entry0 = entry; vm_map_unlock_read(map); return (0); } static int sgx_enclave_find(struct sgx_softc *sc, uint64_t addr, struct sgx_enclave **encl) { struct sgx_vm_handle *vmh; struct sgx_enclave *enclave; vm_map_entry_t entry; vm_object_t object; int ret; ret = sgx_mem_find(sc, addr, &entry, &object); if (ret) return (ret); vmh = object->handle; if (vmh == NULL) { vm_object_deallocate(object); return (EINVAL); } enclave = vmh->enclave; if (enclave == NULL || enclave->object == NULL) { vm_object_deallocate(object); return (EINVAL); } *encl = enclave; return (0); } static int sgx_enclave_alloc(struct sgx_softc *sc, struct secs *secs, struct sgx_enclave **enclave0) { struct sgx_enclave *enclave; enclave = malloc(sizeof(struct sgx_enclave), M_SGX, M_WAITOK | M_ZERO); enclave->base = secs->base; enclave->size = secs->size; *enclave0 = enclave; return (0); } static void sgx_epc_page_remove(struct sgx_softc *sc, struct epc_page *epc) { mtx_lock(&sc->mtx_encls); sgx_eremove((void *)epc->base); mtx_unlock(&sc->mtx_encls); } static void sgx_page_remove(struct sgx_softc *sc, vm_page_t p) { struct epc_page *epc; vm_paddr_t pa; uint64_t offs; (void)vm_page_remove(p); dprintf("%s: p->pidx %ld\n", __func__, p->pindex); pa = VM_PAGE_TO_PHYS(p); epc = &sc->epc_pages[0]; offs = (pa - epc->phys) / PAGE_SIZE; epc = &sc->epc_pages[offs]; sgx_epc_page_remove(sc, epc); sgx_put_epc_page(sc, epc); } static void sgx_enclave_remove(struct sgx_softc *sc, struct sgx_enclave *enclave) { vm_object_t object; vm_page_t p, p_secs, p_next; mtx_lock(&sc->mtx); TAILQ_REMOVE(&sc->enclaves, enclave, next); mtx_unlock(&sc->mtx); object = enclave->object; VM_OBJECT_WLOCK(object); /* * First remove all the pages except SECS, * then remove SECS page. */ p_secs = NULL; TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) { if (p->pindex == SGX_SECS_VM_OBJECT_INDEX) { p_secs = p; continue; } sgx_page_remove(sc, p); } /* Now remove SECS page */ if (p_secs != NULL) sgx_page_remove(sc, p_secs); KASSERT(TAILQ_EMPTY(&object->memq) == 1, ("not empty")); KASSERT(object->resident_page_count == 0, ("count")); VM_OBJECT_WUNLOCK(object); } static int sgx_measure_page(struct sgx_softc *sc, struct epc_page *secs, struct epc_page *epc, uint16_t mrmask) { int i, j; int ret; mtx_lock(&sc->mtx_encls); for (i = 0, j = 1; i < PAGE_SIZE; i += 0x100, j <<= 1) { if (!(j & mrmask)) continue; ret = sgx_eextend((void *)secs->base, (void *)(epc->base + i)); if (ret == SGX_EFAULT) { mtx_unlock(&sc->mtx_encls); return (ret); } } mtx_unlock(&sc->mtx_encls); return (0); } static int sgx_secs_validate(struct sgx_softc *sc, struct secs *secs) { struct secs_attr *attr; int i; if (secs->size == 0) return (EINVAL); /* BASEADDR must be naturally aligned on an SECS.SIZE boundary. */ if (secs->base & (secs->size - 1)) return (EINVAL); /* SECS.SIZE must be at least 2 pages. */ if (secs->size < 2 * PAGE_SIZE) return (EINVAL); if ((secs->size & (secs->size - 1)) != 0) return (EINVAL); attr = &secs->attributes; if (attr->reserved1 != 0 || attr->reserved2 != 0 || attr->reserved3 != 0) return (EINVAL); for (i = 0; i < SECS_ATTR_RSV4_SIZE; i++) if (attr->reserved4[i]) return (EINVAL); /* * Intel® Software Guard Extensions Programming Reference * 6.7.2 Relevant Fields in Various Data Structures * 6.7.2.1 SECS.ATTRIBUTES.XFRM * XFRM[1:0] must be set to 0x3. */ if ((attr->xfrm & 0x3) != 0x3) return (EINVAL); if (!attr->mode64bit) return (EINVAL); if (secs->size > sc->enclave_size_max) return (EINVAL); for (i = 0; i < SECS_RSV1_SIZE; i++) if (secs->reserved1[i]) return (EINVAL); for (i = 0; i < SECS_RSV2_SIZE; i++) if (secs->reserved2[i]) return (EINVAL); for (i = 0; i < SECS_RSV3_SIZE; i++) if (secs->reserved3[i]) return (EINVAL); for (i = 0; i < SECS_RSV4_SIZE; i++) if (secs->reserved4[i]) return (EINVAL); return (0); } static int sgx_tcs_validate(struct tcs *tcs) { int i; if ((tcs->flags) || (tcs->ossa & (PAGE_SIZE - 1)) || (tcs->ofsbasgx & (PAGE_SIZE - 1)) || (tcs->ogsbasgx & (PAGE_SIZE - 1)) || ((tcs->fslimit & 0xfff) != 0xfff) || ((tcs->gslimit & 0xfff) != 0xfff)) return (EINVAL); for (i = 0; i < nitems(tcs->reserved3); i++) if (tcs->reserved3[i]) return (EINVAL); return (0); } static void sgx_tcs_dump(struct sgx_softc *sc, struct tcs *t) { dprintf("t->flags %lx\n", t->flags); dprintf("t->ossa %lx\n", t->ossa); dprintf("t->cssa %x\n", t->cssa); dprintf("t->nssa %x\n", t->nssa); dprintf("t->oentry %lx\n", t->oentry); dprintf("t->ofsbasgx %lx\n", t->ofsbasgx); dprintf("t->ogsbasgx %lx\n", t->ogsbasgx); dprintf("t->fslimit %x\n", t->fslimit); dprintf("t->gslimit %x\n", t->gslimit); } static int sgx_pg_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t foff, struct ucred *cred, u_short *color) { struct sgx_vm_handle *vmh; vmh = handle; if (vmh == NULL) { dprintf("%s: vmh not found.\n", __func__); return (0); } dprintf("%s: vmh->base %lx foff 0x%lx size 0x%lx\n", __func__, vmh->base, foff, size); return (0); } static void sgx_pg_dtor(void *handle) { struct sgx_vm_handle *vmh; struct sgx_softc *sc; vmh = handle; if (vmh == NULL) { dprintf("%s: vmh not found.\n", __func__); return; } sc = vmh->sc; if (sc == NULL) { dprintf("%s: sc is NULL\n", __func__); return; } if (vmh->enclave == NULL) { dprintf("%s: Enclave not found.\n", __func__); return; } sgx_enclave_remove(sc, vmh->enclave); free(vmh->enclave, M_SGX); free(vmh, M_SGX); } static int sgx_pg_fault(vm_object_t object, vm_ooffset_t offset, int prot, vm_page_t *mres) { /* * The purpose of this trivial handler is to handle the race * when user tries to access mmaped region before or during * enclave creation ioctl calls. */ dprintf("%s: offset 0x%lx\n", __func__, offset); return (VM_PAGER_FAIL); } static struct cdev_pager_ops sgx_pg_ops = { .cdev_pg_ctor = sgx_pg_ctor, .cdev_pg_dtor = sgx_pg_dtor, .cdev_pg_fault = sgx_pg_fault, }; static void sgx_insert_epc_page_by_index(vm_page_t page, vm_object_t object, vm_pindex_t pidx) { VM_OBJECT_ASSERT_WLOCKED(object); - vm_page_insert(page, object, pidx); page->valid = VM_PAGE_BITS_ALL; + vm_page_insert(page, object, pidx); } static void sgx_insert_epc_page(struct sgx_enclave *enclave, struct epc_page *epc, uint64_t addr) { vm_pindex_t pidx; vm_page_t page; VM_OBJECT_ASSERT_WLOCKED(enclave->object); pidx = OFF_TO_IDX(addr); page = PHYS_TO_VM_PAGE(epc->phys); sgx_insert_epc_page_by_index(page, enclave->object, pidx); } static int sgx_ioctl_create(struct sgx_softc *sc, struct sgx_enclave_create *param) { struct sgx_vm_handle *vmh; vm_map_entry_t entry; vm_page_t p; struct page_info pginfo; struct secinfo secinfo; struct sgx_enclave *enclave; struct epc_page *epc; struct secs *secs; vm_object_t object; vm_page_t page; int ret; epc = NULL; secs = NULL; enclave = NULL; object = NULL; /* SGX Enclave Control Structure (SECS) */ secs = malloc(PAGE_SIZE, M_SGX, M_WAITOK | M_ZERO); ret = copyin((void *)param->src, secs, sizeof(struct secs)); if (ret) { dprintf("%s: Can't copy SECS.\n", __func__); goto error; } ret = sgx_secs_validate(sc, secs); if (ret) { dprintf("%s: SECS validation failed.\n", __func__); goto error; } ret = sgx_mem_find(sc, secs->base, &entry, &object); if (ret) { dprintf("%s: Can't find vm_map.\n", __func__); goto error; } vmh = object->handle; if (!vmh) { dprintf("%s: Can't find vmh.\n", __func__); ret = ENXIO; goto error; } dprintf("%s: entry start %lx offset %lx\n", __func__, entry->start, entry->offset); vmh->base = (entry->start - entry->offset); ret = sgx_enclave_alloc(sc, secs, &enclave); if (ret) { dprintf("%s: Can't alloc enclave.\n", __func__); goto error; } enclave->object = object; enclave->vmh = vmh; memset(&secinfo, 0, sizeof(struct secinfo)); memset(&pginfo, 0, sizeof(struct page_info)); pginfo.linaddr = 0; pginfo.srcpge = (uint64_t)secs; pginfo.secinfo = &secinfo; pginfo.secs = 0; ret = sgx_get_epc_page(sc, &epc); if (ret) { dprintf("%s: Failed to get free epc page.\n", __func__); goto error; } enclave->secs_epc_page = epc; VM_OBJECT_WLOCK(object); p = vm_page_lookup(object, SGX_SECS_VM_OBJECT_INDEX); if (p) { VM_OBJECT_WUNLOCK(object); /* SECS page already added. */ ret = ENXIO; goto error; } ret = sgx_va_slot_init_by_index(sc, object, - SGX_VA_PAGES_OFFS - SGX_SECS_VM_OBJECT_INDEX); if (ret) { VM_OBJECT_WUNLOCK(object); dprintf("%s: Can't init va slot.\n", __func__); goto error; } mtx_lock(&sc->mtx); if ((sc->state & SGX_STATE_RUNNING) == 0) { mtx_unlock(&sc->mtx); /* Remove VA page that was just created for SECS page. */ p = vm_page_lookup(enclave->object, - SGX_VA_PAGES_OFFS - SGX_SECS_VM_OBJECT_INDEX); sgx_page_remove(sc, p); VM_OBJECT_WUNLOCK(object); goto error; } mtx_lock(&sc->mtx_encls); ret = sgx_ecreate(&pginfo, (void *)epc->base); mtx_unlock(&sc->mtx_encls); if (ret == SGX_EFAULT) { dprintf("%s: gp fault\n", __func__); mtx_unlock(&sc->mtx); /* Remove VA page that was just created for SECS page. */ p = vm_page_lookup(enclave->object, - SGX_VA_PAGES_OFFS - SGX_SECS_VM_OBJECT_INDEX); sgx_page_remove(sc, p); VM_OBJECT_WUNLOCK(object); goto error; } TAILQ_INSERT_TAIL(&sc->enclaves, enclave, next); mtx_unlock(&sc->mtx); vmh->enclave = enclave; page = PHYS_TO_VM_PAGE(epc->phys); sgx_insert_epc_page_by_index(page, enclave->object, SGX_SECS_VM_OBJECT_INDEX); VM_OBJECT_WUNLOCK(object); /* Release the reference. */ vm_object_deallocate(object); free(secs, M_SGX); return (0); error: free(secs, M_SGX); sgx_put_epc_page(sc, epc); free(enclave, M_SGX); vm_object_deallocate(object); return (ret); } static int sgx_ioctl_add_page(struct sgx_softc *sc, struct sgx_enclave_add_page *addp) { struct epc_page *secs_epc_page; struct sgx_enclave *enclave; struct sgx_vm_handle *vmh; struct epc_page *epc; struct page_info pginfo; struct secinfo secinfo; vm_object_t object; void *tmp_vaddr; uint64_t page_type; struct tcs *t; uint64_t addr; uint64_t pidx; vm_page_t p; int ret; tmp_vaddr = NULL; epc = NULL; object = NULL; /* Find and get reference to VM object. */ ret = sgx_enclave_find(sc, addp->addr, &enclave); if (ret) { dprintf("%s: Failed to find enclave.\n", __func__); goto error; } object = enclave->object; KASSERT(object != NULL, ("vm object is NULL\n")); vmh = object->handle; ret = sgx_get_epc_page(sc, &epc); if (ret) { dprintf("%s: Failed to get free epc page.\n", __func__); goto error; } memset(&secinfo, 0, sizeof(struct secinfo)); ret = copyin((void *)addp->secinfo, &secinfo, sizeof(struct secinfo)); if (ret) { dprintf("%s: Failed to copy secinfo.\n", __func__); goto error; } tmp_vaddr = malloc(PAGE_SIZE, M_SGX, M_WAITOK | M_ZERO); ret = copyin((void *)addp->src, tmp_vaddr, PAGE_SIZE); if (ret) { dprintf("%s: Failed to copy page.\n", __func__); goto error; } page_type = (secinfo.flags & SECINFO_FLAGS_PT_M) >> SECINFO_FLAGS_PT_S; if (page_type != SGX_PT_TCS && page_type != SGX_PT_REG) { dprintf("%s: page can't be added.\n", __func__); goto error; } if (page_type == SGX_PT_TCS) { t = (struct tcs *)tmp_vaddr; ret = sgx_tcs_validate(t); if (ret) { dprintf("%s: TCS page validation failed.\n", __func__); goto error; } sgx_tcs_dump(sc, t); } addr = (addp->addr - vmh->base); pidx = OFF_TO_IDX(addr); VM_OBJECT_WLOCK(object); p = vm_page_lookup(object, pidx); if (p) { VM_OBJECT_WUNLOCK(object); /* Page already added. */ ret = ENXIO; goto error; } ret = sgx_va_slot_init(sc, enclave, addr); if (ret) { VM_OBJECT_WUNLOCK(object); dprintf("%s: Can't init va slot.\n", __func__); goto error; } secs_epc_page = enclave->secs_epc_page; memset(&pginfo, 0, sizeof(struct page_info)); pginfo.linaddr = (uint64_t)addp->addr; pginfo.srcpge = (uint64_t)tmp_vaddr; pginfo.secinfo = &secinfo; pginfo.secs = (uint64_t)secs_epc_page->base; mtx_lock(&sc->mtx_encls); ret = sgx_eadd(&pginfo, (void *)epc->base); if (ret == SGX_EFAULT) { dprintf("%s: gp fault on eadd\n", __func__); mtx_unlock(&sc->mtx_encls); VM_OBJECT_WUNLOCK(object); goto error; } mtx_unlock(&sc->mtx_encls); ret = sgx_measure_page(sc, enclave->secs_epc_page, epc, addp->mrmask); if (ret == SGX_EFAULT) { dprintf("%s: gp fault on eextend\n", __func__); sgx_epc_page_remove(sc, epc); VM_OBJECT_WUNLOCK(object); goto error; } sgx_insert_epc_page(enclave, epc, addr); VM_OBJECT_WUNLOCK(object); /* Release the reference. */ vm_object_deallocate(object); free(tmp_vaddr, M_SGX); return (0); error: free(tmp_vaddr, M_SGX); sgx_put_epc_page(sc, epc); vm_object_deallocate(object); return (ret); } static int sgx_ioctl_init(struct sgx_softc *sc, struct sgx_enclave_init *initp) { struct epc_page *secs_epc_page; struct sgx_enclave *enclave; struct thread *td; void *tmp_vaddr; void *einittoken; void *sigstruct; vm_object_t object; int retry; int ret; td = curthread; tmp_vaddr = NULL; object = NULL; dprintf("%s: addr %lx, sigstruct %lx, einittoken %lx\n", __func__, initp->addr, initp->sigstruct, initp->einittoken); /* Find and get reference to VM object. */ ret = sgx_enclave_find(sc, initp->addr, &enclave); if (ret) { dprintf("%s: Failed to find enclave.\n", __func__); goto error; } object = enclave->object; tmp_vaddr = malloc(PAGE_SIZE, M_SGX, M_WAITOK | M_ZERO); sigstruct = tmp_vaddr; einittoken = (void *)((uint64_t)sigstruct + PAGE_SIZE / 2); ret = copyin((void *)initp->sigstruct, sigstruct, SGX_SIGSTRUCT_SIZE); if (ret) { dprintf("%s: Failed to copy SIGSTRUCT page.\n", __func__); goto error; } ret = copyin((void *)initp->einittoken, einittoken, SGX_EINITTOKEN_SIZE); if (ret) { dprintf("%s: Failed to copy EINITTOKEN page.\n", __func__); goto error; } secs_epc_page = enclave->secs_epc_page; retry = 16; do { mtx_lock(&sc->mtx_encls); ret = sgx_einit(sigstruct, (void *)secs_epc_page->base, einittoken); mtx_unlock(&sc->mtx_encls); dprintf("%s: sgx_einit returned %d\n", __func__, ret); } while (ret == SGX_UNMASKED_EVENT && retry--); if (ret) { dprintf("%s: Failed init enclave: %d\n", __func__, ret); td->td_retval[0] = ret; ret = 0; } error: free(tmp_vaddr, M_SGX); /* Release the reference. */ vm_object_deallocate(object); return (ret); } static int sgx_ioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flags, struct thread *td) { struct sgx_enclave_add_page *addp; struct sgx_enclave_create *param; struct sgx_enclave_init *initp; struct sgx_softc *sc; int ret; int len; sc = &sgx_sc; len = IOCPARM_LEN(cmd); dprintf("%s: cmd %lx, addr %lx, len %d\n", __func__, cmd, (uint64_t)addr, len); if (len > SGX_IOCTL_MAX_DATA_LEN) return (EINVAL); switch (cmd) { case SGX_IOC_ENCLAVE_CREATE: param = (struct sgx_enclave_create *)addr; ret = sgx_ioctl_create(sc, param); break; case SGX_IOC_ENCLAVE_ADD_PAGE: addp = (struct sgx_enclave_add_page *)addr; ret = sgx_ioctl_add_page(sc, addp); break; case SGX_IOC_ENCLAVE_INIT: initp = (struct sgx_enclave_init *)addr; ret = sgx_ioctl_init(sc, initp); break; default: return (EINVAL); } return (ret); } static int sgx_mmap_single(struct cdev *cdev, vm_ooffset_t *offset, vm_size_t mapsize, struct vm_object **objp, int nprot) { struct sgx_vm_handle *vmh; struct sgx_softc *sc; sc = &sgx_sc; dprintf("%s: mapsize 0x%lx, offset %lx\n", __func__, mapsize, *offset); vmh = malloc(sizeof(struct sgx_vm_handle), M_SGX, M_WAITOK | M_ZERO); vmh->sc = sc; vmh->size = mapsize; vmh->mem = cdev_pager_allocate(vmh, OBJT_MGTDEVICE, &sgx_pg_ops, mapsize, nprot, *offset, NULL); if (vmh->mem == NULL) { free(vmh, M_SGX); return (ENOMEM); } VM_OBJECT_WLOCK(vmh->mem); vm_object_set_flag(vmh->mem, OBJ_PG_DTOR); VM_OBJECT_WUNLOCK(vmh->mem); *objp = vmh->mem; return (0); } static struct cdevsw sgx_cdevsw = { .d_version = D_VERSION, .d_ioctl = sgx_ioctl, .d_mmap_single = sgx_mmap_single, .d_name = "Intel SGX", }; static int sgx_get_epc_area(struct sgx_softc *sc) { vm_offset_t epc_base_vaddr; u_int cp[4]; int error; int i; cpuid_count(SGX_CPUID, 0x2, cp); sc->epc_base = ((uint64_t)(cp[1] & 0xfffff) << 32) + (cp[0] & 0xfffff000); sc->epc_size = ((uint64_t)(cp[3] & 0xfffff) << 32) + (cp[2] & 0xfffff000); sc->npages = sc->epc_size / SGX_PAGE_SIZE; if (sc->epc_size == 0 || sc->epc_base == 0) { printf("%s: Incorrect EPC data: EPC base %lx, size %lu\n", __func__, sc->epc_base, sc->epc_size); return (EINVAL); } if (cp[3] & 0xffff) sc->enclave_size_max = (1 << ((cp[3] >> 8) & 0xff)); else sc->enclave_size_max = SGX_ENCL_SIZE_MAX_DEF; epc_base_vaddr = (vm_offset_t)pmap_mapdev_attr(sc->epc_base, sc->epc_size, VM_MEMATTR_DEFAULT); sc->epc_pages = malloc(sizeof(struct epc_page) * sc->npages, M_DEVBUF, M_WAITOK | M_ZERO); for (i = 0; i < sc->npages; i++) { sc->epc_pages[i].base = epc_base_vaddr + SGX_PAGE_SIZE * i; sc->epc_pages[i].phys = sc->epc_base + SGX_PAGE_SIZE * i; sc->epc_pages[i].index = i; } sc->vmem_epc = vmem_create("SGX EPC", sc->epc_base, sc->epc_size, PAGE_SIZE, PAGE_SIZE, M_FIRSTFIT | M_WAITOK); if (sc->vmem_epc == NULL) { printf("%s: Can't create vmem arena.\n", __func__); free(sc->epc_pages, M_SGX); return (EINVAL); } error = vm_phys_fictitious_reg_range(sc->epc_base, sc->epc_base + sc->epc_size, VM_MEMATTR_DEFAULT); if (error) { printf("%s: Can't register fictitious space.\n", __func__); free(sc->epc_pages, M_SGX); return (EINVAL); } return (0); } static void sgx_put_epc_area(struct sgx_softc *sc) { vm_phys_fictitious_unreg_range(sc->epc_base, sc->epc_base + sc->epc_size); free(sc->epc_pages, M_SGX); } static int sgx_load(void) { struct sgx_softc *sc; int error; sc = &sgx_sc; if ((cpu_stdext_feature & CPUID_STDEXT_SGX) == 0) return (ENXIO); error = sgx_get_epc_area(sc); if (error) { printf("%s: Failed to get Processor Reserved Memory area.\n", __func__); return (ENXIO); } mtx_init(&sc->mtx_encls, "SGX ENCLS", NULL, MTX_DEF); mtx_init(&sc->mtx, "SGX driver", NULL, MTX_DEF); TAILQ_INIT(&sc->enclaves); sc->sgx_cdev = make_dev(&sgx_cdevsw, 0, UID_ROOT, GID_WHEEL, 0600, "isgx"); sc->state |= SGX_STATE_RUNNING; printf("SGX initialized: EPC base 0x%lx size %ld (%d pages)\n", sc->epc_base, sc->epc_size, sc->npages); return (0); } static int sgx_unload(void) { struct sgx_softc *sc; sc = &sgx_sc; if ((sc->state & SGX_STATE_RUNNING) == 0) return (0); mtx_lock(&sc->mtx); if (!TAILQ_EMPTY(&sc->enclaves)) { mtx_unlock(&sc->mtx); return (EBUSY); } sc->state &= ~SGX_STATE_RUNNING; mtx_unlock(&sc->mtx); destroy_dev(sc->sgx_cdev); vmem_destroy(sc->vmem_epc); sgx_put_epc_area(sc); mtx_destroy(&sc->mtx_encls); mtx_destroy(&sc->mtx); return (0); } static int sgx_handler(module_t mod, int what, void *arg) { int error; switch (what) { case MOD_LOAD: error = sgx_load(); break; case MOD_UNLOAD: error = sgx_unload(); break; default: error = 0; break; } return (error); } static moduledata_t sgx_kmod = { "sgx", sgx_handler, NULL }; DECLARE_MODULE(sgx, sgx_kmod, SI_SUB_LAST, SI_ORDER_ANY); MODULE_VERSION(sgx, 1); Index: head/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dmu.c =================================================================== --- head/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dmu.c (revision 353538) +++ head/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dmu.c (revision 353539) @@ -1,2734 +1,2740 @@ /* * CDDL HEADER START * * The contents of this file are subject to the terms of the * Common Development and Distribution License (the "License"). * You may not use this file except in compliance with the License. * * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE * or http://www.opensolaris.org/os/licensing. * See the License for the specific language governing permissions * and limitations under the License. * * When distributing Covered Code, include this CDDL HEADER in each * file and include the License file at usr/src/OPENSOLARIS.LICENSE. * If applicable, add the following below this CDDL HEADER, with the * fields enclosed by brackets "[]" replaced with your own identifying * information: Portions Copyright [yyyy] [name of copyright owner] * * CDDL HEADER END */ /* * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. * Copyright (c) 2011, 2017 by Delphix. All rights reserved. * Copyright (c) 2019 Datto Inc. */ /* Copyright (c) 2013 by Saso Kiselkov. All rights reserved. */ /* Copyright (c) 2013, Joyent, Inc. All rights reserved. */ /* Copyright 2016 Nexenta Systems, Inc. All rights reserved. */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef _KERNEL #include #include #include #endif /* * Enable/disable nopwrite feature. */ int zfs_nopwrite_enabled = 1; SYSCTL_DECL(_vfs_zfs); SYSCTL_INT(_vfs_zfs, OID_AUTO, nopwrite_enabled, CTLFLAG_RDTUN, &zfs_nopwrite_enabled, 0, "Enable nopwrite feature"); /* * Tunable to control percentage of dirtied L1 blocks from frees allowed into * one TXG. After this threshold is crossed, additional dirty blocks from frees * will wait until the next TXG. * A value of zero will disable this throttle. */ uint32_t zfs_per_txg_dirty_frees_percent = 5; SYSCTL_INT(_vfs_zfs, OID_AUTO, per_txg_dirty_frees_percent, CTLFLAG_RWTUN, &zfs_per_txg_dirty_frees_percent, 0, "Percentage of dirtied indirect blocks from frees allowed in one txg"); /* * This can be used for testing, to ensure that certain actions happen * while in the middle of a remap (which might otherwise complete too * quickly). */ int zfs_object_remap_one_indirect_delay_ticks = 0; const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = { { DMU_BSWAP_UINT8, TRUE, FALSE, "unallocated" }, { DMU_BSWAP_ZAP, TRUE, TRUE, "object directory" }, { DMU_BSWAP_UINT64, TRUE, TRUE, "object array" }, { DMU_BSWAP_UINT8, TRUE, FALSE, "packed nvlist" }, { DMU_BSWAP_UINT64, TRUE, FALSE, "packed nvlist size" }, { DMU_BSWAP_UINT64, TRUE, FALSE, "bpobj" }, { DMU_BSWAP_UINT64, TRUE, FALSE, "bpobj header" }, { DMU_BSWAP_UINT64, TRUE, FALSE, "SPA space map header" }, { DMU_BSWAP_UINT64, TRUE, FALSE, "SPA space map" }, { DMU_BSWAP_UINT64, TRUE, FALSE, "ZIL intent log" }, { DMU_BSWAP_DNODE, TRUE, FALSE, "DMU dnode" }, { DMU_BSWAP_OBJSET, TRUE, TRUE, "DMU objset" }, { DMU_BSWAP_UINT64, TRUE, TRUE, "DSL directory" }, { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL directory child map" }, { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL dataset snap map" }, { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL props" }, { DMU_BSWAP_UINT64, TRUE, TRUE, "DSL dataset" }, { DMU_BSWAP_ZNODE, TRUE, FALSE, "ZFS znode" }, { DMU_BSWAP_OLDACL, TRUE, FALSE, "ZFS V0 ACL" }, { DMU_BSWAP_UINT8, FALSE, FALSE, "ZFS plain file" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS directory" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS master node" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS delete queue" }, { DMU_BSWAP_UINT8, FALSE, FALSE, "zvol object" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "zvol prop" }, { DMU_BSWAP_UINT8, FALSE, FALSE, "other uint8[]" }, { DMU_BSWAP_UINT64, FALSE, FALSE, "other uint64[]" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "other ZAP" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "persistent error log" }, { DMU_BSWAP_UINT8, TRUE, FALSE, "SPA history" }, { DMU_BSWAP_UINT64, TRUE, FALSE, "SPA history offsets" }, { DMU_BSWAP_ZAP, TRUE, TRUE, "Pool properties" }, { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL permissions" }, { DMU_BSWAP_ACL, TRUE, FALSE, "ZFS ACL" }, { DMU_BSWAP_UINT8, TRUE, FALSE, "ZFS SYSACL" }, { DMU_BSWAP_UINT8, TRUE, FALSE, "FUID table" }, { DMU_BSWAP_UINT64, TRUE, FALSE, "FUID table size" }, { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL dataset next clones" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "scan work queue" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS user/group used" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS user/group quota" }, { DMU_BSWAP_ZAP, TRUE, TRUE, "snapshot refcount tags" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "DDT ZAP algorithm" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "DDT statistics" }, { DMU_BSWAP_UINT8, TRUE, FALSE, "System attributes" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "SA master node" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "SA attr registration" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "SA attr layouts" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "scan translations" }, { DMU_BSWAP_UINT8, FALSE, FALSE, "deduplicated block" }, { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL deadlist map" }, { DMU_BSWAP_UINT64, TRUE, TRUE, "DSL deadlist map hdr" }, { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL dir clones" }, { DMU_BSWAP_UINT64, TRUE, FALSE, "bpobj subobj" } }; const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = { { byteswap_uint8_array, "uint8" }, { byteswap_uint16_array, "uint16" }, { byteswap_uint32_array, "uint32" }, { byteswap_uint64_array, "uint64" }, { zap_byteswap, "zap" }, { dnode_buf_byteswap, "dnode" }, { dmu_objset_byteswap, "objset" }, { zfs_znode_byteswap, "znode" }, { zfs_oldacl_byteswap, "oldacl" }, { zfs_acl_byteswap, "acl" } }; int dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset, void *tag, dmu_buf_t **dbp) { uint64_t blkid; dmu_buf_impl_t *db; blkid = dbuf_whichblock(dn, 0, offset); rw_enter(&dn->dn_struct_rwlock, RW_READER); db = dbuf_hold(dn, blkid, tag); rw_exit(&dn->dn_struct_rwlock); if (db == NULL) { *dbp = NULL; return (SET_ERROR(EIO)); } *dbp = &db->db; return (0); } int dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset, void *tag, dmu_buf_t **dbp) { dnode_t *dn; uint64_t blkid; dmu_buf_impl_t *db; int err; err = dnode_hold(os, object, FTAG, &dn); if (err) return (err); blkid = dbuf_whichblock(dn, 0, offset); rw_enter(&dn->dn_struct_rwlock, RW_READER); db = dbuf_hold(dn, blkid, tag); rw_exit(&dn->dn_struct_rwlock); dnode_rele(dn, FTAG); if (db == NULL) { *dbp = NULL; return (SET_ERROR(EIO)); } *dbp = &db->db; return (err); } int dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset, void *tag, dmu_buf_t **dbp, int flags) { int err; int db_flags = DB_RF_CANFAIL; if (flags & DMU_READ_NO_PREFETCH) db_flags |= DB_RF_NOPREFETCH; err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp); if (err == 0) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp); err = dbuf_read(db, NULL, db_flags); if (err != 0) { dbuf_rele(db, tag); *dbp = NULL; } } return (err); } int dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset, void *tag, dmu_buf_t **dbp, int flags) { int err; int db_flags = DB_RF_CANFAIL; if (flags & DMU_READ_NO_PREFETCH) db_flags |= DB_RF_NOPREFETCH; err = dmu_buf_hold_noread(os, object, offset, tag, dbp); if (err == 0) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp); err = dbuf_read(db, NULL, db_flags); if (err != 0) { dbuf_rele(db, tag); *dbp = NULL; } } return (err); } int dmu_bonus_max(void) { return (DN_OLD_MAX_BONUSLEN); } int dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; dnode_t *dn; int error; DB_DNODE_ENTER(db); dn = DB_DNODE(db); if (dn->dn_bonus != db) { error = SET_ERROR(EINVAL); } else if (newsize < 0 || newsize > db_fake->db_size) { error = SET_ERROR(EINVAL); } else { dnode_setbonuslen(dn, newsize, tx); error = 0; } DB_DNODE_EXIT(db); return (error); } int dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; dnode_t *dn; int error; DB_DNODE_ENTER(db); dn = DB_DNODE(db); if (!DMU_OT_IS_VALID(type)) { error = SET_ERROR(EINVAL); } else if (dn->dn_bonus != db) { error = SET_ERROR(EINVAL); } else { dnode_setbonus_type(dn, type, tx); error = 0; } DB_DNODE_EXIT(db); return (error); } dmu_object_type_t dmu_get_bonustype(dmu_buf_t *db_fake) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; dnode_t *dn; dmu_object_type_t type; DB_DNODE_ENTER(db); dn = DB_DNODE(db); type = dn->dn_bonustype; DB_DNODE_EXIT(db); return (type); } int dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx) { dnode_t *dn; int error; error = dnode_hold(os, object, FTAG, &dn); dbuf_rm_spill(dn, tx); rw_enter(&dn->dn_struct_rwlock, RW_WRITER); dnode_rm_spill(dn, tx); rw_exit(&dn->dn_struct_rwlock); dnode_rele(dn, FTAG); return (error); } /* * returns ENOENT, EIO, or 0. */ int dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp) { dnode_t *dn; dmu_buf_impl_t *db; int error; error = dnode_hold(os, object, FTAG, &dn); if (error) return (error); rw_enter(&dn->dn_struct_rwlock, RW_READER); if (dn->dn_bonus == NULL) { rw_exit(&dn->dn_struct_rwlock); rw_enter(&dn->dn_struct_rwlock, RW_WRITER); if (dn->dn_bonus == NULL) dbuf_create_bonus(dn); } db = dn->dn_bonus; /* as long as the bonus buf is held, the dnode will be held */ if (refcount_add(&db->db_holds, tag) == 1) { VERIFY(dnode_add_ref(dn, db)); atomic_inc_32(&dn->dn_dbufs_count); } /* * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's * hold and incrementing the dbuf count to ensure that dnode_move() sees * a dnode hold for every dbuf. */ rw_exit(&dn->dn_struct_rwlock); dnode_rele(dn, FTAG); VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH)); *dbp = &db->db; return (0); } /* * returns ENOENT, EIO, or 0. * * This interface will allocate a blank spill dbuf when a spill blk * doesn't already exist on the dnode. * * if you only want to find an already existing spill db, then * dmu_spill_hold_existing() should be used. */ int dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp) { dmu_buf_impl_t *db = NULL; int err; if ((flags & DB_RF_HAVESTRUCT) == 0) rw_enter(&dn->dn_struct_rwlock, RW_READER); db = dbuf_hold(dn, DMU_SPILL_BLKID, tag); if ((flags & DB_RF_HAVESTRUCT) == 0) rw_exit(&dn->dn_struct_rwlock); ASSERT(db != NULL); err = dbuf_read(db, NULL, flags); if (err == 0) *dbp = &db->db; else dbuf_rele(db, tag); return (err); } int dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus; dnode_t *dn; int err; DB_DNODE_ENTER(db); dn = DB_DNODE(db); if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) { err = SET_ERROR(EINVAL); } else { rw_enter(&dn->dn_struct_rwlock, RW_READER); if (!dn->dn_have_spill) { err = SET_ERROR(ENOENT); } else { err = dmu_spill_hold_by_dnode(dn, DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp); } rw_exit(&dn->dn_struct_rwlock); } DB_DNODE_EXIT(db); return (err); } int dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus; dnode_t *dn; int err; DB_DNODE_ENTER(db); dn = DB_DNODE(db); err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp); DB_DNODE_EXIT(db); return (err); } /* * Note: longer-term, we should modify all of the dmu_buf_*() interfaces * to take a held dnode rather than -- the lookup is wasteful, * and can induce severe lock contention when writing to several files * whose dnodes are in the same block. */ int dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length, boolean_t read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags) { dmu_buf_t **dbp; uint64_t blkid, nblks, i; uint32_t dbuf_flags; int err; zio_t *zio; ASSERT(length <= DMU_MAX_ACCESS); /* * Note: We directly notify the prefetch code of this read, so that * we can tell it about the multi-block read. dbuf_read() only knows * about the one block it is accessing. */ dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH; rw_enter(&dn->dn_struct_rwlock, RW_READER); if (dn->dn_datablkshift) { int blkshift = dn->dn_datablkshift; nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) - P2ALIGN(offset, 1ULL << blkshift)) >> blkshift; } else { if (offset + length > dn->dn_datablksz) { zfs_panic_recover("zfs: accessing past end of object " "%llx/%llx (size=%u access=%llu+%llu)", (longlong_t)dn->dn_objset-> os_dsl_dataset->ds_object, (longlong_t)dn->dn_object, dn->dn_datablksz, (longlong_t)offset, (longlong_t)length); rw_exit(&dn->dn_struct_rwlock); return (SET_ERROR(EIO)); } nblks = 1; } dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP); #if defined(_KERNEL) && defined(RACCT) if (racct_enable && !read) { PROC_LOCK(curproc); racct_add_force(curproc, RACCT_WRITEBPS, length); racct_add_force(curproc, RACCT_WRITEIOPS, nblks); PROC_UNLOCK(curproc); } #endif zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL); blkid = dbuf_whichblock(dn, 0, offset); for (i = 0; i < nblks; i++) { dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag); if (db == NULL) { rw_exit(&dn->dn_struct_rwlock); dmu_buf_rele_array(dbp, nblks, tag); zio_nowait(zio); return (SET_ERROR(EIO)); } /* initiate async i/o */ if (read) (void) dbuf_read(db, zio, dbuf_flags); #ifdef _KERNEL else curthread->td_ru.ru_oublock++; #endif dbp[i] = &db->db; } if ((flags & DMU_READ_NO_PREFETCH) == 0 && DNODE_META_IS_CACHEABLE(dn) && length <= zfetch_array_rd_sz) { dmu_zfetch(&dn->dn_zfetch, blkid, nblks, read && DNODE_IS_CACHEABLE(dn)); } rw_exit(&dn->dn_struct_rwlock); /* wait for async i/o */ err = zio_wait(zio); if (err) { dmu_buf_rele_array(dbp, nblks, tag); return (err); } /* wait for other io to complete */ if (read) { for (i = 0; i < nblks; i++) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i]; mutex_enter(&db->db_mtx); while (db->db_state == DB_READ || db->db_state == DB_FILL) cv_wait(&db->db_changed, &db->db_mtx); if (db->db_state == DB_UNCACHED) err = SET_ERROR(EIO); mutex_exit(&db->db_mtx); if (err) { dmu_buf_rele_array(dbp, nblks, tag); return (err); } } } *numbufsp = nblks; *dbpp = dbp; return (0); } static int dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset, uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp) { dnode_t *dn; int err; err = dnode_hold(os, object, FTAG, &dn); if (err) return (err); err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag, numbufsp, dbpp, DMU_READ_PREFETCH); dnode_rele(dn, FTAG); return (err); } int dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset, uint64_t length, boolean_t read, void *tag, int *numbufsp, dmu_buf_t ***dbpp) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; dnode_t *dn; int err; DB_DNODE_ENTER(db); dn = DB_DNODE(db); err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag, numbufsp, dbpp, DMU_READ_PREFETCH); DB_DNODE_EXIT(db); return (err); } void dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag) { int i; dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake; if (numbufs == 0) return; for (i = 0; i < numbufs; i++) { if (dbp[i]) dbuf_rele(dbp[i], tag); } kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs); } /* * Issue prefetch i/os for the given blocks. If level is greater than 0, the * indirect blocks prefeteched will be those that point to the blocks containing * the data starting at offset, and continuing to offset + len. * * Note that if the indirect blocks above the blocks being prefetched are not in * cache, they will be asychronously read in. */ void dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset, uint64_t len, zio_priority_t pri) { dnode_t *dn; uint64_t blkid; int nblks, err; if (len == 0) { /* they're interested in the bonus buffer */ dn = DMU_META_DNODE(os); if (object == 0 || object >= DN_MAX_OBJECT) return; rw_enter(&dn->dn_struct_rwlock, RW_READER); blkid = dbuf_whichblock(dn, level, object * sizeof (dnode_phys_t)); dbuf_prefetch(dn, level, blkid, pri, 0); rw_exit(&dn->dn_struct_rwlock); return; } /* * XXX - Note, if the dnode for the requested object is not * already cached, we will do a *synchronous* read in the * dnode_hold() call. The same is true for any indirects. */ err = dnode_hold(os, object, FTAG, &dn); if (err != 0) return; rw_enter(&dn->dn_struct_rwlock, RW_READER); /* * offset + len - 1 is the last byte we want to prefetch for, and offset * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the * last block we want to prefetch, and dbuf_whichblock(dn, level, * offset) is the first. Then the number we need to prefetch is the * last - first + 1. */ if (level > 0 || dn->dn_datablkshift != 0) { nblks = dbuf_whichblock(dn, level, offset + len - 1) - dbuf_whichblock(dn, level, offset) + 1; } else { nblks = (offset < dn->dn_datablksz); } if (nblks != 0) { blkid = dbuf_whichblock(dn, level, offset); for (int i = 0; i < nblks; i++) dbuf_prefetch(dn, level, blkid + i, pri, 0); } rw_exit(&dn->dn_struct_rwlock); dnode_rele(dn, FTAG); } /* * Get the next "chunk" of file data to free. We traverse the file from * the end so that the file gets shorter over time (if we crashes in the * middle, this will leave us in a better state). We find allocated file * data by simply searching the allocated level 1 indirects. * * On input, *start should be the first offset that does not need to be * freed (e.g. "offset + length"). On return, *start will be the first * offset that should be freed and l1blks is set to the number of level 1 * indirect blocks found within the chunk. */ static int get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum, uint64_t *l1blks) { uint64_t blks; uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1); /* bytes of data covered by a level-1 indirect block */ uint64_t iblkrange = dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT); ASSERT3U(minimum, <=, *start); /* * Check if we can free the entire range assuming that all of the * L1 blocks in this range have data. If we can, we use this * worst case value as an estimate so we can avoid having to look * at the object's actual data. */ uint64_t total_l1blks = (roundup(*start, iblkrange) - (minimum / iblkrange * iblkrange)) / iblkrange; if (total_l1blks <= maxblks) { *l1blks = total_l1blks; *start = minimum; return (0); } ASSERT(ISP2(iblkrange)); for (blks = 0; *start > minimum && blks < maxblks; blks++) { int err; /* * dnode_next_offset(BACKWARDS) will find an allocated L1 * indirect block at or before the input offset. We must * decrement *start so that it is at the end of the region * to search. */ (*start)--; err = dnode_next_offset(dn, DNODE_FIND_BACKWARDS, start, 2, 1, 0); /* if there are no indirect blocks before start, we are done */ if (err == ESRCH) { *start = minimum; break; } else if (err != 0) { *l1blks = blks; return (err); } /* set start to the beginning of this L1 indirect */ *start = P2ALIGN(*start, iblkrange); } if (*start < minimum) *start = minimum; *l1blks = blks; return (0); } /* * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set, * otherwise return false. * Used below in dmu_free_long_range_impl() to enable abort when unmounting */ /*ARGSUSED*/ static boolean_t dmu_objset_zfs_unmounting(objset_t *os) { #ifdef _KERNEL if (dmu_objset_type(os) == DMU_OST_ZFS) return (zfs_get_vfs_flag_unmounted(os)); #endif return (B_FALSE); } static int dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset, uint64_t length) { uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz; int err; uint64_t dirty_frees_threshold; dsl_pool_t *dp = dmu_objset_pool(os); if (offset >= object_size) return (0); if (zfs_per_txg_dirty_frees_percent <= 100) dirty_frees_threshold = zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100; else dirty_frees_threshold = zfs_dirty_data_max / 20; if (length == DMU_OBJECT_END || offset + length > object_size) length = object_size - offset; while (length != 0) { uint64_t chunk_end, chunk_begin, chunk_len; uint64_t l1blks; dmu_tx_t *tx; if (dmu_objset_zfs_unmounting(dn->dn_objset)) return (SET_ERROR(EINTR)); chunk_end = chunk_begin = offset + length; /* move chunk_begin backwards to the beginning of this chunk */ err = get_next_chunk(dn, &chunk_begin, offset, &l1blks); if (err) return (err); ASSERT3U(chunk_begin, >=, offset); ASSERT3U(chunk_begin, <=, chunk_end); chunk_len = chunk_end - chunk_begin; tx = dmu_tx_create(os); dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len); /* * Mark this transaction as typically resulting in a net * reduction in space used. */ dmu_tx_mark_netfree(tx); err = dmu_tx_assign(tx, TXG_WAIT); if (err) { dmu_tx_abort(tx); return (err); } uint64_t txg = dmu_tx_get_txg(tx); mutex_enter(&dp->dp_lock); uint64_t long_free_dirty = dp->dp_long_free_dirty_pertxg[txg & TXG_MASK]; mutex_exit(&dp->dp_lock); /* * To avoid filling up a TXG with just frees, wait for * the next TXG to open before freeing more chunks if * we have reached the threshold of frees. */ if (dirty_frees_threshold != 0 && long_free_dirty >= dirty_frees_threshold) { dmu_tx_commit(tx); txg_wait_open(dp, 0); continue; } /* * In order to prevent unnecessary write throttling, for each * TXG, we track the cumulative size of L1 blocks being dirtied * in dnode_free_range() below. We compare this number to a * tunable threshold, past which we prevent new L1 dirty freeing * blocks from being added into the open TXG. See * dmu_free_long_range_impl() for details. The threshold * prevents write throttle activation due to dirty freeing L1 * blocks taking up a large percentage of zfs_dirty_data_max. */ mutex_enter(&dp->dp_lock); dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] += l1blks << dn->dn_indblkshift; mutex_exit(&dp->dp_lock); DTRACE_PROBE3(free__long__range, uint64_t, long_free_dirty, uint64_t, chunk_len, uint64_t, txg); dnode_free_range(dn, chunk_begin, chunk_len, tx); dmu_tx_commit(tx); length -= chunk_len; } return (0); } int dmu_free_long_range(objset_t *os, uint64_t object, uint64_t offset, uint64_t length) { dnode_t *dn; int err; err = dnode_hold(os, object, FTAG, &dn); if (err != 0) return (err); err = dmu_free_long_range_impl(os, dn, offset, length); /* * It is important to zero out the maxblkid when freeing the entire * file, so that (a) subsequent calls to dmu_free_long_range_impl() * will take the fast path, and (b) dnode_reallocate() can verify * that the entire file has been freed. */ if (err == 0 && offset == 0 && length == DMU_OBJECT_END) dn->dn_maxblkid = 0; dnode_rele(dn, FTAG); return (err); } int dmu_free_long_object(objset_t *os, uint64_t object) { dmu_tx_t *tx; int err; err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END); if (err != 0) return (err); tx = dmu_tx_create(os); dmu_tx_hold_bonus(tx, object); dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END); dmu_tx_mark_netfree(tx); err = dmu_tx_assign(tx, TXG_WAIT); if (err == 0) { err = dmu_object_free(os, object, tx); dmu_tx_commit(tx); } else { dmu_tx_abort(tx); } return (err); } int dmu_free_range(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, dmu_tx_t *tx) { dnode_t *dn; int err = dnode_hold(os, object, FTAG, &dn); if (err) return (err); ASSERT(offset < UINT64_MAX); ASSERT(size == -1ULL || size <= UINT64_MAX - offset); dnode_free_range(dn, offset, size, tx); dnode_rele(dn, FTAG); return (0); } static int dmu_read_impl(dnode_t *dn, uint64_t offset, uint64_t size, void *buf, uint32_t flags) { dmu_buf_t **dbp; int numbufs, err = 0; /* * Deal with odd block sizes, where there can't be data past the first * block. If we ever do the tail block optimization, we will need to * handle that here as well. */ if (dn->dn_maxblkid == 0) { int newsz = offset > dn->dn_datablksz ? 0 : MIN(size, dn->dn_datablksz - offset); bzero((char *)buf + newsz, size - newsz); size = newsz; } while (size > 0) { uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2); int i; /* * NB: we could do this block-at-a-time, but it's nice * to be reading in parallel. */ err = dmu_buf_hold_array_by_dnode(dn, offset, mylen, TRUE, FTAG, &numbufs, &dbp, flags); if (err) break; for (i = 0; i < numbufs; i++) { int tocpy; int bufoff; dmu_buf_t *db = dbp[i]; ASSERT(size > 0); bufoff = offset - db->db_offset; tocpy = (int)MIN(db->db_size - bufoff, size); bcopy((char *)db->db_data + bufoff, buf, tocpy); offset += tocpy; size -= tocpy; buf = (char *)buf + tocpy; } dmu_buf_rele_array(dbp, numbufs, FTAG); } return (err); } int dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, void *buf, uint32_t flags) { dnode_t *dn; int err; err = dnode_hold(os, object, FTAG, &dn); if (err != 0) return (err); err = dmu_read_impl(dn, offset, size, buf, flags); dnode_rele(dn, FTAG); return (err); } int dmu_read_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, void *buf, uint32_t flags) { return (dmu_read_impl(dn, offset, size, buf, flags)); } static void dmu_write_impl(dmu_buf_t **dbp, int numbufs, uint64_t offset, uint64_t size, const void *buf, dmu_tx_t *tx) { int i; for (i = 0; i < numbufs; i++) { int tocpy; int bufoff; dmu_buf_t *db = dbp[i]; ASSERT(size > 0); bufoff = offset - db->db_offset; tocpy = (int)MIN(db->db_size - bufoff, size); ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); if (tocpy == db->db_size) dmu_buf_will_fill(db, tx); else dmu_buf_will_dirty(db, tx); bcopy(buf, (char *)db->db_data + bufoff, tocpy); if (tocpy == db->db_size) dmu_buf_fill_done(db, tx); offset += tocpy; size -= tocpy; buf = (char *)buf + tocpy; } } void dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, const void *buf, dmu_tx_t *tx) { dmu_buf_t **dbp; int numbufs; if (size == 0) return; VERIFY0(dmu_buf_hold_array(os, object, offset, size, FALSE, FTAG, &numbufs, &dbp)); dmu_write_impl(dbp, numbufs, offset, size, buf, tx); dmu_buf_rele_array(dbp, numbufs, FTAG); } void dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, const void *buf, dmu_tx_t *tx) { dmu_buf_t **dbp; int numbufs; if (size == 0) return; VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size, FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH)); dmu_write_impl(dbp, numbufs, offset, size, buf, tx); dmu_buf_rele_array(dbp, numbufs, FTAG); } static int dmu_object_remap_one_indirect(objset_t *os, dnode_t *dn, uint64_t last_removal_txg, uint64_t offset) { uint64_t l1blkid = dbuf_whichblock(dn, 1, offset); int err = 0; rw_enter(&dn->dn_struct_rwlock, RW_READER); dmu_buf_impl_t *dbuf = dbuf_hold_level(dn, 1, l1blkid, FTAG); ASSERT3P(dbuf, !=, NULL); /* * If the block hasn't been written yet, this default will ensure * we don't try to remap it. */ uint64_t birth = UINT64_MAX; ASSERT3U(last_removal_txg, !=, UINT64_MAX); if (dbuf->db_blkptr != NULL) birth = dbuf->db_blkptr->blk_birth; rw_exit(&dn->dn_struct_rwlock); /* * If this L1 was already written after the last removal, then we've * already tried to remap it. */ if (birth <= last_removal_txg && dbuf_read(dbuf, NULL, DB_RF_MUST_SUCCEED) == 0 && dbuf_can_remap(dbuf)) { dmu_tx_t *tx = dmu_tx_create(os); dmu_tx_hold_remap_l1indirect(tx, dn->dn_object); err = dmu_tx_assign(tx, TXG_WAIT); if (err == 0) { (void) dbuf_dirty(dbuf, tx); dmu_tx_commit(tx); } else { dmu_tx_abort(tx); } } dbuf_rele(dbuf, FTAG); delay(zfs_object_remap_one_indirect_delay_ticks); return (err); } /* * Remap all blockpointers in the object, if possible, so that they reference * only concrete vdevs. * * To do this, iterate over the L0 blockpointers and remap any that reference * an indirect vdev. Note that we only examine L0 blockpointers; since we * cannot guarantee that we can remap all blockpointer anyways (due to split * blocks), we do not want to make the code unnecessarily complicated to * catch the unlikely case that there is an L1 block on an indirect vdev that * contains no indirect blockpointers. */ int dmu_object_remap_indirects(objset_t *os, uint64_t object, uint64_t last_removal_txg) { uint64_t offset, l1span; int err; dnode_t *dn; err = dnode_hold(os, object, FTAG, &dn); if (err != 0) { return (err); } if (dn->dn_nlevels <= 1) { if (issig(JUSTLOOKING) && issig(FORREAL)) { err = SET_ERROR(EINTR); } /* * If the dnode has no indirect blocks, we cannot dirty them. * We still want to remap the blkptr(s) in the dnode if * appropriate, so mark it as dirty. */ if (err == 0 && dnode_needs_remap(dn)) { dmu_tx_t *tx = dmu_tx_create(os); dmu_tx_hold_bonus(tx, dn->dn_object); if ((err = dmu_tx_assign(tx, TXG_WAIT)) == 0) { dnode_setdirty(dn, tx); dmu_tx_commit(tx); } else { dmu_tx_abort(tx); } } dnode_rele(dn, FTAG); return (err); } offset = 0; l1span = 1ULL << (dn->dn_indblkshift - SPA_BLKPTRSHIFT + dn->dn_datablkshift); /* * Find the next L1 indirect that is not a hole. */ while (dnode_next_offset(dn, 0, &offset, 2, 1, 0) == 0) { if (issig(JUSTLOOKING) && issig(FORREAL)) { err = SET_ERROR(EINTR); break; } if ((err = dmu_object_remap_one_indirect(os, dn, last_removal_txg, offset)) != 0) { break; } offset += l1span; } dnode_rele(dn, FTAG); return (err); } void dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, dmu_tx_t *tx) { dmu_buf_t **dbp; int numbufs, i; if (size == 0) return; VERIFY(0 == dmu_buf_hold_array(os, object, offset, size, FALSE, FTAG, &numbufs, &dbp)); for (i = 0; i < numbufs; i++) { dmu_buf_t *db = dbp[i]; dmu_buf_will_not_fill(db, tx); } dmu_buf_rele_array(dbp, numbufs, FTAG); } void dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset, void *data, uint8_t etype, uint8_t comp, int uncompressed_size, int compressed_size, int byteorder, dmu_tx_t *tx) { dmu_buf_t *db; ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES); ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS); VERIFY0(dmu_buf_hold_noread(os, object, offset, FTAG, &db)); dmu_buf_write_embedded(db, data, (bp_embedded_type_t)etype, (enum zio_compress)comp, uncompressed_size, compressed_size, byteorder, tx); dmu_buf_rele(db, FTAG); } /* * DMU support for xuio */ kstat_t *xuio_ksp = NULL; int dmu_xuio_init(xuio_t *xuio, int nblk) { dmu_xuio_t *priv; uio_t *uio = &xuio->xu_uio; uio->uio_iovcnt = nblk; uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP); priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP); priv->cnt = nblk; priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP); priv->iovp = uio->uio_iov; XUIO_XUZC_PRIV(xuio) = priv; if (XUIO_XUZC_RW(xuio) == UIO_READ) XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk); else XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk); return (0); } void dmu_xuio_fini(xuio_t *xuio) { dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); int nblk = priv->cnt; kmem_free(priv->iovp, nblk * sizeof (iovec_t)); kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *)); kmem_free(priv, sizeof (dmu_xuio_t)); if (XUIO_XUZC_RW(xuio) == UIO_READ) XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk); else XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk); } /* * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf } * and increase priv->next by 1. */ int dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n) { struct iovec *iov; uio_t *uio = &xuio->xu_uio; dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); int i = priv->next++; ASSERT(i < priv->cnt); ASSERT(off + n <= arc_buf_lsize(abuf)); iov = uio->uio_iov + i; iov->iov_base = (char *)abuf->b_data + off; iov->iov_len = n; priv->bufs[i] = abuf; return (0); } int dmu_xuio_cnt(xuio_t *xuio) { dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); return (priv->cnt); } arc_buf_t * dmu_xuio_arcbuf(xuio_t *xuio, int i) { dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); ASSERT(i < priv->cnt); return (priv->bufs[i]); } void dmu_xuio_clear(xuio_t *xuio, int i) { dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); ASSERT(i < priv->cnt); priv->bufs[i] = NULL; } static void xuio_stat_init(void) { xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc", KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); if (xuio_ksp != NULL) { xuio_ksp->ks_data = &xuio_stats; kstat_install(xuio_ksp); } } static void xuio_stat_fini(void) { if (xuio_ksp != NULL) { kstat_delete(xuio_ksp); xuio_ksp = NULL; } } void xuio_stat_wbuf_copied(void) { XUIOSTAT_BUMP(xuiostat_wbuf_copied); } void xuio_stat_wbuf_nocopy(void) { XUIOSTAT_BUMP(xuiostat_wbuf_nocopy); } #ifdef _KERNEL int dmu_read_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size) { dmu_buf_t **dbp; int numbufs, i, err; xuio_t *xuio = NULL; /* * NB: we could do this block-at-a-time, but it's nice * to be reading in parallel. */ err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size, TRUE, FTAG, &numbufs, &dbp, 0); if (err) return (err); #ifdef UIO_XUIO if (uio->uio_extflg == UIO_XUIO) xuio = (xuio_t *)uio; #endif for (i = 0; i < numbufs; i++) { int tocpy; int bufoff; dmu_buf_t *db = dbp[i]; ASSERT(size > 0); bufoff = uio->uio_loffset - db->db_offset; tocpy = (int)MIN(db->db_size - bufoff, size); if (xuio) { dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; arc_buf_t *dbuf_abuf = dbi->db_buf; arc_buf_t *abuf = dbuf_loan_arcbuf(dbi); err = dmu_xuio_add(xuio, abuf, bufoff, tocpy); if (!err) { uio->uio_resid -= tocpy; uio->uio_loffset += tocpy; } if (abuf == dbuf_abuf) XUIOSTAT_BUMP(xuiostat_rbuf_nocopy); else XUIOSTAT_BUMP(xuiostat_rbuf_copied); } else { #ifdef illumos err = uiomove((char *)db->db_data + bufoff, tocpy, UIO_READ, uio); #else err = vn_io_fault_uiomove((char *)db->db_data + bufoff, tocpy, uio); #endif } if (err) break; size -= tocpy; } dmu_buf_rele_array(dbp, numbufs, FTAG); return (err); } /* * Read 'size' bytes into the uio buffer. * From object zdb->db_object. * Starting at offset uio->uio_loffset. * * If the caller already has a dbuf in the target object * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(), * because we don't have to find the dnode_t for the object. */ int dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; dnode_t *dn; int err; if (size == 0) return (0); DB_DNODE_ENTER(db); dn = DB_DNODE(db); err = dmu_read_uio_dnode(dn, uio, size); DB_DNODE_EXIT(db); return (err); } /* * Read 'size' bytes into the uio buffer. * From the specified object * Starting at offset uio->uio_loffset. */ int dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size) { dnode_t *dn; int err; if (size == 0) return (0); err = dnode_hold(os, object, FTAG, &dn); if (err) return (err); err = dmu_read_uio_dnode(dn, uio, size); dnode_rele(dn, FTAG); return (err); } int dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx) { dmu_buf_t **dbp; int numbufs; int err = 0; int i; err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size, FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH); if (err) return (err); for (i = 0; i < numbufs; i++) { int tocpy; int bufoff; dmu_buf_t *db = dbp[i]; ASSERT(size > 0); bufoff = uio->uio_loffset - db->db_offset; tocpy = (int)MIN(db->db_size - bufoff, size); ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); if (tocpy == db->db_size) dmu_buf_will_fill(db, tx); else dmu_buf_will_dirty(db, tx); #ifdef illumos /* * XXX uiomove could block forever (eg. nfs-backed * pages). There needs to be a uiolockdown() function * to lock the pages in memory, so that uiomove won't * block. */ err = uiomove((char *)db->db_data + bufoff, tocpy, UIO_WRITE, uio); #else err = vn_io_fault_uiomove((char *)db->db_data + bufoff, tocpy, uio); #endif if (tocpy == db->db_size) dmu_buf_fill_done(db, tx); if (err) break; size -= tocpy; } dmu_buf_rele_array(dbp, numbufs, FTAG); return (err); } /* * Write 'size' bytes from the uio buffer. * To object zdb->db_object. * Starting at offset uio->uio_loffset. * * If the caller already has a dbuf in the target object * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(), * because we don't have to find the dnode_t for the object. */ int dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size, dmu_tx_t *tx) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; dnode_t *dn; int err; if (size == 0) return (0); DB_DNODE_ENTER(db); dn = DB_DNODE(db); err = dmu_write_uio_dnode(dn, uio, size, tx); DB_DNODE_EXIT(db); return (err); } /* * Write 'size' bytes from the uio buffer. * To the specified object. * Starting at offset uio->uio_loffset. */ int dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size, dmu_tx_t *tx) { dnode_t *dn; int err; if (size == 0) return (0); err = dnode_hold(os, object, FTAG, &dn); if (err) return (err); err = dmu_write_uio_dnode(dn, uio, size, tx); dnode_rele(dn, FTAG); return (err); } #ifdef illumos int dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, page_t *pp, dmu_tx_t *tx) { dmu_buf_t **dbp; int numbufs, i; int err; if (size == 0) return (0); err = dmu_buf_hold_array(os, object, offset, size, FALSE, FTAG, &numbufs, &dbp); if (err) return (err); for (i = 0; i < numbufs; i++) { int tocpy, copied, thiscpy; int bufoff; dmu_buf_t *db = dbp[i]; caddr_t va; ASSERT(size > 0); ASSERT3U(db->db_size, >=, PAGESIZE); bufoff = offset - db->db_offset; tocpy = (int)MIN(db->db_size - bufoff, size); ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); if (tocpy == db->db_size) dmu_buf_will_fill(db, tx); else dmu_buf_will_dirty(db, tx); for (copied = 0; copied < tocpy; copied += PAGESIZE) { ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff); thiscpy = MIN(PAGESIZE, tocpy - copied); va = zfs_map_page(pp, S_READ); bcopy(va, (char *)db->db_data + bufoff, thiscpy); zfs_unmap_page(pp, va); pp = pp->p_next; bufoff += PAGESIZE; } if (tocpy == db->db_size) dmu_buf_fill_done(db, tx); offset += tocpy; size -= tocpy; } dmu_buf_rele_array(dbp, numbufs, FTAG); return (err); } #else /* !illumos */ int dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, vm_page_t *ma, dmu_tx_t *tx) { dmu_buf_t **dbp; struct sf_buf *sf; int numbufs, i; int err; if (size == 0) return (0); err = dmu_buf_hold_array(os, object, offset, size, FALSE, FTAG, &numbufs, &dbp); if (err) return (err); for (i = 0; i < numbufs; i++) { int tocpy, copied, thiscpy; int bufoff; dmu_buf_t *db = dbp[i]; caddr_t va; ASSERT(size > 0); ASSERT3U(db->db_size, >=, PAGESIZE); bufoff = offset - db->db_offset; tocpy = (int)MIN(db->db_size - bufoff, size); ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); if (tocpy == db->db_size) dmu_buf_will_fill(db, tx); else dmu_buf_will_dirty(db, tx); for (copied = 0; copied < tocpy; copied += PAGESIZE) { ASSERT3U(ptoa((*ma)->pindex), ==, db->db_offset + bufoff); thiscpy = MIN(PAGESIZE, tocpy - copied); va = zfs_map_page(*ma, &sf); bcopy(va, (char *)db->db_data + bufoff, thiscpy); zfs_unmap_page(sf); ma += 1; bufoff += PAGESIZE; } if (tocpy == db->db_size) dmu_buf_fill_done(db, tx); offset += tocpy; size -= tocpy; } dmu_buf_rele_array(dbp, numbufs, FTAG); return (err); } int dmu_read_pages(objset_t *os, uint64_t object, vm_page_t *ma, int count, int *rbehind, int *rahead, int last_size) { struct sf_buf *sf; vm_object_t vmobj; vm_page_t m; dmu_buf_t **dbp; dmu_buf_t *db; caddr_t va; int numbufs, i; int bufoff, pgoff, tocpy; int mi, di; int err; ASSERT3U(ma[0]->pindex + count - 1, ==, ma[count - 1]->pindex); ASSERT(last_size <= PAGE_SIZE); err = dmu_buf_hold_array(os, object, IDX_TO_OFF(ma[0]->pindex), IDX_TO_OFF(count - 1) + last_size, TRUE, FTAG, &numbufs, &dbp); if (err != 0) return (err); #ifdef DEBUG IMPLY(last_size < PAGE_SIZE, *rahead == 0); if (dbp[0]->db_offset != 0 || numbufs > 1) { for (i = 0; i < numbufs; i++) { ASSERT(ISP2(dbp[i]->db_size)); ASSERT((dbp[i]->db_offset % dbp[i]->db_size) == 0); ASSERT3U(dbp[i]->db_size, ==, dbp[0]->db_size); } } #endif vmobj = ma[0]->object; zfs_vmobject_wlock(vmobj); db = dbp[0]; for (i = 0; i < *rbehind; i++) { m = vm_page_grab(vmobj, ma[0]->pindex - 1 - i, - VM_ALLOC_NORMAL | VM_ALLOC_NOWAIT | VM_ALLOC_NOBUSY); + VM_ALLOC_NORMAL | VM_ALLOC_NOWAIT | + VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY); if (m == NULL) break; - if (m->valid != 0) { + if (!vm_page_none_valid(m)) { ASSERT3U(m->valid, ==, VM_PAGE_BITS_ALL); + vm_page_sunbusy(m); break; } ASSERT(m->dirty == 0); ASSERT(!pmap_page_is_mapped(m)); ASSERT(db->db_size > PAGE_SIZE); bufoff = IDX_TO_OFF(m->pindex) % db->db_size; va = zfs_map_page(m, &sf); bcopy((char *)db->db_data + bufoff, va, PAGESIZE); zfs_unmap_page(sf); - m->valid = VM_PAGE_BITS_ALL; + vm_page_valid(m); vm_page_lock(m); if ((m->busy_lock & VPB_BIT_WAITERS) != 0) vm_page_activate(m); else vm_page_deactivate(m); vm_page_unlock(m); + vm_page_sunbusy(m); } *rbehind = i; bufoff = IDX_TO_OFF(ma[0]->pindex) % db->db_size; pgoff = 0; for (mi = 0, di = 0; mi < count && di < numbufs; ) { if (pgoff == 0) { m = ma[mi]; if (m != bogus_page) { vm_page_assert_xbusied(m); - ASSERT(m->valid == 0); + ASSERT(vm_page_none_valid(m)); ASSERT(m->dirty == 0); ASSERT(!pmap_page_is_mapped(m)); va = zfs_map_page(m, &sf); } } if (bufoff == 0) db = dbp[di]; if (m != bogus_page) { ASSERT3U(IDX_TO_OFF(m->pindex) + pgoff, ==, db->db_offset + bufoff); } /* * We do not need to clamp the copy size by the file * size as the last block is zero-filled beyond the * end of file anyway. */ tocpy = MIN(db->db_size - bufoff, PAGESIZE - pgoff); if (m != bogus_page) bcopy((char *)db->db_data + bufoff, va + pgoff, tocpy); pgoff += tocpy; ASSERT(pgoff <= PAGESIZE); if (pgoff == PAGESIZE) { if (m != bogus_page) { zfs_unmap_page(sf); - m->valid = VM_PAGE_BITS_ALL; + vm_page_valid(m); } ASSERT(mi < count); mi++; pgoff = 0; } bufoff += tocpy; ASSERT(bufoff <= db->db_size); if (bufoff == db->db_size) { ASSERT(di < numbufs); di++; bufoff = 0; } } #ifdef DEBUG /* * Three possibilities: * - last requested page ends at a buffer boundary and , thus, * all pages and buffers have been iterated; * - all requested pages are filled, but the last buffer * has not been exhausted; * the read-ahead is possible only in this case; * - all buffers have been read, but the last page has not been * fully filled; * this is only possible if the file has only a single buffer * with a size that is not a multiple of the page size. */ if (mi == count) { ASSERT(di >= numbufs - 1); IMPLY(*rahead != 0, di == numbufs - 1); IMPLY(*rahead != 0, bufoff != 0); ASSERT(pgoff == 0); } if (di == numbufs) { ASSERT(mi >= count - 1); ASSERT(*rahead == 0); IMPLY(pgoff == 0, mi == count); if (pgoff != 0) { ASSERT(mi == count - 1); ASSERT((dbp[0]->db_size & PAGE_MASK) != 0); } } #endif if (pgoff != 0) { ASSERT(m != bogus_page); bzero(va + pgoff, PAGESIZE - pgoff); zfs_unmap_page(sf); - m->valid = VM_PAGE_BITS_ALL; + vm_page_valid(m); } for (i = 0; i < *rahead; i++) { m = vm_page_grab(vmobj, ma[count - 1]->pindex + 1 + i, - VM_ALLOC_NORMAL | VM_ALLOC_NOWAIT | VM_ALLOC_NOBUSY); + VM_ALLOC_NORMAL | VM_ALLOC_NOWAIT | + VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY); if (m == NULL) break; - if (m->valid != 0) { + if (!vm_page_none_valid(m)) { ASSERT3U(m->valid, ==, VM_PAGE_BITS_ALL); + vm_page_sunbusy(m); break; } ASSERT(m->dirty == 0); ASSERT(!pmap_page_is_mapped(m)); ASSERT(db->db_size > PAGE_SIZE); bufoff = IDX_TO_OFF(m->pindex) % db->db_size; tocpy = MIN(db->db_size - bufoff, PAGESIZE); va = zfs_map_page(m, &sf); bcopy((char *)db->db_data + bufoff, va, tocpy); if (tocpy < PAGESIZE) { ASSERT(i == *rahead - 1); ASSERT((db->db_size & PAGE_MASK) != 0); bzero(va + tocpy, PAGESIZE - tocpy); } zfs_unmap_page(sf); - m->valid = VM_PAGE_BITS_ALL; + vm_page_valid(m); vm_page_lock(m); if ((m->busy_lock & VPB_BIT_WAITERS) != 0) vm_page_activate(m); else vm_page_deactivate(m); vm_page_unlock(m); + vm_page_sunbusy(m); } *rahead = i; zfs_vmobject_wunlock(vmobj); dmu_buf_rele_array(dbp, numbufs, FTAG); return (0); } #endif /* illumos */ #endif /* _KERNEL */ /* * Allocate a loaned anonymous arc buffer. */ arc_buf_t * dmu_request_arcbuf(dmu_buf_t *handle, int size) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle; return (arc_loan_buf(db->db_objset->os_spa, B_FALSE, size)); } /* * Free a loaned arc buffer. */ void dmu_return_arcbuf(arc_buf_t *buf) { arc_return_buf(buf, FTAG); arc_buf_destroy(buf, FTAG); } /* * When possible directly assign passed loaned arc buffer to a dbuf. * If this is not possible copy the contents of passed arc buf via * dmu_write(). */ void dmu_assign_arcbuf_dnode(dnode_t *dn, uint64_t offset, arc_buf_t *buf, dmu_tx_t *tx) { dmu_buf_impl_t *db; uint32_t blksz = (uint32_t)arc_buf_lsize(buf); uint64_t blkid; rw_enter(&dn->dn_struct_rwlock, RW_READER); blkid = dbuf_whichblock(dn, 0, offset); VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL); rw_exit(&dn->dn_struct_rwlock); /* * We can only assign if the offset is aligned, the arc buf is the * same size as the dbuf, and the dbuf is not metadata. */ if (offset == db->db.db_offset && blksz == db->db.db_size) { #ifdef _KERNEL curthread->td_ru.ru_oublock++; #ifdef RACCT if (racct_enable) { PROC_LOCK(curproc); racct_add_force(curproc, RACCT_WRITEBPS, blksz); racct_add_force(curproc, RACCT_WRITEIOPS, 1); PROC_UNLOCK(curproc); } #endif /* RACCT */ #endif /* _KERNEL */ dbuf_assign_arcbuf(db, buf, tx); dbuf_rele(db, FTAG); } else { objset_t *os; uint64_t object; /* compressed bufs must always be assignable to their dbuf */ ASSERT3U(arc_get_compression(buf), ==, ZIO_COMPRESS_OFF); ASSERT(!(buf->b_flags & ARC_BUF_FLAG_COMPRESSED)); os = dn->dn_objset; object = dn->dn_object; dbuf_rele(db, FTAG); dmu_write(os, object, offset, blksz, buf->b_data, tx); dmu_return_arcbuf(buf); XUIOSTAT_BUMP(xuiostat_wbuf_copied); } } void dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf, dmu_tx_t *tx) { dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle; DB_DNODE_ENTER(dbuf); dmu_assign_arcbuf_dnode(DB_DNODE(dbuf), offset, buf, tx); DB_DNODE_EXIT(dbuf); } typedef struct { dbuf_dirty_record_t *dsa_dr; dmu_sync_cb_t *dsa_done; zgd_t *dsa_zgd; dmu_tx_t *dsa_tx; } dmu_sync_arg_t; /* ARGSUSED */ static void dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg) { dmu_sync_arg_t *dsa = varg; dmu_buf_t *db = dsa->dsa_zgd->zgd_db; blkptr_t *bp = zio->io_bp; if (zio->io_error == 0) { if (BP_IS_HOLE(bp)) { /* * A block of zeros may compress to a hole, but the * block size still needs to be known for replay. */ BP_SET_LSIZE(bp, db->db_size); } else if (!BP_IS_EMBEDDED(bp)) { ASSERT(BP_GET_LEVEL(bp) == 0); bp->blk_fill = 1; } } } static void dmu_sync_late_arrival_ready(zio_t *zio) { dmu_sync_ready(zio, NULL, zio->io_private); } /* ARGSUSED */ static void dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg) { dmu_sync_arg_t *dsa = varg; dbuf_dirty_record_t *dr = dsa->dsa_dr; dmu_buf_impl_t *db = dr->dr_dbuf; zgd_t *zgd = dsa->dsa_zgd; /* * Record the vdev(s) backing this blkptr so they can be flushed after * the writes for the lwb have completed. */ if (zio->io_error == 0) { zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp); } mutex_enter(&db->db_mtx); ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC); if (zio->io_error == 0) { dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE); if (dr->dt.dl.dr_nopwrite) { blkptr_t *bp = zio->io_bp; blkptr_t *bp_orig = &zio->io_bp_orig; uint8_t chksum = BP_GET_CHECKSUM(bp_orig); ASSERT(BP_EQUAL(bp, bp_orig)); VERIFY(BP_EQUAL(bp, db->db_blkptr)); ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF); ASSERT(zio_checksum_table[chksum].ci_flags & ZCHECKSUM_FLAG_NOPWRITE); } dr->dt.dl.dr_overridden_by = *zio->io_bp; dr->dt.dl.dr_override_state = DR_OVERRIDDEN; dr->dt.dl.dr_copies = zio->io_prop.zp_copies; /* * Old style holes are filled with all zeros, whereas * new-style holes maintain their lsize, type, level, * and birth time (see zio_write_compress). While we * need to reset the BP_SET_LSIZE() call that happened * in dmu_sync_ready for old style holes, we do *not* * want to wipe out the information contained in new * style holes. Thus, only zero out the block pointer if * it's an old style hole. */ if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) && dr->dt.dl.dr_overridden_by.blk_birth == 0) BP_ZERO(&dr->dt.dl.dr_overridden_by); } else { dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; } cv_broadcast(&db->db_changed); mutex_exit(&db->db_mtx); dsa->dsa_done(dsa->dsa_zgd, zio->io_error); kmem_free(dsa, sizeof (*dsa)); } static void dmu_sync_late_arrival_done(zio_t *zio) { blkptr_t *bp = zio->io_bp; dmu_sync_arg_t *dsa = zio->io_private; blkptr_t *bp_orig = &zio->io_bp_orig; zgd_t *zgd = dsa->dsa_zgd; if (zio->io_error == 0) { /* * Record the vdev(s) backing this blkptr so they can be * flushed after the writes for the lwb have completed. */ zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp); if (!BP_IS_HOLE(bp)) { ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE)); ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig)); ASSERT(zio->io_bp->blk_birth == zio->io_txg); ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa)); zio_free(zio->io_spa, zio->io_txg, zio->io_bp); } } dmu_tx_commit(dsa->dsa_tx); dsa->dsa_done(dsa->dsa_zgd, zio->io_error); abd_put(zio->io_abd); kmem_free(dsa, sizeof (*dsa)); } static int dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd, zio_prop_t *zp, zbookmark_phys_t *zb) { dmu_sync_arg_t *dsa; dmu_tx_t *tx; tx = dmu_tx_create(os); dmu_tx_hold_space(tx, zgd->zgd_db->db_size); if (dmu_tx_assign(tx, TXG_WAIT) != 0) { dmu_tx_abort(tx); /* Make zl_get_data do txg_waited_synced() */ return (SET_ERROR(EIO)); } /* * In order to prevent the zgd's lwb from being free'd prior to * dmu_sync_late_arrival_done() being called, we have to ensure * the lwb's "max txg" takes this tx's txg into account. */ zil_lwb_add_txg(zgd->zgd_lwb, dmu_tx_get_txg(tx)); dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); dsa->dsa_dr = NULL; dsa->dsa_done = done; dsa->dsa_zgd = zgd; dsa->dsa_tx = tx; /* * Since we are currently syncing this txg, it's nontrivial to * determine what BP to nopwrite against, so we disable nopwrite. * * When syncing, the db_blkptr is initially the BP of the previous * txg. We can not nopwrite against it because it will be changed * (this is similar to the non-late-arrival case where the dbuf is * dirty in a future txg). * * Then dbuf_write_ready() sets bp_blkptr to the location we will write. * We can not nopwrite against it because although the BP will not * (typically) be changed, the data has not yet been persisted to this * location. * * Finally, when dbuf_write_done() is called, it is theoretically * possible to always nopwrite, because the data that was written in * this txg is the same data that we are trying to write. However we * would need to check that this dbuf is not dirty in any future * txg's (as we do in the normal dmu_sync() path). For simplicity, we * don't nopwrite in this case. */ zp->zp_nopwrite = B_FALSE; zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp, abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size), zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp, dmu_sync_late_arrival_ready, NULL, NULL, dmu_sync_late_arrival_done, dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb)); return (0); } /* * Intent log support: sync the block associated with db to disk. * N.B. and XXX: the caller is responsible for making sure that the * data isn't changing while dmu_sync() is writing it. * * Return values: * * EEXIST: this txg has already been synced, so there's nothing to do. * The caller should not log the write. * * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do. * The caller should not log the write. * * EALREADY: this block is already in the process of being synced. * The caller should track its progress (somehow). * * EIO: could not do the I/O. * The caller should do a txg_wait_synced(). * * 0: the I/O has been initiated. * The caller should log this blkptr in the done callback. * It is possible that the I/O will fail, in which case * the error will be reported to the done callback and * propagated to pio from zio_done(). */ int dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db; objset_t *os = db->db_objset; dsl_dataset_t *ds = os->os_dsl_dataset; dbuf_dirty_record_t *dr; dmu_sync_arg_t *dsa; zbookmark_phys_t zb; zio_prop_t zp; dnode_t *dn; ASSERT(pio != NULL); ASSERT(txg != 0); SET_BOOKMARK(&zb, ds->ds_object, db->db.db_object, db->db_level, db->db_blkid); DB_DNODE_ENTER(db); dn = DB_DNODE(db); dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp); DB_DNODE_EXIT(db); /* * If we're frozen (running ziltest), we always need to generate a bp. */ if (txg > spa_freeze_txg(os->os_spa)) return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); /* * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf() * and us. If we determine that this txg is not yet syncing, * but it begins to sync a moment later, that's OK because the * sync thread will block in dbuf_sync_leaf() until we drop db_mtx. */ mutex_enter(&db->db_mtx); if (txg <= spa_last_synced_txg(os->os_spa)) { /* * This txg has already synced. There's nothing to do. */ mutex_exit(&db->db_mtx); return (SET_ERROR(EEXIST)); } if (txg <= spa_syncing_txg(os->os_spa)) { /* * This txg is currently syncing, so we can't mess with * the dirty record anymore; just write a new log block. */ mutex_exit(&db->db_mtx); return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); } dr = db->db_last_dirty; while (dr && dr->dr_txg != txg) dr = dr->dr_next; if (dr == NULL) { /* * There's no dr for this dbuf, so it must have been freed. * There's no need to log writes to freed blocks, so we're done. */ mutex_exit(&db->db_mtx); return (SET_ERROR(ENOENT)); } ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg); if (db->db_blkptr != NULL) { /* * We need to fill in zgd_bp with the current blkptr so that * the nopwrite code can check if we're writing the same * data that's already on disk. We can only nopwrite if we * are sure that after making the copy, db_blkptr will not * change until our i/o completes. We ensure this by * holding the db_mtx, and only allowing nopwrite if the * block is not already dirty (see below). This is verified * by dmu_sync_done(), which VERIFYs that the db_blkptr has * not changed. */ *zgd->zgd_bp = *db->db_blkptr; } /* * Assume the on-disk data is X, the current syncing data (in * txg - 1) is Y, and the current in-memory data is Z (currently * in dmu_sync). * * We usually want to perform a nopwrite if X and Z are the * same. However, if Y is different (i.e. the BP is going to * change before this write takes effect), then a nopwrite will * be incorrect - we would override with X, which could have * been freed when Y was written. * * (Note that this is not a concern when we are nop-writing from * syncing context, because X and Y must be identical, because * all previous txgs have been synced.) * * Therefore, we disable nopwrite if the current BP could change * before this TXG. There are two ways it could change: by * being dirty (dr_next is non-NULL), or by being freed * (dnode_block_freed()). This behavior is verified by * zio_done(), which VERIFYs that the override BP is identical * to the on-disk BP. */ DB_DNODE_ENTER(db); dn = DB_DNODE(db); if (dr->dr_next != NULL || dnode_block_freed(dn, db->db_blkid)) zp.zp_nopwrite = B_FALSE; DB_DNODE_EXIT(db); ASSERT(dr->dr_txg == txg); if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC || dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { /* * We have already issued a sync write for this buffer, * or this buffer has already been synced. It could not * have been dirtied since, or we would have cleared the state. */ mutex_exit(&db->db_mtx); return (SET_ERROR(EALREADY)); } ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC; mutex_exit(&db->db_mtx); dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); dsa->dsa_dr = dr; dsa->dsa_done = done; dsa->dsa_zgd = zgd; dsa->dsa_tx = NULL; zio_nowait(arc_write(pio, os->os_spa, txg, zgd->zgd_bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db), &zp, dmu_sync_ready, NULL, NULL, dmu_sync_done, dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb)); return (0); } int dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs, dmu_tx_t *tx) { dnode_t *dn; int err; err = dnode_hold(os, object, FTAG, &dn); if (err) return (err); err = dnode_set_blksz(dn, size, ibs, tx); dnode_rele(dn, FTAG); return (err); } void dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum, dmu_tx_t *tx) { dnode_t *dn; /* * Send streams include each object's checksum function. This * check ensures that the receiving system can understand the * checksum function transmitted. */ ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS); VERIFY0(dnode_hold(os, object, FTAG, &dn)); ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS); dn->dn_checksum = checksum; dnode_setdirty(dn, tx); dnode_rele(dn, FTAG); } void dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress, dmu_tx_t *tx) { dnode_t *dn; /* * Send streams include each object's compression function. This * check ensures that the receiving system can understand the * compression function transmitted. */ ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS); VERIFY0(dnode_hold(os, object, FTAG, &dn)); dn->dn_compress = compress; dnode_setdirty(dn, tx); dnode_rele(dn, FTAG); } int zfs_mdcomp_disable = 0; SYSCTL_INT(_vfs_zfs, OID_AUTO, mdcomp_disable, CTLFLAG_RWTUN, &zfs_mdcomp_disable, 0, "Disable metadata compression"); /* * When the "redundant_metadata" property is set to "most", only indirect * blocks of this level and higher will have an additional ditto block. */ int zfs_redundant_metadata_most_ditto_level = 2; void dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp) { dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET; boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) || (wp & WP_SPILL)); enum zio_checksum checksum = os->os_checksum; enum zio_compress compress = os->os_compress; enum zio_checksum dedup_checksum = os->os_dedup_checksum; boolean_t dedup = B_FALSE; boolean_t nopwrite = B_FALSE; boolean_t dedup_verify = os->os_dedup_verify; int copies = os->os_copies; /* * We maintain different write policies for each of the following * types of data: * 1. metadata * 2. preallocated blocks (i.e. level-0 blocks of a dump device) * 3. all other level 0 blocks */ if (ismd) { if (zfs_mdcomp_disable) { compress = ZIO_COMPRESS_EMPTY; } else { /* * XXX -- we should design a compression algorithm * that specializes in arrays of bps. */ compress = zio_compress_select(os->os_spa, ZIO_COMPRESS_ON, ZIO_COMPRESS_ON); } /* * Metadata always gets checksummed. If the data * checksum is multi-bit correctable, and it's not a * ZBT-style checksum, then it's suitable for metadata * as well. Otherwise, the metadata checksum defaults * to fletcher4. */ if (!(zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_METADATA) || (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED)) checksum = ZIO_CHECKSUM_FLETCHER_4; if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL || (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_MOST && (level >= zfs_redundant_metadata_most_ditto_level || DMU_OT_IS_METADATA(type) || (wp & WP_SPILL)))) copies++; } else if (wp & WP_NOFILL) { ASSERT(level == 0); /* * If we're writing preallocated blocks, we aren't actually * writing them so don't set any policy properties. These * blocks are currently only used by an external subsystem * outside of zfs (i.e. dump) and not written by the zio * pipeline. */ compress = ZIO_COMPRESS_OFF; checksum = ZIO_CHECKSUM_NOPARITY; } else { compress = zio_compress_select(os->os_spa, dn->dn_compress, compress); checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ? zio_checksum_select(dn->dn_checksum, checksum) : dedup_checksum; /* * Determine dedup setting. If we are in dmu_sync(), * we won't actually dedup now because that's all * done in syncing context; but we do want to use the * dedup checkum. If the checksum is not strong * enough to ensure unique signatures, force * dedup_verify. */ if (dedup_checksum != ZIO_CHECKSUM_OFF) { dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE; if (!(zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_DEDUP)) dedup_verify = B_TRUE; } /* * Enable nopwrite if we have secure enough checksum * algorithm (see comment in zio_nop_write) and * compression is enabled. We don't enable nopwrite if * dedup is enabled as the two features are mutually * exclusive. */ nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_NOPWRITE) && compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled); } zp->zp_checksum = checksum; zp->zp_compress = compress; ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT); zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type; zp->zp_level = level; zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa)); zp->zp_dedup = dedup; zp->zp_dedup_verify = dedup && dedup_verify; zp->zp_nopwrite = nopwrite; } int dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off) { dnode_t *dn; int err; /* * Sync any current changes before * we go trundling through the block pointers. */ err = dmu_object_wait_synced(os, object); if (err) { return (err); } err = dnode_hold(os, object, FTAG, &dn); if (err) { return (err); } err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0); dnode_rele(dn, FTAG); return (err); } /* * Given the ZFS object, if it contains any dirty nodes * this function flushes all dirty blocks to disk. This * ensures the DMU object info is updated. A more efficient * future version might just find the TXG with the maximum * ID and wait for that to be synced. */ int dmu_object_wait_synced(objset_t *os, uint64_t object) { dnode_t *dn; int error, i; error = dnode_hold(os, object, FTAG, &dn); if (error) { return (error); } for (i = 0; i < TXG_SIZE; i++) { if (list_link_active(&dn->dn_dirty_link[i])) { break; } } dnode_rele(dn, FTAG); if (i != TXG_SIZE) { txg_wait_synced(dmu_objset_pool(os), 0); } return (0); } void __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi) { dnode_phys_t *dnp = dn->dn_phys; doi->doi_data_block_size = dn->dn_datablksz; doi->doi_metadata_block_size = dn->dn_indblkshift ? 1ULL << dn->dn_indblkshift : 0; doi->doi_type = dn->dn_type; doi->doi_bonus_type = dn->dn_bonustype; doi->doi_bonus_size = dn->dn_bonuslen; doi->doi_dnodesize = dn->dn_num_slots << DNODE_SHIFT; doi->doi_indirection = dn->dn_nlevels; doi->doi_checksum = dn->dn_checksum; doi->doi_compress = dn->dn_compress; doi->doi_nblkptr = dn->dn_nblkptr; doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9; doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz; doi->doi_fill_count = 0; for (int i = 0; i < dnp->dn_nblkptr; i++) doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]); } void dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi) { rw_enter(&dn->dn_struct_rwlock, RW_READER); mutex_enter(&dn->dn_mtx); __dmu_object_info_from_dnode(dn, doi); mutex_exit(&dn->dn_mtx); rw_exit(&dn->dn_struct_rwlock); } /* * Get information on a DMU object. * If doi is NULL, just indicates whether the object exists. */ int dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi) { dnode_t *dn; int err = dnode_hold(os, object, FTAG, &dn); if (err) return (err); if (doi != NULL) dmu_object_info_from_dnode(dn, doi); dnode_rele(dn, FTAG); return (0); } /* * As above, but faster; can be used when you have a held dbuf in hand. */ void dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; DB_DNODE_ENTER(db); dmu_object_info_from_dnode(DB_DNODE(db), doi); DB_DNODE_EXIT(db); } /* * Faster still when you only care about the size. * This is specifically optimized for zfs_getattr(). */ void dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize, u_longlong_t *nblk512) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; dnode_t *dn; DB_DNODE_ENTER(db); dn = DB_DNODE(db); *blksize = dn->dn_datablksz; /* add in number of slots used for the dnode itself */ *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >> SPA_MINBLOCKSHIFT) + dn->dn_num_slots; DB_DNODE_EXIT(db); } void dmu_object_dnsize_from_db(dmu_buf_t *db_fake, int *dnsize) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; dnode_t *dn; DB_DNODE_ENTER(db); dn = DB_DNODE(db); *dnsize = dn->dn_num_slots << DNODE_SHIFT; DB_DNODE_EXIT(db); } void byteswap_uint64_array(void *vbuf, size_t size) { uint64_t *buf = vbuf; size_t count = size >> 3; int i; ASSERT((size & 7) == 0); for (i = 0; i < count; i++) buf[i] = BSWAP_64(buf[i]); } void byteswap_uint32_array(void *vbuf, size_t size) { uint32_t *buf = vbuf; size_t count = size >> 2; int i; ASSERT((size & 3) == 0); for (i = 0; i < count; i++) buf[i] = BSWAP_32(buf[i]); } void byteswap_uint16_array(void *vbuf, size_t size) { uint16_t *buf = vbuf; size_t count = size >> 1; int i; ASSERT((size & 1) == 0); for (i = 0; i < count; i++) buf[i] = BSWAP_16(buf[i]); } /* ARGSUSED */ void byteswap_uint8_array(void *vbuf, size_t size) { } void dmu_init(void) { abd_init(); zfs_dbgmsg_init(); sa_cache_init(); xuio_stat_init(); dmu_objset_init(); dnode_init(); zfetch_init(); zio_compress_init(); l2arc_init(); arc_init(); dbuf_init(); } void dmu_fini(void) { arc_fini(); /* arc depends on l2arc, so arc must go first */ l2arc_fini(); zfetch_fini(); zio_compress_fini(); dbuf_fini(); dnode_fini(); dmu_objset_fini(); xuio_stat_fini(); sa_cache_fini(); zfs_dbgmsg_fini(); abd_fini(); } Index: head/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/zfs_vnops.c =================================================================== --- head/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/zfs_vnops.c (revision 353538) +++ head/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/zfs_vnops.c (revision 353539) @@ -1,6021 +1,6020 @@ /* * CDDL HEADER START * * The contents of this file are subject to the terms of the * Common Development and Distribution License (the "License"). * You may not use this file except in compliance with the License. * * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE * or http://www.opensolaris.org/os/licensing. * See the License for the specific language governing permissions * and limitations under the License. * * When distributing Covered Code, include this CDDL HEADER in each * file and include the License file at usr/src/OPENSOLARIS.LICENSE. * If applicable, add the following below this CDDL HEADER, with the * fields enclosed by brackets "[]" replaced with your own identifying * information: Portions Copyright [yyyy] [name of copyright owner] * * CDDL HEADER END */ /* * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. * Copyright (c) 2012, 2015 by Delphix. All rights reserved. * Copyright (c) 2014 Integros [integros.com] * Copyright 2017 Nexenta Systems, Inc. */ /* Portions Copyright 2007 Jeremy Teo */ /* Portions Copyright 2010 Robert Milkowski */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* * Programming rules. * * Each vnode op performs some logical unit of work. To do this, the ZPL must * properly lock its in-core state, create a DMU transaction, do the work, * record this work in the intent log (ZIL), commit the DMU transaction, * and wait for the intent log to commit if it is a synchronous operation. * Moreover, the vnode ops must work in both normal and log replay context. * The ordering of events is important to avoid deadlocks and references * to freed memory. The example below illustrates the following Big Rules: * * (1) A check must be made in each zfs thread for a mounted file system. * This is done avoiding races using ZFS_ENTER(zfsvfs). * A ZFS_EXIT(zfsvfs) is needed before all returns. Any znodes * must be checked with ZFS_VERIFY_ZP(zp). Both of these macros * can return EIO from the calling function. * * (2) VN_RELE() should always be the last thing except for zil_commit() * (if necessary) and ZFS_EXIT(). This is for 3 reasons: * First, if it's the last reference, the vnode/znode * can be freed, so the zp may point to freed memory. Second, the last * reference will call zfs_zinactive(), which may induce a lot of work -- * pushing cached pages (which acquires range locks) and syncing out * cached atime changes. Third, zfs_zinactive() may require a new tx, * which could deadlock the system if you were already holding one. * If you must call VN_RELE() within a tx then use VN_RELE_ASYNC(). * * (3) All range locks must be grabbed before calling dmu_tx_assign(), * as they can span dmu_tx_assign() calls. * * (4) If ZPL locks are held, pass TXG_NOWAIT as the second argument to * dmu_tx_assign(). This is critical because we don't want to block * while holding locks. * * If no ZPL locks are held (aside from ZFS_ENTER()), use TXG_WAIT. This * reduces lock contention and CPU usage when we must wait (note that if * throughput is constrained by the storage, nearly every transaction * must wait). * * Note, in particular, that if a lock is sometimes acquired before * the tx assigns, and sometimes after (e.g. z_lock), then failing * to use a non-blocking assign can deadlock the system. The scenario: * * Thread A has grabbed a lock before calling dmu_tx_assign(). * Thread B is in an already-assigned tx, and blocks for this lock. * Thread A calls dmu_tx_assign(TXG_WAIT) and blocks in txg_wait_open() * forever, because the previous txg can't quiesce until B's tx commits. * * If dmu_tx_assign() returns ERESTART and zfsvfs->z_assign is TXG_NOWAIT, * then drop all locks, call dmu_tx_wait(), and try again. On subsequent * calls to dmu_tx_assign(), pass TXG_NOTHROTTLE in addition to TXG_NOWAIT, * to indicate that this operation has already called dmu_tx_wait(). * This will ensure that we don't retry forever, waiting a short bit * each time. * * (5) If the operation succeeded, generate the intent log entry for it * before dropping locks. This ensures that the ordering of events * in the intent log matches the order in which they actually occurred. * During ZIL replay the zfs_log_* functions will update the sequence * number to indicate the zil transaction has replayed. * * (6) At the end of each vnode op, the DMU tx must always commit, * regardless of whether there were any errors. * * (7) After dropping all locks, invoke zil_commit(zilog, foid) * to ensure that synchronous semantics are provided when necessary. * * In general, this is how things should be ordered in each vnode op: * * ZFS_ENTER(zfsvfs); // exit if unmounted * top: * zfs_dirent_lookup(&dl, ...) // lock directory entry (may VN_HOLD()) * rw_enter(...); // grab any other locks you need * tx = dmu_tx_create(...); // get DMU tx * dmu_tx_hold_*(); // hold each object you might modify * error = dmu_tx_assign(tx, (waited ? TXG_NOTHROTTLE : 0) | TXG_NOWAIT); * if (error) { * rw_exit(...); // drop locks * zfs_dirent_unlock(dl); // unlock directory entry * VN_RELE(...); // release held vnodes * if (error == ERESTART) { * waited = B_TRUE; * dmu_tx_wait(tx); * dmu_tx_abort(tx); * goto top; * } * dmu_tx_abort(tx); // abort DMU tx * ZFS_EXIT(zfsvfs); // finished in zfs * return (error); // really out of space * } * error = do_real_work(); // do whatever this VOP does * if (error == 0) * zfs_log_*(...); // on success, make ZIL entry * dmu_tx_commit(tx); // commit DMU tx -- error or not * rw_exit(...); // drop locks * zfs_dirent_unlock(dl); // unlock directory entry * VN_RELE(...); // release held vnodes * zil_commit(zilog, foid); // synchronous when necessary * ZFS_EXIT(zfsvfs); // finished in zfs * return (error); // done, report error */ /* ARGSUSED */ static int zfs_open(vnode_t **vpp, int flag, cred_t *cr, caller_context_t *ct) { znode_t *zp = VTOZ(*vpp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); if ((flag & FWRITE) && (zp->z_pflags & ZFS_APPENDONLY) && ((flag & FAPPEND) == 0)) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EPERM)); } if (!zfs_has_ctldir(zp) && zp->z_zfsvfs->z_vscan && ZTOV(zp)->v_type == VREG && !(zp->z_pflags & ZFS_AV_QUARANTINED) && zp->z_size > 0) { if (fs_vscan(*vpp, cr, 0) != 0) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EACCES)); } } /* Keep a count of the synchronous opens in the znode */ if (flag & (FSYNC | FDSYNC)) atomic_inc_32(&zp->z_sync_cnt); ZFS_EXIT(zfsvfs); return (0); } /* ARGSUSED */ static int zfs_close(vnode_t *vp, int flag, int count, offset_t offset, cred_t *cr, caller_context_t *ct) { znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; /* * Clean up any locks held by this process on the vp. */ cleanlocks(vp, ddi_get_pid(), 0); cleanshares(vp, ddi_get_pid()); ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); /* Decrement the synchronous opens in the znode */ if ((flag & (FSYNC | FDSYNC)) && (count == 1)) atomic_dec_32(&zp->z_sync_cnt); if (!zfs_has_ctldir(zp) && zp->z_zfsvfs->z_vscan && ZTOV(zp)->v_type == VREG && !(zp->z_pflags & ZFS_AV_QUARANTINED) && zp->z_size > 0) VERIFY(fs_vscan(vp, cr, 1) == 0); ZFS_EXIT(zfsvfs); return (0); } /* * Lseek support for finding holes (cmd == _FIO_SEEK_HOLE) and * data (cmd == _FIO_SEEK_DATA). "off" is an in/out parameter. */ static int zfs_holey(vnode_t *vp, u_long cmd, offset_t *off) { znode_t *zp = VTOZ(vp); uint64_t noff = (uint64_t)*off; /* new offset */ uint64_t file_sz; int error; boolean_t hole; file_sz = zp->z_size; if (noff >= file_sz) { return (SET_ERROR(ENXIO)); } if (cmd == _FIO_SEEK_HOLE) hole = B_TRUE; else hole = B_FALSE; error = dmu_offset_next(zp->z_zfsvfs->z_os, zp->z_id, hole, &noff); if (error == ESRCH) return (SET_ERROR(ENXIO)); /* * We could find a hole that begins after the logical end-of-file, * because dmu_offset_next() only works on whole blocks. If the * EOF falls mid-block, then indicate that the "virtual hole" * at the end of the file begins at the logical EOF, rather than * at the end of the last block. */ if (noff > file_sz) { ASSERT(hole); noff = file_sz; } if (noff < *off) return (error); *off = noff; return (error); } /* ARGSUSED */ static int zfs_ioctl(vnode_t *vp, u_long com, intptr_t data, int flag, cred_t *cred, int *rvalp, caller_context_t *ct) { offset_t off; offset_t ndata; dmu_object_info_t doi; int error; zfsvfs_t *zfsvfs; znode_t *zp; switch (com) { case _FIOFFS: { return (0); /* * The following two ioctls are used by bfu. Faking out, * necessary to avoid bfu errors. */ } case _FIOGDIO: case _FIOSDIO: { return (0); } case _FIO_SEEK_DATA: case _FIO_SEEK_HOLE: { #ifdef illumos if (ddi_copyin((void *)data, &off, sizeof (off), flag)) return (SET_ERROR(EFAULT)); #else off = *(offset_t *)data; #endif zp = VTOZ(vp); zfsvfs = zp->z_zfsvfs; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); /* offset parameter is in/out */ error = zfs_holey(vp, com, &off); ZFS_EXIT(zfsvfs); if (error) return (error); #ifdef illumos if (ddi_copyout(&off, (void *)data, sizeof (off), flag)) return (SET_ERROR(EFAULT)); #else *(offset_t *)data = off; #endif return (0); } #ifdef illumos case _FIO_COUNT_FILLED: { /* * _FIO_COUNT_FILLED adds a new ioctl command which * exposes the number of filled blocks in a * ZFS object. */ zp = VTOZ(vp); zfsvfs = zp->z_zfsvfs; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); /* * Wait for all dirty blocks for this object * to get synced out to disk, and the DMU info * updated. */ error = dmu_object_wait_synced(zfsvfs->z_os, zp->z_id); if (error) { ZFS_EXIT(zfsvfs); return (error); } /* * Retrieve fill count from DMU object. */ error = dmu_object_info(zfsvfs->z_os, zp->z_id, &doi); if (error) { ZFS_EXIT(zfsvfs); return (error); } ndata = doi.doi_fill_count; ZFS_EXIT(zfsvfs); if (ddi_copyout(&ndata, (void *)data, sizeof (ndata), flag)) return (SET_ERROR(EFAULT)); return (0); } #endif } return (SET_ERROR(ENOTTY)); } static vm_page_t page_busy(vnode_t *vp, int64_t start, int64_t off, int64_t nbytes) { vm_object_t obj; vm_page_t pp; int64_t end; /* * At present vm_page_clear_dirty extends the cleared range to DEV_BSIZE * aligned boundaries, if the range is not aligned. As a result a * DEV_BSIZE subrange with partially dirty data may get marked as clean. * It may happen that all DEV_BSIZE subranges are marked clean and thus * the whole page would be considred clean despite have some dirty data. * For this reason we should shrink the range to DEV_BSIZE aligned * boundaries before calling vm_page_clear_dirty. */ end = rounddown2(off + nbytes, DEV_BSIZE); off = roundup2(off, DEV_BSIZE); nbytes = end - off; obj = vp->v_object; zfs_vmobject_assert_wlocked(obj); vm_page_grab_valid(&pp, obj, OFF_TO_IDX(start), VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_NORMAL | VM_ALLOC_IGN_SBUSY); if (pp != NULL) { ASSERT3U(pp->valid, ==, VM_PAGE_BITS_ALL); vm_object_pip_add(obj, 1); pmap_remove_write(pp); if (nbytes != 0) vm_page_clear_dirty(pp, off, nbytes); } return (pp); } static void page_unbusy(vm_page_t pp) { vm_page_sunbusy(pp); vm_object_pip_wakeup(pp->object); } static vm_page_t page_wire(vnode_t *vp, int64_t start) { vm_object_t obj; vm_page_t m; obj = vp->v_object; zfs_vmobject_assert_wlocked(obj); vm_page_grab_valid(&m, obj, OFF_TO_IDX(start), VM_ALLOC_NOCREAT | VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY | VM_ALLOC_NOBUSY); return (m); } static void page_unwire(vm_page_t pp) { vm_page_unwire(pp, PQ_ACTIVE); } /* * When a file is memory mapped, we must keep the IO data synchronized * between the DMU cache and the memory mapped pages. What this means: * * On Write: If we find a memory mapped page, we write to *both* * the page and the dmu buffer. */ static void update_pages(vnode_t *vp, int64_t start, int len, objset_t *os, uint64_t oid, int segflg, dmu_tx_t *tx) { vm_object_t obj; struct sf_buf *sf; caddr_t va; int off; ASSERT(segflg != UIO_NOCOPY); ASSERT(vp->v_mount != NULL); obj = vp->v_object; ASSERT(obj != NULL); off = start & PAGEOFFSET; zfs_vmobject_wlock(obj); vm_object_pip_add(obj, 1); for (start &= PAGEMASK; len > 0; start += PAGESIZE) { vm_page_t pp; int nbytes = imin(PAGESIZE - off, len); if ((pp = page_busy(vp, start, off, nbytes)) != NULL) { zfs_vmobject_wunlock(obj); va = zfs_map_page(pp, &sf); (void) dmu_read(os, oid, start+off, nbytes, va+off, DMU_READ_PREFETCH);; zfs_unmap_page(sf); zfs_vmobject_wlock(obj); page_unbusy(pp); } len -= nbytes; off = 0; } vm_object_pip_wakeup(obj); zfs_vmobject_wunlock(obj); } /* * Read with UIO_NOCOPY flag means that sendfile(2) requests * ZFS to populate a range of page cache pages with data. * * NOTE: this function could be optimized to pre-allocate * all pages in advance, drain exclusive busy on all of them, * map them into contiguous KVA region and populate them * in one single dmu_read() call. */ static int mappedread_sf(vnode_t *vp, int nbytes, uio_t *uio) { znode_t *zp = VTOZ(vp); objset_t *os = zp->z_zfsvfs->z_os; struct sf_buf *sf; vm_object_t obj; vm_page_t pp; int64_t start; caddr_t va; int len = nbytes; int off; int error = 0; ASSERT(uio->uio_segflg == UIO_NOCOPY); ASSERT(vp->v_mount != NULL); obj = vp->v_object; ASSERT(obj != NULL); ASSERT((uio->uio_loffset & PAGEOFFSET) == 0); zfs_vmobject_wlock(obj); for (start = uio->uio_loffset; len > 0; start += PAGESIZE) { int bytes = MIN(PAGESIZE, len); pp = vm_page_grab(obj, OFF_TO_IDX(start), VM_ALLOC_SBUSY | VM_ALLOC_NORMAL | VM_ALLOC_IGN_SBUSY); - if (pp->valid == 0) { + if (vm_page_none_valid(pp)) { zfs_vmobject_wunlock(obj); va = zfs_map_page(pp, &sf); error = dmu_read(os, zp->z_id, start, bytes, va, DMU_READ_PREFETCH); if (bytes != PAGESIZE && error == 0) bzero(va + bytes, PAGESIZE - bytes); zfs_unmap_page(sf); zfs_vmobject_wlock(obj); - vm_page_sunbusy(pp); - if (error) { - if (!vm_page_busied(pp) && !vm_page_wired(pp) && - pp->valid == 0) - vm_page_free(pp); - } else { - pp->valid = VM_PAGE_BITS_ALL; + if (error == 0) { + vm_page_valid(pp); vm_page_lock(pp); vm_page_activate(pp); vm_page_unlock(pp); } + vm_page_sunbusy(pp); + if (error != 0 && !vm_page_wired(pp) == 0 && + pp->valid == 0 && vm_page_tryxbusy(pp)) + vm_page_free(pp); } else { ASSERT3U(pp->valid, ==, VM_PAGE_BITS_ALL); vm_page_sunbusy(pp); } if (error) break; uio->uio_resid -= bytes; uio->uio_offset += bytes; len -= bytes; } zfs_vmobject_wunlock(obj); return (error); } /* * When a file is memory mapped, we must keep the IO data synchronized * between the DMU cache and the memory mapped pages. What this means: * * On Read: We "read" preferentially from memory mapped pages, * else we default from the dmu buffer. * * NOTE: We will always "break up" the IO into PAGESIZE uiomoves when * the file is memory mapped. */ static int mappedread(vnode_t *vp, int nbytes, uio_t *uio) { znode_t *zp = VTOZ(vp); vm_object_t obj; int64_t start; caddr_t va; int len = nbytes; int off; int error = 0; ASSERT(vp->v_mount != NULL); obj = vp->v_object; ASSERT(obj != NULL); start = uio->uio_loffset; off = start & PAGEOFFSET; zfs_vmobject_wlock(obj); for (start &= PAGEMASK; len > 0; start += PAGESIZE) { vm_page_t pp; uint64_t bytes = MIN(PAGESIZE - off, len); if (pp = page_wire(vp, start)) { struct sf_buf *sf; caddr_t va; zfs_vmobject_wunlock(obj); va = zfs_map_page(pp, &sf); #ifdef illumos error = uiomove(va + off, bytes, UIO_READ, uio); #else error = vn_io_fault_uiomove(va + off, bytes, uio); #endif zfs_unmap_page(sf); zfs_vmobject_wlock(obj); page_unwire(pp); } else { zfs_vmobject_wunlock(obj); error = dmu_read_uio_dbuf(sa_get_db(zp->z_sa_hdl), uio, bytes); zfs_vmobject_wlock(obj); } len -= bytes; off = 0; if (error) break; } zfs_vmobject_wunlock(obj); return (error); } offset_t zfs_read_chunk_size = 1024 * 1024; /* Tunable */ /* * Read bytes from specified file into supplied buffer. * * IN: vp - vnode of file to be read from. * uio - structure supplying read location, range info, * and return buffer. * ioflag - SYNC flags; used to provide FRSYNC semantics. * cr - credentials of caller. * ct - caller context * * OUT: uio - updated offset and range, buffer filled. * * RETURN: 0 on success, error code on failure. * * Side Effects: * vp - atime updated if byte count > 0 */ /* ARGSUSED */ static int zfs_read(vnode_t *vp, uio_t *uio, int ioflag, cred_t *cr, caller_context_t *ct) { znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; ssize_t n, nbytes; int error = 0; rl_t *rl; xuio_t *xuio = NULL; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); if (zp->z_pflags & ZFS_AV_QUARANTINED) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EACCES)); } /* * Validate file offset */ if (uio->uio_loffset < (offset_t)0) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EINVAL)); } /* * Fasttrack empty reads */ if (uio->uio_resid == 0) { ZFS_EXIT(zfsvfs); return (0); } /* * Check for mandatory locks */ if (MANDMODE(zp->z_mode)) { if (error = chklock(vp, FREAD, uio->uio_loffset, uio->uio_resid, uio->uio_fmode, ct)) { ZFS_EXIT(zfsvfs); return (error); } } /* * If we're in FRSYNC mode, sync out this znode before reading it. */ if (zfsvfs->z_log && (ioflag & FRSYNC || zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS)) zil_commit(zfsvfs->z_log, zp->z_id); /* * Lock the range against changes. */ rl = zfs_range_lock(zp, uio->uio_loffset, uio->uio_resid, RL_READER); /* * If we are reading past end-of-file we can skip * to the end; but we might still need to set atime. */ if (uio->uio_loffset >= zp->z_size) { error = 0; goto out; } ASSERT(uio->uio_loffset < zp->z_size); n = MIN(uio->uio_resid, zp->z_size - uio->uio_loffset); #ifdef illumos if ((uio->uio_extflg == UIO_XUIO) && (((xuio_t *)uio)->xu_type == UIOTYPE_ZEROCOPY)) { int nblk; int blksz = zp->z_blksz; uint64_t offset = uio->uio_loffset; xuio = (xuio_t *)uio; if ((ISP2(blksz))) { nblk = (P2ROUNDUP(offset + n, blksz) - P2ALIGN(offset, blksz)) / blksz; } else { ASSERT(offset + n <= blksz); nblk = 1; } (void) dmu_xuio_init(xuio, nblk); if (vn_has_cached_data(vp)) { /* * For simplicity, we always allocate a full buffer * even if we only expect to read a portion of a block. */ while (--nblk >= 0) { (void) dmu_xuio_add(xuio, dmu_request_arcbuf(sa_get_db(zp->z_sa_hdl), blksz), 0, blksz); } } } #endif /* illumos */ while (n > 0) { nbytes = MIN(n, zfs_read_chunk_size - P2PHASE(uio->uio_loffset, zfs_read_chunk_size)); #ifdef __FreeBSD__ if (uio->uio_segflg == UIO_NOCOPY) error = mappedread_sf(vp, nbytes, uio); else #endif /* __FreeBSD__ */ if (vn_has_cached_data(vp)) { error = mappedread(vp, nbytes, uio); } else { error = dmu_read_uio_dbuf(sa_get_db(zp->z_sa_hdl), uio, nbytes); } if (error) { /* convert checksum errors into IO errors */ if (error == ECKSUM) error = SET_ERROR(EIO); break; } n -= nbytes; } out: zfs_range_unlock(rl); ZFS_ACCESSTIME_STAMP(zfsvfs, zp); ZFS_EXIT(zfsvfs); return (error); } /* * Write the bytes to a file. * * IN: vp - vnode of file to be written to. * uio - structure supplying write location, range info, * and data buffer. * ioflag - FAPPEND, FSYNC, and/or FDSYNC. FAPPEND is * set if in append mode. * cr - credentials of caller. * ct - caller context (NFS/CIFS fem monitor only) * * OUT: uio - updated offset and range. * * RETURN: 0 on success, error code on failure. * * Timestamps: * vp - ctime|mtime updated if byte count > 0 */ /* ARGSUSED */ static int zfs_write(vnode_t *vp, uio_t *uio, int ioflag, cred_t *cr, caller_context_t *ct) { znode_t *zp = VTOZ(vp); rlim64_t limit = MAXOFFSET_T; ssize_t start_resid = uio->uio_resid; ssize_t tx_bytes; uint64_t end_size; dmu_tx_t *tx; zfsvfs_t *zfsvfs = zp->z_zfsvfs; zilog_t *zilog; offset_t woff; ssize_t n, nbytes; rl_t *rl; int max_blksz = zfsvfs->z_max_blksz; int error = 0; arc_buf_t *abuf; iovec_t *aiov = NULL; xuio_t *xuio = NULL; int i_iov = 0; int iovcnt = uio->uio_iovcnt; iovec_t *iovp = uio->uio_iov; int write_eof; int count = 0; sa_bulk_attr_t bulk[4]; uint64_t mtime[2], ctime[2]; /* * Fasttrack empty write */ n = start_resid; if (n == 0) return (0); if (limit == RLIM64_INFINITY || limit > MAXOFFSET_T) limit = MAXOFFSET_T; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_MTIME(zfsvfs), NULL, &mtime, 16); SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_CTIME(zfsvfs), NULL, &ctime, 16); SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_SIZE(zfsvfs), NULL, &zp->z_size, 8); SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_FLAGS(zfsvfs), NULL, &zp->z_pflags, 8); /* * In a case vp->v_vfsp != zp->z_zfsvfs->z_vfs (e.g. snapshots) our * callers might not be able to detect properly that we are read-only, * so check it explicitly here. */ if (zfsvfs->z_vfs->vfs_flag & VFS_RDONLY) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EROFS)); } /* * If immutable or not appending then return EPERM. * Intentionally allow ZFS_READONLY through here. * See zfs_zaccess_common() */ if ((zp->z_pflags & ZFS_IMMUTABLE) || ((zp->z_pflags & ZFS_APPENDONLY) && !(ioflag & FAPPEND) && (uio->uio_loffset < zp->z_size))) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EPERM)); } zilog = zfsvfs->z_log; /* * Validate file offset */ woff = ioflag & FAPPEND ? zp->z_size : uio->uio_loffset; if (woff < 0) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EINVAL)); } /* * Check for mandatory locks before calling zfs_range_lock() * in order to prevent a deadlock with locks set via fcntl(). */ if (MANDMODE((mode_t)zp->z_mode) && (error = chklock(vp, FWRITE, woff, n, uio->uio_fmode, ct)) != 0) { ZFS_EXIT(zfsvfs); return (error); } #ifdef illumos /* * Pre-fault the pages to ensure slow (eg NFS) pages * don't hold up txg. * Skip this if uio contains loaned arc_buf. */ if ((uio->uio_extflg == UIO_XUIO) && (((xuio_t *)uio)->xu_type == UIOTYPE_ZEROCOPY)) xuio = (xuio_t *)uio; else uio_prefaultpages(MIN(n, max_blksz), uio); #endif /* * If in append mode, set the io offset pointer to eof. */ if (ioflag & FAPPEND) { /* * Obtain an appending range lock to guarantee file append * semantics. We reset the write offset once we have the lock. */ rl = zfs_range_lock(zp, 0, n, RL_APPEND); woff = rl->r_off; if (rl->r_len == UINT64_MAX) { /* * We overlocked the file because this write will cause * the file block size to increase. * Note that zp_size cannot change with this lock held. */ woff = zp->z_size; } uio->uio_loffset = woff; } else { /* * Note that if the file block size will change as a result of * this write, then this range lock will lock the entire file * so that we can re-write the block safely. */ rl = zfs_range_lock(zp, woff, n, RL_WRITER); } if (vn_rlimit_fsize(vp, uio, uio->uio_td)) { zfs_range_unlock(rl); ZFS_EXIT(zfsvfs); return (EFBIG); } if (woff >= limit) { zfs_range_unlock(rl); ZFS_EXIT(zfsvfs); return (SET_ERROR(EFBIG)); } if ((woff + n) > limit || woff > (limit - n)) n = limit - woff; /* Will this write extend the file length? */ write_eof = (woff + n > zp->z_size); end_size = MAX(zp->z_size, woff + n); /* * Write the file in reasonable size chunks. Each chunk is written * in a separate transaction; this keeps the intent log records small * and allows us to do more fine-grained space accounting. */ while (n > 0) { abuf = NULL; woff = uio->uio_loffset; if (zfs_owner_overquota(zfsvfs, zp, B_FALSE) || zfs_owner_overquota(zfsvfs, zp, B_TRUE)) { if (abuf != NULL) dmu_return_arcbuf(abuf); error = SET_ERROR(EDQUOT); break; } if (xuio && abuf == NULL) { ASSERT(i_iov < iovcnt); aiov = &iovp[i_iov]; abuf = dmu_xuio_arcbuf(xuio, i_iov); dmu_xuio_clear(xuio, i_iov); DTRACE_PROBE3(zfs_cp_write, int, i_iov, iovec_t *, aiov, arc_buf_t *, abuf); ASSERT((aiov->iov_base == abuf->b_data) || ((char *)aiov->iov_base - (char *)abuf->b_data + aiov->iov_len == arc_buf_size(abuf))); i_iov++; } else if (abuf == NULL && n >= max_blksz && woff >= zp->z_size && P2PHASE(woff, max_blksz) == 0 && zp->z_blksz == max_blksz) { /* * This write covers a full block. "Borrow" a buffer * from the dmu so that we can fill it before we enter * a transaction. This avoids the possibility of * holding up the transaction if the data copy hangs * up on a pagefault (e.g., from an NFS server mapping). */ size_t cbytes; abuf = dmu_request_arcbuf(sa_get_db(zp->z_sa_hdl), max_blksz); ASSERT(abuf != NULL); ASSERT(arc_buf_size(abuf) == max_blksz); if (error = uiocopy(abuf->b_data, max_blksz, UIO_WRITE, uio, &cbytes)) { dmu_return_arcbuf(abuf); break; } ASSERT(cbytes == max_blksz); } /* * Start a transaction. */ tx = dmu_tx_create(zfsvfs->z_os); dmu_tx_hold_sa(tx, zp->z_sa_hdl, B_FALSE); dmu_tx_hold_write(tx, zp->z_id, woff, MIN(n, max_blksz)); zfs_sa_upgrade_txholds(tx, zp); error = dmu_tx_assign(tx, TXG_WAIT); if (error) { dmu_tx_abort(tx); if (abuf != NULL) dmu_return_arcbuf(abuf); break; } /* * If zfs_range_lock() over-locked we grow the blocksize * and then reduce the lock range. This will only happen * on the first iteration since zfs_range_reduce() will * shrink down r_len to the appropriate size. */ if (rl->r_len == UINT64_MAX) { uint64_t new_blksz; if (zp->z_blksz > max_blksz) { /* * File's blocksize is already larger than the * "recordsize" property. Only let it grow to * the next power of 2. */ ASSERT(!ISP2(zp->z_blksz)); new_blksz = MIN(end_size, 1 << highbit64(zp->z_blksz)); } else { new_blksz = MIN(end_size, max_blksz); } zfs_grow_blocksize(zp, new_blksz, tx); zfs_range_reduce(rl, woff, n); } /* * XXX - should we really limit each write to z_max_blksz? * Perhaps we should use SPA_MAXBLOCKSIZE chunks? */ nbytes = MIN(n, max_blksz - P2PHASE(woff, max_blksz)); if (woff + nbytes > zp->z_size) vnode_pager_setsize(vp, woff + nbytes); if (abuf == NULL) { tx_bytes = uio->uio_resid; error = dmu_write_uio_dbuf(sa_get_db(zp->z_sa_hdl), uio, nbytes, tx); tx_bytes -= uio->uio_resid; } else { tx_bytes = nbytes; ASSERT(xuio == NULL || tx_bytes == aiov->iov_len); /* * If this is not a full block write, but we are * extending the file past EOF and this data starts * block-aligned, use assign_arcbuf(). Otherwise, * write via dmu_write(). */ if (tx_bytes < max_blksz && (!write_eof || aiov->iov_base != abuf->b_data)) { ASSERT(xuio); dmu_write(zfsvfs->z_os, zp->z_id, woff, aiov->iov_len, aiov->iov_base, tx); dmu_return_arcbuf(abuf); xuio_stat_wbuf_copied(); } else { ASSERT(xuio || tx_bytes == max_blksz); dmu_assign_arcbuf(sa_get_db(zp->z_sa_hdl), woff, abuf, tx); } ASSERT(tx_bytes <= uio->uio_resid); uioskip(uio, tx_bytes); } if (tx_bytes && vn_has_cached_data(vp)) { update_pages(vp, woff, tx_bytes, zfsvfs->z_os, zp->z_id, uio->uio_segflg, tx); } /* * If we made no progress, we're done. If we made even * partial progress, update the znode and ZIL accordingly. */ if (tx_bytes == 0) { (void) sa_update(zp->z_sa_hdl, SA_ZPL_SIZE(zfsvfs), (void *)&zp->z_size, sizeof (uint64_t), tx); dmu_tx_commit(tx); ASSERT(error != 0); break; } /* * Clear Set-UID/Set-GID bits on successful write if not * privileged and at least one of the excute bits is set. * * It would be nice to to this after all writes have * been done, but that would still expose the ISUID/ISGID * to another app after the partial write is committed. * * Note: we don't call zfs_fuid_map_id() here because * user 0 is not an ephemeral uid. */ mutex_enter(&zp->z_acl_lock); if ((zp->z_mode & (S_IXUSR | (S_IXUSR >> 3) | (S_IXUSR >> 6))) != 0 && (zp->z_mode & (S_ISUID | S_ISGID)) != 0 && secpolicy_vnode_setid_retain(vp, cr, (zp->z_mode & S_ISUID) != 0 && zp->z_uid == 0) != 0) { uint64_t newmode; zp->z_mode &= ~(S_ISUID | S_ISGID); newmode = zp->z_mode; (void) sa_update(zp->z_sa_hdl, SA_ZPL_MODE(zfsvfs), (void *)&newmode, sizeof (uint64_t), tx); } mutex_exit(&zp->z_acl_lock); zfs_tstamp_update_setup(zp, CONTENT_MODIFIED, mtime, ctime, B_TRUE); /* * Update the file size (zp_size) if it has changed; * account for possible concurrent updates. */ while ((end_size = zp->z_size) < uio->uio_loffset) { (void) atomic_cas_64(&zp->z_size, end_size, uio->uio_loffset); #ifdef illumos ASSERT(error == 0); #else ASSERT(error == 0 || error == EFAULT); #endif } /* * If we are replaying and eof is non zero then force * the file size to the specified eof. Note, there's no * concurrency during replay. */ if (zfsvfs->z_replay && zfsvfs->z_replay_eof != 0) zp->z_size = zfsvfs->z_replay_eof; if (error == 0) error = sa_bulk_update(zp->z_sa_hdl, bulk, count, tx); else (void) sa_bulk_update(zp->z_sa_hdl, bulk, count, tx); zfs_log_write(zilog, tx, TX_WRITE, zp, woff, tx_bytes, ioflag); dmu_tx_commit(tx); if (error != 0) break; ASSERT(tx_bytes == nbytes); n -= nbytes; #ifdef illumos if (!xuio && n > 0) uio_prefaultpages(MIN(n, max_blksz), uio); #endif } zfs_range_unlock(rl); /* * If we're in replay mode, or we made no progress, return error. * Otherwise, it's at least a partial write, so it's successful. */ if (zfsvfs->z_replay || uio->uio_resid == start_resid) { ZFS_EXIT(zfsvfs); return (error); } #ifdef __FreeBSD__ /* * EFAULT means that at least one page of the source buffer was not * available. VFS will re-try remaining I/O upon this error. */ if (error == EFAULT) { ZFS_EXIT(zfsvfs); return (error); } #endif if (ioflag & (FSYNC | FDSYNC) || zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS) zil_commit(zilog, zp->z_id); ZFS_EXIT(zfsvfs); return (0); } /* ARGSUSED */ void zfs_get_done(zgd_t *zgd, int error) { znode_t *zp = zgd->zgd_private; objset_t *os = zp->z_zfsvfs->z_os; if (zgd->zgd_db) dmu_buf_rele(zgd->zgd_db, zgd); zfs_range_unlock(zgd->zgd_rl); /* * Release the vnode asynchronously as we currently have the * txg stopped from syncing. */ VN_RELE_ASYNC(ZTOV(zp), dsl_pool_vnrele_taskq(dmu_objset_pool(os))); kmem_free(zgd, sizeof (zgd_t)); } #ifdef DEBUG static int zil_fault_io = 0; #endif /* * Get data to generate a TX_WRITE intent log record. */ int zfs_get_data(void *arg, lr_write_t *lr, char *buf, struct lwb *lwb, zio_t *zio) { zfsvfs_t *zfsvfs = arg; objset_t *os = zfsvfs->z_os; znode_t *zp; uint64_t object = lr->lr_foid; uint64_t offset = lr->lr_offset; uint64_t size = lr->lr_length; dmu_buf_t *db; zgd_t *zgd; int error = 0; ASSERT3P(lwb, !=, NULL); ASSERT3P(zio, !=, NULL); ASSERT3U(size, !=, 0); /* * Nothing to do if the file has been removed */ if (zfs_zget(zfsvfs, object, &zp) != 0) return (SET_ERROR(ENOENT)); if (zp->z_unlinked) { /* * Release the vnode asynchronously as we currently have the * txg stopped from syncing. */ VN_RELE_ASYNC(ZTOV(zp), dsl_pool_vnrele_taskq(dmu_objset_pool(os))); return (SET_ERROR(ENOENT)); } zgd = (zgd_t *)kmem_zalloc(sizeof (zgd_t), KM_SLEEP); zgd->zgd_lwb = lwb; zgd->zgd_private = zp; /* * Write records come in two flavors: immediate and indirect. * For small writes it's cheaper to store the data with the * log record (immediate); for large writes it's cheaper to * sync the data and get a pointer to it (indirect) so that * we don't have to write the data twice. */ if (buf != NULL) { /* immediate write */ zgd->zgd_rl = zfs_range_lock(zp, offset, size, RL_READER); /* test for truncation needs to be done while range locked */ if (offset >= zp->z_size) { error = SET_ERROR(ENOENT); } else { error = dmu_read(os, object, offset, size, buf, DMU_READ_NO_PREFETCH); } ASSERT(error == 0 || error == ENOENT); } else { /* indirect write */ /* * Have to lock the whole block to ensure when it's * written out and its checksum is being calculated * that no one can change the data. We need to re-check * blocksize after we get the lock in case it's changed! */ for (;;) { uint64_t blkoff; size = zp->z_blksz; blkoff = ISP2(size) ? P2PHASE(offset, size) : offset; offset -= blkoff; zgd->zgd_rl = zfs_range_lock(zp, offset, size, RL_READER); if (zp->z_blksz == size) break; offset += blkoff; zfs_range_unlock(zgd->zgd_rl); } /* test for truncation needs to be done while range locked */ if (lr->lr_offset >= zp->z_size) error = SET_ERROR(ENOENT); #ifdef DEBUG if (zil_fault_io) { error = SET_ERROR(EIO); zil_fault_io = 0; } #endif if (error == 0) error = dmu_buf_hold(os, object, offset, zgd, &db, DMU_READ_NO_PREFETCH); if (error == 0) { blkptr_t *bp = &lr->lr_blkptr; zgd->zgd_db = db; zgd->zgd_bp = bp; ASSERT(db->db_offset == offset); ASSERT(db->db_size == size); error = dmu_sync(zio, lr->lr_common.lrc_txg, zfs_get_done, zgd); ASSERT(error || lr->lr_length <= size); /* * On success, we need to wait for the write I/O * initiated by dmu_sync() to complete before we can * release this dbuf. We will finish everything up * in the zfs_get_done() callback. */ if (error == 0) return (0); if (error == EALREADY) { lr->lr_common.lrc_txtype = TX_WRITE2; /* * TX_WRITE2 relies on the data previously * written by the TX_WRITE that caused * EALREADY. We zero out the BP because * it is the old, currently-on-disk BP. */ zgd->zgd_bp = NULL; BP_ZERO(bp); error = 0; } } } zfs_get_done(zgd, error); return (error); } /*ARGSUSED*/ static int zfs_access(vnode_t *vp, int mode, int flag, cred_t *cr, caller_context_t *ct) { znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; int error; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); if (flag & V_ACE_MASK) error = zfs_zaccess(zp, mode, flag, B_FALSE, cr); else error = zfs_zaccess_rwx(zp, mode, flag, cr); ZFS_EXIT(zfsvfs); return (error); } static int zfs_dd_callback(struct mount *mp, void *arg, int lkflags, struct vnode **vpp) { int error; *vpp = arg; error = vn_lock(*vpp, lkflags); if (error != 0) vrele(*vpp); return (error); } static int zfs_lookup_lock(vnode_t *dvp, vnode_t *vp, const char *name, int lkflags) { znode_t *zdp = VTOZ(dvp); zfsvfs_t *zfsvfs = zdp->z_zfsvfs; int error; int ltype; ASSERT_VOP_LOCKED(dvp, __func__); #ifdef DIAGNOSTIC if ((zdp->z_pflags & ZFS_XATTR) == 0) VERIFY(!RRM_LOCK_HELD(&zfsvfs->z_teardown_lock)); #endif if (name[0] == 0 || (name[0] == '.' && name[1] == 0)) { ASSERT3P(dvp, ==, vp); vref(dvp); ltype = lkflags & LK_TYPE_MASK; if (ltype != VOP_ISLOCKED(dvp)) { if (ltype == LK_EXCLUSIVE) vn_lock(dvp, LK_UPGRADE | LK_RETRY); else /* if (ltype == LK_SHARED) */ vn_lock(dvp, LK_DOWNGRADE | LK_RETRY); /* * Relock for the "." case could leave us with * reclaimed vnode. */ if (dvp->v_iflag & VI_DOOMED) { vrele(dvp); return (SET_ERROR(ENOENT)); } } return (0); } else if (name[0] == '.' && name[1] == '.' && name[2] == 0) { /* * Note that in this case, dvp is the child vnode, and we * are looking up the parent vnode - exactly reverse from * normal operation. Unlocking dvp requires some rather * tricky unlock/relock dance to prevent mp from being freed; * use vn_vget_ino_gen() which takes care of all that. * * XXX Note that there is a time window when both vnodes are * unlocked. It is possible, although highly unlikely, that * during that window the parent-child relationship between * the vnodes may change, for example, get reversed. * In that case we would have a wrong lock order for the vnodes. * All other filesystems seem to ignore this problem, so we * do the same here. * A potential solution could be implemented as follows: * - using LK_NOWAIT when locking the second vnode and retrying * if necessary * - checking that the parent-child relationship still holds * after locking both vnodes and retrying if it doesn't */ error = vn_vget_ino_gen(dvp, zfs_dd_callback, vp, lkflags, &vp); return (error); } else { error = vn_lock(vp, lkflags); if (error != 0) vrele(vp); return (error); } } /* * Lookup an entry in a directory, or an extended attribute directory. * If it exists, return a held vnode reference for it. * * IN: dvp - vnode of directory to search. * nm - name of entry to lookup. * pnp - full pathname to lookup [UNUSED]. * flags - LOOKUP_XATTR set if looking for an attribute. * rdir - root directory vnode [UNUSED]. * cr - credentials of caller. * ct - caller context * * OUT: vpp - vnode of located entry, NULL if not found. * * RETURN: 0 on success, error code on failure. * * Timestamps: * NA */ /* ARGSUSED */ static int zfs_lookup(vnode_t *dvp, char *nm, vnode_t **vpp, struct componentname *cnp, int nameiop, cred_t *cr, kthread_t *td, int flags) { znode_t *zdp = VTOZ(dvp); znode_t *zp; zfsvfs_t *zfsvfs = zdp->z_zfsvfs; int error = 0; /* * Fast path lookup, however we must skip DNLC lookup * for case folding or normalizing lookups because the * DNLC code only stores the passed in name. This means * creating 'a' and removing 'A' on a case insensitive * file system would work, but DNLC still thinks 'a' * exists and won't let you create it again on the next * pass through fast path. */ if (!(flags & LOOKUP_XATTR)) { if (dvp->v_type != VDIR) { return (SET_ERROR(ENOTDIR)); } else if (zdp->z_sa_hdl == NULL) { return (SET_ERROR(EIO)); } } DTRACE_PROBE2(zfs__fastpath__lookup__miss, vnode_t *, dvp, char *, nm); ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zdp); *vpp = NULL; if (flags & LOOKUP_XATTR) { #ifdef TODO /* * If the xattr property is off, refuse the lookup request. */ if (!(zfsvfs->z_vfs->vfs_flag & VFS_XATTR)) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EINVAL)); } #endif /* * We don't allow recursive attributes.. * Maybe someday we will. */ if (zdp->z_pflags & ZFS_XATTR) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EINVAL)); } if (error = zfs_get_xattrdir(VTOZ(dvp), vpp, cr, flags)) { ZFS_EXIT(zfsvfs); return (error); } /* * Do we have permission to get into attribute directory? */ if (error = zfs_zaccess(VTOZ(*vpp), ACE_EXECUTE, 0, B_FALSE, cr)) { vrele(*vpp); *vpp = NULL; } ZFS_EXIT(zfsvfs); return (error); } /* * Check accessibility of directory. */ if (error = zfs_zaccess(zdp, ACE_EXECUTE, 0, B_FALSE, cr)) { ZFS_EXIT(zfsvfs); return (error); } if (zfsvfs->z_utf8 && u8_validate(nm, strlen(nm), NULL, U8_VALIDATE_ENTIRE, &error) < 0) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EILSEQ)); } /* * First handle the special cases. */ if ((cnp->cn_flags & ISDOTDOT) != 0) { /* * If we are a snapshot mounted under .zfs, return * the vp for the snapshot directory. */ if (zdp->z_id == zfsvfs->z_root && zfsvfs->z_parent != zfsvfs) { struct componentname cn; vnode_t *zfsctl_vp; int ltype; ZFS_EXIT(zfsvfs); ltype = VOP_ISLOCKED(dvp); VOP_UNLOCK(dvp, 0); error = zfsctl_root(zfsvfs->z_parent, LK_SHARED, &zfsctl_vp); if (error == 0) { cn.cn_nameptr = "snapshot"; cn.cn_namelen = strlen(cn.cn_nameptr); cn.cn_nameiop = cnp->cn_nameiop; cn.cn_flags = cnp->cn_flags & ~ISDOTDOT; cn.cn_lkflags = cnp->cn_lkflags; error = VOP_LOOKUP(zfsctl_vp, vpp, &cn); vput(zfsctl_vp); } vn_lock(dvp, ltype | LK_RETRY); return (error); } } if (zfs_has_ctldir(zdp) && strcmp(nm, ZFS_CTLDIR_NAME) == 0) { ZFS_EXIT(zfsvfs); if ((cnp->cn_flags & ISLASTCN) != 0 && nameiop != LOOKUP) return (SET_ERROR(ENOTSUP)); error = zfsctl_root(zfsvfs, cnp->cn_lkflags, vpp); return (error); } /* * The loop is retry the lookup if the parent-child relationship * changes during the dot-dot locking complexities. */ for (;;) { uint64_t parent; error = zfs_dirlook(zdp, nm, &zp); if (error == 0) *vpp = ZTOV(zp); ZFS_EXIT(zfsvfs); if (error != 0) break; error = zfs_lookup_lock(dvp, *vpp, nm, cnp->cn_lkflags); if (error != 0) { /* * If we've got a locking error, then the vnode * got reclaimed because of a force unmount. * We never enter doomed vnodes into the name cache. */ *vpp = NULL; return (error); } if ((cnp->cn_flags & ISDOTDOT) == 0) break; ZFS_ENTER(zfsvfs); if (zdp->z_sa_hdl == NULL) { error = SET_ERROR(EIO); } else { error = sa_lookup(zdp->z_sa_hdl, SA_ZPL_PARENT(zfsvfs), &parent, sizeof (parent)); } if (error != 0) { ZFS_EXIT(zfsvfs); vput(ZTOV(zp)); break; } if (zp->z_id == parent) { ZFS_EXIT(zfsvfs); break; } vput(ZTOV(zp)); } out: if (error != 0) *vpp = NULL; /* Translate errors and add SAVENAME when needed. */ if (cnp->cn_flags & ISLASTCN) { switch (nameiop) { case CREATE: case RENAME: if (error == ENOENT) { error = EJUSTRETURN; cnp->cn_flags |= SAVENAME; break; } /* FALLTHROUGH */ case DELETE: if (error == 0) cnp->cn_flags |= SAVENAME; break; } } /* Insert name into cache (as non-existent) if appropriate. */ if (zfsvfs->z_use_namecache && error == ENOENT && (cnp->cn_flags & MAKEENTRY) != 0) cache_enter(dvp, NULL, cnp); /* Insert name into cache if appropriate. */ if (zfsvfs->z_use_namecache && error == 0 && (cnp->cn_flags & MAKEENTRY)) { if (!(cnp->cn_flags & ISLASTCN) || (nameiop != DELETE && nameiop != RENAME)) { cache_enter(dvp, *vpp, cnp); } } return (error); } /* * Attempt to create a new entry in a directory. If the entry * already exists, truncate the file if permissible, else return * an error. Return the vp of the created or trunc'd file. * * IN: dvp - vnode of directory to put new file entry in. * name - name of new file entry. * vap - attributes of new file. * excl - flag indicating exclusive or non-exclusive mode. * mode - mode to open file with. * cr - credentials of caller. * flag - large file flag [UNUSED]. * ct - caller context * vsecp - ACL to be set * * OUT: vpp - vnode of created or trunc'd entry. * * RETURN: 0 on success, error code on failure. * * Timestamps: * dvp - ctime|mtime updated if new entry created * vp - ctime|mtime always, atime if new */ /* ARGSUSED */ static int zfs_create(vnode_t *dvp, char *name, vattr_t *vap, int excl, int mode, vnode_t **vpp, cred_t *cr, kthread_t *td) { znode_t *zp, *dzp = VTOZ(dvp); zfsvfs_t *zfsvfs = dzp->z_zfsvfs; zilog_t *zilog; objset_t *os; dmu_tx_t *tx; int error; ksid_t *ksid; uid_t uid; gid_t gid = crgetgid(cr); zfs_acl_ids_t acl_ids; boolean_t fuid_dirtied; void *vsecp = NULL; int flag = 0; uint64_t txtype; /* * If we have an ephemeral id, ACL, or XVATTR then * make sure file system is at proper version */ ksid = crgetsid(cr, KSID_OWNER); if (ksid) uid = ksid_getid(ksid); else uid = crgetuid(cr); if (zfsvfs->z_use_fuids == B_FALSE && (vsecp || (vap->va_mask & AT_XVATTR) || IS_EPHEMERAL(uid) || IS_EPHEMERAL(gid))) return (SET_ERROR(EINVAL)); ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(dzp); os = zfsvfs->z_os; zilog = zfsvfs->z_log; if (zfsvfs->z_utf8 && u8_validate(name, strlen(name), NULL, U8_VALIDATE_ENTIRE, &error) < 0) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EILSEQ)); } if (vap->va_mask & AT_XVATTR) { if ((error = secpolicy_xvattr(dvp, (xvattr_t *)vap, crgetuid(cr), cr, vap->va_type)) != 0) { ZFS_EXIT(zfsvfs); return (error); } } *vpp = NULL; if ((vap->va_mode & S_ISVTX) && secpolicy_vnode_stky_modify(cr)) vap->va_mode &= ~S_ISVTX; error = zfs_dirent_lookup(dzp, name, &zp, ZNEW); if (error) { ZFS_EXIT(zfsvfs); return (error); } ASSERT3P(zp, ==, NULL); /* * Create a new file object and update the directory * to reference it. */ if (error = zfs_zaccess(dzp, ACE_ADD_FILE, 0, B_FALSE, cr)) { goto out; } /* * We only support the creation of regular files in * extended attribute directories. */ if ((dzp->z_pflags & ZFS_XATTR) && (vap->va_type != VREG)) { error = SET_ERROR(EINVAL); goto out; } if ((error = zfs_acl_ids_create(dzp, 0, vap, cr, vsecp, &acl_ids)) != 0) goto out; if (zfs_acl_ids_overquota(zfsvfs, &acl_ids)) { zfs_acl_ids_free(&acl_ids); error = SET_ERROR(EDQUOT); goto out; } getnewvnode_reserve(1); tx = dmu_tx_create(os); dmu_tx_hold_sa_create(tx, acl_ids.z_aclp->z_acl_bytes + ZFS_SA_BASE_ATTR_SIZE); fuid_dirtied = zfsvfs->z_fuid_dirty; if (fuid_dirtied) zfs_fuid_txhold(zfsvfs, tx); dmu_tx_hold_zap(tx, dzp->z_id, TRUE, name); dmu_tx_hold_sa(tx, dzp->z_sa_hdl, B_FALSE); if (!zfsvfs->z_use_sa && acl_ids.z_aclp->z_acl_bytes > ZFS_ACE_SPACE) { dmu_tx_hold_write(tx, DMU_NEW_OBJECT, 0, acl_ids.z_aclp->z_acl_bytes); } error = dmu_tx_assign(tx, TXG_WAIT); if (error) { zfs_acl_ids_free(&acl_ids); dmu_tx_abort(tx); getnewvnode_drop_reserve(); ZFS_EXIT(zfsvfs); return (error); } zfs_mknode(dzp, vap, tx, cr, 0, &zp, &acl_ids); if (fuid_dirtied) zfs_fuid_sync(zfsvfs, tx); (void) zfs_link_create(dzp, name, zp, tx, ZNEW); txtype = zfs_log_create_txtype(Z_FILE, vsecp, vap); zfs_log_create(zilog, tx, txtype, dzp, zp, name, vsecp, acl_ids.z_fuidp, vap); zfs_acl_ids_free(&acl_ids); dmu_tx_commit(tx); getnewvnode_drop_reserve(); out: if (error == 0) { *vpp = ZTOV(zp); } if (zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS) zil_commit(zilog, 0); ZFS_EXIT(zfsvfs); return (error); } /* * Remove an entry from a directory. * * IN: dvp - vnode of directory to remove entry from. * name - name of entry to remove. * cr - credentials of caller. * ct - caller context * flags - case flags * * RETURN: 0 on success, error code on failure. * * Timestamps: * dvp - ctime|mtime * vp - ctime (if nlink > 0) */ /*ARGSUSED*/ static int zfs_remove(vnode_t *dvp, vnode_t *vp, char *name, cred_t *cr) { znode_t *dzp = VTOZ(dvp); znode_t *zp = VTOZ(vp); znode_t *xzp; zfsvfs_t *zfsvfs = dzp->z_zfsvfs; zilog_t *zilog; uint64_t acl_obj, xattr_obj; uint64_t obj = 0; dmu_tx_t *tx; boolean_t unlinked, toobig = FALSE; uint64_t txtype; int error; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(dzp); ZFS_VERIFY_ZP(zp); zilog = zfsvfs->z_log; zp = VTOZ(vp); xattr_obj = 0; xzp = NULL; if (error = zfs_zaccess_delete(dzp, zp, cr)) { goto out; } /* * Need to use rmdir for removing directories. */ if (vp->v_type == VDIR) { error = SET_ERROR(EPERM); goto out; } vnevent_remove(vp, dvp, name, ct); obj = zp->z_id; /* are there any extended attributes? */ error = sa_lookup(zp->z_sa_hdl, SA_ZPL_XATTR(zfsvfs), &xattr_obj, sizeof (xattr_obj)); if (error == 0 && xattr_obj) { error = zfs_zget(zfsvfs, xattr_obj, &xzp); ASSERT0(error); } /* * We may delete the znode now, or we may put it in the unlinked set; * it depends on whether we're the last link, and on whether there are * other holds on the vnode. So we dmu_tx_hold() the right things to * allow for either case. */ tx = dmu_tx_create(zfsvfs->z_os); dmu_tx_hold_zap(tx, dzp->z_id, FALSE, name); dmu_tx_hold_sa(tx, zp->z_sa_hdl, B_FALSE); zfs_sa_upgrade_txholds(tx, zp); zfs_sa_upgrade_txholds(tx, dzp); if (xzp) { dmu_tx_hold_sa(tx, zp->z_sa_hdl, B_TRUE); dmu_tx_hold_sa(tx, xzp->z_sa_hdl, B_FALSE); } /* charge as an update -- would be nice not to charge at all */ dmu_tx_hold_zap(tx, zfsvfs->z_unlinkedobj, FALSE, NULL); /* * Mark this transaction as typically resulting in a net free of space */ dmu_tx_mark_netfree(tx); error = dmu_tx_assign(tx, TXG_WAIT); if (error) { dmu_tx_abort(tx); ZFS_EXIT(zfsvfs); return (error); } /* * Remove the directory entry. */ error = zfs_link_destroy(dzp, name, zp, tx, ZEXISTS, &unlinked); if (error) { dmu_tx_commit(tx); goto out; } if (unlinked) { zfs_unlinked_add(zp, tx); vp->v_vflag |= VV_NOSYNC; } txtype = TX_REMOVE; zfs_log_remove(zilog, tx, txtype, dzp, name, obj); dmu_tx_commit(tx); out: if (xzp) vrele(ZTOV(xzp)); if (zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS) zil_commit(zilog, 0); ZFS_EXIT(zfsvfs); return (error); } /* * Create a new directory and insert it into dvp using the name * provided. Return a pointer to the inserted directory. * * IN: dvp - vnode of directory to add subdir to. * dirname - name of new directory. * vap - attributes of new directory. * cr - credentials of caller. * ct - caller context * flags - case flags * vsecp - ACL to be set * * OUT: vpp - vnode of created directory. * * RETURN: 0 on success, error code on failure. * * Timestamps: * dvp - ctime|mtime updated * vp - ctime|mtime|atime updated */ /*ARGSUSED*/ static int zfs_mkdir(vnode_t *dvp, char *dirname, vattr_t *vap, vnode_t **vpp, cred_t *cr) { znode_t *zp, *dzp = VTOZ(dvp); zfsvfs_t *zfsvfs = dzp->z_zfsvfs; zilog_t *zilog; uint64_t txtype; dmu_tx_t *tx; int error; ksid_t *ksid; uid_t uid; gid_t gid = crgetgid(cr); zfs_acl_ids_t acl_ids; boolean_t fuid_dirtied; ASSERT(vap->va_type == VDIR); /* * If we have an ephemeral id, ACL, or XVATTR then * make sure file system is at proper version */ ksid = crgetsid(cr, KSID_OWNER); if (ksid) uid = ksid_getid(ksid); else uid = crgetuid(cr); if (zfsvfs->z_use_fuids == B_FALSE && ((vap->va_mask & AT_XVATTR) || IS_EPHEMERAL(uid) || IS_EPHEMERAL(gid))) return (SET_ERROR(EINVAL)); ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(dzp); zilog = zfsvfs->z_log; if (dzp->z_pflags & ZFS_XATTR) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EINVAL)); } if (zfsvfs->z_utf8 && u8_validate(dirname, strlen(dirname), NULL, U8_VALIDATE_ENTIRE, &error) < 0) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EILSEQ)); } if (vap->va_mask & AT_XVATTR) { if ((error = secpolicy_xvattr(dvp, (xvattr_t *)vap, crgetuid(cr), cr, vap->va_type)) != 0) { ZFS_EXIT(zfsvfs); return (error); } } if ((error = zfs_acl_ids_create(dzp, 0, vap, cr, NULL, &acl_ids)) != 0) { ZFS_EXIT(zfsvfs); return (error); } /* * First make sure the new directory doesn't exist. * * Existence is checked first to make sure we don't return * EACCES instead of EEXIST which can cause some applications * to fail. */ *vpp = NULL; if (error = zfs_dirent_lookup(dzp, dirname, &zp, ZNEW)) { zfs_acl_ids_free(&acl_ids); ZFS_EXIT(zfsvfs); return (error); } ASSERT3P(zp, ==, NULL); if (error = zfs_zaccess(dzp, ACE_ADD_SUBDIRECTORY, 0, B_FALSE, cr)) { zfs_acl_ids_free(&acl_ids); ZFS_EXIT(zfsvfs); return (error); } if (zfs_acl_ids_overquota(zfsvfs, &acl_ids)) { zfs_acl_ids_free(&acl_ids); ZFS_EXIT(zfsvfs); return (SET_ERROR(EDQUOT)); } /* * Add a new entry to the directory. */ getnewvnode_reserve(1); tx = dmu_tx_create(zfsvfs->z_os); dmu_tx_hold_zap(tx, dzp->z_id, TRUE, dirname); dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, FALSE, NULL); fuid_dirtied = zfsvfs->z_fuid_dirty; if (fuid_dirtied) zfs_fuid_txhold(zfsvfs, tx); if (!zfsvfs->z_use_sa && acl_ids.z_aclp->z_acl_bytes > ZFS_ACE_SPACE) { dmu_tx_hold_write(tx, DMU_NEW_OBJECT, 0, acl_ids.z_aclp->z_acl_bytes); } dmu_tx_hold_sa_create(tx, acl_ids.z_aclp->z_acl_bytes + ZFS_SA_BASE_ATTR_SIZE); error = dmu_tx_assign(tx, TXG_WAIT); if (error) { zfs_acl_ids_free(&acl_ids); dmu_tx_abort(tx); getnewvnode_drop_reserve(); ZFS_EXIT(zfsvfs); return (error); } /* * Create new node. */ zfs_mknode(dzp, vap, tx, cr, 0, &zp, &acl_ids); if (fuid_dirtied) zfs_fuid_sync(zfsvfs, tx); /* * Now put new name in parent dir. */ (void) zfs_link_create(dzp, dirname, zp, tx, ZNEW); *vpp = ZTOV(zp); txtype = zfs_log_create_txtype(Z_DIR, NULL, vap); zfs_log_create(zilog, tx, txtype, dzp, zp, dirname, NULL, acl_ids.z_fuidp, vap); zfs_acl_ids_free(&acl_ids); dmu_tx_commit(tx); getnewvnode_drop_reserve(); if (zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS) zil_commit(zilog, 0); ZFS_EXIT(zfsvfs); return (0); } /* * Remove a directory subdir entry. If the current working * directory is the same as the subdir to be removed, the * remove will fail. * * IN: dvp - vnode of directory to remove from. * name - name of directory to be removed. * cwd - vnode of current working directory. * cr - credentials of caller. * ct - caller context * flags - case flags * * RETURN: 0 on success, error code on failure. * * Timestamps: * dvp - ctime|mtime updated */ /*ARGSUSED*/ static int zfs_rmdir(vnode_t *dvp, vnode_t *vp, char *name, cred_t *cr) { znode_t *dzp = VTOZ(dvp); znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = dzp->z_zfsvfs; zilog_t *zilog; dmu_tx_t *tx; int error; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(dzp); ZFS_VERIFY_ZP(zp); zilog = zfsvfs->z_log; if (error = zfs_zaccess_delete(dzp, zp, cr)) { goto out; } if (vp->v_type != VDIR) { error = SET_ERROR(ENOTDIR); goto out; } vnevent_rmdir(vp, dvp, name, ct); tx = dmu_tx_create(zfsvfs->z_os); dmu_tx_hold_zap(tx, dzp->z_id, FALSE, name); dmu_tx_hold_sa(tx, zp->z_sa_hdl, B_FALSE); dmu_tx_hold_zap(tx, zfsvfs->z_unlinkedobj, FALSE, NULL); zfs_sa_upgrade_txholds(tx, zp); zfs_sa_upgrade_txholds(tx, dzp); dmu_tx_mark_netfree(tx); error = dmu_tx_assign(tx, TXG_WAIT); if (error) { dmu_tx_abort(tx); ZFS_EXIT(zfsvfs); return (error); } cache_purge(dvp); error = zfs_link_destroy(dzp, name, zp, tx, ZEXISTS, NULL); if (error == 0) { uint64_t txtype = TX_RMDIR; zfs_log_remove(zilog, tx, txtype, dzp, name, ZFS_NO_OBJECT); } dmu_tx_commit(tx); cache_purge(vp); out: if (zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS) zil_commit(zilog, 0); ZFS_EXIT(zfsvfs); return (error); } /* * Read as many directory entries as will fit into the provided * buffer from the given directory cursor position (specified in * the uio structure). * * IN: vp - vnode of directory to read. * uio - structure supplying read location, range info, * and return buffer. * cr - credentials of caller. * ct - caller context * flags - case flags * * OUT: uio - updated offset and range, buffer filled. * eofp - set to true if end-of-file detected. * * RETURN: 0 on success, error code on failure. * * Timestamps: * vp - atime updated * * Note that the low 4 bits of the cookie returned by zap is always zero. * This allows us to use the low range for "special" directory entries: * We use 0 for '.', and 1 for '..'. If this is the root of the filesystem, * we use the offset 2 for the '.zfs' directory. */ /* ARGSUSED */ static int zfs_readdir(vnode_t *vp, uio_t *uio, cred_t *cr, int *eofp, int *ncookies, u_long **cookies) { znode_t *zp = VTOZ(vp); iovec_t *iovp; edirent_t *eodp; dirent64_t *odp; zfsvfs_t *zfsvfs = zp->z_zfsvfs; objset_t *os; caddr_t outbuf; size_t bufsize; zap_cursor_t zc; zap_attribute_t zap; uint_t bytes_wanted; uint64_t offset; /* must be unsigned; checks for < 1 */ uint64_t parent; int local_eof; int outcount; int error; uint8_t prefetch; boolean_t check_sysattrs; uint8_t type; int ncooks; u_long *cooks = NULL; int flags = 0; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); if ((error = sa_lookup(zp->z_sa_hdl, SA_ZPL_PARENT(zfsvfs), &parent, sizeof (parent))) != 0) { ZFS_EXIT(zfsvfs); return (error); } /* * If we are not given an eof variable, * use a local one. */ if (eofp == NULL) eofp = &local_eof; /* * Check for valid iov_len. */ if (uio->uio_iov->iov_len <= 0) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EINVAL)); } /* * Quit if directory has been removed (posix) */ if ((*eofp = zp->z_unlinked) != 0) { ZFS_EXIT(zfsvfs); return (0); } error = 0; os = zfsvfs->z_os; offset = uio->uio_loffset; prefetch = zp->z_zn_prefetch; /* * Initialize the iterator cursor. */ if (offset <= 3) { /* * Start iteration from the beginning of the directory. */ zap_cursor_init(&zc, os, zp->z_id); } else { /* * The offset is a serialized cursor. */ zap_cursor_init_serialized(&zc, os, zp->z_id, offset); } /* * Get space to change directory entries into fs independent format. */ iovp = uio->uio_iov; bytes_wanted = iovp->iov_len; if (uio->uio_segflg != UIO_SYSSPACE || uio->uio_iovcnt != 1) { bufsize = bytes_wanted; outbuf = kmem_alloc(bufsize, KM_SLEEP); odp = (struct dirent64 *)outbuf; } else { bufsize = bytes_wanted; outbuf = NULL; odp = (struct dirent64 *)iovp->iov_base; } eodp = (struct edirent *)odp; if (ncookies != NULL) { /* * Minimum entry size is dirent size and 1 byte for a file name. */ ncooks = uio->uio_resid / (sizeof(struct dirent) - sizeof(((struct dirent *)NULL)->d_name) + 1); cooks = malloc(ncooks * sizeof(u_long), M_TEMP, M_WAITOK); *cookies = cooks; *ncookies = ncooks; } /* * If this VFS supports the system attribute view interface; and * we're looking at an extended attribute directory; and we care * about normalization conflicts on this vfs; then we must check * for normalization conflicts with the sysattr name space. */ #ifdef TODO check_sysattrs = vfs_has_feature(vp->v_vfsp, VFSFT_SYSATTR_VIEWS) && (vp->v_flag & V_XATTRDIR) && zfsvfs->z_norm && (flags & V_RDDIR_ENTFLAGS); #else check_sysattrs = 0; #endif /* * Transform to file-system independent format */ outcount = 0; while (outcount < bytes_wanted) { ino64_t objnum; ushort_t reclen; off64_t *next = NULL; /* * Special case `.', `..', and `.zfs'. */ if (offset == 0) { (void) strcpy(zap.za_name, "."); zap.za_normalization_conflict = 0; objnum = zp->z_id; type = DT_DIR; } else if (offset == 1) { (void) strcpy(zap.za_name, ".."); zap.za_normalization_conflict = 0; objnum = parent; type = DT_DIR; } else if (offset == 2 && zfs_show_ctldir(zp)) { (void) strcpy(zap.za_name, ZFS_CTLDIR_NAME); zap.za_normalization_conflict = 0; objnum = ZFSCTL_INO_ROOT; type = DT_DIR; } else { /* * Grab next entry. */ if (error = zap_cursor_retrieve(&zc, &zap)) { if ((*eofp = (error == ENOENT)) != 0) break; else goto update; } if (zap.za_integer_length != 8 || zap.za_num_integers != 1) { cmn_err(CE_WARN, "zap_readdir: bad directory " "entry, obj = %lld, offset = %lld\n", (u_longlong_t)zp->z_id, (u_longlong_t)offset); error = SET_ERROR(ENXIO); goto update; } objnum = ZFS_DIRENT_OBJ(zap.za_first_integer); /* * MacOS X can extract the object type here such as: * uint8_t type = ZFS_DIRENT_TYPE(zap.za_first_integer); */ type = ZFS_DIRENT_TYPE(zap.za_first_integer); if (check_sysattrs && !zap.za_normalization_conflict) { #ifdef TODO zap.za_normalization_conflict = xattr_sysattr_casechk(zap.za_name); #else panic("%s:%u: TODO", __func__, __LINE__); #endif } } if (flags & V_RDDIR_ACCFILTER) { /* * If we have no access at all, don't include * this entry in the returned information */ znode_t *ezp; if (zfs_zget(zp->z_zfsvfs, objnum, &ezp) != 0) goto skip_entry; if (!zfs_has_access(ezp, cr)) { vrele(ZTOV(ezp)); goto skip_entry; } vrele(ZTOV(ezp)); } if (flags & V_RDDIR_ENTFLAGS) reclen = EDIRENT_RECLEN(strlen(zap.za_name)); else reclen = DIRENT64_RECLEN(strlen(zap.za_name)); /* * Will this entry fit in the buffer? */ if (outcount + reclen > bufsize) { /* * Did we manage to fit anything in the buffer? */ if (!outcount) { error = SET_ERROR(EINVAL); goto update; } break; } if (flags & V_RDDIR_ENTFLAGS) { /* * Add extended flag entry: */ eodp->ed_ino = objnum; eodp->ed_reclen = reclen; /* NOTE: ed_off is the offset for the *next* entry. */ next = &eodp->ed_off; eodp->ed_eflags = zap.za_normalization_conflict ? ED_CASE_CONFLICT : 0; (void) strncpy(eodp->ed_name, zap.za_name, EDIRENT_NAMELEN(reclen)); eodp = (edirent_t *)((intptr_t)eodp + reclen); } else { /* * Add normal entry: */ odp->d_ino = objnum; odp->d_reclen = reclen; odp->d_namlen = strlen(zap.za_name); /* NOTE: d_off is the offset for the *next* entry. */ next = &odp->d_off; (void) strlcpy(odp->d_name, zap.za_name, odp->d_namlen + 1); odp->d_type = type; dirent_terminate(odp); odp = (dirent64_t *)((intptr_t)odp + reclen); } outcount += reclen; ASSERT(outcount <= bufsize); /* Prefetch znode */ if (prefetch) dmu_prefetch(os, objnum, 0, 0, 0, ZIO_PRIORITY_SYNC_READ); skip_entry: /* * Move to the next entry, fill in the previous offset. */ if (offset > 2 || (offset == 2 && !zfs_show_ctldir(zp))) { zap_cursor_advance(&zc); offset = zap_cursor_serialize(&zc); } else { offset += 1; } /* Fill the offset right after advancing the cursor. */ if (next != NULL) *next = offset; if (cooks != NULL) { *cooks++ = offset; ncooks--; KASSERT(ncooks >= 0, ("ncookies=%d", ncooks)); } } zp->z_zn_prefetch = B_FALSE; /* a lookup will re-enable pre-fetching */ /* Subtract unused cookies */ if (ncookies != NULL) *ncookies -= ncooks; if (uio->uio_segflg == UIO_SYSSPACE && uio->uio_iovcnt == 1) { iovp->iov_base += outcount; iovp->iov_len -= outcount; uio->uio_resid -= outcount; } else if (error = uiomove(outbuf, (long)outcount, UIO_READ, uio)) { /* * Reset the pointer. */ offset = uio->uio_loffset; } update: zap_cursor_fini(&zc); if (uio->uio_segflg != UIO_SYSSPACE || uio->uio_iovcnt != 1) kmem_free(outbuf, bufsize); if (error == ENOENT) error = 0; ZFS_ACCESSTIME_STAMP(zfsvfs, zp); uio->uio_loffset = offset; ZFS_EXIT(zfsvfs); if (error != 0 && cookies != NULL) { free(*cookies, M_TEMP); *cookies = NULL; *ncookies = 0; } return (error); } ulong_t zfs_fsync_sync_cnt = 4; static int zfs_fsync(vnode_t *vp, int syncflag, cred_t *cr, caller_context_t *ct) { znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; (void) tsd_set(zfs_fsyncer_key, (void *)zfs_fsync_sync_cnt); if (zfsvfs->z_os->os_sync != ZFS_SYNC_DISABLED) { ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); zil_commit(zfsvfs->z_log, zp->z_id); ZFS_EXIT(zfsvfs); } return (0); } /* * Get the requested file attributes and place them in the provided * vattr structure. * * IN: vp - vnode of file. * vap - va_mask identifies requested attributes. * If AT_XVATTR set, then optional attrs are requested * flags - ATTR_NOACLCHECK (CIFS server context) * cr - credentials of caller. * ct - caller context * * OUT: vap - attribute values. * * RETURN: 0 (always succeeds). */ /* ARGSUSED */ static int zfs_getattr(vnode_t *vp, vattr_t *vap, int flags, cred_t *cr, caller_context_t *ct) { znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; int error = 0; uint32_t blksize; u_longlong_t nblocks; uint64_t mtime[2], ctime[2], crtime[2], rdev; xvattr_t *xvap = (xvattr_t *)vap; /* vap may be an xvattr_t * */ xoptattr_t *xoap = NULL; boolean_t skipaclchk = (flags & ATTR_NOACLCHECK) ? B_TRUE : B_FALSE; sa_bulk_attr_t bulk[4]; int count = 0; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); zfs_fuid_map_ids(zp, cr, &vap->va_uid, &vap->va_gid); SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_MTIME(zfsvfs), NULL, &mtime, 16); SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_CTIME(zfsvfs), NULL, &ctime, 16); SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_CRTIME(zfsvfs), NULL, &crtime, 16); if (vp->v_type == VBLK || vp->v_type == VCHR) SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_RDEV(zfsvfs), NULL, &rdev, 8); if ((error = sa_bulk_lookup(zp->z_sa_hdl, bulk, count)) != 0) { ZFS_EXIT(zfsvfs); return (error); } /* * If ACL is trivial don't bother looking for ACE_READ_ATTRIBUTES. * Also, if we are the owner don't bother, since owner should * always be allowed to read basic attributes of file. */ if (!(zp->z_pflags & ZFS_ACL_TRIVIAL) && (vap->va_uid != crgetuid(cr))) { if (error = zfs_zaccess(zp, ACE_READ_ATTRIBUTES, 0, skipaclchk, cr)) { ZFS_EXIT(zfsvfs); return (error); } } /* * Return all attributes. It's cheaper to provide the answer * than to determine whether we were asked the question. */ vap->va_type = IFTOVT(zp->z_mode); vap->va_mode = zp->z_mode & ~S_IFMT; #ifdef illumos vap->va_fsid = zp->z_zfsvfs->z_vfs->vfs_dev; #else vn_fsid(vp, vap); #endif vap->va_nodeid = zp->z_id; vap->va_nlink = zp->z_links; if ((vp->v_flag & VROOT) && zfs_show_ctldir(zp) && zp->z_links < ZFS_LINK_MAX) vap->va_nlink++; vap->va_size = zp->z_size; #ifdef illumos vap->va_rdev = vp->v_rdev; #else if (vp->v_type == VBLK || vp->v_type == VCHR) vap->va_rdev = zfs_cmpldev(rdev); #endif vap->va_seq = zp->z_seq; vap->va_flags = 0; /* FreeBSD: Reset chflags(2) flags. */ vap->va_filerev = zp->z_seq; /* * Add in any requested optional attributes and the create time. * Also set the corresponding bits in the returned attribute bitmap. */ if ((xoap = xva_getxoptattr(xvap)) != NULL && zfsvfs->z_use_fuids) { if (XVA_ISSET_REQ(xvap, XAT_ARCHIVE)) { xoap->xoa_archive = ((zp->z_pflags & ZFS_ARCHIVE) != 0); XVA_SET_RTN(xvap, XAT_ARCHIVE); } if (XVA_ISSET_REQ(xvap, XAT_READONLY)) { xoap->xoa_readonly = ((zp->z_pflags & ZFS_READONLY) != 0); XVA_SET_RTN(xvap, XAT_READONLY); } if (XVA_ISSET_REQ(xvap, XAT_SYSTEM)) { xoap->xoa_system = ((zp->z_pflags & ZFS_SYSTEM) != 0); XVA_SET_RTN(xvap, XAT_SYSTEM); } if (XVA_ISSET_REQ(xvap, XAT_HIDDEN)) { xoap->xoa_hidden = ((zp->z_pflags & ZFS_HIDDEN) != 0); XVA_SET_RTN(xvap, XAT_HIDDEN); } if (XVA_ISSET_REQ(xvap, XAT_NOUNLINK)) { xoap->xoa_nounlink = ((zp->z_pflags & ZFS_NOUNLINK) != 0); XVA_SET_RTN(xvap, XAT_NOUNLINK); } if (XVA_ISSET_REQ(xvap, XAT_IMMUTABLE)) { xoap->xoa_immutable = ((zp->z_pflags & ZFS_IMMUTABLE) != 0); XVA_SET_RTN(xvap, XAT_IMMUTABLE); } if (XVA_ISSET_REQ(xvap, XAT_APPENDONLY)) { xoap->xoa_appendonly = ((zp->z_pflags & ZFS_APPENDONLY) != 0); XVA_SET_RTN(xvap, XAT_APPENDONLY); } if (XVA_ISSET_REQ(xvap, XAT_NODUMP)) { xoap->xoa_nodump = ((zp->z_pflags & ZFS_NODUMP) != 0); XVA_SET_RTN(xvap, XAT_NODUMP); } if (XVA_ISSET_REQ(xvap, XAT_OPAQUE)) { xoap->xoa_opaque = ((zp->z_pflags & ZFS_OPAQUE) != 0); XVA_SET_RTN(xvap, XAT_OPAQUE); } if (XVA_ISSET_REQ(xvap, XAT_AV_QUARANTINED)) { xoap->xoa_av_quarantined = ((zp->z_pflags & ZFS_AV_QUARANTINED) != 0); XVA_SET_RTN(xvap, XAT_AV_QUARANTINED); } if (XVA_ISSET_REQ(xvap, XAT_AV_MODIFIED)) { xoap->xoa_av_modified = ((zp->z_pflags & ZFS_AV_MODIFIED) != 0); XVA_SET_RTN(xvap, XAT_AV_MODIFIED); } if (XVA_ISSET_REQ(xvap, XAT_AV_SCANSTAMP) && vp->v_type == VREG) { zfs_sa_get_scanstamp(zp, xvap); } if (XVA_ISSET_REQ(xvap, XAT_REPARSE)) { xoap->xoa_reparse = ((zp->z_pflags & ZFS_REPARSE) != 0); XVA_SET_RTN(xvap, XAT_REPARSE); } if (XVA_ISSET_REQ(xvap, XAT_GEN)) { xoap->xoa_generation = zp->z_gen; XVA_SET_RTN(xvap, XAT_GEN); } if (XVA_ISSET_REQ(xvap, XAT_OFFLINE)) { xoap->xoa_offline = ((zp->z_pflags & ZFS_OFFLINE) != 0); XVA_SET_RTN(xvap, XAT_OFFLINE); } if (XVA_ISSET_REQ(xvap, XAT_SPARSE)) { xoap->xoa_sparse = ((zp->z_pflags & ZFS_SPARSE) != 0); XVA_SET_RTN(xvap, XAT_SPARSE); } } ZFS_TIME_DECODE(&vap->va_atime, zp->z_atime); ZFS_TIME_DECODE(&vap->va_mtime, mtime); ZFS_TIME_DECODE(&vap->va_ctime, ctime); ZFS_TIME_DECODE(&vap->va_birthtime, crtime); sa_object_size(zp->z_sa_hdl, &blksize, &nblocks); vap->va_blksize = blksize; vap->va_bytes = nblocks << 9; /* nblocks * 512 */ if (zp->z_blksz == 0) { /* * Block size hasn't been set; suggest maximal I/O transfers. */ vap->va_blksize = zfsvfs->z_max_blksz; } ZFS_EXIT(zfsvfs); return (0); } /* * Set the file attributes to the values contained in the * vattr structure. * * IN: vp - vnode of file to be modified. * vap - new attribute values. * If AT_XVATTR set, then optional attrs are being set * flags - ATTR_UTIME set if non-default time values provided. * - ATTR_NOACLCHECK (CIFS context only). * cr - credentials of caller. * ct - caller context * * RETURN: 0 on success, error code on failure. * * Timestamps: * vp - ctime updated, mtime updated if size changed. */ /* ARGSUSED */ static int zfs_setattr(vnode_t *vp, vattr_t *vap, int flags, cred_t *cr, caller_context_t *ct) { znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; zilog_t *zilog; dmu_tx_t *tx; vattr_t oldva; xvattr_t tmpxvattr; uint_t mask = vap->va_mask; uint_t saved_mask = 0; uint64_t saved_mode; int trim_mask = 0; uint64_t new_mode; uint64_t new_uid, new_gid; uint64_t xattr_obj; uint64_t mtime[2], ctime[2]; znode_t *attrzp; int need_policy = FALSE; int err, err2; zfs_fuid_info_t *fuidp = NULL; xvattr_t *xvap = (xvattr_t *)vap; /* vap may be an xvattr_t * */ xoptattr_t *xoap; zfs_acl_t *aclp; boolean_t skipaclchk = (flags & ATTR_NOACLCHECK) ? B_TRUE : B_FALSE; boolean_t fuid_dirtied = B_FALSE; sa_bulk_attr_t bulk[7], xattr_bulk[7]; int count = 0, xattr_count = 0; if (mask == 0) return (0); if (mask & AT_NOSET) return (SET_ERROR(EINVAL)); ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); zilog = zfsvfs->z_log; /* * Make sure that if we have ephemeral uid/gid or xvattr specified * that file system is at proper version level */ if (zfsvfs->z_use_fuids == B_FALSE && (((mask & AT_UID) && IS_EPHEMERAL(vap->va_uid)) || ((mask & AT_GID) && IS_EPHEMERAL(vap->va_gid)) || (mask & AT_XVATTR))) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EINVAL)); } if (mask & AT_SIZE && vp->v_type == VDIR) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EISDIR)); } if (mask & AT_SIZE && vp->v_type != VREG && vp->v_type != VFIFO) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EINVAL)); } /* * If this is an xvattr_t, then get a pointer to the structure of * optional attributes. If this is NULL, then we have a vattr_t. */ xoap = xva_getxoptattr(xvap); xva_init(&tmpxvattr); /* * Immutable files can only alter immutable bit and atime */ if ((zp->z_pflags & ZFS_IMMUTABLE) && ((mask & (AT_SIZE|AT_UID|AT_GID|AT_MTIME|AT_MODE)) || ((mask & AT_XVATTR) && XVA_ISSET_REQ(xvap, XAT_CREATETIME)))) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EPERM)); } /* * Note: ZFS_READONLY is handled in zfs_zaccess_common. */ /* * Verify timestamps doesn't overflow 32 bits. * ZFS can handle large timestamps, but 32bit syscalls can't * handle times greater than 2039. This check should be removed * once large timestamps are fully supported. */ if (mask & (AT_ATIME | AT_MTIME)) { if (((mask & AT_ATIME) && TIMESPEC_OVERFLOW(&vap->va_atime)) || ((mask & AT_MTIME) && TIMESPEC_OVERFLOW(&vap->va_mtime))) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EOVERFLOW)); } } if (xoap && (mask & AT_XVATTR) && XVA_ISSET_REQ(xvap, XAT_CREATETIME) && TIMESPEC_OVERFLOW(&vap->va_birthtime)) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EOVERFLOW)); } attrzp = NULL; aclp = NULL; /* Can this be moved to before the top label? */ if (zfsvfs->z_vfs->vfs_flag & VFS_RDONLY) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EROFS)); } /* * First validate permissions */ if (mask & AT_SIZE) { /* * XXX - Note, we are not providing any open * mode flags here (like FNDELAY), so we may * block if there are locks present... this * should be addressed in openat(). */ /* XXX - would it be OK to generate a log record here? */ err = zfs_freesp(zp, vap->va_size, 0, 0, FALSE); if (err) { ZFS_EXIT(zfsvfs); return (err); } } if (mask & (AT_ATIME|AT_MTIME) || ((mask & AT_XVATTR) && (XVA_ISSET_REQ(xvap, XAT_HIDDEN) || XVA_ISSET_REQ(xvap, XAT_READONLY) || XVA_ISSET_REQ(xvap, XAT_ARCHIVE) || XVA_ISSET_REQ(xvap, XAT_OFFLINE) || XVA_ISSET_REQ(xvap, XAT_SPARSE) || XVA_ISSET_REQ(xvap, XAT_CREATETIME) || XVA_ISSET_REQ(xvap, XAT_SYSTEM)))) { need_policy = zfs_zaccess(zp, ACE_WRITE_ATTRIBUTES, 0, skipaclchk, cr); } if (mask & (AT_UID|AT_GID)) { int idmask = (mask & (AT_UID|AT_GID)); int take_owner; int take_group; /* * NOTE: even if a new mode is being set, * we may clear S_ISUID/S_ISGID bits. */ if (!(mask & AT_MODE)) vap->va_mode = zp->z_mode; /* * Take ownership or chgrp to group we are a member of */ take_owner = (mask & AT_UID) && (vap->va_uid == crgetuid(cr)); take_group = (mask & AT_GID) && zfs_groupmember(zfsvfs, vap->va_gid, cr); /* * If both AT_UID and AT_GID are set then take_owner and * take_group must both be set in order to allow taking * ownership. * * Otherwise, send the check through secpolicy_vnode_setattr() * */ if (((idmask == (AT_UID|AT_GID)) && take_owner && take_group) || ((idmask == AT_UID) && take_owner) || ((idmask == AT_GID) && take_group)) { if (zfs_zaccess(zp, ACE_WRITE_OWNER, 0, skipaclchk, cr) == 0) { /* * Remove setuid/setgid for non-privileged users */ secpolicy_setid_clear(vap, vp, cr); trim_mask = (mask & (AT_UID|AT_GID)); } else { need_policy = TRUE; } } else { need_policy = TRUE; } } oldva.va_mode = zp->z_mode; zfs_fuid_map_ids(zp, cr, &oldva.va_uid, &oldva.va_gid); if (mask & AT_XVATTR) { /* * Update xvattr mask to include only those attributes * that are actually changing. * * the bits will be restored prior to actually setting * the attributes so the caller thinks they were set. */ if (XVA_ISSET_REQ(xvap, XAT_APPENDONLY)) { if (xoap->xoa_appendonly != ((zp->z_pflags & ZFS_APPENDONLY) != 0)) { need_policy = TRUE; } else { XVA_CLR_REQ(xvap, XAT_APPENDONLY); XVA_SET_REQ(&tmpxvattr, XAT_APPENDONLY); } } if (XVA_ISSET_REQ(xvap, XAT_NOUNLINK)) { if (xoap->xoa_nounlink != ((zp->z_pflags & ZFS_NOUNLINK) != 0)) { need_policy = TRUE; } else { XVA_CLR_REQ(xvap, XAT_NOUNLINK); XVA_SET_REQ(&tmpxvattr, XAT_NOUNLINK); } } if (XVA_ISSET_REQ(xvap, XAT_IMMUTABLE)) { if (xoap->xoa_immutable != ((zp->z_pflags & ZFS_IMMUTABLE) != 0)) { need_policy = TRUE; } else { XVA_CLR_REQ(xvap, XAT_IMMUTABLE); XVA_SET_REQ(&tmpxvattr, XAT_IMMUTABLE); } } if (XVA_ISSET_REQ(xvap, XAT_NODUMP)) { if (xoap->xoa_nodump != ((zp->z_pflags & ZFS_NODUMP) != 0)) { need_policy = TRUE; } else { XVA_CLR_REQ(xvap, XAT_NODUMP); XVA_SET_REQ(&tmpxvattr, XAT_NODUMP); } } if (XVA_ISSET_REQ(xvap, XAT_AV_MODIFIED)) { if (xoap->xoa_av_modified != ((zp->z_pflags & ZFS_AV_MODIFIED) != 0)) { need_policy = TRUE; } else { XVA_CLR_REQ(xvap, XAT_AV_MODIFIED); XVA_SET_REQ(&tmpxvattr, XAT_AV_MODIFIED); } } if (XVA_ISSET_REQ(xvap, XAT_AV_QUARANTINED)) { if ((vp->v_type != VREG && xoap->xoa_av_quarantined) || xoap->xoa_av_quarantined != ((zp->z_pflags & ZFS_AV_QUARANTINED) != 0)) { need_policy = TRUE; } else { XVA_CLR_REQ(xvap, XAT_AV_QUARANTINED); XVA_SET_REQ(&tmpxvattr, XAT_AV_QUARANTINED); } } if (XVA_ISSET_REQ(xvap, XAT_REPARSE)) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EPERM)); } if (need_policy == FALSE && (XVA_ISSET_REQ(xvap, XAT_AV_SCANSTAMP) || XVA_ISSET_REQ(xvap, XAT_OPAQUE))) { need_policy = TRUE; } } if (mask & AT_MODE) { if (zfs_zaccess(zp, ACE_WRITE_ACL, 0, skipaclchk, cr) == 0) { err = secpolicy_setid_setsticky_clear(vp, vap, &oldva, cr); if (err) { ZFS_EXIT(zfsvfs); return (err); } trim_mask |= AT_MODE; } else { need_policy = TRUE; } } if (need_policy) { /* * If trim_mask is set then take ownership * has been granted or write_acl is present and user * has the ability to modify mode. In that case remove * UID|GID and or MODE from mask so that * secpolicy_vnode_setattr() doesn't revoke it. */ if (trim_mask) { saved_mask = vap->va_mask; vap->va_mask &= ~trim_mask; if (trim_mask & AT_MODE) { /* * Save the mode, as secpolicy_vnode_setattr() * will overwrite it with ova.va_mode. */ saved_mode = vap->va_mode; } } err = secpolicy_vnode_setattr(cr, vp, vap, &oldva, flags, (int (*)(void *, int, cred_t *))zfs_zaccess_unix, zp); if (err) { ZFS_EXIT(zfsvfs); return (err); } if (trim_mask) { vap->va_mask |= saved_mask; if (trim_mask & AT_MODE) { /* * Recover the mode after * secpolicy_vnode_setattr(). */ vap->va_mode = saved_mode; } } } /* * secpolicy_vnode_setattr, or take ownership may have * changed va_mask */ mask = vap->va_mask; if ((mask & (AT_UID | AT_GID))) { err = sa_lookup(zp->z_sa_hdl, SA_ZPL_XATTR(zfsvfs), &xattr_obj, sizeof (xattr_obj)); if (err == 0 && xattr_obj) { err = zfs_zget(zp->z_zfsvfs, xattr_obj, &attrzp); if (err == 0) { err = vn_lock(ZTOV(attrzp), LK_EXCLUSIVE); if (err != 0) vrele(ZTOV(attrzp)); } if (err) goto out2; } if (mask & AT_UID) { new_uid = zfs_fuid_create(zfsvfs, (uint64_t)vap->va_uid, cr, ZFS_OWNER, &fuidp); if (new_uid != zp->z_uid && zfs_fuid_overquota(zfsvfs, B_FALSE, new_uid)) { if (attrzp) vput(ZTOV(attrzp)); err = SET_ERROR(EDQUOT); goto out2; } } if (mask & AT_GID) { new_gid = zfs_fuid_create(zfsvfs, (uint64_t)vap->va_gid, cr, ZFS_GROUP, &fuidp); if (new_gid != zp->z_gid && zfs_fuid_overquota(zfsvfs, B_TRUE, new_gid)) { if (attrzp) vput(ZTOV(attrzp)); err = SET_ERROR(EDQUOT); goto out2; } } } tx = dmu_tx_create(zfsvfs->z_os); if (mask & AT_MODE) { uint64_t pmode = zp->z_mode; uint64_t acl_obj; new_mode = (pmode & S_IFMT) | (vap->va_mode & ~S_IFMT); if (zp->z_zfsvfs->z_acl_mode == ZFS_ACL_RESTRICTED && !(zp->z_pflags & ZFS_ACL_TRIVIAL)) { err = SET_ERROR(EPERM); goto out; } if (err = zfs_acl_chmod_setattr(zp, &aclp, new_mode)) goto out; if (!zp->z_is_sa && ((acl_obj = zfs_external_acl(zp)) != 0)) { /* * Are we upgrading ACL from old V0 format * to V1 format? */ if (zfsvfs->z_version >= ZPL_VERSION_FUID && zfs_znode_acl_version(zp) == ZFS_ACL_VERSION_INITIAL) { dmu_tx_hold_free(tx, acl_obj, 0, DMU_OBJECT_END); dmu_tx_hold_write(tx, DMU_NEW_OBJECT, 0, aclp->z_acl_bytes); } else { dmu_tx_hold_write(tx, acl_obj, 0, aclp->z_acl_bytes); } } else if (!zp->z_is_sa && aclp->z_acl_bytes > ZFS_ACE_SPACE) { dmu_tx_hold_write(tx, DMU_NEW_OBJECT, 0, aclp->z_acl_bytes); } dmu_tx_hold_sa(tx, zp->z_sa_hdl, B_TRUE); } else { if ((mask & AT_XVATTR) && XVA_ISSET_REQ(xvap, XAT_AV_SCANSTAMP)) dmu_tx_hold_sa(tx, zp->z_sa_hdl, B_TRUE); else dmu_tx_hold_sa(tx, zp->z_sa_hdl, B_FALSE); } if (attrzp) { dmu_tx_hold_sa(tx, attrzp->z_sa_hdl, B_FALSE); } fuid_dirtied = zfsvfs->z_fuid_dirty; if (fuid_dirtied) zfs_fuid_txhold(zfsvfs, tx); zfs_sa_upgrade_txholds(tx, zp); err = dmu_tx_assign(tx, TXG_WAIT); if (err) goto out; count = 0; /* * Set each attribute requested. * We group settings according to the locks they need to acquire. * * Note: you cannot set ctime directly, although it will be * updated as a side-effect of calling this function. */ if (mask & (AT_UID|AT_GID|AT_MODE)) mutex_enter(&zp->z_acl_lock); SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_FLAGS(zfsvfs), NULL, &zp->z_pflags, sizeof (zp->z_pflags)); if (attrzp) { if (mask & (AT_UID|AT_GID|AT_MODE)) mutex_enter(&attrzp->z_acl_lock); SA_ADD_BULK_ATTR(xattr_bulk, xattr_count, SA_ZPL_FLAGS(zfsvfs), NULL, &attrzp->z_pflags, sizeof (attrzp->z_pflags)); } if (mask & (AT_UID|AT_GID)) { if (mask & AT_UID) { SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_UID(zfsvfs), NULL, &new_uid, sizeof (new_uid)); zp->z_uid = new_uid; if (attrzp) { SA_ADD_BULK_ATTR(xattr_bulk, xattr_count, SA_ZPL_UID(zfsvfs), NULL, &new_uid, sizeof (new_uid)); attrzp->z_uid = new_uid; } } if (mask & AT_GID) { SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_GID(zfsvfs), NULL, &new_gid, sizeof (new_gid)); zp->z_gid = new_gid; if (attrzp) { SA_ADD_BULK_ATTR(xattr_bulk, xattr_count, SA_ZPL_GID(zfsvfs), NULL, &new_gid, sizeof (new_gid)); attrzp->z_gid = new_gid; } } if (!(mask & AT_MODE)) { SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_MODE(zfsvfs), NULL, &new_mode, sizeof (new_mode)); new_mode = zp->z_mode; } err = zfs_acl_chown_setattr(zp); ASSERT(err == 0); if (attrzp) { err = zfs_acl_chown_setattr(attrzp); ASSERT(err == 0); } } if (mask & AT_MODE) { SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_MODE(zfsvfs), NULL, &new_mode, sizeof (new_mode)); zp->z_mode = new_mode; ASSERT3U((uintptr_t)aclp, !=, 0); err = zfs_aclset_common(zp, aclp, cr, tx); ASSERT0(err); if (zp->z_acl_cached) zfs_acl_free(zp->z_acl_cached); zp->z_acl_cached = aclp; aclp = NULL; } if (mask & AT_ATIME) { ZFS_TIME_ENCODE(&vap->va_atime, zp->z_atime); SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_ATIME(zfsvfs), NULL, &zp->z_atime, sizeof (zp->z_atime)); } if (mask & AT_MTIME) { ZFS_TIME_ENCODE(&vap->va_mtime, mtime); SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_MTIME(zfsvfs), NULL, mtime, sizeof (mtime)); } /* XXX - shouldn't this be done *before* the ATIME/MTIME checks? */ if (mask & AT_SIZE && !(mask & AT_MTIME)) { SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_MTIME(zfsvfs), NULL, mtime, sizeof (mtime)); SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_CTIME(zfsvfs), NULL, &ctime, sizeof (ctime)); zfs_tstamp_update_setup(zp, CONTENT_MODIFIED, mtime, ctime, B_TRUE); } else if (mask != 0) { SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_CTIME(zfsvfs), NULL, &ctime, sizeof (ctime)); zfs_tstamp_update_setup(zp, STATE_CHANGED, mtime, ctime, B_TRUE); if (attrzp) { SA_ADD_BULK_ATTR(xattr_bulk, xattr_count, SA_ZPL_CTIME(zfsvfs), NULL, &ctime, sizeof (ctime)); zfs_tstamp_update_setup(attrzp, STATE_CHANGED, mtime, ctime, B_TRUE); } } /* * Do this after setting timestamps to prevent timestamp * update from toggling bit */ if (xoap && (mask & AT_XVATTR)) { if (XVA_ISSET_REQ(xvap, XAT_CREATETIME)) xoap->xoa_createtime = vap->va_birthtime; /* * restore trimmed off masks * so that return masks can be set for caller. */ if (XVA_ISSET_REQ(&tmpxvattr, XAT_APPENDONLY)) { XVA_SET_REQ(xvap, XAT_APPENDONLY); } if (XVA_ISSET_REQ(&tmpxvattr, XAT_NOUNLINK)) { XVA_SET_REQ(xvap, XAT_NOUNLINK); } if (XVA_ISSET_REQ(&tmpxvattr, XAT_IMMUTABLE)) { XVA_SET_REQ(xvap, XAT_IMMUTABLE); } if (XVA_ISSET_REQ(&tmpxvattr, XAT_NODUMP)) { XVA_SET_REQ(xvap, XAT_NODUMP); } if (XVA_ISSET_REQ(&tmpxvattr, XAT_AV_MODIFIED)) { XVA_SET_REQ(xvap, XAT_AV_MODIFIED); } if (XVA_ISSET_REQ(&tmpxvattr, XAT_AV_QUARANTINED)) { XVA_SET_REQ(xvap, XAT_AV_QUARANTINED); } if (XVA_ISSET_REQ(xvap, XAT_AV_SCANSTAMP)) ASSERT(vp->v_type == VREG); zfs_xvattr_set(zp, xvap, tx); } if (fuid_dirtied) zfs_fuid_sync(zfsvfs, tx); if (mask != 0) zfs_log_setattr(zilog, tx, TX_SETATTR, zp, vap, mask, fuidp); if (mask & (AT_UID|AT_GID|AT_MODE)) mutex_exit(&zp->z_acl_lock); if (attrzp) { if (mask & (AT_UID|AT_GID|AT_MODE)) mutex_exit(&attrzp->z_acl_lock); } out: if (err == 0 && attrzp) { err2 = sa_bulk_update(attrzp->z_sa_hdl, xattr_bulk, xattr_count, tx); ASSERT(err2 == 0); } if (attrzp) vput(ZTOV(attrzp)); if (aclp) zfs_acl_free(aclp); if (fuidp) { zfs_fuid_info_free(fuidp); fuidp = NULL; } if (err) { dmu_tx_abort(tx); } else { err2 = sa_bulk_update(zp->z_sa_hdl, bulk, count, tx); dmu_tx_commit(tx); } out2: if (zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS) zil_commit(zilog, 0); ZFS_EXIT(zfsvfs); return (err); } /* * We acquire all but fdvp locks using non-blocking acquisitions. If we * fail to acquire any lock in the path we will drop all held locks, * acquire the new lock in a blocking fashion, and then release it and * restart the rename. This acquire/release step ensures that we do not * spin on a lock waiting for release. On error release all vnode locks * and decrement references the way tmpfs_rename() would do. */ static int zfs_rename_relock(struct vnode *sdvp, struct vnode **svpp, struct vnode *tdvp, struct vnode **tvpp, const struct componentname *scnp, const struct componentname *tcnp) { zfsvfs_t *zfsvfs; struct vnode *nvp, *svp, *tvp; znode_t *sdzp, *tdzp, *szp, *tzp; const char *snm = scnp->cn_nameptr; const char *tnm = tcnp->cn_nameptr; int error; VOP_UNLOCK(tdvp, 0); if (*tvpp != NULL && *tvpp != tdvp) VOP_UNLOCK(*tvpp, 0); relock: error = vn_lock(sdvp, LK_EXCLUSIVE); if (error) goto out; sdzp = VTOZ(sdvp); error = vn_lock(tdvp, LK_EXCLUSIVE | LK_NOWAIT); if (error != 0) { VOP_UNLOCK(sdvp, 0); if (error != EBUSY) goto out; error = vn_lock(tdvp, LK_EXCLUSIVE); if (error) goto out; VOP_UNLOCK(tdvp, 0); goto relock; } tdzp = VTOZ(tdvp); /* * Before using sdzp and tdzp we must ensure that they are live. * As a porting legacy from illumos we have two things to worry * about. One is typical for FreeBSD and it is that the vnode is * not reclaimed (doomed). The other is that the znode is live. * The current code can invalidate the znode without acquiring the * corresponding vnode lock if the object represented by the znode * and vnode is no longer valid after a rollback or receive operation. * z_teardown_lock hidden behind ZFS_ENTER and ZFS_EXIT is the lock * that protects the znodes from the invalidation. */ zfsvfs = sdzp->z_zfsvfs; ASSERT3P(zfsvfs, ==, tdzp->z_zfsvfs); ZFS_ENTER(zfsvfs); /* * We can not use ZFS_VERIFY_ZP() here because it could directly return * bypassing the cleanup code in the case of an error. */ if (tdzp->z_sa_hdl == NULL || sdzp->z_sa_hdl == NULL) { ZFS_EXIT(zfsvfs); VOP_UNLOCK(sdvp, 0); VOP_UNLOCK(tdvp, 0); error = SET_ERROR(EIO); goto out; } /* * Re-resolve svp to be certain it still exists and fetch the * correct vnode. */ error = zfs_dirent_lookup(sdzp, snm, &szp, ZEXISTS); if (error != 0) { /* Source entry invalid or not there. */ ZFS_EXIT(zfsvfs); VOP_UNLOCK(sdvp, 0); VOP_UNLOCK(tdvp, 0); if ((scnp->cn_flags & ISDOTDOT) != 0 || (scnp->cn_namelen == 1 && scnp->cn_nameptr[0] == '.')) error = SET_ERROR(EINVAL); goto out; } svp = ZTOV(szp); /* * Re-resolve tvp, if it disappeared we just carry on. */ error = zfs_dirent_lookup(tdzp, tnm, &tzp, 0); if (error != 0) { ZFS_EXIT(zfsvfs); VOP_UNLOCK(sdvp, 0); VOP_UNLOCK(tdvp, 0); vrele(svp); if ((tcnp->cn_flags & ISDOTDOT) != 0) error = SET_ERROR(EINVAL); goto out; } if (tzp != NULL) tvp = ZTOV(tzp); else tvp = NULL; /* * At present the vnode locks must be acquired before z_teardown_lock, * although it would be more logical to use the opposite order. */ ZFS_EXIT(zfsvfs); /* * Now try acquire locks on svp and tvp. */ nvp = svp; error = vn_lock(nvp, LK_EXCLUSIVE | LK_NOWAIT); if (error != 0) { VOP_UNLOCK(sdvp, 0); VOP_UNLOCK(tdvp, 0); if (tvp != NULL) vrele(tvp); if (error != EBUSY) { vrele(nvp); goto out; } error = vn_lock(nvp, LK_EXCLUSIVE); if (error != 0) { vrele(nvp); goto out; } VOP_UNLOCK(nvp, 0); /* * Concurrent rename race. * XXX ? */ if (nvp == tdvp) { vrele(nvp); error = SET_ERROR(EINVAL); goto out; } vrele(*svpp); *svpp = nvp; goto relock; } vrele(*svpp); *svpp = nvp; if (*tvpp != NULL) vrele(*tvpp); *tvpp = NULL; if (tvp != NULL) { nvp = tvp; error = vn_lock(nvp, LK_EXCLUSIVE | LK_NOWAIT); if (error != 0) { VOP_UNLOCK(sdvp, 0); VOP_UNLOCK(tdvp, 0); VOP_UNLOCK(*svpp, 0); if (error != EBUSY) { vrele(nvp); goto out; } error = vn_lock(nvp, LK_EXCLUSIVE); if (error != 0) { vrele(nvp); goto out; } vput(nvp); goto relock; } *tvpp = nvp; } return (0); out: return (error); } /* * Note that we must use VRELE_ASYNC in this function as it walks * up the directory tree and vrele may need to acquire an exclusive * lock if a last reference to a vnode is dropped. */ static int zfs_rename_check(znode_t *szp, znode_t *sdzp, znode_t *tdzp) { zfsvfs_t *zfsvfs; znode_t *zp, *zp1; uint64_t parent; int error; zfsvfs = tdzp->z_zfsvfs; if (tdzp == szp) return (SET_ERROR(EINVAL)); if (tdzp == sdzp) return (0); if (tdzp->z_id == zfsvfs->z_root) return (0); zp = tdzp; for (;;) { ASSERT(!zp->z_unlinked); if ((error = sa_lookup(zp->z_sa_hdl, SA_ZPL_PARENT(zfsvfs), &parent, sizeof (parent))) != 0) break; if (parent == szp->z_id) { error = SET_ERROR(EINVAL); break; } if (parent == zfsvfs->z_root) break; if (parent == sdzp->z_id) break; error = zfs_zget(zfsvfs, parent, &zp1); if (error != 0) break; if (zp != tdzp) VN_RELE_ASYNC(ZTOV(zp), dsl_pool_vnrele_taskq(dmu_objset_pool(zfsvfs->z_os))); zp = zp1; } if (error == ENOTDIR) panic("checkpath: .. not a directory\n"); if (zp != tdzp) VN_RELE_ASYNC(ZTOV(zp), dsl_pool_vnrele_taskq(dmu_objset_pool(zfsvfs->z_os))); return (error); } /* * Move an entry from the provided source directory to the target * directory. Change the entry name as indicated. * * IN: sdvp - Source directory containing the "old entry". * snm - Old entry name. * tdvp - Target directory to contain the "new entry". * tnm - New entry name. * cr - credentials of caller. * ct - caller context * flags - case flags * * RETURN: 0 on success, error code on failure. * * Timestamps: * sdvp,tdvp - ctime|mtime updated */ /*ARGSUSED*/ static int zfs_rename(vnode_t *sdvp, vnode_t **svpp, struct componentname *scnp, vnode_t *tdvp, vnode_t **tvpp, struct componentname *tcnp, cred_t *cr) { zfsvfs_t *zfsvfs; znode_t *sdzp, *tdzp, *szp, *tzp; zilog_t *zilog = NULL; dmu_tx_t *tx; char *snm = scnp->cn_nameptr; char *tnm = tcnp->cn_nameptr; int error = 0; /* Reject renames across filesystems. */ if ((*svpp)->v_mount != tdvp->v_mount || ((*tvpp) != NULL && (*svpp)->v_mount != (*tvpp)->v_mount)) { error = SET_ERROR(EXDEV); goto out; } if (zfsctl_is_node(tdvp)) { error = SET_ERROR(EXDEV); goto out; } /* * Lock all four vnodes to ensure safety and semantics of renaming. */ error = zfs_rename_relock(sdvp, svpp, tdvp, tvpp, scnp, tcnp); if (error != 0) { /* no vnodes are locked in the case of error here */ return (error); } tdzp = VTOZ(tdvp); sdzp = VTOZ(sdvp); zfsvfs = tdzp->z_zfsvfs; zilog = zfsvfs->z_log; /* * After we re-enter ZFS_ENTER() we will have to revalidate all * znodes involved. */ ZFS_ENTER(zfsvfs); if (zfsvfs->z_utf8 && u8_validate(tnm, strlen(tnm), NULL, U8_VALIDATE_ENTIRE, &error) < 0) { error = SET_ERROR(EILSEQ); goto unlockout; } /* If source and target are the same file, there is nothing to do. */ if ((*svpp) == (*tvpp)) { error = 0; goto unlockout; } if (((*svpp)->v_type == VDIR && (*svpp)->v_mountedhere != NULL) || ((*tvpp) != NULL && (*tvpp)->v_type == VDIR && (*tvpp)->v_mountedhere != NULL)) { error = SET_ERROR(EXDEV); goto unlockout; } /* * We can not use ZFS_VERIFY_ZP() here because it could directly return * bypassing the cleanup code in the case of an error. */ if (tdzp->z_sa_hdl == NULL || sdzp->z_sa_hdl == NULL) { error = SET_ERROR(EIO); goto unlockout; } szp = VTOZ(*svpp); tzp = *tvpp == NULL ? NULL : VTOZ(*tvpp); if (szp->z_sa_hdl == NULL || (tzp != NULL && tzp->z_sa_hdl == NULL)) { error = SET_ERROR(EIO); goto unlockout; } /* * This is to prevent the creation of links into attribute space * by renaming a linked file into/outof an attribute directory. * See the comment in zfs_link() for why this is considered bad. */ if ((tdzp->z_pflags & ZFS_XATTR) != (sdzp->z_pflags & ZFS_XATTR)) { error = SET_ERROR(EINVAL); goto unlockout; } /* * Must have write access at the source to remove the old entry * and write access at the target to create the new entry. * Note that if target and source are the same, this can be * done in a single check. */ if (error = zfs_zaccess_rename(sdzp, szp, tdzp, tzp, cr)) goto unlockout; if ((*svpp)->v_type == VDIR) { /* * Avoid ".", "..", and aliases of "." for obvious reasons. */ if ((scnp->cn_namelen == 1 && scnp->cn_nameptr[0] == '.') || sdzp == szp || (scnp->cn_flags | tcnp->cn_flags) & ISDOTDOT) { error = EINVAL; goto unlockout; } /* * Check to make sure rename is valid. * Can't do a move like this: /usr/a/b to /usr/a/b/c/d */ if (error = zfs_rename_check(szp, sdzp, tdzp)) goto unlockout; } /* * Does target exist? */ if (tzp) { /* * Source and target must be the same type. */ if ((*svpp)->v_type == VDIR) { if ((*tvpp)->v_type != VDIR) { error = SET_ERROR(ENOTDIR); goto unlockout; } else { cache_purge(tdvp); if (sdvp != tdvp) cache_purge(sdvp); } } else { if ((*tvpp)->v_type == VDIR) { error = SET_ERROR(EISDIR); goto unlockout; } } } vnevent_rename_src(*svpp, sdvp, scnp->cn_nameptr, ct); if (tzp) vnevent_rename_dest(*tvpp, tdvp, tnm, ct); /* * notify the target directory if it is not the same * as source directory. */ if (tdvp != sdvp) { vnevent_rename_dest_dir(tdvp, ct); } tx = dmu_tx_create(zfsvfs->z_os); dmu_tx_hold_sa(tx, szp->z_sa_hdl, B_FALSE); dmu_tx_hold_sa(tx, sdzp->z_sa_hdl, B_FALSE); dmu_tx_hold_zap(tx, sdzp->z_id, FALSE, snm); dmu_tx_hold_zap(tx, tdzp->z_id, TRUE, tnm); if (sdzp != tdzp) { dmu_tx_hold_sa(tx, tdzp->z_sa_hdl, B_FALSE); zfs_sa_upgrade_txholds(tx, tdzp); } if (tzp) { dmu_tx_hold_sa(tx, tzp->z_sa_hdl, B_FALSE); zfs_sa_upgrade_txholds(tx, tzp); } zfs_sa_upgrade_txholds(tx, szp); dmu_tx_hold_zap(tx, zfsvfs->z_unlinkedobj, FALSE, NULL); error = dmu_tx_assign(tx, TXG_WAIT); if (error) { dmu_tx_abort(tx); goto unlockout; } if (tzp) /* Attempt to remove the existing target */ error = zfs_link_destroy(tdzp, tnm, tzp, tx, 0, NULL); if (error == 0) { error = zfs_link_create(tdzp, tnm, szp, tx, ZRENAMING); if (error == 0) { szp->z_pflags |= ZFS_AV_MODIFIED; error = sa_update(szp->z_sa_hdl, SA_ZPL_FLAGS(zfsvfs), (void *)&szp->z_pflags, sizeof (uint64_t), tx); ASSERT0(error); error = zfs_link_destroy(sdzp, snm, szp, tx, ZRENAMING, NULL); if (error == 0) { zfs_log_rename(zilog, tx, TX_RENAME, sdzp, snm, tdzp, tnm, szp); /* * Update path information for the target vnode */ vn_renamepath(tdvp, *svpp, tnm, strlen(tnm)); } else { /* * At this point, we have successfully created * the target name, but have failed to remove * the source name. Since the create was done * with the ZRENAMING flag, there are * complications; for one, the link count is * wrong. The easiest way to deal with this * is to remove the newly created target, and * return the original error. This must * succeed; fortunately, it is very unlikely to * fail, since we just created it. */ VERIFY3U(zfs_link_destroy(tdzp, tnm, szp, tx, ZRENAMING, NULL), ==, 0); } } if (error == 0) { cache_purge(*svpp); if (*tvpp != NULL) cache_purge(*tvpp); cache_purge_negative(tdvp); } } dmu_tx_commit(tx); unlockout: /* all 4 vnodes are locked, ZFS_ENTER called */ ZFS_EXIT(zfsvfs); VOP_UNLOCK(*svpp, 0); VOP_UNLOCK(sdvp, 0); out: /* original two vnodes are locked */ if (error == 0 && zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS) zil_commit(zilog, 0); if (*tvpp != NULL) VOP_UNLOCK(*tvpp, 0); if (tdvp != *tvpp) VOP_UNLOCK(tdvp, 0); return (error); } /* * Insert the indicated symbolic reference entry into the directory. * * IN: dvp - Directory to contain new symbolic link. * link - Name for new symlink entry. * vap - Attributes of new entry. * cr - credentials of caller. * ct - caller context * flags - case flags * * RETURN: 0 on success, error code on failure. * * Timestamps: * dvp - ctime|mtime updated */ /*ARGSUSED*/ static int zfs_symlink(vnode_t *dvp, vnode_t **vpp, char *name, vattr_t *vap, char *link, cred_t *cr, kthread_t *td) { znode_t *zp, *dzp = VTOZ(dvp); dmu_tx_t *tx; zfsvfs_t *zfsvfs = dzp->z_zfsvfs; zilog_t *zilog; uint64_t len = strlen(link); int error; zfs_acl_ids_t acl_ids; boolean_t fuid_dirtied; uint64_t txtype = TX_SYMLINK; int flags = 0; ASSERT(vap->va_type == VLNK); ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(dzp); zilog = zfsvfs->z_log; if (zfsvfs->z_utf8 && u8_validate(name, strlen(name), NULL, U8_VALIDATE_ENTIRE, &error) < 0) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EILSEQ)); } if (len > MAXPATHLEN) { ZFS_EXIT(zfsvfs); return (SET_ERROR(ENAMETOOLONG)); } if ((error = zfs_acl_ids_create(dzp, 0, vap, cr, NULL, &acl_ids)) != 0) { ZFS_EXIT(zfsvfs); return (error); } /* * Attempt to lock directory; fail if entry already exists. */ error = zfs_dirent_lookup(dzp, name, &zp, ZNEW); if (error) { zfs_acl_ids_free(&acl_ids); ZFS_EXIT(zfsvfs); return (error); } if (error = zfs_zaccess(dzp, ACE_ADD_FILE, 0, B_FALSE, cr)) { zfs_acl_ids_free(&acl_ids); ZFS_EXIT(zfsvfs); return (error); } if (zfs_acl_ids_overquota(zfsvfs, &acl_ids)) { zfs_acl_ids_free(&acl_ids); ZFS_EXIT(zfsvfs); return (SET_ERROR(EDQUOT)); } getnewvnode_reserve(1); tx = dmu_tx_create(zfsvfs->z_os); fuid_dirtied = zfsvfs->z_fuid_dirty; dmu_tx_hold_write(tx, DMU_NEW_OBJECT, 0, MAX(1, len)); dmu_tx_hold_zap(tx, dzp->z_id, TRUE, name); dmu_tx_hold_sa_create(tx, acl_ids.z_aclp->z_acl_bytes + ZFS_SA_BASE_ATTR_SIZE + len); dmu_tx_hold_sa(tx, dzp->z_sa_hdl, B_FALSE); if (!zfsvfs->z_use_sa && acl_ids.z_aclp->z_acl_bytes > ZFS_ACE_SPACE) { dmu_tx_hold_write(tx, DMU_NEW_OBJECT, 0, acl_ids.z_aclp->z_acl_bytes); } if (fuid_dirtied) zfs_fuid_txhold(zfsvfs, tx); error = dmu_tx_assign(tx, TXG_WAIT); if (error) { zfs_acl_ids_free(&acl_ids); dmu_tx_abort(tx); getnewvnode_drop_reserve(); ZFS_EXIT(zfsvfs); return (error); } /* * Create a new object for the symlink. * for version 4 ZPL datsets the symlink will be an SA attribute */ zfs_mknode(dzp, vap, tx, cr, 0, &zp, &acl_ids); if (fuid_dirtied) zfs_fuid_sync(zfsvfs, tx); if (zp->z_is_sa) error = sa_update(zp->z_sa_hdl, SA_ZPL_SYMLINK(zfsvfs), link, len, tx); else zfs_sa_symlink(zp, link, len, tx); zp->z_size = len; (void) sa_update(zp->z_sa_hdl, SA_ZPL_SIZE(zfsvfs), &zp->z_size, sizeof (zp->z_size), tx); /* * Insert the new object into the directory. */ (void) zfs_link_create(dzp, name, zp, tx, ZNEW); zfs_log_symlink(zilog, tx, txtype, dzp, zp, name, link); *vpp = ZTOV(zp); zfs_acl_ids_free(&acl_ids); dmu_tx_commit(tx); getnewvnode_drop_reserve(); if (zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS) zil_commit(zilog, 0); ZFS_EXIT(zfsvfs); return (error); } /* * Return, in the buffer contained in the provided uio structure, * the symbolic path referred to by vp. * * IN: vp - vnode of symbolic link. * uio - structure to contain the link path. * cr - credentials of caller. * ct - caller context * * OUT: uio - structure containing the link path. * * RETURN: 0 on success, error code on failure. * * Timestamps: * vp - atime updated */ /* ARGSUSED */ static int zfs_readlink(vnode_t *vp, uio_t *uio, cred_t *cr, caller_context_t *ct) { znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; int error; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); if (zp->z_is_sa) error = sa_lookup_uio(zp->z_sa_hdl, SA_ZPL_SYMLINK(zfsvfs), uio); else error = zfs_sa_readlink(zp, uio); ZFS_ACCESSTIME_STAMP(zfsvfs, zp); ZFS_EXIT(zfsvfs); return (error); } /* * Insert a new entry into directory tdvp referencing svp. * * IN: tdvp - Directory to contain new entry. * svp - vnode of new entry. * name - name of new entry. * cr - credentials of caller. * ct - caller context * * RETURN: 0 on success, error code on failure. * * Timestamps: * tdvp - ctime|mtime updated * svp - ctime updated */ /* ARGSUSED */ static int zfs_link(vnode_t *tdvp, vnode_t *svp, char *name, cred_t *cr, caller_context_t *ct, int flags) { znode_t *dzp = VTOZ(tdvp); znode_t *tzp, *szp; zfsvfs_t *zfsvfs = dzp->z_zfsvfs; zilog_t *zilog; dmu_tx_t *tx; int error; uint64_t parent; uid_t owner; ASSERT(tdvp->v_type == VDIR); ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(dzp); zilog = zfsvfs->z_log; /* * POSIX dictates that we return EPERM here. * Better choices include ENOTSUP or EISDIR. */ if (svp->v_type == VDIR) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EPERM)); } szp = VTOZ(svp); ZFS_VERIFY_ZP(szp); if (szp->z_pflags & (ZFS_APPENDONLY | ZFS_IMMUTABLE | ZFS_READONLY)) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EPERM)); } /* Prevent links to .zfs/shares files */ if ((error = sa_lookup(szp->z_sa_hdl, SA_ZPL_PARENT(zfsvfs), &parent, sizeof (uint64_t))) != 0) { ZFS_EXIT(zfsvfs); return (error); } if (parent == zfsvfs->z_shares_dir) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EPERM)); } if (zfsvfs->z_utf8 && u8_validate(name, strlen(name), NULL, U8_VALIDATE_ENTIRE, &error) < 0) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EILSEQ)); } /* * We do not support links between attributes and non-attributes * because of the potential security risk of creating links * into "normal" file space in order to circumvent restrictions * imposed in attribute space. */ if ((szp->z_pflags & ZFS_XATTR) != (dzp->z_pflags & ZFS_XATTR)) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EINVAL)); } owner = zfs_fuid_map_id(zfsvfs, szp->z_uid, cr, ZFS_OWNER); if (owner != crgetuid(cr) && secpolicy_basic_link(svp, cr) != 0) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EPERM)); } if (error = zfs_zaccess(dzp, ACE_ADD_FILE, 0, B_FALSE, cr)) { ZFS_EXIT(zfsvfs); return (error); } /* * Attempt to lock directory; fail if entry already exists. */ error = zfs_dirent_lookup(dzp, name, &tzp, ZNEW); if (error) { ZFS_EXIT(zfsvfs); return (error); } tx = dmu_tx_create(zfsvfs->z_os); dmu_tx_hold_sa(tx, szp->z_sa_hdl, B_FALSE); dmu_tx_hold_zap(tx, dzp->z_id, TRUE, name); zfs_sa_upgrade_txholds(tx, szp); zfs_sa_upgrade_txholds(tx, dzp); error = dmu_tx_assign(tx, TXG_WAIT); if (error) { dmu_tx_abort(tx); ZFS_EXIT(zfsvfs); return (error); } error = zfs_link_create(dzp, name, szp, tx, 0); if (error == 0) { uint64_t txtype = TX_LINK; zfs_log_link(zilog, tx, txtype, dzp, szp, name); } dmu_tx_commit(tx); if (error == 0) { vnevent_link(svp, ct); } if (zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS) zil_commit(zilog, 0); ZFS_EXIT(zfsvfs); return (error); } /*ARGSUSED*/ void zfs_inactive(vnode_t *vp, cred_t *cr, caller_context_t *ct) { znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; int error; rw_enter(&zfsvfs->z_teardown_inactive_lock, RW_READER); if (zp->z_sa_hdl == NULL) { /* * The fs has been unmounted, or we did a * suspend/resume and this file no longer exists. */ rw_exit(&zfsvfs->z_teardown_inactive_lock); vrecycle(vp); return; } if (zp->z_unlinked) { /* * Fast path to recycle a vnode of a removed file. */ rw_exit(&zfsvfs->z_teardown_inactive_lock); vrecycle(vp); return; } if (zp->z_atime_dirty && zp->z_unlinked == 0) { dmu_tx_t *tx = dmu_tx_create(zfsvfs->z_os); dmu_tx_hold_sa(tx, zp->z_sa_hdl, B_FALSE); zfs_sa_upgrade_txholds(tx, zp); error = dmu_tx_assign(tx, TXG_WAIT); if (error) { dmu_tx_abort(tx); } else { (void) sa_update(zp->z_sa_hdl, SA_ZPL_ATIME(zfsvfs), (void *)&zp->z_atime, sizeof (zp->z_atime), tx); zp->z_atime_dirty = 0; dmu_tx_commit(tx); } } rw_exit(&zfsvfs->z_teardown_inactive_lock); } CTASSERT(sizeof(struct zfid_short) <= sizeof(struct fid)); CTASSERT(sizeof(struct zfid_long) <= sizeof(struct fid)); /*ARGSUSED*/ static int zfs_fid(vnode_t *vp, fid_t *fidp, caller_context_t *ct) { znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; uint32_t gen; uint64_t gen64; uint64_t object = zp->z_id; zfid_short_t *zfid; int size, i, error; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); if ((error = sa_lookup(zp->z_sa_hdl, SA_ZPL_GEN(zfsvfs), &gen64, sizeof (uint64_t))) != 0) { ZFS_EXIT(zfsvfs); return (error); } gen = (uint32_t)gen64; size = (zfsvfs->z_parent != zfsvfs) ? LONG_FID_LEN : SHORT_FID_LEN; #ifdef illumos if (fidp->fid_len < size) { fidp->fid_len = size; ZFS_EXIT(zfsvfs); return (SET_ERROR(ENOSPC)); } #else fidp->fid_len = size; #endif zfid = (zfid_short_t *)fidp; zfid->zf_len = size; for (i = 0; i < sizeof (zfid->zf_object); i++) zfid->zf_object[i] = (uint8_t)(object >> (8 * i)); /* Must have a non-zero generation number to distinguish from .zfs */ if (gen == 0) gen = 1; for (i = 0; i < sizeof (zfid->zf_gen); i++) zfid->zf_gen[i] = (uint8_t)(gen >> (8 * i)); if (size == LONG_FID_LEN) { uint64_t objsetid = dmu_objset_id(zfsvfs->z_os); zfid_long_t *zlfid; zlfid = (zfid_long_t *)fidp; for (i = 0; i < sizeof (zlfid->zf_setid); i++) zlfid->zf_setid[i] = (uint8_t)(objsetid >> (8 * i)); /* XXX - this should be the generation number for the objset */ for (i = 0; i < sizeof (zlfid->zf_setgen); i++) zlfid->zf_setgen[i] = 0; } ZFS_EXIT(zfsvfs); return (0); } static int zfs_pathconf(vnode_t *vp, int cmd, ulong_t *valp, cred_t *cr, caller_context_t *ct) { znode_t *zp, *xzp; zfsvfs_t *zfsvfs; int error; switch (cmd) { case _PC_LINK_MAX: *valp = MIN(LONG_MAX, ZFS_LINK_MAX); return (0); case _PC_FILESIZEBITS: *valp = 64; return (0); #ifdef illumos case _PC_XATTR_EXISTS: zp = VTOZ(vp); zfsvfs = zp->z_zfsvfs; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); *valp = 0; error = zfs_dirent_lookup(zp, "", &xzp, ZXATTR | ZEXISTS | ZSHARED); if (error == 0) { if (!zfs_dirempty(xzp)) *valp = 1; vrele(ZTOV(xzp)); } else if (error == ENOENT) { /* * If there aren't extended attributes, it's the * same as having zero of them. */ error = 0; } ZFS_EXIT(zfsvfs); return (error); case _PC_SATTR_ENABLED: case _PC_SATTR_EXISTS: *valp = vfs_has_feature(vp->v_vfsp, VFSFT_SYSATTR_VIEWS) && (vp->v_type == VREG || vp->v_type == VDIR); return (0); case _PC_ACCESS_FILTERING: *valp = vfs_has_feature(vp->v_vfsp, VFSFT_ACCESS_FILTER) && vp->v_type == VDIR; return (0); case _PC_ACL_ENABLED: *valp = _ACL_ACE_ENABLED; return (0); #endif /* illumos */ case _PC_MIN_HOLE_SIZE: *valp = (int)SPA_MINBLOCKSIZE; return (0); #ifdef illumos case _PC_TIMESTAMP_RESOLUTION: /* nanosecond timestamp resolution */ *valp = 1L; return (0); #endif case _PC_ACL_EXTENDED: *valp = 0; return (0); case _PC_ACL_NFS4: *valp = 1; return (0); case _PC_ACL_PATH_MAX: *valp = ACL_MAX_ENTRIES; return (0); default: return (EOPNOTSUPP); } } /*ARGSUSED*/ static int zfs_getsecattr(vnode_t *vp, vsecattr_t *vsecp, int flag, cred_t *cr, caller_context_t *ct) { znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; int error; boolean_t skipaclchk = (flag & ATTR_NOACLCHECK) ? B_TRUE : B_FALSE; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); error = zfs_getacl(zp, vsecp, skipaclchk, cr); ZFS_EXIT(zfsvfs); return (error); } /*ARGSUSED*/ int zfs_setsecattr(vnode_t *vp, vsecattr_t *vsecp, int flag, cred_t *cr, caller_context_t *ct) { znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; int error; boolean_t skipaclchk = (flag & ATTR_NOACLCHECK) ? B_TRUE : B_FALSE; zilog_t *zilog = zfsvfs->z_log; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); error = zfs_setacl(zp, vsecp, skipaclchk, cr); if (zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS) zil_commit(zilog, 0); ZFS_EXIT(zfsvfs); return (error); } static int zfs_getpages(struct vnode *vp, vm_page_t *ma, int count, int *rbehind, int *rahead) { znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; objset_t *os = zp->z_zfsvfs->z_os; rl_t *rl; vm_object_t object; off_t start, end, obj_size; uint_t blksz; int pgsin_b, pgsin_a; int error; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); start = IDX_TO_OFF(ma[0]->pindex); end = IDX_TO_OFF(ma[count - 1]->pindex + 1); /* * Lock a range covering all required and optional pages. * Note that we need to handle the case of the block size growing. */ for (;;) { blksz = zp->z_blksz; rl = zfs_range_lock(zp, rounddown(start, blksz), roundup(end, blksz) - rounddown(start, blksz), RL_READER); if (blksz == zp->z_blksz) break; zfs_range_unlock(rl); } object = ma[0]->object; zfs_vmobject_wlock(object); obj_size = object->un_pager.vnp.vnp_size; zfs_vmobject_wunlock(object); if (IDX_TO_OFF(ma[count - 1]->pindex) >= obj_size) { zfs_range_unlock(rl); ZFS_EXIT(zfsvfs); return (zfs_vm_pagerret_bad); } pgsin_b = 0; if (rbehind != NULL) { pgsin_b = OFF_TO_IDX(start - rounddown(start, blksz)); pgsin_b = MIN(*rbehind, pgsin_b); } pgsin_a = 0; if (rahead != NULL) { pgsin_a = OFF_TO_IDX(roundup(end, blksz) - end); if (end + IDX_TO_OFF(pgsin_a) >= obj_size) pgsin_a = OFF_TO_IDX(round_page(obj_size) - end); pgsin_a = MIN(*rahead, pgsin_a); } /* * NB: we need to pass the exact byte size of the data that we expect * to read after accounting for the file size. This is required because * ZFS will panic if we request DMU to read beyond the end of the last * allocated block. */ error = dmu_read_pages(os, zp->z_id, ma, count, &pgsin_b, &pgsin_a, MIN(end, obj_size) - (end - PAGE_SIZE)); zfs_range_unlock(rl); ZFS_ACCESSTIME_STAMP(zfsvfs, zp); ZFS_EXIT(zfsvfs); if (error != 0) return (zfs_vm_pagerret_error); VM_CNT_INC(v_vnodein); VM_CNT_ADD(v_vnodepgsin, count + pgsin_b + pgsin_a); if (rbehind != NULL) *rbehind = pgsin_b; if (rahead != NULL) *rahead = pgsin_a; return (zfs_vm_pagerret_ok); } static int zfs_freebsd_getpages(ap) struct vop_getpages_args /* { struct vnode *a_vp; vm_page_t *a_m; int a_count; int *a_rbehind; int *a_rahead; } */ *ap; { return (zfs_getpages(ap->a_vp, ap->a_m, ap->a_count, ap->a_rbehind, ap->a_rahead)); } static int zfs_putpages(struct vnode *vp, vm_page_t *ma, size_t len, int flags, int *rtvals) { znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; rl_t *rl; dmu_tx_t *tx; struct sf_buf *sf; vm_object_t object; vm_page_t m; caddr_t va; size_t tocopy; size_t lo_len; vm_ooffset_t lo_off; vm_ooffset_t off; uint_t blksz; int ncount; int pcount; int err; int i; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); object = vp->v_object; pcount = btoc(len); ncount = pcount; KASSERT(ma[0]->object == object, ("mismatching object")); KASSERT(len > 0 && (len & PAGE_MASK) == 0, ("unexpected length")); for (i = 0; i < pcount; i++) rtvals[i] = zfs_vm_pagerret_error; off = IDX_TO_OFF(ma[0]->pindex); blksz = zp->z_blksz; lo_off = rounddown(off, blksz); lo_len = roundup(len + (off - lo_off), blksz); rl = zfs_range_lock(zp, lo_off, lo_len, RL_WRITER); zfs_vmobject_wlock(object); if (len + off > object->un_pager.vnp.vnp_size) { if (object->un_pager.vnp.vnp_size > off) { int pgoff; len = object->un_pager.vnp.vnp_size - off; ncount = btoc(len); if ((pgoff = (int)len & PAGE_MASK) != 0) { /* * If the object is locked and the following * conditions hold, then the page's dirty * field cannot be concurrently changed by a * pmap operation. */ m = ma[ncount - 1]; vm_page_assert_sbusied(m); KASSERT(!pmap_page_is_write_mapped(m), ("zfs_putpages: page %p is not read-only", m)); vm_page_clear_dirty(m, pgoff, PAGE_SIZE - pgoff); } } else { len = 0; ncount = 0; } if (ncount < pcount) { for (i = ncount; i < pcount; i++) { rtvals[i] = zfs_vm_pagerret_bad; } } } zfs_vmobject_wunlock(object); if (ncount == 0) goto out; if (zfs_owner_overquota(zfsvfs, zp, B_FALSE) || zfs_owner_overquota(zfsvfs, zp, B_TRUE)) { goto out; } tx = dmu_tx_create(zfsvfs->z_os); dmu_tx_hold_write(tx, zp->z_id, off, len); dmu_tx_hold_sa(tx, zp->z_sa_hdl, B_FALSE); zfs_sa_upgrade_txholds(tx, zp); err = dmu_tx_assign(tx, TXG_WAIT); if (err != 0) { dmu_tx_abort(tx); goto out; } if (zp->z_blksz < PAGE_SIZE) { for (i = 0; len > 0; off += tocopy, len -= tocopy, i++) { tocopy = len > PAGE_SIZE ? PAGE_SIZE : len; va = zfs_map_page(ma[i], &sf); dmu_write(zfsvfs->z_os, zp->z_id, off, tocopy, va, tx); zfs_unmap_page(sf); } } else { err = dmu_write_pages(zfsvfs->z_os, zp->z_id, off, len, ma, tx); } if (err == 0) { uint64_t mtime[2], ctime[2]; sa_bulk_attr_t bulk[3]; int count = 0; SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_MTIME(zfsvfs), NULL, &mtime, 16); SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_CTIME(zfsvfs), NULL, &ctime, 16); SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_FLAGS(zfsvfs), NULL, &zp->z_pflags, 8); zfs_tstamp_update_setup(zp, CONTENT_MODIFIED, mtime, ctime, B_TRUE); err = sa_bulk_update(zp->z_sa_hdl, bulk, count, tx); ASSERT0(err); zfs_log_write(zfsvfs->z_log, tx, TX_WRITE, zp, off, len, 0); zfs_vmobject_wlock(object); for (i = 0; i < ncount; i++) { rtvals[i] = zfs_vm_pagerret_ok; vm_page_undirty(ma[i]); } zfs_vmobject_wunlock(object); VM_CNT_INC(v_vnodeout); VM_CNT_ADD(v_vnodepgsout, ncount); } dmu_tx_commit(tx); out: zfs_range_unlock(rl); if ((flags & (zfs_vm_pagerput_sync | zfs_vm_pagerput_inval)) != 0 || zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS) zil_commit(zfsvfs->z_log, zp->z_id); ZFS_EXIT(zfsvfs); return (rtvals[0]); } int zfs_freebsd_putpages(ap) struct vop_putpages_args /* { struct vnode *a_vp; vm_page_t *a_m; int a_count; int a_sync; int *a_rtvals; } */ *ap; { return (zfs_putpages(ap->a_vp, ap->a_m, ap->a_count, ap->a_sync, ap->a_rtvals)); } static int zfs_freebsd_bmap(ap) struct vop_bmap_args /* { struct vnode *a_vp; daddr_t a_bn; struct bufobj **a_bop; daddr_t *a_bnp; int *a_runp; int *a_runb; } */ *ap; { if (ap->a_bop != NULL) *ap->a_bop = &ap->a_vp->v_bufobj; if (ap->a_bnp != NULL) *ap->a_bnp = ap->a_bn; if (ap->a_runp != NULL) *ap->a_runp = 0; if (ap->a_runb != NULL) *ap->a_runb = 0; return (0); } static int zfs_freebsd_open(ap) struct vop_open_args /* { struct vnode *a_vp; int a_mode; struct ucred *a_cred; struct thread *a_td; } */ *ap; { vnode_t *vp = ap->a_vp; znode_t *zp = VTOZ(vp); int error; error = zfs_open(&vp, ap->a_mode, ap->a_cred, NULL); if (error == 0) vnode_create_vobject(vp, zp->z_size, ap->a_td); return (error); } static int zfs_freebsd_close(ap) struct vop_close_args /* { struct vnode *a_vp; int a_fflag; struct ucred *a_cred; struct thread *a_td; } */ *ap; { return (zfs_close(ap->a_vp, ap->a_fflag, 1, 0, ap->a_cred, NULL)); } static int zfs_freebsd_ioctl(ap) struct vop_ioctl_args /* { struct vnode *a_vp; u_long a_command; caddr_t a_data; int a_fflag; struct ucred *cred; struct thread *td; } */ *ap; { return (zfs_ioctl(ap->a_vp, ap->a_command, (intptr_t)ap->a_data, ap->a_fflag, ap->a_cred, NULL, NULL)); } static int ioflags(int ioflags) { int flags = 0; if (ioflags & IO_APPEND) flags |= FAPPEND; if (ioflags & IO_NDELAY) flags |= FNONBLOCK; if (ioflags & IO_SYNC) flags |= (FSYNC | FDSYNC | FRSYNC); return (flags); } static int zfs_freebsd_read(ap) struct vop_read_args /* { struct vnode *a_vp; struct uio *a_uio; int a_ioflag; struct ucred *a_cred; } */ *ap; { return (zfs_read(ap->a_vp, ap->a_uio, ioflags(ap->a_ioflag), ap->a_cred, NULL)); } static int zfs_freebsd_write(ap) struct vop_write_args /* { struct vnode *a_vp; struct uio *a_uio; int a_ioflag; struct ucred *a_cred; } */ *ap; { return (zfs_write(ap->a_vp, ap->a_uio, ioflags(ap->a_ioflag), ap->a_cred, NULL)); } static int zfs_freebsd_access(ap) struct vop_access_args /* { struct vnode *a_vp; accmode_t a_accmode; struct ucred *a_cred; struct thread *a_td; } */ *ap; { vnode_t *vp = ap->a_vp; znode_t *zp = VTOZ(vp); accmode_t accmode; int error = 0; /* * ZFS itself only knowns about VREAD, VWRITE, VEXEC and VAPPEND, */ accmode = ap->a_accmode & (VREAD|VWRITE|VEXEC|VAPPEND); if (accmode != 0) error = zfs_access(ap->a_vp, accmode, 0, ap->a_cred, NULL); /* * VADMIN has to be handled by vaccess(). */ if (error == 0) { accmode = ap->a_accmode & ~(VREAD|VWRITE|VEXEC|VAPPEND); if (accmode != 0) { error = vaccess(vp->v_type, zp->z_mode, zp->z_uid, zp->z_gid, accmode, ap->a_cred, NULL); } } /* * For VEXEC, ensure that at least one execute bit is set for * non-directories. */ if (error == 0 && (ap->a_accmode & VEXEC) != 0 && vp->v_type != VDIR && (zp->z_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) == 0) { error = EACCES; } return (error); } static int zfs_freebsd_lookup(ap) struct vop_lookup_args /* { struct vnode *a_dvp; struct vnode **a_vpp; struct componentname *a_cnp; } */ *ap; { struct componentname *cnp = ap->a_cnp; char nm[NAME_MAX + 1]; ASSERT(cnp->cn_namelen < sizeof(nm)); strlcpy(nm, cnp->cn_nameptr, MIN(cnp->cn_namelen + 1, sizeof(nm))); return (zfs_lookup(ap->a_dvp, nm, ap->a_vpp, cnp, cnp->cn_nameiop, cnp->cn_cred, cnp->cn_thread, 0)); } static int zfs_cache_lookup(ap) struct vop_lookup_args /* { struct vnode *a_dvp; struct vnode **a_vpp; struct componentname *a_cnp; } */ *ap; { zfsvfs_t *zfsvfs; zfsvfs = ap->a_dvp->v_mount->mnt_data; if (zfsvfs->z_use_namecache) return (vfs_cache_lookup(ap)); else return (zfs_freebsd_lookup(ap)); } static int zfs_freebsd_create(ap) struct vop_create_args /* { struct vnode *a_dvp; struct vnode **a_vpp; struct componentname *a_cnp; struct vattr *a_vap; } */ *ap; { zfsvfs_t *zfsvfs; struct componentname *cnp = ap->a_cnp; vattr_t *vap = ap->a_vap; int error, mode; ASSERT(cnp->cn_flags & SAVENAME); vattr_init_mask(vap); mode = vap->va_mode & ALLPERMS; zfsvfs = ap->a_dvp->v_mount->mnt_data; error = zfs_create(ap->a_dvp, cnp->cn_nameptr, vap, !EXCL, mode, ap->a_vpp, cnp->cn_cred, cnp->cn_thread); if (zfsvfs->z_use_namecache && error == 0 && (cnp->cn_flags & MAKEENTRY) != 0) cache_enter(ap->a_dvp, *ap->a_vpp, cnp); return (error); } static int zfs_freebsd_remove(ap) struct vop_remove_args /* { struct vnode *a_dvp; struct vnode *a_vp; struct componentname *a_cnp; } */ *ap; { ASSERT(ap->a_cnp->cn_flags & SAVENAME); return (zfs_remove(ap->a_dvp, ap->a_vp, ap->a_cnp->cn_nameptr, ap->a_cnp->cn_cred)); } static int zfs_freebsd_mkdir(ap) struct vop_mkdir_args /* { struct vnode *a_dvp; struct vnode **a_vpp; struct componentname *a_cnp; struct vattr *a_vap; } */ *ap; { vattr_t *vap = ap->a_vap; ASSERT(ap->a_cnp->cn_flags & SAVENAME); vattr_init_mask(vap); return (zfs_mkdir(ap->a_dvp, ap->a_cnp->cn_nameptr, vap, ap->a_vpp, ap->a_cnp->cn_cred)); } static int zfs_freebsd_rmdir(ap) struct vop_rmdir_args /* { struct vnode *a_dvp; struct vnode *a_vp; struct componentname *a_cnp; } */ *ap; { struct componentname *cnp = ap->a_cnp; ASSERT(cnp->cn_flags & SAVENAME); return (zfs_rmdir(ap->a_dvp, ap->a_vp, cnp->cn_nameptr, cnp->cn_cred)); } static int zfs_freebsd_readdir(ap) struct vop_readdir_args /* { struct vnode *a_vp; struct uio *a_uio; struct ucred *a_cred; int *a_eofflag; int *a_ncookies; u_long **a_cookies; } */ *ap; { return (zfs_readdir(ap->a_vp, ap->a_uio, ap->a_cred, ap->a_eofflag, ap->a_ncookies, ap->a_cookies)); } static int zfs_freebsd_fsync(ap) struct vop_fsync_args /* { struct vnode *a_vp; int a_waitfor; struct thread *a_td; } */ *ap; { vop_stdfsync(ap); return (zfs_fsync(ap->a_vp, 0, ap->a_td->td_ucred, NULL)); } static int zfs_freebsd_getattr(ap) struct vop_getattr_args /* { struct vnode *a_vp; struct vattr *a_vap; struct ucred *a_cred; } */ *ap; { vattr_t *vap = ap->a_vap; xvattr_t xvap; u_long fflags = 0; int error; xva_init(&xvap); xvap.xva_vattr = *vap; xvap.xva_vattr.va_mask |= AT_XVATTR; /* Convert chflags into ZFS-type flags. */ /* XXX: what about SF_SETTABLE?. */ XVA_SET_REQ(&xvap, XAT_IMMUTABLE); XVA_SET_REQ(&xvap, XAT_APPENDONLY); XVA_SET_REQ(&xvap, XAT_NOUNLINK); XVA_SET_REQ(&xvap, XAT_NODUMP); XVA_SET_REQ(&xvap, XAT_READONLY); XVA_SET_REQ(&xvap, XAT_ARCHIVE); XVA_SET_REQ(&xvap, XAT_SYSTEM); XVA_SET_REQ(&xvap, XAT_HIDDEN); XVA_SET_REQ(&xvap, XAT_REPARSE); XVA_SET_REQ(&xvap, XAT_OFFLINE); XVA_SET_REQ(&xvap, XAT_SPARSE); error = zfs_getattr(ap->a_vp, (vattr_t *)&xvap, 0, ap->a_cred, NULL); if (error != 0) return (error); /* Convert ZFS xattr into chflags. */ #define FLAG_CHECK(fflag, xflag, xfield) do { \ if (XVA_ISSET_RTN(&xvap, (xflag)) && (xfield) != 0) \ fflags |= (fflag); \ } while (0) FLAG_CHECK(SF_IMMUTABLE, XAT_IMMUTABLE, xvap.xva_xoptattrs.xoa_immutable); FLAG_CHECK(SF_APPEND, XAT_APPENDONLY, xvap.xva_xoptattrs.xoa_appendonly); FLAG_CHECK(SF_NOUNLINK, XAT_NOUNLINK, xvap.xva_xoptattrs.xoa_nounlink); FLAG_CHECK(UF_ARCHIVE, XAT_ARCHIVE, xvap.xva_xoptattrs.xoa_archive); FLAG_CHECK(UF_NODUMP, XAT_NODUMP, xvap.xva_xoptattrs.xoa_nodump); FLAG_CHECK(UF_READONLY, XAT_READONLY, xvap.xva_xoptattrs.xoa_readonly); FLAG_CHECK(UF_SYSTEM, XAT_SYSTEM, xvap.xva_xoptattrs.xoa_system); FLAG_CHECK(UF_HIDDEN, XAT_HIDDEN, xvap.xva_xoptattrs.xoa_hidden); FLAG_CHECK(UF_REPARSE, XAT_REPARSE, xvap.xva_xoptattrs.xoa_reparse); FLAG_CHECK(UF_OFFLINE, XAT_OFFLINE, xvap.xva_xoptattrs.xoa_offline); FLAG_CHECK(UF_SPARSE, XAT_SPARSE, xvap.xva_xoptattrs.xoa_sparse); #undef FLAG_CHECK *vap = xvap.xva_vattr; vap->va_flags = fflags; return (0); } static int zfs_freebsd_setattr(ap) struct vop_setattr_args /* { struct vnode *a_vp; struct vattr *a_vap; struct ucred *a_cred; } */ *ap; { vnode_t *vp = ap->a_vp; vattr_t *vap = ap->a_vap; cred_t *cred = ap->a_cred; xvattr_t xvap; u_long fflags; uint64_t zflags; vattr_init_mask(vap); vap->va_mask &= ~AT_NOSET; xva_init(&xvap); xvap.xva_vattr = *vap; zflags = VTOZ(vp)->z_pflags; if (vap->va_flags != VNOVAL) { zfsvfs_t *zfsvfs = VTOZ(vp)->z_zfsvfs; int error; if (zfsvfs->z_use_fuids == B_FALSE) return (EOPNOTSUPP); fflags = vap->va_flags; /* * XXX KDM * We need to figure out whether it makes sense to allow * UF_REPARSE through, since we don't really have other * facilities to handle reparse points and zfs_setattr() * doesn't currently allow setting that attribute anyway. */ if ((fflags & ~(SF_IMMUTABLE|SF_APPEND|SF_NOUNLINK|UF_ARCHIVE| UF_NODUMP|UF_SYSTEM|UF_HIDDEN|UF_READONLY|UF_REPARSE| UF_OFFLINE|UF_SPARSE)) != 0) return (EOPNOTSUPP); /* * Unprivileged processes are not permitted to unset system * flags, or modify flags if any system flags are set. * Privileged non-jail processes may not modify system flags * if securelevel > 0 and any existing system flags are set. * Privileged jail processes behave like privileged non-jail * processes if the PR_ALLOW_CHFLAGS permission bit is set; * otherwise, they behave like unprivileged processes. */ if (secpolicy_fs_owner(vp->v_mount, cred) == 0 || priv_check_cred(cred, PRIV_VFS_SYSFLAGS) == 0) { if (zflags & (ZFS_IMMUTABLE | ZFS_APPENDONLY | ZFS_NOUNLINK)) { error = securelevel_gt(cred, 0); if (error != 0) return (error); } } else { /* * Callers may only modify the file flags on objects they * have VADMIN rights for. */ if ((error = VOP_ACCESS(vp, VADMIN, cred, curthread)) != 0) return (error); if (zflags & (ZFS_IMMUTABLE | ZFS_APPENDONLY | ZFS_NOUNLINK)) { return (EPERM); } if (fflags & (SF_IMMUTABLE | SF_APPEND | SF_NOUNLINK)) { return (EPERM); } } #define FLAG_CHANGE(fflag, zflag, xflag, xfield) do { \ if (((fflags & (fflag)) && !(zflags & (zflag))) || \ ((zflags & (zflag)) && !(fflags & (fflag)))) { \ XVA_SET_REQ(&xvap, (xflag)); \ (xfield) = ((fflags & (fflag)) != 0); \ } \ } while (0) /* Convert chflags into ZFS-type flags. */ /* XXX: what about SF_SETTABLE?. */ FLAG_CHANGE(SF_IMMUTABLE, ZFS_IMMUTABLE, XAT_IMMUTABLE, xvap.xva_xoptattrs.xoa_immutable); FLAG_CHANGE(SF_APPEND, ZFS_APPENDONLY, XAT_APPENDONLY, xvap.xva_xoptattrs.xoa_appendonly); FLAG_CHANGE(SF_NOUNLINK, ZFS_NOUNLINK, XAT_NOUNLINK, xvap.xva_xoptattrs.xoa_nounlink); FLAG_CHANGE(UF_ARCHIVE, ZFS_ARCHIVE, XAT_ARCHIVE, xvap.xva_xoptattrs.xoa_archive); FLAG_CHANGE(UF_NODUMP, ZFS_NODUMP, XAT_NODUMP, xvap.xva_xoptattrs.xoa_nodump); FLAG_CHANGE(UF_READONLY, ZFS_READONLY, XAT_READONLY, xvap.xva_xoptattrs.xoa_readonly); FLAG_CHANGE(UF_SYSTEM, ZFS_SYSTEM, XAT_SYSTEM, xvap.xva_xoptattrs.xoa_system); FLAG_CHANGE(UF_HIDDEN, ZFS_HIDDEN, XAT_HIDDEN, xvap.xva_xoptattrs.xoa_hidden); FLAG_CHANGE(UF_REPARSE, ZFS_REPARSE, XAT_REPARSE, xvap.xva_xoptattrs.xoa_reparse); FLAG_CHANGE(UF_OFFLINE, ZFS_OFFLINE, XAT_OFFLINE, xvap.xva_xoptattrs.xoa_offline); FLAG_CHANGE(UF_SPARSE, ZFS_SPARSE, XAT_SPARSE, xvap.xva_xoptattrs.xoa_sparse); #undef FLAG_CHANGE } if (vap->va_birthtime.tv_sec != VNOVAL) { xvap.xva_vattr.va_mask |= AT_XVATTR; XVA_SET_REQ(&xvap, XAT_CREATETIME); } return (zfs_setattr(vp, (vattr_t *)&xvap, 0, cred, NULL)); } static int zfs_freebsd_rename(ap) struct vop_rename_args /* { struct vnode *a_fdvp; struct vnode *a_fvp; struct componentname *a_fcnp; struct vnode *a_tdvp; struct vnode *a_tvp; struct componentname *a_tcnp; } */ *ap; { vnode_t *fdvp = ap->a_fdvp; vnode_t *fvp = ap->a_fvp; vnode_t *tdvp = ap->a_tdvp; vnode_t *tvp = ap->a_tvp; int error; ASSERT(ap->a_fcnp->cn_flags & (SAVENAME|SAVESTART)); ASSERT(ap->a_tcnp->cn_flags & (SAVENAME|SAVESTART)); error = zfs_rename(fdvp, &fvp, ap->a_fcnp, tdvp, &tvp, ap->a_tcnp, ap->a_fcnp->cn_cred); vrele(fdvp); vrele(fvp); vrele(tdvp); if (tvp != NULL) vrele(tvp); return (error); } static int zfs_freebsd_symlink(ap) struct vop_symlink_args /* { struct vnode *a_dvp; struct vnode **a_vpp; struct componentname *a_cnp; struct vattr *a_vap; char *a_target; } */ *ap; { struct componentname *cnp = ap->a_cnp; vattr_t *vap = ap->a_vap; ASSERT(cnp->cn_flags & SAVENAME); vap->va_type = VLNK; /* FreeBSD: Syscall only sets va_mode. */ vattr_init_mask(vap); return (zfs_symlink(ap->a_dvp, ap->a_vpp, cnp->cn_nameptr, vap, __DECONST(char *, ap->a_target), cnp->cn_cred, cnp->cn_thread)); } static int zfs_freebsd_readlink(ap) struct vop_readlink_args /* { struct vnode *a_vp; struct uio *a_uio; struct ucred *a_cred; } */ *ap; { return (zfs_readlink(ap->a_vp, ap->a_uio, ap->a_cred, NULL)); } static int zfs_freebsd_link(ap) struct vop_link_args /* { struct vnode *a_tdvp; struct vnode *a_vp; struct componentname *a_cnp; } */ *ap; { struct componentname *cnp = ap->a_cnp; vnode_t *vp = ap->a_vp; vnode_t *tdvp = ap->a_tdvp; if (tdvp->v_mount != vp->v_mount) return (EXDEV); ASSERT(cnp->cn_flags & SAVENAME); return (zfs_link(tdvp, vp, cnp->cn_nameptr, cnp->cn_cred, NULL, 0)); } static int zfs_freebsd_inactive(ap) struct vop_inactive_args /* { struct vnode *a_vp; struct thread *a_td; } */ *ap; { vnode_t *vp = ap->a_vp; zfs_inactive(vp, ap->a_td->td_ucred, NULL); return (0); } static int zfs_freebsd_reclaim(ap) struct vop_reclaim_args /* { struct vnode *a_vp; struct thread *a_td; } */ *ap; { vnode_t *vp = ap->a_vp; znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; ASSERT(zp != NULL); /* * z_teardown_inactive_lock protects from a race with * zfs_znode_dmu_fini in zfsvfs_teardown during * force unmount. */ rw_enter(&zfsvfs->z_teardown_inactive_lock, RW_READER); if (zp->z_sa_hdl == NULL) zfs_znode_free(zp); else zfs_zinactive(zp); rw_exit(&zfsvfs->z_teardown_inactive_lock); vp->v_data = NULL; return (0); } static int zfs_freebsd_fid(ap) struct vop_fid_args /* { struct vnode *a_vp; struct fid *a_fid; } */ *ap; { return (zfs_fid(ap->a_vp, (void *)ap->a_fid, NULL)); } static int zfs_freebsd_pathconf(ap) struct vop_pathconf_args /* { struct vnode *a_vp; int a_name; register_t *a_retval; } */ *ap; { ulong_t val; int error; error = zfs_pathconf(ap->a_vp, ap->a_name, &val, curthread->td_ucred, NULL); if (error == 0) { *ap->a_retval = val; return (error); } if (error != EOPNOTSUPP) return (error); switch (ap->a_name) { case _PC_NAME_MAX: *ap->a_retval = NAME_MAX; return (0); case _PC_PIPE_BUF: if (ap->a_vp->v_type == VDIR || ap->a_vp->v_type == VFIFO) { *ap->a_retval = PIPE_BUF; return (0); } return (EINVAL); default: return (vop_stdpathconf(ap)); } } /* * FreeBSD's extended attributes namespace defines file name prefix for ZFS' * extended attribute name: * * NAMESPACE PREFIX * system freebsd:system: * user (none, can be used to access ZFS fsattr(5) attributes * created on Solaris) */ static int zfs_create_attrname(int attrnamespace, const char *name, char *attrname, size_t size) { const char *namespace, *prefix, *suffix; /* We don't allow '/' character in attribute name. */ if (strchr(name, '/') != NULL) return (EINVAL); /* We don't allow attribute names that start with "freebsd:" string. */ if (strncmp(name, "freebsd:", 8) == 0) return (EINVAL); bzero(attrname, size); switch (attrnamespace) { case EXTATTR_NAMESPACE_USER: #if 0 prefix = "freebsd:"; namespace = EXTATTR_NAMESPACE_USER_STRING; suffix = ":"; #else /* * This is the default namespace by which we can access all * attributes created on Solaris. */ prefix = namespace = suffix = ""; #endif break; case EXTATTR_NAMESPACE_SYSTEM: prefix = "freebsd:"; namespace = EXTATTR_NAMESPACE_SYSTEM_STRING; suffix = ":"; break; case EXTATTR_NAMESPACE_EMPTY: default: return (EINVAL); } if (snprintf(attrname, size, "%s%s%s%s", prefix, namespace, suffix, name) >= size) { return (ENAMETOOLONG); } return (0); } /* * Vnode operating to retrieve a named extended attribute. */ static int zfs_getextattr(struct vop_getextattr_args *ap) /* vop_getextattr { IN struct vnode *a_vp; IN int a_attrnamespace; IN const char *a_name; INOUT struct uio *a_uio; OUT size_t *a_size; IN struct ucred *a_cred; IN struct thread *a_td; }; */ { zfsvfs_t *zfsvfs = VTOZ(ap->a_vp)->z_zfsvfs; struct thread *td = ap->a_td; struct nameidata nd; char attrname[255]; struct vattr va; vnode_t *xvp = NULL, *vp; int error, flags; error = extattr_check_cred(ap->a_vp, ap->a_attrnamespace, ap->a_cred, ap->a_td, VREAD); if (error != 0) return (error); error = zfs_create_attrname(ap->a_attrnamespace, ap->a_name, attrname, sizeof(attrname)); if (error != 0) return (error); ZFS_ENTER(zfsvfs); error = zfs_lookup(ap->a_vp, NULL, &xvp, NULL, 0, ap->a_cred, td, LOOKUP_XATTR); if (error != 0) { ZFS_EXIT(zfsvfs); return (error); } flags = FREAD; NDINIT_ATVP(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, attrname, xvp, td); error = vn_open_cred(&nd, &flags, 0, 0, ap->a_cred, NULL); vp = nd.ni_vp; NDFREE(&nd, NDF_ONLY_PNBUF); if (error != 0) { ZFS_EXIT(zfsvfs); if (error == ENOENT) error = ENOATTR; return (error); } if (ap->a_size != NULL) { error = VOP_GETATTR(vp, &va, ap->a_cred); if (error == 0) *ap->a_size = (size_t)va.va_size; } else if (ap->a_uio != NULL) error = VOP_READ(vp, ap->a_uio, IO_UNIT, ap->a_cred); VOP_UNLOCK(vp, 0); vn_close(vp, flags, ap->a_cred, td); ZFS_EXIT(zfsvfs); return (error); } /* * Vnode operation to remove a named attribute. */ int zfs_deleteextattr(struct vop_deleteextattr_args *ap) /* vop_deleteextattr { IN struct vnode *a_vp; IN int a_attrnamespace; IN const char *a_name; IN struct ucred *a_cred; IN struct thread *a_td; }; */ { zfsvfs_t *zfsvfs = VTOZ(ap->a_vp)->z_zfsvfs; struct thread *td = ap->a_td; struct nameidata nd; char attrname[255]; struct vattr va; vnode_t *xvp = NULL, *vp; int error, flags; error = extattr_check_cred(ap->a_vp, ap->a_attrnamespace, ap->a_cred, ap->a_td, VWRITE); if (error != 0) return (error); error = zfs_create_attrname(ap->a_attrnamespace, ap->a_name, attrname, sizeof(attrname)); if (error != 0) return (error); ZFS_ENTER(zfsvfs); error = zfs_lookup(ap->a_vp, NULL, &xvp, NULL, 0, ap->a_cred, td, LOOKUP_XATTR); if (error != 0) { ZFS_EXIT(zfsvfs); return (error); } NDINIT_ATVP(&nd, DELETE, NOFOLLOW | LOCKPARENT | LOCKLEAF, UIO_SYSSPACE, attrname, xvp, td); error = namei(&nd); vp = nd.ni_vp; if (error != 0) { ZFS_EXIT(zfsvfs); NDFREE(&nd, NDF_ONLY_PNBUF); if (error == ENOENT) error = ENOATTR; return (error); } error = VOP_REMOVE(nd.ni_dvp, vp, &nd.ni_cnd); NDFREE(&nd, NDF_ONLY_PNBUF); vput(nd.ni_dvp); if (vp == nd.ni_dvp) vrele(vp); else vput(vp); ZFS_EXIT(zfsvfs); return (error); } /* * Vnode operation to set a named attribute. */ static int zfs_setextattr(struct vop_setextattr_args *ap) /* vop_setextattr { IN struct vnode *a_vp; IN int a_attrnamespace; IN const char *a_name; INOUT struct uio *a_uio; IN struct ucred *a_cred; IN struct thread *a_td; }; */ { zfsvfs_t *zfsvfs = VTOZ(ap->a_vp)->z_zfsvfs; struct thread *td = ap->a_td; struct nameidata nd; char attrname[255]; struct vattr va; vnode_t *xvp = NULL, *vp; int error, flags; error = extattr_check_cred(ap->a_vp, ap->a_attrnamespace, ap->a_cred, ap->a_td, VWRITE); if (error != 0) return (error); error = zfs_create_attrname(ap->a_attrnamespace, ap->a_name, attrname, sizeof(attrname)); if (error != 0) return (error); ZFS_ENTER(zfsvfs); error = zfs_lookup(ap->a_vp, NULL, &xvp, NULL, 0, ap->a_cred, td, LOOKUP_XATTR | CREATE_XATTR_DIR); if (error != 0) { ZFS_EXIT(zfsvfs); return (error); } flags = FFLAGS(O_WRONLY | O_CREAT); NDINIT_ATVP(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, attrname, xvp, td); error = vn_open_cred(&nd, &flags, 0600, 0, ap->a_cred, NULL); vp = nd.ni_vp; NDFREE(&nd, NDF_ONLY_PNBUF); if (error != 0) { ZFS_EXIT(zfsvfs); return (error); } VATTR_NULL(&va); va.va_size = 0; error = VOP_SETATTR(vp, &va, ap->a_cred); if (error == 0) VOP_WRITE(vp, ap->a_uio, IO_UNIT, ap->a_cred); VOP_UNLOCK(vp, 0); vn_close(vp, flags, ap->a_cred, td); ZFS_EXIT(zfsvfs); return (error); } /* * Vnode operation to retrieve extended attributes on a vnode. */ static int zfs_listextattr(struct vop_listextattr_args *ap) /* vop_listextattr { IN struct vnode *a_vp; IN int a_attrnamespace; INOUT struct uio *a_uio; OUT size_t *a_size; IN struct ucred *a_cred; IN struct thread *a_td; }; */ { zfsvfs_t *zfsvfs = VTOZ(ap->a_vp)->z_zfsvfs; struct thread *td = ap->a_td; struct nameidata nd; char attrprefix[16]; u_char dirbuf[sizeof(struct dirent)]; struct dirent *dp; struct iovec aiov; struct uio auio, *uio = ap->a_uio; size_t *sizep = ap->a_size; size_t plen; vnode_t *xvp = NULL, *vp; int done, error, eof, pos; error = extattr_check_cred(ap->a_vp, ap->a_attrnamespace, ap->a_cred, ap->a_td, VREAD); if (error != 0) return (error); error = zfs_create_attrname(ap->a_attrnamespace, "", attrprefix, sizeof(attrprefix)); if (error != 0) return (error); plen = strlen(attrprefix); ZFS_ENTER(zfsvfs); if (sizep != NULL) *sizep = 0; error = zfs_lookup(ap->a_vp, NULL, &xvp, NULL, 0, ap->a_cred, td, LOOKUP_XATTR); if (error != 0) { ZFS_EXIT(zfsvfs); /* * ENOATTR means that the EA directory does not yet exist, * i.e. there are no extended attributes there. */ if (error == ENOATTR) error = 0; return (error); } NDINIT_ATVP(&nd, LOOKUP, NOFOLLOW | LOCKLEAF | LOCKSHARED, UIO_SYSSPACE, ".", xvp, td); error = namei(&nd); vp = nd.ni_vp; NDFREE(&nd, NDF_ONLY_PNBUF); if (error != 0) { ZFS_EXIT(zfsvfs); return (error); } auio.uio_iov = &aiov; auio.uio_iovcnt = 1; auio.uio_segflg = UIO_SYSSPACE; auio.uio_td = td; auio.uio_rw = UIO_READ; auio.uio_offset = 0; do { u_char nlen; aiov.iov_base = (void *)dirbuf; aiov.iov_len = sizeof(dirbuf); auio.uio_resid = sizeof(dirbuf); error = VOP_READDIR(vp, &auio, ap->a_cred, &eof, NULL, NULL); done = sizeof(dirbuf) - auio.uio_resid; if (error != 0) break; for (pos = 0; pos < done;) { dp = (struct dirent *)(dirbuf + pos); pos += dp->d_reclen; /* * XXX: Temporarily we also accept DT_UNKNOWN, as this * is what we get when attribute was created on Solaris. */ if (dp->d_type != DT_REG && dp->d_type != DT_UNKNOWN) continue; if (plen == 0 && strncmp(dp->d_name, "freebsd:", 8) == 0) continue; else if (strncmp(dp->d_name, attrprefix, plen) != 0) continue; nlen = dp->d_namlen - plen; if (sizep != NULL) *sizep += 1 + nlen; else if (uio != NULL) { /* * Format of extattr name entry is one byte for * length and the rest for name. */ error = uiomove(&nlen, 1, uio->uio_rw, uio); if (error == 0) { error = uiomove(dp->d_name + plen, nlen, uio->uio_rw, uio); } if (error != 0) break; } } } while (!eof && error == 0); vput(vp); ZFS_EXIT(zfsvfs); return (error); } int zfs_freebsd_getacl(ap) struct vop_getacl_args /* { struct vnode *vp; acl_type_t type; struct acl *aclp; struct ucred *cred; struct thread *td; } */ *ap; { int error; vsecattr_t vsecattr; if (ap->a_type != ACL_TYPE_NFS4) return (EINVAL); vsecattr.vsa_mask = VSA_ACE | VSA_ACECNT; if (error = zfs_getsecattr(ap->a_vp, &vsecattr, 0, ap->a_cred, NULL)) return (error); error = acl_from_aces(ap->a_aclp, vsecattr.vsa_aclentp, vsecattr.vsa_aclcnt); if (vsecattr.vsa_aclentp != NULL) kmem_free(vsecattr.vsa_aclentp, vsecattr.vsa_aclentsz); return (error); } int zfs_freebsd_setacl(ap) struct vop_setacl_args /* { struct vnode *vp; acl_type_t type; struct acl *aclp; struct ucred *cred; struct thread *td; } */ *ap; { int error; vsecattr_t vsecattr; int aclbsize; /* size of acl list in bytes */ aclent_t *aaclp; if (ap->a_type != ACL_TYPE_NFS4) return (EINVAL); if (ap->a_aclp == NULL) return (EINVAL); if (ap->a_aclp->acl_cnt < 1 || ap->a_aclp->acl_cnt > MAX_ACL_ENTRIES) return (EINVAL); /* * With NFSv4 ACLs, chmod(2) may need to add additional entries, * splitting every entry into two and appending "canonical six" * entries at the end. Don't allow for setting an ACL that would * cause chmod(2) to run out of ACL entries. */ if (ap->a_aclp->acl_cnt * 2 + 6 > ACL_MAX_ENTRIES) return (ENOSPC); error = acl_nfs4_check(ap->a_aclp, ap->a_vp->v_type == VDIR); if (error != 0) return (error); vsecattr.vsa_mask = VSA_ACE; aclbsize = ap->a_aclp->acl_cnt * sizeof(ace_t); vsecattr.vsa_aclentp = kmem_alloc(aclbsize, KM_SLEEP); aaclp = vsecattr.vsa_aclentp; vsecattr.vsa_aclentsz = aclbsize; aces_from_acl(vsecattr.vsa_aclentp, &vsecattr.vsa_aclcnt, ap->a_aclp); error = zfs_setsecattr(ap->a_vp, &vsecattr, 0, ap->a_cred, NULL); kmem_free(aaclp, aclbsize); return (error); } int zfs_freebsd_aclcheck(ap) struct vop_aclcheck_args /* { struct vnode *vp; acl_type_t type; struct acl *aclp; struct ucred *cred; struct thread *td; } */ *ap; { return (EOPNOTSUPP); } static int zfs_vptocnp(struct vop_vptocnp_args *ap) { vnode_t *covered_vp; vnode_t *vp = ap->a_vp;; zfsvfs_t *zfsvfs = vp->v_vfsp->vfs_data; znode_t *zp = VTOZ(vp); enum vgetstate vs; int ltype; int error; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); /* * If we are a snapshot mounted under .zfs, run the operation * on the covered vnode. */ if (zp->z_id != zfsvfs->z_root || zfsvfs->z_parent == zfsvfs) { char name[MAXNAMLEN + 1]; znode_t *dzp; size_t len; error = zfs_znode_parent_and_name(zp, &dzp, name); if (error == 0) { len = strlen(name); if (*ap->a_buflen < len) error = SET_ERROR(ENOMEM); } if (error == 0) { *ap->a_buflen -= len; bcopy(name, ap->a_buf + *ap->a_buflen, len); *ap->a_vpp = ZTOV(dzp); } ZFS_EXIT(zfsvfs); return (error); } ZFS_EXIT(zfsvfs); covered_vp = vp->v_mount->mnt_vnodecovered; vs = vget_prep(covered_vp); ltype = VOP_ISLOCKED(vp); VOP_UNLOCK(vp, 0); error = vget_finish(covered_vp, LK_SHARED, vs); if (error == 0) { error = VOP_VPTOCNP(covered_vp, ap->a_vpp, ap->a_cred, ap->a_buf, ap->a_buflen); vput(covered_vp); } vn_lock(vp, ltype | LK_RETRY); if ((vp->v_iflag & VI_DOOMED) != 0) error = SET_ERROR(ENOENT); return (error); } #ifdef DIAGNOSTIC static int zfs_lock(ap) struct vop_lock1_args /* { struct vnode *a_vp; int a_flags; char *file; int line; } */ *ap; { vnode_t *vp; znode_t *zp; int err; err = vop_stdlock(ap); if (err == 0 && (ap->a_flags & LK_NOWAIT) == 0) { vp = ap->a_vp; zp = vp->v_data; if (vp->v_mount != NULL && (vp->v_iflag & VI_DOOMED) == 0 && zp != NULL && (zp->z_pflags & ZFS_XATTR) == 0) VERIFY(!RRM_LOCK_HELD(&zp->z_zfsvfs->z_teardown_lock)); } return (err); } #endif struct vop_vector zfs_vnodeops; struct vop_vector zfs_fifoops; struct vop_vector zfs_shareops; struct vop_vector zfs_vnodeops = { .vop_default = &default_vnodeops, .vop_inactive = zfs_freebsd_inactive, .vop_reclaim = zfs_freebsd_reclaim, .vop_access = zfs_freebsd_access, .vop_allocate = VOP_EINVAL, .vop_lookup = zfs_cache_lookup, .vop_cachedlookup = zfs_freebsd_lookup, .vop_getattr = zfs_freebsd_getattr, .vop_setattr = zfs_freebsd_setattr, .vop_create = zfs_freebsd_create, .vop_mknod = zfs_freebsd_create, .vop_mkdir = zfs_freebsd_mkdir, .vop_readdir = zfs_freebsd_readdir, .vop_fsync = zfs_freebsd_fsync, .vop_open = zfs_freebsd_open, .vop_close = zfs_freebsd_close, .vop_rmdir = zfs_freebsd_rmdir, .vop_ioctl = zfs_freebsd_ioctl, .vop_link = zfs_freebsd_link, .vop_symlink = zfs_freebsd_symlink, .vop_readlink = zfs_freebsd_readlink, .vop_read = zfs_freebsd_read, .vop_write = zfs_freebsd_write, .vop_remove = zfs_freebsd_remove, .vop_rename = zfs_freebsd_rename, .vop_pathconf = zfs_freebsd_pathconf, .vop_bmap = zfs_freebsd_bmap, .vop_fid = zfs_freebsd_fid, .vop_getextattr = zfs_getextattr, .vop_deleteextattr = zfs_deleteextattr, .vop_setextattr = zfs_setextattr, .vop_listextattr = zfs_listextattr, .vop_getacl = zfs_freebsd_getacl, .vop_setacl = zfs_freebsd_setacl, .vop_aclcheck = zfs_freebsd_aclcheck, .vop_getpages = zfs_freebsd_getpages, .vop_putpages = zfs_freebsd_putpages, .vop_vptocnp = zfs_vptocnp, #ifdef DIAGNOSTIC .vop_lock1 = zfs_lock, #endif }; struct vop_vector zfs_fifoops = { .vop_default = &fifo_specops, .vop_fsync = zfs_freebsd_fsync, .vop_access = zfs_freebsd_access, .vop_getattr = zfs_freebsd_getattr, .vop_inactive = zfs_freebsd_inactive, .vop_read = VOP_PANIC, .vop_reclaim = zfs_freebsd_reclaim, .vop_setattr = zfs_freebsd_setattr, .vop_write = VOP_PANIC, .vop_pathconf = zfs_freebsd_pathconf, .vop_fid = zfs_freebsd_fid, .vop_getacl = zfs_freebsd_getacl, .vop_setacl = zfs_freebsd_setacl, .vop_aclcheck = zfs_freebsd_aclcheck, }; /* * special share hidden files vnode operations template */ struct vop_vector zfs_shareops = { .vop_default = &default_vnodeops, .vop_access = zfs_freebsd_access, .vop_inactive = zfs_freebsd_inactive, .vop_reclaim = zfs_freebsd_reclaim, .vop_fid = zfs_freebsd_fid, .vop_pathconf = zfs_freebsd_pathconf, }; Index: head/sys/compat/linuxkpi/common/src/linux_compat.c =================================================================== --- head/sys/compat/linuxkpi/common/src/linux_compat.c (revision 353538) +++ head/sys/compat/linuxkpi/common/src/linux_compat.c (revision 353539) @@ -1,2458 +1,2458 @@ /*- * Copyright (c) 2010 Isilon Systems, Inc. * Copyright (c) 2010 iX Systems, Inc. * Copyright (c) 2010 Panasas, Inc. * Copyright (c) 2013-2018 Mellanox Technologies, Ltd. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice unmodified, this list of conditions, and the following * disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include "opt_stack.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #if defined(__i386__) || defined(__amd64__) #include #endif #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #if defined(__i386__) || defined(__amd64__) #include #endif SYSCTL_NODE(_compat, OID_AUTO, linuxkpi, CTLFLAG_RW, 0, "LinuxKPI parameters"); int linuxkpi_debug; SYSCTL_INT(_compat_linuxkpi, OID_AUTO, debug, CTLFLAG_RWTUN, &linuxkpi_debug, 0, "Set to enable pr_debug() prints. Clear to disable."); MALLOC_DEFINE(M_KMALLOC, "linux", "Linux kmalloc compat"); #include /* Undo Linux compat changes. */ #undef RB_ROOT #undef file #undef cdev #define RB_ROOT(head) (head)->rbh_root static void linux_cdev_deref(struct linux_cdev *ldev); static struct vm_area_struct *linux_cdev_handle_find(void *handle); struct kobject linux_class_root; struct device linux_root_device; struct class linux_class_misc; struct list_head pci_drivers; struct list_head pci_devices; spinlock_t pci_lock; unsigned long linux_timer_hz_mask; int panic_cmp(struct rb_node *one, struct rb_node *two) { panic("no cmp"); } RB_GENERATE(linux_root, rb_node, __entry, panic_cmp); int kobject_set_name_vargs(struct kobject *kobj, const char *fmt, va_list args) { va_list tmp_va; int len; char *old; char *name; char dummy; old = kobj->name; if (old && fmt == NULL) return (0); /* compute length of string */ va_copy(tmp_va, args); len = vsnprintf(&dummy, 0, fmt, tmp_va); va_end(tmp_va); /* account for zero termination */ len++; /* check for error */ if (len < 1) return (-EINVAL); /* allocate memory for string */ name = kzalloc(len, GFP_KERNEL); if (name == NULL) return (-ENOMEM); vsnprintf(name, len, fmt, args); kobj->name = name; /* free old string */ kfree(old); /* filter new string */ for (; *name != '\0'; name++) if (*name == '/') *name = '!'; return (0); } int kobject_set_name(struct kobject *kobj, const char *fmt, ...) { va_list args; int error; va_start(args, fmt); error = kobject_set_name_vargs(kobj, fmt, args); va_end(args); return (error); } static int kobject_add_complete(struct kobject *kobj, struct kobject *parent) { const struct kobj_type *t; int error; kobj->parent = parent; error = sysfs_create_dir(kobj); if (error == 0 && kobj->ktype && kobj->ktype->default_attrs) { struct attribute **attr; t = kobj->ktype; for (attr = t->default_attrs; *attr != NULL; attr++) { error = sysfs_create_file(kobj, *attr); if (error) break; } if (error) sysfs_remove_dir(kobj); } return (error); } int kobject_add(struct kobject *kobj, struct kobject *parent, const char *fmt, ...) { va_list args; int error; va_start(args, fmt); error = kobject_set_name_vargs(kobj, fmt, args); va_end(args); if (error) return (error); return kobject_add_complete(kobj, parent); } void linux_kobject_release(struct kref *kref) { struct kobject *kobj; char *name; kobj = container_of(kref, struct kobject, kref); sysfs_remove_dir(kobj); name = kobj->name; if (kobj->ktype && kobj->ktype->release) kobj->ktype->release(kobj); kfree(name); } static void linux_kobject_kfree(struct kobject *kobj) { kfree(kobj); } static void linux_kobject_kfree_name(struct kobject *kobj) { if (kobj) { kfree(kobj->name); } } const struct kobj_type linux_kfree_type = { .release = linux_kobject_kfree }; static void linux_device_release(struct device *dev) { pr_debug("linux_device_release: %s\n", dev_name(dev)); kfree(dev); } static ssize_t linux_class_show(struct kobject *kobj, struct attribute *attr, char *buf) { struct class_attribute *dattr; ssize_t error; dattr = container_of(attr, struct class_attribute, attr); error = -EIO; if (dattr->show) error = dattr->show(container_of(kobj, struct class, kobj), dattr, buf); return (error); } static ssize_t linux_class_store(struct kobject *kobj, struct attribute *attr, const char *buf, size_t count) { struct class_attribute *dattr; ssize_t error; dattr = container_of(attr, struct class_attribute, attr); error = -EIO; if (dattr->store) error = dattr->store(container_of(kobj, struct class, kobj), dattr, buf, count); return (error); } static void linux_class_release(struct kobject *kobj) { struct class *class; class = container_of(kobj, struct class, kobj); if (class->class_release) class->class_release(class); } static const struct sysfs_ops linux_class_sysfs = { .show = linux_class_show, .store = linux_class_store, }; const struct kobj_type linux_class_ktype = { .release = linux_class_release, .sysfs_ops = &linux_class_sysfs }; static void linux_dev_release(struct kobject *kobj) { struct device *dev; dev = container_of(kobj, struct device, kobj); /* This is the precedence defined by linux. */ if (dev->release) dev->release(dev); else if (dev->class && dev->class->dev_release) dev->class->dev_release(dev); } static ssize_t linux_dev_show(struct kobject *kobj, struct attribute *attr, char *buf) { struct device_attribute *dattr; ssize_t error; dattr = container_of(attr, struct device_attribute, attr); error = -EIO; if (dattr->show) error = dattr->show(container_of(kobj, struct device, kobj), dattr, buf); return (error); } static ssize_t linux_dev_store(struct kobject *kobj, struct attribute *attr, const char *buf, size_t count) { struct device_attribute *dattr; ssize_t error; dattr = container_of(attr, struct device_attribute, attr); error = -EIO; if (dattr->store) error = dattr->store(container_of(kobj, struct device, kobj), dattr, buf, count); return (error); } static const struct sysfs_ops linux_dev_sysfs = { .show = linux_dev_show, .store = linux_dev_store, }; const struct kobj_type linux_dev_ktype = { .release = linux_dev_release, .sysfs_ops = &linux_dev_sysfs }; struct device * device_create(struct class *class, struct device *parent, dev_t devt, void *drvdata, const char *fmt, ...) { struct device *dev; va_list args; dev = kzalloc(sizeof(*dev), M_WAITOK); dev->parent = parent; dev->class = class; dev->devt = devt; dev->driver_data = drvdata; dev->release = linux_device_release; va_start(args, fmt); kobject_set_name_vargs(&dev->kobj, fmt, args); va_end(args); device_register(dev); return (dev); } int kobject_init_and_add(struct kobject *kobj, const struct kobj_type *ktype, struct kobject *parent, const char *fmt, ...) { va_list args; int error; kobject_init(kobj, ktype); kobj->ktype = ktype; kobj->parent = parent; kobj->name = NULL; va_start(args, fmt); error = kobject_set_name_vargs(kobj, fmt, args); va_end(args); if (error) return (error); return kobject_add_complete(kobj, parent); } static void linux_kq_lock(void *arg) { spinlock_t *s = arg; spin_lock(s); } static void linux_kq_unlock(void *arg) { spinlock_t *s = arg; spin_unlock(s); } static void linux_kq_lock_owned(void *arg) { #ifdef INVARIANTS spinlock_t *s = arg; mtx_assert(&s->m, MA_OWNED); #endif } static void linux_kq_lock_unowned(void *arg) { #ifdef INVARIANTS spinlock_t *s = arg; mtx_assert(&s->m, MA_NOTOWNED); #endif } static void linux_file_kqfilter_poll(struct linux_file *, int); struct linux_file * linux_file_alloc(void) { struct linux_file *filp; filp = kzalloc(sizeof(*filp), GFP_KERNEL); /* set initial refcount */ filp->f_count = 1; /* setup fields needed by kqueue support */ spin_lock_init(&filp->f_kqlock); knlist_init(&filp->f_selinfo.si_note, &filp->f_kqlock, linux_kq_lock, linux_kq_unlock, linux_kq_lock_owned, linux_kq_lock_unowned); return (filp); } void linux_file_free(struct linux_file *filp) { if (filp->_file == NULL) { if (filp->f_shmem != NULL) vm_object_deallocate(filp->f_shmem); kfree(filp); } else { /* * The close method of the character device or file * will free the linux_file structure: */ _fdrop(filp->_file, curthread); } } static int linux_cdev_pager_fault(vm_object_t vm_obj, vm_ooffset_t offset, int prot, vm_page_t *mres) { struct vm_area_struct *vmap; vmap = linux_cdev_handle_find(vm_obj->handle); MPASS(vmap != NULL); MPASS(vmap->vm_private_data == vm_obj->handle); if (likely(vmap->vm_ops != NULL && offset < vmap->vm_len)) { vm_paddr_t paddr = IDX_TO_OFF(vmap->vm_pfn) + offset; vm_page_t page; if (((*mres)->flags & PG_FICTITIOUS) != 0) { /* * If the passed in result page is a fake * page, update it with the new physical * address. */ page = *mres; vm_page_updatefake(page, paddr, vm_obj->memattr); } else { /* * Replace the passed in "mres" page with our * own fake page and free up the all of the * original pages. */ VM_OBJECT_WUNLOCK(vm_obj); page = vm_page_getfake(paddr, vm_obj->memattr); VM_OBJECT_WLOCK(vm_obj); vm_page_replace_checked(page, vm_obj, (*mres)->pindex, *mres); vm_page_free(*mres); *mres = page; } - page->valid = VM_PAGE_BITS_ALL; + vm_page_valid(page); return (VM_PAGER_OK); } return (VM_PAGER_FAIL); } static int linux_cdev_pager_populate(vm_object_t vm_obj, vm_pindex_t pidx, int fault_type, vm_prot_t max_prot, vm_pindex_t *first, vm_pindex_t *last) { struct vm_area_struct *vmap; int err; linux_set_current(curthread); /* get VM area structure */ vmap = linux_cdev_handle_find(vm_obj->handle); MPASS(vmap != NULL); MPASS(vmap->vm_private_data == vm_obj->handle); VM_OBJECT_WUNLOCK(vm_obj); down_write(&vmap->vm_mm->mmap_sem); if (unlikely(vmap->vm_ops == NULL)) { err = VM_FAULT_SIGBUS; } else { struct vm_fault vmf; /* fill out VM fault structure */ vmf.virtual_address = (void *)(uintptr_t)IDX_TO_OFF(pidx); vmf.flags = (fault_type & VM_PROT_WRITE) ? FAULT_FLAG_WRITE : 0; vmf.pgoff = 0; vmf.page = NULL; vmf.vma = vmap; vmap->vm_pfn_count = 0; vmap->vm_pfn_pcount = &vmap->vm_pfn_count; vmap->vm_obj = vm_obj; err = vmap->vm_ops->fault(vmap, &vmf); while (vmap->vm_pfn_count == 0 && err == VM_FAULT_NOPAGE) { kern_yield(PRI_USER); err = vmap->vm_ops->fault(vmap, &vmf); } } /* translate return code */ switch (err) { case VM_FAULT_OOM: err = VM_PAGER_AGAIN; break; case VM_FAULT_SIGBUS: err = VM_PAGER_BAD; break; case VM_FAULT_NOPAGE: /* * By contract the fault handler will return having * busied all the pages itself. If pidx is already * found in the object, it will simply xbusy the first * page and return with vm_pfn_count set to 1. */ *first = vmap->vm_pfn_first; *last = *first + vmap->vm_pfn_count - 1; err = VM_PAGER_OK; break; default: err = VM_PAGER_ERROR; break; } up_write(&vmap->vm_mm->mmap_sem); VM_OBJECT_WLOCK(vm_obj); return (err); } static struct rwlock linux_vma_lock; static TAILQ_HEAD(, vm_area_struct) linux_vma_head = TAILQ_HEAD_INITIALIZER(linux_vma_head); static void linux_cdev_handle_free(struct vm_area_struct *vmap) { /* Drop reference on vm_file */ if (vmap->vm_file != NULL) fput(vmap->vm_file); /* Drop reference on mm_struct */ mmput(vmap->vm_mm); kfree(vmap); } static void linux_cdev_handle_remove(struct vm_area_struct *vmap) { rw_wlock(&linux_vma_lock); TAILQ_REMOVE(&linux_vma_head, vmap, vm_entry); rw_wunlock(&linux_vma_lock); } static struct vm_area_struct * linux_cdev_handle_find(void *handle) { struct vm_area_struct *vmap; rw_rlock(&linux_vma_lock); TAILQ_FOREACH(vmap, &linux_vma_head, vm_entry) { if (vmap->vm_private_data == handle) break; } rw_runlock(&linux_vma_lock); return (vmap); } static int linux_cdev_pager_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t foff, struct ucred *cred, u_short *color) { MPASS(linux_cdev_handle_find(handle) != NULL); *color = 0; return (0); } static void linux_cdev_pager_dtor(void *handle) { const struct vm_operations_struct *vm_ops; struct vm_area_struct *vmap; vmap = linux_cdev_handle_find(handle); MPASS(vmap != NULL); /* * Remove handle before calling close operation to prevent * other threads from reusing the handle pointer. */ linux_cdev_handle_remove(vmap); down_write(&vmap->vm_mm->mmap_sem); vm_ops = vmap->vm_ops; if (likely(vm_ops != NULL)) vm_ops->close(vmap); up_write(&vmap->vm_mm->mmap_sem); linux_cdev_handle_free(vmap); } static struct cdev_pager_ops linux_cdev_pager_ops[2] = { { /* OBJT_MGTDEVICE */ .cdev_pg_populate = linux_cdev_pager_populate, .cdev_pg_ctor = linux_cdev_pager_ctor, .cdev_pg_dtor = linux_cdev_pager_dtor }, { /* OBJT_DEVICE */ .cdev_pg_fault = linux_cdev_pager_fault, .cdev_pg_ctor = linux_cdev_pager_ctor, .cdev_pg_dtor = linux_cdev_pager_dtor }, }; int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address, unsigned long size) { vm_object_t obj; vm_page_t m; obj = vma->vm_obj; if (obj == NULL || (obj->flags & OBJ_UNMANAGED) != 0) return (-ENOTSUP); VM_OBJECT_RLOCK(obj); for (m = vm_page_find_least(obj, OFF_TO_IDX(address)); m != NULL && m->pindex < OFF_TO_IDX(address + size); m = TAILQ_NEXT(m, listq)) pmap_remove_all(m); VM_OBJECT_RUNLOCK(obj); return (0); } static struct file_operations dummy_ldev_ops = { /* XXXKIB */ }; static struct linux_cdev dummy_ldev = { .ops = &dummy_ldev_ops, }; #define LDEV_SI_DTR 0x0001 #define LDEV_SI_REF 0x0002 static void linux_get_fop(struct linux_file *filp, const struct file_operations **fop, struct linux_cdev **dev) { struct linux_cdev *ldev; u_int siref; ldev = filp->f_cdev; *fop = filp->f_op; if (ldev != NULL) { for (siref = ldev->siref;;) { if ((siref & LDEV_SI_DTR) != 0) { ldev = &dummy_ldev; siref = ldev->siref; *fop = ldev->ops; MPASS((ldev->siref & LDEV_SI_DTR) == 0); } else if (atomic_fcmpset_int(&ldev->siref, &siref, siref + LDEV_SI_REF)) { break; } } } *dev = ldev; } static void linux_drop_fop(struct linux_cdev *ldev) { if (ldev == NULL) return; MPASS((ldev->siref & ~LDEV_SI_DTR) != 0); atomic_subtract_int(&ldev->siref, LDEV_SI_REF); } #define OPW(fp,td,code) ({ \ struct file *__fpop; \ __typeof(code) __retval; \ \ __fpop = (td)->td_fpop; \ (td)->td_fpop = (fp); \ __retval = (code); \ (td)->td_fpop = __fpop; \ __retval; \ }) static int linux_dev_fdopen(struct cdev *dev, int fflags, struct thread *td, struct file *file) { struct linux_cdev *ldev; struct linux_file *filp; const struct file_operations *fop; int error; ldev = dev->si_drv1; filp = linux_file_alloc(); filp->f_dentry = &filp->f_dentry_store; filp->f_op = ldev->ops; filp->f_mode = file->f_flag; filp->f_flags = file->f_flag; filp->f_vnode = file->f_vnode; filp->_file = file; refcount_acquire(&ldev->refs); filp->f_cdev = ldev; linux_set_current(td); linux_get_fop(filp, &fop, &ldev); if (fop->open != NULL) { error = -fop->open(file->f_vnode, filp); if (error != 0) { linux_drop_fop(ldev); linux_cdev_deref(filp->f_cdev); kfree(filp); return (error); } } /* hold on to the vnode - used for fstat() */ vhold(filp->f_vnode); /* release the file from devfs */ finit(file, filp->f_mode, DTYPE_DEV, filp, &linuxfileops); linux_drop_fop(ldev); return (ENXIO); } #define LINUX_IOCTL_MIN_PTR 0x10000UL #define LINUX_IOCTL_MAX_PTR (LINUX_IOCTL_MIN_PTR + IOCPARM_MAX) static inline int linux_remap_address(void **uaddr, size_t len) { uintptr_t uaddr_val = (uintptr_t)(*uaddr); if (unlikely(uaddr_val >= LINUX_IOCTL_MIN_PTR && uaddr_val < LINUX_IOCTL_MAX_PTR)) { struct task_struct *pts = current; if (pts == NULL) { *uaddr = NULL; return (1); } /* compute data offset */ uaddr_val -= LINUX_IOCTL_MIN_PTR; /* check that length is within bounds */ if ((len > IOCPARM_MAX) || (uaddr_val + len) > pts->bsd_ioctl_len) { *uaddr = NULL; return (1); } /* re-add kernel buffer address */ uaddr_val += (uintptr_t)pts->bsd_ioctl_data; /* update address location */ *uaddr = (void *)uaddr_val; return (1); } return (0); } int linux_copyin(const void *uaddr, void *kaddr, size_t len) { if (linux_remap_address(__DECONST(void **, &uaddr), len)) { if (uaddr == NULL) return (-EFAULT); memcpy(kaddr, uaddr, len); return (0); } return (-copyin(uaddr, kaddr, len)); } int linux_copyout(const void *kaddr, void *uaddr, size_t len) { if (linux_remap_address(&uaddr, len)) { if (uaddr == NULL) return (-EFAULT); memcpy(uaddr, kaddr, len); return (0); } return (-copyout(kaddr, uaddr, len)); } size_t linux_clear_user(void *_uaddr, size_t _len) { uint8_t *uaddr = _uaddr; size_t len = _len; /* make sure uaddr is aligned before going into the fast loop */ while (((uintptr_t)uaddr & 7) != 0 && len > 7) { if (subyte(uaddr, 0)) return (_len); uaddr++; len--; } /* zero 8 bytes at a time */ while (len > 7) { #ifdef __LP64__ if (suword64(uaddr, 0)) return (_len); #else if (suword32(uaddr, 0)) return (_len); if (suword32(uaddr + 4, 0)) return (_len); #endif uaddr += 8; len -= 8; } /* zero fill end, if any */ while (len > 0) { if (subyte(uaddr, 0)) return (_len); uaddr++; len--; } return (0); } int linux_access_ok(const void *uaddr, size_t len) { uintptr_t saddr; uintptr_t eaddr; /* get start and end address */ saddr = (uintptr_t)uaddr; eaddr = (uintptr_t)uaddr + len; /* verify addresses are valid for userspace */ return ((saddr == eaddr) || (eaddr > saddr && eaddr <= VM_MAXUSER_ADDRESS)); } /* * This function should return either EINTR or ERESTART depending on * the signal type sent to this thread: */ static int linux_get_error(struct task_struct *task, int error) { /* check for signal type interrupt code */ if (error == EINTR || error == ERESTARTSYS || error == ERESTART) { error = -linux_schedule_get_interrupt_value(task); if (error == 0) error = EINTR; } return (error); } static int linux_file_ioctl_sub(struct file *fp, struct linux_file *filp, const struct file_operations *fop, u_long cmd, caddr_t data, struct thread *td) { struct task_struct *task = current; unsigned size; int error; size = IOCPARM_LEN(cmd); /* refer to logic in sys_ioctl() */ if (size > 0) { /* * Setup hint for linux_copyin() and linux_copyout(). * * Background: Linux code expects a user-space address * while FreeBSD supplies a kernel-space address. */ task->bsd_ioctl_data = data; task->bsd_ioctl_len = size; data = (void *)LINUX_IOCTL_MIN_PTR; } else { /* fetch user-space pointer */ data = *(void **)data; } #if defined(__amd64__) if (td->td_proc->p_elf_machine == EM_386) { /* try the compat IOCTL handler first */ if (fop->compat_ioctl != NULL) { error = -OPW(fp, td, fop->compat_ioctl(filp, cmd, (u_long)data)); } else { error = ENOTTY; } /* fallback to the regular IOCTL handler, if any */ if (error == ENOTTY && fop->unlocked_ioctl != NULL) { error = -OPW(fp, td, fop->unlocked_ioctl(filp, cmd, (u_long)data)); } } else #endif { if (fop->unlocked_ioctl != NULL) { error = -OPW(fp, td, fop->unlocked_ioctl(filp, cmd, (u_long)data)); } else { error = ENOTTY; } } if (size > 0) { task->bsd_ioctl_data = NULL; task->bsd_ioctl_len = 0; } if (error == EWOULDBLOCK) { /* update kqfilter status, if any */ linux_file_kqfilter_poll(filp, LINUX_KQ_FLAG_HAS_READ | LINUX_KQ_FLAG_HAS_WRITE); } else { error = linux_get_error(task, error); } return (error); } #define LINUX_POLL_TABLE_NORMAL ((poll_table *)1) /* * This function atomically updates the poll wakeup state and returns * the previous state at the time of update. */ static uint8_t linux_poll_wakeup_state(atomic_t *v, const uint8_t *pstate) { int c, old; c = v->counter; while ((old = atomic_cmpxchg(v, c, pstate[c])) != c) c = old; return (c); } static int linux_poll_wakeup_callback(wait_queue_t *wq, unsigned int wq_state, int flags, void *key) { static const uint8_t state[LINUX_FWQ_STATE_MAX] = { [LINUX_FWQ_STATE_INIT] = LINUX_FWQ_STATE_INIT, /* NOP */ [LINUX_FWQ_STATE_NOT_READY] = LINUX_FWQ_STATE_NOT_READY, /* NOP */ [LINUX_FWQ_STATE_QUEUED] = LINUX_FWQ_STATE_READY, [LINUX_FWQ_STATE_READY] = LINUX_FWQ_STATE_READY, /* NOP */ }; struct linux_file *filp = container_of(wq, struct linux_file, f_wait_queue.wq); switch (linux_poll_wakeup_state(&filp->f_wait_queue.state, state)) { case LINUX_FWQ_STATE_QUEUED: linux_poll_wakeup(filp); return (1); default: return (0); } } void linux_poll_wait(struct linux_file *filp, wait_queue_head_t *wqh, poll_table *p) { static const uint8_t state[LINUX_FWQ_STATE_MAX] = { [LINUX_FWQ_STATE_INIT] = LINUX_FWQ_STATE_NOT_READY, [LINUX_FWQ_STATE_NOT_READY] = LINUX_FWQ_STATE_NOT_READY, /* NOP */ [LINUX_FWQ_STATE_QUEUED] = LINUX_FWQ_STATE_QUEUED, /* NOP */ [LINUX_FWQ_STATE_READY] = LINUX_FWQ_STATE_QUEUED, }; /* check if we are called inside the select system call */ if (p == LINUX_POLL_TABLE_NORMAL) selrecord(curthread, &filp->f_selinfo); switch (linux_poll_wakeup_state(&filp->f_wait_queue.state, state)) { case LINUX_FWQ_STATE_INIT: /* NOTE: file handles can only belong to one wait-queue */ filp->f_wait_queue.wqh = wqh; filp->f_wait_queue.wq.func = &linux_poll_wakeup_callback; add_wait_queue(wqh, &filp->f_wait_queue.wq); atomic_set(&filp->f_wait_queue.state, LINUX_FWQ_STATE_QUEUED); break; default: break; } } static void linux_poll_wait_dequeue(struct linux_file *filp) { static const uint8_t state[LINUX_FWQ_STATE_MAX] = { [LINUX_FWQ_STATE_INIT] = LINUX_FWQ_STATE_INIT, /* NOP */ [LINUX_FWQ_STATE_NOT_READY] = LINUX_FWQ_STATE_INIT, [LINUX_FWQ_STATE_QUEUED] = LINUX_FWQ_STATE_INIT, [LINUX_FWQ_STATE_READY] = LINUX_FWQ_STATE_INIT, }; seldrain(&filp->f_selinfo); switch (linux_poll_wakeup_state(&filp->f_wait_queue.state, state)) { case LINUX_FWQ_STATE_NOT_READY: case LINUX_FWQ_STATE_QUEUED: case LINUX_FWQ_STATE_READY: remove_wait_queue(filp->f_wait_queue.wqh, &filp->f_wait_queue.wq); break; default: break; } } void linux_poll_wakeup(struct linux_file *filp) { /* this function should be NULL-safe */ if (filp == NULL) return; selwakeup(&filp->f_selinfo); spin_lock(&filp->f_kqlock); filp->f_kqflags |= LINUX_KQ_FLAG_NEED_READ | LINUX_KQ_FLAG_NEED_WRITE; /* make sure the "knote" gets woken up */ KNOTE_LOCKED(&filp->f_selinfo.si_note, 1); spin_unlock(&filp->f_kqlock); } static void linux_file_kqfilter_detach(struct knote *kn) { struct linux_file *filp = kn->kn_hook; spin_lock(&filp->f_kqlock); knlist_remove(&filp->f_selinfo.si_note, kn, 1); spin_unlock(&filp->f_kqlock); } static int linux_file_kqfilter_read_event(struct knote *kn, long hint) { struct linux_file *filp = kn->kn_hook; mtx_assert(&filp->f_kqlock.m, MA_OWNED); return ((filp->f_kqflags & LINUX_KQ_FLAG_NEED_READ) ? 1 : 0); } static int linux_file_kqfilter_write_event(struct knote *kn, long hint) { struct linux_file *filp = kn->kn_hook; mtx_assert(&filp->f_kqlock.m, MA_OWNED); return ((filp->f_kqflags & LINUX_KQ_FLAG_NEED_WRITE) ? 1 : 0); } static struct filterops linux_dev_kqfiltops_read = { .f_isfd = 1, .f_detach = linux_file_kqfilter_detach, .f_event = linux_file_kqfilter_read_event, }; static struct filterops linux_dev_kqfiltops_write = { .f_isfd = 1, .f_detach = linux_file_kqfilter_detach, .f_event = linux_file_kqfilter_write_event, }; static void linux_file_kqfilter_poll(struct linux_file *filp, int kqflags) { struct thread *td; const struct file_operations *fop; struct linux_cdev *ldev; int temp; if ((filp->f_kqflags & kqflags) == 0) return; td = curthread; linux_get_fop(filp, &fop, &ldev); /* get the latest polling state */ temp = OPW(filp->_file, td, fop->poll(filp, NULL)); linux_drop_fop(ldev); spin_lock(&filp->f_kqlock); /* clear kqflags */ filp->f_kqflags &= ~(LINUX_KQ_FLAG_NEED_READ | LINUX_KQ_FLAG_NEED_WRITE); /* update kqflags */ if ((temp & (POLLIN | POLLOUT)) != 0) { if ((temp & POLLIN) != 0) filp->f_kqflags |= LINUX_KQ_FLAG_NEED_READ; if ((temp & POLLOUT) != 0) filp->f_kqflags |= LINUX_KQ_FLAG_NEED_WRITE; /* make sure the "knote" gets woken up */ KNOTE_LOCKED(&filp->f_selinfo.si_note, 0); } spin_unlock(&filp->f_kqlock); } static int linux_file_kqfilter(struct file *file, struct knote *kn) { struct linux_file *filp; struct thread *td; int error; td = curthread; filp = (struct linux_file *)file->f_data; filp->f_flags = file->f_flag; if (filp->f_op->poll == NULL) return (EINVAL); spin_lock(&filp->f_kqlock); switch (kn->kn_filter) { case EVFILT_READ: filp->f_kqflags |= LINUX_KQ_FLAG_HAS_READ; kn->kn_fop = &linux_dev_kqfiltops_read; kn->kn_hook = filp; knlist_add(&filp->f_selinfo.si_note, kn, 1); error = 0; break; case EVFILT_WRITE: filp->f_kqflags |= LINUX_KQ_FLAG_HAS_WRITE; kn->kn_fop = &linux_dev_kqfiltops_write; kn->kn_hook = filp; knlist_add(&filp->f_selinfo.si_note, kn, 1); error = 0; break; default: error = EINVAL; break; } spin_unlock(&filp->f_kqlock); if (error == 0) { linux_set_current(td); /* update kqfilter status, if any */ linux_file_kqfilter_poll(filp, LINUX_KQ_FLAG_HAS_READ | LINUX_KQ_FLAG_HAS_WRITE); } return (error); } static int linux_file_mmap_single(struct file *fp, const struct file_operations *fop, vm_ooffset_t *offset, vm_size_t size, struct vm_object **object, int nprot, struct thread *td) { struct task_struct *task; struct vm_area_struct *vmap; struct mm_struct *mm; struct linux_file *filp; vm_memattr_t attr; int error; filp = (struct linux_file *)fp->f_data; filp->f_flags = fp->f_flag; if (fop->mmap == NULL) return (EOPNOTSUPP); linux_set_current(td); /* * The same VM object might be shared by multiple processes * and the mm_struct is usually freed when a process exits. * * The atomic reference below makes sure the mm_struct is * available as long as the vmap is in the linux_vma_head. */ task = current; mm = task->mm; if (atomic_inc_not_zero(&mm->mm_users) == 0) return (EINVAL); vmap = kzalloc(sizeof(*vmap), GFP_KERNEL); vmap->vm_start = 0; vmap->vm_end = size; vmap->vm_pgoff = *offset / PAGE_SIZE; vmap->vm_pfn = 0; vmap->vm_flags = vmap->vm_page_prot = (nprot & VM_PROT_ALL); vmap->vm_ops = NULL; vmap->vm_file = get_file(filp); vmap->vm_mm = mm; if (unlikely(down_write_killable(&vmap->vm_mm->mmap_sem))) { error = linux_get_error(task, EINTR); } else { error = -OPW(fp, td, fop->mmap(filp, vmap)); error = linux_get_error(task, error); up_write(&vmap->vm_mm->mmap_sem); } if (error != 0) { linux_cdev_handle_free(vmap); return (error); } attr = pgprot2cachemode(vmap->vm_page_prot); if (vmap->vm_ops != NULL) { struct vm_area_struct *ptr; void *vm_private_data; bool vm_no_fault; if (vmap->vm_ops->open == NULL || vmap->vm_ops->close == NULL || vmap->vm_private_data == NULL) { /* free allocated VM area struct */ linux_cdev_handle_free(vmap); return (EINVAL); } vm_private_data = vmap->vm_private_data; rw_wlock(&linux_vma_lock); TAILQ_FOREACH(ptr, &linux_vma_head, vm_entry) { if (ptr->vm_private_data == vm_private_data) break; } /* check if there is an existing VM area struct */ if (ptr != NULL) { /* check if the VM area structure is invalid */ if (ptr->vm_ops == NULL || ptr->vm_ops->open == NULL || ptr->vm_ops->close == NULL) { error = ESTALE; vm_no_fault = 1; } else { error = EEXIST; vm_no_fault = (ptr->vm_ops->fault == NULL); } } else { /* insert VM area structure into list */ TAILQ_INSERT_TAIL(&linux_vma_head, vmap, vm_entry); error = 0; vm_no_fault = (vmap->vm_ops->fault == NULL); } rw_wunlock(&linux_vma_lock); if (error != 0) { /* free allocated VM area struct */ linux_cdev_handle_free(vmap); /* check for stale VM area struct */ if (error != EEXIST) return (error); } /* check if there is no fault handler */ if (vm_no_fault) { *object = cdev_pager_allocate(vm_private_data, OBJT_DEVICE, &linux_cdev_pager_ops[1], size, nprot, *offset, td->td_ucred); } else { *object = cdev_pager_allocate(vm_private_data, OBJT_MGTDEVICE, &linux_cdev_pager_ops[0], size, nprot, *offset, td->td_ucred); } /* check if allocating the VM object failed */ if (*object == NULL) { if (error == 0) { /* remove VM area struct from list */ linux_cdev_handle_remove(vmap); /* free allocated VM area struct */ linux_cdev_handle_free(vmap); } return (EINVAL); } } else { struct sglist *sg; sg = sglist_alloc(1, M_WAITOK); sglist_append_phys(sg, (vm_paddr_t)vmap->vm_pfn << PAGE_SHIFT, vmap->vm_len); *object = vm_pager_allocate(OBJT_SG, sg, vmap->vm_len, nprot, 0, td->td_ucred); linux_cdev_handle_free(vmap); if (*object == NULL) { sglist_free(sg); return (EINVAL); } } if (attr != VM_MEMATTR_DEFAULT) { VM_OBJECT_WLOCK(*object); vm_object_set_memattr(*object, attr); VM_OBJECT_WUNLOCK(*object); } *offset = 0; return (0); } struct cdevsw linuxcdevsw = { .d_version = D_VERSION, .d_fdopen = linux_dev_fdopen, .d_name = "lkpidev", }; static int linux_file_read(struct file *file, struct uio *uio, struct ucred *active_cred, int flags, struct thread *td) { struct linux_file *filp; const struct file_operations *fop; struct linux_cdev *ldev; ssize_t bytes; int error; error = 0; filp = (struct linux_file *)file->f_data; filp->f_flags = file->f_flag; /* XXX no support for I/O vectors currently */ if (uio->uio_iovcnt != 1) return (EOPNOTSUPP); if (uio->uio_resid > DEVFS_IOSIZE_MAX) return (EINVAL); linux_set_current(td); linux_get_fop(filp, &fop, &ldev); if (fop->read != NULL) { bytes = OPW(file, td, fop->read(filp, uio->uio_iov->iov_base, uio->uio_iov->iov_len, &uio->uio_offset)); if (bytes >= 0) { uio->uio_iov->iov_base = ((uint8_t *)uio->uio_iov->iov_base) + bytes; uio->uio_iov->iov_len -= bytes; uio->uio_resid -= bytes; } else { error = linux_get_error(current, -bytes); } } else error = ENXIO; /* update kqfilter status, if any */ linux_file_kqfilter_poll(filp, LINUX_KQ_FLAG_HAS_READ); linux_drop_fop(ldev); return (error); } static int linux_file_write(struct file *file, struct uio *uio, struct ucred *active_cred, int flags, struct thread *td) { struct linux_file *filp; const struct file_operations *fop; struct linux_cdev *ldev; ssize_t bytes; int error; filp = (struct linux_file *)file->f_data; filp->f_flags = file->f_flag; /* XXX no support for I/O vectors currently */ if (uio->uio_iovcnt != 1) return (EOPNOTSUPP); if (uio->uio_resid > DEVFS_IOSIZE_MAX) return (EINVAL); linux_set_current(td); linux_get_fop(filp, &fop, &ldev); if (fop->write != NULL) { bytes = OPW(file, td, fop->write(filp, uio->uio_iov->iov_base, uio->uio_iov->iov_len, &uio->uio_offset)); if (bytes >= 0) { uio->uio_iov->iov_base = ((uint8_t *)uio->uio_iov->iov_base) + bytes; uio->uio_iov->iov_len -= bytes; uio->uio_resid -= bytes; error = 0; } else { error = linux_get_error(current, -bytes); } } else error = ENXIO; /* update kqfilter status, if any */ linux_file_kqfilter_poll(filp, LINUX_KQ_FLAG_HAS_WRITE); linux_drop_fop(ldev); return (error); } static int linux_file_poll(struct file *file, int events, struct ucred *active_cred, struct thread *td) { struct linux_file *filp; const struct file_operations *fop; struct linux_cdev *ldev; int revents; filp = (struct linux_file *)file->f_data; filp->f_flags = file->f_flag; linux_set_current(td); linux_get_fop(filp, &fop, &ldev); if (fop->poll != NULL) { revents = OPW(file, td, fop->poll(filp, LINUX_POLL_TABLE_NORMAL)) & events; } else { revents = 0; } linux_drop_fop(ldev); return (revents); } static int linux_file_close(struct file *file, struct thread *td) { struct linux_file *filp; const struct file_operations *fop; struct linux_cdev *ldev; int error; filp = (struct linux_file *)file->f_data; KASSERT(file_count(filp) == 0, ("File refcount(%d) is not zero", file_count(filp))); error = 0; filp->f_flags = file->f_flag; linux_set_current(td); linux_poll_wait_dequeue(filp); linux_get_fop(filp, &fop, &ldev); if (fop->release != NULL) error = -OPW(file, td, fop->release(filp->f_vnode, filp)); funsetown(&filp->f_sigio); if (filp->f_vnode != NULL) vdrop(filp->f_vnode); linux_drop_fop(ldev); if (filp->f_cdev != NULL) linux_cdev_deref(filp->f_cdev); kfree(filp); return (error); } static int linux_file_ioctl(struct file *fp, u_long cmd, void *data, struct ucred *cred, struct thread *td) { struct linux_file *filp; const struct file_operations *fop; struct linux_cdev *ldev; int error; error = 0; filp = (struct linux_file *)fp->f_data; filp->f_flags = fp->f_flag; linux_get_fop(filp, &fop, &ldev); linux_set_current(td); switch (cmd) { case FIONBIO: break; case FIOASYNC: if (fop->fasync == NULL) break; error = -OPW(fp, td, fop->fasync(0, filp, fp->f_flag & FASYNC)); break; case FIOSETOWN: error = fsetown(*(int *)data, &filp->f_sigio); if (error == 0) { if (fop->fasync == NULL) break; error = -OPW(fp, td, fop->fasync(0, filp, fp->f_flag & FASYNC)); } break; case FIOGETOWN: *(int *)data = fgetown(&filp->f_sigio); break; default: error = linux_file_ioctl_sub(fp, filp, fop, cmd, data, td); break; } linux_drop_fop(ldev); return (error); } static int linux_file_mmap_sub(struct thread *td, vm_size_t objsize, vm_prot_t prot, vm_prot_t *maxprotp, int *flagsp, struct file *fp, vm_ooffset_t *foff, const struct file_operations *fop, vm_object_t *objp) { /* * Character devices do not provide private mappings * of any kind: */ if ((*maxprotp & VM_PROT_WRITE) == 0 && (prot & VM_PROT_WRITE) != 0) return (EACCES); if ((*flagsp & (MAP_PRIVATE | MAP_COPY)) != 0) return (EINVAL); return (linux_file_mmap_single(fp, fop, foff, objsize, objp, (int)prot, td)); } static int linux_file_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size, vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff, struct thread *td) { struct linux_file *filp; const struct file_operations *fop; struct linux_cdev *ldev; struct mount *mp; struct vnode *vp; vm_object_t object; vm_prot_t maxprot; int error; filp = (struct linux_file *)fp->f_data; vp = filp->f_vnode; if (vp == NULL) return (EOPNOTSUPP); /* * Ensure that file and memory protections are * compatible. */ mp = vp->v_mount; if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) { maxprot = VM_PROT_NONE; if ((prot & VM_PROT_EXECUTE) != 0) return (EACCES); } else maxprot = VM_PROT_EXECUTE; if ((fp->f_flag & FREAD) != 0) maxprot |= VM_PROT_READ; else if ((prot & VM_PROT_READ) != 0) return (EACCES); /* * If we are sharing potential changes via MAP_SHARED and we * are trying to get write permission although we opened it * without asking for it, bail out. * * Note that most character devices always share mappings. * * Rely on linux_file_mmap_sub() to fail invalid MAP_PRIVATE * requests rather than doing it here. */ if ((flags & MAP_SHARED) != 0) { if ((fp->f_flag & FWRITE) != 0) maxprot |= VM_PROT_WRITE; else if ((prot & VM_PROT_WRITE) != 0) return (EACCES); } maxprot &= cap_maxprot; linux_get_fop(filp, &fop, &ldev); error = linux_file_mmap_sub(td, size, prot, &maxprot, &flags, fp, &foff, fop, &object); if (error != 0) goto out; error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object, foff, FALSE, td); if (error != 0) vm_object_deallocate(object); out: linux_drop_fop(ldev); return (error); } static int linux_file_stat(struct file *fp, struct stat *sb, struct ucred *active_cred, struct thread *td) { struct linux_file *filp; struct vnode *vp; int error; filp = (struct linux_file *)fp->f_data; if (filp->f_vnode == NULL) return (EOPNOTSUPP); vp = filp->f_vnode; vn_lock(vp, LK_SHARED | LK_RETRY); error = vn_stat(vp, sb, td->td_ucred, NOCRED, td); VOP_UNLOCK(vp, 0); return (error); } static int linux_file_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp) { struct linux_file *filp; struct vnode *vp; int error; filp = fp->f_data; vp = filp->f_vnode; if (vp == NULL) { error = 0; kif->kf_type = KF_TYPE_DEV; } else { vref(vp); FILEDESC_SUNLOCK(fdp); error = vn_fill_kinfo_vnode(vp, kif); vrele(vp); kif->kf_type = KF_TYPE_VNODE; FILEDESC_SLOCK(fdp); } return (error); } unsigned int linux_iminor(struct inode *inode) { struct linux_cdev *ldev; if (inode == NULL || inode->v_rdev == NULL || inode->v_rdev->si_devsw != &linuxcdevsw) return (-1U); ldev = inode->v_rdev->si_drv1; if (ldev == NULL) return (-1U); return (minor(ldev->dev)); } struct fileops linuxfileops = { .fo_read = linux_file_read, .fo_write = linux_file_write, .fo_truncate = invfo_truncate, .fo_kqfilter = linux_file_kqfilter, .fo_stat = linux_file_stat, .fo_fill_kinfo = linux_file_fill_kinfo, .fo_poll = linux_file_poll, .fo_close = linux_file_close, .fo_ioctl = linux_file_ioctl, .fo_mmap = linux_file_mmap, .fo_chmod = invfo_chmod, .fo_chown = invfo_chown, .fo_sendfile = invfo_sendfile, .fo_flags = DFLAG_PASSABLE, }; /* * Hash of vmmap addresses. This is infrequently accessed and does not * need to be particularly large. This is done because we must store the * caller's idea of the map size to properly unmap. */ struct vmmap { LIST_ENTRY(vmmap) vm_next; void *vm_addr; unsigned long vm_size; }; struct vmmaphd { struct vmmap *lh_first; }; #define VMMAP_HASH_SIZE 64 #define VMMAP_HASH_MASK (VMMAP_HASH_SIZE - 1) #define VM_HASH(addr) ((uintptr_t)(addr) >> PAGE_SHIFT) & VMMAP_HASH_MASK static struct vmmaphd vmmaphead[VMMAP_HASH_SIZE]; static struct mtx vmmaplock; static void vmmap_add(void *addr, unsigned long size) { struct vmmap *vmmap; vmmap = kmalloc(sizeof(*vmmap), GFP_KERNEL); mtx_lock(&vmmaplock); vmmap->vm_size = size; vmmap->vm_addr = addr; LIST_INSERT_HEAD(&vmmaphead[VM_HASH(addr)], vmmap, vm_next); mtx_unlock(&vmmaplock); } static struct vmmap * vmmap_remove(void *addr) { struct vmmap *vmmap; mtx_lock(&vmmaplock); LIST_FOREACH(vmmap, &vmmaphead[VM_HASH(addr)], vm_next) if (vmmap->vm_addr == addr) break; if (vmmap) LIST_REMOVE(vmmap, vm_next); mtx_unlock(&vmmaplock); return (vmmap); } #if defined(__i386__) || defined(__amd64__) || defined(__powerpc__) || defined(__aarch64__) void * _ioremap_attr(vm_paddr_t phys_addr, unsigned long size, int attr) { void *addr; addr = pmap_mapdev_attr(phys_addr, size, attr); if (addr == NULL) return (NULL); vmmap_add(addr, size); return (addr); } #endif void iounmap(void *addr) { struct vmmap *vmmap; vmmap = vmmap_remove(addr); if (vmmap == NULL) return; #if defined(__i386__) || defined(__amd64__) || defined(__powerpc__) || defined(__aarch64__) pmap_unmapdev((vm_offset_t)addr, vmmap->vm_size); #endif kfree(vmmap); } void * vmap(struct page **pages, unsigned int count, unsigned long flags, int prot) { vm_offset_t off; size_t size; size = count * PAGE_SIZE; off = kva_alloc(size); if (off == 0) return (NULL); vmmap_add((void *)off, size); pmap_qenter(off, pages, count); return ((void *)off); } void vunmap(void *addr) { struct vmmap *vmmap; vmmap = vmmap_remove(addr); if (vmmap == NULL) return; pmap_qremove((vm_offset_t)addr, vmmap->vm_size / PAGE_SIZE); kva_free((vm_offset_t)addr, vmmap->vm_size); kfree(vmmap); } char * kvasprintf(gfp_t gfp, const char *fmt, va_list ap) { unsigned int len; char *p; va_list aq; va_copy(aq, ap); len = vsnprintf(NULL, 0, fmt, aq); va_end(aq); p = kmalloc(len + 1, gfp); if (p != NULL) vsnprintf(p, len + 1, fmt, ap); return (p); } char * kasprintf(gfp_t gfp, const char *fmt, ...) { va_list ap; char *p; va_start(ap, fmt); p = kvasprintf(gfp, fmt, ap); va_end(ap); return (p); } static void linux_timer_callback_wrapper(void *context) { struct timer_list *timer; linux_set_current(curthread); timer = context; timer->function(timer->data); } void mod_timer(struct timer_list *timer, int expires) { timer->expires = expires; callout_reset(&timer->callout, linux_timer_jiffies_until(expires), &linux_timer_callback_wrapper, timer); } void add_timer(struct timer_list *timer) { callout_reset(&timer->callout, linux_timer_jiffies_until(timer->expires), &linux_timer_callback_wrapper, timer); } void add_timer_on(struct timer_list *timer, int cpu) { callout_reset_on(&timer->callout, linux_timer_jiffies_until(timer->expires), &linux_timer_callback_wrapper, timer, cpu); } int del_timer(struct timer_list *timer) { if (callout_stop(&(timer)->callout) == -1) return (0); return (1); } static void linux_timer_init(void *arg) { /* * Compute an internal HZ value which can divide 2**32 to * avoid timer rounding problems when the tick value wraps * around 2**32: */ linux_timer_hz_mask = 1; while (linux_timer_hz_mask < (unsigned long)hz) linux_timer_hz_mask *= 2; linux_timer_hz_mask--; } SYSINIT(linux_timer, SI_SUB_DRIVERS, SI_ORDER_FIRST, linux_timer_init, NULL); void linux_complete_common(struct completion *c, int all) { int wakeup_swapper; sleepq_lock(c); if (all) { c->done = UINT_MAX; wakeup_swapper = sleepq_broadcast(c, SLEEPQ_SLEEP, 0, 0); } else { if (c->done != UINT_MAX) c->done++; wakeup_swapper = sleepq_signal(c, SLEEPQ_SLEEP, 0, 0); } sleepq_release(c); if (wakeup_swapper) kick_proc0(); } /* * Indefinite wait for done != 0 with or without signals. */ int linux_wait_for_common(struct completion *c, int flags) { struct task_struct *task; int error; if (SCHEDULER_STOPPED()) return (0); task = current; if (flags != 0) flags = SLEEPQ_INTERRUPTIBLE | SLEEPQ_SLEEP; else flags = SLEEPQ_SLEEP; error = 0; for (;;) { sleepq_lock(c); if (c->done) break; sleepq_add(c, NULL, "completion", flags, 0); if (flags & SLEEPQ_INTERRUPTIBLE) { DROP_GIANT(); error = -sleepq_wait_sig(c, 0); PICKUP_GIANT(); if (error != 0) { linux_schedule_save_interrupt_value(task, error); error = -ERESTARTSYS; goto intr; } } else { DROP_GIANT(); sleepq_wait(c, 0); PICKUP_GIANT(); } } if (c->done != UINT_MAX) c->done--; sleepq_release(c); intr: return (error); } /* * Time limited wait for done != 0 with or without signals. */ int linux_wait_for_timeout_common(struct completion *c, int timeout, int flags) { struct task_struct *task; int end = jiffies + timeout; int error; if (SCHEDULER_STOPPED()) return (0); task = current; if (flags != 0) flags = SLEEPQ_INTERRUPTIBLE | SLEEPQ_SLEEP; else flags = SLEEPQ_SLEEP; for (;;) { sleepq_lock(c); if (c->done) break; sleepq_add(c, NULL, "completion", flags, 0); sleepq_set_timeout(c, linux_timer_jiffies_until(end)); DROP_GIANT(); if (flags & SLEEPQ_INTERRUPTIBLE) error = -sleepq_timedwait_sig(c, 0); else error = -sleepq_timedwait(c, 0); PICKUP_GIANT(); if (error != 0) { /* check for timeout */ if (error == -EWOULDBLOCK) { error = 0; /* timeout */ } else { /* signal happened */ linux_schedule_save_interrupt_value(task, error); error = -ERESTARTSYS; } goto done; } } if (c->done != UINT_MAX) c->done--; sleepq_release(c); /* return how many jiffies are left */ error = linux_timer_jiffies_until(end); done: return (error); } int linux_try_wait_for_completion(struct completion *c) { int isdone; sleepq_lock(c); isdone = (c->done != 0); if (c->done != 0 && c->done != UINT_MAX) c->done--; sleepq_release(c); return (isdone); } int linux_completion_done(struct completion *c) { int isdone; sleepq_lock(c); isdone = (c->done != 0); sleepq_release(c); return (isdone); } static void linux_cdev_deref(struct linux_cdev *ldev) { if (refcount_release(&ldev->refs)) kfree(ldev); } static void linux_cdev_release(struct kobject *kobj) { struct linux_cdev *cdev; struct kobject *parent; cdev = container_of(kobj, struct linux_cdev, kobj); parent = kobj->parent; linux_destroy_dev(cdev); linux_cdev_deref(cdev); kobject_put(parent); } static void linux_cdev_static_release(struct kobject *kobj) { struct linux_cdev *cdev; struct kobject *parent; cdev = container_of(kobj, struct linux_cdev, kobj); parent = kobj->parent; linux_destroy_dev(cdev); kobject_put(parent); } void linux_destroy_dev(struct linux_cdev *ldev) { if (ldev->cdev == NULL) return; MPASS((ldev->siref & LDEV_SI_DTR) == 0); atomic_set_int(&ldev->siref, LDEV_SI_DTR); while ((atomic_load_int(&ldev->siref) & ~LDEV_SI_DTR) != 0) pause("ldevdtr", hz / 4); destroy_dev(ldev->cdev); ldev->cdev = NULL; } const struct kobj_type linux_cdev_ktype = { .release = linux_cdev_release, }; const struct kobj_type linux_cdev_static_ktype = { .release = linux_cdev_static_release, }; static void linux_handle_ifnet_link_event(void *arg, struct ifnet *ifp, int linkstate) { struct notifier_block *nb; nb = arg; if (linkstate == LINK_STATE_UP) nb->notifier_call(nb, NETDEV_UP, ifp); else nb->notifier_call(nb, NETDEV_DOWN, ifp); } static void linux_handle_ifnet_arrival_event(void *arg, struct ifnet *ifp) { struct notifier_block *nb; nb = arg; nb->notifier_call(nb, NETDEV_REGISTER, ifp); } static void linux_handle_ifnet_departure_event(void *arg, struct ifnet *ifp) { struct notifier_block *nb; nb = arg; nb->notifier_call(nb, NETDEV_UNREGISTER, ifp); } static void linux_handle_iflladdr_event(void *arg, struct ifnet *ifp) { struct notifier_block *nb; nb = arg; nb->notifier_call(nb, NETDEV_CHANGEADDR, ifp); } static void linux_handle_ifaddr_event(void *arg, struct ifnet *ifp) { struct notifier_block *nb; nb = arg; nb->notifier_call(nb, NETDEV_CHANGEIFADDR, ifp); } int register_netdevice_notifier(struct notifier_block *nb) { nb->tags[NETDEV_UP] = EVENTHANDLER_REGISTER( ifnet_link_event, linux_handle_ifnet_link_event, nb, 0); nb->tags[NETDEV_REGISTER] = EVENTHANDLER_REGISTER( ifnet_arrival_event, linux_handle_ifnet_arrival_event, nb, 0); nb->tags[NETDEV_UNREGISTER] = EVENTHANDLER_REGISTER( ifnet_departure_event, linux_handle_ifnet_departure_event, nb, 0); nb->tags[NETDEV_CHANGEADDR] = EVENTHANDLER_REGISTER( iflladdr_event, linux_handle_iflladdr_event, nb, 0); return (0); } int register_inetaddr_notifier(struct notifier_block *nb) { nb->tags[NETDEV_CHANGEIFADDR] = EVENTHANDLER_REGISTER( ifaddr_event, linux_handle_ifaddr_event, nb, 0); return (0); } int unregister_netdevice_notifier(struct notifier_block *nb) { EVENTHANDLER_DEREGISTER(ifnet_link_event, nb->tags[NETDEV_UP]); EVENTHANDLER_DEREGISTER(ifnet_arrival_event, nb->tags[NETDEV_REGISTER]); EVENTHANDLER_DEREGISTER(ifnet_departure_event, nb->tags[NETDEV_UNREGISTER]); EVENTHANDLER_DEREGISTER(iflladdr_event, nb->tags[NETDEV_CHANGEADDR]); return (0); } int unregister_inetaddr_notifier(struct notifier_block *nb) { EVENTHANDLER_DEREGISTER(ifaddr_event, nb->tags[NETDEV_CHANGEIFADDR]); return (0); } struct list_sort_thunk { int (*cmp)(void *, struct list_head *, struct list_head *); void *priv; }; static inline int linux_le_cmp(void *priv, const void *d1, const void *d2) { struct list_head *le1, *le2; struct list_sort_thunk *thunk; thunk = priv; le1 = *(__DECONST(struct list_head **, d1)); le2 = *(__DECONST(struct list_head **, d2)); return ((thunk->cmp)(thunk->priv, le1, le2)); } void list_sort(void *priv, struct list_head *head, int (*cmp)(void *priv, struct list_head *a, struct list_head *b)) { struct list_sort_thunk thunk; struct list_head **ar, *le; size_t count, i; count = 0; list_for_each(le, head) count++; ar = malloc(sizeof(struct list_head *) * count, M_KMALLOC, M_WAITOK); i = 0; list_for_each(le, head) ar[i++] = le; thunk.cmp = cmp; thunk.priv = priv; qsort_r(ar, count, sizeof(struct list_head *), &thunk, linux_le_cmp); INIT_LIST_HEAD(head); for (i = 0; i < count; i++) list_add_tail(ar[i], head); free(ar, M_KMALLOC); } void linux_irq_handler(void *ent) { struct irq_ent *irqe; linux_set_current(curthread); irqe = ent; irqe->handler(irqe->irq, irqe->arg); } #if defined(__i386__) || defined(__amd64__) int linux_wbinvd_on_all_cpus(void) { pmap_invalidate_cache(); return (0); } #endif int linux_on_each_cpu(void callback(void *), void *data) { smp_rendezvous(smp_no_rendezvous_barrier, callback, smp_no_rendezvous_barrier, data); return (0); } int linux_in_atomic(void) { return ((curthread->td_pflags & TDP_NOFAULTING) != 0); } struct linux_cdev * linux_find_cdev(const char *name, unsigned major, unsigned minor) { dev_t dev = MKDEV(major, minor); struct cdev *cdev; dev_lock(); LIST_FOREACH(cdev, &linuxcdevsw.d_devs, si_list) { struct linux_cdev *ldev = cdev->si_drv1; if (ldev->dev == dev && strcmp(kobject_name(&ldev->kobj), name) == 0) { break; } } dev_unlock(); return (cdev != NULL ? cdev->si_drv1 : NULL); } int __register_chrdev(unsigned int major, unsigned int baseminor, unsigned int count, const char *name, const struct file_operations *fops) { struct linux_cdev *cdev; int ret = 0; int i; for (i = baseminor; i < baseminor + count; i++) { cdev = cdev_alloc(); cdev->ops = fops; kobject_set_name(&cdev->kobj, name); ret = cdev_add(cdev, makedev(major, i), 1); if (ret != 0) break; } return (ret); } int __register_chrdev_p(unsigned int major, unsigned int baseminor, unsigned int count, const char *name, const struct file_operations *fops, uid_t uid, gid_t gid, int mode) { struct linux_cdev *cdev; int ret = 0; int i; for (i = baseminor; i < baseminor + count; i++) { cdev = cdev_alloc(); cdev->ops = fops; kobject_set_name(&cdev->kobj, name); ret = cdev_add_ext(cdev, makedev(major, i), uid, gid, mode); if (ret != 0) break; } return (ret); } void __unregister_chrdev(unsigned int major, unsigned int baseminor, unsigned int count, const char *name) { struct linux_cdev *cdevp; int i; for (i = baseminor; i < baseminor + count; i++) { cdevp = linux_find_cdev(name, major, i); if (cdevp != NULL) cdev_del(cdevp); } } void linux_dump_stack(void) { #ifdef STACK struct stack st; stack_zero(&st); stack_save(&st); stack_print(&st); #endif } #if defined(__i386__) || defined(__amd64__) bool linux_cpu_has_clflush; #endif static void linux_compat_init(void *arg) { struct sysctl_oid *rootoid; int i; #if defined(__i386__) || defined(__amd64__) linux_cpu_has_clflush = (cpu_feature & CPUID_CLFSH); #endif rw_init(&linux_vma_lock, "lkpi-vma-lock"); rootoid = SYSCTL_ADD_ROOT_NODE(NULL, OID_AUTO, "sys", CTLFLAG_RD|CTLFLAG_MPSAFE, NULL, "sys"); kobject_init(&linux_class_root, &linux_class_ktype); kobject_set_name(&linux_class_root, "class"); linux_class_root.oidp = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(rootoid), OID_AUTO, "class", CTLFLAG_RD|CTLFLAG_MPSAFE, NULL, "class"); kobject_init(&linux_root_device.kobj, &linux_dev_ktype); kobject_set_name(&linux_root_device.kobj, "device"); linux_root_device.kobj.oidp = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(rootoid), OID_AUTO, "device", CTLFLAG_RD, NULL, "device"); linux_root_device.bsddev = root_bus; linux_class_misc.name = "misc"; class_register(&linux_class_misc); INIT_LIST_HEAD(&pci_drivers); INIT_LIST_HEAD(&pci_devices); spin_lock_init(&pci_lock); mtx_init(&vmmaplock, "IO Map lock", NULL, MTX_DEF); for (i = 0; i < VMMAP_HASH_SIZE; i++) LIST_INIT(&vmmaphead[i]); } SYSINIT(linux_compat, SI_SUB_DRIVERS, SI_ORDER_SECOND, linux_compat_init, NULL); static void linux_compat_uninit(void *arg) { linux_kobject_kfree_name(&linux_class_root); linux_kobject_kfree_name(&linux_root_device.kobj); linux_kobject_kfree_name(&linux_class_misc.kobj); mtx_destroy(&vmmaplock); spin_lock_destroy(&pci_lock); rw_destroy(&linux_vma_lock); } SYSUNINIT(linux_compat, SI_SUB_DRIVERS, SI_ORDER_SECOND, linux_compat_uninit, NULL); /* * NOTE: Linux frequently uses "unsigned long" for pointer to integer * conversion and vice versa, where in FreeBSD "uintptr_t" would be * used. Assert these types have the same size, else some parts of the * LinuxKPI may not work like expected: */ CTASSERT(sizeof(unsigned long) == sizeof(uintptr_t)); Index: head/sys/dev/drm2/ttm/ttm_bo_vm.c =================================================================== --- head/sys/dev/drm2/ttm/ttm_bo_vm.c (revision 353538) +++ head/sys/dev/drm2/ttm/ttm_bo_vm.c (revision 353539) @@ -1,555 +1,555 @@ /************************************************************************** * * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA * All Rights Reserved. * * Permission is hereby granted, free of charge, to any person obtaining a * copy of this software and associated documentation files (the * "Software"), to deal in the Software without restriction, including * without limitation the rights to use, copy, modify, merge, publish, * distribute, sub license, and/or sell copies of the Software, and to * permit persons to whom the Software is furnished to do so, subject to * the following conditions: * * The above copyright notice and this permission notice (including the * next paragraph) shall be included in all copies or substantial portions * of the Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE * USE OR OTHER DEALINGS IN THE SOFTWARE. * **************************************************************************/ /* * Authors: Thomas Hellstrom */ /* * Copyright (c) 2013 The FreeBSD Foundation * All rights reserved. * * Portions of this software were developed by Konstantin Belousov * under sponsorship from the FreeBSD Foundation. */ #include __FBSDID("$FreeBSD$"); #include "opt_vm.h" #include #include #include #include #include #include #include #define TTM_BO_VM_NUM_PREFAULT 16 RB_GENERATE(ttm_bo_device_buffer_objects, ttm_buffer_object, vm_rb, ttm_bo_cmp_rb_tree_items); int ttm_bo_cmp_rb_tree_items(struct ttm_buffer_object *a, struct ttm_buffer_object *b) { if (a->vm_node->start < b->vm_node->start) { return (-1); } else if (a->vm_node->start > b->vm_node->start) { return (1); } else { return (0); } } static struct ttm_buffer_object *ttm_bo_vm_lookup_rb(struct ttm_bo_device *bdev, unsigned long page_start, unsigned long num_pages) { unsigned long cur_offset; struct ttm_buffer_object *bo; struct ttm_buffer_object *best_bo = NULL; bo = RB_ROOT(&bdev->addr_space_rb); while (bo != NULL) { cur_offset = bo->vm_node->start; if (page_start >= cur_offset) { best_bo = bo; if (page_start == cur_offset) break; bo = RB_RIGHT(bo, vm_rb); } else bo = RB_LEFT(bo, vm_rb); } if (unlikely(best_bo == NULL)) return NULL; if (unlikely((best_bo->vm_node->start + best_bo->num_pages) < (page_start + num_pages))) return NULL; return best_bo; } static int ttm_bo_vm_fault(vm_object_t vm_obj, vm_ooffset_t offset, int prot, vm_page_t *mres) { struct ttm_buffer_object *bo = vm_obj->handle; struct ttm_bo_device *bdev = bo->bdev; struct ttm_tt *ttm = NULL; vm_page_t m, m1; int ret; int retval = VM_PAGER_OK; struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; vm_object_pip_add(vm_obj, 1); if (*mres != NULL) { (void)vm_page_remove(*mres); } retry: VM_OBJECT_WUNLOCK(vm_obj); m = NULL; reserve: ret = ttm_bo_reserve(bo, false, false, false, 0); if (unlikely(ret != 0)) { if (ret == -EBUSY) { kern_yield(PRI_USER); goto reserve; } } if (bdev->driver->fault_reserve_notify) { ret = bdev->driver->fault_reserve_notify(bo); switch (ret) { case 0: break; case -EBUSY: case -ERESTARTSYS: case -EINTR: kern_yield(PRI_USER); goto reserve; default: retval = VM_PAGER_ERROR; goto out_unlock; } } /* * Wait for buffer data in transit, due to a pipelined * move. */ mtx_lock(&bdev->fence_lock); if (test_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags)) { /* * Here, the behavior differs between Linux and FreeBSD. * * On Linux, the wait is interruptible (3rd argument to * ttm_bo_wait). There must be some mechanism to resume * page fault handling, once the signal is processed. * * On FreeBSD, the wait is uninteruptible. This is not a * problem as we can't end up with an unkillable process * here, because the wait will eventually time out. * * An example of this situation is the Xorg process * which uses SIGALRM internally. The signal could * interrupt the wait, causing the page fault to fail * and the process to receive SIGSEGV. */ ret = ttm_bo_wait(bo, false, false, false); mtx_unlock(&bdev->fence_lock); if (unlikely(ret != 0)) { retval = VM_PAGER_ERROR; goto out_unlock; } } else mtx_unlock(&bdev->fence_lock); ret = ttm_mem_io_lock(man, true); if (unlikely(ret != 0)) { retval = VM_PAGER_ERROR; goto out_unlock; } ret = ttm_mem_io_reserve_vm(bo); if (unlikely(ret != 0)) { retval = VM_PAGER_ERROR; goto out_io_unlock; } /* * Strictly, we're not allowed to modify vma->vm_page_prot here, * since the mmap_sem is only held in read mode. However, we * modify only the caching bits of vma->vm_page_prot and * consider those bits protected by * the bo->mutex, as we should be the only writers. * There shouldn't really be any readers of these bits except * within vm_insert_mixed()? fork? * * TODO: Add a list of vmas to the bo, and change the * vma->vm_page_prot when the object changes caching policy, with * the correct locks held. */ if (!bo->mem.bus.is_iomem) { /* Allocate all page at once, most common usage */ ttm = bo->ttm; if (ttm->bdev->driver->ttm_tt_populate(ttm)) { retval = VM_PAGER_ERROR; goto out_io_unlock; } } if (bo->mem.bus.is_iomem) { m = PHYS_TO_VM_PAGE(bo->mem.bus.base + bo->mem.bus.offset + offset); KASSERT((m->flags & PG_FICTITIOUS) != 0, ("physical address %#jx not fictitious", (uintmax_t)(bo->mem.bus.base + bo->mem.bus.offset + offset))); pmap_page_set_memattr(m, ttm_io_prot(bo->mem.placement)); } else { ttm = bo->ttm; m = ttm->pages[OFF_TO_IDX(offset)]; if (unlikely(!m)) { retval = VM_PAGER_ERROR; goto out_io_unlock; } pmap_page_set_memattr(m, (bo->mem.placement & TTM_PL_FLAG_CACHED) ? VM_MEMATTR_WRITE_BACK : ttm_io_prot(bo->mem.placement)); } VM_OBJECT_WLOCK(vm_obj); if (vm_page_busy_acquire(m, VM_ALLOC_WAITFAIL) == 0) { ttm_mem_io_unlock(man); ttm_bo_unreserve(bo); goto retry; } m1 = vm_page_lookup(vm_obj, OFF_TO_IDX(offset)); if (m1 == NULL) { if (vm_page_insert(m, vm_obj, OFF_TO_IDX(offset))) { vm_page_xunbusy(m); VM_OBJECT_WUNLOCK(vm_obj); vm_wait(vm_obj); VM_OBJECT_WLOCK(vm_obj); ttm_mem_io_unlock(man); ttm_bo_unreserve(bo); goto retry; } } else { KASSERT(m == m1, ("inconsistent insert bo %p m %p m1 %p offset %jx", bo, m, m1, (uintmax_t)offset)); } - m->valid = VM_PAGE_BITS_ALL; + vm_page_valid(m); if (*mres != NULL) { KASSERT(*mres != m, ("losing %p %p", *mres, m)); vm_page_free(*mres); } *mres = m; out_io_unlock1: ttm_mem_io_unlock(man); out_unlock1: ttm_bo_unreserve(bo); vm_object_pip_wakeup(vm_obj); return (retval); out_io_unlock: VM_OBJECT_WLOCK(vm_obj); goto out_io_unlock1; out_unlock: VM_OBJECT_WLOCK(vm_obj); goto out_unlock1; } static int ttm_bo_vm_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t foff, struct ucred *cred, u_short *color) { /* * On Linux, a reference to the buffer object is acquired here. * The reason is that this function is not called when the * mmap() is initialized, but only when a process forks for * instance. Therefore on Linux, the reference on the bo is * acquired either in ttm_bo_mmap() or ttm_bo_vm_open(). It's * then released in ttm_bo_vm_close(). * * Here, this function is called during mmap() initialization. * Thus, the reference acquired in ttm_bo_mmap_single() is * sufficient. */ *color = 0; return (0); } static void ttm_bo_vm_dtor(void *handle) { struct ttm_buffer_object *bo = handle; ttm_bo_unref(&bo); } static struct cdev_pager_ops ttm_pager_ops = { .cdev_pg_fault = ttm_bo_vm_fault, .cdev_pg_ctor = ttm_bo_vm_ctor, .cdev_pg_dtor = ttm_bo_vm_dtor }; int ttm_bo_mmap_single(struct ttm_bo_device *bdev, vm_ooffset_t *offset, vm_size_t size, struct vm_object **obj_res, int nprot) { struct ttm_bo_driver *driver; struct ttm_buffer_object *bo; struct vm_object *vm_obj; int ret; rw_wlock(&bdev->vm_lock); bo = ttm_bo_vm_lookup_rb(bdev, OFF_TO_IDX(*offset), OFF_TO_IDX(size)); if (likely(bo != NULL)) refcount_acquire(&bo->kref); rw_wunlock(&bdev->vm_lock); if (unlikely(bo == NULL)) { printf("[TTM] Could not find buffer object to map\n"); return (-EINVAL); } driver = bo->bdev->driver; if (unlikely(!driver->verify_access)) { ret = -EPERM; goto out_unref; } ret = driver->verify_access(bo); if (unlikely(ret != 0)) goto out_unref; vm_obj = cdev_pager_allocate(bo, OBJT_MGTDEVICE, &ttm_pager_ops, size, nprot, 0, curthread->td_ucred); if (vm_obj == NULL) { ret = -EINVAL; goto out_unref; } /* * Note: We're transferring the bo reference to vm_obj->handle here. */ *offset = 0; *obj_res = vm_obj; return 0; out_unref: ttm_bo_unref(&bo); return ret; } void ttm_bo_release_mmap(struct ttm_buffer_object *bo) { vm_object_t vm_obj; vm_page_t m; int i; vm_obj = cdev_pager_lookup(bo); if (vm_obj == NULL) return; VM_OBJECT_WLOCK(vm_obj); retry: for (i = 0; i < bo->num_pages; i++) { m = vm_page_lookup(vm_obj, i); if (m == NULL) continue; if (vm_page_busy_acquire(m, VM_ALLOC_WAITFAIL) == 0) goto retry; cdev_pager_free_page(vm_obj, m); } VM_OBJECT_WUNLOCK(vm_obj); vm_object_deallocate(vm_obj); } #if 0 int ttm_fbdev_mmap(struct vm_area_struct *vma, struct ttm_buffer_object *bo) { if (vma->vm_pgoff != 0) return -EACCES; vma->vm_ops = &ttm_bo_vm_ops; vma->vm_private_data = ttm_bo_reference(bo); vma->vm_flags |= VM_IO | VM_MIXEDMAP | VM_DONTEXPAND; return 0; } ssize_t ttm_bo_io(struct ttm_bo_device *bdev, struct file *filp, const char __user *wbuf, char __user *rbuf, size_t count, loff_t *f_pos, bool write) { struct ttm_buffer_object *bo; struct ttm_bo_driver *driver; struct ttm_bo_kmap_obj map; unsigned long dev_offset = (*f_pos >> PAGE_SHIFT); unsigned long kmap_offset; unsigned long kmap_end; unsigned long kmap_num; size_t io_size; unsigned int page_offset; char *virtual; int ret; bool no_wait = false; bool dummy; read_lock(&bdev->vm_lock); bo = ttm_bo_vm_lookup_rb(bdev, dev_offset, 1); if (likely(bo != NULL)) ttm_bo_reference(bo); read_unlock(&bdev->vm_lock); if (unlikely(bo == NULL)) return -EFAULT; driver = bo->bdev->driver; if (unlikely(!driver->verify_access)) { ret = -EPERM; goto out_unref; } ret = driver->verify_access(bo, filp); if (unlikely(ret != 0)) goto out_unref; kmap_offset = dev_offset - bo->vm_node->start; if (unlikely(kmap_offset >= bo->num_pages)) { ret = -EFBIG; goto out_unref; } page_offset = *f_pos & ~PAGE_MASK; io_size = bo->num_pages - kmap_offset; io_size = (io_size << PAGE_SHIFT) - page_offset; if (count < io_size) io_size = count; kmap_end = (*f_pos + count - 1) >> PAGE_SHIFT; kmap_num = kmap_end - kmap_offset + 1; ret = ttm_bo_reserve(bo, true, no_wait, false, 0); switch (ret) { case 0: break; case -EBUSY: ret = -EAGAIN; goto out_unref; default: goto out_unref; } ret = ttm_bo_kmap(bo, kmap_offset, kmap_num, &map); if (unlikely(ret != 0)) { ttm_bo_unreserve(bo); goto out_unref; } virtual = ttm_kmap_obj_virtual(&map, &dummy); virtual += page_offset; if (write) ret = copy_from_user(virtual, wbuf, io_size); else ret = copy_to_user(rbuf, virtual, io_size); ttm_bo_kunmap(&map); ttm_bo_unreserve(bo); ttm_bo_unref(&bo); if (unlikely(ret != 0)) return -EFBIG; *f_pos += io_size; return io_size; out_unref: ttm_bo_unref(&bo); return ret; } ssize_t ttm_bo_fbdev_io(struct ttm_buffer_object *bo, const char __user *wbuf, char __user *rbuf, size_t count, loff_t *f_pos, bool write) { struct ttm_bo_kmap_obj map; unsigned long kmap_offset; unsigned long kmap_end; unsigned long kmap_num; size_t io_size; unsigned int page_offset; char *virtual; int ret; bool no_wait = false; bool dummy; kmap_offset = (*f_pos >> PAGE_SHIFT); if (unlikely(kmap_offset >= bo->num_pages)) return -EFBIG; page_offset = *f_pos & ~PAGE_MASK; io_size = bo->num_pages - kmap_offset; io_size = (io_size << PAGE_SHIFT) - page_offset; if (count < io_size) io_size = count; kmap_end = (*f_pos + count - 1) >> PAGE_SHIFT; kmap_num = kmap_end - kmap_offset + 1; ret = ttm_bo_reserve(bo, true, no_wait, false, 0); switch (ret) { case 0: break; case -EBUSY: return -EAGAIN; default: return ret; } ret = ttm_bo_kmap(bo, kmap_offset, kmap_num, &map); if (unlikely(ret != 0)) { ttm_bo_unreserve(bo); return ret; } virtual = ttm_kmap_obj_virtual(&map, &dummy); virtual += page_offset; if (write) ret = copy_from_user(virtual, wbuf, io_size); else ret = copy_to_user(rbuf, virtual, io_size); ttm_bo_kunmap(&map); ttm_bo_unreserve(bo); ttm_bo_unref(&bo); if (unlikely(ret != 0)) return ret; *f_pos += io_size; return io_size; } #endif Index: head/sys/dev/drm2/ttm/ttm_tt.c =================================================================== --- head/sys/dev/drm2/ttm/ttm_tt.c (revision 353538) +++ head/sys/dev/drm2/ttm/ttm_tt.c (revision 353539) @@ -1,360 +1,360 @@ /************************************************************************** * * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA * All Rights Reserved. * * Permission is hereby granted, free of charge, to any person obtaining a * copy of this software and associated documentation files (the * "Software"), to deal in the Software without restriction, including * without limitation the rights to use, copy, modify, merge, publish, * distribute, sub license, and/or sell copies of the Software, and to * permit persons to whom the Software is furnished to do so, subject to * the following conditions: * * The above copyright notice and this permission notice (including the * next paragraph) shall be included in all copies or substantial portions * of the Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE * USE OR OTHER DEALINGS IN THE SOFTWARE. * **************************************************************************/ /* * Authors: Thomas Hellstrom */ /* * Copyright (c) 2013 The FreeBSD Foundation * All rights reserved. * * Portions of this software were developed by Konstantin Belousov * under sponsorship from the FreeBSD Foundation. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include MALLOC_DEFINE(M_TTM_PD, "ttm_pd", "TTM Page Directories"); /** * Allocates storage for pointers to the pages that back the ttm. */ static void ttm_tt_alloc_page_directory(struct ttm_tt *ttm) { ttm->pages = malloc(ttm->num_pages * sizeof(void *), M_TTM_PD, M_WAITOK | M_ZERO); } static void ttm_dma_tt_alloc_page_directory(struct ttm_dma_tt *ttm) { ttm->ttm.pages = malloc(ttm->ttm.num_pages * sizeof(void *), M_TTM_PD, M_WAITOK | M_ZERO); ttm->dma_address = malloc(ttm->ttm.num_pages * sizeof(*ttm->dma_address), M_TTM_PD, M_WAITOK); } #if defined(__i386__) || defined(__amd64__) static inline int ttm_tt_set_page_caching(vm_page_t p, enum ttm_caching_state c_old, enum ttm_caching_state c_new) { /* XXXKIB our VM does not need this. */ #if 0 if (c_old != tt_cached) { /* p isn't in the default caching state, set it to * writeback first to free its current memtype. */ pmap_page_set_memattr(p, VM_MEMATTR_WRITE_BACK); } #endif if (c_new == tt_wc) pmap_page_set_memattr(p, VM_MEMATTR_WRITE_COMBINING); else if (c_new == tt_uncached) pmap_page_set_memattr(p, VM_MEMATTR_UNCACHEABLE); return (0); } #else static inline int ttm_tt_set_page_caching(vm_page_t p, enum ttm_caching_state c_old, enum ttm_caching_state c_new) { return 0; } #endif /* * Change caching policy for the linear kernel map * for range of pages in a ttm. */ static int ttm_tt_set_caching(struct ttm_tt *ttm, enum ttm_caching_state c_state) { int i, j; vm_page_t cur_page; int ret; if (ttm->caching_state == c_state) return 0; if (ttm->state == tt_unpopulated) { /* Change caching but don't populate */ ttm->caching_state = c_state; return 0; } if (ttm->caching_state == tt_cached) drm_clflush_pages(ttm->pages, ttm->num_pages); for (i = 0; i < ttm->num_pages; ++i) { cur_page = ttm->pages[i]; if (likely(cur_page != NULL)) { ret = ttm_tt_set_page_caching(cur_page, ttm->caching_state, c_state); if (unlikely(ret != 0)) goto out_err; } } ttm->caching_state = c_state; return 0; out_err: for (j = 0; j < i; ++j) { cur_page = ttm->pages[j]; if (cur_page != NULL) { (void)ttm_tt_set_page_caching(cur_page, c_state, ttm->caching_state); } } return ret; } int ttm_tt_set_placement_caching(struct ttm_tt *ttm, uint32_t placement) { enum ttm_caching_state state; if (placement & TTM_PL_FLAG_WC) state = tt_wc; else if (placement & TTM_PL_FLAG_UNCACHED) state = tt_uncached; else state = tt_cached; return ttm_tt_set_caching(ttm, state); } void ttm_tt_destroy(struct ttm_tt *ttm) { if (unlikely(ttm == NULL)) return; if (ttm->state == tt_bound) { ttm_tt_unbind(ttm); } if (likely(ttm->pages != NULL)) { ttm->bdev->driver->ttm_tt_unpopulate(ttm); } if (!(ttm->page_flags & TTM_PAGE_FLAG_PERSISTENT_SWAP) && ttm->swap_storage) vm_object_deallocate(ttm->swap_storage); ttm->swap_storage = NULL; ttm->func->destroy(ttm); } int ttm_tt_init(struct ttm_tt *ttm, struct ttm_bo_device *bdev, unsigned long size, uint32_t page_flags, vm_page_t dummy_read_page) { ttm->bdev = bdev; ttm->glob = bdev->glob; ttm->num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT; ttm->caching_state = tt_cached; ttm->page_flags = page_flags; ttm->dummy_read_page = dummy_read_page; ttm->state = tt_unpopulated; ttm->swap_storage = NULL; ttm_tt_alloc_page_directory(ttm); if (!ttm->pages) { ttm_tt_destroy(ttm); printf("Failed allocating page table\n"); return -ENOMEM; } return 0; } void ttm_tt_fini(struct ttm_tt *ttm) { free(ttm->pages, M_TTM_PD); ttm->pages = NULL; } int ttm_dma_tt_init(struct ttm_dma_tt *ttm_dma, struct ttm_bo_device *bdev, unsigned long size, uint32_t page_flags, vm_page_t dummy_read_page) { struct ttm_tt *ttm = &ttm_dma->ttm; ttm->bdev = bdev; ttm->glob = bdev->glob; ttm->num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT; ttm->caching_state = tt_cached; ttm->page_flags = page_flags; ttm->dummy_read_page = dummy_read_page; ttm->state = tt_unpopulated; ttm->swap_storage = NULL; INIT_LIST_HEAD(&ttm_dma->pages_list); ttm_dma_tt_alloc_page_directory(ttm_dma); if (!ttm->pages || !ttm_dma->dma_address) { ttm_tt_destroy(ttm); printf("Failed allocating page table\n"); return -ENOMEM; } return 0; } void ttm_dma_tt_fini(struct ttm_dma_tt *ttm_dma) { struct ttm_tt *ttm = &ttm_dma->ttm; free(ttm->pages, M_TTM_PD); ttm->pages = NULL; free(ttm_dma->dma_address, M_TTM_PD); ttm_dma->dma_address = NULL; } void ttm_tt_unbind(struct ttm_tt *ttm) { int ret; if (ttm->state == tt_bound) { ret = ttm->func->unbind(ttm); MPASS(ret == 0); ttm->state = tt_unbound; } } int ttm_tt_bind(struct ttm_tt *ttm, struct ttm_mem_reg *bo_mem) { int ret = 0; if (!ttm) return -EINVAL; if (ttm->state == tt_bound) return 0; ret = ttm->bdev->driver->ttm_tt_populate(ttm); if (ret) return ret; ret = ttm->func->bind(ttm, bo_mem); if (unlikely(ret != 0)) return ret; ttm->state = tt_bound; return 0; } int ttm_tt_swapin(struct ttm_tt *ttm) { vm_object_t obj; vm_page_t from_page, to_page; int i, ret, rv; obj = ttm->swap_storage; VM_OBJECT_WLOCK(obj); vm_object_pip_add(obj, 1); for (i = 0; i < ttm->num_pages; ++i) { rv = vm_page_grab_valid(&from_page, obj, i, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY); if (rv != VM_PAGER_OK) { ret = -EIO; goto err_ret; } to_page = ttm->pages[i]; if (unlikely(to_page == NULL)) { ret = -ENOMEM; goto err_ret; } pmap_copy_page(from_page, to_page); } vm_object_pip_wakeup(obj); VM_OBJECT_WUNLOCK(obj); if (!(ttm->page_flags & TTM_PAGE_FLAG_PERSISTENT_SWAP)) vm_object_deallocate(obj); ttm->swap_storage = NULL; ttm->page_flags &= ~TTM_PAGE_FLAG_SWAPPED; return (0); err_ret: vm_object_pip_wakeup(obj); VM_OBJECT_WUNLOCK(obj); return (ret); } int ttm_tt_swapout(struct ttm_tt *ttm, vm_object_t persistent_swap_storage) { vm_object_t obj; vm_page_t from_page, to_page; int i; MPASS(ttm->state == tt_unbound || ttm->state == tt_unpopulated); MPASS(ttm->caching_state == tt_cached); if (persistent_swap_storage == NULL) { obj = vm_pager_allocate(OBJT_SWAP, NULL, IDX_TO_OFF(ttm->num_pages), VM_PROT_DEFAULT, 0, curthread->td_ucred); if (obj == NULL) { printf("[TTM] Failed allocating swap storage\n"); return (-ENOMEM); } } else obj = persistent_swap_storage; VM_OBJECT_WLOCK(obj); vm_object_pip_add(obj, 1); for (i = 0; i < ttm->num_pages; ++i) { from_page = ttm->pages[i]; if (unlikely(from_page == NULL)) continue; to_page = vm_page_grab(obj, i, VM_ALLOC_NORMAL); pmap_copy_page(from_page, to_page); - to_page->valid = VM_PAGE_BITS_ALL; + vm_page_valid(to_page); vm_page_dirty(to_page); vm_page_xunbusy(to_page); } vm_object_pip_wakeup(obj); VM_OBJECT_WUNLOCK(obj); ttm->bdev->driver->ttm_tt_unpopulate(ttm); ttm->swap_storage = obj; ttm->page_flags |= TTM_PAGE_FLAG_SWAPPED; if (persistent_swap_storage != NULL) ttm->page_flags |= TTM_PAGE_FLAG_PERSISTENT_SWAP; return (0); } Index: head/sys/dev/md/md.c =================================================================== --- head/sys/dev/md/md.c (revision 353538) +++ head/sys/dev/md/md.c (revision 353539) @@ -1,2182 +1,2182 @@ /*- * SPDX-License-Identifier: (Beerware AND BSD-3-Clause) * * ---------------------------------------------------------------------------- * "THE BEER-WARE LICENSE" (Revision 42): * wrote this file. As long as you retain this notice you * can do whatever you want with this stuff. If we meet some day, and you think * this stuff is worth it, you can buy me a beer in return. Poul-Henning Kamp * ---------------------------------------------------------------------------- * * $FreeBSD$ * */ /*- * The following functions are based in the vn(4) driver: mdstart_swap(), * mdstart_vnode(), mdcreate_swap(), mdcreate_vnode() and mddestroy(), * and as such under the following copyright: * * Copyright (c) 1988 University of Utah. * Copyright (c) 1990, 1993 * The Regents of the University of California. All rights reserved. * Copyright (c) 2013 The FreeBSD Foundation * All rights reserved. * * This code is derived from software contributed to Berkeley by * the Systems Programming Group of the University of Utah Computer * Science Department. * * Portions of this software were developed by Konstantin Belousov * under sponsorship from the FreeBSD Foundation. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: Utah Hdr: vn.c 1.13 94/04/02 * * from: @(#)vn.c 8.6 (Berkeley) 4/1/94 * From: src/sys/dev/vn/vn.c,v 1.122 2000/12/16 16:06:03 */ #include "opt_rootdevname.h" #include "opt_geom.h" #include "opt_md.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #define MD_MODVER 1 #define MD_SHUTDOWN 0x10000 /* Tell worker thread to terminate. */ #define MD_EXITING 0x20000 /* Worker thread is exiting. */ #define MD_PROVIDERGONE 0x40000 /* Safe to free the softc */ #ifndef MD_NSECT #define MD_NSECT (10000 * 2) #endif struct md_req { unsigned md_unit; /* unit number */ enum md_types md_type; /* type of disk */ off_t md_mediasize; /* size of disk in bytes */ unsigned md_sectorsize; /* sectorsize */ unsigned md_options; /* options */ int md_fwheads; /* firmware heads */ int md_fwsectors; /* firmware sectors */ char *md_file; /* pathname of file to mount */ enum uio_seg md_file_seg; /* location of md_file */ char *md_label; /* label of the device (userspace) */ int *md_units; /* pointer to units array (kernel) */ size_t md_units_nitems; /* items in md_units array */ }; #ifdef COMPAT_FREEBSD32 struct md_ioctl32 { unsigned md_version; unsigned md_unit; enum md_types md_type; uint32_t md_file; off_t md_mediasize; unsigned md_sectorsize; unsigned md_options; uint64_t md_base; int md_fwheads; int md_fwsectors; uint32_t md_label; int md_pad[MDNPAD]; } __attribute__((__packed__)); CTASSERT((sizeof(struct md_ioctl32)) == 436); #define MDIOCATTACH_32 _IOC_NEWTYPE(MDIOCATTACH, struct md_ioctl32) #define MDIOCDETACH_32 _IOC_NEWTYPE(MDIOCDETACH, struct md_ioctl32) #define MDIOCQUERY_32 _IOC_NEWTYPE(MDIOCQUERY, struct md_ioctl32) #define MDIOCRESIZE_32 _IOC_NEWTYPE(MDIOCRESIZE, struct md_ioctl32) #endif /* COMPAT_FREEBSD32 */ static MALLOC_DEFINE(M_MD, "md_disk", "Memory Disk"); static MALLOC_DEFINE(M_MDSECT, "md_sectors", "Memory Disk Sectors"); static int md_debug; SYSCTL_INT(_debug, OID_AUTO, mddebug, CTLFLAG_RW, &md_debug, 0, "Enable md(4) debug messages"); static int md_malloc_wait; SYSCTL_INT(_vm, OID_AUTO, md_malloc_wait, CTLFLAG_RW, &md_malloc_wait, 0, "Allow malloc to wait for memory allocations"); #if defined(MD_ROOT) && !defined(MD_ROOT_FSTYPE) #define MD_ROOT_FSTYPE "ufs" #endif #if defined(MD_ROOT) /* * Preloaded image gets put here. */ #if defined(MD_ROOT_SIZE) /* * We put the mfs_root symbol into the oldmfs section of the kernel object file. * Applications that patch the object with the image can determine * the size looking at the oldmfs section size within the kernel. */ u_char mfs_root[MD_ROOT_SIZE*1024] __attribute__ ((section ("oldmfs"))); const int mfs_root_size = sizeof(mfs_root); #elif defined(MD_ROOT_MEM) /* MD region already mapped in the memory */ u_char *mfs_root; int mfs_root_size; #else extern volatile u_char __weak_symbol mfs_root; extern volatile u_char __weak_symbol mfs_root_end; __GLOBL(mfs_root); __GLOBL(mfs_root_end); #define mfs_root_size ((uintptr_t)(&mfs_root_end - &mfs_root)) #endif #endif static g_init_t g_md_init; static g_fini_t g_md_fini; static g_start_t g_md_start; static g_access_t g_md_access; static void g_md_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp, struct g_consumer *cp __unused, struct g_provider *pp); static g_provgone_t g_md_providergone; static struct cdev *status_dev = NULL; static struct sx md_sx; static struct unrhdr *md_uh; static d_ioctl_t mdctlioctl; static struct cdevsw mdctl_cdevsw = { .d_version = D_VERSION, .d_ioctl = mdctlioctl, .d_name = MD_NAME, }; struct g_class g_md_class = { .name = "MD", .version = G_VERSION, .init = g_md_init, .fini = g_md_fini, .start = g_md_start, .access = g_md_access, .dumpconf = g_md_dumpconf, .providergone = g_md_providergone, }; DECLARE_GEOM_CLASS(g_md_class, g_md); static LIST_HEAD(, md_s) md_softc_list = LIST_HEAD_INITIALIZER(md_softc_list); #define NINDIR (PAGE_SIZE / sizeof(uintptr_t)) #define NMASK (NINDIR-1) static int nshift; static uma_zone_t md_pbuf_zone; struct indir { uintptr_t *array; u_int total; u_int used; u_int shift; }; struct md_s { int unit; LIST_ENTRY(md_s) list; struct bio_queue_head bio_queue; struct mtx queue_mtx; struct mtx stat_mtx; struct cdev *dev; enum md_types type; off_t mediasize; unsigned sectorsize; unsigned opencount; unsigned fwheads; unsigned fwsectors; char ident[32]; unsigned flags; char name[20]; struct proc *procp; struct g_geom *gp; struct g_provider *pp; int (*start)(struct md_s *sc, struct bio *bp); struct devstat *devstat; /* MD_MALLOC related fields */ struct indir *indir; uma_zone_t uma; /* MD_PRELOAD related fields */ u_char *pl_ptr; size_t pl_len; /* MD_VNODE related fields */ struct vnode *vnode; char file[PATH_MAX]; char label[PATH_MAX]; struct ucred *cred; /* MD_SWAP related fields */ vm_object_t object; }; static struct indir * new_indir(u_int shift) { struct indir *ip; ip = malloc(sizeof *ip, M_MD, (md_malloc_wait ? M_WAITOK : M_NOWAIT) | M_ZERO); if (ip == NULL) return (NULL); ip->array = malloc(sizeof(uintptr_t) * NINDIR, M_MDSECT, (md_malloc_wait ? M_WAITOK : M_NOWAIT) | M_ZERO); if (ip->array == NULL) { free(ip, M_MD); return (NULL); } ip->total = NINDIR; ip->shift = shift; return (ip); } static void del_indir(struct indir *ip) { free(ip->array, M_MDSECT); free(ip, M_MD); } static void destroy_indir(struct md_s *sc, struct indir *ip) { int i; for (i = 0; i < NINDIR; i++) { if (!ip->array[i]) continue; if (ip->shift) destroy_indir(sc, (struct indir*)(ip->array[i])); else if (ip->array[i] > 255) uma_zfree(sc->uma, (void *)(ip->array[i])); } del_indir(ip); } /* * This function does the math and allocates the top level "indir" structure * for a device of "size" sectors. */ static struct indir * dimension(off_t size) { off_t rcnt; struct indir *ip; int layer; rcnt = size; layer = 0; while (rcnt > NINDIR) { rcnt /= NINDIR; layer++; } /* * XXX: the top layer is probably not fully populated, so we allocate * too much space for ip->array in here. */ ip = malloc(sizeof *ip, M_MD, M_WAITOK | M_ZERO); ip->array = malloc(sizeof(uintptr_t) * NINDIR, M_MDSECT, M_WAITOK | M_ZERO); ip->total = NINDIR; ip->shift = layer * nshift; return (ip); } /* * Read a given sector */ static uintptr_t s_read(struct indir *ip, off_t offset) { struct indir *cip; int idx; uintptr_t up; if (md_debug > 1) printf("s_read(%jd)\n", (intmax_t)offset); up = 0; for (cip = ip; cip != NULL;) { if (cip->shift) { idx = (offset >> cip->shift) & NMASK; up = cip->array[idx]; cip = (struct indir *)up; continue; } idx = offset & NMASK; return (cip->array[idx]); } return (0); } /* * Write a given sector, prune the tree if the value is 0 */ static int s_write(struct indir *ip, off_t offset, uintptr_t ptr) { struct indir *cip, *lip[10]; int idx, li; uintptr_t up; if (md_debug > 1) printf("s_write(%jd, %p)\n", (intmax_t)offset, (void *)ptr); up = 0; li = 0; cip = ip; for (;;) { lip[li++] = cip; if (cip->shift) { idx = (offset >> cip->shift) & NMASK; up = cip->array[idx]; if (up != 0) { cip = (struct indir *)up; continue; } /* Allocate branch */ cip->array[idx] = (uintptr_t)new_indir(cip->shift - nshift); if (cip->array[idx] == 0) return (ENOSPC); cip->used++; up = cip->array[idx]; cip = (struct indir *)up; continue; } /* leafnode */ idx = offset & NMASK; up = cip->array[idx]; if (up != 0) cip->used--; cip->array[idx] = ptr; if (ptr != 0) cip->used++; break; } if (cip->used != 0 || li == 1) return (0); li--; while (cip->used == 0 && cip != ip) { li--; idx = (offset >> lip[li]->shift) & NMASK; up = lip[li]->array[idx]; KASSERT(up == (uintptr_t)cip, ("md screwed up")); del_indir(cip); lip[li]->array[idx] = 0; lip[li]->used--; cip = lip[li]; } return (0); } static int g_md_access(struct g_provider *pp, int r, int w, int e) { struct md_s *sc; sc = pp->geom->softc; if (sc == NULL) { if (r <= 0 && w <= 0 && e <= 0) return (0); return (ENXIO); } r += pp->acr; w += pp->acw; e += pp->ace; if ((sc->flags & MD_READONLY) != 0 && w > 0) return (EROFS); if ((pp->acr + pp->acw + pp->ace) == 0 && (r + w + e) > 0) { sc->opencount = 1; } else if ((pp->acr + pp->acw + pp->ace) > 0 && (r + w + e) == 0) { sc->opencount = 0; } return (0); } static void g_md_start(struct bio *bp) { struct md_s *sc; sc = bp->bio_to->geom->softc; if ((bp->bio_cmd == BIO_READ) || (bp->bio_cmd == BIO_WRITE)) { mtx_lock(&sc->stat_mtx); devstat_start_transaction_bio(sc->devstat, bp); mtx_unlock(&sc->stat_mtx); } mtx_lock(&sc->queue_mtx); bioq_disksort(&sc->bio_queue, bp); wakeup(sc); mtx_unlock(&sc->queue_mtx); } #define MD_MALLOC_MOVE_ZERO 1 #define MD_MALLOC_MOVE_FILL 2 #define MD_MALLOC_MOVE_READ 3 #define MD_MALLOC_MOVE_WRITE 4 #define MD_MALLOC_MOVE_CMP 5 static int md_malloc_move_ma(vm_page_t **mp, int *ma_offs, unsigned sectorsize, void *ptr, u_char fill, int op) { struct sf_buf *sf; vm_page_t m, *mp1; char *p, first; off_t *uc; unsigned n; int error, i, ma_offs1, sz, first_read; m = NULL; error = 0; sf = NULL; /* if (op == MD_MALLOC_MOVE_CMP) { gcc */ first = 0; first_read = 0; uc = ptr; mp1 = *mp; ma_offs1 = *ma_offs; /* } */ sched_pin(); for (n = sectorsize; n != 0; n -= sz) { sz = imin(PAGE_SIZE - *ma_offs, n); if (m != **mp) { if (sf != NULL) sf_buf_free(sf); m = **mp; sf = sf_buf_alloc(m, SFB_CPUPRIVATE | (md_malloc_wait ? 0 : SFB_NOWAIT)); if (sf == NULL) { error = ENOMEM; break; } } p = (char *)sf_buf_kva(sf) + *ma_offs; switch (op) { case MD_MALLOC_MOVE_ZERO: bzero(p, sz); break; case MD_MALLOC_MOVE_FILL: memset(p, fill, sz); break; case MD_MALLOC_MOVE_READ: bcopy(ptr, p, sz); cpu_flush_dcache(p, sz); break; case MD_MALLOC_MOVE_WRITE: bcopy(p, ptr, sz); break; case MD_MALLOC_MOVE_CMP: for (i = 0; i < sz; i++, p++) { if (!first_read) { *uc = (u_char)*p; first = *p; first_read = 1; } else if (*p != first) { error = EDOOFUS; break; } } break; default: KASSERT(0, ("md_malloc_move_ma unknown op %d\n", op)); break; } if (error != 0) break; *ma_offs += sz; *ma_offs %= PAGE_SIZE; if (*ma_offs == 0) (*mp)++; ptr = (char *)ptr + sz; } if (sf != NULL) sf_buf_free(sf); sched_unpin(); if (op == MD_MALLOC_MOVE_CMP && error != 0) { *mp = mp1; *ma_offs = ma_offs1; } return (error); } static int md_malloc_move_vlist(bus_dma_segment_t **pvlist, int *pma_offs, unsigned len, void *ptr, u_char fill, int op) { bus_dma_segment_t *vlist; uint8_t *p, *end, first; off_t *uc; int ma_offs, seg_len; vlist = *pvlist; ma_offs = *pma_offs; uc = ptr; for (; len != 0; len -= seg_len) { seg_len = imin(vlist->ds_len - ma_offs, len); p = (uint8_t *)(uintptr_t)vlist->ds_addr + ma_offs; switch (op) { case MD_MALLOC_MOVE_ZERO: bzero(p, seg_len); break; case MD_MALLOC_MOVE_FILL: memset(p, fill, seg_len); break; case MD_MALLOC_MOVE_READ: bcopy(ptr, p, seg_len); cpu_flush_dcache(p, seg_len); break; case MD_MALLOC_MOVE_WRITE: bcopy(p, ptr, seg_len); break; case MD_MALLOC_MOVE_CMP: end = p + seg_len; first = *uc = *p; /* Confirm all following bytes match the first */ while (++p < end) { if (*p != first) return (EDOOFUS); } break; default: KASSERT(0, ("md_malloc_move_vlist unknown op %d\n", op)); break; } ma_offs += seg_len; if (ma_offs == vlist->ds_len) { ma_offs = 0; vlist++; } ptr = (uint8_t *)ptr + seg_len; } *pvlist = vlist; *pma_offs = ma_offs; return (0); } static int mdstart_malloc(struct md_s *sc, struct bio *bp) { u_char *dst; vm_page_t *m; bus_dma_segment_t *vlist; int i, error, error1, ma_offs, notmapped; off_t secno, nsec, uc; uintptr_t sp, osp; switch (bp->bio_cmd) { case BIO_READ: case BIO_WRITE: case BIO_DELETE: break; default: return (EOPNOTSUPP); } notmapped = (bp->bio_flags & BIO_UNMAPPED) != 0; vlist = (bp->bio_flags & BIO_VLIST) != 0 ? (bus_dma_segment_t *)bp->bio_data : NULL; if (notmapped) { m = bp->bio_ma; ma_offs = bp->bio_ma_offset; dst = NULL; KASSERT(vlist == NULL, ("vlists cannot be unmapped")); } else if (vlist != NULL) { ma_offs = bp->bio_ma_offset; dst = NULL; } else { dst = bp->bio_data; } nsec = bp->bio_length / sc->sectorsize; secno = bp->bio_offset / sc->sectorsize; error = 0; while (nsec--) { osp = s_read(sc->indir, secno); if (bp->bio_cmd == BIO_DELETE) { if (osp != 0) error = s_write(sc->indir, secno, 0); } else if (bp->bio_cmd == BIO_READ) { if (osp == 0) { if (notmapped) { error = md_malloc_move_ma(&m, &ma_offs, sc->sectorsize, NULL, 0, MD_MALLOC_MOVE_ZERO); } else if (vlist != NULL) { error = md_malloc_move_vlist(&vlist, &ma_offs, sc->sectorsize, NULL, 0, MD_MALLOC_MOVE_ZERO); } else bzero(dst, sc->sectorsize); } else if (osp <= 255) { if (notmapped) { error = md_malloc_move_ma(&m, &ma_offs, sc->sectorsize, NULL, osp, MD_MALLOC_MOVE_FILL); } else if (vlist != NULL) { error = md_malloc_move_vlist(&vlist, &ma_offs, sc->sectorsize, NULL, osp, MD_MALLOC_MOVE_FILL); } else memset(dst, osp, sc->sectorsize); } else { if (notmapped) { error = md_malloc_move_ma(&m, &ma_offs, sc->sectorsize, (void *)osp, 0, MD_MALLOC_MOVE_READ); } else if (vlist != NULL) { error = md_malloc_move_vlist(&vlist, &ma_offs, sc->sectorsize, (void *)osp, 0, MD_MALLOC_MOVE_READ); } else { bcopy((void *)osp, dst, sc->sectorsize); cpu_flush_dcache(dst, sc->sectorsize); } } osp = 0; } else if (bp->bio_cmd == BIO_WRITE) { if (sc->flags & MD_COMPRESS) { if (notmapped) { error1 = md_malloc_move_ma(&m, &ma_offs, sc->sectorsize, &uc, 0, MD_MALLOC_MOVE_CMP); i = error1 == 0 ? sc->sectorsize : 0; } else if (vlist != NULL) { error1 = md_malloc_move_vlist(&vlist, &ma_offs, sc->sectorsize, &uc, 0, MD_MALLOC_MOVE_CMP); i = error1 == 0 ? sc->sectorsize : 0; } else { uc = dst[0]; for (i = 1; i < sc->sectorsize; i++) { if (dst[i] != uc) break; } } } else { i = 0; uc = 0; } if (i == sc->sectorsize) { if (osp != uc) error = s_write(sc->indir, secno, uc); } else { if (osp <= 255) { sp = (uintptr_t)uma_zalloc(sc->uma, md_malloc_wait ? M_WAITOK : M_NOWAIT); if (sp == 0) { error = ENOSPC; break; } if (notmapped) { error = md_malloc_move_ma(&m, &ma_offs, sc->sectorsize, (void *)sp, 0, MD_MALLOC_MOVE_WRITE); } else if (vlist != NULL) { error = md_malloc_move_vlist( &vlist, &ma_offs, sc->sectorsize, (void *)sp, 0, MD_MALLOC_MOVE_WRITE); } else { bcopy(dst, (void *)sp, sc->sectorsize); } error = s_write(sc->indir, secno, sp); } else { if (notmapped) { error = md_malloc_move_ma(&m, &ma_offs, sc->sectorsize, (void *)osp, 0, MD_MALLOC_MOVE_WRITE); } else if (vlist != NULL) { error = md_malloc_move_vlist( &vlist, &ma_offs, sc->sectorsize, (void *)osp, 0, MD_MALLOC_MOVE_WRITE); } else { bcopy(dst, (void *)osp, sc->sectorsize); } osp = 0; } } } else { error = EOPNOTSUPP; } if (osp > 255) uma_zfree(sc->uma, (void*)osp); if (error != 0) break; secno++; if (!notmapped && vlist == NULL) dst += sc->sectorsize; } bp->bio_resid = 0; return (error); } static void mdcopyto_vlist(void *src, bus_dma_segment_t *vlist, off_t offset, off_t len) { off_t seg_len; while (offset >= vlist->ds_len) { offset -= vlist->ds_len; vlist++; } while (len != 0) { seg_len = omin(len, vlist->ds_len - offset); bcopy(src, (void *)(uintptr_t)(vlist->ds_addr + offset), seg_len); offset = 0; src = (uint8_t *)src + seg_len; len -= seg_len; vlist++; } } static void mdcopyfrom_vlist(bus_dma_segment_t *vlist, off_t offset, void *dst, off_t len) { off_t seg_len; while (offset >= vlist->ds_len) { offset -= vlist->ds_len; vlist++; } while (len != 0) { seg_len = omin(len, vlist->ds_len - offset); bcopy((void *)(uintptr_t)(vlist->ds_addr + offset), dst, seg_len); offset = 0; dst = (uint8_t *)dst + seg_len; len -= seg_len; vlist++; } } static int mdstart_preload(struct md_s *sc, struct bio *bp) { uint8_t *p; p = sc->pl_ptr + bp->bio_offset; switch (bp->bio_cmd) { case BIO_READ: if ((bp->bio_flags & BIO_VLIST) != 0) { mdcopyto_vlist(p, (bus_dma_segment_t *)bp->bio_data, bp->bio_ma_offset, bp->bio_length); } else { bcopy(p, bp->bio_data, bp->bio_length); } cpu_flush_dcache(bp->bio_data, bp->bio_length); break; case BIO_WRITE: if ((bp->bio_flags & BIO_VLIST) != 0) { mdcopyfrom_vlist((bus_dma_segment_t *)bp->bio_data, bp->bio_ma_offset, p, bp->bio_length); } else { bcopy(bp->bio_data, p, bp->bio_length); } break; } bp->bio_resid = 0; return (0); } static int mdstart_vnode(struct md_s *sc, struct bio *bp) { int error; struct uio auio; struct iovec aiov; struct iovec *piov; struct mount *mp; struct vnode *vp; struct buf *pb; bus_dma_segment_t *vlist; struct thread *td; off_t iolen, iostart, len, zerosize; int ma_offs, npages; switch (bp->bio_cmd) { case BIO_READ: auio.uio_rw = UIO_READ; break; case BIO_WRITE: case BIO_DELETE: auio.uio_rw = UIO_WRITE; break; case BIO_FLUSH: break; default: return (EOPNOTSUPP); } td = curthread; vp = sc->vnode; pb = NULL; piov = NULL; ma_offs = bp->bio_ma_offset; len = bp->bio_length; /* * VNODE I/O * * If an error occurs, we set BIO_ERROR but we do not set * B_INVAL because (for a write anyway), the buffer is * still valid. */ if (bp->bio_cmd == BIO_FLUSH) { (void) vn_start_write(vp, &mp, V_WAIT); vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); error = VOP_FSYNC(vp, MNT_WAIT, td); VOP_UNLOCK(vp, 0); vn_finished_write(mp); return (error); } auio.uio_offset = (vm_ooffset_t)bp->bio_offset; auio.uio_resid = bp->bio_length; auio.uio_segflg = UIO_SYSSPACE; auio.uio_td = td; if (bp->bio_cmd == BIO_DELETE) { /* * Emulate BIO_DELETE by writing zeros. */ zerosize = ZERO_REGION_SIZE - (ZERO_REGION_SIZE % sc->sectorsize); auio.uio_iovcnt = howmany(bp->bio_length, zerosize); piov = malloc(sizeof(*piov) * auio.uio_iovcnt, M_MD, M_WAITOK); auio.uio_iov = piov; while (len > 0) { piov->iov_base = __DECONST(void *, zero_region); piov->iov_len = len; if (len > zerosize) piov->iov_len = zerosize; len -= piov->iov_len; piov++; } piov = auio.uio_iov; } else if ((bp->bio_flags & BIO_VLIST) != 0) { piov = malloc(sizeof(*piov) * bp->bio_ma_n, M_MD, M_WAITOK); auio.uio_iov = piov; vlist = (bus_dma_segment_t *)bp->bio_data; while (len > 0) { piov->iov_base = (void *)(uintptr_t)(vlist->ds_addr + ma_offs); piov->iov_len = vlist->ds_len - ma_offs; if (piov->iov_len > len) piov->iov_len = len; len -= piov->iov_len; ma_offs = 0; vlist++; piov++; } auio.uio_iovcnt = piov - auio.uio_iov; piov = auio.uio_iov; } else if ((bp->bio_flags & BIO_UNMAPPED) != 0) { pb = uma_zalloc(md_pbuf_zone, M_WAITOK); bp->bio_resid = len; unmapped_step: npages = atop(min(MAXPHYS, round_page(len + (ma_offs & PAGE_MASK)))); iolen = min(ptoa(npages) - (ma_offs & PAGE_MASK), len); KASSERT(iolen > 0, ("zero iolen")); pmap_qenter((vm_offset_t)pb->b_data, &bp->bio_ma[atop(ma_offs)], npages); aiov.iov_base = (void *)((vm_offset_t)pb->b_data + (ma_offs & PAGE_MASK)); aiov.iov_len = iolen; auio.uio_iov = &aiov; auio.uio_iovcnt = 1; auio.uio_resid = iolen; } else { aiov.iov_base = bp->bio_data; aiov.iov_len = bp->bio_length; auio.uio_iov = &aiov; auio.uio_iovcnt = 1; } iostart = auio.uio_offset; if (auio.uio_rw == UIO_READ) { vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); error = VOP_READ(vp, &auio, 0, sc->cred); VOP_UNLOCK(vp, 0); } else { (void) vn_start_write(vp, &mp, V_WAIT); vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); error = VOP_WRITE(vp, &auio, sc->flags & MD_ASYNC ? 0 : IO_SYNC, sc->cred); VOP_UNLOCK(vp, 0); vn_finished_write(mp); if (error == 0) sc->flags &= ~MD_VERIFY; } /* When MD_CACHE is set, try to avoid double-caching the data. */ if (error == 0 && (sc->flags & MD_CACHE) == 0) VOP_ADVISE(vp, iostart, auio.uio_offset - 1, POSIX_FADV_DONTNEED); if (pb != NULL) { pmap_qremove((vm_offset_t)pb->b_data, npages); if (error == 0) { len -= iolen; bp->bio_resid -= iolen; ma_offs += iolen; if (len > 0) goto unmapped_step; } uma_zfree(md_pbuf_zone, pb); } free(piov, M_MD); if (pb == NULL) bp->bio_resid = auio.uio_resid; return (error); } static void md_swap_page_free(vm_page_t m) { vm_page_xunbusy(m); vm_page_free(m); } static int mdstart_swap(struct md_s *sc, struct bio *bp) { vm_page_t m; u_char *p; vm_pindex_t i, lastp; bus_dma_segment_t *vlist; int rv, ma_offs, offs, len, lastend; switch (bp->bio_cmd) { case BIO_READ: case BIO_WRITE: case BIO_DELETE: break; default: return (EOPNOTSUPP); } p = bp->bio_data; ma_offs = (bp->bio_flags & (BIO_UNMAPPED|BIO_VLIST)) != 0 ? bp->bio_ma_offset : 0; vlist = (bp->bio_flags & BIO_VLIST) != 0 ? (bus_dma_segment_t *)bp->bio_data : NULL; /* * offs is the offset at which to start operating on the * next (ie, first) page. lastp is the last page on * which we're going to operate. lastend is the ending * position within that last page (ie, PAGE_SIZE if * we're operating on complete aligned pages). */ offs = bp->bio_offset % PAGE_SIZE; lastp = (bp->bio_offset + bp->bio_length - 1) / PAGE_SIZE; lastend = (bp->bio_offset + bp->bio_length - 1) % PAGE_SIZE + 1; rv = VM_PAGER_OK; VM_OBJECT_WLOCK(sc->object); vm_object_pip_add(sc->object, 1); for (i = bp->bio_offset / PAGE_SIZE; i <= lastp; i++) { len = ((i == lastp) ? lastend : PAGE_SIZE) - offs; m = vm_page_grab(sc->object, i, VM_ALLOC_SYSTEM); if (bp->bio_cmd == BIO_READ) { - if (m->valid == VM_PAGE_BITS_ALL) + if (vm_page_all_valid(m)) rv = VM_PAGER_OK; else rv = vm_pager_get_pages(sc->object, &m, 1, NULL, NULL); if (rv == VM_PAGER_ERROR) { md_swap_page_free(m); break; } else if (rv == VM_PAGER_FAIL) { /* * Pager does not have the page. Zero * the allocated page, and mark it as * valid. Do not set dirty, the page * can be recreated if thrown out. */ pmap_zero_page(m); - m->valid = VM_PAGE_BITS_ALL; + vm_page_valid(m); } if ((bp->bio_flags & BIO_UNMAPPED) != 0) { pmap_copy_pages(&m, offs, bp->bio_ma, ma_offs, len); } else if ((bp->bio_flags & BIO_VLIST) != 0) { physcopyout_vlist(VM_PAGE_TO_PHYS(m) + offs, vlist, ma_offs, len); cpu_flush_dcache(p, len); } else { physcopyout(VM_PAGE_TO_PHYS(m) + offs, p, len); cpu_flush_dcache(p, len); } } else if (bp->bio_cmd == BIO_WRITE) { - if (len == PAGE_SIZE || m->valid == VM_PAGE_BITS_ALL) + if (len == PAGE_SIZE || vm_page_all_valid(m)) rv = VM_PAGER_OK; else rv = vm_pager_get_pages(sc->object, &m, 1, NULL, NULL); if (rv == VM_PAGER_ERROR) { md_swap_page_free(m); break; } else if (rv == VM_PAGER_FAIL) pmap_zero_page(m); if ((bp->bio_flags & BIO_UNMAPPED) != 0) { pmap_copy_pages(bp->bio_ma, ma_offs, &m, offs, len); } else if ((bp->bio_flags & BIO_VLIST) != 0) { physcopyin_vlist(vlist, ma_offs, VM_PAGE_TO_PHYS(m) + offs, len); } else { physcopyin(p, VM_PAGE_TO_PHYS(m) + offs, len); } - m->valid = VM_PAGE_BITS_ALL; + vm_page_valid(m); if (m->dirty != VM_PAGE_BITS_ALL) { vm_page_dirty(m); vm_pager_page_unswapped(m); } } else if (bp->bio_cmd == BIO_DELETE) { - if (len == PAGE_SIZE || m->valid == VM_PAGE_BITS_ALL) + if (len == PAGE_SIZE || vm_page_all_valid(m)) rv = VM_PAGER_OK; else rv = vm_pager_get_pages(sc->object, &m, 1, NULL, NULL); if (rv == VM_PAGER_ERROR) { md_swap_page_free(m); break; } else if (rv == VM_PAGER_FAIL) { md_swap_page_free(m); m = NULL; } else { /* Page is valid. */ if (len != PAGE_SIZE) { pmap_zero_page_area(m, offs, len); if (m->dirty != VM_PAGE_BITS_ALL) { vm_page_dirty(m); vm_pager_page_unswapped(m); } } else { vm_pager_page_unswapped(m); md_swap_page_free(m); m = NULL; } } } if (m != NULL) { vm_page_xunbusy(m); vm_page_lock(m); if (vm_page_active(m)) vm_page_reference(m); else vm_page_activate(m); vm_page_unlock(m); } /* Actions on further pages start at offset 0 */ p += PAGE_SIZE - offs; offs = 0; ma_offs += len; } vm_object_pip_wakeup(sc->object); VM_OBJECT_WUNLOCK(sc->object); return (rv != VM_PAGER_ERROR ? 0 : ENOSPC); } static int mdstart_null(struct md_s *sc, struct bio *bp) { switch (bp->bio_cmd) { case BIO_READ: bzero(bp->bio_data, bp->bio_length); cpu_flush_dcache(bp->bio_data, bp->bio_length); break; case BIO_WRITE: break; } bp->bio_resid = 0; return (0); } static void md_kthread(void *arg) { struct md_s *sc; struct bio *bp; int error; sc = arg; thread_lock(curthread); sched_prio(curthread, PRIBIO); thread_unlock(curthread); if (sc->type == MD_VNODE) curthread->td_pflags |= TDP_NORUNNINGBUF; for (;;) { mtx_lock(&sc->queue_mtx); if (sc->flags & MD_SHUTDOWN) { sc->flags |= MD_EXITING; mtx_unlock(&sc->queue_mtx); kproc_exit(0); } bp = bioq_takefirst(&sc->bio_queue); if (!bp) { msleep(sc, &sc->queue_mtx, PRIBIO | PDROP, "mdwait", 0); continue; } mtx_unlock(&sc->queue_mtx); if (bp->bio_cmd == BIO_GETATTR) { int isv = ((sc->flags & MD_VERIFY) != 0); if ((sc->fwsectors && sc->fwheads && (g_handleattr_int(bp, "GEOM::fwsectors", sc->fwsectors) || g_handleattr_int(bp, "GEOM::fwheads", sc->fwheads))) || g_handleattr_int(bp, "GEOM::candelete", 1)) error = -1; else if (sc->ident[0] != '\0' && g_handleattr_str(bp, "GEOM::ident", sc->ident)) error = -1; else if (g_handleattr_int(bp, "MNT::verified", isv)) error = -1; else error = EOPNOTSUPP; } else { error = sc->start(sc, bp); } if (bp->bio_cmd == BIO_READ || bp->bio_cmd == BIO_WRITE) { /* * Devstat uses (bio_bcount, bio_resid) for * determining the length of the completed part of * the i/o. g_io_deliver() will translate from * bio_completed to that, but it also destroys the * bio so we must do our own translation. */ bp->bio_bcount = bp->bio_length; bp->bio_resid = (error == -1 ? bp->bio_bcount : 0); devstat_end_transaction_bio(sc->devstat, bp); } if (error != -1) { bp->bio_completed = bp->bio_length; g_io_deliver(bp, error); } } } static struct md_s * mdfind(int unit) { struct md_s *sc; LIST_FOREACH(sc, &md_softc_list, list) { if (sc->unit == unit) break; } return (sc); } static struct md_s * mdnew(int unit, int *errp, enum md_types type) { struct md_s *sc; int error; *errp = 0; if (unit == -1) unit = alloc_unr(md_uh); else unit = alloc_unr_specific(md_uh, unit); if (unit == -1) { *errp = EBUSY; return (NULL); } sc = (struct md_s *)malloc(sizeof *sc, M_MD, M_WAITOK | M_ZERO); sc->type = type; bioq_init(&sc->bio_queue); mtx_init(&sc->queue_mtx, "md bio queue", NULL, MTX_DEF); mtx_init(&sc->stat_mtx, "md stat", NULL, MTX_DEF); sc->unit = unit; sprintf(sc->name, "md%d", unit); LIST_INSERT_HEAD(&md_softc_list, sc, list); error = kproc_create(md_kthread, sc, &sc->procp, 0, 0,"%s", sc->name); if (error == 0) return (sc); LIST_REMOVE(sc, list); mtx_destroy(&sc->stat_mtx); mtx_destroy(&sc->queue_mtx); free_unr(md_uh, sc->unit); free(sc, M_MD); *errp = error; return (NULL); } static void mdinit(struct md_s *sc) { struct g_geom *gp; struct g_provider *pp; g_topology_lock(); gp = g_new_geomf(&g_md_class, "md%d", sc->unit); gp->softc = sc; pp = g_new_providerf(gp, "md%d", sc->unit); pp->flags |= G_PF_DIRECT_SEND | G_PF_DIRECT_RECEIVE; pp->mediasize = sc->mediasize; pp->sectorsize = sc->sectorsize; switch (sc->type) { case MD_MALLOC: case MD_VNODE: case MD_SWAP: pp->flags |= G_PF_ACCEPT_UNMAPPED; break; case MD_PRELOAD: case MD_NULL: break; } sc->gp = gp; sc->pp = pp; g_error_provider(pp, 0); g_topology_unlock(); sc->devstat = devstat_new_entry("md", sc->unit, sc->sectorsize, DEVSTAT_ALL_SUPPORTED, DEVSTAT_TYPE_DIRECT, DEVSTAT_PRIORITY_MAX); } static int mdcreate_malloc(struct md_s *sc, struct md_req *mdr) { uintptr_t sp; int error; off_t u; error = 0; if (mdr->md_options & ~(MD_AUTOUNIT | MD_COMPRESS | MD_RESERVE)) return (EINVAL); if (mdr->md_sectorsize != 0 && !powerof2(mdr->md_sectorsize)) return (EINVAL); /* Compression doesn't make sense if we have reserved space */ if (mdr->md_options & MD_RESERVE) mdr->md_options &= ~MD_COMPRESS; if (mdr->md_fwsectors != 0) sc->fwsectors = mdr->md_fwsectors; if (mdr->md_fwheads != 0) sc->fwheads = mdr->md_fwheads; sc->flags = mdr->md_options & (MD_COMPRESS | MD_FORCE); sc->indir = dimension(sc->mediasize / sc->sectorsize); sc->uma = uma_zcreate(sc->name, sc->sectorsize, NULL, NULL, NULL, NULL, 0x1ff, 0); if (mdr->md_options & MD_RESERVE) { off_t nsectors; nsectors = sc->mediasize / sc->sectorsize; for (u = 0; u < nsectors; u++) { sp = (uintptr_t)uma_zalloc(sc->uma, (md_malloc_wait ? M_WAITOK : M_NOWAIT) | M_ZERO); if (sp != 0) error = s_write(sc->indir, u, sp); else error = ENOMEM; if (error != 0) break; } } return (error); } static int mdsetcred(struct md_s *sc, struct ucred *cred) { char *tmpbuf; int error = 0; /* * Set credits in our softc */ if (sc->cred) crfree(sc->cred); sc->cred = crhold(cred); /* * Horrible kludge to establish credentials for NFS XXX. */ if (sc->vnode) { struct uio auio; struct iovec aiov; tmpbuf = malloc(sc->sectorsize, M_TEMP, M_WAITOK); bzero(&auio, sizeof(auio)); aiov.iov_base = tmpbuf; aiov.iov_len = sc->sectorsize; auio.uio_iov = &aiov; auio.uio_iovcnt = 1; auio.uio_offset = 0; auio.uio_rw = UIO_READ; auio.uio_segflg = UIO_SYSSPACE; auio.uio_resid = aiov.iov_len; vn_lock(sc->vnode, LK_EXCLUSIVE | LK_RETRY); error = VOP_READ(sc->vnode, &auio, 0, sc->cred); VOP_UNLOCK(sc->vnode, 0); free(tmpbuf, M_TEMP); } return (error); } static int mdcreate_vnode(struct md_s *sc, struct md_req *mdr, struct thread *td) { struct vattr vattr; struct nameidata nd; char *fname; int error, flags; fname = mdr->md_file; if (mdr->md_file_seg == UIO_USERSPACE) { error = copyinstr(fname, sc->file, sizeof(sc->file), NULL); if (error != 0) return (error); } else if (mdr->md_file_seg == UIO_SYSSPACE) strlcpy(sc->file, fname, sizeof(sc->file)); else return (EDOOFUS); /* * If the user specified that this is a read only device, don't * set the FWRITE mask before trying to open the backing store. */ flags = FREAD | ((mdr->md_options & MD_READONLY) ? 0 : FWRITE) \ | ((mdr->md_options & MD_VERIFY) ? O_VERIFY : 0); NDINIT(&nd, LOOKUP, FOLLOW, UIO_SYSSPACE, sc->file, td); error = vn_open(&nd, &flags, 0, NULL); if (error != 0) return (error); NDFREE(&nd, NDF_ONLY_PNBUF); if (nd.ni_vp->v_type != VREG) { error = EINVAL; goto bad; } error = VOP_GETATTR(nd.ni_vp, &vattr, td->td_ucred); if (error != 0) goto bad; if (VOP_ISLOCKED(nd.ni_vp) != LK_EXCLUSIVE) { vn_lock(nd.ni_vp, LK_UPGRADE | LK_RETRY); if (nd.ni_vp->v_iflag & VI_DOOMED) { /* Forced unmount. */ error = EBADF; goto bad; } } nd.ni_vp->v_vflag |= VV_MD; VOP_UNLOCK(nd.ni_vp, 0); if (mdr->md_fwsectors != 0) sc->fwsectors = mdr->md_fwsectors; if (mdr->md_fwheads != 0) sc->fwheads = mdr->md_fwheads; snprintf(sc->ident, sizeof(sc->ident), "MD-DEV%ju-INO%ju", (uintmax_t)vattr.va_fsid, (uintmax_t)vattr.va_fileid); sc->flags = mdr->md_options & (MD_ASYNC | MD_CACHE | MD_FORCE | MD_VERIFY); if (!(flags & FWRITE)) sc->flags |= MD_READONLY; sc->vnode = nd.ni_vp; error = mdsetcred(sc, td->td_ucred); if (error != 0) { sc->vnode = NULL; vn_lock(nd.ni_vp, LK_EXCLUSIVE | LK_RETRY); nd.ni_vp->v_vflag &= ~VV_MD; goto bad; } return (0); bad: VOP_UNLOCK(nd.ni_vp, 0); (void)vn_close(nd.ni_vp, flags, td->td_ucred, td); return (error); } static void g_md_providergone(struct g_provider *pp) { struct md_s *sc = pp->geom->softc; mtx_lock(&sc->queue_mtx); sc->flags |= MD_PROVIDERGONE; wakeup(&sc->flags); mtx_unlock(&sc->queue_mtx); } static int mddestroy(struct md_s *sc, struct thread *td) { if (sc->gp) { g_topology_lock(); g_wither_geom(sc->gp, ENXIO); g_topology_unlock(); mtx_lock(&sc->queue_mtx); while (!(sc->flags & MD_PROVIDERGONE)) msleep(&sc->flags, &sc->queue_mtx, PRIBIO, "mddestroy", 0); mtx_unlock(&sc->queue_mtx); } if (sc->devstat) { devstat_remove_entry(sc->devstat); sc->devstat = NULL; } mtx_lock(&sc->queue_mtx); sc->flags |= MD_SHUTDOWN; wakeup(sc); while (!(sc->flags & MD_EXITING)) msleep(sc->procp, &sc->queue_mtx, PRIBIO, "mddestroy", hz / 10); mtx_unlock(&sc->queue_mtx); mtx_destroy(&sc->stat_mtx); mtx_destroy(&sc->queue_mtx); if (sc->vnode != NULL) { vn_lock(sc->vnode, LK_EXCLUSIVE | LK_RETRY); sc->vnode->v_vflag &= ~VV_MD; VOP_UNLOCK(sc->vnode, 0); (void)vn_close(sc->vnode, sc->flags & MD_READONLY ? FREAD : (FREAD|FWRITE), sc->cred, td); } if (sc->cred != NULL) crfree(sc->cred); if (sc->object != NULL) vm_object_deallocate(sc->object); if (sc->indir) destroy_indir(sc, sc->indir); if (sc->uma) uma_zdestroy(sc->uma); LIST_REMOVE(sc, list); free_unr(md_uh, sc->unit); free(sc, M_MD); return (0); } static int mdresize(struct md_s *sc, struct md_req *mdr) { int error, res; vm_pindex_t oldpages, newpages; switch (sc->type) { case MD_VNODE: case MD_NULL: break; case MD_SWAP: if (mdr->md_mediasize <= 0 || (mdr->md_mediasize % PAGE_SIZE) != 0) return (EDOM); oldpages = OFF_TO_IDX(round_page(sc->mediasize)); newpages = OFF_TO_IDX(round_page(mdr->md_mediasize)); if (newpages < oldpages) { VM_OBJECT_WLOCK(sc->object); vm_object_page_remove(sc->object, newpages, 0, 0); swap_pager_freespace(sc->object, newpages, oldpages - newpages); swap_release_by_cred(IDX_TO_OFF(oldpages - newpages), sc->cred); sc->object->charge = IDX_TO_OFF(newpages); sc->object->size = newpages; VM_OBJECT_WUNLOCK(sc->object); } else if (newpages > oldpages) { res = swap_reserve_by_cred(IDX_TO_OFF(newpages - oldpages), sc->cred); if (!res) return (ENOMEM); if ((mdr->md_options & MD_RESERVE) || (sc->flags & MD_RESERVE)) { error = swap_pager_reserve(sc->object, oldpages, newpages - oldpages); if (error < 0) { swap_release_by_cred( IDX_TO_OFF(newpages - oldpages), sc->cred); return (EDOM); } } VM_OBJECT_WLOCK(sc->object); sc->object->charge = IDX_TO_OFF(newpages); sc->object->size = newpages; VM_OBJECT_WUNLOCK(sc->object); } break; default: return (EOPNOTSUPP); } sc->mediasize = mdr->md_mediasize; g_topology_lock(); g_resize_provider(sc->pp, sc->mediasize); g_topology_unlock(); return (0); } static int mdcreate_swap(struct md_s *sc, struct md_req *mdr, struct thread *td) { vm_ooffset_t npage; int error; /* * Range check. Disallow negative sizes and sizes not being * multiple of page size. */ if (sc->mediasize <= 0 || (sc->mediasize % PAGE_SIZE) != 0) return (EDOM); /* * Allocate an OBJT_SWAP object. * * Note the truncation. */ if ((mdr->md_options & MD_VERIFY) != 0) return (EINVAL); npage = mdr->md_mediasize / PAGE_SIZE; if (mdr->md_fwsectors != 0) sc->fwsectors = mdr->md_fwsectors; if (mdr->md_fwheads != 0) sc->fwheads = mdr->md_fwheads; sc->object = vm_pager_allocate(OBJT_SWAP, NULL, PAGE_SIZE * npage, VM_PROT_DEFAULT, 0, td->td_ucred); if (sc->object == NULL) return (ENOMEM); sc->flags = mdr->md_options & (MD_FORCE | MD_RESERVE); if (mdr->md_options & MD_RESERVE) { if (swap_pager_reserve(sc->object, 0, npage) < 0) { error = EDOM; goto finish; } } error = mdsetcred(sc, td->td_ucred); finish: if (error != 0) { vm_object_deallocate(sc->object); sc->object = NULL; } return (error); } static int mdcreate_null(struct md_s *sc, struct md_req *mdr, struct thread *td) { /* * Range check. Disallow negative sizes and sizes not being * multiple of page size. */ if (sc->mediasize <= 0 || (sc->mediasize % PAGE_SIZE) != 0) return (EDOM); return (0); } static int kern_mdattach_locked(struct thread *td, struct md_req *mdr) { struct md_s *sc; unsigned sectsize; int error, i; sx_assert(&md_sx, SA_XLOCKED); switch (mdr->md_type) { case MD_MALLOC: case MD_PRELOAD: case MD_VNODE: case MD_SWAP: case MD_NULL: break; default: return (EINVAL); } if (mdr->md_sectorsize == 0) sectsize = DEV_BSIZE; else sectsize = mdr->md_sectorsize; if (sectsize > MAXPHYS || mdr->md_mediasize < sectsize) return (EINVAL); if (mdr->md_options & MD_AUTOUNIT) sc = mdnew(-1, &error, mdr->md_type); else { if (mdr->md_unit > INT_MAX) return (EINVAL); sc = mdnew(mdr->md_unit, &error, mdr->md_type); } if (sc == NULL) return (error); if (mdr->md_label != NULL) error = copyinstr(mdr->md_label, sc->label, sizeof(sc->label), NULL); if (error != 0) goto err_after_new; if (mdr->md_options & MD_AUTOUNIT) mdr->md_unit = sc->unit; sc->mediasize = mdr->md_mediasize; sc->sectorsize = sectsize; error = EDOOFUS; switch (sc->type) { case MD_MALLOC: sc->start = mdstart_malloc; error = mdcreate_malloc(sc, mdr); break; case MD_PRELOAD: /* * We disallow attaching preloaded memory disks via * ioctl. Preloaded memory disks are automatically * attached in g_md_init(). */ error = EOPNOTSUPP; break; case MD_VNODE: sc->start = mdstart_vnode; error = mdcreate_vnode(sc, mdr, td); break; case MD_SWAP: sc->start = mdstart_swap; error = mdcreate_swap(sc, mdr, td); break; case MD_NULL: sc->start = mdstart_null; error = mdcreate_null(sc, mdr, td); break; } err_after_new: if (error != 0) { mddestroy(sc, td); return (error); } /* Prune off any residual fractional sector */ i = sc->mediasize % sc->sectorsize; sc->mediasize -= i; mdinit(sc); return (0); } static int kern_mdattach(struct thread *td, struct md_req *mdr) { int error; sx_xlock(&md_sx); error = kern_mdattach_locked(td, mdr); sx_xunlock(&md_sx); return (error); } static int kern_mddetach_locked(struct thread *td, struct md_req *mdr) { struct md_s *sc; sx_assert(&md_sx, SA_XLOCKED); if (mdr->md_mediasize != 0 || (mdr->md_options & ~MD_FORCE) != 0) return (EINVAL); sc = mdfind(mdr->md_unit); if (sc == NULL) return (ENOENT); if (sc->opencount != 0 && !(sc->flags & MD_FORCE) && !(mdr->md_options & MD_FORCE)) return (EBUSY); return (mddestroy(sc, td)); } static int kern_mddetach(struct thread *td, struct md_req *mdr) { int error; sx_xlock(&md_sx); error = kern_mddetach_locked(td, mdr); sx_xunlock(&md_sx); return (error); } static int kern_mdresize_locked(struct md_req *mdr) { struct md_s *sc; sx_assert(&md_sx, SA_XLOCKED); if ((mdr->md_options & ~(MD_FORCE | MD_RESERVE)) != 0) return (EINVAL); sc = mdfind(mdr->md_unit); if (sc == NULL) return (ENOENT); if (mdr->md_mediasize < sc->sectorsize) return (EINVAL); if (mdr->md_mediasize < sc->mediasize && !(sc->flags & MD_FORCE) && !(mdr->md_options & MD_FORCE)) return (EBUSY); return (mdresize(sc, mdr)); } static int kern_mdresize(struct md_req *mdr) { int error; sx_xlock(&md_sx); error = kern_mdresize_locked(mdr); sx_xunlock(&md_sx); return (error); } static int kern_mdquery_locked(struct md_req *mdr) { struct md_s *sc; int error; sx_assert(&md_sx, SA_XLOCKED); sc = mdfind(mdr->md_unit); if (sc == NULL) return (ENOENT); mdr->md_type = sc->type; mdr->md_options = sc->flags; mdr->md_mediasize = sc->mediasize; mdr->md_sectorsize = sc->sectorsize; error = 0; if (mdr->md_label != NULL) { error = copyout(sc->label, mdr->md_label, strlen(sc->label) + 1); if (error != 0) return (error); } if (sc->type == MD_VNODE || (sc->type == MD_PRELOAD && mdr->md_file != NULL)) error = copyout(sc->file, mdr->md_file, strlen(sc->file) + 1); return (error); } static int kern_mdquery(struct md_req *mdr) { int error; sx_xlock(&md_sx); error = kern_mdquery_locked(mdr); sx_xunlock(&md_sx); return (error); } /* Copy members that are not userspace pointers. */ #define MD_IOCTL2REQ(mdio, mdr) do { \ (mdr)->md_unit = (mdio)->md_unit; \ (mdr)->md_type = (mdio)->md_type; \ (mdr)->md_mediasize = (mdio)->md_mediasize; \ (mdr)->md_sectorsize = (mdio)->md_sectorsize; \ (mdr)->md_options = (mdio)->md_options; \ (mdr)->md_fwheads = (mdio)->md_fwheads; \ (mdr)->md_fwsectors = (mdio)->md_fwsectors; \ (mdr)->md_units = &(mdio)->md_pad[0]; \ (mdr)->md_units_nitems = nitems((mdio)->md_pad); \ } while(0) /* Copy members that might have been updated */ #define MD_REQ2IOCTL(mdr, mdio) do { \ (mdio)->md_unit = (mdr)->md_unit; \ (mdio)->md_type = (mdr)->md_type; \ (mdio)->md_mediasize = (mdr)->md_mediasize; \ (mdio)->md_sectorsize = (mdr)->md_sectorsize; \ (mdio)->md_options = (mdr)->md_options; \ (mdio)->md_fwheads = (mdr)->md_fwheads; \ (mdio)->md_fwsectors = (mdr)->md_fwsectors; \ } while(0) static int mdctlioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flags, struct thread *td) { struct md_req mdr; int error; if (md_debug) printf("mdctlioctl(%s %lx %p %x %p)\n", devtoname(dev), cmd, addr, flags, td); bzero(&mdr, sizeof(mdr)); switch (cmd) { case MDIOCATTACH: case MDIOCDETACH: case MDIOCRESIZE: case MDIOCQUERY: { struct md_ioctl *mdio = (struct md_ioctl *)addr; if (mdio->md_version != MDIOVERSION) return (EINVAL); MD_IOCTL2REQ(mdio, &mdr); mdr.md_file = mdio->md_file; mdr.md_file_seg = UIO_USERSPACE; /* If the file is adjacent to the md_ioctl it's in kernel. */ if ((void *)mdio->md_file == (void *)(mdio + 1)) mdr.md_file_seg = UIO_SYSSPACE; mdr.md_label = mdio->md_label; break; } #ifdef COMPAT_FREEBSD32 case MDIOCATTACH_32: case MDIOCDETACH_32: case MDIOCRESIZE_32: case MDIOCQUERY_32: { struct md_ioctl32 *mdio = (struct md_ioctl32 *)addr; if (mdio->md_version != MDIOVERSION) return (EINVAL); MD_IOCTL2REQ(mdio, &mdr); mdr.md_file = (void *)(uintptr_t)mdio->md_file; mdr.md_file_seg = UIO_USERSPACE; mdr.md_label = (void *)(uintptr_t)mdio->md_label; break; } #endif default: /* Fall through to handler switch. */ break; } error = 0; switch (cmd) { case MDIOCATTACH: #ifdef COMPAT_FREEBSD32 case MDIOCATTACH_32: #endif error = kern_mdattach(td, &mdr); break; case MDIOCDETACH: #ifdef COMPAT_FREEBSD32 case MDIOCDETACH_32: #endif error = kern_mddetach(td, &mdr); break; case MDIOCRESIZE: #ifdef COMPAT_FREEBSD32 case MDIOCRESIZE_32: #endif error = kern_mdresize(&mdr); break; case MDIOCQUERY: #ifdef COMPAT_FREEBSD32 case MDIOCQUERY_32: #endif error = kern_mdquery(&mdr); break; default: error = ENOIOCTL; } switch (cmd) { case MDIOCATTACH: case MDIOCQUERY: { struct md_ioctl *mdio = (struct md_ioctl *)addr; MD_REQ2IOCTL(&mdr, mdio); break; } #ifdef COMPAT_FREEBSD32 case MDIOCATTACH_32: case MDIOCQUERY_32: { struct md_ioctl32 *mdio = (struct md_ioctl32 *)addr; MD_REQ2IOCTL(&mdr, mdio); break; } #endif default: /* Other commands to not alter mdr. */ break; } return (error); } static void md_preloaded(u_char *image, size_t length, const char *name) { struct md_s *sc; int error; sc = mdnew(-1, &error, MD_PRELOAD); if (sc == NULL) return; sc->mediasize = length; sc->sectorsize = DEV_BSIZE; sc->pl_ptr = image; sc->pl_len = length; sc->start = mdstart_preload; if (name != NULL) strlcpy(sc->file, name, sizeof(sc->file)); #ifdef MD_ROOT if (sc->unit == 0) { #ifndef ROOTDEVNAME rootdevnames[0] = MD_ROOT_FSTYPE ":/dev/md0"; #endif #ifdef MD_ROOT_READONLY sc->flags |= MD_READONLY; #endif } #endif mdinit(sc); if (name != NULL) { printf("%s%d: Preloaded image <%s> %zd bytes at %p\n", MD_NAME, sc->unit, name, length, image); } else { printf("%s%d: Embedded image %zd bytes at %p\n", MD_NAME, sc->unit, length, image); } } static void g_md_init(struct g_class *mp __unused) { caddr_t mod; u_char *ptr, *name, *type; unsigned len; int i; /* figure out log2(NINDIR) */ for (i = NINDIR, nshift = -1; i; nshift++) i >>= 1; mod = NULL; sx_init(&md_sx, "MD config lock"); g_topology_unlock(); md_uh = new_unrhdr(0, INT_MAX, NULL); #ifdef MD_ROOT if (mfs_root_size != 0) { sx_xlock(&md_sx); #ifdef MD_ROOT_MEM md_preloaded(mfs_root, mfs_root_size, NULL); #else md_preloaded(__DEVOLATILE(u_char *, &mfs_root), mfs_root_size, NULL); #endif sx_xunlock(&md_sx); } #endif /* XXX: are preload_* static or do they need Giant ? */ while ((mod = preload_search_next_name(mod)) != NULL) { name = (char *)preload_search_info(mod, MODINFO_NAME); if (name == NULL) continue; type = (char *)preload_search_info(mod, MODINFO_TYPE); if (type == NULL) continue; if (strcmp(type, "md_image") && strcmp(type, "mfs_root")) continue; ptr = preload_fetch_addr(mod); len = preload_fetch_size(mod); if (ptr != NULL && len != 0) { sx_xlock(&md_sx); md_preloaded(ptr, len, name); sx_xunlock(&md_sx); } } md_pbuf_zone = pbuf_zsecond_create("mdpbuf", nswbuf / 10); status_dev = make_dev(&mdctl_cdevsw, INT_MAX, UID_ROOT, GID_WHEEL, 0600, MDCTL_NAME); g_topology_lock(); } static void g_md_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp, struct g_consumer *cp __unused, struct g_provider *pp) { struct md_s *mp; char *type; mp = gp->softc; if (mp == NULL) return; switch (mp->type) { case MD_MALLOC: type = "malloc"; break; case MD_PRELOAD: type = "preload"; break; case MD_VNODE: type = "vnode"; break; case MD_SWAP: type = "swap"; break; case MD_NULL: type = "null"; break; default: type = "unknown"; break; } if (pp != NULL) { if (indent == NULL) { sbuf_printf(sb, " u %d", mp->unit); sbuf_printf(sb, " s %ju", (uintmax_t) mp->sectorsize); sbuf_printf(sb, " f %ju", (uintmax_t) mp->fwheads); sbuf_printf(sb, " fs %ju", (uintmax_t) mp->fwsectors); sbuf_printf(sb, " l %ju", (uintmax_t) mp->mediasize); sbuf_printf(sb, " t %s", type); if ((mp->type == MD_VNODE && mp->vnode != NULL) || (mp->type == MD_PRELOAD && mp->file[0] != '\0')) sbuf_printf(sb, " file %s", mp->file); sbuf_printf(sb, " label %s", mp->label); } else { sbuf_printf(sb, "%s%d\n", indent, mp->unit); sbuf_printf(sb, "%s%ju\n", indent, (uintmax_t) mp->sectorsize); sbuf_printf(sb, "%s%ju\n", indent, (uintmax_t) mp->fwheads); sbuf_printf(sb, "%s%ju\n", indent, (uintmax_t) mp->fwsectors); if (mp->ident[0] != '\0') { sbuf_printf(sb, "%s", indent); g_conf_printf_escaped(sb, "%s", mp->ident); sbuf_printf(sb, "\n"); } sbuf_printf(sb, "%s%ju\n", indent, (uintmax_t) mp->mediasize); sbuf_printf(sb, "%s%s\n", indent, (mp->flags & MD_COMPRESS) == 0 ? "off": "on"); sbuf_printf(sb, "%s%s\n", indent, (mp->flags & MD_READONLY) == 0 ? "read-write": "read-only"); sbuf_printf(sb, "%s%s\n", indent, type); if ((mp->type == MD_VNODE && mp->vnode != NULL) || (mp->type == MD_PRELOAD && mp->file[0] != '\0')) { sbuf_printf(sb, "%s", indent); g_conf_printf_escaped(sb, "%s", mp->file); sbuf_printf(sb, "\n"); } if (mp->type == MD_VNODE) sbuf_printf(sb, "%s%s\n", indent, (mp->flags & MD_CACHE) == 0 ? "off": "on"); sbuf_printf(sb, "%s\n"); } } } static void g_md_fini(struct g_class *mp __unused) { sx_destroy(&md_sx); if (status_dev != NULL) destroy_dev(status_dev); uma_zdestroy(md_pbuf_zone); delete_unrhdr(md_uh); } Index: head/sys/dev/netmap/netmap_freebsd.c =================================================================== --- head/sys/dev/netmap/netmap_freebsd.c (revision 353538) +++ head/sys/dev/netmap/netmap_freebsd.c (revision 353539) @@ -1,1620 +1,1620 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (C) 2013-2014 Universita` di Pisa. All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ /* $FreeBSD$ */ #include "opt_inet.h" #include "opt_inet6.h" #include #include #include #include #include #include /* POLLIN, POLLOUT */ #include /* types used in module initialization */ #include /* DEV_MODULE_ORDERED */ #include #include /* kern_ioctl() */ #include #include /* vtophys */ #include /* vtophys */ #include #include #include #include #include #include #include /* sockaddrs */ #include #include /* kthread_add() */ #include /* PROC_LOCK() */ #include /* RFNOWAIT */ #include /* sched_bind() */ #include /* mp_maxid */ #include /* taskqueue_enqueue(), taskqueue_create(), ... */ #include #include #include /* IFT_ETHER */ #include /* ether_ifdetach */ #include /* LLADDR */ #include /* bus_dmamap_* */ #include /* in6_cksum_pseudo() */ #include /* in_pseudo(), in_cksum_hdr() */ #include #include #include #include /* ======================== FREEBSD-SPECIFIC ROUTINES ================== */ static void nm_kqueue_notify(void *opaque, int pending) { struct nm_selinfo *si = opaque; /* We use a non-zero hint to distinguish this notification call * from the call done in kqueue_scan(), which uses hint=0. */ KNOTE_UNLOCKED(&si->si.si_note, /*hint=*/0x100); } int nm_os_selinfo_init(NM_SELINFO_T *si, const char *name) { int err; TASK_INIT(&si->ntfytask, 0, nm_kqueue_notify, si); si->ntfytq = taskqueue_create(name, M_NOWAIT, taskqueue_thread_enqueue, &si->ntfytq); if (si->ntfytq == NULL) return -ENOMEM; err = taskqueue_start_threads(&si->ntfytq, 1, PI_NET, "tq %s", name); if (err) { taskqueue_free(si->ntfytq); si->ntfytq = NULL; return err; } snprintf(si->mtxname, sizeof(si->mtxname), "nmkl%s", name); mtx_init(&si->m, si->mtxname, NULL, MTX_DEF); knlist_init_mtx(&si->si.si_note, &si->m); si->kqueue_users = 0; return (0); } void nm_os_selinfo_uninit(NM_SELINFO_T *si) { if (si->ntfytq == NULL) { return; /* si was not initialized */ } taskqueue_drain(si->ntfytq, &si->ntfytask); taskqueue_free(si->ntfytq); si->ntfytq = NULL; knlist_delete(&si->si.si_note, curthread, /*islocked=*/0); knlist_destroy(&si->si.si_note); /* now we don't need the mutex anymore */ mtx_destroy(&si->m); } void * nm_os_malloc(size_t size) { return malloc(size, M_DEVBUF, M_NOWAIT | M_ZERO); } void * nm_os_realloc(void *addr, size_t new_size, size_t old_size __unused) { return realloc(addr, new_size, M_DEVBUF, M_NOWAIT | M_ZERO); } void nm_os_free(void *addr) { free(addr, M_DEVBUF); } void nm_os_ifnet_lock(void) { IFNET_RLOCK(); } void nm_os_ifnet_unlock(void) { IFNET_RUNLOCK(); } static int netmap_use_count = 0; void nm_os_get_module(void) { netmap_use_count++; } void nm_os_put_module(void) { netmap_use_count--; } static void netmap_ifnet_arrival_handler(void *arg __unused, struct ifnet *ifp) { netmap_undo_zombie(ifp); } static void netmap_ifnet_departure_handler(void *arg __unused, struct ifnet *ifp) { netmap_make_zombie(ifp); } static eventhandler_tag nm_ifnet_ah_tag; static eventhandler_tag nm_ifnet_dh_tag; int nm_os_ifnet_init(void) { nm_ifnet_ah_tag = EVENTHANDLER_REGISTER(ifnet_arrival_event, netmap_ifnet_arrival_handler, NULL, EVENTHANDLER_PRI_ANY); nm_ifnet_dh_tag = EVENTHANDLER_REGISTER(ifnet_departure_event, netmap_ifnet_departure_handler, NULL, EVENTHANDLER_PRI_ANY); return 0; } void nm_os_ifnet_fini(void) { EVENTHANDLER_DEREGISTER(ifnet_arrival_event, nm_ifnet_ah_tag); EVENTHANDLER_DEREGISTER(ifnet_departure_event, nm_ifnet_dh_tag); } unsigned nm_os_ifnet_mtu(struct ifnet *ifp) { #if __FreeBSD_version < 1100030 return ifp->if_data.ifi_mtu; #else /* __FreeBSD_version >= 1100030 */ return ifp->if_mtu; #endif } rawsum_t nm_os_csum_raw(uint8_t *data, size_t len, rawsum_t cur_sum) { /* TODO XXX please use the FreeBSD implementation for this. */ uint16_t *words = (uint16_t *)data; int nw = len / 2; int i; for (i = 0; i < nw; i++) cur_sum += be16toh(words[i]); if (len & 1) cur_sum += (data[len-1] << 8); return cur_sum; } /* Fold a raw checksum: 'cur_sum' is in host byte order, while the * return value is in network byte order. */ uint16_t nm_os_csum_fold(rawsum_t cur_sum) { /* TODO XXX please use the FreeBSD implementation for this. */ while (cur_sum >> 16) cur_sum = (cur_sum & 0xFFFF) + (cur_sum >> 16); return htobe16((~cur_sum) & 0xFFFF); } uint16_t nm_os_csum_ipv4(struct nm_iphdr *iph) { #if 0 return in_cksum_hdr((void *)iph); #else return nm_os_csum_fold(nm_os_csum_raw((uint8_t*)iph, sizeof(struct nm_iphdr), 0)); #endif } void nm_os_csum_tcpudp_ipv4(struct nm_iphdr *iph, void *data, size_t datalen, uint16_t *check) { #ifdef INET uint16_t pseudolen = datalen + iph->protocol; /* Compute and insert the pseudo-header cheksum. */ *check = in_pseudo(iph->saddr, iph->daddr, htobe16(pseudolen)); /* Compute the checksum on TCP/UDP header + payload * (includes the pseudo-header). */ *check = nm_os_csum_fold(nm_os_csum_raw(data, datalen, 0)); #else static int notsupported = 0; if (!notsupported) { notsupported = 1; nm_prerr("inet4 segmentation not supported"); } #endif } void nm_os_csum_tcpudp_ipv6(struct nm_ipv6hdr *ip6h, void *data, size_t datalen, uint16_t *check) { #ifdef INET6 *check = in6_cksum_pseudo((void*)ip6h, datalen, ip6h->nexthdr, 0); *check = nm_os_csum_fold(nm_os_csum_raw(data, datalen, 0)); #else static int notsupported = 0; if (!notsupported) { notsupported = 1; nm_prerr("inet6 segmentation not supported"); } #endif } /* on FreeBSD we send up one packet at a time */ void * nm_os_send_up(struct ifnet *ifp, struct mbuf *m, struct mbuf *prev) { NA(ifp)->if_input(ifp, m); return NULL; } int nm_os_mbuf_has_csum_offld(struct mbuf *m) { return m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_SCTP | CSUM_TCP_IPV6 | CSUM_UDP_IPV6 | CSUM_SCTP_IPV6); } int nm_os_mbuf_has_seg_offld(struct mbuf *m) { return m->m_pkthdr.csum_flags & CSUM_TSO; } static void freebsd_generic_rx_handler(struct ifnet *ifp, struct mbuf *m) { int stolen; if (unlikely(!NM_NA_VALID(ifp))) { nm_prlim(1, "Warning: RX packet intercepted, but no" " emulated adapter"); return; } stolen = generic_rx_handler(ifp, m); if (!stolen) { struct netmap_generic_adapter *gna = (struct netmap_generic_adapter *)NA(ifp); gna->save_if_input(ifp, m); } } /* * Intercept the rx routine in the standard device driver. * Second argument is non-zero to intercept, 0 to restore */ int nm_os_catch_rx(struct netmap_generic_adapter *gna, int intercept) { struct netmap_adapter *na = &gna->up.up; struct ifnet *ifp = na->ifp; int ret = 0; nm_os_ifnet_lock(); if (intercept) { if (gna->save_if_input) { nm_prerr("RX on %s already intercepted", na->name); ret = EBUSY; /* already set */ goto out; } gna->save_if_input = ifp->if_input; ifp->if_input = freebsd_generic_rx_handler; } else { if (!gna->save_if_input) { nm_prerr("Failed to undo RX intercept on %s", na->name); ret = EINVAL; /* not saved */ goto out; } ifp->if_input = gna->save_if_input; gna->save_if_input = NULL; } out: nm_os_ifnet_unlock(); return ret; } /* * Intercept the packet steering routine in the tx path, * so that we can decide which queue is used for an mbuf. * Second argument is non-zero to intercept, 0 to restore. * On freebsd we just intercept if_transmit. */ int nm_os_catch_tx(struct netmap_generic_adapter *gna, int intercept) { struct netmap_adapter *na = &gna->up.up; struct ifnet *ifp = netmap_generic_getifp(gna); nm_os_ifnet_lock(); if (intercept) { na->if_transmit = ifp->if_transmit; ifp->if_transmit = netmap_transmit; } else { ifp->if_transmit = na->if_transmit; } nm_os_ifnet_unlock(); return 0; } /* * Transmit routine used by generic_netmap_txsync(). Returns 0 on success * and non-zero on error (which may be packet drops or other errors). * addr and len identify the netmap buffer, m is the (preallocated) * mbuf to use for transmissions. * * We should add a reference to the mbuf so the m_freem() at the end * of the transmission does not consume resources. * * On FreeBSD, and on multiqueue cards, we can force the queue using * if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) * i = m->m_pkthdr.flowid % adapter->num_queues; * else * i = curcpu % adapter->num_queues; * */ int nm_os_generic_xmit_frame(struct nm_os_gen_arg *a) { int ret; u_int len = a->len; struct ifnet *ifp = a->ifp; struct mbuf *m = a->m; #if __FreeBSD_version < 1100000 /* * Old FreeBSD versions. The mbuf has a cluster attached, * we need to copy from the cluster to the netmap buffer. */ if (MBUF_REFCNT(m) != 1) { nm_prerr("invalid refcnt %d for %p", MBUF_REFCNT(m), m); panic("in generic_xmit_frame"); } if (m->m_ext.ext_size < len) { nm_prlim(2, "size %d < len %d", m->m_ext.ext_size, len); len = m->m_ext.ext_size; } bcopy(a->addr, m->m_data, len); #else /* __FreeBSD_version >= 1100000 */ /* New FreeBSD versions. Link the external storage to * the netmap buffer, so that no copy is necessary. */ m->m_ext.ext_buf = m->m_data = a->addr; m->m_ext.ext_size = len; #endif /* __FreeBSD_version >= 1100000 */ m->m_flags |= M_PKTHDR; m->m_len = m->m_pkthdr.len = len; /* mbuf refcnt is not contended, no need to use atomic * (a memory barrier is enough). */ SET_MBUF_REFCNT(m, 2); M_HASHTYPE_SET(m, M_HASHTYPE_OPAQUE); m->m_pkthdr.flowid = a->ring_nr; m->m_pkthdr.rcvif = ifp; /* used for tx notification */ CURVNET_SET(ifp->if_vnet); ret = NA(ifp)->if_transmit(ifp, m); CURVNET_RESTORE(); return ret ? -1 : 0; } #if __FreeBSD_version >= 1100005 struct netmap_adapter * netmap_getna(if_t ifp) { return (NA((struct ifnet *)ifp)); } #endif /* __FreeBSD_version >= 1100005 */ /* * The following two functions are empty until we have a generic * way to extract the info from the ifp */ int nm_os_generic_find_num_desc(struct ifnet *ifp, unsigned int *tx, unsigned int *rx) { return 0; } void nm_os_generic_find_num_queues(struct ifnet *ifp, u_int *txq, u_int *rxq) { unsigned num_rings = netmap_generic_rings ? netmap_generic_rings : 1; *txq = num_rings; *rxq = num_rings; } void nm_os_generic_set_features(struct netmap_generic_adapter *gna) { gna->rxsg = 1; /* Supported through m_copydata. */ gna->txqdisc = 0; /* Not supported. */ } void nm_os_mitigation_init(struct nm_generic_mit *mit, int idx, struct netmap_adapter *na) { mit->mit_pending = 0; mit->mit_ring_idx = idx; mit->mit_na = na; } void nm_os_mitigation_start(struct nm_generic_mit *mit) { } void nm_os_mitigation_restart(struct nm_generic_mit *mit) { } int nm_os_mitigation_active(struct nm_generic_mit *mit) { return 0; } void nm_os_mitigation_cleanup(struct nm_generic_mit *mit) { } static int nm_vi_dummy(struct ifnet *ifp, u_long cmd, caddr_t addr) { return EINVAL; } static void nm_vi_start(struct ifnet *ifp) { panic("nm_vi_start() must not be called"); } /* * Index manager of persistent virtual interfaces. * It is used to decide the lowest byte of the MAC address. * We use the same algorithm with management of bridge port index. */ #define NM_VI_MAX 255 static struct { uint8_t index[NM_VI_MAX]; /* XXX just for a reasonable number */ uint8_t active; struct mtx lock; } nm_vi_indices; void nm_os_vi_init_index(void) { int i; for (i = 0; i < NM_VI_MAX; i++) nm_vi_indices.index[i] = i; nm_vi_indices.active = 0; mtx_init(&nm_vi_indices.lock, "nm_vi_indices_lock", NULL, MTX_DEF); } /* return -1 if no index available */ static int nm_vi_get_index(void) { int ret; mtx_lock(&nm_vi_indices.lock); ret = nm_vi_indices.active == NM_VI_MAX ? -1 : nm_vi_indices.index[nm_vi_indices.active++]; mtx_unlock(&nm_vi_indices.lock); return ret; } static void nm_vi_free_index(uint8_t val) { int i, lim; mtx_lock(&nm_vi_indices.lock); lim = nm_vi_indices.active; for (i = 0; i < lim; i++) { if (nm_vi_indices.index[i] == val) { /* swap index[lim-1] and j */ int tmp = nm_vi_indices.index[lim-1]; nm_vi_indices.index[lim-1] = val; nm_vi_indices.index[i] = tmp; nm_vi_indices.active--; break; } } if (lim == nm_vi_indices.active) nm_prerr("Index %u not found", val); mtx_unlock(&nm_vi_indices.lock); } #undef NM_VI_MAX /* * Implementation of a netmap-capable virtual interface that * registered to the system. * It is based on if_tap.c and ip_fw_log.c in FreeBSD 9. * * Note: Linux sets refcount to 0 on allocation of net_device, * then increments it on registration to the system. * FreeBSD sets refcount to 1 on if_alloc(), and does not * increment this refcount on if_attach(). */ int nm_os_vi_persist(const char *name, struct ifnet **ret) { struct ifnet *ifp; u_short macaddr_hi; uint32_t macaddr_mid; u_char eaddr[6]; int unit = nm_vi_get_index(); /* just to decide MAC address */ if (unit < 0) return EBUSY; /* * We use the same MAC address generation method with tap * except for the highest octet is 00:be instead of 00:bd */ macaddr_hi = htons(0x00be); /* XXX tap + 1 */ macaddr_mid = (uint32_t) ticks; bcopy(&macaddr_hi, eaddr, sizeof(short)); bcopy(&macaddr_mid, &eaddr[2], sizeof(uint32_t)); eaddr[5] = (uint8_t)unit; ifp = if_alloc(IFT_ETHER); if (ifp == NULL) { nm_prerr("if_alloc failed"); return ENOMEM; } if_initname(ifp, name, IF_DUNIT_NONE); ifp->if_mtu = 65536; ifp->if_flags = IFF_UP | IFF_SIMPLEX | IFF_MULTICAST; ifp->if_init = (void *)nm_vi_dummy; ifp->if_ioctl = nm_vi_dummy; ifp->if_start = nm_vi_start; ifp->if_mtu = ETHERMTU; IFQ_SET_MAXLEN(&ifp->if_snd, ifqmaxlen); ifp->if_capabilities |= IFCAP_LINKSTATE; ifp->if_capenable |= IFCAP_LINKSTATE; ether_ifattach(ifp, eaddr); *ret = ifp; return 0; } /* unregister from the system and drop the final refcount */ void nm_os_vi_detach(struct ifnet *ifp) { nm_vi_free_index(((char *)IF_LLADDR(ifp))[5]); ether_ifdetach(ifp); if_free(ifp); } #ifdef WITH_EXTMEM #include #include struct nm_os_extmem { vm_object_t obj; vm_offset_t kva; vm_offset_t size; uintptr_t scan; }; void nm_os_extmem_delete(struct nm_os_extmem *e) { nm_prinf("freeing %zx bytes", (size_t)e->size); vm_map_remove(kernel_map, e->kva, e->kva + e->size); nm_os_free(e); } char * nm_os_extmem_nextpage(struct nm_os_extmem *e) { char *rv = NULL; if (e->scan < e->kva + e->size) { rv = (char *)e->scan; e->scan += PAGE_SIZE; } return rv; } int nm_os_extmem_isequal(struct nm_os_extmem *e1, struct nm_os_extmem *e2) { return (e1->obj == e2->obj); } int nm_os_extmem_nr_pages(struct nm_os_extmem *e) { return e->size >> PAGE_SHIFT; } struct nm_os_extmem * nm_os_extmem_create(unsigned long p, struct nmreq_pools_info *pi, int *perror) { vm_map_t map; vm_map_entry_t entry; vm_object_t obj; vm_prot_t prot; vm_pindex_t index; boolean_t wired; struct nm_os_extmem *e = NULL; int rv, error = 0; e = nm_os_malloc(sizeof(*e)); if (e == NULL) { error = ENOMEM; goto out; } map = &curthread->td_proc->p_vmspace->vm_map; rv = vm_map_lookup(&map, p, VM_PROT_RW, &entry, &obj, &index, &prot, &wired); if (rv != KERN_SUCCESS) { nm_prerr("address %lx not found", p); goto out_free; } /* check that we are given the whole vm_object ? */ vm_map_lookup_done(map, entry); // XXX can we really use obj after releasing the map lock? e->obj = obj; vm_object_reference(obj); /* wire the memory and add the vm_object to the kernel map, * to make sure that it is not fred even if the processes that * are mmap()ing it all exit */ e->kva = vm_map_min(kernel_map); e->size = obj->size << PAGE_SHIFT; rv = vm_map_find(kernel_map, obj, 0, &e->kva, e->size, 0, VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE, VM_PROT_READ | VM_PROT_WRITE, 0); if (rv != KERN_SUCCESS) { nm_prerr("vm_map_find(%zx) failed", (size_t)e->size); goto out_rel; } rv = vm_map_wire(kernel_map, e->kva, e->kva + e->size, VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES); if (rv != KERN_SUCCESS) { nm_prerr("vm_map_wire failed"); goto out_rem; } e->scan = e->kva; return e; out_rem: vm_map_remove(kernel_map, e->kva, e->kva + e->size); e->obj = NULL; out_rel: vm_object_deallocate(e->obj); out_free: nm_os_free(e); out: if (perror) *perror = error; return NULL; } #endif /* WITH_EXTMEM */ /* ================== PTNETMAP GUEST SUPPORT ==================== */ #ifdef WITH_PTNETMAP #include #include #include /* bus_dmamap_* */ #include #include #include /* * ptnetmap memory device (memdev) for freebsd guest, * ssed to expose host netmap memory to the guest through a PCI BAR. */ /* * ptnetmap memdev private data structure */ struct ptnetmap_memdev { device_t dev; struct resource *pci_io; struct resource *pci_mem; struct netmap_mem_d *nm_mem; }; static int ptn_memdev_probe(device_t); static int ptn_memdev_attach(device_t); static int ptn_memdev_detach(device_t); static int ptn_memdev_shutdown(device_t); static device_method_t ptn_memdev_methods[] = { DEVMETHOD(device_probe, ptn_memdev_probe), DEVMETHOD(device_attach, ptn_memdev_attach), DEVMETHOD(device_detach, ptn_memdev_detach), DEVMETHOD(device_shutdown, ptn_memdev_shutdown), DEVMETHOD_END }; static driver_t ptn_memdev_driver = { PTNETMAP_MEMDEV_NAME, ptn_memdev_methods, sizeof(struct ptnetmap_memdev), }; /* We use (SI_ORDER_MIDDLE+1) here, see DEV_MODULE_ORDERED() invocation * below. */ static devclass_t ptnetmap_devclass; DRIVER_MODULE_ORDERED(ptn_memdev, pci, ptn_memdev_driver, ptnetmap_devclass, NULL, NULL, SI_ORDER_MIDDLE + 1); /* * Map host netmap memory through PCI-BAR in the guest OS, * returning physical (nm_paddr) and virtual (nm_addr) addresses * of the netmap memory mapped in the guest. */ int nm_os_pt_memdev_iomap(struct ptnetmap_memdev *ptn_dev, vm_paddr_t *nm_paddr, void **nm_addr, uint64_t *mem_size) { int rid; nm_prinf("ptn_memdev_driver iomap"); rid = PCIR_BAR(PTNETMAP_MEM_PCI_BAR); *mem_size = bus_read_4(ptn_dev->pci_io, PTNET_MDEV_IO_MEMSIZE_HI); *mem_size = bus_read_4(ptn_dev->pci_io, PTNET_MDEV_IO_MEMSIZE_LO) | (*mem_size << 32); /* map memory allocator */ ptn_dev->pci_mem = bus_alloc_resource(ptn_dev->dev, SYS_RES_MEMORY, &rid, 0, ~0, *mem_size, RF_ACTIVE); if (ptn_dev->pci_mem == NULL) { *nm_paddr = 0; *nm_addr = NULL; return ENOMEM; } *nm_paddr = rman_get_start(ptn_dev->pci_mem); *nm_addr = rman_get_virtual(ptn_dev->pci_mem); nm_prinf("=== BAR %d start %lx len %lx mem_size %lx ===", PTNETMAP_MEM_PCI_BAR, (unsigned long)(*nm_paddr), (unsigned long)rman_get_size(ptn_dev->pci_mem), (unsigned long)*mem_size); return (0); } uint32_t nm_os_pt_memdev_ioread(struct ptnetmap_memdev *ptn_dev, unsigned int reg) { return bus_read_4(ptn_dev->pci_io, reg); } /* Unmap host netmap memory. */ void nm_os_pt_memdev_iounmap(struct ptnetmap_memdev *ptn_dev) { nm_prinf("ptn_memdev_driver iounmap"); if (ptn_dev->pci_mem) { bus_release_resource(ptn_dev->dev, SYS_RES_MEMORY, PCIR_BAR(PTNETMAP_MEM_PCI_BAR), ptn_dev->pci_mem); ptn_dev->pci_mem = NULL; } } /* Device identification routine, return BUS_PROBE_DEFAULT on success, * positive on failure */ static int ptn_memdev_probe(device_t dev) { char desc[256]; if (pci_get_vendor(dev) != PTNETMAP_PCI_VENDOR_ID) return (ENXIO); if (pci_get_device(dev) != PTNETMAP_PCI_DEVICE_ID) return (ENXIO); snprintf(desc, sizeof(desc), "%s PCI adapter", PTNETMAP_MEMDEV_NAME); device_set_desc_copy(dev, desc); return (BUS_PROBE_DEFAULT); } /* Device initialization routine. */ static int ptn_memdev_attach(device_t dev) { struct ptnetmap_memdev *ptn_dev; int rid; uint16_t mem_id; ptn_dev = device_get_softc(dev); ptn_dev->dev = dev; pci_enable_busmaster(dev); rid = PCIR_BAR(PTNETMAP_IO_PCI_BAR); ptn_dev->pci_io = bus_alloc_resource_any(dev, SYS_RES_IOPORT, &rid, RF_ACTIVE); if (ptn_dev->pci_io == NULL) { device_printf(dev, "cannot map I/O space\n"); return (ENXIO); } mem_id = bus_read_4(ptn_dev->pci_io, PTNET_MDEV_IO_MEMID); /* create guest allocator */ ptn_dev->nm_mem = netmap_mem_pt_guest_attach(ptn_dev, mem_id); if (ptn_dev->nm_mem == NULL) { ptn_memdev_detach(dev); return (ENOMEM); } netmap_mem_get(ptn_dev->nm_mem); nm_prinf("ptnetmap memdev attached, host memid: %u", mem_id); return (0); } /* Device removal routine. */ static int ptn_memdev_detach(device_t dev) { struct ptnetmap_memdev *ptn_dev; ptn_dev = device_get_softc(dev); if (ptn_dev->nm_mem) { nm_prinf("ptnetmap memdev detached, host memid %u", netmap_mem_get_id(ptn_dev->nm_mem)); netmap_mem_put(ptn_dev->nm_mem); ptn_dev->nm_mem = NULL; } if (ptn_dev->pci_mem) { bus_release_resource(dev, SYS_RES_MEMORY, PCIR_BAR(PTNETMAP_MEM_PCI_BAR), ptn_dev->pci_mem); ptn_dev->pci_mem = NULL; } if (ptn_dev->pci_io) { bus_release_resource(dev, SYS_RES_IOPORT, PCIR_BAR(PTNETMAP_IO_PCI_BAR), ptn_dev->pci_io); ptn_dev->pci_io = NULL; } return (0); } static int ptn_memdev_shutdown(device_t dev) { return bus_generic_shutdown(dev); } #endif /* WITH_PTNETMAP */ /* * In order to track whether pages are still mapped, we hook into * the standard cdev_pager and intercept the constructor and * destructor. */ struct netmap_vm_handle_t { struct cdev *dev; struct netmap_priv_d *priv; }; static int netmap_dev_pager_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t foff, struct ucred *cred, u_short *color) { struct netmap_vm_handle_t *vmh = handle; if (netmap_verbose) nm_prinf("handle %p size %jd prot %d foff %jd", handle, (intmax_t)size, prot, (intmax_t)foff); if (color) *color = 0; dev_ref(vmh->dev); return 0; } static void netmap_dev_pager_dtor(void *handle) { struct netmap_vm_handle_t *vmh = handle; struct cdev *dev = vmh->dev; struct netmap_priv_d *priv = vmh->priv; if (netmap_verbose) nm_prinf("handle %p", handle); netmap_dtor(priv); free(vmh, M_DEVBUF); dev_rel(dev); } static int netmap_dev_pager_fault(vm_object_t object, vm_ooffset_t offset, int prot, vm_page_t *mres) { struct netmap_vm_handle_t *vmh = object->handle; struct netmap_priv_d *priv = vmh->priv; struct netmap_adapter *na = priv->np_na; vm_paddr_t paddr; vm_page_t page; vm_memattr_t memattr; vm_pindex_t pidx; nm_prdis("object %p offset %jd prot %d mres %p", object, (intmax_t)offset, prot, mres); memattr = object->memattr; pidx = OFF_TO_IDX(offset); paddr = netmap_mem_ofstophys(na->nm_mem, offset); if (paddr == 0) return VM_PAGER_FAIL; if (((*mres)->flags & PG_FICTITIOUS) != 0) { /* * If the passed in result page is a fake page, update it with * the new physical address. */ page = *mres; vm_page_updatefake(page, paddr, memattr); } else { /* * Replace the passed in reqpage page with our own fake page and * free up the all of the original pages. */ #ifndef VM_OBJECT_WUNLOCK /* FreeBSD < 10.x */ #define VM_OBJECT_WUNLOCK VM_OBJECT_UNLOCK #define VM_OBJECT_WLOCK VM_OBJECT_LOCK #endif /* VM_OBJECT_WUNLOCK */ VM_OBJECT_WUNLOCK(object); page = vm_page_getfake(paddr, memattr); VM_OBJECT_WLOCK(object); vm_page_free(*mres); *mres = page; vm_page_insert(page, object, pidx); } - page->valid = VM_PAGE_BITS_ALL; + vm_page_valid(page); return (VM_PAGER_OK); } static struct cdev_pager_ops netmap_cdev_pager_ops = { .cdev_pg_ctor = netmap_dev_pager_ctor, .cdev_pg_dtor = netmap_dev_pager_dtor, .cdev_pg_fault = netmap_dev_pager_fault, }; static int netmap_mmap_single(struct cdev *cdev, vm_ooffset_t *foff, vm_size_t objsize, vm_object_t *objp, int prot) { int error; struct netmap_vm_handle_t *vmh; struct netmap_priv_d *priv; vm_object_t obj; if (netmap_verbose) nm_prinf("cdev %p foff %jd size %jd objp %p prot %d", cdev, (intmax_t )*foff, (intmax_t )objsize, objp, prot); vmh = malloc(sizeof(struct netmap_vm_handle_t), M_DEVBUF, M_NOWAIT | M_ZERO); if (vmh == NULL) return ENOMEM; vmh->dev = cdev; NMG_LOCK(); error = devfs_get_cdevpriv((void**)&priv); if (error) goto err_unlock; if (priv->np_nifp == NULL) { error = EINVAL; goto err_unlock; } vmh->priv = priv; priv->np_refs++; NMG_UNLOCK(); obj = cdev_pager_allocate(vmh, OBJT_DEVICE, &netmap_cdev_pager_ops, objsize, prot, *foff, NULL); if (obj == NULL) { nm_prerr("cdev_pager_allocate failed"); error = EINVAL; goto err_deref; } *objp = obj; return 0; err_deref: NMG_LOCK(); priv->np_refs--; err_unlock: NMG_UNLOCK(); // err: free(vmh, M_DEVBUF); return error; } /* * On FreeBSD the close routine is only called on the last close on * the device (/dev/netmap) so we cannot do anything useful. * To track close() on individual file descriptors we pass netmap_dtor() to * devfs_set_cdevpriv() on open(). The FreeBSD kernel will call the destructor * when the last fd pointing to the device is closed. * * Note that FreeBSD does not even munmap() on close() so we also have * to track mmap() ourselves, and postpone the call to * netmap_dtor() is called when the process has no open fds and no active * memory maps on /dev/netmap, as in linux. */ static int netmap_close(struct cdev *dev, int fflag, int devtype, struct thread *td) { if (netmap_verbose) nm_prinf("dev %p fflag 0x%x devtype %d td %p", dev, fflag, devtype, td); return 0; } static int netmap_open(struct cdev *dev, int oflags, int devtype, struct thread *td) { struct netmap_priv_d *priv; int error; (void)dev; (void)oflags; (void)devtype; (void)td; NMG_LOCK(); priv = netmap_priv_new(); if (priv == NULL) { error = ENOMEM; goto out; } error = devfs_set_cdevpriv(priv, netmap_dtor); if (error) { netmap_priv_delete(priv); } out: NMG_UNLOCK(); return error; } /******************** kthread wrapper ****************/ #include u_int nm_os_ncpus(void) { return mp_maxid + 1; } struct nm_kctx_ctx { /* Userspace thread (kthread creator). */ struct thread *user_td; /* worker function and parameter */ nm_kctx_worker_fn_t worker_fn; void *worker_private; struct nm_kctx *nmk; /* integer to manage multiple worker contexts (e.g., RX or TX on ptnetmap) */ long type; }; struct nm_kctx { struct thread *worker; struct mtx worker_lock; struct nm_kctx_ctx worker_ctx; int run; /* used to stop kthread */ int attach_user; /* kthread attached to user_process */ int affinity; }; static void nm_kctx_worker(void *data) { struct nm_kctx *nmk = data; struct nm_kctx_ctx *ctx = &nmk->worker_ctx; if (nmk->affinity >= 0) { thread_lock(curthread); sched_bind(curthread, nmk->affinity); thread_unlock(curthread); } while (nmk->run) { /* * check if the parent process dies * (when kthread is attached to user process) */ if (ctx->user_td) { PROC_LOCK(curproc); thread_suspend_check(0); PROC_UNLOCK(curproc); } else { kthread_suspend_check(); } /* Continuously execute worker process. */ ctx->worker_fn(ctx->worker_private); /* worker body */ } kthread_exit(); } void nm_os_kctx_worker_setaff(struct nm_kctx *nmk, int affinity) { nmk->affinity = affinity; } struct nm_kctx * nm_os_kctx_create(struct nm_kctx_cfg *cfg, void *opaque) { struct nm_kctx *nmk = NULL; nmk = malloc(sizeof(*nmk), M_DEVBUF, M_NOWAIT | M_ZERO); if (!nmk) return NULL; mtx_init(&nmk->worker_lock, "nm_kthread lock", NULL, MTX_DEF); nmk->worker_ctx.worker_fn = cfg->worker_fn; nmk->worker_ctx.worker_private = cfg->worker_private; nmk->worker_ctx.type = cfg->type; nmk->affinity = -1; /* attach kthread to user process (ptnetmap) */ nmk->attach_user = cfg->attach_user; return nmk; } int nm_os_kctx_worker_start(struct nm_kctx *nmk) { struct proc *p = NULL; int error = 0; /* Temporarily disable this function as it is currently broken * and causes kernel crashes. The failure can be triggered by * the "vale_polling_enable_disable" test in ctrl-api-test.c. */ return EOPNOTSUPP; if (nmk->worker) return EBUSY; /* check if we want to attach kthread to user process */ if (nmk->attach_user) { nmk->worker_ctx.user_td = curthread; p = curthread->td_proc; } /* enable kthread main loop */ nmk->run = 1; /* create kthread */ if((error = kthread_add(nm_kctx_worker, nmk, p, &nmk->worker, RFNOWAIT /* to be checked */, 0, "nm-kthread-%ld", nmk->worker_ctx.type))) { goto err; } nm_prinf("nm_kthread started td %p", nmk->worker); return 0; err: nm_prerr("nm_kthread start failed err %d", error); nmk->worker = NULL; return error; } void nm_os_kctx_worker_stop(struct nm_kctx *nmk) { if (!nmk->worker) return; /* tell to kthread to exit from main loop */ nmk->run = 0; /* wake up kthread if it sleeps */ kthread_resume(nmk->worker); nmk->worker = NULL; } void nm_os_kctx_destroy(struct nm_kctx *nmk) { if (!nmk) return; if (nmk->worker) nm_os_kctx_worker_stop(nmk); free(nmk, M_DEVBUF); } /******************** kqueue support ****************/ /* * In addition to calling selwakeuppri(), nm_os_selwakeup() also * needs to call knote() to wake up kqueue listeners. * This operation is deferred to a taskqueue in order to avoid possible * lock order reversals; these may happen because knote() grabs a * private lock associated to the 'si' (see struct selinfo, * struct nm_selinfo, and nm_os_selinfo_init), and nm_os_selwakeup() * can be called while holding the lock associated to a different * 'si'. * When calling knote() we use a non-zero 'hint' argument to inform * the netmap_knrw() function that it is being called from * 'nm_os_selwakeup'; this is necessary because when netmap_knrw() is * called by the kevent subsystem (i.e. kevent_scan()) we also need to * call netmap_poll(). * * The netmap_kqfilter() function registers one or another f_event * depending on read or write mode. A pointer to the struct * 'netmap_priv_d' is stored into kn->kn_hook, so that it can later * be passed to netmap_poll(). We pass NULL as a third argument to * netmap_poll(), so that the latter only runs the txsync/rxsync * (if necessary), and skips the nm_os_selrecord() calls. */ void nm_os_selwakeup(struct nm_selinfo *si) { selwakeuppri(&si->si, PI_NET); if (si->kqueue_users > 0) { taskqueue_enqueue(si->ntfytq, &si->ntfytask); } } void nm_os_selrecord(struct thread *td, struct nm_selinfo *si) { selrecord(td, &si->si); } static void netmap_knrdetach(struct knote *kn) { struct netmap_priv_d *priv = (struct netmap_priv_d *)kn->kn_hook; struct nm_selinfo *si = priv->np_si[NR_RX]; knlist_remove(&si->si.si_note, kn, /*islocked=*/0); NMG_LOCK(); KASSERT(si->kqueue_users > 0, ("kqueue_user underflow on %s", si->mtxname)); si->kqueue_users--; nm_prinf("kqueue users for %s: %d", si->mtxname, si->kqueue_users); NMG_UNLOCK(); } static void netmap_knwdetach(struct knote *kn) { struct netmap_priv_d *priv = (struct netmap_priv_d *)kn->kn_hook; struct nm_selinfo *si = priv->np_si[NR_TX]; knlist_remove(&si->si.si_note, kn, /*islocked=*/0); NMG_LOCK(); si->kqueue_users--; nm_prinf("kqueue users for %s: %d", si->mtxname, si->kqueue_users); NMG_UNLOCK(); } /* * Callback triggered by netmap notifications (see netmap_notify()), * and by the application calling kevent(). In the former case we * just return 1 (events ready), since we are not able to do better. * In the latter case we use netmap_poll() to see which events are * ready. */ static int netmap_knrw(struct knote *kn, long hint, int events) { struct netmap_priv_d *priv; int revents; if (hint != 0) { /* Called from netmap_notify(), typically from a * thread different from the one issuing kevent(). * Assume we are ready. */ return 1; } /* Called from kevent(). */ priv = kn->kn_hook; revents = netmap_poll(priv, events, /*thread=*/NULL); return (events & revents) ? 1 : 0; } static int netmap_knread(struct knote *kn, long hint) { return netmap_knrw(kn, hint, POLLIN); } static int netmap_knwrite(struct knote *kn, long hint) { return netmap_knrw(kn, hint, POLLOUT); } static struct filterops netmap_rfiltops = { .f_isfd = 1, .f_detach = netmap_knrdetach, .f_event = netmap_knread, }; static struct filterops netmap_wfiltops = { .f_isfd = 1, .f_detach = netmap_knwdetach, .f_event = netmap_knwrite, }; /* * This is called when a thread invokes kevent() to record * a change in the configuration of the kqueue(). * The 'priv' is the one associated to the open netmap device. */ static int netmap_kqfilter(struct cdev *dev, struct knote *kn) { struct netmap_priv_d *priv; int error; struct netmap_adapter *na; struct nm_selinfo *si; int ev = kn->kn_filter; if (ev != EVFILT_READ && ev != EVFILT_WRITE) { nm_prerr("bad filter request %d", ev); return 1; } error = devfs_get_cdevpriv((void**)&priv); if (error) { nm_prerr("device not yet setup"); return 1; } na = priv->np_na; if (na == NULL) { nm_prerr("no netmap adapter for this file descriptor"); return 1; } /* the si is indicated in the priv */ si = priv->np_si[(ev == EVFILT_WRITE) ? NR_TX : NR_RX]; kn->kn_fop = (ev == EVFILT_WRITE) ? &netmap_wfiltops : &netmap_rfiltops; kn->kn_hook = priv; NMG_LOCK(); si->kqueue_users++; nm_prinf("kqueue users for %s: %d", si->mtxname, si->kqueue_users); NMG_UNLOCK(); knlist_add(&si->si.si_note, kn, /*islocked=*/0); return 0; } static int freebsd_netmap_poll(struct cdev *cdevi __unused, int events, struct thread *td) { struct netmap_priv_d *priv; if (devfs_get_cdevpriv((void **)&priv)) { return POLLERR; } return netmap_poll(priv, events, td); } static int freebsd_netmap_ioctl(struct cdev *dev __unused, u_long cmd, caddr_t data, int ffla __unused, struct thread *td) { int error; struct netmap_priv_d *priv; CURVNET_SET(TD_TO_VNET(td)); error = devfs_get_cdevpriv((void **)&priv); if (error) { /* XXX ENOENT should be impossible, since the priv * is now created in the open */ if (error == ENOENT) error = ENXIO; goto out; } error = netmap_ioctl(priv, cmd, data, td, /*nr_body_is_user=*/1); out: CURVNET_RESTORE(); return error; } void nm_os_onattach(struct ifnet *ifp) { ifp->if_capabilities |= IFCAP_NETMAP; } void nm_os_onenter(struct ifnet *ifp) { struct netmap_adapter *na = NA(ifp); na->if_transmit = ifp->if_transmit; ifp->if_transmit = netmap_transmit; ifp->if_capenable |= IFCAP_NETMAP; } void nm_os_onexit(struct ifnet *ifp) { struct netmap_adapter *na = NA(ifp); ifp->if_transmit = na->if_transmit; ifp->if_capenable &= ~IFCAP_NETMAP; } extern struct cdevsw netmap_cdevsw; /* XXX used in netmap.c, should go elsewhere */ struct cdevsw netmap_cdevsw = { .d_version = D_VERSION, .d_name = "netmap", .d_open = netmap_open, .d_mmap_single = netmap_mmap_single, .d_ioctl = freebsd_netmap_ioctl, .d_poll = freebsd_netmap_poll, .d_kqfilter = netmap_kqfilter, .d_close = netmap_close, }; /*--- end of kqueue support ----*/ /* * Kernel entry point. * * Initialize/finalize the module and return. * * Return 0 on success, errno on failure. */ static int netmap_loader(__unused struct module *module, int event, __unused void *arg) { int error = 0; switch (event) { case MOD_LOAD: error = netmap_init(); break; case MOD_UNLOAD: /* * if some one is still using netmap, * then the module can not be unloaded. */ if (netmap_use_count) { nm_prerr("netmap module can not be unloaded - netmap_use_count: %d", netmap_use_count); error = EBUSY; break; } netmap_fini(); break; default: error = EOPNOTSUPP; break; } return (error); } #ifdef DEV_MODULE_ORDERED /* * The netmap module contains three drivers: (i) the netmap character device * driver; (ii) the ptnetmap memdev PCI device driver, (iii) the ptnet PCI * device driver. The attach() routines of both (ii) and (iii) need the * lock of the global allocator, and such lock is initialized in netmap_init(), * which is part of (i). * Therefore, we make sure that (i) is loaded before (ii) and (iii), using * the 'order' parameter of driver declaration macros. For (i), we specify * SI_ORDER_MIDDLE, while higher orders are used with the DRIVER_MODULE_ORDERED * macros for (ii) and (iii). */ DEV_MODULE_ORDERED(netmap, netmap_loader, NULL, SI_ORDER_MIDDLE); #else /* !DEV_MODULE_ORDERED */ DEV_MODULE(netmap, netmap_loader, NULL); #endif /* DEV_MODULE_ORDERED */ MODULE_DEPEND(netmap, pci, 1, 1, 1); MODULE_VERSION(netmap, 1); /* reduce conditional code */ // linux API, use for the knlist in FreeBSD /* use a private mutex for the knlist */ Index: head/sys/dev/xen/gntdev/gntdev.c =================================================================== --- head/sys/dev/xen/gntdev/gntdev.c (revision 353538) +++ head/sys/dev/xen/gntdev/gntdev.c (revision 353539) @@ -1,1291 +1,1291 @@ /*- * Copyright (c) 2016 Akshay Jaggi * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * gntdev.c * * Interface to /dev/xen/gntdev. * */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include MALLOC_DEFINE(M_GNTDEV, "gntdev", "Xen grant-table user-space device"); #define MAX_OFFSET_COUNT ((0xffffffffffffffffull >> PAGE_SHIFT) + 1) static d_open_t gntdev_open; static d_ioctl_t gntdev_ioctl; static d_mmap_single_t gntdev_mmap_single; static struct cdevsw gntdev_devsw = { .d_version = D_VERSION, .d_open = gntdev_open, .d_ioctl = gntdev_ioctl, .d_mmap_single = gntdev_mmap_single, .d_name = "gntdev", }; static device_t gntdev_dev = NULL; struct gntdev_gref; struct gntdev_gmap; STAILQ_HEAD(gref_list_head, gntdev_gref); STAILQ_HEAD(gmap_list_head, gntdev_gmap); RB_HEAD(gref_tree_head, gntdev_gref); RB_HEAD(gmap_tree_head, gntdev_gmap); struct file_offset_struct { RB_ENTRY(file_offset_struct) next; uint64_t file_offset; uint64_t count; }; static int offset_cmp(struct file_offset_struct *f1, struct file_offset_struct *f2) { return (f1->file_offset - f2->file_offset); } RB_HEAD(file_offset_head, file_offset_struct); RB_GENERATE_STATIC(file_offset_head, file_offset_struct, next, offset_cmp); struct per_user_data { struct mtx user_data_lock; struct gref_tree_head gref_tree; struct gmap_tree_head gmap_tree; struct file_offset_head file_offset; }; /* * Get offset into the file which will be used while mmapping the * appropriate pages by the userspace program. */ static int get_file_offset(struct per_user_data *priv_user, uint32_t count, uint64_t *file_offset) { struct file_offset_struct *offset, *offset_tmp; if (count == 0) return (EINVAL); mtx_lock(&priv_user->user_data_lock); RB_FOREACH_SAFE(offset, file_offset_head, &priv_user->file_offset, offset_tmp) { if (offset->count >= count) { offset->count -= count; *file_offset = offset->file_offset + offset->count * PAGE_SIZE; if (offset->count == 0) { RB_REMOVE(file_offset_head, &priv_user->file_offset, offset); free(offset, M_GNTDEV); } mtx_unlock(&priv_user->user_data_lock); return (0); } } mtx_unlock(&priv_user->user_data_lock); return (ENOSPC); } static void put_file_offset(struct per_user_data *priv_user, uint32_t count, uint64_t file_offset) { struct file_offset_struct *offset, *offset_nxt, *offset_prv; offset = malloc(sizeof(*offset), M_GNTDEV, M_WAITOK | M_ZERO); offset->file_offset = file_offset; offset->count = count; mtx_lock(&priv_user->user_data_lock); RB_INSERT(file_offset_head, &priv_user->file_offset, offset); offset_nxt = RB_NEXT(file_offset_head, &priv_user->file_offset, offset); offset_prv = RB_PREV(file_offset_head, &priv_user->file_offset, offset); if (offset_nxt != NULL && offset_nxt->file_offset == offset->file_offset + offset->count * PAGE_SIZE) { offset->count += offset_nxt->count; RB_REMOVE(file_offset_head, &priv_user->file_offset, offset_nxt); free(offset_nxt, M_GNTDEV); } if (offset_prv != NULL && offset->file_offset == offset_prv->file_offset + offset_prv->count * PAGE_SIZE) { offset_prv->count += offset->count; RB_REMOVE(file_offset_head, &priv_user->file_offset, offset); free(offset, M_GNTDEV); } mtx_unlock(&priv_user->user_data_lock); } static int gntdev_gmap_pg_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t foff, struct ucred *cred, u_short *color); static void gntdev_gmap_pg_dtor(void *handle); static int gntdev_gmap_pg_fault(vm_object_t object, vm_ooffset_t offset, int prot, vm_page_t *mres); static struct cdev_pager_ops gntdev_gmap_pg_ops = { .cdev_pg_fault = gntdev_gmap_pg_fault, .cdev_pg_ctor = gntdev_gmap_pg_ctor, .cdev_pg_dtor = gntdev_gmap_pg_dtor, }; struct cleanup_data_struct { struct mtx to_kill_grefs_mtx; struct mtx to_kill_gmaps_mtx; struct gref_list_head to_kill_grefs; struct gmap_list_head to_kill_gmaps; }; static struct cleanup_data_struct cleanup_data = { .to_kill_grefs = STAILQ_HEAD_INITIALIZER(cleanup_data.to_kill_grefs), .to_kill_gmaps = STAILQ_HEAD_INITIALIZER(cleanup_data.to_kill_gmaps), }; MTX_SYSINIT(to_kill_grefs_mtx, &cleanup_data.to_kill_grefs_mtx, "gntdev to_kill_grefs mutex", MTX_DEF); MTX_SYSINIT(to_kill_gmaps_mtx, &cleanup_data.to_kill_gmaps_mtx, "gntdev to_kill_gmaps mutex", MTX_DEF); static void cleanup_function(void *arg, __unused int pending); static struct task cleanup_task = TASK_INITIALIZER(0, cleanup_function, &cleanup_data); struct notify_data { uint64_t index; uint32_t action; uint32_t event_channel_port; xen_intr_handle_t notify_evtchn_handle; }; static void notify(struct notify_data *notify, vm_page_t page); /*-------------------- Grant Allocation Methods -----------------------------*/ struct gntdev_gref { union gref_next_union { STAILQ_ENTRY(gntdev_gref) list; RB_ENTRY(gntdev_gref) tree; } gref_next; uint64_t file_index; grant_ref_t gref_id; vm_page_t page; struct notify_data *notify; }; static int gref_cmp(struct gntdev_gref *g1, struct gntdev_gref *g2) { return (g1->file_index - g2->file_index); } RB_GENERATE_STATIC(gref_tree_head, gntdev_gref, gref_next.tree, gref_cmp); /* * Traverse over the device-list of to-be-deleted grants allocated, and * if all accesses, both local mmaps and foreign maps, to them have ended, * destroy them. */ static void gref_list_dtor(struct cleanup_data_struct *cleanup_data) { struct gref_list_head tmp_grefs; struct gntdev_gref *gref, *gref_tmp, *gref_previous; STAILQ_INIT(&tmp_grefs); mtx_lock(&cleanup_data->to_kill_grefs_mtx); STAILQ_SWAP(&cleanup_data->to_kill_grefs, &tmp_grefs, gntdev_gref); mtx_unlock(&cleanup_data->to_kill_grefs_mtx); gref_previous = NULL; STAILQ_FOREACH_SAFE(gref, &tmp_grefs, gref_next.list, gref_tmp) { if (gref->page && gref->page->object == NULL) { if (gref->notify) { notify(gref->notify, gref->page); } if (gref->gref_id != GRANT_REF_INVALID) { if (gnttab_query_foreign_access(gref->gref_id)) continue; if (gnttab_end_foreign_access_ref(gref->gref_id) == 0) continue; gnttab_free_grant_reference(gref->gref_id); } vm_page_unwire_noq(gref->page); vm_page_free(gref->page); gref->page = NULL; } if (gref->page == NULL) { if (gref_previous == NULL) STAILQ_REMOVE_HEAD(&tmp_grefs, gref_next.list); else STAILQ_REMOVE_AFTER(&tmp_grefs, gref_previous, gref_next.list); if (gref->notify) free(gref->notify, M_GNTDEV); free(gref, M_GNTDEV); } else gref_previous = gref; } if (!STAILQ_EMPTY(&tmp_grefs)) { mtx_lock(&cleanup_data->to_kill_grefs_mtx); STAILQ_CONCAT(&cleanup_data->to_kill_grefs, &tmp_grefs); mtx_unlock(&cleanup_data->to_kill_grefs_mtx); } } /* * Find count number of contiguous allocated grants for a given userspace * program by file-offset (index). */ static struct gntdev_gref* gntdev_find_grefs(struct per_user_data *priv_user, uint64_t index, uint32_t count) { struct gntdev_gref find_gref, *gref, *gref_start = NULL; find_gref.file_index = index; mtx_lock(&priv_user->user_data_lock); gref_start = RB_FIND(gref_tree_head, &priv_user->gref_tree, &find_gref); for (gref = gref_start; gref != NULL && count > 0; gref = RB_NEXT(gref_tree_head, &priv_user->gref_tree, gref)) { if (index != gref->file_index) break; index += PAGE_SIZE; count--; } mtx_unlock(&priv_user->user_data_lock); if (count) return (NULL); return (gref_start); } /* * IOCTL_GNTDEV_ALLOC_GREF * Allocate required number of wired pages for the request, grant foreign * access to the physical frames for these pages, and add details about * this allocation to the per user private data, so that these pages can * be mmapped by the userspace program. */ static int gntdev_alloc_gref(struct ioctl_gntdev_alloc_gref *arg) { uint32_t i; int error, readonly; uint64_t file_offset; struct gntdev_gref *grefs; struct per_user_data *priv_user; readonly = !(arg->flags & GNTDEV_ALLOC_FLAG_WRITABLE); error = devfs_get_cdevpriv((void**) &priv_user); if (error != 0) return (EINVAL); /* Cleanup grefs and free pages. */ taskqueue_enqueue(taskqueue_thread, &cleanup_task); /* Get file offset for this request. */ error = get_file_offset(priv_user, arg->count, &file_offset); if (error != 0) return (error); /* Allocate grefs. */ grefs = malloc(sizeof(*grefs) * arg->count, M_GNTDEV, M_WAITOK); for (i = 0; i < arg->count; i++) { grefs[i].file_index = file_offset + i * PAGE_SIZE; grefs[i].gref_id = GRANT_REF_INVALID; grefs[i].notify = NULL; grefs[i].page = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO); if (grefs[i].page == NULL) { log(LOG_ERR, "Page allocation failed."); error = ENOMEM; break; } if ((grefs[i].page->flags & PG_ZERO) == 0) { /* * Zero the allocated page, as we don't want to * leak our memory to other domains. */ pmap_zero_page(grefs[i].page); } grefs[i].page->valid = VM_PAGE_BITS_ALL; error = gnttab_grant_foreign_access(arg->domid, (VM_PAGE_TO_PHYS(grefs[i].page) >> PAGE_SHIFT), readonly, &grefs[i].gref_id); if (error != 0) { log(LOG_ERR, "Grant Table Hypercall failed."); break; } } if (error != 0) { /* * If target domain maps the gref (by guessing the gref-id), * then we can't clean it up yet and we have to leave the * page in place so as to not leak our memory to that domain. * Add it to a global list to be cleaned up later. */ mtx_lock(&cleanup_data.to_kill_grefs_mtx); for (i = 0; i < arg->count; i++) STAILQ_INSERT_TAIL(&cleanup_data.to_kill_grefs, &grefs[i], gref_next.list); mtx_unlock(&cleanup_data.to_kill_grefs_mtx); taskqueue_enqueue(taskqueue_thread, &cleanup_task); return (error); } /* Copy the output values. */ arg->index = file_offset; for (i = 0; i < arg->count; i++) suword32(&arg->gref_ids[i], grefs[i].gref_id); /* Modify the per user private data. */ mtx_lock(&priv_user->user_data_lock); for (i = 0; i < arg->count; i++) RB_INSERT(gref_tree_head, &priv_user->gref_tree, &grefs[i]); mtx_unlock(&priv_user->user_data_lock); return (error); } /* * IOCTL_GNTDEV_DEALLOC_GREF * Remove grant allocation information from the per user private data, so * that it can't be mmapped anymore by the userspace program, and add it * to the to-be-deleted grants global device-list. */ static int gntdev_dealloc_gref(struct ioctl_gntdev_dealloc_gref *arg) { int error; uint32_t count; struct gntdev_gref *gref, *gref_tmp; struct per_user_data *priv_user; error = devfs_get_cdevpriv((void**) &priv_user); if (error != 0) return (EINVAL); gref = gntdev_find_grefs(priv_user, arg->index, arg->count); if (gref == NULL) { log(LOG_ERR, "Can't find requested grant-refs."); return (EINVAL); } /* Remove the grefs from user private data. */ count = arg->count; mtx_lock(&priv_user->user_data_lock); mtx_lock(&cleanup_data.to_kill_grefs_mtx); for (; gref != NULL && count > 0; gref = gref_tmp) { gref_tmp = RB_NEXT(gref_tree_head, &priv_user->gref_tree, gref); RB_REMOVE(gref_tree_head, &priv_user->gref_tree, gref); STAILQ_INSERT_TAIL(&cleanup_data.to_kill_grefs, gref, gref_next.list); count--; } mtx_unlock(&cleanup_data.to_kill_grefs_mtx); mtx_unlock(&priv_user->user_data_lock); taskqueue_enqueue(taskqueue_thread, &cleanup_task); put_file_offset(priv_user, arg->count, arg->index); return (0); } /*-------------------- Grant Mapping Methods --------------------------------*/ struct gntdev_gmap_map { vm_object_t mem; struct resource *pseudo_phys_res; int pseudo_phys_res_id; vm_paddr_t phys_base_addr; }; struct gntdev_gmap { union gmap_next_union { STAILQ_ENTRY(gntdev_gmap) list; RB_ENTRY(gntdev_gmap) tree; } gmap_next; uint64_t file_index; uint32_t count; struct gnttab_map_grant_ref *grant_map_ops; struct gntdev_gmap_map *map; struct notify_data *notify; }; static int gmap_cmp(struct gntdev_gmap *g1, struct gntdev_gmap *g2) { return (g1->file_index - g2->file_index); } RB_GENERATE_STATIC(gmap_tree_head, gntdev_gmap, gmap_next.tree, gmap_cmp); /* * Traverse over the device-list of to-be-deleted grant mappings, and if * the region is no longer mmapped by anyone, free the memory used to * store information about the mapping. */ static void gmap_list_dtor(struct cleanup_data_struct *cleanup_data) { struct gmap_list_head tmp_gmaps; struct gntdev_gmap *gmap, *gmap_tmp, *gmap_previous; STAILQ_INIT(&tmp_gmaps); mtx_lock(&cleanup_data->to_kill_gmaps_mtx); STAILQ_SWAP(&cleanup_data->to_kill_gmaps, &tmp_gmaps, gntdev_gmap); mtx_unlock(&cleanup_data->to_kill_gmaps_mtx); gmap_previous = NULL; STAILQ_FOREACH_SAFE(gmap, &tmp_gmaps, gmap_next.list, gmap_tmp) { if (gmap->map == NULL) { if (gmap_previous == NULL) STAILQ_REMOVE_HEAD(&tmp_gmaps, gmap_next.list); else STAILQ_REMOVE_AFTER(&tmp_gmaps, gmap_previous, gmap_next.list); if (gmap->notify) free(gmap->notify, M_GNTDEV); free(gmap->grant_map_ops, M_GNTDEV); free(gmap, M_GNTDEV); } else gmap_previous = gmap; } if (!STAILQ_EMPTY(&tmp_gmaps)) { mtx_lock(&cleanup_data->to_kill_gmaps_mtx); STAILQ_CONCAT(&cleanup_data->to_kill_gmaps, &tmp_gmaps); mtx_unlock(&cleanup_data->to_kill_gmaps_mtx); } } /* * Find mapped grants for a given userspace program, by file-offset (index) * and count, as supplied during the map-ioctl. */ static struct gntdev_gmap* gntdev_find_gmap(struct per_user_data *priv_user, uint64_t index, uint32_t count) { struct gntdev_gmap find_gmap, *gmap; find_gmap.file_index = index; mtx_lock(&priv_user->user_data_lock); gmap = RB_FIND(gmap_tree_head, &priv_user->gmap_tree, &find_gmap); mtx_unlock(&priv_user->user_data_lock); if (gmap != NULL && gmap->count == count) return (gmap); return (NULL); } /* * Remove the pages from the mgtdevice pager, call the unmap hypercall, * free the xenmem resource. This function is called during the * destruction of the mgtdevice pager, which happens when all mmaps to * it have been removed, and the unmap-ioctl has been performed. */ static int notify_unmap_cleanup(struct gntdev_gmap *gmap) { uint32_t i; int error, count; vm_page_t m; struct gnttab_unmap_grant_ref *unmap_ops; unmap_ops = malloc(sizeof(struct gnttab_unmap_grant_ref) * gmap->count, M_GNTDEV, M_WAITOK); /* Enumerate freeable maps. */ count = 0; for (i = 0; i < gmap->count; i++) { if (gmap->grant_map_ops[i].handle != -1) { unmap_ops[count].handle = gmap->grant_map_ops[i].handle; unmap_ops[count].host_addr = gmap->grant_map_ops[i].host_addr; unmap_ops[count].dev_bus_addr = 0; count++; } } /* Perform notification. */ if (count > 0 && gmap->notify) { vm_page_t page; uint64_t page_offset; page_offset = gmap->notify->index - gmap->file_index; page = PHYS_TO_VM_PAGE(gmap->map->phys_base_addr + page_offset); notify(gmap->notify, page); } /* Free the pages. */ VM_OBJECT_WLOCK(gmap->map->mem); retry: for (i = 0; i < gmap->count; i++) { m = vm_page_lookup(gmap->map->mem, i); if (m == NULL) continue; if (vm_page_busy_acquire(m, VM_ALLOC_WAITFAIL) == 0) goto retry; cdev_pager_free_page(gmap->map->mem, m); } VM_OBJECT_WUNLOCK(gmap->map->mem); /* Perform unmap hypercall. */ error = HYPERVISOR_grant_table_op(GNTTABOP_unmap_grant_ref, unmap_ops, count); for (i = 0; i < gmap->count; i++) { gmap->grant_map_ops[i].handle = -1; gmap->grant_map_ops[i].host_addr = 0; } if (gmap->map) { error = xenmem_free(gntdev_dev, gmap->map->pseudo_phys_res_id, gmap->map->pseudo_phys_res); KASSERT(error == 0, ("Unable to release memory resource: %d", error)); free(gmap->map, M_GNTDEV); gmap->map = NULL; } free(unmap_ops, M_GNTDEV); return (error); } /* * IOCTL_GNTDEV_MAP_GRANT_REF * Populate structures for mapping the grant reference in the per user * private data. Actual resource allocation and map hypercall is performed * during the mmap. */ static int gntdev_map_grant_ref(struct ioctl_gntdev_map_grant_ref *arg) { uint32_t i; int error; struct gntdev_gmap *gmap; struct per_user_data *priv_user; error = devfs_get_cdevpriv((void**) &priv_user); if (error != 0) return (EINVAL); gmap = malloc(sizeof(*gmap), M_GNTDEV, M_WAITOK | M_ZERO); gmap->count = arg->count; gmap->grant_map_ops = malloc(sizeof(struct gnttab_map_grant_ref) * arg->count, M_GNTDEV, M_WAITOK | M_ZERO); for (i = 0; i < arg->count; i++) { struct ioctl_gntdev_grant_ref ref; error = copyin(&arg->refs[i], &ref, sizeof(ref)); if (error != 0) { free(gmap->grant_map_ops, M_GNTDEV); free(gmap, M_GNTDEV); return (error); } gmap->grant_map_ops[i].dom = ref.domid; gmap->grant_map_ops[i].ref = ref.ref; gmap->grant_map_ops[i].handle = -1; gmap->grant_map_ops[i].flags = GNTMAP_host_map; } error = get_file_offset(priv_user, arg->count, &gmap->file_index); if (error != 0) { free(gmap->grant_map_ops, M_GNTDEV); free(gmap, M_GNTDEV); return (error); } mtx_lock(&priv_user->user_data_lock); RB_INSERT(gmap_tree_head, &priv_user->gmap_tree, gmap); mtx_unlock(&priv_user->user_data_lock); arg->index = gmap->file_index; return (error); } /* * IOCTL_GNTDEV_UNMAP_GRANT_REF * Remove the map information from the per user private data and add it * to the global device-list of mappings to be deleted. A reference to * the mgtdevice pager is also decreased, the reason for which is * explained in mmap_gmap(). */ static int gntdev_unmap_grant_ref(struct ioctl_gntdev_unmap_grant_ref *arg) { int error; struct gntdev_gmap *gmap; struct per_user_data *priv_user; error = devfs_get_cdevpriv((void**) &priv_user); if (error != 0) return (EINVAL); gmap = gntdev_find_gmap(priv_user, arg->index, arg->count); if (gmap == NULL) { log(LOG_ERR, "Can't find requested grant-map."); return (EINVAL); } mtx_lock(&priv_user->user_data_lock); mtx_lock(&cleanup_data.to_kill_gmaps_mtx); RB_REMOVE(gmap_tree_head, &priv_user->gmap_tree, gmap); STAILQ_INSERT_TAIL(&cleanup_data.to_kill_gmaps, gmap, gmap_next.list); mtx_unlock(&cleanup_data.to_kill_gmaps_mtx); mtx_unlock(&priv_user->user_data_lock); if (gmap->map) vm_object_deallocate(gmap->map->mem); taskqueue_enqueue(taskqueue_thread, &cleanup_task); put_file_offset(priv_user, arg->count, arg->index); return (0); } /* * IOCTL_GNTDEV_GET_OFFSET_FOR_VADDR * Get file-offset and count for a given mapping, from the virtual address * where the mapping is mmapped. * Please note, this only works for grants mapped by this domain, and not * grants allocated. Count doesn't make much sense in reference to grants * allocated. Also, because this function is present in the linux gntdev * device, but not in the linux gntalloc one, most userspace code only use * it for mapped grants. */ static int gntdev_get_offset_for_vaddr(struct ioctl_gntdev_get_offset_for_vaddr *arg, struct thread *td) { int error; vm_map_t map; vm_map_entry_t entry; vm_object_t mem; vm_pindex_t pindex; vm_prot_t prot; boolean_t wired; struct gntdev_gmap *gmap; int rc; map = &td->td_proc->p_vmspace->vm_map; error = vm_map_lookup(&map, arg->vaddr, VM_PROT_NONE, &entry, &mem, &pindex, &prot, &wired); if (error != KERN_SUCCESS) return (EINVAL); if ((mem->type != OBJT_MGTDEVICE) || (mem->un_pager.devp.ops != &gntdev_gmap_pg_ops)) { rc = EINVAL; goto out; } gmap = mem->handle; if (gmap == NULL || (entry->end - entry->start) != (gmap->count * PAGE_SIZE)) { rc = EINVAL; goto out; } arg->count = gmap->count; arg->offset = gmap->file_index; rc = 0; out: vm_map_lookup_done(map, entry); return (rc); } /*-------------------- Grant Mapping Pager ----------------------------------*/ static int gntdev_gmap_pg_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t foff, struct ucred *cred, u_short *color) { return (0); } static void gntdev_gmap_pg_dtor(void *handle) { notify_unmap_cleanup((struct gntdev_gmap *)handle); } static int gntdev_gmap_pg_fault(vm_object_t object, vm_ooffset_t offset, int prot, vm_page_t *mres) { struct gntdev_gmap *gmap = object->handle; vm_pindex_t pidx, ridx; vm_page_t page, oldm; vm_ooffset_t relative_offset; if (gmap->map == NULL) return (VM_PAGER_FAIL); relative_offset = offset - gmap->file_index; pidx = OFF_TO_IDX(offset); ridx = OFF_TO_IDX(relative_offset); if (ridx >= gmap->count || gmap->grant_map_ops[ridx].status != GNTST_okay) return (VM_PAGER_FAIL); page = PHYS_TO_VM_PAGE(gmap->map->phys_base_addr + relative_offset); if (page == NULL) return (VM_PAGER_FAIL); KASSERT((page->flags & PG_FICTITIOUS) != 0, ("not fictitious %p", page)); KASSERT(vm_page_wired(page), ("page %p is not wired", page)); KASSERT(!vm_page_busied(page), ("page %p is busy", page)); if (*mres != NULL) { oldm = *mres; vm_page_free(oldm); *mres = NULL; } vm_page_busy_acquire(page, 0); + vm_page_valid(page); vm_page_insert(page, object, pidx); - page->valid = VM_PAGE_BITS_ALL; *mres = page; return (VM_PAGER_OK); } /*------------------ Grant Table Methods ------------------------------------*/ static void notify(struct notify_data *notify, vm_page_t page) { if (notify->action & UNMAP_NOTIFY_CLEAR_BYTE) { uint8_t *mem; uint64_t offset; offset = notify->index & PAGE_MASK; mem = (uint8_t *)pmap_quick_enter_page(page); mem[offset] = 0; pmap_quick_remove_page((vm_offset_t)mem); } if (notify->action & UNMAP_NOTIFY_SEND_EVENT) { xen_intr_signal(notify->notify_evtchn_handle); xen_intr_unbind(¬ify->notify_evtchn_handle); } notify->action = 0; } /* * Helper to copy new arguments from the notify ioctl into * the existing notify data. */ static int copy_notify_helper(struct notify_data *destination, struct ioctl_gntdev_unmap_notify *source) { xen_intr_handle_t handlep = NULL; /* * "Get" before "Put"ting previous reference, as we might be * holding the last reference to the event channel port. */ if (source->action & UNMAP_NOTIFY_SEND_EVENT) if (xen_intr_get_evtchn_from_port(source->event_channel_port, &handlep) != 0) return (EINVAL); if (destination->action & UNMAP_NOTIFY_SEND_EVENT) xen_intr_unbind(&destination->notify_evtchn_handle); destination->action = source->action; destination->event_channel_port = source->event_channel_port; destination->index = source->index; destination->notify_evtchn_handle = handlep; return (0); } /* * IOCTL_GNTDEV_SET_UNMAP_NOTIFY * Set unmap notification inside the appropriate grant. It sends a * notification when the grant is completely munmapped by this domain * and ready for destruction. */ static int gntdev_set_unmap_notify(struct ioctl_gntdev_unmap_notify *arg) { int error; uint64_t index; struct per_user_data *priv_user; struct gntdev_gref *gref = NULL; struct gntdev_gmap *gmap; error = devfs_get_cdevpriv((void**) &priv_user); if (error != 0) return (EINVAL); if (arg->action & ~(UNMAP_NOTIFY_CLEAR_BYTE|UNMAP_NOTIFY_SEND_EVENT)) return (EINVAL); index = arg->index & ~PAGE_MASK; gref = gntdev_find_grefs(priv_user, index, 1); if (gref) { if (gref->notify == NULL) gref->notify = malloc(sizeof(*arg), M_GNTDEV, M_WAITOK | M_ZERO); return (copy_notify_helper(gref->notify, arg)); } error = EINVAL; mtx_lock(&priv_user->user_data_lock); RB_FOREACH(gmap, gmap_tree_head, &priv_user->gmap_tree) { if (arg->index >= gmap->file_index && arg->index < gmap->file_index + gmap->count * PAGE_SIZE) { if (gmap->notify == NULL) gmap->notify = malloc(sizeof(*arg), M_GNTDEV, M_WAITOK | M_ZERO); error = copy_notify_helper(gmap->notify, arg); break; } } mtx_unlock(&priv_user->user_data_lock); return (error); } /*------------------ Gntdev Char Device Methods -----------------------------*/ static void cleanup_function(void *arg, __unused int pending) { gref_list_dtor((struct cleanup_data_struct *) arg); gmap_list_dtor((struct cleanup_data_struct *) arg); } static void per_user_data_dtor(void *arg) { struct gntdev_gref *gref, *gref_tmp; struct gntdev_gmap *gmap, *gmap_tmp; struct file_offset_struct *offset, *offset_tmp; struct per_user_data *priv_user; priv_user = (struct per_user_data *) arg; mtx_lock(&priv_user->user_data_lock); mtx_lock(&cleanup_data.to_kill_grefs_mtx); RB_FOREACH_SAFE(gref, gref_tree_head, &priv_user->gref_tree, gref_tmp) { RB_REMOVE(gref_tree_head, &priv_user->gref_tree, gref); STAILQ_INSERT_TAIL(&cleanup_data.to_kill_grefs, gref, gref_next.list); } mtx_unlock(&cleanup_data.to_kill_grefs_mtx); mtx_lock(&cleanup_data.to_kill_gmaps_mtx); RB_FOREACH_SAFE(gmap, gmap_tree_head, &priv_user->gmap_tree, gmap_tmp) { RB_REMOVE(gmap_tree_head, &priv_user->gmap_tree, gmap); STAILQ_INSERT_TAIL(&cleanup_data.to_kill_gmaps, gmap, gmap_next.list); if (gmap->map) vm_object_deallocate(gmap->map->mem); } mtx_unlock(&cleanup_data.to_kill_gmaps_mtx); RB_FOREACH_SAFE(offset, file_offset_head, &priv_user->file_offset, offset_tmp) { RB_REMOVE(file_offset_head, &priv_user->file_offset, offset); free(offset, M_GNTDEV); } mtx_unlock(&priv_user->user_data_lock); taskqueue_enqueue(taskqueue_thread, &cleanup_task); mtx_destroy(&priv_user->user_data_lock); free(priv_user, M_GNTDEV); } static int gntdev_open(struct cdev *dev, int flag, int otyp, struct thread *td) { int error; struct per_user_data *priv_user; struct file_offset_struct *offset; priv_user = malloc(sizeof(*priv_user), M_GNTDEV, M_WAITOK | M_ZERO); RB_INIT(&priv_user->gref_tree); RB_INIT(&priv_user->gmap_tree); RB_INIT(&priv_user->file_offset); offset = malloc(sizeof(*offset), M_GNTDEV, M_WAITOK | M_ZERO); offset->file_offset = 0; offset->count = MAX_OFFSET_COUNT; RB_INSERT(file_offset_head, &priv_user->file_offset, offset); mtx_init(&priv_user->user_data_lock, "per user data mutex", NULL, MTX_DEF); error = devfs_set_cdevpriv(priv_user, per_user_data_dtor); if (error != 0) per_user_data_dtor(priv_user); return (error); } static int gntdev_ioctl(struct cdev *dev, u_long cmd, caddr_t data, int fflag, struct thread *td) { int error; switch (cmd) { case IOCTL_GNTDEV_SET_UNMAP_NOTIFY: error = gntdev_set_unmap_notify( (struct ioctl_gntdev_unmap_notify*) data); break; case IOCTL_GNTDEV_ALLOC_GREF: error = gntdev_alloc_gref( (struct ioctl_gntdev_alloc_gref*) data); break; case IOCTL_GNTDEV_DEALLOC_GREF: error = gntdev_dealloc_gref( (struct ioctl_gntdev_dealloc_gref*) data); break; case IOCTL_GNTDEV_MAP_GRANT_REF: error = gntdev_map_grant_ref( (struct ioctl_gntdev_map_grant_ref*) data); break; case IOCTL_GNTDEV_UNMAP_GRANT_REF: error = gntdev_unmap_grant_ref( (struct ioctl_gntdev_unmap_grant_ref*) data); break; case IOCTL_GNTDEV_GET_OFFSET_FOR_VADDR: error = gntdev_get_offset_for_vaddr( (struct ioctl_gntdev_get_offset_for_vaddr*) data, td); break; default: error = ENOSYS; break; } return (error); } /* * MMAP an allocated grant into user memory. * Please note, that the grants must not already be mmapped, otherwise * this function will fail. */ static int mmap_gref(struct per_user_data *priv_user, struct gntdev_gref *gref_start, uint32_t count, vm_size_t size, struct vm_object **object) { vm_object_t mem_obj; struct gntdev_gref *gref; mem_obj = vm_object_allocate(OBJT_PHYS, size); if (mem_obj == NULL) return (ENOMEM); mtx_lock(&priv_user->user_data_lock); VM_OBJECT_WLOCK(mem_obj); for (gref = gref_start; gref != NULL && count > 0; gref = RB_NEXT(gref_tree_head, &priv_user->gref_tree, gref)) { if (gref->page->object) break; vm_page_insert(gref->page, mem_obj, OFF_TO_IDX(gref->file_index)); count--; } VM_OBJECT_WUNLOCK(mem_obj); mtx_unlock(&priv_user->user_data_lock); if (count) { vm_object_deallocate(mem_obj); return (EINVAL); } *object = mem_obj; return (0); } /* * MMAP a mapped grant into user memory. */ static int mmap_gmap(struct per_user_data *priv_user, struct gntdev_gmap *gmap_start, vm_ooffset_t *offset, vm_size_t size, struct vm_object **object, int nprot) { uint32_t i; int error; /* * The grant map hypercall might already be done. * If that is the case, increase a reference to the * vm object and return the already allocated object. */ if (gmap_start->map) { vm_object_reference(gmap_start->map->mem); *object = gmap_start->map->mem; return (0); } gmap_start->map = malloc(sizeof(*(gmap_start->map)), M_GNTDEV, M_WAITOK | M_ZERO); /* Allocate the xen pseudo physical memory resource. */ gmap_start->map->pseudo_phys_res_id = 0; gmap_start->map->pseudo_phys_res = xenmem_alloc(gntdev_dev, &gmap_start->map->pseudo_phys_res_id, size); if (gmap_start->map->pseudo_phys_res == NULL) { free(gmap_start->map, M_GNTDEV); gmap_start->map = NULL; return (ENOMEM); } gmap_start->map->phys_base_addr = rman_get_start(gmap_start->map->pseudo_phys_res); /* Allocate the mgtdevice pager. */ gmap_start->map->mem = cdev_pager_allocate(gmap_start, OBJT_MGTDEVICE, &gntdev_gmap_pg_ops, size, nprot, *offset, NULL); if (gmap_start->map->mem == NULL) { xenmem_free(gntdev_dev, gmap_start->map->pseudo_phys_res_id, gmap_start->map->pseudo_phys_res); free(gmap_start->map, M_GNTDEV); gmap_start->map = NULL; return (ENOMEM); } for (i = 0; i < gmap_start->count; i++) { gmap_start->grant_map_ops[i].host_addr = gmap_start->map->phys_base_addr + i * PAGE_SIZE; if ((nprot & PROT_WRITE) == 0) gmap_start->grant_map_ops[i].flags |= GNTMAP_readonly; } /* Make the MAP hypercall. */ error = HYPERVISOR_grant_table_op(GNTTABOP_map_grant_ref, gmap_start->grant_map_ops, gmap_start->count); if (error != 0) { /* * Deallocate pager. * Pager deallocation will automatically take care of * xenmem deallocation, etc. */ vm_object_deallocate(gmap_start->map->mem); return (EINVAL); } /* Retry EAGAIN maps. */ for (i = 0; i < gmap_start->count; i++) { int delay = 1; while (delay < 256 && gmap_start->grant_map_ops[i].status == GNTST_eagain) { HYPERVISOR_grant_table_op( GNTTABOP_map_grant_ref, &gmap_start->grant_map_ops[i], 1); pause(("gntmap"), delay * SBT_1MS); delay++; } if (gmap_start->grant_map_ops[i].status == GNTST_eagain) gmap_start->grant_map_ops[i].status = GNTST_bad_page; if (gmap_start->grant_map_ops[i].status != GNTST_okay) { /* * Deallocate pager. * Pager deallocation will automatically take care of * xenmem deallocation, notification, unmap hypercall, * etc. */ vm_object_deallocate(gmap_start->map->mem); return (EINVAL); } } /* * Add a reference to the vm object. We do not want * the vm object to be deleted when all the mmaps are * unmapped, because it may be re-mmapped. Instead, * we want the object to be deleted, when along with * munmaps, we have also processed the unmap-ioctl. */ vm_object_reference(gmap_start->map->mem); *object = gmap_start->map->mem; return (0); } static int gntdev_mmap_single(struct cdev *cdev, vm_ooffset_t *offset, vm_size_t size, struct vm_object **object, int nprot) { int error; uint32_t count; struct gntdev_gref *gref_start; struct gntdev_gmap *gmap_start; struct per_user_data *priv_user; error = devfs_get_cdevpriv((void**) &priv_user); if (error != 0) return (EINVAL); count = OFF_TO_IDX(size); gref_start = gntdev_find_grefs(priv_user, *offset, count); if (gref_start) { error = mmap_gref(priv_user, gref_start, count, size, object); return (error); } gmap_start = gntdev_find_gmap(priv_user, *offset, count); if (gmap_start) { error = mmap_gmap(priv_user, gmap_start, offset, size, object, nprot); return (error); } return (EINVAL); } /*------------------ Private Device Attachment Functions --------------------*/ static void gntdev_identify(driver_t *driver, device_t parent) { KASSERT((xen_domain()), ("Trying to attach gntdev device on non Xen domain")); if (BUS_ADD_CHILD(parent, 0, "gntdev", 0) == NULL) panic("unable to attach gntdev user-space device"); } static int gntdev_probe(device_t dev) { gntdev_dev = dev; device_set_desc(dev, "Xen grant-table user-space device"); return (BUS_PROBE_NOWILDCARD); } static int gntdev_attach(device_t dev) { make_dev_credf(MAKEDEV_ETERNAL, &gntdev_devsw, 0, NULL, UID_ROOT, GID_WHEEL, 0600, "xen/gntdev"); return (0); } /*-------------------- Private Device Attachment Data -----------------------*/ static device_method_t gntdev_methods[] = { DEVMETHOD(device_identify, gntdev_identify), DEVMETHOD(device_probe, gntdev_probe), DEVMETHOD(device_attach, gntdev_attach), DEVMETHOD_END }; static driver_t gntdev_driver = { "gntdev", gntdev_methods, 0, }; devclass_t gntdev_devclass; DRIVER_MODULE(gntdev, xenpv, gntdev_driver, gntdev_devclass, 0, 0); MODULE_DEPEND(gntdev, xenpv, 1, 1, 1); Index: head/sys/dev/xen/privcmd/privcmd.c =================================================================== --- head/sys/dev/xen/privcmd/privcmd.c (revision 353538) +++ head/sys/dev/xen/privcmd/privcmd.c (revision 353539) @@ -1,429 +1,429 @@ /* * Copyright (c) 2014 Roger Pau Monné * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include MALLOC_DEFINE(M_PRIVCMD, "privcmd_dev", "Xen privcmd user-space device"); struct privcmd_map { vm_object_t mem; vm_size_t size; struct resource *pseudo_phys_res; int pseudo_phys_res_id; vm_paddr_t phys_base_addr; boolean_t mapped; BITSET_DEFINE_VAR() *err; }; static d_ioctl_t privcmd_ioctl; static d_mmap_single_t privcmd_mmap_single; static struct cdevsw privcmd_devsw = { .d_version = D_VERSION, .d_ioctl = privcmd_ioctl, .d_mmap_single = privcmd_mmap_single, .d_name = "privcmd", }; static int privcmd_pg_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t foff, struct ucred *cred, u_short *color); static void privcmd_pg_dtor(void *handle); static int privcmd_pg_fault(vm_object_t object, vm_ooffset_t offset, int prot, vm_page_t *mres); static struct cdev_pager_ops privcmd_pg_ops = { .cdev_pg_fault = privcmd_pg_fault, .cdev_pg_ctor = privcmd_pg_ctor, .cdev_pg_dtor = privcmd_pg_dtor, }; static device_t privcmd_dev = NULL; /*------------------------- Privcmd Pager functions --------------------------*/ static int privcmd_pg_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t foff, struct ucred *cred, u_short *color) { return (0); } static void privcmd_pg_dtor(void *handle) { struct xen_remove_from_physmap rm = { .domid = DOMID_SELF }; struct privcmd_map *map = handle; int error; vm_size_t i; vm_page_t m; /* * Remove the mappings from the used pages. This will remove the * underlying p2m bindings in Xen second stage translation. */ if (map->mapped == true) { VM_OBJECT_WLOCK(map->mem); retry: for (i = 0; i < map->size; i++) { m = vm_page_lookup(map->mem, i); if (m == NULL) continue; if (vm_page_busy_acquire(m, VM_ALLOC_WAITFAIL) == 0) goto retry; cdev_pager_free_page(map->mem, m); } VM_OBJECT_WUNLOCK(map->mem); for (i = 0; i < map->size; i++) { rm.gpfn = atop(map->phys_base_addr) + i; HYPERVISOR_memory_op(XENMEM_remove_from_physmap, &rm); } free(map->err, M_PRIVCMD); } error = xenmem_free(privcmd_dev, map->pseudo_phys_res_id, map->pseudo_phys_res); KASSERT(error == 0, ("Unable to release memory resource: %d", error)); free(map, M_PRIVCMD); } static int privcmd_pg_fault(vm_object_t object, vm_ooffset_t offset, int prot, vm_page_t *mres) { struct privcmd_map *map = object->handle; vm_pindex_t pidx; vm_page_t page, oldm; if (map->mapped != true) return (VM_PAGER_FAIL); pidx = OFF_TO_IDX(offset); if (pidx >= map->size || BIT_ISSET(map->size, pidx, map->err)) return (VM_PAGER_FAIL); page = PHYS_TO_VM_PAGE(map->phys_base_addr + offset); if (page == NULL) return (VM_PAGER_FAIL); KASSERT((page->flags & PG_FICTITIOUS) != 0, ("not fictitious %p", page)); KASSERT(vm_page_wired(page), ("page %p not wired", page)); KASSERT(!vm_page_busied(page), ("page %p is busy", page)); if (*mres != NULL) { oldm = *mres; vm_page_free(oldm); *mres = NULL; } vm_page_busy_acquire(page, 0); + vm_page_valid(page); vm_page_insert(page, object, pidx); - page->valid = VM_PAGE_BITS_ALL; *mres = page; return (VM_PAGER_OK); } /*----------------------- Privcmd char device methods ------------------------*/ static int privcmd_mmap_single(struct cdev *cdev, vm_ooffset_t *offset, vm_size_t size, vm_object_t *object, int nprot) { struct privcmd_map *map; map = malloc(sizeof(*map), M_PRIVCMD, M_WAITOK | M_ZERO); map->size = OFF_TO_IDX(size); map->pseudo_phys_res_id = 0; map->pseudo_phys_res = xenmem_alloc(privcmd_dev, &map->pseudo_phys_res_id, size); if (map->pseudo_phys_res == NULL) { free(map, M_PRIVCMD); return (ENOMEM); } map->phys_base_addr = rman_get_start(map->pseudo_phys_res); map->mem = cdev_pager_allocate(map, OBJT_MGTDEVICE, &privcmd_pg_ops, size, nprot, *offset, NULL); if (map->mem == NULL) { xenmem_free(privcmd_dev, map->pseudo_phys_res_id, map->pseudo_phys_res); free(map, M_PRIVCMD); return (ENOMEM); } *object = map->mem; return (0); } static int privcmd_ioctl(struct cdev *dev, unsigned long cmd, caddr_t arg, int mode, struct thread *td) { int error, i; switch (cmd) { case IOCTL_PRIVCMD_HYPERCALL: { struct ioctl_privcmd_hypercall *hcall; hcall = (struct ioctl_privcmd_hypercall *)arg; #ifdef __amd64__ /* * The hypervisor page table walker will refuse to access * user-space pages if SMAP is enabled, so temporary disable it * while performing the hypercall. */ if (cpu_stdext_feature & CPUID_STDEXT_SMAP) stac(); #endif error = privcmd_hypercall(hcall->op, hcall->arg[0], hcall->arg[1], hcall->arg[2], hcall->arg[3], hcall->arg[4]); #ifdef __amd64__ if (cpu_stdext_feature & CPUID_STDEXT_SMAP) clac(); #endif if (error >= 0) { hcall->retval = error; error = 0; } else { error = xen_translate_error(error); hcall->retval = 0; } break; } case IOCTL_PRIVCMD_MMAPBATCH: { struct ioctl_privcmd_mmapbatch *mmap; vm_map_t map; vm_map_entry_t entry; vm_object_t mem; vm_pindex_t pindex; vm_prot_t prot; boolean_t wired; struct xen_add_to_physmap_range add; xen_ulong_t *idxs; xen_pfn_t *gpfns; int *errs, index; struct privcmd_map *umap; uint16_t num; mmap = (struct ioctl_privcmd_mmapbatch *)arg; if ((mmap->num == 0) || ((mmap->addr & PAGE_MASK) != 0)) { error = EINVAL; break; } map = &td->td_proc->p_vmspace->vm_map; error = vm_map_lookup(&map, mmap->addr, VM_PROT_NONE, &entry, &mem, &pindex, &prot, &wired); if (error != KERN_SUCCESS) { error = EINVAL; break; } if ((entry->start != mmap->addr) || (entry->end != mmap->addr + (mmap->num * PAGE_SIZE))) { vm_map_lookup_done(map, entry); error = EINVAL; break; } vm_map_lookup_done(map, entry); if ((mem->type != OBJT_MGTDEVICE) || (mem->un_pager.devp.ops != &privcmd_pg_ops)) { error = EINVAL; break; } umap = mem->handle; add.domid = DOMID_SELF; add.space = XENMAPSPACE_gmfn_foreign; add.foreign_domid = mmap->dom; /* * The 'size' field in the xen_add_to_physmap_range only * allows for UINT16_MAX mappings in a single hypercall. */ num = MIN(mmap->num, UINT16_MAX); idxs = malloc(sizeof(*idxs) * num, M_PRIVCMD, M_WAITOK); gpfns = malloc(sizeof(*gpfns) * num, M_PRIVCMD, M_WAITOK); errs = malloc(sizeof(*errs) * num, M_PRIVCMD, M_WAITOK); set_xen_guest_handle(add.idxs, idxs); set_xen_guest_handle(add.gpfns, gpfns); set_xen_guest_handle(add.errs, errs); /* Allocate a bitset to store broken page mappings. */ umap->err = BITSET_ALLOC(mmap->num, M_PRIVCMD, M_WAITOK | M_ZERO); for (index = 0; index < mmap->num; index += num) { num = MIN(mmap->num - index, UINT16_MAX); add.size = num; error = copyin(&mmap->arr[index], idxs, sizeof(idxs[0]) * num); if (error != 0) goto mmap_out; for (i = 0; i < num; i++) gpfns[i] = atop(umap->phys_base_addr + (i + index) * PAGE_SIZE); bzero(errs, sizeof(*errs) * num); error = HYPERVISOR_memory_op( XENMEM_add_to_physmap_range, &add); if (error != 0) { error = xen_translate_error(error); goto mmap_out; } for (i = 0; i < num; i++) { if (errs[i] != 0) { errs[i] = xen_translate_error(errs[i]); /* Mark the page as invalid. */ BIT_SET(mmap->num, index + i, umap->err); } } error = copyout(errs, &mmap->err[index], sizeof(errs[0]) * num); if (error != 0) goto mmap_out; } umap->mapped = true; mmap_out: free(idxs, M_PRIVCMD); free(gpfns, M_PRIVCMD); free(errs, M_PRIVCMD); if (!umap->mapped) free(umap->err, M_PRIVCMD); break; } default: error = ENOSYS; break; } return (error); } /*------------------ Private Device Attachment Functions --------------------*/ static void privcmd_identify(driver_t *driver, device_t parent) { KASSERT(xen_domain(), ("Trying to attach privcmd device on non Xen domain")); if (BUS_ADD_CHILD(parent, 0, "privcmd", 0) == NULL) panic("unable to attach privcmd user-space device"); } static int privcmd_probe(device_t dev) { privcmd_dev = dev; device_set_desc(dev, "Xen privileged interface user-space device"); return (BUS_PROBE_NOWILDCARD); } static int privcmd_attach(device_t dev) { make_dev_credf(MAKEDEV_ETERNAL, &privcmd_devsw, 0, NULL, UID_ROOT, GID_WHEEL, 0600, "xen/privcmd"); return (0); } /*-------------------- Private Device Attachment Data -----------------------*/ static device_method_t privcmd_methods[] = { DEVMETHOD(device_identify, privcmd_identify), DEVMETHOD(device_probe, privcmd_probe), DEVMETHOD(device_attach, privcmd_attach), DEVMETHOD_END }; static driver_t privcmd_driver = { "privcmd", privcmd_methods, 0, }; devclass_t privcmd_devclass; DRIVER_MODULE(privcmd, xenpv, privcmd_driver, privcmd_devclass, 0, 0); MODULE_DEPEND(privcmd, xenpv, 1, 1, 1); Index: head/sys/fs/nfsclient/nfs_clbio.c =================================================================== --- head/sys/fs/nfsclient/nfs_clbio.c (revision 353538) +++ head/sys/fs/nfsclient/nfs_clbio.c (revision 353539) @@ -1,1874 +1,1874 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1989, 1993 * The Regents of the University of California. All rights reserved. * * This code is derived from software contributed to Berkeley by * Rick Macklem at The University of Guelph. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)nfs_bio.c 8.9 (Berkeley) 3/30/95 */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include extern int newnfs_directio_allow_mmap; extern struct nfsstatsv1 nfsstatsv1; extern struct mtx ncl_iod_mutex; extern int ncl_numasync; extern enum nfsiod_state ncl_iodwant[NFS_MAXASYNCDAEMON]; extern struct nfsmount *ncl_iodmount[NFS_MAXASYNCDAEMON]; extern int newnfs_directio_enable; extern int nfs_keep_dirty_on_error; uma_zone_t ncl_pbuf_zone; static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td); static int nfs_directio_write(struct vnode *vp, struct uio *uiop, struct ucred *cred, int ioflag); /* * Vnode op for VM getpages. */ SYSCTL_DECL(_vfs_nfs); static int use_buf_pager = 1; SYSCTL_INT(_vfs_nfs, OID_AUTO, use_buf_pager, CTLFLAG_RWTUN, &use_buf_pager, 0, "Use buffer pager instead of direct readrpc call"); static daddr_t ncl_gbp_getblkno(struct vnode *vp, vm_ooffset_t off) { return (off / vp->v_bufobj.bo_bsize); } static int ncl_gbp_getblksz(struct vnode *vp, daddr_t lbn) { struct nfsnode *np; u_quad_t nsize; int biosize, bcount; np = VTONFS(vp); NFSLOCKNODE(np); nsize = np->n_size; NFSUNLOCKNODE(np); biosize = vp->v_bufobj.bo_bsize; bcount = biosize; if ((off_t)lbn * biosize >= nsize) bcount = 0; else if ((off_t)(lbn + 1) * biosize > nsize) bcount = nsize - (off_t)lbn * biosize; return (bcount); } int ncl_getpages(struct vop_getpages_args *ap) { int i, error, nextoff, size, toff, count, npages; struct uio uio; struct iovec iov; vm_offset_t kva; struct buf *bp; struct vnode *vp; struct thread *td; struct ucred *cred; struct nfsmount *nmp; vm_object_t object; vm_page_t *pages; struct nfsnode *np; vp = ap->a_vp; np = VTONFS(vp); td = curthread; cred = curthread->td_ucred; nmp = VFSTONFS(vp->v_mount); pages = ap->a_m; npages = ap->a_count; if ((object = vp->v_object) == NULL) { printf("ncl_getpages: called with non-merged cache vnode\n"); return (VM_PAGER_ERROR); } if (newnfs_directio_enable && !newnfs_directio_allow_mmap) { NFSLOCKNODE(np); if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { NFSUNLOCKNODE(np); printf("ncl_getpages: called on non-cacheable vnode\n"); return (VM_PAGER_ERROR); } else NFSUNLOCKNODE(np); } mtx_lock(&nmp->nm_mtx); if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { mtx_unlock(&nmp->nm_mtx); /* We'll never get here for v4, because we always have fsinfo */ (void)ncl_fsinfo(nmp, vp, cred, td); } else mtx_unlock(&nmp->nm_mtx); if (use_buf_pager) return (vfs_bio_getpages(vp, pages, npages, ap->a_rbehind, ap->a_rahead, ncl_gbp_getblkno, ncl_gbp_getblksz)); /* * If the requested page is partially valid, just return it and * allow the pager to zero-out the blanks. Partially valid pages * can only occur at the file EOF. * * XXXGL: is that true for NFS, where short read can occur??? */ VM_OBJECT_WLOCK(object); - if (pages[npages - 1]->valid != 0 && --npages == 0) + if (!vm_page_none_valid(pages[npages - 1]) && --npages == 0) goto out; VM_OBJECT_WUNLOCK(object); /* * We use only the kva address for the buffer, but this is extremely * convenient and fast. */ bp = uma_zalloc(ncl_pbuf_zone, M_WAITOK); kva = (vm_offset_t) bp->b_data; pmap_qenter(kva, pages, npages); VM_CNT_INC(v_vnodein); VM_CNT_ADD(v_vnodepgsin, npages); count = npages << PAGE_SHIFT; iov.iov_base = (caddr_t) kva; iov.iov_len = count; uio.uio_iov = &iov; uio.uio_iovcnt = 1; uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); uio.uio_resid = count; uio.uio_segflg = UIO_SYSSPACE; uio.uio_rw = UIO_READ; uio.uio_td = td; error = ncl_readrpc(vp, &uio, cred); pmap_qremove(kva, npages); uma_zfree(ncl_pbuf_zone, bp); if (error && (uio.uio_resid == count)) { printf("ncl_getpages: error %d\n", error); return (VM_PAGER_ERROR); } /* * Calculate the number of bytes read and validate only that number * of bytes. Note that due to pending writes, size may be 0. This * does not mean that the remaining data is invalid! */ size = count - uio.uio_resid; VM_OBJECT_WLOCK(object); for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { vm_page_t m; nextoff = toff + PAGE_SIZE; m = pages[i]; if (nextoff <= size) { /* * Read operation filled an entire page */ - m->valid = VM_PAGE_BITS_ALL; + vm_page_valid(m); KASSERT(m->dirty == 0, ("nfs_getpages: page %p is dirty", m)); } else if (size > toff) { /* * Read operation filled a partial page. */ - m->valid = 0; + vm_page_invalid(m); vm_page_set_valid_range(m, 0, size - toff); KASSERT(m->dirty == 0, ("nfs_getpages: page %p is dirty", m)); } else { /* * Read operation was short. If no error * occurred we may have hit a zero-fill * section. We leave valid set to 0, and page * is freed by vm_page_readahead_finish() if * its index is not equal to requested, or * page is zeroed and set valid by * vm_pager_get_pages() for requested page. */ ; } } out: VM_OBJECT_WUNLOCK(object); if (ap->a_rbehind) *ap->a_rbehind = 0; if (ap->a_rahead) *ap->a_rahead = 0; return (VM_PAGER_OK); } /* * Vnode op for VM putpages. */ int ncl_putpages(struct vop_putpages_args *ap) { struct uio uio; struct iovec iov; int i, error, npages, count; off_t offset; int *rtvals; struct vnode *vp; struct thread *td; struct ucred *cred; struct nfsmount *nmp; struct nfsnode *np; vm_page_t *pages; vp = ap->a_vp; np = VTONFS(vp); td = curthread; /* XXX */ /* Set the cred to n_writecred for the write rpcs. */ if (np->n_writecred != NULL) cred = crhold(np->n_writecred); else cred = crhold(curthread->td_ucred); /* XXX */ nmp = VFSTONFS(vp->v_mount); pages = ap->a_m; count = ap->a_count; rtvals = ap->a_rtvals; npages = btoc(count); offset = IDX_TO_OFF(pages[0]->pindex); mtx_lock(&nmp->nm_mtx); if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { mtx_unlock(&nmp->nm_mtx); (void)ncl_fsinfo(nmp, vp, cred, td); } else mtx_unlock(&nmp->nm_mtx); NFSLOCKNODE(np); if (newnfs_directio_enable && !newnfs_directio_allow_mmap && (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { NFSUNLOCKNODE(np); printf("ncl_putpages: called on noncache-able vnode\n"); NFSLOCKNODE(np); } /* * When putting pages, do not extend file past EOF. */ if (offset + count > np->n_size) { count = np->n_size - offset; if (count < 0) count = 0; } NFSUNLOCKNODE(np); for (i = 0; i < npages; i++) rtvals[i] = VM_PAGER_ERROR; VM_CNT_INC(v_vnodeout); VM_CNT_ADD(v_vnodepgsout, count); iov.iov_base = unmapped_buf; iov.iov_len = count; uio.uio_iov = &iov; uio.uio_iovcnt = 1; uio.uio_offset = offset; uio.uio_resid = count; uio.uio_segflg = UIO_NOCOPY; uio.uio_rw = UIO_WRITE; uio.uio_td = td; error = VOP_WRITE(vp, &uio, vnode_pager_putpages_ioflags(ap->a_sync), cred); crfree(cred); if (error == 0 || !nfs_keep_dirty_on_error) { vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid, np->n_size - offset, npages * PAGE_SIZE); } return (rtvals[0]); } /* * For nfs, cache consistency can only be maintained approximately. * Although RFC1094 does not specify the criteria, the following is * believed to be compatible with the reference port. * For nfs: * If the file's modify time on the server has changed since the * last read rpc or you have written to the file, * you may have lost data cache consistency with the * server, so flush all of the file's data out of the cache. * Then force a getattr rpc to ensure that you have up to date * attributes. * NB: This implies that cache data can be read when up to * NFS_ATTRTIMEO seconds out of date. If you find that you need current * attributes this could be forced by setting n_attrstamp to 0 before * the VOP_GETATTR() call. */ static inline int nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred) { int error = 0; struct vattr vattr; struct nfsnode *np = VTONFS(vp); bool old_lock; /* * Ensure the exclusove access to the node before checking * whether the cache is consistent. */ old_lock = ncl_excl_start(vp); NFSLOCKNODE(np); if (np->n_flag & NMODIFIED) { NFSUNLOCKNODE(np); if (vp->v_type != VREG) { if (vp->v_type != VDIR) panic("nfs: bioread, not dir"); ncl_invaldir(vp); error = ncl_vinvalbuf(vp, V_SAVE | V_ALLOWCLEAN, td, 1); if (error != 0) goto out; } np->n_attrstamp = 0; KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); error = VOP_GETATTR(vp, &vattr, cred); if (error) goto out; NFSLOCKNODE(np); np->n_mtime = vattr.va_mtime; NFSUNLOCKNODE(np); } else { NFSUNLOCKNODE(np); error = VOP_GETATTR(vp, &vattr, cred); if (error) goto out; NFSLOCKNODE(np); if ((np->n_flag & NSIZECHANGED) || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) { NFSUNLOCKNODE(np); if (vp->v_type == VDIR) ncl_invaldir(vp); error = ncl_vinvalbuf(vp, V_SAVE | V_ALLOWCLEAN, td, 1); if (error != 0) goto out; NFSLOCKNODE(np); np->n_mtime = vattr.va_mtime; np->n_flag &= ~NSIZECHANGED; } NFSUNLOCKNODE(np); } out: ncl_excl_finish(vp, old_lock); return (error); } /* * Vnode op for read using bio */ int ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred) { struct nfsnode *np = VTONFS(vp); int biosize, i; struct buf *bp, *rabp; struct thread *td; struct nfsmount *nmp = VFSTONFS(vp->v_mount); daddr_t lbn, rabn; int bcount; int seqcount; int nra, error = 0, n = 0, on = 0; off_t tmp_off; KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode")); if (uio->uio_resid == 0) return (0); if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ return (EINVAL); td = uio->uio_td; mtx_lock(&nmp->nm_mtx); if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { mtx_unlock(&nmp->nm_mtx); (void)ncl_fsinfo(nmp, vp, cred, td); mtx_lock(&nmp->nm_mtx); } if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0) (void) newnfs_iosize(nmp); tmp_off = uio->uio_offset + uio->uio_resid; if (vp->v_type != VDIR && (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) { mtx_unlock(&nmp->nm_mtx); return (EFBIG); } mtx_unlock(&nmp->nm_mtx); if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG)) /* No caching/ no readaheads. Just read data into the user buffer */ return ncl_readrpc(vp, uio, cred); biosize = vp->v_bufobj.bo_bsize; seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE); error = nfs_bioread_check_cons(vp, td, cred); if (error) return error; do { u_quad_t nsize; NFSLOCKNODE(np); nsize = np->n_size; NFSUNLOCKNODE(np); switch (vp->v_type) { case VREG: NFSINCRGLOBAL(nfsstatsv1.biocache_reads); lbn = uio->uio_offset / biosize; on = uio->uio_offset - (lbn * biosize); /* * Start the read ahead(s), as required. */ if (nmp->nm_readahead > 0) { for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) { rabn = lbn + 1 + nra; if (incore(&vp->v_bufobj, rabn) == NULL) { rabp = nfs_getcacheblk(vp, rabn, biosize, td); if (!rabp) { error = newnfs_sigintr(nmp, td); return (error ? error : EINTR); } if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { rabp->b_flags |= B_ASYNC; rabp->b_iocmd = BIO_READ; vfs_busy_pages(rabp, 0); if (ncl_asyncio(nmp, rabp, cred, td)) { rabp->b_flags |= B_INVAL; rabp->b_ioflags |= BIO_ERROR; vfs_unbusy_pages(rabp); brelse(rabp); break; } } else { brelse(rabp); } } } } /* Note that bcount is *not* DEV_BSIZE aligned. */ bcount = biosize; if ((off_t)lbn * biosize >= nsize) { bcount = 0; } else if ((off_t)(lbn + 1) * biosize > nsize) { bcount = nsize - (off_t)lbn * biosize; } bp = nfs_getcacheblk(vp, lbn, bcount, td); if (!bp) { error = newnfs_sigintr(nmp, td); return (error ? error : EINTR); } /* * If B_CACHE is not set, we must issue the read. If this * fails, we return an error. */ if ((bp->b_flags & B_CACHE) == 0) { bp->b_iocmd = BIO_READ; vfs_busy_pages(bp, 0); error = ncl_doio(vp, bp, cred, td, 0); if (error) { brelse(bp); return (error); } } /* * on is the offset into the current bp. Figure out how many * bytes we can copy out of the bp. Note that bcount is * NOT DEV_BSIZE aligned. * * Then figure out how many bytes we can copy into the uio. */ n = 0; if (on < bcount) n = MIN((unsigned)(bcount - on), uio->uio_resid); break; case VLNK: NFSINCRGLOBAL(nfsstatsv1.biocache_readlinks); bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td); if (!bp) { error = newnfs_sigintr(nmp, td); return (error ? error : EINTR); } if ((bp->b_flags & B_CACHE) == 0) { bp->b_iocmd = BIO_READ; vfs_busy_pages(bp, 0); error = ncl_doio(vp, bp, cred, td, 0); if (error) { bp->b_ioflags |= BIO_ERROR; brelse(bp); return (error); } } n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); on = 0; break; case VDIR: NFSINCRGLOBAL(nfsstatsv1.biocache_readdirs); if (np->n_direofoffset && uio->uio_offset >= np->n_direofoffset) { return (0); } lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td); if (!bp) { error = newnfs_sigintr(nmp, td); return (error ? error : EINTR); } if ((bp->b_flags & B_CACHE) == 0) { bp->b_iocmd = BIO_READ; vfs_busy_pages(bp, 0); error = ncl_doio(vp, bp, cred, td, 0); if (error) { brelse(bp); } while (error == NFSERR_BAD_COOKIE) { ncl_invaldir(vp); error = ncl_vinvalbuf(vp, 0, td, 1); /* * Yuck! The directory has been modified on the * server. The only way to get the block is by * reading from the beginning to get all the * offset cookies. * * Leave the last bp intact unless there is an error. * Loop back up to the while if the error is another * NFSERR_BAD_COOKIE (double yuch!). */ for (i = 0; i <= lbn && !error; i++) { if (np->n_direofoffset && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) return (0); bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td); if (!bp) { error = newnfs_sigintr(nmp, td); return (error ? error : EINTR); } if ((bp->b_flags & B_CACHE) == 0) { bp->b_iocmd = BIO_READ; vfs_busy_pages(bp, 0); error = ncl_doio(vp, bp, cred, td, 0); /* * no error + B_INVAL == directory EOF, * use the block. */ if (error == 0 && (bp->b_flags & B_INVAL)) break; } /* * An error will throw away the block and the * for loop will break out. If no error and this * is not the block we want, we throw away the * block and go for the next one via the for loop. */ if (error || i < lbn) brelse(bp); } } /* * The above while is repeated if we hit another cookie * error. If we hit an error and it wasn't a cookie error, * we give up. */ if (error) return (error); } /* * If not eof and read aheads are enabled, start one. * (You need the current block first, so that you have the * directory offset cookie of the next block.) */ if (nmp->nm_readahead > 0 && (bp->b_flags & B_INVAL) == 0 && (np->n_direofoffset == 0 || (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && incore(&vp->v_bufobj, lbn + 1) == NULL) { rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td); if (rabp) { if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { rabp->b_flags |= B_ASYNC; rabp->b_iocmd = BIO_READ; vfs_busy_pages(rabp, 0); if (ncl_asyncio(nmp, rabp, cred, td)) { rabp->b_flags |= B_INVAL; rabp->b_ioflags |= BIO_ERROR; vfs_unbusy_pages(rabp); brelse(rabp); } } else { brelse(rabp); } } } /* * Unlike VREG files, whos buffer size ( bp->b_bcount ) is * chopped for the EOF condition, we cannot tell how large * NFS directories are going to be until we hit EOF. So * an NFS directory buffer is *not* chopped to its EOF. Now, * it just so happens that b_resid will effectively chop it * to EOF. *BUT* this information is lost if the buffer goes * away and is reconstituted into a B_CACHE state ( due to * being VMIO ) later. So we keep track of the directory eof * in np->n_direofoffset and chop it off as an extra step * right here. */ n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) n = np->n_direofoffset - uio->uio_offset; break; default: printf(" ncl_bioread: type %x unexpected\n", vp->v_type); bp = NULL; break; } if (n > 0) { error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio); } if (vp->v_type == VLNK) n = 0; if (bp != NULL) brelse(bp); } while (error == 0 && uio->uio_resid > 0 && n > 0); return (error); } /* * The NFS write path cannot handle iovecs with len > 1. So we need to * break up iovecs accordingly (restricting them to wsize). * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf). * For the ASYNC case, 2 copies are needed. The first a copy from the * user buffer to a staging buffer and then a second copy from the staging * buffer to mbufs. This can be optimized by copying from the user buffer * directly into mbufs and passing the chain down, but that requires a * fair amount of re-working of the relevant codepaths (and can be done * later). */ static int nfs_directio_write(vp, uiop, cred, ioflag) struct vnode *vp; struct uio *uiop; struct ucred *cred; int ioflag; { int error; struct nfsmount *nmp = VFSTONFS(vp->v_mount); struct thread *td = uiop->uio_td; int size; int wsize; mtx_lock(&nmp->nm_mtx); wsize = nmp->nm_wsize; mtx_unlock(&nmp->nm_mtx); if (ioflag & IO_SYNC) { int iomode, must_commit; struct uio uio; struct iovec iov; do_sync: while (uiop->uio_resid > 0) { size = MIN(uiop->uio_resid, wsize); size = MIN(uiop->uio_iov->iov_len, size); iov.iov_base = uiop->uio_iov->iov_base; iov.iov_len = size; uio.uio_iov = &iov; uio.uio_iovcnt = 1; uio.uio_offset = uiop->uio_offset; uio.uio_resid = size; uio.uio_segflg = UIO_USERSPACE; uio.uio_rw = UIO_WRITE; uio.uio_td = td; iomode = NFSWRITE_FILESYNC; error = ncl_writerpc(vp, &uio, cred, &iomode, &must_commit, 0); KASSERT((must_commit == 0), ("ncl_directio_write: Did not commit write")); if (error) return (error); uiop->uio_offset += size; uiop->uio_resid -= size; if (uiop->uio_iov->iov_len <= size) { uiop->uio_iovcnt--; uiop->uio_iov++; } else { uiop->uio_iov->iov_base = (char *)uiop->uio_iov->iov_base + size; uiop->uio_iov->iov_len -= size; } } } else { struct uio *t_uio; struct iovec *t_iov; struct buf *bp; /* * Break up the write into blocksize chunks and hand these * over to nfsiod's for write back. * Unfortunately, this incurs a copy of the data. Since * the user could modify the buffer before the write is * initiated. * * The obvious optimization here is that one of the 2 copies * in the async write path can be eliminated by copying the * data here directly into mbufs and passing the mbuf chain * down. But that will require a fair amount of re-working * of the code and can be done if there's enough interest * in NFS directio access. */ while (uiop->uio_resid > 0) { size = MIN(uiop->uio_resid, wsize); size = MIN(uiop->uio_iov->iov_len, size); bp = uma_zalloc(ncl_pbuf_zone, M_WAITOK); t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK); t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK); t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK); t_iov->iov_len = size; t_uio->uio_iov = t_iov; t_uio->uio_iovcnt = 1; t_uio->uio_offset = uiop->uio_offset; t_uio->uio_resid = size; t_uio->uio_segflg = UIO_SYSSPACE; t_uio->uio_rw = UIO_WRITE; t_uio->uio_td = td; KASSERT(uiop->uio_segflg == UIO_USERSPACE || uiop->uio_segflg == UIO_SYSSPACE, ("nfs_directio_write: Bad uio_segflg")); if (uiop->uio_segflg == UIO_USERSPACE) { error = copyin(uiop->uio_iov->iov_base, t_iov->iov_base, size); if (error != 0) goto err_free; } else /* * UIO_SYSSPACE may never happen, but handle * it just in case it does. */ bcopy(uiop->uio_iov->iov_base, t_iov->iov_base, size); bp->b_flags |= B_DIRECT; bp->b_iocmd = BIO_WRITE; if (cred != NOCRED) { crhold(cred); bp->b_wcred = cred; } else bp->b_wcred = NOCRED; bp->b_caller1 = (void *)t_uio; bp->b_vp = vp; error = ncl_asyncio(nmp, bp, NOCRED, td); err_free: if (error) { free(t_iov->iov_base, M_NFSDIRECTIO); free(t_iov, M_NFSDIRECTIO); free(t_uio, M_NFSDIRECTIO); bp->b_vp = NULL; uma_zfree(ncl_pbuf_zone, bp); if (error == EINTR) return (error); goto do_sync; } uiop->uio_offset += size; uiop->uio_resid -= size; if (uiop->uio_iov->iov_len <= size) { uiop->uio_iovcnt--; uiop->uio_iov++; } else { uiop->uio_iov->iov_base = (char *)uiop->uio_iov->iov_base + size; uiop->uio_iov->iov_len -= size; } } } return (0); } /* * Vnode op for write using bio */ int ncl_write(struct vop_write_args *ap) { int biosize; struct uio *uio = ap->a_uio; struct thread *td = uio->uio_td; struct vnode *vp = ap->a_vp; struct nfsnode *np = VTONFS(vp); struct ucred *cred = ap->a_cred; int ioflag = ap->a_ioflag; struct buf *bp; struct vattr vattr; struct nfsmount *nmp = VFSTONFS(vp->v_mount); daddr_t lbn; int bcount, noncontig_write, obcount; int bp_cached, n, on, error = 0, error1, wouldcommit; size_t orig_resid, local_resid; off_t orig_size, tmp_off; KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode")); KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread, ("ncl_write proc")); if (vp->v_type != VREG) return (EIO); NFSLOCKNODE(np); if (np->n_flag & NWRITEERR) { np->n_flag &= ~NWRITEERR; NFSUNLOCKNODE(np); return (np->n_error); } else NFSUNLOCKNODE(np); mtx_lock(&nmp->nm_mtx); if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { mtx_unlock(&nmp->nm_mtx); (void)ncl_fsinfo(nmp, vp, cred, td); mtx_lock(&nmp->nm_mtx); } if (nmp->nm_wsize == 0) (void) newnfs_iosize(nmp); mtx_unlock(&nmp->nm_mtx); /* * Synchronously flush pending buffers if we are in synchronous * mode or if we are appending. */ if (ioflag & (IO_APPEND | IO_SYNC)) { NFSLOCKNODE(np); if (np->n_flag & NMODIFIED) { NFSUNLOCKNODE(np); #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */ /* * Require non-blocking, synchronous writes to * dirty files to inform the program it needs * to fsync(2) explicitly. */ if (ioflag & IO_NDELAY) return (EAGAIN); #endif np->n_attrstamp = 0; KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag & IO_VMIO) != 0 ? V_VMIO : 0), td, 1); if (error != 0) return (error); } else NFSUNLOCKNODE(np); } orig_resid = uio->uio_resid; NFSLOCKNODE(np); orig_size = np->n_size; NFSUNLOCKNODE(np); /* * If IO_APPEND then load uio_offset. We restart here if we cannot * get the append lock. */ if (ioflag & IO_APPEND) { np->n_attrstamp = 0; KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); error = VOP_GETATTR(vp, &vattr, cred); if (error) return (error); NFSLOCKNODE(np); uio->uio_offset = np->n_size; NFSUNLOCKNODE(np); } if (uio->uio_offset < 0) return (EINVAL); tmp_off = uio->uio_offset + uio->uio_resid; if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset) return (EFBIG); if (uio->uio_resid == 0) return (0); if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG) return nfs_directio_write(vp, uio, cred, ioflag); /* * Maybe this should be above the vnode op call, but so long as * file servers have no limits, i don't think it matters */ if (vn_rlimit_fsize(vp, uio, td)) return (EFBIG); biosize = vp->v_bufobj.bo_bsize; /* * Find all of this file's B_NEEDCOMMIT buffers. If our writes * would exceed the local maximum per-file write commit size when * combined with those, we must decide whether to flush, * go synchronous, or return error. We don't bother checking * IO_UNIT -- we just make all writes atomic anyway, as there's * no point optimizing for something that really won't ever happen. */ wouldcommit = 0; if (!(ioflag & IO_SYNC)) { int nflag; NFSLOCKNODE(np); nflag = np->n_flag; NFSUNLOCKNODE(np); if (nflag & NMODIFIED) { BO_LOCK(&vp->v_bufobj); if (vp->v_bufobj.bo_dirty.bv_cnt != 0) { TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { if (bp->b_flags & B_NEEDCOMMIT) wouldcommit += bp->b_bcount; } } BO_UNLOCK(&vp->v_bufobj); } } do { if (!(ioflag & IO_SYNC)) { wouldcommit += biosize; if (wouldcommit > nmp->nm_wcommitsize) { np->n_attrstamp = 0; KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag & IO_VMIO) != 0 ? V_VMIO : 0), td, 1); if (error != 0) return (error); wouldcommit = biosize; } } NFSINCRGLOBAL(nfsstatsv1.biocache_writes); lbn = uio->uio_offset / biosize; on = uio->uio_offset - (lbn * biosize); n = MIN((unsigned)(biosize - on), uio->uio_resid); again: /* * Handle direct append and file extension cases, calculate * unaligned buffer size. */ NFSLOCKNODE(np); if ((np->n_flag & NHASBEENLOCKED) == 0 && (nmp->nm_flag & NFSMNT_NONCONTIGWR) != 0) noncontig_write = 1; else noncontig_write = 0; if ((uio->uio_offset == np->n_size || (noncontig_write != 0 && lbn == (np->n_size / biosize) && uio->uio_offset + n > np->n_size)) && n) { NFSUNLOCKNODE(np); /* * Get the buffer (in its pre-append state to maintain * B_CACHE if it was previously set). Resize the * nfsnode after we have locked the buffer to prevent * readers from reading garbage. */ obcount = np->n_size - (lbn * biosize); bp = nfs_getcacheblk(vp, lbn, obcount, td); if (bp != NULL) { long save; NFSLOCKNODE(np); np->n_size = uio->uio_offset + n; np->n_flag |= NMODIFIED; vnode_pager_setsize(vp, np->n_size); NFSUNLOCKNODE(np); save = bp->b_flags & B_CACHE; bcount = on + n; allocbuf(bp, bcount); bp->b_flags |= save; if (noncontig_write != 0 && on > obcount) vfs_bio_bzero_buf(bp, obcount, on - obcount); } } else { /* * Obtain the locked cache block first, and then * adjust the file's size as appropriate. */ bcount = on + n; if ((off_t)lbn * biosize + bcount < np->n_size) { if ((off_t)(lbn + 1) * biosize < np->n_size) bcount = biosize; else bcount = np->n_size - (off_t)lbn * biosize; } NFSUNLOCKNODE(np); bp = nfs_getcacheblk(vp, lbn, bcount, td); NFSLOCKNODE(np); if (uio->uio_offset + n > np->n_size) { np->n_size = uio->uio_offset + n; np->n_flag |= NMODIFIED; vnode_pager_setsize(vp, np->n_size); } NFSUNLOCKNODE(np); } if (!bp) { error = newnfs_sigintr(nmp, td); if (!error) error = EINTR; break; } /* * Issue a READ if B_CACHE is not set. In special-append * mode, B_CACHE is based on the buffer prior to the write * op and is typically set, avoiding the read. If a read * is required in special append mode, the server will * probably send us a short-read since we extended the file * on our end, resulting in b_resid == 0 and, thusly, * B_CACHE getting set. * * We can also avoid issuing the read if the write covers * the entire buffer. We have to make sure the buffer state * is reasonable in this case since we will not be initiating * I/O. See the comments in kern/vfs_bio.c's getblk() for * more information. * * B_CACHE may also be set due to the buffer being cached * normally. */ bp_cached = 1; if (on == 0 && n == bcount) { if ((bp->b_flags & B_CACHE) == 0) bp_cached = 0; bp->b_flags |= B_CACHE; bp->b_flags &= ~B_INVAL; bp->b_ioflags &= ~BIO_ERROR; } if ((bp->b_flags & B_CACHE) == 0) { bp->b_iocmd = BIO_READ; vfs_busy_pages(bp, 0); error = ncl_doio(vp, bp, cred, td, 0); if (error) { brelse(bp); break; } } if (bp->b_wcred == NOCRED) bp->b_wcred = crhold(cred); NFSLOCKNODE(np); np->n_flag |= NMODIFIED; NFSUNLOCKNODE(np); /* * If dirtyend exceeds file size, chop it down. This should * not normally occur but there is an append race where it * might occur XXX, so we log it. * * If the chopping creates a reverse-indexed or degenerate * situation with dirtyoff/end, we 0 both of them. */ if (bp->b_dirtyend > bcount) { printf("NFS append race @%lx:%d\n", (long)bp->b_blkno * DEV_BSIZE, bp->b_dirtyend - bcount); bp->b_dirtyend = bcount; } if (bp->b_dirtyoff >= bp->b_dirtyend) bp->b_dirtyoff = bp->b_dirtyend = 0; /* * If the new write will leave a contiguous dirty * area, just update the b_dirtyoff and b_dirtyend, * otherwise force a write rpc of the old dirty area. * * If there has been a file lock applied to this file * or vfs.nfs.old_noncontig_writing is set, do the following: * While it is possible to merge discontiguous writes due to * our having a B_CACHE buffer ( and thus valid read data * for the hole), we don't because it could lead to * significant cache coherency problems with multiple clients, * especially if locking is implemented later on. * * If vfs.nfs.old_noncontig_writing is not set and there has * not been file locking done on this file: * Relax coherency a bit for the sake of performance and * expand the current dirty region to contain the new * write even if it means we mark some non-dirty data as * dirty. */ if (noncontig_write == 0 && bp->b_dirtyend > 0 && (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { if (bwrite(bp) == EINTR) { error = EINTR; break; } goto again; } local_resid = uio->uio_resid; error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio); if (error != 0 && !bp_cached) { /* * This block has no other content then what * possibly was written by the faulty uiomove. * Release it, forgetting the data pages, to * prevent the leak of uninitialized data to * usermode. */ bp->b_ioflags |= BIO_ERROR; brelse(bp); uio->uio_offset -= local_resid - uio->uio_resid; uio->uio_resid = local_resid; break; } /* * Since this block is being modified, it must be written * again and not just committed. Since write clustering does * not work for the stage 1 data write, only the stage 2 * commit rpc, we have to clear B_CLUSTEROK as well. */ bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); /* * Get the partial update on the progress made from * uiomove, if an error occurred. */ if (error != 0) n = local_resid - uio->uio_resid; /* * Only update dirtyoff/dirtyend if not a degenerate * condition. */ if (n > 0) { if (bp->b_dirtyend > 0) { bp->b_dirtyoff = min(on, bp->b_dirtyoff); bp->b_dirtyend = max((on + n), bp->b_dirtyend); } else { bp->b_dirtyoff = on; bp->b_dirtyend = on + n; } vfs_bio_set_valid(bp, on, n); } /* * If IO_SYNC do bwrite(). * * IO_INVAL appears to be unused. The idea appears to be * to turn off caching in this case. Very odd. XXX */ if ((ioflag & IO_SYNC)) { if (ioflag & IO_INVAL) bp->b_flags |= B_NOCACHE; error1 = bwrite(bp); if (error1 != 0) { if (error == 0) error = error1; break; } } else if ((n + on) == biosize || (ioflag & IO_ASYNC) != 0) { bp->b_flags |= B_ASYNC; (void) ncl_writebp(bp, 0, NULL); } else { bdwrite(bp); } if (error != 0) break; } while (uio->uio_resid > 0 && n > 0); if (error != 0) { if (ioflag & IO_UNIT) { VATTR_NULL(&vattr); vattr.va_size = orig_size; /* IO_SYNC is handled implicitely */ (void)VOP_SETATTR(vp, &vattr, cred); uio->uio_offset -= orig_resid - uio->uio_resid; uio->uio_resid = orig_resid; } } return (error); } /* * Get an nfs cache block. * * Allocate a new one if the block isn't currently in the cache * and return the block marked busy. If the calling process is * interrupted by a signal for an interruptible mount point, return * NULL. * * The caller must carefully deal with the possible B_INVAL state of * the buffer. ncl_doio() clears B_INVAL (and ncl_asyncio() clears it * indirectly), so synchronous reads can be issued without worrying about * the B_INVAL state. We have to be a little more careful when dealing * with writes (see comments in nfs_write()) when extending a file past * its EOF. */ static struct buf * nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td) { struct buf *bp; struct mount *mp; struct nfsmount *nmp; mp = vp->v_mount; nmp = VFSTONFS(mp); if (nmp->nm_flag & NFSMNT_INT) { sigset_t oldset; newnfs_set_sigmask(td, &oldset); bp = getblk(vp, bn, size, PCATCH, 0, 0); newnfs_restore_sigmask(td, &oldset); while (bp == NULL) { if (newnfs_sigintr(nmp, td)) return (NULL); bp = getblk(vp, bn, size, 0, 2 * hz, 0); } } else { bp = getblk(vp, bn, size, 0, 0, 0); } if (vp->v_type == VREG) bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE); return (bp); } /* * Flush and invalidate all dirty buffers. If another process is already * doing the flush, just wait for completion. */ int ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg) { struct nfsnode *np = VTONFS(vp); struct nfsmount *nmp = VFSTONFS(vp->v_mount); int error = 0, slpflag, slptimeo; bool old_lock; ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf"); if ((nmp->nm_flag & NFSMNT_INT) == 0) intrflg = 0; if (NFSCL_FORCEDISM(nmp->nm_mountp)) intrflg = 1; if (intrflg) { slpflag = PCATCH; slptimeo = 2 * hz; } else { slpflag = 0; slptimeo = 0; } old_lock = ncl_excl_start(vp); if (old_lock) flags |= V_ALLOWCLEAN; /* * Now, flush as required. */ if ((flags & (V_SAVE | V_VMIO)) == V_SAVE && vp->v_bufobj.bo_object != NULL) { VM_OBJECT_WLOCK(vp->v_bufobj.bo_object); vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC); VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object); /* * If the page clean was interrupted, fail the invalidation. * Not doing so, we run the risk of losing dirty pages in the * vinvalbuf() call below. */ if (intrflg && (error = newnfs_sigintr(nmp, td))) goto out; } error = vinvalbuf(vp, flags, slpflag, 0); while (error) { if (intrflg && (error = newnfs_sigintr(nmp, td))) goto out; error = vinvalbuf(vp, flags, 0, slptimeo); } if (NFSHASPNFS(nmp)) { nfscl_layoutcommit(vp, td); /* * Invalidate the attribute cache, since writes to a DS * won't update the size attribute. */ NFSLOCKNODE(np); np->n_attrstamp = 0; } else NFSLOCKNODE(np); if (np->n_directio_asyncwr == 0) np->n_flag &= ~NMODIFIED; NFSUNLOCKNODE(np); out: ncl_excl_finish(vp, old_lock); return error; } /* * Initiate asynchronous I/O. Return an error if no nfsiods are available. * This is mainly to avoid queueing async I/O requests when the nfsiods * are all hung on a dead server. * * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp * is eventually dequeued by the async daemon, ncl_doio() *will*. */ int ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td) { int iod; int gotiod; int slpflag = 0; int slptimeo = 0; int error, error2; /* * Commits are usually short and sweet so lets save some cpu and * leave the async daemons for more important rpc's (such as reads * and writes). * * Readdirplus RPCs do vget()s to acquire the vnodes for entries * in the directory in order to update attributes. This can deadlock * with another thread that is waiting for async I/O to be done by * an nfsiod thread while holding a lock on one of these vnodes. * To avoid this deadlock, don't allow the async nfsiod threads to * perform Readdirplus RPCs. */ NFSLOCKIOD(); if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) && (nmp->nm_bufqiods > ncl_numasync / 2)) || (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) { NFSUNLOCKIOD(); return(EIO); } again: if (nmp->nm_flag & NFSMNT_INT) slpflag = PCATCH; gotiod = FALSE; /* * Find a free iod to process this request. */ for (iod = 0; iod < ncl_numasync; iod++) if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) { gotiod = TRUE; break; } /* * Try to create one if none are free. */ if (!gotiod) ncl_nfsiodnew(); else { /* * Found one, so wake it up and tell it which * mount to process. */ NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n", iod, nmp)); ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE; ncl_iodmount[iod] = nmp; nmp->nm_bufqiods++; wakeup(&ncl_iodwant[iod]); } /* * If none are free, we may already have an iod working on this mount * point. If so, it will process our request. */ if (!gotiod) { if (nmp->nm_bufqiods > 0) { NFS_DPF(ASYNCIO, ("ncl_asyncio: %d iods are already processing mount %p\n", nmp->nm_bufqiods, nmp)); gotiod = TRUE; } } /* * If we have an iod which can process the request, then queue * the buffer. */ if (gotiod) { /* * Ensure that the queue never grows too large. We still want * to asynchronize so we block rather then return EIO. */ while (nmp->nm_bufqlen >= 2*ncl_numasync) { NFS_DPF(ASYNCIO, ("ncl_asyncio: waiting for mount %p queue to drain\n", nmp)); nmp->nm_bufqwant = TRUE; error = newnfs_msleep(td, &nmp->nm_bufq, &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio", slptimeo); if (error) { error2 = newnfs_sigintr(nmp, td); if (error2) { NFSUNLOCKIOD(); return (error2); } if (slpflag == PCATCH) { slpflag = 0; slptimeo = 2 * hz; } } /* * We might have lost our iod while sleeping, * so check and loop if necessary. */ goto again; } /* We might have lost our nfsiod */ if (nmp->nm_bufqiods == 0) { NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); goto again; } if (bp->b_iocmd == BIO_READ) { if (bp->b_rcred == NOCRED && cred != NOCRED) bp->b_rcred = crhold(cred); } else { if (bp->b_wcred == NOCRED && cred != NOCRED) bp->b_wcred = crhold(cred); } if (bp->b_flags & B_REMFREE) bremfreef(bp); BUF_KERNPROC(bp); TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); nmp->nm_bufqlen++; if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { NFSLOCKNODE(VTONFS(bp->b_vp)); VTONFS(bp->b_vp)->n_flag |= NMODIFIED; VTONFS(bp->b_vp)->n_directio_asyncwr++; NFSUNLOCKNODE(VTONFS(bp->b_vp)); } NFSUNLOCKIOD(); return (0); } NFSUNLOCKIOD(); /* * All the iods are busy on other mounts, so return EIO to * force the caller to process the i/o synchronously. */ NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n")); return (EIO); } void ncl_doio_directwrite(struct buf *bp) { int iomode, must_commit; struct uio *uiop = (struct uio *)bp->b_caller1; char *iov_base = uiop->uio_iov->iov_base; iomode = NFSWRITE_FILESYNC; uiop->uio_td = NULL; /* NULL since we're in nfsiod */ ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0); KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write")); free(iov_base, M_NFSDIRECTIO); free(uiop->uio_iov, M_NFSDIRECTIO); free(uiop, M_NFSDIRECTIO); if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { struct nfsnode *np = VTONFS(bp->b_vp); NFSLOCKNODE(np); if (NFSHASPNFS(VFSTONFS(vnode_mount(bp->b_vp)))) { /* * Invalidate the attribute cache, since writes to a DS * won't update the size attribute. */ np->n_attrstamp = 0; } np->n_directio_asyncwr--; if (np->n_directio_asyncwr == 0) { np->n_flag &= ~NMODIFIED; if ((np->n_flag & NFSYNCWAIT)) { np->n_flag &= ~NFSYNCWAIT; wakeup((caddr_t)&np->n_directio_asyncwr); } } NFSUNLOCKNODE(np); } bp->b_vp = NULL; uma_zfree(ncl_pbuf_zone, bp); } /* * Do an I/O operation to/from a cache block. This may be called * synchronously or from an nfsiod. */ int ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td, int called_from_strategy) { struct uio *uiop; struct nfsnode *np; struct nfsmount *nmp; int error = 0, iomode, must_commit = 0; struct uio uio; struct iovec io; struct proc *p = td ? td->td_proc : NULL; uint8_t iocmd; np = VTONFS(vp); nmp = VFSTONFS(vp->v_mount); uiop = &uio; uiop->uio_iov = &io; uiop->uio_iovcnt = 1; uiop->uio_segflg = UIO_SYSSPACE; uiop->uio_td = td; /* * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We * do this here so we do not have to do it in all the code that * calls us. */ bp->b_flags &= ~B_INVAL; bp->b_ioflags &= ~BIO_ERROR; KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp)); iocmd = bp->b_iocmd; if (iocmd == BIO_READ) { io.iov_len = uiop->uio_resid = bp->b_bcount; io.iov_base = bp->b_data; uiop->uio_rw = UIO_READ; switch (vp->v_type) { case VREG: uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; NFSINCRGLOBAL(nfsstatsv1.read_bios); error = ncl_readrpc(vp, uiop, cr); if (!error) { if (uiop->uio_resid) { /* * If we had a short read with no error, we must have * hit a file hole. We should zero-fill the remainder. * This can also occur if the server hits the file EOF. * * Holes used to be able to occur due to pending * writes, but that is not possible any longer. */ int nread = bp->b_bcount - uiop->uio_resid; ssize_t left = uiop->uio_resid; if (left > 0) bzero((char *)bp->b_data + nread, left); uiop->uio_resid = 0; } } /* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */ if (p && vp->v_writecount <= -1) { NFSLOCKNODE(np); if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) { NFSUNLOCKNODE(np); PROC_LOCK(p); killproc(p, "text file modification"); PROC_UNLOCK(p); } else NFSUNLOCKNODE(np); } break; case VLNK: uiop->uio_offset = (off_t)0; NFSINCRGLOBAL(nfsstatsv1.readlink_bios); error = ncl_readlinkrpc(vp, uiop, cr); break; case VDIR: NFSINCRGLOBAL(nfsstatsv1.readdir_bios); uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) { error = ncl_readdirplusrpc(vp, uiop, cr, td); if (error == NFSERR_NOTSUPP) nmp->nm_flag &= ~NFSMNT_RDIRPLUS; } if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) error = ncl_readdirrpc(vp, uiop, cr, td); /* * end-of-directory sets B_INVAL but does not generate an * error. */ if (error == 0 && uiop->uio_resid == bp->b_bcount) bp->b_flags |= B_INVAL; break; default: printf("ncl_doio: type %x unexpected\n", vp->v_type); break; } if (error) { bp->b_ioflags |= BIO_ERROR; bp->b_error = error; } } else { /* * If we only need to commit, try to commit */ if (bp->b_flags & B_NEEDCOMMIT) { int retv; off_t off; off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff, bp->b_wcred, td); if (retv == 0) { bp->b_dirtyoff = bp->b_dirtyend = 0; bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); bp->b_resid = 0; bufdone(bp); return (0); } if (retv == NFSERR_STALEWRITEVERF) { ncl_clearcommit(vp->v_mount); } } /* * Setup for actual write */ NFSLOCKNODE(np); if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; NFSUNLOCKNODE(np); if (bp->b_dirtyend > bp->b_dirtyoff) { io.iov_len = uiop->uio_resid = bp->b_dirtyend - bp->b_dirtyoff; uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyoff; io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; uiop->uio_rw = UIO_WRITE; NFSINCRGLOBAL(nfsstatsv1.write_bios); if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) iomode = NFSWRITE_UNSTABLE; else iomode = NFSWRITE_FILESYNC; error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit, called_from_strategy); /* * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try * to cluster the buffers needing commit. This will allow * the system to submit a single commit rpc for the whole * cluster. We can do this even if the buffer is not 100% * dirty (relative to the NFS blocksize), so we optimize the * append-to-file-case. * * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be * cleared because write clustering only works for commit * rpc's, not for the data portion of the write). */ if (!error && iomode == NFSWRITE_UNSTABLE) { bp->b_flags |= B_NEEDCOMMIT; if (bp->b_dirtyoff == 0 && bp->b_dirtyend == bp->b_bcount) bp->b_flags |= B_CLUSTEROK; } else { bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); } /* * For an interrupted write, the buffer is still valid * and the write hasn't been pushed to the server yet, * so we can't set BIO_ERROR and report the interruption * by setting B_EINTR. For the B_ASYNC case, B_EINTR * is not relevant, so the rpc attempt is essentially * a noop. For the case of a V3 write rpc not being * committed to stable storage, the block is still * dirty and requires either a commit rpc or another * write rpc with iomode == NFSV3WRITE_FILESYNC before * the block is reused. This is indicated by setting * the B_DELWRI and B_NEEDCOMMIT flags. * * EIO is returned by ncl_writerpc() to indicate a recoverable * write error and is handled as above, except that * B_EINTR isn't set. One cause of this is a stale stateid * error for the RPC that indicates recovery is required, * when called with called_from_strategy != 0. * * If the buffer is marked B_PAGING, it does not reside on * the vp's paging queues so we cannot call bdirty(). The * bp in this case is not an NFS cache block so we should * be safe. XXX * * The logic below breaks up errors into recoverable and * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE * and keep the buffer around for potential write retries. * For the latter (eg ESTALE), we toss the buffer away (B_INVAL) * and save the error in the nfsnode. This is less than ideal * but necessary. Keeping such buffers around could potentially * cause buffer exhaustion eventually (they can never be written * out, so will get constantly be re-dirtied). It also causes * all sorts of vfs panics. For non-recoverable write errors, * also invalidate the attrcache, so we'll be forced to go over * the wire for this object, returning an error to user on next * call (most of the time). */ if (error == EINTR || error == EIO || error == ETIMEDOUT || (!error && (bp->b_flags & B_NEEDCOMMIT))) { bp->b_flags &= ~(B_INVAL|B_NOCACHE); if ((bp->b_flags & B_PAGING) == 0) { bdirty(bp); bp->b_flags &= ~B_DONE; } if ((error == EINTR || error == ETIMEDOUT) && (bp->b_flags & B_ASYNC) == 0) bp->b_flags |= B_EINTR; } else { if (error) { bp->b_ioflags |= BIO_ERROR; bp->b_flags |= B_INVAL; bp->b_error = np->n_error = error; NFSLOCKNODE(np); np->n_flag |= NWRITEERR; np->n_attrstamp = 0; KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); NFSUNLOCKNODE(np); } bp->b_dirtyoff = bp->b_dirtyend = 0; } } else { bp->b_resid = 0; bufdone(bp); return (0); } } bp->b_resid = uiop->uio_resid; if (must_commit) ncl_clearcommit(vp->v_mount); bufdone(bp); return (error); } /* * Used to aid in handling ftruncate() operations on the NFS client side. * Truncation creates a number of special problems for NFS. We have to * throw away VM pages and buffer cache buffers that are beyond EOF, and * we have to properly handle VM pages or (potentially dirty) buffers * that straddle the truncation point. */ int ncl_meta_setsize(struct vnode *vp, struct thread *td, u_quad_t nsize) { struct nfsnode *np = VTONFS(vp); u_quad_t tsize; int biosize = vp->v_bufobj.bo_bsize; int error = 0; NFSLOCKNODE(np); tsize = np->n_size; np->n_size = nsize; NFSUNLOCKNODE(np); if (nsize < tsize) { struct buf *bp; daddr_t lbn; int bufsize; /* * vtruncbuf() doesn't get the buffer overlapping the * truncation point. We may have a B_DELWRI and/or B_CACHE * buffer that now needs to be truncated. */ error = vtruncbuf(vp, nsize, biosize); lbn = nsize / biosize; bufsize = nsize - (lbn * biosize); bp = nfs_getcacheblk(vp, lbn, bufsize, td); if (!bp) return EINTR; if (bp->b_dirtyoff > bp->b_bcount) bp->b_dirtyoff = bp->b_bcount; if (bp->b_dirtyend > bp->b_bcount) bp->b_dirtyend = bp->b_bcount; bp->b_flags |= B_RELBUF; /* don't leave garbage around */ brelse(bp); } else { vnode_pager_setsize(vp, nsize); } return(error); } Index: head/sys/fs/smbfs/smbfs_io.c =================================================================== --- head/sys/fs/smbfs/smbfs_io.c (revision 353538) +++ head/sys/fs/smbfs/smbfs_io.c (revision 353539) @@ -1,676 +1,676 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2000-2001 Boris Popov * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * $FreeBSD$ * */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* #include */ #include #include #include #include #include #include /*#define SMBFS_RWGENERIC*/ extern uma_zone_t smbfs_pbuf_zone; static int smbfs_fastlookup = 1; SYSCTL_DECL(_vfs_smbfs); SYSCTL_INT(_vfs_smbfs, OID_AUTO, fastlookup, CTLFLAG_RW, &smbfs_fastlookup, 0, ""); #define DE_SIZE (sizeof(struct dirent)) static int smbfs_readvdir(struct vnode *vp, struct uio *uio, struct ucred *cred) { struct dirent de; struct componentname cn; struct smb_cred *scred; struct smbfs_fctx *ctx; struct vnode *newvp; struct smbnode *np = VTOSMB(vp); int error/*, *eofflag = ap->a_eofflag*/; long offset, limit; np = VTOSMB(vp); SMBVDEBUG("dirname='%s'\n", np->n_name); scred = smbfs_malloc_scred(); smb_makescred(scred, uio->uio_td, cred); offset = uio->uio_offset / DE_SIZE; /* offset in the directory */ limit = uio->uio_resid / DE_SIZE; if (uio->uio_resid < DE_SIZE || uio->uio_offset < 0) { error = EINVAL; goto out; } while (limit && offset < 2) { limit--; bzero((caddr_t)&de, DE_SIZE); de.d_reclen = DE_SIZE; de.d_fileno = (offset == 0) ? np->n_ino : (np->n_parent ? np->n_parentino : 2); if (de.d_fileno == 0) de.d_fileno = 0x7ffffffd + offset; de.d_namlen = offset + 1; de.d_name[0] = '.'; de.d_name[1] = '.'; de.d_type = DT_DIR; dirent_terminate(&de); error = uiomove(&de, DE_SIZE, uio); if (error) goto out; offset++; uio->uio_offset += DE_SIZE; } if (limit == 0) { error = 0; goto out; } if (offset != np->n_dirofs || np->n_dirseq == NULL) { SMBVDEBUG("Reopening search %ld:%ld\n", offset, np->n_dirofs); if (np->n_dirseq) { smbfs_findclose(np->n_dirseq, scred); np->n_dirseq = NULL; } np->n_dirofs = 2; error = smbfs_findopen(np, "*", 1, SMB_FA_SYSTEM | SMB_FA_HIDDEN | SMB_FA_DIR, scred, &ctx); if (error) { SMBVDEBUG("can not open search, error = %d", error); goto out; } np->n_dirseq = ctx; } else ctx = np->n_dirseq; while (np->n_dirofs < offset) { error = smbfs_findnext(ctx, offset - np->n_dirofs++, scred); if (error) { smbfs_findclose(np->n_dirseq, scred); np->n_dirseq = NULL; error = ENOENT ? 0 : error; goto out; } } error = 0; for (; limit; limit--, offset++) { error = smbfs_findnext(ctx, limit, scred); if (error) break; np->n_dirofs++; bzero((caddr_t)&de, DE_SIZE); de.d_reclen = DE_SIZE; de.d_fileno = ctx->f_attr.fa_ino; de.d_type = (ctx->f_attr.fa_attr & SMB_FA_DIR) ? DT_DIR : DT_REG; de.d_namlen = ctx->f_nmlen; bcopy(ctx->f_name, de.d_name, de.d_namlen); dirent_terminate(&de); if (smbfs_fastlookup) { error = smbfs_nget(vp->v_mount, vp, ctx->f_name, ctx->f_nmlen, &ctx->f_attr, &newvp); if (!error) { cn.cn_nameptr = de.d_name; cn.cn_namelen = de.d_namlen; cache_enter(vp, newvp, &cn); vput(newvp); } } error = uiomove(&de, DE_SIZE, uio); if (error) break; } if (error == ENOENT) error = 0; uio->uio_offset = offset * DE_SIZE; out: smbfs_free_scred(scred); return error; } int smbfs_readvnode(struct vnode *vp, struct uio *uiop, struct ucred *cred) { struct smbmount *smp = VFSTOSMBFS(vp->v_mount); struct smbnode *np = VTOSMB(vp); struct thread *td; struct vattr vattr; struct smb_cred *scred; int error, lks; /* * Protect against method which is not supported for now */ if (uiop->uio_segflg == UIO_NOCOPY) return EOPNOTSUPP; if (vp->v_type != VREG && vp->v_type != VDIR) { SMBFSERR("vn types other than VREG or VDIR are unsupported !\n"); return EIO; } if (uiop->uio_resid == 0) return 0; if (uiop->uio_offset < 0) return EINVAL; /* if (uiop->uio_offset + uiop->uio_resid > smp->nm_maxfilesize) return EFBIG;*/ td = uiop->uio_td; if (vp->v_type == VDIR) { lks = LK_EXCLUSIVE; /* lockstatus(vp->v_vnlock); */ if (lks == LK_SHARED) vn_lock(vp, LK_UPGRADE | LK_RETRY); error = smbfs_readvdir(vp, uiop, cred); if (lks == LK_SHARED) vn_lock(vp, LK_DOWNGRADE | LK_RETRY); return error; } /* biosize = SSTOCN(smp->sm_share)->sc_txmax;*/ if (np->n_flag & NMODIFIED) { smbfs_attr_cacheremove(vp); error = VOP_GETATTR(vp, &vattr, cred); if (error) return error; np->n_mtime.tv_sec = vattr.va_mtime.tv_sec; } else { error = VOP_GETATTR(vp, &vattr, cred); if (error) return error; if (np->n_mtime.tv_sec != vattr.va_mtime.tv_sec) { error = smbfs_vinvalbuf(vp, td); if (error) return error; np->n_mtime.tv_sec = vattr.va_mtime.tv_sec; } } scred = smbfs_malloc_scred(); smb_makescred(scred, td, cred); error = smb_read(smp->sm_share, np->n_fid, uiop, scred); smbfs_free_scred(scred); return (error); } int smbfs_writevnode(struct vnode *vp, struct uio *uiop, struct ucred *cred, int ioflag) { struct smbmount *smp = VTOSMBFS(vp); struct smbnode *np = VTOSMB(vp); struct smb_cred *scred; struct thread *td; int error = 0; if (vp->v_type != VREG) { SMBERROR("vn types other than VREG unsupported !\n"); return EIO; } SMBVDEBUG("ofs=%jd,resid=%zd\n", (intmax_t)uiop->uio_offset, uiop->uio_resid); if (uiop->uio_offset < 0) return EINVAL; /* if (uiop->uio_offset + uiop->uio_resid > smp->nm_maxfilesize) return (EFBIG);*/ td = uiop->uio_td; if (ioflag & (IO_APPEND | IO_SYNC)) { if (np->n_flag & NMODIFIED) { smbfs_attr_cacheremove(vp); error = smbfs_vinvalbuf(vp, td); if (error) return error; } if (ioflag & IO_APPEND) { #ifdef notyet /* * File size can be changed by another client */ smbfs_attr_cacheremove(vp); error = VOP_GETATTR(vp, &vattr, cred); if (error) return (error); #endif uiop->uio_offset = np->n_size; } } if (uiop->uio_resid == 0) return 0; if (vn_rlimit_fsize(vp, uiop, td)) return (EFBIG); scred = smbfs_malloc_scred(); smb_makescred(scred, td, cred); error = smb_write(smp->sm_share, np->n_fid, uiop, scred); smbfs_free_scred(scred); SMBVDEBUG("after: ofs=%jd,resid=%zd\n", (intmax_t)uiop->uio_offset, uiop->uio_resid); if (!error) { if (uiop->uio_offset > np->n_size) { np->n_size = uiop->uio_offset; vnode_pager_setsize(vp, np->n_size); } } return error; } /* * Do an I/O operation to/from a cache block. */ int smbfs_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td) { struct smbmount *smp = VFSTOSMBFS(vp->v_mount); struct smbnode *np = VTOSMB(vp); struct uio *uiop; struct iovec io; struct smb_cred *scred; int error = 0; uiop = malloc(sizeof(struct uio), M_SMBFSDATA, M_WAITOK); uiop->uio_iov = &io; uiop->uio_iovcnt = 1; uiop->uio_segflg = UIO_SYSSPACE; uiop->uio_td = td; scred = smbfs_malloc_scred(); smb_makescred(scred, td, cr); if (bp->b_iocmd == BIO_READ) { io.iov_len = uiop->uio_resid = bp->b_bcount; io.iov_base = bp->b_data; uiop->uio_rw = UIO_READ; switch (vp->v_type) { case VREG: uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; error = smb_read(smp->sm_share, np->n_fid, uiop, scred); if (error) break; if (uiop->uio_resid) { int left = uiop->uio_resid; int nread = bp->b_bcount - left; if (left > 0) bzero((char *)bp->b_data + nread, left); } break; default: printf("smbfs_doio: type %x unexpected\n",vp->v_type); break; } if (error) { bp->b_error = error; bp->b_ioflags |= BIO_ERROR; } } else { /* write */ if (((bp->b_blkno * DEV_BSIZE) + bp->b_dirtyend) > np->n_size) bp->b_dirtyend = np->n_size - (bp->b_blkno * DEV_BSIZE); if (bp->b_dirtyend > bp->b_dirtyoff) { io.iov_len = uiop->uio_resid = bp->b_dirtyend - bp->b_dirtyoff; uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; uiop->uio_rw = UIO_WRITE; error = smb_write(smp->sm_share, np->n_fid, uiop, scred); /* * For an interrupted write, the buffer is still valid * and the write hasn't been pushed to the server yet, * so we can't set BIO_ERROR and report the interruption * by setting B_EINTR. For the B_ASYNC case, B_EINTR * is not relevant, so the rpc attempt is essentially * a noop. For the case of a V3 write rpc not being * committed to stable storage, the block is still * dirty and requires either a commit rpc or another * write rpc with iomode == NFSV3WRITE_FILESYNC before * the block is reused. This is indicated by setting * the B_DELWRI and B_NEEDCOMMIT flags. */ if (error == EINTR || (!error && (bp->b_flags & B_NEEDCOMMIT))) { bp->b_flags &= ~(B_INVAL|B_NOCACHE); if ((bp->b_flags & B_ASYNC) == 0) bp->b_flags |= B_EINTR; if ((bp->b_flags & B_PAGING) == 0) { bdirty(bp); bp->b_flags &= ~B_DONE; } if ((bp->b_flags & B_ASYNC) == 0) bp->b_flags |= B_EINTR; } else { if (error) { bp->b_ioflags |= BIO_ERROR; bp->b_error = error; } bp->b_dirtyoff = bp->b_dirtyend = 0; } } else { bp->b_resid = 0; bufdone(bp); free(uiop, M_SMBFSDATA); smbfs_free_scred(scred); return 0; } } bp->b_resid = uiop->uio_resid; bufdone(bp); free(uiop, M_SMBFSDATA); smbfs_free_scred(scred); return error; } /* * Vnode op for VM getpages. * Wish wish .... get rid from multiple IO routines */ int smbfs_getpages(ap) struct vop_getpages_args /* { struct vnode *a_vp; vm_page_t *a_m; int a_count; int a_reqpage; } */ *ap; { #ifdef SMBFS_RWGENERIC return vop_stdgetpages(ap); #else int i, error, nextoff, size, toff, npages, count; struct uio uio; struct iovec iov; vm_offset_t kva; struct buf *bp; struct vnode *vp; struct thread *td; struct ucred *cred; struct smbmount *smp; struct smbnode *np; struct smb_cred *scred; vm_object_t object; vm_page_t *pages; vp = ap->a_vp; if ((object = vp->v_object) == NULL) { printf("smbfs_getpages: called with non-merged cache vnode??\n"); return VM_PAGER_ERROR; } td = curthread; /* XXX */ cred = td->td_ucred; /* XXX */ np = VTOSMB(vp); smp = VFSTOSMBFS(vp->v_mount); pages = ap->a_m; npages = ap->a_count; /* * If the requested page is partially valid, just return it and * allow the pager to zero-out the blanks. Partially valid pages * can only occur at the file EOF. * * XXXGL: is that true for SMB filesystem? */ VM_OBJECT_WLOCK(object); - if (pages[npages - 1]->valid != 0 && --npages == 0) + if (!vm_page_none_valid(pages[npages - 1]) && --npages == 0) goto out; VM_OBJECT_WUNLOCK(object); scred = smbfs_malloc_scred(); smb_makescred(scred, td, cred); bp = uma_zalloc(smbfs_pbuf_zone, M_WAITOK); kva = (vm_offset_t) bp->b_data; pmap_qenter(kva, pages, npages); VM_CNT_INC(v_vnodein); VM_CNT_ADD(v_vnodepgsin, npages); count = npages << PAGE_SHIFT; iov.iov_base = (caddr_t) kva; iov.iov_len = count; uio.uio_iov = &iov; uio.uio_iovcnt = 1; uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); uio.uio_resid = count; uio.uio_segflg = UIO_SYSSPACE; uio.uio_rw = UIO_READ; uio.uio_td = td; error = smb_read(smp->sm_share, np->n_fid, &uio, scred); smbfs_free_scred(scred); pmap_qremove(kva, npages); uma_zfree(smbfs_pbuf_zone, bp); if (error && (uio.uio_resid == count)) { printf("smbfs_getpages: error %d\n",error); return VM_PAGER_ERROR; } size = count - uio.uio_resid; VM_OBJECT_WLOCK(object); for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { vm_page_t m; nextoff = toff + PAGE_SIZE; m = pages[i]; if (nextoff <= size) { /* * Read operation filled an entire page */ - m->valid = VM_PAGE_BITS_ALL; + vm_page_valid(m); KASSERT(m->dirty == 0, ("smbfs_getpages: page %p is dirty", m)); } else if (size > toff) { /* * Read operation filled a partial page. */ - m->valid = 0; + vm_page_invalid(m); vm_page_set_valid_range(m, 0, size - toff); KASSERT(m->dirty == 0, ("smbfs_getpages: page %p is dirty", m)); } else { /* * Read operation was short. If no error occurred * we may have hit a zero-fill section. We simply * leave valid set to 0. */ ; } } out: VM_OBJECT_WUNLOCK(object); if (ap->a_rbehind) *ap->a_rbehind = 0; if (ap->a_rahead) *ap->a_rahead = 0; return (VM_PAGER_OK); #endif /* SMBFS_RWGENERIC */ } /* * Vnode op for VM putpages. * possible bug: all IO done in sync mode * Note that vop_close always invalidate pages before close, so it's * not necessary to open vnode. */ int smbfs_putpages(ap) struct vop_putpages_args /* { struct vnode *a_vp; vm_page_t *a_m; int a_count; int a_sync; int *a_rtvals; } */ *ap; { int error; struct vnode *vp = ap->a_vp; struct thread *td; struct ucred *cred; #ifdef SMBFS_RWGENERIC td = curthread; /* XXX */ cred = td->td_ucred; /* XXX */ VOP_OPEN(vp, FWRITE, cred, td, NULL); error = vop_stdputpages(ap); VOP_CLOSE(vp, FWRITE, cred, td); return error; #else struct uio uio; struct iovec iov; vm_offset_t kva; struct buf *bp; int i, npages, count; int *rtvals; struct smbmount *smp; struct smbnode *np; struct smb_cred *scred; vm_page_t *pages; td = curthread; /* XXX */ cred = td->td_ucred; /* XXX */ /* VOP_OPEN(vp, FWRITE, cred, td, NULL);*/ np = VTOSMB(vp); smp = VFSTOSMBFS(vp->v_mount); pages = ap->a_m; count = ap->a_count; rtvals = ap->a_rtvals; npages = btoc(count); for (i = 0; i < npages; i++) { rtvals[i] = VM_PAGER_ERROR; } bp = uma_zalloc(smbfs_pbuf_zone, M_WAITOK); kva = (vm_offset_t) bp->b_data; pmap_qenter(kva, pages, npages); VM_CNT_INC(v_vnodeout); VM_CNT_ADD(v_vnodepgsout, count); iov.iov_base = (caddr_t) kva; iov.iov_len = count; uio.uio_iov = &iov; uio.uio_iovcnt = 1; uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); uio.uio_resid = count; uio.uio_segflg = UIO_SYSSPACE; uio.uio_rw = UIO_WRITE; uio.uio_td = td; SMBVDEBUG("ofs=%jd,resid=%zd\n", (intmax_t)uio.uio_offset, uio.uio_resid); scred = smbfs_malloc_scred(); smb_makescred(scred, td, cred); error = smb_write(smp->sm_share, np->n_fid, &uio, scred); smbfs_free_scred(scred); /* VOP_CLOSE(vp, FWRITE, cred, td);*/ SMBVDEBUG("paged write done: %d\n", error); pmap_qremove(kva, npages); uma_zfree(smbfs_pbuf_zone, bp); if (error == 0) { vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid, npages * PAGE_SIZE, npages * PAGE_SIZE); } return (rtvals[0]); #endif /* SMBFS_RWGENERIC */ } /* * Flush and invalidate all dirty buffers. If another process is already * doing the flush, just wait for completion. */ int smbfs_vinvalbuf(struct vnode *vp, struct thread *td) { struct smbnode *np = VTOSMB(vp); int error = 0; if (vp->v_iflag & VI_DOOMED) return 0; while (np->n_flag & NFLUSHINPROG) { np->n_flag |= NFLUSHWANT; error = tsleep(&np->n_flag, PRIBIO + 2, "smfsvinv", 2 * hz); error = smb_td_intr(td); if (error == EINTR) return EINTR; } np->n_flag |= NFLUSHINPROG; if (vp->v_bufobj.bo_object != NULL) { VM_OBJECT_WLOCK(vp->v_bufobj.bo_object); vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC); VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object); } error = vinvalbuf(vp, V_SAVE, PCATCH, 0); while (error) { if (error == ERESTART || error == EINTR) { np->n_flag &= ~NFLUSHINPROG; if (np->n_flag & NFLUSHWANT) { np->n_flag &= ~NFLUSHWANT; wakeup(&np->n_flag); } return EINTR; } error = vinvalbuf(vp, V_SAVE, PCATCH, 0); } np->n_flag &= ~(NMODIFIED | NFLUSHINPROG); if (np->n_flag & NFLUSHWANT) { np->n_flag &= ~NFLUSHWANT; wakeup(&np->n_flag); } return (error); } Index: head/sys/fs/tmpfs/tmpfs_subr.c =================================================================== --- head/sys/fs/tmpfs/tmpfs_subr.c (revision 353538) +++ head/sys/fs/tmpfs/tmpfs_subr.c (revision 353539) @@ -1,1873 +1,1873 @@ /* $NetBSD: tmpfs_subr.c,v 1.35 2007/07/09 21:10:50 ad Exp $ */ /*- * SPDX-License-Identifier: BSD-2-Clause-NetBSD * * Copyright (c) 2005 The NetBSD Foundation, Inc. * All rights reserved. * * This code is derived from software contributed to The NetBSD Foundation * by Julio M. Merino Vidal, developed as part of Google's Summer of Code * 2005 program. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE * POSSIBILITY OF SUCH DAMAGE. */ /* * Efficient memory file system supporting functions. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include SYSCTL_NODE(_vfs, OID_AUTO, tmpfs, CTLFLAG_RW, 0, "tmpfs file system"); static long tmpfs_pages_reserved = TMPFS_PAGES_MINRESERVED; static int sysctl_mem_reserved(SYSCTL_HANDLER_ARGS) { int error; long pages, bytes; pages = *(long *)arg1; bytes = pages * PAGE_SIZE; error = sysctl_handle_long(oidp, &bytes, 0, req); if (error || !req->newptr) return (error); pages = bytes / PAGE_SIZE; if (pages < TMPFS_PAGES_MINRESERVED) return (EINVAL); *(long *)arg1 = pages; return (0); } SYSCTL_PROC(_vfs_tmpfs, OID_AUTO, memory_reserved, CTLTYPE_LONG|CTLFLAG_RW, &tmpfs_pages_reserved, 0, sysctl_mem_reserved, "L", "Amount of available memory and swap below which tmpfs growth stops"); static __inline int tmpfs_dirtree_cmp(struct tmpfs_dirent *a, struct tmpfs_dirent *b); RB_PROTOTYPE_STATIC(tmpfs_dir, tmpfs_dirent, uh.td_entries, tmpfs_dirtree_cmp); size_t tmpfs_mem_avail(void) { vm_ooffset_t avail; avail = swap_pager_avail + vm_free_count() - tmpfs_pages_reserved; if (__predict_false(avail < 0)) avail = 0; return (avail); } size_t tmpfs_pages_used(struct tmpfs_mount *tmp) { const size_t node_size = sizeof(struct tmpfs_node) + sizeof(struct tmpfs_dirent); size_t meta_pages; meta_pages = howmany((uintmax_t)tmp->tm_nodes_inuse * node_size, PAGE_SIZE); return (meta_pages + tmp->tm_pages_used); } static size_t tmpfs_pages_check_avail(struct tmpfs_mount *tmp, size_t req_pages) { if (tmpfs_mem_avail() < req_pages) return (0); if (tmp->tm_pages_max != ULONG_MAX && tmp->tm_pages_max < req_pages + tmpfs_pages_used(tmp)) return (0); return (1); } void tmpfs_ref_node(struct tmpfs_node *node) { TMPFS_NODE_LOCK(node); tmpfs_ref_node_locked(node); TMPFS_NODE_UNLOCK(node); } void tmpfs_ref_node_locked(struct tmpfs_node *node) { TMPFS_NODE_ASSERT_LOCKED(node); KASSERT(node->tn_refcount > 0, ("node %p zero refcount", node)); KASSERT(node->tn_refcount < UINT_MAX, ("node %p refcount %u", node, node->tn_refcount)); node->tn_refcount++; } /* * Allocates a new node of type 'type' inside the 'tmp' mount point, with * its owner set to 'uid', its group to 'gid' and its mode set to 'mode', * using the credentials of the process 'p'. * * If the node type is set to 'VDIR', then the parent parameter must point * to the parent directory of the node being created. It may only be NULL * while allocating the root node. * * If the node type is set to 'VBLK' or 'VCHR', then the rdev parameter * specifies the device the node represents. * * If the node type is set to 'VLNK', then the parameter target specifies * the file name of the target file for the symbolic link that is being * created. * * Note that new nodes are retrieved from the available list if it has * items or, if it is empty, from the node pool as long as there is enough * space to create them. * * Returns zero on success or an appropriate error code on failure. */ int tmpfs_alloc_node(struct mount *mp, struct tmpfs_mount *tmp, enum vtype type, uid_t uid, gid_t gid, mode_t mode, struct tmpfs_node *parent, const char *target, dev_t rdev, struct tmpfs_node **node) { struct tmpfs_node *nnode; vm_object_t obj; /* If the root directory of the 'tmp' file system is not yet * allocated, this must be the request to do it. */ MPASS(IMPLIES(tmp->tm_root == NULL, parent == NULL && type == VDIR)); MPASS(IFF(type == VLNK, target != NULL)); MPASS(IFF(type == VBLK || type == VCHR, rdev != VNOVAL)); if (tmp->tm_nodes_inuse >= tmp->tm_nodes_max) return (ENOSPC); if (tmpfs_pages_check_avail(tmp, 1) == 0) return (ENOSPC); if ((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) { /* * When a new tmpfs node is created for fully * constructed mount point, there must be a parent * node, which vnode is locked exclusively. As * consequence, if the unmount is executing in * parallel, vflush() cannot reclaim the parent vnode. * Due to this, the check for MNTK_UNMOUNT flag is not * racy: if we did not see MNTK_UNMOUNT flag, then tmp * cannot be destroyed until node construction is * finished and the parent vnode unlocked. * * Tmpfs does not need to instantiate new nodes during * unmount. */ return (EBUSY); } if ((mp->mnt_kern_flag & MNT_RDONLY) != 0) return (EROFS); nnode = (struct tmpfs_node *)uma_zalloc_arg(tmp->tm_node_pool, tmp, M_WAITOK); /* Generic initialization. */ nnode->tn_type = type; vfs_timestamp(&nnode->tn_atime); nnode->tn_birthtime = nnode->tn_ctime = nnode->tn_mtime = nnode->tn_atime; nnode->tn_uid = uid; nnode->tn_gid = gid; nnode->tn_mode = mode; nnode->tn_id = alloc_unr64(&tmp->tm_ino_unr); nnode->tn_refcount = 1; /* Type-specific initialization. */ switch (nnode->tn_type) { case VBLK: case VCHR: nnode->tn_rdev = rdev; break; case VDIR: RB_INIT(&nnode->tn_dir.tn_dirhead); LIST_INIT(&nnode->tn_dir.tn_dupindex); MPASS(parent != nnode); MPASS(IMPLIES(parent == NULL, tmp->tm_root == NULL)); nnode->tn_dir.tn_parent = (parent == NULL) ? nnode : parent; nnode->tn_dir.tn_readdir_lastn = 0; nnode->tn_dir.tn_readdir_lastp = NULL; nnode->tn_links++; TMPFS_NODE_LOCK(nnode->tn_dir.tn_parent); nnode->tn_dir.tn_parent->tn_links++; TMPFS_NODE_UNLOCK(nnode->tn_dir.tn_parent); break; case VFIFO: /* FALLTHROUGH */ case VSOCK: break; case VLNK: MPASS(strlen(target) < MAXPATHLEN); nnode->tn_size = strlen(target); nnode->tn_link = malloc(nnode->tn_size, M_TMPFSNAME, M_WAITOK); memcpy(nnode->tn_link, target, nnode->tn_size); break; case VREG: obj = nnode->tn_reg.tn_aobj = vm_pager_allocate(OBJT_SWAP, NULL, 0, VM_PROT_DEFAULT, 0, NULL /* XXXKIB - tmpfs needs swap reservation */); VM_OBJECT_WLOCK(obj); /* OBJ_TMPFS is set together with the setting of vp->v_object */ vm_object_set_flag(obj, OBJ_NOSPLIT | OBJ_TMPFS_NODE); vm_object_clear_flag(obj, OBJ_ONEMAPPING); VM_OBJECT_WUNLOCK(obj); break; default: panic("tmpfs_alloc_node: type %p %d", nnode, (int)nnode->tn_type); } TMPFS_LOCK(tmp); LIST_INSERT_HEAD(&tmp->tm_nodes_used, nnode, tn_entries); nnode->tn_attached = true; tmp->tm_nodes_inuse++; tmp->tm_refcount++; TMPFS_UNLOCK(tmp); *node = nnode; return (0); } /* * Destroys the node pointed to by node from the file system 'tmp'. * If the node references a directory, no entries are allowed. */ void tmpfs_free_node(struct tmpfs_mount *tmp, struct tmpfs_node *node) { TMPFS_LOCK(tmp); TMPFS_NODE_LOCK(node); if (!tmpfs_free_node_locked(tmp, node, false)) { TMPFS_NODE_UNLOCK(node); TMPFS_UNLOCK(tmp); } } bool tmpfs_free_node_locked(struct tmpfs_mount *tmp, struct tmpfs_node *node, bool detach) { vm_object_t uobj; TMPFS_MP_ASSERT_LOCKED(tmp); TMPFS_NODE_ASSERT_LOCKED(node); KASSERT(node->tn_refcount > 0, ("node %p refcount zero", node)); node->tn_refcount--; if (node->tn_attached && (detach || node->tn_refcount == 0)) { MPASS(tmp->tm_nodes_inuse > 0); tmp->tm_nodes_inuse--; LIST_REMOVE(node, tn_entries); node->tn_attached = false; } if (node->tn_refcount > 0) return (false); #ifdef INVARIANTS MPASS(node->tn_vnode == NULL); MPASS((node->tn_vpstate & TMPFS_VNODE_ALLOCATING) == 0); #endif TMPFS_NODE_UNLOCK(node); TMPFS_UNLOCK(tmp); switch (node->tn_type) { case VBLK: /* FALLTHROUGH */ case VCHR: /* FALLTHROUGH */ case VDIR: /* FALLTHROUGH */ case VFIFO: /* FALLTHROUGH */ case VSOCK: break; case VLNK: free(node->tn_link, M_TMPFSNAME); break; case VREG: uobj = node->tn_reg.tn_aobj; if (uobj != NULL) { if (uobj->size != 0) atomic_subtract_long(&tmp->tm_pages_used, uobj->size); KASSERT((uobj->flags & OBJ_TMPFS) == 0, ("leaked OBJ_TMPFS node %p vm_obj %p", node, uobj)); vm_object_deallocate(uobj); } break; default: panic("tmpfs_free_node: type %p %d", node, (int)node->tn_type); } uma_zfree(tmp->tm_node_pool, node); TMPFS_LOCK(tmp); tmpfs_free_tmp(tmp); return (true); } static __inline uint32_t tmpfs_dirent_hash(const char *name, u_int len) { uint32_t hash; hash = fnv_32_buf(name, len, FNV1_32_INIT + len) & TMPFS_DIRCOOKIE_MASK; #ifdef TMPFS_DEBUG_DIRCOOKIE_DUP hash &= 0xf; #endif if (hash < TMPFS_DIRCOOKIE_MIN) hash += TMPFS_DIRCOOKIE_MIN; return (hash); } static __inline off_t tmpfs_dirent_cookie(struct tmpfs_dirent *de) { if (de == NULL) return (TMPFS_DIRCOOKIE_EOF); MPASS(de->td_cookie >= TMPFS_DIRCOOKIE_MIN); return (de->td_cookie); } static __inline boolean_t tmpfs_dirent_dup(struct tmpfs_dirent *de) { return ((de->td_cookie & TMPFS_DIRCOOKIE_DUP) != 0); } static __inline boolean_t tmpfs_dirent_duphead(struct tmpfs_dirent *de) { return ((de->td_cookie & TMPFS_DIRCOOKIE_DUPHEAD) != 0); } void tmpfs_dirent_init(struct tmpfs_dirent *de, const char *name, u_int namelen) { de->td_hash = de->td_cookie = tmpfs_dirent_hash(name, namelen); memcpy(de->ud.td_name, name, namelen); de->td_namelen = namelen; } /* * Allocates a new directory entry for the node node with a name of name. * The new directory entry is returned in *de. * * The link count of node is increased by one to reflect the new object * referencing it. * * Returns zero on success or an appropriate error code on failure. */ int tmpfs_alloc_dirent(struct tmpfs_mount *tmp, struct tmpfs_node *node, const char *name, u_int len, struct tmpfs_dirent **de) { struct tmpfs_dirent *nde; nde = uma_zalloc(tmp->tm_dirent_pool, M_WAITOK); nde->td_node = node; if (name != NULL) { nde->ud.td_name = malloc(len, M_TMPFSNAME, M_WAITOK); tmpfs_dirent_init(nde, name, len); } else nde->td_namelen = 0; if (node != NULL) node->tn_links++; *de = nde; return 0; } /* * Frees a directory entry. It is the caller's responsibility to destroy * the node referenced by it if needed. * * The link count of node is decreased by one to reflect the removal of an * object that referenced it. This only happens if 'node_exists' is true; * otherwise the function will not access the node referred to by the * directory entry, as it may already have been released from the outside. */ void tmpfs_free_dirent(struct tmpfs_mount *tmp, struct tmpfs_dirent *de) { struct tmpfs_node *node; node = de->td_node; if (node != NULL) { MPASS(node->tn_links > 0); node->tn_links--; } if (!tmpfs_dirent_duphead(de) && de->ud.td_name != NULL) free(de->ud.td_name, M_TMPFSNAME); uma_zfree(tmp->tm_dirent_pool, de); } void tmpfs_destroy_vobject(struct vnode *vp, vm_object_t obj) { ASSERT_VOP_ELOCKED(vp, "tmpfs_destroy_vobject"); if (vp->v_type != VREG || obj == NULL) return; VM_OBJECT_WLOCK(obj); VI_LOCK(vp); vm_object_clear_flag(obj, OBJ_TMPFS); obj->un_pager.swp.swp_tmpfs = NULL; if (vp->v_writecount < 0) vp->v_writecount = 0; VI_UNLOCK(vp); VM_OBJECT_WUNLOCK(obj); } /* * Need to clear v_object for insmntque failure. */ static void tmpfs_insmntque_dtr(struct vnode *vp, void *dtr_arg) { tmpfs_destroy_vobject(vp, vp->v_object); vp->v_object = NULL; vp->v_data = NULL; vp->v_op = &dead_vnodeops; vgone(vp); vput(vp); } /* * Allocates a new vnode for the node node or returns a new reference to * an existing one if the node had already a vnode referencing it. The * resulting locked vnode is returned in *vpp. * * Returns zero on success or an appropriate error code on failure. */ int tmpfs_alloc_vp(struct mount *mp, struct tmpfs_node *node, int lkflag, struct vnode **vpp) { struct vnode *vp; struct tmpfs_mount *tm; vm_object_t object; int error; error = 0; tm = VFS_TO_TMPFS(mp); TMPFS_NODE_LOCK(node); tmpfs_ref_node_locked(node); loop: TMPFS_NODE_ASSERT_LOCKED(node); if ((vp = node->tn_vnode) != NULL) { MPASS((node->tn_vpstate & TMPFS_VNODE_DOOMED) == 0); VI_LOCK(vp); if ((node->tn_type == VDIR && node->tn_dir.tn_parent == NULL) || ((vp->v_iflag & VI_DOOMED) != 0 && (lkflag & LK_NOWAIT) != 0)) { VI_UNLOCK(vp); TMPFS_NODE_UNLOCK(node); error = ENOENT; vp = NULL; goto out; } if ((vp->v_iflag & VI_DOOMED) != 0) { VI_UNLOCK(vp); node->tn_vpstate |= TMPFS_VNODE_WRECLAIM; while ((node->tn_vpstate & TMPFS_VNODE_WRECLAIM) != 0) { msleep(&node->tn_vnode, TMPFS_NODE_MTX(node), 0, "tmpfsE", 0); } goto loop; } TMPFS_NODE_UNLOCK(node); error = vget(vp, lkflag | LK_INTERLOCK, curthread); if (error == ENOENT) { TMPFS_NODE_LOCK(node); goto loop; } if (error != 0) { vp = NULL; goto out; } /* * Make sure the vnode is still there after * getting the interlock to avoid racing a free. */ if (node->tn_vnode == NULL || node->tn_vnode != vp) { vput(vp); TMPFS_NODE_LOCK(node); goto loop; } goto out; } if ((node->tn_vpstate & TMPFS_VNODE_DOOMED) || (node->tn_type == VDIR && node->tn_dir.tn_parent == NULL)) { TMPFS_NODE_UNLOCK(node); error = ENOENT; vp = NULL; goto out; } /* * otherwise lock the vp list while we call getnewvnode * since that can block. */ if (node->tn_vpstate & TMPFS_VNODE_ALLOCATING) { node->tn_vpstate |= TMPFS_VNODE_WANT; error = msleep((caddr_t) &node->tn_vpstate, TMPFS_NODE_MTX(node), 0, "tmpfs_alloc_vp", 0); if (error != 0) goto out; goto loop; } else node->tn_vpstate |= TMPFS_VNODE_ALLOCATING; TMPFS_NODE_UNLOCK(node); /* Get a new vnode and associate it with our node. */ error = getnewvnode("tmpfs", mp, VFS_TO_TMPFS(mp)->tm_nonc ? &tmpfs_vnodeop_nonc_entries : &tmpfs_vnodeop_entries, &vp); if (error != 0) goto unlock; MPASS(vp != NULL); /* lkflag is ignored, the lock is exclusive */ (void) vn_lock(vp, lkflag | LK_RETRY); vp->v_data = node; vp->v_type = node->tn_type; /* Type-specific initialization. */ switch (node->tn_type) { case VBLK: /* FALLTHROUGH */ case VCHR: /* FALLTHROUGH */ case VLNK: /* FALLTHROUGH */ case VSOCK: break; case VFIFO: vp->v_op = &tmpfs_fifoop_entries; break; case VREG: object = node->tn_reg.tn_aobj; VM_OBJECT_WLOCK(object); VI_LOCK(vp); KASSERT(vp->v_object == NULL, ("Not NULL v_object in tmpfs")); vp->v_object = object; object->un_pager.swp.swp_tmpfs = vp; vm_object_set_flag(object, OBJ_TMPFS); VI_UNLOCK(vp); VM_OBJECT_WUNLOCK(object); break; case VDIR: MPASS(node->tn_dir.tn_parent != NULL); if (node->tn_dir.tn_parent == node) vp->v_vflag |= VV_ROOT; break; default: panic("tmpfs_alloc_vp: type %p %d", node, (int)node->tn_type); } if (vp->v_type != VFIFO) VN_LOCK_ASHARE(vp); error = insmntque1(vp, mp, tmpfs_insmntque_dtr, NULL); if (error != 0) vp = NULL; unlock: TMPFS_NODE_LOCK(node); MPASS(node->tn_vpstate & TMPFS_VNODE_ALLOCATING); node->tn_vpstate &= ~TMPFS_VNODE_ALLOCATING; node->tn_vnode = vp; if (node->tn_vpstate & TMPFS_VNODE_WANT) { node->tn_vpstate &= ~TMPFS_VNODE_WANT; TMPFS_NODE_UNLOCK(node); wakeup((caddr_t) &node->tn_vpstate); } else TMPFS_NODE_UNLOCK(node); out: if (error == 0) { *vpp = vp; #ifdef INVARIANTS MPASS(*vpp != NULL && VOP_ISLOCKED(*vpp)); TMPFS_NODE_LOCK(node); MPASS(*vpp == node->tn_vnode); TMPFS_NODE_UNLOCK(node); #endif } tmpfs_free_node(tm, node); return (error); } /* * Destroys the association between the vnode vp and the node it * references. */ void tmpfs_free_vp(struct vnode *vp) { struct tmpfs_node *node; node = VP_TO_TMPFS_NODE(vp); TMPFS_NODE_ASSERT_LOCKED(node); node->tn_vnode = NULL; if ((node->tn_vpstate & TMPFS_VNODE_WRECLAIM) != 0) wakeup(&node->tn_vnode); node->tn_vpstate &= ~TMPFS_VNODE_WRECLAIM; vp->v_data = NULL; } /* * Allocates a new file of type 'type' and adds it to the parent directory * 'dvp'; this addition is done using the component name given in 'cnp'. * The ownership of the new file is automatically assigned based on the * credentials of the caller (through 'cnp'), the group is set based on * the parent directory and the mode is determined from the 'vap' argument. * If successful, *vpp holds a vnode to the newly created file and zero * is returned. Otherwise *vpp is NULL and the function returns an * appropriate error code. */ int tmpfs_alloc_file(struct vnode *dvp, struct vnode **vpp, struct vattr *vap, struct componentname *cnp, const char *target) { int error; struct tmpfs_dirent *de; struct tmpfs_mount *tmp; struct tmpfs_node *dnode; struct tmpfs_node *node; struct tmpfs_node *parent; ASSERT_VOP_ELOCKED(dvp, "tmpfs_alloc_file"); MPASS(cnp->cn_flags & HASBUF); tmp = VFS_TO_TMPFS(dvp->v_mount); dnode = VP_TO_TMPFS_DIR(dvp); *vpp = NULL; /* If the entry we are creating is a directory, we cannot overflow * the number of links of its parent, because it will get a new * link. */ if (vap->va_type == VDIR) { /* Ensure that we do not overflow the maximum number of links * imposed by the system. */ MPASS(dnode->tn_links <= TMPFS_LINK_MAX); if (dnode->tn_links == TMPFS_LINK_MAX) { return (EMLINK); } parent = dnode; MPASS(parent != NULL); } else parent = NULL; /* Allocate a node that represents the new file. */ error = tmpfs_alloc_node(dvp->v_mount, tmp, vap->va_type, cnp->cn_cred->cr_uid, dnode->tn_gid, vap->va_mode, parent, target, vap->va_rdev, &node); if (error != 0) return (error); /* Allocate a directory entry that points to the new file. */ error = tmpfs_alloc_dirent(tmp, node, cnp->cn_nameptr, cnp->cn_namelen, &de); if (error != 0) { tmpfs_free_node(tmp, node); return (error); } /* Allocate a vnode for the new file. */ error = tmpfs_alloc_vp(dvp->v_mount, node, LK_EXCLUSIVE, vpp); if (error != 0) { tmpfs_free_dirent(tmp, de); tmpfs_free_node(tmp, node); return (error); } /* Now that all required items are allocated, we can proceed to * insert the new node into the directory, an operation that * cannot fail. */ if (cnp->cn_flags & ISWHITEOUT) tmpfs_dir_whiteout_remove(dvp, cnp); tmpfs_dir_attach(dvp, de); return (0); } struct tmpfs_dirent * tmpfs_dir_first(struct tmpfs_node *dnode, struct tmpfs_dir_cursor *dc) { struct tmpfs_dirent *de; de = RB_MIN(tmpfs_dir, &dnode->tn_dir.tn_dirhead); dc->tdc_tree = de; if (de != NULL && tmpfs_dirent_duphead(de)) de = LIST_FIRST(&de->ud.td_duphead); dc->tdc_current = de; return (dc->tdc_current); } struct tmpfs_dirent * tmpfs_dir_next(struct tmpfs_node *dnode, struct tmpfs_dir_cursor *dc) { struct tmpfs_dirent *de; MPASS(dc->tdc_tree != NULL); if (tmpfs_dirent_dup(dc->tdc_current)) { dc->tdc_current = LIST_NEXT(dc->tdc_current, uh.td_dup.entries); if (dc->tdc_current != NULL) return (dc->tdc_current); } dc->tdc_tree = dc->tdc_current = RB_NEXT(tmpfs_dir, &dnode->tn_dir.tn_dirhead, dc->tdc_tree); if ((de = dc->tdc_current) != NULL && tmpfs_dirent_duphead(de)) { dc->tdc_current = LIST_FIRST(&de->ud.td_duphead); MPASS(dc->tdc_current != NULL); } return (dc->tdc_current); } /* Lookup directory entry in RB-Tree. Function may return duphead entry. */ static struct tmpfs_dirent * tmpfs_dir_xlookup_hash(struct tmpfs_node *dnode, uint32_t hash) { struct tmpfs_dirent *de, dekey; dekey.td_hash = hash; de = RB_FIND(tmpfs_dir, &dnode->tn_dir.tn_dirhead, &dekey); return (de); } /* Lookup directory entry by cookie, initialize directory cursor accordingly. */ static struct tmpfs_dirent * tmpfs_dir_lookup_cookie(struct tmpfs_node *node, off_t cookie, struct tmpfs_dir_cursor *dc) { struct tmpfs_dir *dirhead = &node->tn_dir.tn_dirhead; struct tmpfs_dirent *de, dekey; MPASS(cookie >= TMPFS_DIRCOOKIE_MIN); if (cookie == node->tn_dir.tn_readdir_lastn && (de = node->tn_dir.tn_readdir_lastp) != NULL) { /* Protect against possible race, tn_readdir_last[pn] * may be updated with only shared vnode lock held. */ if (cookie == tmpfs_dirent_cookie(de)) goto out; } if ((cookie & TMPFS_DIRCOOKIE_DUP) != 0) { LIST_FOREACH(de, &node->tn_dir.tn_dupindex, uh.td_dup.index_entries) { MPASS(tmpfs_dirent_dup(de)); if (de->td_cookie == cookie) goto out; /* dupindex list is sorted. */ if (de->td_cookie < cookie) { de = NULL; goto out; } } MPASS(de == NULL); goto out; } if ((cookie & TMPFS_DIRCOOKIE_MASK) != cookie) { de = NULL; } else { dekey.td_hash = cookie; /* Recover if direntry for cookie was removed */ de = RB_NFIND(tmpfs_dir, dirhead, &dekey); } dc->tdc_tree = de; dc->tdc_current = de; if (de != NULL && tmpfs_dirent_duphead(de)) { dc->tdc_current = LIST_FIRST(&de->ud.td_duphead); MPASS(dc->tdc_current != NULL); } return (dc->tdc_current); out: dc->tdc_tree = de; dc->tdc_current = de; if (de != NULL && tmpfs_dirent_dup(de)) dc->tdc_tree = tmpfs_dir_xlookup_hash(node, de->td_hash); return (dc->tdc_current); } /* * Looks for a directory entry in the directory represented by node. * 'cnp' describes the name of the entry to look for. Note that the . * and .. components are not allowed as they do not physically exist * within directories. * * Returns a pointer to the entry when found, otherwise NULL. */ struct tmpfs_dirent * tmpfs_dir_lookup(struct tmpfs_node *node, struct tmpfs_node *f, struct componentname *cnp) { struct tmpfs_dir_duphead *duphead; struct tmpfs_dirent *de; uint32_t hash; MPASS(IMPLIES(cnp->cn_namelen == 1, cnp->cn_nameptr[0] != '.')); MPASS(IMPLIES(cnp->cn_namelen == 2, !(cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.'))); TMPFS_VALIDATE_DIR(node); hash = tmpfs_dirent_hash(cnp->cn_nameptr, cnp->cn_namelen); de = tmpfs_dir_xlookup_hash(node, hash); if (de != NULL && tmpfs_dirent_duphead(de)) { duphead = &de->ud.td_duphead; LIST_FOREACH(de, duphead, uh.td_dup.entries) { if (TMPFS_DIRENT_MATCHES(de, cnp->cn_nameptr, cnp->cn_namelen)) break; } } else if (de != NULL) { if (!TMPFS_DIRENT_MATCHES(de, cnp->cn_nameptr, cnp->cn_namelen)) de = NULL; } if (de != NULL && f != NULL && de->td_node != f) de = NULL; return (de); } /* * Attach duplicate-cookie directory entry nde to dnode and insert to dupindex * list, allocate new cookie value. */ static void tmpfs_dir_attach_dup(struct tmpfs_node *dnode, struct tmpfs_dir_duphead *duphead, struct tmpfs_dirent *nde) { struct tmpfs_dir_duphead *dupindex; struct tmpfs_dirent *de, *pde; dupindex = &dnode->tn_dir.tn_dupindex; de = LIST_FIRST(dupindex); if (de == NULL || de->td_cookie < TMPFS_DIRCOOKIE_DUP_MAX) { if (de == NULL) nde->td_cookie = TMPFS_DIRCOOKIE_DUP_MIN; else nde->td_cookie = de->td_cookie + 1; MPASS(tmpfs_dirent_dup(nde)); LIST_INSERT_HEAD(dupindex, nde, uh.td_dup.index_entries); LIST_INSERT_HEAD(duphead, nde, uh.td_dup.entries); return; } /* * Cookie numbers are near exhaustion. Scan dupindex list for unused * numbers. dupindex list is sorted in descending order. Keep it so * after inserting nde. */ while (1) { pde = de; de = LIST_NEXT(de, uh.td_dup.index_entries); if (de == NULL && pde->td_cookie != TMPFS_DIRCOOKIE_DUP_MIN) { /* * Last element of the index doesn't have minimal cookie * value, use it. */ nde->td_cookie = TMPFS_DIRCOOKIE_DUP_MIN; LIST_INSERT_AFTER(pde, nde, uh.td_dup.index_entries); LIST_INSERT_HEAD(duphead, nde, uh.td_dup.entries); return; } else if (de == NULL) { /* * We are so lucky have 2^30 hash duplicates in single * directory :) Return largest possible cookie value. * It should be fine except possible issues with * VOP_READDIR restart. */ nde->td_cookie = TMPFS_DIRCOOKIE_DUP_MAX; LIST_INSERT_HEAD(dupindex, nde, uh.td_dup.index_entries); LIST_INSERT_HEAD(duphead, nde, uh.td_dup.entries); return; } if (de->td_cookie + 1 == pde->td_cookie || de->td_cookie >= TMPFS_DIRCOOKIE_DUP_MAX) continue; /* No hole or invalid cookie. */ nde->td_cookie = de->td_cookie + 1; MPASS(tmpfs_dirent_dup(nde)); MPASS(pde->td_cookie > nde->td_cookie); MPASS(nde->td_cookie > de->td_cookie); LIST_INSERT_BEFORE(de, nde, uh.td_dup.index_entries); LIST_INSERT_HEAD(duphead, nde, uh.td_dup.entries); return; } } /* * Attaches the directory entry de to the directory represented by vp. * Note that this does not change the link count of the node pointed by * the directory entry, as this is done by tmpfs_alloc_dirent. */ void tmpfs_dir_attach(struct vnode *vp, struct tmpfs_dirent *de) { struct tmpfs_node *dnode; struct tmpfs_dirent *xde, *nde; ASSERT_VOP_ELOCKED(vp, __func__); MPASS(de->td_namelen > 0); MPASS(de->td_hash >= TMPFS_DIRCOOKIE_MIN); MPASS(de->td_cookie == de->td_hash); dnode = VP_TO_TMPFS_DIR(vp); dnode->tn_dir.tn_readdir_lastn = 0; dnode->tn_dir.tn_readdir_lastp = NULL; MPASS(!tmpfs_dirent_dup(de)); xde = RB_INSERT(tmpfs_dir, &dnode->tn_dir.tn_dirhead, de); if (xde != NULL && tmpfs_dirent_duphead(xde)) tmpfs_dir_attach_dup(dnode, &xde->ud.td_duphead, de); else if (xde != NULL) { /* * Allocate new duphead. Swap xde with duphead to avoid * adding/removing elements with the same hash. */ MPASS(!tmpfs_dirent_dup(xde)); tmpfs_alloc_dirent(VFS_TO_TMPFS(vp->v_mount), NULL, NULL, 0, &nde); /* *nde = *xde; XXX gcc 4.2.1 may generate invalid code. */ memcpy(nde, xde, sizeof(*xde)); xde->td_cookie |= TMPFS_DIRCOOKIE_DUPHEAD; LIST_INIT(&xde->ud.td_duphead); xde->td_namelen = 0; xde->td_node = NULL; tmpfs_dir_attach_dup(dnode, &xde->ud.td_duphead, nde); tmpfs_dir_attach_dup(dnode, &xde->ud.td_duphead, de); } dnode->tn_size += sizeof(struct tmpfs_dirent); dnode->tn_status |= TMPFS_NODE_ACCESSED | TMPFS_NODE_CHANGED | \ TMPFS_NODE_MODIFIED; tmpfs_update(vp); } /* * Detaches the directory entry de from the directory represented by vp. * Note that this does not change the link count of the node pointed by * the directory entry, as this is done by tmpfs_free_dirent. */ void tmpfs_dir_detach(struct vnode *vp, struct tmpfs_dirent *de) { struct tmpfs_mount *tmp; struct tmpfs_dir *head; struct tmpfs_node *dnode; struct tmpfs_dirent *xde; ASSERT_VOP_ELOCKED(vp, __func__); dnode = VP_TO_TMPFS_DIR(vp); head = &dnode->tn_dir.tn_dirhead; dnode->tn_dir.tn_readdir_lastn = 0; dnode->tn_dir.tn_readdir_lastp = NULL; if (tmpfs_dirent_dup(de)) { /* Remove duphead if de was last entry. */ if (LIST_NEXT(de, uh.td_dup.entries) == NULL) { xde = tmpfs_dir_xlookup_hash(dnode, de->td_hash); MPASS(tmpfs_dirent_duphead(xde)); } else xde = NULL; LIST_REMOVE(de, uh.td_dup.entries); LIST_REMOVE(de, uh.td_dup.index_entries); if (xde != NULL) { if (LIST_EMPTY(&xde->ud.td_duphead)) { RB_REMOVE(tmpfs_dir, head, xde); tmp = VFS_TO_TMPFS(vp->v_mount); MPASS(xde->td_node == NULL); tmpfs_free_dirent(tmp, xde); } } de->td_cookie = de->td_hash; } else RB_REMOVE(tmpfs_dir, head, de); dnode->tn_size -= sizeof(struct tmpfs_dirent); dnode->tn_status |= TMPFS_NODE_ACCESSED | TMPFS_NODE_CHANGED | \ TMPFS_NODE_MODIFIED; tmpfs_update(vp); } void tmpfs_dir_destroy(struct tmpfs_mount *tmp, struct tmpfs_node *dnode) { struct tmpfs_dirent *de, *dde, *nde; RB_FOREACH_SAFE(de, tmpfs_dir, &dnode->tn_dir.tn_dirhead, nde) { RB_REMOVE(tmpfs_dir, &dnode->tn_dir.tn_dirhead, de); /* Node may already be destroyed. */ de->td_node = NULL; if (tmpfs_dirent_duphead(de)) { while ((dde = LIST_FIRST(&de->ud.td_duphead)) != NULL) { LIST_REMOVE(dde, uh.td_dup.entries); dde->td_node = NULL; tmpfs_free_dirent(tmp, dde); } } tmpfs_free_dirent(tmp, de); } } /* * Helper function for tmpfs_readdir. Creates a '.' entry for the given * directory and returns it in the uio space. The function returns 0 * on success, -1 if there was not enough space in the uio structure to * hold the directory entry or an appropriate error code if another * error happens. */ static int tmpfs_dir_getdotdent(struct tmpfs_mount *tm, struct tmpfs_node *node, struct uio *uio) { int error; struct dirent dent; TMPFS_VALIDATE_DIR(node); MPASS(uio->uio_offset == TMPFS_DIRCOOKIE_DOT); dent.d_fileno = node->tn_id; dent.d_type = DT_DIR; dent.d_namlen = 1; dent.d_name[0] = '.'; dent.d_reclen = GENERIC_DIRSIZ(&dent); dirent_terminate(&dent); if (dent.d_reclen > uio->uio_resid) error = EJUSTRETURN; else error = uiomove(&dent, dent.d_reclen, uio); tmpfs_set_status(tm, node, TMPFS_NODE_ACCESSED); return (error); } /* * Helper function for tmpfs_readdir. Creates a '..' entry for the given * directory and returns it in the uio space. The function returns 0 * on success, -1 if there was not enough space in the uio structure to * hold the directory entry or an appropriate error code if another * error happens. */ static int tmpfs_dir_getdotdotdent(struct tmpfs_mount *tm, struct tmpfs_node *node, struct uio *uio) { struct tmpfs_node *parent; struct dirent dent; int error; TMPFS_VALIDATE_DIR(node); MPASS(uio->uio_offset == TMPFS_DIRCOOKIE_DOTDOT); /* * Return ENOENT if the current node is already removed. */ TMPFS_ASSERT_LOCKED(node); parent = node->tn_dir.tn_parent; if (parent == NULL) return (ENOENT); TMPFS_NODE_LOCK(parent); dent.d_fileno = parent->tn_id; TMPFS_NODE_UNLOCK(parent); dent.d_type = DT_DIR; dent.d_namlen = 2; dent.d_name[0] = '.'; dent.d_name[1] = '.'; dent.d_reclen = GENERIC_DIRSIZ(&dent); dirent_terminate(&dent); if (dent.d_reclen > uio->uio_resid) error = EJUSTRETURN; else error = uiomove(&dent, dent.d_reclen, uio); tmpfs_set_status(tm, node, TMPFS_NODE_ACCESSED); return (error); } /* * Helper function for tmpfs_readdir. Returns as much directory entries * as can fit in the uio space. The read starts at uio->uio_offset. * The function returns 0 on success, -1 if there was not enough space * in the uio structure to hold the directory entry or an appropriate * error code if another error happens. */ int tmpfs_dir_getdents(struct tmpfs_mount *tm, struct tmpfs_node *node, struct uio *uio, int maxcookies, u_long *cookies, int *ncookies) { struct tmpfs_dir_cursor dc; struct tmpfs_dirent *de; off_t off; int error; TMPFS_VALIDATE_DIR(node); off = 0; /* * Lookup the node from the current offset. The starting offset of * 0 will lookup both '.' and '..', and then the first real entry, * or EOF if there are none. Then find all entries for the dir that * fit into the buffer. Once no more entries are found (de == NULL), * the offset is set to TMPFS_DIRCOOKIE_EOF, which will cause the next * call to return 0. */ switch (uio->uio_offset) { case TMPFS_DIRCOOKIE_DOT: error = tmpfs_dir_getdotdent(tm, node, uio); if (error != 0) return (error); uio->uio_offset = TMPFS_DIRCOOKIE_DOTDOT; if (cookies != NULL) cookies[(*ncookies)++] = off = uio->uio_offset; /* FALLTHROUGH */ case TMPFS_DIRCOOKIE_DOTDOT: error = tmpfs_dir_getdotdotdent(tm, node, uio); if (error != 0) return (error); de = tmpfs_dir_first(node, &dc); uio->uio_offset = tmpfs_dirent_cookie(de); if (cookies != NULL) cookies[(*ncookies)++] = off = uio->uio_offset; /* EOF. */ if (de == NULL) return (0); break; case TMPFS_DIRCOOKIE_EOF: return (0); default: de = tmpfs_dir_lookup_cookie(node, uio->uio_offset, &dc); if (de == NULL) return (EINVAL); if (cookies != NULL) off = tmpfs_dirent_cookie(de); } /* Read as much entries as possible; i.e., until we reach the end of * the directory or we exhaust uio space. */ do { struct dirent d; /* Create a dirent structure representing the current * tmpfs_node and fill it. */ if (de->td_node == NULL) { d.d_fileno = 1; d.d_type = DT_WHT; } else { d.d_fileno = de->td_node->tn_id; switch (de->td_node->tn_type) { case VBLK: d.d_type = DT_BLK; break; case VCHR: d.d_type = DT_CHR; break; case VDIR: d.d_type = DT_DIR; break; case VFIFO: d.d_type = DT_FIFO; break; case VLNK: d.d_type = DT_LNK; break; case VREG: d.d_type = DT_REG; break; case VSOCK: d.d_type = DT_SOCK; break; default: panic("tmpfs_dir_getdents: type %p %d", de->td_node, (int)de->td_node->tn_type); } } d.d_namlen = de->td_namelen; MPASS(de->td_namelen < sizeof(d.d_name)); (void)memcpy(d.d_name, de->ud.td_name, de->td_namelen); d.d_reclen = GENERIC_DIRSIZ(&d); dirent_terminate(&d); /* Stop reading if the directory entry we are treating is * bigger than the amount of data that can be returned. */ if (d.d_reclen > uio->uio_resid) { error = EJUSTRETURN; break; } /* Copy the new dirent structure into the output buffer and * advance pointers. */ error = uiomove(&d, d.d_reclen, uio); if (error == 0) { de = tmpfs_dir_next(node, &dc); if (cookies != NULL) { off = tmpfs_dirent_cookie(de); MPASS(*ncookies < maxcookies); cookies[(*ncookies)++] = off; } } } while (error == 0 && uio->uio_resid > 0 && de != NULL); /* Skip setting off when using cookies as it is already done above. */ if (cookies == NULL) off = tmpfs_dirent_cookie(de); /* Update the offset and cache. */ uio->uio_offset = off; node->tn_dir.tn_readdir_lastn = off; node->tn_dir.tn_readdir_lastp = de; tmpfs_set_status(tm, node, TMPFS_NODE_ACCESSED); return error; } int tmpfs_dir_whiteout_add(struct vnode *dvp, struct componentname *cnp) { struct tmpfs_dirent *de; int error; error = tmpfs_alloc_dirent(VFS_TO_TMPFS(dvp->v_mount), NULL, cnp->cn_nameptr, cnp->cn_namelen, &de); if (error != 0) return (error); tmpfs_dir_attach(dvp, de); return (0); } void tmpfs_dir_whiteout_remove(struct vnode *dvp, struct componentname *cnp) { struct tmpfs_dirent *de; de = tmpfs_dir_lookup(VP_TO_TMPFS_DIR(dvp), NULL, cnp); MPASS(de != NULL && de->td_node == NULL); tmpfs_dir_detach(dvp, de); tmpfs_free_dirent(VFS_TO_TMPFS(dvp->v_mount), de); } /* * Resizes the aobj associated with the regular file pointed to by 'vp' to the * size 'newsize'. 'vp' must point to a vnode that represents a regular file. * 'newsize' must be positive. * * Returns zero on success or an appropriate error code on failure. */ int tmpfs_reg_resize(struct vnode *vp, off_t newsize, boolean_t ignerr) { struct tmpfs_mount *tmp; struct tmpfs_node *node; vm_object_t uobj; vm_page_t m; vm_pindex_t idx, newpages, oldpages; off_t oldsize; int base, rv; MPASS(vp->v_type == VREG); MPASS(newsize >= 0); node = VP_TO_TMPFS_NODE(vp); uobj = node->tn_reg.tn_aobj; tmp = VFS_TO_TMPFS(vp->v_mount); /* * Convert the old and new sizes to the number of pages needed to * store them. It may happen that we do not need to do anything * because the last allocated page can accommodate the change on * its own. */ oldsize = node->tn_size; oldpages = OFF_TO_IDX(oldsize + PAGE_MASK); MPASS(oldpages == uobj->size); newpages = OFF_TO_IDX(newsize + PAGE_MASK); if (__predict_true(newpages == oldpages && newsize >= oldsize)) { node->tn_size = newsize; return (0); } if (newpages > oldpages && tmpfs_pages_check_avail(tmp, newpages - oldpages) == 0) return (ENOSPC); VM_OBJECT_WLOCK(uobj); if (newsize < oldsize) { /* * Zero the truncated part of the last page. */ base = newsize & PAGE_MASK; if (base != 0) { idx = OFF_TO_IDX(newsize); retry: m = vm_page_grab(uobj, idx, VM_ALLOC_NOCREAT); if (m != NULL) { - MPASS(m->valid == VM_PAGE_BITS_ALL); + MPASS(vm_page_all_valid(m)); } else if (vm_pager_has_page(uobj, idx, NULL, NULL)) { m = vm_page_alloc(uobj, idx, VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL); if (m == NULL) goto retry; rv = vm_pager_get_pages(uobj, &m, 1, NULL, NULL); if (rv == VM_PAGER_OK) { /* * Since the page was not resident, * and therefore not recently * accessed, immediately enqueue it * for asynchronous laundering. The * current operation is not regarded * as an access. */ vm_page_lock(m); vm_page_launder(m); vm_page_unlock(m); } else { vm_page_free(m); if (ignerr) m = NULL; else { VM_OBJECT_WUNLOCK(uobj); return (EIO); } } } if (m != NULL) { pmap_zero_page_area(m, base, PAGE_SIZE - base); vm_page_dirty(m); vm_page_xunbusy(m); vm_pager_page_unswapped(m); } } /* * Release any swap space and free any whole pages. */ if (newpages < oldpages) { swap_pager_freespace(uobj, newpages, oldpages - newpages); vm_object_page_remove(uobj, newpages, 0, 0); } } uobj->size = newpages; VM_OBJECT_WUNLOCK(uobj); atomic_add_long(&tmp->tm_pages_used, newpages - oldpages); node->tn_size = newsize; return (0); } void tmpfs_check_mtime(struct vnode *vp) { struct tmpfs_node *node; struct vm_object *obj; ASSERT_VOP_ELOCKED(vp, "check_mtime"); if (vp->v_type != VREG) return; obj = vp->v_object; KASSERT((obj->flags & (OBJ_TMPFS_NODE | OBJ_TMPFS)) == (OBJ_TMPFS_NODE | OBJ_TMPFS), ("non-tmpfs obj")); /* unlocked read */ if ((obj->flags & OBJ_TMPFS_DIRTY) != 0) { VM_OBJECT_WLOCK(obj); if ((obj->flags & OBJ_TMPFS_DIRTY) != 0) { obj->flags &= ~OBJ_TMPFS_DIRTY; node = VP_TO_TMPFS_NODE(vp); node->tn_status |= TMPFS_NODE_MODIFIED | TMPFS_NODE_CHANGED; } VM_OBJECT_WUNLOCK(obj); } } /* * Change flags of the given vnode. * Caller should execute tmpfs_update on vp after a successful execution. * The vnode must be locked on entry and remain locked on exit. */ int tmpfs_chflags(struct vnode *vp, u_long flags, struct ucred *cred, struct thread *p) { int error; struct tmpfs_node *node; ASSERT_VOP_ELOCKED(vp, "chflags"); node = VP_TO_TMPFS_NODE(vp); if ((flags & ~(SF_APPEND | SF_ARCHIVED | SF_IMMUTABLE | SF_NOUNLINK | UF_APPEND | UF_ARCHIVE | UF_HIDDEN | UF_IMMUTABLE | UF_NODUMP | UF_NOUNLINK | UF_OFFLINE | UF_OPAQUE | UF_READONLY | UF_REPARSE | UF_SPARSE | UF_SYSTEM)) != 0) return (EOPNOTSUPP); /* Disallow this operation if the file system is mounted read-only. */ if (vp->v_mount->mnt_flag & MNT_RDONLY) return EROFS; /* * Callers may only modify the file flags on objects they * have VADMIN rights for. */ if ((error = VOP_ACCESS(vp, VADMIN, cred, p))) return (error); /* * Unprivileged processes are not permitted to unset system * flags, or modify flags if any system flags are set. */ if (!priv_check_cred(cred, PRIV_VFS_SYSFLAGS)) { if (node->tn_flags & (SF_NOUNLINK | SF_IMMUTABLE | SF_APPEND)) { error = securelevel_gt(cred, 0); if (error) return (error); } } else { if (node->tn_flags & (SF_NOUNLINK | SF_IMMUTABLE | SF_APPEND) || ((flags ^ node->tn_flags) & SF_SETTABLE)) return (EPERM); } node->tn_flags = flags; node->tn_status |= TMPFS_NODE_CHANGED; ASSERT_VOP_ELOCKED(vp, "chflags2"); return (0); } /* * Change access mode on the given vnode. * Caller should execute tmpfs_update on vp after a successful execution. * The vnode must be locked on entry and remain locked on exit. */ int tmpfs_chmod(struct vnode *vp, mode_t mode, struct ucred *cred, struct thread *p) { int error; struct tmpfs_node *node; ASSERT_VOP_ELOCKED(vp, "chmod"); node = VP_TO_TMPFS_NODE(vp); /* Disallow this operation if the file system is mounted read-only. */ if (vp->v_mount->mnt_flag & MNT_RDONLY) return EROFS; /* Immutable or append-only files cannot be modified, either. */ if (node->tn_flags & (IMMUTABLE | APPEND)) return EPERM; /* * To modify the permissions on a file, must possess VADMIN * for that file. */ if ((error = VOP_ACCESS(vp, VADMIN, cred, p))) return (error); /* * Privileged processes may set the sticky bit on non-directories, * as well as set the setgid bit on a file with a group that the * process is not a member of. */ if (vp->v_type != VDIR && (mode & S_ISTXT)) { if (priv_check_cred(cred, PRIV_VFS_STICKYFILE)) return (EFTYPE); } if (!groupmember(node->tn_gid, cred) && (mode & S_ISGID)) { error = priv_check_cred(cred, PRIV_VFS_SETGID); if (error) return (error); } node->tn_mode &= ~ALLPERMS; node->tn_mode |= mode & ALLPERMS; node->tn_status |= TMPFS_NODE_CHANGED; ASSERT_VOP_ELOCKED(vp, "chmod2"); return (0); } /* * Change ownership of the given vnode. At least one of uid or gid must * be different than VNOVAL. If one is set to that value, the attribute * is unchanged. * Caller should execute tmpfs_update on vp after a successful execution. * The vnode must be locked on entry and remain locked on exit. */ int tmpfs_chown(struct vnode *vp, uid_t uid, gid_t gid, struct ucred *cred, struct thread *p) { int error; struct tmpfs_node *node; uid_t ouid; gid_t ogid; ASSERT_VOP_ELOCKED(vp, "chown"); node = VP_TO_TMPFS_NODE(vp); /* Assign default values if they are unknown. */ MPASS(uid != VNOVAL || gid != VNOVAL); if (uid == VNOVAL) uid = node->tn_uid; if (gid == VNOVAL) gid = node->tn_gid; MPASS(uid != VNOVAL && gid != VNOVAL); /* Disallow this operation if the file system is mounted read-only. */ if (vp->v_mount->mnt_flag & MNT_RDONLY) return EROFS; /* Immutable or append-only files cannot be modified, either. */ if (node->tn_flags & (IMMUTABLE | APPEND)) return EPERM; /* * To modify the ownership of a file, must possess VADMIN for that * file. */ if ((error = VOP_ACCESS(vp, VADMIN, cred, p))) return (error); /* * To change the owner of a file, or change the group of a file to a * group of which we are not a member, the caller must have * privilege. */ if ((uid != node->tn_uid || (gid != node->tn_gid && !groupmember(gid, cred))) && (error = priv_check_cred(cred, PRIV_VFS_CHOWN))) return (error); ogid = node->tn_gid; ouid = node->tn_uid; node->tn_uid = uid; node->tn_gid = gid; node->tn_status |= TMPFS_NODE_CHANGED; if ((node->tn_mode & (S_ISUID | S_ISGID)) && (ouid != uid || ogid != gid)) { if (priv_check_cred(cred, PRIV_VFS_RETAINSUGID)) node->tn_mode &= ~(S_ISUID | S_ISGID); } ASSERT_VOP_ELOCKED(vp, "chown2"); return (0); } /* * Change size of the given vnode. * Caller should execute tmpfs_update on vp after a successful execution. * The vnode must be locked on entry and remain locked on exit. */ int tmpfs_chsize(struct vnode *vp, u_quad_t size, struct ucred *cred, struct thread *p) { int error; struct tmpfs_node *node; ASSERT_VOP_ELOCKED(vp, "chsize"); node = VP_TO_TMPFS_NODE(vp); /* Decide whether this is a valid operation based on the file type. */ error = 0; switch (vp->v_type) { case VDIR: return EISDIR; case VREG: if (vp->v_mount->mnt_flag & MNT_RDONLY) return EROFS; break; case VBLK: /* FALLTHROUGH */ case VCHR: /* FALLTHROUGH */ case VFIFO: /* Allow modifications of special files even if in the file * system is mounted read-only (we are not modifying the * files themselves, but the objects they represent). */ return 0; default: /* Anything else is unsupported. */ return EOPNOTSUPP; } /* Immutable or append-only files cannot be modified, either. */ if (node->tn_flags & (IMMUTABLE | APPEND)) return EPERM; error = tmpfs_truncate(vp, size); /* tmpfs_truncate will raise the NOTE_EXTEND and NOTE_ATTRIB kevents * for us, as will update tn_status; no need to do that here. */ ASSERT_VOP_ELOCKED(vp, "chsize2"); return (error); } /* * Change access and modification times of the given vnode. * Caller should execute tmpfs_update on vp after a successful execution. * The vnode must be locked on entry and remain locked on exit. */ int tmpfs_chtimes(struct vnode *vp, struct vattr *vap, struct ucred *cred, struct thread *l) { int error; struct tmpfs_node *node; ASSERT_VOP_ELOCKED(vp, "chtimes"); node = VP_TO_TMPFS_NODE(vp); /* Disallow this operation if the file system is mounted read-only. */ if (vp->v_mount->mnt_flag & MNT_RDONLY) return EROFS; /* Immutable or append-only files cannot be modified, either. */ if (node->tn_flags & (IMMUTABLE | APPEND)) return EPERM; error = vn_utimes_perm(vp, vap, cred, l); if (error != 0) return (error); if (vap->va_atime.tv_sec != VNOVAL) node->tn_status |= TMPFS_NODE_ACCESSED; if (vap->va_mtime.tv_sec != VNOVAL) node->tn_status |= TMPFS_NODE_MODIFIED; if (vap->va_birthtime.tv_sec != VNOVAL) node->tn_status |= TMPFS_NODE_MODIFIED; tmpfs_itimes(vp, &vap->va_atime, &vap->va_mtime); if (vap->va_birthtime.tv_sec != VNOVAL) node->tn_birthtime = vap->va_birthtime; ASSERT_VOP_ELOCKED(vp, "chtimes2"); return (0); } void tmpfs_set_status(struct tmpfs_mount *tm, struct tmpfs_node *node, int status) { if ((node->tn_status & status) == status || tm->tm_ronly) return; TMPFS_NODE_LOCK(node); node->tn_status |= status; TMPFS_NODE_UNLOCK(node); } /* Sync timestamps */ void tmpfs_itimes(struct vnode *vp, const struct timespec *acc, const struct timespec *mod) { struct tmpfs_node *node; struct timespec now; ASSERT_VOP_LOCKED(vp, "tmpfs_itimes"); node = VP_TO_TMPFS_NODE(vp); if ((node->tn_status & (TMPFS_NODE_ACCESSED | TMPFS_NODE_MODIFIED | TMPFS_NODE_CHANGED)) == 0) return; vfs_timestamp(&now); TMPFS_NODE_LOCK(node); if (node->tn_status & TMPFS_NODE_ACCESSED) { if (acc == NULL) acc = &now; node->tn_atime = *acc; } if (node->tn_status & TMPFS_NODE_MODIFIED) { if (mod == NULL) mod = &now; node->tn_mtime = *mod; } if (node->tn_status & TMPFS_NODE_CHANGED) node->tn_ctime = now; node->tn_status &= ~(TMPFS_NODE_ACCESSED | TMPFS_NODE_MODIFIED | TMPFS_NODE_CHANGED); TMPFS_NODE_UNLOCK(node); /* XXX: FIX? The entropy here is desirable, but the harvesting may be expensive */ random_harvest_queue(node, sizeof(*node), RANDOM_FS_ATIME); } void tmpfs_update(struct vnode *vp) { tmpfs_itimes(vp, NULL, NULL); } int tmpfs_truncate(struct vnode *vp, off_t length) { int error; struct tmpfs_node *node; node = VP_TO_TMPFS_NODE(vp); if (length < 0) { error = EINVAL; goto out; } if (node->tn_size == length) { error = 0; goto out; } if (length > VFS_TO_TMPFS(vp->v_mount)->tm_maxfilesize) return (EFBIG); error = tmpfs_reg_resize(vp, length, FALSE); if (error == 0) node->tn_status |= TMPFS_NODE_CHANGED | TMPFS_NODE_MODIFIED; out: tmpfs_update(vp); return (error); } static __inline int tmpfs_dirtree_cmp(struct tmpfs_dirent *a, struct tmpfs_dirent *b) { if (a->td_hash > b->td_hash) return (1); else if (a->td_hash < b->td_hash) return (-1); return (0); } RB_GENERATE_STATIC(tmpfs_dir, tmpfs_dirent, uh.td_entries, tmpfs_dirtree_cmp); Index: head/sys/kern/kern_exec.c =================================================================== --- head/sys/kern/kern_exec.c (revision 353538) +++ head/sys/kern/kern_exec.c (revision 353539) @@ -1,1831 +1,1837 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 1993, David Greenman * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include "opt_capsicum.h" #include "opt_hwpmc_hooks.h" #include "opt_ktrace.h" #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef KTRACE #include #endif #include #include #include #include #include #include #include #include #include #ifdef HWPMC_HOOKS #include #endif #include #include #include #ifdef KDTRACE_HOOKS #include dtrace_execexit_func_t dtrace_fasttrap_exec; #endif SDT_PROVIDER_DECLARE(proc); SDT_PROBE_DEFINE1(proc, , , exec, "char *"); SDT_PROBE_DEFINE1(proc, , , exec__failure, "int"); SDT_PROBE_DEFINE1(proc, , , exec__success, "char *"); MALLOC_DEFINE(M_PARGS, "proc-args", "Process arguments"); int coredump_pack_fileinfo = 1; SYSCTL_INT(_kern, OID_AUTO, coredump_pack_fileinfo, CTLFLAG_RWTUN, &coredump_pack_fileinfo, 0, "Enable file path packing in 'procstat -f' coredump notes"); int coredump_pack_vmmapinfo = 1; SYSCTL_INT(_kern, OID_AUTO, coredump_pack_vmmapinfo, CTLFLAG_RWTUN, &coredump_pack_vmmapinfo, 0, "Enable file path packing in 'procstat -v' coredump notes"); static int sysctl_kern_ps_strings(SYSCTL_HANDLER_ARGS); static int sysctl_kern_usrstack(SYSCTL_HANDLER_ARGS); static int sysctl_kern_stackprot(SYSCTL_HANDLER_ARGS); static int do_execve(struct thread *td, struct image_args *args, struct mac *mac_p); /* XXX This should be vm_size_t. */ SYSCTL_PROC(_kern, KERN_PS_STRINGS, ps_strings, CTLTYPE_ULONG|CTLFLAG_RD| CTLFLAG_CAPRD|CTLFLAG_MPSAFE, NULL, 0, sysctl_kern_ps_strings, "LU", ""); /* XXX This should be vm_size_t. */ SYSCTL_PROC(_kern, KERN_USRSTACK, usrstack, CTLTYPE_ULONG|CTLFLAG_RD| CTLFLAG_CAPRD|CTLFLAG_MPSAFE, NULL, 0, sysctl_kern_usrstack, "LU", ""); SYSCTL_PROC(_kern, OID_AUTO, stackprot, CTLTYPE_INT|CTLFLAG_RD|CTLFLAG_MPSAFE, NULL, 0, sysctl_kern_stackprot, "I", ""); u_long ps_arg_cache_limit = PAGE_SIZE / 16; SYSCTL_ULONG(_kern, OID_AUTO, ps_arg_cache_limit, CTLFLAG_RW, &ps_arg_cache_limit, 0, ""); static int disallow_high_osrel; SYSCTL_INT(_kern, OID_AUTO, disallow_high_osrel, CTLFLAG_RW, &disallow_high_osrel, 0, "Disallow execution of binaries built for higher version of the world"); static int map_at_zero = 0; SYSCTL_INT(_security_bsd, OID_AUTO, map_at_zero, CTLFLAG_RWTUN, &map_at_zero, 0, "Permit processes to map an object at virtual address 0."); static int sysctl_kern_ps_strings(SYSCTL_HANDLER_ARGS) { struct proc *p; int error; p = curproc; #ifdef SCTL_MASK32 if (req->flags & SCTL_MASK32) { unsigned int val; val = (unsigned int)p->p_sysent->sv_psstrings; error = SYSCTL_OUT(req, &val, sizeof(val)); } else #endif error = SYSCTL_OUT(req, &p->p_sysent->sv_psstrings, sizeof(p->p_sysent->sv_psstrings)); return error; } static int sysctl_kern_usrstack(SYSCTL_HANDLER_ARGS) { struct proc *p; int error; p = curproc; #ifdef SCTL_MASK32 if (req->flags & SCTL_MASK32) { unsigned int val; val = (unsigned int)p->p_sysent->sv_usrstack; error = SYSCTL_OUT(req, &val, sizeof(val)); } else #endif error = SYSCTL_OUT(req, &p->p_sysent->sv_usrstack, sizeof(p->p_sysent->sv_usrstack)); return error; } static int sysctl_kern_stackprot(SYSCTL_HANDLER_ARGS) { struct proc *p; p = curproc; return (SYSCTL_OUT(req, &p->p_sysent->sv_stackprot, sizeof(p->p_sysent->sv_stackprot))); } /* * Each of the items is a pointer to a `const struct execsw', hence the * double pointer here. */ static const struct execsw **execsw; #ifndef _SYS_SYSPROTO_H_ struct execve_args { char *fname; char **argv; char **envv; }; #endif int sys_execve(struct thread *td, struct execve_args *uap) { struct image_args args; struct vmspace *oldvmspace; int error; error = pre_execve(td, &oldvmspace); if (error != 0) return (error); error = exec_copyin_args(&args, uap->fname, UIO_USERSPACE, uap->argv, uap->envv); if (error == 0) error = kern_execve(td, &args, NULL); post_execve(td, error, oldvmspace); return (error); } #ifndef _SYS_SYSPROTO_H_ struct fexecve_args { int fd; char **argv; char **envv; } #endif int sys_fexecve(struct thread *td, struct fexecve_args *uap) { struct image_args args; struct vmspace *oldvmspace; int error; error = pre_execve(td, &oldvmspace); if (error != 0) return (error); error = exec_copyin_args(&args, NULL, UIO_SYSSPACE, uap->argv, uap->envv); if (error == 0) { args.fd = uap->fd; error = kern_execve(td, &args, NULL); } post_execve(td, error, oldvmspace); return (error); } #ifndef _SYS_SYSPROTO_H_ struct __mac_execve_args { char *fname; char **argv; char **envv; struct mac *mac_p; }; #endif int sys___mac_execve(struct thread *td, struct __mac_execve_args *uap) { #ifdef MAC struct image_args args; struct vmspace *oldvmspace; int error; error = pre_execve(td, &oldvmspace); if (error != 0) return (error); error = exec_copyin_args(&args, uap->fname, UIO_USERSPACE, uap->argv, uap->envv); if (error == 0) error = kern_execve(td, &args, uap->mac_p); post_execve(td, error, oldvmspace); return (error); #else return (ENOSYS); #endif } int pre_execve(struct thread *td, struct vmspace **oldvmspace) { struct proc *p; int error; KASSERT(td == curthread, ("non-current thread %p", td)); error = 0; p = td->td_proc; if ((p->p_flag & P_HADTHREADS) != 0) { PROC_LOCK(p); if (thread_single(p, SINGLE_BOUNDARY) != 0) error = ERESTART; PROC_UNLOCK(p); } KASSERT(error != 0 || (td->td_pflags & TDP_EXECVMSPC) == 0, ("nested execve")); *oldvmspace = p->p_vmspace; return (error); } void post_execve(struct thread *td, int error, struct vmspace *oldvmspace) { struct proc *p; KASSERT(td == curthread, ("non-current thread %p", td)); p = td->td_proc; if ((p->p_flag & P_HADTHREADS) != 0) { PROC_LOCK(p); /* * If success, we upgrade to SINGLE_EXIT state to * force other threads to suicide. */ if (error == EJUSTRETURN) thread_single(p, SINGLE_EXIT); else thread_single_end(p, SINGLE_BOUNDARY); PROC_UNLOCK(p); } if ((td->td_pflags & TDP_EXECVMSPC) != 0) { KASSERT(p->p_vmspace != oldvmspace, ("oldvmspace still used")); vmspace_free(oldvmspace); td->td_pflags &= ~TDP_EXECVMSPC; } } /* * XXX: kern_execve has the astonishing property of not always returning to * the caller. If sufficiently bad things happen during the call to * do_execve(), it can end up calling exit1(); as a result, callers must * avoid doing anything which they might need to undo (e.g., allocating * memory). */ int kern_execve(struct thread *td, struct image_args *args, struct mac *mac_p) { AUDIT_ARG_ARGV(args->begin_argv, args->argc, exec_args_get_begin_envv(args) - args->begin_argv); AUDIT_ARG_ENVV(exec_args_get_begin_envv(args), args->envc, args->endp - exec_args_get_begin_envv(args)); return (do_execve(td, args, mac_p)); } /* * In-kernel implementation of execve(). All arguments are assumed to be * userspace pointers from the passed thread. */ static int do_execve(struct thread *td, struct image_args *args, struct mac *mac_p) { struct proc *p = td->td_proc; struct nameidata nd; struct ucred *oldcred; struct uidinfo *euip = NULL; register_t *stack_base; int error, i; struct image_params image_params, *imgp; struct vattr attr; int (*img_first)(struct image_params *); struct pargs *oldargs = NULL, *newargs = NULL; struct sigacts *oldsigacts = NULL, *newsigacts = NULL; #ifdef KTRACE struct vnode *tracevp = NULL; struct ucred *tracecred = NULL; #endif struct vnode *oldtextvp = NULL, *newtextvp; int credential_changing; #ifdef MAC struct label *interpvplabel = NULL; int will_transition; #endif #ifdef HWPMC_HOOKS struct pmckern_procexec pe; #endif static const char fexecv_proc_title[] = "(fexecv)"; imgp = &image_params; /* * Lock the process and set the P_INEXEC flag to indicate that * it should be left alone until we're done here. This is * necessary to avoid race conditions - e.g. in ptrace() - * that might allow a local user to illicitly obtain elevated * privileges. */ PROC_LOCK(p); KASSERT((p->p_flag & P_INEXEC) == 0, ("%s(): process already has P_INEXEC flag", __func__)); p->p_flag |= P_INEXEC; PROC_UNLOCK(p); /* * Initialize part of the common data */ bzero(imgp, sizeof(*imgp)); imgp->proc = p; imgp->attr = &attr; imgp->args = args; oldcred = p->p_ucred; #ifdef MAC error = mac_execve_enter(imgp, mac_p); if (error) goto exec_fail; #endif /* * Translate the file name. namei() returns a vnode pointer * in ni_vp among other things. * * XXXAUDIT: It would be desirable to also audit the name of the * interpreter if this is an interpreted binary. */ if (args->fname != NULL) { NDINIT(&nd, LOOKUP, ISOPEN | LOCKLEAF | LOCKSHARED | FOLLOW | SAVENAME | AUDITVNODE1, UIO_SYSSPACE, args->fname, td); } SDT_PROBE1(proc, , , exec, args->fname); interpret: if (args->fname != NULL) { #ifdef CAPABILITY_MODE /* * While capability mode can't reach this point via direct * path arguments to execve(), we also don't allow * interpreters to be used in capability mode (for now). * Catch indirect lookups and return a permissions error. */ if (IN_CAPABILITY_MODE(td)) { error = ECAPMODE; goto exec_fail; } #endif error = namei(&nd); if (error) goto exec_fail; newtextvp = nd.ni_vp; imgp->vp = newtextvp; } else { AUDIT_ARG_FD(args->fd); /* * Descriptors opened only with O_EXEC or O_RDONLY are allowed. */ error = fgetvp_exec(td, args->fd, &cap_fexecve_rights, &newtextvp); if (error) goto exec_fail; vn_lock(newtextvp, LK_SHARED | LK_RETRY); AUDIT_ARG_VNODE1(newtextvp); imgp->vp = newtextvp; } /* * Check file permissions. Also 'opens' file and sets its vnode to * text mode. */ error = exec_check_permissions(imgp); if (error) goto exec_fail_dealloc; imgp->object = imgp->vp->v_object; if (imgp->object != NULL) vm_object_reference(imgp->object); error = exec_map_first_page(imgp); if (error) goto exec_fail_dealloc; imgp->proc->p_osrel = 0; imgp->proc->p_fctl0 = 0; /* * Implement image setuid/setgid. * * Determine new credentials before attempting image activators * so that it can be used by process_exec handlers to determine * credential/setid changes. * * Don't honor setuid/setgid if the filesystem prohibits it or if * the process is being traced. * * We disable setuid/setgid/etc in capability mode on the basis * that most setugid applications are not written with that * environment in mind, and will therefore almost certainly operate * incorrectly. In principle there's no reason that setugid * applications might not be useful in capability mode, so we may want * to reconsider this conservative design choice in the future. * * XXXMAC: For the time being, use NOSUID to also prohibit * transitions on the file system. */ credential_changing = 0; credential_changing |= (attr.va_mode & S_ISUID) && oldcred->cr_uid != attr.va_uid; credential_changing |= (attr.va_mode & S_ISGID) && oldcred->cr_gid != attr.va_gid; #ifdef MAC will_transition = mac_vnode_execve_will_transition(oldcred, imgp->vp, interpvplabel, imgp); credential_changing |= will_transition; #endif /* Don't inherit PROC_PDEATHSIG_CTL value if setuid/setgid. */ if (credential_changing) imgp->proc->p_pdeathsig = 0; if (credential_changing && #ifdef CAPABILITY_MODE ((oldcred->cr_flags & CRED_FLAG_CAPMODE) == 0) && #endif (imgp->vp->v_mount->mnt_flag & MNT_NOSUID) == 0 && (p->p_flag & P_TRACED) == 0) { imgp->credential_setid = true; VOP_UNLOCK(imgp->vp, 0); imgp->newcred = crdup(oldcred); if (attr.va_mode & S_ISUID) { euip = uifind(attr.va_uid); change_euid(imgp->newcred, euip); } vn_lock(imgp->vp, LK_SHARED | LK_RETRY); if (attr.va_mode & S_ISGID) change_egid(imgp->newcred, attr.va_gid); /* * Implement correct POSIX saved-id behavior. * * XXXMAC: Note that the current logic will save the * uid and gid if a MAC domain transition occurs, even * though maybe it shouldn't. */ change_svuid(imgp->newcred, imgp->newcred->cr_uid); change_svgid(imgp->newcred, imgp->newcred->cr_gid); } else { /* * Implement correct POSIX saved-id behavior. * * XXX: It's not clear that the existing behavior is * POSIX-compliant. A number of sources indicate that the * saved uid/gid should only be updated if the new ruid is * not equal to the old ruid, or the new euid is not equal * to the old euid and the new euid is not equal to the old * ruid. The FreeBSD code always updates the saved uid/gid. * Also, this code uses the new (replaced) euid and egid as * the source, which may or may not be the right ones to use. */ if (oldcred->cr_svuid != oldcred->cr_uid || oldcred->cr_svgid != oldcred->cr_gid) { VOP_UNLOCK(imgp->vp, 0); imgp->newcred = crdup(oldcred); vn_lock(imgp->vp, LK_SHARED | LK_RETRY); change_svuid(imgp->newcred, imgp->newcred->cr_uid); change_svgid(imgp->newcred, imgp->newcred->cr_gid); } } /* The new credentials are installed into the process later. */ /* * Do the best to calculate the full path to the image file. */ if (args->fname != NULL && args->fname[0] == '/') imgp->execpath = args->fname; else { VOP_UNLOCK(imgp->vp, 0); if (vn_fullpath(td, imgp->vp, &imgp->execpath, &imgp->freepath) != 0) imgp->execpath = args->fname; vn_lock(imgp->vp, LK_SHARED | LK_RETRY); } /* * If the current process has a special image activator it * wants to try first, call it. For example, emulating shell * scripts differently. */ error = -1; if ((img_first = imgp->proc->p_sysent->sv_imgact_try) != NULL) error = img_first(imgp); /* * Loop through the list of image activators, calling each one. * An activator returns -1 if there is no match, 0 on success, * and an error otherwise. */ for (i = 0; error == -1 && execsw[i]; ++i) { if (execsw[i]->ex_imgact == NULL || execsw[i]->ex_imgact == img_first) { continue; } error = (*execsw[i]->ex_imgact)(imgp); } if (error) { if (error == -1) error = ENOEXEC; goto exec_fail_dealloc; } /* * Special interpreter operation, cleanup and loop up to try to * activate the interpreter. */ if (imgp->interpreted) { exec_unmap_first_page(imgp); /* * The text reference needs to be removed for scripts. * There is a short period before we determine that * something is a script where text reference is active. * The vnode lock is held over this entire period * so nothing should illegitimately be blocked. */ MPASS(imgp->textset); VOP_UNSET_TEXT_CHECKED(newtextvp); imgp->textset = false; /* free name buffer and old vnode */ if (args->fname != NULL) NDFREE(&nd, NDF_ONLY_PNBUF); #ifdef MAC mac_execve_interpreter_enter(newtextvp, &interpvplabel); #endif if (imgp->opened) { VOP_CLOSE(newtextvp, FREAD, td->td_ucred, td); imgp->opened = 0; } vput(newtextvp); vm_object_deallocate(imgp->object); imgp->object = NULL; imgp->credential_setid = false; if (imgp->newcred != NULL) { crfree(imgp->newcred); imgp->newcred = NULL; } imgp->execpath = NULL; free(imgp->freepath, M_TEMP); imgp->freepath = NULL; /* set new name to that of the interpreter */ NDINIT(&nd, LOOKUP, ISOPEN | LOCKLEAF | FOLLOW | SAVENAME, UIO_SYSSPACE, imgp->interpreter_name, td); args->fname = imgp->interpreter_name; goto interpret; } /* * NB: We unlock the vnode here because it is believed that none * of the sv_copyout_strings/sv_fixup operations require the vnode. */ VOP_UNLOCK(imgp->vp, 0); if (disallow_high_osrel && P_OSREL_MAJOR(p->p_osrel) > P_OSREL_MAJOR(__FreeBSD_version)) { error = ENOEXEC; uprintf("Osrel %d for image %s too high\n", p->p_osrel, imgp->execpath != NULL ? imgp->execpath : ""); vn_lock(imgp->vp, LK_SHARED | LK_RETRY); goto exec_fail_dealloc; } /* ABI enforces the use of Capsicum. Switch into capabilities mode. */ if (SV_PROC_FLAG(p, SV_CAPSICUM)) sys_cap_enter(td, NULL); /* * Copy out strings (args and env) and initialize stack base. */ stack_base = (*p->p_sysent->sv_copyout_strings)(imgp); /* * Stack setup. */ error = (*p->p_sysent->sv_fixup)(&stack_base, imgp); if (error != 0) { vn_lock(imgp->vp, LK_SHARED | LK_RETRY); goto exec_fail_dealloc; } if (args->fdp != NULL) { /* Install a brand new file descriptor table. */ fdinstall_remapped(td, args->fdp); args->fdp = NULL; } else { /* * Keep on using the existing file descriptor table. For * security and other reasons, the file descriptor table * cannot be shared after an exec. */ fdunshare(td); /* close files on exec */ fdcloseexec(td); } /* * Malloc things before we need locks. */ i = exec_args_get_begin_envv(imgp->args) - imgp->args->begin_argv; /* Cache arguments if they fit inside our allowance */ if (ps_arg_cache_limit >= i + sizeof(struct pargs)) { newargs = pargs_alloc(i); bcopy(imgp->args->begin_argv, newargs->ar_args, i); } /* * For security and other reasons, signal handlers cannot * be shared after an exec. The new process gets a copy of the old * handlers. In execsigs(), the new process will have its signals * reset. */ if (sigacts_shared(p->p_sigacts)) { oldsigacts = p->p_sigacts; newsigacts = sigacts_alloc(); sigacts_copy(newsigacts, oldsigacts); } vn_lock(imgp->vp, LK_SHARED | LK_RETRY); PROC_LOCK(p); if (oldsigacts) p->p_sigacts = newsigacts; /* Stop profiling */ stopprofclock(p); /* reset caught signals */ execsigs(p); /* name this process - nameiexec(p, ndp) */ bzero(p->p_comm, sizeof(p->p_comm)); if (args->fname) bcopy(nd.ni_cnd.cn_nameptr, p->p_comm, min(nd.ni_cnd.cn_namelen, MAXCOMLEN)); else if (vn_commname(newtextvp, p->p_comm, sizeof(p->p_comm)) != 0) bcopy(fexecv_proc_title, p->p_comm, sizeof(fexecv_proc_title)); bcopy(p->p_comm, td->td_name, sizeof(td->td_name)); #ifdef KTR sched_clear_tdname(td); #endif /* * mark as execed, wakeup the process that vforked (if any) and tell * it that it now has its own resources back */ p->p_flag |= P_EXEC; if ((p->p_flag2 & P2_NOTRACE_EXEC) == 0) p->p_flag2 &= ~P2_NOTRACE; if ((p->p_flag2 & P2_STKGAP_DISABLE_EXEC) == 0) p->p_flag2 &= ~P2_STKGAP_DISABLE; if (p->p_flag & P_PPWAIT) { p->p_flag &= ~(P_PPWAIT | P_PPTRACE); cv_broadcast(&p->p_pwait); /* STOPs are no longer ignored, arrange for AST */ signotify(td); } /* * Implement image setuid/setgid installation. */ if (imgp->credential_setid) { /* * Turn off syscall tracing for set-id programs, except for * root. Record any set-id flags first to make sure that * we do not regain any tracing during a possible block. */ setsugid(p); #ifdef KTRACE if (p->p_tracecred != NULL && priv_check_cred(p->p_tracecred, PRIV_DEBUG_DIFFCRED)) ktrprocexec(p, &tracecred, &tracevp); #endif /* * Close any file descriptors 0..2 that reference procfs, * then make sure file descriptors 0..2 are in use. * * Both fdsetugidsafety() and fdcheckstd() may call functions * taking sleepable locks, so temporarily drop our locks. */ PROC_UNLOCK(p); VOP_UNLOCK(imgp->vp, 0); fdsetugidsafety(td); error = fdcheckstd(td); vn_lock(imgp->vp, LK_SHARED | LK_RETRY); if (error != 0) goto exec_fail_dealloc; PROC_LOCK(p); #ifdef MAC if (will_transition) { mac_vnode_execve_transition(oldcred, imgp->newcred, imgp->vp, interpvplabel, imgp); } #endif } else { if (oldcred->cr_uid == oldcred->cr_ruid && oldcred->cr_gid == oldcred->cr_rgid) p->p_flag &= ~P_SUGID; } /* * Set the new credentials. */ if (imgp->newcred != NULL) { proc_set_cred(p, imgp->newcred); crfree(oldcred); oldcred = NULL; } /* * Store the vp for use in procfs. This vnode was referenced by namei * or fgetvp_exec. */ oldtextvp = p->p_textvp; p->p_textvp = newtextvp; #ifdef KDTRACE_HOOKS /* * Tell the DTrace fasttrap provider about the exec if it * has declared an interest. */ if (dtrace_fasttrap_exec) dtrace_fasttrap_exec(p); #endif /* * Notify others that we exec'd, and clear the P_INEXEC flag * as we're now a bona fide freshly-execed process. */ KNOTE_LOCKED(p->p_klist, NOTE_EXEC); p->p_flag &= ~P_INEXEC; /* clear "fork but no exec" flag, as we _are_ execing */ p->p_acflag &= ~AFORK; /* * Free any previous argument cache and replace it with * the new argument cache, if any. */ oldargs = p->p_args; p->p_args = newargs; newargs = NULL; PROC_UNLOCK(p); #ifdef HWPMC_HOOKS /* * Check if system-wide sampling is in effect or if the * current process is using PMCs. If so, do exec() time * processing. This processing needs to happen AFTER the * P_INEXEC flag is cleared. */ if (PMC_SYSTEM_SAMPLING_ACTIVE() || PMC_PROC_IS_USING_PMCS(p)) { VOP_UNLOCK(imgp->vp, 0); pe.pm_credentialschanged = credential_changing; pe.pm_entryaddr = imgp->entry_addr; PMC_CALL_HOOK_X(td, PMC_FN_PROCESS_EXEC, (void *) &pe); vn_lock(imgp->vp, LK_SHARED | LK_RETRY); } #endif /* Set values passed into the program in registers. */ (*p->p_sysent->sv_setregs)(td, imgp, (u_long)(uintptr_t)stack_base); vfs_mark_atime(imgp->vp, td->td_ucred); SDT_PROBE1(proc, , , exec__success, args->fname); exec_fail_dealloc: if (imgp->firstpage != NULL) exec_unmap_first_page(imgp); if (imgp->vp != NULL) { if (args->fname) NDFREE(&nd, NDF_ONLY_PNBUF); if (imgp->opened) VOP_CLOSE(imgp->vp, FREAD, td->td_ucred, td); if (imgp->textset) VOP_UNSET_TEXT_CHECKED(imgp->vp); if (error != 0) vput(imgp->vp); else VOP_UNLOCK(imgp->vp, 0); } if (imgp->object != NULL) vm_object_deallocate(imgp->object); free(imgp->freepath, M_TEMP); if (error == 0) { if (p->p_ptevents & PTRACE_EXEC) { PROC_LOCK(p); if (p->p_ptevents & PTRACE_EXEC) td->td_dbgflags |= TDB_EXEC; PROC_UNLOCK(p); } /* * Stop the process here if its stop event mask has * the S_EXEC bit set. */ STOPEVENT(p, S_EXEC, 0); } else { exec_fail: /* we're done here, clear P_INEXEC */ PROC_LOCK(p); p->p_flag &= ~P_INEXEC; PROC_UNLOCK(p); SDT_PROBE1(proc, , , exec__failure, error); } if (imgp->newcred != NULL && oldcred != NULL) crfree(imgp->newcred); #ifdef MAC mac_execve_exit(imgp); mac_execve_interpreter_exit(interpvplabel); #endif exec_free_args(args); /* * Handle deferred decrement of ref counts. */ if (oldtextvp != NULL) vrele(oldtextvp); #ifdef KTRACE if (tracevp != NULL) vrele(tracevp); if (tracecred != NULL) crfree(tracecred); #endif pargs_drop(oldargs); pargs_drop(newargs); if (oldsigacts != NULL) sigacts_free(oldsigacts); if (euip != NULL) uifree(euip); if (error && imgp->vmspace_destroyed) { /* sorry, no more process anymore. exit gracefully */ exit1(td, 0, SIGABRT); /* NOT REACHED */ } #ifdef KTRACE if (error == 0) ktrprocctor(p); #endif /* * We don't want cpu_set_syscall_retval() to overwrite any of * the register values put in place by exec_setregs(). * Implementations of cpu_set_syscall_retval() will leave * registers unmodified when returning EJUSTRETURN. */ return (error == 0 ? EJUSTRETURN : error); } int exec_map_first_page(struct image_params *imgp) { int rv, i, after, initial_pagein; vm_page_t ma[VM_INITIAL_PAGEIN]; vm_object_t object; if (imgp->firstpage != NULL) exec_unmap_first_page(imgp); object = imgp->vp->v_object; if (object == NULL) return (EACCES); VM_OBJECT_WLOCK(object); #if VM_NRESERVLEVEL > 0 vm_object_color(object, 0); #endif retry: ma[0] = vm_page_grab(object, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY | VM_ALLOC_WIRED); - if (ma[0]->valid != VM_PAGE_BITS_ALL) { + if (!vm_page_all_valid(ma[0])) { if (vm_page_busy_acquire(ma[0], VM_ALLOC_WAITFAIL) == 0) { vm_page_unwire_noq(ma[0]); goto retry; } + if (vm_page_all_valid(ma[0])) { + vm_page_xunbusy(ma[0]); + goto out; + } if (!vm_pager_has_page(object, 0, NULL, &after)) { if (vm_page_unwire_noq(ma[0])) vm_page_free(ma[0]); else vm_page_xunbusy(ma[0]); VM_OBJECT_WUNLOCK(object); return (EIO); } initial_pagein = min(after, VM_INITIAL_PAGEIN); KASSERT(initial_pagein <= object->size, ("%s: initial_pagein %d object->size %ju", __func__, initial_pagein, (uintmax_t )object->size)); for (i = 1; i < initial_pagein; i++) { if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) { if (ma[i]->valid) break; if (!vm_page_tryxbusy(ma[i])) break; } else { ma[i] = vm_page_alloc(object, i, VM_ALLOC_NORMAL); if (ma[i] == NULL) break; } } initial_pagein = i; rv = vm_pager_get_pages(object, ma, initial_pagein, NULL, NULL); if (rv != VM_PAGER_OK) { if (vm_page_unwire_noq(ma[0])) vm_page_free(ma[0]); else vm_page_xunbusy(ma[0]); for (i = 1; i < initial_pagein; i++) { if (!vm_page_wired(ma[i])) vm_page_free(ma[i]); else vm_page_xunbusy(ma[i]); } VM_OBJECT_WUNLOCK(object); return (EIO); } vm_page_xunbusy(ma[0]); for (i = 1; i < initial_pagein; i++) vm_page_readahead_finish(ma[i]); } + +out: VM_OBJECT_WUNLOCK(object); imgp->firstpage = sf_buf_alloc(ma[0], 0); imgp->image_header = (char *)sf_buf_kva(imgp->firstpage); return (0); } void exec_unmap_first_page(struct image_params *imgp) { vm_page_t m; if (imgp->firstpage != NULL) { m = sf_buf_page(imgp->firstpage); sf_buf_free(imgp->firstpage); imgp->firstpage = NULL; vm_page_unwire(m, PQ_ACTIVE); } } /* * Destroy old address space, and allocate a new stack. * The new stack is only sgrowsiz large because it is grown * automatically on a page fault. */ int exec_new_vmspace(struct image_params *imgp, struct sysentvec *sv) { int error; struct proc *p = imgp->proc; struct vmspace *vmspace = p->p_vmspace; vm_object_t obj; struct rlimit rlim_stack; vm_offset_t sv_minuser, stack_addr; vm_map_t map; u_long ssiz; imgp->vmspace_destroyed = 1; imgp->sysent = sv; /* May be called with Giant held */ EVENTHANDLER_DIRECT_INVOKE(process_exec, p, imgp); /* * Blow away entire process VM, if address space not shared, * otherwise, create a new VM space so that other threads are * not disrupted */ map = &vmspace->vm_map; if (map_at_zero) sv_minuser = sv->sv_minuser; else sv_minuser = MAX(sv->sv_minuser, PAGE_SIZE); if (vmspace->vm_refcnt == 1 && vm_map_min(map) == sv_minuser && vm_map_max(map) == sv->sv_maxuser && cpu_exec_vmspace_reuse(p, map)) { shmexit(vmspace); pmap_remove_pages(vmspace_pmap(vmspace)); vm_map_remove(map, vm_map_min(map), vm_map_max(map)); /* * An exec terminates mlockall(MCL_FUTURE), ASLR state * must be re-evaluated. */ vm_map_lock(map); vm_map_modflags(map, 0, MAP_WIREFUTURE | MAP_ASLR | MAP_ASLR_IGNSTART); vm_map_unlock(map); } else { error = vmspace_exec(p, sv_minuser, sv->sv_maxuser); if (error) return (error); vmspace = p->p_vmspace; map = &vmspace->vm_map; } map->flags |= imgp->map_flags; /* Map a shared page */ obj = sv->sv_shared_page_obj; if (obj != NULL) { vm_object_reference(obj); error = vm_map_fixed(map, obj, 0, sv->sv_shared_page_base, sv->sv_shared_page_len, VM_PROT_READ | VM_PROT_EXECUTE, VM_PROT_READ | VM_PROT_EXECUTE, MAP_INHERIT_SHARE | MAP_ACC_NO_CHARGE); if (error != KERN_SUCCESS) { vm_object_deallocate(obj); return (vm_mmap_to_errno(error)); } } /* Allocate a new stack */ if (imgp->stack_sz != 0) { ssiz = trunc_page(imgp->stack_sz); PROC_LOCK(p); lim_rlimit_proc(p, RLIMIT_STACK, &rlim_stack); PROC_UNLOCK(p); if (ssiz > rlim_stack.rlim_max) ssiz = rlim_stack.rlim_max; if (ssiz > rlim_stack.rlim_cur) { rlim_stack.rlim_cur = ssiz; kern_setrlimit(curthread, RLIMIT_STACK, &rlim_stack); } } else if (sv->sv_maxssiz != NULL) { ssiz = *sv->sv_maxssiz; } else { ssiz = maxssiz; } imgp->eff_stack_sz = lim_cur(curthread, RLIMIT_STACK); if (ssiz < imgp->eff_stack_sz) imgp->eff_stack_sz = ssiz; stack_addr = sv->sv_usrstack - ssiz; error = vm_map_stack(map, stack_addr, (vm_size_t)ssiz, obj != NULL && imgp->stack_prot != 0 ? imgp->stack_prot : sv->sv_stackprot, VM_PROT_ALL, MAP_STACK_GROWS_DOWN); if (error != KERN_SUCCESS) return (vm_mmap_to_errno(error)); /* * vm_ssize and vm_maxsaddr are somewhat antiquated concepts, but they * are still used to enforce the stack rlimit on the process stack. */ vmspace->vm_ssize = sgrowsiz >> PAGE_SHIFT; vmspace->vm_maxsaddr = (char *)stack_addr; return (0); } /* * Copy out argument and environment strings from the old process address * space into the temporary string buffer. */ int exec_copyin_args(struct image_args *args, const char *fname, enum uio_seg segflg, char **argv, char **envv) { u_long arg, env; int error; bzero(args, sizeof(*args)); if (argv == NULL) return (EFAULT); /* * Allocate demand-paged memory for the file name, argument, and * environment strings. */ error = exec_alloc_args(args); if (error != 0) return (error); /* * Copy the file name. */ error = exec_args_add_fname(args, fname, segflg); if (error != 0) goto err_exit; /* * extract arguments first */ for (;;) { error = fueword(argv++, &arg); if (error == -1) { error = EFAULT; goto err_exit; } if (arg == 0) break; error = exec_args_add_arg(args, (char *)(uintptr_t)arg, UIO_USERSPACE); if (error != 0) goto err_exit; } /* * extract environment strings */ if (envv) { for (;;) { error = fueword(envv++, &env); if (error == -1) { error = EFAULT; goto err_exit; } if (env == 0) break; error = exec_args_add_env(args, (char *)(uintptr_t)env, UIO_USERSPACE); if (error != 0) goto err_exit; } } return (0); err_exit: exec_free_args(args); return (error); } int exec_copyin_data_fds(struct thread *td, struct image_args *args, const void *data, size_t datalen, const int *fds, size_t fdslen) { struct filedesc *ofdp; const char *p; int *kfds; int error; memset(args, '\0', sizeof(*args)); ofdp = td->td_proc->p_fd; if (datalen >= ARG_MAX || fdslen > ofdp->fd_lastfile + 1) return (E2BIG); error = exec_alloc_args(args); if (error != 0) return (error); args->begin_argv = args->buf; args->stringspace = ARG_MAX; if (datalen > 0) { /* * Argument buffer has been provided. Copy it into the * kernel as a single string and add a terminating null * byte. */ error = copyin(data, args->begin_argv, datalen); if (error != 0) goto err_exit; args->begin_argv[datalen] = '\0'; args->endp = args->begin_argv + datalen + 1; args->stringspace -= datalen + 1; /* * Traditional argument counting. Count the number of * null bytes. */ for (p = args->begin_argv; p < args->endp; ++p) if (*p == '\0') ++args->argc; } else { /* No argument buffer provided. */ args->endp = args->begin_argv; } /* Create new file descriptor table. */ kfds = malloc(fdslen * sizeof(int), M_TEMP, M_WAITOK); error = copyin(fds, kfds, fdslen * sizeof(int)); if (error != 0) { free(kfds, M_TEMP); goto err_exit; } error = fdcopy_remapped(ofdp, kfds, fdslen, &args->fdp); free(kfds, M_TEMP); if (error != 0) goto err_exit; return (0); err_exit: exec_free_args(args); return (error); } struct exec_args_kva { vm_offset_t addr; u_int gen; SLIST_ENTRY(exec_args_kva) next; }; DPCPU_DEFINE_STATIC(struct exec_args_kva *, exec_args_kva); static SLIST_HEAD(, exec_args_kva) exec_args_kva_freelist; static struct mtx exec_args_kva_mtx; static u_int exec_args_gen; static void exec_prealloc_args_kva(void *arg __unused) { struct exec_args_kva *argkva; u_int i; SLIST_INIT(&exec_args_kva_freelist); mtx_init(&exec_args_kva_mtx, "exec args kva", NULL, MTX_DEF); for (i = 0; i < exec_map_entries; i++) { argkva = malloc(sizeof(*argkva), M_PARGS, M_WAITOK); argkva->addr = kmap_alloc_wait(exec_map, exec_map_entry_size); argkva->gen = exec_args_gen; SLIST_INSERT_HEAD(&exec_args_kva_freelist, argkva, next); } } SYSINIT(exec_args_kva, SI_SUB_EXEC, SI_ORDER_ANY, exec_prealloc_args_kva, NULL); static vm_offset_t exec_alloc_args_kva(void **cookie) { struct exec_args_kva *argkva; argkva = (void *)atomic_readandclear_ptr( (uintptr_t *)DPCPU_PTR(exec_args_kva)); if (argkva == NULL) { mtx_lock(&exec_args_kva_mtx); while ((argkva = SLIST_FIRST(&exec_args_kva_freelist)) == NULL) (void)mtx_sleep(&exec_args_kva_freelist, &exec_args_kva_mtx, 0, "execkva", 0); SLIST_REMOVE_HEAD(&exec_args_kva_freelist, next); mtx_unlock(&exec_args_kva_mtx); } *(struct exec_args_kva **)cookie = argkva; return (argkva->addr); } static void exec_release_args_kva(struct exec_args_kva *argkva, u_int gen) { vm_offset_t base; base = argkva->addr; if (argkva->gen != gen) { (void)vm_map_madvise(exec_map, base, base + exec_map_entry_size, MADV_FREE); argkva->gen = gen; } if (!atomic_cmpset_ptr((uintptr_t *)DPCPU_PTR(exec_args_kva), (uintptr_t)NULL, (uintptr_t)argkva)) { mtx_lock(&exec_args_kva_mtx); SLIST_INSERT_HEAD(&exec_args_kva_freelist, argkva, next); wakeup_one(&exec_args_kva_freelist); mtx_unlock(&exec_args_kva_mtx); } } static void exec_free_args_kva(void *cookie) { exec_release_args_kva(cookie, exec_args_gen); } static void exec_args_kva_lowmem(void *arg __unused) { SLIST_HEAD(, exec_args_kva) head; struct exec_args_kva *argkva; u_int gen; int i; gen = atomic_fetchadd_int(&exec_args_gen, 1) + 1; /* * Force an madvise of each KVA range. Any currently allocated ranges * will have MADV_FREE applied once they are freed. */ SLIST_INIT(&head); mtx_lock(&exec_args_kva_mtx); SLIST_SWAP(&head, &exec_args_kva_freelist, exec_args_kva); mtx_unlock(&exec_args_kva_mtx); while ((argkva = SLIST_FIRST(&head)) != NULL) { SLIST_REMOVE_HEAD(&head, next); exec_release_args_kva(argkva, gen); } CPU_FOREACH(i) { argkva = (void *)atomic_readandclear_ptr( (uintptr_t *)DPCPU_ID_PTR(i, exec_args_kva)); if (argkva != NULL) exec_release_args_kva(argkva, gen); } } EVENTHANDLER_DEFINE(vm_lowmem, exec_args_kva_lowmem, NULL, EVENTHANDLER_PRI_ANY); /* * Allocate temporary demand-paged, zero-filled memory for the file name, * argument, and environment strings. */ int exec_alloc_args(struct image_args *args) { args->buf = (char *)exec_alloc_args_kva(&args->bufkva); return (0); } void exec_free_args(struct image_args *args) { if (args->buf != NULL) { exec_free_args_kva(args->bufkva); args->buf = NULL; } if (args->fname_buf != NULL) { free(args->fname_buf, M_TEMP); args->fname_buf = NULL; } if (args->fdp != NULL) fdescfree_remapped(args->fdp); } /* * A set to functions to fill struct image args. * * NOTE: exec_args_add_fname() must be called (possibly with a NULL * fname) before the other functions. All exec_args_add_arg() calls must * be made before any exec_args_add_env() calls. exec_args_adjust_args() * may be called any time after exec_args_add_fname(). * * exec_args_add_fname() - install path to be executed * exec_args_add_arg() - append an argument string * exec_args_add_env() - append an env string * exec_args_adjust_args() - adjust location of the argument list to * allow new arguments to be prepended */ int exec_args_add_fname(struct image_args *args, const char *fname, enum uio_seg segflg) { int error; size_t length; KASSERT(args->fname == NULL, ("fname already appended")); KASSERT(args->endp == NULL, ("already appending to args")); if (fname != NULL) { args->fname = args->buf; error = segflg == UIO_SYSSPACE ? copystr(fname, args->fname, PATH_MAX, &length) : copyinstr(fname, args->fname, PATH_MAX, &length); if (error != 0) return (error == ENAMETOOLONG ? E2BIG : error); } else length = 0; /* Set up for _arg_*()/_env_*() */ args->endp = args->buf + length; /* begin_argv must be set and kept updated */ args->begin_argv = args->endp; KASSERT(exec_map_entry_size - length >= ARG_MAX, ("too little space remaining for arguments %zu < %zu", exec_map_entry_size - length, (size_t)ARG_MAX)); args->stringspace = ARG_MAX; return (0); } static int exec_args_add_str(struct image_args *args, const char *str, enum uio_seg segflg, int *countp) { int error; size_t length; KASSERT(args->endp != NULL, ("endp not initialized")); KASSERT(args->begin_argv != NULL, ("begin_argp not initialized")); error = (segflg == UIO_SYSSPACE) ? copystr(str, args->endp, args->stringspace, &length) : copyinstr(str, args->endp, args->stringspace, &length); if (error != 0) return (error == ENAMETOOLONG ? E2BIG : error); args->stringspace -= length; args->endp += length; (*countp)++; return (0); } int exec_args_add_arg(struct image_args *args, const char *argp, enum uio_seg segflg) { KASSERT(args->envc == 0, ("appending args after env")); return (exec_args_add_str(args, argp, segflg, &args->argc)); } int exec_args_add_env(struct image_args *args, const char *envp, enum uio_seg segflg) { if (args->envc == 0) args->begin_envv = args->endp; return (exec_args_add_str(args, envp, segflg, &args->envc)); } int exec_args_adjust_args(struct image_args *args, size_t consume, ssize_t extend) { ssize_t offset; KASSERT(args->endp != NULL, ("endp not initialized")); KASSERT(args->begin_argv != NULL, ("begin_argp not initialized")); offset = extend - consume; if (args->stringspace < offset) return (E2BIG); memmove(args->begin_argv + extend, args->begin_argv + consume, args->endp - args->begin_argv + consume); if (args->envc > 0) args->begin_envv += offset; args->endp += offset; args->stringspace -= offset; return (0); } char * exec_args_get_begin_envv(struct image_args *args) { KASSERT(args->endp != NULL, ("endp not initialized")); if (args->envc > 0) return (args->begin_envv); return (args->endp); } /* * Copy strings out to the new process address space, constructing new arg * and env vector tables. Return a pointer to the base so that it can be used * as the initial stack pointer. */ register_t * exec_copyout_strings(struct image_params *imgp) { int argc, envc; char **vectp; char *stringp; uintptr_t destp; register_t *stack_base; struct ps_strings *arginfo; struct proc *p; size_t execpath_len; int szsigcode, szps; char canary[sizeof(long) * 8]; szps = sizeof(pagesizes[0]) * MAXPAGESIZES; /* * Calculate string base and vector table pointers. * Also deal with signal trampoline code for this exec type. */ if (imgp->execpath != NULL && imgp->auxargs != NULL) execpath_len = strlen(imgp->execpath) + 1; else execpath_len = 0; p = imgp->proc; szsigcode = 0; arginfo = (struct ps_strings *)p->p_sysent->sv_psstrings; if (p->p_sysent->sv_sigcode_base == 0) { if (p->p_sysent->sv_szsigcode != NULL) szsigcode = *(p->p_sysent->sv_szsigcode); } destp = (uintptr_t)arginfo; /* * install sigcode */ if (szsigcode != 0) { destp -= szsigcode; destp = rounddown2(destp, sizeof(void *)); copyout(p->p_sysent->sv_sigcode, (void *)destp, szsigcode); } /* * Copy the image path for the rtld. */ if (execpath_len != 0) { destp -= execpath_len; destp = rounddown2(destp, sizeof(void *)); imgp->execpathp = destp; copyout(imgp->execpath, (void *)destp, execpath_len); } /* * Prepare the canary for SSP. */ arc4rand(canary, sizeof(canary), 0); destp -= sizeof(canary); imgp->canary = destp; copyout(canary, (void *)destp, sizeof(canary)); imgp->canarylen = sizeof(canary); /* * Prepare the pagesizes array. */ destp -= szps; destp = rounddown2(destp, sizeof(void *)); imgp->pagesizes = destp; copyout(pagesizes, (void *)destp, szps); imgp->pagesizeslen = szps; destp -= ARG_MAX - imgp->args->stringspace; destp = rounddown2(destp, sizeof(void *)); vectp = (char **)destp; if (imgp->sysent->sv_stackgap != NULL) imgp->sysent->sv_stackgap(imgp, (u_long *)&vectp); if (imgp->auxargs) { /* * Allocate room on the stack for the ELF auxargs * array. It has up to AT_COUNT entries. */ vectp -= howmany(AT_COUNT * sizeof(Elf_Auxinfo), sizeof(*vectp)); } /* * Allocate room for the argv[] and env vectors including the * terminating NULL pointers. */ vectp -= imgp->args->argc + 1 + imgp->args->envc + 1; /* * vectp also becomes our initial stack base */ stack_base = (register_t *)vectp; stringp = imgp->args->begin_argv; argc = imgp->args->argc; envc = imgp->args->envc; /* * Copy out strings - arguments and environment. */ copyout(stringp, (void *)destp, ARG_MAX - imgp->args->stringspace); /* * Fill in "ps_strings" struct for ps, w, etc. */ suword(&arginfo->ps_argvstr, (long)(intptr_t)vectp); suword32(&arginfo->ps_nargvstr, argc); /* * Fill in argument portion of vector table. */ for (; argc > 0; --argc) { suword(vectp++, (long)(intptr_t)destp); while (*stringp++ != 0) destp++; destp++; } /* a null vector table pointer separates the argp's from the envp's */ suword(vectp++, 0); suword(&arginfo->ps_envstr, (long)(intptr_t)vectp); suword32(&arginfo->ps_nenvstr, envc); /* * Fill in environment portion of vector table. */ for (; envc > 0; --envc) { suword(vectp++, (long)(intptr_t)destp); while (*stringp++ != 0) destp++; destp++; } /* end of vector table is a null pointer */ suword(vectp, 0); return (stack_base); } /* * Check permissions of file to execute. * Called with imgp->vp locked. * Return 0 for success or error code on failure. */ int exec_check_permissions(struct image_params *imgp) { struct vnode *vp = imgp->vp; struct vattr *attr = imgp->attr; struct thread *td; int error; td = curthread; /* Get file attributes */ error = VOP_GETATTR(vp, attr, td->td_ucred); if (error) return (error); #ifdef MAC error = mac_vnode_check_exec(td->td_ucred, imgp->vp, imgp); if (error) return (error); #endif /* * 1) Check if file execution is disabled for the filesystem that * this file resides on. * 2) Ensure that at least one execute bit is on. Otherwise, a * privileged user will always succeed, and we don't want this * to happen unless the file really is executable. * 3) Ensure that the file is a regular file. */ if ((vp->v_mount->mnt_flag & MNT_NOEXEC) || (attr->va_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) == 0 || (attr->va_type != VREG)) return (EACCES); /* * Zero length files can't be exec'd */ if (attr->va_size == 0) return (ENOEXEC); /* * Check for execute permission to file based on current credentials. */ error = VOP_ACCESS(vp, VEXEC, td->td_ucred, td); if (error) return (error); /* * Check number of open-for-writes on the file and deny execution * if there are any. * * Add a text reference now so no one can write to the * executable while we're activating it. * * Remember if this was set before and unset it in case this is not * actually an executable image. */ error = VOP_SET_TEXT(vp); if (error != 0) return (error); imgp->textset = true; /* * Call filesystem specific open routine (which does nothing in the * general case). */ error = VOP_OPEN(vp, FREAD, td->td_ucred, td, NULL); if (error == 0) imgp->opened = 1; return (error); } /* * Exec handler registration */ int exec_register(const struct execsw *execsw_arg) { const struct execsw **es, **xs, **newexecsw; u_int count = 2; /* New slot and trailing NULL */ if (execsw) for (es = execsw; *es; es++) count++; newexecsw = malloc(count * sizeof(*es), M_TEMP, M_WAITOK); xs = newexecsw; if (execsw) for (es = execsw; *es; es++) *xs++ = *es; *xs++ = execsw_arg; *xs = NULL; if (execsw) free(execsw, M_TEMP); execsw = newexecsw; return (0); } int exec_unregister(const struct execsw *execsw_arg) { const struct execsw **es, **xs, **newexecsw; int count = 1; if (execsw == NULL) panic("unregister with no handlers left?\n"); for (es = execsw; *es; es++) { if (*es == execsw_arg) break; } if (*es == NULL) return (ENOENT); for (es = execsw; *es; es++) if (*es != execsw_arg) count++; newexecsw = malloc(count * sizeof(*es), M_TEMP, M_WAITOK); xs = newexecsw; for (es = execsw; *es; es++) if (*es != execsw_arg) *xs++ = *es; *xs = NULL; if (execsw) free(execsw, M_TEMP); execsw = newexecsw; return (0); } Index: head/sys/kern/uipc_shm.c =================================================================== --- head/sys/kern/uipc_shm.c (revision 353538) +++ head/sys/kern/uipc_shm.c (revision 353539) @@ -1,1505 +1,1505 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2006, 2011, 2016-2017 Robert N. M. Watson * All rights reserved. * * Portions of this software were developed by BAE Systems, the University of * Cambridge Computer Laboratory, and Memorial University under DARPA/AFRL * contract FA8650-15-C-7558 ("CADETS"), as part of the DARPA Transparent * Computing (TC) research program. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ /* * Support for shared swap-backed anonymous memory objects via * shm_open(2), shm_rename(2), and shm_unlink(2). * While most of the implementation is here, vm_mmap.c contains * mapping logic changes. * * posixshmcontrol(1) allows users to inspect the state of the memory * objects. Per-uid swap resource limit controls total amount of * memory that user can consume for anonymous objects, including * shared. */ #include __FBSDID("$FreeBSD$"); #include "opt_capsicum.h" #include "opt_ktrace.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include struct shm_mapping { char *sm_path; Fnv32_t sm_fnv; struct shmfd *sm_shmfd; LIST_ENTRY(shm_mapping) sm_link; }; static MALLOC_DEFINE(M_SHMFD, "shmfd", "shared memory file descriptor"); static LIST_HEAD(, shm_mapping) *shm_dictionary; static struct sx shm_dict_lock; static struct mtx shm_timestamp_lock; static u_long shm_hash; static struct unrhdr64 shm_ino_unr; static dev_t shm_dev_ino; #define SHM_HASH(fnv) (&shm_dictionary[(fnv) & shm_hash]) static void shm_init(void *arg); static void shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd); static struct shmfd *shm_lookup(char *path, Fnv32_t fnv); static int shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred); static int shm_dotruncate_locked(struct shmfd *shmfd, off_t length, void *rl_cookie); static fo_rdwr_t shm_read; static fo_rdwr_t shm_write; static fo_truncate_t shm_truncate; static fo_ioctl_t shm_ioctl; static fo_stat_t shm_stat; static fo_close_t shm_close; static fo_chmod_t shm_chmod; static fo_chown_t shm_chown; static fo_seek_t shm_seek; static fo_fill_kinfo_t shm_fill_kinfo; static fo_mmap_t shm_mmap; static fo_get_seals_t shm_get_seals; static fo_add_seals_t shm_add_seals; /* File descriptor operations. */ struct fileops shm_ops = { .fo_read = shm_read, .fo_write = shm_write, .fo_truncate = shm_truncate, .fo_ioctl = shm_ioctl, .fo_poll = invfo_poll, .fo_kqfilter = invfo_kqfilter, .fo_stat = shm_stat, .fo_close = shm_close, .fo_chmod = shm_chmod, .fo_chown = shm_chown, .fo_sendfile = vn_sendfile, .fo_seek = shm_seek, .fo_fill_kinfo = shm_fill_kinfo, .fo_mmap = shm_mmap, .fo_get_seals = shm_get_seals, .fo_add_seals = shm_add_seals, .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE }; FEATURE(posix_shm, "POSIX shared memory"); static int uiomove_object_page(vm_object_t obj, size_t len, struct uio *uio) { vm_page_t m; vm_pindex_t idx; size_t tlen; int error, offset, rv; idx = OFF_TO_IDX(uio->uio_offset); offset = uio->uio_offset & PAGE_MASK; tlen = MIN(PAGE_SIZE - offset, len); VM_OBJECT_WLOCK(obj); /* * Read I/O without either a corresponding resident page or swap * page: use zero_region. This is intended to avoid instantiating * pages on read from a sparse region. */ if (uio->uio_rw == UIO_READ && vm_page_lookup(obj, idx) == NULL && !vm_pager_has_page(obj, idx, NULL, NULL)) { VM_OBJECT_WUNLOCK(obj); return (uiomove(__DECONST(void *, zero_region), tlen, uio)); } /* * Parallel reads of the page content from disk are prevented * by exclusive busy. * * Although the tmpfs vnode lock is held here, it is * nonetheless safe to sleep waiting for a free page. The * pageout daemon does not need to acquire the tmpfs vnode * lock to page out tobj's pages because tobj is a OBJT_SWAP * type object. */ rv = vm_page_grab_valid(&m, obj, idx, VM_ALLOC_NORMAL | VM_ALLOC_WIRED | VM_ALLOC_NOBUSY); if (rv != VM_PAGER_OK) { VM_OBJECT_WUNLOCK(obj); printf("uiomove_object: vm_obj %p idx %jd pager error %d\n", obj, idx, rv); return (EIO); } VM_OBJECT_WUNLOCK(obj); error = uiomove_fromphys(&m, offset, tlen, uio); if (uio->uio_rw == UIO_WRITE && error == 0) { VM_OBJECT_WLOCK(obj); vm_page_dirty(m); vm_pager_page_unswapped(m); VM_OBJECT_WUNLOCK(obj); } vm_page_unwire(m, PQ_ACTIVE); return (error); } int uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio) { ssize_t resid; size_t len; int error; error = 0; while ((resid = uio->uio_resid) > 0) { if (obj_size <= uio->uio_offset) break; len = MIN(obj_size - uio->uio_offset, resid); if (len == 0) break; error = uiomove_object_page(obj, len, uio); if (error != 0 || resid == uio->uio_resid) break; } return (error); } static int shm_seek(struct file *fp, off_t offset, int whence, struct thread *td) { struct shmfd *shmfd; off_t foffset; int error; shmfd = fp->f_data; foffset = foffset_lock(fp, 0); error = 0; switch (whence) { case L_INCR: if (foffset < 0 || (offset > 0 && foffset > OFF_MAX - offset)) { error = EOVERFLOW; break; } offset += foffset; break; case L_XTND: if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) { error = EOVERFLOW; break; } offset += shmfd->shm_size; break; case L_SET: break; default: error = EINVAL; } if (error == 0) { if (offset < 0 || offset > shmfd->shm_size) error = EINVAL; else td->td_uretoff.tdu_off = offset; } foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0); return (error); } static int shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, struct thread *td) { struct shmfd *shmfd; void *rl_cookie; int error; shmfd = fp->f_data; #ifdef MAC error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd); if (error) return (error); #endif foffset_lock_uio(fp, uio, flags); rl_cookie = rangelock_rlock(&shmfd->shm_rl, uio->uio_offset, uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx); error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio); rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); foffset_unlock_uio(fp, uio, flags); return (error); } static int shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, struct thread *td) { struct shmfd *shmfd; void *rl_cookie; int error; shmfd = fp->f_data; #ifdef MAC error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd); if (error) return (error); #endif foffset_lock_uio(fp, uio, flags); if ((flags & FOF_OFFSET) == 0) { rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX, &shmfd->shm_mtx); } else { rl_cookie = rangelock_wlock(&shmfd->shm_rl, uio->uio_offset, uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx); } if ((shmfd->shm_seals & F_SEAL_WRITE) != 0) error = EPERM; else error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio); rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); foffset_unlock_uio(fp, uio, flags); return (error); } static int shm_truncate(struct file *fp, off_t length, struct ucred *active_cred, struct thread *td) { struct shmfd *shmfd; #ifdef MAC int error; #endif shmfd = fp->f_data; #ifdef MAC error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd); if (error) return (error); #endif return (shm_dotruncate(shmfd, length)); } int shm_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred, struct thread *td) { switch (com) { case FIONBIO: case FIOASYNC: /* * Allow fcntl(fd, F_SETFL, O_NONBLOCK) to work, * just like it would on an unlinked regular file */ return (0); default: return (ENOTTY); } } static int shm_stat(struct file *fp, struct stat *sb, struct ucred *active_cred, struct thread *td) { struct shmfd *shmfd; #ifdef MAC int error; #endif shmfd = fp->f_data; #ifdef MAC error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd); if (error) return (error); #endif /* * Attempt to return sanish values for fstat() on a memory file * descriptor. */ bzero(sb, sizeof(*sb)); sb->st_blksize = PAGE_SIZE; sb->st_size = shmfd->shm_size; sb->st_blocks = howmany(sb->st_size, sb->st_blksize); mtx_lock(&shm_timestamp_lock); sb->st_atim = shmfd->shm_atime; sb->st_ctim = shmfd->shm_ctime; sb->st_mtim = shmfd->shm_mtime; sb->st_birthtim = shmfd->shm_birthtime; sb->st_mode = S_IFREG | shmfd->shm_mode; /* XXX */ sb->st_uid = shmfd->shm_uid; sb->st_gid = shmfd->shm_gid; mtx_unlock(&shm_timestamp_lock); sb->st_dev = shm_dev_ino; sb->st_ino = shmfd->shm_ino; sb->st_nlink = shmfd->shm_object->ref_count; return (0); } static int shm_close(struct file *fp, struct thread *td) { struct shmfd *shmfd; shmfd = fp->f_data; fp->f_data = NULL; shm_drop(shmfd); return (0); } static int shm_dotruncate_locked(struct shmfd *shmfd, off_t length, void *rl_cookie) { vm_object_t object; vm_page_t m; vm_pindex_t idx, nobjsize; vm_ooffset_t delta; int base, rv; KASSERT(length >= 0, ("shm_dotruncate: length < 0")); object = shmfd->shm_object; VM_OBJECT_ASSERT_WLOCKED(object); rangelock_cookie_assert(rl_cookie, RA_WLOCKED); if (length == shmfd->shm_size) return (0); nobjsize = OFF_TO_IDX(length + PAGE_MASK); /* Are we shrinking? If so, trim the end. */ if (length < shmfd->shm_size) { if ((shmfd->shm_seals & F_SEAL_SHRINK) != 0) return (EPERM); /* * Disallow any requests to shrink the size if this * object is mapped into the kernel. */ if (shmfd->shm_kmappings > 0) return (EBUSY); /* * Zero the truncated part of the last page. */ base = length & PAGE_MASK; if (base != 0) { idx = OFF_TO_IDX(length); retry: m = vm_page_grab(object, idx, VM_ALLOC_NOCREAT); if (m != NULL) { - MPASS(m->valid == VM_PAGE_BITS_ALL); + MPASS(vm_page_all_valid(m)); } else if (vm_pager_has_page(object, idx, NULL, NULL)) { m = vm_page_alloc(object, idx, VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL); if (m == NULL) goto retry; rv = vm_pager_get_pages(object, &m, 1, NULL, NULL); if (rv == VM_PAGER_OK) { /* * Since the page was not resident, * and therefore not recently * accessed, immediately enqueue it * for asynchronous laundering. The * current operation is not regarded * as an access. */ vm_page_launder(m); } else { vm_page_free(m); VM_OBJECT_WUNLOCK(object); return (EIO); } } if (m != NULL) { pmap_zero_page_area(m, base, PAGE_SIZE - base); - KASSERT(m->valid == VM_PAGE_BITS_ALL, + KASSERT(vm_page_all_valid(m), ("shm_dotruncate: page %p is invalid", m)); vm_page_dirty(m); vm_page_xunbusy(m); vm_pager_page_unswapped(m); } } delta = IDX_TO_OFF(object->size - nobjsize); /* Toss in memory pages. */ if (nobjsize < object->size) vm_object_page_remove(object, nobjsize, object->size, 0); /* Toss pages from swap. */ if (object->type == OBJT_SWAP) swap_pager_freespace(object, nobjsize, delta); /* Free the swap accounted for shm */ swap_release_by_cred(delta, object->cred); object->charge -= delta; } else { if ((shmfd->shm_seals & F_SEAL_GROW) != 0) return (EPERM); /* Try to reserve additional swap space. */ delta = IDX_TO_OFF(nobjsize - object->size); if (!swap_reserve_by_cred(delta, object->cred)) return (ENOMEM); object->charge += delta; } shmfd->shm_size = length; mtx_lock(&shm_timestamp_lock); vfs_timestamp(&shmfd->shm_ctime); shmfd->shm_mtime = shmfd->shm_ctime; mtx_unlock(&shm_timestamp_lock); object->size = nobjsize; return (0); } int shm_dotruncate(struct shmfd *shmfd, off_t length) { void *rl_cookie; int error; rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX, &shmfd->shm_mtx); VM_OBJECT_WLOCK(shmfd->shm_object); error = shm_dotruncate_locked(shmfd, length, rl_cookie); VM_OBJECT_WUNLOCK(shmfd->shm_object); rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); return (error); } /* * shmfd object management including creation and reference counting * routines. */ struct shmfd * shm_alloc(struct ucred *ucred, mode_t mode) { struct shmfd *shmfd; shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO); shmfd->shm_size = 0; shmfd->shm_uid = ucred->cr_uid; shmfd->shm_gid = ucred->cr_gid; shmfd->shm_mode = mode; shmfd->shm_object = vm_pager_allocate(OBJT_SWAP, NULL, shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred); KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate")); shmfd->shm_object->pg_color = 0; VM_OBJECT_WLOCK(shmfd->shm_object); vm_object_clear_flag(shmfd->shm_object, OBJ_ONEMAPPING); vm_object_set_flag(shmfd->shm_object, OBJ_COLORED | OBJ_NOSPLIT); VM_OBJECT_WUNLOCK(shmfd->shm_object); vfs_timestamp(&shmfd->shm_birthtime); shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime = shmfd->shm_birthtime; shmfd->shm_ino = alloc_unr64(&shm_ino_unr); refcount_init(&shmfd->shm_refs, 1); mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF); rangelock_init(&shmfd->shm_rl); #ifdef MAC mac_posixshm_init(shmfd); mac_posixshm_create(ucred, shmfd); #endif return (shmfd); } struct shmfd * shm_hold(struct shmfd *shmfd) { refcount_acquire(&shmfd->shm_refs); return (shmfd); } void shm_drop(struct shmfd *shmfd) { if (refcount_release(&shmfd->shm_refs)) { #ifdef MAC mac_posixshm_destroy(shmfd); #endif rangelock_destroy(&shmfd->shm_rl); mtx_destroy(&shmfd->shm_mtx); vm_object_deallocate(shmfd->shm_object); free(shmfd, M_SHMFD); } } /* * Determine if the credentials have sufficient permissions for a * specified combination of FREAD and FWRITE. */ int shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags) { accmode_t accmode; int error; accmode = 0; if (flags & FREAD) accmode |= VREAD; if (flags & FWRITE) accmode |= VWRITE; mtx_lock(&shm_timestamp_lock); error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid, accmode, ucred, NULL); mtx_unlock(&shm_timestamp_lock); return (error); } /* * Dictionary management. We maintain an in-kernel dictionary to map * paths to shmfd objects. We use the FNV hash on the path to store * the mappings in a hash table. */ static void shm_init(void *arg) { mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF); sx_init(&shm_dict_lock, "shm dictionary"); shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash); new_unrhdr64(&shm_ino_unr, 1); shm_dev_ino = devfs_alloc_cdp_inode(); KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized")); } SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL); static struct shmfd * shm_lookup(char *path, Fnv32_t fnv) { struct shm_mapping *map; LIST_FOREACH(map, SHM_HASH(fnv), sm_link) { if (map->sm_fnv != fnv) continue; if (strcmp(map->sm_path, path) == 0) return (map->sm_shmfd); } return (NULL); } static void shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd) { struct shm_mapping *map; map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK); map->sm_path = path; map->sm_fnv = fnv; map->sm_shmfd = shm_hold(shmfd); shmfd->shm_path = path; LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link); } static int shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred) { struct shm_mapping *map; int error; LIST_FOREACH(map, SHM_HASH(fnv), sm_link) { if (map->sm_fnv != fnv) continue; if (strcmp(map->sm_path, path) == 0) { #ifdef MAC error = mac_posixshm_check_unlink(ucred, map->sm_shmfd); if (error) return (error); #endif error = shm_access(map->sm_shmfd, ucred, FREAD | FWRITE); if (error) return (error); map->sm_shmfd->shm_path = NULL; LIST_REMOVE(map, sm_link); shm_drop(map->sm_shmfd); free(map->sm_path, M_SHMFD); free(map, M_SHMFD); return (0); } } return (ENOENT); } int kern_shm_open(struct thread *td, const char *userpath, int flags, mode_t mode, struct filecaps *fcaps, int initial_seals) { struct filedesc *fdp; struct shmfd *shmfd; struct file *fp; char *path; const char *pr_path; void *rl_cookie; size_t pr_pathlen; Fnv32_t fnv; mode_t cmode; int fd, error; #ifdef CAPABILITY_MODE /* * shm_open(2) is only allowed for anonymous objects. */ if (IN_CAPABILITY_MODE(td) && (userpath != SHM_ANON)) return (ECAPMODE); #endif AUDIT_ARG_FFLAGS(flags); AUDIT_ARG_MODE(mode); if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR) return (EINVAL); if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0) return (EINVAL); /* * Currently only F_SEAL_SEAL may be set when creating or opening shmfd. * If the decision is made later to allow additional seals, care must be * taken below to ensure that the seals are properly set if the shmfd * already existed -- this currently assumes that only F_SEAL_SEAL can * be set and doesn't take further precautions to ensure the validity of * the seals being added with respect to current mappings. */ if ((initial_seals & ~F_SEAL_SEAL) != 0) return (EINVAL); fdp = td->td_proc->p_fd; cmode = (mode & ~fdp->fd_cmask) & ACCESSPERMS; /* * shm_open(2) created shm should always have O_CLOEXEC set, as mandated * by POSIX. We allow it to be unset here so that an in-kernel * interface may be written as a thin layer around shm, optionally not * setting CLOEXEC. For shm_open(2), O_CLOEXEC is set unconditionally * in sys_shm_open() to keep this implementation compliant. */ error = falloc_caps(td, &fp, &fd, flags & O_CLOEXEC, fcaps); if (error) return (error); /* A SHM_ANON path pointer creates an anonymous object. */ if (userpath == SHM_ANON) { /* A read-only anonymous object is pointless. */ if ((flags & O_ACCMODE) == O_RDONLY) { fdclose(td, fp, fd); fdrop(fp, td); return (EINVAL); } shmfd = shm_alloc(td->td_ucred, cmode); shmfd->shm_seals = initial_seals; } else { path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK); pr_path = td->td_ucred->cr_prison->pr_path; /* Construct a full pathname for jailed callers. */ pr_pathlen = strcmp(pr_path, "/") == 0 ? 0 : strlcpy(path, pr_path, MAXPATHLEN); error = copyinstr(userpath, path + pr_pathlen, MAXPATHLEN - pr_pathlen, NULL); #ifdef KTRACE if (error == 0 && KTRPOINT(curthread, KTR_NAMEI)) ktrnamei(path); #endif /* Require paths to start with a '/' character. */ if (error == 0 && path[pr_pathlen] != '/') error = EINVAL; if (error) { fdclose(td, fp, fd); fdrop(fp, td); free(path, M_SHMFD); return (error); } AUDIT_ARG_UPATH1_CANON(path); fnv = fnv_32_str(path, FNV1_32_INIT); sx_xlock(&shm_dict_lock); shmfd = shm_lookup(path, fnv); if (shmfd == NULL) { /* Object does not yet exist, create it if requested. */ if (flags & O_CREAT) { #ifdef MAC error = mac_posixshm_check_create(td->td_ucred, path); if (error == 0) { #endif shmfd = shm_alloc(td->td_ucred, cmode); shmfd->shm_seals = initial_seals; shm_insert(path, fnv, shmfd); #ifdef MAC } #endif } else { free(path, M_SHMFD); error = ENOENT; } } else { rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX, &shmfd->shm_mtx); /* * kern_shm_open() likely shouldn't ever error out on * trying to set a seal that already exists, unlike * F_ADD_SEALS. This would break terribly as * shm_open(2) actually sets F_SEAL_SEAL to maintain * historical behavior where the underlying file could * not be sealed. */ initial_seals &= ~shmfd->shm_seals; /* * Object already exists, obtain a new * reference if requested and permitted. */ free(path, M_SHMFD); /* * initial_seals can't set additional seals if we've * already been set F_SEAL_SEAL. If F_SEAL_SEAL is set, * then we've already removed that one from * initial_seals. This is currently redundant as we * only allow setting F_SEAL_SEAL at creation time, but * it's cheap to check and decreases the effort required * to allow additional seals. */ if ((shmfd->shm_seals & F_SEAL_SEAL) != 0 && initial_seals != 0) error = EPERM; else if ((flags & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL)) error = EEXIST; else { #ifdef MAC error = mac_posixshm_check_open(td->td_ucred, shmfd, FFLAGS(flags & O_ACCMODE)); if (error == 0) #endif error = shm_access(shmfd, td->td_ucred, FFLAGS(flags & O_ACCMODE)); } /* * Truncate the file back to zero length if * O_TRUNC was specified and the object was * opened with read/write. */ if (error == 0 && (flags & (O_ACCMODE | O_TRUNC)) == (O_RDWR | O_TRUNC)) { VM_OBJECT_WLOCK(shmfd->shm_object); #ifdef MAC error = mac_posixshm_check_truncate( td->td_ucred, fp->f_cred, shmfd); if (error == 0) #endif error = shm_dotruncate_locked(shmfd, 0, rl_cookie); VM_OBJECT_WUNLOCK(shmfd->shm_object); } if (error == 0) { /* * Currently we only allow F_SEAL_SEAL to be * set initially. As noted above, this would * need to be reworked should that change. */ shmfd->shm_seals |= initial_seals; shm_hold(shmfd); } rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); } sx_xunlock(&shm_dict_lock); if (error) { fdclose(td, fp, fd); fdrop(fp, td); return (error); } } finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops); td->td_retval[0] = fd; fdrop(fp, td); return (0); } /* System calls. */ #ifdef COMPAT_FREEBSD12 int freebsd12_shm_open(struct thread *td, struct freebsd12_shm_open_args *uap) { return (kern_shm_open(td, uap->path, uap->flags | O_CLOEXEC, uap->mode, NULL, F_SEAL_SEAL)); } #endif int sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap) { char *path; const char *pr_path; size_t pr_pathlen; Fnv32_t fnv; int error; path = malloc(MAXPATHLEN, M_TEMP, M_WAITOK); pr_path = td->td_ucred->cr_prison->pr_path; pr_pathlen = strcmp(pr_path, "/") == 0 ? 0 : strlcpy(path, pr_path, MAXPATHLEN); error = copyinstr(uap->path, path + pr_pathlen, MAXPATHLEN - pr_pathlen, NULL); if (error) { free(path, M_TEMP); return (error); } #ifdef KTRACE if (KTRPOINT(curthread, KTR_NAMEI)) ktrnamei(path); #endif AUDIT_ARG_UPATH1_CANON(path); fnv = fnv_32_str(path, FNV1_32_INIT); sx_xlock(&shm_dict_lock); error = shm_remove(path, fnv, td->td_ucred); sx_xunlock(&shm_dict_lock); free(path, M_TEMP); return (error); } int sys_shm_rename(struct thread *td, struct shm_rename_args *uap) { char *path_from = NULL, *path_to = NULL; Fnv32_t fnv_from, fnv_to; struct shmfd *fd_from; struct shmfd *fd_to; int error; int flags; flags = uap->flags; /* * Make sure the user passed only valid flags. * If you add a new flag, please add a new term here. */ if ((flags & ~( SHM_RENAME_NOREPLACE | SHM_RENAME_EXCHANGE )) != 0) { error = EINVAL; goto out; } /* * EXCHANGE and NOREPLACE don't quite make sense together. Let's * force the user to choose one or the other. */ if ((flags & SHM_RENAME_NOREPLACE) != 0 && (flags & SHM_RENAME_EXCHANGE) != 0) { error = EINVAL; goto out; } /* * Malloc zone M_SHMFD, since this path may end up freed later from * M_SHMFD if we end up doing an insert. */ path_from = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK); error = copyinstr(uap->path_from, path_from, MAXPATHLEN, NULL); if (error) goto out; path_to = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK); error = copyinstr(uap->path_to, path_to, MAXPATHLEN, NULL); if (error) goto out; /* Rename with from/to equal is a no-op */ if (strncmp(path_from, path_to, MAXPATHLEN) == 0) goto out; fnv_from = fnv_32_str(path_from, FNV1_32_INIT); fnv_to = fnv_32_str(path_to, FNV1_32_INIT); sx_xlock(&shm_dict_lock); fd_from = shm_lookup(path_from, fnv_from); if (fd_from == NULL) { sx_xunlock(&shm_dict_lock); error = ENOENT; goto out; } fd_to = shm_lookup(path_to, fnv_to); if ((flags & SHM_RENAME_NOREPLACE) != 0 && fd_to != NULL) { sx_xunlock(&shm_dict_lock); error = EEXIST; goto out; } /* * Unconditionally prevents shm_remove from invalidating the 'from' * shm's state. */ shm_hold(fd_from); error = shm_remove(path_from, fnv_from, td->td_ucred); /* * One of my assumptions failed if ENOENT (e.g. locking didn't * protect us) */ KASSERT(error != ENOENT, ("Our shm disappeared during shm_rename: %s", path_from)); if (error) { shm_drop(fd_from); sx_xunlock(&shm_dict_lock); goto out; } /* * If we are exchanging, we need to ensure the shm_remove below * doesn't invalidate the dest shm's state. */ if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) shm_hold(fd_to); /* * NOTE: if path_to is not already in the hash, c'est la vie; * it simply means we have nothing already at path_to to unlink. * That is the ENOENT case. * * If we somehow don't have access to unlink this guy, but * did for the shm at path_from, then relink the shm to path_from * and abort with EACCES. * * All other errors: that is weird; let's relink and abort the * operation. */ error = shm_remove(path_to, fnv_to, td->td_ucred); if (error && error != ENOENT) { shm_insert(path_from, fnv_from, fd_from); shm_drop(fd_from); /* Don't free path_from now, since the hash references it */ path_from = NULL; sx_xunlock(&shm_dict_lock); goto out; } shm_insert(path_to, fnv_to, fd_from); /* Don't free path_to now, since the hash references it */ path_to = NULL; /* We kept a ref when we removed, and incremented again in insert */ shm_drop(fd_from); #ifdef DEBUG KASSERT(fd_from->shm_refs > 0, ("Expected >0 refs; got: %d\n", fd_from->shm_refs)); #endif if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) { shm_insert(path_from, fnv_from, fd_to); path_from = NULL; shm_drop(fd_to); #ifdef DEBUG KASSERT(fd_to->shm_refs > 0, ("Expected >0 refs; got: %d\n", fd_to->shm_refs)); #endif } error = 0; sx_xunlock(&shm_dict_lock); out: if (path_from != NULL) free(path_from, M_SHMFD); if (path_to != NULL) free(path_to, M_SHMFD); return(error); } int shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize, vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff, struct thread *td) { struct shmfd *shmfd; vm_prot_t maxprot; int error; bool writecnt; void *rl_cookie; shmfd = fp->f_data; maxprot = VM_PROT_NONE; rl_cookie = rangelock_rlock(&shmfd->shm_rl, 0, objsize, &shmfd->shm_mtx); /* FREAD should always be set. */ if ((fp->f_flag & FREAD) != 0) maxprot |= VM_PROT_EXECUTE | VM_PROT_READ; if ((fp->f_flag & FWRITE) != 0) maxprot |= VM_PROT_WRITE; writecnt = (flags & MAP_SHARED) != 0 && (prot & VM_PROT_WRITE) != 0; if (writecnt && (shmfd->shm_seals & F_SEAL_WRITE) != 0) { error = EPERM; goto out; } /* Don't permit shared writable mappings on read-only descriptors. */ if (writecnt && (maxprot & VM_PROT_WRITE) == 0) { error = EACCES; goto out; } maxprot &= cap_maxprot; /* See comment in vn_mmap(). */ if ( #ifdef _LP64 objsize > OFF_MAX || #endif foff < 0 || foff > OFF_MAX - objsize) { error = EINVAL; goto out; } #ifdef MAC error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags); if (error != 0) goto out; #endif mtx_lock(&shm_timestamp_lock); vfs_timestamp(&shmfd->shm_atime); mtx_unlock(&shm_timestamp_lock); vm_object_reference(shmfd->shm_object); if (writecnt) vm_pager_update_writecount(shmfd->shm_object, 0, objsize); error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags, shmfd->shm_object, foff, writecnt, td); if (error != 0) { if (writecnt) vm_pager_release_writecount(shmfd->shm_object, 0, objsize); vm_object_deallocate(shmfd->shm_object); } out: rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); return (error); } static int shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td) { struct shmfd *shmfd; int error; error = 0; shmfd = fp->f_data; mtx_lock(&shm_timestamp_lock); /* * SUSv4 says that x bits of permission need not be affected. * Be consistent with our shm_open there. */ #ifdef MAC error = mac_posixshm_check_setmode(active_cred, shmfd, mode); if (error != 0) goto out; #endif error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid, VADMIN, active_cred, NULL); if (error != 0) goto out; shmfd->shm_mode = mode & ACCESSPERMS; out: mtx_unlock(&shm_timestamp_lock); return (error); } static int shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred, struct thread *td) { struct shmfd *shmfd; int error; error = 0; shmfd = fp->f_data; mtx_lock(&shm_timestamp_lock); #ifdef MAC error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid); if (error != 0) goto out; #endif if (uid == (uid_t)-1) uid = shmfd->shm_uid; if (gid == (gid_t)-1) gid = shmfd->shm_gid; if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) || (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) && (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN))) goto out; shmfd->shm_uid = uid; shmfd->shm_gid = gid; out: mtx_unlock(&shm_timestamp_lock); return (error); } /* * Helper routines to allow the backing object of a shared memory file * descriptor to be mapped in the kernel. */ int shm_map(struct file *fp, size_t size, off_t offset, void **memp) { struct shmfd *shmfd; vm_offset_t kva, ofs; vm_object_t obj; int rv; if (fp->f_type != DTYPE_SHM) return (EINVAL); shmfd = fp->f_data; obj = shmfd->shm_object; VM_OBJECT_WLOCK(obj); /* * XXXRW: This validation is probably insufficient, and subject to * sign errors. It should be fixed. */ if (offset >= shmfd->shm_size || offset + size > round_page(shmfd->shm_size)) { VM_OBJECT_WUNLOCK(obj); return (EINVAL); } shmfd->shm_kmappings++; vm_object_reference_locked(obj); VM_OBJECT_WUNLOCK(obj); /* Map the object into the kernel_map and wire it. */ kva = vm_map_min(kernel_map); ofs = offset & PAGE_MASK; offset = trunc_page(offset); size = round_page(size + ofs); rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0, VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE, VM_PROT_READ | VM_PROT_WRITE, 0); if (rv == KERN_SUCCESS) { rv = vm_map_wire(kernel_map, kva, kva + size, VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES); if (rv == KERN_SUCCESS) { *memp = (void *)(kva + ofs); return (0); } vm_map_remove(kernel_map, kva, kva + size); } else vm_object_deallocate(obj); /* On failure, drop our mapping reference. */ VM_OBJECT_WLOCK(obj); shmfd->shm_kmappings--; VM_OBJECT_WUNLOCK(obj); return (vm_mmap_to_errno(rv)); } /* * We require the caller to unmap the entire entry. This allows us to * safely decrement shm_kmappings when a mapping is removed. */ int shm_unmap(struct file *fp, void *mem, size_t size) { struct shmfd *shmfd; vm_map_entry_t entry; vm_offset_t kva, ofs; vm_object_t obj; vm_pindex_t pindex; vm_prot_t prot; boolean_t wired; vm_map_t map; int rv; if (fp->f_type != DTYPE_SHM) return (EINVAL); shmfd = fp->f_data; kva = (vm_offset_t)mem; ofs = kva & PAGE_MASK; kva = trunc_page(kva); size = round_page(size + ofs); map = kernel_map; rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry, &obj, &pindex, &prot, &wired); if (rv != KERN_SUCCESS) return (EINVAL); if (entry->start != kva || entry->end != kva + size) { vm_map_lookup_done(map, entry); return (EINVAL); } vm_map_lookup_done(map, entry); if (obj != shmfd->shm_object) return (EINVAL); vm_map_remove(map, kva, kva + size); VM_OBJECT_WLOCK(obj); KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped")); shmfd->shm_kmappings--; VM_OBJECT_WUNLOCK(obj); return (0); } static int shm_fill_kinfo_locked(struct shmfd *shmfd, struct kinfo_file *kif, bool list) { const char *path, *pr_path; size_t pr_pathlen; bool visible; sx_assert(&shm_dict_lock, SA_LOCKED); kif->kf_type = KF_TYPE_SHM; kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode; kif->kf_un.kf_file.kf_file_size = shmfd->shm_size; if (shmfd->shm_path != NULL) { if (shmfd->shm_path != NULL) { path = shmfd->shm_path; pr_path = curthread->td_ucred->cr_prison->pr_path; if (strcmp(pr_path, "/") != 0) { /* Return the jail-rooted pathname. */ pr_pathlen = strlen(pr_path); visible = strncmp(path, pr_path, pr_pathlen) == 0 && path[pr_pathlen] == '/'; if (list && !visible) return (EPERM); if (visible) path += pr_pathlen; } strlcpy(kif->kf_path, path, sizeof(kif->kf_path)); } } return (0); } static int shm_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp __unused) { int res; sx_slock(&shm_dict_lock); res = shm_fill_kinfo_locked(fp->f_data, kif, false); sx_sunlock(&shm_dict_lock); return (res); } static int shm_add_seals(struct file *fp, int seals) { struct shmfd *shmfd; void *rl_cookie; vm_ooffset_t writemappings; int error, nseals; error = 0; shmfd = fp->f_data; rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX, &shmfd->shm_mtx); /* Even already-set seals should result in EPERM. */ if ((shmfd->shm_seals & F_SEAL_SEAL) != 0) { error = EPERM; goto out; } nseals = seals & ~shmfd->shm_seals; if ((nseals & F_SEAL_WRITE) != 0) { /* * The rangelock above prevents writable mappings from being * added after we've started applying seals. The RLOCK here * is to avoid torn reads on ILP32 arches as unmapping/reducing * writemappings will be done without a rangelock. */ VM_OBJECT_RLOCK(shmfd->shm_object); writemappings = shmfd->shm_object->un_pager.swp.writemappings; VM_OBJECT_RUNLOCK(shmfd->shm_object); /* kmappings are also writable */ if (writemappings > 0) { error = EBUSY; goto out; } } shmfd->shm_seals |= nseals; out: rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); return (error); } static int shm_get_seals(struct file *fp, int *seals) { struct shmfd *shmfd; shmfd = fp->f_data; *seals = shmfd->shm_seals; return (0); } static int sysctl_posix_shm_list(SYSCTL_HANDLER_ARGS) { struct shm_mapping *shmm; struct sbuf sb; struct kinfo_file kif; u_long i; ssize_t curlen; int error, error2; sbuf_new_for_sysctl(&sb, NULL, sizeof(struct kinfo_file) * 5, req); sbuf_clear_flags(&sb, SBUF_INCLUDENUL); curlen = 0; error = 0; sx_slock(&shm_dict_lock); for (i = 0; i < shm_hash + 1; i++) { LIST_FOREACH(shmm, &shm_dictionary[i], sm_link) { error = shm_fill_kinfo_locked(shmm->sm_shmfd, &kif, true); if (error == EPERM) continue; if (error != 0) break; pack_kinfo(&kif); if (req->oldptr != NULL && kif.kf_structsize + curlen > req->oldlen) break; error = sbuf_bcat(&sb, &kif, kif.kf_structsize) == 0 ? 0 : ENOMEM; if (error != 0) break; curlen += kif.kf_structsize; } } sx_sunlock(&shm_dict_lock); error2 = sbuf_finish(&sb); sbuf_delete(&sb); return (error != 0 ? error : error2); } SYSCTL_PROC(_kern_ipc, OID_AUTO, posix_shm_list, CTLFLAG_RD | CTLFLAG_MPSAFE | CTLTYPE_OPAQUE, NULL, 0, sysctl_posix_shm_list, "", "POSIX SHM list"); int kern_shm_open2(struct thread *td, const char *path, int flags, mode_t mode, int shmflags, const char *name __unused) { int initial_seals; if ((shmflags & ~SHM_ALLOW_SEALING) != 0) return (EINVAL); initial_seals = F_SEAL_SEAL; if ((shmflags & SHM_ALLOW_SEALING) != 0) initial_seals &= ~F_SEAL_SEAL; return (kern_shm_open(td, path, flags, mode, NULL, initial_seals)); } /* * This version of the shm_open() interface leaves CLOEXEC behavior up to the * caller, and libc will enforce it for the traditional shm_open() call. This * allows other consumers, like memfd_create(), to opt-in for CLOEXEC. This * interface also includes a 'name' argument that is currently unused, but could * potentially be exported later via some interface for debugging purposes. * From the kernel's perspective, it is optional. Individual consumers like * memfd_create() may require it in order to be compatible with other systems * implementing the same function. */ int sys_shm_open2(struct thread *td, struct shm_open2_args *uap) { return (kern_shm_open2(td, uap->path, uap->flags, uap->mode, uap->shmflags, uap->name)); } Index: head/sys/kern/vfs_bio.c =================================================================== --- head/sys/kern/vfs_bio.c (revision 353538) +++ head/sys/kern/vfs_bio.c (revision 353539) @@ -1,5463 +1,5480 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2004 Poul-Henning Kamp * Copyright (c) 1994,1997 John S. Dyson * Copyright (c) 2013 The FreeBSD Foundation * All rights reserved. * * Portions of this software were developed by Konstantin Belousov * under sponsorship from the FreeBSD Foundation. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ /* * this file contains a new buffer I/O scheme implementing a coherent * VM object and buffer cache scheme. Pains have been taken to make * sure that the performance degradation associated with schemes such * as this is not realized. * * Author: John S. Dyson * Significant help during the development and debugging phases * had been provided by David Greenman, also of the FreeBSD core team. * * see man buf(9) for more info. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer"); struct bio_ops bioops; /* I/O operation notification */ struct buf_ops buf_ops_bio = { .bop_name = "buf_ops_bio", .bop_write = bufwrite, .bop_strategy = bufstrategy, .bop_sync = bufsync, .bop_bdflush = bufbdflush, }; struct bufqueue { struct mtx_padalign bq_lock; TAILQ_HEAD(, buf) bq_queue; uint8_t bq_index; uint16_t bq_subqueue; int bq_len; } __aligned(CACHE_LINE_SIZE); #define BQ_LOCKPTR(bq) (&(bq)->bq_lock) #define BQ_LOCK(bq) mtx_lock(BQ_LOCKPTR((bq))) #define BQ_UNLOCK(bq) mtx_unlock(BQ_LOCKPTR((bq))) #define BQ_ASSERT_LOCKED(bq) mtx_assert(BQ_LOCKPTR((bq)), MA_OWNED) struct bufdomain { struct bufqueue bd_subq[MAXCPU + 1]; /* Per-cpu sub queues + global */ struct bufqueue bd_dirtyq; struct bufqueue *bd_cleanq; struct mtx_padalign bd_run_lock; /* Constants */ long bd_maxbufspace; long bd_hibufspace; long bd_lobufspace; long bd_bufspacethresh; int bd_hifreebuffers; int bd_lofreebuffers; int bd_hidirtybuffers; int bd_lodirtybuffers; int bd_dirtybufthresh; int bd_lim; /* atomics */ int bd_wanted; int __aligned(CACHE_LINE_SIZE) bd_numdirtybuffers; int __aligned(CACHE_LINE_SIZE) bd_running; long __aligned(CACHE_LINE_SIZE) bd_bufspace; int __aligned(CACHE_LINE_SIZE) bd_freebuffers; } __aligned(CACHE_LINE_SIZE); #define BD_LOCKPTR(bd) (&(bd)->bd_cleanq->bq_lock) #define BD_LOCK(bd) mtx_lock(BD_LOCKPTR((bd))) #define BD_UNLOCK(bd) mtx_unlock(BD_LOCKPTR((bd))) #define BD_ASSERT_LOCKED(bd) mtx_assert(BD_LOCKPTR((bd)), MA_OWNED) #define BD_RUN_LOCKPTR(bd) (&(bd)->bd_run_lock) #define BD_RUN_LOCK(bd) mtx_lock(BD_RUN_LOCKPTR((bd))) #define BD_RUN_UNLOCK(bd) mtx_unlock(BD_RUN_LOCKPTR((bd))) #define BD_DOMAIN(bd) (bd - bdomain) static struct buf *buf; /* buffer header pool */ extern struct buf *swbuf; /* Swap buffer header pool. */ caddr_t unmapped_buf; /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */ struct proc *bufdaemonproc; static int inmem(struct vnode *vp, daddr_t blkno); static void vm_hold_free_pages(struct buf *bp, int newbsize); static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to); static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m); static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m); static void vfs_clean_pages_dirty_buf(struct buf *bp); static void vfs_setdirty_locked_object(struct buf *bp); static void vfs_vmio_invalidate(struct buf *bp); static void vfs_vmio_truncate(struct buf *bp, int npages); static void vfs_vmio_extend(struct buf *bp, int npages, int size); static int vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno); static void breada(struct vnode *, daddr_t *, int *, int, struct ucred *, int, void (*)(struct buf *)); static int buf_flush(struct vnode *vp, struct bufdomain *, int); static int flushbufqueues(struct vnode *, struct bufdomain *, int, int); static void buf_daemon(void); static __inline void bd_wakeup(void); static int sysctl_runningspace(SYSCTL_HANDLER_ARGS); static void bufkva_reclaim(vmem_t *, int); static void bufkva_free(struct buf *); static int buf_import(void *, void **, int, int, int); static void buf_release(void *, void **, int); static void maxbcachebuf_adjust(void); static inline struct bufdomain *bufdomain(struct buf *); static void bq_remove(struct bufqueue *bq, struct buf *bp); static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock); static int buf_recycle(struct bufdomain *, bool kva); static void bq_init(struct bufqueue *bq, int qindex, int cpu, const char *lockname); static void bd_init(struct bufdomain *bd); static int bd_flushall(struct bufdomain *bd); static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS); static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS); static int sysctl_bufspace(SYSCTL_HANDLER_ARGS); int vmiodirenable = TRUE; SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, "Use the VM system for directory writes"); long runningbufspace; SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, "Amount of presently outstanding async buffer io"); SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_bufspace, "L", "Physical memory used for buffers"); static counter_u64_t bufkvaspace; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace, "Kernel virtual memory used for buffers"); static long maxbufspace; SYSCTL_PROC(_vfs, OID_AUTO, maxbufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &maxbufspace, __offsetof(struct bufdomain, bd_maxbufspace), sysctl_bufdomain_long, "L", "Maximum allowed value of bufspace (including metadata)"); static long bufmallocspace; SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, "Amount of malloced memory for buffers"); static long maxbufmallocspace; SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, "Maximum amount of malloced memory for buffers"); static long lobufspace; SYSCTL_PROC(_vfs, OID_AUTO, lobufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lobufspace, __offsetof(struct bufdomain, bd_lobufspace), sysctl_bufdomain_long, "L", "Minimum amount of buffers we want to have"); long hibufspace; SYSCTL_PROC(_vfs, OID_AUTO, hibufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hibufspace, __offsetof(struct bufdomain, bd_hibufspace), sysctl_bufdomain_long, "L", "Maximum allowed value of bufspace (excluding metadata)"); long bufspacethresh; SYSCTL_PROC(_vfs, OID_AUTO, bufspacethresh, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &bufspacethresh, __offsetof(struct bufdomain, bd_bufspacethresh), sysctl_bufdomain_long, "L", "Bufspace consumed before waking the daemon to free some"); static counter_u64_t buffreekvacnt; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, "Number of times we have freed the KVA space from some buffer"); static counter_u64_t bufdefragcnt; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, "Number of times we have had to repeat buffer allocation to defragment"); static long lorunningspace; SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE | CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L", "Minimum preferred space used for in-progress I/O"); static long hirunningspace; SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE | CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L", "Maximum amount of space to use for in-progress I/O"); int dirtybufferflushes; SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); int bdwriteskip; SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip, 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk"); int altbufferflushes; SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW | CTLFLAG_STATS, &altbufferflushes, 0, "Number of fsync flushes to limit dirty buffers"); static int recursiveflushes; SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW | CTLFLAG_STATS, &recursiveflushes, 0, "Number of flushes skipped due to being recursive"); static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS); SYSCTL_PROC(_vfs, OID_AUTO, numdirtybuffers, CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_numdirtybuffers, "I", "Number of buffers that are dirty (has unwritten changes) at the moment"); static int lodirtybuffers; SYSCTL_PROC(_vfs, OID_AUTO, lodirtybuffers, CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lodirtybuffers, __offsetof(struct bufdomain, bd_lodirtybuffers), sysctl_bufdomain_int, "I", "How many buffers we want to have free before bufdaemon can sleep"); static int hidirtybuffers; SYSCTL_PROC(_vfs, OID_AUTO, hidirtybuffers, CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hidirtybuffers, __offsetof(struct bufdomain, bd_hidirtybuffers), sysctl_bufdomain_int, "I", "When the number of dirty buffers is considered severe"); int dirtybufthresh; SYSCTL_PROC(_vfs, OID_AUTO, dirtybufthresh, CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &dirtybufthresh, __offsetof(struct bufdomain, bd_dirtybufthresh), sysctl_bufdomain_int, "I", "Number of bdwrite to bawrite conversions to clear dirty buffers"); static int numfreebuffers; SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, "Number of free buffers"); static int lofreebuffers; SYSCTL_PROC(_vfs, OID_AUTO, lofreebuffers, CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lofreebuffers, __offsetof(struct bufdomain, bd_lofreebuffers), sysctl_bufdomain_int, "I", "Target number of free buffers"); static int hifreebuffers; SYSCTL_PROC(_vfs, OID_AUTO, hifreebuffers, CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hifreebuffers, __offsetof(struct bufdomain, bd_hifreebuffers), sysctl_bufdomain_int, "I", "Threshold for clean buffer recycling"); static counter_u64_t getnewbufcalls; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, "Number of calls to getnewbuf"); static counter_u64_t getnewbufrestarts; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts, "Number of times getnewbuf has had to restart a buffer acquisition"); static counter_u64_t mappingrestarts; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RD, &mappingrestarts, "Number of times getblk has had to restart a buffer mapping for " "unmapped buffer"); static counter_u64_t numbufallocfails; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW, &numbufallocfails, "Number of times buffer allocations failed"); static int flushbufqtarget = 100; SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0, "Amount of work to do in flushbufqueues when helping bufdaemon"); static counter_u64_t notbufdflushes; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, ¬bufdflushes, "Number of dirty buffer flushes done by the bufdaemon helpers"); static long barrierwrites; SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW | CTLFLAG_STATS, &barrierwrites, 0, "Number of barrier writes"); SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD, &unmapped_buf_allowed, 0, "Permit the use of the unmapped i/o"); int maxbcachebuf = MAXBCACHEBUF; SYSCTL_INT(_vfs, OID_AUTO, maxbcachebuf, CTLFLAG_RDTUN, &maxbcachebuf, 0, "Maximum size of a buffer cache block"); /* * This lock synchronizes access to bd_request. */ static struct mtx_padalign __exclusive_cache_line bdlock; /* * This lock protects the runningbufreq and synchronizes runningbufwakeup and * waitrunningbufspace(). */ static struct mtx_padalign __exclusive_cache_line rbreqlock; /* * Lock that protects bdirtywait. */ static struct mtx_padalign __exclusive_cache_line bdirtylock; /* * Wakeup point for bufdaemon, as well as indicator of whether it is already * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it * is idling. */ static int bd_request; /* * Request for the buf daemon to write more buffers than is indicated by * lodirtybuf. This may be necessary to push out excess dependencies or * defragment the address space where a simple count of the number of dirty * buffers is insufficient to characterize the demand for flushing them. */ static int bd_speedupreq; /* * Synchronization (sleep/wakeup) variable for active buffer space requests. * Set when wait starts, cleared prior to wakeup(). * Used in runningbufwakeup() and waitrunningbufspace(). */ static int runningbufreq; /* * Synchronization for bwillwrite() waiters. */ static int bdirtywait; /* * Definitions for the buffer free lists. */ #define QUEUE_NONE 0 /* on no queue */ #define QUEUE_EMPTY 1 /* empty buffer headers */ #define QUEUE_DIRTY 2 /* B_DELWRI buffers */ #define QUEUE_CLEAN 3 /* non-B_DELWRI buffers */ #define QUEUE_SENTINEL 4 /* not an queue index, but mark for sentinel */ /* Maximum number of buffer domains. */ #define BUF_DOMAINS 8 struct bufdomainset bdlodirty; /* Domains > lodirty */ struct bufdomainset bdhidirty; /* Domains > hidirty */ /* Configured number of clean queues. */ static int __read_mostly buf_domains; BITSET_DEFINE(bufdomainset, BUF_DOMAINS); struct bufdomain __exclusive_cache_line bdomain[BUF_DOMAINS]; struct bufqueue __exclusive_cache_line bqempty; /* * per-cpu empty buffer cache. */ uma_zone_t buf_zone; /* * Single global constant for BUF_WMESG, to avoid getting multiple references. * buf_wmesg is referred from macros. */ const char *buf_wmesg = BUF_WMESG; static int sysctl_runningspace(SYSCTL_HANDLER_ARGS) { long value; int error; value = *(long *)arg1; error = sysctl_handle_long(oidp, &value, 0, req); if (error != 0 || req->newptr == NULL) return (error); mtx_lock(&rbreqlock); if (arg1 == &hirunningspace) { if (value < lorunningspace) error = EINVAL; else hirunningspace = value; } else { KASSERT(arg1 == &lorunningspace, ("%s: unknown arg1", __func__)); if (value > hirunningspace) error = EINVAL; else lorunningspace = value; } mtx_unlock(&rbreqlock); return (error); } static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS) { int error; int value; int i; value = *(int *)arg1; error = sysctl_handle_int(oidp, &value, 0, req); if (error != 0 || req->newptr == NULL) return (error); *(int *)arg1 = value; for (i = 0; i < buf_domains; i++) *(int *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) = value / buf_domains; return (error); } static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS) { long value; int error; int i; value = *(long *)arg1; error = sysctl_handle_long(oidp, &value, 0, req); if (error != 0 || req->newptr == NULL) return (error); *(long *)arg1 = value; for (i = 0; i < buf_domains; i++) *(long *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) = value / buf_domains; return (error); } #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) static int sysctl_bufspace(SYSCTL_HANDLER_ARGS) { long lvalue; int ivalue; int i; lvalue = 0; for (i = 0; i < buf_domains; i++) lvalue += bdomain[i].bd_bufspace; if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long)) return (sysctl_handle_long(oidp, &lvalue, 0, req)); if (lvalue > INT_MAX) /* On overflow, still write out a long to trigger ENOMEM. */ return (sysctl_handle_long(oidp, &lvalue, 0, req)); ivalue = lvalue; return (sysctl_handle_int(oidp, &ivalue, 0, req)); } #else static int sysctl_bufspace(SYSCTL_HANDLER_ARGS) { long lvalue; int i; lvalue = 0; for (i = 0; i < buf_domains; i++) lvalue += bdomain[i].bd_bufspace; return (sysctl_handle_long(oidp, &lvalue, 0, req)); } #endif static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS) { int value; int i; value = 0; for (i = 0; i < buf_domains; i++) value += bdomain[i].bd_numdirtybuffers; return (sysctl_handle_int(oidp, &value, 0, req)); } /* * bdirtywakeup: * * Wakeup any bwillwrite() waiters. */ static void bdirtywakeup(void) { mtx_lock(&bdirtylock); if (bdirtywait) { bdirtywait = 0; wakeup(&bdirtywait); } mtx_unlock(&bdirtylock); } /* * bd_clear: * * Clear a domain from the appropriate bitsets when dirtybuffers * is decremented. */ static void bd_clear(struct bufdomain *bd) { mtx_lock(&bdirtylock); if (bd->bd_numdirtybuffers <= bd->bd_lodirtybuffers) BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty); if (bd->bd_numdirtybuffers <= bd->bd_hidirtybuffers) BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty); mtx_unlock(&bdirtylock); } /* * bd_set: * * Set a domain in the appropriate bitsets when dirtybuffers * is incremented. */ static void bd_set(struct bufdomain *bd) { mtx_lock(&bdirtylock); if (bd->bd_numdirtybuffers > bd->bd_lodirtybuffers) BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty); if (bd->bd_numdirtybuffers > bd->bd_hidirtybuffers) BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty); mtx_unlock(&bdirtylock); } /* * bdirtysub: * * Decrement the numdirtybuffers count by one and wakeup any * threads blocked in bwillwrite(). */ static void bdirtysub(struct buf *bp) { struct bufdomain *bd; int num; bd = bufdomain(bp); num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, -1); if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2) bdirtywakeup(); if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers) bd_clear(bd); } /* * bdirtyadd: * * Increment the numdirtybuffers count by one and wakeup the buf * daemon if needed. */ static void bdirtyadd(struct buf *bp) { struct bufdomain *bd; int num; /* * Only do the wakeup once as we cross the boundary. The * buf daemon will keep running until the condition clears. */ bd = bufdomain(bp); num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, 1); if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2) bd_wakeup(); if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers) bd_set(bd); } /* * bufspace_daemon_wakeup: * * Wakeup the daemons responsible for freeing clean bufs. */ static void bufspace_daemon_wakeup(struct bufdomain *bd) { /* * avoid the lock if the daemon is running. */ if (atomic_fetchadd_int(&bd->bd_running, 1) == 0) { BD_RUN_LOCK(bd); atomic_store_int(&bd->bd_running, 1); wakeup(&bd->bd_running); BD_RUN_UNLOCK(bd); } } /* * bufspace_daemon_wait: * * Sleep until the domain falls below a limit or one second passes. */ static void bufspace_daemon_wait(struct bufdomain *bd) { /* * Re-check our limits and sleep. bd_running must be * cleared prior to checking the limits to avoid missed * wakeups. The waker will adjust one of bufspace or * freebuffers prior to checking bd_running. */ BD_RUN_LOCK(bd); atomic_store_int(&bd->bd_running, 0); if (bd->bd_bufspace < bd->bd_bufspacethresh && bd->bd_freebuffers > bd->bd_lofreebuffers) { msleep(&bd->bd_running, BD_RUN_LOCKPTR(bd), PRIBIO|PDROP, "-", hz); } else { /* Avoid spurious wakeups while running. */ atomic_store_int(&bd->bd_running, 1); BD_RUN_UNLOCK(bd); } } /* * bufspace_adjust: * * Adjust the reported bufspace for a KVA managed buffer, possibly * waking any waiters. */ static void bufspace_adjust(struct buf *bp, int bufsize) { struct bufdomain *bd; long space; int diff; KASSERT((bp->b_flags & B_MALLOC) == 0, ("bufspace_adjust: malloc buf %p", bp)); bd = bufdomain(bp); diff = bufsize - bp->b_bufsize; if (diff < 0) { atomic_subtract_long(&bd->bd_bufspace, -diff); } else if (diff > 0) { space = atomic_fetchadd_long(&bd->bd_bufspace, diff); /* Wake up the daemon on the transition. */ if (space < bd->bd_bufspacethresh && space + diff >= bd->bd_bufspacethresh) bufspace_daemon_wakeup(bd); } bp->b_bufsize = bufsize; } /* * bufspace_reserve: * * Reserve bufspace before calling allocbuf(). metadata has a * different space limit than data. */ static int bufspace_reserve(struct bufdomain *bd, int size, bool metadata) { long limit, new; long space; if (metadata) limit = bd->bd_maxbufspace; else limit = bd->bd_hibufspace; space = atomic_fetchadd_long(&bd->bd_bufspace, size); new = space + size; if (new > limit) { atomic_subtract_long(&bd->bd_bufspace, size); return (ENOSPC); } /* Wake up the daemon on the transition. */ if (space < bd->bd_bufspacethresh && new >= bd->bd_bufspacethresh) bufspace_daemon_wakeup(bd); return (0); } /* * bufspace_release: * * Release reserved bufspace after bufspace_adjust() has consumed it. */ static void bufspace_release(struct bufdomain *bd, int size) { atomic_subtract_long(&bd->bd_bufspace, size); } /* * bufspace_wait: * * Wait for bufspace, acting as the buf daemon if a locked vnode is * supplied. bd_wanted must be set prior to polling for space. The * operation must be re-tried on return. */ static void bufspace_wait(struct bufdomain *bd, struct vnode *vp, int gbflags, int slpflag, int slptimeo) { struct thread *td; int error, fl, norunbuf; if ((gbflags & GB_NOWAIT_BD) != 0) return; td = curthread; BD_LOCK(bd); while (bd->bd_wanted) { if (vp != NULL && vp->v_type != VCHR && (td->td_pflags & TDP_BUFNEED) == 0) { BD_UNLOCK(bd); /* * getblk() is called with a vnode locked, and * some majority of the dirty buffers may as * well belong to the vnode. Flushing the * buffers there would make a progress that * cannot be achieved by the buf_daemon, that * cannot lock the vnode. */ norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) | (td->td_pflags & TDP_NORUNNINGBUF); /* * Play bufdaemon. The getnewbuf() function * may be called while the thread owns lock * for another dirty buffer for the same * vnode, which makes it impossible to use * VOP_FSYNC() there, due to the buffer lock * recursion. */ td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF; fl = buf_flush(vp, bd, flushbufqtarget); td->td_pflags &= norunbuf; BD_LOCK(bd); if (fl != 0) continue; if (bd->bd_wanted == 0) break; } error = msleep(&bd->bd_wanted, BD_LOCKPTR(bd), (PRIBIO + 4) | slpflag, "newbuf", slptimeo); if (error != 0) break; } BD_UNLOCK(bd); } /* * bufspace_daemon: * * buffer space management daemon. Tries to maintain some marginal * amount of free buffer space so that requesting processes neither * block nor work to reclaim buffers. */ static void bufspace_daemon(void *arg) { struct bufdomain *bd; EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread, SHUTDOWN_PRI_LAST + 100); bd = arg; for (;;) { kthread_suspend_check(); /* * Free buffers from the clean queue until we meet our * targets. * * Theory of operation: The buffer cache is most efficient * when some free buffer headers and space are always * available to getnewbuf(). This daemon attempts to prevent * the excessive blocking and synchronization associated * with shortfall. It goes through three phases according * demand: * * 1) The daemon wakes up voluntarily once per-second * during idle periods when the counters are below * the wakeup thresholds (bufspacethresh, lofreebuffers). * * 2) The daemon wakes up as we cross the thresholds * ahead of any potential blocking. This may bounce * slightly according to the rate of consumption and * release. * * 3) The daemon and consumers are starved for working * clean buffers. This is the 'bufspace' sleep below * which will inefficiently trade bufs with bqrelse * until we return to condition 2. */ while (bd->bd_bufspace > bd->bd_lobufspace || bd->bd_freebuffers < bd->bd_hifreebuffers) { if (buf_recycle(bd, false) != 0) { if (bd_flushall(bd)) continue; /* * Speedup dirty if we've run out of clean * buffers. This is possible in particular * because softdep may held many bufs locked * pending writes to other bufs which are * marked for delayed write, exhausting * clean space until they are written. */ bd_speedup(); BD_LOCK(bd); if (bd->bd_wanted) { msleep(&bd->bd_wanted, BD_LOCKPTR(bd), PRIBIO|PDROP, "bufspace", hz/10); } else BD_UNLOCK(bd); } maybe_yield(); } bufspace_daemon_wait(bd); } } /* * bufmallocadjust: * * Adjust the reported bufspace for a malloc managed buffer, possibly * waking any waiters. */ static void bufmallocadjust(struct buf *bp, int bufsize) { int diff; KASSERT((bp->b_flags & B_MALLOC) != 0, ("bufmallocadjust: non-malloc buf %p", bp)); diff = bufsize - bp->b_bufsize; if (diff < 0) atomic_subtract_long(&bufmallocspace, -diff); else atomic_add_long(&bufmallocspace, diff); bp->b_bufsize = bufsize; } /* * runningwakeup: * * Wake up processes that are waiting on asynchronous writes to fall * below lorunningspace. */ static void runningwakeup(void) { mtx_lock(&rbreqlock); if (runningbufreq) { runningbufreq = 0; wakeup(&runningbufreq); } mtx_unlock(&rbreqlock); } /* * runningbufwakeup: * * Decrement the outstanding write count according. */ void runningbufwakeup(struct buf *bp) { long space, bspace; bspace = bp->b_runningbufspace; if (bspace == 0) return; space = atomic_fetchadd_long(&runningbufspace, -bspace); KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld", space, bspace)); bp->b_runningbufspace = 0; /* * Only acquire the lock and wakeup on the transition from exceeding * the threshold to falling below it. */ if (space < lorunningspace) return; if (space - bspace > lorunningspace) return; runningwakeup(); } /* * waitrunningbufspace() * * runningbufspace is a measure of the amount of I/O currently * running. This routine is used in async-write situations to * prevent creating huge backups of pending writes to a device. * Only asynchronous writes are governed by this function. * * This does NOT turn an async write into a sync write. It waits * for earlier writes to complete and generally returns before the * caller's write has reached the device. */ void waitrunningbufspace(void) { mtx_lock(&rbreqlock); while (runningbufspace > hirunningspace) { runningbufreq = 1; msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); } mtx_unlock(&rbreqlock); } /* * vfs_buf_test_cache: * * Called when a buffer is extended. This function clears the B_CACHE * bit if the newly extended portion of the buffer does not contain * valid data. */ static __inline void vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, vm_page_t m) { VM_OBJECT_ASSERT_LOCKED(m->object); + + /* + * This function and its results are protected by higher level + * synchronization requiring vnode and buf locks to page in and + * validate pages. + */ if (bp->b_flags & B_CACHE) { int base = (foff + off) & PAGE_MASK; if (vm_page_is_valid(m, base, size) == 0) bp->b_flags &= ~B_CACHE; } } /* Wake up the buffer daemon if necessary */ static void bd_wakeup(void) { mtx_lock(&bdlock); if (bd_request == 0) { bd_request = 1; wakeup(&bd_request); } mtx_unlock(&bdlock); } /* * Adjust the maxbcachbuf tunable. */ static void maxbcachebuf_adjust(void) { int i; /* * maxbcachebuf must be a power of 2 >= MAXBSIZE. */ i = 2; while (i * 2 <= maxbcachebuf) i *= 2; maxbcachebuf = i; if (maxbcachebuf < MAXBSIZE) maxbcachebuf = MAXBSIZE; if (maxbcachebuf > MAXPHYS) maxbcachebuf = MAXPHYS; if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF) printf("maxbcachebuf=%d\n", maxbcachebuf); } /* * bd_speedup - speedup the buffer cache flushing code */ void bd_speedup(void) { int needwake; mtx_lock(&bdlock); needwake = 0; if (bd_speedupreq == 0 || bd_request == 0) needwake = 1; bd_speedupreq = 1; bd_request = 1; if (needwake) wakeup(&bd_request); mtx_unlock(&bdlock); } #ifdef __i386__ #define TRANSIENT_DENOM 5 #else #define TRANSIENT_DENOM 10 #endif /* * Calculating buffer cache scaling values and reserve space for buffer * headers. This is called during low level kernel initialization and * may be called more then once. We CANNOT write to the memory area * being reserved at this time. */ caddr_t kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) { int tuned_nbuf; long maxbuf, maxbuf_sz, buf_sz, biotmap_sz; /* * physmem_est is in pages. Convert it to kilobytes (assumes * PAGE_SIZE is >= 1K) */ physmem_est = physmem_est * (PAGE_SIZE / 1024); maxbcachebuf_adjust(); /* * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. * For the first 64MB of ram nominally allocate sufficient buffers to * cover 1/4 of our ram. Beyond the first 64MB allocate additional * buffers to cover 1/10 of our ram over 64MB. When auto-sizing * the buffer cache we limit the eventual kva reservation to * maxbcache bytes. * * factor represents the 1/4 x ram conversion. */ if (nbuf == 0) { int factor = 4 * BKVASIZE / 1024; nbuf = 50; if (physmem_est > 4096) nbuf += min((physmem_est - 4096) / factor, 65536 / factor); if (physmem_est > 65536) nbuf += min((physmem_est - 65536) * 2 / (factor * 5), 32 * 1024 * 1024 / (factor * 5)); if (maxbcache && nbuf > maxbcache / BKVASIZE) nbuf = maxbcache / BKVASIZE; tuned_nbuf = 1; } else tuned_nbuf = 0; /* XXX Avoid unsigned long overflows later on with maxbufspace. */ maxbuf = (LONG_MAX / 3) / BKVASIZE; if (nbuf > maxbuf) { if (!tuned_nbuf) printf("Warning: nbufs lowered from %d to %ld\n", nbuf, maxbuf); nbuf = maxbuf; } /* * Ideal allocation size for the transient bio submap is 10% * of the maximal space buffer map. This roughly corresponds * to the amount of the buffer mapped for typical UFS load. * * Clip the buffer map to reserve space for the transient * BIOs, if its extent is bigger than 90% (80% on i386) of the * maximum buffer map extent on the platform. * * The fall-back to the maxbuf in case of maxbcache unset, * allows to not trim the buffer KVA for the architectures * with ample KVA space. */ if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) { maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE; buf_sz = (long)nbuf * BKVASIZE; if (buf_sz < maxbuf_sz / TRANSIENT_DENOM * (TRANSIENT_DENOM - 1)) { /* * There is more KVA than memory. Do not * adjust buffer map size, and assign the rest * of maxbuf to transient map. */ biotmap_sz = maxbuf_sz - buf_sz; } else { /* * Buffer map spans all KVA we could afford on * this platform. Give 10% (20% on i386) of * the buffer map to the transient bio map. */ biotmap_sz = buf_sz / TRANSIENT_DENOM; buf_sz -= biotmap_sz; } if (biotmap_sz / INT_MAX > MAXPHYS) bio_transient_maxcnt = INT_MAX; else bio_transient_maxcnt = biotmap_sz / MAXPHYS; /* * Artificially limit to 1024 simultaneous in-flight I/Os * using the transient mapping. */ if (bio_transient_maxcnt > 1024) bio_transient_maxcnt = 1024; if (tuned_nbuf) nbuf = buf_sz / BKVASIZE; } if (nswbuf == 0) { nswbuf = min(nbuf / 4, 256); if (nswbuf < NSWBUF_MIN) nswbuf = NSWBUF_MIN; } /* * Reserve space for the buffer cache buffers */ buf = (void *)v; v = (caddr_t)(buf + nbuf); return(v); } /* Initialize the buffer subsystem. Called before use of any buffers. */ void bufinit(void) { struct buf *bp; int i; KASSERT(maxbcachebuf >= MAXBSIZE, ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf, MAXBSIZE)); bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock"); mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF); unmapped_buf = (caddr_t)kva_alloc(MAXPHYS); /* finally, initialize each buffer header and stick on empty q */ for (i = 0; i < nbuf; i++) { bp = &buf[i]; bzero(bp, sizeof *bp); bp->b_flags = B_INVAL; bp->b_rcred = NOCRED; bp->b_wcred = NOCRED; bp->b_qindex = QUEUE_NONE; bp->b_domain = -1; bp->b_subqueue = mp_maxid + 1; bp->b_xflags = 0; bp->b_data = bp->b_kvabase = unmapped_buf; LIST_INIT(&bp->b_dep); BUF_LOCKINIT(bp); bq_insert(&bqempty, bp, false); } /* * maxbufspace is the absolute maximum amount of buffer space we are * allowed to reserve in KVM and in real terms. The absolute maximum * is nominally used by metadata. hibufspace is the nominal maximum * used by most other requests. The differential is required to * ensure that metadata deadlocks don't occur. * * maxbufspace is based on BKVASIZE. Allocating buffers larger then * this may result in KVM fragmentation which is not handled optimally * by the system. XXX This is less true with vmem. We could use * PAGE_SIZE. */ maxbufspace = (long)nbuf * BKVASIZE; hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10); lobufspace = (hibufspace / 20) * 19; /* 95% */ bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2; /* * Note: The 16 MiB upper limit for hirunningspace was chosen * arbitrarily and may need further tuning. It corresponds to * 128 outstanding write IO requests (if IO size is 128 KiB), * which fits with many RAID controllers' tagged queuing limits. * The lower 1 MiB limit is the historical upper limit for * hirunningspace. */ hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf), 16 * 1024 * 1024), 1024 * 1024); lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf); /* * Limit the amount of malloc memory since it is wired permanently into * the kernel space. Even though this is accounted for in the buffer * allocation, we don't want the malloced region to grow uncontrolled. * The malloc scheme improves memory utilization significantly on * average (small) directories. */ maxbufmallocspace = hibufspace / 20; /* * Reduce the chance of a deadlock occurring by limiting the number * of delayed-write dirty buffers we allow to stack up. */ hidirtybuffers = nbuf / 4 + 20; dirtybufthresh = hidirtybuffers * 9 / 10; /* * To support extreme low-memory systems, make sure hidirtybuffers * cannot eat up all available buffer space. This occurs when our * minimum cannot be met. We try to size hidirtybuffers to 3/4 our * buffer space assuming BKVASIZE'd buffers. */ while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { hidirtybuffers >>= 1; } lodirtybuffers = hidirtybuffers / 2; /* * lofreebuffers should be sufficient to avoid stalling waiting on * buf headers under heavy utilization. The bufs in per-cpu caches * are counted as free but will be unavailable to threads executing * on other cpus. * * hifreebuffers is the free target for the bufspace daemon. This * should be set appropriately to limit work per-iteration. */ lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus); hifreebuffers = (3 * lofreebuffers) / 2; numfreebuffers = nbuf; /* Setup the kva and free list allocators. */ vmem_set_reclaim(buffer_arena, bufkva_reclaim); buf_zone = uma_zcache_create("buf free cache", sizeof(struct buf), NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0); /* * Size the clean queue according to the amount of buffer space. * One queue per-256mb up to the max. More queues gives better * concurrency but less accurate LRU. */ buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS); for (i = 0 ; i < buf_domains; i++) { struct bufdomain *bd; bd = &bdomain[i]; bd_init(bd); bd->bd_freebuffers = nbuf / buf_domains; bd->bd_hifreebuffers = hifreebuffers / buf_domains; bd->bd_lofreebuffers = lofreebuffers / buf_domains; bd->bd_bufspace = 0; bd->bd_maxbufspace = maxbufspace / buf_domains; bd->bd_hibufspace = hibufspace / buf_domains; bd->bd_lobufspace = lobufspace / buf_domains; bd->bd_bufspacethresh = bufspacethresh / buf_domains; bd->bd_numdirtybuffers = 0; bd->bd_hidirtybuffers = hidirtybuffers / buf_domains; bd->bd_lodirtybuffers = lodirtybuffers / buf_domains; bd->bd_dirtybufthresh = dirtybufthresh / buf_domains; /* Don't allow more than 2% of bufs in the per-cpu caches. */ bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus; } getnewbufcalls = counter_u64_alloc(M_WAITOK); getnewbufrestarts = counter_u64_alloc(M_WAITOK); mappingrestarts = counter_u64_alloc(M_WAITOK); numbufallocfails = counter_u64_alloc(M_WAITOK); notbufdflushes = counter_u64_alloc(M_WAITOK); buffreekvacnt = counter_u64_alloc(M_WAITOK); bufdefragcnt = counter_u64_alloc(M_WAITOK); bufkvaspace = counter_u64_alloc(M_WAITOK); } #ifdef INVARIANTS static inline void vfs_buf_check_mapped(struct buf *bp) { KASSERT(bp->b_kvabase != unmapped_buf, ("mapped buf: b_kvabase was not updated %p", bp)); KASSERT(bp->b_data != unmapped_buf, ("mapped buf: b_data was not updated %p", bp)); KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf + MAXPHYS, ("b_data + b_offset unmapped %p", bp)); } static inline void vfs_buf_check_unmapped(struct buf *bp) { KASSERT(bp->b_data == unmapped_buf, ("unmapped buf: corrupted b_data %p", bp)); } #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp) #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp) #else #define BUF_CHECK_MAPPED(bp) do {} while (0) #define BUF_CHECK_UNMAPPED(bp) do {} while (0) #endif static int isbufbusy(struct buf *bp) { if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) || ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI)) return (1); return (0); } /* * Shutdown the system cleanly to prepare for reboot, halt, or power off. */ void bufshutdown(int show_busybufs) { static int first_buf_printf = 1; struct buf *bp; int iter, nbusy, pbusy; #ifndef PREEMPTION int subiter; #endif /* * Sync filesystems for shutdown */ wdog_kern_pat(WD_LASTVAL); sys_sync(curthread, NULL); /* * With soft updates, some buffers that are * written will be remarked as dirty until other * buffers are written. */ for (iter = pbusy = 0; iter < 20; iter++) { nbusy = 0; for (bp = &buf[nbuf]; --bp >= buf; ) if (isbufbusy(bp)) nbusy++; if (nbusy == 0) { if (first_buf_printf) printf("All buffers synced."); break; } if (first_buf_printf) { printf("Syncing disks, buffers remaining... "); first_buf_printf = 0; } printf("%d ", nbusy); if (nbusy < pbusy) iter = 0; pbusy = nbusy; wdog_kern_pat(WD_LASTVAL); sys_sync(curthread, NULL); #ifdef PREEMPTION /* * Spin for a while to allow interrupt threads to run. */ DELAY(50000 * iter); #else /* * Context switch several times to allow interrupt * threads to run. */ for (subiter = 0; subiter < 50 * iter; subiter++) { thread_lock(curthread); mi_switch(SW_VOL, NULL); thread_unlock(curthread); DELAY(1000); } #endif } printf("\n"); /* * Count only busy local buffers to prevent forcing * a fsck if we're just a client of a wedged NFS server */ nbusy = 0; for (bp = &buf[nbuf]; --bp >= buf; ) { if (isbufbusy(bp)) { #if 0 /* XXX: This is bogus. We should probably have a BO_REMOTE flag instead */ if (bp->b_dev == NULL) { TAILQ_REMOVE(&mountlist, bp->b_vp->v_mount, mnt_list); continue; } #endif nbusy++; if (show_busybufs > 0) { printf( "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:", nbusy, bp, bp->b_vp, bp->b_flags, (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno); BUF_LOCKPRINTINFO(bp); if (show_busybufs > 1) vn_printf(bp->b_vp, "vnode content: "); } } } if (nbusy) { /* * Failed to sync all blocks. Indicate this and don't * unmount filesystems (thus forcing an fsck on reboot). */ printf("Giving up on %d buffers\n", nbusy); DELAY(5000000); /* 5 seconds */ } else { if (!first_buf_printf) printf("Final sync complete\n"); /* * Unmount filesystems */ if (panicstr == NULL) vfs_unmountall(); } swapoff_all(); DELAY(100000); /* wait for console output to finish */ } static void bpmap_qenter(struct buf *bp) { BUF_CHECK_MAPPED(bp); /* * bp->b_data is relative to bp->b_offset, but * bp->b_offset may be offset into the first page. */ bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data); pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages); bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | (vm_offset_t)(bp->b_offset & PAGE_MASK)); } static inline struct bufdomain * bufdomain(struct buf *bp) { return (&bdomain[bp->b_domain]); } static struct bufqueue * bufqueue(struct buf *bp) { switch (bp->b_qindex) { case QUEUE_NONE: /* FALLTHROUGH */ case QUEUE_SENTINEL: return (NULL); case QUEUE_EMPTY: return (&bqempty); case QUEUE_DIRTY: return (&bufdomain(bp)->bd_dirtyq); case QUEUE_CLEAN: return (&bufdomain(bp)->bd_subq[bp->b_subqueue]); default: break; } panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex); } /* * Return the locked bufqueue that bp is a member of. */ static struct bufqueue * bufqueue_acquire(struct buf *bp) { struct bufqueue *bq, *nbq; /* * bp can be pushed from a per-cpu queue to the * cleanq while we're waiting on the lock. Retry * if the queues don't match. */ bq = bufqueue(bp); BQ_LOCK(bq); for (;;) { nbq = bufqueue(bp); if (bq == nbq) break; BQ_UNLOCK(bq); BQ_LOCK(nbq); bq = nbq; } return (bq); } /* * binsfree: * * Insert the buffer into the appropriate free list. Requires a * locked buffer on entry and buffer is unlocked before return. */ static void binsfree(struct buf *bp, int qindex) { struct bufdomain *bd; struct bufqueue *bq; KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY, ("binsfree: Invalid qindex %d", qindex)); BUF_ASSERT_XLOCKED(bp); /* * Handle delayed bremfree() processing. */ if (bp->b_flags & B_REMFREE) { if (bp->b_qindex == qindex) { bp->b_flags |= B_REUSE; bp->b_flags &= ~B_REMFREE; BUF_UNLOCK(bp); return; } bq = bufqueue_acquire(bp); bq_remove(bq, bp); BQ_UNLOCK(bq); } bd = bufdomain(bp); if (qindex == QUEUE_CLEAN) { if (bd->bd_lim != 0) bq = &bd->bd_subq[PCPU_GET(cpuid)]; else bq = bd->bd_cleanq; } else bq = &bd->bd_dirtyq; bq_insert(bq, bp, true); } /* * buf_free: * * Free a buffer to the buf zone once it no longer has valid contents. */ static void buf_free(struct buf *bp) { if (bp->b_flags & B_REMFREE) bremfreef(bp); if (bp->b_vflags & BV_BKGRDINPROG) panic("losing buffer 1"); if (bp->b_rcred != NOCRED) { crfree(bp->b_rcred); bp->b_rcred = NOCRED; } if (bp->b_wcred != NOCRED) { crfree(bp->b_wcred); bp->b_wcred = NOCRED; } if (!LIST_EMPTY(&bp->b_dep)) buf_deallocate(bp); bufkva_free(bp); atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1); BUF_UNLOCK(bp); uma_zfree(buf_zone, bp); } /* * buf_import: * * Import bufs into the uma cache from the buf list. The system still * expects a static array of bufs and much of the synchronization * around bufs assumes type stable storage. As a result, UMA is used * only as a per-cpu cache of bufs still maintained on a global list. */ static int buf_import(void *arg, void **store, int cnt, int domain, int flags) { struct buf *bp; int i; BQ_LOCK(&bqempty); for (i = 0; i < cnt; i++) { bp = TAILQ_FIRST(&bqempty.bq_queue); if (bp == NULL) break; bq_remove(&bqempty, bp); store[i] = bp; } BQ_UNLOCK(&bqempty); return (i); } /* * buf_release: * * Release bufs from the uma cache back to the buffer queues. */ static void buf_release(void *arg, void **store, int cnt) { struct bufqueue *bq; struct buf *bp; int i; bq = &bqempty; BQ_LOCK(bq); for (i = 0; i < cnt; i++) { bp = store[i]; /* Inline bq_insert() to batch locking. */ TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist); bp->b_flags &= ~(B_AGE | B_REUSE); bq->bq_len++; bp->b_qindex = bq->bq_index; } BQ_UNLOCK(bq); } /* * buf_alloc: * * Allocate an empty buffer header. */ static struct buf * buf_alloc(struct bufdomain *bd) { struct buf *bp; int freebufs; /* * We can only run out of bufs in the buf zone if the average buf * is less than BKVASIZE. In this case the actual wait/block will * come from buf_reycle() failing to flush one of these small bufs. */ bp = NULL; freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1); if (freebufs > 0) bp = uma_zalloc(buf_zone, M_NOWAIT); if (bp == NULL) { atomic_add_int(&bd->bd_freebuffers, 1); bufspace_daemon_wakeup(bd); counter_u64_add(numbufallocfails, 1); return (NULL); } /* * Wake-up the bufspace daemon on transition below threshold. */ if (freebufs == bd->bd_lofreebuffers) bufspace_daemon_wakeup(bd); if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) panic("getnewbuf_empty: Locked buf %p on free queue.", bp); KASSERT(bp->b_vp == NULL, ("bp: %p still has vnode %p.", bp, bp->b_vp)); KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0, ("invalid buffer %p flags %#x", bp, bp->b_flags)); KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags)); KASSERT(bp->b_npages == 0, ("bp: %p still has %d vm pages\n", bp, bp->b_npages)); KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp)); KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp)); bp->b_domain = BD_DOMAIN(bd); bp->b_flags = 0; bp->b_ioflags = 0; bp->b_xflags = 0; bp->b_vflags = 0; bp->b_vp = NULL; bp->b_blkno = bp->b_lblkno = 0; bp->b_offset = NOOFFSET; bp->b_iodone = 0; bp->b_error = 0; bp->b_resid = 0; bp->b_bcount = 0; bp->b_npages = 0; bp->b_dirtyoff = bp->b_dirtyend = 0; bp->b_bufobj = NULL; bp->b_data = bp->b_kvabase = unmapped_buf; bp->b_fsprivate1 = NULL; bp->b_fsprivate2 = NULL; bp->b_fsprivate3 = NULL; LIST_INIT(&bp->b_dep); return (bp); } /* * buf_recycle: * * Free a buffer from the given bufqueue. kva controls whether the * freed buf must own some kva resources. This is used for * defragmenting. */ static int buf_recycle(struct bufdomain *bd, bool kva) { struct bufqueue *bq; struct buf *bp, *nbp; if (kva) counter_u64_add(bufdefragcnt, 1); nbp = NULL; bq = bd->bd_cleanq; BQ_LOCK(bq); KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd), ("buf_recycle: Locks don't match")); nbp = TAILQ_FIRST(&bq->bq_queue); /* * Run scan, possibly freeing data and/or kva mappings on the fly * depending. */ while ((bp = nbp) != NULL) { /* * Calculate next bp (we can only use it if we do not * release the bqlock). */ nbp = TAILQ_NEXT(bp, b_freelist); /* * If we are defragging then we need a buffer with * some kva to reclaim. */ if (kva && bp->b_kvasize == 0) continue; if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) continue; /* * Implement a second chance algorithm for frequently * accessed buffers. */ if ((bp->b_flags & B_REUSE) != 0) { TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist); TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist); bp->b_flags &= ~B_REUSE; BUF_UNLOCK(bp); continue; } /* * Skip buffers with background writes in progress. */ if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { BUF_UNLOCK(bp); continue; } KASSERT(bp->b_qindex == QUEUE_CLEAN, ("buf_recycle: inconsistent queue %d bp %p", bp->b_qindex, bp)); KASSERT(bp->b_domain == BD_DOMAIN(bd), ("getnewbuf: queue domain %d doesn't match request %d", bp->b_domain, (int)BD_DOMAIN(bd))); /* * NOTE: nbp is now entirely invalid. We can only restart * the scan from this point on. */ bq_remove(bq, bp); BQ_UNLOCK(bq); /* * Requeue the background write buffer with error and * restart the scan. */ if ((bp->b_vflags & BV_BKGRDERR) != 0) { bqrelse(bp); BQ_LOCK(bq); nbp = TAILQ_FIRST(&bq->bq_queue); continue; } bp->b_flags |= B_INVAL; brelse(bp); return (0); } bd->bd_wanted = 1; BQ_UNLOCK(bq); return (ENOBUFS); } /* * bremfree: * * Mark the buffer for removal from the appropriate free list. * */ void bremfree(struct buf *bp) { CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT((bp->b_flags & B_REMFREE) == 0, ("bremfree: buffer %p already marked for delayed removal.", bp)); KASSERT(bp->b_qindex != QUEUE_NONE, ("bremfree: buffer %p not on a queue.", bp)); BUF_ASSERT_XLOCKED(bp); bp->b_flags |= B_REMFREE; } /* * bremfreef: * * Force an immediate removal from a free list. Used only in nfs when * it abuses the b_freelist pointer. */ void bremfreef(struct buf *bp) { struct bufqueue *bq; bq = bufqueue_acquire(bp); bq_remove(bq, bp); BQ_UNLOCK(bq); } static void bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname) { mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF); TAILQ_INIT(&bq->bq_queue); bq->bq_len = 0; bq->bq_index = qindex; bq->bq_subqueue = subqueue; } static void bd_init(struct bufdomain *bd) { int i; bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1]; bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock"); bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock"); for (i = 0; i <= mp_maxid; i++) bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i, "bufq clean subqueue lock"); mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF); } /* * bq_remove: * * Removes a buffer from the free list, must be called with the * correct qlock held. */ static void bq_remove(struct bufqueue *bq, struct buf *bp) { CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT(bp->b_qindex != QUEUE_NONE, ("bq_remove: buffer %p not on a queue.", bp)); KASSERT(bufqueue(bp) == bq, ("bq_remove: Remove buffer %p from wrong queue.", bp)); BQ_ASSERT_LOCKED(bq); if (bp->b_qindex != QUEUE_EMPTY) { BUF_ASSERT_XLOCKED(bp); } KASSERT(bq->bq_len >= 1, ("queue %d underflow", bp->b_qindex)); TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist); bq->bq_len--; bp->b_qindex = QUEUE_NONE; bp->b_flags &= ~(B_REMFREE | B_REUSE); } static void bd_flush(struct bufdomain *bd, struct bufqueue *bq) { struct buf *bp; BQ_ASSERT_LOCKED(bq); if (bq != bd->bd_cleanq) { BD_LOCK(bd); while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) { TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist); TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp, b_freelist); bp->b_subqueue = bd->bd_cleanq->bq_subqueue; } bd->bd_cleanq->bq_len += bq->bq_len; bq->bq_len = 0; } if (bd->bd_wanted) { bd->bd_wanted = 0; wakeup(&bd->bd_wanted); } if (bq != bd->bd_cleanq) BD_UNLOCK(bd); } static int bd_flushall(struct bufdomain *bd) { struct bufqueue *bq; int flushed; int i; if (bd->bd_lim == 0) return (0); flushed = 0; for (i = 0; i <= mp_maxid; i++) { bq = &bd->bd_subq[i]; if (bq->bq_len == 0) continue; BQ_LOCK(bq); bd_flush(bd, bq); BQ_UNLOCK(bq); flushed++; } return (flushed); } static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock) { struct bufdomain *bd; if (bp->b_qindex != QUEUE_NONE) panic("bq_insert: free buffer %p onto another queue?", bp); bd = bufdomain(bp); if (bp->b_flags & B_AGE) { /* Place this buf directly on the real queue. */ if (bq->bq_index == QUEUE_CLEAN) bq = bd->bd_cleanq; BQ_LOCK(bq); TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist); } else { BQ_LOCK(bq); TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist); } bp->b_flags &= ~(B_AGE | B_REUSE); bq->bq_len++; bp->b_qindex = bq->bq_index; bp->b_subqueue = bq->bq_subqueue; /* * Unlock before we notify so that we don't wakeup a waiter that * fails a trylock on the buf and sleeps again. */ if (unlock) BUF_UNLOCK(bp); if (bp->b_qindex == QUEUE_CLEAN) { /* * Flush the per-cpu queue and notify any waiters. */ if (bd->bd_wanted || (bq != bd->bd_cleanq && bq->bq_len >= bd->bd_lim)) bd_flush(bd, bq); } BQ_UNLOCK(bq); } /* * bufkva_free: * * Free the kva allocation for a buffer. * */ static void bufkva_free(struct buf *bp) { #ifdef INVARIANTS if (bp->b_kvasize == 0) { KASSERT(bp->b_kvabase == unmapped_buf && bp->b_data == unmapped_buf, ("Leaked KVA space on %p", bp)); } else if (buf_mapped(bp)) BUF_CHECK_MAPPED(bp); else BUF_CHECK_UNMAPPED(bp); #endif if (bp->b_kvasize == 0) return; vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize); counter_u64_add(bufkvaspace, -bp->b_kvasize); counter_u64_add(buffreekvacnt, 1); bp->b_data = bp->b_kvabase = unmapped_buf; bp->b_kvasize = 0; } /* * bufkva_alloc: * * Allocate the buffer KVA and set b_kvasize and b_kvabase. */ static int bufkva_alloc(struct buf *bp, int maxsize, int gbflags) { vm_offset_t addr; int error; KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0, ("Invalid gbflags 0x%x in %s", gbflags, __func__)); bufkva_free(bp); addr = 0; error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr); if (error != 0) { /* * Buffer map is too fragmented. Request the caller * to defragment the map. */ return (error); } bp->b_kvabase = (caddr_t)addr; bp->b_kvasize = maxsize; counter_u64_add(bufkvaspace, bp->b_kvasize); if ((gbflags & GB_UNMAPPED) != 0) { bp->b_data = unmapped_buf; BUF_CHECK_UNMAPPED(bp); } else { bp->b_data = bp->b_kvabase; BUF_CHECK_MAPPED(bp); } return (0); } /* * bufkva_reclaim: * * Reclaim buffer kva by freeing buffers holding kva. This is a vmem * callback that fires to avoid returning failure. */ static void bufkva_reclaim(vmem_t *vmem, int flags) { bool done; int q; int i; done = false; for (i = 0; i < 5; i++) { for (q = 0; q < buf_domains; q++) if (buf_recycle(&bdomain[q], true) != 0) done = true; if (done) break; } return; } /* * Attempt to initiate asynchronous I/O on read-ahead blocks. We must * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, * the buffer is valid and we do not have to do anything. */ static void breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt, struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *)) { struct buf *rabp; struct thread *td; int i; td = curthread; for (i = 0; i < cnt; i++, rablkno++, rabsize++) { if (inmem(vp, *rablkno)) continue; rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); if ((rabp->b_flags & B_CACHE) != 0) { brelse(rabp); continue; } #ifdef RACCT if (racct_enable) { PROC_LOCK(curproc); racct_add_buf(curproc, rabp, 0); PROC_UNLOCK(curproc); } #endif /* RACCT */ td->td_ru.ru_inblock++; rabp->b_flags |= B_ASYNC; rabp->b_flags &= ~B_INVAL; if ((flags & GB_CKHASH) != 0) { rabp->b_flags |= B_CKHASH; rabp->b_ckhashcalc = ckhashfunc; } rabp->b_ioflags &= ~BIO_ERROR; rabp->b_iocmd = BIO_READ; if (rabp->b_rcred == NOCRED && cred != NOCRED) rabp->b_rcred = crhold(cred); vfs_busy_pages(rabp, 0); BUF_KERNPROC(rabp); rabp->b_iooffset = dbtob(rabp->b_blkno); bstrategy(rabp); } } /* * Entry point for bread() and breadn() via #defines in sys/buf.h. * * Get a buffer with the specified data. Look in the cache first. We * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE * is set, the buffer is valid and we do not have to do anything, see * getblk(). Also starts asynchronous I/O on read-ahead blocks. * * Always return a NULL buffer pointer (in bpp) when returning an error. */ int breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno, int *rabsize, int cnt, struct ucred *cred, int flags, void (*ckhashfunc)(struct buf *), struct buf **bpp) { struct buf *bp; struct thread *td; int error, readwait, rv; CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); td = curthread; /* * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags * are specified. */ error = getblkx(vp, blkno, size, 0, 0, flags, &bp); if (error != 0) { *bpp = NULL; return (error); } flags &= ~GB_NOSPARSE; *bpp = bp; /* * If not found in cache, do some I/O */ readwait = 0; if ((bp->b_flags & B_CACHE) == 0) { #ifdef RACCT if (racct_enable) { PROC_LOCK(td->td_proc); racct_add_buf(td->td_proc, bp, 0); PROC_UNLOCK(td->td_proc); } #endif /* RACCT */ td->td_ru.ru_inblock++; bp->b_iocmd = BIO_READ; bp->b_flags &= ~B_INVAL; if ((flags & GB_CKHASH) != 0) { bp->b_flags |= B_CKHASH; bp->b_ckhashcalc = ckhashfunc; } bp->b_ioflags &= ~BIO_ERROR; if (bp->b_rcred == NOCRED && cred != NOCRED) bp->b_rcred = crhold(cred); vfs_busy_pages(bp, 0); bp->b_iooffset = dbtob(bp->b_blkno); bstrategy(bp); ++readwait; } /* * Attempt to initiate asynchronous I/O on read-ahead blocks. */ breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc); rv = 0; if (readwait) { rv = bufwait(bp); if (rv != 0) { brelse(bp); *bpp = NULL; } } return (rv); } /* * Write, release buffer on completion. (Done by iodone * if async). Do not bother writing anything if the buffer * is invalid. * * Note that we set B_CACHE here, indicating that buffer is * fully valid and thus cacheable. This is true even of NFS * now so we set it generally. This could be set either here * or in biodone() since the I/O is synchronous. We put it * here. */ int bufwrite(struct buf *bp) { int oldflags; struct vnode *vp; long space; int vp_md; CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) { bp->b_flags |= B_INVAL | B_RELBUF; bp->b_flags &= ~B_CACHE; brelse(bp); return (ENXIO); } if (bp->b_flags & B_INVAL) { brelse(bp); return (0); } if (bp->b_flags & B_BARRIER) atomic_add_long(&barrierwrites, 1); oldflags = bp->b_flags; KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), ("FFS background buffer should not get here %p", bp)); vp = bp->b_vp; if (vp) vp_md = vp->v_vflag & VV_MD; else vp_md = 0; /* * Mark the buffer clean. Increment the bufobj write count * before bundirty() call, to prevent other thread from seeing * empty dirty list and zero counter for writes in progress, * falsely indicating that the bufobj is clean. */ bufobj_wref(bp->b_bufobj); bundirty(bp); bp->b_flags &= ~B_DONE; bp->b_ioflags &= ~BIO_ERROR; bp->b_flags |= B_CACHE; bp->b_iocmd = BIO_WRITE; vfs_busy_pages(bp, 1); /* * Normal bwrites pipeline writes */ bp->b_runningbufspace = bp->b_bufsize; space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace); #ifdef RACCT if (racct_enable) { PROC_LOCK(curproc); racct_add_buf(curproc, bp, 1); PROC_UNLOCK(curproc); } #endif /* RACCT */ curthread->td_ru.ru_oublock++; if (oldflags & B_ASYNC) BUF_KERNPROC(bp); bp->b_iooffset = dbtob(bp->b_blkno); buf_track(bp, __func__); bstrategy(bp); if ((oldflags & B_ASYNC) == 0) { int rtval = bufwait(bp); brelse(bp); return (rtval); } else if (space > hirunningspace) { /* * don't allow the async write to saturate the I/O * system. We will not deadlock here because * we are blocking waiting for I/O that is already in-progress * to complete. We do not block here if it is the update * or syncer daemon trying to clean up as that can lead * to deadlock. */ if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) waitrunningbufspace(); } return (0); } void bufbdflush(struct bufobj *bo, struct buf *bp) { struct buf *nbp; if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); altbufferflushes++; } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { BO_LOCK(bo); /* * Try to find a buffer to flush. */ TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { if ((nbp->b_vflags & BV_BKGRDINPROG) || BUF_LOCK(nbp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) continue; if (bp == nbp) panic("bdwrite: found ourselves"); BO_UNLOCK(bo); /* Don't countdeps with the bo lock held. */ if (buf_countdeps(nbp, 0)) { BO_LOCK(bo); BUF_UNLOCK(nbp); continue; } if (nbp->b_flags & B_CLUSTEROK) { vfs_bio_awrite(nbp); } else { bremfree(nbp); bawrite(nbp); } dirtybufferflushes++; break; } if (nbp == NULL) BO_UNLOCK(bo); } } /* * Delayed write. (Buffer is marked dirty). Do not bother writing * anything if the buffer is marked invalid. * * Note that since the buffer must be completely valid, we can safely * set B_CACHE. In fact, we have to set B_CACHE here rather then in * biodone() in order to prevent getblk from writing the buffer * out synchronously. */ void bdwrite(struct buf *bp) { struct thread *td = curthread; struct vnode *vp; struct bufobj *bo; CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); KASSERT((bp->b_flags & B_BARRIER) == 0, ("Barrier request in delayed write %p", bp)); if (bp->b_flags & B_INVAL) { brelse(bp); return; } /* * If we have too many dirty buffers, don't create any more. * If we are wildly over our limit, then force a complete * cleanup. Otherwise, just keep the situation from getting * out of control. Note that we have to avoid a recursive * disaster and not try to clean up after our own cleanup! */ vp = bp->b_vp; bo = bp->b_bufobj; if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { td->td_pflags |= TDP_INBDFLUSH; BO_BDFLUSH(bo, bp); td->td_pflags &= ~TDP_INBDFLUSH; } else recursiveflushes++; bdirty(bp); /* * Set B_CACHE, indicating that the buffer is fully valid. This is * true even of NFS now. */ bp->b_flags |= B_CACHE; /* * This bmap keeps the system from needing to do the bmap later, * perhaps when the system is attempting to do a sync. Since it * is likely that the indirect block -- or whatever other datastructure * that the filesystem needs is still in memory now, it is a good * thing to do this. Note also, that if the pageout daemon is * requesting a sync -- there might not be enough memory to do * the bmap then... So, this is important to do. */ if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); } buf_track(bp, __func__); /* * Set the *dirty* buffer range based upon the VM system dirty * pages. * * Mark the buffer pages as clean. We need to do this here to * satisfy the vnode_pager and the pageout daemon, so that it * thinks that the pages have been "cleaned". Note that since * the pages are in a delayed write buffer -- the VFS layer * "will" see that the pages get written out on the next sync, * or perhaps the cluster will be completed. */ vfs_clean_pages_dirty_buf(bp); bqrelse(bp); /* * note: we cannot initiate I/O from a bdwrite even if we wanted to, * due to the softdep code. */ } /* * bdirty: * * Turn buffer into delayed write request. We must clear BIO_READ and * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to * itself to properly update it in the dirty/clean lists. We mark it * B_DONE to ensure that any asynchronization of the buffer properly * clears B_DONE ( else a panic will occur later ). * * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() * should only be called if the buffer is known-good. * * Since the buffer is not on a queue, we do not update the numfreebuffers * count. * * The buffer must be on QUEUE_NONE. */ void bdirty(struct buf *bp) { CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); bp->b_flags &= ~(B_RELBUF); bp->b_iocmd = BIO_WRITE; if ((bp->b_flags & B_DELWRI) == 0) { bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; reassignbuf(bp); bdirtyadd(bp); } } /* * bundirty: * * Clear B_DELWRI for buffer. * * Since the buffer is not on a queue, we do not update the numfreebuffers * count. * * The buffer must be on QUEUE_NONE. */ void bundirty(struct buf *bp) { CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); if (bp->b_flags & B_DELWRI) { bp->b_flags &= ~B_DELWRI; reassignbuf(bp); bdirtysub(bp); } /* * Since it is now being written, we can clear its deferred write flag. */ bp->b_flags &= ~B_DEFERRED; } /* * bawrite: * * Asynchronous write. Start output on a buffer, but do not wait for * it to complete. The buffer is released when the output completes. * * bwrite() ( or the VOP routine anyway ) is responsible for handling * B_INVAL buffers. Not us. */ void bawrite(struct buf *bp) { bp->b_flags |= B_ASYNC; (void) bwrite(bp); } /* * babarrierwrite: * * Asynchronous barrier write. Start output on a buffer, but do not * wait for it to complete. Place a write barrier after this write so * that this buffer and all buffers written before it are committed to * the disk before any buffers written after this write are committed * to the disk. The buffer is released when the output completes. */ void babarrierwrite(struct buf *bp) { bp->b_flags |= B_ASYNC | B_BARRIER; (void) bwrite(bp); } /* * bbarrierwrite: * * Synchronous barrier write. Start output on a buffer and wait for * it to complete. Place a write barrier after this write so that * this buffer and all buffers written before it are committed to * the disk before any buffers written after this write are committed * to the disk. The buffer is released when the output completes. */ int bbarrierwrite(struct buf *bp) { bp->b_flags |= B_BARRIER; return (bwrite(bp)); } /* * bwillwrite: * * Called prior to the locking of any vnodes when we are expecting to * write. We do not want to starve the buffer cache with too many * dirty buffers so we block here. By blocking prior to the locking * of any vnodes we attempt to avoid the situation where a locked vnode * prevents the various system daemons from flushing related buffers. */ void bwillwrite(void) { if (buf_dirty_count_severe()) { mtx_lock(&bdirtylock); while (buf_dirty_count_severe()) { bdirtywait = 1; msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4), "flswai", 0); } mtx_unlock(&bdirtylock); } } /* * Return true if we have too many dirty buffers. */ int buf_dirty_count_severe(void) { return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty)); } /* * brelse: * * Release a busy buffer and, if requested, free its resources. The * buffer will be stashed in the appropriate bufqueue[] allowing it * to be accessed later as a cache entity or reused for other purposes. */ void brelse(struct buf *bp) { struct mount *v_mnt; int qindex; /* * Many functions erroneously call brelse with a NULL bp under rare * error conditions. Simply return when called with a NULL bp. */ if (bp == NULL) return; CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0, ("brelse: non-VMIO buffer marked NOREUSE")); if (BUF_LOCKRECURSED(bp)) { /* * Do not process, in particular, do not handle the * B_INVAL/B_RELBUF and do not release to free list. */ BUF_UNLOCK(bp); return; } if (bp->b_flags & B_MANAGED) { bqrelse(bp); return; } if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) { BO_LOCK(bp->b_bufobj); bp->b_vflags &= ~BV_BKGRDERR; BO_UNLOCK(bp->b_bufobj); bdirty(bp); } if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && (bp->b_flags & B_INVALONERR)) { /* * Forced invalidation of dirty buffer contents, to be used * after a failed write in the rare case that the loss of the * contents is acceptable. The buffer is invalidated and * freed. */ bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE; bp->b_flags &= ~(B_ASYNC | B_CACHE); } if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) && !(bp->b_flags & B_INVAL)) { /* * Failed write, redirty. All errors except ENXIO (which * means the device is gone) are treated as being * transient. * * XXX Treating EIO as transient is not correct; the * contract with the local storage device drivers is that * they will only return EIO once the I/O is no longer * retriable. Network I/O also respects this through the * guarantees of TCP and/or the internal retries of NFS. * ENOMEM might be transient, but we also have no way of * knowing when its ok to retry/reschedule. In general, * this entire case should be made obsolete through better * error handling/recovery and resource scheduling. * * Do this also for buffers that failed with ENXIO, but have * non-empty dependencies - the soft updates code might need * to access the buffer to untangle them. * * Must clear BIO_ERROR to prevent pages from being scrapped. */ bp->b_ioflags &= ~BIO_ERROR; bdirty(bp); } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { /* * Either a failed read I/O, or we were asked to free or not * cache the buffer, or we failed to write to a device that's * no longer present. */ bp->b_flags |= B_INVAL; if (!LIST_EMPTY(&bp->b_dep)) buf_deallocate(bp); if (bp->b_flags & B_DELWRI) bdirtysub(bp); bp->b_flags &= ~(B_DELWRI | B_CACHE); if ((bp->b_flags & B_VMIO) == 0) { allocbuf(bp, 0); if (bp->b_vp) brelvp(bp); } } /* * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_truncate() * is called with B_DELWRI set, the underlying pages may wind up * getting freed causing a previous write (bdwrite()) to get 'lost' * because pages associated with a B_DELWRI bp are marked clean. * * We still allow the B_INVAL case to call vfs_vmio_truncate(), even * if B_DELWRI is set. */ if (bp->b_flags & B_DELWRI) bp->b_flags &= ~B_RELBUF; /* * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer * constituted, not even NFS buffers now. Two flags effect this. If * B_INVAL, the struct buf is invalidated but the VM object is kept * around ( i.e. so it is trivial to reconstitute the buffer later ). * * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be * invalidated. BIO_ERROR cannot be set for a failed write unless the * buffer is also B_INVAL because it hits the re-dirtying code above. * * Normally we can do this whether a buffer is B_DELWRI or not. If * the buffer is an NFS buffer, it is tracking piecemeal writes or * the commit state and we cannot afford to lose the buffer. If the * buffer has a background write in progress, we need to keep it * around to prevent it from being reconstituted and starting a second * background write. */ v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL; if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE || (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) && (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 || vn_isdisk(bp->b_vp, NULL) || (bp->b_flags & B_DELWRI) == 0)) { vfs_vmio_invalidate(bp); allocbuf(bp, 0); } if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 || (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) { allocbuf(bp, 0); bp->b_flags &= ~B_NOREUSE; if (bp->b_vp != NULL) brelvp(bp); } /* * If the buffer has junk contents signal it and eventually * clean up B_DELWRI and diassociate the vnode so that gbincore() * doesn't find it. */ if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 || (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0) bp->b_flags |= B_INVAL; if (bp->b_flags & B_INVAL) { if (bp->b_flags & B_DELWRI) bundirty(bp); if (bp->b_vp) brelvp(bp); } buf_track(bp, __func__); /* buffers with no memory */ if (bp->b_bufsize == 0) { buf_free(bp); return; } /* buffers with junk contents */ if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || (bp->b_ioflags & BIO_ERROR)) { bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); if (bp->b_vflags & BV_BKGRDINPROG) panic("losing buffer 2"); qindex = QUEUE_CLEAN; bp->b_flags |= B_AGE; /* remaining buffers */ } else if (bp->b_flags & B_DELWRI) qindex = QUEUE_DIRTY; else qindex = QUEUE_CLEAN; if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) panic("brelse: not dirty"); bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT); /* binsfree unlocks bp. */ binsfree(bp, qindex); } /* * Release a buffer back to the appropriate queue but do not try to free * it. The buffer is expected to be used again soon. * * bqrelse() is used by bdwrite() to requeue a delayed write, and used by * biodone() to requeue an async I/O on completion. It is also used when * known good buffers need to be requeued but we think we may need the data * again soon. * * XXX we should be able to leave the B_RELBUF hint set on completion. */ void bqrelse(struct buf *bp) { int qindex; CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); qindex = QUEUE_NONE; if (BUF_LOCKRECURSED(bp)) { /* do not release to free list */ BUF_UNLOCK(bp); return; } bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); if (bp->b_flags & B_MANAGED) { if (bp->b_flags & B_REMFREE) bremfreef(bp); goto out; } /* buffers with stale but valid contents */ if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) { BO_LOCK(bp->b_bufobj); bp->b_vflags &= ~BV_BKGRDERR; BO_UNLOCK(bp->b_bufobj); qindex = QUEUE_DIRTY; } else { if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) panic("bqrelse: not dirty"); if ((bp->b_flags & B_NOREUSE) != 0) { brelse(bp); return; } qindex = QUEUE_CLEAN; } buf_track(bp, __func__); /* binsfree unlocks bp. */ binsfree(bp, qindex); return; out: buf_track(bp, __func__); /* unlock */ BUF_UNLOCK(bp); } /* * Complete I/O to a VMIO backed page. Validate the pages as appropriate, * restore bogus pages. */ static void vfs_vmio_iodone(struct buf *bp) { vm_ooffset_t foff; vm_page_t m; vm_object_t obj; struct vnode *vp __unused; int i, iosize, resid; bool bogus; obj = bp->b_bufobj->bo_object; KASSERT(REFCOUNT_COUNT(obj->paging_in_progress) >= bp->b_npages, ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)", REFCOUNT_COUNT(obj->paging_in_progress), bp->b_npages)); vp = bp->b_vp; KASSERT(vp->v_holdcnt > 0, ("vfs_vmio_iodone: vnode %p has zero hold count", vp)); KASSERT(vp->v_object != NULL, ("vfs_vmio_iodone: vnode %p has no vm_object", vp)); foff = bp->b_offset; KASSERT(bp->b_offset != NOOFFSET, ("vfs_vmio_iodone: bp %p has no buffer offset", bp)); bogus = false; iosize = bp->b_bcount - bp->b_resid; VM_OBJECT_WLOCK(obj); for (i = 0; i < bp->b_npages; i++) { resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; if (resid > iosize) resid = iosize; /* * cleanup bogus pages, restoring the originals */ m = bp->b_pages[i]; if (m == bogus_page) { bogus = true; m = vm_page_lookup(obj, OFF_TO_IDX(foff)); if (m == NULL) panic("biodone: page disappeared!"); bp->b_pages[i] = m; } else if ((bp->b_iocmd == BIO_READ) && resid > 0) { /* * In the write case, the valid and clean bits are * already changed correctly ( see bdwrite() ), so we * only need to do this here in the read case. */ KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK, resid)) == 0, ("vfs_vmio_iodone: page %p " "has unexpected dirty bits", m)); vfs_page_set_valid(bp, foff, m); } KASSERT(OFF_TO_IDX(foff) == m->pindex, ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch", (intmax_t)foff, (uintmax_t)m->pindex)); vm_page_sunbusy(m); foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; iosize -= resid; } vm_object_pip_wakeupn(obj, bp->b_npages); VM_OBJECT_WUNLOCK(obj); if (bogus && buf_mapped(bp)) { BUF_CHECK_MAPPED(bp); pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); } } /* * Perform page invalidation when a buffer is released. The fully invalid * pages will be reclaimed later in vfs_vmio_truncate(). */ static void vfs_vmio_invalidate(struct buf *bp) { vm_object_t obj; vm_page_t m; int flags, i, resid, poffset, presid; if (buf_mapped(bp)) { BUF_CHECK_MAPPED(bp); pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages); } else BUF_CHECK_UNMAPPED(bp); /* * Get the base offset and length of the buffer. Note that * in the VMIO case if the buffer block size is not * page-aligned then b_data pointer may not be page-aligned. * But our b_pages[] array *IS* page aligned. * * block sizes less then DEV_BSIZE (usually 512) are not * supported due to the page granularity bits (m->valid, * m->dirty, etc...). * * See man buf(9) for more information */ flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0; obj = bp->b_bufobj->bo_object; resid = bp->b_bufsize; poffset = bp->b_offset & PAGE_MASK; VM_OBJECT_WLOCK(obj); for (i = 0; i < bp->b_npages; i++) { m = bp->b_pages[i]; if (m == bogus_page) panic("vfs_vmio_invalidate: Unexpected bogus page."); bp->b_pages[i] = NULL; presid = resid > (PAGE_SIZE - poffset) ? (PAGE_SIZE - poffset) : resid; KASSERT(presid >= 0, ("brelse: extra page")); vm_page_busy_acquire(m, VM_ALLOC_SBUSY); if (pmap_page_wired_mappings(m) == 0) vm_page_set_invalid(m, poffset, presid); vm_page_sunbusy(m); vm_page_release_locked(m, flags); resid -= presid; poffset = 0; } VM_OBJECT_WUNLOCK(obj); bp->b_npages = 0; } /* * Page-granular truncation of an existing VMIO buffer. */ static void vfs_vmio_truncate(struct buf *bp, int desiredpages) { vm_object_t obj; vm_page_t m; int flags, i; if (bp->b_npages == desiredpages) return; if (buf_mapped(bp)) { BUF_CHECK_MAPPED(bp); pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) + (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages); } else BUF_CHECK_UNMAPPED(bp); /* * The object lock is needed only if we will attempt to free pages. */ flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0; if ((bp->b_flags & B_DIRECT) != 0) { flags |= VPR_TRYFREE; obj = bp->b_bufobj->bo_object; VM_OBJECT_WLOCK(obj); } else { obj = NULL; } for (i = desiredpages; i < bp->b_npages; i++) { m = bp->b_pages[i]; KASSERT(m != bogus_page, ("allocbuf: bogus page found")); bp->b_pages[i] = NULL; if (obj != NULL) vm_page_release_locked(m, flags); else vm_page_release(m, flags); } if (obj != NULL) VM_OBJECT_WUNLOCK(obj); bp->b_npages = desiredpages; } /* * Byte granular extension of VMIO buffers. */ static void vfs_vmio_extend(struct buf *bp, int desiredpages, int size) { /* * We are growing the buffer, possibly in a * byte-granular fashion. */ vm_object_t obj; vm_offset_t toff; vm_offset_t tinc; vm_page_t m; /* * Step 1, bring in the VM pages from the object, allocating * them if necessary. We must clear B_CACHE if these pages * are not valid for the range covered by the buffer. */ obj = bp->b_bufobj->bo_object; VM_OBJECT_WLOCK(obj); if (bp->b_npages < desiredpages) { /* * We must allocate system pages since blocking * here could interfere with paging I/O, no * matter which process we are. * * Only exclusive busy can be tested here. * Blocking on shared busy might lead to * deadlocks once allocbuf() is called after * pages are vfs_busy_pages(). */ (void)vm_page_grab_pages(obj, OFF_TO_IDX(bp->b_offset) + bp->b_npages, VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY | VM_ALLOC_NOBUSY | VM_ALLOC_WIRED, &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages); bp->b_npages = desiredpages; } /* * Step 2. We've loaded the pages into the buffer, * we have to figure out if we can still have B_CACHE * set. Note that B_CACHE is set according to the * byte-granular range ( bcount and size ), not the * aligned range ( newbsize ). * * The VM test is against m->valid, which is DEV_BSIZE * aligned. Needless to say, the validity of the data * needs to also be DEV_BSIZE aligned. Note that this * fails with NFS if the server or some other client * extends the file's EOF. If our buffer is resized, * B_CACHE may remain set! XXX */ toff = bp->b_bcount; tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); while ((bp->b_flags & B_CACHE) && toff < size) { vm_pindex_t pi; if (tinc > (size - toff)) tinc = size - toff; pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT; m = bp->b_pages[pi]; vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m); toff += tinc; tinc = PAGE_SIZE; } VM_OBJECT_WUNLOCK(obj); /* * Step 3, fixup the KVA pmap. */ if (buf_mapped(bp)) bpmap_qenter(bp); else BUF_CHECK_UNMAPPED(bp); } /* * Check to see if a block at a particular lbn is available for a clustered * write. */ static int vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) { struct buf *bpa; int match; match = 0; /* If the buf isn't in core skip it */ if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) return (0); /* If the buf is busy we don't want to wait for it */ if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) return (0); /* Only cluster with valid clusterable delayed write buffers */ if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != (B_DELWRI | B_CLUSTEROK)) goto done; if (bpa->b_bufsize != size) goto done; /* * Check to see if it is in the expected place on disk and that the * block has been mapped. */ if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) match = 1; done: BUF_UNLOCK(bpa); return (match); } /* * vfs_bio_awrite: * * Implement clustered async writes for clearing out B_DELWRI buffers. * This is much better then the old way of writing only one buffer at * a time. Note that we may not be presented with the buffers in the * correct order, so we search for the cluster in both directions. */ int vfs_bio_awrite(struct buf *bp) { struct bufobj *bo; int i; int j; daddr_t lblkno = bp->b_lblkno; struct vnode *vp = bp->b_vp; int ncl; int nwritten; int size; int maxcl; int gbflags; bo = &vp->v_bufobj; gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0; /* * right now we support clustered writing only to regular files. If * we find a clusterable block we could be in the middle of a cluster * rather then at the beginning. */ if ((vp->v_type == VREG) && (vp->v_mount != 0) && /* Only on nodes that have the size info */ (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { size = vp->v_mount->mnt_stat.f_iosize; maxcl = MAXPHYS / size; BO_RLOCK(bo); for (i = 1; i < maxcl; i++) if (vfs_bio_clcheck(vp, size, lblkno + i, bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) break; for (j = 1; i + j <= maxcl && j <= lblkno; j++) if (vfs_bio_clcheck(vp, size, lblkno - j, bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) break; BO_RUNLOCK(bo); --j; ncl = i + j; /* * this is a possible cluster write */ if (ncl != 1) { BUF_UNLOCK(bp); nwritten = cluster_wbuild(vp, size, lblkno - j, ncl, gbflags); return (nwritten); } } bremfree(bp); bp->b_flags |= B_ASYNC; /* * default (old) behavior, writing out only one block * * XXX returns b_bufsize instead of b_bcount for nwritten? */ nwritten = bp->b_bufsize; (void) bwrite(bp); return (nwritten); } /* * getnewbuf_kva: * * Allocate KVA for an empty buf header according to gbflags. */ static int getnewbuf_kva(struct buf *bp, int gbflags, int maxsize) { if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) { /* * In order to keep fragmentation sane we only allocate kva * in BKVASIZE chunks. XXX with vmem we can do page size. */ maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; if (maxsize != bp->b_kvasize && bufkva_alloc(bp, maxsize, gbflags)) return (ENOSPC); } return (0); } /* * getnewbuf: * * Find and initialize a new buffer header, freeing up existing buffers * in the bufqueues as necessary. The new buffer is returned locked. * * We block if: * We have insufficient buffer headers * We have insufficient buffer space * buffer_arena is too fragmented ( space reservation fails ) * If we have to flush dirty buffers ( but we try to avoid this ) * * The caller is responsible for releasing the reserved bufspace after * allocbuf() is called. */ static struct buf * getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags) { struct bufdomain *bd; struct buf *bp; bool metadata, reserved; bp = NULL; KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); if (!unmapped_buf_allowed) gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC); if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 || vp->v_type == VCHR) metadata = true; else metadata = false; if (vp == NULL) bd = &bdomain[0]; else bd = &bdomain[vp->v_bufobj.bo_domain]; counter_u64_add(getnewbufcalls, 1); reserved = false; do { if (reserved == false && bufspace_reserve(bd, maxsize, metadata) != 0) { counter_u64_add(getnewbufrestarts, 1); continue; } reserved = true; if ((bp = buf_alloc(bd)) == NULL) { counter_u64_add(getnewbufrestarts, 1); continue; } if (getnewbuf_kva(bp, gbflags, maxsize) == 0) return (bp); break; } while (buf_recycle(bd, false) == 0); if (reserved) bufspace_release(bd, maxsize); if (bp != NULL) { bp->b_flags |= B_INVAL; brelse(bp); } bufspace_wait(bd, vp, gbflags, slpflag, slptimeo); return (NULL); } /* * buf_daemon: * * buffer flushing daemon. Buffers are normally flushed by the * update daemon but if it cannot keep up this process starts to * take the load in an attempt to prevent getnewbuf() from blocking. */ static struct kproc_desc buf_kp = { "bufdaemon", buf_daemon, &bufdaemonproc }; SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); static int buf_flush(struct vnode *vp, struct bufdomain *bd, int target) { int flushed; flushed = flushbufqueues(vp, bd, target, 0); if (flushed == 0) { /* * Could not find any buffers without rollback * dependencies, so just write the first one * in the hopes of eventually making progress. */ if (vp != NULL && target > 2) target /= 2; flushbufqueues(vp, bd, target, 1); } return (flushed); } static void buf_daemon() { struct bufdomain *bd; int speedupreq; int lodirty; int i; /* * This process needs to be suspended prior to shutdown sync. */ EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread, SHUTDOWN_PRI_LAST + 100); /* * Start the buf clean daemons as children threads. */ for (i = 0 ; i < buf_domains; i++) { int error; error = kthread_add((void (*)(void *))bufspace_daemon, &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i); if (error) panic("error %d spawning bufspace daemon", error); } /* * This process is allowed to take the buffer cache to the limit */ curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; mtx_lock(&bdlock); for (;;) { bd_request = 0; mtx_unlock(&bdlock); kthread_suspend_check(); /* * Save speedupreq for this pass and reset to capture new * requests. */ speedupreq = bd_speedupreq; bd_speedupreq = 0; /* * Flush each domain sequentially according to its level and * the speedup request. */ for (i = 0; i < buf_domains; i++) { bd = &bdomain[i]; if (speedupreq) lodirty = bd->bd_numdirtybuffers / 2; else lodirty = bd->bd_lodirtybuffers; while (bd->bd_numdirtybuffers > lodirty) { if (buf_flush(NULL, bd, bd->bd_numdirtybuffers - lodirty) == 0) break; kern_yield(PRI_USER); } } /* * Only clear bd_request if we have reached our low water * mark. The buf_daemon normally waits 1 second and * then incrementally flushes any dirty buffers that have * built up, within reason. * * If we were unable to hit our low water mark and couldn't * find any flushable buffers, we sleep for a short period * to avoid endless loops on unlockable buffers. */ mtx_lock(&bdlock); if (!BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) { /* * We reached our low water mark, reset the * request and sleep until we are needed again. * The sleep is just so the suspend code works. */ bd_request = 0; /* * Do an extra wakeup in case dirty threshold * changed via sysctl and the explicit transition * out of shortfall was missed. */ bdirtywakeup(); if (runningbufspace <= lorunningspace) runningwakeup(); msleep(&bd_request, &bdlock, PVM, "psleep", hz); } else { /* * We couldn't find any flushable dirty buffers but * still have too many dirty buffers, we * have to sleep and try again. (rare) */ msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); } } } /* * flushbufqueues: * * Try to flush a buffer in the dirty queue. We must be careful to * free up B_INVAL buffers instead of write them, which NFS is * particularly sensitive to. */ static int flushwithdeps = 0; SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW | CTLFLAG_STATS, &flushwithdeps, 0, "Number of buffers flushed with dependecies that require rollbacks"); static int flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target, int flushdeps) { struct bufqueue *bq; struct buf *sentinel; struct vnode *vp; struct mount *mp; struct buf *bp; int hasdeps; int flushed; int error; bool unlock; flushed = 0; bq = &bd->bd_dirtyq; bp = NULL; sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); sentinel->b_qindex = QUEUE_SENTINEL; BQ_LOCK(bq); TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist); BQ_UNLOCK(bq); while (flushed != target) { maybe_yield(); BQ_LOCK(bq); bp = TAILQ_NEXT(sentinel, b_freelist); if (bp != NULL) { TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist); TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel, b_freelist); } else { BQ_UNLOCK(bq); break; } /* * Skip sentinels inserted by other invocations of the * flushbufqueues(), taking care to not reorder them. * * Only flush the buffers that belong to the * vnode locked by the curthread. */ if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL && bp->b_vp != lvp)) { BQ_UNLOCK(bq); continue; } error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL); BQ_UNLOCK(bq); if (error != 0) continue; /* * BKGRDINPROG can only be set with the buf and bufobj * locks both held. We tolerate a race to clear it here. */ if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || (bp->b_flags & B_DELWRI) == 0) { BUF_UNLOCK(bp); continue; } if (bp->b_flags & B_INVAL) { bremfreef(bp); brelse(bp); flushed++; continue; } if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { if (flushdeps == 0) { BUF_UNLOCK(bp); continue; } hasdeps = 1; } else hasdeps = 0; /* * We must hold the lock on a vnode before writing * one of its buffers. Otherwise we may confuse, or * in the case of a snapshot vnode, deadlock the * system. * * The lock order here is the reverse of the normal * of vnode followed by buf lock. This is ok because * the NOWAIT will prevent deadlock. */ vp = bp->b_vp; if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { BUF_UNLOCK(bp); continue; } if (lvp == NULL) { unlock = true; error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT); } else { ASSERT_VOP_LOCKED(vp, "getbuf"); unlock = false; error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 : vn_lock(vp, LK_TRYUPGRADE); } if (error == 0) { CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); if (curproc == bufdaemonproc) { vfs_bio_awrite(bp); } else { bremfree(bp); bwrite(bp); counter_u64_add(notbufdflushes, 1); } vn_finished_write(mp); if (unlock) VOP_UNLOCK(vp, 0); flushwithdeps += hasdeps; flushed++; /* * Sleeping on runningbufspace while holding * vnode lock leads to deadlock. */ if (curproc == bufdaemonproc && runningbufspace > hirunningspace) waitrunningbufspace(); continue; } vn_finished_write(mp); BUF_UNLOCK(bp); } BQ_LOCK(bq); TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist); BQ_UNLOCK(bq); free(sentinel, M_TEMP); return (flushed); } /* * Check to see if a block is currently memory resident. */ struct buf * incore(struct bufobj *bo, daddr_t blkno) { struct buf *bp; BO_RLOCK(bo); bp = gbincore(bo, blkno); BO_RUNLOCK(bo); return (bp); } /* * Returns true if no I/O is needed to access the * associated VM object. This is like incore except * it also hunts around in the VM system for the data. */ static int inmem(struct vnode * vp, daddr_t blkno) { vm_object_t obj; vm_offset_t toff, tinc, size; vm_page_t m; vm_ooffset_t off; ASSERT_VOP_LOCKED(vp, "inmem"); if (incore(&vp->v_bufobj, blkno)) return 1; if (vp->v_mount == NULL) return 0; obj = vp->v_object; if (obj == NULL) return (0); size = PAGE_SIZE; if (size > vp->v_mount->mnt_stat.f_iosize) size = vp->v_mount->mnt_stat.f_iosize; off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; VM_OBJECT_RLOCK(obj); for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); if (!m) goto notinmem; tinc = size; if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); if (vm_page_is_valid(m, (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) goto notinmem; } VM_OBJECT_RUNLOCK(obj); return 1; notinmem: VM_OBJECT_RUNLOCK(obj); return (0); } /* * Set the dirty range for a buffer based on the status of the dirty * bits in the pages comprising the buffer. The range is limited * to the size of the buffer. * * Tell the VM system that the pages associated with this buffer * are clean. This is used for delayed writes where the data is * going to go to disk eventually without additional VM intevention. * * Note that while we only really need to clean through to b_bcount, we * just go ahead and clean through to b_bufsize. */ static void vfs_clean_pages_dirty_buf(struct buf *bp) { vm_ooffset_t foff, noff, eoff; vm_page_t m; int i; if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0) return; foff = bp->b_offset; KASSERT(bp->b_offset != NOOFFSET, ("vfs_clean_pages_dirty_buf: no buffer offset")); VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); vfs_busy_pages_acquire(bp); vfs_setdirty_locked_object(bp); for (i = 0; i < bp->b_npages; i++) { noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; eoff = noff; if (eoff > bp->b_offset + bp->b_bufsize) eoff = bp->b_offset + bp->b_bufsize; m = bp->b_pages[i]; vfs_page_set_validclean(bp, foff, m); /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ foff = noff; } vfs_busy_pages_release(bp); VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); } static void vfs_setdirty_locked_object(struct buf *bp) { vm_object_t object; int i; object = bp->b_bufobj->bo_object; VM_OBJECT_ASSERT_WLOCKED(object); /* * We qualify the scan for modified pages on whether the * object has been flushed yet. */ if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) { vm_offset_t boffset; vm_offset_t eoffset; /* * test the pages to see if they have been modified directly * by users through the VM system. */ for (i = 0; i < bp->b_npages; i++) vm_page_test_dirty(bp->b_pages[i]); /* * Calculate the encompassing dirty range, boffset and eoffset, * (eoffset - boffset) bytes. */ for (i = 0; i < bp->b_npages; i++) { if (bp->b_pages[i]->dirty) break; } boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); for (i = bp->b_npages - 1; i >= 0; --i) { if (bp->b_pages[i]->dirty) { break; } } eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); /* * Fit it to the buffer. */ if (eoffset > bp->b_bcount) eoffset = bp->b_bcount; /* * If we have a good dirty range, merge with the existing * dirty range. */ if (boffset < eoffset) { if (bp->b_dirtyoff > boffset) bp->b_dirtyoff = boffset; if (bp->b_dirtyend < eoffset) bp->b_dirtyend = eoffset; } } } /* * Allocate the KVA mapping for an existing buffer. * If an unmapped buffer is provided but a mapped buffer is requested, take * also care to properly setup mappings between pages and KVA. */ static void bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags) { int bsize, maxsize, need_mapping, need_kva; off_t offset; need_mapping = bp->b_data == unmapped_buf && (gbflags & GB_UNMAPPED) == 0; need_kva = bp->b_kvabase == unmapped_buf && bp->b_data == unmapped_buf && (gbflags & GB_KVAALLOC) != 0; if (!need_mapping && !need_kva) return; BUF_CHECK_UNMAPPED(bp); if (need_mapping && bp->b_kvabase != unmapped_buf) { /* * Buffer is not mapped, but the KVA was already * reserved at the time of the instantiation. Use the * allocated space. */ goto has_addr; } /* * Calculate the amount of the address space we would reserve * if the buffer was mapped. */ bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize; KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize")); offset = blkno * bsize; maxsize = size + (offset & PAGE_MASK); maxsize = imax(maxsize, bsize); while (bufkva_alloc(bp, maxsize, gbflags) != 0) { if ((gbflags & GB_NOWAIT_BD) != 0) { /* * XXXKIB: defragmentation cannot * succeed, not sure what else to do. */ panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp); } counter_u64_add(mappingrestarts, 1); bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0); } has_addr: if (need_mapping) { /* b_offset is handled by bpmap_qenter. */ bp->b_data = bp->b_kvabase; BUF_CHECK_MAPPED(bp); bpmap_qenter(bp); } } struct buf * getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo, int flags) { struct buf *bp; int error; error = getblkx(vp, blkno, size, slpflag, slptimeo, flags, &bp); if (error != 0) return (NULL); return (bp); } /* * getblkx: * * Get a block given a specified block and offset into a file/device. * The buffers B_DONE bit will be cleared on return, making it almost * ready for an I/O initiation. B_INVAL may or may not be set on * return. The caller should clear B_INVAL prior to initiating a * READ. * * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for * an existing buffer. * * For a VMIO buffer, B_CACHE is modified according to the backing VM. * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set * and then cleared based on the backing VM. If the previous buffer is * non-0-sized but invalid, B_CACHE will be cleared. * * If getblk() must create a new buffer, the new buffer is returned with * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which * case it is returned with B_INVAL clear and B_CACHE set based on the * backing VM. * * getblk() also forces a bwrite() for any B_DELWRI buffer whos * B_CACHE bit is clear. * * What this means, basically, is that the caller should use B_CACHE to * determine whether the buffer is fully valid or not and should clear * B_INVAL prior to issuing a read. If the caller intends to validate * the buffer by loading its data area with something, the caller needs * to clear B_INVAL. If the caller does this without issuing an I/O, * the caller should set B_CACHE ( as an optimization ), else the caller * should issue the I/O and biodone() will set B_CACHE if the I/O was * a write attempt or if it was a successful read. If the caller * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR * prior to issuing the READ. biodone() will *not* clear B_INVAL. */ int getblkx(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo, int flags, struct buf **bpp) { struct buf *bp; struct bufobj *bo; daddr_t d_blkno; int bsize, error, maxsize, vmio; off_t offset; CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); ASSERT_VOP_LOCKED(vp, "getblk"); if (size > maxbcachebuf) panic("getblk: size(%d) > maxbcachebuf(%d)\n", size, maxbcachebuf); if (!unmapped_buf_allowed) flags &= ~(GB_UNMAPPED | GB_KVAALLOC); bo = &vp->v_bufobj; d_blkno = blkno; loop: BO_RLOCK(bo); bp = gbincore(bo, blkno); if (bp != NULL) { int lockflags; /* * Buffer is in-core. If the buffer is not busy nor managed, * it must be on a queue. */ lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; if ((flags & GB_LOCK_NOWAIT) != 0) lockflags |= LK_NOWAIT; error = BUF_TIMELOCK(bp, lockflags, BO_LOCKPTR(bo), "getblk", slpflag, slptimeo); /* * If we slept and got the lock we have to restart in case * the buffer changed identities. */ if (error == ENOLCK) goto loop; /* We timed out or were interrupted. */ else if (error != 0) return (error); /* If recursed, assume caller knows the rules. */ else if (BUF_LOCKRECURSED(bp)) goto end; /* * The buffer is locked. B_CACHE is cleared if the buffer is * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set * and for a VMIO buffer B_CACHE is adjusted according to the * backing VM cache. */ if (bp->b_flags & B_INVAL) bp->b_flags &= ~B_CACHE; else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) bp->b_flags |= B_CACHE; if (bp->b_flags & B_MANAGED) MPASS(bp->b_qindex == QUEUE_NONE); else bremfree(bp); /* * check for size inconsistencies for non-VMIO case. */ if (bp->b_bcount != size) { if ((bp->b_flags & B_VMIO) == 0 || (size > bp->b_kvasize)) { if (bp->b_flags & B_DELWRI) { bp->b_flags |= B_NOCACHE; bwrite(bp); } else { if (LIST_EMPTY(&bp->b_dep)) { bp->b_flags |= B_RELBUF; brelse(bp); } else { bp->b_flags |= B_NOCACHE; bwrite(bp); } } goto loop; } } /* * Handle the case of unmapped buffer which should * become mapped, or the buffer for which KVA * reservation is requested. */ bp_unmapped_get_kva(bp, blkno, size, flags); /* * If the size is inconsistent in the VMIO case, we can resize * the buffer. This might lead to B_CACHE getting set or * cleared. If the size has not changed, B_CACHE remains * unchanged from its previous state. */ allocbuf(bp, size); KASSERT(bp->b_offset != NOOFFSET, ("getblk: no buffer offset")); /* * A buffer with B_DELWRI set and B_CACHE clear must * be committed before we can return the buffer in * order to prevent the caller from issuing a read * ( due to B_CACHE not being set ) and overwriting * it. * * Most callers, including NFS and FFS, need this to * operate properly either because they assume they * can issue a read if B_CACHE is not set, or because * ( for example ) an uncached B_DELWRI might loop due * to softupdates re-dirtying the buffer. In the latter * case, B_CACHE is set after the first write completes, * preventing further loops. * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE * above while extending the buffer, we cannot allow the * buffer to remain with B_CACHE set after the write * completes or it will represent a corrupt state. To * deal with this we set B_NOCACHE to scrap the buffer * after the write. * * We might be able to do something fancy, like setting * B_CACHE in bwrite() except if B_DELWRI is already set, * so the below call doesn't set B_CACHE, but that gets real * confusing. This is much easier. */ if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { bp->b_flags |= B_NOCACHE; bwrite(bp); goto loop; } bp->b_flags &= ~B_DONE; } else { /* * Buffer is not in-core, create new buffer. The buffer * returned by getnewbuf() is locked. Note that the returned * buffer is also considered valid (not marked B_INVAL). */ BO_RUNLOCK(bo); /* * If the user does not want us to create the buffer, bail out * here. */ if (flags & GB_NOCREAT) return (EEXIST); bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize; KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize")); offset = blkno * bsize; vmio = vp->v_object != NULL; if (vmio) { maxsize = size + (offset & PAGE_MASK); } else { maxsize = size; /* Do not allow non-VMIO notmapped buffers. */ flags &= ~(GB_UNMAPPED | GB_KVAALLOC); } maxsize = imax(maxsize, bsize); if ((flags & GB_NOSPARSE) != 0 && vmio && !vn_isdisk(vp, NULL)) { error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0); KASSERT(error != EOPNOTSUPP, ("GB_NOSPARSE from fs not supporting bmap, vp %p", vp)); if (error != 0) return (error); if (d_blkno == -1) return (EJUSTRETURN); } bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags); if (bp == NULL) { if (slpflag || slptimeo) return (ETIMEDOUT); /* * XXX This is here until the sleep path is diagnosed * enough to work under very low memory conditions. * * There's an issue on low memory, 4BSD+non-preempt * systems (eg MIPS routers with 32MB RAM) where buffer * exhaustion occurs without sleeping for buffer * reclaimation. This just sticks in a loop and * constantly attempts to allocate a buffer, which * hits exhaustion and tries to wakeup bufdaemon. * This never happens because we never yield. * * The real solution is to identify and fix these cases * so we aren't effectively busy-waiting in a loop * until the reclaimation path has cycles to run. */ kern_yield(PRI_USER); goto loop; } /* * This code is used to make sure that a buffer is not * created while the getnewbuf routine is blocked. * This can be a problem whether the vnode is locked or not. * If the buffer is created out from under us, we have to * throw away the one we just created. * * Note: this must occur before we associate the buffer * with the vp especially considering limitations in * the splay tree implementation when dealing with duplicate * lblkno's. */ BO_LOCK(bo); if (gbincore(bo, blkno)) { BO_UNLOCK(bo); bp->b_flags |= B_INVAL; bufspace_release(bufdomain(bp), maxsize); brelse(bp); goto loop; } /* * Insert the buffer into the hash, so that it can * be found by incore. */ bp->b_lblkno = blkno; bp->b_blkno = d_blkno; bp->b_offset = offset; bgetvp(vp, bp); BO_UNLOCK(bo); /* * set B_VMIO bit. allocbuf() the buffer bigger. Since the * buffer size starts out as 0, B_CACHE will be set by * allocbuf() for the VMIO case prior to it testing the * backing store for validity. */ if (vmio) { bp->b_flags |= B_VMIO; KASSERT(vp->v_object == bp->b_bufobj->bo_object, ("ARGH! different b_bufobj->bo_object %p %p %p\n", bp, vp->v_object, bp->b_bufobj->bo_object)); } else { bp->b_flags &= ~B_VMIO; KASSERT(bp->b_bufobj->bo_object == NULL, ("ARGH! has b_bufobj->bo_object %p %p\n", bp, bp->b_bufobj->bo_object)); BUF_CHECK_MAPPED(bp); } allocbuf(bp, size); bufspace_release(bufdomain(bp), maxsize); bp->b_flags &= ~B_DONE; } CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); end: buf_track(bp, __func__); KASSERT(bp->b_bufobj == bo, ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); *bpp = bp; return (0); } /* * Get an empty, disassociated buffer of given size. The buffer is initially * set to B_INVAL. */ struct buf * geteblk(int size, int flags) { struct buf *bp; int maxsize; maxsize = (size + BKVAMASK) & ~BKVAMASK; while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) { if ((flags & GB_NOWAIT_BD) && (curthread->td_pflags & TDP_BUFNEED) != 0) return (NULL); } allocbuf(bp, size); bufspace_release(bufdomain(bp), maxsize); bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ return (bp); } /* * Truncate the backing store for a non-vmio buffer. */ static void vfs_nonvmio_truncate(struct buf *bp, int newbsize) { if (bp->b_flags & B_MALLOC) { /* * malloced buffers are not shrunk */ if (newbsize == 0) { bufmallocadjust(bp, 0); free(bp->b_data, M_BIOBUF); bp->b_data = bp->b_kvabase; bp->b_flags &= ~B_MALLOC; } return; } vm_hold_free_pages(bp, newbsize); bufspace_adjust(bp, newbsize); } /* * Extend the backing for a non-VMIO buffer. */ static void vfs_nonvmio_extend(struct buf *bp, int newbsize) { caddr_t origbuf; int origbufsize; /* * We only use malloced memory on the first allocation. * and revert to page-allocated memory when the buffer * grows. * * There is a potential smp race here that could lead * to bufmallocspace slightly passing the max. It * is probably extremely rare and not worth worrying * over. */ if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 && bufmallocspace < maxbufmallocspace) { bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK); bp->b_flags |= B_MALLOC; bufmallocadjust(bp, newbsize); return; } /* * If the buffer is growing on its other-than-first * allocation then we revert to the page-allocation * scheme. */ origbuf = NULL; origbufsize = 0; if (bp->b_flags & B_MALLOC) { origbuf = bp->b_data; origbufsize = bp->b_bufsize; bp->b_data = bp->b_kvabase; bufmallocadjust(bp, 0); bp->b_flags &= ~B_MALLOC; newbsize = round_page(newbsize); } vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize, (vm_offset_t) bp->b_data + newbsize); if (origbuf != NULL) { bcopy(origbuf, bp->b_data, origbufsize); free(origbuf, M_BIOBUF); } bufspace_adjust(bp, newbsize); } /* * This code constitutes the buffer memory from either anonymous system * memory (in the case of non-VMIO operations) or from an associated * VM object (in the case of VMIO operations). This code is able to * resize a buffer up or down. * * Note that this code is tricky, and has many complications to resolve * deadlock or inconsistent data situations. Tread lightly!!! * There are B_CACHE and B_DELWRI interactions that must be dealt with by * the caller. Calling this code willy nilly can result in the loss of data. * * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with * B_CACHE for the non-VMIO case. */ int allocbuf(struct buf *bp, int size) { int newbsize; if (bp->b_bcount == size) return (1); if (bp->b_kvasize != 0 && bp->b_kvasize < size) panic("allocbuf: buffer too small"); newbsize = roundup2(size, DEV_BSIZE); if ((bp->b_flags & B_VMIO) == 0) { if ((bp->b_flags & B_MALLOC) == 0) newbsize = round_page(newbsize); /* * Just get anonymous memory from the kernel. Don't * mess with B_CACHE. */ if (newbsize < bp->b_bufsize) vfs_nonvmio_truncate(bp, newbsize); else if (newbsize > bp->b_bufsize) vfs_nonvmio_extend(bp, newbsize); } else { int desiredpages; desiredpages = (size == 0) ? 0 : num_pages((bp->b_offset & PAGE_MASK) + newbsize); if (bp->b_flags & B_MALLOC) panic("allocbuf: VMIO buffer can't be malloced"); /* * Set B_CACHE initially if buffer is 0 length or will become * 0-length. */ if (size == 0 || bp->b_bufsize == 0) bp->b_flags |= B_CACHE; if (newbsize < bp->b_bufsize) vfs_vmio_truncate(bp, desiredpages); /* XXX This looks as if it should be newbsize > b_bufsize */ else if (size > bp->b_bcount) vfs_vmio_extend(bp, desiredpages, size); bufspace_adjust(bp, newbsize); } bp->b_bcount = size; /* requested buffer size. */ return (1); } extern int inflight_transient_maps; static struct bio_queue nondump_bios; void biodone(struct bio *bp) { struct mtx *mtxp; void (*done)(struct bio *); vm_offset_t start, end; biotrack(bp, __func__); /* * Avoid completing I/O when dumping after a panic since that may * result in a deadlock in the filesystem or pager code. Note that * this doesn't affect dumps that were started manually since we aim * to keep the system usable after it has been resumed. */ if (__predict_false(dumping && SCHEDULER_STOPPED())) { TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue); return; } if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) { bp->bio_flags &= ~BIO_TRANSIENT_MAPPING; bp->bio_flags |= BIO_UNMAPPED; start = trunc_page((vm_offset_t)bp->bio_data); end = round_page((vm_offset_t)bp->bio_data + bp->bio_length); bp->bio_data = unmapped_buf; pmap_qremove(start, atop(end - start)); vmem_free(transient_arena, start, end - start); atomic_add_int(&inflight_transient_maps, -1); } done = bp->bio_done; if (done == NULL) { mtxp = mtx_pool_find(mtxpool_sleep, bp); mtx_lock(mtxp); bp->bio_flags |= BIO_DONE; wakeup(bp); mtx_unlock(mtxp); } else done(bp); } /* * Wait for a BIO to finish. */ int biowait(struct bio *bp, const char *wchan) { struct mtx *mtxp; mtxp = mtx_pool_find(mtxpool_sleep, bp); mtx_lock(mtxp); while ((bp->bio_flags & BIO_DONE) == 0) msleep(bp, mtxp, PRIBIO, wchan, 0); mtx_unlock(mtxp); if (bp->bio_error != 0) return (bp->bio_error); if (!(bp->bio_flags & BIO_ERROR)) return (0); return (EIO); } void biofinish(struct bio *bp, struct devstat *stat, int error) { if (error) { bp->bio_error = error; bp->bio_flags |= BIO_ERROR; } if (stat != NULL) devstat_end_transaction_bio(stat, bp); biodone(bp); } #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING) void biotrack_buf(struct bio *bp, const char *location) { buf_track(bp->bio_track_bp, location); } #endif /* * bufwait: * * Wait for buffer I/O completion, returning error status. The buffer * is left locked and B_DONE on return. B_EINTR is converted into an EINTR * error and cleared. */ int bufwait(struct buf *bp) { if (bp->b_iocmd == BIO_READ) bwait(bp, PRIBIO, "biord"); else bwait(bp, PRIBIO, "biowr"); if (bp->b_flags & B_EINTR) { bp->b_flags &= ~B_EINTR; return (EINTR); } if (bp->b_ioflags & BIO_ERROR) { return (bp->b_error ? bp->b_error : EIO); } else { return (0); } } /* * bufdone: * * Finish I/O on a buffer, optionally calling a completion function. * This is usually called from an interrupt so process blocking is * not allowed. * * biodone is also responsible for setting B_CACHE in a B_VMIO bp. * In a non-VMIO bp, B_CACHE will be set on the next getblk() * assuming B_INVAL is clear. * * For the VMIO case, we set B_CACHE if the op was a read and no * read error occurred, or if the op was a write. B_CACHE is never * set if the buffer is invalid or otherwise uncacheable. * * bufdone does not mess with B_INVAL, allowing the I/O routine or the * initiator to leave B_INVAL set to brelse the buffer out of existence * in the biodone routine. */ void bufdone(struct buf *bp) { struct bufobj *dropobj; void (*biodone)(struct buf *); buf_track(bp, __func__); CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); dropobj = NULL; KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); runningbufwakeup(bp); if (bp->b_iocmd == BIO_WRITE) dropobj = bp->b_bufobj; /* call optional completion function if requested */ if (bp->b_iodone != NULL) { biodone = bp->b_iodone; bp->b_iodone = NULL; (*biodone) (bp); if (dropobj) bufobj_wdrop(dropobj); return; } if (bp->b_flags & B_VMIO) { /* * Set B_CACHE if the op was a normal read and no error * occurred. B_CACHE is set for writes in the b*write() * routines. */ if (bp->b_iocmd == BIO_READ && !(bp->b_flags & (B_INVAL|B_NOCACHE)) && !(bp->b_ioflags & BIO_ERROR)) bp->b_flags |= B_CACHE; vfs_vmio_iodone(bp); } if (!LIST_EMPTY(&bp->b_dep)) buf_complete(bp); if ((bp->b_flags & B_CKHASH) != 0) { KASSERT(bp->b_iocmd == BIO_READ, ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd)); KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp)); (*bp->b_ckhashcalc)(bp); } /* * For asynchronous completions, release the buffer now. The brelse * will do a wakeup there if necessary - so no need to do a wakeup * here in the async case. The sync case always needs to do a wakeup. */ if (bp->b_flags & B_ASYNC) { if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) brelse(bp); else bqrelse(bp); } else bdone(bp); if (dropobj) bufobj_wdrop(dropobj); } /* * This routine is called in lieu of iodone in the case of * incomplete I/O. This keeps the busy status for pages * consistent. */ void vfs_unbusy_pages(struct buf *bp) { int i; vm_object_t obj; vm_page_t m; runningbufwakeup(bp); if (!(bp->b_flags & B_VMIO)) return; obj = bp->b_bufobj->bo_object; VM_OBJECT_WLOCK(obj); for (i = 0; i < bp->b_npages; i++) { m = bp->b_pages[i]; if (m == bogus_page) { m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); if (!m) panic("vfs_unbusy_pages: page missing\n"); bp->b_pages[i] = m; if (buf_mapped(bp)) { BUF_CHECK_MAPPED(bp); pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); } else BUF_CHECK_UNMAPPED(bp); } vm_page_sunbusy(m); } vm_object_pip_wakeupn(obj, bp->b_npages); VM_OBJECT_WUNLOCK(obj); } /* * vfs_page_set_valid: * * Set the valid bits in a page based on the supplied offset. The * range is restricted to the buffer's size. * * This routine is typically called after a read completes. */ static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) { vm_ooffset_t eoff; /* * Compute the end offset, eoff, such that [off, eoff) does not span a * page boundary and eoff is not greater than the end of the buffer. * The end of the buffer, in this case, is our file EOF, not the * allocation size of the buffer. */ eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; if (eoff > bp->b_offset + bp->b_bcount) eoff = bp->b_offset + bp->b_bcount; /* * Set valid range. This is typically the entire buffer and thus the * entire page. */ if (eoff > off) vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off); } /* * vfs_page_set_validclean: * * Set the valid bits and clear the dirty bits in a page based on the * supplied offset. The range is restricted to the buffer's size. */ static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) { vm_ooffset_t soff, eoff; /* * Start and end offsets in buffer. eoff - soff may not cross a * page boundary or cross the end of the buffer. The end of the * buffer, in this case, is our file EOF, not the allocation size * of the buffer. */ soff = off; eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; if (eoff > bp->b_offset + bp->b_bcount) eoff = bp->b_offset + bp->b_bcount; /* * Set valid range. This is typically the entire buffer and thus the * entire page. */ if (eoff > soff) { vm_page_set_validclean( m, (vm_offset_t) (soff & PAGE_MASK), (vm_offset_t) (eoff - soff) ); } } /* * Acquire a shared busy on all pages in the buf. */ void vfs_busy_pages_acquire(struct buf *bp) { int i; VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object); for (i = 0; i < bp->b_npages; i++) vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY); } void vfs_busy_pages_release(struct buf *bp) { int i; VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object); for (i = 0; i < bp->b_npages; i++) vm_page_sunbusy(bp->b_pages[i]); } /* * This routine is called before a device strategy routine. * It is used to tell the VM system that paging I/O is in * progress, and treat the pages associated with the buffer * almost as being exclusive busy. Also the object paging_in_progress * flag is handled to make sure that the object doesn't become * inconsistent. * * Since I/O has not been initiated yet, certain buffer flags * such as BIO_ERROR or B_INVAL may be in an inconsistent state * and should be ignored. */ void vfs_busy_pages(struct buf *bp, int clear_modify) { vm_object_t obj; vm_ooffset_t foff; vm_page_t m; int i; bool bogus; if (!(bp->b_flags & B_VMIO)) return; obj = bp->b_bufobj->bo_object; foff = bp->b_offset; KASSERT(bp->b_offset != NOOFFSET, ("vfs_busy_pages: no buffer offset")); VM_OBJECT_WLOCK(obj); if ((bp->b_flags & B_CLUSTER) == 0) { vm_object_pip_add(obj, bp->b_npages); vfs_busy_pages_acquire(bp); } if (bp->b_bufsize != 0) vfs_setdirty_locked_object(bp); bogus = false; for (i = 0; i < bp->b_npages; i++) { m = bp->b_pages[i]; vm_page_assert_sbusied(m); /* * When readying a buffer for a read ( i.e * clear_modify == 0 ), it is important to do * bogus_page replacement for valid pages in * partially instantiated buffers. Partially * instantiated buffers can, in turn, occur when * reconstituting a buffer from its VM backing store * base. We only have to do this if B_CACHE is * clear ( which causes the I/O to occur in the * first place ). The replacement prevents the read * I/O from overwriting potentially dirty VM-backed * pages. XXX bogus page replacement is, uh, bogus. * It may not work properly with small-block devices. * We need to find a better way. */ if (clear_modify) { pmap_remove_write(m); vfs_page_set_validclean(bp, foff, m); - } else if (m->valid == VM_PAGE_BITS_ALL && + } else if (vm_page_all_valid(m) && (bp->b_flags & B_CACHE) == 0) { bp->b_pages[i] = bogus_page; bogus = true; } foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; } VM_OBJECT_WUNLOCK(obj); if (bogus && buf_mapped(bp)) { BUF_CHECK_MAPPED(bp); pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); } } /* * vfs_bio_set_valid: * * Set the range within the buffer to valid. The range is * relative to the beginning of the buffer, b_offset. Note that * b_offset itself may be offset from the beginning of the first * page. */ void vfs_bio_set_valid(struct buf *bp, int base, int size) { int i, n; vm_page_t m; if (!(bp->b_flags & B_VMIO)) return; /* * Fixup base to be relative to beginning of first page. * Set initial n to be the maximum number of bytes in the * first page that can be validated. */ base += (bp->b_offset & PAGE_MASK); n = PAGE_SIZE - (base & PAGE_MASK); VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); + + /* + * Busy may not be strictly necessary here because the pages are + * unlikely to be fully valid and the vnode lock will synchronize + * their access via getpages. It is grabbed for consistency with + * other page validation. + */ + vfs_busy_pages_acquire(bp); for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { m = bp->b_pages[i]; if (n > size) n = size; vm_page_set_valid_range(m, base & PAGE_MASK, n); base += n; size -= n; n = PAGE_SIZE; } + vfs_busy_pages_release(bp); VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); } /* * vfs_bio_clrbuf: * * If the specified buffer is a non-VMIO buffer, clear the entire * buffer. If the specified buffer is a VMIO buffer, clear and * validate only the previously invalid portions of the buffer. * This routine essentially fakes an I/O, so we need to clear * BIO_ERROR and B_INVAL. * * Note that while we only theoretically need to clear through b_bcount, * we go ahead and clear through b_bufsize. */ void vfs_bio_clrbuf(struct buf *bp) { int i, j, mask, sa, ea, slide; if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { clrbuf(bp); return; } bp->b_flags &= ~B_INVAL; bp->b_ioflags &= ~BIO_ERROR; VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); + vfs_busy_pages_acquire(bp); if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && (bp->b_offset & PAGE_MASK) == 0) { if (bp->b_pages[0] == bogus_page) goto unlock; mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object); if ((bp->b_pages[0]->valid & mask) == mask) goto unlock; if ((bp->b_pages[0]->valid & mask) == 0) { pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize); bp->b_pages[0]->valid |= mask; goto unlock; } } sa = bp->b_offset & PAGE_MASK; slide = 0; for (i = 0; i < bp->b_npages; i++, sa = 0) { slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize); ea = slide & PAGE_MASK; if (ea == 0) ea = PAGE_SIZE; if (bp->b_pages[i] == bogus_page) continue; j = sa / DEV_BSIZE; mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object); if ((bp->b_pages[i]->valid & mask) == mask) continue; if ((bp->b_pages[i]->valid & mask) == 0) pmap_zero_page_area(bp->b_pages[i], sa, ea - sa); else { for (; sa < ea; sa += DEV_BSIZE, j++) { if ((bp->b_pages[i]->valid & (1 << j)) == 0) { pmap_zero_page_area(bp->b_pages[i], sa, DEV_BSIZE); } } } bp->b_pages[i]->valid |= mask; } unlock: + vfs_busy_pages_release(bp); VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); bp->b_resid = 0; } void vfs_bio_bzero_buf(struct buf *bp, int base, int size) { vm_page_t m; int i, n; if (buf_mapped(bp)) { BUF_CHECK_MAPPED(bp); bzero(bp->b_data + base, size); } else { BUF_CHECK_UNMAPPED(bp); n = PAGE_SIZE - (base & PAGE_MASK); for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { m = bp->b_pages[i]; if (n > size) n = size; pmap_zero_page_area(m, base & PAGE_MASK, n); base += n; size -= n; n = PAGE_SIZE; } } } /* * Update buffer flags based on I/O request parameters, optionally releasing the * buffer. If it's VMIO or direct I/O, the buffer pages are released to the VM, * where they may be placed on a page queue (VMIO) or freed immediately (direct * I/O). Otherwise the buffer is released to the cache. */ static void b_io_dismiss(struct buf *bp, int ioflag, bool release) { KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0, ("buf %p non-VMIO noreuse", bp)); if ((ioflag & IO_DIRECT) != 0) bp->b_flags |= B_DIRECT; if ((ioflag & IO_EXT) != 0) bp->b_xflags |= BX_ALTDATA; if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) { bp->b_flags |= B_RELBUF; if ((ioflag & IO_NOREUSE) != 0) bp->b_flags |= B_NOREUSE; if (release) brelse(bp); } else if (release) bqrelse(bp); } void vfs_bio_brelse(struct buf *bp, int ioflag) { b_io_dismiss(bp, ioflag, true); } void vfs_bio_set_flags(struct buf *bp, int ioflag) { b_io_dismiss(bp, ioflag, false); } /* * vm_hold_load_pages and vm_hold_free_pages get pages into * a buffers address space. The pages are anonymous and are * not associated with a file object. */ static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) { vm_offset_t pg; vm_page_t p; int index; BUF_CHECK_MAPPED(bp); to = round_page(to); from = round_page(from); index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; for (pg = from; pg < to; pg += PAGE_SIZE, index++) { /* * note: must allocate system pages since blocking here * could interfere with paging I/O, no matter which * process we are. */ p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) | VM_ALLOC_WAITOK); pmap_qenter(pg, &p, 1); bp->b_pages[index] = p; } bp->b_npages = index; } /* Return pages associated with this buf to the vm system */ static void vm_hold_free_pages(struct buf *bp, int newbsize) { vm_offset_t from; vm_page_t p; int index, newnpages; BUF_CHECK_MAPPED(bp); from = round_page((vm_offset_t)bp->b_data + newbsize); newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; if (bp->b_npages > newnpages) pmap_qremove(from, bp->b_npages - newnpages); for (index = newnpages; index < bp->b_npages; index++) { p = bp->b_pages[index]; bp->b_pages[index] = NULL; vm_page_unwire_noq(p); vm_page_free(p); } bp->b_npages = newnpages; } /* * Map an IO request into kernel virtual address space. * * All requests are (re)mapped into kernel VA space. * Notice that we use b_bufsize for the size of the buffer * to be mapped. b_bcount might be modified by the driver. * * Note that even if the caller determines that the address space should * be valid, a race or a smaller-file mapped into a larger space may * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST * check the return value. * * This function only works with pager buffers. */ int vmapbuf(struct buf *bp, int mapbuf) { vm_prot_t prot; int pidx; if (bp->b_bufsize < 0) return (-1); prot = VM_PROT_READ; if (bp->b_iocmd == BIO_READ) prot |= VM_PROT_WRITE; /* Less backwards than it looks */ if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages, btoc(MAXPHYS))) < 0) return (-1); bp->b_npages = pidx; bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK; if (mapbuf || !unmapped_buf_allowed) { pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx); bp->b_data = bp->b_kvabase + bp->b_offset; } else bp->b_data = unmapped_buf; return(0); } /* * Free the io map PTEs associated with this IO operation. * We also invalidate the TLB entries and restore the original b_addr. * * This function only works with pager buffers. */ void vunmapbuf(struct buf *bp) { int npages; npages = bp->b_npages; if (buf_mapped(bp)) pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); vm_page_unhold_pages(bp->b_pages, npages); bp->b_data = unmapped_buf; } void bdone(struct buf *bp) { struct mtx *mtxp; mtxp = mtx_pool_find(mtxpool_sleep, bp); mtx_lock(mtxp); bp->b_flags |= B_DONE; wakeup(bp); mtx_unlock(mtxp); } void bwait(struct buf *bp, u_char pri, const char *wchan) { struct mtx *mtxp; mtxp = mtx_pool_find(mtxpool_sleep, bp); mtx_lock(mtxp); while ((bp->b_flags & B_DONE) == 0) msleep(bp, mtxp, pri, wchan, 0); mtx_unlock(mtxp); } int bufsync(struct bufobj *bo, int waitfor) { return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread)); } void bufstrategy(struct bufobj *bo, struct buf *bp) { int i __unused; struct vnode *vp; vp = bp->b_vp; KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); i = VOP_STRATEGY(vp, bp); KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); } /* * Initialize a struct bufobj before use. Memory is assumed zero filled. */ void bufobj_init(struct bufobj *bo, void *private) { static volatile int bufobj_cleanq; bo->bo_domain = atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains; rw_init(BO_LOCKPTR(bo), "bufobj interlock"); bo->bo_private = private; TAILQ_INIT(&bo->bo_clean.bv_hd); TAILQ_INIT(&bo->bo_dirty.bv_hd); } void bufobj_wrefl(struct bufobj *bo) { KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); ASSERT_BO_WLOCKED(bo); bo->bo_numoutput++; } void bufobj_wref(struct bufobj *bo) { KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); BO_LOCK(bo); bo->bo_numoutput++; BO_UNLOCK(bo); } void bufobj_wdrop(struct bufobj *bo) { KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); BO_LOCK(bo); KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { bo->bo_flag &= ~BO_WWAIT; wakeup(&bo->bo_numoutput); } BO_UNLOCK(bo); } int bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) { int error; KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); ASSERT_BO_WLOCKED(bo); error = 0; while (bo->bo_numoutput) { bo->bo_flag |= BO_WWAIT; error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo), slpflag | (PRIBIO + 1), "bo_wwait", timeo); if (error) break; } return (error); } /* * Set bio_data or bio_ma for struct bio from the struct buf. */ void bdata2bio(struct buf *bp, struct bio *bip) { if (!buf_mapped(bp)) { KASSERT(unmapped_buf_allowed, ("unmapped")); bip->bio_ma = bp->b_pages; bip->bio_ma_n = bp->b_npages; bip->bio_data = unmapped_buf; bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; bip->bio_flags |= BIO_UNMAPPED; KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) / PAGE_SIZE == bp->b_npages, ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset, (long long)bip->bio_length, bip->bio_ma_n)); } else { bip->bio_data = bp->b_data; bip->bio_ma = NULL; } } /* * The MIPS pmap code currently doesn't handle aliased pages. * The VIPT caches may not handle page aliasing themselves, leading * to data corruption. * * As such, this code makes a system extremely unhappy if said * system doesn't support unaliasing the above situation in hardware. * Some "recent" systems (eg some mips24k/mips74k cores) don't enable * this feature at build time, so it has to be handled in software. * * Once the MIPS pmap/cache code grows to support this function on * earlier chips, it should be flipped back off. */ #ifdef __mips__ static int buf_pager_relbuf = 1; #else static int buf_pager_relbuf = 0; #endif SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN, &buf_pager_relbuf, 0, "Make buffer pager release buffers after reading"); /* * The buffer pager. It uses buffer reads to validate pages. * * In contrast to the generic local pager from vm/vnode_pager.c, this * pager correctly and easily handles volumes where the underlying * device block size is greater than the machine page size. The * buffer cache transparently extends the requested page run to be * aligned at the block boundary, and does the necessary bogus page * replacements in the addends to avoid obliterating already valid * pages. * * The only non-trivial issue is that the exclusive busy state for * pages, which is assumed by the vm_pager_getpages() interface, is * incompatible with the VMIO buffer cache's desire to share-busy the * pages. This function performs a trivial downgrade of the pages' * state before reading buffers, and a less trivial upgrade from the * shared-busy to excl-busy state after the read. */ int vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count, int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno, vbg_get_blksize_t get_blksize) { vm_page_t m; vm_object_t object; struct buf *bp; struct mount *mp; daddr_t lbn, lbnp; vm_ooffset_t la, lb, poff, poffe; long bsize; int bo_bs, br_flags, error, i, pgsin, pgsin_a, pgsin_b; bool redo, lpart; object = vp->v_object; mp = vp->v_mount; error = 0; la = IDX_TO_OFF(ma[count - 1]->pindex); if (la >= object->un_pager.vnp.vnp_size) return (VM_PAGER_BAD); /* * Change the meaning of la from where the last requested page starts * to where it ends, because that's the end of the requested region * and the start of the potential read-ahead region. */ la += PAGE_SIZE; lpart = la > object->un_pager.vnp.vnp_size; bo_bs = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex))); /* * Calculate read-ahead, behind and total pages. */ pgsin = count; lb = IDX_TO_OFF(ma[0]->pindex); pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs)); pgsin += pgsin_b; if (rbehind != NULL) *rbehind = pgsin_b; pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la); if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size) pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size, PAGE_SIZE) - la); pgsin += pgsin_a; if (rahead != NULL) *rahead = pgsin_a; VM_CNT_INC(v_vnodein); VM_CNT_ADD(v_vnodepgsin, pgsin); br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS) != 0) ? GB_UNMAPPED : 0; VM_OBJECT_WLOCK(object); again: for (i = 0; i < count; i++) vm_page_busy_downgrade(ma[i]); VM_OBJECT_WUNLOCK(object); lbnp = -1; for (i = 0; i < count; i++) { m = ma[i]; /* * Pages are shared busy and the object lock is not * owned, which together allow for the pages' * invalidation. The racy test for validity avoids * useless creation of the buffer for the most typical * case when invalidation is not used in redo or for * parallel read. The shared->excl upgrade loop at * the end of the function catches the race in a * reliable way (protected by the object lock). */ - if (m->valid == VM_PAGE_BITS_ALL) + if (vm_page_all_valid(m)) continue; poff = IDX_TO_OFF(m->pindex); poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size); for (; poff < poffe; poff += bsize) { lbn = get_lblkno(vp, poff); if (lbn == lbnp) goto next_page; lbnp = lbn; bsize = get_blksize(vp, lbn); error = bread_gb(vp, lbn, bsize, curthread->td_ucred, br_flags, &bp); if (error != 0) goto end_pages; if (LIST_EMPTY(&bp->b_dep)) { /* * Invalidation clears m->valid, but * may leave B_CACHE flag if the * buffer existed at the invalidation * time. In this case, recycle the * buffer to do real read on next * bread() after redo. * * Otherwise B_RELBUF is not strictly * necessary, enable to reduce buf * cache pressure. */ if (buf_pager_relbuf || - m->valid != VM_PAGE_BITS_ALL) + !vm_page_all_valid(m)) bp->b_flags |= B_RELBUF; bp->b_flags &= ~B_NOCACHE; brelse(bp); } else { bqrelse(bp); } } KASSERT(1 /* racy, enable for debugging */ || - m->valid == VM_PAGE_BITS_ALL || i == count - 1, + vm_page_all_valid(m) || i == count - 1, ("buf %d %p invalid", i, m)); if (i == count - 1 && lpart) { VM_OBJECT_WLOCK(object); - if (m->valid != 0 && - m->valid != VM_PAGE_BITS_ALL) + if (!vm_page_none_valid(m) && + !vm_page_all_valid(m)) vm_page_zero_invalid(m, TRUE); VM_OBJECT_WUNLOCK(object); } next_page:; } end_pages: VM_OBJECT_WLOCK(object); redo = false; for (i = 0; i < count; i++) { vm_page_sunbusy(ma[i]); ma[i] = vm_page_grab(object, ma[i]->pindex, VM_ALLOC_NORMAL); /* * Since the pages were only sbusy while neither the * buffer nor the object lock was held by us, or * reallocated while vm_page_grab() slept for busy * relinguish, they could have been invalidated. * Recheck the valid bits and re-read as needed. * * Note that the last page is made fully valid in the * read loop, and partial validity for the page at * index count - 1 could mean that the page was * invalidated or removed, so we must restart for * safety as well. */ - if (ma[i]->valid != VM_PAGE_BITS_ALL) + if (!vm_page_all_valid(ma[i])) redo = true; } if (redo && error == 0) goto again; VM_OBJECT_WUNLOCK(object); return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK); } #include "opt_ddb.h" #ifdef DDB #include /* DDB command to show buffer data */ DB_SHOW_COMMAND(buffer, db_show_buffer) { /* get args */ struct buf *bp = (struct buf *)addr; #ifdef FULL_BUF_TRACKING uint32_t i, j; #endif if (!have_addr) { db_printf("usage: show buffer \n"); return; } db_printf("buf at %p\n", bp); db_printf("b_flags = 0x%b, b_xflags=0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags, PRINT_BUF_XFLAGS); db_printf("b_vflags=0x%b b_ioflags0x%b\n", (u_int)bp->b_vflags, PRINT_BUF_VFLAGS, (u_int)bp->b_ioflags, PRINT_BIO_FLAGS); db_printf( "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, " "b_vp = %p, b_dep = %p\n", bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first); db_printf("b_kvabase = %p, b_kvasize = %d\n", bp->b_kvabase, bp->b_kvasize); if (bp->b_npages) { int i; db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); for (i = 0; i < bp->b_npages; i++) { vm_page_t m; m = bp->b_pages[i]; if (m != NULL) db_printf("(%p, 0x%lx, 0x%lx)", m->object, (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); else db_printf("( ??? )"); if ((i + 1) < bp->b_npages) db_printf(","); } db_printf("\n"); } BUF_LOCKPRINTINFO(bp); #if defined(FULL_BUF_TRACKING) db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt); i = bp->b_io_tcnt % BUF_TRACKING_SIZE; for (j = 1; j <= BUF_TRACKING_SIZE; j++) { if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL) continue; db_printf(" %2u: %s\n", j, bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]); } #elif defined(BUF_TRACKING) db_printf("b_io_tracking: %s\n", bp->b_io_tracking); #endif db_printf(" "); } DB_SHOW_COMMAND(bufqueues, bufqueues) { struct bufdomain *bd; struct buf *bp; long total; int i, j, cnt; db_printf("bqempty: %d\n", bqempty.bq_len); for (i = 0; i < buf_domains; i++) { bd = &bdomain[i]; db_printf("Buf domain %d\n", i); db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers); db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers); db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers); db_printf("\n"); db_printf("\tbufspace\t%ld\n", bd->bd_bufspace); db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace); db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace); db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace); db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh); db_printf("\n"); db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers); db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers); db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers); db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh); db_printf("\n"); total = 0; TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist) total += bp->b_bufsize; db_printf("\tcleanq count\t%d (%ld)\n", bd->bd_cleanq->bq_len, total); total = 0; TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist) total += bp->b_bufsize; db_printf("\tdirtyq count\t%d (%ld)\n", bd->bd_dirtyq.bq_len, total); db_printf("\twakeup\t\t%d\n", bd->bd_wanted); db_printf("\tlim\t\t%d\n", bd->bd_lim); db_printf("\tCPU "); for (j = 0; j <= mp_maxid; j++) db_printf("%d, ", bd->bd_subq[j].bq_len); db_printf("\n"); cnt = 0; total = 0; for (j = 0; j < nbuf; j++) if (buf[j].b_domain == i && BUF_ISLOCKED(&buf[j])) { cnt++; total += buf[j].b_bufsize; } db_printf("\tLocked buffers: %d space %ld\n", cnt, total); cnt = 0; total = 0; for (j = 0; j < nbuf; j++) if (buf[j].b_domain == i) { cnt++; total += buf[j].b_bufsize; } db_printf("\tTotal buffers: %d space %ld\n", cnt, total); } } DB_SHOW_COMMAND(lockedbufs, lockedbufs) { struct buf *bp; int i; for (i = 0; i < nbuf; i++) { bp = &buf[i]; if (BUF_ISLOCKED(bp)) { db_show_buffer((uintptr_t)bp, 1, 0, NULL); db_printf("\n"); if (db_pager_quit) break; } } } DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) { struct vnode *vp; struct buf *bp; if (!have_addr) { db_printf("usage: show vnodebufs \n"); return; } vp = (struct vnode *)addr; db_printf("Clean buffers:\n"); TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { db_show_buffer((uintptr_t)bp, 1, 0, NULL); db_printf("\n"); } db_printf("Dirty buffers:\n"); TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { db_show_buffer((uintptr_t)bp, 1, 0, NULL); db_printf("\n"); } } DB_COMMAND(countfreebufs, db_coundfreebufs) { struct buf *bp; int i, used = 0, nfree = 0; if (have_addr) { db_printf("usage: countfreebufs\n"); return; } for (i = 0; i < nbuf; i++) { bp = &buf[i]; if (bp->b_qindex == QUEUE_EMPTY) nfree++; else used++; } db_printf("Counted %d free, %d used (%d tot)\n", nfree, used, nfree + used); db_printf("numfreebuffers is %d\n", numfreebuffers); } #endif /* DDB */ Index: head/sys/kern/vfs_cluster.c =================================================================== --- head/sys/kern/vfs_cluster.c (revision 353538) +++ head/sys/kern/vfs_cluster.c (revision 353539) @@ -1,1097 +1,1099 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1993 * The Regents of the University of California. All rights reserved. * Modifications/enhancements: * Copyright (c) 1995 John S. Dyson. All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)vfs_cluster.c 8.7 (Berkeley) 2/13/94 */ #include __FBSDID("$FreeBSD$"); #include "opt_debug_cluster.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #if defined(CLUSTERDEBUG) static int rcluster= 0; SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0, "Debug VFS clustering code"); #endif static MALLOC_DEFINE(M_SEGMENT, "cl_savebuf", "cluster_save buffer"); static uma_zone_t cluster_pbuf_zone; static void cluster_init(void *); static struct cluster_save *cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int gbflags); static struct buf *cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn, daddr_t blkno, long size, int run, int gbflags, struct buf *fbp); static void cluster_callback(struct buf *); static int write_behind = 1; SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0, "Cluster write-behind; 0: disable, 1: enable, 2: backed off"); static int read_max = 64; SYSCTL_INT(_vfs, OID_AUTO, read_max, CTLFLAG_RW, &read_max, 0, "Cluster read-ahead max block count"); static int read_min = 1; SYSCTL_INT(_vfs, OID_AUTO, read_min, CTLFLAG_RW, &read_min, 0, "Cluster read min block count"); SYSINIT(cluster, SI_SUB_CPU, SI_ORDER_ANY, cluster_init, NULL); static void cluster_init(void *dummy) { cluster_pbuf_zone = pbuf_zsecond_create("clpbuf", nswbuf / 2); } /* * Read data to a buf, including read-ahead if we find this to be beneficial. * cluster_read replaces bread. */ int cluster_read(struct vnode *vp, u_quad_t filesize, daddr_t lblkno, long size, struct ucred *cred, long totread, int seqcount, int gbflags, struct buf **bpp) { struct buf *bp, *rbp, *reqbp; struct bufobj *bo; struct thread *td; daddr_t blkno, origblkno; int maxra, racluster; int error, ncontig; int i; error = 0; td = curthread; bo = &vp->v_bufobj; if (!unmapped_buf_allowed) gbflags &= ~GB_UNMAPPED; /* * Try to limit the amount of read-ahead by a few * ad-hoc parameters. This needs work!!! */ racluster = vp->v_mount->mnt_iosize_max / size; maxra = seqcount; maxra = min(read_max, maxra); maxra = min(nbuf/8, maxra); if (((u_quad_t)(lblkno + maxra + 1) * size) > filesize) maxra = (filesize / size) - lblkno; /* * get the requested block */ error = getblkx(vp, lblkno, size, 0, 0, gbflags, &bp); if (error != 0) { *bpp = NULL; return (error); } gbflags &= ~GB_NOSPARSE; origblkno = lblkno; *bpp = reqbp = bp; /* * if it is in the cache, then check to see if the reads have been * sequential. If they have, then try some read-ahead, otherwise * back-off on prospective read-aheads. */ if (bp->b_flags & B_CACHE) { if (!seqcount) { return 0; } else if ((bp->b_flags & B_RAM) == 0) { return 0; } else { bp->b_flags &= ~B_RAM; BO_RLOCK(bo); for (i = 1; i < maxra; i++) { /* * Stop if the buffer does not exist or it * is invalid (about to go away?) */ rbp = gbincore(&vp->v_bufobj, lblkno+i); if (rbp == NULL || (rbp->b_flags & B_INVAL)) break; /* * Set another read-ahead mark so we know * to check again. (If we can lock the * buffer without waiting) */ if ((((i % racluster) == (racluster - 1)) || (i == (maxra - 1))) && (0 == BUF_LOCK(rbp, LK_EXCLUSIVE | LK_NOWAIT, NULL))) { rbp->b_flags |= B_RAM; BUF_UNLOCK(rbp); } } BO_RUNLOCK(bo); if (i >= maxra) { return 0; } lblkno += i; } reqbp = bp = NULL; /* * If it isn't in the cache, then get a chunk from * disk if sequential, otherwise just get the block. */ } else { off_t firstread = bp->b_offset; int nblks; long minread; KASSERT(bp->b_offset != NOOFFSET, ("cluster_read: no buffer offset")); ncontig = 0; /* * Adjust totread if needed */ minread = read_min * size; if (minread > totread) totread = minread; /* * Compute the total number of blocks that we should read * synchronously. */ if (firstread + totread > filesize) totread = filesize - firstread; nblks = howmany(totread, size); if (nblks > racluster) nblks = racluster; /* * Now compute the number of contiguous blocks. */ if (nblks > 1) { error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL); /* * If this failed to map just do the original block. */ if (error || blkno == -1) ncontig = 0; } /* * If we have contiguous data available do a cluster * otherwise just read the requested block. */ if (ncontig) { /* Account for our first block. */ ncontig = min(ncontig + 1, nblks); if (ncontig < nblks) nblks = ncontig; bp = cluster_rbuild(vp, filesize, lblkno, blkno, size, nblks, gbflags, bp); lblkno += (bp->b_bufsize / size); } else { bp->b_flags |= B_RAM; bp->b_iocmd = BIO_READ; lblkno += 1; } } /* * handle the synchronous read so that it is available ASAP. */ if (bp) { if ((bp->b_flags & B_CLUSTER) == 0) { vfs_busy_pages(bp, 0); } bp->b_flags &= ~B_INVAL; bp->b_ioflags &= ~BIO_ERROR; if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL) BUF_KERNPROC(bp); bp->b_iooffset = dbtob(bp->b_blkno); bstrategy(bp); #ifdef RACCT if (racct_enable) { PROC_LOCK(td->td_proc); racct_add_buf(td->td_proc, bp, 0); PROC_UNLOCK(td->td_proc); } #endif /* RACCT */ td->td_ru.ru_inblock++; } /* * If we have been doing sequential I/O, then do some read-ahead. */ while (lblkno < (origblkno + maxra)) { error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL); if (error) break; if (blkno == -1) break; /* * We could throttle ncontig here by maxra but we might as * well read the data if it is contiguous. We're throttled * by racluster anyway. */ if (ncontig) { ncontig = min(ncontig + 1, racluster); rbp = cluster_rbuild(vp, filesize, lblkno, blkno, size, ncontig, gbflags, NULL); lblkno += (rbp->b_bufsize / size); if (rbp->b_flags & B_DELWRI) { bqrelse(rbp); continue; } } else { rbp = getblk(vp, lblkno, size, 0, 0, gbflags); lblkno += 1; if (rbp->b_flags & B_DELWRI) { bqrelse(rbp); continue; } rbp->b_flags |= B_ASYNC | B_RAM; rbp->b_iocmd = BIO_READ; rbp->b_blkno = blkno; } if (rbp->b_flags & B_CACHE) { rbp->b_flags &= ~B_ASYNC; bqrelse(rbp); continue; } if ((rbp->b_flags & B_CLUSTER) == 0) { vfs_busy_pages(rbp, 0); } rbp->b_flags &= ~B_INVAL; rbp->b_ioflags &= ~BIO_ERROR; if ((rbp->b_flags & B_ASYNC) || rbp->b_iodone != NULL) BUF_KERNPROC(rbp); rbp->b_iooffset = dbtob(rbp->b_blkno); bstrategy(rbp); #ifdef RACCT if (racct_enable) { PROC_LOCK(td->td_proc); racct_add_buf(td->td_proc, rbp, 0); PROC_UNLOCK(td->td_proc); } #endif /* RACCT */ td->td_ru.ru_inblock++; } if (reqbp) { /* * Like bread, always brelse() the buffer when * returning an error. */ error = bufwait(reqbp); if (error != 0) { brelse(reqbp); *bpp = NULL; } } return (error); } /* * If blocks are contiguous on disk, use this to provide clustered * read ahead. We will read as many blocks as possible sequentially * and then parcel them up into logical blocks in the buffer hash table. */ static struct buf * cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn, daddr_t blkno, long size, int run, int gbflags, struct buf *fbp) { struct buf *bp, *tbp; daddr_t bn; off_t off; long tinc, tsize; int i, inc, j, k, toff; KASSERT(size == vp->v_mount->mnt_stat.f_iosize, ("cluster_rbuild: size %ld != f_iosize %jd\n", size, (intmax_t)vp->v_mount->mnt_stat.f_iosize)); /* * avoid a division */ while ((u_quad_t) size * (lbn + run) > filesize) { --run; } if (fbp) { tbp = fbp; tbp->b_iocmd = BIO_READ; } else { tbp = getblk(vp, lbn, size, 0, 0, gbflags); if (tbp->b_flags & B_CACHE) return tbp; tbp->b_flags |= B_ASYNC | B_RAM; tbp->b_iocmd = BIO_READ; } tbp->b_blkno = blkno; if( (tbp->b_flags & B_MALLOC) || ((tbp->b_flags & B_VMIO) == 0) || (run <= 1) ) return tbp; bp = uma_zalloc(cluster_pbuf_zone, M_NOWAIT); if (bp == NULL) return tbp; /* * We are synthesizing a buffer out of vm_page_t's, but * if the block size is not page aligned then the starting * address may not be either. Inherit the b_data offset * from the original buffer. */ bp->b_flags = B_ASYNC | B_CLUSTER | B_VMIO; if ((gbflags & GB_UNMAPPED) != 0) { bp->b_data = unmapped_buf; } else { bp->b_data = (char *)((vm_offset_t)bp->b_data | ((vm_offset_t)tbp->b_data & PAGE_MASK)); } bp->b_iocmd = BIO_READ; bp->b_iodone = cluster_callback; bp->b_blkno = blkno; bp->b_lblkno = lbn; bp->b_offset = tbp->b_offset; KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset")); pbgetvp(vp, bp); TAILQ_INIT(&bp->b_cluster.cluster_head); bp->b_bcount = 0; bp->b_bufsize = 0; bp->b_npages = 0; inc = btodb(size); for (bn = blkno, i = 0; i < run; ++i, bn += inc) { if (i == 0) { VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object); vm_object_pip_add(tbp->b_bufobj->bo_object, tbp->b_npages); vfs_busy_pages_acquire(tbp); VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object); } else { if ((bp->b_npages * PAGE_SIZE) + round_page(size) > vp->v_mount->mnt_iosize_max) { break; } tbp = getblk(vp, lbn + i, size, 0, 0, GB_LOCK_NOWAIT | (gbflags & GB_UNMAPPED)); /* Don't wait around for locked bufs. */ if (tbp == NULL) break; /* * Stop scanning if the buffer is fully valid * (marked B_CACHE), or locked (may be doing a * background write), or if the buffer is not * VMIO backed. The clustering code can only deal * with VMIO-backed buffers. The bo lock is not * required for the BKGRDINPROG check since it * can not be set without the buf lock. */ if ((tbp->b_vflags & BV_BKGRDINPROG) || (tbp->b_flags & B_CACHE) || (tbp->b_flags & B_VMIO) == 0) { bqrelse(tbp); break; } /* * The buffer must be completely invalid in order to * take part in the cluster. If it is partially valid * then we stop. */ off = tbp->b_offset; tsize = size; VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object); for (j = 0; tsize > 0; j++) { toff = off & PAGE_MASK; tinc = tsize; if (toff + tinc > PAGE_SIZE) tinc = PAGE_SIZE - toff; VM_OBJECT_ASSERT_WLOCKED(tbp->b_pages[j]->object); - if ((tbp->b_pages[j]->valid & - vm_page_bits(toff, tinc)) != 0) - break; if (vm_page_trysbusy(tbp->b_pages[j]) == 0) break; + if ((tbp->b_pages[j]->valid & + vm_page_bits(toff, tinc)) != 0) { + vm_page_sunbusy(tbp->b_pages[j]); + break; + } vm_object_pip_add(tbp->b_bufobj->bo_object, 1); off += tinc; tsize -= tinc; } if (tsize > 0) { clean_sbusy: vm_object_pip_wakeupn(tbp->b_bufobj->bo_object, j); for (k = 0; k < j; k++) vm_page_sunbusy(tbp->b_pages[k]); VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object); bqrelse(tbp); break; } VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object); /* * Set a read-ahead mark as appropriate */ if ((fbp && (i == 1)) || (i == (run - 1))) tbp->b_flags |= B_RAM; /* * Set the buffer up for an async read (XXX should * we do this only if we do not wind up brelse()ing?). * Set the block number if it isn't set, otherwise * if it is make sure it matches the block number we * expect. */ tbp->b_flags |= B_ASYNC; tbp->b_iocmd = BIO_READ; if (tbp->b_blkno == tbp->b_lblkno) { tbp->b_blkno = bn; } else if (tbp->b_blkno != bn) { VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object); goto clean_sbusy; } } /* * XXX fbp from caller may not be B_ASYNC, but we are going * to biodone() it in cluster_callback() anyway */ BUF_KERNPROC(tbp); TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head, tbp, b_cluster.cluster_entry); VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object); for (j = 0; j < tbp->b_npages; j += 1) { vm_page_t m; m = tbp->b_pages[j]; if ((bp->b_npages == 0) || (bp->b_pages[bp->b_npages-1] != m)) { bp->b_pages[bp->b_npages] = m; bp->b_npages++; } - if (m->valid == VM_PAGE_BITS_ALL) + if (vm_page_all_valid(m)) tbp->b_pages[j] = bogus_page; } VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object); /* * Don't inherit tbp->b_bufsize as it may be larger due to * a non-page-aligned size. Instead just aggregate using * 'size'. */ if (tbp->b_bcount != size) printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size); if (tbp->b_bufsize != size) printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size); bp->b_bcount += size; bp->b_bufsize += size; } /* * Fully valid pages in the cluster are already good and do not need * to be re-read from disk. Replace the page with bogus_page */ VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); for (j = 0; j < bp->b_npages; j++) { VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[j]->object); - if (bp->b_pages[j]->valid == VM_PAGE_BITS_ALL) + if (vm_page_all_valid(bp->b_pages[j])) bp->b_pages[j] = bogus_page; } VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); if (bp->b_bufsize > bp->b_kvasize) panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n", bp->b_bufsize, bp->b_kvasize); if (buf_mapped(bp)) { pmap_qenter(trunc_page((vm_offset_t) bp->b_data), (vm_page_t *)bp->b_pages, bp->b_npages); } return (bp); } /* * Cleanup after a clustered read or write. * This is complicated by the fact that any of the buffers might have * extra memory (if there were no empty buffer headers at allocbuf time) * that we will need to shift around. */ static void cluster_callback(struct buf *bp) { struct buf *nbp, *tbp; int error = 0; /* * Must propagate errors to all the components. */ if (bp->b_ioflags & BIO_ERROR) error = bp->b_error; if (buf_mapped(bp)) { pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); } /* * Move memory from the large cluster buffer into the component * buffers and mark IO as done on these. */ for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head); tbp; tbp = nbp) { nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry); if (error) { tbp->b_ioflags |= BIO_ERROR; tbp->b_error = error; } else { tbp->b_dirtyoff = tbp->b_dirtyend = 0; tbp->b_flags &= ~B_INVAL; tbp->b_ioflags &= ~BIO_ERROR; /* * XXX the bdwrite()/bqrelse() issued during * cluster building clears B_RELBUF (see bqrelse() * comment). If direct I/O was specified, we have * to restore it here to allow the buffer and VM * to be freed. */ if (tbp->b_flags & B_DIRECT) tbp->b_flags |= B_RELBUF; } bufdone(tbp); } pbrelvp(bp); uma_zfree(cluster_pbuf_zone, bp); } /* * cluster_wbuild_wb: * * Implement modified write build for cluster. * * write_behind = 0 write behind disabled * write_behind = 1 write behind normal (default) * write_behind = 2 write behind backed-off */ static __inline int cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len, int gbflags) { int r = 0; switch (write_behind) { case 2: if (start_lbn < len) break; start_lbn -= len; /* FALLTHROUGH */ case 1: r = cluster_wbuild(vp, size, start_lbn, len, gbflags); /* FALLTHROUGH */ default: /* FALLTHROUGH */ break; } return(r); } /* * Do clustered write for FFS. * * Three cases: * 1. Write is not sequential (write asynchronously) * Write is sequential: * 2. beginning of cluster - begin cluster * 3. middle of a cluster - add to cluster * 4. end of a cluster - asynchronously write cluster */ void cluster_write(struct vnode *vp, struct buf *bp, u_quad_t filesize, int seqcount, int gbflags) { daddr_t lbn; int maxclen, cursize; int lblocksize; int async; if (!unmapped_buf_allowed) gbflags &= ~GB_UNMAPPED; if (vp->v_type == VREG) { async = DOINGASYNC(vp); lblocksize = vp->v_mount->mnt_stat.f_iosize; } else { async = 0; lblocksize = bp->b_bufsize; } lbn = bp->b_lblkno; KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset")); /* Initialize vnode to beginning of file. */ if (lbn == 0) vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0; if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 || (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) { maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1; if (vp->v_clen != 0) { /* * Next block is not sequential. * * If we are not writing at end of file, the process * seeked to another point in the file since its last * write, or we have reached our maximum cluster size, * then push the previous cluster. Otherwise try * reallocating to make it sequential. * * Change to algorithm: only push previous cluster if * it was sequential from the point of view of the * seqcount heuristic, otherwise leave the buffer * intact so we can potentially optimize the I/O * later on in the buf_daemon or update daemon * flush. */ cursize = vp->v_lastw - vp->v_cstart + 1; if (((u_quad_t) bp->b_offset + lblocksize) != filesize || lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) { if (!async && seqcount > 0) { cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, cursize, gbflags); } } else { struct buf **bpp, **endbp; struct cluster_save *buflist; buflist = cluster_collectbufs(vp, bp, gbflags); if (buflist == NULL) { /* * Cluster build failed so just write * it now. */ bawrite(bp); return; } endbp = &buflist->bs_children [buflist->bs_nchildren - 1]; if (VOP_REALLOCBLKS(vp, buflist)) { /* * Failed, push the previous cluster * if *really* writing sequentially * in the logical file (seqcount > 1), * otherwise delay it in the hopes that * the low level disk driver can * optimize the write ordering. */ for (bpp = buflist->bs_children; bpp < endbp; bpp++) brelse(*bpp); free(buflist, M_SEGMENT); if (seqcount > 1) { cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, cursize, gbflags); } } else { /* * Succeeded, keep building cluster. */ for (bpp = buflist->bs_children; bpp <= endbp; bpp++) bdwrite(*bpp); free(buflist, M_SEGMENT); vp->v_lastw = lbn; vp->v_lasta = bp->b_blkno; return; } } } /* * Consider beginning a cluster. If at end of file, make * cluster as large as possible, otherwise find size of * existing cluster. */ if ((vp->v_type == VREG) && ((u_quad_t) bp->b_offset + lblocksize) != filesize && (bp->b_blkno == bp->b_lblkno) && (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) || bp->b_blkno == -1)) { bawrite(bp); vp->v_clen = 0; vp->v_lasta = bp->b_blkno; vp->v_cstart = lbn + 1; vp->v_lastw = lbn; return; } vp->v_clen = maxclen; if (!async && maxclen == 0) { /* I/O not contiguous */ vp->v_cstart = lbn + 1; bawrite(bp); } else { /* Wait for rest of cluster */ vp->v_cstart = lbn; bdwrite(bp); } } else if (lbn == vp->v_cstart + vp->v_clen) { /* * At end of cluster, write it out if seqcount tells us we * are operating sequentially, otherwise let the buf or * update daemon handle it. */ bdwrite(bp); if (seqcount > 1) { cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, vp->v_clen + 1, gbflags); } vp->v_clen = 0; vp->v_cstart = lbn + 1; } else if (vm_page_count_severe()) { /* * We are low on memory, get it going NOW */ bawrite(bp); } else { /* * In the middle of a cluster, so just delay the I/O for now. */ bdwrite(bp); } vp->v_lastw = lbn; vp->v_lasta = bp->b_blkno; } /* * This is an awful lot like cluster_rbuild...wish they could be combined. * The last lbn argument is the current block on which I/O is being * performed. Check to see that it doesn't fall in the middle of * the current block (if last_bp == NULL). */ int cluster_wbuild(struct vnode *vp, long size, daddr_t start_lbn, int len, int gbflags) { struct buf *bp, *tbp; struct bufobj *bo; int i, j; int totalwritten = 0; int dbsize = btodb(size); if (!unmapped_buf_allowed) gbflags &= ~GB_UNMAPPED; bo = &vp->v_bufobj; while (len > 0) { /* * If the buffer is not delayed-write (i.e. dirty), or it * is delayed-write but either locked or inval, it cannot * partake in the clustered write. */ BO_LOCK(bo); if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL || (tbp->b_vflags & BV_BKGRDINPROG)) { BO_UNLOCK(bo); ++start_lbn; --len; continue; } if (BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, BO_LOCKPTR(bo))) { ++start_lbn; --len; continue; } if ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) { BUF_UNLOCK(tbp); ++start_lbn; --len; continue; } bremfree(tbp); tbp->b_flags &= ~B_DONE; /* * Extra memory in the buffer, punt on this buffer. * XXX we could handle this in most cases, but we would * have to push the extra memory down to after our max * possible cluster size and then potentially pull it back * up if the cluster was terminated prematurely--too much * hassle. */ if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) != (B_CLUSTEROK | B_VMIO)) || (tbp->b_bcount != tbp->b_bufsize) || (tbp->b_bcount != size) || (len == 1) || ((bp = uma_zalloc(cluster_pbuf_zone, (vp->v_vflag & VV_MD) != 0 ? M_NOWAIT : M_WAITOK)) == NULL)) { totalwritten += tbp->b_bufsize; bawrite(tbp); ++start_lbn; --len; continue; } /* * We got a pbuf to make the cluster in. * so initialise it. */ TAILQ_INIT(&bp->b_cluster.cluster_head); bp->b_bcount = 0; bp->b_bufsize = 0; bp->b_npages = 0; if (tbp->b_wcred != NOCRED) bp->b_wcred = crhold(tbp->b_wcred); bp->b_blkno = tbp->b_blkno; bp->b_lblkno = tbp->b_lblkno; bp->b_offset = tbp->b_offset; /* * We are synthesizing a buffer out of vm_page_t's, but * if the block size is not page aligned then the starting * address may not be either. Inherit the b_data offset * from the original buffer. */ if ((gbflags & GB_UNMAPPED) == 0 || (tbp->b_flags & B_VMIO) == 0) { bp->b_data = (char *)((vm_offset_t)bp->b_data | ((vm_offset_t)tbp->b_data & PAGE_MASK)); } else { bp->b_data = unmapped_buf; } bp->b_flags |= B_CLUSTER | (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT)); bp->b_iodone = cluster_callback; pbgetvp(vp, bp); /* * From this location in the file, scan forward to see * if there are buffers with adjacent data that need to * be written as well. */ for (i = 0; i < len; ++i, ++start_lbn) { if (i != 0) { /* If not the first buffer */ /* * If the adjacent data is not even in core it * can't need to be written. */ BO_LOCK(bo); if ((tbp = gbincore(bo, start_lbn)) == NULL || (tbp->b_vflags & BV_BKGRDINPROG)) { BO_UNLOCK(bo); break; } /* * If it IS in core, but has different * characteristics, or is locked (which * means it could be undergoing a background * I/O or be in a weird state), then don't * cluster with it. */ if (BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, BO_LOCKPTR(bo))) break; if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK | B_INVAL | B_DELWRI | B_NEEDCOMMIT)) != (B_DELWRI | B_CLUSTEROK | (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) || tbp->b_wcred != bp->b_wcred) { BUF_UNLOCK(tbp); break; } /* * Check that the combined cluster * would make sense with regard to pages * and would not be too large */ if ((tbp->b_bcount != size) || ((bp->b_blkno + (dbsize * i)) != tbp->b_blkno) || ((tbp->b_npages + bp->b_npages) > (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) { BUF_UNLOCK(tbp); break; } /* * Ok, it's passed all the tests, * so remove it from the free list * and mark it busy. We will use it. */ bremfree(tbp); tbp->b_flags &= ~B_DONE; } /* end of code for non-first buffers only */ /* * If the IO is via the VM then we do some * special VM hackery (yuck). Since the buffer's * block size may not be page-aligned it is possible * for a page to be shared between two buffers. We * have to get rid of the duplication when building * the cluster. */ if (tbp->b_flags & B_VMIO) { vm_page_t m; VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object); if (i == 0) { vfs_busy_pages_acquire(tbp); } else { /* if not first buffer */ for (j = 0; j < tbp->b_npages; j += 1) { m = tbp->b_pages[j]; if (vm_page_trysbusy(m) == 0) { for (j--; j >= 0; j--) vm_page_sunbusy( tbp->b_pages[j]); VM_OBJECT_WUNLOCK( tbp->b_object); bqrelse(tbp); goto finishcluster; } } } vm_object_pip_add(tbp->b_bufobj->bo_object, tbp->b_npages); for (j = 0; j < tbp->b_npages; j += 1) { m = tbp->b_pages[j]; if ((bp->b_npages == 0) || (bp->b_pages[bp->b_npages - 1] != m)) { bp->b_pages[bp->b_npages] = m; bp->b_npages++; } } VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object); } bp->b_bcount += size; bp->b_bufsize += size; /* * If any of the clustered buffers have their * B_BARRIER flag set, transfer that request to * the cluster. */ bp->b_flags |= (tbp->b_flags & B_BARRIER); tbp->b_flags &= ~(B_DONE | B_BARRIER); tbp->b_flags |= B_ASYNC; tbp->b_ioflags &= ~BIO_ERROR; tbp->b_iocmd = BIO_WRITE; bundirty(tbp); reassignbuf(tbp); /* put on clean list */ bufobj_wref(tbp->b_bufobj); BUF_KERNPROC(tbp); buf_track(tbp, __func__); TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head, tbp, b_cluster.cluster_entry); } finishcluster: if (buf_mapped(bp)) { pmap_qenter(trunc_page((vm_offset_t) bp->b_data), (vm_page_t *)bp->b_pages, bp->b_npages); } if (bp->b_bufsize > bp->b_kvasize) panic( "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n", bp->b_bufsize, bp->b_kvasize); totalwritten += bp->b_bufsize; bp->b_dirtyoff = 0; bp->b_dirtyend = bp->b_bufsize; bawrite(bp); len -= i; } return totalwritten; } /* * Collect together all the buffers in a cluster. * Plus add one additional buffer. */ static struct cluster_save * cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int gbflags) { struct cluster_save *buflist; struct buf *bp; daddr_t lbn; int i, j, len, error; len = vp->v_lastw - vp->v_cstart + 1; buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist), M_SEGMENT, M_WAITOK); buflist->bs_nchildren = 0; buflist->bs_children = (struct buf **) (buflist + 1); for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) { error = bread_gb(vp, lbn, last_bp->b_bcount, NOCRED, gbflags, &bp); if (error != 0) { /* * If read fails, release collected buffers * and return failure. */ for (j = 0; j < i; j++) brelse(buflist->bs_children[j]); free(buflist, M_SEGMENT); return (NULL); } buflist->bs_children[i] = bp; if (bp->b_blkno == bp->b_lblkno) VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); } buflist->bs_children[i] = bp = last_bp; if (bp->b_blkno == bp->b_lblkno) VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); buflist->bs_nchildren = i + 1; return (buflist); } Index: head/sys/vm/device_pager.c =================================================================== --- head/sys/vm/device_pager.c (revision 353538) +++ head/sys/vm/device_pager.c (revision 353539) @@ -1,468 +1,468 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1990 University of Utah. * Copyright (c) 1991, 1993 * The Regents of the University of California. All rights reserved. * * This code is derived from software contributed to Berkeley by * the Systems Programming Group of the University of Utah Computer * Science Department. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)device_pager.c 8.1 (Berkeley) 6/11/93 */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include static void dev_pager_init(void); static vm_object_t dev_pager_alloc(void *, vm_ooffset_t, vm_prot_t, vm_ooffset_t, struct ucred *); static void dev_pager_dealloc(vm_object_t); static int dev_pager_getpages(vm_object_t, vm_page_t *, int, int *, int *); static void dev_pager_putpages(vm_object_t, vm_page_t *, int, int, int *); static boolean_t dev_pager_haspage(vm_object_t, vm_pindex_t, int *, int *); static void dev_pager_free_page(vm_object_t object, vm_page_t m); static int dev_pager_populate(vm_object_t object, vm_pindex_t pidx, int fault_type, vm_prot_t, vm_pindex_t *first, vm_pindex_t *last); /* list of device pager objects */ static struct pagerlst dev_pager_object_list; /* protect list manipulation */ static struct mtx dev_pager_mtx; struct pagerops devicepagerops = { .pgo_init = dev_pager_init, .pgo_alloc = dev_pager_alloc, .pgo_dealloc = dev_pager_dealloc, .pgo_getpages = dev_pager_getpages, .pgo_putpages = dev_pager_putpages, .pgo_haspage = dev_pager_haspage, }; struct pagerops mgtdevicepagerops = { .pgo_alloc = dev_pager_alloc, .pgo_dealloc = dev_pager_dealloc, .pgo_getpages = dev_pager_getpages, .pgo_putpages = dev_pager_putpages, .pgo_haspage = dev_pager_haspage, .pgo_populate = dev_pager_populate, }; static int old_dev_pager_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t foff, struct ucred *cred, u_short *color); static void old_dev_pager_dtor(void *handle); static int old_dev_pager_fault(vm_object_t object, vm_ooffset_t offset, int prot, vm_page_t *mres); static struct cdev_pager_ops old_dev_pager_ops = { .cdev_pg_ctor = old_dev_pager_ctor, .cdev_pg_dtor = old_dev_pager_dtor, .cdev_pg_fault = old_dev_pager_fault }; static void dev_pager_init(void) { TAILQ_INIT(&dev_pager_object_list); mtx_init(&dev_pager_mtx, "dev_pager list", NULL, MTX_DEF); } vm_object_t cdev_pager_lookup(void *handle) { vm_object_t object; mtx_lock(&dev_pager_mtx); object = vm_pager_object_lookup(&dev_pager_object_list, handle); mtx_unlock(&dev_pager_mtx); return (object); } vm_object_t cdev_pager_allocate(void *handle, enum obj_type tp, struct cdev_pager_ops *ops, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t foff, struct ucred *cred) { vm_object_t object, object1; vm_pindex_t pindex; u_short color; if (tp != OBJT_DEVICE && tp != OBJT_MGTDEVICE) return (NULL); KASSERT(tp == OBJT_MGTDEVICE || ops->cdev_pg_populate == NULL, ("populate on unmanaged device pager")); /* * Offset should be page aligned. */ if (foff & PAGE_MASK) return (NULL); /* * Treat the mmap(2) file offset as an unsigned value for a * device mapping. This, in effect, allows a user to pass all * possible off_t values as the mapping cookie to the driver. At * this point, we know that both foff and size are a multiple * of the page size. Do a check to avoid wrap. */ size = round_page(size); pindex = OFF_TO_IDX(foff) + OFF_TO_IDX(size); if (pindex > OBJ_MAX_SIZE || pindex < OFF_TO_IDX(foff) || pindex < OFF_TO_IDX(size)) return (NULL); if (ops->cdev_pg_ctor(handle, size, prot, foff, cred, &color) != 0) return (NULL); mtx_lock(&dev_pager_mtx); /* * Look up pager, creating as necessary. */ object1 = NULL; object = vm_pager_object_lookup(&dev_pager_object_list, handle); if (object == NULL) { /* * Allocate object and associate it with the pager. Initialize * the object's pg_color based upon the physical address of the * device's memory. */ mtx_unlock(&dev_pager_mtx); object1 = vm_object_allocate(tp, pindex); object1->flags |= OBJ_COLORED; object1->pg_color = color; object1->handle = handle; object1->un_pager.devp.ops = ops; object1->un_pager.devp.dev = handle; TAILQ_INIT(&object1->un_pager.devp.devp_pglist); mtx_lock(&dev_pager_mtx); object = vm_pager_object_lookup(&dev_pager_object_list, handle); if (object != NULL) { /* * We raced with other thread while allocating object. */ if (pindex > object->size) object->size = pindex; KASSERT(object->type == tp, ("Inconsistent device pager type %p %d", object, tp)); KASSERT(object->un_pager.devp.ops == ops, ("Inconsistent devops %p %p", object, ops)); } else { object = object1; object1 = NULL; object->handle = handle; TAILQ_INSERT_TAIL(&dev_pager_object_list, object, pager_object_list); if (ops->cdev_pg_populate != NULL) vm_object_set_flag(object, OBJ_POPULATE); } } else { if (pindex > object->size) object->size = pindex; KASSERT(object->type == tp, ("Inconsistent device pager type %p %d", object, tp)); } mtx_unlock(&dev_pager_mtx); if (object1 != NULL) { object1->handle = object1; mtx_lock(&dev_pager_mtx); TAILQ_INSERT_TAIL(&dev_pager_object_list, object1, pager_object_list); mtx_unlock(&dev_pager_mtx); vm_object_deallocate(object1); } return (object); } static vm_object_t dev_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t foff, struct ucred *cred) { return (cdev_pager_allocate(handle, OBJT_DEVICE, &old_dev_pager_ops, size, prot, foff, cred)); } void cdev_pager_free_page(vm_object_t object, vm_page_t m) { VM_OBJECT_ASSERT_WLOCKED(object); if (object->type == OBJT_MGTDEVICE) { KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("unmanaged %p", m)); pmap_remove_all(m); (void)vm_page_remove(m); vm_page_xunbusy(m); } else if (object->type == OBJT_DEVICE) dev_pager_free_page(object, m); } static void dev_pager_free_page(vm_object_t object, vm_page_t m) { VM_OBJECT_ASSERT_WLOCKED(object); KASSERT((object->type == OBJT_DEVICE && (m->oflags & VPO_UNMANAGED) != 0), ("Managed device or page obj %p m %p", object, m)); TAILQ_REMOVE(&object->un_pager.devp.devp_pglist, m, plinks.q); vm_page_putfake(m); } static void dev_pager_dealloc(vm_object_t object) { vm_page_t m; VM_OBJECT_WUNLOCK(object); object->un_pager.devp.ops->cdev_pg_dtor(object->un_pager.devp.dev); mtx_lock(&dev_pager_mtx); TAILQ_REMOVE(&dev_pager_object_list, object, pager_object_list); mtx_unlock(&dev_pager_mtx); VM_OBJECT_WLOCK(object); if (object->type == OBJT_DEVICE) { /* * Free up our fake pages. */ while ((m = TAILQ_FIRST(&object->un_pager.devp.devp_pglist)) != NULL) dev_pager_free_page(object, m); } object->handle = NULL; object->type = OBJT_DEAD; } static int dev_pager_getpages(vm_object_t object, vm_page_t *ma, int count, int *rbehind, int *rahead) { int error; /* Since our haspage reports zero after/before, the count is 1. */ KASSERT(count == 1, ("%s: count %d", __func__, count)); VM_OBJECT_ASSERT_WLOCKED(object); if (object->un_pager.devp.ops->cdev_pg_fault == NULL) return (VM_PAGER_FAIL); error = object->un_pager.devp.ops->cdev_pg_fault(object, IDX_TO_OFF(ma[0]->pindex), PROT_READ, &ma[0]); VM_OBJECT_ASSERT_WLOCKED(object); if (error == VM_PAGER_OK) { KASSERT((object->type == OBJT_DEVICE && (ma[0]->oflags & VPO_UNMANAGED) != 0) || (object->type == OBJT_MGTDEVICE && (ma[0]->oflags & VPO_UNMANAGED) == 0), ("Wrong page type %p %p", ma[0], object)); if (object->type == OBJT_DEVICE) { TAILQ_INSERT_TAIL(&object->un_pager.devp.devp_pglist, ma[0], plinks.q); } if (rbehind) *rbehind = 0; if (rahead) *rahead = 0; } return (error); } static int dev_pager_populate(vm_object_t object, vm_pindex_t pidx, int fault_type, vm_prot_t max_prot, vm_pindex_t *first, vm_pindex_t *last) { VM_OBJECT_ASSERT_WLOCKED(object); if (object->un_pager.devp.ops->cdev_pg_populate == NULL) return (VM_PAGER_FAIL); return (object->un_pager.devp.ops->cdev_pg_populate(object, pidx, fault_type, max_prot, first, last)); } static int old_dev_pager_fault(vm_object_t object, vm_ooffset_t offset, int prot, vm_page_t *mres) { vm_paddr_t paddr; vm_page_t m_paddr, page; struct cdev *dev; struct cdevsw *csw; struct file *fpop; struct thread *td; vm_memattr_t memattr, memattr1; int ref, ret; memattr = object->memattr; VM_OBJECT_WUNLOCK(object); dev = object->handle; csw = dev_refthread(dev, &ref); if (csw == NULL) { VM_OBJECT_WLOCK(object); return (VM_PAGER_FAIL); } td = curthread; fpop = td->td_fpop; td->td_fpop = NULL; ret = csw->d_mmap(dev, offset, &paddr, prot, &memattr); td->td_fpop = fpop; dev_relthread(dev, ref); if (ret != 0) { printf( "WARNING: dev_pager_getpage: map function returns error %d", ret); VM_OBJECT_WLOCK(object); return (VM_PAGER_FAIL); } /* If "paddr" is a real page, perform a sanity check on "memattr". */ if ((m_paddr = vm_phys_paddr_to_vm_page(paddr)) != NULL && (memattr1 = pmap_page_get_memattr(m_paddr)) != memattr) { /* * For the /dev/mem d_mmap routine to return the * correct memattr, pmap_page_get_memattr() needs to * be called, which we do there. */ if ((csw->d_flags & D_MEM) == 0) { printf("WARNING: Device driver %s has set " "\"memattr\" inconsistently (drv %u pmap %u).\n", csw->d_name, memattr, memattr1); } memattr = memattr1; } if (((*mres)->flags & PG_FICTITIOUS) != 0) { /* * If the passed in result page is a fake page, update it with * the new physical address. */ page = *mres; VM_OBJECT_WLOCK(object); vm_page_updatefake(page, paddr, memattr); } else { /* * Replace the passed in reqpage page with our own fake page and * free up the all of the original pages. */ page = vm_page_getfake(paddr, memattr); VM_OBJECT_WLOCK(object); vm_page_replace_checked(page, object, (*mres)->pindex, *mres); vm_page_free(*mres); *mres = page; } - page->valid = VM_PAGE_BITS_ALL; + vm_page_valid(page); return (VM_PAGER_OK); } static void dev_pager_putpages(vm_object_t object, vm_page_t *m, int count, int flags, int *rtvals) { panic("dev_pager_putpage called"); } static boolean_t dev_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after) { if (before != NULL) *before = 0; if (after != NULL) *after = 0; return (TRUE); } static int old_dev_pager_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t foff, struct ucred *cred, u_short *color) { struct cdev *dev; struct cdevsw *csw; vm_memattr_t dummy; vm_ooffset_t off; vm_paddr_t paddr; unsigned int npages; int ref; /* * Make sure this device can be mapped. */ dev = handle; csw = dev_refthread(dev, &ref); if (csw == NULL) return (ENXIO); /* * Check that the specified range of the device allows the desired * protection. * * XXX assumes VM_PROT_* == PROT_* */ npages = OFF_TO_IDX(size); paddr = 0; /* Make paddr initialized for the case of size == 0. */ for (off = foff; npages--; off += PAGE_SIZE) { if (csw->d_mmap(dev, off, &paddr, (int)prot, &dummy) != 0) { dev_relthread(dev, ref); return (EINVAL); } } dev_ref(dev); dev_relthread(dev, ref); *color = atop(paddr) - OFF_TO_IDX(off - PAGE_SIZE); return (0); } static void old_dev_pager_dtor(void *handle) { dev_rel(handle); } Index: head/sys/vm/phys_pager.c =================================================================== --- head/sys/vm/phys_pager.c (revision 353538) +++ head/sys/vm/phys_pager.c (revision 353539) @@ -1,253 +1,251 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2000 Peter Wemm * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* list of phys pager objects */ static struct pagerlst phys_pager_object_list; /* protect access to phys_pager_object_list */ static struct mtx phys_pager_mtx; static void phys_pager_init(void) { TAILQ_INIT(&phys_pager_object_list); mtx_init(&phys_pager_mtx, "phys_pager list", NULL, MTX_DEF); } static vm_object_t phys_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t foff, struct ucred *cred) { vm_object_t object, object1; vm_pindex_t pindex; /* * Offset should be page aligned. */ if (foff & PAGE_MASK) return (NULL); pindex = OFF_TO_IDX(foff + PAGE_MASK + size); if (handle != NULL) { mtx_lock(&phys_pager_mtx); /* * Look up pager, creating as necessary. */ object1 = NULL; object = vm_pager_object_lookup(&phys_pager_object_list, handle); if (object == NULL) { /* * Allocate object and associate it with the pager. */ mtx_unlock(&phys_pager_mtx); object1 = vm_object_allocate(OBJT_PHYS, pindex); mtx_lock(&phys_pager_mtx); object = vm_pager_object_lookup(&phys_pager_object_list, handle); if (object != NULL) { /* * We raced with other thread while * allocating object. */ if (pindex > object->size) object->size = pindex; } else { object = object1; object1 = NULL; object->handle = handle; vm_object_set_flag(object, OBJ_POPULATE); TAILQ_INSERT_TAIL(&phys_pager_object_list, object, pager_object_list); } } else { if (pindex > object->size) object->size = pindex; } mtx_unlock(&phys_pager_mtx); vm_object_deallocate(object1); } else { object = vm_object_allocate(OBJT_PHYS, pindex); vm_object_set_flag(object, OBJ_POPULATE); } return (object); } static void phys_pager_dealloc(vm_object_t object) { if (object->handle != NULL) { VM_OBJECT_WUNLOCK(object); mtx_lock(&phys_pager_mtx); TAILQ_REMOVE(&phys_pager_object_list, object, pager_object_list); mtx_unlock(&phys_pager_mtx); VM_OBJECT_WLOCK(object); } object->handle = NULL; object->type = OBJT_DEAD; } /* * Fill as many pages as vm_fault has allocated for us. */ static int phys_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind, int *rahead) { int i; VM_OBJECT_ASSERT_WLOCKED(object); for (i = 0; i < count; i++) { - if (m[i]->valid == 0) { + if (vm_page_none_valid(m[i])) { if ((m[i]->flags & PG_ZERO) == 0) pmap_zero_page(m[i]); - m[i]->valid = VM_PAGE_BITS_ALL; + vm_page_valid(m[i]); } - KASSERT(m[i]->valid == VM_PAGE_BITS_ALL, + KASSERT(vm_page_all_valid(m[i]), ("phys_pager_getpages: partially valid page %p", m[i])); KASSERT(m[i]->dirty == 0, ("phys_pager_getpages: dirty page %p", m[i])); } if (rbehind) *rbehind = 0; if (rahead) *rahead = 0; return (VM_PAGER_OK); } /* * Implement a pretty aggressive clustered getpages strategy. Hint that * everything in an entire 4MB window should be prefaulted at once. * * 4MB (1024 slots per page table page) is convenient for x86, * but may not be for other arches. */ #ifndef PHYSCLUSTER #define PHYSCLUSTER 1024 #endif static int phys_pager_cluster = PHYSCLUSTER; SYSCTL_INT(_vm, OID_AUTO, phys_pager_cluster, CTLFLAG_RWTUN, &phys_pager_cluster, 0, "prefault window size for phys pager"); /* * Max hint to vm_page_alloc() about the further allocation needs * inside the phys_pager_populate() loop. The number of bits used to * implement VM_ALLOC_COUNT() determines the hard limit on this value. * That limit is currently 65535. */ #define PHYSALLOC 16 static int phys_pager_populate(vm_object_t object, vm_pindex_t pidx, int fault_type __unused, vm_prot_t max_prot __unused, vm_pindex_t *first, vm_pindex_t *last) { vm_page_t m; vm_pindex_t base, end, i; int ahead; base = rounddown(pidx, phys_pager_cluster); end = base + phys_pager_cluster - 1; if (end >= object->size) end = object->size - 1; if (*first > base) base = *first; if (end > *last) end = *last; *first = base; *last = end; for (i = base; i <= end; i++) { ahead = MIN(end - i, PHYSALLOC); m = vm_page_grab(object, i, VM_ALLOC_NORMAL | VM_ALLOC_COUNT(ahead)); - if (m->valid != VM_PAGE_BITS_ALL) { + if (!vm_page_all_valid(m)) vm_page_zero_invalid(m, TRUE); - m->valid = VM_PAGE_BITS_ALL; - } KASSERT(m->dirty == 0, ("phys_pager_populate: dirty page %p", m)); } return (VM_PAGER_OK); } static void phys_pager_putpages(vm_object_t object, vm_page_t *m, int count, boolean_t sync, int *rtvals) { panic("phys_pager_putpage called"); } static boolean_t phys_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after) { vm_pindex_t base, end; base = rounddown(pindex, phys_pager_cluster); end = base + phys_pager_cluster - 1; if (before != NULL) *before = pindex - base; if (after != NULL) *after = end - pindex; return (TRUE); } struct pagerops physpagerops = { .pgo_init = phys_pager_init, .pgo_alloc = phys_pager_alloc, .pgo_dealloc = phys_pager_dealloc, .pgo_getpages = phys_pager_getpages, .pgo_putpages = phys_pager_putpages, .pgo_haspage = phys_pager_haspage, .pgo_populate = phys_pager_populate, }; Index: head/sys/vm/sg_pager.c =================================================================== --- head/sys/vm/sg_pager.c (revision 353538) +++ head/sys/vm/sg_pager.c (revision 353539) @@ -1,229 +1,229 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2009 Hudson River Trading LLC * Written by: John H. Baldwin * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); /* * This pager manages OBJT_SG objects. These objects are backed by * a scatter/gather list of physical address ranges. */ #include #include #include #include #include #include #include #include #include #include #include #include #include static vm_object_t sg_pager_alloc(void *, vm_ooffset_t, vm_prot_t, vm_ooffset_t, struct ucred *); static void sg_pager_dealloc(vm_object_t); static int sg_pager_getpages(vm_object_t, vm_page_t *, int, int *, int *); static void sg_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *); static boolean_t sg_pager_haspage(vm_object_t, vm_pindex_t, int *, int *); struct pagerops sgpagerops = { .pgo_alloc = sg_pager_alloc, .pgo_dealloc = sg_pager_dealloc, .pgo_getpages = sg_pager_getpages, .pgo_putpages = sg_pager_putpages, .pgo_haspage = sg_pager_haspage, }; static vm_object_t sg_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t foff, struct ucred *cred) { struct sglist *sg; vm_object_t object; vm_pindex_t npages, pindex; int i; /* * Offset should be page aligned. */ if (foff & PAGE_MASK) return (NULL); /* * The scatter/gather list must only include page-aligned * ranges. */ npages = 0; sg = handle; for (i = 0; i < sg->sg_nseg; i++) { if ((sg->sg_segs[i].ss_paddr % PAGE_SIZE) != 0 || (sg->sg_segs[i].ss_len % PAGE_SIZE) != 0) return (NULL); npages += sg->sg_segs[i].ss_len / PAGE_SIZE; } /* * The scatter/gather list has a fixed size. Refuse requests * to map beyond that. */ size = round_page(size); pindex = OFF_TO_IDX(foff) + OFF_TO_IDX(size); if (pindex > npages || pindex < OFF_TO_IDX(foff) || pindex < OFF_TO_IDX(size)) return (NULL); /* * Allocate a new object and associate it with the * scatter/gather list. It is ok for our purposes to have * multiple VM objects associated with the same scatter/gather * list because scatter/gather lists are static. This is also * simpler than ensuring a unique object per scatter/gather * list. */ object = vm_object_allocate(OBJT_SG, npages); object->handle = sglist_hold(sg); TAILQ_INIT(&object->un_pager.sgp.sgp_pglist); return (object); } static void sg_pager_dealloc(vm_object_t object) { struct sglist *sg; vm_page_t m; /* * Free up our fake pages. */ while ((m = TAILQ_FIRST(&object->un_pager.sgp.sgp_pglist)) != 0) { TAILQ_REMOVE(&object->un_pager.sgp.sgp_pglist, m, plinks.q); vm_page_putfake(m); } sg = object->handle; sglist_free(sg); object->handle = NULL; object->type = OBJT_DEAD; } static int sg_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind, int *rahead) { struct sglist *sg; vm_page_t m_paddr, page; vm_pindex_t offset; vm_paddr_t paddr; vm_memattr_t memattr; size_t space; int i; /* Since our haspage reports zero after/before, the count is 1. */ KASSERT(count == 1, ("%s: count %d", __func__, count)); VM_OBJECT_ASSERT_WLOCKED(object); sg = object->handle; memattr = object->memattr; VM_OBJECT_WUNLOCK(object); offset = m[0]->pindex; /* * Lookup the physical address of the requested page. An initial * value of '1' instead of '0' is used so we can assert that the * page is found since '0' can be a valid page-aligned physical * address. */ space = 0; paddr = 1; for (i = 0; i < sg->sg_nseg; i++) { if (space + sg->sg_segs[i].ss_len <= (offset * PAGE_SIZE)) { space += sg->sg_segs[i].ss_len; continue; } paddr = sg->sg_segs[i].ss_paddr + offset * PAGE_SIZE - space; break; } KASSERT(paddr != 1, ("invalid SG page index")); /* If "paddr" is a real page, perform a sanity check on "memattr". */ if ((m_paddr = vm_phys_paddr_to_vm_page(paddr)) != NULL && pmap_page_get_memattr(m_paddr) != memattr) { memattr = pmap_page_get_memattr(m_paddr); printf( "WARNING: A device driver has set \"memattr\" inconsistently.\n"); } /* Return a fake page for the requested page. */ KASSERT(!(m[0]->flags & PG_FICTITIOUS), ("backing page for SG is fake")); /* Construct a new fake page. */ page = vm_page_getfake(paddr, memattr); VM_OBJECT_WLOCK(object); TAILQ_INSERT_TAIL(&object->un_pager.sgp.sgp_pglist, page, plinks.q); vm_page_replace_checked(page, object, offset, m[0]); vm_page_lock(m[0]); vm_page_free(m[0]); vm_page_unlock(m[0]); m[0] = page; - page->valid = VM_PAGE_BITS_ALL; + vm_page_valid(page); if (rbehind) *rbehind = 0; if (rahead) *rahead = 0; return (VM_PAGER_OK); } static void sg_pager_putpages(vm_object_t object, vm_page_t *m, int count, boolean_t sync, int *rtvals) { panic("sg_pager_putpage called"); } static boolean_t sg_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after) { if (before != NULL) *before = 0; if (after != NULL) *after = 0; return (TRUE); } Index: head/sys/vm/swap_pager.c =================================================================== --- head/sys/vm/swap_pager.c (revision 353538) +++ head/sys/vm/swap_pager.c (revision 353539) @@ -1,3025 +1,3025 @@ /*- * SPDX-License-Identifier: BSD-4-Clause * * Copyright (c) 1998 Matthew Dillon, * Copyright (c) 1994 John S. Dyson * Copyright (c) 1990 University of Utah. * Copyright (c) 1982, 1986, 1989, 1993 * The Regents of the University of California. All rights reserved. * * This code is derived from software contributed to Berkeley by * the Systems Programming Group of the University of Utah Computer * Science Department. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. All advertising materials mentioning features or use of this software * must display the following acknowledgement: * This product includes software developed by the University of * California, Berkeley and its contributors. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * New Swap System * Matthew Dillon * * Radix Bitmap 'blists'. * * - The new swapper uses the new radix bitmap code. This should scale * to arbitrarily small or arbitrarily large swap spaces and an almost * arbitrary degree of fragmentation. * * Features: * * - on the fly reallocation of swap during putpages. The new system * does not try to keep previously allocated swap blocks for dirty * pages. * * - on the fly deallocation of swap * * - No more garbage collection required. Unnecessarily allocated swap * blocks only exist for dirty vm_page_t's now and these are already * cycled (in a high-load system) by the pager. We also do on-the-fly * removal of invalidated swap blocks when a page is destroyed * or renamed. * * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$ * * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94 * @(#)vm_swap.c 8.5 (Berkeley) 2/17/94 */ #include __FBSDID("$FreeBSD$"); #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* * MAX_PAGEOUT_CLUSTER must be a power of 2 between 1 and 64. * The 64-page limit is due to the radix code (kern/subr_blist.c). */ #ifndef MAX_PAGEOUT_CLUSTER #define MAX_PAGEOUT_CLUSTER 32 #endif #if !defined(SWB_NPAGES) #define SWB_NPAGES MAX_PAGEOUT_CLUSTER #endif #define SWAP_META_PAGES PCTRIE_COUNT /* * A swblk structure maps each page index within a * SWAP_META_PAGES-aligned and sized range to the address of an * on-disk swap block (or SWAPBLK_NONE). The collection of these * mappings for an entire vm object is implemented as a pc-trie. */ struct swblk { vm_pindex_t p; daddr_t d[SWAP_META_PAGES]; }; static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data"); static struct mtx sw_dev_mtx; static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq); static struct swdevt *swdevhd; /* Allocate from here next */ static int nswapdev; /* Number of swap devices */ int swap_pager_avail; static struct sx swdev_syscall_lock; /* serialize swap(on|off) */ static u_long swap_reserved; static u_long swap_total; static int sysctl_page_shift(SYSCTL_HANDLER_ARGS); SYSCTL_PROC(_vm, OID_AUTO, swap_reserved, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, &swap_reserved, 0, sysctl_page_shift, "A", "Amount of swap storage needed to back all allocated anonymous memory."); SYSCTL_PROC(_vm, OID_AUTO, swap_total, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, &swap_total, 0, sysctl_page_shift, "A", "Total amount of available swap storage."); static int overcommit = 0; SYSCTL_INT(_vm, VM_OVERCOMMIT, overcommit, CTLFLAG_RW, &overcommit, 0, "Configure virtual memory overcommit behavior. See tuning(7) " "for details."); static unsigned long swzone; SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0, "Actual size of swap metadata zone"); static unsigned long swap_maxpages; SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0, "Maximum amount of swap supported"); /* bits from overcommit */ #define SWAP_RESERVE_FORCE_ON (1 << 0) #define SWAP_RESERVE_RLIMIT_ON (1 << 1) #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2) static int sysctl_page_shift(SYSCTL_HANDLER_ARGS) { uint64_t newval; u_long value = *(u_long *)arg1; newval = ((uint64_t)value) << PAGE_SHIFT; return (sysctl_handle_64(oidp, &newval, 0, req)); } int swap_reserve(vm_ooffset_t incr) { return (swap_reserve_by_cred(incr, curthread->td_ucred)); } int swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred) { u_long r, s, prev, pincr; int res, error; static int curfail; static struct timeval lastfail; struct uidinfo *uip; uip = cred->cr_ruidinfo; KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__, (uintmax_t)incr)); #ifdef RACCT if (racct_enable) { PROC_LOCK(curproc); error = racct_add(curproc, RACCT_SWAP, incr); PROC_UNLOCK(curproc); if (error != 0) return (0); } #endif pincr = atop(incr); res = 0; prev = atomic_fetchadd_long(&swap_reserved, pincr); r = prev + pincr; if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) { s = vm_cnt.v_page_count - vm_cnt.v_free_reserved - vm_wire_count(); } else s = 0; s += swap_total; if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s || (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) { res = 1; } else { prev = atomic_fetchadd_long(&swap_reserved, -pincr); if (prev < pincr) panic("swap_reserved < incr on overcommit fail"); } if (res) { prev = atomic_fetchadd_long(&uip->ui_vmsize, pincr); if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 && prev + pincr > lim_cur(curthread, RLIMIT_SWAP) && priv_check(curthread, PRIV_VM_SWAP_NORLIMIT)) { res = 0; prev = atomic_fetchadd_long(&uip->ui_vmsize, -pincr); if (prev < pincr) panic("uip->ui_vmsize < incr on overcommit fail"); } } if (!res && ppsratecheck(&lastfail, &curfail, 1)) { printf("uid %d, pid %d: swap reservation for %jd bytes failed\n", uip->ui_uid, curproc->p_pid, incr); } #ifdef RACCT if (racct_enable && !res) { PROC_LOCK(curproc); racct_sub(curproc, RACCT_SWAP, incr); PROC_UNLOCK(curproc); } #endif return (res); } void swap_reserve_force(vm_ooffset_t incr) { struct uidinfo *uip; u_long pincr; KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__, (uintmax_t)incr)); PROC_LOCK(curproc); #ifdef RACCT if (racct_enable) racct_add_force(curproc, RACCT_SWAP, incr); #endif pincr = atop(incr); atomic_add_long(&swap_reserved, pincr); uip = curproc->p_ucred->cr_ruidinfo; atomic_add_long(&uip->ui_vmsize, pincr); PROC_UNLOCK(curproc); } void swap_release(vm_ooffset_t decr) { struct ucred *cred; PROC_LOCK(curproc); cred = curproc->p_ucred; swap_release_by_cred(decr, cred); PROC_UNLOCK(curproc); } void swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred) { u_long prev, pdecr; struct uidinfo *uip; uip = cred->cr_ruidinfo; KASSERT((decr & PAGE_MASK) == 0, ("%s: decr: %ju & PAGE_MASK", __func__, (uintmax_t)decr)); pdecr = atop(decr); prev = atomic_fetchadd_long(&swap_reserved, -pdecr); if (prev < pdecr) panic("swap_reserved < decr"); prev = atomic_fetchadd_long(&uip->ui_vmsize, -pdecr); if (prev < pdecr) printf("negative vmsize for uid = %d\n", uip->ui_uid); #ifdef RACCT if (racct_enable) racct_sub_cred(cred, RACCT_SWAP, decr); #endif } #define SWM_POP 0x01 /* pop out */ static int swap_pager_full = 2; /* swap space exhaustion (task killing) */ static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/ static struct mtx swbuf_mtx; /* to sync nsw_wcount_async */ static int nsw_wcount_async; /* limit async write buffers */ static int nsw_wcount_async_max;/* assigned maximum */ static int nsw_cluster_max; /* maximum VOP I/O allowed */ static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS); SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I", "Maximum running async swap ops"); static int sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS); SYSCTL_PROC(_vm, OID_AUTO, swap_fragmentation, CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_fragmentation, "A", "Swap Fragmentation Info"); static struct sx sw_alloc_sx; /* * "named" and "unnamed" anon region objects. Try to reduce the overhead * of searching a named list by hashing it just a little. */ #define NOBJLISTS 8 #define NOBJLIST(handle) \ (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)]) static struct pagerlst swap_pager_object_list[NOBJLISTS]; static uma_zone_t swwbuf_zone; static uma_zone_t swrbuf_zone; static uma_zone_t swblk_zone; static uma_zone_t swpctrie_zone; /* * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure * calls hooked from other parts of the VM system and do not appear here. * (see vm/swap_pager.h). */ static vm_object_t swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t offset, struct ucred *); static void swap_pager_dealloc(vm_object_t object); static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int *, int *); static int swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *, int *, pgo_getpages_iodone_t, void *); static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *); static boolean_t swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after); static void swap_pager_init(void); static void swap_pager_unswapped(vm_page_t); static void swap_pager_swapoff(struct swdevt *sp); static void swap_pager_update_writecount(vm_object_t object, vm_offset_t start, vm_offset_t end); static void swap_pager_release_writecount(vm_object_t object, vm_offset_t start, vm_offset_t end); struct pagerops swappagerops = { .pgo_init = swap_pager_init, /* early system initialization of pager */ .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */ .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */ .pgo_getpages = swap_pager_getpages, /* pagein */ .pgo_getpages_async = swap_pager_getpages_async, /* pagein (async) */ .pgo_putpages = swap_pager_putpages, /* pageout */ .pgo_haspage = swap_pager_haspage, /* get backing store status for page */ .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */ .pgo_update_writecount = swap_pager_update_writecount, .pgo_release_writecount = swap_pager_release_writecount, }; /* * swap_*() routines are externally accessible. swp_*() routines are * internal. */ static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */ static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */ SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0, "Maximum size of a swap block in pages"); static void swp_sizecheck(void); static void swp_pager_async_iodone(struct buf *bp); static bool swp_pager_swblk_empty(struct swblk *sb, int start, int limit); static int swapongeom(struct vnode *); static int swaponvp(struct thread *, struct vnode *, u_long); static int swapoff_one(struct swdevt *sp, struct ucred *cred); /* * Swap bitmap functions */ static void swp_pager_freeswapspace(daddr_t blk, daddr_t npages); static daddr_t swp_pager_getswapspace(int *npages, int limit); /* * Metadata functions */ static daddr_t swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t); static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t); static void swp_pager_meta_free_all(vm_object_t); static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int); static void swp_pager_init_freerange(daddr_t *start, daddr_t *num) { *start = SWAPBLK_NONE; *num = 0; } static void swp_pager_update_freerange(daddr_t *start, daddr_t *num, daddr_t addr) { if (*start + *num == addr) { (*num)++; } else { swp_pager_freeswapspace(*start, *num); *start = addr; *num = 1; } } static void * swblk_trie_alloc(struct pctrie *ptree) { return (uma_zalloc(swpctrie_zone, M_NOWAIT | (curproc == pageproc ? M_USE_RESERVE : 0))); } static void swblk_trie_free(struct pctrie *ptree, void *node) { uma_zfree(swpctrie_zone, node); } PCTRIE_DEFINE(SWAP, swblk, p, swblk_trie_alloc, swblk_trie_free); /* * SWP_SIZECHECK() - update swap_pager_full indication * * update the swap_pager_almost_full indication and warn when we are * about to run out of swap space, using lowat/hiwat hysteresis. * * Clear swap_pager_full ( task killing ) indication when lowat is met. * * No restrictions on call * This routine may not block. */ static void swp_sizecheck(void) { if (swap_pager_avail < nswap_lowat) { if (swap_pager_almost_full == 0) { printf("swap_pager: out of swap space\n"); swap_pager_almost_full = 1; } } else { swap_pager_full = 0; if (swap_pager_avail > nswap_hiwat) swap_pager_almost_full = 0; } } /* * SWAP_PAGER_INIT() - initialize the swap pager! * * Expected to be started from system init. NOTE: This code is run * before much else so be careful what you depend on. Most of the VM * system has yet to be initialized at this point. */ static void swap_pager_init(void) { /* * Initialize object lists */ int i; for (i = 0; i < NOBJLISTS; ++i) TAILQ_INIT(&swap_pager_object_list[i]); mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF); sx_init(&sw_alloc_sx, "swspsx"); sx_init(&swdev_syscall_lock, "swsysc"); } /* * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process * * Expected to be started from pageout process once, prior to entering * its main loop. */ void swap_pager_swap_init(void) { unsigned long n, n2; /* * Number of in-transit swap bp operations. Don't * exhaust the pbufs completely. Make sure we * initialize workable values (0 will work for hysteresis * but it isn't very efficient). * * The nsw_cluster_max is constrained by the bp->b_pages[] * array, which has MAXPHYS / PAGE_SIZE entries, and our locally * defined MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are * constrained by the swap device interleave stripe size. * * Currently we hardwire nsw_wcount_async to 4. This limit is * designed to prevent other I/O from having high latencies due to * our pageout I/O. The value 4 works well for one or two active swap * devices but is probably a little low if you have more. Even so, * a higher value would probably generate only a limited improvement * with three or four active swap devices since the system does not * typically have to pageout at extreme bandwidths. We will want * at least 2 per swap devices, and 4 is a pretty good value if you * have one NFS swap device due to the command/ack latency over NFS. * So it all works out pretty well. */ nsw_cluster_max = min(MAXPHYS / PAGE_SIZE, MAX_PAGEOUT_CLUSTER); nsw_wcount_async = 4; nsw_wcount_async_max = nsw_wcount_async; mtx_init(&swbuf_mtx, "async swbuf mutex", NULL, MTX_DEF); swwbuf_zone = pbuf_zsecond_create("swwbuf", nswbuf / 4); swrbuf_zone = pbuf_zsecond_create("swrbuf", nswbuf / 2); /* * Initialize our zone, taking the user's requested size or * estimating the number we need based on the number of pages * in the system. */ n = maxswzone != 0 ? maxswzone / sizeof(struct swblk) : vm_cnt.v_page_count / 2; swpctrie_zone = uma_zcreate("swpctrie", pctrie_node_size(), NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM); if (swpctrie_zone == NULL) panic("failed to create swap pctrie zone."); swblk_zone = uma_zcreate("swblk", sizeof(struct swblk), NULL, NULL, NULL, NULL, _Alignof(struct swblk) - 1, UMA_ZONE_VM); if (swblk_zone == NULL) panic("failed to create swap blk zone."); n2 = n; do { if (uma_zone_reserve_kva(swblk_zone, n)) break; /* * if the allocation failed, try a zone two thirds the * size of the previous attempt. */ n -= ((n + 2) / 3); } while (n > 0); /* * Often uma_zone_reserve_kva() cannot reserve exactly the * requested size. Account for the difference when * calculating swap_maxpages. */ n = uma_zone_get_max(swblk_zone); if (n < n2) printf("Swap blk zone entries changed from %lu to %lu.\n", n2, n); swap_maxpages = n * SWAP_META_PAGES; swzone = n * sizeof(struct swblk); if (!uma_zone_reserve_kva(swpctrie_zone, n)) printf("Cannot reserve swap pctrie zone, " "reduce kern.maxswzone.\n"); } static vm_object_t swap_pager_alloc_init(void *handle, struct ucred *cred, vm_ooffset_t size, vm_ooffset_t offset) { vm_object_t object; if (cred != NULL) { if (!swap_reserve_by_cred(size, cred)) return (NULL); crhold(cred); } /* * The un_pager.swp.swp_blks trie is initialized by * vm_object_allocate() to ensure the correct order of * visibility to other threads. */ object = vm_object_allocate(OBJT_SWAP, OFF_TO_IDX(offset + PAGE_MASK + size)); object->un_pager.swp.writemappings = 0; object->handle = handle; if (cred != NULL) { object->cred = cred; object->charge = size; } return (object); } /* * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate * its metadata structures. * * This routine is called from the mmap and fork code to create a new * OBJT_SWAP object. * * This routine must ensure that no live duplicate is created for * the named object request, which is protected against by * holding the sw_alloc_sx lock in case handle != NULL. */ static vm_object_t swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t offset, struct ucred *cred) { vm_object_t object; if (handle != NULL) { /* * Reference existing named region or allocate new one. There * should not be a race here against swp_pager_meta_build() * as called from vm_page_remove() in regards to the lookup * of the handle. */ sx_xlock(&sw_alloc_sx); object = vm_pager_object_lookup(NOBJLIST(handle), handle); if (object == NULL) { object = swap_pager_alloc_init(handle, cred, size, offset); if (object != NULL) { TAILQ_INSERT_TAIL(NOBJLIST(object->handle), object, pager_object_list); } } sx_xunlock(&sw_alloc_sx); } else { object = swap_pager_alloc_init(handle, cred, size, offset); } return (object); } /* * SWAP_PAGER_DEALLOC() - remove swap metadata from object * * The swap backing for the object is destroyed. The code is * designed such that we can reinstantiate it later, but this * routine is typically called only when the entire object is * about to be destroyed. * * The object must be locked. */ static void swap_pager_dealloc(vm_object_t object) { VM_OBJECT_ASSERT_WLOCKED(object); KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj")); /* * Remove from list right away so lookups will fail if we block for * pageout completion. */ if (object->handle != NULL) { VM_OBJECT_WUNLOCK(object); sx_xlock(&sw_alloc_sx); TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list); sx_xunlock(&sw_alloc_sx); VM_OBJECT_WLOCK(object); } vm_object_pip_wait(object, "swpdea"); /* * Free all remaining metadata. We only bother to free it from * the swap meta data. We do not attempt to free swapblk's still * associated with vm_page_t's for this object. We do not care * if paging is still in progress on some objects. */ swp_pager_meta_free_all(object); object->handle = NULL; object->type = OBJT_DEAD; } /************************************************************************ * SWAP PAGER BITMAP ROUTINES * ************************************************************************/ /* * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space * * Allocate swap for up to the requested number of pages, and at * least a minimum number of pages. The starting swap block number * (a page index) is returned or SWAPBLK_NONE if the allocation * failed. * * Also has the side effect of advising that somebody made a mistake * when they configured swap and didn't configure enough. * * This routine may not sleep. * * We allocate in round-robin fashion from the configured devices. */ static daddr_t swp_pager_getswapspace(int *io_npages, int limit) { daddr_t blk; struct swdevt *sp; int mpages, npages; blk = SWAPBLK_NONE; mpages = *io_npages; npages = imin(BLIST_MAX_ALLOC, mpages); mtx_lock(&sw_dev_mtx); sp = swdevhd; while (!TAILQ_EMPTY(&swtailq)) { if (sp == NULL) sp = TAILQ_FIRST(&swtailq); if ((sp->sw_flags & SW_CLOSING) == 0) blk = blist_alloc(sp->sw_blist, &npages, mpages); if (blk != SWAPBLK_NONE) break; sp = TAILQ_NEXT(sp, sw_list); if (swdevhd == sp) { if (npages <= limit) break; mpages = npages - 1; npages >>= 1; } } if (blk != SWAPBLK_NONE) { *io_npages = npages; blk += sp->sw_first; sp->sw_used += npages; swap_pager_avail -= npages; swp_sizecheck(); swdevhd = TAILQ_NEXT(sp, sw_list); } else { if (swap_pager_full != 2) { printf("swp_pager_getswapspace(%d): failed\n", *io_npages); swap_pager_full = 2; swap_pager_almost_full = 1; } swdevhd = NULL; } mtx_unlock(&sw_dev_mtx); return (blk); } static bool swp_pager_isondev(daddr_t blk, struct swdevt *sp) { return (blk >= sp->sw_first && blk < sp->sw_end); } static void swp_pager_strategy(struct buf *bp) { struct swdevt *sp; mtx_lock(&sw_dev_mtx); TAILQ_FOREACH(sp, &swtailq, sw_list) { if (swp_pager_isondev(bp->b_blkno, sp)) { mtx_unlock(&sw_dev_mtx); if ((sp->sw_flags & SW_UNMAPPED) != 0 && unmapped_buf_allowed) { bp->b_data = unmapped_buf; bp->b_offset = 0; } else { pmap_qenter((vm_offset_t)bp->b_data, &bp->b_pages[0], bp->b_bcount / PAGE_SIZE); } sp->sw_strategy(bp, sp); return; } } panic("Swapdev not found"); } /* * SWP_PAGER_FREESWAPSPACE() - free raw swap space * * This routine returns the specified swap blocks back to the bitmap. * * This routine may not sleep. */ static void swp_pager_freeswapspace(daddr_t blk, daddr_t npages) { struct swdevt *sp; if (npages == 0) return; mtx_lock(&sw_dev_mtx); TAILQ_FOREACH(sp, &swtailq, sw_list) { if (swp_pager_isondev(blk, sp)) { sp->sw_used -= npages; /* * If we are attempting to stop swapping on * this device, we don't want to mark any * blocks free lest they be reused. */ if ((sp->sw_flags & SW_CLOSING) == 0) { blist_free(sp->sw_blist, blk - sp->sw_first, npages); swap_pager_avail += npages; swp_sizecheck(); } mtx_unlock(&sw_dev_mtx); return; } } panic("Swapdev not found"); } /* * SYSCTL_SWAP_FRAGMENTATION() - produce raw swap space stats */ static int sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS) { struct sbuf sbuf; struct swdevt *sp; const char *devname; int error; error = sysctl_wire_old_buffer(req, 0); if (error != 0) return (error); sbuf_new_for_sysctl(&sbuf, NULL, 128, req); mtx_lock(&sw_dev_mtx); TAILQ_FOREACH(sp, &swtailq, sw_list) { if (vn_isdisk(sp->sw_vp, NULL)) devname = devtoname(sp->sw_vp->v_rdev); else devname = "[file]"; sbuf_printf(&sbuf, "\nFree space on device %s:\n", devname); blist_stats(sp->sw_blist, &sbuf); } mtx_unlock(&sw_dev_mtx); error = sbuf_finish(&sbuf); sbuf_delete(&sbuf); return (error); } /* * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page * range within an object. * * This is a globally accessible routine. * * This routine removes swapblk assignments from swap metadata. * * The external callers of this routine typically have already destroyed * or renamed vm_page_t's associated with this range in the object so * we should be ok. * * The object must be locked. */ void swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size) { swp_pager_meta_free(object, start, size); } /* * SWAP_PAGER_RESERVE() - reserve swap blocks in object * * Assigns swap blocks to the specified range within the object. The * swap blocks are not zeroed. Any previous swap assignment is destroyed. * * Returns 0 on success, -1 on failure. */ int swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size) { daddr_t addr, blk, n_free, s_free; int i, j, n; swp_pager_init_freerange(&s_free, &n_free); VM_OBJECT_WLOCK(object); for (i = 0; i < size; i += n) { n = size - i; blk = swp_pager_getswapspace(&n, 1); if (blk == SWAPBLK_NONE) { swp_pager_meta_free(object, start, i); VM_OBJECT_WUNLOCK(object); return (-1); } for (j = 0; j < n; ++j) { addr = swp_pager_meta_build(object, start + i + j, blk + j); if (addr != SWAPBLK_NONE) swp_pager_update_freerange(&s_free, &n_free, addr); } } swp_pager_freeswapspace(s_free, n_free); VM_OBJECT_WUNLOCK(object); return (0); } /* * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager * and destroy the source. * * Copy any valid swapblks from the source to the destination. In * cases where both the source and destination have a valid swapblk, * we keep the destination's. * * This routine is allowed to sleep. It may sleep allocating metadata * indirectly through swp_pager_meta_build() or if paging is still in * progress on the source. * * The source object contains no vm_page_t's (which is just as well) * * The source object is of type OBJT_SWAP. * * The source and destination objects must be locked. * Both object locks may temporarily be released. */ void swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject, vm_pindex_t offset, int destroysource) { vm_pindex_t i; daddr_t dstaddr, n_free, s_free, srcaddr; VM_OBJECT_ASSERT_WLOCKED(srcobject); VM_OBJECT_ASSERT_WLOCKED(dstobject); /* * If destroysource is set, we remove the source object from the * swap_pager internal queue now. */ if (destroysource && srcobject->handle != NULL) { vm_object_pip_add(srcobject, 1); VM_OBJECT_WUNLOCK(srcobject); vm_object_pip_add(dstobject, 1); VM_OBJECT_WUNLOCK(dstobject); sx_xlock(&sw_alloc_sx); TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject, pager_object_list); sx_xunlock(&sw_alloc_sx); VM_OBJECT_WLOCK(dstobject); vm_object_pip_wakeup(dstobject); VM_OBJECT_WLOCK(srcobject); vm_object_pip_wakeup(srcobject); } /* * Transfer source to destination. */ swp_pager_init_freerange(&s_free, &n_free); for (i = 0; i < dstobject->size; ++i) { srcaddr = swp_pager_meta_ctl(srcobject, i + offset, SWM_POP); if (srcaddr == SWAPBLK_NONE) continue; dstaddr = swp_pager_meta_ctl(dstobject, i, 0); if (dstaddr != SWAPBLK_NONE) { /* * Destination has valid swapblk or it is represented * by a resident page. We destroy the source block. */ swp_pager_update_freerange(&s_free, &n_free, srcaddr); continue; } /* * Destination has no swapblk and is not resident, * copy source. * * swp_pager_meta_build() can sleep. */ vm_object_pip_add(srcobject, 1); VM_OBJECT_WUNLOCK(srcobject); vm_object_pip_add(dstobject, 1); dstaddr = swp_pager_meta_build(dstobject, i, srcaddr); KASSERT(dstaddr == SWAPBLK_NONE, ("Unexpected destination swapblk")); vm_object_pip_wakeup(dstobject); VM_OBJECT_WLOCK(srcobject); vm_object_pip_wakeup(srcobject); } swp_pager_freeswapspace(s_free, n_free); /* * Free left over swap blocks in source. * * We have to revert the type to OBJT_DEFAULT so we do not accidentally * double-remove the object from the swap queues. */ if (destroysource) { swp_pager_meta_free_all(srcobject); /* * Reverting the type is not necessary, the caller is going * to destroy srcobject directly, but I'm doing it here * for consistency since we've removed the object from its * queues. */ srcobject->type = OBJT_DEFAULT; } } /* * SWAP_PAGER_HASPAGE() - determine if we have good backing store for * the requested page. * * We determine whether good backing store exists for the requested * page and return TRUE if it does, FALSE if it doesn't. * * If TRUE, we also try to determine how much valid, contiguous backing * store exists before and after the requested page. */ static boolean_t swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after) { daddr_t blk, blk0; int i; VM_OBJECT_ASSERT_LOCKED(object); /* * do we have good backing store at the requested index ? */ blk0 = swp_pager_meta_ctl(object, pindex, 0); if (blk0 == SWAPBLK_NONE) { if (before) *before = 0; if (after) *after = 0; return (FALSE); } /* * find backwards-looking contiguous good backing store */ if (before != NULL) { for (i = 1; i < SWB_NPAGES; i++) { if (i > pindex) break; blk = swp_pager_meta_ctl(object, pindex - i, 0); if (blk != blk0 - i) break; } *before = i - 1; } /* * find forward-looking contiguous good backing store */ if (after != NULL) { for (i = 1; i < SWB_NPAGES; i++) { blk = swp_pager_meta_ctl(object, pindex + i, 0); if (blk != blk0 + i) break; } *after = i - 1; } return (TRUE); } /* * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page * * This removes any associated swap backing store, whether valid or * not, from the page. * * This routine is typically called when a page is made dirty, at * which point any associated swap can be freed. MADV_FREE also * calls us in a special-case situation * * NOTE!!! If the page is clean and the swap was valid, the caller * should make the page dirty before calling this routine. This routine * does NOT change the m->dirty status of the page. Also: MADV_FREE * depends on it. * * This routine may not sleep. * * The object containing the page must be locked. */ static void swap_pager_unswapped(vm_page_t m) { daddr_t srcaddr; srcaddr = swp_pager_meta_ctl(m->object, m->pindex, SWM_POP); if (srcaddr != SWAPBLK_NONE) swp_pager_freeswapspace(srcaddr, 1); } /* * swap_pager_getpages() - bring pages in from swap * * Attempt to page in the pages in array "ma" of length "count". The * caller may optionally specify that additional pages preceding and * succeeding the specified range be paged in. The number of such pages * is returned in the "rbehind" and "rahead" parameters, and they will * be in the inactive queue upon return. * * The pages in "ma" must be busied and will remain busied upon return. */ static int swap_pager_getpages(vm_object_t object, vm_page_t *ma, int count, int *rbehind, int *rahead) { struct buf *bp; vm_page_t bm, mpred, msucc, p; vm_pindex_t pindex; daddr_t blk; int i, maxahead, maxbehind, reqcount; reqcount = count; /* * Determine the final number of read-behind pages and * allocate them BEFORE releasing the object lock. Otherwise, * there can be a problematic race with vm_object_split(). * Specifically, vm_object_split() might first transfer pages * that precede ma[0] in the current object to a new object, * and then this function incorrectly recreates those pages as * read-behind pages in the current object. */ if (!swap_pager_haspage(object, ma[0]->pindex, &maxbehind, &maxahead)) return (VM_PAGER_FAIL); /* * Clip the readahead and readbehind ranges to exclude resident pages. */ if (rahead != NULL) { KASSERT(reqcount - 1 <= maxahead, ("page count %d extends beyond swap block", reqcount)); *rahead = imin(*rahead, maxahead - (reqcount - 1)); pindex = ma[reqcount - 1]->pindex; msucc = TAILQ_NEXT(ma[reqcount - 1], listq); if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead) *rahead = msucc->pindex - pindex - 1; } if (rbehind != NULL) { *rbehind = imin(*rbehind, maxbehind); pindex = ma[0]->pindex; mpred = TAILQ_PREV(ma[0], pglist, listq); if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind) *rbehind = pindex - mpred->pindex - 1; } bm = ma[0]; for (i = 0; i < count; i++) ma[i]->oflags |= VPO_SWAPINPROG; /* * Allocate readahead and readbehind pages. */ if (rbehind != NULL) { for (i = 1; i <= *rbehind; i++) { p = vm_page_alloc(object, ma[0]->pindex - i, VM_ALLOC_NORMAL); if (p == NULL) break; p->oflags |= VPO_SWAPINPROG; bm = p; } *rbehind = i - 1; } if (rahead != NULL) { for (i = 0; i < *rahead; i++) { p = vm_page_alloc(object, ma[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL); if (p == NULL) break; p->oflags |= VPO_SWAPINPROG; } *rahead = i; } if (rbehind != NULL) count += *rbehind; if (rahead != NULL) count += *rahead; vm_object_pip_add(object, count); pindex = bm->pindex; blk = swp_pager_meta_ctl(object, pindex, 0); KASSERT(blk != SWAPBLK_NONE, ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex)); VM_OBJECT_WUNLOCK(object); bp = uma_zalloc(swrbuf_zone, M_WAITOK); /* Pages cannot leave the object while busy. */ for (i = 0, p = bm; i < count; i++, p = TAILQ_NEXT(p, listq)) { MPASS(p->pindex == bm->pindex + i); bp->b_pages[i] = p; } bp->b_flags |= B_PAGING; bp->b_iocmd = BIO_READ; bp->b_iodone = swp_pager_async_iodone; bp->b_rcred = crhold(thread0.td_ucred); bp->b_wcred = crhold(thread0.td_ucred); bp->b_blkno = blk; bp->b_bcount = PAGE_SIZE * count; bp->b_bufsize = PAGE_SIZE * count; bp->b_npages = count; bp->b_pgbefore = rbehind != NULL ? *rbehind : 0; bp->b_pgafter = rahead != NULL ? *rahead : 0; VM_CNT_INC(v_swapin); VM_CNT_ADD(v_swappgsin, count); /* * perform the I/O. NOTE!!! bp cannot be considered valid after * this point because we automatically release it on completion. * Instead, we look at the one page we are interested in which we * still hold a lock on even through the I/O completion. * * The other pages in our ma[] array are also released on completion, * so we cannot assume they are valid anymore either. * * NOTE: b_blkno is destroyed by the call to swapdev_strategy */ BUF_KERNPROC(bp); swp_pager_strategy(bp); /* * Wait for the pages we want to complete. VPO_SWAPINPROG is always * cleared on completion. If an I/O error occurs, SWAPBLK_NONE * is set in the metadata for each page in the request. */ VM_OBJECT_WLOCK(object); while ((ma[0]->oflags & VPO_SWAPINPROG) != 0) { ma[0]->oflags |= VPO_SWAPSLEEP; VM_CNT_INC(v_intrans); if (VM_OBJECT_SLEEP(object, &object->handle, PSWP, "swread", hz * 20)) { printf( "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n", bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount); } } /* * If we had an unrecoverable read error pages will not be valid. */ for (i = 0; i < reqcount; i++) if (ma[i]->valid != VM_PAGE_BITS_ALL) return (VM_PAGER_ERROR); return (VM_PAGER_OK); /* * A final note: in a low swap situation, we cannot deallocate swap * and mark a page dirty here because the caller is likely to mark * the page clean when we return, causing the page to possibly revert * to all-zero's later. */ } /* * swap_pager_getpages_async(): * * Right now this is emulation of asynchronous operation on top of * swap_pager_getpages(). */ static int swap_pager_getpages_async(vm_object_t object, vm_page_t *ma, int count, int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg) { int r, error; r = swap_pager_getpages(object, ma, count, rbehind, rahead); VM_OBJECT_WUNLOCK(object); switch (r) { case VM_PAGER_OK: error = 0; break; case VM_PAGER_ERROR: error = EIO; break; case VM_PAGER_FAIL: error = EINVAL; break; default: panic("unhandled swap_pager_getpages() error %d", r); } (iodone)(arg, ma, count, error); VM_OBJECT_WLOCK(object); return (r); } /* * swap_pager_putpages: * * Assign swap (if necessary) and initiate I/O on the specified pages. * * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects * are automatically converted to SWAP objects. * * In a low memory situation we may block in VOP_STRATEGY(), but the new * vm_page reservation system coupled with properly written VFS devices * should ensure that no low-memory deadlock occurs. This is an area * which needs work. * * The parent has N vm_object_pip_add() references prior to * calling us and will remove references for rtvals[] that are * not set to VM_PAGER_PEND. We need to remove the rest on I/O * completion. * * The parent has soft-busy'd the pages it passes us and will unbusy * those whose rtvals[] entry is not set to VM_PAGER_PEND on return. * We need to unbusy the rest on I/O completion. */ static void swap_pager_putpages(vm_object_t object, vm_page_t *ma, int count, int flags, int *rtvals) { struct buf *bp; daddr_t addr, blk, n_free, s_free; vm_page_t mreq; int i, j, n; bool async; KASSERT(count == 0 || ma[0]->object == object, ("%s: object mismatch %p/%p", __func__, object, ma[0]->object)); /* * Step 1 * * Turn object into OBJT_SWAP. Force sync if not a pageout process. */ if (object->type != OBJT_SWAP) { addr = swp_pager_meta_build(object, 0, SWAPBLK_NONE); KASSERT(addr == SWAPBLK_NONE, ("unexpected object swap block")); } VM_OBJECT_WUNLOCK(object); async = curproc == pageproc && (flags & VM_PAGER_PUT_SYNC) == 0; swp_pager_init_freerange(&s_free, &n_free); /* * Step 2 * * Assign swap blocks and issue I/O. We reallocate swap on the fly. * The page is left dirty until the pageout operation completes * successfully. */ for (i = 0; i < count; i += n) { /* Maximum I/O size is limited by maximum swap block size. */ n = min(count - i, nsw_cluster_max); /* Get a block of swap of size up to size n. */ blk = swp_pager_getswapspace(&n, 4); if (blk == SWAPBLK_NONE) { for (j = 0; j < n; ++j) rtvals[i + j] = VM_PAGER_FAIL; continue; } /* * All I/O parameters have been satisfied. Build the I/O * request and assign the swap space. */ if (async) { mtx_lock(&swbuf_mtx); while (nsw_wcount_async == 0) msleep(&nsw_wcount_async, &swbuf_mtx, PVM, "swbufa", 0); nsw_wcount_async--; mtx_unlock(&swbuf_mtx); } bp = uma_zalloc(swwbuf_zone, M_WAITOK); if (async) bp->b_flags = B_ASYNC; bp->b_flags |= B_PAGING; bp->b_iocmd = BIO_WRITE; bp->b_rcred = crhold(thread0.td_ucred); bp->b_wcred = crhold(thread0.td_ucred); bp->b_bcount = PAGE_SIZE * n; bp->b_bufsize = PAGE_SIZE * n; bp->b_blkno = blk; VM_OBJECT_WLOCK(object); for (j = 0; j < n; ++j) { mreq = ma[i + j]; addr = swp_pager_meta_build(mreq->object, mreq->pindex, blk + j); if (addr != SWAPBLK_NONE) swp_pager_update_freerange(&s_free, &n_free, addr); MPASS(mreq->dirty == VM_PAGE_BITS_ALL); mreq->oflags |= VPO_SWAPINPROG; bp->b_pages[j] = mreq; } VM_OBJECT_WUNLOCK(object); bp->b_npages = n; /* * Must set dirty range for NFS to work. */ bp->b_dirtyoff = 0; bp->b_dirtyend = bp->b_bcount; VM_CNT_INC(v_swapout); VM_CNT_ADD(v_swappgsout, bp->b_npages); /* * We unconditionally set rtvals[] to VM_PAGER_PEND so that we * can call the async completion routine at the end of a * synchronous I/O operation. Otherwise, our caller would * perform duplicate unbusy and wakeup operations on the page * and object, respectively. */ for (j = 0; j < n; j++) rtvals[i + j] = VM_PAGER_PEND; /* * asynchronous * * NOTE: b_blkno is destroyed by the call to swapdev_strategy. */ if (async) { bp->b_iodone = swp_pager_async_iodone; BUF_KERNPROC(bp); swp_pager_strategy(bp); continue; } /* * synchronous * * NOTE: b_blkno is destroyed by the call to swapdev_strategy. */ bp->b_iodone = bdone; swp_pager_strategy(bp); /* * Wait for the sync I/O to complete. */ bwait(bp, PVM, "swwrt"); /* * Now that we are through with the bp, we can call the * normal async completion, which frees everything up. */ swp_pager_async_iodone(bp); } swp_pager_freeswapspace(s_free, n_free); VM_OBJECT_WLOCK(object); } /* * swp_pager_async_iodone: * * Completion routine for asynchronous reads and writes from/to swap. * Also called manually by synchronous code to finish up a bp. * * This routine may not sleep. */ static void swp_pager_async_iodone(struct buf *bp) { int i; vm_object_t object = NULL; /* * Report error - unless we ran out of memory, in which case * we've already logged it in swapgeom_strategy(). */ if (bp->b_ioflags & BIO_ERROR && bp->b_error != ENOMEM) { printf( "swap_pager: I/O error - %s failed; blkno %ld," "size %ld, error %d\n", ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"), (long)bp->b_blkno, (long)bp->b_bcount, bp->b_error ); } /* * remove the mapping for kernel virtual */ if (buf_mapped(bp)) pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages); else bp->b_data = bp->b_kvabase; if (bp->b_npages) { object = bp->b_pages[0]->object; VM_OBJECT_WLOCK(object); } /* * cleanup pages. If an error occurs writing to swap, we are in * very serious trouble. If it happens to be a disk error, though, * we may be able to recover by reassigning the swap later on. So * in this case we remove the m->swapblk assignment for the page * but do not free it in the rlist. The errornous block(s) are thus * never reallocated as swap. Redirty the page and continue. */ for (i = 0; i < bp->b_npages; ++i) { vm_page_t m = bp->b_pages[i]; m->oflags &= ~VPO_SWAPINPROG; if (m->oflags & VPO_SWAPSLEEP) { m->oflags &= ~VPO_SWAPSLEEP; wakeup(&object->handle); } if (bp->b_ioflags & BIO_ERROR) { /* * If an error occurs I'd love to throw the swapblk * away without freeing it back to swapspace, so it * can never be used again. But I can't from an * interrupt. */ if (bp->b_iocmd == BIO_READ) { /* * NOTE: for reads, m->dirty will probably * be overridden by the original caller of * getpages so don't play cute tricks here. */ - m->valid = 0; + vm_page_invalid(m); } else { /* * If a write error occurs, reactivate page * so it doesn't clog the inactive list, * then finish the I/O. */ MPASS(m->dirty == VM_PAGE_BITS_ALL); vm_page_lock(m); vm_page_activate(m); vm_page_unlock(m); vm_page_sunbusy(m); } } else if (bp->b_iocmd == BIO_READ) { /* * NOTE: for reads, m->dirty will probably be * overridden by the original caller of getpages so * we cannot set them in order to free the underlying * swap in a low-swap situation. I don't think we'd * want to do that anyway, but it was an optimization * that existed in the old swapper for a time before * it got ripped out due to precisely this problem. */ KASSERT(!pmap_page_is_mapped(m), ("swp_pager_async_iodone: page %p is mapped", m)); KASSERT(m->dirty == 0, ("swp_pager_async_iodone: page %p is dirty", m)); - m->valid = VM_PAGE_BITS_ALL; + vm_page_valid(m); if (i < bp->b_pgbefore || i >= bp->b_npages - bp->b_pgafter) vm_page_readahead_finish(m); } else { /* * For write success, clear the dirty * status, then finish the I/O ( which decrements the * busy count and possibly wakes waiter's up ). * A page is only written to swap after a period of * inactivity. Therefore, we do not expect it to be * reused. */ KASSERT(!pmap_page_is_write_mapped(m), ("swp_pager_async_iodone: page %p is not write" " protected", m)); vm_page_undirty(m); vm_page_lock(m); vm_page_deactivate_noreuse(m); vm_page_unlock(m); vm_page_sunbusy(m); } } /* * adjust pip. NOTE: the original parent may still have its own * pip refs on the object. */ if (object != NULL) { vm_object_pip_wakeupn(object, bp->b_npages); VM_OBJECT_WUNLOCK(object); } /* * swapdev_strategy() manually sets b_vp and b_bufobj before calling * bstrategy(). Set them back to NULL now we're done with it, or we'll * trigger a KASSERT in relpbuf(). */ if (bp->b_vp) { bp->b_vp = NULL; bp->b_bufobj = NULL; } /* * release the physical I/O buffer */ if (bp->b_flags & B_ASYNC) { mtx_lock(&swbuf_mtx); if (++nsw_wcount_async == 1) wakeup(&nsw_wcount_async); mtx_unlock(&swbuf_mtx); } uma_zfree((bp->b_iocmd == BIO_READ) ? swrbuf_zone : swwbuf_zone, bp); } int swap_pager_nswapdev(void) { return (nswapdev); } static void swp_pager_force_dirty(vm_page_t m) { vm_page_dirty(m); #ifdef INVARIANTS vm_page_lock(m); if (!vm_page_wired(m) && m->queue == PQ_NONE) panic("page %p is neither wired nor queued", m); vm_page_unlock(m); #endif vm_page_xunbusy(m); swap_pager_unswapped(m); } static void swp_pager_force_launder(vm_page_t m) { vm_page_dirty(m); vm_page_lock(m); vm_page_launder(m); vm_page_unlock(m); vm_page_xunbusy(m); swap_pager_unswapped(m); } /* * SWP_PAGER_FORCE_PAGEIN() - force swap blocks to be paged in * * This routine dissociates pages starting at the given index within an * object from their backing store, paging them in if they do not reside * in memory. Pages that are paged in are marked dirty and placed in the * laundry queue. Pages are marked dirty because they no longer have * backing store. They are placed in the laundry queue because they have * not been accessed recently. Otherwise, they would already reside in * memory. */ static void swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex, int npages) { vm_page_t ma[npages]; int i, j; KASSERT(npages > 0, ("%s: No pages", __func__)); KASSERT(npages <= MAXPHYS / PAGE_SIZE, ("%s: Too many pages: %d", __func__, npages)); vm_object_pip_add(object, npages); vm_page_grab_pages(object, pindex, VM_ALLOC_NORMAL, ma, npages); for (i = j = 0;; i++) { /* Count nonresident pages, to page-in all at once. */ if (i < npages && ma[i]->valid != VM_PAGE_BITS_ALL) continue; if (j < i) { /* Page-in nonresident pages. Mark for laundering. */ if (swap_pager_getpages(object, &ma[j], i - j, NULL, NULL) != VM_PAGER_OK) panic("%s: read from swap failed", __func__); do { swp_pager_force_launder(ma[j]); } while (++j < i); } if (i == npages) break; /* Mark dirty a resident page. */ swp_pager_force_dirty(ma[j++]); } vm_object_pip_wakeupn(object, npages); } /* * swap_pager_swapoff_object: * * Page in all of the pages that have been paged out for an object * to a swap device. */ static void swap_pager_swapoff_object(struct swdevt *sp, vm_object_t object) { struct swblk *sb; vm_pindex_t pi, s_pindex; daddr_t blk, n_blks, s_blk; int i; n_blks = 0; for (pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE( &object->un_pager.swp.swp_blks, pi)) != NULL; ) { for (i = 0; i < SWAP_META_PAGES; i++) { blk = sb->d[i]; if (!swp_pager_isondev(blk, sp)) blk = SWAPBLK_NONE; /* * If there are no blocks/pages accumulated, start a new * accumulation here. */ if (n_blks == 0) { if (blk != SWAPBLK_NONE) { s_blk = blk; s_pindex = sb->p + i; n_blks = 1; } continue; } /* * If the accumulation can be extended without breaking * the sequence of consecutive blocks and pages that * swp_pager_force_pagein() depends on, do so. */ if (n_blks < MAXPHYS / PAGE_SIZE && s_blk + n_blks == blk && s_pindex + n_blks == sb->p + i) { ++n_blks; continue; } /* * The sequence of consecutive blocks and pages cannot * be extended, so page them all in here. Then, * because doing so involves releasing and reacquiring * a lock that protects the swap block pctrie, do not * rely on the current swap block. Break this loop and * re-fetch the same pindex from the pctrie again. */ swp_pager_force_pagein(object, s_pindex, n_blks); n_blks = 0; break; } if (i == SWAP_META_PAGES) pi = sb->p + SWAP_META_PAGES; } if (n_blks > 0) swp_pager_force_pagein(object, s_pindex, n_blks); } /* * swap_pager_swapoff: * * Page in all of the pages that have been paged out to the * given device. The corresponding blocks in the bitmap must be * marked as allocated and the device must be flagged SW_CLOSING. * There may be no processes swapped out to the device. * * This routine may block. */ static void swap_pager_swapoff(struct swdevt *sp) { vm_object_t object; int retries; sx_assert(&swdev_syscall_lock, SA_XLOCKED); retries = 0; full_rescan: mtx_lock(&vm_object_list_mtx); TAILQ_FOREACH(object, &vm_object_list, object_list) { if (object->type != OBJT_SWAP) continue; mtx_unlock(&vm_object_list_mtx); /* Depends on type-stability. */ VM_OBJECT_WLOCK(object); /* * Dead objects are eventually terminated on their own. */ if ((object->flags & OBJ_DEAD) != 0) goto next_obj; /* * Sync with fences placed after pctrie * initialization. We must not access pctrie below * unless we checked that our object is swap and not * dead. */ atomic_thread_fence_acq(); if (object->type != OBJT_SWAP) goto next_obj; swap_pager_swapoff_object(sp, object); next_obj: VM_OBJECT_WUNLOCK(object); mtx_lock(&vm_object_list_mtx); } mtx_unlock(&vm_object_list_mtx); if (sp->sw_used) { /* * Objects may be locked or paging to the device being * removed, so we will miss their pages and need to * make another pass. We have marked this device as * SW_CLOSING, so the activity should finish soon. */ retries++; if (retries > 100) { panic("swapoff: failed to locate %d swap blocks", sp->sw_used); } pause("swpoff", hz / 20); goto full_rescan; } EVENTHANDLER_INVOKE(swapoff, sp); } /************************************************************************ * SWAP META DATA * ************************************************************************ * * These routines manipulate the swap metadata stored in the * OBJT_SWAP object. * * Swap metadata is implemented with a global hash and not directly * linked into the object. Instead the object simply contains * appropriate tracking counters. */ /* * SWP_PAGER_SWBLK_EMPTY() - is a range of blocks free? */ static bool swp_pager_swblk_empty(struct swblk *sb, int start, int limit) { int i; MPASS(0 <= start && start <= limit && limit <= SWAP_META_PAGES); for (i = start; i < limit; i++) { if (sb->d[i] != SWAPBLK_NONE) return (false); } return (true); } /* * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object * * We first convert the object to a swap object if it is a default * object. * * The specified swapblk is added to the object's swap metadata. If * the swapblk is not valid, it is freed instead. Any previously * assigned swapblk is returned. */ static daddr_t swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk) { static volatile int swblk_zone_exhausted, swpctrie_zone_exhausted; struct swblk *sb, *sb1; vm_pindex_t modpi, rdpi; daddr_t prev_swapblk; int error, i; VM_OBJECT_ASSERT_WLOCKED(object); /* * Convert default object to swap object if necessary */ if (object->type != OBJT_SWAP) { pctrie_init(&object->un_pager.swp.swp_blks); /* * Ensure that swap_pager_swapoff()'s iteration over * object_list does not see a garbage pctrie. */ atomic_thread_fence_rel(); object->type = OBJT_SWAP; object->un_pager.swp.writemappings = 0; KASSERT(object->handle == NULL, ("default pager with handle")); } rdpi = rounddown(pindex, SWAP_META_PAGES); sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi); if (sb == NULL) { if (swapblk == SWAPBLK_NONE) return (SWAPBLK_NONE); for (;;) { sb = uma_zalloc(swblk_zone, M_NOWAIT | (curproc == pageproc ? M_USE_RESERVE : 0)); if (sb != NULL) { sb->p = rdpi; for (i = 0; i < SWAP_META_PAGES; i++) sb->d[i] = SWAPBLK_NONE; if (atomic_cmpset_int(&swblk_zone_exhausted, 1, 0)) printf("swblk zone ok\n"); break; } VM_OBJECT_WUNLOCK(object); if (uma_zone_exhausted(swblk_zone)) { if (atomic_cmpset_int(&swblk_zone_exhausted, 0, 1)) printf("swap blk zone exhausted, " "increase kern.maxswzone\n"); vm_pageout_oom(VM_OOM_SWAPZ); pause("swzonxb", 10); } else uma_zwait(swblk_zone); VM_OBJECT_WLOCK(object); sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi); if (sb != NULL) /* * Somebody swapped out a nearby page, * allocating swblk at the rdpi index, * while we dropped the object lock. */ goto allocated; } for (;;) { error = SWAP_PCTRIE_INSERT( &object->un_pager.swp.swp_blks, sb); if (error == 0) { if (atomic_cmpset_int(&swpctrie_zone_exhausted, 1, 0)) printf("swpctrie zone ok\n"); break; } VM_OBJECT_WUNLOCK(object); if (uma_zone_exhausted(swpctrie_zone)) { if (atomic_cmpset_int(&swpctrie_zone_exhausted, 0, 1)) printf("swap pctrie zone exhausted, " "increase kern.maxswzone\n"); vm_pageout_oom(VM_OOM_SWAPZ); pause("swzonxp", 10); } else uma_zwait(swpctrie_zone); VM_OBJECT_WLOCK(object); sb1 = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi); if (sb1 != NULL) { uma_zfree(swblk_zone, sb); sb = sb1; goto allocated; } } } allocated: MPASS(sb->p == rdpi); modpi = pindex % SWAP_META_PAGES; /* Return prior contents of metadata. */ prev_swapblk = sb->d[modpi]; /* Enter block into metadata. */ sb->d[modpi] = swapblk; /* * Free the swblk if we end up with the empty page run. */ if (swapblk == SWAPBLK_NONE && swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) { SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, rdpi); uma_zfree(swblk_zone, sb); } return (prev_swapblk); } /* * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata * * The requested range of blocks is freed, with any associated swap * returned to the swap bitmap. * * This routine will free swap metadata structures as they are cleaned * out. This routine does *NOT* operate on swap metadata associated * with resident pages. */ static void swp_pager_meta_free(vm_object_t object, vm_pindex_t pindex, vm_pindex_t count) { struct swblk *sb; daddr_t n_free, s_free; vm_pindex_t last; int i, limit, start; VM_OBJECT_ASSERT_WLOCKED(object); if (object->type != OBJT_SWAP || count == 0) return; swp_pager_init_freerange(&s_free, &n_free); last = pindex + count; for (;;) { sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks, rounddown(pindex, SWAP_META_PAGES)); if (sb == NULL || sb->p >= last) break; start = pindex > sb->p ? pindex - sb->p : 0; limit = last - sb->p < SWAP_META_PAGES ? last - sb->p : SWAP_META_PAGES; for (i = start; i < limit; i++) { if (sb->d[i] == SWAPBLK_NONE) continue; swp_pager_update_freerange(&s_free, &n_free, sb->d[i]); sb->d[i] = SWAPBLK_NONE; } pindex = sb->p + SWAP_META_PAGES; if (swp_pager_swblk_empty(sb, 0, start) && swp_pager_swblk_empty(sb, limit, SWAP_META_PAGES)) { SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p); uma_zfree(swblk_zone, sb); } } swp_pager_freeswapspace(s_free, n_free); } /* * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object * * This routine locates and destroys all swap metadata associated with * an object. */ static void swp_pager_meta_free_all(vm_object_t object) { struct swblk *sb; daddr_t n_free, s_free; vm_pindex_t pindex; int i; VM_OBJECT_ASSERT_WLOCKED(object); if (object->type != OBJT_SWAP) return; swp_pager_init_freerange(&s_free, &n_free); for (pindex = 0; (sb = SWAP_PCTRIE_LOOKUP_GE( &object->un_pager.swp.swp_blks, pindex)) != NULL;) { pindex = sb->p + SWAP_META_PAGES; for (i = 0; i < SWAP_META_PAGES; i++) { if (sb->d[i] == SWAPBLK_NONE) continue; swp_pager_update_freerange(&s_free, &n_free, sb->d[i]); } SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p); uma_zfree(swblk_zone, sb); } swp_pager_freeswapspace(s_free, n_free); } /* * SWP_PAGER_METACTL() - misc control of swap meta data. * * This routine is capable of looking up, or removing swapblk * assignments in the swap meta data. It returns the swapblk being * looked-up, popped, or SWAPBLK_NONE if the block was invalid. * * When acting on a busy resident page and paging is in progress, we * have to wait until paging is complete but otherwise can act on the * busy page. * * SWM_POP remove from meta data but do not free it */ static daddr_t swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags) { struct swblk *sb; daddr_t r1; if ((flags & SWM_POP) != 0) VM_OBJECT_ASSERT_WLOCKED(object); else VM_OBJECT_ASSERT_LOCKED(object); /* * The meta data only exists if the object is OBJT_SWAP * and even then might not be allocated yet. */ if (object->type != OBJT_SWAP) return (SWAPBLK_NONE); sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rounddown(pindex, SWAP_META_PAGES)); if (sb == NULL) return (SWAPBLK_NONE); r1 = sb->d[pindex % SWAP_META_PAGES]; if (r1 == SWAPBLK_NONE) return (SWAPBLK_NONE); if ((flags & SWM_POP) != 0) { sb->d[pindex % SWAP_META_PAGES] = SWAPBLK_NONE; if (swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) { SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, rounddown(pindex, SWAP_META_PAGES)); uma_zfree(swblk_zone, sb); } } return (r1); } /* * Returns the least page index which is greater than or equal to the * parameter pindex and for which there is a swap block allocated. * Returns object's size if the object's type is not swap or if there * are no allocated swap blocks for the object after the requested * pindex. */ vm_pindex_t swap_pager_find_least(vm_object_t object, vm_pindex_t pindex) { struct swblk *sb; int i; VM_OBJECT_ASSERT_LOCKED(object); if (object->type != OBJT_SWAP) return (object->size); sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks, rounddown(pindex, SWAP_META_PAGES)); if (sb == NULL) return (object->size); if (sb->p < pindex) { for (i = pindex % SWAP_META_PAGES; i < SWAP_META_PAGES; i++) { if (sb->d[i] != SWAPBLK_NONE) return (sb->p + i); } sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks, roundup(pindex, SWAP_META_PAGES)); if (sb == NULL) return (object->size); } for (i = 0; i < SWAP_META_PAGES; i++) { if (sb->d[i] != SWAPBLK_NONE) return (sb->p + i); } /* * We get here if a swblk is present in the trie but it * doesn't map any blocks. */ MPASS(0); return (object->size); } /* * System call swapon(name) enables swapping on device name, * which must be in the swdevsw. Return EBUSY * if already swapping on this device. */ #ifndef _SYS_SYSPROTO_H_ struct swapon_args { char *name; }; #endif /* * MPSAFE */ /* ARGSUSED */ int sys_swapon(struct thread *td, struct swapon_args *uap) { struct vattr attr; struct vnode *vp; struct nameidata nd; int error; error = priv_check(td, PRIV_SWAPON); if (error) return (error); sx_xlock(&swdev_syscall_lock); /* * Swap metadata may not fit in the KVM if we have physical * memory of >1GB. */ if (swblk_zone == NULL) { error = ENOMEM; goto done; } NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name, td); error = namei(&nd); if (error) goto done; NDFREE(&nd, NDF_ONLY_PNBUF); vp = nd.ni_vp; if (vn_isdisk(vp, &error)) { error = swapongeom(vp); } else if (vp->v_type == VREG && (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) { /* * Allow direct swapping to NFS regular files in the same * way that nfs_mountroot() sets up diskless swapping. */ error = swaponvp(td, vp, attr.va_size / DEV_BSIZE); } if (error) vrele(vp); done: sx_xunlock(&swdev_syscall_lock); return (error); } /* * Check that the total amount of swap currently configured does not * exceed half the theoretical maximum. If it does, print a warning * message. */ static void swapon_check_swzone(void) { unsigned long maxpages, npages; npages = swap_total; /* absolute maximum we can handle assuming 100% efficiency */ maxpages = uma_zone_get_max(swblk_zone) * SWAP_META_PAGES; /* recommend using no more than half that amount */ if (npages > maxpages / 2) { printf("warning: total configured swap (%lu pages) " "exceeds maximum recommended amount (%lu pages).\n", npages, maxpages / 2); printf("warning: increase kern.maxswzone " "or reduce amount of swap.\n"); } } static void swaponsomething(struct vnode *vp, void *id, u_long nblks, sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags) { struct swdevt *sp, *tsp; swblk_t dvbase; u_long mblocks; /* * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks. * First chop nblks off to page-align it, then convert. * * sw->sw_nblks is in page-sized chunks now too. */ nblks &= ~(ctodb(1) - 1); nblks = dbtoc(nblks); /* * If we go beyond this, we get overflows in the radix * tree bitmap code. */ mblocks = 0x40000000 / BLIST_META_RADIX; if (nblks > mblocks) { printf( "WARNING: reducing swap size to maximum of %luMB per unit\n", mblocks / 1024 / 1024 * PAGE_SIZE); nblks = mblocks; } sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO); sp->sw_vp = vp; sp->sw_id = id; sp->sw_dev = dev; sp->sw_nblks = nblks; sp->sw_used = 0; sp->sw_strategy = strategy; sp->sw_close = close; sp->sw_flags = flags; sp->sw_blist = blist_create(nblks, M_WAITOK); /* * Do not free the first blocks in order to avoid overwriting * any bsd label at the front of the partition */ blist_free(sp->sw_blist, howmany(BBSIZE, PAGE_SIZE), nblks - howmany(BBSIZE, PAGE_SIZE)); dvbase = 0; mtx_lock(&sw_dev_mtx); TAILQ_FOREACH(tsp, &swtailq, sw_list) { if (tsp->sw_end >= dvbase) { /* * We put one uncovered page between the devices * in order to definitively prevent any cross-device * I/O requests */ dvbase = tsp->sw_end + 1; } } sp->sw_first = dvbase; sp->sw_end = dvbase + nblks; TAILQ_INSERT_TAIL(&swtailq, sp, sw_list); nswapdev++; swap_pager_avail += nblks - howmany(BBSIZE, PAGE_SIZE); swap_total += nblks; swapon_check_swzone(); swp_sizecheck(); mtx_unlock(&sw_dev_mtx); EVENTHANDLER_INVOKE(swapon, sp); } /* * SYSCALL: swapoff(devname) * * Disable swapping on the given device. * * XXX: Badly designed system call: it should use a device index * rather than filename as specification. We keep sw_vp around * only to make this work. */ #ifndef _SYS_SYSPROTO_H_ struct swapoff_args { char *name; }; #endif /* * MPSAFE */ /* ARGSUSED */ int sys_swapoff(struct thread *td, struct swapoff_args *uap) { struct vnode *vp; struct nameidata nd; struct swdevt *sp; int error; error = priv_check(td, PRIV_SWAPOFF); if (error) return (error); sx_xlock(&swdev_syscall_lock); NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name, td); error = namei(&nd); if (error) goto done; NDFREE(&nd, NDF_ONLY_PNBUF); vp = nd.ni_vp; mtx_lock(&sw_dev_mtx); TAILQ_FOREACH(sp, &swtailq, sw_list) { if (sp->sw_vp == vp) break; } mtx_unlock(&sw_dev_mtx); if (sp == NULL) { error = EINVAL; goto done; } error = swapoff_one(sp, td->td_ucred); done: sx_xunlock(&swdev_syscall_lock); return (error); } static int swapoff_one(struct swdevt *sp, struct ucred *cred) { u_long nblks; #ifdef MAC int error; #endif sx_assert(&swdev_syscall_lock, SA_XLOCKED); #ifdef MAC (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY); error = mac_system_check_swapoff(cred, sp->sw_vp); (void) VOP_UNLOCK(sp->sw_vp, 0); if (error != 0) return (error); #endif nblks = sp->sw_nblks; /* * We can turn off this swap device safely only if the * available virtual memory in the system will fit the amount * of data we will have to page back in, plus an epsilon so * the system doesn't become critically low on swap space. */ if (vm_free_count() + swap_pager_avail < nblks + nswap_lowat) return (ENOMEM); /* * Prevent further allocations on this device. */ mtx_lock(&sw_dev_mtx); sp->sw_flags |= SW_CLOSING; swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks); swap_total -= nblks; mtx_unlock(&sw_dev_mtx); /* * Page in the contents of the device and close it. */ swap_pager_swapoff(sp); sp->sw_close(curthread, sp); mtx_lock(&sw_dev_mtx); sp->sw_id = NULL; TAILQ_REMOVE(&swtailq, sp, sw_list); nswapdev--; if (nswapdev == 0) { swap_pager_full = 2; swap_pager_almost_full = 1; } if (swdevhd == sp) swdevhd = NULL; mtx_unlock(&sw_dev_mtx); blist_destroy(sp->sw_blist); free(sp, M_VMPGDATA); return (0); } void swapoff_all(void) { struct swdevt *sp, *spt; const char *devname; int error; sx_xlock(&swdev_syscall_lock); mtx_lock(&sw_dev_mtx); TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) { mtx_unlock(&sw_dev_mtx); if (vn_isdisk(sp->sw_vp, NULL)) devname = devtoname(sp->sw_vp->v_rdev); else devname = "[file]"; error = swapoff_one(sp, thread0.td_ucred); if (error != 0) { printf("Cannot remove swap device %s (error=%d), " "skipping.\n", devname, error); } else if (bootverbose) { printf("Swap device %s removed.\n", devname); } mtx_lock(&sw_dev_mtx); } mtx_unlock(&sw_dev_mtx); sx_xunlock(&swdev_syscall_lock); } void swap_pager_status(int *total, int *used) { struct swdevt *sp; *total = 0; *used = 0; mtx_lock(&sw_dev_mtx); TAILQ_FOREACH(sp, &swtailq, sw_list) { *total += sp->sw_nblks; *used += sp->sw_used; } mtx_unlock(&sw_dev_mtx); } int swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len) { struct swdevt *sp; const char *tmp_devname; int error, n; n = 0; error = ENOENT; mtx_lock(&sw_dev_mtx); TAILQ_FOREACH(sp, &swtailq, sw_list) { if (n != name) { n++; continue; } xs->xsw_version = XSWDEV_VERSION; xs->xsw_dev = sp->sw_dev; xs->xsw_flags = sp->sw_flags; xs->xsw_nblks = sp->sw_nblks; xs->xsw_used = sp->sw_used; if (devname != NULL) { if (vn_isdisk(sp->sw_vp, NULL)) tmp_devname = devtoname(sp->sw_vp->v_rdev); else tmp_devname = "[file]"; strncpy(devname, tmp_devname, len); } error = 0; break; } mtx_unlock(&sw_dev_mtx); return (error); } #if defined(COMPAT_FREEBSD11) #define XSWDEV_VERSION_11 1 struct xswdev11 { u_int xsw_version; uint32_t xsw_dev; int xsw_flags; int xsw_nblks; int xsw_used; }; #endif #if defined(__amd64__) && defined(COMPAT_FREEBSD32) struct xswdev32 { u_int xsw_version; u_int xsw_dev1, xsw_dev2; int xsw_flags; int xsw_nblks; int xsw_used; }; #endif static int sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS) { struct xswdev xs; #if defined(__amd64__) && defined(COMPAT_FREEBSD32) struct xswdev32 xs32; #endif #if defined(COMPAT_FREEBSD11) struct xswdev11 xs11; #endif int error; if (arg2 != 1) /* name length */ return (EINVAL); error = swap_dev_info(*(int *)arg1, &xs, NULL, 0); if (error != 0) return (error); #if defined(__amd64__) && defined(COMPAT_FREEBSD32) if (req->oldlen == sizeof(xs32)) { xs32.xsw_version = XSWDEV_VERSION; xs32.xsw_dev1 = xs.xsw_dev; xs32.xsw_dev2 = xs.xsw_dev >> 32; xs32.xsw_flags = xs.xsw_flags; xs32.xsw_nblks = xs.xsw_nblks; xs32.xsw_used = xs.xsw_used; error = SYSCTL_OUT(req, &xs32, sizeof(xs32)); return (error); } #endif #if defined(COMPAT_FREEBSD11) if (req->oldlen == sizeof(xs11)) { xs11.xsw_version = XSWDEV_VERSION_11; xs11.xsw_dev = xs.xsw_dev; /* truncation */ xs11.xsw_flags = xs.xsw_flags; xs11.xsw_nblks = xs.xsw_nblks; xs11.xsw_used = xs.xsw_used; error = SYSCTL_OUT(req, &xs11, sizeof(xs11)); return (error); } #endif error = SYSCTL_OUT(req, &xs, sizeof(xs)); return (error); } SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0, "Number of swap devices"); SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE, sysctl_vm_swap_info, "Swap statistics by device"); /* * Count the approximate swap usage in pages for a vmspace. The * shadowed or not yet copied on write swap blocks are not accounted. * The map must be locked. */ long vmspace_swap_count(struct vmspace *vmspace) { vm_map_t map; vm_map_entry_t cur; vm_object_t object; struct swblk *sb; vm_pindex_t e, pi; long count; int i; map = &vmspace->vm_map; count = 0; VM_MAP_ENTRY_FOREACH(cur, map) { if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) continue; object = cur->object.vm_object; if (object == NULL || object->type != OBJT_SWAP) continue; VM_OBJECT_RLOCK(object); if (object->type != OBJT_SWAP) goto unlock; pi = OFF_TO_IDX(cur->offset); e = pi + OFF_TO_IDX(cur->end - cur->start); for (;; pi = sb->p + SWAP_META_PAGES) { sb = SWAP_PCTRIE_LOOKUP_GE( &object->un_pager.swp.swp_blks, pi); if (sb == NULL || sb->p >= e) break; for (i = 0; i < SWAP_META_PAGES; i++) { if (sb->p + i < e && sb->d[i] != SWAPBLK_NONE) count++; } } unlock: VM_OBJECT_RUNLOCK(object); } return (count); } /* * GEOM backend * * Swapping onto disk devices. * */ static g_orphan_t swapgeom_orphan; static struct g_class g_swap_class = { .name = "SWAP", .version = G_VERSION, .orphan = swapgeom_orphan, }; DECLARE_GEOM_CLASS(g_swap_class, g_class); static void swapgeom_close_ev(void *arg, int flags) { struct g_consumer *cp; cp = arg; g_access(cp, -1, -1, 0); g_detach(cp); g_destroy_consumer(cp); } /* * Add a reference to the g_consumer for an inflight transaction. */ static void swapgeom_acquire(struct g_consumer *cp) { mtx_assert(&sw_dev_mtx, MA_OWNED); cp->index++; } /* * Remove a reference from the g_consumer. Post a close event if all * references go away, since the function might be called from the * biodone context. */ static void swapgeom_release(struct g_consumer *cp, struct swdevt *sp) { mtx_assert(&sw_dev_mtx, MA_OWNED); cp->index--; if (cp->index == 0) { if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0) sp->sw_id = NULL; } } static void swapgeom_done(struct bio *bp2) { struct swdevt *sp; struct buf *bp; struct g_consumer *cp; bp = bp2->bio_caller2; cp = bp2->bio_from; bp->b_ioflags = bp2->bio_flags; if (bp2->bio_error) bp->b_ioflags |= BIO_ERROR; bp->b_resid = bp->b_bcount - bp2->bio_completed; bp->b_error = bp2->bio_error; bp->b_caller1 = NULL; bufdone(bp); sp = bp2->bio_caller1; mtx_lock(&sw_dev_mtx); swapgeom_release(cp, sp); mtx_unlock(&sw_dev_mtx); g_destroy_bio(bp2); } static void swapgeom_strategy(struct buf *bp, struct swdevt *sp) { struct bio *bio; struct g_consumer *cp; mtx_lock(&sw_dev_mtx); cp = sp->sw_id; if (cp == NULL) { mtx_unlock(&sw_dev_mtx); bp->b_error = ENXIO; bp->b_ioflags |= BIO_ERROR; bufdone(bp); return; } swapgeom_acquire(cp); mtx_unlock(&sw_dev_mtx); if (bp->b_iocmd == BIO_WRITE) bio = g_new_bio(); else bio = g_alloc_bio(); if (bio == NULL) { mtx_lock(&sw_dev_mtx); swapgeom_release(cp, sp); mtx_unlock(&sw_dev_mtx); bp->b_error = ENOMEM; bp->b_ioflags |= BIO_ERROR; printf("swap_pager: cannot allocate bio\n"); bufdone(bp); return; } bp->b_caller1 = bio; bio->bio_caller1 = sp; bio->bio_caller2 = bp; bio->bio_cmd = bp->b_iocmd; bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE; bio->bio_length = bp->b_bcount; bio->bio_done = swapgeom_done; if (!buf_mapped(bp)) { bio->bio_ma = bp->b_pages; bio->bio_data = unmapped_buf; bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; bio->bio_ma_n = bp->b_npages; bio->bio_flags |= BIO_UNMAPPED; } else { bio->bio_data = bp->b_data; bio->bio_ma = NULL; } g_io_request(bio, cp); return; } static void swapgeom_orphan(struct g_consumer *cp) { struct swdevt *sp; int destroy; mtx_lock(&sw_dev_mtx); TAILQ_FOREACH(sp, &swtailq, sw_list) { if (sp->sw_id == cp) { sp->sw_flags |= SW_CLOSING; break; } } /* * Drop reference we were created with. Do directly since we're in a * special context where we don't have to queue the call to * swapgeom_close_ev(). */ cp->index--; destroy = ((sp != NULL) && (cp->index == 0)); if (destroy) sp->sw_id = NULL; mtx_unlock(&sw_dev_mtx); if (destroy) swapgeom_close_ev(cp, 0); } static void swapgeom_close(struct thread *td, struct swdevt *sw) { struct g_consumer *cp; mtx_lock(&sw_dev_mtx); cp = sw->sw_id; sw->sw_id = NULL; mtx_unlock(&sw_dev_mtx); /* * swapgeom_close() may be called from the biodone context, * where we cannot perform topology changes. Delegate the * work to the events thread. */ if (cp != NULL) g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL); } static int swapongeom_locked(struct cdev *dev, struct vnode *vp) { struct g_provider *pp; struct g_consumer *cp; static struct g_geom *gp; struct swdevt *sp; u_long nblks; int error; pp = g_dev_getprovider(dev); if (pp == NULL) return (ENODEV); mtx_lock(&sw_dev_mtx); TAILQ_FOREACH(sp, &swtailq, sw_list) { cp = sp->sw_id; if (cp != NULL && cp->provider == pp) { mtx_unlock(&sw_dev_mtx); return (EBUSY); } } mtx_unlock(&sw_dev_mtx); if (gp == NULL) gp = g_new_geomf(&g_swap_class, "swap"); cp = g_new_consumer(gp); cp->index = 1; /* Number of active I/Os, plus one for being active. */ cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE; g_attach(cp, pp); /* * XXX: Every time you think you can improve the margin for * footshooting, somebody depends on the ability to do so: * savecore(8) wants to write to our swapdev so we cannot * set an exclusive count :-( */ error = g_access(cp, 1, 1, 0); if (error != 0) { g_detach(cp); g_destroy_consumer(cp); return (error); } nblks = pp->mediasize / DEV_BSIZE; swaponsomething(vp, cp, nblks, swapgeom_strategy, swapgeom_close, dev2udev(dev), (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0); return (0); } static int swapongeom(struct vnode *vp) { int error; vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); if (vp->v_type != VCHR || (vp->v_iflag & VI_DOOMED) != 0) { error = ENOENT; } else { g_topology_lock(); error = swapongeom_locked(vp->v_rdev, vp); g_topology_unlock(); } VOP_UNLOCK(vp, 0); return (error); } /* * VNODE backend * * This is used mainly for network filesystem (read: probably only tested * with NFS) swapfiles. * */ static void swapdev_strategy(struct buf *bp, struct swdevt *sp) { struct vnode *vp2; bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first); vp2 = sp->sw_id; vhold(vp2); if (bp->b_iocmd == BIO_WRITE) { if (bp->b_bufobj) bufobj_wdrop(bp->b_bufobj); bufobj_wref(&vp2->v_bufobj); } if (bp->b_bufobj != &vp2->v_bufobj) bp->b_bufobj = &vp2->v_bufobj; bp->b_vp = vp2; bp->b_iooffset = dbtob(bp->b_blkno); bstrategy(bp); return; } static void swapdev_close(struct thread *td, struct swdevt *sp) { VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td); vrele(sp->sw_vp); } static int swaponvp(struct thread *td, struct vnode *vp, u_long nblks) { struct swdevt *sp; int error; if (nblks == 0) return (ENXIO); mtx_lock(&sw_dev_mtx); TAILQ_FOREACH(sp, &swtailq, sw_list) { if (sp->sw_id == vp) { mtx_unlock(&sw_dev_mtx); return (EBUSY); } } mtx_unlock(&sw_dev_mtx); (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); #ifdef MAC error = mac_system_check_swapon(td->td_ucred, vp); if (error == 0) #endif error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL); (void) VOP_UNLOCK(vp, 0); if (error) return (error); swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close, NODEV, 0); return (0); } static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS) { int error, new, n; new = nsw_wcount_async_max; error = sysctl_handle_int(oidp, &new, 0, req); if (error != 0 || req->newptr == NULL) return (error); if (new > nswbuf / 2 || new < 1) return (EINVAL); mtx_lock(&swbuf_mtx); while (nsw_wcount_async_max != new) { /* * Adjust difference. If the current async count is too low, * we will need to sqeeze our update slowly in. Sleep with a * higher priority than getpbuf() to finish faster. */ n = new - nsw_wcount_async_max; if (nsw_wcount_async + n >= 0) { nsw_wcount_async += n; nsw_wcount_async_max += n; wakeup(&nsw_wcount_async); } else { nsw_wcount_async_max -= nsw_wcount_async; nsw_wcount_async = 0; msleep(&nsw_wcount_async, &swbuf_mtx, PSWP, "swpsysctl", 0); } } mtx_unlock(&swbuf_mtx); return (0); } static void swap_pager_update_writecount(vm_object_t object, vm_offset_t start, vm_offset_t end) { VM_OBJECT_WLOCK(object); KASSERT((object->flags & OBJ_NOSPLIT) != 0, ("Splittable object with writecount")); object->un_pager.swp.writemappings += (vm_ooffset_t)end - start; VM_OBJECT_WUNLOCK(object); } static void swap_pager_release_writecount(vm_object_t object, vm_offset_t start, vm_offset_t end) { VM_OBJECT_WLOCK(object); KASSERT((object->flags & OBJ_NOSPLIT) != 0, ("Splittable object with writecount")); object->un_pager.swp.writemappings -= (vm_ooffset_t)end - start; VM_OBJECT_WUNLOCK(object); } Index: head/sys/vm/vm_fault.c =================================================================== --- head/sys/vm/vm_fault.c (revision 353538) +++ head/sys/vm/vm_fault.c (revision 353539) @@ -1,1905 +1,1906 @@ /*- * SPDX-License-Identifier: (BSD-4-Clause AND MIT-CMU) * * Copyright (c) 1991, 1993 * The Regents of the University of California. All rights reserved. * Copyright (c) 1994 John S. Dyson * All rights reserved. * Copyright (c) 1994 David Greenman * All rights reserved. * * * This code is derived from software contributed to Berkeley by * The Mach Operating System project at Carnegie-Mellon University. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. All advertising materials mentioning features or use of this software * must display the following acknowledgement: * This product includes software developed by the University of * California, Berkeley and its contributors. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)vm_fault.c 8.4 (Berkeley) 1/12/94 * * * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Authors: Avadis Tevanian, Jr., Michael Wayne Young * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. */ /* * Page fault handling module. */ #include __FBSDID("$FreeBSD$"); #include "opt_ktrace.h" #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef KTRACE #include #endif #include #include #include #include #include #include #include #include #include #include #include #define PFBAK 4 #define PFFOR 4 #define VM_FAULT_READ_DEFAULT (1 + VM_FAULT_READ_AHEAD_INIT) #define VM_FAULT_READ_MAX (1 + VM_FAULT_READ_AHEAD_MAX) #define VM_FAULT_DONTNEED_MIN 1048576 struct faultstate { vm_page_t m; vm_object_t object; vm_pindex_t pindex; vm_page_t first_m; vm_object_t first_object; vm_pindex_t first_pindex; vm_map_t map; vm_map_entry_t entry; int map_generation; bool lookup_still_valid; struct vnode *vp; }; static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead); static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra, int backward, int forward, bool obj_locked); static int vm_pfault_oom_attempts = 3; SYSCTL_INT(_vm, OID_AUTO, pfault_oom_attempts, CTLFLAG_RWTUN, &vm_pfault_oom_attempts, 0, "Number of page allocation attempts in page fault handler before it " "triggers OOM handling"); static int vm_pfault_oom_wait = 10; SYSCTL_INT(_vm, OID_AUTO, pfault_oom_wait, CTLFLAG_RWTUN, &vm_pfault_oom_wait, 0, "Number of seconds to wait for free pages before retrying " "the page fault handler"); static inline void release_page(struct faultstate *fs) { vm_page_xunbusy(fs->m); vm_page_lock(fs->m); vm_page_deactivate(fs->m); vm_page_unlock(fs->m); fs->m = NULL; } static inline void unlock_map(struct faultstate *fs) { if (fs->lookup_still_valid) { vm_map_lookup_done(fs->map, fs->entry); fs->lookup_still_valid = false; } } static void unlock_vp(struct faultstate *fs) { if (fs->vp != NULL) { vput(fs->vp); fs->vp = NULL; } } static void unlock_and_deallocate(struct faultstate *fs) { vm_object_pip_wakeup(fs->object); VM_OBJECT_WUNLOCK(fs->object); if (fs->object != fs->first_object) { VM_OBJECT_WLOCK(fs->first_object); vm_page_free(fs->first_m); vm_object_pip_wakeup(fs->first_object); VM_OBJECT_WUNLOCK(fs->first_object); fs->first_m = NULL; } vm_object_deallocate(fs->first_object); unlock_map(fs); unlock_vp(fs); } static void vm_fault_dirty(vm_map_entry_t entry, vm_page_t m, vm_prot_t prot, vm_prot_t fault_type, int fault_flags, bool set_wd) { bool need_dirty; if (((prot & VM_PROT_WRITE) == 0 && (fault_flags & VM_FAULT_DIRTY) == 0) || (m->oflags & VPO_UNMANAGED) != 0) return; VM_OBJECT_ASSERT_LOCKED(m->object); + VM_PAGE_OBJECT_BUSY_ASSERT(m); need_dirty = ((fault_type & VM_PROT_WRITE) != 0 && (fault_flags & VM_FAULT_WIRE) == 0) || (fault_flags & VM_FAULT_DIRTY) != 0; if (set_wd) vm_object_set_writeable_dirty(m->object); else /* * If two callers of vm_fault_dirty() with set_wd == * FALSE, one for the map entry with MAP_ENTRY_NOSYNC * flag set, other with flag clear, race, it is * possible for the no-NOSYNC thread to see m->dirty * != 0 and not clear VPO_NOSYNC. Take vm_page lock * around manipulation of VPO_NOSYNC and * vm_page_dirty() call, to avoid the race and keep * m->oflags consistent. */ vm_page_lock(m); /* * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC * if the page is already dirty to prevent data written with * the expectation of being synced from not being synced. * Likewise if this entry does not request NOSYNC then make * sure the page isn't marked NOSYNC. Applications sharing * data should use the same flags to avoid ping ponging. */ if ((entry->eflags & MAP_ENTRY_NOSYNC) != 0) { if (m->dirty == 0) { m->oflags |= VPO_NOSYNC; } } else { m->oflags &= ~VPO_NOSYNC; } /* * If the fault is a write, we know that this page is being * written NOW so dirty it explicitly to save on * pmap_is_modified() calls later. * * Also, since the page is now dirty, we can possibly tell * the pager to release any swap backing the page. Calling * the pager requires a write lock on the object. */ if (need_dirty) vm_page_dirty(m); if (!set_wd) vm_page_unlock(m); else if (need_dirty) vm_pager_page_unswapped(m); } /* * Unlocks fs.first_object and fs.map on success. */ static int vm_fault_soft_fast(struct faultstate *fs, vm_offset_t vaddr, vm_prot_t prot, int fault_type, int fault_flags, boolean_t wired, vm_page_t *m_hold) { vm_page_t m, m_map; #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \ __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \ VM_NRESERVLEVEL > 0 vm_page_t m_super; int flags; #endif int psind, rv; MPASS(fs->vp == NULL); vm_object_busy(fs->first_object); m = vm_page_lookup(fs->first_object, fs->first_pindex); /* A busy page can be mapped for read|execute access. */ if (m == NULL || ((prot & VM_PROT_WRITE) != 0 && - vm_page_busied(m)) || m->valid != VM_PAGE_BITS_ALL) { + vm_page_busied(m)) || !vm_page_all_valid(m)) { rv = KERN_FAILURE; goto out; } m_map = m; psind = 0; #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \ __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \ VM_NRESERVLEVEL > 0 if ((m->flags & PG_FICTITIOUS) == 0 && (m_super = vm_reserv_to_superpage(m)) != NULL && rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start && roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end && (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) & (pagesizes[m_super->psind] - 1)) && !wired && pmap_ps_enabled(fs->map->pmap)) { flags = PS_ALL_VALID; if ((prot & VM_PROT_WRITE) != 0) { /* * Create a superpage mapping allowing write access * only if none of the constituent pages are busy and * all of them are already dirty (except possibly for * the page that was faulted on). */ flags |= PS_NONE_BUSY; if ((fs->first_object->flags & OBJ_UNMANAGED) == 0) flags |= PS_ALL_DIRTY; } if (vm_page_ps_test(m_super, flags, m)) { m_map = m_super; psind = m_super->psind; vaddr = rounddown2(vaddr, pagesizes[psind]); /* Preset the modified bit for dirty superpages. */ if ((flags & PS_ALL_DIRTY) != 0) fault_type |= VM_PROT_WRITE; } } #endif rv = pmap_enter(fs->map->pmap, vaddr, m_map, prot, fault_type | PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED : 0), psind); if (rv != KERN_SUCCESS) goto out; if (m_hold != NULL) { *m_hold = m; vm_page_wire(m); } vm_fault_dirty(fs->entry, m, prot, fault_type, fault_flags, false); if (psind == 0 && !wired) vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true); VM_OBJECT_RUNLOCK(fs->first_object); vm_map_lookup_done(fs->map, fs->entry); curthread->td_ru.ru_minflt++; out: vm_object_unbusy(fs->first_object); return (rv); } static void vm_fault_restore_map_lock(struct faultstate *fs) { VM_OBJECT_ASSERT_WLOCKED(fs->first_object); MPASS(REFCOUNT_COUNT(fs->first_object->paging_in_progress) > 0); if (!vm_map_trylock_read(fs->map)) { VM_OBJECT_WUNLOCK(fs->first_object); vm_map_lock_read(fs->map); VM_OBJECT_WLOCK(fs->first_object); } fs->lookup_still_valid = true; } static void vm_fault_populate_check_page(vm_page_t m) { /* * Check each page to ensure that the pager is obeying the * interface: the page must be installed in the object, fully * valid, and exclusively busied. */ MPASS(m != NULL); - MPASS(m->valid == VM_PAGE_BITS_ALL); + MPASS(vm_page_all_valid(m)); MPASS(vm_page_xbusied(m)); } static void vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first, vm_pindex_t last) { vm_page_t m; vm_pindex_t pidx; VM_OBJECT_ASSERT_WLOCKED(object); MPASS(first <= last); for (pidx = first, m = vm_page_lookup(object, pidx); pidx <= last; pidx++, m = vm_page_next(m)) { vm_fault_populate_check_page(m); vm_page_lock(m); vm_page_deactivate(m); vm_page_unlock(m); vm_page_xunbusy(m); } } static int vm_fault_populate(struct faultstate *fs, vm_prot_t prot, int fault_type, int fault_flags, boolean_t wired, vm_page_t *m_hold) { struct mtx *m_mtx; vm_offset_t vaddr; vm_page_t m; vm_pindex_t map_first, map_last, pager_first, pager_last, pidx; int i, npages, psind, rv; MPASS(fs->object == fs->first_object); VM_OBJECT_ASSERT_WLOCKED(fs->first_object); MPASS(REFCOUNT_COUNT(fs->first_object->paging_in_progress) > 0); MPASS(fs->first_object->backing_object == NULL); MPASS(fs->lookup_still_valid); pager_first = OFF_TO_IDX(fs->entry->offset); pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1; unlock_map(fs); unlock_vp(fs); /* * Call the pager (driver) populate() method. * * There is no guarantee that the method will be called again * if the current fault is for read, and a future fault is * for write. Report the entry's maximum allowed protection * to the driver. */ rv = vm_pager_populate(fs->first_object, fs->first_pindex, fault_type, fs->entry->max_protection, &pager_first, &pager_last); VM_OBJECT_ASSERT_WLOCKED(fs->first_object); if (rv == VM_PAGER_BAD) { /* * VM_PAGER_BAD is the backdoor for a pager to request * normal fault handling. */ vm_fault_restore_map_lock(fs); if (fs->map->timestamp != fs->map_generation) return (KERN_RESOURCE_SHORTAGE); /* RetryFault */ return (KERN_NOT_RECEIVER); } if (rv != VM_PAGER_OK) return (KERN_FAILURE); /* AKA SIGSEGV */ /* Ensure that the driver is obeying the interface. */ MPASS(pager_first <= pager_last); MPASS(fs->first_pindex <= pager_last); MPASS(fs->first_pindex >= pager_first); MPASS(pager_last < fs->first_object->size); vm_fault_restore_map_lock(fs); if (fs->map->timestamp != fs->map_generation) { vm_fault_populate_cleanup(fs->first_object, pager_first, pager_last); return (KERN_RESOURCE_SHORTAGE); /* RetryFault */ } /* * The map is unchanged after our last unlock. Process the fault. * * The range [pager_first, pager_last] that is given to the * pager is only a hint. The pager may populate any range * within the object that includes the requested page index. * In case the pager expanded the range, clip it to fit into * the map entry. */ map_first = OFF_TO_IDX(fs->entry->offset); if (map_first > pager_first) { vm_fault_populate_cleanup(fs->first_object, pager_first, map_first - 1); pager_first = map_first; } map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1; if (map_last < pager_last) { vm_fault_populate_cleanup(fs->first_object, map_last + 1, pager_last); pager_last = map_last; } for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx); pidx <= pager_last; pidx += npages, m = vm_page_next(&m[npages - 1])) { vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset; #if defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \ __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv) psind = m->psind; if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 || pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last || !pmap_ps_enabled(fs->map->pmap) || wired)) psind = 0; #else psind = 0; #endif npages = atop(pagesizes[psind]); for (i = 0; i < npages; i++) { vm_fault_populate_check_page(&m[i]); vm_fault_dirty(fs->entry, &m[i], prot, fault_type, fault_flags, true); } VM_OBJECT_WUNLOCK(fs->first_object); rv = pmap_enter(fs->map->pmap, vaddr, m, prot, fault_type | (wired ? PMAP_ENTER_WIRED : 0), psind); #if defined(__amd64__) if (psind > 0 && rv == KERN_FAILURE) { for (i = 0; i < npages; i++) { rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i), &m[i], prot, fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0); MPASS(rv == KERN_SUCCESS); } } #else MPASS(rv == KERN_SUCCESS); #endif VM_OBJECT_WLOCK(fs->first_object); m_mtx = NULL; for (i = 0; i < npages; i++) { if ((fault_flags & VM_FAULT_WIRE) != 0) { vm_page_wire(&m[i]); } else { vm_page_change_lock(&m[i], &m_mtx); vm_page_activate(&m[i]); } if (m_hold != NULL && m[i].pindex == fs->first_pindex) { *m_hold = &m[i]; vm_page_wire(&m[i]); } vm_page_xunbusy(&m[i]); } if (m_mtx != NULL) mtx_unlock(m_mtx); } curthread->td_ru.ru_majflt++; return (KERN_SUCCESS); } static int prot_fault_translation; SYSCTL_INT(_machdep, OID_AUTO, prot_fault_translation, CTLFLAG_RWTUN, &prot_fault_translation, 0, "Control signal to deliver on protection fault"); /* compat definition to keep common code for signal translation */ #define UCODE_PAGEFLT 12 #ifdef T_PAGEFLT _Static_assert(UCODE_PAGEFLT == T_PAGEFLT, "T_PAGEFLT"); #endif /* * vm_fault_trap: * * Handle a page fault occurring at the given address, * requiring the given permissions, in the map specified. * If successful, the page is inserted into the * associated physical map. * * NOTE: the given address should be truncated to the * proper page address. * * KERN_SUCCESS is returned if the page fault is handled; otherwise, * a standard error specifying why the fault is fatal is returned. * * The map in question must be referenced, and remains so. * Caller may hold no locks. */ int vm_fault_trap(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags, int *signo, int *ucode) { int result; MPASS(signo == NULL || ucode != NULL); #ifdef KTRACE if (map != kernel_map && KTRPOINT(curthread, KTR_FAULT)) ktrfault(vaddr, fault_type); #endif result = vm_fault(map, trunc_page(vaddr), fault_type, fault_flags, NULL); KASSERT(result == KERN_SUCCESS || result == KERN_FAILURE || result == KERN_INVALID_ADDRESS || result == KERN_RESOURCE_SHORTAGE || result == KERN_PROTECTION_FAILURE || result == KERN_OUT_OF_BOUNDS, ("Unexpected Mach error %d from vm_fault()", result)); #ifdef KTRACE if (map != kernel_map && KTRPOINT(curthread, KTR_FAULTEND)) ktrfaultend(result); #endif if (result != KERN_SUCCESS && signo != NULL) { switch (result) { case KERN_FAILURE: case KERN_INVALID_ADDRESS: *signo = SIGSEGV; *ucode = SEGV_MAPERR; break; case KERN_RESOURCE_SHORTAGE: *signo = SIGBUS; *ucode = BUS_OOMERR; break; case KERN_OUT_OF_BOUNDS: *signo = SIGBUS; *ucode = BUS_OBJERR; break; case KERN_PROTECTION_FAILURE: if (prot_fault_translation == 0) { /* * Autodetect. This check also covers * the images without the ABI-tag ELF * note. */ if (SV_CURPROC_ABI() == SV_ABI_FREEBSD && curproc->p_osrel >= P_OSREL_SIGSEGV) { *signo = SIGSEGV; *ucode = SEGV_ACCERR; } else { *signo = SIGBUS; *ucode = UCODE_PAGEFLT; } } else if (prot_fault_translation == 1) { /* Always compat mode. */ *signo = SIGBUS; *ucode = UCODE_PAGEFLT; } else { /* Always SIGSEGV mode. */ *signo = SIGSEGV; *ucode = SEGV_ACCERR; } break; default: KASSERT(0, ("Unexpected Mach error %d from vm_fault()", result)); break; } } return (result); } int vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags, vm_page_t *m_hold) { struct faultstate fs; struct vnode *vp; struct domainset *dset; vm_object_t next_object, retry_object; vm_offset_t e_end, e_start; vm_pindex_t retry_pindex; vm_prot_t prot, retry_prot; int ahead, alloc_req, behind, cluster_offset, error, era, faultcount; int locked, nera, oom, result, rv; u_char behavior; boolean_t wired; /* Passed by reference. */ bool dead, hardfault, is_first_object_locked; VM_CNT_INC(v_vm_faults); if ((curthread->td_pflags & TDP_NOFAULTING) != 0) return (KERN_PROTECTION_FAILURE); fs.vp = NULL; faultcount = 0; nera = -1; hardfault = false; RetryFault: oom = 0; RetryFault_oom: /* * Find the backing store object and offset into it to begin the * search. */ fs.map = map; result = vm_map_lookup(&fs.map, vaddr, fault_type | VM_PROT_FAULT_LOOKUP, &fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired); if (result != KERN_SUCCESS) { unlock_vp(&fs); return (result); } fs.map_generation = fs.map->timestamp; if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { panic("%s: fault on nofault entry, addr: %#lx", __func__, (u_long)vaddr); } if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION && fs.entry->wiring_thread != curthread) { vm_map_unlock_read(fs.map); vm_map_lock(fs.map); if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) && (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) { unlock_vp(&fs); fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; vm_map_unlock_and_wait(fs.map, 0); } else vm_map_unlock(fs.map); goto RetryFault; } MPASS((fs.entry->eflags & MAP_ENTRY_GUARD) == 0); if (wired) fault_type = prot | (fault_type & VM_PROT_COPY); else KASSERT((fault_flags & VM_FAULT_WIRE) == 0, ("!wired && VM_FAULT_WIRE")); /* * Try to avoid lock contention on the top-level object through * special-case handling of some types of page faults, specifically, * those that are both (1) mapping an existing page from the top- * level object and (2) not having to mark that object as containing * dirty pages. Under these conditions, a read lock on the top-level * object suffices, allowing multiple page faults of a similar type to * run in parallel on the same top-level object. */ if (fs.vp == NULL /* avoid locked vnode leak */ && (fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0 && /* avoid calling vm_object_set_writeable_dirty() */ ((prot & VM_PROT_WRITE) == 0 || (fs.first_object->type != OBJT_VNODE && (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) || (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0)) { VM_OBJECT_RLOCK(fs.first_object); if ((prot & VM_PROT_WRITE) == 0 || (fs.first_object->type != OBJT_VNODE && (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) || (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0) { rv = vm_fault_soft_fast(&fs, vaddr, prot, fault_type, fault_flags, wired, m_hold); if (rv == KERN_SUCCESS) return (rv); } if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) { VM_OBJECT_RUNLOCK(fs.first_object); VM_OBJECT_WLOCK(fs.first_object); } } else { VM_OBJECT_WLOCK(fs.first_object); } /* * Make a reference to this object to prevent its disposal while we * are messing with it. Once we have the reference, the map is free * to be diddled. Since objects reference their shadows (and copies), * they will stay around as well. * * Bump the paging-in-progress count to prevent size changes (e.g. * truncation operations) during I/O. */ vm_object_reference_locked(fs.first_object); vm_object_pip_add(fs.first_object, 1); fs.lookup_still_valid = true; fs.first_m = NULL; /* * Search for the page at object/offset. */ fs.object = fs.first_object; fs.pindex = fs.first_pindex; while (TRUE) { /* * If the object is marked for imminent termination, * we retry here, since the collapse pass has raced * with us. Otherwise, if we see terminally dead * object, return fail. */ if ((fs.object->flags & OBJ_DEAD) != 0) { dead = fs.object->type == OBJT_DEAD; unlock_and_deallocate(&fs); if (dead) return (KERN_PROTECTION_FAILURE); pause("vmf_de", 1); goto RetryFault; } /* * See if page is resident */ fs.m = vm_page_lookup(fs.object, fs.pindex); if (fs.m != NULL) { /* * Wait/Retry if the page is busy. We have to do this * if the page is either exclusive or shared busy * because the vm_pager may be using read busy for * pageouts (and even pageins if it is the vnode * pager), and we could end up trying to pagein and * pageout the same page simultaneously. * * We can theoretically allow the busy case on a read * fault if the page is marked valid, but since such * pages are typically already pmap'd, putting that * special case in might be more effort then it is * worth. We cannot under any circumstances mess * around with a shared busied page except, perhaps, * to pmap it. */ if (vm_page_tryxbusy(fs.m) == 0) { /* * Reference the page before unlocking and * sleeping so that the page daemon is less * likely to reclaim it. */ vm_page_aflag_set(fs.m, PGA_REFERENCED); if (fs.object != fs.first_object) { if (!VM_OBJECT_TRYWLOCK( fs.first_object)) { VM_OBJECT_WUNLOCK(fs.object); VM_OBJECT_WLOCK(fs.first_object); VM_OBJECT_WLOCK(fs.object); } vm_page_free(fs.first_m); vm_object_pip_wakeup(fs.first_object); VM_OBJECT_WUNLOCK(fs.first_object); fs.first_m = NULL; } unlock_map(&fs); if (fs.m == vm_page_lookup(fs.object, fs.pindex)) { vm_page_sleep_if_busy(fs.m, "vmpfw"); } vm_object_pip_wakeup(fs.object); VM_OBJECT_WUNLOCK(fs.object); VM_CNT_INC(v_intrans); vm_object_deallocate(fs.first_object); goto RetryFault; } /* * The page is marked busy for other processes and the * pagedaemon. If it still isn't completely valid * (readable), jump to readrest, else break-out ( we * found the page ). */ - if (fs.m->valid != VM_PAGE_BITS_ALL) + if (!vm_page_all_valid(fs.m)) goto readrest; break; /* break to PAGE HAS BEEN FOUND */ } KASSERT(fs.m == NULL, ("fs.m should be NULL, not %p", fs.m)); /* * Page is not resident. If the pager might contain the page * or this is the beginning of the search, allocate a new * page. (Default objects are zero-fill, so there is no real * pager for them.) */ if (fs.object->type != OBJT_DEFAULT || fs.object == fs.first_object) { if (fs.pindex >= fs.object->size) { unlock_and_deallocate(&fs); return (KERN_OUT_OF_BOUNDS); } if (fs.object == fs.first_object && (fs.first_object->flags & OBJ_POPULATE) != 0 && fs.first_object->shadow_count == 0) { rv = vm_fault_populate(&fs, prot, fault_type, fault_flags, wired, m_hold); switch (rv) { case KERN_SUCCESS: case KERN_FAILURE: unlock_and_deallocate(&fs); return (rv); case KERN_RESOURCE_SHORTAGE: unlock_and_deallocate(&fs); goto RetryFault; case KERN_NOT_RECEIVER: /* * Pager's populate() method * returned VM_PAGER_BAD. */ break; default: panic("inconsistent return codes"); } } /* * Allocate a new page for this object/offset pair. * * Unlocked read of the p_flag is harmless. At * worst, the P_KILLED might be not observed * there, and allocation can fail, causing * restart and new reading of the p_flag. */ dset = fs.object->domain.dr_policy; if (dset == NULL) dset = curthread->td_domain.dr_policy; if (!vm_page_count_severe_set(&dset->ds_mask) || P_KILLED(curproc)) { #if VM_NRESERVLEVEL > 0 vm_object_color(fs.object, atop(vaddr) - fs.pindex); #endif alloc_req = P_KILLED(curproc) ? VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL; if (fs.object->type != OBJT_VNODE && fs.object->backing_object == NULL) alloc_req |= VM_ALLOC_ZERO; fs.m = vm_page_alloc(fs.object, fs.pindex, alloc_req); } if (fs.m == NULL) { unlock_and_deallocate(&fs); if (vm_pfault_oom_attempts < 0 || oom < vm_pfault_oom_attempts) { oom++; vm_waitpfault(dset, vm_pfault_oom_wait * hz); goto RetryFault_oom; } if (bootverbose) printf( "proc %d (%s) failed to alloc page on fault, starting OOM\n", curproc->p_pid, curproc->p_comm); vm_pageout_oom(VM_OOM_MEM_PF); goto RetryFault; } } readrest: /* * At this point, we have either allocated a new page or found * an existing page that is only partially valid. * * We hold a reference on the current object and the page is * exclusive busied. */ /* * If the pager for the current object might have the page, * then determine the number of additional pages to read and * potentially reprioritize previously read pages for earlier * reclamation. These operations should only be performed * once per page fault. Even if the current pager doesn't * have the page, the number of additional pages to read will * apply to subsequent objects in the shadow chain. */ if (fs.object->type != OBJT_DEFAULT && nera == -1 && !P_KILLED(curproc)) { KASSERT(fs.lookup_still_valid, ("map unlocked")); era = fs.entry->read_ahead; behavior = vm_map_entry_behavior(fs.entry); if (behavior == MAP_ENTRY_BEHAV_RANDOM) { nera = 0; } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) { nera = VM_FAULT_READ_AHEAD_MAX; if (vaddr == fs.entry->next_read) vm_fault_dontneed(&fs, vaddr, nera); } else if (vaddr == fs.entry->next_read) { /* * This is a sequential fault. Arithmetically * increase the requested number of pages in * the read-ahead window. The requested * number of pages is "# of sequential faults * x (read ahead min + 1) + read ahead min" */ nera = VM_FAULT_READ_AHEAD_MIN; if (era > 0) { nera += era + 1; if (nera > VM_FAULT_READ_AHEAD_MAX) nera = VM_FAULT_READ_AHEAD_MAX; } if (era == VM_FAULT_READ_AHEAD_MAX) vm_fault_dontneed(&fs, vaddr, nera); } else { /* * This is a non-sequential fault. */ nera = 0; } if (era != nera) { /* * A read lock on the map suffices to update * the read ahead count safely. */ fs.entry->read_ahead = nera; } /* * Prepare for unlocking the map. Save the map * entry's start and end addresses, which are used to * optimize the size of the pager operation below. * Even if the map entry's addresses change after * unlocking the map, using the saved addresses is * safe. */ e_start = fs.entry->start; e_end = fs.entry->end; } /* * Call the pager to retrieve the page if there is a chance * that the pager has it, and potentially retrieve additional * pages at the same time. */ if (fs.object->type != OBJT_DEFAULT) { /* * Release the map lock before locking the vnode or * sleeping in the pager. (If the current object has * a shadow, then an earlier iteration of this loop * may have already unlocked the map.) */ unlock_map(&fs); if (fs.object->type == OBJT_VNODE && (vp = fs.object->handle) != fs.vp) { /* * Perform an unlock in case the desired vnode * changed while the map was unlocked during a * retry. */ unlock_vp(&fs); locked = VOP_ISLOCKED(vp); if (locked != LK_EXCLUSIVE) locked = LK_SHARED; /* * We must not sleep acquiring the vnode lock * while we have the page exclusive busied or * the object's paging-in-progress count * incremented. Otherwise, we could deadlock. */ error = vget(vp, locked | LK_CANRECURSE | LK_NOWAIT, curthread); if (error != 0) { vhold(vp); release_page(&fs); unlock_and_deallocate(&fs); error = vget(vp, locked | LK_RETRY | LK_CANRECURSE, curthread); vdrop(vp); fs.vp = vp; KASSERT(error == 0, ("vm_fault: vget failed")); goto RetryFault; } fs.vp = vp; } KASSERT(fs.vp == NULL || !fs.map->system_map, ("vm_fault: vnode-backed object mapped by system map")); /* * Page in the requested page and hint the pager, * that it may bring up surrounding pages. */ if (nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM || P_KILLED(curproc)) { behind = 0; ahead = 0; } else { /* Is this a sequential fault? */ if (nera > 0) { behind = 0; ahead = nera; } else { /* * Request a cluster of pages that is * aligned to a VM_FAULT_READ_DEFAULT * page offset boundary within the * object. Alignment to a page offset * boundary is more likely to coincide * with the underlying file system * block than alignment to a virtual * address boundary. */ cluster_offset = fs.pindex % VM_FAULT_READ_DEFAULT; behind = ulmin(cluster_offset, atop(vaddr - e_start)); ahead = VM_FAULT_READ_DEFAULT - 1 - cluster_offset; } ahead = ulmin(ahead, atop(e_end - vaddr) - 1); } rv = vm_pager_get_pages(fs.object, &fs.m, 1, &behind, &ahead); if (rv == VM_PAGER_OK) { faultcount = behind + 1 + ahead; hardfault = true; break; /* break to PAGE HAS BEEN FOUND */ } if (rv == VM_PAGER_ERROR) printf("vm_fault: pager read error, pid %d (%s)\n", curproc->p_pid, curproc->p_comm); /* * If an I/O error occurred or the requested page was * outside the range of the pager, clean up and return * an error. */ if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) { if (!vm_page_wired(fs.m)) vm_page_free(fs.m); else vm_page_xunbusy(fs.m); fs.m = NULL; unlock_and_deallocate(&fs); return (KERN_OUT_OF_BOUNDS); } /* * The requested page does not exist at this object/ * offset. Remove the invalid page from the object, * waking up anyone waiting for it, and continue on to * the next object. However, if this is the top-level * object, we must leave the busy page in place to * prevent another process from rushing past us, and * inserting the page in that object at the same time * that we are. */ if (fs.object != fs.first_object) { if (!vm_page_wired(fs.m)) vm_page_free(fs.m); else vm_page_xunbusy(fs.m); fs.m = NULL; } } /* * We get here if the object has default pager (or unwiring) * or the pager doesn't have the page. */ if (fs.object == fs.first_object) fs.first_m = fs.m; /* * Move on to the next object. Lock the next object before * unlocking the current one. */ next_object = fs.object->backing_object; if (next_object == NULL) { /* * If there's no object left, fill the page in the top * object with zeros. */ if (fs.object != fs.first_object) { vm_object_pip_wakeup(fs.object); VM_OBJECT_WUNLOCK(fs.object); fs.object = fs.first_object; fs.pindex = fs.first_pindex; fs.m = fs.first_m; VM_OBJECT_WLOCK(fs.object); } fs.first_m = NULL; /* * Zero the page if necessary and mark it valid. */ if ((fs.m->flags & PG_ZERO) == 0) { pmap_zero_page(fs.m); } else { VM_CNT_INC(v_ozfod); } VM_CNT_INC(v_zfod); - fs.m->valid = VM_PAGE_BITS_ALL; + vm_page_valid(fs.m); /* Don't try to prefault neighboring pages. */ faultcount = 1; break; /* break to PAGE HAS BEEN FOUND */ } else { KASSERT(fs.object != next_object, ("object loop %p", next_object)); VM_OBJECT_WLOCK(next_object); vm_object_pip_add(next_object, 1); if (fs.object != fs.first_object) vm_object_pip_wakeup(fs.object); fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset); VM_OBJECT_WUNLOCK(fs.object); fs.object = next_object; } } vm_page_assert_xbusied(fs.m); /* * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock * is held.] */ /* * If the page is being written, but isn't already owned by the * top-level object, we have to copy it into a new page owned by the * top-level object. */ if (fs.object != fs.first_object) { /* * We only really need to copy if we want to write it. */ if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) { /* * This allows pages to be virtually copied from a * backing_object into the first_object, where the * backing object has no other refs to it, and cannot * gain any more refs. Instead of a bcopy, we just * move the page from the backing object to the * first object. Note that we must mark the page * dirty in the first object so that it will go out * to swap when needed. */ is_first_object_locked = false; if ( /* * Only one shadow object */ (fs.object->shadow_count == 1) && /* * No COW refs, except us */ (fs.object->ref_count == 1) && /* * No one else can look this object up */ (fs.object->handle == NULL) && /* * No other ways to look the object up */ ((fs.object->type == OBJT_DEFAULT) || (fs.object->type == OBJT_SWAP)) && (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) && /* * We don't chase down the shadow chain */ fs.object == fs.first_object->backing_object) { (void)vm_page_remove(fs.m); vm_page_replace_checked(fs.m, fs.first_object, fs.first_pindex, fs.first_m); vm_page_free(fs.first_m); vm_page_dirty(fs.m); #if VM_NRESERVLEVEL > 0 /* * Rename the reservation. */ vm_reserv_rename(fs.m, fs.first_object, fs.object, OFF_TO_IDX( fs.first_object->backing_object_offset)); #endif fs.first_m = fs.m; fs.m = NULL; VM_CNT_INC(v_cow_optim); } else { /* * Oh, well, lets copy it. */ pmap_copy_page(fs.m, fs.first_m); - fs.first_m->valid = VM_PAGE_BITS_ALL; + vm_page_valid(fs.first_m); if (wired && (fault_flags & VM_FAULT_WIRE) == 0) { vm_page_wire(fs.first_m); vm_page_unwire(fs.m, PQ_INACTIVE); } /* * We no longer need the old page or object. */ release_page(&fs); } /* * fs.object != fs.first_object due to above * conditional */ vm_object_pip_wakeup(fs.object); VM_OBJECT_WUNLOCK(fs.object); /* * We only try to prefault read-only mappings to the * neighboring pages when this copy-on-write fault is * a hard fault. In other cases, trying to prefault * is typically wasted effort. */ if (faultcount == 0) faultcount = 1; /* * Only use the new page below... */ fs.object = fs.first_object; fs.pindex = fs.first_pindex; fs.m = fs.first_m; if (!is_first_object_locked) VM_OBJECT_WLOCK(fs.object); VM_CNT_INC(v_cow_faults); curthread->td_cow++; } else { prot &= ~VM_PROT_WRITE; } } /* * We must verify that the maps have not changed since our last * lookup. */ if (!fs.lookup_still_valid) { if (!vm_map_trylock_read(fs.map)) { release_page(&fs); unlock_and_deallocate(&fs); goto RetryFault; } fs.lookup_still_valid = true; if (fs.map->timestamp != fs.map_generation) { result = vm_map_lookup_locked(&fs.map, vaddr, fault_type, &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired); /* * If we don't need the page any longer, put it on the inactive * list (the easiest thing to do here). If no one needs it, * pageout will grab it eventually. */ if (result != KERN_SUCCESS) { release_page(&fs); unlock_and_deallocate(&fs); /* * If retry of map lookup would have blocked then * retry fault from start. */ if (result == KERN_FAILURE) goto RetryFault; return (result); } if ((retry_object != fs.first_object) || (retry_pindex != fs.first_pindex)) { release_page(&fs); unlock_and_deallocate(&fs); goto RetryFault; } /* * Check whether the protection has changed or the object has * been copied while we left the map unlocked. Changing from * read to write permission is OK - we leave the page * write-protected, and catch the write fault. Changing from * write to read permission means that we can't mark the page * write-enabled after all. */ prot &= retry_prot; fault_type &= retry_prot; if (prot == 0) { release_page(&fs); unlock_and_deallocate(&fs); goto RetryFault; } /* Reassert because wired may have changed. */ KASSERT(wired || (fault_flags & VM_FAULT_WIRE) == 0, ("!wired && VM_FAULT_WIRE")); } } /* * If the page was filled by a pager, save the virtual address that * should be faulted on next under a sequential access pattern to the * map entry. A read lock on the map suffices to update this address * safely. */ if (hardfault) fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE; vm_page_assert_xbusied(fs.m); vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags, true); /* * Page must be completely valid or it is not fit to * map into user space. vm_pager_get_pages() ensures this. */ - KASSERT(fs.m->valid == VM_PAGE_BITS_ALL, + KASSERT(vm_page_all_valid(fs.m), ("vm_fault: page %p partially invalid", fs.m)); VM_OBJECT_WUNLOCK(fs.object); /* * Put this page into the physical map. We had to do the unlock above * because pmap_enter() may sleep. We don't put the page * back on the active queue until later so that the pageout daemon * won't find it (yet). */ pmap_enter(fs.map->pmap, vaddr, fs.m, prot, fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0); if (faultcount != 1 && (fault_flags & VM_FAULT_WIRE) == 0 && wired == 0) vm_fault_prefault(&fs, vaddr, faultcount > 0 ? behind : PFBAK, faultcount > 0 ? ahead : PFFOR, false); VM_OBJECT_WLOCK(fs.object); /* * If the page is not wired down, then put it where the pageout daemon * can find it. */ if ((fault_flags & VM_FAULT_WIRE) != 0) { vm_page_wire(fs.m); } else { vm_page_lock(fs.m); vm_page_activate(fs.m); vm_page_unlock(fs.m); } if (m_hold != NULL) { *m_hold = fs.m; vm_page_wire(fs.m); } vm_page_xunbusy(fs.m); /* * Unlock everything, and return */ unlock_and_deallocate(&fs); if (hardfault) { VM_CNT_INC(v_io_faults); curthread->td_ru.ru_majflt++; #ifdef RACCT if (racct_enable && fs.object->type == OBJT_VNODE) { PROC_LOCK(curproc); if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) { racct_add_force(curproc, RACCT_WRITEBPS, PAGE_SIZE + behind * PAGE_SIZE); racct_add_force(curproc, RACCT_WRITEIOPS, 1); } else { racct_add_force(curproc, RACCT_READBPS, PAGE_SIZE + ahead * PAGE_SIZE); racct_add_force(curproc, RACCT_READIOPS, 1); } PROC_UNLOCK(curproc); } #endif } else curthread->td_ru.ru_minflt++; return (KERN_SUCCESS); } /* * Speed up the reclamation of pages that precede the faulting pindex within * the first object of the shadow chain. Essentially, perform the equivalent * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes * the faulting pindex by the cluster size when the pages read by vm_fault() * cross a cluster-size boundary. The cluster size is the greater of the * smallest superpage size and VM_FAULT_DONTNEED_MIN. * * When "fs->first_object" is a shadow object, the pages in the backing object * that precede the faulting pindex are deactivated by vm_fault(). So, this * function must only be concerned with pages in the first object. */ static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead) { vm_map_entry_t entry; vm_object_t first_object, object; vm_offset_t end, start; vm_page_t m, m_next; vm_pindex_t pend, pstart; vm_size_t size; object = fs->object; VM_OBJECT_ASSERT_WLOCKED(object); first_object = fs->first_object; if (first_object != object) { if (!VM_OBJECT_TRYWLOCK(first_object)) { VM_OBJECT_WUNLOCK(object); VM_OBJECT_WLOCK(first_object); VM_OBJECT_WLOCK(object); } } /* Neither fictitious nor unmanaged pages can be reclaimed. */ if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) { size = VM_FAULT_DONTNEED_MIN; if (MAXPAGESIZES > 1 && size < pagesizes[1]) size = pagesizes[1]; end = rounddown2(vaddr, size); if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) && (entry = fs->entry)->start < end) { if (end - entry->start < size) start = entry->start; else start = end - size; pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED); pstart = OFF_TO_IDX(entry->offset) + atop(start - entry->start); m_next = vm_page_find_least(first_object, pstart); pend = OFF_TO_IDX(entry->offset) + atop(end - entry->start); while ((m = m_next) != NULL && m->pindex < pend) { m_next = TAILQ_NEXT(m, listq); - if (m->valid != VM_PAGE_BITS_ALL || + if (!vm_page_all_valid(m) || vm_page_busied(m)) continue; /* * Don't clear PGA_REFERENCED, since it would * likely represent a reference by a different * process. * * Typically, at this point, prefetched pages * are still in the inactive queue. Only * pages that triggered page faults are in the * active queue. */ vm_page_lock(m); if (!vm_page_inactive(m)) vm_page_deactivate(m); vm_page_unlock(m); } } } if (first_object != object) VM_OBJECT_WUNLOCK(first_object); } /* * vm_fault_prefault provides a quick way of clustering * pagefaults into a processes address space. It is a "cousin" * of vm_map_pmap_enter, except it runs at page fault time instead * of mmap time. */ static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra, int backward, int forward, bool obj_locked) { pmap_t pmap; vm_map_entry_t entry; vm_object_t backing_object, lobject; vm_offset_t addr, starta; vm_pindex_t pindex; vm_page_t m; int i; pmap = fs->map->pmap; if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)) return; entry = fs->entry; if (addra < backward * PAGE_SIZE) { starta = entry->start; } else { starta = addra - backward * PAGE_SIZE; if (starta < entry->start) starta = entry->start; } /* * Generate the sequence of virtual addresses that are candidates for * prefaulting in an outward spiral from the faulting virtual address, * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ... * If the candidate address doesn't have a backing physical page, then * the loop immediately terminates. */ for (i = 0; i < 2 * imax(backward, forward); i++) { addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE : PAGE_SIZE); if (addr > addra + forward * PAGE_SIZE) addr = 0; if (addr < starta || addr >= entry->end) continue; if (!pmap_is_prefaultable(pmap, addr)) continue; pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; lobject = entry->object.vm_object; if (!obj_locked) VM_OBJECT_RLOCK(lobject); while ((m = vm_page_lookup(lobject, pindex)) == NULL && lobject->type == OBJT_DEFAULT && (backing_object = lobject->backing_object) != NULL) { KASSERT((lobject->backing_object_offset & PAGE_MASK) == 0, ("vm_fault_prefault: unaligned object offset")); pindex += lobject->backing_object_offset >> PAGE_SHIFT; VM_OBJECT_RLOCK(backing_object); if (!obj_locked || lobject != entry->object.vm_object) VM_OBJECT_RUNLOCK(lobject); lobject = backing_object; } if (m == NULL) { if (!obj_locked || lobject != entry->object.vm_object) VM_OBJECT_RUNLOCK(lobject); break; } - if (m->valid == VM_PAGE_BITS_ALL && + if (vm_page_all_valid(m) && (m->flags & PG_FICTITIOUS) == 0) pmap_enter_quick(pmap, addr, m, entry->protection); if (!obj_locked || lobject != entry->object.vm_object) VM_OBJECT_RUNLOCK(lobject); } } /* * Hold each of the physical pages that are mapped by the specified range of * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid * and allow the specified types of access, "prot". If all of the implied * pages are successfully held, then the number of held pages is returned * together with pointers to those pages in the array "ma". However, if any * of the pages cannot be held, -1 is returned. */ int vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len, vm_prot_t prot, vm_page_t *ma, int max_count) { vm_offset_t end, va; vm_page_t *mp; int count; boolean_t pmap_failed; if (len == 0) return (0); end = round_page(addr + len); addr = trunc_page(addr); /* * Check for illegal addresses. */ if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map)) return (-1); if (atop(end - addr) > max_count) panic("vm_fault_quick_hold_pages: count > max_count"); count = atop(end - addr); /* * Most likely, the physical pages are resident in the pmap, so it is * faster to try pmap_extract_and_hold() first. */ pmap_failed = FALSE; for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) { *mp = pmap_extract_and_hold(map->pmap, va, prot); if (*mp == NULL) pmap_failed = TRUE; else if ((prot & VM_PROT_WRITE) != 0 && (*mp)->dirty != VM_PAGE_BITS_ALL) { /* * Explicitly dirty the physical page. Otherwise, the * caller's changes may go unnoticed because they are * performed through an unmanaged mapping or by a DMA * operation. * * The object lock is not held here. * See vm_page_clear_dirty_mask(). */ vm_page_dirty(*mp); } } if (pmap_failed) { /* * One or more pages could not be held by the pmap. Either no * page was mapped at the specified virtual address or that * mapping had insufficient permissions. Attempt to fault in * and hold these pages. * * If vm_fault_disable_pagefaults() was called, * i.e., TDP_NOFAULTING is set, we must not sleep nor * acquire MD VM locks, which means we must not call * vm_fault(). Some (out of tree) callers mark * too wide a code area with vm_fault_disable_pagefaults() * already, use the VM_PROT_QUICK_NOFAULT flag to request * the proper behaviour explicitly. */ if ((prot & VM_PROT_QUICK_NOFAULT) != 0 && (curthread->td_pflags & TDP_NOFAULTING) != 0) goto error; for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) if (*mp == NULL && vm_fault(map, va, prot, VM_FAULT_NORMAL, mp) != KERN_SUCCESS) goto error; } return (count); error: for (mp = ma; mp < ma + count; mp++) if (*mp != NULL) vm_page_unwire(*mp, PQ_INACTIVE); return (-1); } /* * Routine: * vm_fault_copy_entry * Function: * Create new shadow object backing dst_entry with private copy of * all underlying pages. When src_entry is equal to dst_entry, * function implements COW for wired-down map entry. Otherwise, * it forks wired entry into dst_map. * * In/out conditions: * The source and destination maps must be locked for write. * The source map entry must be wired down (or be a sharing map * entry corresponding to a main map entry that is wired down). */ void vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map, vm_map_entry_t dst_entry, vm_map_entry_t src_entry, vm_ooffset_t *fork_charge) { vm_object_t backing_object, dst_object, object, src_object; vm_pindex_t dst_pindex, pindex, src_pindex; vm_prot_t access, prot; vm_offset_t vaddr; vm_page_t dst_m; vm_page_t src_m; boolean_t upgrade; #ifdef lint src_map++; #endif /* lint */ upgrade = src_entry == dst_entry; access = prot = dst_entry->protection; src_object = src_entry->object.vm_object; src_pindex = OFF_TO_IDX(src_entry->offset); if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) { dst_object = src_object; vm_object_reference(dst_object); } else { /* * Create the top-level object for the destination entry. (Doesn't * actually shadow anything - we copy the pages directly.) */ dst_object = vm_object_allocate(OBJT_DEFAULT, atop(dst_entry->end - dst_entry->start)); #if VM_NRESERVLEVEL > 0 dst_object->flags |= OBJ_COLORED; dst_object->pg_color = atop(dst_entry->start); #endif dst_object->domain = src_object->domain; dst_object->charge = dst_entry->end - dst_entry->start; } VM_OBJECT_WLOCK(dst_object); KASSERT(upgrade || dst_entry->object.vm_object == NULL, ("vm_fault_copy_entry: vm_object not NULL")); if (src_object != dst_object) { dst_entry->object.vm_object = dst_object; dst_entry->offset = 0; dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC; } if (fork_charge != NULL) { KASSERT(dst_entry->cred == NULL, ("vm_fault_copy_entry: leaked swp charge")); dst_object->cred = curthread->td_ucred; crhold(dst_object->cred); *fork_charge += dst_object->charge; } else if ((dst_object->type == OBJT_DEFAULT || dst_object->type == OBJT_SWAP) && dst_object->cred == NULL) { KASSERT(dst_entry->cred != NULL, ("no cred for entry %p", dst_entry)); dst_object->cred = dst_entry->cred; dst_entry->cred = NULL; } /* * If not an upgrade, then enter the mappings in the pmap as * read and/or execute accesses. Otherwise, enter them as * write accesses. * * A writeable large page mapping is only created if all of * the constituent small page mappings are modified. Marking * PTEs as modified on inception allows promotion to happen * without taking potentially large number of soft faults. */ if (!upgrade) access &= ~VM_PROT_WRITE; /* * Loop through all of the virtual pages within the entry's * range, copying each page from the source object to the * destination object. Since the source is wired, those pages * must exist. In contrast, the destination is pageable. * Since the destination object doesn't share any backing storage * with the source object, all of its pages must be dirtied, * regardless of whether they can be written. */ for (vaddr = dst_entry->start, dst_pindex = 0; vaddr < dst_entry->end; vaddr += PAGE_SIZE, dst_pindex++) { again: /* * Find the page in the source object, and copy it in. * Because the source is wired down, the page will be * in memory. */ if (src_object != dst_object) VM_OBJECT_RLOCK(src_object); object = src_object; pindex = src_pindex + dst_pindex; while ((src_m = vm_page_lookup(object, pindex)) == NULL && (backing_object = object->backing_object) != NULL) { /* * Unless the source mapping is read-only or * it is presently being upgraded from * read-only, the first object in the shadow * chain should provide all of the pages. In * other words, this loop body should never be * executed when the source mapping is already * read/write. */ KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 || upgrade, ("vm_fault_copy_entry: main object missing page")); VM_OBJECT_RLOCK(backing_object); pindex += OFF_TO_IDX(object->backing_object_offset); if (object != dst_object) VM_OBJECT_RUNLOCK(object); object = backing_object; } KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing")); if (object != dst_object) { /* * Allocate a page in the destination object. */ dst_m = vm_page_alloc(dst_object, (src_object == dst_object ? src_pindex : 0) + dst_pindex, VM_ALLOC_NORMAL); if (dst_m == NULL) { VM_OBJECT_WUNLOCK(dst_object); VM_OBJECT_RUNLOCK(object); vm_wait(dst_object); VM_OBJECT_WLOCK(dst_object); goto again; } pmap_copy_page(src_m, dst_m); VM_OBJECT_RUNLOCK(object); dst_m->dirty = dst_m->valid = src_m->valid; } else { dst_m = src_m; if (vm_page_busy_acquire(dst_m, VM_ALLOC_WAITFAIL) == 0) goto again; if (dst_m->pindex >= dst_object->size) { /* * We are upgrading. Index can occur * out of bounds if the object type is * vnode and the file was truncated. */ vm_page_xunbusy(dst_m); break; } } VM_OBJECT_WUNLOCK(dst_object); /* * Enter it in the pmap. If a wired, copy-on-write * mapping is being replaced by a write-enabled * mapping, then wire that new mapping. * * The page can be invalid if the user called * msync(MS_INVALIDATE) or truncated the backing vnode * or shared memory object. In this case, do not * insert it into pmap, but still do the copy so that * all copies of the wired map entry have similar * backing pages. */ - if (dst_m->valid == VM_PAGE_BITS_ALL) { + if (vm_page_all_valid(dst_m)) { pmap_enter(dst_map->pmap, vaddr, dst_m, prot, access | (upgrade ? PMAP_ENTER_WIRED : 0), 0); } /* * Mark it no longer busy, and put it on the active list. */ VM_OBJECT_WLOCK(dst_object); if (upgrade) { if (src_m != dst_m) { vm_page_unwire(src_m, PQ_INACTIVE); vm_page_wire(dst_m); } else { KASSERT(vm_page_wired(dst_m), ("dst_m %p is not wired", dst_m)); } } else { vm_page_lock(dst_m); vm_page_activate(dst_m); vm_page_unlock(dst_m); } vm_page_xunbusy(dst_m); } VM_OBJECT_WUNLOCK(dst_object); if (upgrade) { dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY); vm_object_deallocate(src_object); } } /* * Block entry into the machine-independent layer's page fault handler by * the calling thread. Subsequent calls to vm_fault() by that thread will * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of * spurious page faults. */ int vm_fault_disable_pagefaults(void) { return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR)); } void vm_fault_enable_pagefaults(int save) { curthread_pflags_restore(save); } Index: head/sys/vm/vm_map.c =================================================================== --- head/sys/vm/vm_map.c (revision 353538) +++ head/sys/vm/vm_map.c (revision 353539) @@ -1,4931 +1,4931 @@ /*- * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU) * * Copyright (c) 1991, 1993 * The Regents of the University of California. All rights reserved. * * This code is derived from software contributed to Berkeley by * The Mach Operating System project at Carnegie-Mellon University. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)vm_map.c 8.3 (Berkeley) 1/12/94 * * * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Authors: Avadis Tevanian, Jr., Michael Wayne Young * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. */ /* * Virtual memory mapping module. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* * Virtual memory maps provide for the mapping, protection, * and sharing of virtual memory objects. In addition, * this module provides for an efficient virtual copy of * memory from one map to another. * * Synchronization is required prior to most operations. * * Maps consist of an ordered doubly-linked list of simple * entries; a self-adjusting binary search tree of these * entries is used to speed up lookups. * * Since portions of maps are specified by start/end addresses, * which may not align with existing map entries, all * routines merely "clip" entries to these start/end values. * [That is, an entry is split into two, bordering at a * start or end value.] Note that these clippings may not * always be necessary (as the two resulting entries are then * not changed); however, the clipping is done for convenience. * * As mentioned above, virtual copy operations are performed * by copying VM object references from one map to * another, and then marking both regions as copy-on-write. */ static struct mtx map_sleep_mtx; static uma_zone_t mapentzone; static uma_zone_t kmapentzone; static uma_zone_t mapzone; static uma_zone_t vmspace_zone; static int vmspace_zinit(void *mem, int size, int flags); static int vm_map_zinit(void *mem, int ize, int flags); static void _vm_map_init(vm_map_t map, pmap_t pmap, vm_offset_t min, vm_offset_t max); static void vm_map_entry_deallocate(vm_map_entry_t entry, boolean_t system_map); static void vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry); static void vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry); static int vm_map_growstack(vm_map_t map, vm_offset_t addr, vm_map_entry_t gap_entry); static void vm_map_pmap_enter(vm_map_t map, vm_offset_t addr, vm_prot_t prot, vm_object_t object, vm_pindex_t pindex, vm_size_t size, int flags); #ifdef INVARIANTS static void vm_map_zdtor(void *mem, int size, void *arg); static void vmspace_zdtor(void *mem, int size, void *arg); #endif static int vm_map_stack_locked(vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize, vm_size_t growsize, vm_prot_t prot, vm_prot_t max, int cow); static void vm_map_wire_entry_failure(vm_map_t map, vm_map_entry_t entry, vm_offset_t failed_addr); #define ENTRY_CHARGED(e) ((e)->cred != NULL || \ ((e)->object.vm_object != NULL && (e)->object.vm_object->cred != NULL && \ !((e)->eflags & MAP_ENTRY_NEEDS_COPY))) /* * PROC_VMSPACE_{UN,}LOCK() can be a noop as long as vmspaces are type * stable. */ #define PROC_VMSPACE_LOCK(p) do { } while (0) #define PROC_VMSPACE_UNLOCK(p) do { } while (0) /* * VM_MAP_RANGE_CHECK: [ internal use only ] * * Asserts that the starting and ending region * addresses fall within the valid range of the map. */ #define VM_MAP_RANGE_CHECK(map, start, end) \ { \ if (start < vm_map_min(map)) \ start = vm_map_min(map); \ if (end > vm_map_max(map)) \ end = vm_map_max(map); \ if (start > end) \ start = end; \ } /* * vm_map_startup: * * Initialize the vm_map module. Must be called before * any other vm_map routines. * * Map and entry structures are allocated from the general * purpose memory pool with some exceptions: * * - The kernel map and kmem submap are allocated statically. * - Kernel map entries are allocated out of a static pool. * * These restrictions are necessary since malloc() uses the * maps and requires map entries. */ void vm_map_startup(void) { mtx_init(&map_sleep_mtx, "vm map sleep mutex", NULL, MTX_DEF); mapzone = uma_zcreate("MAP", sizeof(struct vm_map), NULL, #ifdef INVARIANTS vm_map_zdtor, #else NULL, #endif vm_map_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); uma_prealloc(mapzone, MAX_KMAP); kmapentzone = uma_zcreate("KMAP ENTRY", sizeof(struct vm_map_entry), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_MTXCLASS | UMA_ZONE_VM); mapentzone = uma_zcreate("MAP ENTRY", sizeof(struct vm_map_entry), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); vmspace_zone = uma_zcreate("VMSPACE", sizeof(struct vmspace), NULL, #ifdef INVARIANTS vmspace_zdtor, #else NULL, #endif vmspace_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); } static int vmspace_zinit(void *mem, int size, int flags) { struct vmspace *vm; vm = (struct vmspace *)mem; vm->vm_map.pmap = NULL; (void)vm_map_zinit(&vm->vm_map, sizeof(vm->vm_map), flags); PMAP_LOCK_INIT(vmspace_pmap(vm)); return (0); } static int vm_map_zinit(void *mem, int size, int flags) { vm_map_t map; map = (vm_map_t)mem; memset(map, 0, sizeof(*map)); mtx_init(&map->system_mtx, "vm map (system)", NULL, MTX_DEF | MTX_DUPOK); sx_init(&map->lock, "vm map (user)"); return (0); } #ifdef INVARIANTS static void vmspace_zdtor(void *mem, int size, void *arg) { struct vmspace *vm; vm = (struct vmspace *)mem; vm_map_zdtor(&vm->vm_map, sizeof(vm->vm_map), arg); } static void vm_map_zdtor(void *mem, int size, void *arg) { vm_map_t map; map = (vm_map_t)mem; KASSERT(map->nentries == 0, ("map %p nentries == %d on free.", map, map->nentries)); KASSERT(map->size == 0, ("map %p size == %lu on free.", map, (unsigned long)map->size)); } #endif /* INVARIANTS */ /* * Allocate a vmspace structure, including a vm_map and pmap, * and initialize those structures. The refcnt is set to 1. * * If 'pinit' is NULL then the embedded pmap is initialized via pmap_pinit(). */ struct vmspace * vmspace_alloc(vm_offset_t min, vm_offset_t max, pmap_pinit_t pinit) { struct vmspace *vm; vm = uma_zalloc(vmspace_zone, M_WAITOK); KASSERT(vm->vm_map.pmap == NULL, ("vm_map.pmap must be NULL")); if (!pinit(vmspace_pmap(vm))) { uma_zfree(vmspace_zone, vm); return (NULL); } CTR1(KTR_VM, "vmspace_alloc: %p", vm); _vm_map_init(&vm->vm_map, vmspace_pmap(vm), min, max); vm->vm_refcnt = 1; vm->vm_shm = NULL; vm->vm_swrss = 0; vm->vm_tsize = 0; vm->vm_dsize = 0; vm->vm_ssize = 0; vm->vm_taddr = 0; vm->vm_daddr = 0; vm->vm_maxsaddr = 0; return (vm); } #ifdef RACCT static void vmspace_container_reset(struct proc *p) { PROC_LOCK(p); racct_set(p, RACCT_DATA, 0); racct_set(p, RACCT_STACK, 0); racct_set(p, RACCT_RSS, 0); racct_set(p, RACCT_MEMLOCK, 0); racct_set(p, RACCT_VMEM, 0); PROC_UNLOCK(p); } #endif static inline void vmspace_dofree(struct vmspace *vm) { CTR1(KTR_VM, "vmspace_free: %p", vm); /* * Make sure any SysV shm is freed, it might not have been in * exit1(). */ shmexit(vm); /* * Lock the map, to wait out all other references to it. * Delete all of the mappings and pages they hold, then call * the pmap module to reclaim anything left. */ (void)vm_map_remove(&vm->vm_map, vm_map_min(&vm->vm_map), vm_map_max(&vm->vm_map)); pmap_release(vmspace_pmap(vm)); vm->vm_map.pmap = NULL; uma_zfree(vmspace_zone, vm); } void vmspace_free(struct vmspace *vm) { WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "vmspace_free() called"); if (vm->vm_refcnt == 0) panic("vmspace_free: attempt to free already freed vmspace"); if (atomic_fetchadd_int(&vm->vm_refcnt, -1) == 1) vmspace_dofree(vm); } void vmspace_exitfree(struct proc *p) { struct vmspace *vm; PROC_VMSPACE_LOCK(p); vm = p->p_vmspace; p->p_vmspace = NULL; PROC_VMSPACE_UNLOCK(p); KASSERT(vm == &vmspace0, ("vmspace_exitfree: wrong vmspace")); vmspace_free(vm); } void vmspace_exit(struct thread *td) { int refcnt; struct vmspace *vm; struct proc *p; /* * Release user portion of address space. * This releases references to vnodes, * which could cause I/O if the file has been unlinked. * Need to do this early enough that we can still sleep. * * The last exiting process to reach this point releases as * much of the environment as it can. vmspace_dofree() is the * slower fallback in case another process had a temporary * reference to the vmspace. */ p = td->td_proc; vm = p->p_vmspace; atomic_add_int(&vmspace0.vm_refcnt, 1); refcnt = vm->vm_refcnt; do { if (refcnt > 1 && p->p_vmspace != &vmspace0) { /* Switch now since other proc might free vmspace */ PROC_VMSPACE_LOCK(p); p->p_vmspace = &vmspace0; PROC_VMSPACE_UNLOCK(p); pmap_activate(td); } } while (!atomic_fcmpset_int(&vm->vm_refcnt, &refcnt, refcnt - 1)); if (refcnt == 1) { if (p->p_vmspace != vm) { /* vmspace not yet freed, switch back */ PROC_VMSPACE_LOCK(p); p->p_vmspace = vm; PROC_VMSPACE_UNLOCK(p); pmap_activate(td); } pmap_remove_pages(vmspace_pmap(vm)); /* Switch now since this proc will free vmspace */ PROC_VMSPACE_LOCK(p); p->p_vmspace = &vmspace0; PROC_VMSPACE_UNLOCK(p); pmap_activate(td); vmspace_dofree(vm); } #ifdef RACCT if (racct_enable) vmspace_container_reset(p); #endif } /* Acquire reference to vmspace owned by another process. */ struct vmspace * vmspace_acquire_ref(struct proc *p) { struct vmspace *vm; int refcnt; PROC_VMSPACE_LOCK(p); vm = p->p_vmspace; if (vm == NULL) { PROC_VMSPACE_UNLOCK(p); return (NULL); } refcnt = vm->vm_refcnt; do { if (refcnt <= 0) { /* Avoid 0->1 transition */ PROC_VMSPACE_UNLOCK(p); return (NULL); } } while (!atomic_fcmpset_int(&vm->vm_refcnt, &refcnt, refcnt + 1)); if (vm != p->p_vmspace) { PROC_VMSPACE_UNLOCK(p); vmspace_free(vm); return (NULL); } PROC_VMSPACE_UNLOCK(p); return (vm); } /* * Switch between vmspaces in an AIO kernel process. * * The new vmspace is either the vmspace of a user process obtained * from an active AIO request or the initial vmspace of the AIO kernel * process (when it is idling). Because user processes will block to * drain any active AIO requests before proceeding in exit() or * execve(), the reference count for vmspaces from AIO requests can * never be 0. Similarly, AIO kernel processes hold an extra * reference on their initial vmspace for the life of the process. As * a result, the 'newvm' vmspace always has a non-zero reference * count. This permits an additional reference on 'newvm' to be * acquired via a simple atomic increment rather than the loop in * vmspace_acquire_ref() above. */ void vmspace_switch_aio(struct vmspace *newvm) { struct vmspace *oldvm; /* XXX: Need some way to assert that this is an aio daemon. */ KASSERT(newvm->vm_refcnt > 0, ("vmspace_switch_aio: newvm unreferenced")); oldvm = curproc->p_vmspace; if (oldvm == newvm) return; /* * Point to the new address space and refer to it. */ curproc->p_vmspace = newvm; atomic_add_int(&newvm->vm_refcnt, 1); /* Activate the new mapping. */ pmap_activate(curthread); vmspace_free(oldvm); } void _vm_map_lock(vm_map_t map, const char *file, int line) { if (map->system_map) mtx_lock_flags_(&map->system_mtx, 0, file, line); else sx_xlock_(&map->lock, file, line); map->timestamp++; } void vm_map_entry_set_vnode_text(vm_map_entry_t entry, bool add) { vm_object_t object, object1; struct vnode *vp; if ((entry->eflags & MAP_ENTRY_VN_EXEC) == 0) return; KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0, ("Submap with execs")); object = entry->object.vm_object; KASSERT(object != NULL, ("No object for text, entry %p", entry)); VM_OBJECT_RLOCK(object); while ((object1 = object->backing_object) != NULL) { VM_OBJECT_RLOCK(object1); VM_OBJECT_RUNLOCK(object); object = object1; } vp = NULL; if (object->type == OBJT_DEAD) { /* * For OBJT_DEAD objects, v_writecount was handled in * vnode_pager_dealloc(). */ } else if (object->type == OBJT_VNODE) { vp = object->handle; } else if (object->type == OBJT_SWAP) { KASSERT((object->flags & OBJ_TMPFS_NODE) != 0, ("vm_map_entry_set_vnode_text: swap and !TMPFS " "entry %p, object %p, add %d", entry, object, add)); /* * Tmpfs VREG node, which was reclaimed, has * OBJ_TMPFS_NODE flag set, but not OBJ_TMPFS. In * this case there is no v_writecount to adjust. */ if ((object->flags & OBJ_TMPFS) != 0) vp = object->un_pager.swp.swp_tmpfs; } else { KASSERT(0, ("vm_map_entry_set_vnode_text: wrong object type, " "entry %p, object %p, add %d", entry, object, add)); } if (vp != NULL) { if (add) { VOP_SET_TEXT_CHECKED(vp); VM_OBJECT_RUNLOCK(object); } else { vhold(vp); VM_OBJECT_RUNLOCK(object); vn_lock(vp, LK_SHARED | LK_RETRY); VOP_UNSET_TEXT_CHECKED(vp); VOP_UNLOCK(vp, 0); vdrop(vp); } } else { VM_OBJECT_RUNLOCK(object); } } static void vm_map_process_deferred(void) { struct thread *td; vm_map_entry_t entry, next; vm_object_t object; td = curthread; entry = td->td_map_def_user; td->td_map_def_user = NULL; while (entry != NULL) { next = entry->next; MPASS((entry->eflags & (MAP_ENTRY_WRITECNT | MAP_ENTRY_VN_EXEC)) != (MAP_ENTRY_WRITECNT | MAP_ENTRY_VN_EXEC)); if ((entry->eflags & MAP_ENTRY_WRITECNT) != 0) { /* * Decrement the object's writemappings and * possibly the vnode's v_writecount. */ KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0, ("Submap with writecount")); object = entry->object.vm_object; KASSERT(object != NULL, ("No object for writecount")); vm_pager_release_writecount(object, entry->start, entry->end); } vm_map_entry_set_vnode_text(entry, false); vm_map_entry_deallocate(entry, FALSE); entry = next; } } void _vm_map_unlock(vm_map_t map, const char *file, int line) { if (map->system_map) mtx_unlock_flags_(&map->system_mtx, 0, file, line); else { sx_xunlock_(&map->lock, file, line); vm_map_process_deferred(); } } void _vm_map_lock_read(vm_map_t map, const char *file, int line) { if (map->system_map) mtx_lock_flags_(&map->system_mtx, 0, file, line); else sx_slock_(&map->lock, file, line); } void _vm_map_unlock_read(vm_map_t map, const char *file, int line) { if (map->system_map) mtx_unlock_flags_(&map->system_mtx, 0, file, line); else { sx_sunlock_(&map->lock, file, line); vm_map_process_deferred(); } } int _vm_map_trylock(vm_map_t map, const char *file, int line) { int error; error = map->system_map ? !mtx_trylock_flags_(&map->system_mtx, 0, file, line) : !sx_try_xlock_(&map->lock, file, line); if (error == 0) map->timestamp++; return (error == 0); } int _vm_map_trylock_read(vm_map_t map, const char *file, int line) { int error; error = map->system_map ? !mtx_trylock_flags_(&map->system_mtx, 0, file, line) : !sx_try_slock_(&map->lock, file, line); return (error == 0); } /* * _vm_map_lock_upgrade: [ internal use only ] * * Tries to upgrade a read (shared) lock on the specified map to a write * (exclusive) lock. Returns the value "0" if the upgrade succeeds and a * non-zero value if the upgrade fails. If the upgrade fails, the map is * returned without a read or write lock held. * * Requires that the map be read locked. */ int _vm_map_lock_upgrade(vm_map_t map, const char *file, int line) { unsigned int last_timestamp; if (map->system_map) { mtx_assert_(&map->system_mtx, MA_OWNED, file, line); } else { if (!sx_try_upgrade_(&map->lock, file, line)) { last_timestamp = map->timestamp; sx_sunlock_(&map->lock, file, line); vm_map_process_deferred(); /* * If the map's timestamp does not change while the * map is unlocked, then the upgrade succeeds. */ sx_xlock_(&map->lock, file, line); if (last_timestamp != map->timestamp) { sx_xunlock_(&map->lock, file, line); return (1); } } } map->timestamp++; return (0); } void _vm_map_lock_downgrade(vm_map_t map, const char *file, int line) { if (map->system_map) { mtx_assert_(&map->system_mtx, MA_OWNED, file, line); } else sx_downgrade_(&map->lock, file, line); } /* * vm_map_locked: * * Returns a non-zero value if the caller holds a write (exclusive) lock * on the specified map and the value "0" otherwise. */ int vm_map_locked(vm_map_t map) { if (map->system_map) return (mtx_owned(&map->system_mtx)); else return (sx_xlocked(&map->lock)); } #ifdef INVARIANTS static void _vm_map_assert_locked(vm_map_t map, const char *file, int line) { if (map->system_map) mtx_assert_(&map->system_mtx, MA_OWNED, file, line); else sx_assert_(&map->lock, SA_XLOCKED, file, line); } #define VM_MAP_ASSERT_LOCKED(map) \ _vm_map_assert_locked(map, LOCK_FILE, LOCK_LINE) #ifdef DIAGNOSTIC static int enable_vmmap_check = 1; #else static int enable_vmmap_check = 0; #endif SYSCTL_INT(_debug, OID_AUTO, vmmap_check, CTLFLAG_RWTUN, &enable_vmmap_check, 0, "Enable vm map consistency checking"); static void _vm_map_assert_consistent(vm_map_t map); #define VM_MAP_ASSERT_CONSISTENT(map) \ _vm_map_assert_consistent(map) #else #define VM_MAP_ASSERT_LOCKED(map) #define VM_MAP_ASSERT_CONSISTENT(map) #endif /* INVARIANTS */ /* * _vm_map_unlock_and_wait: * * Atomically releases the lock on the specified map and puts the calling * thread to sleep. The calling thread will remain asleep until either * vm_map_wakeup() is performed on the map or the specified timeout is * exceeded. * * WARNING! This function does not perform deferred deallocations of * objects and map entries. Therefore, the calling thread is expected to * reacquire the map lock after reawakening and later perform an ordinary * unlock operation, such as vm_map_unlock(), before completing its * operation on the map. */ int _vm_map_unlock_and_wait(vm_map_t map, int timo, const char *file, int line) { mtx_lock(&map_sleep_mtx); if (map->system_map) mtx_unlock_flags_(&map->system_mtx, 0, file, line); else sx_xunlock_(&map->lock, file, line); return (msleep(&map->root, &map_sleep_mtx, PDROP | PVM, "vmmaps", timo)); } /* * vm_map_wakeup: * * Awaken any threads that have slept on the map using * vm_map_unlock_and_wait(). */ void vm_map_wakeup(vm_map_t map) { /* * Acquire and release map_sleep_mtx to prevent a wakeup() * from being performed (and lost) between the map unlock * and the msleep() in _vm_map_unlock_and_wait(). */ mtx_lock(&map_sleep_mtx); mtx_unlock(&map_sleep_mtx); wakeup(&map->root); } void vm_map_busy(vm_map_t map) { VM_MAP_ASSERT_LOCKED(map); map->busy++; } void vm_map_unbusy(vm_map_t map) { VM_MAP_ASSERT_LOCKED(map); KASSERT(map->busy, ("vm_map_unbusy: not busy")); if (--map->busy == 0 && (map->flags & MAP_BUSY_WAKEUP)) { vm_map_modflags(map, 0, MAP_BUSY_WAKEUP); wakeup(&map->busy); } } void vm_map_wait_busy(vm_map_t map) { VM_MAP_ASSERT_LOCKED(map); while (map->busy) { vm_map_modflags(map, MAP_BUSY_WAKEUP, 0); if (map->system_map) msleep(&map->busy, &map->system_mtx, 0, "mbusy", 0); else sx_sleep(&map->busy, &map->lock, 0, "mbusy", 0); } map->timestamp++; } long vmspace_resident_count(struct vmspace *vmspace) { return pmap_resident_count(vmspace_pmap(vmspace)); } /* * vm_map_create: * * Creates and returns a new empty VM map with * the given physical map structure, and having * the given lower and upper address bounds. */ vm_map_t vm_map_create(pmap_t pmap, vm_offset_t min, vm_offset_t max) { vm_map_t result; result = uma_zalloc(mapzone, M_WAITOK); CTR1(KTR_VM, "vm_map_create: %p", result); _vm_map_init(result, pmap, min, max); return (result); } /* * Initialize an existing vm_map structure * such as that in the vmspace structure. */ static void _vm_map_init(vm_map_t map, pmap_t pmap, vm_offset_t min, vm_offset_t max) { map->header.next = map->header.prev = &map->header; map->header.eflags = MAP_ENTRY_HEADER; map->needs_wakeup = FALSE; map->system_map = 0; map->pmap = pmap; map->header.end = min; map->header.start = max; map->flags = 0; map->root = NULL; map->timestamp = 0; map->busy = 0; map->anon_loc = 0; } void vm_map_init(vm_map_t map, pmap_t pmap, vm_offset_t min, vm_offset_t max) { _vm_map_init(map, pmap, min, max); mtx_init(&map->system_mtx, "system map", NULL, MTX_DEF | MTX_DUPOK); sx_init(&map->lock, "user map"); } /* * vm_map_entry_dispose: [ internal use only ] * * Inverse of vm_map_entry_create. */ static void vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry) { uma_zfree(map->system_map ? kmapentzone : mapentzone, entry); } /* * vm_map_entry_create: [ internal use only ] * * Allocates a VM map entry for insertion. * No entry fields are filled in. */ static vm_map_entry_t vm_map_entry_create(vm_map_t map) { vm_map_entry_t new_entry; if (map->system_map) new_entry = uma_zalloc(kmapentzone, M_NOWAIT); else new_entry = uma_zalloc(mapentzone, M_WAITOK); if (new_entry == NULL) panic("vm_map_entry_create: kernel resources exhausted"); return (new_entry); } /* * vm_map_entry_set_behavior: * * Set the expected access behavior, either normal, random, or * sequential. */ static inline void vm_map_entry_set_behavior(vm_map_entry_t entry, u_char behavior) { entry->eflags = (entry->eflags & ~MAP_ENTRY_BEHAV_MASK) | (behavior & MAP_ENTRY_BEHAV_MASK); } /* * vm_map_entry_max_free_{left,right}: * * Compute the size of the largest free gap between two entries, * one the root of a tree and the other the ancestor of that root * that is the least or greatest ancestor found on the search path. */ static inline vm_size_t vm_map_entry_max_free_left(vm_map_entry_t root, vm_map_entry_t left_ancestor) { return (root->left != NULL ? root->left->max_free : root->start - left_ancestor->end); } static inline vm_size_t vm_map_entry_max_free_right(vm_map_entry_t root, vm_map_entry_t right_ancestor) { return (root->right != NULL ? root->right->max_free : right_ancestor->start - root->end); } #define SPLAY_LEFT_STEP(root, y, rlist, test) do { \ vm_size_t max_free; \ \ /* \ * Infer root->right->max_free == root->max_free when \ * y->max_free < root->max_free || root->max_free == 0. \ * Otherwise, look right to find it. \ */ \ y = root->left; \ max_free = root->max_free; \ KASSERT(max_free >= vm_map_entry_max_free_right(root, rlist), \ ("%s: max_free invariant fails", __func__)); \ if (y == NULL ? max_free > 0 : max_free - 1 < y->max_free) \ max_free = vm_map_entry_max_free_right(root, rlist); \ if (y != NULL && (test)) { \ /* Rotate right and make y root. */ \ root->left = y->right; \ y->right = root; \ if (max_free < y->max_free) \ root->max_free = max_free = MAX(max_free, \ vm_map_entry_max_free_left(root, y)); \ root = y; \ y = root->left; \ } \ /* Copy right->max_free. Put root on rlist. */ \ root->max_free = max_free; \ KASSERT(max_free == vm_map_entry_max_free_right(root, rlist), \ ("%s: max_free not copied from right", __func__)); \ root->left = rlist; \ rlist = root; \ root = y; \ } while (0) #define SPLAY_RIGHT_STEP(root, y, llist, test) do { \ vm_size_t max_free; \ \ /* \ * Infer root->left->max_free == root->max_free when \ * y->max_free < root->max_free || root->max_free == 0. \ * Otherwise, look left to find it. \ */ \ y = root->right; \ max_free = root->max_free; \ KASSERT(max_free >= vm_map_entry_max_free_left(root, llist), \ ("%s: max_free invariant fails", __func__)); \ if (y == NULL ? max_free > 0 : max_free - 1 < y->max_free) \ max_free = vm_map_entry_max_free_left(root, llist); \ if (y != NULL && (test)) { \ /* Rotate left and make y root. */ \ root->right = y->left; \ y->left = root; \ if (max_free < y->max_free) \ root->max_free = max_free = MAX(max_free, \ vm_map_entry_max_free_right(root, y)); \ root = y; \ y = root->right; \ } \ /* Copy left->max_free. Put root on llist. */ \ root->max_free = max_free; \ KASSERT(max_free == vm_map_entry_max_free_left(root, llist), \ ("%s: max_free not copied from left", __func__)); \ root->right = llist; \ llist = root; \ root = y; \ } while (0) /* * Walk down the tree until we find addr or a NULL pointer where addr would go, * breaking off left and right subtrees of nodes less than, or greater than * addr. Treat pointers to nodes with max_free < length as NULL pointers. * llist and rlist are the two sides in reverse order (bottom-up), with llist * linked by the right pointer and rlist linked by the left pointer in the * vm_map_entry, and both lists terminated by &map->header. This function, and * the subsequent call to vm_map_splay_merge, rely on the start and end address * values in &map->header. */ static vm_map_entry_t vm_map_splay_split(vm_map_t map, vm_offset_t addr, vm_size_t length, vm_map_entry_t *out_llist, vm_map_entry_t *out_rlist) { vm_map_entry_t llist, rlist, root, y; llist = rlist = &map->header; root = map->root; while (root != NULL && root->max_free >= length) { KASSERT(llist->end <= root->start && root->end <= rlist->start, ("%s: root not within tree bounds", __func__)); if (addr < root->start) { SPLAY_LEFT_STEP(root, y, rlist, y->max_free >= length && addr < y->start); } else if (addr >= root->end) { SPLAY_RIGHT_STEP(root, y, llist, y->max_free >= length && addr >= y->end); } else break; } *out_llist = llist; *out_rlist = rlist; return (root); } static void vm_map_splay_findnext(vm_map_entry_t root, vm_map_entry_t *iolist) { vm_map_entry_t rlist, y; root = root->right; rlist = *iolist; while (root != NULL) SPLAY_LEFT_STEP(root, y, rlist, true); *iolist = rlist; } static void vm_map_splay_findprev(vm_map_entry_t root, vm_map_entry_t *iolist) { vm_map_entry_t llist, y; root = root->left; llist = *iolist; while (root != NULL) SPLAY_RIGHT_STEP(root, y, llist, true); *iolist = llist; } static inline void vm_map_entry_swap(vm_map_entry_t *a, vm_map_entry_t *b) { vm_map_entry_t tmp; tmp = *b; *b = *a; *a = tmp; } /* * Walk back up the two spines, flip the pointers and set max_free. The * subtrees of the root go at the bottom of llist and rlist. */ static void vm_map_splay_merge(vm_map_t map, vm_map_entry_t root, vm_map_entry_t llist, vm_map_entry_t rlist) { vm_map_entry_t prev; vm_size_t max_free_left, max_free_right; max_free_left = vm_map_entry_max_free_left(root, llist); if (llist != &map->header) { prev = root->left; do { /* * The max_free values of the children of llist are in * llist->max_free and max_free_left. Update with the * max value. */ llist->max_free = max_free_left = MAX(llist->max_free, max_free_left); vm_map_entry_swap(&llist->right, &prev); vm_map_entry_swap(&prev, &llist); } while (llist != &map->header); root->left = prev; } max_free_right = vm_map_entry_max_free_right(root, rlist); if (rlist != &map->header) { prev = root->right; do { /* * The max_free values of the children of rlist are in * rlist->max_free and max_free_right. Update with the * max value. */ rlist->max_free = max_free_right = MAX(rlist->max_free, max_free_right); vm_map_entry_swap(&rlist->left, &prev); vm_map_entry_swap(&prev, &rlist); } while (rlist != &map->header); root->right = prev; } root->max_free = MAX(max_free_left, max_free_right); map->root = root; } /* * vm_map_splay: * * The Sleator and Tarjan top-down splay algorithm with the * following variation. Max_free must be computed bottom-up, so * on the downward pass, maintain the left and right spines in * reverse order. Then, make a second pass up each side to fix * the pointers and compute max_free. The time bound is O(log n) * amortized. * * The new root is the vm_map_entry containing "addr", or else an * adjacent entry (lower if possible) if addr is not in the tree. * * The map must be locked, and leaves it so. * * Returns: the new root. */ static vm_map_entry_t vm_map_splay(vm_map_t map, vm_offset_t addr) { vm_map_entry_t llist, rlist, root; root = vm_map_splay_split(map, addr, 0, &llist, &rlist); if (root != NULL) { /* do nothing */ } else if (llist != &map->header) { /* * Recover the greatest node in the left * subtree and make it the root. */ root = llist; llist = root->right; root->right = NULL; } else if (rlist != &map->header) { /* * Recover the least node in the right * subtree and make it the root. */ root = rlist; rlist = root->left; root->left = NULL; } else { /* There is no root. */ return (NULL); } vm_map_splay_merge(map, root, llist, rlist); VM_MAP_ASSERT_CONSISTENT(map); return (root); } /* * vm_map_entry_{un,}link: * * Insert/remove entries from maps. */ static void vm_map_entry_link(vm_map_t map, vm_map_entry_t entry) { vm_map_entry_t llist, rlist, root; CTR3(KTR_VM, "vm_map_entry_link: map %p, nentries %d, entry %p", map, map->nentries, entry); VM_MAP_ASSERT_LOCKED(map); map->nentries++; root = vm_map_splay_split(map, entry->start, 0, &llist, &rlist); KASSERT(root == NULL, ("vm_map_entry_link: link object already mapped")); entry->prev = llist; entry->next = rlist; llist->next = rlist->prev = entry; entry->left = entry->right = NULL; vm_map_splay_merge(map, entry, llist, rlist); VM_MAP_ASSERT_CONSISTENT(map); } enum unlink_merge_type { UNLINK_MERGE_NONE, UNLINK_MERGE_NEXT }; static void vm_map_entry_unlink(vm_map_t map, vm_map_entry_t entry, enum unlink_merge_type op) { vm_map_entry_t llist, rlist, root, y; VM_MAP_ASSERT_LOCKED(map); root = vm_map_splay_split(map, entry->start, 0, &llist, &rlist); KASSERT(root != NULL, ("vm_map_entry_unlink: unlink object not mapped")); vm_map_splay_findnext(root, &rlist); switch (op) { case UNLINK_MERGE_NEXT: rlist->start = root->start; rlist->offset = root->offset; y = root->left; root = rlist; rlist = root->left; root->left = y; break; case UNLINK_MERGE_NONE: vm_map_splay_findprev(root, &llist); if (llist != &map->header) { root = llist; llist = root->right; root->right = NULL; } else if (rlist != &map->header) { root = rlist; rlist = root->left; root->left = NULL; } else root = NULL; break; } y = entry->next; y->prev = entry->prev; y->prev->next = y; if (root != NULL) vm_map_splay_merge(map, root, llist, rlist); else map->root = NULL; VM_MAP_ASSERT_CONSISTENT(map); map->nentries--; CTR3(KTR_VM, "vm_map_entry_unlink: map %p, nentries %d, entry %p", map, map->nentries, entry); } /* * vm_map_entry_resize: * * Resize a vm_map_entry, recompute the amount of free space that * follows it and propagate that value up the tree. * * The map must be locked, and leaves it so. */ static void vm_map_entry_resize(vm_map_t map, vm_map_entry_t entry, vm_size_t grow_amount) { vm_map_entry_t llist, rlist, root; VM_MAP_ASSERT_LOCKED(map); root = vm_map_splay_split(map, entry->start, 0, &llist, &rlist); KASSERT(root != NULL, ("%s: resize object not mapped", __func__)); vm_map_splay_findnext(root, &rlist); root->right = NULL; entry->end += grow_amount; vm_map_splay_merge(map, root, llist, rlist); VM_MAP_ASSERT_CONSISTENT(map); CTR4(KTR_VM, "%s: map %p, nentries %d, entry %p", __func__, map, map->nentries, entry); } /* * vm_map_lookup_entry: [ internal use only ] * * Finds the map entry containing (or * immediately preceding) the specified address * in the given map; the entry is returned * in the "entry" parameter. The boolean * result indicates whether the address is * actually contained in the map. */ boolean_t vm_map_lookup_entry( vm_map_t map, vm_offset_t address, vm_map_entry_t *entry) /* OUT */ { vm_map_entry_t cur, lbound; boolean_t locked; /* * If the map is empty, then the map entry immediately preceding * "address" is the map's header. */ cur = map->root; if (cur == NULL) { *entry = &map->header; return (FALSE); } if (address >= cur->start && cur->end > address) { *entry = cur; return (TRUE); } if ((locked = vm_map_locked(map)) || sx_try_upgrade(&map->lock)) { /* * Splay requires a write lock on the map. However, it only * restructures the binary search tree; it does not otherwise * change the map. Thus, the map's timestamp need not change * on a temporary upgrade. */ cur = vm_map_splay(map, address); if (!locked) sx_downgrade(&map->lock); /* * If "address" is contained within a map entry, the new root * is that map entry. Otherwise, the new root is a map entry * immediately before or after "address". */ if (address < cur->start) { *entry = &map->header; return (FALSE); } *entry = cur; return (address < cur->end); } /* * Since the map is only locked for read access, perform a * standard binary search tree lookup for "address". */ lbound = &map->header; do { if (address < cur->start) { cur = cur->left; } else if (cur->end <= address) { lbound = cur; cur = cur->right; } else { *entry = cur; return (TRUE); } } while (cur != NULL); *entry = lbound; return (FALSE); } /* * vm_map_insert: * * Inserts the given whole VM object into the target * map at the specified address range. The object's * size should match that of the address range. * * Requires that the map be locked, and leaves it so. * * If object is non-NULL, ref count must be bumped by caller * prior to making call to account for the new entry. */ int vm_map_insert(vm_map_t map, vm_object_t object, vm_ooffset_t offset, vm_offset_t start, vm_offset_t end, vm_prot_t prot, vm_prot_t max, int cow) { vm_map_entry_t new_entry, prev_entry; struct ucred *cred; vm_eflags_t protoeflags; vm_inherit_t inheritance; VM_MAP_ASSERT_LOCKED(map); KASSERT(object != kernel_object || (cow & MAP_COPY_ON_WRITE) == 0, ("vm_map_insert: kernel object and COW")); KASSERT(object == NULL || (cow & MAP_NOFAULT) == 0, ("vm_map_insert: paradoxical MAP_NOFAULT request")); KASSERT((prot & ~max) == 0, ("prot %#x is not subset of max_prot %#x", prot, max)); /* * Check that the start and end points are not bogus. */ if (start < vm_map_min(map) || end > vm_map_max(map) || start >= end) return (KERN_INVALID_ADDRESS); /* * Find the entry prior to the proposed starting address; if it's part * of an existing entry, this range is bogus. */ if (vm_map_lookup_entry(map, start, &prev_entry)) return (KERN_NO_SPACE); /* * Assert that the next entry doesn't overlap the end point. */ if (prev_entry->next->start < end) return (KERN_NO_SPACE); if ((cow & MAP_CREATE_GUARD) != 0 && (object != NULL || max != VM_PROT_NONE)) return (KERN_INVALID_ARGUMENT); protoeflags = 0; if (cow & MAP_COPY_ON_WRITE) protoeflags |= MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY; if (cow & MAP_NOFAULT) protoeflags |= MAP_ENTRY_NOFAULT; if (cow & MAP_DISABLE_SYNCER) protoeflags |= MAP_ENTRY_NOSYNC; if (cow & MAP_DISABLE_COREDUMP) protoeflags |= MAP_ENTRY_NOCOREDUMP; if (cow & MAP_STACK_GROWS_DOWN) protoeflags |= MAP_ENTRY_GROWS_DOWN; if (cow & MAP_STACK_GROWS_UP) protoeflags |= MAP_ENTRY_GROWS_UP; if (cow & MAP_WRITECOUNT) protoeflags |= MAP_ENTRY_WRITECNT; if (cow & MAP_VN_EXEC) protoeflags |= MAP_ENTRY_VN_EXEC; if ((cow & MAP_CREATE_GUARD) != 0) protoeflags |= MAP_ENTRY_GUARD; if ((cow & MAP_CREATE_STACK_GAP_DN) != 0) protoeflags |= MAP_ENTRY_STACK_GAP_DN; if ((cow & MAP_CREATE_STACK_GAP_UP) != 0) protoeflags |= MAP_ENTRY_STACK_GAP_UP; if (cow & MAP_INHERIT_SHARE) inheritance = VM_INHERIT_SHARE; else inheritance = VM_INHERIT_DEFAULT; cred = NULL; if ((cow & (MAP_ACC_NO_CHARGE | MAP_NOFAULT | MAP_CREATE_GUARD)) != 0) goto charged; if ((cow & MAP_ACC_CHARGED) || ((prot & VM_PROT_WRITE) && ((protoeflags & MAP_ENTRY_NEEDS_COPY) || object == NULL))) { if (!(cow & MAP_ACC_CHARGED) && !swap_reserve(end - start)) return (KERN_RESOURCE_SHORTAGE); KASSERT(object == NULL || (protoeflags & MAP_ENTRY_NEEDS_COPY) != 0 || object->cred == NULL, ("overcommit: vm_map_insert o %p", object)); cred = curthread->td_ucred; } charged: /* Expand the kernel pmap, if necessary. */ if (map == kernel_map && end > kernel_vm_end) pmap_growkernel(end); if (object != NULL) { /* * OBJ_ONEMAPPING must be cleared unless this mapping * is trivially proven to be the only mapping for any * of the object's pages. (Object granularity * reference counting is insufficient to recognize * aliases with precision.) */ VM_OBJECT_WLOCK(object); if (object->ref_count > 1 || object->shadow_count != 0) vm_object_clear_flag(object, OBJ_ONEMAPPING); VM_OBJECT_WUNLOCK(object); } else if ((prev_entry->eflags & ~MAP_ENTRY_USER_WIRED) == protoeflags && (cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP | MAP_VN_EXEC)) == 0 && prev_entry->end == start && (prev_entry->cred == cred || (prev_entry->object.vm_object != NULL && prev_entry->object.vm_object->cred == cred)) && vm_object_coalesce(prev_entry->object.vm_object, prev_entry->offset, (vm_size_t)(prev_entry->end - prev_entry->start), (vm_size_t)(end - prev_entry->end), cred != NULL && (protoeflags & MAP_ENTRY_NEEDS_COPY) == 0)) { /* * We were able to extend the object. Determine if we * can extend the previous map entry to include the * new range as well. */ if (prev_entry->inheritance == inheritance && prev_entry->protection == prot && prev_entry->max_protection == max && prev_entry->wired_count == 0) { KASSERT((prev_entry->eflags & MAP_ENTRY_USER_WIRED) == 0, ("prev_entry %p has incoherent wiring", prev_entry)); if ((prev_entry->eflags & MAP_ENTRY_GUARD) == 0) map->size += end - prev_entry->end; vm_map_entry_resize(map, prev_entry, end - prev_entry->end); vm_map_try_merge_entries(map, prev_entry, prev_entry->next); return (KERN_SUCCESS); } /* * If we can extend the object but cannot extend the * map entry, we have to create a new map entry. We * must bump the ref count on the extended object to * account for it. object may be NULL. */ object = prev_entry->object.vm_object; offset = prev_entry->offset + (prev_entry->end - prev_entry->start); vm_object_reference(object); if (cred != NULL && object != NULL && object->cred != NULL && !(prev_entry->eflags & MAP_ENTRY_NEEDS_COPY)) { /* Object already accounts for this uid. */ cred = NULL; } } if (cred != NULL) crhold(cred); /* * Create a new entry */ new_entry = vm_map_entry_create(map); new_entry->start = start; new_entry->end = end; new_entry->cred = NULL; new_entry->eflags = protoeflags; new_entry->object.vm_object = object; new_entry->offset = offset; new_entry->inheritance = inheritance; new_entry->protection = prot; new_entry->max_protection = max; new_entry->wired_count = 0; new_entry->wiring_thread = NULL; new_entry->read_ahead = VM_FAULT_READ_AHEAD_INIT; new_entry->next_read = start; KASSERT(cred == NULL || !ENTRY_CHARGED(new_entry), ("overcommit: vm_map_insert leaks vm_map %p", new_entry)); new_entry->cred = cred; /* * Insert the new entry into the list */ vm_map_entry_link(map, new_entry); if ((new_entry->eflags & MAP_ENTRY_GUARD) == 0) map->size += new_entry->end - new_entry->start; /* * Try to coalesce the new entry with both the previous and next * entries in the list. Previously, we only attempted to coalesce * with the previous entry when object is NULL. Here, we handle the * other cases, which are less common. */ vm_map_try_merge_entries(map, prev_entry, new_entry); vm_map_try_merge_entries(map, new_entry, new_entry->next); if ((cow & (MAP_PREFAULT | MAP_PREFAULT_PARTIAL)) != 0) { vm_map_pmap_enter(map, start, prot, object, OFF_TO_IDX(offset), end - start, cow & MAP_PREFAULT_PARTIAL); } return (KERN_SUCCESS); } /* * vm_map_findspace: * * Find the first fit (lowest VM address) for "length" free bytes * beginning at address >= start in the given map. * * In a vm_map_entry, "max_free" is the maximum amount of * contiguous free space between an entry in its subtree and a * neighbor of that entry. This allows finding a free region in * one path down the tree, so O(log n) amortized with splay * trees. * * The map must be locked, and leaves it so. * * Returns: starting address if sufficient space, * vm_map_max(map)-length+1 if insufficient space. */ vm_offset_t vm_map_findspace(vm_map_t map, vm_offset_t start, vm_size_t length) { vm_map_entry_t llist, rlist, root, y; vm_size_t left_length; vm_offset_t gap_end; /* * Request must fit within min/max VM address and must avoid * address wrap. */ start = MAX(start, vm_map_min(map)); if (start >= vm_map_max(map) || length > vm_map_max(map) - start) return (vm_map_max(map) - length + 1); /* Empty tree means wide open address space. */ if (map->root == NULL) return (start); /* * After splay_split, if start is within an entry, push it to the start * of the following gap. If rlist is at the end of the gap containing * start, save the end of that gap in gap_end to see if the gap is big * enough; otherwise set gap_end to start skip gap-checking and move * directly to a search of the right subtree. */ root = vm_map_splay_split(map, start, length, &llist, &rlist); gap_end = rlist->start; if (root != NULL) { start = root->end; if (root->right != NULL) gap_end = start; } else if (rlist != &map->header) { root = rlist; rlist = root->left; root->left = NULL; } else { root = llist; llist = root->right; root->right = NULL; } vm_map_splay_merge(map, root, llist, rlist); VM_MAP_ASSERT_CONSISTENT(map); if (length <= gap_end - start) return (start); /* With max_free, can immediately tell if no solution. */ if (root->right == NULL || length > root->right->max_free) return (vm_map_max(map) - length + 1); /* * Splay for the least large-enough gap in the right subtree. */ llist = rlist = &map->header; for (left_length = 0;; left_length = vm_map_entry_max_free_left(root, llist)) { if (length <= left_length) SPLAY_LEFT_STEP(root, y, rlist, length <= vm_map_entry_max_free_left(y, llist)); else SPLAY_RIGHT_STEP(root, y, llist, length > vm_map_entry_max_free_left(y, root)); if (root == NULL) break; } root = llist; llist = root->right; root->right = NULL; if (rlist != &map->header) { y = rlist; rlist = y->left; y->left = NULL; vm_map_splay_merge(map, y, &map->header, rlist); y->max_free = MAX( vm_map_entry_max_free_left(y, root), vm_map_entry_max_free_right(y, &map->header)); root->right = y; } vm_map_splay_merge(map, root, llist, &map->header); VM_MAP_ASSERT_CONSISTENT(map); return (root->end); } int vm_map_fixed(vm_map_t map, vm_object_t object, vm_ooffset_t offset, vm_offset_t start, vm_size_t length, vm_prot_t prot, vm_prot_t max, int cow) { vm_offset_t end; int result; end = start + length; KASSERT((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) == 0 || object == NULL, ("vm_map_fixed: non-NULL backing object for stack")); vm_map_lock(map); VM_MAP_RANGE_CHECK(map, start, end); if ((cow & MAP_CHECK_EXCL) == 0) vm_map_delete(map, start, end); if ((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) != 0) { result = vm_map_stack_locked(map, start, length, sgrowsiz, prot, max, cow); } else { result = vm_map_insert(map, object, offset, start, end, prot, max, cow); } vm_map_unlock(map); return (result); } static const int aslr_pages_rnd_64[2] = {0x1000, 0x10}; static const int aslr_pages_rnd_32[2] = {0x100, 0x4}; static int cluster_anon = 1; SYSCTL_INT(_vm, OID_AUTO, cluster_anon, CTLFLAG_RW, &cluster_anon, 0, "Cluster anonymous mappings: 0 = no, 1 = yes if no hint, 2 = always"); static bool clustering_anon_allowed(vm_offset_t addr) { switch (cluster_anon) { case 0: return (false); case 1: return (addr == 0); case 2: default: return (true); } } static long aslr_restarts; SYSCTL_LONG(_vm, OID_AUTO, aslr_restarts, CTLFLAG_RD, &aslr_restarts, 0, "Number of aslr failures"); #define MAP_32BIT_MAX_ADDR ((vm_offset_t)1 << 31) /* * Searches for the specified amount of free space in the given map with the * specified alignment. Performs an address-ordered, first-fit search from * the given address "*addr", with an optional upper bound "max_addr". If the * parameter "alignment" is zero, then the alignment is computed from the * given (object, offset) pair so as to enable the greatest possible use of * superpage mappings. Returns KERN_SUCCESS and the address of the free space * in "*addr" if successful. Otherwise, returns KERN_NO_SPACE. * * The map must be locked. Initially, there must be at least "length" bytes * of free space at the given address. */ static int vm_map_alignspace(vm_map_t map, vm_object_t object, vm_ooffset_t offset, vm_offset_t *addr, vm_size_t length, vm_offset_t max_addr, vm_offset_t alignment) { vm_offset_t aligned_addr, free_addr; VM_MAP_ASSERT_LOCKED(map); free_addr = *addr; KASSERT(free_addr == vm_map_findspace(map, free_addr, length), ("caller failed to provide space %#jx at address %p", (uintmax_t)length, (void *)free_addr)); for (;;) { /* * At the start of every iteration, the free space at address * "*addr" is at least "length" bytes. */ if (alignment == 0) pmap_align_superpage(object, offset, addr, length); else if ((*addr & (alignment - 1)) != 0) { *addr &= ~(alignment - 1); *addr += alignment; } aligned_addr = *addr; if (aligned_addr == free_addr) { /* * Alignment did not change "*addr", so "*addr" must * still provide sufficient free space. */ return (KERN_SUCCESS); } /* * Test for address wrap on "*addr". A wrapped "*addr" could * be a valid address, in which case vm_map_findspace() cannot * be relied upon to fail. */ if (aligned_addr < free_addr) return (KERN_NO_SPACE); *addr = vm_map_findspace(map, aligned_addr, length); if (*addr + length > vm_map_max(map) || (max_addr != 0 && *addr + length > max_addr)) return (KERN_NO_SPACE); free_addr = *addr; if (free_addr == aligned_addr) { /* * If a successful call to vm_map_findspace() did not * change "*addr", then "*addr" must still be aligned * and provide sufficient free space. */ return (KERN_SUCCESS); } } } /* * vm_map_find finds an unallocated region in the target address * map with the given length. The search is defined to be * first-fit from the specified address; the region found is * returned in the same parameter. * * If object is non-NULL, ref count must be bumped by caller * prior to making call to account for the new entry. */ int vm_map_find(vm_map_t map, vm_object_t object, vm_ooffset_t offset, vm_offset_t *addr, /* IN/OUT */ vm_size_t length, vm_offset_t max_addr, int find_space, vm_prot_t prot, vm_prot_t max, int cow) { vm_offset_t alignment, curr_min_addr, min_addr; int gap, pidx, rv, try; bool cluster, en_aslr, update_anon; KASSERT((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) == 0 || object == NULL, ("vm_map_find: non-NULL backing object for stack")); MPASS((cow & MAP_REMAP) == 0 || (find_space == VMFS_NO_SPACE && (cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) == 0)); if (find_space == VMFS_OPTIMAL_SPACE && (object == NULL || (object->flags & OBJ_COLORED) == 0)) find_space = VMFS_ANY_SPACE; if (find_space >> 8 != 0) { KASSERT((find_space & 0xff) == 0, ("bad VMFS flags")); alignment = (vm_offset_t)1 << (find_space >> 8); } else alignment = 0; en_aslr = (map->flags & MAP_ASLR) != 0; update_anon = cluster = clustering_anon_allowed(*addr) && (map->flags & MAP_IS_SUB_MAP) == 0 && max_addr == 0 && find_space != VMFS_NO_SPACE && object == NULL && (cow & (MAP_INHERIT_SHARE | MAP_STACK_GROWS_UP | MAP_STACK_GROWS_DOWN)) == 0 && prot != PROT_NONE; curr_min_addr = min_addr = *addr; if (en_aslr && min_addr == 0 && !cluster && find_space != VMFS_NO_SPACE && (map->flags & MAP_ASLR_IGNSTART) != 0) curr_min_addr = min_addr = vm_map_min(map); try = 0; vm_map_lock(map); if (cluster) { curr_min_addr = map->anon_loc; if (curr_min_addr == 0) cluster = false; } if (find_space != VMFS_NO_SPACE) { KASSERT(find_space == VMFS_ANY_SPACE || find_space == VMFS_OPTIMAL_SPACE || find_space == VMFS_SUPER_SPACE || alignment != 0, ("unexpected VMFS flag")); again: /* * When creating an anonymous mapping, try clustering * with an existing anonymous mapping first. * * We make up to two attempts to find address space * for a given find_space value. The first attempt may * apply randomization or may cluster with an existing * anonymous mapping. If this first attempt fails, * perform a first-fit search of the available address * space. * * If all tries failed, and find_space is * VMFS_OPTIMAL_SPACE, fallback to VMFS_ANY_SPACE. * Again enable clustering and randomization. */ try++; MPASS(try <= 2); if (try == 2) { /* * Second try: we failed either to find a * suitable region for randomizing the * allocation, or to cluster with an existing * mapping. Retry with free run. */ curr_min_addr = (map->flags & MAP_ASLR_IGNSTART) != 0 ? vm_map_min(map) : min_addr; atomic_add_long(&aslr_restarts, 1); } if (try == 1 && en_aslr && !cluster) { /* * Find space for allocation, including * gap needed for later randomization. */ pidx = MAXPAGESIZES > 1 && pagesizes[1] != 0 && (find_space == VMFS_SUPER_SPACE || find_space == VMFS_OPTIMAL_SPACE) ? 1 : 0; gap = vm_map_max(map) > MAP_32BIT_MAX_ADDR && (max_addr == 0 || max_addr > MAP_32BIT_MAX_ADDR) ? aslr_pages_rnd_64[pidx] : aslr_pages_rnd_32[pidx]; *addr = vm_map_findspace(map, curr_min_addr, length + gap * pagesizes[pidx]); if (*addr + length + gap * pagesizes[pidx] > vm_map_max(map)) goto again; /* And randomize the start address. */ *addr += (arc4random() % gap) * pagesizes[pidx]; if (max_addr != 0 && *addr + length > max_addr) goto again; } else { *addr = vm_map_findspace(map, curr_min_addr, length); if (*addr + length > vm_map_max(map) || (max_addr != 0 && *addr + length > max_addr)) { if (cluster) { cluster = false; MPASS(try == 1); goto again; } rv = KERN_NO_SPACE; goto done; } } if (find_space != VMFS_ANY_SPACE && (rv = vm_map_alignspace(map, object, offset, addr, length, max_addr, alignment)) != KERN_SUCCESS) { if (find_space == VMFS_OPTIMAL_SPACE) { find_space = VMFS_ANY_SPACE; curr_min_addr = min_addr; cluster = update_anon; try = 0; goto again; } goto done; } } else if ((cow & MAP_REMAP) != 0) { if (*addr < vm_map_min(map) || *addr + length > vm_map_max(map) || *addr + length <= length) { rv = KERN_INVALID_ADDRESS; goto done; } vm_map_delete(map, *addr, *addr + length); } if ((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) != 0) { rv = vm_map_stack_locked(map, *addr, length, sgrowsiz, prot, max, cow); } else { rv = vm_map_insert(map, object, offset, *addr, *addr + length, prot, max, cow); } if (rv == KERN_SUCCESS && update_anon) map->anon_loc = *addr + length; done: vm_map_unlock(map); return (rv); } /* * vm_map_find_min() is a variant of vm_map_find() that takes an * additional parameter (min_addr) and treats the given address * (*addr) differently. Specifically, it treats *addr as a hint * and not as the minimum address where the mapping is created. * * This function works in two phases. First, it tries to * allocate above the hint. If that fails and the hint is * greater than min_addr, it performs a second pass, replacing * the hint with min_addr as the minimum address for the * allocation. */ int vm_map_find_min(vm_map_t map, vm_object_t object, vm_ooffset_t offset, vm_offset_t *addr, vm_size_t length, vm_offset_t min_addr, vm_offset_t max_addr, int find_space, vm_prot_t prot, vm_prot_t max, int cow) { vm_offset_t hint; int rv; hint = *addr; for (;;) { rv = vm_map_find(map, object, offset, addr, length, max_addr, find_space, prot, max, cow); if (rv == KERN_SUCCESS || min_addr >= hint) return (rv); *addr = hint = min_addr; } } /* * A map entry with any of the following flags set must not be merged with * another entry. */ #define MAP_ENTRY_NOMERGE_MASK (MAP_ENTRY_GROWS_DOWN | MAP_ENTRY_GROWS_UP | \ MAP_ENTRY_IN_TRANSITION | MAP_ENTRY_IS_SUB_MAP | MAP_ENTRY_VN_EXEC) static bool vm_map_mergeable_neighbors(vm_map_entry_t prev, vm_map_entry_t entry) { KASSERT((prev->eflags & MAP_ENTRY_NOMERGE_MASK) == 0 || (entry->eflags & MAP_ENTRY_NOMERGE_MASK) == 0, ("vm_map_mergeable_neighbors: neither %p nor %p are mergeable", prev, entry)); return (prev->end == entry->start && prev->object.vm_object == entry->object.vm_object && (prev->object.vm_object == NULL || prev->offset + (prev->end - prev->start) == entry->offset) && prev->eflags == entry->eflags && prev->protection == entry->protection && prev->max_protection == entry->max_protection && prev->inheritance == entry->inheritance && prev->wired_count == entry->wired_count && prev->cred == entry->cred); } static void vm_map_merged_neighbor_dispose(vm_map_t map, vm_map_entry_t entry) { /* * If the backing object is a vnode object, vm_object_deallocate() * calls vrele(). However, vrele() does not lock the vnode because * the vnode has additional references. Thus, the map lock can be * kept without causing a lock-order reversal with the vnode lock. * * Since we count the number of virtual page mappings in * object->un_pager.vnp.writemappings, the writemappings value * should not be adjusted when the entry is disposed of. */ if (entry->object.vm_object != NULL) vm_object_deallocate(entry->object.vm_object); if (entry->cred != NULL) crfree(entry->cred); vm_map_entry_dispose(map, entry); } /* * vm_map_try_merge_entries: * * Compare the given map entry to its predecessor, and merge its precessor * into it if possible. The entry remains valid, and may be extended. * The predecessor may be deleted. * * The map must be locked. */ void vm_map_try_merge_entries(vm_map_t map, vm_map_entry_t prev, vm_map_entry_t entry) { VM_MAP_ASSERT_LOCKED(map); if ((entry->eflags & MAP_ENTRY_NOMERGE_MASK) == 0 && vm_map_mergeable_neighbors(prev, entry)) { vm_map_entry_unlink(map, prev, UNLINK_MERGE_NEXT); vm_map_merged_neighbor_dispose(map, prev); } } /* * vm_map_entry_back: * * Allocate an object to back a map entry. */ static inline void vm_map_entry_back(vm_map_entry_t entry) { vm_object_t object; KASSERT(entry->object.vm_object == NULL, ("map entry %p has backing object", entry)); KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0, ("map entry %p is a submap", entry)); object = vm_object_allocate(OBJT_DEFAULT, atop(entry->end - entry->start)); entry->object.vm_object = object; entry->offset = 0; if (entry->cred != NULL) { object->cred = entry->cred; object->charge = entry->end - entry->start; entry->cred = NULL; } } /* * vm_map_entry_charge_object * * If there is no object backing this entry, create one. Otherwise, if * the entry has cred, give it to the backing object. */ static inline void vm_map_entry_charge_object(vm_map_t map, vm_map_entry_t entry) { VM_MAP_ASSERT_LOCKED(map); KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0, ("map entry %p is a submap", entry)); if (entry->object.vm_object == NULL && !map->system_map && (entry->eflags & MAP_ENTRY_GUARD) == 0) vm_map_entry_back(entry); else if (entry->object.vm_object != NULL && ((entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) && entry->cred != NULL) { VM_OBJECT_WLOCK(entry->object.vm_object); KASSERT(entry->object.vm_object->cred == NULL, ("OVERCOMMIT: %s: both cred e %p", __func__, entry)); entry->object.vm_object->cred = entry->cred; entry->object.vm_object->charge = entry->end - entry->start; VM_OBJECT_WUNLOCK(entry->object.vm_object); entry->cred = NULL; } } /* * vm_map_clip_start: [ internal use only ] * * Asserts that the given entry begins at or after * the specified address; if necessary, * it splits the entry into two. */ #define vm_map_clip_start(map, entry, startaddr) \ { \ if (startaddr > entry->start) \ _vm_map_clip_start(map, entry, startaddr); \ } /* * This routine is called only when it is known that * the entry must be split. */ static void _vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t start) { vm_map_entry_t new_entry; VM_MAP_ASSERT_LOCKED(map); KASSERT(entry->end > start && entry->start < start, ("_vm_map_clip_start: invalid clip of entry %p", entry)); /* * Create a backing object now, if none exists, so that more individual * objects won't be created after the map entry is split. */ vm_map_entry_charge_object(map, entry); /* Clone the entry. */ new_entry = vm_map_entry_create(map); *new_entry = *entry; /* * Split off the front portion. Insert the new entry BEFORE this one, * so that this entry has the specified starting address. */ new_entry->end = start; entry->offset += (start - entry->start); entry->start = start; if (new_entry->cred != NULL) crhold(entry->cred); vm_map_entry_link(map, new_entry); if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { vm_object_reference(new_entry->object.vm_object); vm_map_entry_set_vnode_text(new_entry, true); /* * The object->un_pager.vnp.writemappings for the * object of MAP_ENTRY_WRITECNT type entry shall be * kept as is here. The virtual pages are * re-distributed among the clipped entries, so the sum is * left the same. */ } } /* * vm_map_clip_end: [ internal use only ] * * Asserts that the given entry ends at or before * the specified address; if necessary, * it splits the entry into two. */ #define vm_map_clip_end(map, entry, endaddr) \ { \ if ((endaddr) < (entry->end)) \ _vm_map_clip_end((map), (entry), (endaddr)); \ } /* * This routine is called only when it is known that * the entry must be split. */ static void _vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t end) { vm_map_entry_t new_entry; VM_MAP_ASSERT_LOCKED(map); KASSERT(entry->start < end && entry->end > end, ("_vm_map_clip_end: invalid clip of entry %p", entry)); /* * Create a backing object now, if none exists, so that more individual * objects won't be created after the map entry is split. */ vm_map_entry_charge_object(map, entry); /* Clone the entry. */ new_entry = vm_map_entry_create(map); *new_entry = *entry; /* * Split off the back portion. Insert the new entry AFTER this one, * so that this entry has the specified ending address. */ new_entry->start = entry->end = end; new_entry->offset += (end - entry->start); if (new_entry->cred != NULL) crhold(entry->cred); vm_map_entry_link(map, new_entry); if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { vm_object_reference(new_entry->object.vm_object); vm_map_entry_set_vnode_text(new_entry, true); } } /* * vm_map_submap: [ kernel use only ] * * Mark the given range as handled by a subordinate map. * * This range must have been created with vm_map_find, * and no other operations may have been performed on this * range prior to calling vm_map_submap. * * Only a limited number of operations can be performed * within this rage after calling vm_map_submap: * vm_fault * [Don't try vm_map_copy!] * * To remove a submapping, one must first remove the * range from the superior map, and then destroy the * submap (if desired). [Better yet, don't try it.] */ int vm_map_submap( vm_map_t map, vm_offset_t start, vm_offset_t end, vm_map_t submap) { vm_map_entry_t entry; int result; result = KERN_INVALID_ARGUMENT; vm_map_lock(submap); submap->flags |= MAP_IS_SUB_MAP; vm_map_unlock(submap); vm_map_lock(map); VM_MAP_RANGE_CHECK(map, start, end); if (vm_map_lookup_entry(map, start, &entry)) { vm_map_clip_start(map, entry, start); } else entry = entry->next; vm_map_clip_end(map, entry, end); if ((entry->start == start) && (entry->end == end) && ((entry->eflags & MAP_ENTRY_COW) == 0) && (entry->object.vm_object == NULL)) { entry->object.sub_map = submap; entry->eflags |= MAP_ENTRY_IS_SUB_MAP; result = KERN_SUCCESS; } vm_map_unlock(map); if (result != KERN_SUCCESS) { vm_map_lock(submap); submap->flags &= ~MAP_IS_SUB_MAP; vm_map_unlock(submap); } return (result); } /* * The maximum number of pages to map if MAP_PREFAULT_PARTIAL is specified */ #define MAX_INIT_PT 96 /* * vm_map_pmap_enter: * * Preload the specified map's pmap with mappings to the specified * object's memory-resident pages. No further physical pages are * allocated, and no further virtual pages are retrieved from secondary * storage. If the specified flags include MAP_PREFAULT_PARTIAL, then a * limited number of page mappings are created at the low-end of the * specified address range. (For this purpose, a superpage mapping * counts as one page mapping.) Otherwise, all resident pages within * the specified address range are mapped. */ static void vm_map_pmap_enter(vm_map_t map, vm_offset_t addr, vm_prot_t prot, vm_object_t object, vm_pindex_t pindex, vm_size_t size, int flags) { vm_offset_t start; vm_page_t p, p_start; vm_pindex_t mask, psize, threshold, tmpidx; if ((prot & (VM_PROT_READ | VM_PROT_EXECUTE)) == 0 || object == NULL) return; VM_OBJECT_RLOCK(object); if (object->type == OBJT_DEVICE || object->type == OBJT_SG) { VM_OBJECT_RUNLOCK(object); VM_OBJECT_WLOCK(object); if (object->type == OBJT_DEVICE || object->type == OBJT_SG) { pmap_object_init_pt(map->pmap, addr, object, pindex, size); VM_OBJECT_WUNLOCK(object); return; } VM_OBJECT_LOCK_DOWNGRADE(object); } psize = atop(size); if (psize + pindex > object->size) { if (object->size < pindex) { VM_OBJECT_RUNLOCK(object); return; } psize = object->size - pindex; } start = 0; p_start = NULL; threshold = MAX_INIT_PT; p = vm_page_find_least(object, pindex); /* * Assert: the variable p is either (1) the page with the * least pindex greater than or equal to the parameter pindex * or (2) NULL. */ for (; p != NULL && (tmpidx = p->pindex - pindex) < psize; p = TAILQ_NEXT(p, listq)) { /* * don't allow an madvise to blow away our really * free pages allocating pv entries. */ if (((flags & MAP_PREFAULT_MADVISE) != 0 && vm_page_count_severe()) || ((flags & MAP_PREFAULT_PARTIAL) != 0 && tmpidx >= threshold)) { psize = tmpidx; break; } - if (p->valid == VM_PAGE_BITS_ALL) { + if (vm_page_all_valid(p)) { if (p_start == NULL) { start = addr + ptoa(tmpidx); p_start = p; } /* Jump ahead if a superpage mapping is possible. */ if (p->psind > 0 && ((addr + ptoa(tmpidx)) & (pagesizes[p->psind] - 1)) == 0) { mask = atop(pagesizes[p->psind]) - 1; if (tmpidx + mask < psize && vm_page_ps_test(p, PS_ALL_VALID, NULL)) { p += mask; threshold += mask; } } } else if (p_start != NULL) { pmap_enter_object(map->pmap, start, addr + ptoa(tmpidx), p_start, prot); p_start = NULL; } } if (p_start != NULL) pmap_enter_object(map->pmap, start, addr + ptoa(psize), p_start, prot); VM_OBJECT_RUNLOCK(object); } /* * vm_map_protect: * * Sets the protection of the specified address * region in the target map. If "set_max" is * specified, the maximum protection is to be set; * otherwise, only the current protection is affected. */ int vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end, vm_prot_t new_prot, boolean_t set_max) { vm_map_entry_t current, entry, in_tran; vm_object_t obj; struct ucred *cred; vm_prot_t old_prot; int rv; if (start == end) return (KERN_SUCCESS); again: in_tran = NULL; vm_map_lock(map); /* * Ensure that we are not concurrently wiring pages. vm_map_wire() may * need to fault pages into the map and will drop the map lock while * doing so, and the VM object may end up in an inconsistent state if we * update the protection on the map entry in between faults. */ vm_map_wait_busy(map); VM_MAP_RANGE_CHECK(map, start, end); if (!vm_map_lookup_entry(map, start, &entry)) entry = entry->next; /* * Make a first pass to check for protection violations. */ for (current = entry; current->start < end; current = current->next) { if ((current->eflags & MAP_ENTRY_GUARD) != 0) continue; if (current->eflags & MAP_ENTRY_IS_SUB_MAP) { vm_map_unlock(map); return (KERN_INVALID_ARGUMENT); } if ((new_prot & current->max_protection) != new_prot) { vm_map_unlock(map); return (KERN_PROTECTION_FAILURE); } if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0) in_tran = entry; } /* * Postpone the operation until all in transition map entries * are stabilized. In-transition entry might already have its * pages wired and wired_count incremented, but * MAP_ENTRY_USER_WIRED flag not yet set, and visible to other * threads because the map lock is dropped. In this case we * would miss our call to vm_fault_copy_entry(). */ if (in_tran != NULL) { in_tran->eflags |= MAP_ENTRY_NEEDS_WAKEUP; vm_map_unlock_and_wait(map, 0); goto again; } /* * Before changing the protections, try to reserve swap space for any * private (i.e., copy-on-write) mappings that are transitioning from * read-only to read/write access. If a reservation fails, break out * of this loop early and let the next loop simplify the entries, since * some may now be mergeable. */ rv = KERN_SUCCESS; vm_map_clip_start(map, entry, start); for (current = entry; current->start < end; current = current->next) { vm_map_clip_end(map, current, end); if (set_max || ((new_prot & ~(current->protection)) & VM_PROT_WRITE) == 0 || ENTRY_CHARGED(current) || (current->eflags & MAP_ENTRY_GUARD) != 0) { continue; } cred = curthread->td_ucred; obj = current->object.vm_object; if (obj == NULL || (current->eflags & MAP_ENTRY_NEEDS_COPY)) { if (!swap_reserve(current->end - current->start)) { rv = KERN_RESOURCE_SHORTAGE; end = current->end; break; } crhold(cred); current->cred = cred; continue; } VM_OBJECT_WLOCK(obj); if (obj->type != OBJT_DEFAULT && obj->type != OBJT_SWAP) { VM_OBJECT_WUNLOCK(obj); continue; } /* * Charge for the whole object allocation now, since * we cannot distinguish between non-charged and * charged clipped mapping of the same object later. */ KASSERT(obj->charge == 0, ("vm_map_protect: object %p overcharged (entry %p)", obj, current)); if (!swap_reserve(ptoa(obj->size))) { VM_OBJECT_WUNLOCK(obj); rv = KERN_RESOURCE_SHORTAGE; end = current->end; break; } crhold(cred); obj->cred = cred; obj->charge = ptoa(obj->size); VM_OBJECT_WUNLOCK(obj); } /* * If enough swap space was available, go back and fix up protections. * Otherwise, just simplify entries, since some may have been modified. * [Note that clipping is not necessary the second time.] */ for (current = entry; current->start < end; vm_map_try_merge_entries(map, current->prev, current), current = current->next) { if (rv != KERN_SUCCESS || (current->eflags & MAP_ENTRY_GUARD) != 0) continue; old_prot = current->protection; if (set_max) current->protection = (current->max_protection = new_prot) & old_prot; else current->protection = new_prot; /* * For user wired map entries, the normal lazy evaluation of * write access upgrades through soft page faults is * undesirable. Instead, immediately copy any pages that are * copy-on-write and enable write access in the physical map. */ if ((current->eflags & MAP_ENTRY_USER_WIRED) != 0 && (current->protection & VM_PROT_WRITE) != 0 && (old_prot & VM_PROT_WRITE) == 0) vm_fault_copy_entry(map, map, current, current, NULL); /* * When restricting access, update the physical map. Worry * about copy-on-write here. */ if ((old_prot & ~current->protection) != 0) { #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \ VM_PROT_ALL) pmap_protect(map->pmap, current->start, current->end, current->protection & MASK(current)); #undef MASK } } vm_map_try_merge_entries(map, current->prev, current); vm_map_unlock(map); return (rv); } /* * vm_map_madvise: * * This routine traverses a processes map handling the madvise * system call. Advisories are classified as either those effecting * the vm_map_entry structure, or those effecting the underlying * objects. */ int vm_map_madvise( vm_map_t map, vm_offset_t start, vm_offset_t end, int behav) { vm_map_entry_t current, entry; bool modify_map; /* * Some madvise calls directly modify the vm_map_entry, in which case * we need to use an exclusive lock on the map and we need to perform * various clipping operations. Otherwise we only need a read-lock * on the map. */ switch(behav) { case MADV_NORMAL: case MADV_SEQUENTIAL: case MADV_RANDOM: case MADV_NOSYNC: case MADV_AUTOSYNC: case MADV_NOCORE: case MADV_CORE: if (start == end) return (0); modify_map = true; vm_map_lock(map); break; case MADV_WILLNEED: case MADV_DONTNEED: case MADV_FREE: if (start == end) return (0); modify_map = false; vm_map_lock_read(map); break; default: return (EINVAL); } /* * Locate starting entry and clip if necessary. */ VM_MAP_RANGE_CHECK(map, start, end); if (vm_map_lookup_entry(map, start, &entry)) { if (modify_map) vm_map_clip_start(map, entry, start); } else { entry = entry->next; } if (modify_map) { /* * madvise behaviors that are implemented in the vm_map_entry. * * We clip the vm_map_entry so that behavioral changes are * limited to the specified address range. */ for (current = entry; current->start < end; current = current->next) { if (current->eflags & MAP_ENTRY_IS_SUB_MAP) continue; vm_map_clip_end(map, current, end); switch (behav) { case MADV_NORMAL: vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_NORMAL); break; case MADV_SEQUENTIAL: vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_SEQUENTIAL); break; case MADV_RANDOM: vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_RANDOM); break; case MADV_NOSYNC: current->eflags |= MAP_ENTRY_NOSYNC; break; case MADV_AUTOSYNC: current->eflags &= ~MAP_ENTRY_NOSYNC; break; case MADV_NOCORE: current->eflags |= MAP_ENTRY_NOCOREDUMP; break; case MADV_CORE: current->eflags &= ~MAP_ENTRY_NOCOREDUMP; break; default: break; } vm_map_try_merge_entries(map, current->prev, current); } vm_map_try_merge_entries(map, current->prev, current); vm_map_unlock(map); } else { vm_pindex_t pstart, pend; /* * madvise behaviors that are implemented in the underlying * vm_object. * * Since we don't clip the vm_map_entry, we have to clip * the vm_object pindex and count. */ for (current = entry; current->start < end; current = current->next) { vm_offset_t useEnd, useStart; if (current->eflags & MAP_ENTRY_IS_SUB_MAP) continue; /* * MADV_FREE would otherwise rewind time to * the creation of the shadow object. Because * we hold the VM map read-locked, neither the * entry's object nor the presence of a * backing object can change. */ if (behav == MADV_FREE && current->object.vm_object != NULL && current->object.vm_object->backing_object != NULL) continue; pstart = OFF_TO_IDX(current->offset); pend = pstart + atop(current->end - current->start); useStart = current->start; useEnd = current->end; if (current->start < start) { pstart += atop(start - current->start); useStart = start; } if (current->end > end) { pend -= atop(current->end - end); useEnd = end; } if (pstart >= pend) continue; /* * Perform the pmap_advise() before clearing * PGA_REFERENCED in vm_page_advise(). Otherwise, a * concurrent pmap operation, such as pmap_remove(), * could clear a reference in the pmap and set * PGA_REFERENCED on the page before the pmap_advise() * had completed. Consequently, the page would appear * referenced based upon an old reference that * occurred before this pmap_advise() ran. */ if (behav == MADV_DONTNEED || behav == MADV_FREE) pmap_advise(map->pmap, useStart, useEnd, behav); vm_object_madvise(current->object.vm_object, pstart, pend, behav); /* * Pre-populate paging structures in the * WILLNEED case. For wired entries, the * paging structures are already populated. */ if (behav == MADV_WILLNEED && current->wired_count == 0) { vm_map_pmap_enter(map, useStart, current->protection, current->object.vm_object, pstart, ptoa(pend - pstart), MAP_PREFAULT_MADVISE ); } } vm_map_unlock_read(map); } return (0); } /* * vm_map_inherit: * * Sets the inheritance of the specified address * range in the target map. Inheritance * affects how the map will be shared with * child maps at the time of vmspace_fork. */ int vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end, vm_inherit_t new_inheritance) { vm_map_entry_t entry; vm_map_entry_t temp_entry; switch (new_inheritance) { case VM_INHERIT_NONE: case VM_INHERIT_COPY: case VM_INHERIT_SHARE: case VM_INHERIT_ZERO: break; default: return (KERN_INVALID_ARGUMENT); } if (start == end) return (KERN_SUCCESS); vm_map_lock(map); VM_MAP_RANGE_CHECK(map, start, end); if (vm_map_lookup_entry(map, start, &temp_entry)) { entry = temp_entry; vm_map_clip_start(map, entry, start); } else entry = temp_entry->next; while (entry->start < end) { vm_map_clip_end(map, entry, end); if ((entry->eflags & MAP_ENTRY_GUARD) == 0 || new_inheritance != VM_INHERIT_ZERO) entry->inheritance = new_inheritance; vm_map_try_merge_entries(map, entry->prev, entry); entry = entry->next; } vm_map_try_merge_entries(map, entry->prev, entry); vm_map_unlock(map); return (KERN_SUCCESS); } /* * vm_map_entry_in_transition: * * Release the map lock, and sleep until the entry is no longer in * transition. Awake and acquire the map lock. If the map changed while * another held the lock, lookup a possibly-changed entry at or after the * 'start' position of the old entry. */ static vm_map_entry_t vm_map_entry_in_transition(vm_map_t map, vm_offset_t in_start, vm_offset_t *io_end, bool holes_ok, vm_map_entry_t in_entry) { vm_map_entry_t entry; vm_offset_t start; u_int last_timestamp; VM_MAP_ASSERT_LOCKED(map); KASSERT((in_entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0, ("not in-tranition map entry %p", in_entry)); /* * We have not yet clipped the entry. */ start = MAX(in_start, in_entry->start); in_entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; last_timestamp = map->timestamp; if (vm_map_unlock_and_wait(map, 0)) { /* * Allow interruption of user wiring/unwiring? */ } vm_map_lock(map); if (last_timestamp + 1 == map->timestamp) return (in_entry); /* * Look again for the entry because the map was modified while it was * unlocked. Specifically, the entry may have been clipped, merged, or * deleted. */ if (!vm_map_lookup_entry(map, start, &entry)) { if (!holes_ok) { *io_end = start; return (NULL); } entry = entry->next; } return (entry); } /* * vm_map_unwire: * * Implements both kernel and user unwiring. */ int vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end, int flags) { vm_map_entry_t entry, first_entry; int rv; bool first_iteration, holes_ok, need_wakeup, user_unwire; if (start == end) return (KERN_SUCCESS); holes_ok = (flags & VM_MAP_WIRE_HOLESOK) != 0; user_unwire = (flags & VM_MAP_WIRE_USER) != 0; vm_map_lock(map); VM_MAP_RANGE_CHECK(map, start, end); if (!vm_map_lookup_entry(map, start, &first_entry)) { if (holes_ok) first_entry = first_entry->next; else { vm_map_unlock(map); return (KERN_INVALID_ADDRESS); } } first_iteration = true; entry = first_entry; rv = KERN_SUCCESS; while (entry->start < end) { if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { /* * We have not yet clipped the entry. */ entry = vm_map_entry_in_transition(map, start, &end, holes_ok, entry); if (entry == NULL) { if (first_iteration) { vm_map_unlock(map); return (KERN_INVALID_ADDRESS); } rv = KERN_INVALID_ADDRESS; break; } first_entry = first_iteration ? entry : NULL; continue; } first_iteration = false; vm_map_clip_start(map, entry, start); vm_map_clip_end(map, entry, end); /* * Mark the entry in case the map lock is released. (See * above.) */ KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 && entry->wiring_thread == NULL, ("owned map entry %p", entry)); entry->eflags |= MAP_ENTRY_IN_TRANSITION; entry->wiring_thread = curthread; /* * Check the map for holes in the specified region. * If holes_ok, skip this check. */ if (!holes_ok && (entry->end < end && entry->next->start > entry->end)) { end = entry->end; rv = KERN_INVALID_ADDRESS; break; } /* * If system unwiring, require that the entry is system wired. */ if (!user_unwire && vm_map_entry_system_wired_count(entry) == 0) { end = entry->end; rv = KERN_INVALID_ARGUMENT; break; } entry = entry->next; } need_wakeup = false; if (first_entry == NULL && !vm_map_lookup_entry(map, start, &first_entry)) { KASSERT(holes_ok, ("vm_map_unwire: lookup failed")); first_entry = first_entry->next; } for (entry = first_entry; entry->start < end; entry = entry->next) { /* * If holes_ok was specified, an empty * space in the unwired region could have been mapped * while the map lock was dropped for draining * MAP_ENTRY_IN_TRANSITION. Moreover, another thread * could be simultaneously wiring this new mapping * entry. Detect these cases and skip any entries * marked as in transition by us. */ if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 || entry->wiring_thread != curthread) { KASSERT(holes_ok, ("vm_map_unwire: !HOLESOK and new/changed entry")); continue; } if (rv == KERN_SUCCESS && (!user_unwire || (entry->eflags & MAP_ENTRY_USER_WIRED))) { if (entry->wired_count == 1) vm_map_entry_unwire(map, entry); else entry->wired_count--; if (user_unwire) entry->eflags &= ~MAP_ENTRY_USER_WIRED; } KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0, ("vm_map_unwire: in-transition flag missing %p", entry)); KASSERT(entry->wiring_thread == curthread, ("vm_map_unwire: alien wire %p", entry)); entry->eflags &= ~MAP_ENTRY_IN_TRANSITION; entry->wiring_thread = NULL; if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) { entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP; need_wakeup = true; } vm_map_try_merge_entries(map, entry->prev, entry); } vm_map_try_merge_entries(map, entry->prev, entry); vm_map_unlock(map); if (need_wakeup) vm_map_wakeup(map); return (rv); } static void vm_map_wire_user_count_sub(u_long npages) { atomic_subtract_long(&vm_user_wire_count, npages); } static bool vm_map_wire_user_count_add(u_long npages) { u_long wired; wired = vm_user_wire_count; do { if (npages + wired > vm_page_max_user_wired) return (false); } while (!atomic_fcmpset_long(&vm_user_wire_count, &wired, npages + wired)); return (true); } /* * vm_map_wire_entry_failure: * * Handle a wiring failure on the given entry. * * The map should be locked. */ static void vm_map_wire_entry_failure(vm_map_t map, vm_map_entry_t entry, vm_offset_t failed_addr) { VM_MAP_ASSERT_LOCKED(map); KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0 && entry->wired_count == 1, ("vm_map_wire_entry_failure: entry %p isn't being wired", entry)); KASSERT(failed_addr < entry->end, ("vm_map_wire_entry_failure: entry %p was fully wired", entry)); /* * If any pages at the start of this entry were successfully wired, * then unwire them. */ if (failed_addr > entry->start) { pmap_unwire(map->pmap, entry->start, failed_addr); vm_object_unwire(entry->object.vm_object, entry->offset, failed_addr - entry->start, PQ_ACTIVE); } /* * Assign an out-of-range value to represent the failure to wire this * entry. */ entry->wired_count = -1; } int vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t end, int flags) { int rv; vm_map_lock(map); rv = vm_map_wire_locked(map, start, end, flags); vm_map_unlock(map); return (rv); } /* * vm_map_wire_locked: * * Implements both kernel and user wiring. Returns with the map locked, * the map lock may be dropped. */ int vm_map_wire_locked(vm_map_t map, vm_offset_t start, vm_offset_t end, int flags) { vm_map_entry_t entry, first_entry, tmp_entry; vm_offset_t faddr, saved_end, saved_start; u_long npages; u_int last_timestamp; int rv; bool first_iteration, holes_ok, need_wakeup, user_wire; vm_prot_t prot; VM_MAP_ASSERT_LOCKED(map); if (start == end) return (KERN_SUCCESS); prot = 0; if (flags & VM_MAP_WIRE_WRITE) prot |= VM_PROT_WRITE; holes_ok = (flags & VM_MAP_WIRE_HOLESOK) != 0; user_wire = (flags & VM_MAP_WIRE_USER) != 0; VM_MAP_RANGE_CHECK(map, start, end); if (!vm_map_lookup_entry(map, start, &first_entry)) { if (holes_ok) first_entry = first_entry->next; else return (KERN_INVALID_ADDRESS); } first_iteration = true; entry = first_entry; while (entry->start < end) { if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { /* * We have not yet clipped the entry. */ entry = vm_map_entry_in_transition(map, start, &end, holes_ok, entry); if (entry == NULL) { if (first_iteration) return (KERN_INVALID_ADDRESS); rv = KERN_INVALID_ADDRESS; goto done; } first_entry = first_iteration ? entry : NULL; continue; } first_iteration = false; vm_map_clip_start(map, entry, start); vm_map_clip_end(map, entry, end); /* * Mark the entry in case the map lock is released. (See * above.) */ KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 && entry->wiring_thread == NULL, ("owned map entry %p", entry)); entry->eflags |= MAP_ENTRY_IN_TRANSITION; entry->wiring_thread = curthread; if ((entry->protection & (VM_PROT_READ | VM_PROT_EXECUTE)) == 0 || (entry->protection & prot) != prot) { entry->eflags |= MAP_ENTRY_WIRE_SKIPPED; if (!holes_ok) { end = entry->end; rv = KERN_INVALID_ADDRESS; goto done; } } else if (entry->wired_count == 0) { entry->wired_count++; npages = atop(entry->end - entry->start); if (user_wire && !vm_map_wire_user_count_add(npages)) { vm_map_wire_entry_failure(map, entry, entry->start); end = entry->end; rv = KERN_RESOURCE_SHORTAGE; goto done; } /* * Release the map lock, relying on the in-transition * mark. Mark the map busy for fork. */ saved_start = entry->start; saved_end = entry->end; last_timestamp = map->timestamp; vm_map_busy(map); vm_map_unlock(map); faddr = saved_start; do { /* * Simulate a fault to get the page and enter * it into the physical map. */ if ((rv = vm_fault(map, faddr, VM_PROT_NONE, VM_FAULT_WIRE, NULL)) != KERN_SUCCESS) break; } while ((faddr += PAGE_SIZE) < saved_end); vm_map_lock(map); vm_map_unbusy(map); if (last_timestamp + 1 != map->timestamp) { /* * Look again for the entry because the map was * modified while it was unlocked. The entry * may have been clipped, but NOT merged or * deleted. */ if (!vm_map_lookup_entry(map, saved_start, &tmp_entry)) KASSERT(false, ("vm_map_wire: lookup failed")); if (entry == first_entry) first_entry = tmp_entry; else first_entry = NULL; entry = tmp_entry; while (entry->end < saved_end) { /* * In case of failure, handle entries * that were not fully wired here; * fully wired entries are handled * later. */ if (rv != KERN_SUCCESS && faddr < entry->end) vm_map_wire_entry_failure(map, entry, faddr); entry = entry->next; } } if (rv != KERN_SUCCESS) { vm_map_wire_entry_failure(map, entry, faddr); if (user_wire) vm_map_wire_user_count_sub(npages); end = entry->end; goto done; } } else if (!user_wire || (entry->eflags & MAP_ENTRY_USER_WIRED) == 0) { entry->wired_count++; } /* * Check the map for holes in the specified region. * If holes_ok was specified, skip this check. */ if (!holes_ok && entry->end < end && entry->next->start > entry->end) { end = entry->end; rv = KERN_INVALID_ADDRESS; goto done; } entry = entry->next; } rv = KERN_SUCCESS; done: need_wakeup = false; if (first_entry == NULL && !vm_map_lookup_entry(map, start, &first_entry)) { KASSERT(holes_ok, ("vm_map_wire: lookup failed")); first_entry = first_entry->next; } for (entry = first_entry; entry->start < end; entry = entry->next) { /* * If holes_ok was specified, an empty * space in the unwired region could have been mapped * while the map lock was dropped for faulting in the * pages or draining MAP_ENTRY_IN_TRANSITION. * Moreover, another thread could be simultaneously * wiring this new mapping entry. Detect these cases * and skip any entries marked as in transition not by us. */ if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 || entry->wiring_thread != curthread) { KASSERT(holes_ok, ("vm_map_wire: !HOLESOK and new/changed entry")); continue; } if ((entry->eflags & MAP_ENTRY_WIRE_SKIPPED) != 0) { /* do nothing */ } else if (rv == KERN_SUCCESS) { if (user_wire) entry->eflags |= MAP_ENTRY_USER_WIRED; } else if (entry->wired_count == -1) { /* * Wiring failed on this entry. Thus, unwiring is * unnecessary. */ entry->wired_count = 0; } else if (!user_wire || (entry->eflags & MAP_ENTRY_USER_WIRED) == 0) { /* * Undo the wiring. Wiring succeeded on this entry * but failed on a later entry. */ if (entry->wired_count == 1) { vm_map_entry_unwire(map, entry); if (user_wire) vm_map_wire_user_count_sub( atop(entry->end - entry->start)); } else entry->wired_count--; } KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0, ("vm_map_wire: in-transition flag missing %p", entry)); KASSERT(entry->wiring_thread == curthread, ("vm_map_wire: alien wire %p", entry)); entry->eflags &= ~(MAP_ENTRY_IN_TRANSITION | MAP_ENTRY_WIRE_SKIPPED); entry->wiring_thread = NULL; if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) { entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP; need_wakeup = true; } vm_map_try_merge_entries(map, entry->prev, entry); } vm_map_try_merge_entries(map, entry->prev, entry); if (need_wakeup) vm_map_wakeup(map); return (rv); } /* * vm_map_sync * * Push any dirty cached pages in the address range to their pager. * If syncio is TRUE, dirty pages are written synchronously. * If invalidate is TRUE, any cached pages are freed as well. * * If the size of the region from start to end is zero, we are * supposed to flush all modified pages within the region containing * start. Unfortunately, a region can be split or coalesced with * neighboring regions, making it difficult to determine what the * original region was. Therefore, we approximate this requirement by * flushing the current region containing start. * * Returns an error if any part of the specified range is not mapped. */ int vm_map_sync( vm_map_t map, vm_offset_t start, vm_offset_t end, boolean_t syncio, boolean_t invalidate) { vm_map_entry_t current; vm_map_entry_t entry; vm_size_t size; vm_object_t object; vm_ooffset_t offset; unsigned int last_timestamp; boolean_t failed; vm_map_lock_read(map); VM_MAP_RANGE_CHECK(map, start, end); if (!vm_map_lookup_entry(map, start, &entry)) { vm_map_unlock_read(map); return (KERN_INVALID_ADDRESS); } else if (start == end) { start = entry->start; end = entry->end; } /* * Make a first pass to check for user-wired memory and holes. */ for (current = entry; current->start < end; current = current->next) { if (invalidate && (current->eflags & MAP_ENTRY_USER_WIRED)) { vm_map_unlock_read(map); return (KERN_INVALID_ARGUMENT); } if (end > current->end && current->end != current->next->start) { vm_map_unlock_read(map); return (KERN_INVALID_ADDRESS); } } if (invalidate) pmap_remove(map->pmap, start, end); failed = FALSE; /* * Make a second pass, cleaning/uncaching pages from the indicated * objects as we go. */ for (current = entry; current->start < end;) { offset = current->offset + (start - current->start); size = (end <= current->end ? end : current->end) - start; if (current->eflags & MAP_ENTRY_IS_SUB_MAP) { vm_map_t smap; vm_map_entry_t tentry; vm_size_t tsize; smap = current->object.sub_map; vm_map_lock_read(smap); (void) vm_map_lookup_entry(smap, offset, &tentry); tsize = tentry->end - offset; if (tsize < size) size = tsize; object = tentry->object.vm_object; offset = tentry->offset + (offset - tentry->start); vm_map_unlock_read(smap); } else { object = current->object.vm_object; } vm_object_reference(object); last_timestamp = map->timestamp; vm_map_unlock_read(map); if (!vm_object_sync(object, offset, size, syncio, invalidate)) failed = TRUE; start += size; vm_object_deallocate(object); vm_map_lock_read(map); if (last_timestamp == map->timestamp || !vm_map_lookup_entry(map, start, ¤t)) current = current->next; } vm_map_unlock_read(map); return (failed ? KERN_FAILURE : KERN_SUCCESS); } /* * vm_map_entry_unwire: [ internal use only ] * * Make the region specified by this entry pageable. * * The map in question should be locked. * [This is the reason for this routine's existence.] */ static void vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry) { vm_size_t size; VM_MAP_ASSERT_LOCKED(map); KASSERT(entry->wired_count > 0, ("vm_map_entry_unwire: entry %p isn't wired", entry)); size = entry->end - entry->start; if ((entry->eflags & MAP_ENTRY_USER_WIRED) != 0) vm_map_wire_user_count_sub(atop(size)); pmap_unwire(map->pmap, entry->start, entry->end); vm_object_unwire(entry->object.vm_object, entry->offset, size, PQ_ACTIVE); entry->wired_count = 0; } static void vm_map_entry_deallocate(vm_map_entry_t entry, boolean_t system_map) { if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) vm_object_deallocate(entry->object.vm_object); uma_zfree(system_map ? kmapentzone : mapentzone, entry); } /* * vm_map_entry_delete: [ internal use only ] * * Deallocate the given entry from the target map. */ static void vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry) { vm_object_t object; vm_pindex_t offidxstart, offidxend, count, size1; vm_size_t size; vm_map_entry_unlink(map, entry, UNLINK_MERGE_NONE); object = entry->object.vm_object; if ((entry->eflags & MAP_ENTRY_GUARD) != 0) { MPASS(entry->cred == NULL); MPASS((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0); MPASS(object == NULL); vm_map_entry_deallocate(entry, map->system_map); return; } size = entry->end - entry->start; map->size -= size; if (entry->cred != NULL) { swap_release_by_cred(size, entry->cred); crfree(entry->cred); } if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 && (object != NULL)) { KASSERT(entry->cred == NULL || object->cred == NULL || (entry->eflags & MAP_ENTRY_NEEDS_COPY), ("OVERCOMMIT vm_map_entry_delete: both cred %p", entry)); count = atop(size); offidxstart = OFF_TO_IDX(entry->offset); offidxend = offidxstart + count; VM_OBJECT_WLOCK(object); if (object->ref_count != 1 && ((object->flags & (OBJ_NOSPLIT | OBJ_ONEMAPPING)) == OBJ_ONEMAPPING || object == kernel_object)) { vm_object_collapse(object); /* * The option OBJPR_NOTMAPPED can be passed here * because vm_map_delete() already performed * pmap_remove() on the only mapping to this range * of pages. */ vm_object_page_remove(object, offidxstart, offidxend, OBJPR_NOTMAPPED); if (object->type == OBJT_SWAP) swap_pager_freespace(object, offidxstart, count); if (offidxend >= object->size && offidxstart < object->size) { size1 = object->size; object->size = offidxstart; if (object->cred != NULL) { size1 -= object->size; KASSERT(object->charge >= ptoa(size1), ("object %p charge < 0", object)); swap_release_by_cred(ptoa(size1), object->cred); object->charge -= ptoa(size1); } } } VM_OBJECT_WUNLOCK(object); } else entry->object.vm_object = NULL; if (map->system_map) vm_map_entry_deallocate(entry, TRUE); else { entry->next = curthread->td_map_def_user; curthread->td_map_def_user = entry; } } /* * vm_map_delete: [ internal use only ] * * Deallocates the given address range from the target * map. */ int vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end) { vm_map_entry_t entry; vm_map_entry_t first_entry; VM_MAP_ASSERT_LOCKED(map); if (start == end) return (KERN_SUCCESS); /* * Find the start of the region, and clip it */ if (!vm_map_lookup_entry(map, start, &first_entry)) entry = first_entry->next; else { entry = first_entry; vm_map_clip_start(map, entry, start); } /* * Step through all entries in this region */ while (entry->start < end) { vm_map_entry_t next; /* * Wait for wiring or unwiring of an entry to complete. * Also wait for any system wirings to disappear on * user maps. */ if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0 || (vm_map_pmap(map) != kernel_pmap && vm_map_entry_system_wired_count(entry) != 0)) { unsigned int last_timestamp; vm_offset_t saved_start; vm_map_entry_t tmp_entry; saved_start = entry->start; entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; last_timestamp = map->timestamp; (void) vm_map_unlock_and_wait(map, 0); vm_map_lock(map); if (last_timestamp + 1 != map->timestamp) { /* * Look again for the entry because the map was * modified while it was unlocked. * Specifically, the entry may have been * clipped, merged, or deleted. */ if (!vm_map_lookup_entry(map, saved_start, &tmp_entry)) entry = tmp_entry->next; else { entry = tmp_entry; vm_map_clip_start(map, entry, saved_start); } } continue; } vm_map_clip_end(map, entry, end); next = entry->next; /* * Unwire before removing addresses from the pmap; otherwise, * unwiring will put the entries back in the pmap. */ if (entry->wired_count != 0) vm_map_entry_unwire(map, entry); /* * Remove mappings for the pages, but only if the * mappings could exist. For instance, it does not * make sense to call pmap_remove() for guard entries. */ if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0 || entry->object.vm_object != NULL) pmap_remove(map->pmap, entry->start, entry->end); if (entry->end == map->anon_loc) map->anon_loc = entry->start; /* * Delete the entry only after removing all pmap * entries pointing to its pages. (Otherwise, its * page frames may be reallocated, and any modify bits * will be set in the wrong object!) */ vm_map_entry_delete(map, entry); entry = next; } return (KERN_SUCCESS); } /* * vm_map_remove: * * Remove the given address range from the target map. * This is the exported form of vm_map_delete. */ int vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end) { int result; vm_map_lock(map); VM_MAP_RANGE_CHECK(map, start, end); result = vm_map_delete(map, start, end); vm_map_unlock(map); return (result); } /* * vm_map_check_protection: * * Assert that the target map allows the specified privilege on the * entire address region given. The entire region must be allocated. * * WARNING! This code does not and should not check whether the * contents of the region is accessible. For example a smaller file * might be mapped into a larger address space. * * NOTE! This code is also called by munmap(). * * The map must be locked. A read lock is sufficient. */ boolean_t vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end, vm_prot_t protection) { vm_map_entry_t entry; vm_map_entry_t tmp_entry; if (!vm_map_lookup_entry(map, start, &tmp_entry)) return (FALSE); entry = tmp_entry; while (start < end) { /* * No holes allowed! */ if (start < entry->start) return (FALSE); /* * Check protection associated with entry. */ if ((entry->protection & protection) != protection) return (FALSE); /* go to next entry */ start = entry->end; entry = entry->next; } return (TRUE); } /* * vm_map_copy_entry: * * Copies the contents of the source entry to the destination * entry. The entries *must* be aligned properly. */ static void vm_map_copy_entry( vm_map_t src_map, vm_map_t dst_map, vm_map_entry_t src_entry, vm_map_entry_t dst_entry, vm_ooffset_t *fork_charge) { vm_object_t src_object; vm_map_entry_t fake_entry; vm_offset_t size; struct ucred *cred; int charged; VM_MAP_ASSERT_LOCKED(dst_map); if ((dst_entry->eflags|src_entry->eflags) & MAP_ENTRY_IS_SUB_MAP) return; if (src_entry->wired_count == 0 || (src_entry->protection & VM_PROT_WRITE) == 0) { /* * If the source entry is marked needs_copy, it is already * write-protected. */ if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0 && (src_entry->protection & VM_PROT_WRITE) != 0) { pmap_protect(src_map->pmap, src_entry->start, src_entry->end, src_entry->protection & ~VM_PROT_WRITE); } /* * Make a copy of the object. */ size = src_entry->end - src_entry->start; if ((src_object = src_entry->object.vm_object) != NULL) { VM_OBJECT_WLOCK(src_object); charged = ENTRY_CHARGED(src_entry); if (src_object->handle == NULL && (src_object->type == OBJT_DEFAULT || src_object->type == OBJT_SWAP)) { vm_object_collapse(src_object); if ((src_object->flags & (OBJ_NOSPLIT | OBJ_ONEMAPPING)) == OBJ_ONEMAPPING) { vm_object_split(src_entry); src_object = src_entry->object.vm_object; } } vm_object_reference_locked(src_object); vm_object_clear_flag(src_object, OBJ_ONEMAPPING); if (src_entry->cred != NULL && !(src_entry->eflags & MAP_ENTRY_NEEDS_COPY)) { KASSERT(src_object->cred == NULL, ("OVERCOMMIT: vm_map_copy_entry: cred %p", src_object)); src_object->cred = src_entry->cred; src_object->charge = size; } VM_OBJECT_WUNLOCK(src_object); dst_entry->object.vm_object = src_object; if (charged) { cred = curthread->td_ucred; crhold(cred); dst_entry->cred = cred; *fork_charge += size; if (!(src_entry->eflags & MAP_ENTRY_NEEDS_COPY)) { crhold(cred); src_entry->cred = cred; *fork_charge += size; } } src_entry->eflags |= MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY; dst_entry->eflags |= MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY; dst_entry->offset = src_entry->offset; if (src_entry->eflags & MAP_ENTRY_WRITECNT) { /* * MAP_ENTRY_WRITECNT cannot * indicate write reference from * src_entry, since the entry is * marked as needs copy. Allocate a * fake entry that is used to * decrement object->un_pager writecount * at the appropriate time. Attach * fake_entry to the deferred list. */ fake_entry = vm_map_entry_create(dst_map); fake_entry->eflags = MAP_ENTRY_WRITECNT; src_entry->eflags &= ~MAP_ENTRY_WRITECNT; vm_object_reference(src_object); fake_entry->object.vm_object = src_object; fake_entry->start = src_entry->start; fake_entry->end = src_entry->end; fake_entry->next = curthread->td_map_def_user; curthread->td_map_def_user = fake_entry; } pmap_copy(dst_map->pmap, src_map->pmap, dst_entry->start, dst_entry->end - dst_entry->start, src_entry->start); } else { dst_entry->object.vm_object = NULL; dst_entry->offset = 0; if (src_entry->cred != NULL) { dst_entry->cred = curthread->td_ucred; crhold(dst_entry->cred); *fork_charge += size; } } } else { /* * We don't want to make writeable wired pages copy-on-write. * Immediately copy these pages into the new map by simulating * page faults. The new pages are pageable. */ vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry, fork_charge); } } /* * vmspace_map_entry_forked: * Update the newly-forked vmspace each time a map entry is inherited * or copied. The values for vm_dsize and vm_tsize are approximate * (and mostly-obsolete ideas in the face of mmap(2) et al.) */ static void vmspace_map_entry_forked(const struct vmspace *vm1, struct vmspace *vm2, vm_map_entry_t entry) { vm_size_t entrysize; vm_offset_t newend; if ((entry->eflags & MAP_ENTRY_GUARD) != 0) return; entrysize = entry->end - entry->start; vm2->vm_map.size += entrysize; if (entry->eflags & (MAP_ENTRY_GROWS_DOWN | MAP_ENTRY_GROWS_UP)) { vm2->vm_ssize += btoc(entrysize); } else if (entry->start >= (vm_offset_t)vm1->vm_daddr && entry->start < (vm_offset_t)vm1->vm_daddr + ctob(vm1->vm_dsize)) { newend = MIN(entry->end, (vm_offset_t)vm1->vm_daddr + ctob(vm1->vm_dsize)); vm2->vm_dsize += btoc(newend - entry->start); } else if (entry->start >= (vm_offset_t)vm1->vm_taddr && entry->start < (vm_offset_t)vm1->vm_taddr + ctob(vm1->vm_tsize)) { newend = MIN(entry->end, (vm_offset_t)vm1->vm_taddr + ctob(vm1->vm_tsize)); vm2->vm_tsize += btoc(newend - entry->start); } } /* * vmspace_fork: * Create a new process vmspace structure and vm_map * based on those of an existing process. The new map * is based on the old map, according to the inheritance * values on the regions in that map. * * XXX It might be worth coalescing the entries added to the new vmspace. * * The source map must not be locked. */ struct vmspace * vmspace_fork(struct vmspace *vm1, vm_ooffset_t *fork_charge) { struct vmspace *vm2; vm_map_t new_map, old_map; vm_map_entry_t new_entry, old_entry; vm_object_t object; int error, locked; vm_inherit_t inh; old_map = &vm1->vm_map; /* Copy immutable fields of vm1 to vm2. */ vm2 = vmspace_alloc(vm_map_min(old_map), vm_map_max(old_map), pmap_pinit); if (vm2 == NULL) return (NULL); vm2->vm_taddr = vm1->vm_taddr; vm2->vm_daddr = vm1->vm_daddr; vm2->vm_maxsaddr = vm1->vm_maxsaddr; vm_map_lock(old_map); if (old_map->busy) vm_map_wait_busy(old_map); new_map = &vm2->vm_map; locked = vm_map_trylock(new_map); /* trylock to silence WITNESS */ KASSERT(locked, ("vmspace_fork: lock failed")); error = pmap_vmspace_copy(new_map->pmap, old_map->pmap); if (error != 0) { sx_xunlock(&old_map->lock); sx_xunlock(&new_map->lock); vm_map_process_deferred(); vmspace_free(vm2); return (NULL); } new_map->anon_loc = old_map->anon_loc; old_entry = old_map->header.next; while (old_entry != &old_map->header) { if (old_entry->eflags & MAP_ENTRY_IS_SUB_MAP) panic("vm_map_fork: encountered a submap"); inh = old_entry->inheritance; if ((old_entry->eflags & MAP_ENTRY_GUARD) != 0 && inh != VM_INHERIT_NONE) inh = VM_INHERIT_COPY; switch (inh) { case VM_INHERIT_NONE: break; case VM_INHERIT_SHARE: /* * Clone the entry, creating the shared object if necessary. */ object = old_entry->object.vm_object; if (object == NULL) { vm_map_entry_back(old_entry); object = old_entry->object.vm_object; } /* * Add the reference before calling vm_object_shadow * to insure that a shadow object is created. */ vm_object_reference(object); if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) { vm_object_shadow(&old_entry->object.vm_object, &old_entry->offset, old_entry->end - old_entry->start); old_entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; /* Transfer the second reference too. */ vm_object_reference( old_entry->object.vm_object); /* * As in vm_map_merged_neighbor_dispose(), * the vnode lock will not be acquired in * this call to vm_object_deallocate(). */ vm_object_deallocate(object); object = old_entry->object.vm_object; } VM_OBJECT_WLOCK(object); vm_object_clear_flag(object, OBJ_ONEMAPPING); if (old_entry->cred != NULL) { KASSERT(object->cred == NULL, ("vmspace_fork both cred")); object->cred = old_entry->cred; object->charge = old_entry->end - old_entry->start; old_entry->cred = NULL; } /* * Assert the correct state of the vnode * v_writecount while the object is locked, to * not relock it later for the assertion * correctness. */ if (old_entry->eflags & MAP_ENTRY_WRITECNT && object->type == OBJT_VNODE) { KASSERT(((struct vnode *)object->handle)-> v_writecount > 0, ("vmspace_fork: v_writecount %p", object)); KASSERT(object->un_pager.vnp.writemappings > 0, ("vmspace_fork: vnp.writecount %p", object)); } VM_OBJECT_WUNLOCK(object); /* * Clone the entry, referencing the shared object. */ new_entry = vm_map_entry_create(new_map); *new_entry = *old_entry; new_entry->eflags &= ~(MAP_ENTRY_USER_WIRED | MAP_ENTRY_IN_TRANSITION); new_entry->wiring_thread = NULL; new_entry->wired_count = 0; if (new_entry->eflags & MAP_ENTRY_WRITECNT) { vm_pager_update_writecount(object, new_entry->start, new_entry->end); } vm_map_entry_set_vnode_text(new_entry, true); /* * Insert the entry into the new map -- we know we're * inserting at the end of the new map. */ vm_map_entry_link(new_map, new_entry); vmspace_map_entry_forked(vm1, vm2, new_entry); /* * Update the physical map */ pmap_copy(new_map->pmap, old_map->pmap, new_entry->start, (old_entry->end - old_entry->start), old_entry->start); break; case VM_INHERIT_COPY: /* * Clone the entry and link into the map. */ new_entry = vm_map_entry_create(new_map); *new_entry = *old_entry; /* * Copied entry is COW over the old object. */ new_entry->eflags &= ~(MAP_ENTRY_USER_WIRED | MAP_ENTRY_IN_TRANSITION | MAP_ENTRY_WRITECNT); new_entry->wiring_thread = NULL; new_entry->wired_count = 0; new_entry->object.vm_object = NULL; new_entry->cred = NULL; vm_map_entry_link(new_map, new_entry); vmspace_map_entry_forked(vm1, vm2, new_entry); vm_map_copy_entry(old_map, new_map, old_entry, new_entry, fork_charge); vm_map_entry_set_vnode_text(new_entry, true); break; case VM_INHERIT_ZERO: /* * Create a new anonymous mapping entry modelled from * the old one. */ new_entry = vm_map_entry_create(new_map); memset(new_entry, 0, sizeof(*new_entry)); new_entry->start = old_entry->start; new_entry->end = old_entry->end; new_entry->eflags = old_entry->eflags & ~(MAP_ENTRY_USER_WIRED | MAP_ENTRY_IN_TRANSITION | MAP_ENTRY_WRITECNT | MAP_ENTRY_VN_EXEC); new_entry->protection = old_entry->protection; new_entry->max_protection = old_entry->max_protection; new_entry->inheritance = VM_INHERIT_ZERO; vm_map_entry_link(new_map, new_entry); vmspace_map_entry_forked(vm1, vm2, new_entry); new_entry->cred = curthread->td_ucred; crhold(new_entry->cred); *fork_charge += (new_entry->end - new_entry->start); break; } old_entry = old_entry->next; } /* * Use inlined vm_map_unlock() to postpone handling the deferred * map entries, which cannot be done until both old_map and * new_map locks are released. */ sx_xunlock(&old_map->lock); sx_xunlock(&new_map->lock); vm_map_process_deferred(); return (vm2); } /* * Create a process's stack for exec_new_vmspace(). This function is never * asked to wire the newly created stack. */ int vm_map_stack(vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize, vm_prot_t prot, vm_prot_t max, int cow) { vm_size_t growsize, init_ssize; rlim_t vmemlim; int rv; MPASS((map->flags & MAP_WIREFUTURE) == 0); growsize = sgrowsiz; init_ssize = (max_ssize < growsize) ? max_ssize : growsize; vm_map_lock(map); vmemlim = lim_cur(curthread, RLIMIT_VMEM); /* If we would blow our VMEM resource limit, no go */ if (map->size + init_ssize > vmemlim) { rv = KERN_NO_SPACE; goto out; } rv = vm_map_stack_locked(map, addrbos, max_ssize, growsize, prot, max, cow); out: vm_map_unlock(map); return (rv); } static int stack_guard_page = 1; SYSCTL_INT(_security_bsd, OID_AUTO, stack_guard_page, CTLFLAG_RWTUN, &stack_guard_page, 0, "Specifies the number of guard pages for a stack that grows"); static int vm_map_stack_locked(vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize, vm_size_t growsize, vm_prot_t prot, vm_prot_t max, int cow) { vm_map_entry_t new_entry, prev_entry; vm_offset_t bot, gap_bot, gap_top, top; vm_size_t init_ssize, sgp; int orient, rv; /* * The stack orientation is piggybacked with the cow argument. * Extract it into orient and mask the cow argument so that we * don't pass it around further. */ orient = cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP); KASSERT(orient != 0, ("No stack grow direction")); KASSERT(orient != (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP), ("bi-dir stack")); if (addrbos < vm_map_min(map) || addrbos + max_ssize > vm_map_max(map) || addrbos + max_ssize <= addrbos) return (KERN_INVALID_ADDRESS); sgp = (curproc->p_flag2 & P2_STKGAP_DISABLE) != 0 ? 0 : (vm_size_t)stack_guard_page * PAGE_SIZE; if (sgp >= max_ssize) return (KERN_INVALID_ARGUMENT); init_ssize = growsize; if (max_ssize < init_ssize + sgp) init_ssize = max_ssize - sgp; /* If addr is already mapped, no go */ if (vm_map_lookup_entry(map, addrbos, &prev_entry)) return (KERN_NO_SPACE); /* * If we can't accommodate max_ssize in the current mapping, no go. */ if (prev_entry->next->start < addrbos + max_ssize) return (KERN_NO_SPACE); /* * We initially map a stack of only init_ssize. We will grow as * needed later. Depending on the orientation of the stack (i.e. * the grow direction) we either map at the top of the range, the * bottom of the range or in the middle. * * Note: we would normally expect prot and max to be VM_PROT_ALL, * and cow to be 0. Possibly we should eliminate these as input * parameters, and just pass these values here in the insert call. */ if (orient == MAP_STACK_GROWS_DOWN) { bot = addrbos + max_ssize - init_ssize; top = bot + init_ssize; gap_bot = addrbos; gap_top = bot; } else /* if (orient == MAP_STACK_GROWS_UP) */ { bot = addrbos; top = bot + init_ssize; gap_bot = top; gap_top = addrbos + max_ssize; } rv = vm_map_insert(map, NULL, 0, bot, top, prot, max, cow); if (rv != KERN_SUCCESS) return (rv); new_entry = prev_entry->next; KASSERT(new_entry->end == top || new_entry->start == bot, ("Bad entry start/end for new stack entry")); KASSERT((orient & MAP_STACK_GROWS_DOWN) == 0 || (new_entry->eflags & MAP_ENTRY_GROWS_DOWN) != 0, ("new entry lacks MAP_ENTRY_GROWS_DOWN")); KASSERT((orient & MAP_STACK_GROWS_UP) == 0 || (new_entry->eflags & MAP_ENTRY_GROWS_UP) != 0, ("new entry lacks MAP_ENTRY_GROWS_UP")); if (gap_bot == gap_top) return (KERN_SUCCESS); rv = vm_map_insert(map, NULL, 0, gap_bot, gap_top, VM_PROT_NONE, VM_PROT_NONE, MAP_CREATE_GUARD | (orient == MAP_STACK_GROWS_DOWN ? MAP_CREATE_STACK_GAP_DN : MAP_CREATE_STACK_GAP_UP)); if (rv == KERN_SUCCESS) { /* * Gap can never successfully handle a fault, so * read-ahead logic is never used for it. Re-use * next_read of the gap entry to store * stack_guard_page for vm_map_growstack(). */ if (orient == MAP_STACK_GROWS_DOWN) new_entry->prev->next_read = sgp; else new_entry->next->next_read = sgp; } else { (void)vm_map_delete(map, bot, top); } return (rv); } /* * Attempts to grow a vm stack entry. Returns KERN_SUCCESS if we * successfully grow the stack. */ static int vm_map_growstack(vm_map_t map, vm_offset_t addr, vm_map_entry_t gap_entry) { vm_map_entry_t stack_entry; struct proc *p; struct vmspace *vm; struct ucred *cred; vm_offset_t gap_end, gap_start, grow_start; vm_size_t grow_amount, guard, max_grow; rlim_t lmemlim, stacklim, vmemlim; int rv, rv1; bool gap_deleted, grow_down, is_procstack; #ifdef notyet uint64_t limit; #endif #ifdef RACCT int error; #endif p = curproc; vm = p->p_vmspace; /* * Disallow stack growth when the access is performed by a * debugger or AIO daemon. The reason is that the wrong * resource limits are applied. */ if (p != initproc && (map != &p->p_vmspace->vm_map || p->p_textvp == NULL)) return (KERN_FAILURE); MPASS(!map->system_map); lmemlim = lim_cur(curthread, RLIMIT_MEMLOCK); stacklim = lim_cur(curthread, RLIMIT_STACK); vmemlim = lim_cur(curthread, RLIMIT_VMEM); retry: /* If addr is not in a hole for a stack grow area, no need to grow. */ if (gap_entry == NULL && !vm_map_lookup_entry(map, addr, &gap_entry)) return (KERN_FAILURE); if ((gap_entry->eflags & MAP_ENTRY_GUARD) == 0) return (KERN_SUCCESS); if ((gap_entry->eflags & MAP_ENTRY_STACK_GAP_DN) != 0) { stack_entry = gap_entry->next; if ((stack_entry->eflags & MAP_ENTRY_GROWS_DOWN) == 0 || stack_entry->start != gap_entry->end) return (KERN_FAILURE); grow_amount = round_page(stack_entry->start - addr); grow_down = true; } else if ((gap_entry->eflags & MAP_ENTRY_STACK_GAP_UP) != 0) { stack_entry = gap_entry->prev; if ((stack_entry->eflags & MAP_ENTRY_GROWS_UP) == 0 || stack_entry->end != gap_entry->start) return (KERN_FAILURE); grow_amount = round_page(addr + 1 - stack_entry->end); grow_down = false; } else { return (KERN_FAILURE); } guard = (curproc->p_flag2 & P2_STKGAP_DISABLE) != 0 ? 0 : gap_entry->next_read; max_grow = gap_entry->end - gap_entry->start; if (guard > max_grow) return (KERN_NO_SPACE); max_grow -= guard; if (grow_amount > max_grow) return (KERN_NO_SPACE); /* * If this is the main process stack, see if we're over the stack * limit. */ is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr && addr < (vm_offset_t)p->p_sysent->sv_usrstack; if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > stacklim)) return (KERN_NO_SPACE); #ifdef RACCT if (racct_enable) { PROC_LOCK(p); if (is_procstack && racct_set(p, RACCT_STACK, ctob(vm->vm_ssize) + grow_amount)) { PROC_UNLOCK(p); return (KERN_NO_SPACE); } PROC_UNLOCK(p); } #endif grow_amount = roundup(grow_amount, sgrowsiz); if (grow_amount > max_grow) grow_amount = max_grow; if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > stacklim)) { grow_amount = trunc_page((vm_size_t)stacklim) - ctob(vm->vm_ssize); } #ifdef notyet PROC_LOCK(p); limit = racct_get_available(p, RACCT_STACK); PROC_UNLOCK(p); if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > limit)) grow_amount = limit - ctob(vm->vm_ssize); #endif if (!old_mlock && (map->flags & MAP_WIREFUTURE) != 0) { if (ptoa(pmap_wired_count(map->pmap)) + grow_amount > lmemlim) { rv = KERN_NO_SPACE; goto out; } #ifdef RACCT if (racct_enable) { PROC_LOCK(p); if (racct_set(p, RACCT_MEMLOCK, ptoa(pmap_wired_count(map->pmap)) + grow_amount)) { PROC_UNLOCK(p); rv = KERN_NO_SPACE; goto out; } PROC_UNLOCK(p); } #endif } /* If we would blow our VMEM resource limit, no go */ if (map->size + grow_amount > vmemlim) { rv = KERN_NO_SPACE; goto out; } #ifdef RACCT if (racct_enable) { PROC_LOCK(p); if (racct_set(p, RACCT_VMEM, map->size + grow_amount)) { PROC_UNLOCK(p); rv = KERN_NO_SPACE; goto out; } PROC_UNLOCK(p); } #endif if (vm_map_lock_upgrade(map)) { gap_entry = NULL; vm_map_lock_read(map); goto retry; } if (grow_down) { grow_start = gap_entry->end - grow_amount; if (gap_entry->start + grow_amount == gap_entry->end) { gap_start = gap_entry->start; gap_end = gap_entry->end; vm_map_entry_delete(map, gap_entry); gap_deleted = true; } else { MPASS(gap_entry->start < gap_entry->end - grow_amount); vm_map_entry_resize(map, gap_entry, -grow_amount); gap_deleted = false; } rv = vm_map_insert(map, NULL, 0, grow_start, grow_start + grow_amount, stack_entry->protection, stack_entry->max_protection, MAP_STACK_GROWS_DOWN); if (rv != KERN_SUCCESS) { if (gap_deleted) { rv1 = vm_map_insert(map, NULL, 0, gap_start, gap_end, VM_PROT_NONE, VM_PROT_NONE, MAP_CREATE_GUARD | MAP_CREATE_STACK_GAP_DN); MPASS(rv1 == KERN_SUCCESS); } else vm_map_entry_resize(map, gap_entry, grow_amount); } } else { grow_start = stack_entry->end; cred = stack_entry->cred; if (cred == NULL && stack_entry->object.vm_object != NULL) cred = stack_entry->object.vm_object->cred; if (cred != NULL && !swap_reserve_by_cred(grow_amount, cred)) rv = KERN_NO_SPACE; /* Grow the underlying object if applicable. */ else if (stack_entry->object.vm_object == NULL || vm_object_coalesce(stack_entry->object.vm_object, stack_entry->offset, (vm_size_t)(stack_entry->end - stack_entry->start), grow_amount, cred != NULL)) { if (gap_entry->start + grow_amount == gap_entry->end) { vm_map_entry_delete(map, gap_entry); vm_map_entry_resize(map, stack_entry, grow_amount); } else { gap_entry->start += grow_amount; stack_entry->end += grow_amount; } map->size += grow_amount; rv = KERN_SUCCESS; } else rv = KERN_FAILURE; } if (rv == KERN_SUCCESS && is_procstack) vm->vm_ssize += btoc(grow_amount); /* * Heed the MAP_WIREFUTURE flag if it was set for this process. */ if (rv == KERN_SUCCESS && (map->flags & MAP_WIREFUTURE) != 0) { rv = vm_map_wire_locked(map, grow_start, grow_start + grow_amount, VM_MAP_WIRE_USER | VM_MAP_WIRE_NOHOLES); } vm_map_lock_downgrade(map); out: #ifdef RACCT if (racct_enable && rv != KERN_SUCCESS) { PROC_LOCK(p); error = racct_set(p, RACCT_VMEM, map->size); KASSERT(error == 0, ("decreasing RACCT_VMEM failed")); if (!old_mlock) { error = racct_set(p, RACCT_MEMLOCK, ptoa(pmap_wired_count(map->pmap))); KASSERT(error == 0, ("decreasing RACCT_MEMLOCK failed")); } error = racct_set(p, RACCT_STACK, ctob(vm->vm_ssize)); KASSERT(error == 0, ("decreasing RACCT_STACK failed")); PROC_UNLOCK(p); } #endif return (rv); } /* * Unshare the specified VM space for exec. If other processes are * mapped to it, then create a new one. The new vmspace is null. */ int vmspace_exec(struct proc *p, vm_offset_t minuser, vm_offset_t maxuser) { struct vmspace *oldvmspace = p->p_vmspace; struct vmspace *newvmspace; KASSERT((curthread->td_pflags & TDP_EXECVMSPC) == 0, ("vmspace_exec recursed")); newvmspace = vmspace_alloc(minuser, maxuser, pmap_pinit); if (newvmspace == NULL) return (ENOMEM); newvmspace->vm_swrss = oldvmspace->vm_swrss; /* * This code is written like this for prototype purposes. The * goal is to avoid running down the vmspace here, but let the * other process's that are still using the vmspace to finally * run it down. Even though there is little or no chance of blocking * here, it is a good idea to keep this form for future mods. */ PROC_VMSPACE_LOCK(p); p->p_vmspace = newvmspace; PROC_VMSPACE_UNLOCK(p); if (p == curthread->td_proc) pmap_activate(curthread); curthread->td_pflags |= TDP_EXECVMSPC; return (0); } /* * Unshare the specified VM space for forcing COW. This * is called by rfork, for the (RFMEM|RFPROC) == 0 case. */ int vmspace_unshare(struct proc *p) { struct vmspace *oldvmspace = p->p_vmspace; struct vmspace *newvmspace; vm_ooffset_t fork_charge; if (oldvmspace->vm_refcnt == 1) return (0); fork_charge = 0; newvmspace = vmspace_fork(oldvmspace, &fork_charge); if (newvmspace == NULL) return (ENOMEM); if (!swap_reserve_by_cred(fork_charge, p->p_ucred)) { vmspace_free(newvmspace); return (ENOMEM); } PROC_VMSPACE_LOCK(p); p->p_vmspace = newvmspace; PROC_VMSPACE_UNLOCK(p); if (p == curthread->td_proc) pmap_activate(curthread); vmspace_free(oldvmspace); return (0); } /* * vm_map_lookup: * * Finds the VM object, offset, and * protection for a given virtual address in the * specified map, assuming a page fault of the * type specified. * * Leaves the map in question locked for read; return * values are guaranteed until a vm_map_lookup_done * call is performed. Note that the map argument * is in/out; the returned map must be used in * the call to vm_map_lookup_done. * * A handle (out_entry) is returned for use in * vm_map_lookup_done, to make that fast. * * If a lookup is requested with "write protection" * specified, the map may be changed to perform virtual * copying operations, although the data referenced will * remain the same. */ int vm_map_lookup(vm_map_t *var_map, /* IN/OUT */ vm_offset_t vaddr, vm_prot_t fault_typea, vm_map_entry_t *out_entry, /* OUT */ vm_object_t *object, /* OUT */ vm_pindex_t *pindex, /* OUT */ vm_prot_t *out_prot, /* OUT */ boolean_t *wired) /* OUT */ { vm_map_entry_t entry; vm_map_t map = *var_map; vm_prot_t prot; vm_prot_t fault_type = fault_typea; vm_object_t eobject; vm_size_t size; struct ucred *cred; RetryLookup: vm_map_lock_read(map); RetryLookupLocked: /* * Lookup the faulting address. */ if (!vm_map_lookup_entry(map, vaddr, out_entry)) { vm_map_unlock_read(map); return (KERN_INVALID_ADDRESS); } entry = *out_entry; /* * Handle submaps. */ if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { vm_map_t old_map = map; *var_map = map = entry->object.sub_map; vm_map_unlock_read(old_map); goto RetryLookup; } /* * Check whether this task is allowed to have this page. */ prot = entry->protection; if ((fault_typea & VM_PROT_FAULT_LOOKUP) != 0) { fault_typea &= ~VM_PROT_FAULT_LOOKUP; if (prot == VM_PROT_NONE && map != kernel_map && (entry->eflags & MAP_ENTRY_GUARD) != 0 && (entry->eflags & (MAP_ENTRY_STACK_GAP_DN | MAP_ENTRY_STACK_GAP_UP)) != 0 && vm_map_growstack(map, vaddr, entry) == KERN_SUCCESS) goto RetryLookupLocked; } fault_type &= VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE; if ((fault_type & prot) != fault_type || prot == VM_PROT_NONE) { vm_map_unlock_read(map); return (KERN_PROTECTION_FAILURE); } KASSERT((prot & VM_PROT_WRITE) == 0 || (entry->eflags & (MAP_ENTRY_USER_WIRED | MAP_ENTRY_NEEDS_COPY)) != (MAP_ENTRY_USER_WIRED | MAP_ENTRY_NEEDS_COPY), ("entry %p flags %x", entry, entry->eflags)); if ((fault_typea & VM_PROT_COPY) != 0 && (entry->max_protection & VM_PROT_WRITE) == 0 && (entry->eflags & MAP_ENTRY_COW) == 0) { vm_map_unlock_read(map); return (KERN_PROTECTION_FAILURE); } /* * If this page is not pageable, we have to get it for all possible * accesses. */ *wired = (entry->wired_count != 0); if (*wired) fault_type = entry->protection; size = entry->end - entry->start; /* * If the entry was copy-on-write, we either ... */ if (entry->eflags & MAP_ENTRY_NEEDS_COPY) { /* * If we want to write the page, we may as well handle that * now since we've got the map locked. * * If we don't need to write the page, we just demote the * permissions allowed. */ if ((fault_type & VM_PROT_WRITE) != 0 || (fault_typea & VM_PROT_COPY) != 0) { /* * Make a new object, and place it in the object * chain. Note that no new references have appeared * -- one just moved from the map to the new * object. */ if (vm_map_lock_upgrade(map)) goto RetryLookup; if (entry->cred == NULL) { /* * The debugger owner is charged for * the memory. */ cred = curthread->td_ucred; crhold(cred); if (!swap_reserve_by_cred(size, cred)) { crfree(cred); vm_map_unlock(map); return (KERN_RESOURCE_SHORTAGE); } entry->cred = cred; } vm_object_shadow(&entry->object.vm_object, &entry->offset, size); entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; eobject = entry->object.vm_object; if (eobject->cred != NULL) { /* * The object was not shadowed. */ swap_release_by_cred(size, entry->cred); crfree(entry->cred); entry->cred = NULL; } else if (entry->cred != NULL) { VM_OBJECT_WLOCK(eobject); eobject->cred = entry->cred; eobject->charge = size; VM_OBJECT_WUNLOCK(eobject); entry->cred = NULL; } vm_map_lock_downgrade(map); } else { /* * We're attempting to read a copy-on-write page -- * don't allow writes. */ prot &= ~VM_PROT_WRITE; } } /* * Create an object if necessary. */ if (entry->object.vm_object == NULL && !map->system_map) { if (vm_map_lock_upgrade(map)) goto RetryLookup; entry->object.vm_object = vm_object_allocate(OBJT_DEFAULT, atop(size)); entry->offset = 0; if (entry->cred != NULL) { VM_OBJECT_WLOCK(entry->object.vm_object); entry->object.vm_object->cred = entry->cred; entry->object.vm_object->charge = size; VM_OBJECT_WUNLOCK(entry->object.vm_object); entry->cred = NULL; } vm_map_lock_downgrade(map); } /* * Return the object/offset from this entry. If the entry was * copy-on-write or empty, it has been fixed up. */ *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset); *object = entry->object.vm_object; *out_prot = prot; return (KERN_SUCCESS); } /* * vm_map_lookup_locked: * * Lookup the faulting address. A version of vm_map_lookup that returns * KERN_FAILURE instead of blocking on map lock or memory allocation. */ int vm_map_lookup_locked(vm_map_t *var_map, /* IN/OUT */ vm_offset_t vaddr, vm_prot_t fault_typea, vm_map_entry_t *out_entry, /* OUT */ vm_object_t *object, /* OUT */ vm_pindex_t *pindex, /* OUT */ vm_prot_t *out_prot, /* OUT */ boolean_t *wired) /* OUT */ { vm_map_entry_t entry; vm_map_t map = *var_map; vm_prot_t prot; vm_prot_t fault_type = fault_typea; /* * Lookup the faulting address. */ if (!vm_map_lookup_entry(map, vaddr, out_entry)) return (KERN_INVALID_ADDRESS); entry = *out_entry; /* * Fail if the entry refers to a submap. */ if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) return (KERN_FAILURE); /* * Check whether this task is allowed to have this page. */ prot = entry->protection; fault_type &= VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE; if ((fault_type & prot) != fault_type) return (KERN_PROTECTION_FAILURE); /* * If this page is not pageable, we have to get it for all possible * accesses. */ *wired = (entry->wired_count != 0); if (*wired) fault_type = entry->protection; if (entry->eflags & MAP_ENTRY_NEEDS_COPY) { /* * Fail if the entry was copy-on-write for a write fault. */ if (fault_type & VM_PROT_WRITE) return (KERN_FAILURE); /* * We're attempting to read a copy-on-write page -- * don't allow writes. */ prot &= ~VM_PROT_WRITE; } /* * Fail if an object should be created. */ if (entry->object.vm_object == NULL && !map->system_map) return (KERN_FAILURE); /* * Return the object/offset from this entry. If the entry was * copy-on-write or empty, it has been fixed up. */ *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset); *object = entry->object.vm_object; *out_prot = prot; return (KERN_SUCCESS); } /* * vm_map_lookup_done: * * Releases locks acquired by a vm_map_lookup * (according to the handle returned by that lookup). */ void vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry) { /* * Unlock the main-level map */ vm_map_unlock_read(map); } vm_offset_t vm_map_max_KBI(const struct vm_map *map) { return (vm_map_max(map)); } vm_offset_t vm_map_min_KBI(const struct vm_map *map) { return (vm_map_min(map)); } pmap_t vm_map_pmap_KBI(vm_map_t map) { return (map->pmap); } #ifdef INVARIANTS static void _vm_map_assert_consistent(vm_map_t map) { vm_map_entry_t entry, prev; vm_size_t max_left, max_right; if (!enable_vmmap_check) return; prev = &map->header; VM_MAP_ENTRY_FOREACH(entry, map) { KASSERT(prev->end <= entry->start, ("map %p prev->end = %jx, start = %jx", map, (uintmax_t)prev->end, (uintmax_t)entry->start)); KASSERT(entry->start < entry->end, ("map %p start = %jx, end = %jx", map, (uintmax_t)entry->start, (uintmax_t)entry->end)); KASSERT(entry->end <= entry->next->start, ("map %p end = %jx, next->start = %jx", map, (uintmax_t)entry->end, (uintmax_t)entry->next->start)); KASSERT(entry->left == NULL || entry->left->start < entry->start, ("map %p left->start = %jx, start = %jx", map, (uintmax_t)entry->left->start, (uintmax_t)entry->start)); KASSERT(entry->right == NULL || entry->start < entry->right->start, ("map %p start = %jx, right->start = %jx", map, (uintmax_t)entry->start, (uintmax_t)entry->right->start)); max_left = vm_map_entry_max_free_left(entry, entry->prev); max_right = vm_map_entry_max_free_right(entry, entry->next); KASSERT(entry->max_free == MAX(max_left, max_right), ("map %p max = %jx, max_left = %jx, max_right = %jx", map, (uintmax_t)entry->max_free, (uintmax_t)max_left, (uintmax_t)max_right)); prev = entry; } KASSERT(prev->end <= entry->start, ("map %p prev->end = %jx, start = %jx", map, (uintmax_t)prev->end, (uintmax_t)entry->start)); } #endif #include "opt_ddb.h" #ifdef DDB #include #include static void vm_map_print(vm_map_t map) { vm_map_entry_t entry, prev; db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n", (void *)map, (void *)map->pmap, map->nentries, map->timestamp); db_indent += 2; prev = &map->header; VM_MAP_ENTRY_FOREACH(entry, map) { db_iprintf("map entry %p: start=%p, end=%p, eflags=%#x, \n", (void *)entry, (void *)entry->start, (void *)entry->end, entry->eflags); { static char *inheritance_name[4] = {"share", "copy", "none", "donate_copy"}; db_iprintf(" prot=%x/%x/%s", entry->protection, entry->max_protection, inheritance_name[(int)(unsigned char) entry->inheritance]); if (entry->wired_count != 0) db_printf(", wired"); } if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { db_printf(", share=%p, offset=0x%jx\n", (void *)entry->object.sub_map, (uintmax_t)entry->offset); if (prev == &map->header || prev->object.sub_map != entry->object.sub_map) { db_indent += 2; vm_map_print((vm_map_t)entry->object.sub_map); db_indent -= 2; } } else { if (entry->cred != NULL) db_printf(", ruid %d", entry->cred->cr_ruid); db_printf(", object=%p, offset=0x%jx", (void *)entry->object.vm_object, (uintmax_t)entry->offset); if (entry->object.vm_object && entry->object.vm_object->cred) db_printf(", obj ruid %d charge %jx", entry->object.vm_object->cred->cr_ruid, (uintmax_t)entry->object.vm_object->charge); if (entry->eflags & MAP_ENTRY_COW) db_printf(", copy (%s)", (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done"); db_printf("\n"); if (prev == &map->header || prev->object.vm_object != entry->object.vm_object) { db_indent += 2; vm_object_print((db_expr_t)(intptr_t) entry->object.vm_object, 0, 0, (char *)0); db_indent -= 2; } } prev = entry; } db_indent -= 2; } DB_SHOW_COMMAND(map, map) { if (!have_addr) { db_printf("usage: show map \n"); return; } vm_map_print((vm_map_t)addr); } DB_SHOW_COMMAND(procvm, procvm) { struct proc *p; if (have_addr) { p = db_lookup_proc(addr); } else { p = curproc; } db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n", (void *)p, (void *)p->p_vmspace, (void *)&p->p_vmspace->vm_map, (void *)vmspace_pmap(p->p_vmspace)); vm_map_print((vm_map_t)&p->p_vmspace->vm_map); } #endif /* DDB */ Index: head/sys/vm/vm_mmap.c =================================================================== --- head/sys/vm/vm_mmap.c (revision 353538) +++ head/sys/vm/vm_mmap.c (revision 353539) @@ -1,1647 +1,1647 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1988 University of Utah. * Copyright (c) 1991, 1993 * The Regents of the University of California. All rights reserved. * * This code is derived from software contributed to Berkeley by * the Systems Programming Group of the University of Utah Computer * Science Department. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: Utah $Hdr: vm_mmap.c 1.6 91/10/21$ * * @(#)vm_mmap.c 8.4 (Berkeley) 1/12/94 */ /* * Mapped file (mmap) interface to VM */ #include __FBSDID("$FreeBSD$"); #include "opt_hwpmc_hooks.h" #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #if defined(__amd64__) || defined(__i386__) /* for i386_read_exec */ #include #endif #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef HWPMC_HOOKS #include #endif int old_mlock = 0; SYSCTL_INT(_vm, OID_AUTO, old_mlock, CTLFLAG_RWTUN, &old_mlock, 0, "Do not apply RLIMIT_MEMLOCK on mlockall"); static int mincore_mapped = 1; SYSCTL_INT(_vm, OID_AUTO, mincore_mapped, CTLFLAG_RWTUN, &mincore_mapped, 0, "mincore reports mappings, not residency"); static int imply_prot_max = 0; SYSCTL_INT(_vm, OID_AUTO, imply_prot_max, CTLFLAG_RWTUN, &imply_prot_max, 0, "Imply maximum page permissions in mmap() when none are specified"); #ifdef MAP_32BIT #define MAP_32BIT_MAX_ADDR ((vm_offset_t)1 << 31) #endif #ifndef _SYS_SYSPROTO_H_ struct sbrk_args { int incr; }; #endif int sys_sbrk(struct thread *td, struct sbrk_args *uap) { /* Not yet implemented */ return (EOPNOTSUPP); } #ifndef _SYS_SYSPROTO_H_ struct sstk_args { int incr; }; #endif int sys_sstk(struct thread *td, struct sstk_args *uap) { /* Not yet implemented */ return (EOPNOTSUPP); } #if defined(COMPAT_43) int ogetpagesize(struct thread *td, struct ogetpagesize_args *uap) { td->td_retval[0] = PAGE_SIZE; return (0); } #endif /* COMPAT_43 */ /* * Memory Map (mmap) system call. Note that the file offset * and address are allowed to be NOT page aligned, though if * the MAP_FIXED flag it set, both must have the same remainder * modulo the PAGE_SIZE (POSIX 1003.1b). If the address is not * page-aligned, the actual mapping starts at trunc_page(addr) * and the return value is adjusted up by the page offset. * * Generally speaking, only character devices which are themselves * memory-based, such as a video framebuffer, can be mmap'd. Otherwise * there would be no cache coherency between a descriptor and a VM mapping * both to the same character device. */ #ifndef _SYS_SYSPROTO_H_ struct mmap_args { void *addr; size_t len; int prot; int flags; int fd; long pad; off_t pos; }; #endif int sys_mmap(struct thread *td, struct mmap_args *uap) { return (kern_mmap(td, (uintptr_t)uap->addr, uap->len, uap->prot, uap->flags, uap->fd, uap->pos)); } int kern_mmap_maxprot(struct proc *p, int prot) { if ((p->p_flag2 & P2_PROTMAX_DISABLE) != 0 || (p->p_fctl0 & NT_FREEBSD_FCTL_PROTMAX_DISABLE) != 0) return (_PROT_ALL); if (((p->p_flag2 & P2_PROTMAX_ENABLE) != 0 || imply_prot_max) && prot != PROT_NONE) return (prot); return (_PROT_ALL); } int kern_mmap(struct thread *td, uintptr_t addr0, size_t len, int prot, int flags, int fd, off_t pos) { struct vmspace *vms; struct file *fp; struct proc *p; vm_offset_t addr; vm_size_t pageoff, size; vm_prot_t cap_maxprot; int align, error, max_prot; cap_rights_t rights; if ((prot & ~(_PROT_ALL | PROT_MAX(_PROT_ALL))) != 0) return (EINVAL); max_prot = PROT_MAX_EXTRACT(prot); prot = PROT_EXTRACT(prot); if (max_prot != 0 && (max_prot & prot) != prot) return (EINVAL); p = td->td_proc; /* * Always honor PROT_MAX if set. If not, default to all * permissions unless we're implying maximum permissions. */ if (max_prot == 0) max_prot = kern_mmap_maxprot(p, prot); vms = p->p_vmspace; fp = NULL; AUDIT_ARG_FD(fd); addr = addr0; /* * Ignore old flags that used to be defined but did not do anything. */ flags &= ~(MAP_RESERVED0020 | MAP_RESERVED0040); /* * Enforce the constraints. * Mapping of length 0 is only allowed for old binaries. * Anonymous mapping shall specify -1 as filedescriptor and * zero position for new code. Be nice to ancient a.out * binaries and correct pos for anonymous mapping, since old * ld.so sometimes issues anonymous map requests with non-zero * pos. */ if (!SV_CURPROC_FLAG(SV_AOUT)) { if ((len == 0 && p->p_osrel >= P_OSREL_MAP_ANON) || ((flags & MAP_ANON) != 0 && (fd != -1 || pos != 0))) return (EINVAL); } else { if ((flags & MAP_ANON) != 0) pos = 0; } if (flags & MAP_STACK) { if ((fd != -1) || ((prot & (PROT_READ | PROT_WRITE)) != (PROT_READ | PROT_WRITE))) return (EINVAL); flags |= MAP_ANON; pos = 0; } if ((flags & ~(MAP_SHARED | MAP_PRIVATE | MAP_FIXED | MAP_HASSEMAPHORE | MAP_STACK | MAP_NOSYNC | MAP_ANON | MAP_EXCL | MAP_NOCORE | MAP_PREFAULT_READ | MAP_GUARD | #ifdef MAP_32BIT MAP_32BIT | #endif MAP_ALIGNMENT_MASK)) != 0) return (EINVAL); if ((flags & (MAP_EXCL | MAP_FIXED)) == MAP_EXCL) return (EINVAL); if ((flags & (MAP_SHARED | MAP_PRIVATE)) == (MAP_SHARED | MAP_PRIVATE)) return (EINVAL); if (prot != PROT_NONE && (prot & ~(PROT_READ | PROT_WRITE | PROT_EXEC)) != 0) return (EINVAL); if ((flags & MAP_GUARD) != 0 && (prot != PROT_NONE || fd != -1 || pos != 0 || (flags & ~(MAP_FIXED | MAP_GUARD | MAP_EXCL | #ifdef MAP_32BIT MAP_32BIT | #endif MAP_ALIGNMENT_MASK)) != 0)) return (EINVAL); /* * Align the file position to a page boundary, * and save its page offset component. */ pageoff = (pos & PAGE_MASK); pos -= pageoff; /* Compute size from len by rounding (on both ends). */ size = len + pageoff; /* low end... */ size = round_page(size); /* hi end */ /* Check for rounding up to zero. */ if (len > size) return (ENOMEM); /* Ensure alignment is at least a page and fits in a pointer. */ align = flags & MAP_ALIGNMENT_MASK; if (align != 0 && align != MAP_ALIGNED_SUPER && (align >> MAP_ALIGNMENT_SHIFT >= sizeof(void *) * NBBY || align >> MAP_ALIGNMENT_SHIFT < PAGE_SHIFT)) return (EINVAL); /* * Check for illegal addresses. Watch out for address wrap... Note * that VM_*_ADDRESS are not constants due to casts (argh). */ if (flags & MAP_FIXED) { /* * The specified address must have the same remainder * as the file offset taken modulo PAGE_SIZE, so it * should be aligned after adjustment by pageoff. */ addr -= pageoff; if (addr & PAGE_MASK) return (EINVAL); /* Address range must be all in user VM space. */ if (addr < vm_map_min(&vms->vm_map) || addr + size > vm_map_max(&vms->vm_map)) return (EINVAL); if (addr + size < addr) return (EINVAL); #ifdef MAP_32BIT if (flags & MAP_32BIT && addr + size > MAP_32BIT_MAX_ADDR) return (EINVAL); } else if (flags & MAP_32BIT) { /* * For MAP_32BIT, override the hint if it is too high and * do not bother moving the mapping past the heap (since * the heap is usually above 2GB). */ if (addr + size > MAP_32BIT_MAX_ADDR) addr = 0; #endif } else { /* * XXX for non-fixed mappings where no hint is provided or * the hint would fall in the potential heap space, * place it after the end of the largest possible heap. * * There should really be a pmap call to determine a reasonable * location. */ if (addr == 0 || (addr >= round_page((vm_offset_t)vms->vm_taddr) && addr < round_page((vm_offset_t)vms->vm_daddr + lim_max(td, RLIMIT_DATA)))) addr = round_page((vm_offset_t)vms->vm_daddr + lim_max(td, RLIMIT_DATA)); } if (len == 0) { /* * Return success without mapping anything for old * binaries that request a page-aligned mapping of * length 0. For modern binaries, this function * returns an error earlier. */ error = 0; } else if ((flags & MAP_GUARD) != 0) { error = vm_mmap_object(&vms->vm_map, &addr, size, VM_PROT_NONE, VM_PROT_NONE, flags, NULL, pos, FALSE, td); } else if ((flags & MAP_ANON) != 0) { /* * Mapping blank space is trivial. * * This relies on VM_PROT_* matching PROT_*. */ error = vm_mmap_object(&vms->vm_map, &addr, size, prot, max_prot, flags, NULL, pos, FALSE, td); } else { /* * Mapping file, get fp for validation and don't let the * descriptor disappear on us if we block. Check capability * rights, but also return the maximum rights to be combined * with maxprot later. */ cap_rights_init(&rights, CAP_MMAP); if (prot & PROT_READ) cap_rights_set(&rights, CAP_MMAP_R); if ((flags & MAP_SHARED) != 0) { if (prot & PROT_WRITE) cap_rights_set(&rights, CAP_MMAP_W); } if (prot & PROT_EXEC) cap_rights_set(&rights, CAP_MMAP_X); error = fget_mmap(td, fd, &rights, &cap_maxprot, &fp); if (error != 0) goto done; if ((flags & (MAP_SHARED | MAP_PRIVATE)) == 0 && p->p_osrel >= P_OSREL_MAP_FSTRICT) { error = EINVAL; goto done; } /* This relies on VM_PROT_* matching PROT_*. */ error = fo_mmap(fp, &vms->vm_map, &addr, size, prot, max_prot & cap_maxprot, flags, pos, td); } if (error == 0) td->td_retval[0] = (register_t) (addr + pageoff); done: if (fp) fdrop(fp, td); return (error); } #if defined(COMPAT_FREEBSD6) int freebsd6_mmap(struct thread *td, struct freebsd6_mmap_args *uap) { return (kern_mmap(td, (uintptr_t)uap->addr, uap->len, uap->prot, uap->flags, uap->fd, uap->pos)); } #endif #ifdef COMPAT_43 #ifndef _SYS_SYSPROTO_H_ struct ommap_args { caddr_t addr; int len; int prot; int flags; int fd; long pos; }; #endif int ommap(struct thread *td, struct ommap_args *uap) { static const char cvtbsdprot[8] = { 0, PROT_EXEC, PROT_WRITE, PROT_EXEC | PROT_WRITE, PROT_READ, PROT_EXEC | PROT_READ, PROT_WRITE | PROT_READ, PROT_EXEC | PROT_WRITE | PROT_READ, }; int flags, prot; #define OMAP_ANON 0x0002 #define OMAP_COPY 0x0020 #define OMAP_SHARED 0x0010 #define OMAP_FIXED 0x0100 prot = cvtbsdprot[uap->prot & 0x7]; #if (defined(COMPAT_FREEBSD32) && defined(__amd64__)) || defined(__i386__) if (i386_read_exec && SV_PROC_FLAG(td->td_proc, SV_ILP32) && prot != 0) prot |= PROT_EXEC; #endif flags = 0; if (uap->flags & OMAP_ANON) flags |= MAP_ANON; if (uap->flags & OMAP_COPY) flags |= MAP_COPY; if (uap->flags & OMAP_SHARED) flags |= MAP_SHARED; else flags |= MAP_PRIVATE; if (uap->flags & OMAP_FIXED) flags |= MAP_FIXED; return (kern_mmap(td, (uintptr_t)uap->addr, uap->len, prot, flags, uap->fd, uap->pos)); } #endif /* COMPAT_43 */ #ifndef _SYS_SYSPROTO_H_ struct msync_args { void *addr; size_t len; int flags; }; #endif int sys_msync(struct thread *td, struct msync_args *uap) { return (kern_msync(td, (uintptr_t)uap->addr, uap->len, uap->flags)); } int kern_msync(struct thread *td, uintptr_t addr0, size_t size, int flags) { vm_offset_t addr; vm_size_t pageoff; vm_map_t map; int rv; addr = addr0; pageoff = (addr & PAGE_MASK); addr -= pageoff; size += pageoff; size = (vm_size_t) round_page(size); if (addr + size < addr) return (EINVAL); if ((flags & (MS_ASYNC|MS_INVALIDATE)) == (MS_ASYNC|MS_INVALIDATE)) return (EINVAL); map = &td->td_proc->p_vmspace->vm_map; /* * Clean the pages and interpret the return value. */ rv = vm_map_sync(map, addr, addr + size, (flags & MS_ASYNC) == 0, (flags & MS_INVALIDATE) != 0); switch (rv) { case KERN_SUCCESS: return (0); case KERN_INVALID_ADDRESS: return (ENOMEM); case KERN_INVALID_ARGUMENT: return (EBUSY); case KERN_FAILURE: return (EIO); default: return (EINVAL); } } #ifndef _SYS_SYSPROTO_H_ struct munmap_args { void *addr; size_t len; }; #endif int sys_munmap(struct thread *td, struct munmap_args *uap) { return (kern_munmap(td, (uintptr_t)uap->addr, uap->len)); } int kern_munmap(struct thread *td, uintptr_t addr0, size_t size) { #ifdef HWPMC_HOOKS struct pmckern_map_out pkm; vm_map_entry_t entry; bool pmc_handled; #endif vm_offset_t addr; vm_size_t pageoff; vm_map_t map; if (size == 0) return (EINVAL); addr = addr0; pageoff = (addr & PAGE_MASK); addr -= pageoff; size += pageoff; size = (vm_size_t) round_page(size); if (addr + size < addr) return (EINVAL); /* * Check for illegal addresses. Watch out for address wrap... */ map = &td->td_proc->p_vmspace->vm_map; if (addr < vm_map_min(map) || addr + size > vm_map_max(map)) return (EINVAL); vm_map_lock(map); #ifdef HWPMC_HOOKS pmc_handled = false; if (PMC_HOOK_INSTALLED(PMC_FN_MUNMAP)) { pmc_handled = true; /* * Inform hwpmc if the address range being unmapped contains * an executable region. */ pkm.pm_address = (uintptr_t) NULL; if (vm_map_lookup_entry(map, addr, &entry)) { for (; entry->start < addr + size; entry = entry->next) { if (vm_map_check_protection(map, entry->start, entry->end, VM_PROT_EXECUTE) == TRUE) { pkm.pm_address = (uintptr_t) addr; pkm.pm_size = (size_t) size; break; } } } } #endif vm_map_delete(map, addr, addr + size); #ifdef HWPMC_HOOKS if (__predict_false(pmc_handled)) { /* downgrade the lock to prevent a LOR with the pmc-sx lock */ vm_map_lock_downgrade(map); if (pkm.pm_address != (uintptr_t) NULL) PMC_CALL_HOOK(td, PMC_FN_MUNMAP, (void *) &pkm); vm_map_unlock_read(map); } else #endif vm_map_unlock(map); /* vm_map_delete returns nothing but KERN_SUCCESS anyway */ return (0); } #ifndef _SYS_SYSPROTO_H_ struct mprotect_args { const void *addr; size_t len; int prot; }; #endif int sys_mprotect(struct thread *td, struct mprotect_args *uap) { return (kern_mprotect(td, (uintptr_t)uap->addr, uap->len, uap->prot)); } int kern_mprotect(struct thread *td, uintptr_t addr0, size_t size, int prot) { vm_offset_t addr; vm_size_t pageoff; int vm_error, max_prot; addr = addr0; if ((prot & ~(_PROT_ALL | PROT_MAX(_PROT_ALL))) != 0) return (EINVAL); max_prot = PROT_MAX_EXTRACT(prot); prot = PROT_EXTRACT(prot); pageoff = (addr & PAGE_MASK); addr -= pageoff; size += pageoff; size = (vm_size_t) round_page(size); #ifdef COMPAT_FREEBSD32 if (SV_PROC_FLAG(td->td_proc, SV_ILP32)) { if (((addr + size) & 0xffffffff) < addr) return (EINVAL); } else #endif if (addr + size < addr) return (EINVAL); vm_error = KERN_SUCCESS; if (max_prot != 0) { if ((max_prot & prot) != prot) return (EINVAL); vm_error = vm_map_protect(&td->td_proc->p_vmspace->vm_map, addr, addr + size, max_prot, TRUE); } if (vm_error == KERN_SUCCESS) vm_error = vm_map_protect(&td->td_proc->p_vmspace->vm_map, addr, addr + size, prot, FALSE); switch (vm_error) { case KERN_SUCCESS: return (0); case KERN_PROTECTION_FAILURE: return (EACCES); case KERN_RESOURCE_SHORTAGE: return (ENOMEM); } return (EINVAL); } #ifndef _SYS_SYSPROTO_H_ struct minherit_args { void *addr; size_t len; int inherit; }; #endif int sys_minherit(struct thread *td, struct minherit_args *uap) { vm_offset_t addr; vm_size_t size, pageoff; vm_inherit_t inherit; addr = (vm_offset_t)uap->addr; size = uap->len; inherit = uap->inherit; pageoff = (addr & PAGE_MASK); addr -= pageoff; size += pageoff; size = (vm_size_t) round_page(size); if (addr + size < addr) return (EINVAL); switch (vm_map_inherit(&td->td_proc->p_vmspace->vm_map, addr, addr + size, inherit)) { case KERN_SUCCESS: return (0); case KERN_PROTECTION_FAILURE: return (EACCES); } return (EINVAL); } #ifndef _SYS_SYSPROTO_H_ struct madvise_args { void *addr; size_t len; int behav; }; #endif int sys_madvise(struct thread *td, struct madvise_args *uap) { return (kern_madvise(td, (uintptr_t)uap->addr, uap->len, uap->behav)); } int kern_madvise(struct thread *td, uintptr_t addr0, size_t len, int behav) { vm_map_t map; vm_offset_t addr, end, start; int flags; /* * Check for our special case, advising the swap pager we are * "immortal." */ if (behav == MADV_PROTECT) { flags = PPROT_SET; return (kern_procctl(td, P_PID, td->td_proc->p_pid, PROC_SPROTECT, &flags)); } /* * Check for illegal addresses. Watch out for address wrap... Note * that VM_*_ADDRESS are not constants due to casts (argh). */ map = &td->td_proc->p_vmspace->vm_map; addr = addr0; if (addr < vm_map_min(map) || addr + len > vm_map_max(map)) return (EINVAL); if ((addr + len) < addr) return (EINVAL); /* * Since this routine is only advisory, we default to conservative * behavior. */ start = trunc_page(addr); end = round_page(addr + len); /* * vm_map_madvise() checks for illegal values of behav. */ return (vm_map_madvise(map, start, end, behav)); } #ifndef _SYS_SYSPROTO_H_ struct mincore_args { const void *addr; size_t len; char *vec; }; #endif int sys_mincore(struct thread *td, struct mincore_args *uap) { return (kern_mincore(td, (uintptr_t)uap->addr, uap->len, uap->vec)); } int kern_mincore(struct thread *td, uintptr_t addr0, size_t len, char *vec) { vm_offset_t addr, first_addr; vm_offset_t end, cend; pmap_t pmap; vm_map_t map; int error = 0; int vecindex, lastvecindex; vm_map_entry_t current; vm_map_entry_t entry; vm_object_t object; vm_paddr_t locked_pa; vm_page_t m; vm_pindex_t pindex; int mincoreinfo; unsigned int timestamp; boolean_t locked; /* * Make sure that the addresses presented are valid for user * mode. */ first_addr = addr = trunc_page(addr0); end = addr + (vm_size_t)round_page(len); map = &td->td_proc->p_vmspace->vm_map; if (end > vm_map_max(map) || end < addr) return (ENOMEM); pmap = vmspace_pmap(td->td_proc->p_vmspace); vm_map_lock_read(map); RestartScan: timestamp = map->timestamp; if (!vm_map_lookup_entry(map, addr, &entry)) { vm_map_unlock_read(map); return (ENOMEM); } /* * Do this on a map entry basis so that if the pages are not * in the current processes address space, we can easily look * up the pages elsewhere. */ lastvecindex = -1; for (current = entry; current->start < end; current = current->next) { /* * check for contiguity */ if (current->end < end && current->next->start > current->end) { vm_map_unlock_read(map); return (ENOMEM); } /* * ignore submaps (for now) or null objects */ if ((current->eflags & MAP_ENTRY_IS_SUB_MAP) || current->object.vm_object == NULL) continue; /* * limit this scan to the current map entry and the * limits for the mincore call */ if (addr < current->start) addr = current->start; cend = current->end; if (cend > end) cend = end; /* * scan this entry one page at a time */ while (addr < cend) { /* * Check pmap first, it is likely faster, also * it can provide info as to whether we are the * one referencing or modifying the page. */ object = NULL; locked_pa = 0; retry: m = NULL; mincoreinfo = pmap_mincore(pmap, addr, &locked_pa); if (mincore_mapped) { /* * We only care about this pmap's * mapping of the page, if any. */ if (locked_pa != 0) { vm_page_unlock(PHYS_TO_VM_PAGE( locked_pa)); } } else if (locked_pa != 0) { /* * The page is mapped by this process but not * both accessed and modified. It is also * managed. Acquire the object lock so that * other mappings might be examined. */ m = PHYS_TO_VM_PAGE(locked_pa); if (m->object != object) { if (object != NULL) VM_OBJECT_WUNLOCK(object); object = m->object; locked = VM_OBJECT_TRYWLOCK(object); vm_page_unlock(m); if (!locked) { VM_OBJECT_WLOCK(object); vm_page_lock(m); goto retry; } } else vm_page_unlock(m); - KASSERT(m->valid == VM_PAGE_BITS_ALL, + KASSERT(vm_page_all_valid(m), ("mincore: page %p is mapped but invalid", m)); } else if (mincoreinfo == 0) { /* * The page is not mapped by this process. If * the object implements managed pages, then * determine if the page is resident so that * the mappings might be examined. */ if (current->object.vm_object != object) { if (object != NULL) VM_OBJECT_WUNLOCK(object); object = current->object.vm_object; VM_OBJECT_WLOCK(object); } if (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP || object->type == OBJT_VNODE) { pindex = OFF_TO_IDX(current->offset + (addr - current->start)); m = vm_page_lookup(object, pindex); - if (m != NULL && m->valid == 0) + if (m != NULL && vm_page_none_valid(m)) m = NULL; if (m != NULL) mincoreinfo = MINCORE_INCORE; } } if (m != NULL) { /* Examine other mappings to the page. */ if (m->dirty == 0 && pmap_is_modified(m)) vm_page_dirty(m); if (m->dirty != 0) mincoreinfo |= MINCORE_MODIFIED_OTHER; /* * The first test for PGA_REFERENCED is an * optimization. The second test is * required because a concurrent pmap * operation could clear the last reference * and set PGA_REFERENCED before the call to * pmap_is_referenced(). */ if ((m->aflags & PGA_REFERENCED) != 0 || pmap_is_referenced(m) || (m->aflags & PGA_REFERENCED) != 0) mincoreinfo |= MINCORE_REFERENCED_OTHER; } if (object != NULL) VM_OBJECT_WUNLOCK(object); /* * subyte may page fault. In case it needs to modify * the map, we release the lock. */ vm_map_unlock_read(map); /* * calculate index into user supplied byte vector */ vecindex = atop(addr - first_addr); /* * If we have skipped map entries, we need to make sure that * the byte vector is zeroed for those skipped entries. */ while ((lastvecindex + 1) < vecindex) { ++lastvecindex; error = subyte(vec + lastvecindex, 0); if (error) { error = EFAULT; goto done2; } } /* * Pass the page information to the user */ error = subyte(vec + vecindex, mincoreinfo); if (error) { error = EFAULT; goto done2; } /* * If the map has changed, due to the subyte, the previous * output may be invalid. */ vm_map_lock_read(map); if (timestamp != map->timestamp) goto RestartScan; lastvecindex = vecindex; addr += PAGE_SIZE; } } /* * subyte may page fault. In case it needs to modify * the map, we release the lock. */ vm_map_unlock_read(map); /* * Zero the last entries in the byte vector. */ vecindex = atop(end - first_addr); while ((lastvecindex + 1) < vecindex) { ++lastvecindex; error = subyte(vec + lastvecindex, 0); if (error) { error = EFAULT; goto done2; } } /* * If the map has changed, due to the subyte, the previous * output may be invalid. */ vm_map_lock_read(map); if (timestamp != map->timestamp) goto RestartScan; vm_map_unlock_read(map); done2: return (error); } #ifndef _SYS_SYSPROTO_H_ struct mlock_args { const void *addr; size_t len; }; #endif int sys_mlock(struct thread *td, struct mlock_args *uap) { return (kern_mlock(td->td_proc, td->td_ucred, __DECONST(uintptr_t, uap->addr), uap->len)); } int kern_mlock(struct proc *proc, struct ucred *cred, uintptr_t addr0, size_t len) { vm_offset_t addr, end, last, start; vm_size_t npages, size; vm_map_t map; unsigned long nsize; int error; error = priv_check_cred(cred, PRIV_VM_MLOCK); if (error) return (error); addr = addr0; size = len; last = addr + size; start = trunc_page(addr); end = round_page(last); if (last < addr || end < addr) return (EINVAL); npages = atop(end - start); if (npages > vm_page_max_user_wired) return (ENOMEM); map = &proc->p_vmspace->vm_map; PROC_LOCK(proc); nsize = ptoa(npages + pmap_wired_count(map->pmap)); if (nsize > lim_cur_proc(proc, RLIMIT_MEMLOCK)) { PROC_UNLOCK(proc); return (ENOMEM); } PROC_UNLOCK(proc); #ifdef RACCT if (racct_enable) { PROC_LOCK(proc); error = racct_set(proc, RACCT_MEMLOCK, nsize); PROC_UNLOCK(proc); if (error != 0) return (ENOMEM); } #endif error = vm_map_wire(map, start, end, VM_MAP_WIRE_USER | VM_MAP_WIRE_NOHOLES); #ifdef RACCT if (racct_enable && error != KERN_SUCCESS) { PROC_LOCK(proc); racct_set(proc, RACCT_MEMLOCK, ptoa(pmap_wired_count(map->pmap))); PROC_UNLOCK(proc); } #endif return (error == KERN_SUCCESS ? 0 : ENOMEM); } #ifndef _SYS_SYSPROTO_H_ struct mlockall_args { int how; }; #endif int sys_mlockall(struct thread *td, struct mlockall_args *uap) { vm_map_t map; int error; map = &td->td_proc->p_vmspace->vm_map; error = priv_check(td, PRIV_VM_MLOCK); if (error) return (error); if ((uap->how == 0) || ((uap->how & ~(MCL_CURRENT|MCL_FUTURE)) != 0)) return (EINVAL); /* * If wiring all pages in the process would cause it to exceed * a hard resource limit, return ENOMEM. */ if (!old_mlock && uap->how & MCL_CURRENT) { if (map->size > lim_cur(td, RLIMIT_MEMLOCK)) return (ENOMEM); } #ifdef RACCT if (racct_enable) { PROC_LOCK(td->td_proc); error = racct_set(td->td_proc, RACCT_MEMLOCK, map->size); PROC_UNLOCK(td->td_proc); if (error != 0) return (ENOMEM); } #endif if (uap->how & MCL_FUTURE) { vm_map_lock(map); vm_map_modflags(map, MAP_WIREFUTURE, 0); vm_map_unlock(map); error = 0; } if (uap->how & MCL_CURRENT) { /* * P1003.1-2001 mandates that all currently mapped pages * will be memory resident and locked (wired) upon return * from mlockall(). vm_map_wire() will wire pages, by * calling vm_fault_wire() for each page in the region. */ error = vm_map_wire(map, vm_map_min(map), vm_map_max(map), VM_MAP_WIRE_USER|VM_MAP_WIRE_HOLESOK); if (error == KERN_SUCCESS) error = 0; else if (error == KERN_RESOURCE_SHORTAGE) error = ENOMEM; else error = EAGAIN; } #ifdef RACCT if (racct_enable && error != KERN_SUCCESS) { PROC_LOCK(td->td_proc); racct_set(td->td_proc, RACCT_MEMLOCK, ptoa(pmap_wired_count(map->pmap))); PROC_UNLOCK(td->td_proc); } #endif return (error); } #ifndef _SYS_SYSPROTO_H_ struct munlockall_args { register_t dummy; }; #endif int sys_munlockall(struct thread *td, struct munlockall_args *uap) { vm_map_t map; int error; map = &td->td_proc->p_vmspace->vm_map; error = priv_check(td, PRIV_VM_MUNLOCK); if (error) return (error); /* Clear the MAP_WIREFUTURE flag from this vm_map. */ vm_map_lock(map); vm_map_modflags(map, 0, MAP_WIREFUTURE); vm_map_unlock(map); /* Forcibly unwire all pages. */ error = vm_map_unwire(map, vm_map_min(map), vm_map_max(map), VM_MAP_WIRE_USER|VM_MAP_WIRE_HOLESOK); #ifdef RACCT if (racct_enable && error == KERN_SUCCESS) { PROC_LOCK(td->td_proc); racct_set(td->td_proc, RACCT_MEMLOCK, 0); PROC_UNLOCK(td->td_proc); } #endif return (error); } #ifndef _SYS_SYSPROTO_H_ struct munlock_args { const void *addr; size_t len; }; #endif int sys_munlock(struct thread *td, struct munlock_args *uap) { return (kern_munlock(td, (uintptr_t)uap->addr, uap->len)); } int kern_munlock(struct thread *td, uintptr_t addr0, size_t size) { vm_offset_t addr, end, last, start; #ifdef RACCT vm_map_t map; #endif int error; error = priv_check(td, PRIV_VM_MUNLOCK); if (error) return (error); addr = addr0; last = addr + size; start = trunc_page(addr); end = round_page(last); if (last < addr || end < addr) return (EINVAL); error = vm_map_unwire(&td->td_proc->p_vmspace->vm_map, start, end, VM_MAP_WIRE_USER | VM_MAP_WIRE_NOHOLES); #ifdef RACCT if (racct_enable && error == KERN_SUCCESS) { PROC_LOCK(td->td_proc); map = &td->td_proc->p_vmspace->vm_map; racct_set(td->td_proc, RACCT_MEMLOCK, ptoa(pmap_wired_count(map->pmap))); PROC_UNLOCK(td->td_proc); } #endif return (error == KERN_SUCCESS ? 0 : ENOMEM); } /* * vm_mmap_vnode() * * Helper function for vm_mmap. Perform sanity check specific for mmap * operations on vnodes. */ int vm_mmap_vnode(struct thread *td, vm_size_t objsize, vm_prot_t prot, vm_prot_t *maxprotp, int *flagsp, struct vnode *vp, vm_ooffset_t *foffp, vm_object_t *objp, boolean_t *writecounted) { struct vattr va; vm_object_t obj; vm_ooffset_t foff; struct ucred *cred; int error, flags; bool writex; cred = td->td_ucred; writex = (*maxprotp & VM_PROT_WRITE) != 0 && (*flagsp & MAP_SHARED) != 0; if ((error = vget(vp, LK_SHARED, td)) != 0) return (error); AUDIT_ARG_VNODE1(vp); foff = *foffp; flags = *flagsp; obj = vp->v_object; if (vp->v_type == VREG) { /* * Get the proper underlying object */ if (obj == NULL) { error = EINVAL; goto done; } if (obj->type == OBJT_VNODE && obj->handle != vp) { vput(vp); vp = (struct vnode *)obj->handle; /* * Bypass filesystems obey the mpsafety of the * underlying fs. Tmpfs never bypasses. */ error = vget(vp, LK_SHARED, td); if (error != 0) return (error); } if (writex) { *writecounted = TRUE; vm_pager_update_writecount(obj, 0, objsize); } } else { error = EINVAL; goto done; } if ((error = VOP_GETATTR(vp, &va, cred))) goto done; #ifdef MAC /* This relies on VM_PROT_* matching PROT_*. */ error = mac_vnode_check_mmap(cred, vp, (int)prot, flags); if (error != 0) goto done; #endif if ((flags & MAP_SHARED) != 0) { if ((va.va_flags & (SF_SNAPSHOT|IMMUTABLE|APPEND)) != 0) { if (prot & VM_PROT_WRITE) { error = EPERM; goto done; } *maxprotp &= ~VM_PROT_WRITE; } } /* * If it is a regular file without any references * we do not need to sync it. * Adjust object size to be the size of actual file. */ objsize = round_page(va.va_size); if (va.va_nlink == 0) flags |= MAP_NOSYNC; if (obj->type == OBJT_VNODE) { obj = vm_pager_allocate(OBJT_VNODE, vp, objsize, prot, foff, cred); if (obj == NULL) { error = ENOMEM; goto done; } } else { KASSERT(obj->type == OBJT_DEFAULT || obj->type == OBJT_SWAP, ("wrong object type")); VM_OBJECT_WLOCK(obj); vm_object_reference_locked(obj); #if VM_NRESERVLEVEL > 0 vm_object_color(obj, 0); #endif VM_OBJECT_WUNLOCK(obj); } *objp = obj; *flagsp = flags; vfs_mark_atime(vp, cred); done: if (error != 0 && *writecounted) { *writecounted = FALSE; vm_pager_update_writecount(obj, objsize, 0); } vput(vp); return (error); } /* * vm_mmap_cdev() * * Helper function for vm_mmap. Perform sanity check specific for mmap * operations on cdevs. */ int vm_mmap_cdev(struct thread *td, vm_size_t objsize, vm_prot_t prot, vm_prot_t *maxprotp, int *flagsp, struct cdev *cdev, struct cdevsw *dsw, vm_ooffset_t *foff, vm_object_t *objp) { vm_object_t obj; int error, flags; flags = *flagsp; if (dsw->d_flags & D_MMAP_ANON) { *objp = NULL; *foff = 0; *maxprotp = VM_PROT_ALL; *flagsp |= MAP_ANON; return (0); } /* * cdevs do not provide private mappings of any kind. */ if ((*maxprotp & VM_PROT_WRITE) == 0 && (prot & VM_PROT_WRITE) != 0) return (EACCES); if (flags & (MAP_PRIVATE|MAP_COPY)) return (EINVAL); /* * Force device mappings to be shared. */ flags |= MAP_SHARED; #ifdef MAC_XXX error = mac_cdev_check_mmap(td->td_ucred, cdev, (int)prot); if (error != 0) return (error); #endif /* * First, try d_mmap_single(). If that is not implemented * (returns ENODEV), fall back to using the device pager. * Note that d_mmap_single() must return a reference to the * object (it needs to bump the reference count of the object * it returns somehow). * * XXX assumes VM_PROT_* == PROT_* */ error = dsw->d_mmap_single(cdev, foff, objsize, objp, (int)prot); if (error != ENODEV) return (error); obj = vm_pager_allocate(OBJT_DEVICE, cdev, objsize, prot, *foff, td->td_ucred); if (obj == NULL) return (EINVAL); *objp = obj; *flagsp = flags; return (0); } /* * vm_mmap() * * Internal version of mmap used by exec, sys5 shared memory, and * various device drivers. Handle is either a vnode pointer, a * character device, or NULL for MAP_ANON. */ int vm_mmap(vm_map_t map, vm_offset_t *addr, vm_size_t size, vm_prot_t prot, vm_prot_t maxprot, int flags, objtype_t handle_type, void *handle, vm_ooffset_t foff) { vm_object_t object; struct thread *td = curthread; int error; boolean_t writecounted; if (size == 0) return (EINVAL); size = round_page(size); object = NULL; writecounted = FALSE; /* * Lookup/allocate object. */ switch (handle_type) { case OBJT_DEVICE: { struct cdevsw *dsw; struct cdev *cdev; int ref; cdev = handle; dsw = dev_refthread(cdev, &ref); if (dsw == NULL) return (ENXIO); error = vm_mmap_cdev(td, size, prot, &maxprot, &flags, cdev, dsw, &foff, &object); dev_relthread(cdev, ref); break; } case OBJT_VNODE: error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, handle, &foff, &object, &writecounted); break; case OBJT_DEFAULT: if (handle == NULL) { error = 0; break; } /* FALLTHROUGH */ default: error = EINVAL; break; } if (error) return (error); error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object, foff, writecounted, td); if (error != 0 && object != NULL) { /* * If this mapping was accounted for in the vnode's * writecount, then undo that now. */ if (writecounted) vm_pager_release_writecount(object, 0, size); vm_object_deallocate(object); } return (error); } /* * Internal version of mmap that maps a specific VM object into an * map. Called by mmap for MAP_ANON, vm_mmap, shm_mmap, and vn_mmap. */ int vm_mmap_object(vm_map_t map, vm_offset_t *addr, vm_size_t size, vm_prot_t prot, vm_prot_t maxprot, int flags, vm_object_t object, vm_ooffset_t foff, boolean_t writecounted, struct thread *td) { boolean_t curmap, fitit; vm_offset_t max_addr; int docow, error, findspace, rv; curmap = map == &td->td_proc->p_vmspace->vm_map; if (curmap) { RACCT_PROC_LOCK(td->td_proc); if (map->size + size > lim_cur(td, RLIMIT_VMEM)) { RACCT_PROC_UNLOCK(td->td_proc); return (ENOMEM); } if (racct_set(td->td_proc, RACCT_VMEM, map->size + size)) { RACCT_PROC_UNLOCK(td->td_proc); return (ENOMEM); } if (!old_mlock && map->flags & MAP_WIREFUTURE) { if (ptoa(pmap_wired_count(map->pmap)) + size > lim_cur(td, RLIMIT_MEMLOCK)) { racct_set_force(td->td_proc, RACCT_VMEM, map->size); RACCT_PROC_UNLOCK(td->td_proc); return (ENOMEM); } error = racct_set(td->td_proc, RACCT_MEMLOCK, ptoa(pmap_wired_count(map->pmap)) + size); if (error != 0) { racct_set_force(td->td_proc, RACCT_VMEM, map->size); RACCT_PROC_UNLOCK(td->td_proc); return (error); } } RACCT_PROC_UNLOCK(td->td_proc); } /* * We currently can only deal with page aligned file offsets. * The mmap() system call already enforces this by subtracting * the page offset from the file offset, but checking here * catches errors in device drivers (e.g. d_single_mmap() * callbacks) and other internal mapping requests (such as in * exec). */ if (foff & PAGE_MASK) return (EINVAL); if ((flags & MAP_FIXED) == 0) { fitit = TRUE; *addr = round_page(*addr); } else { if (*addr != trunc_page(*addr)) return (EINVAL); fitit = FALSE; } if (flags & MAP_ANON) { if (object != NULL || foff != 0) return (EINVAL); docow = 0; } else if (flags & MAP_PREFAULT_READ) docow = MAP_PREFAULT; else docow = MAP_PREFAULT_PARTIAL; if ((flags & (MAP_ANON|MAP_SHARED)) == 0) docow |= MAP_COPY_ON_WRITE; if (flags & MAP_NOSYNC) docow |= MAP_DISABLE_SYNCER; if (flags & MAP_NOCORE) docow |= MAP_DISABLE_COREDUMP; /* Shared memory is also shared with children. */ if (flags & MAP_SHARED) docow |= MAP_INHERIT_SHARE; if (writecounted) docow |= MAP_WRITECOUNT; if (flags & MAP_STACK) { if (object != NULL) return (EINVAL); docow |= MAP_STACK_GROWS_DOWN; } if ((flags & MAP_EXCL) != 0) docow |= MAP_CHECK_EXCL; if ((flags & MAP_GUARD) != 0) docow |= MAP_CREATE_GUARD; if (fitit) { if ((flags & MAP_ALIGNMENT_MASK) == MAP_ALIGNED_SUPER) findspace = VMFS_SUPER_SPACE; else if ((flags & MAP_ALIGNMENT_MASK) != 0) findspace = VMFS_ALIGNED_SPACE(flags >> MAP_ALIGNMENT_SHIFT); else findspace = VMFS_OPTIMAL_SPACE; max_addr = 0; #ifdef MAP_32BIT if ((flags & MAP_32BIT) != 0) max_addr = MAP_32BIT_MAX_ADDR; #endif if (curmap) { rv = vm_map_find_min(map, object, foff, addr, size, round_page((vm_offset_t)td->td_proc->p_vmspace-> vm_daddr + lim_max(td, RLIMIT_DATA)), max_addr, findspace, prot, maxprot, docow); } else { rv = vm_map_find(map, object, foff, addr, size, max_addr, findspace, prot, maxprot, docow); } } else { rv = vm_map_fixed(map, object, foff, *addr, size, prot, maxprot, docow); } if (rv == KERN_SUCCESS) { /* * If the process has requested that all future mappings * be wired, then heed this. */ if ((map->flags & MAP_WIREFUTURE) != 0) { vm_map_lock(map); if ((map->flags & MAP_WIREFUTURE) != 0) (void)vm_map_wire_locked(map, *addr, *addr + size, VM_MAP_WIRE_USER | ((flags & MAP_STACK) ? VM_MAP_WIRE_HOLESOK : VM_MAP_WIRE_NOHOLES)); vm_map_unlock(map); } } return (vm_mmap_to_errno(rv)); } /* * Translate a Mach VM return code to zero on success or the appropriate errno * on failure. */ int vm_mmap_to_errno(int rv) { switch (rv) { case KERN_SUCCESS: return (0); case KERN_INVALID_ADDRESS: case KERN_NO_SPACE: return (ENOMEM); case KERN_PROTECTION_FAILURE: return (EACCES); default: return (EINVAL); } } Index: head/sys/vm/vm_object.c =================================================================== --- head/sys/vm/vm_object.c (revision 353538) +++ head/sys/vm/vm_object.c (revision 353539) @@ -1,2656 +1,2661 @@ /*- * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU) * * Copyright (c) 1991, 1993 * The Regents of the University of California. All rights reserved. * * This code is derived from software contributed to Berkeley by * The Mach Operating System project at Carnegie-Mellon University. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94 * * * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Authors: Avadis Tevanian, Jr., Michael Wayne Young * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. */ /* * Virtual memory object module. */ #include __FBSDID("$FreeBSD$"); #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include /* for curproc, pageproc */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include static int old_msync; SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0, "Use old (insecure) msync behavior"); static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags, int flags, boolean_t *clearobjflags, boolean_t *eio); static boolean_t vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags); static void vm_object_qcollapse(vm_object_t object); static void vm_object_vndeallocate(vm_object_t object); /* * Virtual memory objects maintain the actual data * associated with allocated virtual memory. A given * page of memory exists within exactly one object. * * An object is only deallocated when all "references" * are given up. Only one "reference" to a given * region of an object should be writeable. * * Associated with each object is a list of all resident * memory pages belonging to that object; this list is * maintained by the "vm_page" module, and locked by the object's * lock. * * Each object also records a "pager" routine which is * used to retrieve (and store) pages to the proper backing * storage. In addition, objects may be backed by other * objects from which they were virtual-copied. * * The only items within the object structure which are * modified after time of creation are: * reference count locked by object's lock * pager routine locked by object's lock * */ struct object_q vm_object_list; struct mtx vm_object_list_mtx; /* lock for object list and count */ struct vm_object kernel_object_store; static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0, "VM object stats"); static counter_u64_t object_collapses = EARLY_COUNTER; SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD, &object_collapses, "VM object collapses"); static counter_u64_t object_bypasses = EARLY_COUNTER; SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD, &object_bypasses, "VM object bypasses"); static void counter_startup(void) { object_collapses = counter_u64_alloc(M_WAITOK); object_bypasses = counter_u64_alloc(M_WAITOK); } SYSINIT(object_counters, SI_SUB_CPU, SI_ORDER_ANY, counter_startup, NULL); static uma_zone_t obj_zone; static int vm_object_zinit(void *mem, int size, int flags); #ifdef INVARIANTS static void vm_object_zdtor(void *mem, int size, void *arg); static void vm_object_zdtor(void *mem, int size, void *arg) { vm_object_t object; object = (vm_object_t)mem; KASSERT(object->ref_count == 0, ("object %p ref_count = %d", object, object->ref_count)); KASSERT(TAILQ_EMPTY(&object->memq), ("object %p has resident pages in its memq", object)); KASSERT(vm_radix_is_empty(&object->rtree), ("object %p has resident pages in its trie", object)); #if VM_NRESERVLEVEL > 0 KASSERT(LIST_EMPTY(&object->rvq), ("object %p has reservations", object)); #endif KASSERT(REFCOUNT_COUNT(object->paging_in_progress) == 0, ("object %p paging_in_progress = %d", object, REFCOUNT_COUNT(object->paging_in_progress))); KASSERT(object->busy == 0, ("object %p busy = %d", object, object->busy)); KASSERT(object->resident_page_count == 0, ("object %p resident_page_count = %d", object, object->resident_page_count)); KASSERT(object->shadow_count == 0, ("object %p shadow_count = %d", object, object->shadow_count)); KASSERT(object->type == OBJT_DEAD, ("object %p has non-dead type %d", object, object->type)); } #endif static int vm_object_zinit(void *mem, int size, int flags) { vm_object_t object; object = (vm_object_t)mem; rw_init_flags(&object->lock, "vm object", RW_DUPOK | RW_NEW); /* These are true for any object that has been freed */ object->type = OBJT_DEAD; object->ref_count = 0; vm_radix_init(&object->rtree); refcount_init(&object->paging_in_progress, 0); refcount_init(&object->busy, 0); object->resident_page_count = 0; object->shadow_count = 0; object->flags = OBJ_DEAD; mtx_lock(&vm_object_list_mtx); TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); mtx_unlock(&vm_object_list_mtx); return (0); } static void _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object) { TAILQ_INIT(&object->memq); LIST_INIT(&object->shadow_head); object->type = type; if (type == OBJT_SWAP) pctrie_init(&object->un_pager.swp.swp_blks); /* * Ensure that swap_pager_swapoff() iteration over object_list * sees up to date type and pctrie head if it observed * non-dead object. */ atomic_thread_fence_rel(); switch (type) { case OBJT_DEAD: panic("_vm_object_allocate: can't create OBJT_DEAD"); case OBJT_DEFAULT: case OBJT_SWAP: object->flags = OBJ_ONEMAPPING; break; case OBJT_DEVICE: case OBJT_SG: object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED; break; case OBJT_MGTDEVICE: object->flags = OBJ_FICTITIOUS; break; case OBJT_PHYS: object->flags = OBJ_UNMANAGED; break; case OBJT_VNODE: object->flags = 0; break; default: panic("_vm_object_allocate: type %d is undefined", type); } object->size = size; object->domain.dr_policy = NULL; object->generation = 1; object->ref_count = 1; object->memattr = VM_MEMATTR_DEFAULT; object->cred = NULL; object->charge = 0; object->handle = NULL; object->backing_object = NULL; object->backing_object_offset = (vm_ooffset_t) 0; #if VM_NRESERVLEVEL > 0 LIST_INIT(&object->rvq); #endif umtx_shm_object_init(object); } /* * vm_object_init: * * Initialize the VM objects module. */ void vm_object_init(void) { TAILQ_INIT(&vm_object_list); mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); rw_init(&kernel_object->lock, "kernel vm object"); _vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), kernel_object); #if VM_NRESERVLEVEL > 0 kernel_object->flags |= OBJ_COLORED; kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); #endif /* * The lock portion of struct vm_object must be type stable due * to vm_pageout_fallback_object_lock locking a vm object * without holding any references to it. */ obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, #ifdef INVARIANTS vm_object_zdtor, #else NULL, #endif vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); vm_radix_zinit(); } void vm_object_clear_flag(vm_object_t object, u_short bits) { VM_OBJECT_ASSERT_WLOCKED(object); object->flags &= ~bits; } /* * Sets the default memory attribute for the specified object. Pages * that are allocated to this object are by default assigned this memory * attribute. * * Presently, this function must be called before any pages are allocated * to the object. In the future, this requirement may be relaxed for * "default" and "swap" objects. */ int vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr) { VM_OBJECT_ASSERT_WLOCKED(object); switch (object->type) { case OBJT_DEFAULT: case OBJT_DEVICE: case OBJT_MGTDEVICE: case OBJT_PHYS: case OBJT_SG: case OBJT_SWAP: case OBJT_VNODE: if (!TAILQ_EMPTY(&object->memq)) return (KERN_FAILURE); break; case OBJT_DEAD: return (KERN_INVALID_ARGUMENT); default: panic("vm_object_set_memattr: object %p is of undefined type", object); } object->memattr = memattr; return (KERN_SUCCESS); } void vm_object_pip_add(vm_object_t object, short i) { refcount_acquiren(&object->paging_in_progress, i); } void vm_object_pip_wakeup(vm_object_t object) { refcount_release(&object->paging_in_progress); } void vm_object_pip_wakeupn(vm_object_t object, short i) { refcount_releasen(&object->paging_in_progress, i); } void vm_object_pip_wait(vm_object_t object, char *waitid) { VM_OBJECT_ASSERT_WLOCKED(object); while (REFCOUNT_COUNT(object->paging_in_progress) > 0) { VM_OBJECT_WUNLOCK(object); refcount_wait(&object->paging_in_progress, waitid, PVM); VM_OBJECT_WLOCK(object); } } void vm_object_pip_wait_unlocked(vm_object_t object, char *waitid) { VM_OBJECT_ASSERT_UNLOCKED(object); while (REFCOUNT_COUNT(object->paging_in_progress) > 0) refcount_wait(&object->paging_in_progress, waitid, PVM); } /* * vm_object_allocate: * * Returns a new object with the given size. */ vm_object_t vm_object_allocate(objtype_t type, vm_pindex_t size) { vm_object_t object; object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK); _vm_object_allocate(type, size, object); return (object); } /* * vm_object_reference: * * Gets another reference to the given object. Note: OBJ_DEAD * objects can be referenced during final cleaning. */ void vm_object_reference(vm_object_t object) { if (object == NULL) return; VM_OBJECT_WLOCK(object); vm_object_reference_locked(object); VM_OBJECT_WUNLOCK(object); } /* * vm_object_reference_locked: * * Gets another reference to the given object. * * The object must be locked. */ void vm_object_reference_locked(vm_object_t object) { struct vnode *vp; VM_OBJECT_ASSERT_WLOCKED(object); object->ref_count++; if (object->type == OBJT_VNODE) { vp = object->handle; vref(vp); } } /* * Handle deallocating an object of type OBJT_VNODE. */ static void vm_object_vndeallocate(vm_object_t object) { struct vnode *vp = (struct vnode *) object->handle; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(object->type == OBJT_VNODE, ("vm_object_vndeallocate: not a vnode object")); KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); #ifdef INVARIANTS if (object->ref_count == 0) { vn_printf(vp, "vm_object_vndeallocate "); panic("vm_object_vndeallocate: bad object reference count"); } #endif if (!umtx_shm_vnobj_persistent && object->ref_count == 1) umtx_shm_object_terminated(object); object->ref_count--; /* vrele may need the vnode lock. */ VM_OBJECT_WUNLOCK(object); vrele(vp); } /* * vm_object_deallocate: * * Release a reference to the specified object, * gained either through a vm_object_allocate * or a vm_object_reference call. When all references * are gone, storage associated with this object * may be relinquished. * * No object may be locked. */ void vm_object_deallocate(vm_object_t object) { vm_object_t temp; while (object != NULL) { VM_OBJECT_WLOCK(object); if (object->type == OBJT_VNODE) { vm_object_vndeallocate(object); return; } KASSERT(object->ref_count != 0, ("vm_object_deallocate: object deallocated too many times: %d", object->type)); /* * If the reference count goes to 0 we start calling * vm_object_terminate() on the object chain. * A ref count of 1 may be a special case depending on the * shadow count being 0 or 1. */ object->ref_count--; if (object->ref_count > 1) { VM_OBJECT_WUNLOCK(object); return; } else if (object->ref_count == 1) { if (object->shadow_count == 0 && object->handle == NULL && (object->type == OBJT_DEFAULT || (object->type == OBJT_SWAP && (object->flags & OBJ_TMPFS_NODE) == 0))) { vm_object_set_flag(object, OBJ_ONEMAPPING); } else if ((object->shadow_count == 1) && (object->handle == NULL) && (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { vm_object_t robject; robject = LIST_FIRST(&object->shadow_head); KASSERT(robject != NULL, ("vm_object_deallocate: ref_count: %d, shadow_count: %d", object->ref_count, object->shadow_count)); KASSERT((robject->flags & OBJ_TMPFS_NODE) == 0, ("shadowed tmpfs v_object %p", object)); if (!VM_OBJECT_TRYWLOCK(robject)) { /* * Avoid a potential deadlock. */ object->ref_count++; VM_OBJECT_WUNLOCK(object); /* * More likely than not the thread * holding robject's lock has lower * priority than the current thread. * Let the lower priority thread run. */ pause("vmo_de", 1); continue; } /* * Collapse object into its shadow unless its * shadow is dead. In that case, object will * be deallocated by the thread that is * deallocating its shadow. */ if ((robject->flags & OBJ_DEAD) == 0 && (robject->handle == NULL) && (robject->type == OBJT_DEFAULT || robject->type == OBJT_SWAP)) { robject->ref_count++; retry: if (REFCOUNT_COUNT(robject->paging_in_progress) > 0) { VM_OBJECT_WUNLOCK(object); vm_object_pip_wait(robject, "objde1"); temp = robject->backing_object; if (object == temp) { VM_OBJECT_WLOCK(object); goto retry; } } else if (REFCOUNT_COUNT(object->paging_in_progress) > 0) { VM_OBJECT_WUNLOCK(robject); VM_OBJECT_WUNLOCK(object); refcount_wait( &object->paging_in_progress, "objde2", PVM); VM_OBJECT_WLOCK(robject); temp = robject->backing_object; if (object == temp) { VM_OBJECT_WLOCK(object); goto retry; } } else VM_OBJECT_WUNLOCK(object); if (robject->ref_count == 1) { robject->ref_count--; object = robject; goto doterm; } object = robject; vm_object_collapse(object); VM_OBJECT_WUNLOCK(object); continue; } VM_OBJECT_WUNLOCK(robject); } VM_OBJECT_WUNLOCK(object); return; } doterm: umtx_shm_object_terminated(object); temp = object->backing_object; if (temp != NULL) { KASSERT((object->flags & OBJ_TMPFS_NODE) == 0, ("shadowed tmpfs v_object 2 %p", object)); VM_OBJECT_WLOCK(temp); LIST_REMOVE(object, shadow_list); temp->shadow_count--; VM_OBJECT_WUNLOCK(temp); object->backing_object = NULL; } /* * Don't double-terminate, we could be in a termination * recursion due to the terminate having to sync data * to disk. */ if ((object->flags & OBJ_DEAD) == 0) { vm_object_set_flag(object, OBJ_DEAD); vm_object_terminate(object); } else VM_OBJECT_WUNLOCK(object); object = temp; } } /* * vm_object_destroy removes the object from the global object list * and frees the space for the object. */ void vm_object_destroy(vm_object_t object) { /* * Release the allocation charge. */ if (object->cred != NULL) { swap_release_by_cred(object->charge, object->cred); object->charge = 0; crfree(object->cred); object->cred = NULL; } /* * Free the space for the object. */ uma_zfree(obj_zone, object); } /* * vm_object_terminate_pages removes any remaining pageable pages * from the object and resets the object to an empty state. */ static void vm_object_terminate_pages(vm_object_t object) { vm_page_t p, p_next; VM_OBJECT_ASSERT_WLOCKED(object); /* * Free any remaining pageable pages. This also removes them from the * paging queues. However, don't free wired pages, just remove them * from the object. Rather than incrementally removing each page from * the object, the page and object are reset to any empty state. */ TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) { vm_page_assert_unbusied(p); KASSERT(p->object == object && (p->ref_count & VPRC_OBJREF) != 0, ("vm_object_terminate_pages: page %p is inconsistent", p)); p->object = NULL; if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) { VM_CNT_INC(v_pfree); vm_page_free(p); } } /* * If the object contained any pages, then reset it to an empty state. * None of the object's fields, including "resident_page_count", were * modified by the preceding loop. */ if (object->resident_page_count != 0) { vm_radix_reclaim_allnodes(&object->rtree); TAILQ_INIT(&object->memq); object->resident_page_count = 0; if (object->type == OBJT_VNODE) vdrop(object->handle); } } /* * vm_object_terminate actually destroys the specified object, freeing * up all previously used resources. * * The object must be locked. * This routine may block. */ void vm_object_terminate(vm_object_t object) { VM_OBJECT_ASSERT_WLOCKED(object); KASSERT((object->flags & OBJ_DEAD) != 0, ("terminating non-dead obj %p", object)); /* * wait for the pageout daemon to be done with the object */ vm_object_pip_wait(object, "objtrm"); KASSERT(!REFCOUNT_COUNT(object->paging_in_progress), ("vm_object_terminate: pageout in progress")); KASSERT(object->ref_count == 0, ("vm_object_terminate: object with references, ref_count=%d", object->ref_count)); if ((object->flags & OBJ_PG_DTOR) == 0) vm_object_terminate_pages(object); #if VM_NRESERVLEVEL > 0 if (__predict_false(!LIST_EMPTY(&object->rvq))) vm_reserv_break_all(object); #endif KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT || object->type == OBJT_SWAP, ("%s: non-swap obj %p has cred", __func__, object)); /* * Let the pager know object is dead. */ vm_pager_deallocate(object); VM_OBJECT_WUNLOCK(object); vm_object_destroy(object); } /* * Make the page read-only so that we can clear the object flags. However, if * this is a nosync mmap then the object is likely to stay dirty so do not * mess with the page and do not clear the object flags. Returns TRUE if the * page should be flushed, and FALSE otherwise. */ static boolean_t vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags) { /* * If we have been asked to skip nosync pages and this is a * nosync page, skip it. Note that the object flags were not * cleared in this case so we do not have to set them. */ if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) { *clearobjflags = FALSE; return (FALSE); } else { pmap_remove_write(p); return (p->dirty != 0); } } /* * vm_object_page_clean * * Clean all dirty pages in the specified range of object. Leaves page * on whatever queue it is currently on. If NOSYNC is set then do not * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC), * leaving the object dirty. * * When stuffing pages asynchronously, allow clustering. XXX we need a * synchronous clustering mode implementation. * * Odd semantics: if start == end, we clean everything. * * The object must be locked. * * Returns FALSE if some page from the range was not written, as * reported by the pager, and TRUE otherwise. */ boolean_t vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end, int flags) { vm_page_t np, p; vm_pindex_t pi, tend, tstart; int curgeneration, n, pagerflags; boolean_t clearobjflags, eio, res; VM_OBJECT_ASSERT_WLOCKED(object); /* * The OBJ_MIGHTBEDIRTY flag is only set for OBJT_VNODE * objects. The check below prevents the function from * operating on non-vnode objects. */ if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 || object->resident_page_count == 0) return (TRUE); pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0; tstart = OFF_TO_IDX(start); tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK); clearobjflags = tstart == 0 && tend >= object->size; res = TRUE; rescan: curgeneration = object->generation; for (p = vm_page_find_least(object, tstart); p != NULL; p = np) { pi = p->pindex; if (pi >= tend) break; np = TAILQ_NEXT(p, listq); - if (p->valid == 0) + if (vm_page_none_valid(p)) continue; if (vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL) == 0) { if (object->generation != curgeneration) { if ((flags & OBJPC_SYNC) != 0) goto rescan; else clearobjflags = FALSE; } np = vm_page_find_least(object, pi); continue; } if (!vm_object_page_remove_write(p, flags, &clearobjflags)) { vm_page_xunbusy(p); continue; } n = vm_object_page_collect_flush(object, p, pagerflags, flags, &clearobjflags, &eio); if (eio) { res = FALSE; clearobjflags = FALSE; } if (object->generation != curgeneration) { if ((flags & OBJPC_SYNC) != 0) goto rescan; else clearobjflags = FALSE; } /* * If the VOP_PUTPAGES() did a truncated write, so * that even the first page of the run is not fully * written, vm_pageout_flush() returns 0 as the run * length. Since the condition that caused truncated * write may be permanent, e.g. exhausted free space, * accepting n == 0 would cause an infinite loop. * * Forwarding the iterator leaves the unwritten page * behind, but there is not much we can do there if * filesystem refuses to write it. */ if (n == 0) { n = 1; clearobjflags = FALSE; } np = vm_page_find_least(object, pi + n); } #if 0 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0); #endif if (clearobjflags) vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY); return (res); } static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags, int flags, boolean_t *clearobjflags, boolean_t *eio) { vm_page_t ma[vm_pageout_page_count], p_first, tp; int count, i, mreq, runlen; vm_page_lock_assert(p, MA_NOTOWNED); vm_page_assert_xbusied(p); VM_OBJECT_ASSERT_WLOCKED(object); count = 1; mreq = 0; for (tp = p; count < vm_pageout_page_count; count++) { tp = vm_page_next(tp); if (tp == NULL || vm_page_tryxbusy(tp) == 0) break; if (!vm_object_page_remove_write(tp, flags, clearobjflags)) { vm_page_xunbusy(tp); break; } } for (p_first = p; count < vm_pageout_page_count; count++) { tp = vm_page_prev(p_first); if (tp == NULL || vm_page_tryxbusy(tp) == 0) break; if (!vm_object_page_remove_write(tp, flags, clearobjflags)) { vm_page_xunbusy(tp); break; } p_first = tp; mreq++; } for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++) ma[i] = tp; vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio); return (runlen); } /* * Note that there is absolutely no sense in writing out * anonymous objects, so we track down the vnode object * to write out. * We invalidate (remove) all pages from the address space * for semantic correctness. * * If the backing object is a device object with unmanaged pages, then any * mappings to the specified range of pages must be removed before this * function is called. * * Note: certain anonymous maps, such as MAP_NOSYNC maps, * may start out with a NULL object. */ boolean_t vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size, boolean_t syncio, boolean_t invalidate) { vm_object_t backing_object; struct vnode *vp; struct mount *mp; int error, flags, fsync_after; boolean_t res; if (object == NULL) return (TRUE); res = TRUE; error = 0; VM_OBJECT_WLOCK(object); while ((backing_object = object->backing_object) != NULL) { VM_OBJECT_WLOCK(backing_object); offset += object->backing_object_offset; VM_OBJECT_WUNLOCK(object); object = backing_object; if (object->size < OFF_TO_IDX(offset + size)) size = IDX_TO_OFF(object->size) - offset; } /* * Flush pages if writing is allowed, invalidate them * if invalidation requested. Pages undergoing I/O * will be ignored by vm_object_page_remove(). * * We cannot lock the vnode and then wait for paging * to complete without deadlocking against vm_fault. * Instead we simply call vm_object_page_remove() and * allow it to block internally on a page-by-page * basis when it encounters pages undergoing async * I/O. */ if (object->type == OBJT_VNODE && (object->flags & OBJ_MIGHTBEDIRTY) != 0 && ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) { VM_OBJECT_WUNLOCK(object); (void) vn_start_write(vp, &mp, V_WAIT); vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); if (syncio && !invalidate && offset == 0 && atop(size) == object->size) { /* * If syncing the whole mapping of the file, * it is faster to schedule all the writes in * async mode, also allowing the clustering, * and then wait for i/o to complete. */ flags = 0; fsync_after = TRUE; } else { flags = (syncio || invalidate) ? OBJPC_SYNC : 0; flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0; fsync_after = FALSE; } VM_OBJECT_WLOCK(object); res = vm_object_page_clean(object, offset, offset + size, flags); VM_OBJECT_WUNLOCK(object); if (fsync_after) error = VOP_FSYNC(vp, MNT_WAIT, curthread); VOP_UNLOCK(vp, 0); vn_finished_write(mp); if (error != 0) res = FALSE; VM_OBJECT_WLOCK(object); } if ((object->type == OBJT_VNODE || object->type == OBJT_DEVICE) && invalidate) { if (object->type == OBJT_DEVICE) /* * The option OBJPR_NOTMAPPED must be passed here * because vm_object_page_remove() cannot remove * unmanaged mappings. */ flags = OBJPR_NOTMAPPED; else if (old_msync) flags = 0; else flags = OBJPR_CLEANONLY; vm_object_page_remove(object, OFF_TO_IDX(offset), OFF_TO_IDX(offset + size + PAGE_MASK), flags); } VM_OBJECT_WUNLOCK(object); return (res); } /* * Determine whether the given advice can be applied to the object. Advice is * not applied to unmanaged pages since they never belong to page queues, and * since MADV_FREE is destructive, it can apply only to anonymous pages that * have been mapped at most once. */ static bool vm_object_advice_applies(vm_object_t object, int advice) { if ((object->flags & OBJ_UNMANAGED) != 0) return (false); if (advice != MADV_FREE) return (true); return ((object->type == OBJT_DEFAULT || object->type == OBJT_SWAP) && (object->flags & OBJ_ONEMAPPING) != 0); } static void vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex, vm_size_t size) { if (advice == MADV_FREE && object->type == OBJT_SWAP) swap_pager_freespace(object, pindex, size); } /* * vm_object_madvise: * * Implements the madvise function at the object/page level. * * MADV_WILLNEED (any object) * * Activate the specified pages if they are resident. * * MADV_DONTNEED (any object) * * Deactivate the specified pages if they are resident. * * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, * OBJ_ONEMAPPING only) * * Deactivate and clean the specified pages if they are * resident. This permits the process to reuse the pages * without faulting or the kernel to reclaim the pages * without I/O. */ void vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end, int advice) { vm_pindex_t tpindex; vm_object_t backing_object, tobject; vm_page_t m, tm; if (object == NULL) return; relookup: VM_OBJECT_WLOCK(object); if (!vm_object_advice_applies(object, advice)) { VM_OBJECT_WUNLOCK(object); return; } for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) { tobject = object; /* * If the next page isn't resident in the top-level object, we * need to search the shadow chain. When applying MADV_FREE, we * take care to release any swap space used to store * non-resident pages. */ if (m == NULL || pindex < m->pindex) { /* * Optimize a common case: if the top-level object has * no backing object, we can skip over the non-resident * range in constant time. */ if (object->backing_object == NULL) { tpindex = (m != NULL && m->pindex < end) ? m->pindex : end; vm_object_madvise_freespace(object, advice, pindex, tpindex - pindex); if ((pindex = tpindex) == end) break; goto next_page; } tpindex = pindex; do { vm_object_madvise_freespace(tobject, advice, tpindex, 1); /* * Prepare to search the next object in the * chain. */ backing_object = tobject->backing_object; if (backing_object == NULL) goto next_pindex; VM_OBJECT_WLOCK(backing_object); tpindex += OFF_TO_IDX(tobject->backing_object_offset); if (tobject != object) VM_OBJECT_WUNLOCK(tobject); tobject = backing_object; if (!vm_object_advice_applies(tobject, advice)) goto next_pindex; } while ((tm = vm_page_lookup(tobject, tpindex)) == NULL); } else { next_page: tm = m; m = TAILQ_NEXT(m, listq); } /* - * If the page is not in a normal state, skip it. + * If the page is not in a normal state, skip it. The page + * can not be invalidated while the object lock is held. */ - if (tm->valid != VM_PAGE_BITS_ALL || - vm_page_wired(tm)) + if (!vm_page_all_valid(tm) || vm_page_wired(tm)) goto next_pindex; KASSERT((tm->flags & PG_FICTITIOUS) == 0, ("vm_object_madvise: page %p is fictitious", tm)); KASSERT((tm->oflags & VPO_UNMANAGED) == 0, ("vm_object_madvise: page %p is not managed", tm)); if (vm_page_tryxbusy(tm) == 0) { if (object != tobject) VM_OBJECT_WUNLOCK(object); if (advice == MADV_WILLNEED) { /* * Reference the page before unlocking and * sleeping so that the page daemon is less * likely to reclaim it. */ vm_page_aflag_set(tm, PGA_REFERENCED); } vm_page_busy_sleep(tm, "madvpo", false); goto relookup; } vm_page_lock(tm); vm_page_advise(tm, advice); vm_page_unlock(tm); vm_page_xunbusy(tm); vm_object_madvise_freespace(tobject, advice, tm->pindex, 1); next_pindex: if (tobject != object) VM_OBJECT_WUNLOCK(tobject); } VM_OBJECT_WUNLOCK(object); } /* * vm_object_shadow: * * Create a new object which is backed by the * specified existing object range. The source * object reference is deallocated. * * The new object and offset into that object * are returned in the source parameters. */ void vm_object_shadow( vm_object_t *object, /* IN/OUT */ vm_ooffset_t *offset, /* IN/OUT */ vm_size_t length) { vm_object_t source; vm_object_t result; source = *object; /* * Don't create the new object if the old object isn't shared. */ if (source != NULL) { VM_OBJECT_WLOCK(source); if (source->ref_count == 1 && source->handle == NULL && (source->type == OBJT_DEFAULT || source->type == OBJT_SWAP)) { VM_OBJECT_WUNLOCK(source); return; } VM_OBJECT_WUNLOCK(source); } /* * Allocate a new object with the given length. */ result = vm_object_allocate(OBJT_DEFAULT, atop(length)); /* * The new object shadows the source object, adding a reference to it. * Our caller changes his reference to point to the new object, * removing a reference to the source object. Net result: no change * of reference count. * * Try to optimize the result object's page color when shadowing * in order to maintain page coloring consistency in the combined * shadowed object. */ result->backing_object = source; /* * Store the offset into the source object, and fix up the offset into * the new object. */ result->backing_object_offset = *offset; if (source != NULL) { VM_OBJECT_WLOCK(source); result->domain = source->domain; LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list); source->shadow_count++; #if VM_NRESERVLEVEL > 0 result->flags |= source->flags & OBJ_COLORED; result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER - 1)) - 1); #endif VM_OBJECT_WUNLOCK(source); } /* * Return the new things */ *offset = 0; *object = result; } /* * vm_object_split: * * Split the pages in a map entry into a new object. This affords * easier removal of unused pages, and keeps object inheritance from * being a negative impact on memory usage. */ void vm_object_split(vm_map_entry_t entry) { vm_page_t m, m_next; vm_object_t orig_object, new_object, source; vm_pindex_t idx, offidxstart; vm_size_t size; orig_object = entry->object.vm_object; if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP) return; if (orig_object->ref_count <= 1) return; VM_OBJECT_WUNLOCK(orig_object); offidxstart = OFF_TO_IDX(entry->offset); size = atop(entry->end - entry->start); /* * If swap_pager_copy() is later called, it will convert new_object * into a swap object. */ new_object = vm_object_allocate(OBJT_DEFAULT, size); /* * At this point, the new object is still private, so the order in * which the original and new objects are locked does not matter. */ VM_OBJECT_WLOCK(new_object); VM_OBJECT_WLOCK(orig_object); new_object->domain = orig_object->domain; source = orig_object->backing_object; if (source != NULL) { VM_OBJECT_WLOCK(source); if ((source->flags & OBJ_DEAD) != 0) { VM_OBJECT_WUNLOCK(source); VM_OBJECT_WUNLOCK(orig_object); VM_OBJECT_WUNLOCK(new_object); vm_object_deallocate(new_object); VM_OBJECT_WLOCK(orig_object); return; } LIST_INSERT_HEAD(&source->shadow_head, new_object, shadow_list); source->shadow_count++; vm_object_reference_locked(source); /* for new_object */ vm_object_clear_flag(source, OBJ_ONEMAPPING); VM_OBJECT_WUNLOCK(source); new_object->backing_object_offset = orig_object->backing_object_offset + entry->offset; new_object->backing_object = source; } if (orig_object->cred != NULL) { new_object->cred = orig_object->cred; crhold(orig_object->cred); new_object->charge = ptoa(size); KASSERT(orig_object->charge >= ptoa(size), ("orig_object->charge < 0")); orig_object->charge -= ptoa(size); } retry: m = vm_page_find_least(orig_object, offidxstart); for (; m != NULL && (idx = m->pindex - offidxstart) < size; m = m_next) { m_next = TAILQ_NEXT(m, listq); /* * We must wait for pending I/O to complete before we can * rename the page. * * We do not have to VM_PROT_NONE the page as mappings should * not be changed by this operation. */ if (vm_page_tryxbusy(m) == 0) { VM_OBJECT_WUNLOCK(new_object); vm_page_sleep_if_busy(m, "spltwt"); VM_OBJECT_WLOCK(new_object); goto retry; } /* vm_page_rename() will dirty the page. */ if (vm_page_rename(m, new_object, idx)) { vm_page_xunbusy(m); VM_OBJECT_WUNLOCK(new_object); VM_OBJECT_WUNLOCK(orig_object); vm_radix_wait(); VM_OBJECT_WLOCK(orig_object); VM_OBJECT_WLOCK(new_object); goto retry; } #if VM_NRESERVLEVEL > 0 /* * If some of the reservation's allocated pages remain with * the original object, then transferring the reservation to * the new object is neither particularly beneficial nor * particularly harmful as compared to leaving the reservation * with the original object. If, however, all of the * reservation's allocated pages are transferred to the new * object, then transferring the reservation is typically * beneficial. Determining which of these two cases applies * would be more costly than unconditionally renaming the * reservation. */ vm_reserv_rename(m, new_object, orig_object, offidxstart); #endif if (orig_object->type != OBJT_SWAP) vm_page_xunbusy(m); } if (orig_object->type == OBJT_SWAP) { /* * swap_pager_copy() can sleep, in which case the orig_object's * and new_object's locks are released and reacquired. */ swap_pager_copy(orig_object, new_object, offidxstart, 0); TAILQ_FOREACH(m, &new_object->memq, listq) vm_page_xunbusy(m); } VM_OBJECT_WUNLOCK(orig_object); VM_OBJECT_WUNLOCK(new_object); entry->object.vm_object = new_object; entry->offset = 0LL; vm_object_deallocate(orig_object); VM_OBJECT_WLOCK(new_object); } #define OBSC_COLLAPSE_NOWAIT 0x0002 #define OBSC_COLLAPSE_WAIT 0x0004 static vm_page_t vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p, vm_page_t next, int op) { vm_object_t backing_object; VM_OBJECT_ASSERT_WLOCKED(object); backing_object = object->backing_object; VM_OBJECT_ASSERT_WLOCKED(backing_object); KASSERT(p == NULL || p->object == object || p->object == backing_object, ("invalid ownership %p %p %p", p, object, backing_object)); if ((op & OBSC_COLLAPSE_NOWAIT) != 0) return (next); /* The page is only NULL when rename fails. */ if (p == NULL) { vm_radix_wait(); } else { if (p->object == object) VM_OBJECT_WUNLOCK(backing_object); else VM_OBJECT_WUNLOCK(object); vm_page_busy_sleep(p, "vmocol", false); } VM_OBJECT_WLOCK(object); VM_OBJECT_WLOCK(backing_object); return (TAILQ_FIRST(&backing_object->memq)); } static bool vm_object_scan_all_shadowed(vm_object_t object) { vm_object_t backing_object; vm_page_t p, pp; vm_pindex_t backing_offset_index, new_pindex, pi, ps; VM_OBJECT_ASSERT_WLOCKED(object); VM_OBJECT_ASSERT_WLOCKED(object->backing_object); backing_object = object->backing_object; if (backing_object->type != OBJT_DEFAULT && backing_object->type != OBJT_SWAP) return (false); pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset); p = vm_page_find_least(backing_object, pi); ps = swap_pager_find_least(backing_object, pi); /* * Only check pages inside the parent object's range and * inside the parent object's mapping of the backing object. */ for (;; pi++) { if (p != NULL && p->pindex < pi) p = TAILQ_NEXT(p, listq); if (ps < pi) ps = swap_pager_find_least(backing_object, pi); if (p == NULL && ps >= backing_object->size) break; else if (p == NULL) pi = ps; else pi = MIN(p->pindex, ps); new_pindex = pi - backing_offset_index; if (new_pindex >= object->size) break; /* * See if the parent has the page or if the parent's object * pager has the page. If the parent has the page but the page * is not valid, the parent's object pager must have the page. * * If this fails, the parent does not completely shadow the * object and we might as well give up now. */ pp = vm_page_lookup(object, new_pindex); - if ((pp == NULL || pp->valid == 0) && + /* + * The valid check here is stable due to object lock being + * required to clear valid and initiate paging. + */ + if ((pp == NULL || vm_page_none_valid(pp)) && !vm_pager_has_page(object, new_pindex, NULL, NULL)) return (false); } return (true); } static bool vm_object_collapse_scan(vm_object_t object, int op) { vm_object_t backing_object; vm_page_t next, p, pp; vm_pindex_t backing_offset_index, new_pindex; VM_OBJECT_ASSERT_WLOCKED(object); VM_OBJECT_ASSERT_WLOCKED(object->backing_object); backing_object = object->backing_object; backing_offset_index = OFF_TO_IDX(object->backing_object_offset); /* * Initial conditions */ if ((op & OBSC_COLLAPSE_WAIT) != 0) vm_object_set_flag(backing_object, OBJ_DEAD); /* * Our scan */ for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) { next = TAILQ_NEXT(p, listq); new_pindex = p->pindex - backing_offset_index; /* * Check for busy page */ if (vm_page_tryxbusy(p) == 0) { next = vm_object_collapse_scan_wait(object, p, next, op); continue; } KASSERT(p->object == backing_object, ("vm_object_collapse_scan: object mismatch")); if (p->pindex < backing_offset_index || new_pindex >= object->size) { if (backing_object->type == OBJT_SWAP) swap_pager_freespace(backing_object, p->pindex, 1); KASSERT(!pmap_page_is_mapped(p), ("freeing mapped page %p", p)); if (vm_page_remove(p)) vm_page_free(p); else vm_page_xunbusy(p); continue; } pp = vm_page_lookup(object, new_pindex); if (pp != NULL && vm_page_tryxbusy(pp) == 0) { vm_page_xunbusy(p); /* * The page in the parent is busy and possibly not * (yet) valid. Until its state is finalized by the * busy bit owner, we can't tell whether it shadows the * original page. Therefore, we must either skip it * and the original (backing_object) page or wait for * its state to be finalized. * * This is due to a race with vm_fault() where we must * unbusy the original (backing_obj) page before we can * (re)lock the parent. Hence we can get here. */ next = vm_object_collapse_scan_wait(object, pp, next, op); continue; } - KASSERT(pp == NULL || pp->valid != 0, + KASSERT(pp == NULL || !vm_page_none_valid(pp), ("unbusy invalid page %p", pp)); if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL, NULL)) { /* * The page already exists in the parent OR swap exists * for this location in the parent. Leave the parent's * page alone. Destroy the original page from the * backing object. */ if (backing_object->type == OBJT_SWAP) swap_pager_freespace(backing_object, p->pindex, 1); KASSERT(!pmap_page_is_mapped(p), ("freeing mapped page %p", p)); if (vm_page_remove(p)) vm_page_free(p); else vm_page_xunbusy(p); if (pp != NULL) vm_page_xunbusy(pp); continue; } /* * Page does not exist in parent, rename the page from the * backing object to the main object. * * If the page was mapped to a process, it can remain mapped * through the rename. vm_page_rename() will dirty the page. */ if (vm_page_rename(p, object, new_pindex)) { vm_page_xunbusy(p); if (pp != NULL) vm_page_xunbusy(pp); next = vm_object_collapse_scan_wait(object, NULL, next, op); continue; } /* Use the old pindex to free the right page. */ if (backing_object->type == OBJT_SWAP) swap_pager_freespace(backing_object, new_pindex + backing_offset_index, 1); #if VM_NRESERVLEVEL > 0 /* * Rename the reservation. */ vm_reserv_rename(p, object, backing_object, backing_offset_index); #endif vm_page_xunbusy(p); } return (true); } /* * this version of collapse allows the operation to occur earlier and * when paging_in_progress is true for an object... This is not a complete * operation, but should plug 99.9% of the rest of the leaks. */ static void vm_object_qcollapse(vm_object_t object) { vm_object_t backing_object = object->backing_object; VM_OBJECT_ASSERT_WLOCKED(object); VM_OBJECT_ASSERT_WLOCKED(backing_object); if (backing_object->ref_count != 1) return; vm_object_collapse_scan(object, OBSC_COLLAPSE_NOWAIT); } /* * vm_object_collapse: * * Collapse an object with the object backing it. * Pages in the backing object are moved into the * parent, and the backing object is deallocated. */ void vm_object_collapse(vm_object_t object) { vm_object_t backing_object, new_backing_object; VM_OBJECT_ASSERT_WLOCKED(object); while (TRUE) { /* * Verify that the conditions are right for collapse: * * The object exists and the backing object exists. */ if ((backing_object = object->backing_object) == NULL) break; /* * we check the backing object first, because it is most likely * not collapsable. */ VM_OBJECT_WLOCK(backing_object); if (backing_object->handle != NULL || (backing_object->type != OBJT_DEFAULT && backing_object->type != OBJT_SWAP) || (backing_object->flags & (OBJ_DEAD | OBJ_NOSPLIT)) != 0 || object->handle != NULL || (object->type != OBJT_DEFAULT && object->type != OBJT_SWAP) || (object->flags & OBJ_DEAD)) { VM_OBJECT_WUNLOCK(backing_object); break; } if (REFCOUNT_COUNT(object->paging_in_progress) > 0 || REFCOUNT_COUNT(backing_object->paging_in_progress) > 0) { vm_object_qcollapse(object); VM_OBJECT_WUNLOCK(backing_object); break; } /* * We know that we can either collapse the backing object (if * the parent is the only reference to it) or (perhaps) have * the parent bypass the object if the parent happens to shadow * all the resident pages in the entire backing object. * * This is ignoring pager-backed pages such as swap pages. * vm_object_collapse_scan fails the shadowing test in this * case. */ if (backing_object->ref_count == 1) { vm_object_pip_add(object, 1); vm_object_pip_add(backing_object, 1); /* * If there is exactly one reference to the backing * object, we can collapse it into the parent. */ vm_object_collapse_scan(object, OBSC_COLLAPSE_WAIT); #if VM_NRESERVLEVEL > 0 /* * Break any reservations from backing_object. */ if (__predict_false(!LIST_EMPTY(&backing_object->rvq))) vm_reserv_break_all(backing_object); #endif /* * Move the pager from backing_object to object. */ if (backing_object->type == OBJT_SWAP) { /* * swap_pager_copy() can sleep, in which case * the backing_object's and object's locks are * released and reacquired. * Since swap_pager_copy() is being asked to * destroy the source, it will change the * backing_object's type to OBJT_DEFAULT. */ swap_pager_copy( backing_object, object, OFF_TO_IDX(object->backing_object_offset), TRUE); } /* * Object now shadows whatever backing_object did. * Note that the reference to * backing_object->backing_object moves from within * backing_object to within object. */ LIST_REMOVE(object, shadow_list); backing_object->shadow_count--; if (backing_object->backing_object) { VM_OBJECT_WLOCK(backing_object->backing_object); LIST_REMOVE(backing_object, shadow_list); LIST_INSERT_HEAD( &backing_object->backing_object->shadow_head, object, shadow_list); /* * The shadow_count has not changed. */ VM_OBJECT_WUNLOCK(backing_object->backing_object); } object->backing_object = backing_object->backing_object; object->backing_object_offset += backing_object->backing_object_offset; /* * Discard backing_object. * * Since the backing object has no pages, no pager left, * and no object references within it, all that is * necessary is to dispose of it. */ KASSERT(backing_object->ref_count == 1, ( "backing_object %p was somehow re-referenced during collapse!", backing_object)); vm_object_pip_wakeup(backing_object); backing_object->type = OBJT_DEAD; backing_object->ref_count = 0; VM_OBJECT_WUNLOCK(backing_object); vm_object_destroy(backing_object); vm_object_pip_wakeup(object); counter_u64_add(object_collapses, 1); } else { /* * If we do not entirely shadow the backing object, * there is nothing we can do so we give up. */ if (object->resident_page_count != object->size && !vm_object_scan_all_shadowed(object)) { VM_OBJECT_WUNLOCK(backing_object); break; } /* * Make the parent shadow the next object in the * chain. Deallocating backing_object will not remove * it, since its reference count is at least 2. */ LIST_REMOVE(object, shadow_list); backing_object->shadow_count--; new_backing_object = backing_object->backing_object; if ((object->backing_object = new_backing_object) != NULL) { VM_OBJECT_WLOCK(new_backing_object); LIST_INSERT_HEAD( &new_backing_object->shadow_head, object, shadow_list ); new_backing_object->shadow_count++; vm_object_reference_locked(new_backing_object); VM_OBJECT_WUNLOCK(new_backing_object); object->backing_object_offset += backing_object->backing_object_offset; } /* * Drop the reference count on backing_object. Since * its ref_count was at least 2, it will not vanish. */ backing_object->ref_count--; VM_OBJECT_WUNLOCK(backing_object); counter_u64_add(object_bypasses, 1); } /* * Try again with this object's new backing object. */ } } /* * vm_object_page_remove: * * For the given object, either frees or invalidates each of the * specified pages. In general, a page is freed. However, if a page is * wired for any reason other than the existence of a managed, wired * mapping, then it may be invalidated but not removed from the object. * Pages are specified by the given range ["start", "end") and the option * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range * extends from "start" to the end of the object. If the option * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the * specified range are affected. If the option OBJPR_NOTMAPPED is * specified, then the pages within the specified range must have no * mappings. Otherwise, if this option is not specified, any mappings to * the specified pages are removed before the pages are freed or * invalidated. * * In general, this operation should only be performed on objects that * contain managed pages. There are, however, two exceptions. First, it * is performed on the kernel and kmem objects by vm_map_entry_delete(). * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device- * backed pages. In both of these cases, the option OBJPR_CLEANONLY must * not be specified and the option OBJPR_NOTMAPPED must be specified. * * The object must be locked. */ void vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int options) { vm_page_t p, next; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT((object->flags & OBJ_UNMANAGED) == 0 || (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED, ("vm_object_page_remove: illegal options for object %p", object)); if (object->resident_page_count == 0) return; vm_object_pip_add(object, 1); again: p = vm_page_find_least(object, start); /* * Here, the variable "p" is either (1) the page with the least pindex * greater than or equal to the parameter "start" or (2) NULL. */ for (; p != NULL && (p->pindex < end || end == 0); p = next) { next = TAILQ_NEXT(p, listq); /* * If the page is wired for any reason besides the existence * of managed, wired mappings, then it cannot be freed. For * example, fictitious pages, which represent device memory, * are inherently wired and cannot be freed. They can, * however, be invalidated if the option OBJPR_CLEANONLY is * not specified. */ if (vm_page_tryxbusy(p) == 0) { vm_page_sleep_if_busy(p, "vmopar"); goto again; } if (vm_page_wired(p)) { wired: if ((options & OBJPR_NOTMAPPED) == 0 && object->ref_count != 0) pmap_remove_all(p); if ((options & OBJPR_CLEANONLY) == 0) { - p->valid = 0; + vm_page_invalid(p); vm_page_undirty(p); } vm_page_xunbusy(p); continue; } KASSERT((p->flags & PG_FICTITIOUS) == 0, ("vm_object_page_remove: page %p is fictitious", p)); - if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) { + if ((options & OBJPR_CLEANONLY) != 0 && + !vm_page_none_valid(p)) { if ((options & OBJPR_NOTMAPPED) == 0 && object->ref_count != 0 && !vm_page_try_remove_write(p)) goto wired; if (p->dirty != 0) { vm_page_xunbusy(p); continue; } } if ((options & OBJPR_NOTMAPPED) == 0 && object->ref_count != 0 && !vm_page_try_remove_all(p)) goto wired; vm_page_free(p); } vm_object_pip_wakeup(object); } /* * vm_object_page_noreuse: * * For the given object, attempt to move the specified pages to * the head of the inactive queue. This bypasses regular LRU * operation and allows the pages to be reused quickly under memory * pressure. If a page is wired for any reason, then it will not * be queued. Pages are specified by the range ["start", "end"). * As a special case, if "end" is zero, then the range extends from * "start" to the end of the object. * * This operation should only be performed on objects that * contain non-fictitious, managed pages. * * The object must be locked. */ void vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end) { struct mtx *mtx; vm_page_t p, next; VM_OBJECT_ASSERT_LOCKED(object); KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0, ("vm_object_page_noreuse: illegal object %p", object)); if (object->resident_page_count == 0) return; p = vm_page_find_least(object, start); /* * Here, the variable "p" is either (1) the page with the least pindex * greater than or equal to the parameter "start" or (2) NULL. */ mtx = NULL; for (; p != NULL && (p->pindex < end || end == 0); p = next) { next = TAILQ_NEXT(p, listq); vm_page_change_lock(p, &mtx); vm_page_deactivate_noreuse(p); } if (mtx != NULL) mtx_unlock(mtx); } /* * Populate the specified range of the object with valid pages. Returns * TRUE if the range is successfully populated and FALSE otherwise. * * Note: This function should be optimized to pass a larger array of * pages to vm_pager_get_pages() before it is applied to a non- * OBJT_DEVICE object. * * The object must be locked. */ boolean_t vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end) { vm_page_t m; vm_pindex_t pindex; int rv; VM_OBJECT_ASSERT_WLOCKED(object); for (pindex = start; pindex < end; pindex++) { rv = vm_page_grab_valid(&m, object, pindex, VM_ALLOC_NORMAL); if (rv != VM_PAGER_OK) break; /* * Keep "m" busy because a subsequent iteration may unlock * the object. */ } if (pindex > start) { m = vm_page_lookup(object, start); while (m != NULL && m->pindex < pindex) { vm_page_xunbusy(m); m = TAILQ_NEXT(m, listq); } } return (pindex == end); } /* * Routine: vm_object_coalesce * Function: Coalesces two objects backing up adjoining * regions of memory into a single object. * * returns TRUE if objects were combined. * * NOTE: Only works at the moment if the second object is NULL - * if it's not, which object do we lock first? * * Parameters: * prev_object First object to coalesce * prev_offset Offset into prev_object * prev_size Size of reference to prev_object * next_size Size of reference to the second object * reserved Indicator that extension region has * swap accounted for * * Conditions: * The object must *not* be locked. */ boolean_t vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset, vm_size_t prev_size, vm_size_t next_size, boolean_t reserved) { vm_pindex_t next_pindex; if (prev_object == NULL) return (TRUE); VM_OBJECT_WLOCK(prev_object); if ((prev_object->type != OBJT_DEFAULT && prev_object->type != OBJT_SWAP) || (prev_object->flags & OBJ_NOSPLIT) != 0) { VM_OBJECT_WUNLOCK(prev_object); return (FALSE); } /* * Try to collapse the object first */ vm_object_collapse(prev_object); /* * Can't coalesce if: . more than one reference . paged out . shadows * another object . has a copy elsewhere (any of which mean that the * pages not mapped to prev_entry may be in use anyway) */ if (prev_object->backing_object != NULL) { VM_OBJECT_WUNLOCK(prev_object); return (FALSE); } prev_size >>= PAGE_SHIFT; next_size >>= PAGE_SHIFT; next_pindex = OFF_TO_IDX(prev_offset) + prev_size; if (prev_object->ref_count > 1 && prev_object->size != next_pindex && (prev_object->flags & OBJ_ONEMAPPING) == 0) { VM_OBJECT_WUNLOCK(prev_object); return (FALSE); } /* * Account for the charge. */ if (prev_object->cred != NULL) { /* * If prev_object was charged, then this mapping, * although not charged now, may become writable * later. Non-NULL cred in the object would prevent * swap reservation during enabling of the write * access, so reserve swap now. Failed reservation * cause allocation of the separate object for the map * entry, and swap reservation for this entry is * managed in appropriate time. */ if (!reserved && !swap_reserve_by_cred(ptoa(next_size), prev_object->cred)) { VM_OBJECT_WUNLOCK(prev_object); return (FALSE); } prev_object->charge += ptoa(next_size); } /* * Remove any pages that may still be in the object from a previous * deallocation. */ if (next_pindex < prev_object->size) { vm_object_page_remove(prev_object, next_pindex, next_pindex + next_size, 0); if (prev_object->type == OBJT_SWAP) swap_pager_freespace(prev_object, next_pindex, next_size); #if 0 if (prev_object->cred != NULL) { KASSERT(prev_object->charge >= ptoa(prev_object->size - next_pindex), ("object %p overcharged 1 %jx %jx", prev_object, (uintmax_t)next_pindex, (uintmax_t)next_size)); prev_object->charge -= ptoa(prev_object->size - next_pindex); } #endif } /* * Extend the object if necessary. */ if (next_pindex + next_size > prev_object->size) prev_object->size = next_pindex + next_size; VM_OBJECT_WUNLOCK(prev_object); return (TRUE); } void vm_object_set_writeable_dirty(vm_object_t object) { VM_OBJECT_ASSERT_WLOCKED(object); if (object->type != OBJT_VNODE) { if ((object->flags & OBJ_TMPFS_NODE) != 0) { KASSERT(object->type == OBJT_SWAP, ("non-swap tmpfs")); vm_object_set_flag(object, OBJ_TMPFS_DIRTY); } return; } object->generation++; if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) return; vm_object_set_flag(object, OBJ_MIGHTBEDIRTY); } /* * vm_object_unwire: * * For each page offset within the specified range of the given object, * find the highest-level page in the shadow chain and unwire it. A page * must exist at every page offset, and the highest-level page must be * wired. */ void vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length, uint8_t queue) { vm_object_t tobject, t1object; vm_page_t m, tm; vm_pindex_t end_pindex, pindex, tpindex; int depth, locked_depth; KASSERT((offset & PAGE_MASK) == 0, ("vm_object_unwire: offset is not page aligned")); KASSERT((length & PAGE_MASK) == 0, ("vm_object_unwire: length is not a multiple of PAGE_SIZE")); /* The wired count of a fictitious page never changes. */ if ((object->flags & OBJ_FICTITIOUS) != 0) return; pindex = OFF_TO_IDX(offset); end_pindex = pindex + atop(length); again: locked_depth = 1; VM_OBJECT_RLOCK(object); m = vm_page_find_least(object, pindex); while (pindex < end_pindex) { if (m == NULL || pindex < m->pindex) { /* * The first object in the shadow chain doesn't * contain a page at the current index. Therefore, * the page must exist in a backing object. */ tobject = object; tpindex = pindex; depth = 0; do { tpindex += OFF_TO_IDX(tobject->backing_object_offset); tobject = tobject->backing_object; KASSERT(tobject != NULL, ("vm_object_unwire: missing page")); if ((tobject->flags & OBJ_FICTITIOUS) != 0) goto next_page; depth++; if (depth == locked_depth) { locked_depth++; VM_OBJECT_RLOCK(tobject); } } while ((tm = vm_page_lookup(tobject, tpindex)) == NULL); } else { tm = m; m = TAILQ_NEXT(m, listq); } if (vm_page_trysbusy(tm) == 0) { for (tobject = object; locked_depth >= 1; locked_depth--) { t1object = tobject->backing_object; if (tm->object != tobject) VM_OBJECT_RUNLOCK(tobject); tobject = t1object; } vm_page_busy_sleep(tm, "unwbo", true); goto again; } vm_page_unwire(tm, queue); vm_page_sunbusy(tm); next_page: pindex++; } /* Release the accumulated object locks. */ for (tobject = object; locked_depth >= 1; locked_depth--) { t1object = tobject->backing_object; VM_OBJECT_RUNLOCK(tobject); tobject = t1object; } } /* * Return the vnode for the given object, or NULL if none exists. * For tmpfs objects, the function may return NULL if there is * no vnode allocated at the time of the call. */ struct vnode * vm_object_vnode(vm_object_t object) { struct vnode *vp; VM_OBJECT_ASSERT_LOCKED(object); if (object->type == OBJT_VNODE) { vp = object->handle; KASSERT(vp != NULL, ("%s: OBJT_VNODE has no vnode", __func__)); } else if (object->type == OBJT_SWAP && (object->flags & OBJ_TMPFS) != 0) { vp = object->un_pager.swp.swp_tmpfs; KASSERT(vp != NULL, ("%s: OBJT_TMPFS has no vnode", __func__)); } else { vp = NULL; } return (vp); } /* * Busy the vm object. This prevents new pages belonging to the object from * becoming busy. Existing pages persist as busy. Callers are responsible * for checking page state before proceeding. */ void vm_object_busy(vm_object_t obj) { VM_OBJECT_ASSERT_LOCKED(obj); refcount_acquire(&obj->busy); /* The fence is required to order loads of page busy. */ atomic_thread_fence_acq_rel(); } void vm_object_unbusy(vm_object_t obj) { VM_OBJECT_ASSERT_LOCKED(obj); refcount_release(&obj->busy); } void vm_object_busy_wait(vm_object_t obj, const char *wmesg) { VM_OBJECT_ASSERT_UNLOCKED(obj); if (obj->busy) refcount_sleep(&obj->busy, wmesg, PVM); } /* * Return the kvme type of the given object. * If vpp is not NULL, set it to the object's vm_object_vnode() or NULL. */ int vm_object_kvme_type(vm_object_t object, struct vnode **vpp) { VM_OBJECT_ASSERT_LOCKED(object); if (vpp != NULL) *vpp = vm_object_vnode(object); switch (object->type) { case OBJT_DEFAULT: return (KVME_TYPE_DEFAULT); case OBJT_VNODE: return (KVME_TYPE_VNODE); case OBJT_SWAP: if ((object->flags & OBJ_TMPFS_NODE) != 0) return (KVME_TYPE_VNODE); return (KVME_TYPE_SWAP); case OBJT_DEVICE: return (KVME_TYPE_DEVICE); case OBJT_PHYS: return (KVME_TYPE_PHYS); case OBJT_DEAD: return (KVME_TYPE_DEAD); case OBJT_SG: return (KVME_TYPE_SG); case OBJT_MGTDEVICE: return (KVME_TYPE_MGTDEVICE); default: return (KVME_TYPE_UNKNOWN); } } static int sysctl_vm_object_list(SYSCTL_HANDLER_ARGS) { struct kinfo_vmobject *kvo; char *fullpath, *freepath; struct vnode *vp; struct vattr va; vm_object_t obj; vm_page_t m; int count, error; if (req->oldptr == NULL) { /* * If an old buffer has not been provided, generate an * estimate of the space needed for a subsequent call. */ mtx_lock(&vm_object_list_mtx); count = 0; TAILQ_FOREACH(obj, &vm_object_list, object_list) { if (obj->type == OBJT_DEAD) continue; count++; } mtx_unlock(&vm_object_list_mtx); return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) * count * 11 / 10)); } kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK); error = 0; /* * VM objects are type stable and are never removed from the * list once added. This allows us to safely read obj->object_list * after reacquiring the VM object lock. */ mtx_lock(&vm_object_list_mtx); TAILQ_FOREACH(obj, &vm_object_list, object_list) { if (obj->type == OBJT_DEAD) continue; VM_OBJECT_RLOCK(obj); if (obj->type == OBJT_DEAD) { VM_OBJECT_RUNLOCK(obj); continue; } mtx_unlock(&vm_object_list_mtx); kvo->kvo_size = ptoa(obj->size); kvo->kvo_resident = obj->resident_page_count; kvo->kvo_ref_count = obj->ref_count; kvo->kvo_shadow_count = obj->shadow_count; kvo->kvo_memattr = obj->memattr; kvo->kvo_active = 0; kvo->kvo_inactive = 0; TAILQ_FOREACH(m, &obj->memq, listq) { /* * A page may belong to the object but be * dequeued and set to PQ_NONE while the * object lock is not held. This makes the * reads of m->queue below racy, and we do not * count pages set to PQ_NONE. However, this * sysctl is only meant to give an * approximation of the system anyway. */ if (m->queue == PQ_ACTIVE) kvo->kvo_active++; else if (m->queue == PQ_INACTIVE) kvo->kvo_inactive++; } kvo->kvo_vn_fileid = 0; kvo->kvo_vn_fsid = 0; kvo->kvo_vn_fsid_freebsd11 = 0; freepath = NULL; fullpath = ""; kvo->kvo_type = vm_object_kvme_type(obj, &vp); if (vp != NULL) vref(vp); VM_OBJECT_RUNLOCK(obj); if (vp != NULL) { vn_fullpath(curthread, vp, &fullpath, &freepath); vn_lock(vp, LK_SHARED | LK_RETRY); if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) { kvo->kvo_vn_fileid = va.va_fileid; kvo->kvo_vn_fsid = va.va_fsid; kvo->kvo_vn_fsid_freebsd11 = va.va_fsid; /* truncate */ } vput(vp); } strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path)); if (freepath != NULL) free(freepath, M_TEMP); /* Pack record size down */ kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path) + strlen(kvo->kvo_path) + 1; kvo->kvo_structsize = roundup(kvo->kvo_structsize, sizeof(uint64_t)); error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize); mtx_lock(&vm_object_list_mtx); if (error) break; } mtx_unlock(&vm_object_list_mtx); free(kvo, M_TEMP); return (error); } SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject", "List of VM objects"); #include "opt_ddb.h" #ifdef DDB #include #include #include static int _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) { vm_map_t tmpm; vm_map_entry_t tmpe; vm_object_t obj; if (map == 0) return 0; if (entry == 0) { VM_MAP_ENTRY_FOREACH(tmpe, map) { if (_vm_object_in_map(map, object, tmpe)) { return 1; } } } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { tmpm = entry->object.sub_map; VM_MAP_ENTRY_FOREACH(tmpe, tmpm) { if (_vm_object_in_map(tmpm, object, tmpe)) { return 1; } } } else if ((obj = entry->object.vm_object) != NULL) { for (; obj; obj = obj->backing_object) if (obj == object) { return 1; } } return 0; } static int vm_object_in_map(vm_object_t object) { struct proc *p; /* sx_slock(&allproc_lock); */ FOREACH_PROC_IN_SYSTEM(p) { if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) continue; if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { /* sx_sunlock(&allproc_lock); */ return 1; } } /* sx_sunlock(&allproc_lock); */ if (_vm_object_in_map(kernel_map, object, 0)) return 1; return 0; } DB_SHOW_COMMAND(vmochk, vm_object_check) { vm_object_t object; /* * make sure that internal objs are in a map somewhere * and none have zero ref counts. */ TAILQ_FOREACH(object, &vm_object_list, object_list) { if (object->handle == NULL && (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { if (object->ref_count == 0) { db_printf("vmochk: internal obj has zero ref count: %ld\n", (long)object->size); } if (!vm_object_in_map(object)) { db_printf( "vmochk: internal obj is not in a map: " "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", object->ref_count, (u_long)object->size, (u_long)object->size, (void *)object->backing_object); } } } } /* * vm_object_print: [ debug ] */ DB_SHOW_COMMAND(object, vm_object_print_static) { /* XXX convert args. */ vm_object_t object = (vm_object_t)addr; boolean_t full = have_addr; vm_page_t p; /* XXX count is an (unused) arg. Avoid shadowing it. */ #define count was_count int count; if (object == NULL) return; db_iprintf( "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n", object, (int)object->type, (uintmax_t)object->size, object->resident_page_count, object->ref_count, object->flags, object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge); db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n", object->shadow_count, object->backing_object ? object->backing_object->ref_count : 0, object->backing_object, (uintmax_t)object->backing_object_offset); if (!full) return; db_indent += 2; count = 0; TAILQ_FOREACH(p, &object->memq, listq) { if (count == 0) db_iprintf("memory:="); else if (count == 6) { db_printf("\n"); db_iprintf(" ..."); count = 0; } else db_printf(","); count++; db_printf("(off=0x%jx,page=0x%jx)", (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p)); } if (count != 0) db_printf("\n"); db_indent -= 2; } /* XXX. */ #undef count /* XXX need this non-static entry for calling from vm_map_print. */ void vm_object_print( /* db_expr_t */ long addr, boolean_t have_addr, /* db_expr_t */ long count, char *modif) { vm_object_print_static(addr, have_addr, count, modif); } DB_SHOW_COMMAND(vmopag, vm_object_print_pages) { vm_object_t object; vm_pindex_t fidx; vm_paddr_t pa; vm_page_t m, prev_m; int rcount, nl, c; nl = 0; TAILQ_FOREACH(object, &vm_object_list, object_list) { db_printf("new object: %p\n", (void *)object); if (nl > 18) { c = cngetc(); if (c != ' ') return; nl = 0; } nl++; rcount = 0; fidx = 0; pa = -1; TAILQ_FOREACH(m, &object->memq, listq) { if (m->pindex > 128) break; if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL && prev_m->pindex + 1 != m->pindex) { if (rcount) { db_printf(" index(%ld)run(%d)pa(0x%lx)\n", (long)fidx, rcount, (long)pa); if (nl > 18) { c = cngetc(); if (c != ' ') return; nl = 0; } nl++; rcount = 0; } } if (rcount && (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { ++rcount; continue; } if (rcount) { db_printf(" index(%ld)run(%d)pa(0x%lx)\n", (long)fidx, rcount, (long)pa); if (nl > 18) { c = cngetc(); if (c != ' ') return; nl = 0; } nl++; } fidx = m->pindex; pa = VM_PAGE_TO_PHYS(m); rcount = 1; } if (rcount) { db_printf(" index(%ld)run(%d)pa(0x%lx)\n", (long)fidx, rcount, (long)pa); if (nl > 18) { c = cngetc(); if (c != ' ') return; nl = 0; } nl++; } } } #endif /* DDB */ Index: head/sys/vm/vm_page.c =================================================================== --- head/sys/vm/vm_page.c (revision 353538) +++ head/sys/vm/vm_page.c (revision 353539) @@ -1,5026 +1,5115 @@ /*- * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU) * * Copyright (c) 1991 Regents of the University of California. * All rights reserved. * Copyright (c) 1998 Matthew Dillon. All Rights Reserved. * * This code is derived from software contributed to Berkeley by * The Mach Operating System project at Carnegie-Mellon University. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91 */ /*- * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Authors: Avadis Tevanian, Jr., Michael Wayne Young * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. */ /* * Resident memory management module. */ #include __FBSDID("$FreeBSD$"); #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include extern int uma_startup_count(int); extern void uma_startup(void *, int); extern int vmem_startup_count(void); struct vm_domain vm_dom[MAXMEMDOM]; DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]); struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT]; struct mtx_padalign __exclusive_cache_line vm_domainset_lock; /* The following fields are protected by the domainset lock. */ domainset_t __exclusive_cache_line vm_min_domains; domainset_t __exclusive_cache_line vm_severe_domains; static int vm_min_waiters; static int vm_severe_waiters; static int vm_pageproc_waiters; static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD, 0, "VM page statistics"); static counter_u64_t queue_ops = EARLY_COUNTER; SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops, CTLFLAG_RD, &queue_ops, "Number of batched queue operations"); static counter_u64_t queue_nops = EARLY_COUNTER; SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops, CTLFLAG_RD, &queue_nops, "Number of batched queue operations with no effects"); static void counter_startup(void) { queue_ops = counter_u64_alloc(M_WAITOK); queue_nops = counter_u64_alloc(M_WAITOK); } SYSINIT(page_counters, SI_SUB_CPU, SI_ORDER_ANY, counter_startup, NULL); /* * bogus page -- for I/O to/from partially complete buffers, * or for paging into sparsely invalid regions. */ vm_page_t bogus_page; vm_page_t vm_page_array; long vm_page_array_size; long first_page; static int boot_pages; SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &boot_pages, 0, "number of pages allocated for bootstrapping the VM system"); static int pa_tryrelock_restart; SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD, &pa_tryrelock_restart, 0, "Number of tryrelock restarts"); static TAILQ_HEAD(, vm_page) blacklist_head; static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS); SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages"); static uma_zone_t fakepg_zone; static void vm_page_alloc_check(vm_page_t m); static void _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, const char *wmesg, bool nonshared, bool locked); static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); static void vm_page_dequeue_complete(vm_page_t m); static void vm_page_enqueue(vm_page_t m, uint8_t queue); static void vm_page_init(void *dummy); static int vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, vm_page_t mpred); static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred); static void vm_page_mvqueue(vm_page_t m, uint8_t queue); static int vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run, vm_paddr_t high); static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req); static int vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags); static void vm_page_zone_release(void *arg, void **store, int cnt); SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL); static void vm_page_init(void *dummy) { fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM); bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | VM_ALLOC_NORMAL | VM_ALLOC_WIRED); } /* * The cache page zone is initialized later since we need to be able to allocate * pages before UMA is fully initialized. */ static void vm_page_init_cache_zones(void *dummy __unused) { struct vm_domain *vmd; struct vm_pgcache *pgcache; int domain, pool; for (domain = 0; domain < vm_ndomains; domain++) { vmd = VM_DOMAIN(domain); /* * Don't allow the page caches to take up more than .25% of * memory. */ if (vmd->vmd_page_count / 400 < 256 * mp_ncpus * VM_NFREEPOOL) continue; for (pool = 0; pool < VM_NFREEPOOL; pool++) { pgcache = &vmd->vmd_pgcache[pool]; pgcache->domain = domain; pgcache->pool = pool; pgcache->zone = uma_zcache_create("vm pgcache", sizeof(struct vm_page), NULL, NULL, NULL, NULL, vm_page_zone_import, vm_page_zone_release, pgcache, UMA_ZONE_MAXBUCKET | UMA_ZONE_VM); (void)uma_zone_set_maxcache(pgcache->zone, 0); } } } SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL); /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ #if PAGE_SIZE == 32768 #ifdef CTASSERT CTASSERT(sizeof(u_long) >= 8); #endif #endif /* * Try to acquire a physical address lock while a pmap is locked. If we * fail to trylock we unlock and lock the pmap directly and cache the * locked pa in *locked. The caller should then restart their loop in case * the virtual to physical mapping has changed. */ int vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked) { vm_paddr_t lockpa; lockpa = *locked; *locked = pa; if (lockpa) { PA_LOCK_ASSERT(lockpa, MA_OWNED); if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa)) return (0); PA_UNLOCK(lockpa); } if (PA_TRYLOCK(pa)) return (0); PMAP_UNLOCK(pmap); atomic_add_int(&pa_tryrelock_restart, 1); PA_LOCK(pa); PMAP_LOCK(pmap); return (EAGAIN); } /* * vm_set_page_size: * * Sets the page size, perhaps based upon the memory * size. Must be called before any use of page-size * dependent functions. */ void vm_set_page_size(void) { if (vm_cnt.v_page_size == 0) vm_cnt.v_page_size = PAGE_SIZE; if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0) panic("vm_set_page_size: page size not a power of two"); } /* * vm_page_blacklist_next: * * Find the next entry in the provided string of blacklist * addresses. Entries are separated by space, comma, or newline. * If an invalid integer is encountered then the rest of the * string is skipped. Updates the list pointer to the next * character, or NULL if the string is exhausted or invalid. */ static vm_paddr_t vm_page_blacklist_next(char **list, char *end) { vm_paddr_t bad; char *cp, *pos; if (list == NULL || *list == NULL) return (0); if (**list =='\0') { *list = NULL; return (0); } /* * If there's no end pointer then the buffer is coming from * the kenv and we know it's null-terminated. */ if (end == NULL) end = *list + strlen(*list); /* Ensure that strtoq() won't walk off the end */ if (*end != '\0') { if (*end == '\n' || *end == ' ' || *end == ',') *end = '\0'; else { printf("Blacklist not terminated, skipping\n"); *list = NULL; return (0); } } for (pos = *list; *pos != '\0'; pos = cp) { bad = strtoq(pos, &cp, 0); if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') { if (bad == 0) { if (++cp < end) continue; else break; } } else break; if (*cp == '\0' || ++cp >= end) *list = NULL; else *list = cp; return (trunc_page(bad)); } printf("Garbage in RAM blacklist, skipping\n"); *list = NULL; return (0); } bool vm_page_blacklist_add(vm_paddr_t pa, bool verbose) { struct vm_domain *vmd; vm_page_t m; int ret; m = vm_phys_paddr_to_vm_page(pa); if (m == NULL) return (true); /* page does not exist, no failure */ vmd = vm_pagequeue_domain(m); vm_domain_free_lock(vmd); ret = vm_phys_unfree_page(m); vm_domain_free_unlock(vmd); if (ret != 0) { vm_domain_freecnt_inc(vmd, -1); TAILQ_INSERT_TAIL(&blacklist_head, m, listq); if (verbose) printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa); } return (ret); } /* * vm_page_blacklist_check: * * Iterate through the provided string of blacklist addresses, pulling * each entry out of the physical allocator free list and putting it * onto a list for reporting via the vm.page_blacklist sysctl. */ static void vm_page_blacklist_check(char *list, char *end) { vm_paddr_t pa; char *next; next = list; while (next != NULL) { if ((pa = vm_page_blacklist_next(&next, end)) == 0) continue; vm_page_blacklist_add(pa, bootverbose); } } /* * vm_page_blacklist_load: * * Search for a special module named "ram_blacklist". It'll be a * plain text file provided by the user via the loader directive * of the same name. */ static void vm_page_blacklist_load(char **list, char **end) { void *mod; u_char *ptr; u_int len; mod = NULL; ptr = NULL; mod = preload_search_by_type("ram_blacklist"); if (mod != NULL) { ptr = preload_fetch_addr(mod); len = preload_fetch_size(mod); } *list = ptr; if (ptr != NULL) *end = ptr + len; else *end = NULL; return; } static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS) { vm_page_t m; struct sbuf sbuf; int error, first; first = 1; error = sysctl_wire_old_buffer(req, 0); if (error != 0) return (error); sbuf_new_for_sysctl(&sbuf, NULL, 128, req); TAILQ_FOREACH(m, &blacklist_head, listq) { sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",", (uintmax_t)m->phys_addr); first = 0; } error = sbuf_finish(&sbuf); sbuf_delete(&sbuf); return (error); } /* * Initialize a dummy page for use in scans of the specified paging queue. * In principle, this function only needs to set the flag PG_MARKER. * Nonetheless, it write busies the page as a safety precaution. */ static void vm_page_init_marker(vm_page_t marker, int queue, uint8_t aflags) { bzero(marker, sizeof(*marker)); marker->flags = PG_MARKER; marker->aflags = aflags; marker->busy_lock = VPB_SINGLE_EXCLUSIVER; marker->queue = queue; } static void vm_page_domain_init(int domain) { struct vm_domain *vmd; struct vm_pagequeue *pq; int i; vmd = VM_DOMAIN(domain); bzero(vmd, sizeof(*vmd)); *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) = "vm inactive pagequeue"; *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) = "vm active pagequeue"; *__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) = "vm laundry pagequeue"; *__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) = "vm unswappable pagequeue"; vmd->vmd_domain = domain; vmd->vmd_page_count = 0; vmd->vmd_free_count = 0; vmd->vmd_segs = 0; vmd->vmd_oom = FALSE; for (i = 0; i < PQ_COUNT; i++) { pq = &vmd->vmd_pagequeues[i]; TAILQ_INIT(&pq->pq_pl); mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue", MTX_DEF | MTX_DUPOK); pq->pq_pdpages = 0; vm_page_init_marker(&vmd->vmd_markers[i], i, 0); } mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF); mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF); snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain); /* * inacthead is used to provide FIFO ordering for LRU-bypassing * insertions. */ vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED); TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl, &vmd->vmd_inacthead, plinks.q); /* * The clock pages are used to implement active queue scanning without * requeues. Scans start at clock[0], which is advanced after the scan * ends. When the two clock hands meet, they are reset and scanning * resumes from the head of the queue. */ vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED); vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED); TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, &vmd->vmd_clock[0], plinks.q); TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, &vmd->vmd_clock[1], plinks.q); } /* * Initialize a physical page in preparation for adding it to the free * lists. */ static void vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind) { m->object = NULL; m->ref_count = 0; m->busy_lock = VPB_UNBUSIED; m->flags = m->aflags = 0; m->phys_addr = pa; m->queue = PQ_NONE; m->psind = 0; m->segind = segind; m->order = VM_NFREEORDER; m->pool = VM_FREEPOOL_DEFAULT; m->valid = m->dirty = 0; pmap_page_init(m); } #ifndef PMAP_HAS_PAGE_ARRAY static vm_paddr_t vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range) { vm_paddr_t new_end; /* * Reserve an unmapped guard page to trap access to vm_page_array[-1]. * However, because this page is allocated from KVM, out-of-bounds * accesses using the direct map will not be trapped. */ *vaddr += PAGE_SIZE; /* * Allocate physical memory for the page structures, and map it. */ new_end = trunc_page(end - page_range * sizeof(struct vm_page)); vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE); vm_page_array_size = page_range; return (new_end); } #endif /* * vm_page_startup: * * Initializes the resident memory module. Allocates physical memory for * bootstrapping UMA and some data structures that are used to manage * physical pages. Initializes these structures, and populates the free * page queues. */ vm_offset_t vm_page_startup(vm_offset_t vaddr) { struct vm_phys_seg *seg; vm_page_t m; char *list, *listend; vm_offset_t mapped; vm_paddr_t end, high_avail, low_avail, new_end, page_range, size; vm_paddr_t last_pa, pa; u_long pagecount; int biggestone, i, segind; #ifdef WITNESS int witness_size; #endif #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) long ii; #endif vaddr = round_page(vaddr); vm_phys_early_startup(); biggestone = vm_phys_avail_largest(); end = phys_avail[biggestone+1]; /* * Initialize the page and queue locks. */ mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF); for (i = 0; i < PA_LOCK_COUNT; i++) mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF); for (i = 0; i < vm_ndomains; i++) vm_page_domain_init(i); /* * Allocate memory for use when boot strapping the kernel memory * allocator. Tell UMA how many zones we are going to create * before going fully functional. UMA will add its zones. * * VM startup zones: vmem, vmem_btag, VM OBJECT, RADIX NODE, MAP, * KMAP ENTRY, MAP ENTRY, VMSPACE. */ boot_pages = uma_startup_count(8); #ifndef UMA_MD_SMALL_ALLOC /* vmem_startup() calls uma_prealloc(). */ boot_pages += vmem_startup_count(); /* vm_map_startup() calls uma_prealloc(). */ boot_pages += howmany(MAX_KMAP, UMA_SLAB_SPACE / sizeof(struct vm_map)); /* * Before going fully functional kmem_init() does allocation * from "KMAP ENTRY" and vmem_create() does allocation from "vmem". */ boot_pages += 2; #endif /* * CTFLAG_RDTUN doesn't work during the early boot process, so we must * manually fetch the value. */ TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages); new_end = end - (boot_pages * UMA_SLAB_SIZE); new_end = trunc_page(new_end); mapped = pmap_map(&vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE); bzero((void *)mapped, end - new_end); uma_startup((void *)mapped, boot_pages); #ifdef WITNESS witness_size = round_page(witness_startup_count()); new_end -= witness_size; mapped = pmap_map(&vaddr, new_end, new_end + witness_size, VM_PROT_READ | VM_PROT_WRITE); bzero((void *)mapped, witness_size); witness_startup((void *)mapped); #endif #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \ defined(__i386__) || defined(__mips__) || defined(__riscv) || \ defined(__powerpc64__) /* * Allocate a bitmap to indicate that a random physical page * needs to be included in a minidump. * * The amd64 port needs this to indicate which direct map pages * need to be dumped, via calls to dump_add_page()/dump_drop_page(). * * However, i386 still needs this workspace internally within the * minidump code. In theory, they are not needed on i386, but are * included should the sf_buf code decide to use them. */ last_pa = 0; for (i = 0; dump_avail[i + 1] != 0; i += 2) if (dump_avail[i + 1] > last_pa) last_pa = dump_avail[i + 1]; page_range = last_pa / PAGE_SIZE; vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); new_end -= vm_page_dump_size; vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); bzero((void *)vm_page_dump, vm_page_dump_size); #else (void)last_pa; #endif #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \ defined(__riscv) || defined(__powerpc64__) /* * Include the UMA bootstrap pages, witness pages and vm_page_dump * in a crash dump. When pmap_map() uses the direct map, they are * not automatically included. */ for (pa = new_end; pa < end; pa += PAGE_SIZE) dump_add_page(pa); #endif phys_avail[biggestone + 1] = new_end; #ifdef __amd64__ /* * Request that the physical pages underlying the message buffer be * included in a crash dump. Since the message buffer is accessed * through the direct map, they are not automatically included. */ pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); last_pa = pa + round_page(msgbufsize); while (pa < last_pa) { dump_add_page(pa); pa += PAGE_SIZE; } #endif /* * Compute the number of pages of memory that will be available for * use, taking into account the overhead of a page structure per page. * In other words, solve * "available physical memory" - round_page(page_range * * sizeof(struct vm_page)) = page_range * PAGE_SIZE * for page_range. */ low_avail = phys_avail[0]; high_avail = phys_avail[1]; for (i = 0; i < vm_phys_nsegs; i++) { if (vm_phys_segs[i].start < low_avail) low_avail = vm_phys_segs[i].start; if (vm_phys_segs[i].end > high_avail) high_avail = vm_phys_segs[i].end; } /* Skip the first chunk. It is already accounted for. */ for (i = 2; phys_avail[i + 1] != 0; i += 2) { if (phys_avail[i] < low_avail) low_avail = phys_avail[i]; if (phys_avail[i + 1] > high_avail) high_avail = phys_avail[i + 1]; } first_page = low_avail / PAGE_SIZE; #ifdef VM_PHYSSEG_SPARSE size = 0; for (i = 0; i < vm_phys_nsegs; i++) size += vm_phys_segs[i].end - vm_phys_segs[i].start; for (i = 0; phys_avail[i + 1] != 0; i += 2) size += phys_avail[i + 1] - phys_avail[i]; #elif defined(VM_PHYSSEG_DENSE) size = high_avail - low_avail; #else #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." #endif #ifdef PMAP_HAS_PAGE_ARRAY pmap_page_array_startup(size / PAGE_SIZE); biggestone = vm_phys_avail_largest(); end = new_end = phys_avail[biggestone + 1]; #else #ifdef VM_PHYSSEG_DENSE /* * In the VM_PHYSSEG_DENSE case, the number of pages can account for * the overhead of a page structure per page only if vm_page_array is * allocated from the last physical memory chunk. Otherwise, we must * allocate page structures representing the physical memory * underlying vm_page_array, even though they will not be used. */ if (new_end != high_avail) page_range = size / PAGE_SIZE; else #endif { page_range = size / (PAGE_SIZE + sizeof(struct vm_page)); /* * If the partial bytes remaining are large enough for * a page (PAGE_SIZE) without a corresponding * 'struct vm_page', then new_end will contain an * extra page after subtracting the length of the VM * page array. Compensate by subtracting an extra * page from new_end. */ if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) { if (new_end == high_avail) high_avail -= PAGE_SIZE; new_end -= PAGE_SIZE; } } end = new_end; new_end = vm_page_array_alloc(&vaddr, end, page_range); #endif #if VM_NRESERVLEVEL > 0 /* * Allocate physical memory for the reservation management system's * data structures, and map it. */ new_end = vm_reserv_startup(&vaddr, new_end); #endif #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \ defined(__riscv) || defined(__powerpc64__) /* * Include vm_page_array and vm_reserv_array in a crash dump. */ for (pa = new_end; pa < end; pa += PAGE_SIZE) dump_add_page(pa); #endif phys_avail[biggestone + 1] = new_end; /* * Add physical memory segments corresponding to the available * physical pages. */ for (i = 0; phys_avail[i + 1] != 0; i += 2) if (vm_phys_avail_size(i) != 0) vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]); /* * Initialize the physical memory allocator. */ vm_phys_init(); /* * Initialize the page structures and add every available page to the * physical memory allocator's free lists. */ #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) for (ii = 0; ii < vm_page_array_size; ii++) { m = &vm_page_array[ii]; vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0); m->flags = PG_FICTITIOUS; } #endif vm_cnt.v_page_count = 0; for (segind = 0; segind < vm_phys_nsegs; segind++) { seg = &vm_phys_segs[segind]; for (m = seg->first_page, pa = seg->start; pa < seg->end; m++, pa += PAGE_SIZE) vm_page_init_page(m, pa, segind); /* * Add the segment to the free lists only if it is covered by * one of the ranges in phys_avail. Because we've added the * ranges to the vm_phys_segs array, we can assume that each * segment is either entirely contained in one of the ranges, * or doesn't overlap any of them. */ for (i = 0; phys_avail[i + 1] != 0; i += 2) { struct vm_domain *vmd; if (seg->start < phys_avail[i] || seg->end > phys_avail[i + 1]) continue; m = seg->first_page; pagecount = (u_long)atop(seg->end - seg->start); vmd = VM_DOMAIN(seg->domain); vm_domain_free_lock(vmd); vm_phys_enqueue_contig(m, pagecount); vm_domain_free_unlock(vmd); vm_domain_freecnt_inc(vmd, pagecount); vm_cnt.v_page_count += (u_int)pagecount; vmd = VM_DOMAIN(seg->domain); vmd->vmd_page_count += (u_int)pagecount; vmd->vmd_segs |= 1UL << m->segind; break; } } /* * Remove blacklisted pages from the physical memory allocator. */ TAILQ_INIT(&blacklist_head); vm_page_blacklist_load(&list, &listend); vm_page_blacklist_check(list, listend); list = kern_getenv("vm.blacklist"); vm_page_blacklist_check(list, NULL); freeenv(list); #if VM_NRESERVLEVEL > 0 /* * Initialize the reservation management system. */ vm_reserv_init(); #endif return (vaddr); } void vm_page_reference(vm_page_t m) { vm_page_aflag_set(m, PGA_REFERENCED); } /* * vm_page_busy_acquire: * * Acquire the busy lock as described by VM_ALLOC_* flags. Will loop * and drop the object lock if necessary. */ int vm_page_busy_acquire(vm_page_t m, int allocflags) { vm_object_t obj; bool locked; /* * The page-specific object must be cached because page * identity can change during the sleep, causing the * re-lock of a different object. * It is assumed that a reference to the object is already * held by the callers. */ obj = m->object; for (;;) { if ((allocflags & VM_ALLOC_SBUSY) == 0) { if (vm_page_tryxbusy(m)) return (TRUE); } else { if (vm_page_trysbusy(m)) return (TRUE); } if ((allocflags & VM_ALLOC_NOWAIT) != 0) return (FALSE); if (obj != NULL) locked = VM_OBJECT_WOWNED(obj); else locked = FALSE; MPASS(locked || vm_page_wired(m)); _vm_page_busy_sleep(obj, m, "vmpba", (allocflags & VM_ALLOC_SBUSY) != 0, locked); if (locked) VM_OBJECT_WLOCK(obj); MPASS(m->object == obj || m->object == NULL); if ((allocflags & VM_ALLOC_WAITFAIL) != 0) return (FALSE); } } /* * vm_page_busy_downgrade: * * Downgrade an exclusive busy page into a single shared busy page. */ void vm_page_busy_downgrade(vm_page_t m) { u_int x; vm_page_assert_xbusied(m); x = m->busy_lock; for (;;) { if (atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_SHARERS_WORD(1))) break; } if ((x & VPB_BIT_WAITERS) != 0) wakeup(m); } /* * * vm_page_busy_tryupgrade: * * Attempt to upgrade a single shared busy into an exclusive busy. */ int vm_page_busy_tryupgrade(vm_page_t m) { u_int x; vm_page_assert_sbusied(m); x = m->busy_lock; for (;;) { if (VPB_SHARERS(x) > 1) return (0); KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), ("vm_page_busy_tryupgrade: invalid lock state")); if (!atomic_fcmpset_acq_int(&m->busy_lock, &x, VPB_SINGLE_EXCLUSIVER | (x & VPB_BIT_WAITERS))) continue; return (1); } } /* * vm_page_sbusied: * * Return a positive value if the page is shared busied, 0 otherwise. */ int vm_page_sbusied(vm_page_t m) { u_int x; x = m->busy_lock; return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); } /* * vm_page_sunbusy: * * Shared unbusy a page. */ void vm_page_sunbusy(vm_page_t m) { u_int x; vm_page_assert_sbusied(m); x = m->busy_lock; for (;;) { if (VPB_SHARERS(x) > 1) { if (atomic_fcmpset_int(&m->busy_lock, &x, x - VPB_ONE_SHARER)) break; continue; } KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), ("vm_page_sunbusy: invalid lock state")); if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) continue; if ((x & VPB_BIT_WAITERS) == 0) break; wakeup(m); break; } } /* * vm_page_busy_sleep: * * Sleep if the page is busy, using the page pointer as wchan. * This is used to implement the hard-path of busying mechanism. * * If nonshared is true, sleep only if the page is xbusy. * * The object lock must be held on entry and will be released on exit. */ void vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared) { vm_object_t obj; obj = m->object; VM_OBJECT_ASSERT_LOCKED(obj); vm_page_lock_assert(m, MA_NOTOWNED); _vm_page_busy_sleep(obj, m, wmesg, nonshared, true); } static void _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, const char *wmesg, bool nonshared, bool locked) { u_int x; /* * If the object is busy we must wait for that to drain to zero * before trying the page again. */ if (obj != NULL && vm_object_busied(obj)) { if (locked) VM_OBJECT_DROP(obj); vm_object_busy_wait(obj, wmesg); return; } sleepq_lock(m); x = m->busy_lock; if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) || ((x & VPB_BIT_WAITERS) == 0 && !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) { if (locked) VM_OBJECT_DROP(obj); sleepq_release(m); return; } if (locked) VM_OBJECT_DROP(obj); sleepq_add(m, NULL, wmesg, 0, 0); sleepq_wait(m, PVM); } /* * vm_page_trysbusy: * * Try to shared busy a page. * If the operation succeeds 1 is returned otherwise 0. * The operation never sleeps. */ int vm_page_trysbusy(vm_page_t m) { vm_object_t obj; u_int x; obj = m->object; x = m->busy_lock; for (;;) { if ((x & VPB_BIT_SHARED) == 0) return (0); /* * Reduce the window for transient busies that will trigger * false negatives in vm_page_ps_test(). */ if (obj != NULL && vm_object_busied(obj)) return (0); if (atomic_fcmpset_acq_int(&m->busy_lock, &x, x + VPB_ONE_SHARER)) break; } /* Refetch the object now that we're guaranteed that it is stable. */ obj = m->object; if (obj != NULL && vm_object_busied(obj)) { vm_page_sunbusy(m); return (0); } return (1); } /* * vm_page_tryxbusy: * * Try to exclusive busy a page. * If the operation succeeds 1 is returned otherwise 0. * The operation never sleeps. */ int vm_page_tryxbusy(vm_page_t m) { vm_object_t obj; if (atomic_cmpset_acq_int(&(m)->busy_lock, VPB_UNBUSIED, VPB_SINGLE_EXCLUSIVER) == 0) return (0); obj = m->object; if (obj != NULL && vm_object_busied(obj)) { vm_page_xunbusy(m); return (0); } return (1); } /* * vm_page_xunbusy_hard: * * Called when unbusy has failed because there is a waiter. */ void vm_page_xunbusy_hard(vm_page_t m) { vm_page_assert_xbusied(m); /* * Wake the waiter. */ atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); wakeup(m); } /* * Avoid releasing and reacquiring the same page lock. */ void vm_page_change_lock(vm_page_t m, struct mtx **mtx) { struct mtx *mtx1; mtx1 = vm_page_lockptr(m); if (*mtx == mtx1) return; if (*mtx != NULL) mtx_unlock(*mtx); *mtx = mtx1; mtx_lock(mtx1); } /* * vm_page_unhold_pages: * * Unhold each of the pages that is referenced by the given array. */ void vm_page_unhold_pages(vm_page_t *ma, int count) { for (; count != 0; count--) { vm_page_unwire(*ma, PQ_ACTIVE); ma++; } } vm_page_t PHYS_TO_VM_PAGE(vm_paddr_t pa) { vm_page_t m; #ifdef VM_PHYSSEG_SPARSE m = vm_phys_paddr_to_vm_page(pa); if (m == NULL) m = vm_phys_fictitious_to_vm_page(pa); return (m); #elif defined(VM_PHYSSEG_DENSE) long pi; pi = atop(pa); if (pi >= first_page && (pi - first_page) < vm_page_array_size) { m = &vm_page_array[pi - first_page]; return (m); } return (vm_phys_fictitious_to_vm_page(pa)); #else #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." #endif } /* * vm_page_getfake: * * Create a fictitious page with the specified physical address and * memory attribute. The memory attribute is the only the machine- * dependent aspect of a fictitious page that must be initialized. */ vm_page_t vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) { vm_page_t m; m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); vm_page_initfake(m, paddr, memattr); return (m); } void vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) { if ((m->flags & PG_FICTITIOUS) != 0) { /* * The page's memattr might have changed since the * previous initialization. Update the pmap to the * new memattr. */ goto memattr; } m->phys_addr = paddr; m->queue = PQ_NONE; /* Fictitious pages don't use "segind". */ m->flags = PG_FICTITIOUS; /* Fictitious pages don't use "order" or "pool". */ m->oflags = VPO_UNMANAGED; m->busy_lock = VPB_SINGLE_EXCLUSIVER; /* Fictitious pages are unevictable. */ m->ref_count = 1; pmap_page_init(m); memattr: pmap_page_set_memattr(m, memattr); } /* * vm_page_putfake: * * Release a fictitious page. */ void vm_page_putfake(vm_page_t m) { KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); KASSERT((m->flags & PG_FICTITIOUS) != 0, ("vm_page_putfake: bad page %p", m)); if (vm_page_xbusied(m)) vm_page_xunbusy(m); uma_zfree(fakepg_zone, m); } /* * vm_page_updatefake: * * Update the given fictitious page to the specified physical address and * memory attribute. */ void vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) { KASSERT((m->flags & PG_FICTITIOUS) != 0, ("vm_page_updatefake: bad page %p", m)); m->phys_addr = paddr; pmap_page_set_memattr(m, memattr); } /* * vm_page_free: * * Free a page. */ void vm_page_free(vm_page_t m) { m->flags &= ~PG_ZERO; vm_page_free_toq(m); } /* * vm_page_free_zero: * * Free a page to the zerod-pages queue */ void vm_page_free_zero(vm_page_t m) { m->flags |= PG_ZERO; vm_page_free_toq(m); } /* * Unbusy and handle the page queueing for a page from a getpages request that * was optionally read ahead or behind. */ void vm_page_readahead_finish(vm_page_t m) { /* We shouldn't put invalid pages on queues. */ - KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m)); + KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m)); /* * Since the page is not the actually needed one, whether it should * be activated or deactivated is not obvious. Empirical results * have shown that deactivating the page is usually the best choice, * unless the page is wanted by another thread. */ vm_page_lock(m); if ((m->busy_lock & VPB_BIT_WAITERS) != 0) vm_page_activate(m); else vm_page_deactivate(m); vm_page_unlock(m); vm_page_xunbusy(m); } /* * vm_page_sleep_if_busy: * * Sleep and release the object lock if the page is busied. * Returns TRUE if the thread slept. * * The given page must be unlocked and object containing it must * be locked. */ int vm_page_sleep_if_busy(vm_page_t m, const char *msg) { vm_object_t obj; vm_page_lock_assert(m, MA_NOTOWNED); VM_OBJECT_ASSERT_WLOCKED(m->object); /* * The page-specific object must be cached because page * identity can change during the sleep, causing the * re-lock of a different object. * It is assumed that a reference to the object is already * held by the callers. */ obj = m->object; if (vm_page_busied(m) || (obj != NULL && obj->busy)) { vm_page_busy_sleep(m, msg, false); VM_OBJECT_WLOCK(obj); return (TRUE); } return (FALSE); } /* * vm_page_sleep_if_xbusy: * * Sleep and release the object lock if the page is xbusied. * Returns TRUE if the thread slept. * * The given page must be unlocked and object containing it must * be locked. */ int vm_page_sleep_if_xbusy(vm_page_t m, const char *msg) { vm_object_t obj; vm_page_lock_assert(m, MA_NOTOWNED); VM_OBJECT_ASSERT_WLOCKED(m->object); /* * The page-specific object must be cached because page * identity can change during the sleep, causing the * re-lock of a different object. * It is assumed that a reference to the object is already * held by the callers. */ obj = m->object; if (vm_page_xbusied(m) || (obj != NULL && obj->busy)) { vm_page_busy_sleep(m, msg, true); VM_OBJECT_WLOCK(obj); return (TRUE); } return (FALSE); } /* * vm_page_dirty_KBI: [ internal use only ] * * Set all bits in the page's dirty field. * * The object containing the specified page must be locked if the * call is made from the machine-independent layer. * * See vm_page_clear_dirty_mask(). * * This function should only be called by vm_page_dirty(). */ void vm_page_dirty_KBI(vm_page_t m) { /* Refer to this operation by its public name. */ - KASSERT(m->valid == VM_PAGE_BITS_ALL, - ("vm_page_dirty: page is invalid!")); + KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!")); m->dirty = VM_PAGE_BITS_ALL; } /* * vm_page_insert: [ internal use only ] * * Inserts the given mem entry into the object and object list. * * The object must be locked. */ int vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) { vm_page_t mpred; VM_OBJECT_ASSERT_WLOCKED(object); mpred = vm_radix_lookup_le(&object->rtree, pindex); return (vm_page_insert_after(m, object, pindex, mpred)); } /* * vm_page_insert_after: * * Inserts the page "m" into the specified object at offset "pindex". * * The page "mpred" must immediately precede the offset "pindex" within * the specified object. * * The object must be locked. */ static int vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, vm_page_t mpred) { vm_page_t msucc; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(m->object == NULL, ("vm_page_insert_after: page already inserted")); if (mpred != NULL) { KASSERT(mpred->object == object, ("vm_page_insert_after: object doesn't contain mpred")); KASSERT(mpred->pindex < pindex, ("vm_page_insert_after: mpred doesn't precede pindex")); msucc = TAILQ_NEXT(mpred, listq); } else msucc = TAILQ_FIRST(&object->memq); if (msucc != NULL) KASSERT(msucc->pindex > pindex, ("vm_page_insert_after: msucc doesn't succeed pindex")); /* * Record the object/offset pair in this page. */ m->object = object; m->pindex = pindex; m->ref_count |= VPRC_OBJREF; /* * Now link into the object's ordered list of backed pages. */ if (vm_radix_insert(&object->rtree, m)) { m->object = NULL; m->pindex = 0; m->ref_count &= ~VPRC_OBJREF; return (1); } vm_page_insert_radixdone(m, object, mpred); return (0); } /* * vm_page_insert_radixdone: * * Complete page "m" insertion into the specified object after the * radix trie hooking. * * The page "mpred" must precede the offset "m->pindex" within the * specified object. * * The object must be locked. */ static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) { VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(object != NULL && m->object == object, ("vm_page_insert_radixdone: page %p has inconsistent object", m)); KASSERT((m->ref_count & VPRC_OBJREF) != 0, ("vm_page_insert_radixdone: page %p is missing object ref", m)); if (mpred != NULL) { KASSERT(mpred->object == object, ("vm_page_insert_radixdone: object doesn't contain mpred")); KASSERT(mpred->pindex < m->pindex, ("vm_page_insert_radixdone: mpred doesn't precede pindex")); } if (mpred != NULL) TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); else TAILQ_INSERT_HEAD(&object->memq, m, listq); /* * Show that the object has one more resident page. */ object->resident_page_count++; /* * Hold the vnode until the last page is released. */ if (object->resident_page_count == 1 && object->type == OBJT_VNODE) vhold(object->handle); /* * Since we are inserting a new and possibly dirty page, * update the object's OBJ_MIGHTBEDIRTY flag. */ if (pmap_page_is_write_mapped(m)) vm_object_set_writeable_dirty(object); } /* * Do the work to remove a page from its object. The caller is responsible for * updating the page's fields to reflect this removal. */ static void vm_page_object_remove(vm_page_t m) { vm_object_t object; vm_page_t mrem; object = m->object; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT((m->ref_count & VPRC_OBJREF) != 0, ("page %p is missing its object ref", m)); mrem = vm_radix_remove(&object->rtree, m->pindex); KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m)); /* * Now remove from the object's list of backed pages. */ TAILQ_REMOVE(&object->memq, m, listq); /* * And show that the object has one fewer resident page. */ object->resident_page_count--; /* * The vnode may now be recycled. */ if (object->resident_page_count == 0 && object->type == OBJT_VNODE) vdrop(object->handle); } /* * vm_page_remove: * * Removes the specified page from its containing object, but does not * invalidate any backing storage. Returns true if the object's reference * was the last reference to the page, and false otherwise. * * The object must be locked. */ bool vm_page_remove(vm_page_t m) { vm_page_object_remove(m); m->object = NULL; return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF); } /* * vm_page_lookup: * * Returns the page associated with the object/offset * pair specified; if none is found, NULL is returned. * * The object must be locked. */ vm_page_t vm_page_lookup(vm_object_t object, vm_pindex_t pindex) { VM_OBJECT_ASSERT_LOCKED(object); return (vm_radix_lookup(&object->rtree, pindex)); } /* * vm_page_find_least: * * Returns the page associated with the object with least pindex * greater than or equal to the parameter pindex, or NULL. * * The object must be locked. */ vm_page_t vm_page_find_least(vm_object_t object, vm_pindex_t pindex) { vm_page_t m; VM_OBJECT_ASSERT_LOCKED(object); if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) m = vm_radix_lookup_ge(&object->rtree, pindex); return (m); } /* * Returns the given page's successor (by pindex) within the object if it is * resident; if none is found, NULL is returned. * * The object must be locked. */ vm_page_t vm_page_next(vm_page_t m) { vm_page_t next; VM_OBJECT_ASSERT_LOCKED(m->object); if ((next = TAILQ_NEXT(m, listq)) != NULL) { MPASS(next->object == m->object); if (next->pindex != m->pindex + 1) next = NULL; } return (next); } /* * Returns the given page's predecessor (by pindex) within the object if it is * resident; if none is found, NULL is returned. * * The object must be locked. */ vm_page_t vm_page_prev(vm_page_t m) { vm_page_t prev; VM_OBJECT_ASSERT_LOCKED(m->object); if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) { MPASS(prev->object == m->object); if (prev->pindex != m->pindex - 1) prev = NULL; } return (prev); } /* * Uses the page mnew as a replacement for an existing page at index * pindex which must be already present in the object. */ vm_page_t vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex) { vm_page_t mold; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0, ("vm_page_replace: page %p already in object", mnew)); /* * This function mostly follows vm_page_insert() and * vm_page_remove() without the radix, object count and vnode * dance. Double check such functions for more comments. */ mnew->object = object; mnew->pindex = pindex; atomic_set_int(&mnew->ref_count, VPRC_OBJREF); mold = vm_radix_replace(&object->rtree, mnew); KASSERT(mold->queue == PQ_NONE, ("vm_page_replace: old page %p is on a paging queue", mold)); /* Keep the resident page list in sorted order. */ TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq); TAILQ_REMOVE(&object->memq, mold, listq); mold->object = NULL; atomic_clear_int(&mold->ref_count, VPRC_OBJREF); vm_page_xunbusy(mold); /* * The object's resident_page_count does not change because we have * swapped one page for another, but OBJ_MIGHTBEDIRTY. */ if (pmap_page_is_write_mapped(mnew)) vm_object_set_writeable_dirty(object); return (mold); } /* * vm_page_rename: * * Move the given memory entry from its * current object to the specified target object/offset. * * Note: swap associated with the page must be invalidated by the move. We * have to do this for several reasons: (1) we aren't freeing the * page, (2) we are dirtying the page, (3) the VM system is probably * moving the page from object A to B, and will then later move * the backing store from A to B and we can't have a conflict. * * Note: we *always* dirty the page. It is necessary both for the * fact that we moved it, and because we may be invalidating * swap. * * The objects must be locked. */ int vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) { vm_page_t mpred; vm_pindex_t opidx; VM_OBJECT_ASSERT_WLOCKED(new_object); KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m)); mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex); KASSERT(mpred == NULL || mpred->pindex != new_pindex, ("vm_page_rename: pindex already renamed")); /* * Create a custom version of vm_page_insert() which does not depend * by m_prev and can cheat on the implementation aspects of the * function. */ opidx = m->pindex; m->pindex = new_pindex; if (vm_radix_insert(&new_object->rtree, m)) { m->pindex = opidx; return (1); } /* * The operation cannot fail anymore. The removal must happen before * the listq iterator is tainted. */ m->pindex = opidx; vm_page_object_remove(m); /* Return back to the new pindex to complete vm_page_insert(). */ m->pindex = new_pindex; m->object = new_object; vm_page_insert_radixdone(m, new_object, mpred); vm_page_dirty(m); return (0); } /* * vm_page_alloc: * * Allocate and return a page that is associated with the specified * object and offset pair. By default, this page is exclusive busied. * * The caller must always specify an allocation class. * * allocation classes: * VM_ALLOC_NORMAL normal process request * VM_ALLOC_SYSTEM system *really* needs a page * VM_ALLOC_INTERRUPT interrupt time request * * optional allocation flags: * VM_ALLOC_COUNT(number) the number of additional pages that the caller * intends to allocate * VM_ALLOC_NOBUSY do not exclusive busy the page * VM_ALLOC_NODUMP do not include the page in a kernel core dump * VM_ALLOC_NOOBJ page is not associated with an object and * should not be exclusive busy * VM_ALLOC_SBUSY shared busy the allocated page * VM_ALLOC_WIRED wire the allocated page * VM_ALLOC_ZERO prefer a zeroed page */ vm_page_t vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) { return (vm_page_alloc_after(object, pindex, req, object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) : NULL)); } vm_page_t vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain, int req) { return (vm_page_alloc_domain_after(object, pindex, domain, req, object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) : NULL)); } /* * Allocate a page in the specified object with the given page index. To * optimize insertion of the page into the object, the caller must also specifiy * the resident page in the object with largest index smaller than the given * page index, or NULL if no such page exists. */ vm_page_t vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex, int req, vm_page_t mpred) { struct vm_domainset_iter di; vm_page_t m; int domain; vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); do { m = vm_page_alloc_domain_after(object, pindex, domain, req, mpred); if (m != NULL) break; } while (vm_domainset_iter_page(&di, object, &domain) == 0); return (m); } /* * Returns true if the number of free pages exceeds the minimum * for the request class and false otherwise. */ int vm_domain_allocate(struct vm_domain *vmd, int req, int npages) { u_int limit, old, new; req = req & VM_ALLOC_CLASS_MASK; /* * The page daemon is allowed to dig deeper into the free page list. */ if (curproc == pageproc && req != VM_ALLOC_INTERRUPT) req = VM_ALLOC_SYSTEM; if (req == VM_ALLOC_INTERRUPT) limit = 0; else if (req == VM_ALLOC_SYSTEM) limit = vmd->vmd_interrupt_free_min; else limit = vmd->vmd_free_reserved; /* * Attempt to reserve the pages. Fail if we're below the limit. */ limit += npages; old = vmd->vmd_free_count; do { if (old < limit) return (0); new = old - npages; } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0); /* Wake the page daemon if we've crossed the threshold. */ if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old)) pagedaemon_wakeup(vmd->vmd_domain); /* Only update bitsets on transitions. */ if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) || (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe)) vm_domain_set(vmd); return (1); } vm_page_t vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain, int req, vm_page_t mpred) { struct vm_domain *vmd; vm_page_t m; int flags, pool; KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), ("inconsistent object(%p)/req(%x)", object, req)); KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0, ("Can't sleep and retry object insertion.")); KASSERT(mpred == NULL || mpred->pindex < pindex, ("mpred %p doesn't precede pindex 0x%jx", mpred, (uintmax_t)pindex)); if (object != NULL) VM_OBJECT_ASSERT_WLOCKED(object); flags = 0; m = NULL; pool = object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT; again: #if VM_NRESERVLEVEL > 0 /* * Can we allocate the page from a reservation? */ if (vm_object_reserv(object) && (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) != NULL) { domain = vm_phys_domain(m); vmd = VM_DOMAIN(domain); goto found; } #endif vmd = VM_DOMAIN(domain); if (vmd->vmd_pgcache[pool].zone != NULL) { m = uma_zalloc(vmd->vmd_pgcache[pool].zone, M_NOWAIT); if (m != NULL) { flags |= PG_PCPU_CACHE; goto found; } } if (vm_domain_allocate(vmd, req, 1)) { /* * If not, allocate it from the free page queues. */ vm_domain_free_lock(vmd); m = vm_phys_alloc_pages(domain, pool, 0); vm_domain_free_unlock(vmd); if (m == NULL) { vm_domain_freecnt_inc(vmd, 1); #if VM_NRESERVLEVEL > 0 if (vm_reserv_reclaim_inactive(domain)) goto again; #endif } } if (m == NULL) { /* * Not allocatable, give up. */ if (vm_domain_alloc_fail(vmd, object, req)) goto again; return (NULL); } /* * At this point we had better have found a good page. */ found: vm_page_dequeue(m); vm_page_alloc_check(m); /* * Initialize the page. Only the PG_ZERO flag is inherited. */ if ((req & VM_ALLOC_ZERO) != 0) flags |= (m->flags & PG_ZERO); if ((req & VM_ALLOC_NODUMP) != 0) flags |= PG_NODUMP; m->flags = flags; m->aflags = 0; m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; m->busy_lock = VPB_UNBUSIED; if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) m->busy_lock = VPB_SINGLE_EXCLUSIVER; if ((req & VM_ALLOC_SBUSY) != 0) m->busy_lock = VPB_SHARERS_WORD(1); if (req & VM_ALLOC_WIRED) { /* * The page lock is not required for wiring a page until that * page is inserted into the object. */ vm_wire_add(1); m->ref_count = 1; } m->act_count = 0; if (object != NULL) { if (vm_page_insert_after(m, object, pindex, mpred)) { if (req & VM_ALLOC_WIRED) { vm_wire_sub(1); m->ref_count = 0; } KASSERT(m->object == NULL, ("page %p has object", m)); m->oflags = VPO_UNMANAGED; m->busy_lock = VPB_UNBUSIED; /* Don't change PG_ZERO. */ vm_page_free_toq(m); if (req & VM_ALLOC_WAITFAIL) { VM_OBJECT_WUNLOCK(object); vm_radix_wait(); VM_OBJECT_WLOCK(object); } return (NULL); } /* Ignore device objects; the pager sets "memattr" for them. */ if (object->memattr != VM_MEMATTR_DEFAULT && (object->flags & OBJ_FICTITIOUS) == 0) pmap_page_set_memattr(m, object->memattr); } else m->pindex = pindex; return (m); } /* * vm_page_alloc_contig: * * Allocate a contiguous set of physical pages of the given size "npages" * from the free lists. All of the physical pages must be at or above * the given physical address "low" and below the given physical address * "high". The given value "alignment" determines the alignment of the * first physical page in the set. If the given value "boundary" is * non-zero, then the set of physical pages cannot cross any physical * address boundary that is a multiple of that value. Both "alignment" * and "boundary" must be a power of two. * * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, * then the memory attribute setting for the physical pages is configured * to the object's memory attribute setting. Otherwise, the memory * attribute setting for the physical pages is configured to "memattr", * overriding the object's memory attribute setting. However, if the * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the * memory attribute setting for the physical pages cannot be configured * to VM_MEMATTR_DEFAULT. * * The specified object may not contain fictitious pages. * * The caller must always specify an allocation class. * * allocation classes: * VM_ALLOC_NORMAL normal process request * VM_ALLOC_SYSTEM system *really* needs a page * VM_ALLOC_INTERRUPT interrupt time request * * optional allocation flags: * VM_ALLOC_NOBUSY do not exclusive busy the page * VM_ALLOC_NODUMP do not include the page in a kernel core dump * VM_ALLOC_NOOBJ page is not associated with an object and * should not be exclusive busy * VM_ALLOC_SBUSY shared busy the allocated page * VM_ALLOC_WIRED wire the allocated page * VM_ALLOC_ZERO prefer a zeroed page */ vm_page_t vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr) { struct vm_domainset_iter di; vm_page_t m; int domain; vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); do { m = vm_page_alloc_contig_domain(object, pindex, domain, req, npages, low, high, alignment, boundary, memattr); if (m != NULL) break; } while (vm_domainset_iter_page(&di, object, &domain) == 0); return (m); } vm_page_t vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr) { struct vm_domain *vmd; vm_page_t m, m_ret, mpred; u_int busy_lock, flags, oflags; mpred = NULL; /* XXX: pacify gcc */ KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object, req)); KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0, ("Can't sleep and retry object insertion.")); if (object != NULL) { VM_OBJECT_ASSERT_WLOCKED(object); KASSERT((object->flags & OBJ_FICTITIOUS) == 0, ("vm_page_alloc_contig: object %p has fictitious pages", object)); } KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); if (object != NULL) { mpred = vm_radix_lookup_le(&object->rtree, pindex); KASSERT(mpred == NULL || mpred->pindex != pindex, ("vm_page_alloc_contig: pindex already allocated")); } /* * Can we allocate the pages without the number of free pages falling * below the lower bound for the allocation class? */ m_ret = NULL; again: #if VM_NRESERVLEVEL > 0 /* * Can we allocate the pages from a reservation? */ if (vm_object_reserv(object) && (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req, mpred, npages, low, high, alignment, boundary)) != NULL) { domain = vm_phys_domain(m_ret); vmd = VM_DOMAIN(domain); goto found; } #endif vmd = VM_DOMAIN(domain); if (vm_domain_allocate(vmd, req, npages)) { /* * allocate them from the free page queues. */ vm_domain_free_lock(vmd); m_ret = vm_phys_alloc_contig(domain, npages, low, high, alignment, boundary); vm_domain_free_unlock(vmd); if (m_ret == NULL) { vm_domain_freecnt_inc(vmd, npages); #if VM_NRESERVLEVEL > 0 if (vm_reserv_reclaim_contig(domain, npages, low, high, alignment, boundary)) goto again; #endif } } if (m_ret == NULL) { if (vm_domain_alloc_fail(vmd, object, req)) goto again; return (NULL); } #if VM_NRESERVLEVEL > 0 found: #endif for (m = m_ret; m < &m_ret[npages]; m++) { vm_page_dequeue(m); vm_page_alloc_check(m); } /* * Initialize the pages. Only the PG_ZERO flag is inherited. */ flags = 0; if ((req & VM_ALLOC_ZERO) != 0) flags = PG_ZERO; if ((req & VM_ALLOC_NODUMP) != 0) flags |= PG_NODUMP; oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; busy_lock = VPB_UNBUSIED; if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) busy_lock = VPB_SINGLE_EXCLUSIVER; if ((req & VM_ALLOC_SBUSY) != 0) busy_lock = VPB_SHARERS_WORD(1); if ((req & VM_ALLOC_WIRED) != 0) vm_wire_add(npages); if (object != NULL) { if (object->memattr != VM_MEMATTR_DEFAULT && memattr == VM_MEMATTR_DEFAULT) memattr = object->memattr; } for (m = m_ret; m < &m_ret[npages]; m++) { m->aflags = 0; m->flags = (m->flags | PG_NODUMP) & flags; m->busy_lock = busy_lock; if ((req & VM_ALLOC_WIRED) != 0) m->ref_count = 1; m->act_count = 0; m->oflags = oflags; if (object != NULL) { if (vm_page_insert_after(m, object, pindex, mpred)) { if ((req & VM_ALLOC_WIRED) != 0) vm_wire_sub(npages); KASSERT(m->object == NULL, ("page %p has object", m)); mpred = m; for (m = m_ret; m < &m_ret[npages]; m++) { if (m <= mpred && (req & VM_ALLOC_WIRED) != 0) m->ref_count = 0; m->oflags = VPO_UNMANAGED; m->busy_lock = VPB_UNBUSIED; /* Don't change PG_ZERO. */ vm_page_free_toq(m); } if (req & VM_ALLOC_WAITFAIL) { VM_OBJECT_WUNLOCK(object); vm_radix_wait(); VM_OBJECT_WLOCK(object); } return (NULL); } mpred = m; } else m->pindex = pindex; if (memattr != VM_MEMATTR_DEFAULT) pmap_page_set_memattr(m, memattr); pindex++; } return (m_ret); } /* * Check a page that has been freshly dequeued from a freelist. */ static void vm_page_alloc_check(vm_page_t m) { KASSERT(m->object == NULL, ("page %p has object", m)); KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0, ("page %p has unexpected queue %d, flags %#x", m, m->queue, (m->aflags & PGA_QUEUE_STATE_MASK))); KASSERT(m->ref_count == 0, ("page %p has references", m)); KASSERT(!vm_page_busied(m), ("page %p is busy", m)); KASSERT(m->dirty == 0, ("page %p is dirty", m)); KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, ("page %p has unexpected memattr %d", m, pmap_page_get_memattr(m))); KASSERT(m->valid == 0, ("free page %p is valid", m)); } /* * vm_page_alloc_freelist: * * Allocate a physical page from the specified free page list. * * The caller must always specify an allocation class. * * allocation classes: * VM_ALLOC_NORMAL normal process request * VM_ALLOC_SYSTEM system *really* needs a page * VM_ALLOC_INTERRUPT interrupt time request * * optional allocation flags: * VM_ALLOC_COUNT(number) the number of additional pages that the caller * intends to allocate * VM_ALLOC_WIRED wire the allocated page * VM_ALLOC_ZERO prefer a zeroed page */ vm_page_t vm_page_alloc_freelist(int freelist, int req) { struct vm_domainset_iter di; vm_page_t m; int domain; vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); do { m = vm_page_alloc_freelist_domain(domain, freelist, req); if (m != NULL) break; } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); return (m); } vm_page_t vm_page_alloc_freelist_domain(int domain, int freelist, int req) { struct vm_domain *vmd; vm_page_t m; u_int flags; m = NULL; vmd = VM_DOMAIN(domain); again: if (vm_domain_allocate(vmd, req, 1)) { vm_domain_free_lock(vmd); m = vm_phys_alloc_freelist_pages(domain, freelist, VM_FREEPOOL_DIRECT, 0); vm_domain_free_unlock(vmd); if (m == NULL) vm_domain_freecnt_inc(vmd, 1); } if (m == NULL) { if (vm_domain_alloc_fail(vmd, NULL, req)) goto again; return (NULL); } vm_page_dequeue(m); vm_page_alloc_check(m); /* * Initialize the page. Only the PG_ZERO flag is inherited. */ m->aflags = 0; flags = 0; if ((req & VM_ALLOC_ZERO) != 0) flags = PG_ZERO; m->flags &= flags; if ((req & VM_ALLOC_WIRED) != 0) { /* * The page lock is not required for wiring a page that does * not belong to an object. */ vm_wire_add(1); m->ref_count = 1; } /* Unmanaged pages don't use "act_count". */ m->oflags = VPO_UNMANAGED; return (m); } static int vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags) { struct vm_domain *vmd; struct vm_pgcache *pgcache; int i; pgcache = arg; vmd = VM_DOMAIN(pgcache->domain); /* Only import if we can bring in a full bucket. */ if (cnt == 1 || !vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt)) return (0); domain = vmd->vmd_domain; vm_domain_free_lock(vmd); i = vm_phys_alloc_npages(domain, pgcache->pool, cnt, (vm_page_t *)store); vm_domain_free_unlock(vmd); if (cnt != i) vm_domain_freecnt_inc(vmd, cnt - i); return (i); } static void vm_page_zone_release(void *arg, void **store, int cnt) { struct vm_domain *vmd; struct vm_pgcache *pgcache; vm_page_t m; int i; pgcache = arg; vmd = VM_DOMAIN(pgcache->domain); vm_domain_free_lock(vmd); for (i = 0; i < cnt; i++) { m = (vm_page_t)store[i]; vm_phys_free_pages(m, 0); } vm_domain_free_unlock(vmd); vm_domain_freecnt_inc(vmd, cnt); } #define VPSC_ANY 0 /* No restrictions. */ #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */ #define VPSC_NOSUPER 2 /* Skip superpages. */ /* * vm_page_scan_contig: * * Scan vm_page_array[] between the specified entries "m_start" and * "m_end" for a run of contiguous physical pages that satisfy the * specified conditions, and return the lowest page in the run. The * specified "alignment" determines the alignment of the lowest physical * page in the run. If the specified "boundary" is non-zero, then the * run of physical pages cannot span a physical address that is a * multiple of "boundary". * * "m_end" is never dereferenced, so it need not point to a vm_page * structure within vm_page_array[]. * * "npages" must be greater than zero. "m_start" and "m_end" must not * span a hole (or discontiguity) in the physical address space. Both * "alignment" and "boundary" must be a power of two. */ vm_page_t vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, u_long alignment, vm_paddr_t boundary, int options) { struct mtx *m_mtx; vm_object_t object; vm_paddr_t pa; vm_page_t m, m_run; #if VM_NRESERVLEVEL > 0 int level; #endif int m_inc, order, run_ext, run_len; KASSERT(npages > 0, ("npages is 0")); KASSERT(powerof2(alignment), ("alignment is not a power of 2")); KASSERT(powerof2(boundary), ("boundary is not a power of 2")); m_run = NULL; run_len = 0; m_mtx = NULL; for (m = m_start; m < m_end && run_len < npages; m += m_inc) { KASSERT((m->flags & PG_MARKER) == 0, ("page %p is PG_MARKER", m)); KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1, ("fictitious page %p has invalid ref count", m)); /* * If the current page would be the start of a run, check its * physical address against the end, alignment, and boundary * conditions. If it doesn't satisfy these conditions, either * terminate the scan or advance to the next page that * satisfies the failed condition. */ if (run_len == 0) { KASSERT(m_run == NULL, ("m_run != NULL")); if (m + npages > m_end) break; pa = VM_PAGE_TO_PHYS(m); if ((pa & (alignment - 1)) != 0) { m_inc = atop(roundup2(pa, alignment) - pa); continue; } if (rounddown2(pa ^ (pa + ptoa(npages) - 1), boundary) != 0) { m_inc = atop(roundup2(pa, boundary) - pa); continue; } } else KASSERT(m_run != NULL, ("m_run == NULL")); vm_page_change_lock(m, &m_mtx); m_inc = 1; retry: if (vm_page_wired(m)) run_ext = 0; #if VM_NRESERVLEVEL > 0 else if ((level = vm_reserv_level(m)) >= 0 && (options & VPSC_NORESERV) != 0) { run_ext = 0; /* Advance to the end of the reservation. */ pa = VM_PAGE_TO_PHYS(m); m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - pa); } #endif else if ((object = m->object) != NULL) { /* * The page is considered eligible for relocation if * and only if it could be laundered or reclaimed by * the page daemon. */ if (!VM_OBJECT_TRYRLOCK(object)) { mtx_unlock(m_mtx); VM_OBJECT_RLOCK(object); mtx_lock(m_mtx); if (m->object != object) { /* * The page may have been freed. */ VM_OBJECT_RUNLOCK(object); goto retry; } } /* Don't care: PG_NODUMP, PG_ZERO. */ if (object->type != OBJT_DEFAULT && object->type != OBJT_SWAP && object->type != OBJT_VNODE) { run_ext = 0; #if VM_NRESERVLEVEL > 0 } else if ((options & VPSC_NOSUPER) != 0 && (level = vm_reserv_level_iffullpop(m)) >= 0) { run_ext = 0; /* Advance to the end of the superpage. */ pa = VM_PAGE_TO_PHYS(m); m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - pa); #endif } else if (object->memattr == VM_MEMATTR_DEFAULT && vm_page_queue(m) != PQ_NONE && !vm_page_busied(m) && !vm_page_wired(m)) { /* * The page is allocated but eligible for * relocation. Extend the current run by one * page. */ KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, ("page %p has an unexpected memattr", m)); KASSERT((m->oflags & (VPO_SWAPINPROG | VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, ("page %p has unexpected oflags", m)); /* Don't care: VPO_NOSYNC. */ run_ext = 1; } else run_ext = 0; VM_OBJECT_RUNLOCK(object); #if VM_NRESERVLEVEL > 0 } else if (level >= 0) { /* * The page is reserved but not yet allocated. In * other words, it is still free. Extend the current * run by one page. */ run_ext = 1; #endif } else if ((order = m->order) < VM_NFREEORDER) { /* * The page is enqueued in the physical memory * allocator's free page queues. Moreover, it is the * first page in a power-of-two-sized run of * contiguous free pages. Add these pages to the end * of the current run, and jump ahead. */ run_ext = 1 << order; m_inc = 1 << order; } else { /* * Skip the page for one of the following reasons: (1) * It is enqueued in the physical memory allocator's * free page queues. However, it is not the first * page in a run of contiguous free pages. (This case * rarely occurs because the scan is performed in * ascending order.) (2) It is not reserved, and it is * transitioning from free to allocated. (Conversely, * the transition from allocated to free for managed * pages is blocked by the page lock.) (3) It is * allocated but not contained by an object and not * wired, e.g., allocated by Xen's balloon driver. */ run_ext = 0; } /* * Extend or reset the current run of pages. */ if (run_ext > 0) { if (run_len == 0) m_run = m; run_len += run_ext; } else { if (run_len > 0) { m_run = NULL; run_len = 0; } } } if (m_mtx != NULL) mtx_unlock(m_mtx); if (run_len >= npages) return (m_run); return (NULL); } /* * vm_page_reclaim_run: * * Try to relocate each of the allocated virtual pages within the * specified run of physical pages to a new physical address. Free the * physical pages underlying the relocated virtual pages. A virtual page * is relocatable if and only if it could be laundered or reclaimed by * the page daemon. Whenever possible, a virtual page is relocated to a * physical address above "high". * * Returns 0 if every physical page within the run was already free or * just freed by a successful relocation. Otherwise, returns a non-zero * value indicating why the last attempt to relocate a virtual page was * unsuccessful. * * "req_class" must be an allocation class. */ static int vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run, vm_paddr_t high) { struct vm_domain *vmd; struct mtx *m_mtx; struct spglist free; vm_object_t object; vm_paddr_t pa; vm_page_t m, m_end, m_new; int error, order, req; KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class, ("req_class is not an allocation class")); SLIST_INIT(&free); error = 0; m = m_run; m_end = m_run + npages; m_mtx = NULL; for (; error == 0 && m < m_end; m++) { KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, ("page %p is PG_FICTITIOUS or PG_MARKER", m)); /* * Avoid releasing and reacquiring the same page lock. */ vm_page_change_lock(m, &m_mtx); retry: /* * Racily check for wirings. Races are handled below. */ if (vm_page_wired(m)) error = EBUSY; else if ((object = m->object) != NULL) { /* * The page is relocated if and only if it could be * laundered or reclaimed by the page daemon. */ if (!VM_OBJECT_TRYWLOCK(object)) { mtx_unlock(m_mtx); VM_OBJECT_WLOCK(object); mtx_lock(m_mtx); if (m->object != object) { /* * The page may have been freed. */ VM_OBJECT_WUNLOCK(object); goto retry; } } /* Don't care: PG_NODUMP, PG_ZERO. */ if (object->type != OBJT_DEFAULT && object->type != OBJT_SWAP && object->type != OBJT_VNODE) error = EINVAL; else if (object->memattr != VM_MEMATTR_DEFAULT) error = EINVAL; else if (vm_page_queue(m) != PQ_NONE && vm_page_tryxbusy(m) != 0) { if (vm_page_wired(m)) { vm_page_xunbusy(m); error = EBUSY; goto unlock; } KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, ("page %p has an unexpected memattr", m)); KASSERT((m->oflags & (VPO_SWAPINPROG | VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, ("page %p has unexpected oflags", m)); /* Don't care: VPO_NOSYNC. */ - if (m->valid != 0) { + if (!vm_page_none_valid(m)) { /* * First, try to allocate a new page * that is above "high". Failing * that, try to allocate a new page * that is below "m_run". Allocate * the new page between the end of * "m_run" and "high" only as a last * resort. */ req = req_class | VM_ALLOC_NOOBJ; if ((m->flags & PG_NODUMP) != 0) req |= VM_ALLOC_NODUMP; if (trunc_page(high) != ~(vm_paddr_t)PAGE_MASK) { m_new = vm_page_alloc_contig( NULL, 0, req, 1, round_page(high), ~(vm_paddr_t)0, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT); } else m_new = NULL; if (m_new == NULL) { pa = VM_PAGE_TO_PHYS(m_run); m_new = vm_page_alloc_contig( NULL, 0, req, 1, 0, pa - 1, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT); } if (m_new == NULL) { pa += ptoa(npages); m_new = vm_page_alloc_contig( NULL, 0, req, 1, pa, high, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT); } if (m_new == NULL) { vm_page_xunbusy(m); error = ENOMEM; goto unlock; } /* * Unmap the page and check for new * wirings that may have been acquired * through a pmap lookup. */ if (object->ref_count != 0 && !vm_page_try_remove_all(m)) { vm_page_free(m_new); error = EBUSY; goto unlock; } /* * Replace "m" with the new page. For * vm_page_replace(), "m" must be busy * and dequeued. Finally, change "m" * as if vm_page_free() was called. */ m_new->aflags = m->aflags & ~PGA_QUEUE_STATE_MASK; KASSERT(m_new->oflags == VPO_UNMANAGED, ("page %p is managed", m_new)); m_new->oflags = m->oflags & VPO_NOSYNC; pmap_copy_page(m, m_new); m_new->valid = m->valid; m_new->dirty = m->dirty; m->flags &= ~PG_ZERO; vm_page_dequeue(m); vm_page_replace_checked(m_new, object, m->pindex, m); if (vm_page_free_prep(m)) SLIST_INSERT_HEAD(&free, m, plinks.s.ss); /* * The new page must be deactivated * before the object is unlocked. */ vm_page_change_lock(m_new, &m_mtx); vm_page_deactivate(m_new); } else { m->flags &= ~PG_ZERO; vm_page_dequeue(m); if (vm_page_free_prep(m)) SLIST_INSERT_HEAD(&free, m, plinks.s.ss); KASSERT(m->dirty == 0, ("page %p is dirty", m)); } } else error = EBUSY; unlock: VM_OBJECT_WUNLOCK(object); } else { MPASS(vm_phys_domain(m) == domain); vmd = VM_DOMAIN(domain); vm_domain_free_lock(vmd); order = m->order; if (order < VM_NFREEORDER) { /* * The page is enqueued in the physical memory * allocator's free page queues. Moreover, it * is the first page in a power-of-two-sized * run of contiguous free pages. Jump ahead * to the last page within that run, and * continue from there. */ m += (1 << order) - 1; } #if VM_NRESERVLEVEL > 0 else if (vm_reserv_is_page_free(m)) order = 0; #endif vm_domain_free_unlock(vmd); if (order == VM_NFREEORDER) error = EINVAL; } } if (m_mtx != NULL) mtx_unlock(m_mtx); if ((m = SLIST_FIRST(&free)) != NULL) { int cnt; vmd = VM_DOMAIN(domain); cnt = 0; vm_domain_free_lock(vmd); do { MPASS(vm_phys_domain(m) == domain); SLIST_REMOVE_HEAD(&free, plinks.s.ss); vm_phys_free_pages(m, 0); cnt++; } while ((m = SLIST_FIRST(&free)) != NULL); vm_domain_free_unlock(vmd); vm_domain_freecnt_inc(vmd, cnt); } return (error); } #define NRUNS 16 CTASSERT(powerof2(NRUNS)); #define RUN_INDEX(count) ((count) & (NRUNS - 1)) #define MIN_RECLAIM 8 /* * vm_page_reclaim_contig: * * Reclaim allocated, contiguous physical memory satisfying the specified * conditions by relocating the virtual pages using that physical memory. * Returns true if reclamation is successful and false otherwise. Since * relocation requires the allocation of physical pages, reclamation may * fail due to a shortage of free pages. When reclamation fails, callers * are expected to perform vm_wait() before retrying a failed allocation * operation, e.g., vm_page_alloc_contig(). * * The caller must always specify an allocation class through "req". * * allocation classes: * VM_ALLOC_NORMAL normal process request * VM_ALLOC_SYSTEM system *really* needs a page * VM_ALLOC_INTERRUPT interrupt time request * * The optional allocation flags are ignored. * * "npages" must be greater than zero. Both "alignment" and "boundary" * must be a power of two. */ bool vm_page_reclaim_contig_domain(int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) { struct vm_domain *vmd; vm_paddr_t curr_low; vm_page_t m_run, m_runs[NRUNS]; u_long count, reclaimed; int error, i, options, req_class; KASSERT(npages > 0, ("npages is 0")); KASSERT(powerof2(alignment), ("alignment is not a power of 2")); KASSERT(powerof2(boundary), ("boundary is not a power of 2")); req_class = req & VM_ALLOC_CLASS_MASK; /* * The page daemon is allowed to dig deeper into the free page list. */ if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) req_class = VM_ALLOC_SYSTEM; /* * Return if the number of free pages cannot satisfy the requested * allocation. */ vmd = VM_DOMAIN(domain); count = vmd->vmd_free_count; if (count < npages + vmd->vmd_free_reserved || (count < npages + vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) || (count < npages && req_class == VM_ALLOC_INTERRUPT)) return (false); /* * Scan up to three times, relaxing the restrictions ("options") on * the reclamation of reservations and superpages each time. */ for (options = VPSC_NORESERV;;) { /* * Find the highest runs that satisfy the given constraints * and restrictions, and record them in "m_runs". */ curr_low = low; count = 0; for (;;) { m_run = vm_phys_scan_contig(domain, npages, curr_low, high, alignment, boundary, options); if (m_run == NULL) break; curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages); m_runs[RUN_INDEX(count)] = m_run; count++; } /* * Reclaim the highest runs in LIFO (descending) order until * the number of reclaimed pages, "reclaimed", is at least * MIN_RECLAIM. Reset "reclaimed" each time because each * reclamation is idempotent, and runs will (likely) recur * from one scan to the next as restrictions are relaxed. */ reclaimed = 0; for (i = 0; count > 0 && i < NRUNS; i++) { count--; m_run = m_runs[RUN_INDEX(count)]; error = vm_page_reclaim_run(req_class, domain, npages, m_run, high); if (error == 0) { reclaimed += npages; if (reclaimed >= MIN_RECLAIM) return (true); } } /* * Either relax the restrictions on the next scan or return if * the last scan had no restrictions. */ if (options == VPSC_NORESERV) options = VPSC_NOSUPER; else if (options == VPSC_NOSUPER) options = VPSC_ANY; else if (options == VPSC_ANY) return (reclaimed != 0); } } bool vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) { struct vm_domainset_iter di; int domain; bool ret; vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); do { ret = vm_page_reclaim_contig_domain(domain, req, npages, low, high, alignment, boundary); if (ret) break; } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); return (ret); } /* * Set the domain in the appropriate page level domainset. */ void vm_domain_set(struct vm_domain *vmd) { mtx_lock(&vm_domainset_lock); if (!vmd->vmd_minset && vm_paging_min(vmd)) { vmd->vmd_minset = 1; DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains); } if (!vmd->vmd_severeset && vm_paging_severe(vmd)) { vmd->vmd_severeset = 1; DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains); } mtx_unlock(&vm_domainset_lock); } /* * Clear the domain from the appropriate page level domainset. */ void vm_domain_clear(struct vm_domain *vmd) { mtx_lock(&vm_domainset_lock); if (vmd->vmd_minset && !vm_paging_min(vmd)) { vmd->vmd_minset = 0; DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains); if (vm_min_waiters != 0) { vm_min_waiters = 0; wakeup(&vm_min_domains); } } if (vmd->vmd_severeset && !vm_paging_severe(vmd)) { vmd->vmd_severeset = 0; DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains); if (vm_severe_waiters != 0) { vm_severe_waiters = 0; wakeup(&vm_severe_domains); } } /* * If pageout daemon needs pages, then tell it that there are * some free. */ if (vmd->vmd_pageout_pages_needed && vmd->vmd_free_count >= vmd->vmd_pageout_free_min) { wakeup(&vmd->vmd_pageout_pages_needed); vmd->vmd_pageout_pages_needed = 0; } /* See comments in vm_wait_doms(). */ if (vm_pageproc_waiters) { vm_pageproc_waiters = 0; wakeup(&vm_pageproc_waiters); } mtx_unlock(&vm_domainset_lock); } /* * Wait for free pages to exceed the min threshold globally. */ void vm_wait_min(void) { mtx_lock(&vm_domainset_lock); while (vm_page_count_min()) { vm_min_waiters++; msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0); } mtx_unlock(&vm_domainset_lock); } /* * Wait for free pages to exceed the severe threshold globally. */ void vm_wait_severe(void) { mtx_lock(&vm_domainset_lock); while (vm_page_count_severe()) { vm_severe_waiters++; msleep(&vm_severe_domains, &vm_domainset_lock, PVM, "vmwait", 0); } mtx_unlock(&vm_domainset_lock); } u_int vm_wait_count(void) { return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters); } void vm_wait_doms(const domainset_t *wdoms) { /* * We use racey wakeup synchronization to avoid expensive global * locking for the pageproc when sleeping with a non-specific vm_wait. * To handle this, we only sleep for one tick in this instance. It * is expected that most allocations for the pageproc will come from * kmem or vm_page_grab* which will use the more specific and * race-free vm_wait_domain(). */ if (curproc == pageproc) { mtx_lock(&vm_domainset_lock); vm_pageproc_waiters++; msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP, "pageprocwait", 1); } else { /* * XXX Ideally we would wait only until the allocation could * be satisfied. This condition can cause new allocators to * consume all freed pages while old allocators wait. */ mtx_lock(&vm_domainset_lock); if (vm_page_count_min_set(wdoms)) { vm_min_waiters++; msleep(&vm_min_domains, &vm_domainset_lock, PVM | PDROP, "vmwait", 0); } else mtx_unlock(&vm_domainset_lock); } } /* * vm_wait_domain: * * Sleep until free pages are available for allocation. * - Called in various places after failed memory allocations. */ void vm_wait_domain(int domain) { struct vm_domain *vmd; domainset_t wdom; vmd = VM_DOMAIN(domain); vm_domain_free_assert_unlocked(vmd); if (curproc == pageproc) { mtx_lock(&vm_domainset_lock); if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) { vmd->vmd_pageout_pages_needed = 1; msleep(&vmd->vmd_pageout_pages_needed, &vm_domainset_lock, PDROP | PSWP, "VMWait", 0); } else mtx_unlock(&vm_domainset_lock); } else { if (pageproc == NULL) panic("vm_wait in early boot"); DOMAINSET_ZERO(&wdom); DOMAINSET_SET(vmd->vmd_domain, &wdom); vm_wait_doms(&wdom); } } /* * vm_wait: * * Sleep until free pages are available for allocation in the * affinity domains of the obj. If obj is NULL, the domain set * for the calling thread is used. * Called in various places after failed memory allocations. */ void vm_wait(vm_object_t obj) { struct domainset *d; d = NULL; /* * Carefully fetch pointers only once: the struct domainset * itself is ummutable but the pointer might change. */ if (obj != NULL) d = obj->domain.dr_policy; if (d == NULL) d = curthread->td_domain.dr_policy; vm_wait_doms(&d->ds_mask); } /* * vm_domain_alloc_fail: * * Called when a page allocation function fails. Informs the * pagedaemon and performs the requested wait. Requires the * domain_free and object lock on entry. Returns with the * object lock held and free lock released. Returns an error when * retry is necessary. * */ static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req) { vm_domain_free_assert_unlocked(vmd); atomic_add_int(&vmd->vmd_pageout_deficit, max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) { if (object != NULL) VM_OBJECT_WUNLOCK(object); vm_wait_domain(vmd->vmd_domain); if (object != NULL) VM_OBJECT_WLOCK(object); if (req & VM_ALLOC_WAITOK) return (EAGAIN); } return (0); } /* * vm_waitpfault: * * Sleep until free pages are available for allocation. * - Called only in vm_fault so that processes page faulting * can be easily tracked. * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing * processes will be able to grab memory first. Do not change * this balance without careful testing first. */ void vm_waitpfault(struct domainset *dset, int timo) { /* * XXX Ideally we would wait only until the allocation could * be satisfied. This condition can cause new allocators to * consume all freed pages while old allocators wait. */ mtx_lock(&vm_domainset_lock); if (vm_page_count_min_set(&dset->ds_mask)) { vm_min_waiters++; msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP, "pfault", timo); } else mtx_unlock(&vm_domainset_lock); } static struct vm_pagequeue * vm_page_pagequeue(vm_page_t m) { uint8_t queue; if ((queue = atomic_load_8(&m->queue)) == PQ_NONE) return (NULL); return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]); } static inline void vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m) { struct vm_domain *vmd; uint8_t qflags; CRITICAL_ASSERT(curthread); vm_pagequeue_assert_locked(pq); /* * The page daemon is allowed to set m->queue = PQ_NONE without * the page queue lock held. In this case it is about to free the page, * which must not have any queue state. */ qflags = atomic_load_8(&m->aflags); KASSERT(pq == vm_page_pagequeue(m) || (qflags & PGA_QUEUE_STATE_MASK) == 0, ("page %p doesn't belong to queue %p but has aflags %#x", m, pq, qflags)); if ((qflags & PGA_DEQUEUE) != 0) { if (__predict_true((qflags & PGA_ENQUEUED) != 0)) vm_pagequeue_remove(pq, m); vm_page_dequeue_complete(m); counter_u64_add(queue_ops, 1); } else if ((qflags & (PGA_REQUEUE | PGA_REQUEUE_HEAD)) != 0) { if ((qflags & PGA_ENQUEUED) != 0) TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); else { vm_pagequeue_cnt_inc(pq); vm_page_aflag_set(m, PGA_ENQUEUED); } /* * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE. * In particular, if both flags are set in close succession, * only PGA_REQUEUE_HEAD will be applied, even if it was set * first. */ if ((qflags & PGA_REQUEUE_HEAD) != 0) { KASSERT(m->queue == PQ_INACTIVE, ("head enqueue not supported for page %p", m)); vmd = vm_pagequeue_domain(m); TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q); } else TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); vm_page_aflag_clear(m, qflags & (PGA_REQUEUE | PGA_REQUEUE_HEAD)); counter_u64_add(queue_ops, 1); } else { counter_u64_add(queue_nops, 1); } } static void vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq, uint8_t queue) { vm_page_t m; int i; for (i = 0; i < bq->bq_cnt; i++) { m = bq->bq_pa[i]; if (__predict_false(m->queue != queue)) continue; vm_pqbatch_process_page(pq, m); } vm_batchqueue_init(bq); } /* * vm_page_pqbatch_submit: [ internal use only ] * * Enqueue a page in the specified page queue's batched work queue. * The caller must have encoded the requested operation in the page * structure's aflags field. */ void vm_page_pqbatch_submit(vm_page_t m, uint8_t queue) { struct vm_batchqueue *bq; struct vm_pagequeue *pq; int domain; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("page %p is unmanaged", m)); KASSERT(mtx_owned(vm_page_lockptr(m)) || m->object == NULL, ("missing synchronization for page %p", m)); KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue)); domain = vm_phys_domain(m); pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue]; critical_enter(); bq = DPCPU_PTR(pqbatch[domain][queue]); if (vm_batchqueue_insert(bq, m)) { critical_exit(); return; } critical_exit(); vm_pagequeue_lock(pq); critical_enter(); bq = DPCPU_PTR(pqbatch[domain][queue]); vm_pqbatch_process(pq, bq, queue); /* * The page may have been logically dequeued before we acquired the * page queue lock. In this case, since we either hold the page lock * or the page is being freed, a different thread cannot be concurrently * enqueuing the page. */ if (__predict_true(m->queue == queue)) vm_pqbatch_process_page(pq, m); else { KASSERT(m->queue == PQ_NONE, ("invalid queue transition for page %p", m)); KASSERT((m->aflags & PGA_ENQUEUED) == 0, ("page %p is enqueued with invalid queue index", m)); } vm_pagequeue_unlock(pq); critical_exit(); } /* * vm_page_pqbatch_drain: [ internal use only ] * * Force all per-CPU page queue batch queues to be drained. This is * intended for use in severe memory shortages, to ensure that pages * do not remain stuck in the batch queues. */ void vm_page_pqbatch_drain(void) { struct thread *td; struct vm_domain *vmd; struct vm_pagequeue *pq; int cpu, domain, queue; td = curthread; CPU_FOREACH(cpu) { thread_lock(td); sched_bind(td, cpu); thread_unlock(td); for (domain = 0; domain < vm_ndomains; domain++) { vmd = VM_DOMAIN(domain); for (queue = 0; queue < PQ_COUNT; queue++) { pq = &vmd->vmd_pagequeues[queue]; vm_pagequeue_lock(pq); critical_enter(); vm_pqbatch_process(pq, DPCPU_PTR(pqbatch[domain][queue]), queue); critical_exit(); vm_pagequeue_unlock(pq); } } } thread_lock(td); sched_unbind(td); thread_unlock(td); } /* * Complete the logical removal of a page from a page queue. We must be * careful to synchronize with the page daemon, which may be concurrently * examining the page with only the page lock held. The page must not be * in a state where it appears to be logically enqueued. */ static void vm_page_dequeue_complete(vm_page_t m) { m->queue = PQ_NONE; atomic_thread_fence_rel(); vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK); } /* * vm_page_dequeue_deferred: [ internal use only ] * * Request removal of the given page from its current page * queue. Physical removal from the queue may be deferred * indefinitely. * * The page must be locked. */ void vm_page_dequeue_deferred(vm_page_t m) { uint8_t queue; vm_page_assert_locked(m); if ((queue = vm_page_queue(m)) == PQ_NONE) return; /* * Set PGA_DEQUEUE if it is not already set to handle a concurrent call * to vm_page_dequeue_deferred_free(). In particular, avoid modifying * the page's queue state once vm_page_dequeue_deferred_free() has been * called. In the event of a race, two batch queue entries for the page * will be created, but the second will have no effect. */ if (vm_page_pqstate_cmpset(m, queue, queue, PGA_DEQUEUE, PGA_DEQUEUE)) vm_page_pqbatch_submit(m, queue); } /* * A variant of vm_page_dequeue_deferred() that does not assert the page * lock and is only to be called from vm_page_free_prep(). Because the * page is being freed, we can assume that nothing other than the page * daemon is scheduling queue operations on this page, so we get for * free the mutual exclusion that is otherwise provided by the page lock. * To handle races, the page daemon must take care to atomically check * for PGA_DEQUEUE when updating queue state. */ static void vm_page_dequeue_deferred_free(vm_page_t m) { uint8_t queue; KASSERT(m->ref_count == 0, ("page %p has references", m)); for (;;) { if ((m->aflags & PGA_DEQUEUE) != 0) return; atomic_thread_fence_acq(); if ((queue = atomic_load_8(&m->queue)) == PQ_NONE) return; if (vm_page_pqstate_cmpset(m, queue, queue, PGA_DEQUEUE, PGA_DEQUEUE)) { vm_page_pqbatch_submit(m, queue); break; } } } /* * vm_page_dequeue: * * Remove the page from whichever page queue it's in, if any. * The page must either be locked or unallocated. This constraint * ensures that the queue state of the page will remain consistent * after this function returns. */ void vm_page_dequeue(vm_page_t m) { struct vm_pagequeue *pq, *pq1; uint8_t aflags; KASSERT(mtx_owned(vm_page_lockptr(m)) || m->object == NULL, ("page %p is allocated and unlocked", m)); for (pq = vm_page_pagequeue(m);; pq = pq1) { if (pq == NULL) { /* * A thread may be concurrently executing * vm_page_dequeue_complete(). Ensure that all queue * state is cleared before we return. */ aflags = atomic_load_8(&m->aflags); if ((aflags & PGA_QUEUE_STATE_MASK) == 0) return; KASSERT((aflags & PGA_DEQUEUE) != 0, ("page %p has unexpected queue state flags %#x", m, aflags)); /* * Busy wait until the thread updating queue state is * finished. Such a thread must be executing in a * critical section. */ cpu_spinwait(); pq1 = vm_page_pagequeue(m); continue; } vm_pagequeue_lock(pq); if ((pq1 = vm_page_pagequeue(m)) == pq) break; vm_pagequeue_unlock(pq); } KASSERT(pq == vm_page_pagequeue(m), ("%s: page %p migrated directly between queues", __func__, m)); KASSERT((m->aflags & PGA_DEQUEUE) != 0 || mtx_owned(vm_page_lockptr(m)), ("%s: queued unlocked page %p", __func__, m)); if ((m->aflags & PGA_ENQUEUED) != 0) vm_pagequeue_remove(pq, m); vm_page_dequeue_complete(m); vm_pagequeue_unlock(pq); } /* * Schedule the given page for insertion into the specified page queue. * Physical insertion of the page may be deferred indefinitely. */ static void vm_page_enqueue(vm_page_t m, uint8_t queue) { vm_page_assert_locked(m); KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0, ("%s: page %p is already enqueued", __func__, m)); m->queue = queue; if ((m->aflags & PGA_REQUEUE) == 0) vm_page_aflag_set(m, PGA_REQUEUE); vm_page_pqbatch_submit(m, queue); } /* * vm_page_requeue: [ internal use only ] * * Schedule a requeue of the given page. * * The page must be locked. */ void vm_page_requeue(vm_page_t m) { vm_page_assert_locked(m); KASSERT(vm_page_queue(m) != PQ_NONE, ("%s: page %p is not logically enqueued", __func__, m)); if ((m->aflags & PGA_REQUEUE) == 0) vm_page_aflag_set(m, PGA_REQUEUE); vm_page_pqbatch_submit(m, atomic_load_8(&m->queue)); } /* * vm_page_swapqueue: [ internal use only ] * * Move the page from one queue to another, or to the tail of its * current queue, in the face of a possible concurrent call to * vm_page_dequeue_deferred_free(). */ void vm_page_swapqueue(vm_page_t m, uint8_t oldq, uint8_t newq) { struct vm_pagequeue *pq; vm_page_t next; bool queued; KASSERT(oldq < PQ_COUNT && newq < PQ_COUNT && oldq != newq, ("vm_page_swapqueue: invalid queues (%d, %d)", oldq, newq)); vm_page_assert_locked(m); pq = &vm_pagequeue_domain(m)->vmd_pagequeues[oldq]; vm_pagequeue_lock(pq); /* * The physical queue state might change at any point before the page * queue lock is acquired, so we must verify that we hold the correct * lock before proceeding. */ if (__predict_false(m->queue != oldq)) { vm_pagequeue_unlock(pq); return; } /* * Once the queue index of the page changes, there is nothing * synchronizing with further updates to the physical queue state. * Therefore we must remove the page from the queue now in anticipation * of a successful commit, and be prepared to roll back. */ if (__predict_true((m->aflags & PGA_ENQUEUED) != 0)) { next = TAILQ_NEXT(m, plinks.q); TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); vm_page_aflag_clear(m, PGA_ENQUEUED); queued = true; } else { queued = false; } /* * Atomically update the queue field and set PGA_REQUEUE while * ensuring that PGA_DEQUEUE has not been set. */ if (__predict_false(!vm_page_pqstate_cmpset(m, oldq, newq, PGA_DEQUEUE, PGA_REQUEUE))) { if (queued) { vm_page_aflag_set(m, PGA_ENQUEUED); if (next != NULL) TAILQ_INSERT_BEFORE(next, m, plinks.q); else TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); } vm_pagequeue_unlock(pq); return; } vm_pagequeue_cnt_dec(pq); vm_pagequeue_unlock(pq); vm_page_pqbatch_submit(m, newq); } /* * vm_page_free_prep: * * Prepares the given page to be put on the free list, * disassociating it from any VM object. The caller may return * the page to the free list only if this function returns true. * * The object must be locked. The page must be locked if it is * managed. */ bool vm_page_free_prep(vm_page_t m) { /* * Synchronize with threads that have dropped a reference to this * page. */ atomic_thread_fence_acq(); #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP) if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) { uint64_t *p; int i; p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++) KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx", m, i, (uintmax_t)*p)); } #endif if ((m->oflags & VPO_UNMANAGED) == 0) { KASSERT(!pmap_page_is_mapped(m), ("vm_page_free_prep: freeing mapped page %p", m)); KASSERT((m->aflags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0, ("vm_page_free_prep: mapping flags set in page %p", m)); } else { KASSERT(m->queue == PQ_NONE, ("vm_page_free_prep: unmanaged page %p is queued", m)); } VM_CNT_INC(v_tfree); if (vm_page_sbusied(m)) panic("vm_page_free_prep: freeing shared busy page %p", m); if (m->object != NULL) { vm_page_object_remove(m); /* * The object reference can be released without an atomic * operation. */ KASSERT((m->flags & PG_FICTITIOUS) != 0 || m->ref_count == VPRC_OBJREF, ("vm_page_free_prep: page %p has unexpected ref_count %u", m, m->ref_count)); m->object = NULL; m->ref_count -= VPRC_OBJREF; } if (vm_page_xbusied(m)) vm_page_xunbusy(m); /* * If fictitious remove object association and * return. */ if ((m->flags & PG_FICTITIOUS) != 0) { KASSERT(m->ref_count == 1, ("fictitious page %p is referenced", m)); KASSERT(m->queue == PQ_NONE, ("fictitious page %p is queued", m)); return (false); } /* * Pages need not be dequeued before they are returned to the physical * memory allocator, but they must at least be marked for a deferred * dequeue. */ if ((m->oflags & VPO_UNMANAGED) == 0) vm_page_dequeue_deferred_free(m); m->valid = 0; vm_page_undirty(m); if (m->ref_count != 0) panic("vm_page_free_prep: page %p has references", m); /* * Restore the default memory attribute to the page. */ if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); #if VM_NRESERVLEVEL > 0 /* * Determine whether the page belongs to a reservation. If the page was * allocated from a per-CPU cache, it cannot belong to a reservation, so * as an optimization, we avoid the check in that case. */ if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m)) return (false); #endif return (true); } /* * vm_page_free_toq: * * Returns the given page to the free list, disassociating it * from any VM object. * * The object must be locked. The page must be locked if it is * managed. */ void vm_page_free_toq(vm_page_t m) { struct vm_domain *vmd; uma_zone_t zone; if (!vm_page_free_prep(m)) return; vmd = vm_pagequeue_domain(m); zone = vmd->vmd_pgcache[m->pool].zone; if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) { uma_zfree(zone, m); return; } vm_domain_free_lock(vmd); vm_phys_free_pages(m, 0); vm_domain_free_unlock(vmd); vm_domain_freecnt_inc(vmd, 1); } /* * vm_page_free_pages_toq: * * Returns a list of pages to the free list, disassociating it * from any VM object. In other words, this is equivalent to * calling vm_page_free_toq() for each page of a list of VM objects. * * The objects must be locked. The pages must be locked if it is * managed. */ void vm_page_free_pages_toq(struct spglist *free, bool update_wire_count) { vm_page_t m; int count; if (SLIST_EMPTY(free)) return; count = 0; while ((m = SLIST_FIRST(free)) != NULL) { count++; SLIST_REMOVE_HEAD(free, plinks.s.ss); vm_page_free_toq(m); } if (update_wire_count) vm_wire_sub(count); } /* * Mark this page as wired down, preventing reclamation by the page daemon * or when the containing object is destroyed. */ void vm_page_wire(vm_page_t m) { u_int old; KASSERT(m->object != NULL, ("vm_page_wire: page %p does not belong to an object", m)); if (!vm_page_busied(m)) VM_OBJECT_ASSERT_LOCKED(m->object); KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(m->ref_count) >= 1, ("vm_page_wire: fictitious page %p has zero wirings", m)); old = atomic_fetchadd_int(&m->ref_count, 1); KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX, ("vm_page_wire: counter overflow for page %p", m)); if (VPRC_WIRE_COUNT(old) == 0) vm_wire_add(1); } /* * Attempt to wire a mapped page following a pmap lookup of that page. * This may fail if a thread is concurrently tearing down mappings of the page. */ bool vm_page_wire_mapped(vm_page_t m) { u_int old; old = m->ref_count; do { KASSERT(old > 0, ("vm_page_wire_mapped: wiring unreferenced page %p", m)); if ((old & VPRC_BLOCKED) != 0) return (false); } while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1)); if (VPRC_WIRE_COUNT(old) == 0) vm_wire_add(1); return (true); } /* * Release one wiring of the specified page, potentially allowing it to be * paged out. * * Only managed pages belonging to an object can be paged out. If the number * of wirings transitions to zero and the page is eligible for page out, then * the page is added to the specified paging queue. If the released wiring * represented the last reference to the page, the page is freed. * * A managed page must be locked. */ void vm_page_unwire(vm_page_t m, uint8_t queue) { u_int old; bool locked; KASSERT(queue < PQ_COUNT, ("vm_page_unwire: invalid queue %u request for page %p", queue, m)); if ((m->oflags & VPO_UNMANAGED) != 0) { if (vm_page_unwire_noq(m) && m->ref_count == 0) vm_page_free(m); return; } /* * Update LRU state before releasing the wiring reference. * We only need to do this once since we hold the page lock. * Use a release store when updating the reference count to * synchronize with vm_page_free_prep(). */ old = m->ref_count; locked = false; do { KASSERT(VPRC_WIRE_COUNT(old) > 0, ("vm_page_unwire: wire count underflow for page %p", m)); if (!locked && VPRC_WIRE_COUNT(old) == 1) { vm_page_lock(m); locked = true; if (queue == PQ_ACTIVE && vm_page_queue(m) == PQ_ACTIVE) vm_page_reference(m); else vm_page_mvqueue(m, queue); } } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1)); /* * Release the lock only after the wiring is released, to ensure that * the page daemon does not encounter and dequeue the page while it is * still wired. */ if (locked) vm_page_unlock(m); if (VPRC_WIRE_COUNT(old) == 1) { vm_wire_sub(1); if (old == 1) vm_page_free(m); } } /* * Unwire a page without (re-)inserting it into a page queue. It is up * to the caller to enqueue, requeue, or free the page as appropriate. * In most cases involving managed pages, vm_page_unwire() should be used * instead. */ bool vm_page_unwire_noq(vm_page_t m) { u_int old; old = vm_page_drop(m, 1); KASSERT(VPRC_WIRE_COUNT(old) != 0, ("vm_page_unref: counter underflow for page %p", m)); KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1, ("vm_page_unref: missing ref on fictitious page %p", m)); if (VPRC_WIRE_COUNT(old) > 1) return (false); vm_wire_sub(1); return (true); } /* * Ensure that the page is in the specified page queue. If the page is * active or being moved to the active queue, ensure that its act_count is * at least ACT_INIT but do not otherwise mess with it. Otherwise, ensure that * the page is at the tail of its page queue. * * The page may be wired. The caller should release its wiring reference * before releasing the page lock, otherwise the page daemon may immediately * dequeue the page. * * A managed page must be locked. */ static __always_inline void vm_page_mvqueue(vm_page_t m, const uint8_t nqueue) { vm_page_assert_locked(m); KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("vm_page_mvqueue: page %p is unmanaged", m)); if (vm_page_queue(m) != nqueue) { vm_page_dequeue(m); vm_page_enqueue(m, nqueue); } else if (nqueue != PQ_ACTIVE) { vm_page_requeue(m); } if (nqueue == PQ_ACTIVE && m->act_count < ACT_INIT) m->act_count = ACT_INIT; } /* * Put the specified page on the active list (if appropriate). */ void vm_page_activate(vm_page_t m) { if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m)) return; vm_page_mvqueue(m, PQ_ACTIVE); } /* * Move the specified page to the tail of the inactive queue, or requeue * the page if it is already in the inactive queue. */ void vm_page_deactivate(vm_page_t m) { if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m)) return; vm_page_mvqueue(m, PQ_INACTIVE); } /* * Move the specified page close to the head of the inactive queue, * bypassing LRU. A marker page is used to maintain FIFO ordering. * As with regular enqueues, we use a per-CPU batch queue to reduce * contention on the page queue lock. */ static void _vm_page_deactivate_noreuse(vm_page_t m) { vm_page_assert_locked(m); if (!vm_page_inactive(m)) { vm_page_dequeue(m); m->queue = PQ_INACTIVE; } if ((m->aflags & PGA_REQUEUE_HEAD) == 0) vm_page_aflag_set(m, PGA_REQUEUE_HEAD); vm_page_pqbatch_submit(m, PQ_INACTIVE); } void vm_page_deactivate_noreuse(vm_page_t m) { KASSERT(m->object != NULL, ("vm_page_deactivate_noreuse: page %p has no object", m)); if ((m->oflags & VPO_UNMANAGED) == 0 && !vm_page_wired(m)) _vm_page_deactivate_noreuse(m); } /* * Put a page in the laundry, or requeue it if it is already there. */ void vm_page_launder(vm_page_t m) { if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m)) return; vm_page_mvqueue(m, PQ_LAUNDRY); } /* * Put a page in the PQ_UNSWAPPABLE holding queue. */ void vm_page_unswappable(vm_page_t m) { vm_page_assert_locked(m); KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0, ("page %p already unswappable", m)); vm_page_dequeue(m); vm_page_enqueue(m, PQ_UNSWAPPABLE); } static void vm_page_release_toq(vm_page_t m, int flags) { vm_page_assert_locked(m); /* * Use a check of the valid bits to determine whether we should * accelerate reclamation of the page. The object lock might not be * held here, in which case the check is racy. At worst we will either * accelerate reclamation of a valid page and violate LRU, or * unnecessarily defer reclamation of an invalid page. * * If we were asked to not cache the page, place it near the head of the * inactive queue so that is reclaimed sooner. */ if ((flags & (VPR_TRYFREE | VPR_NOREUSE)) != 0 || m->valid == 0) _vm_page_deactivate_noreuse(m); else if (vm_page_active(m)) vm_page_reference(m); else vm_page_mvqueue(m, PQ_INACTIVE); } /* * Unwire a page and either attempt to free it or re-add it to the page queues. */ void vm_page_release(vm_page_t m, int flags) { vm_object_t object; u_int old; bool locked; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("vm_page_release: page %p is unmanaged", m)); if ((flags & VPR_TRYFREE) != 0) { for (;;) { object = (vm_object_t)atomic_load_ptr(&m->object); if (object == NULL) break; /* Depends on type-stability. */ if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object)) { object = NULL; break; } if (object == m->object) break; VM_OBJECT_WUNLOCK(object); } if (__predict_true(object != NULL)) { vm_page_release_locked(m, flags); VM_OBJECT_WUNLOCK(object); return; } } /* * Update LRU state before releasing the wiring reference. * Use a release store when updating the reference count to * synchronize with vm_page_free_prep(). */ old = m->ref_count; locked = false; do { KASSERT(VPRC_WIRE_COUNT(old) > 0, ("vm_page_unwire: wire count underflow for page %p", m)); if (!locked && VPRC_WIRE_COUNT(old) == 1) { vm_page_lock(m); locked = true; vm_page_release_toq(m, flags); } } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1)); /* * Release the lock only after the wiring is released, to ensure that * the page daemon does not encounter and dequeue the page while it is * still wired. */ if (locked) vm_page_unlock(m); if (VPRC_WIRE_COUNT(old) == 1) { vm_wire_sub(1); if (old == 1) vm_page_free(m); } } /* See vm_page_release(). */ void vm_page_release_locked(vm_page_t m, int flags) { VM_OBJECT_ASSERT_WLOCKED(m->object); KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("vm_page_release_locked: page %p is unmanaged", m)); if (vm_page_unwire_noq(m)) { if ((flags & VPR_TRYFREE) != 0 && (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) && m->dirty == 0 && !vm_page_busied(m)) { vm_page_free(m); } else { vm_page_lock(m); vm_page_release_toq(m, flags); vm_page_unlock(m); } } } static bool vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t)) { u_int old; KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0, ("vm_page_try_blocked_op: page %p has no object", m)); KASSERT(vm_page_busied(m), ("vm_page_try_blocked_op: page %p is not busy", m)); VM_OBJECT_ASSERT_LOCKED(m->object); old = m->ref_count; do { KASSERT(old != 0, ("vm_page_try_blocked_op: page %p has no references", m)); if (VPRC_WIRE_COUNT(old) != 0) return (false); } while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED)); (op)(m); /* * If the object is read-locked, new wirings may be created via an * object lookup. */ old = vm_page_drop(m, VPRC_BLOCKED); KASSERT(!VM_OBJECT_WOWNED(m->object) || old == (VPRC_BLOCKED | VPRC_OBJREF), ("vm_page_try_blocked_op: unexpected refcount value %u for %p", old, m)); return (true); } /* * Atomically check for wirings and remove all mappings of the page. */ bool vm_page_try_remove_all(vm_page_t m) { return (vm_page_try_blocked_op(m, pmap_remove_all)); } /* * Atomically check for wirings and remove all writeable mappings of the page. */ bool vm_page_try_remove_write(vm_page_t m) { return (vm_page_try_blocked_op(m, pmap_remove_write)); } /* * vm_page_advise * * Apply the specified advice to the given page. * * The object and page must be locked. */ void vm_page_advise(vm_page_t m, int advice) { vm_page_assert_locked(m); VM_OBJECT_ASSERT_WLOCKED(m->object); if (advice == MADV_FREE) /* * Mark the page clean. This will allow the page to be freed * without first paging it out. MADV_FREE pages are often * quickly reused by malloc(3), so we do not do anything that * would result in a page fault on a later access. */ vm_page_undirty(m); else if (advice != MADV_DONTNEED) { if (advice == MADV_WILLNEED) vm_page_activate(m); return; } /* * Clear any references to the page. Otherwise, the page daemon will * immediately reactivate the page. */ vm_page_aflag_clear(m, PGA_REFERENCED); if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) vm_page_dirty(m); /* * Place clean pages near the head of the inactive queue rather than * the tail, thus defeating the queue's LRU operation and ensuring that * the page will be reused quickly. Dirty pages not already in the * laundry are moved there. */ if (m->dirty == 0) vm_page_deactivate_noreuse(m); else if (!vm_page_in_laundry(m)) vm_page_launder(m); } /* * Grab a page, waiting until we are waken up due to the page * changing state. We keep on waiting, if the page continues * to be in the object. If the page doesn't exist, first allocate it * and then conditionally zero it. * * This routine may sleep. * * The object must be locked on entry. The lock will, however, be released * and reacquired if the routine sleeps. */ vm_page_t vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) { vm_page_t m; int sleep; int pflags; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || (allocflags & VM_ALLOC_IGN_SBUSY) != 0, ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); pflags = allocflags & ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL | VM_ALLOC_NOBUSY); if ((allocflags & VM_ALLOC_NOWAIT) == 0) pflags |= VM_ALLOC_WAITFAIL; if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) pflags |= VM_ALLOC_SBUSY; retrylookup: if ((m = vm_page_lookup(object, pindex)) != NULL) { if ((allocflags & (VM_ALLOC_IGN_SBUSY | VM_ALLOC_SBUSY)) != 0) sleep = !vm_page_trysbusy(m); else sleep = !vm_page_tryxbusy(m); if (sleep) { if ((allocflags & VM_ALLOC_NOWAIT) != 0) return (NULL); /* * Reference the page before unlocking and * sleeping so that the page daemon is less * likely to reclaim it. */ if ((allocflags & VM_ALLOC_NOCREAT) == 0) vm_page_aflag_set(m, PGA_REFERENCED); vm_page_busy_sleep(m, "pgrbwt", (allocflags & VM_ALLOC_IGN_SBUSY) != 0); VM_OBJECT_WLOCK(object); if ((allocflags & VM_ALLOC_WAITFAIL) != 0) return (NULL); goto retrylookup; } else { if ((allocflags & VM_ALLOC_WIRED) != 0) vm_page_wire(m); goto out; } } if ((allocflags & VM_ALLOC_NOCREAT) != 0) return (NULL); m = vm_page_alloc(object, pindex, pflags); if (m == NULL) { if ((allocflags & VM_ALLOC_NOWAIT) != 0) return (NULL); goto retrylookup; } if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) pmap_zero_page(m); out: if ((allocflags & VM_ALLOC_NOBUSY) != 0) { if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) vm_page_sunbusy(m); else vm_page_xunbusy(m); } return (m); } /* * Grab a page and make it valid, paging in if necessary. Pages missing from * their pager are zero filled and validated. */ int vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags) { vm_page_t m; bool sleep, xbusy; int pflags; int rv; KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || (allocflags & VM_ALLOC_IGN_SBUSY) != 0, ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); KASSERT((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, ("vm_page_grab_valid: Invalid flags 0x%X", allocflags)); VM_OBJECT_ASSERT_WLOCKED(object); pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY); pflags |= VM_ALLOC_WAITFAIL; retrylookup: xbusy = false; if ((m = vm_page_lookup(object, pindex)) != NULL) { /* * If the page is fully valid it can only become invalid * with the object lock held. If it is not valid it can * become valid with the busy lock held. Therefore, we * may unnecessarily lock the exclusive busy here if we * race with I/O completion not using the object lock. * However, we will not end up with an invalid page and a * shared lock. */ - if (m->valid != VM_PAGE_BITS_ALL || + if (!vm_page_all_valid(m) || (allocflags & (VM_ALLOC_IGN_SBUSY | VM_ALLOC_SBUSY)) == 0) { sleep = !vm_page_tryxbusy(m); xbusy = true; } else sleep = !vm_page_trysbusy(m); if (sleep) { /* * Reference the page before unlocking and * sleeping so that the page daemon is less * likely to reclaim it. */ if ((allocflags & VM_ALLOC_NOCREAT) == 0) vm_page_aflag_set(m, PGA_REFERENCED); vm_page_busy_sleep(m, "pgrbwt", (allocflags & VM_ALLOC_IGN_SBUSY) != 0); VM_OBJECT_WLOCK(object); goto retrylookup; } if ((allocflags & VM_ALLOC_NOCREAT) != 0 && - m->valid != VM_PAGE_BITS_ALL) { + !vm_page_all_valid(m)) { if (xbusy) vm_page_xunbusy(m); else vm_page_sunbusy(m); *mp = NULL; return (VM_PAGER_FAIL); } if ((allocflags & VM_ALLOC_WIRED) != 0) vm_page_wire(m); - if (m->valid == VM_PAGE_BITS_ALL) + if (vm_page_all_valid(m)) goto out; } else if ((allocflags & VM_ALLOC_NOCREAT) != 0) { *mp = NULL; return (VM_PAGER_FAIL); } else if ((m = vm_page_alloc(object, pindex, pflags)) != NULL) { xbusy = true; } else { goto retrylookup; } vm_page_assert_xbusied(m); MPASS(xbusy); if (vm_pager_has_page(object, pindex, NULL, NULL)) { rv = vm_pager_get_pages(object, &m, 1, NULL, NULL); if (rv != VM_PAGER_OK) { if (allocflags & VM_ALLOC_WIRED) vm_page_unwire_noq(m); vm_page_free(m); *mp = NULL; return (rv); } - MPASS(m->valid == VM_PAGE_BITS_ALL); + MPASS(vm_page_all_valid(m)); } else { vm_page_zero_invalid(m, TRUE); } out: if ((allocflags & VM_ALLOC_NOBUSY) != 0) { if (xbusy) vm_page_xunbusy(m); else vm_page_sunbusy(m); } if ((allocflags & VM_ALLOC_SBUSY) != 0 && xbusy) vm_page_busy_downgrade(m); *mp = m; return (VM_PAGER_OK); } /* * Return the specified range of pages from the given object. For each * page offset within the range, if a page already exists within the object * at that offset and it is busy, then wait for it to change state. If, * instead, the page doesn't exist, then allocate it. * * The caller must always specify an allocation class. * * allocation classes: * VM_ALLOC_NORMAL normal process request * VM_ALLOC_SYSTEM system *really* needs the pages * * The caller must always specify that the pages are to be busied and/or * wired. * * optional allocation flags: * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages * VM_ALLOC_NOBUSY do not exclusive busy the page * VM_ALLOC_NOWAIT do not sleep * VM_ALLOC_SBUSY set page to sbusy state * VM_ALLOC_WIRED wire the pages * VM_ALLOC_ZERO zero and validate any invalid pages * * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it * may return a partial prefix of the requested range. */ int vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, vm_page_t *ma, int count) { vm_page_t m, mpred; int pflags; int i; bool sleep; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0, ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed")); KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 || (allocflags & VM_ALLOC_WIRED) != 0, ("vm_page_grab_pages: the pages must be busied or wired")); KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || (allocflags & VM_ALLOC_IGN_SBUSY) != 0, ("vm_page_grab_pages: VM_ALLOC_SBUSY/IGN_SBUSY mismatch")); if (count == 0) return (0); pflags = allocflags & ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL | VM_ALLOC_NOBUSY); if ((allocflags & VM_ALLOC_NOWAIT) == 0) pflags |= VM_ALLOC_WAITFAIL; if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) pflags |= VM_ALLOC_SBUSY; i = 0; retrylookup: m = vm_radix_lookup_le(&object->rtree, pindex + i); if (m == NULL || m->pindex != pindex + i) { mpred = m; m = NULL; } else mpred = TAILQ_PREV(m, pglist, listq); for (; i < count; i++) { if (m != NULL) { if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0) sleep = !vm_page_trysbusy(m); else sleep = !vm_page_tryxbusy(m); if (sleep) { if ((allocflags & VM_ALLOC_NOWAIT) != 0) break; /* * Reference the page before unlocking and * sleeping so that the page daemon is less * likely to reclaim it. */ if ((allocflags & VM_ALLOC_NOCREAT) == 0) vm_page_aflag_set(m, PGA_REFERENCED); vm_page_busy_sleep(m, "grbmaw", (allocflags & VM_ALLOC_IGN_SBUSY) != 0); VM_OBJECT_WLOCK(object); goto retrylookup; } if ((allocflags & VM_ALLOC_WIRED) != 0) vm_page_wire(m); } else { if ((allocflags & VM_ALLOC_NOCREAT) != 0) break; m = vm_page_alloc_after(object, pindex + i, pflags | VM_ALLOC_COUNT(count - i), mpred); if (m == NULL) { if ((allocflags & VM_ALLOC_NOWAIT) != 0) break; goto retrylookup; } } - if (m->valid == 0 && (allocflags & VM_ALLOC_ZERO) != 0) { + if (vm_page_none_valid(m) && + (allocflags & VM_ALLOC_ZERO) != 0) { if ((m->flags & PG_ZERO) == 0) pmap_zero_page(m); - m->valid = VM_PAGE_BITS_ALL; + vm_page_valid(m); } if ((allocflags & VM_ALLOC_NOBUSY) != 0) { if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) vm_page_sunbusy(m); else vm_page_xunbusy(m); } ma[i] = mpred = m; m = vm_page_next(m); } return (i); } /* * Mapping function for valid or dirty bits in a page. * * Inputs are required to range within a page. */ vm_page_bits_t vm_page_bits(int base, int size) { int first_bit; int last_bit; KASSERT( base + size <= PAGE_SIZE, ("vm_page_bits: illegal base/size %d/%d", base, size) ); if (size == 0) /* handle degenerate case */ return (0); first_bit = base >> DEV_BSHIFT; last_bit = (base + size - 1) >> DEV_BSHIFT; return (((vm_page_bits_t)2 << last_bit) - ((vm_page_bits_t)1 << first_bit)); } +static inline void +vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set) +{ + +#if PAGE_SIZE == 32768 + atomic_set_64((uint64_t *)bits, set); +#elif PAGE_SIZE == 16384 + atomic_set_32((uint32_t *)bits, set); +#elif (PAGE_SIZE == 8192) && defined(atomic_set_16) + atomic_set_16((uint16_t *)bits, set); +#elif (PAGE_SIZE == 4096) && defined(atomic_set_8) + atomic_set_8((uint8_t *)bits, set); +#else /* PAGE_SIZE <= 8192 */ + uintptr_t addr; + int shift; + + addr = (uintptr_t)bits; + /* + * Use a trick to perform a 32-bit atomic on the + * containing aligned word, to not depend on the existence + * of atomic_{set, clear}_{8, 16}. + */ + shift = addr & (sizeof(uint32_t) - 1); +#if BYTE_ORDER == BIG_ENDIAN + shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; +#else + shift *= NBBY; +#endif + addr &= ~(sizeof(uint32_t) - 1); + atomic_set_32((uint32_t *)addr, set << shift); +#endif /* PAGE_SIZE */ +} + +static inline void +vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear) +{ + +#if PAGE_SIZE == 32768 + atomic_clear_64((uint64_t *)bits, clear); +#elif PAGE_SIZE == 16384 + atomic_clear_32((uint32_t *)bits, clear); +#elif (PAGE_SIZE == 8192) && defined(atomic_clear_16) + atomic_clear_16((uint16_t *)bits, clear); +#elif (PAGE_SIZE == 4096) && defined(atomic_clear_8) + atomic_clear_8((uint8_t *)bits, clear); +#else /* PAGE_SIZE <= 8192 */ + uintptr_t addr; + int shift; + + addr = (uintptr_t)bits; + /* + * Use a trick to perform a 32-bit atomic on the + * containing aligned word, to not depend on the existence + * of atomic_{set, clear}_{8, 16}. + */ + shift = addr & (sizeof(uint32_t) - 1); +#if BYTE_ORDER == BIG_ENDIAN + shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; +#else + shift *= NBBY; +#endif + addr &= ~(sizeof(uint32_t) - 1); + atomic_clear_32((uint32_t *)addr, clear << shift); +#endif /* PAGE_SIZE */ +} + /* * vm_page_set_valid_range: * * Sets portions of a page valid. The arguments are expected * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive * of any partial chunks touched by the range. The invalid portion of * such chunks will be zeroed. * * (base + size) must be less then or equal to PAGE_SIZE. */ void vm_page_set_valid_range(vm_page_t m, int base, int size) { int endoff, frag; + vm_page_bits_t pagebits; - VM_OBJECT_ASSERT_WLOCKED(m->object); + vm_page_assert_busied(m); if (size == 0) /* handle degenerate case */ return; /* * If the base is not DEV_BSIZE aligned and the valid * bit is clear, we have to zero out a portion of the * first block. */ if ((frag = rounddown2(base, DEV_BSIZE)) != base && (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) pmap_zero_page_area(m, frag, base - frag); /* * If the ending offset is not DEV_BSIZE aligned and the * valid bit is clear, we have to zero out a portion of * the last block. */ endoff = base + size; if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) pmap_zero_page_area(m, endoff, DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); /* * Assert that no previously invalid block that is now being validated * is already dirty. */ KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, ("vm_page_set_valid_range: page %p is dirty", m)); /* * Set valid bits inclusive of any overlap. */ - m->valid |= vm_page_bits(base, size); + pagebits = vm_page_bits(base, size); + if (vm_page_xbusied(m)) + m->valid |= pagebits; + else + vm_page_bits_set(m, &m->valid, pagebits); } /* * Clear the given bits from the specified page's dirty field. */ static __inline void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) { - uintptr_t addr; -#if PAGE_SIZE < 16384 - int shift; -#endif + vm_page_assert_busied(m); + /* - * If the object is locked and the page is neither exclusive busy nor - * write mapped, then the page's dirty field cannot possibly be - * set by a concurrent pmap operation. + * If the page is xbusied and not write mapped we are the + * only thread that can modify dirty bits. Otherwise, The pmap + * layer can call vm_page_dirty() without holding a distinguished + * lock. The combination of page busy and atomic operations + * suffice to guarantee consistency of the page dirty field. */ - VM_OBJECT_ASSERT_WLOCKED(m->object); - if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) + if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) m->dirty &= ~pagebits; - else { - /* - * The pmap layer can call vm_page_dirty() without - * holding a distinguished lock. The combination of - * the object's lock and an atomic operation suffice - * to guarantee consistency of the page dirty field. - * - * For PAGE_SIZE == 32768 case, compiler already - * properly aligns the dirty field, so no forcible - * alignment is needed. Only require existence of - * atomic_clear_64 when page size is 32768. - */ - addr = (uintptr_t)&m->dirty; -#if PAGE_SIZE == 32768 - atomic_clear_64((uint64_t *)addr, pagebits); -#elif PAGE_SIZE == 16384 - atomic_clear_32((uint32_t *)addr, pagebits); -#else /* PAGE_SIZE <= 8192 */ - /* - * Use a trick to perform a 32-bit atomic on the - * containing aligned word, to not depend on the existence - * of atomic_clear_{8, 16}. - */ - shift = addr & (sizeof(uint32_t) - 1); -#if BYTE_ORDER == BIG_ENDIAN - shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY; -#else - shift *= NBBY; -#endif - addr &= ~(sizeof(uint32_t) - 1); - atomic_clear_32((uint32_t *)addr, pagebits << shift); -#endif /* PAGE_SIZE */ - } + else + vm_page_bits_clear(m, &m->dirty, pagebits); } /* * vm_page_set_validclean: * * Sets portions of a page valid and clean. The arguments are expected * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive * of any partial chunks touched by the range. The invalid portion of * such chunks will be zero'd. * * (base + size) must be less then or equal to PAGE_SIZE. */ void vm_page_set_validclean(vm_page_t m, int base, int size) { vm_page_bits_t oldvalid, pagebits; int endoff, frag; + /* Object lock for VPO_NOSYNC */ VM_OBJECT_ASSERT_WLOCKED(m->object); + vm_page_assert_busied(m); if (size == 0) /* handle degenerate case */ return; /* * If the base is not DEV_BSIZE aligned and the valid * bit is clear, we have to zero out a portion of the * first block. */ if ((frag = rounddown2(base, DEV_BSIZE)) != base && (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) pmap_zero_page_area(m, frag, base - frag); /* * If the ending offset is not DEV_BSIZE aligned and the * valid bit is clear, we have to zero out a portion of * the last block. */ endoff = base + size; if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) pmap_zero_page_area(m, endoff, DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); /* * Set valid, clear dirty bits. If validating the entire * page we can safely clear the pmap modify bit. We also * use this opportunity to clear the VPO_NOSYNC flag. If a process * takes a write fault on a MAP_NOSYNC memory area the flag will * be set again. * * We set valid bits inclusive of any overlap, but we can only * clear dirty bits for DEV_BSIZE chunks that are fully within * the range. */ oldvalid = m->valid; pagebits = vm_page_bits(base, size); - m->valid |= pagebits; + if (vm_page_xbusied(m)) + m->valid |= pagebits; + else + vm_page_bits_set(m, &m->valid, pagebits); #if 0 /* NOT YET */ if ((frag = base & (DEV_BSIZE - 1)) != 0) { frag = DEV_BSIZE - frag; base += frag; size -= frag; if (size < 0) size = 0; } pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); #endif if (base == 0 && size == PAGE_SIZE) { /* * The page can only be modified within the pmap if it is * mapped, and it can only be mapped if it was previously * fully valid. */ if (oldvalid == VM_PAGE_BITS_ALL) /* * Perform the pmap_clear_modify() first. Otherwise, * a concurrent pmap operation, such as * pmap_protect(), could clear a modification in the * pmap and set the dirty field on the page before * pmap_clear_modify() had begun and after the dirty * field was cleared here. */ pmap_clear_modify(m); m->dirty = 0; m->oflags &= ~VPO_NOSYNC; - } else if (oldvalid != VM_PAGE_BITS_ALL) + } else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m)) m->dirty &= ~pagebits; else vm_page_clear_dirty_mask(m, pagebits); } void vm_page_clear_dirty(vm_page_t m, int base, int size) { vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); } /* * vm_page_set_invalid: * * Invalidates DEV_BSIZE'd chunks within a page. Both the * valid and dirty bits for the effected areas are cleared. */ void vm_page_set_invalid(vm_page_t m, int base, int size) { vm_page_bits_t bits; vm_object_t object; + /* + * The object lock is required so that pages can't be mapped + * read-only while we're in the process of invalidating them. + */ object = m->object; VM_OBJECT_ASSERT_WLOCKED(object); + vm_page_assert_busied(m); + if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + size >= object->un_pager.vnp.vnp_size) bits = VM_PAGE_BITS_ALL; else bits = vm_page_bits(base, size); - if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL && - bits != 0) + if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0) pmap_remove_all(m); - KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) || + KASSERT((bits == 0 && vm_page_all_valid(m)) || !pmap_page_is_mapped(m), ("vm_page_set_invalid: page %p is mapped", m)); - m->valid &= ~bits; - m->dirty &= ~bits; + if (vm_page_xbusied(m)) { + m->valid &= ~bits; + m->dirty &= ~bits; + } else { + vm_page_bits_clear(m, &m->valid, bits); + vm_page_bits_clear(m, &m->dirty, bits); + } } /* + * vm_page_invalid: + * + * Invalidates the entire page. The page must be busy, unmapped, and + * the enclosing object must be locked. The object locks protects + * against concurrent read-only pmap enter which is done without + * busy. + */ +void +vm_page_invalid(vm_page_t m) +{ + + vm_page_assert_busied(m); + VM_OBJECT_ASSERT_LOCKED(m->object); + MPASS(!pmap_page_is_mapped(m)); + + if (vm_page_xbusied(m)) + m->valid = 0; + else + vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL); +} + +/* * vm_page_zero_invalid() * * The kernel assumes that the invalid portions of a page contain * garbage, but such pages can be mapped into memory by user code. * When this occurs, we must zero out the non-valid portions of the * page so user code sees what it expects. * * Pages are most often semi-valid when the end of a file is mapped * into memory and the file's size is not page aligned. */ void vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) { int b; int i; - VM_OBJECT_ASSERT_WLOCKED(m->object); /* * Scan the valid bits looking for invalid sections that * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the * valid bit may be set ) have already been zeroed by * vm_page_set_validclean(). */ for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { if (i == (PAGE_SIZE / DEV_BSIZE) || (m->valid & ((vm_page_bits_t)1 << i))) { if (i > b) { pmap_zero_page_area(m, b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); } b = i + 1; } } /* * setvalid is TRUE when we can safely set the zero'd areas * as being valid. We can do this if there are no cache consistancy * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. */ if (setvalid) - m->valid = VM_PAGE_BITS_ALL; + vm_page_valid(m); } /* * vm_page_is_valid: * * Is (partial) page valid? Note that the case where size == 0 * will return FALSE in the degenerate case where the page is * entirely invalid, and TRUE otherwise. + * + * Some callers envoke this routine without the busy lock held and + * handle races via higher level locks. Typical callers should + * hold a busy lock to prevent invalidation. */ int vm_page_is_valid(vm_page_t m, int base, int size) { vm_page_bits_t bits; - VM_OBJECT_ASSERT_LOCKED(m->object); bits = vm_page_bits(base, size); return (m->valid != 0 && (m->valid & bits) == bits); } /* * Returns true if all of the specified predicates are true for the entire * (super)page and false otherwise. */ bool vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m) { vm_object_t object; int i, npages; object = m->object; if (skip_m != NULL && skip_m->object != object) return (false); VM_OBJECT_ASSERT_LOCKED(object); npages = atop(pagesizes[m->psind]); /* * The physically contiguous pages that make up a superpage, i.e., a * page with a page size index ("psind") greater than zero, will * occupy adjacent entries in vm_page_array[]. */ for (i = 0; i < npages; i++) { /* Always test object consistency, including "skip_m". */ if (m[i].object != object) return (false); if (&m[i] == skip_m) continue; if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i])) return (false); if ((flags & PS_ALL_DIRTY) != 0) { /* * Calling vm_page_test_dirty() or pmap_is_modified() * might stop this case from spuriously returning * "false". However, that would require a write lock * on the object containing "m[i]". */ if (m[i].dirty != VM_PAGE_BITS_ALL) return (false); } if ((flags & PS_ALL_VALID) != 0 && m[i].valid != VM_PAGE_BITS_ALL) return (false); } return (true); } /* * Set the page's dirty bits if the page is modified. */ void vm_page_test_dirty(vm_page_t m) { - VM_OBJECT_ASSERT_WLOCKED(m->object); + vm_page_assert_busied(m); if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) vm_page_dirty(m); +} + +void +vm_page_valid(vm_page_t m) +{ + + vm_page_assert_busied(m); + if (vm_page_xbusied(m)) + m->valid = VM_PAGE_BITS_ALL; + else + vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL); } void vm_page_lock_KBI(vm_page_t m, const char *file, int line) { mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); } void vm_page_unlock_KBI(vm_page_t m, const char *file, int line) { mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); } int vm_page_trylock_KBI(vm_page_t m, const char *file, int line) { return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); } #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) void vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) { vm_page_lock_assert_KBI(m, MA_OWNED, file, line); } void vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) { mtx_assert_(vm_page_lockptr(m), a, file, line); } #endif #ifdef INVARIANTS void vm_page_object_busy_assert(vm_page_t m) { /* * Certain of the page's fields may only be modified by the * holder of a page or object busy. */ if (m->object != NULL && !vm_page_busied(m)) VM_OBJECT_ASSERT_BUSY(m->object); } void vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits) { if ((bits & PGA_WRITEABLE) == 0) return; /* * The PGA_WRITEABLE flag can only be set if the page is * managed, is exclusively busied or the object is locked. * Currently, this flag is only set by pmap_enter(). */ KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("PGA_WRITEABLE on unmanaged page")); if (!vm_page_xbusied(m)) VM_OBJECT_ASSERT_BUSY(m->object); } #endif #include "opt_ddb.h" #ifdef DDB #include #include DB_SHOW_COMMAND(page, vm_page_print_page_info) { db_printf("vm_cnt.v_free_count: %d\n", vm_free_count()); db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count()); db_printf("vm_cnt.v_active_count: %d\n", vm_active_count()); db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count()); db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count()); db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); } DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) { int dom; db_printf("pq_free %d\n", vm_free_count()); for (dom = 0; dom < vm_ndomains; dom++) { db_printf( "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n", dom, vm_dom[dom].vmd_page_count, vm_dom[dom].vmd_free_count, vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt, vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt); } } DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) { vm_page_t m; boolean_t phys, virt; if (!have_addr) { db_printf("show pginfo addr\n"); return; } phys = strchr(modif, 'p') != NULL; virt = strchr(modif, 'v') != NULL; if (virt) m = PHYS_TO_VM_PAGE(pmap_kextract(addr)); else if (phys) m = PHYS_TO_VM_PAGE(addr); else m = (vm_page_t)addr; db_printf( "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref %u\n" " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, m->queue, m->ref_count, m->aflags, m->oflags, m->flags, m->act_count, m->busy_lock, m->valid, m->dirty); } #endif /* DDB */ Index: head/sys/vm/vm_page.h =================================================================== --- head/sys/vm/vm_page.h (revision 353538) +++ head/sys/vm/vm_page.h (revision 353539) @@ -1,932 +1,975 @@ /*- * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU) * * Copyright (c) 1991, 1993 * The Regents of the University of California. All rights reserved. * * This code is derived from software contributed to Berkeley by * The Mach Operating System project at Carnegie-Mellon University. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)vm_page.h 8.2 (Berkeley) 12/13/93 * * * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Authors: Avadis Tevanian, Jr., Michael Wayne Young * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. * * $FreeBSD$ */ /* * Resident memory system definitions. */ #ifndef _VM_PAGE_ #define _VM_PAGE_ #include /* * Management of resident (logical) pages. * * A small structure is kept for each resident * page, indexed by page number. Each structure * is an element of several collections: * * A radix tree used to quickly * perform object/offset lookups * * A list of all pages for a given object, * so they can be quickly deactivated at * time of deallocation. * * An ordered list of pages due for pageout. * * In addition, the structure contains the object * and offset to which this page belongs (for pageout), * and sundry status bits. * * In general, operations on this structure's mutable fields are - * synchronized using either one of or a combination of the lock on the - * object that the page belongs to (O), the page lock (P), - * the per-domain lock for the free queues (F), or the page's queue - * lock (Q). The physical address of a page is used to select its page - * lock from a pool. The queue lock for a page depends on the value of - * its queue field and described in detail below. If a field is - * annotated below with two of these locks, then holding either lock is - * sufficient for read access, but both locks are required for write - * access. An annotation of (C) indicates that the field is immutable. - * An annotation of (A) indicates that modifications to the field must - * be atomic. Accesses to such fields may require additional - * synchronization depending on the context. + * synchronized using either one of or a combination of locks. If a + * field is annotated with two of these locks then holding either is + * sufficient for read access but both are required for write access. + * The physical address of a page is used to select its page lock from + * a pool. The queue lock for a page depends on the value of its queue + * field and is described in detail below. * + * The following annotations are possible: + * (A) the field is atomic and may require additional synchronization. + * (B) the page busy lock. + * (C) the field is immutable. + * (F) the per-domain lock for the free queues + * (M) Machine dependent, defined by pmap layer. + * (O) the object that the page belongs to. + * (P) the page lock. + * (Q) the page's queue lock. + * + * The busy lock is an embedded reader-writer lock that protects the + * page's contents and identity (i.e., its tuple) as + * well as certain valid/dirty modifications. To avoid bloating the + * the page structure, the busy lock lacks some of the features available + * the kernel's general-purpose synchronization primitives. As a result, + * busy lock ordering rules are not verified, lock recursion is not + * detected, and an attempt to xbusy a busy page or sbusy an xbusy page + * results will trigger a panic rather than causing the thread to block. + * vm_page_sleep_if_busy() can be used to sleep until the page's busy + * state changes, after which the caller must re-lookup the page and + * re-evaluate its state. vm_page_busy_acquire() will block until + * the lock is acquired. + * + * The valid field is protected by the page busy lock (B) and object + * lock (O). Transitions from invalid to valid are generally done + * via I/O or zero filling and do not require the object lock. + * These must be protected with the busy lock to prevent page-in or + * creation races. Page invalidation generally happens as a result + * of truncate or msync. When invalidated, pages must not be present + * in pmap and must hold the object lock to prevent concurrent + * speculative read-only mappings that do not require busy. I/O + * routines may check for validity without a lock if they are prepared + * to handle invalidation races with higher level locks (vnode) or are + * unconcerned with races so long as they hold a reference to prevent + * recycling. When a valid bit is set while holding a shared busy + * lock (A) atomic operations are used to protect against concurrent + * modification. + * * In contrast, the synchronization of accesses to the page's - * dirty field is machine dependent (M). In the - * machine-independent layer, the lock on the object that the - * page belongs to must be held in order to operate on the field. - * However, the pmap layer is permitted to set all bits within - * the field without holding that lock. If the underlying - * architecture does not support atomic read-modify-write + * dirty field is a mix of machine dependent (M) and busy (B). In + * the machine-independent layer, the page busy must be held to + * operate on the field. However, the pmap layer is permitted to + * set all bits within the field without holding that lock. If the + * underlying architecture does not support atomic read-modify-write * operations on the field's type, then the machine-independent * layer uses a 32-bit atomic on the aligned 32-bit word that * contains the dirty field. In the machine-independent layer, * the implementation of read-modify-write operations on the - * field is encapsulated in vm_page_clear_dirty_mask(). + * field is encapsulated in vm_page_clear_dirty_mask(). An + * exclusive busy lock combined with pmap_remove_{write/all}() is the + * only way to ensure a page can not become dirty. I/O generally + * removes the page from pmap to ensure exclusive access and atomic + * writes. * * The ref_count field tracks references to the page. References that * prevent the page from being reclaimable are called wirings and are * counted in the low bits of ref_count. The containing object's * reference, if one exists, is counted using the VPRC_OBJREF bit in the * ref_count field. Additionally, the VPRC_BLOCKED bit is used to * atomically check for wirings and prevent new wirings via * pmap_extract_and_hold(). When a page belongs to an object, it may be * wired only when the object is locked, or the page is busy, or by * pmap_extract_and_hold(). As a result, if the object is locked and the * page is not busy (or is exclusively busied by the current thread), and * the page is unmapped, its wire count will not increase. The ref_count * field is updated using atomic operations in most cases, except when it * is known that no other references to the page exist, such as in the page * allocator. A page may be present in the page queues, or even actively * scanned by the page daemon, without an explicitly counted referenced. * The page daemon must therefore handle the possibility of a concurrent * free of the page. * - * The busy lock is an embedded reader-writer lock which protects the - * page's contents and identity (i.e., its tuple) and - * interlocks with the object lock (O). In particular, a page may be - * busied or unbusied only with the object write lock held. To avoid - * bloating the page structure, the busy lock lacks some of the - * features available to the kernel's general-purpose synchronization - * primitives. As a result, busy lock ordering rules are not verified, - * lock recursion is not detected, and an attempt to xbusy a busy page - * or sbusy an xbusy page results will trigger a panic rather than - * causing the thread to block. vm_page_sleep_if_busy() can be used to - * sleep until the page's busy state changes, after which the caller - * must re-lookup the page and re-evaluate its state. - * * The queue field is the index of the page queue containing the page, * or PQ_NONE if the page is not enqueued. The queue lock of a page is * the page queue lock corresponding to the page queue index, or the * page lock (P) for the page if it is not enqueued. To modify the * queue field, the queue lock for the old value of the field must be * held. There is one exception to this rule: the page daemon may * transition the queue field from PQ_INACTIVE to PQ_NONE immediately * prior to freeing a page during an inactive queue scan. At that * point the page has already been physically dequeued and no other * references to that vm_page structure exist. * * To avoid contention on page queue locks, page queue operations * (enqueue, dequeue, requeue) are batched using per-CPU queues. A * deferred operation is requested by inserting an entry into a batch * queue; the entry is simply a pointer to the page, and the request * type is encoded in the page's aflags field using the values in * PGA_QUEUE_STATE_MASK. The type-stability of struct vm_pages is * crucial to this scheme since the processing of entries in a given * batch queue may be deferred indefinitely. In particular, a page may * be freed before its pending batch queue entries have been processed. * The page lock (P) must be held to schedule a batched queue * operation, and the page queue lock must be held in order to process * batch queue entries for the page queue. There is one exception to * this rule: the thread freeing a page may schedule a dequeue without * holding the page lock. In this scenario the only other thread which * may hold a reference to the page is the page daemon, which is * careful to avoid modifying the page's queue state once the dequeue * has been requested by setting PGA_DEQUEUE. */ #if PAGE_SIZE == 4096 #define VM_PAGE_BITS_ALL 0xffu typedef uint8_t vm_page_bits_t; #elif PAGE_SIZE == 8192 #define VM_PAGE_BITS_ALL 0xffffu typedef uint16_t vm_page_bits_t; #elif PAGE_SIZE == 16384 #define VM_PAGE_BITS_ALL 0xffffffffu typedef uint32_t vm_page_bits_t; #elif PAGE_SIZE == 32768 #define VM_PAGE_BITS_ALL 0xfffffffffffffffflu typedef uint64_t vm_page_bits_t; #endif struct vm_page { union { TAILQ_ENTRY(vm_page) q; /* page queue or free list (Q) */ struct { SLIST_ENTRY(vm_page) ss; /* private slists */ void *pv; } s; struct { u_long p; u_long v; } memguard; } plinks; TAILQ_ENTRY(vm_page) listq; /* pages in same object (O) */ vm_object_t object; /* which object am I in (O) */ vm_pindex_t pindex; /* offset into object (O,P) */ vm_paddr_t phys_addr; /* physical address of page (C) */ struct md_page md; /* machine dependent stuff */ u_int ref_count; /* page references (A) */ volatile u_int busy_lock; /* busy owners lock */ uint16_t flags; /* page PG_* flags (P) */ uint8_t order; /* index of the buddy queue (F) */ uint8_t pool; /* vm_phys freepool index (F) */ - uint8_t aflags; /* access is atomic */ + uint8_t aflags; /* atomic flags (A) */ uint8_t oflags; /* page VPO_* flags (O) */ uint8_t queue; /* page queue index (Q) */ int8_t psind; /* pagesizes[] index (O) */ int8_t segind; /* vm_phys segment index (C) */ u_char act_count; /* page usage count (P) */ /* NOTE that these must support one bit per DEV_BSIZE in a page */ /* so, on normal X86 kernels, they must be at least 8 bits wide */ - vm_page_bits_t valid; /* map of valid DEV_BSIZE chunks (O) */ - vm_page_bits_t dirty; /* map of dirty DEV_BSIZE chunks (M) */ + vm_page_bits_t valid; /* valid DEV_BSIZE chunk map (O,B) */ + vm_page_bits_t dirty; /* dirty DEV_BSIZE chunk map (M,B) */ }; /* * Special bits used in the ref_count field. * * ref_count is normally used to count wirings that prevent the page from being * reclaimed, but also supports several special types of references that do not * prevent reclamation. Accesses to the ref_count field must be atomic unless * the page is unallocated. * * VPRC_OBJREF is the reference held by the containing object. It can set or * cleared only when the corresponding object's write lock is held. * * VPRC_BLOCKED is used to atomically block wirings via pmap lookups while * attempting to tear down all mappings of a given page. The page lock and * object write lock must both be held in order to set or clear this bit. */ #define VPRC_BLOCKED 0x40000000u /* mappings are being removed */ #define VPRC_OBJREF 0x80000000u /* object reference, cleared with (O) */ #define VPRC_WIRE_COUNT(c) ((c) & ~(VPRC_BLOCKED | VPRC_OBJREF)) #define VPRC_WIRE_COUNT_MAX (~(VPRC_BLOCKED | VPRC_OBJREF)) /* * Page flags stored in oflags: * * Access to these page flags is synchronized by the lock on the object * containing the page (O). * * Note: VPO_UNMANAGED (used by OBJT_DEVICE, OBJT_PHYS and OBJT_SG) * indicates that the page is not under PV management but * otherwise should be treated as a normal page. Pages not * under PV management cannot be paged out via the * object/vm_page_t because there is no knowledge of their pte * mappings, and such pages are also not on any PQ queue. * */ #define VPO_KMEM_EXEC 0x01 /* kmem mapping allows execution */ #define VPO_SWAPSLEEP 0x02 /* waiting for swap to finish */ #define VPO_UNMANAGED 0x04 /* no PV management for page */ #define VPO_SWAPINPROG 0x08 /* swap I/O in progress on page */ #define VPO_NOSYNC 0x10 /* do not collect for syncer */ /* * Busy page implementation details. * The algorithm is taken mostly by rwlock(9) and sx(9) locks implementation, * even if the support for owner identity is removed because of size * constraints. Checks on lock recursion are then not possible, while the * lock assertions effectiveness is someway reduced. */ #define VPB_BIT_SHARED 0x01 #define VPB_BIT_EXCLUSIVE 0x02 #define VPB_BIT_WAITERS 0x04 #define VPB_BIT_FLAGMASK \ (VPB_BIT_SHARED | VPB_BIT_EXCLUSIVE | VPB_BIT_WAITERS) #define VPB_SHARERS_SHIFT 3 #define VPB_SHARERS(x) \ (((x) & ~VPB_BIT_FLAGMASK) >> VPB_SHARERS_SHIFT) #define VPB_SHARERS_WORD(x) ((x) << VPB_SHARERS_SHIFT | VPB_BIT_SHARED) #define VPB_ONE_SHARER (1 << VPB_SHARERS_SHIFT) #define VPB_SINGLE_EXCLUSIVER VPB_BIT_EXCLUSIVE #define VPB_UNBUSIED VPB_SHARERS_WORD(0) #define PQ_NONE 255 #define PQ_INACTIVE 0 #define PQ_ACTIVE 1 #define PQ_LAUNDRY 2 #define PQ_UNSWAPPABLE 3 #define PQ_COUNT 4 #ifndef VM_PAGE_HAVE_PGLIST TAILQ_HEAD(pglist, vm_page); #define VM_PAGE_HAVE_PGLIST #endif SLIST_HEAD(spglist, vm_page); #ifdef _KERNEL extern vm_page_t bogus_page; #endif /* _KERNEL */ extern struct mtx_padalign pa_lock[]; #if defined(__arm__) #define PDRSHIFT PDR_SHIFT #elif !defined(PDRSHIFT) #define PDRSHIFT 21 #endif #define pa_index(pa) ((pa) >> PDRSHIFT) #define PA_LOCKPTR(pa) ((struct mtx *)(&pa_lock[pa_index(pa) % PA_LOCK_COUNT])) #define PA_LOCKOBJPTR(pa) ((struct lock_object *)PA_LOCKPTR((pa))) #define PA_LOCK(pa) mtx_lock(PA_LOCKPTR(pa)) #define PA_TRYLOCK(pa) mtx_trylock(PA_LOCKPTR(pa)) #define PA_UNLOCK(pa) mtx_unlock(PA_LOCKPTR(pa)) #define PA_UNLOCK_COND(pa) \ do { \ if ((pa) != 0) { \ PA_UNLOCK((pa)); \ (pa) = 0; \ } \ } while (0) #define PA_LOCK_ASSERT(pa, a) mtx_assert(PA_LOCKPTR(pa), (a)) #if defined(KLD_MODULE) && !defined(KLD_TIED) #define vm_page_lock(m) vm_page_lock_KBI((m), LOCK_FILE, LOCK_LINE) #define vm_page_unlock(m) vm_page_unlock_KBI((m), LOCK_FILE, LOCK_LINE) #define vm_page_trylock(m) vm_page_trylock_KBI((m), LOCK_FILE, LOCK_LINE) #else /* !KLD_MODULE */ #define vm_page_lockptr(m) (PA_LOCKPTR(VM_PAGE_TO_PHYS((m)))) #define vm_page_lock(m) mtx_lock(vm_page_lockptr((m))) #define vm_page_unlock(m) mtx_unlock(vm_page_lockptr((m))) #define vm_page_trylock(m) mtx_trylock(vm_page_lockptr((m))) #endif #if defined(INVARIANTS) #define vm_page_assert_locked(m) \ vm_page_assert_locked_KBI((m), __FILE__, __LINE__) #define vm_page_lock_assert(m, a) \ vm_page_lock_assert_KBI((m), (a), __FILE__, __LINE__) #else #define vm_page_assert_locked(m) #define vm_page_lock_assert(m, a) #endif /* * The vm_page's aflags are updated using atomic operations. To set or clear * these flags, the functions vm_page_aflag_set() and vm_page_aflag_clear() * must be used. Neither these flags nor these functions are part of the KBI. * * PGA_REFERENCED may be cleared only if the page is locked. It is set by * both the MI and MD VM layers. However, kernel loadable modules should not * directly set this flag. They should call vm_page_reference() instead. * * PGA_WRITEABLE is set exclusively on managed pages by pmap_enter(). * When it does so, the object must be locked, or the page must be * exclusive busied. The MI VM layer must never access this flag * directly. Instead, it should call pmap_page_is_write_mapped(). * * PGA_EXECUTABLE may be set by pmap routines, and indicates that a page has * at least one executable mapping. It is not consumed by the MI VM layer. * * PGA_ENQUEUED is set and cleared when a page is inserted into or removed * from a page queue, respectively. It determines whether the plinks.q field * of the page is valid. To set or clear this flag, the queue lock for the * page must be held: the page queue lock corresponding to the page's "queue" * field if its value is not PQ_NONE, and the page lock otherwise. * * PGA_DEQUEUE is set when the page is scheduled to be dequeued from a page * queue, and cleared when the dequeue request is processed. A page may * have PGA_DEQUEUE set and PGA_ENQUEUED cleared, for instance if a dequeue * is requested after the page is scheduled to be enqueued but before it is * actually inserted into the page queue. For allocated pages, the page lock * must be held to set this flag, but it may be set by vm_page_free_prep() * without the page lock held. The page queue lock must be held to clear the * PGA_DEQUEUE flag. * * PGA_REQUEUE is set when the page is scheduled to be enqueued or requeued * in its page queue. The page lock must be held to set this flag, and the * queue lock for the page must be held to clear it. * * PGA_REQUEUE_HEAD is a special flag for enqueuing pages near the head of * the inactive queue, thus bypassing LRU. The page lock must be held to * set this flag, and the queue lock for the page must be held to clear it. */ #define PGA_WRITEABLE 0x01 /* page may be mapped writeable */ #define PGA_REFERENCED 0x02 /* page has been referenced */ #define PGA_EXECUTABLE 0x04 /* page may be mapped executable */ #define PGA_ENQUEUED 0x08 /* page is enqueued in a page queue */ #define PGA_DEQUEUE 0x10 /* page is due to be dequeued */ #define PGA_REQUEUE 0x20 /* page is due to be requeued */ #define PGA_REQUEUE_HEAD 0x40 /* page requeue should bypass LRU */ #define PGA_QUEUE_STATE_MASK (PGA_ENQUEUED | PGA_DEQUEUE | PGA_REQUEUE | \ PGA_REQUEUE_HEAD) /* * Page flags. If changed at any other time than page allocation or * freeing, the modification must be protected by the vm_page lock. * * The PG_PCPU_CACHE flag is set at allocation time if the page was * allocated from a per-CPU cache. It is cleared the next time that the * page is allocated from the physical memory allocator. */ #define PG_PCPU_CACHE 0x0001 /* was allocated from per-CPU caches */ #define PG_FICTITIOUS 0x0004 /* physical page doesn't exist */ #define PG_ZERO 0x0008 /* page is zeroed */ #define PG_MARKER 0x0010 /* special queue marker page */ #define PG_NODUMP 0x0080 /* don't include this page in a dump */ /* * Misc constants. */ #define ACT_DECLINE 1 #define ACT_ADVANCE 3 #define ACT_INIT 5 #define ACT_MAX 64 #ifdef _KERNEL #include #include /* * Each pageable resident page falls into one of five lists: * * free * Available for allocation now. * * inactive * Low activity, candidates for reclamation. * This list is approximately LRU ordered. * * laundry * This is the list of pages that should be * paged out next. * * unswappable * Dirty anonymous pages that cannot be paged * out because no swap device is configured. * * active * Pages that are "active", i.e., they have been * recently referenced. * */ extern vm_page_t vm_page_array; /* First resident page in table */ extern long vm_page_array_size; /* number of vm_page_t's */ extern long first_page; /* first physical page number */ #define VM_PAGE_TO_PHYS(entry) ((entry)->phys_addr) /* * PHYS_TO_VM_PAGE() returns the vm_page_t object that represents a memory * page to which the given physical address belongs. The correct vm_page_t * object is returned for addresses that are not page-aligned. */ vm_page_t PHYS_TO_VM_PAGE(vm_paddr_t pa); /* * Page allocation parameters for vm_page for the functions * vm_page_alloc(), vm_page_grab(), vm_page_alloc_contig() and * vm_page_alloc_freelist(). Some functions support only a subset * of the flags, and ignore others, see the flags legend. * * The meaning of VM_ALLOC_ZERO differs slightly between the vm_page_alloc*() * and the vm_page_grab*() functions. See these functions for details. * * Bits 0 - 1 define class. * Bits 2 - 15 dedicated for flags. * Legend: * (a) - vm_page_alloc() supports the flag. * (c) - vm_page_alloc_contig() supports the flag. * (f) - vm_page_alloc_freelist() supports the flag. * (g) - vm_page_grab() supports the flag. * (p) - vm_page_grab_pages() supports the flag. * Bits above 15 define the count of additional pages that the caller * intends to allocate. */ #define VM_ALLOC_NORMAL 0 #define VM_ALLOC_INTERRUPT 1 #define VM_ALLOC_SYSTEM 2 #define VM_ALLOC_CLASS_MASK 3 #define VM_ALLOC_WAITOK 0x0008 /* (acf) Sleep and retry */ #define VM_ALLOC_WAITFAIL 0x0010 /* (acf) Sleep and return error */ #define VM_ALLOC_WIRED 0x0020 /* (acfgp) Allocate a wired page */ #define VM_ALLOC_ZERO 0x0040 /* (acfgp) Allocate a prezeroed page */ #define VM_ALLOC_NOOBJ 0x0100 /* (acg) No associated object */ #define VM_ALLOC_NOBUSY 0x0200 /* (acgp) Do not excl busy the page */ #define VM_ALLOC_NOCREAT 0x0400 /* (gp) Don't create a page */ #define VM_ALLOC_IGN_SBUSY 0x1000 /* (gp) Ignore shared busy flag */ #define VM_ALLOC_NODUMP 0x2000 /* (ag) don't include in dump */ #define VM_ALLOC_SBUSY 0x4000 /* (acgp) Shared busy the page */ #define VM_ALLOC_NOWAIT 0x8000 /* (acfgp) Do not sleep */ #define VM_ALLOC_COUNT_SHIFT 16 #define VM_ALLOC_COUNT(count) ((count) << VM_ALLOC_COUNT_SHIFT) #ifdef M_NOWAIT static inline int malloc2vm_flags(int malloc_flags) { int pflags; KASSERT((malloc_flags & M_USE_RESERVE) == 0 || (malloc_flags & M_NOWAIT) != 0, ("M_USE_RESERVE requires M_NOWAIT")); pflags = (malloc_flags & M_USE_RESERVE) != 0 ? VM_ALLOC_INTERRUPT : VM_ALLOC_SYSTEM; if ((malloc_flags & M_ZERO) != 0) pflags |= VM_ALLOC_ZERO; if ((malloc_flags & M_NODUMP) != 0) pflags |= VM_ALLOC_NODUMP; if ((malloc_flags & M_NOWAIT)) pflags |= VM_ALLOC_NOWAIT; if ((malloc_flags & M_WAITOK)) pflags |= VM_ALLOC_WAITOK; return (pflags); } #endif /* * Predicates supported by vm_page_ps_test(): * * PS_ALL_DIRTY is true only if the entire (super)page is dirty. * However, it can be spuriously false when the (super)page has become * dirty in the pmap but that information has not been propagated to the * machine-independent layer. */ #define PS_ALL_DIRTY 0x1 #define PS_ALL_VALID 0x2 #define PS_NONE_BUSY 0x4 int vm_page_busy_acquire(vm_page_t m, int allocflags); void vm_page_busy_downgrade(vm_page_t m); int vm_page_busy_tryupgrade(vm_page_t m); void vm_page_busy_sleep(vm_page_t m, const char *msg, bool nonshared); void vm_page_free(vm_page_t m); void vm_page_free_zero(vm_page_t m); void vm_page_activate (vm_page_t); void vm_page_advise(vm_page_t m, int advice); vm_page_t vm_page_alloc(vm_object_t, vm_pindex_t, int); vm_page_t vm_page_alloc_domain(vm_object_t, vm_pindex_t, int, int); vm_page_t vm_page_alloc_after(vm_object_t, vm_pindex_t, int, vm_page_t); vm_page_t vm_page_alloc_domain_after(vm_object_t, vm_pindex_t, int, int, vm_page_t); vm_page_t vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr); vm_page_t vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr); vm_page_t vm_page_alloc_freelist(int, int); vm_page_t vm_page_alloc_freelist_domain(int, int, int); bool vm_page_blacklist_add(vm_paddr_t pa, bool verbose); void vm_page_change_lock(vm_page_t m, struct mtx **mtx); vm_page_t vm_page_grab (vm_object_t, vm_pindex_t, int); int vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, vm_page_t *ma, int count); int vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags); void vm_page_deactivate(vm_page_t); void vm_page_deactivate_noreuse(vm_page_t); void vm_page_dequeue(vm_page_t m); void vm_page_dequeue_deferred(vm_page_t m); vm_page_t vm_page_find_least(vm_object_t, vm_pindex_t); bool vm_page_free_prep(vm_page_t m); vm_page_t vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr); void vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr); int vm_page_insert (vm_page_t, vm_object_t, vm_pindex_t); +void vm_page_invalid(vm_page_t m); void vm_page_launder(vm_page_t m); vm_page_t vm_page_lookup (vm_object_t, vm_pindex_t); vm_page_t vm_page_next(vm_page_t m); int vm_page_pa_tryrelock(pmap_t, vm_paddr_t, vm_paddr_t *); void vm_page_pqbatch_drain(void); void vm_page_pqbatch_submit(vm_page_t m, uint8_t queue); vm_page_t vm_page_prev(vm_page_t m); bool vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m); void vm_page_putfake(vm_page_t m); void vm_page_readahead_finish(vm_page_t m); bool vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary); bool vm_page_reclaim_contig_domain(int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary); void vm_page_reference(vm_page_t m); #define VPR_TRYFREE 0x01 #define VPR_NOREUSE 0x02 void vm_page_release(vm_page_t m, int flags); void vm_page_release_locked(vm_page_t m, int flags); bool vm_page_remove(vm_page_t); int vm_page_rename(vm_page_t, vm_object_t, vm_pindex_t); vm_page_t vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex); void vm_page_requeue(vm_page_t m); int vm_page_sbusied(vm_page_t m); vm_page_t vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, u_long alignment, vm_paddr_t boundary, int options); void vm_page_set_valid_range(vm_page_t m, int base, int size); int vm_page_sleep_if_busy(vm_page_t m, const char *msg); int vm_page_sleep_if_xbusy(vm_page_t m, const char *msg); vm_offset_t vm_page_startup(vm_offset_t vaddr); void vm_page_sunbusy(vm_page_t m); void vm_page_swapqueue(vm_page_t m, uint8_t oldq, uint8_t newq); bool vm_page_try_remove_all(vm_page_t m); bool vm_page_try_remove_write(vm_page_t m); int vm_page_trysbusy(vm_page_t m); int vm_page_tryxbusy(vm_page_t m); void vm_page_unhold_pages(vm_page_t *ma, int count); void vm_page_unswappable(vm_page_t m); void vm_page_unwire(vm_page_t m, uint8_t queue); bool vm_page_unwire_noq(vm_page_t m); void vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr); void vm_page_wire(vm_page_t); bool vm_page_wire_mapped(vm_page_t m); void vm_page_xunbusy_hard(vm_page_t m); void vm_page_set_validclean (vm_page_t, int, int); -void vm_page_clear_dirty (vm_page_t, int, int); -void vm_page_set_invalid (vm_page_t, int, int); -int vm_page_is_valid (vm_page_t, int, int); -void vm_page_test_dirty (vm_page_t); +void vm_page_clear_dirty(vm_page_t, int, int); +void vm_page_set_invalid(vm_page_t, int, int); +void vm_page_valid(vm_page_t m); +int vm_page_is_valid(vm_page_t, int, int); +void vm_page_test_dirty(vm_page_t); vm_page_bits_t vm_page_bits(int base, int size); void vm_page_zero_invalid(vm_page_t m, boolean_t setvalid); void vm_page_free_toq(vm_page_t m); void vm_page_free_pages_toq(struct spglist *free, bool update_wire_count); void vm_page_dirty_KBI(vm_page_t m); void vm_page_lock_KBI(vm_page_t m, const char *file, int line); void vm_page_unlock_KBI(vm_page_t m, const char *file, int line); int vm_page_trylock_KBI(vm_page_t m, const char *file, int line); #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) void vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line); void vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line); #endif +#define vm_page_assert_busied(m) \ + KASSERT(vm_page_busied(m), \ + ("vm_page_assert_busied: page %p not busy @ %s:%d", \ + (m), __FILE__, __LINE__)) + #define vm_page_assert_sbusied(m) \ KASSERT(vm_page_sbusied(m), \ ("vm_page_assert_sbusied: page %p not shared busy @ %s:%d", \ (m), __FILE__, __LINE__)) #define vm_page_assert_unbusied(m) \ KASSERT(!vm_page_busied(m), \ ("vm_page_assert_unbusied: page %p busy @ %s:%d", \ (m), __FILE__, __LINE__)) #define vm_page_assert_xbusied(m) \ KASSERT(vm_page_xbusied(m), \ ("vm_page_assert_xbusied: page %p not exclusive busy @ %s:%d", \ (m), __FILE__, __LINE__)) #define vm_page_busied(m) \ ((m)->busy_lock != VPB_UNBUSIED) #define vm_page_sbusy(m) do { \ if (!vm_page_trysbusy(m)) \ panic("%s: page %p failed shared busying", __func__, \ (m)); \ } while (0) #define vm_page_xbusied(m) \ (((m)->busy_lock & VPB_SINGLE_EXCLUSIVER) != 0) #define vm_page_xbusy(m) do { \ if (!vm_page_tryxbusy(m)) \ panic("%s: page %p failed exclusive busying", __func__, \ (m)); \ } while (0) /* Note: page m's lock must not be owned by the caller. */ #define vm_page_xunbusy(m) do { \ if (!atomic_cmpset_rel_int(&(m)->busy_lock, \ VPB_SINGLE_EXCLUSIVER, VPB_UNBUSIED)) \ vm_page_xunbusy_hard(m); \ } while (0) #ifdef INVARIANTS void vm_page_object_busy_assert(vm_page_t m); #define VM_PAGE_OBJECT_BUSY_ASSERT(m) vm_page_object_busy_assert(m) void vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits); #define VM_PAGE_ASSERT_PGA_WRITEABLE(m, bits) \ vm_page_assert_pga_writeable(m, bits) #else #define VM_PAGE_OBJECT_BUSY_ASSERT(m) (void)0 #define VM_PAGE_ASSERT_PGA_WRITEABLE(m, bits) (void)0 #endif /* * We want to use atomic updates for the aflags field, which is 8 bits wide. * However, not all architectures support atomic operations on 8-bit * destinations. In order that we can easily use a 32-bit operation, we * require that the aflags field be 32-bit aligned. */ _Static_assert(offsetof(struct vm_page, aflags) % sizeof(uint32_t) == 0, "aflags field is not 32-bit aligned"); /* * We want to be able to update the aflags and queue fields atomically in * the same operation. */ _Static_assert(offsetof(struct vm_page, aflags) / sizeof(uint32_t) == offsetof(struct vm_page, queue) / sizeof(uint32_t), "aflags and queue fields do not belong to the same 32-bit word"); _Static_assert(offsetof(struct vm_page, queue) % sizeof(uint32_t) == 2, "queue field is at an unexpected offset"); _Static_assert(sizeof(((struct vm_page *)NULL)->queue) == 1, "queue field has an unexpected size"); #if BYTE_ORDER == LITTLE_ENDIAN #define VM_PAGE_AFLAG_SHIFT 0 #define VM_PAGE_QUEUE_SHIFT 16 #else #define VM_PAGE_AFLAG_SHIFT 24 #define VM_PAGE_QUEUE_SHIFT 8 #endif #define VM_PAGE_QUEUE_MASK (0xff << VM_PAGE_QUEUE_SHIFT) /* * Clear the given bits in the specified page. */ static inline void vm_page_aflag_clear(vm_page_t m, uint8_t bits) { uint32_t *addr, val; /* * The PGA_REFERENCED flag can only be cleared if the page is locked. */ if ((bits & PGA_REFERENCED) != 0) vm_page_assert_locked(m); /* * Access the whole 32-bit word containing the aflags field with an * atomic update. Parallel non-atomic updates to the other fields * within this word are handled properly by the atomic update. */ addr = (void *)&m->aflags; val = bits << VM_PAGE_AFLAG_SHIFT; atomic_clear_32(addr, val); } /* * Set the given bits in the specified page. */ static inline void vm_page_aflag_set(vm_page_t m, uint8_t bits) { uint32_t *addr, val; VM_PAGE_ASSERT_PGA_WRITEABLE(m, bits); /* * Access the whole 32-bit word containing the aflags field with an * atomic update. Parallel non-atomic updates to the other fields * within this word are handled properly by the atomic update. */ addr = (void *)&m->aflags; val = bits << VM_PAGE_AFLAG_SHIFT; atomic_set_32(addr, val); } /* * Atomically update the queue state of the page. The operation fails if * any of the queue flags in "fflags" are set or if the "queue" field of * the page does not match the expected value; if the operation is * successful, the flags in "nflags" are set and all other queue state * flags are cleared. */ static inline bool vm_page_pqstate_cmpset(vm_page_t m, uint32_t oldq, uint32_t newq, uint32_t fflags, uint32_t nflags) { uint32_t *addr, nval, oval, qsmask; fflags <<= VM_PAGE_AFLAG_SHIFT; nflags <<= VM_PAGE_AFLAG_SHIFT; newq <<= VM_PAGE_QUEUE_SHIFT; oldq <<= VM_PAGE_QUEUE_SHIFT; qsmask = ((PGA_DEQUEUE | PGA_REQUEUE | PGA_REQUEUE_HEAD) << VM_PAGE_AFLAG_SHIFT) | VM_PAGE_QUEUE_MASK; addr = (void *)&m->aflags; oval = atomic_load_32(addr); do { if ((oval & fflags) != 0) return (false); if ((oval & VM_PAGE_QUEUE_MASK) != oldq) return (false); nval = (oval & ~qsmask) | nflags | newq; } while (!atomic_fcmpset_32(addr, &oval, nval)); return (true); } /* * vm_page_dirty: * * Set all bits in the page's dirty field. * * The object containing the specified page must be locked if the * call is made from the machine-independent layer. * * See vm_page_clear_dirty_mask(). */ static __inline void vm_page_dirty(vm_page_t m) { /* Use vm_page_dirty_KBI() under INVARIANTS to save memory. */ #if (defined(KLD_MODULE) && !defined(KLD_TIED)) || defined(INVARIANTS) vm_page_dirty_KBI(m); #else m->dirty = VM_PAGE_BITS_ALL; #endif } /* * vm_page_undirty: * * Set page to not be dirty. Note: does not clear pmap modify bits */ static __inline void vm_page_undirty(vm_page_t m) { VM_PAGE_OBJECT_BUSY_ASSERT(m); m->dirty = 0; } static inline void vm_page_replace_checked(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, vm_page_t mold) { vm_page_t mret; mret = vm_page_replace(mnew, object, pindex); KASSERT(mret == mold, ("invalid page replacement, mold=%p, mret=%p", mold, mret)); /* Unused if !INVARIANTS. */ (void)mold; (void)mret; } /* * vm_page_queue: * * Return the index of the queue containing m. This index is guaranteed * not to change while the page lock is held. */ static inline uint8_t vm_page_queue(vm_page_t m) { vm_page_assert_locked(m); if ((m->aflags & PGA_DEQUEUE) != 0) return (PQ_NONE); atomic_thread_fence_acq(); return (m->queue); } static inline bool vm_page_active(vm_page_t m) { return (vm_page_queue(m) == PQ_ACTIVE); } static inline bool vm_page_inactive(vm_page_t m) { return (vm_page_queue(m) == PQ_INACTIVE); } static inline bool vm_page_in_laundry(vm_page_t m) { uint8_t queue; queue = vm_page_queue(m); return (queue == PQ_LAUNDRY || queue == PQ_UNSWAPPABLE); } /* * vm_page_drop: * * Release a reference to a page and return the old reference count. */ static inline u_int vm_page_drop(vm_page_t m, u_int val) { u_int old; /* * Synchronize with vm_page_free_prep(): ensure that all updates to the * page structure are visible before it is freed. */ atomic_thread_fence_rel(); old = atomic_fetchadd_int(&m->ref_count, -val); KASSERT(old != VPRC_BLOCKED, ("vm_page_drop: page %p has an invalid refcount value", m)); return (old); } /* * vm_page_wired: * * Perform a racy check to determine whether a reference prevents the page * from being reclaimable. If the page's object is locked, and the page is * unmapped and unbusied or exclusively busied by the current thread, no * new wirings may be created. */ static inline bool vm_page_wired(vm_page_t m) { return (VPRC_WIRE_COUNT(m->ref_count) > 0); +} + +static inline bool +vm_page_all_valid(vm_page_t m) +{ + + return (m->valid == VM_PAGE_BITS_ALL); +} + +static inline bool +vm_page_none_valid(vm_page_t m) +{ + + return (m->valid == 0); } #endif /* _KERNEL */ #endif /* !_VM_PAGE_ */ Index: head/sys/vm/vm_pageout.c =================================================================== --- head/sys/vm/vm_pageout.c (revision 353538) +++ head/sys/vm/vm_pageout.c (revision 353539) @@ -1,2245 +1,2245 @@ /*- * SPDX-License-Identifier: (BSD-4-Clause AND MIT-CMU) * * Copyright (c) 1991 Regents of the University of California. * All rights reserved. * Copyright (c) 1994 John S. Dyson * All rights reserved. * Copyright (c) 1994 David Greenman * All rights reserved. * Copyright (c) 2005 Yahoo! Technologies Norway AS * All rights reserved. * * This code is derived from software contributed to Berkeley by * The Mach Operating System project at Carnegie-Mellon University. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. All advertising materials mentioning features or use of this software * must display the following acknowledgement: * This product includes software developed by the University of * California, Berkeley and its contributors. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91 * * * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Authors: Avadis Tevanian, Jr., Michael Wayne Young * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. */ /* * The proverbial page-out daemon. */ #include __FBSDID("$FreeBSD$"); #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* * System initialization */ /* the kernel process "vm_pageout"*/ static void vm_pageout(void); static void vm_pageout_init(void); static int vm_pageout_clean(vm_page_t m, int *numpagedout); static int vm_pageout_cluster(vm_page_t m); static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage, int starting_page_shortage); SYSINIT(pagedaemon_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, vm_pageout_init, NULL); struct proc *pageproc; static struct kproc_desc page_kp = { "pagedaemon", vm_pageout, &pageproc }; SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start, &page_kp); SDT_PROVIDER_DEFINE(vm); SDT_PROBE_DEFINE(vm, , , vm__lowmem_scan); /* Pagedaemon activity rates, in subdivisions of one second. */ #define VM_LAUNDER_RATE 10 #define VM_INACT_SCAN_RATE 10 static int vm_pageout_oom_seq = 12; static int vm_pageout_update_period; static int disable_swap_pageouts; static int lowmem_period = 10; static int swapdev_enabled; static int vm_panic_on_oom = 0; SYSCTL_INT(_vm, OID_AUTO, panic_on_oom, CTLFLAG_RWTUN, &vm_panic_on_oom, 0, "panic on out of memory instead of killing the largest process"); SYSCTL_INT(_vm, OID_AUTO, pageout_update_period, CTLFLAG_RWTUN, &vm_pageout_update_period, 0, "Maximum active LRU update period"); SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RWTUN, &lowmem_period, 0, "Low memory callback period"); SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, CTLFLAG_RWTUN, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); static int pageout_lock_miss; SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); SYSCTL_INT(_vm, OID_AUTO, pageout_oom_seq, CTLFLAG_RWTUN, &vm_pageout_oom_seq, 0, "back-to-back calls to oom detector to start OOM"); static int act_scan_laundry_weight = 3; SYSCTL_INT(_vm, OID_AUTO, act_scan_laundry_weight, CTLFLAG_RWTUN, &act_scan_laundry_weight, 0, "weight given to clean vs. dirty pages in active queue scans"); static u_int vm_background_launder_rate = 4096; SYSCTL_UINT(_vm, OID_AUTO, background_launder_rate, CTLFLAG_RWTUN, &vm_background_launder_rate, 0, "background laundering rate, in kilobytes per second"); static u_int vm_background_launder_max = 20 * 1024; SYSCTL_UINT(_vm, OID_AUTO, background_launder_max, CTLFLAG_RWTUN, &vm_background_launder_max, 0, "background laundering cap, in kilobytes"); int vm_pageout_page_count = 32; u_long vm_page_max_user_wired; SYSCTL_ULONG(_vm, OID_AUTO, max_user_wired, CTLFLAG_RW, &vm_page_max_user_wired, 0, "system-wide limit to user-wired page count"); static u_int isqrt(u_int num); static int vm_pageout_launder(struct vm_domain *vmd, int launder, bool in_shortfall); static void vm_pageout_laundry_worker(void *arg); struct scan_state { struct vm_batchqueue bq; struct vm_pagequeue *pq; vm_page_t marker; int maxscan; int scanned; }; static void vm_pageout_init_scan(struct scan_state *ss, struct vm_pagequeue *pq, vm_page_t marker, vm_page_t after, int maxscan) { vm_pagequeue_assert_locked(pq); KASSERT((marker->aflags & PGA_ENQUEUED) == 0, ("marker %p already enqueued", marker)); if (after == NULL) TAILQ_INSERT_HEAD(&pq->pq_pl, marker, plinks.q); else TAILQ_INSERT_AFTER(&pq->pq_pl, after, marker, plinks.q); vm_page_aflag_set(marker, PGA_ENQUEUED); vm_batchqueue_init(&ss->bq); ss->pq = pq; ss->marker = marker; ss->maxscan = maxscan; ss->scanned = 0; vm_pagequeue_unlock(pq); } static void vm_pageout_end_scan(struct scan_state *ss) { struct vm_pagequeue *pq; pq = ss->pq; vm_pagequeue_assert_locked(pq); KASSERT((ss->marker->aflags & PGA_ENQUEUED) != 0, ("marker %p not enqueued", ss->marker)); TAILQ_REMOVE(&pq->pq_pl, ss->marker, plinks.q); vm_page_aflag_clear(ss->marker, PGA_ENQUEUED); pq->pq_pdpages += ss->scanned; } /* * Add a small number of queued pages to a batch queue for later processing * without the corresponding queue lock held. The caller must have enqueued a * marker page at the desired start point for the scan. Pages will be * physically dequeued if the caller so requests. Otherwise, the returned * batch may contain marker pages, and it is up to the caller to handle them. * * When processing the batch queue, vm_page_queue() must be used to * determine whether the page has been logically dequeued by another thread. * Once this check is performed, the page lock guarantees that the page will * not be disassociated from the queue. */ static __always_inline void vm_pageout_collect_batch(struct scan_state *ss, const bool dequeue) { struct vm_pagequeue *pq; vm_page_t m, marker, n; marker = ss->marker; pq = ss->pq; KASSERT((marker->aflags & PGA_ENQUEUED) != 0, ("marker %p not enqueued", ss->marker)); vm_pagequeue_lock(pq); for (m = TAILQ_NEXT(marker, plinks.q); m != NULL && ss->scanned < ss->maxscan && ss->bq.bq_cnt < VM_BATCHQUEUE_SIZE; m = n, ss->scanned++) { n = TAILQ_NEXT(m, plinks.q); if ((m->flags & PG_MARKER) == 0) { KASSERT((m->aflags & PGA_ENQUEUED) != 0, ("page %p not enqueued", m)); KASSERT((m->flags & PG_FICTITIOUS) == 0, ("Fictitious page %p cannot be in page queue", m)); KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("Unmanaged page %p cannot be in page queue", m)); } else if (dequeue) continue; (void)vm_batchqueue_insert(&ss->bq, m); if (dequeue) { TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); vm_page_aflag_clear(m, PGA_ENQUEUED); } } TAILQ_REMOVE(&pq->pq_pl, marker, plinks.q); if (__predict_true(m != NULL)) TAILQ_INSERT_BEFORE(m, marker, plinks.q); else TAILQ_INSERT_TAIL(&pq->pq_pl, marker, plinks.q); if (dequeue) vm_pagequeue_cnt_add(pq, -ss->bq.bq_cnt); vm_pagequeue_unlock(pq); } /* * Return the next page to be scanned, or NULL if the scan is complete. */ static __always_inline vm_page_t vm_pageout_next(struct scan_state *ss, const bool dequeue) { if (ss->bq.bq_cnt == 0) vm_pageout_collect_batch(ss, dequeue); return (vm_batchqueue_pop(&ss->bq)); } /* * Scan for pages at adjacent offsets within the given page's object that are * eligible for laundering, form a cluster of these pages and the given page, * and launder that cluster. */ static int vm_pageout_cluster(vm_page_t m) { vm_object_t object; vm_page_t mc[2 * vm_pageout_page_count], p, pb, ps; vm_pindex_t pindex; int ib, is, page_base, pageout_count; object = m->object; VM_OBJECT_ASSERT_WLOCKED(object); pindex = m->pindex; vm_page_assert_xbusied(m); mc[vm_pageout_page_count] = pb = ps = m; pageout_count = 1; page_base = vm_pageout_page_count; ib = 1; is = 1; /* * We can cluster only if the page is not clean, busy, or held, and * the page is in the laundry queue. * * During heavy mmap/modification loads the pageout * daemon can really fragment the underlying file * due to flushing pages out of order and not trying to * align the clusters (which leaves sporadic out-of-order * holes). To solve this problem we do the reverse scan * first and attempt to align our cluster, then do a * forward scan if room remains. */ more: while (ib != 0 && pageout_count < vm_pageout_page_count) { if (ib > pindex) { ib = 0; break; } if ((p = vm_page_prev(pb)) == NULL || vm_page_tryxbusy(p) == 0) { ib = 0; break; } if (vm_page_wired(p)) { ib = 0; vm_page_xunbusy(p); break; } vm_page_test_dirty(p); if (p->dirty == 0) { ib = 0; vm_page_xunbusy(p); break; } vm_page_lock(p); if (!vm_page_in_laundry(p) || !vm_page_try_remove_write(p)) { vm_page_unlock(p); vm_page_xunbusy(p); ib = 0; break; } vm_page_unlock(p); mc[--page_base] = pb = p; ++pageout_count; ++ib; /* * We are at an alignment boundary. Stop here, and switch * directions. Do not clear ib. */ if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) break; } while (pageout_count < vm_pageout_page_count && pindex + is < object->size) { if ((p = vm_page_next(ps)) == NULL || vm_page_tryxbusy(p) == 0) break; if (vm_page_wired(p)) { vm_page_xunbusy(p); break; } vm_page_test_dirty(p); if (p->dirty == 0) { vm_page_xunbusy(p); break; } vm_page_lock(p); if (!vm_page_in_laundry(p) || !vm_page_try_remove_write(p)) { vm_page_unlock(p); vm_page_xunbusy(p); break; } vm_page_unlock(p); mc[page_base + pageout_count] = ps = p; ++pageout_count; ++is; } /* * If we exhausted our forward scan, continue with the reverse scan * when possible, even past an alignment boundary. This catches * boundary conditions. */ if (ib != 0 && pageout_count < vm_pageout_page_count) goto more; return (vm_pageout_flush(&mc[page_base], pageout_count, VM_PAGER_PUT_NOREUSE, 0, NULL, NULL)); } /* * vm_pageout_flush() - launder the given pages * * The given pages are laundered. Note that we setup for the start of * I/O ( i.e. busy the page ), mark it read-only, and bump the object * reference count all in here rather then in the parent. If we want * the parent to do more sophisticated things we may have to change * the ordering. * * Returned runlen is the count of pages between mreq and first * page after mreq with status VM_PAGER_AGAIN. * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL * for any page in runlen set. */ int vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen, boolean_t *eio) { vm_object_t object = mc[0]->object; int pageout_status[count]; int numpagedout = 0; int i, runlen; VM_OBJECT_ASSERT_WLOCKED(object); /* * Initiate I/O. Mark the pages shared busy and verify that they're * valid and read-only. * * We do not have to fixup the clean/dirty bits here... we can * allow the pager to do it after the I/O completes. * * NOTE! mc[i]->dirty may be partial or fragmented due to an * edge case with file fragments. */ for (i = 0; i < count; i++) { - KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, + KASSERT(vm_page_all_valid(mc[i]), ("vm_pageout_flush: partially invalid page %p index %d/%d", mc[i], i, count)); KASSERT((mc[i]->aflags & PGA_WRITEABLE) == 0, ("vm_pageout_flush: writeable page %p", mc[i])); vm_page_busy_downgrade(mc[i]); } vm_object_pip_add(object, count); vm_pager_put_pages(object, mc, count, flags, pageout_status); runlen = count - mreq; if (eio != NULL) *eio = FALSE; for (i = 0; i < count; i++) { vm_page_t mt = mc[i]; KASSERT(pageout_status[i] == VM_PAGER_PEND || !pmap_page_is_write_mapped(mt), ("vm_pageout_flush: page %p is not write protected", mt)); switch (pageout_status[i]) { case VM_PAGER_OK: vm_page_lock(mt); if (vm_page_in_laundry(mt)) vm_page_deactivate_noreuse(mt); vm_page_unlock(mt); /* FALLTHROUGH */ case VM_PAGER_PEND: numpagedout++; break; case VM_PAGER_BAD: /* * The page is outside the object's range. We pretend * that the page out worked and clean the page, so the * changes will be lost if the page is reclaimed by * the page daemon. */ vm_page_undirty(mt); vm_page_lock(mt); if (vm_page_in_laundry(mt)) vm_page_deactivate_noreuse(mt); vm_page_unlock(mt); break; case VM_PAGER_ERROR: case VM_PAGER_FAIL: /* * If the page couldn't be paged out to swap because the * pager wasn't able to find space, place the page in * the PQ_UNSWAPPABLE holding queue. This is an * optimization that prevents the page daemon from * wasting CPU cycles on pages that cannot be reclaimed * becase no swap device is configured. * * Otherwise, reactivate the page so that it doesn't * clog the laundry and inactive queues. (We will try * paging it out again later.) */ vm_page_lock(mt); if (object->type == OBJT_SWAP && pageout_status[i] == VM_PAGER_FAIL) { vm_page_unswappable(mt); numpagedout++; } else vm_page_activate(mt); vm_page_unlock(mt); if (eio != NULL && i >= mreq && i - mreq < runlen) *eio = TRUE; break; case VM_PAGER_AGAIN: if (i >= mreq && i - mreq < runlen) runlen = i - mreq; break; } /* * If the operation is still going, leave the page busy to * block all other accesses. Also, leave the paging in * progress indicator set so that we don't attempt an object * collapse. */ if (pageout_status[i] != VM_PAGER_PEND) { vm_object_pip_wakeup(object); vm_page_sunbusy(mt); } } if (prunlen != NULL) *prunlen = runlen; return (numpagedout); } static void vm_pageout_swapon(void *arg __unused, struct swdevt *sp __unused) { atomic_store_rel_int(&swapdev_enabled, 1); } static void vm_pageout_swapoff(void *arg __unused, struct swdevt *sp __unused) { if (swap_pager_nswapdev() == 1) atomic_store_rel_int(&swapdev_enabled, 0); } /* * Attempt to acquire all of the necessary locks to launder a page and * then call through the clustering layer to PUTPAGES. Wait a short * time for a vnode lock. * * Requires the page and object lock on entry, releases both before return. * Returns 0 on success and an errno otherwise. */ static int vm_pageout_clean(vm_page_t m, int *numpagedout) { struct vnode *vp; struct mount *mp; vm_object_t object; vm_pindex_t pindex; int error, lockmode; vm_page_assert_locked(m); object = m->object; VM_OBJECT_ASSERT_WLOCKED(object); error = 0; vp = NULL; mp = NULL; /* * The object is already known NOT to be dead. It * is possible for the vget() to block the whole * pageout daemon, but the new low-memory handling * code should prevent it. * * We can't wait forever for the vnode lock, we might * deadlock due to a vn_read() getting stuck in * vm_wait while holding this vnode. We skip the * vnode if we can't get it in a reasonable amount * of time. */ if (object->type == OBJT_VNODE) { vm_page_unlock(m); vm_page_xunbusy(m); vp = object->handle; if (vp->v_type == VREG && vn_start_write(vp, &mp, V_NOWAIT) != 0) { mp = NULL; error = EDEADLK; goto unlock_all; } KASSERT(mp != NULL, ("vp %p with NULL v_mount", vp)); vm_object_reference_locked(object); pindex = m->pindex; VM_OBJECT_WUNLOCK(object); lockmode = MNT_SHARED_WRITES(vp->v_mount) ? LK_SHARED : LK_EXCLUSIVE; if (vget(vp, lockmode | LK_TIMELOCK, curthread)) { vp = NULL; error = EDEADLK; goto unlock_mp; } VM_OBJECT_WLOCK(object); /* * Ensure that the object and vnode were not disassociated * while locks were dropped. */ if (vp->v_object != object) { error = ENOENT; goto unlock_all; } vm_page_lock(m); /* * While the object and page were unlocked, the page * may have been: * (1) moved to a different queue, * (2) reallocated to a different object, * (3) reallocated to a different offset, or * (4) cleaned. */ if (!vm_page_in_laundry(m) || m->object != object || m->pindex != pindex || m->dirty == 0) { vm_page_unlock(m); error = ENXIO; goto unlock_all; } /* * The page may have been busied while the object and page * locks were released. */ if (vm_page_tryxbusy(m) == 0) { vm_page_unlock(m); error = EBUSY; goto unlock_all; } } /* * Remove all writeable mappings, failing if the page is wired. */ if (!vm_page_try_remove_write(m)) { vm_page_xunbusy(m); vm_page_unlock(m); error = EBUSY; goto unlock_all; } vm_page_unlock(m); /* * If a page is dirty, then it is either being washed * (but not yet cleaned) or it is still in the * laundry. If it is still in the laundry, then we * start the cleaning operation. */ if ((*numpagedout = vm_pageout_cluster(m)) == 0) error = EIO; unlock_all: VM_OBJECT_WUNLOCK(object); unlock_mp: vm_page_lock_assert(m, MA_NOTOWNED); if (mp != NULL) { if (vp != NULL) vput(vp); vm_object_deallocate(object); vn_finished_write(mp); } return (error); } /* * Attempt to launder the specified number of pages. * * Returns the number of pages successfully laundered. */ static int vm_pageout_launder(struct vm_domain *vmd, int launder, bool in_shortfall) { struct scan_state ss; struct vm_pagequeue *pq; struct mtx *mtx; vm_object_t object; vm_page_t m, marker; int act_delta, error, numpagedout, queue, starting_target; int vnodes_skipped; bool pageout_ok; mtx = NULL; object = NULL; starting_target = launder; vnodes_skipped = 0; /* * Scan the laundry queues for pages eligible to be laundered. We stop * once the target number of dirty pages have been laundered, or once * we've reached the end of the queue. A single iteration of this loop * may cause more than one page to be laundered because of clustering. * * As an optimization, we avoid laundering from PQ_UNSWAPPABLE when no * swap devices are configured. */ if (atomic_load_acq_int(&swapdev_enabled)) queue = PQ_UNSWAPPABLE; else queue = PQ_LAUNDRY; scan: marker = &vmd->vmd_markers[queue]; pq = &vmd->vmd_pagequeues[queue]; vm_pagequeue_lock(pq); vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt); while (launder > 0 && (m = vm_pageout_next(&ss, false)) != NULL) { if (__predict_false((m->flags & PG_MARKER) != 0)) continue; vm_page_change_lock(m, &mtx); recheck: /* * The page may have been disassociated from the queue * or even freed while locks were dropped. We thus must be * careful whenever modifying page state. Once the object lock * has been acquired, we have a stable reference to the page. */ if (vm_page_queue(m) != queue) continue; /* * A requeue was requested, so this page gets a second * chance. */ if ((m->aflags & PGA_REQUEUE) != 0) { vm_page_pqbatch_submit(m, queue); continue; } /* * Wired pages may not be freed. Complete their removal * from the queue now to avoid needless revisits during * future scans. This check is racy and must be reverified once * we hold the object lock and have verified that the page * is not busy. */ if (vm_page_wired(m)) { vm_page_dequeue_deferred(m); continue; } if (object != m->object) { if (object != NULL) VM_OBJECT_WUNLOCK(object); /* * A page's object pointer may be set to NULL before * the object lock is acquired. */ object = (vm_object_t)atomic_load_ptr(&m->object); if (object != NULL && !VM_OBJECT_TRYWLOCK(object)) { mtx_unlock(mtx); /* Depends on type-stability. */ VM_OBJECT_WLOCK(object); mtx_lock(mtx); goto recheck; } } if (__predict_false(m->object == NULL)) /* * The page has been removed from its object. */ continue; KASSERT(m->object == object, ("page %p does not belong to %p", m, object)); if (vm_page_tryxbusy(m) == 0) continue; /* * Re-check for wirings now that we hold the object lock and * have verified that the page is unbusied. If the page is * mapped, it may still be wired by pmap lookups. The call to * vm_page_try_remove_all() below atomically checks for such * wirings and removes mappings. If the page is unmapped, the * wire count is guaranteed not to increase. */ if (__predict_false(vm_page_wired(m))) { vm_page_xunbusy(m); vm_page_dequeue_deferred(m); continue; } /* * Invalid pages can be easily freed. They cannot be * mapped; vm_page_free() asserts this. */ - if (m->valid == 0) + if (vm_page_none_valid(m)) goto free_page; /* * If the page has been referenced and the object is not dead, * reactivate or requeue the page depending on whether the * object is mapped. * * Test PGA_REFERENCED after calling pmap_ts_referenced() so * that a reference from a concurrently destroyed mapping is * observed here and now. */ if (object->ref_count != 0) act_delta = pmap_ts_referenced(m); else { KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m)); act_delta = 0; } if ((m->aflags & PGA_REFERENCED) != 0) { vm_page_aflag_clear(m, PGA_REFERENCED); act_delta++; } if (act_delta != 0) { if (object->ref_count != 0) { vm_page_xunbusy(m); VM_CNT_INC(v_reactivated); vm_page_activate(m); /* * Increase the activation count if the page * was referenced while in the laundry queue. * This makes it less likely that the page will * be returned prematurely to the inactive * queue. */ m->act_count += act_delta + ACT_ADVANCE; /* * If this was a background laundering, count * activated pages towards our target. The * purpose of background laundering is to ensure * that pages are eventually cycled through the * laundry queue, and an activation is a valid * way out. */ if (!in_shortfall) launder--; continue; } else if ((object->flags & OBJ_DEAD) == 0) { vm_page_xunbusy(m); vm_page_requeue(m); continue; } } /* * If the page appears to be clean at the machine-independent * layer, then remove all of its mappings from the pmap in * anticipation of freeing it. If, however, any of the page's * mappings allow write access, then the page may still be * modified until the last of those mappings are removed. */ if (object->ref_count != 0) { vm_page_test_dirty(m); if (m->dirty == 0 && !vm_page_try_remove_all(m)) { vm_page_xunbusy(m); vm_page_dequeue_deferred(m); continue; } } /* * Clean pages are freed, and dirty pages are paged out unless * they belong to a dead object. Requeueing dirty pages from * dead objects is pointless, as they are being paged out and * freed by the thread that destroyed the object. */ if (m->dirty == 0) { free_page: vm_page_free(m); VM_CNT_INC(v_dfree); } else if ((object->flags & OBJ_DEAD) == 0) { if (object->type != OBJT_SWAP && object->type != OBJT_DEFAULT) pageout_ok = true; else if (disable_swap_pageouts) pageout_ok = false; else pageout_ok = true; if (!pageout_ok) { vm_page_xunbusy(m); vm_page_requeue(m); continue; } /* * Form a cluster with adjacent, dirty pages from the * same object, and page out that entire cluster. * * The adjacent, dirty pages must also be in the * laundry. However, their mappings are not checked * for new references. Consequently, a recently * referenced page may be paged out. However, that * page will not be prematurely reclaimed. After page * out, the page will be placed in the inactive queue, * where any new references will be detected and the * page reactivated. */ error = vm_pageout_clean(m, &numpagedout); if (error == 0) { launder -= numpagedout; ss.scanned += numpagedout; } else if (error == EDEADLK) { pageout_lock_miss++; vnodes_skipped++; } mtx = NULL; object = NULL; } else vm_page_xunbusy(m); } if (mtx != NULL) { mtx_unlock(mtx); mtx = NULL; } if (object != NULL) { VM_OBJECT_WUNLOCK(object); object = NULL; } vm_pagequeue_lock(pq); vm_pageout_end_scan(&ss); vm_pagequeue_unlock(pq); if (launder > 0 && queue == PQ_UNSWAPPABLE) { queue = PQ_LAUNDRY; goto scan; } /* * Wakeup the sync daemon if we skipped a vnode in a writeable object * and we didn't launder enough pages. */ if (vnodes_skipped > 0 && launder > 0) (void)speedup_syncer(); return (starting_target - launder); } /* * Compute the integer square root. */ static u_int isqrt(u_int num) { u_int bit, root, tmp; bit = num != 0 ? (1u << ((fls(num) - 1) & ~1)) : 0; root = 0; while (bit != 0) { tmp = root + bit; root >>= 1; if (num >= tmp) { num -= tmp; root += bit; } bit >>= 2; } return (root); } /* * Perform the work of the laundry thread: periodically wake up and determine * whether any pages need to be laundered. If so, determine the number of pages * that need to be laundered, and launder them. */ static void vm_pageout_laundry_worker(void *arg) { struct vm_domain *vmd; struct vm_pagequeue *pq; uint64_t nclean, ndirty, nfreed; int domain, last_target, launder, shortfall, shortfall_cycle, target; bool in_shortfall; domain = (uintptr_t)arg; vmd = VM_DOMAIN(domain); pq = &vmd->vmd_pagequeues[PQ_LAUNDRY]; KASSERT(vmd->vmd_segs != 0, ("domain without segments")); shortfall = 0; in_shortfall = false; shortfall_cycle = 0; last_target = target = 0; nfreed = 0; /* * Calls to these handlers are serialized by the swap syscall lock. */ (void)EVENTHANDLER_REGISTER(swapon, vm_pageout_swapon, vmd, EVENTHANDLER_PRI_ANY); (void)EVENTHANDLER_REGISTER(swapoff, vm_pageout_swapoff, vmd, EVENTHANDLER_PRI_ANY); /* * The pageout laundry worker is never done, so loop forever. */ for (;;) { KASSERT(target >= 0, ("negative target %d", target)); KASSERT(shortfall_cycle >= 0, ("negative cycle %d", shortfall_cycle)); launder = 0; /* * First determine whether we need to launder pages to meet a * shortage of free pages. */ if (shortfall > 0) { in_shortfall = true; shortfall_cycle = VM_LAUNDER_RATE / VM_INACT_SCAN_RATE; target = shortfall; } else if (!in_shortfall) goto trybackground; else if (shortfall_cycle == 0 || vm_laundry_target(vmd) <= 0) { /* * We recently entered shortfall and began laundering * pages. If we have completed that laundering run * (and we are no longer in shortfall) or we have met * our laundry target through other activity, then we * can stop laundering pages. */ in_shortfall = false; target = 0; goto trybackground; } launder = target / shortfall_cycle--; goto dolaundry; /* * There's no immediate need to launder any pages; see if we * meet the conditions to perform background laundering: * * 1. The ratio of dirty to clean inactive pages exceeds the * background laundering threshold, or * 2. we haven't yet reached the target of the current * background laundering run. * * The background laundering threshold is not a constant. * Instead, it is a slowly growing function of the number of * clean pages freed by the page daemon since the last * background laundering. Thus, as the ratio of dirty to * clean inactive pages grows, the amount of memory pressure * required to trigger laundering decreases. We ensure * that the threshold is non-zero after an inactive queue * scan, even if that scan failed to free a single clean page. */ trybackground: nclean = vmd->vmd_free_count + vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt; ndirty = vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt; if (target == 0 && ndirty * isqrt(howmany(nfreed + 1, vmd->vmd_free_target - vmd->vmd_free_min)) >= nclean) { target = vmd->vmd_background_launder_target; } /* * We have a non-zero background laundering target. If we've * laundered up to our maximum without observing a page daemon * request, just stop. This is a safety belt that ensures we * don't launder an excessive amount if memory pressure is low * and the ratio of dirty to clean pages is large. Otherwise, * proceed at the background laundering rate. */ if (target > 0) { if (nfreed > 0) { nfreed = 0; last_target = target; } else if (last_target - target >= vm_background_launder_max * PAGE_SIZE / 1024) { target = 0; } launder = vm_background_launder_rate * PAGE_SIZE / 1024; launder /= VM_LAUNDER_RATE; if (launder > target) launder = target; } dolaundry: if (launder > 0) { /* * Because of I/O clustering, the number of laundered * pages could exceed "target" by the maximum size of * a cluster minus one. */ target -= min(vm_pageout_launder(vmd, launder, in_shortfall), target); pause("laundp", hz / VM_LAUNDER_RATE); } /* * If we're not currently laundering pages and the page daemon * hasn't posted a new request, sleep until the page daemon * kicks us. */ vm_pagequeue_lock(pq); if (target == 0 && vmd->vmd_laundry_request == VM_LAUNDRY_IDLE) (void)mtx_sleep(&vmd->vmd_laundry_request, vm_pagequeue_lockptr(pq), PVM, "launds", 0); /* * If the pagedaemon has indicated that it's in shortfall, start * a shortfall laundering unless we're already in the middle of * one. This may preempt a background laundering. */ if (vmd->vmd_laundry_request == VM_LAUNDRY_SHORTFALL && (!in_shortfall || shortfall_cycle == 0)) { shortfall = vm_laundry_target(vmd) + vmd->vmd_pageout_deficit; target = 0; } else shortfall = 0; if (target == 0) vmd->vmd_laundry_request = VM_LAUNDRY_IDLE; nfreed += vmd->vmd_clean_pages_freed; vmd->vmd_clean_pages_freed = 0; vm_pagequeue_unlock(pq); } } /* * Compute the number of pages we want to try to move from the * active queue to either the inactive or laundry queue. * * When scanning active pages during a shortage, we make clean pages * count more heavily towards the page shortage than dirty pages. * This is because dirty pages must be laundered before they can be * reused and thus have less utility when attempting to quickly * alleviate a free page shortage. However, this weighting also * causes the scan to deactivate dirty pages more aggressively, * improving the effectiveness of clustering. */ static int vm_pageout_active_target(struct vm_domain *vmd) { int shortage; shortage = vmd->vmd_inactive_target + vm_paging_target(vmd) - (vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt + vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt / act_scan_laundry_weight); shortage *= act_scan_laundry_weight; return (shortage); } /* * Scan the active queue. If there is no shortage of inactive pages, scan a * small portion of the queue in order to maintain quasi-LRU. */ static void vm_pageout_scan_active(struct vm_domain *vmd, int page_shortage) { struct scan_state ss; struct mtx *mtx; vm_object_t object; vm_page_t m, marker; struct vm_pagequeue *pq; long min_scan; int act_delta, max_scan, scan_tick; marker = &vmd->vmd_markers[PQ_ACTIVE]; pq = &vmd->vmd_pagequeues[PQ_ACTIVE]; vm_pagequeue_lock(pq); /* * If we're just idle polling attempt to visit every * active page within 'update_period' seconds. */ scan_tick = ticks; if (vm_pageout_update_period != 0) { min_scan = pq->pq_cnt; min_scan *= scan_tick - vmd->vmd_last_active_scan; min_scan /= hz * vm_pageout_update_period; } else min_scan = 0; if (min_scan > 0 || (page_shortage > 0 && pq->pq_cnt > 0)) vmd->vmd_last_active_scan = scan_tick; /* * Scan the active queue for pages that can be deactivated. Update * the per-page activity counter and use it to identify deactivation * candidates. Held pages may be deactivated. * * To avoid requeuing each page that remains in the active queue, we * implement the CLOCK algorithm. To keep the implementation of the * enqueue operation consistent for all page queues, we use two hands, * represented by marker pages. Scans begin at the first hand, which * precedes the second hand in the queue. When the two hands meet, * they are moved back to the head and tail of the queue, respectively, * and scanning resumes. */ max_scan = page_shortage > 0 ? pq->pq_cnt : min_scan; mtx = NULL; act_scan: vm_pageout_init_scan(&ss, pq, marker, &vmd->vmd_clock[0], max_scan); while ((m = vm_pageout_next(&ss, false)) != NULL) { if (__predict_false(m == &vmd->vmd_clock[1])) { vm_pagequeue_lock(pq); TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q); TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[1], plinks.q); TAILQ_INSERT_HEAD(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q); TAILQ_INSERT_TAIL(&pq->pq_pl, &vmd->vmd_clock[1], plinks.q); max_scan -= ss.scanned; vm_pageout_end_scan(&ss); goto act_scan; } if (__predict_false((m->flags & PG_MARKER) != 0)) continue; vm_page_change_lock(m, &mtx); /* * The page may have been disassociated from the queue * or even freed while locks were dropped. We thus must be * careful whenever modifying page state. Once the object lock * has been acquired, we have a stable reference to the page. */ if (vm_page_queue(m) != PQ_ACTIVE) continue; /* * Wired pages are dequeued lazily. */ if (vm_page_wired(m)) { vm_page_dequeue_deferred(m); continue; } /* * A page's object pointer may be set to NULL before * the object lock is acquired. */ object = (vm_object_t)atomic_load_ptr(&m->object); if (__predict_false(object == NULL)) /* * The page has been removed from its object. */ continue; /* * Check to see "how much" the page has been used. * * Test PGA_REFERENCED after calling pmap_ts_referenced() so * that a reference from a concurrently destroyed mapping is * observed here and now. * * Perform an unsynchronized object ref count check. While * the page lock ensures that the page is not reallocated to * another object, in particular, one with unmanaged mappings * that cannot support pmap_ts_referenced(), two races are, * nonetheless, possible: * 1) The count was transitioning to zero, but we saw a non- * zero value. pmap_ts_referenced() will return zero * because the page is not mapped. * 2) The count was transitioning to one, but we saw zero. * This race delays the detection of a new reference. At * worst, we will deactivate and reactivate the page. */ if (object->ref_count != 0) act_delta = pmap_ts_referenced(m); else act_delta = 0; if ((m->aflags & PGA_REFERENCED) != 0) { vm_page_aflag_clear(m, PGA_REFERENCED); act_delta++; } /* * Advance or decay the act_count based on recent usage. */ if (act_delta != 0) { m->act_count += ACT_ADVANCE + act_delta; if (m->act_count > ACT_MAX) m->act_count = ACT_MAX; } else m->act_count -= min(m->act_count, ACT_DECLINE); if (m->act_count == 0) { /* * When not short for inactive pages, let dirty pages go * through the inactive queue before moving to the * laundry queues. This gives them some extra time to * be reactivated, potentially avoiding an expensive * pageout. However, during a page shortage, the * inactive queue is necessarily small, and so dirty * pages would only spend a trivial amount of time in * the inactive queue. Therefore, we might as well * place them directly in the laundry queue to reduce * queuing overhead. */ if (page_shortage <= 0) { vm_page_swapqueue(m, PQ_ACTIVE, PQ_INACTIVE); } else { /* * Calling vm_page_test_dirty() here would * require acquisition of the object's write * lock. However, during a page shortage, * directing dirty pages into the laundry * queue is only an optimization and not a * requirement. Therefore, we simply rely on * the opportunistic updates to the page's * dirty field by the pmap. */ if (m->dirty == 0) { vm_page_swapqueue(m, PQ_ACTIVE, PQ_INACTIVE); page_shortage -= act_scan_laundry_weight; } else { vm_page_swapqueue(m, PQ_ACTIVE, PQ_LAUNDRY); page_shortage--; } } } } if (mtx != NULL) { mtx_unlock(mtx); mtx = NULL; } vm_pagequeue_lock(pq); TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q); TAILQ_INSERT_AFTER(&pq->pq_pl, marker, &vmd->vmd_clock[0], plinks.q); vm_pageout_end_scan(&ss); vm_pagequeue_unlock(pq); } static int vm_pageout_reinsert_inactive_page(struct scan_state *ss, vm_page_t m) { struct vm_domain *vmd; if (m->queue != PQ_INACTIVE || (m->aflags & PGA_ENQUEUED) != 0) return (0); vm_page_aflag_set(m, PGA_ENQUEUED); if ((m->aflags & PGA_REQUEUE_HEAD) != 0) { vmd = vm_pagequeue_domain(m); TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q); vm_page_aflag_clear(m, PGA_REQUEUE | PGA_REQUEUE_HEAD); } else if ((m->aflags & PGA_REQUEUE) != 0) { TAILQ_INSERT_TAIL(&ss->pq->pq_pl, m, plinks.q); vm_page_aflag_clear(m, PGA_REQUEUE | PGA_REQUEUE_HEAD); } else TAILQ_INSERT_BEFORE(ss->marker, m, plinks.q); return (1); } /* * Re-add stuck pages to the inactive queue. We will examine them again * during the next scan. If the queue state of a page has changed since * it was physically removed from the page queue in * vm_pageout_collect_batch(), don't do anything with that page. */ static void vm_pageout_reinsert_inactive(struct scan_state *ss, struct vm_batchqueue *bq, vm_page_t m) { struct vm_pagequeue *pq; int delta; delta = 0; pq = ss->pq; if (m != NULL) { if (vm_batchqueue_insert(bq, m)) return; vm_pagequeue_lock(pq); delta += vm_pageout_reinsert_inactive_page(ss, m); } else vm_pagequeue_lock(pq); while ((m = vm_batchqueue_pop(bq)) != NULL) delta += vm_pageout_reinsert_inactive_page(ss, m); vm_pagequeue_cnt_add(pq, delta); vm_pagequeue_unlock(pq); vm_batchqueue_init(bq); } /* * Attempt to reclaim the requested number of pages from the inactive queue. * Returns true if the shortage was addressed. */ static int vm_pageout_scan_inactive(struct vm_domain *vmd, int shortage, int *addl_shortage) { struct scan_state ss; struct vm_batchqueue rq; struct mtx *mtx; vm_page_t m, marker; struct vm_pagequeue *pq; vm_object_t object; int act_delta, addl_page_shortage, deficit, page_shortage; int starting_page_shortage; /* * The addl_page_shortage is an estimate of the number of temporarily * stuck pages in the inactive queue. In other words, the * number of pages from the inactive count that should be * discounted in setting the target for the active queue scan. */ addl_page_shortage = 0; /* * vmd_pageout_deficit counts the number of pages requested in * allocations that failed because of a free page shortage. We assume * that the allocations will be reattempted and thus include the deficit * in our scan target. */ deficit = atomic_readandclear_int(&vmd->vmd_pageout_deficit); starting_page_shortage = page_shortage = shortage + deficit; mtx = NULL; object = NULL; vm_batchqueue_init(&rq); /* * Start scanning the inactive queue for pages that we can free. The * scan will stop when we reach the target or we have scanned the * entire queue. (Note that m->act_count is not used to make * decisions for the inactive queue, only for the active queue.) */ marker = &vmd->vmd_markers[PQ_INACTIVE]; pq = &vmd->vmd_pagequeues[PQ_INACTIVE]; vm_pagequeue_lock(pq); vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt); while (page_shortage > 0 && (m = vm_pageout_next(&ss, true)) != NULL) { KASSERT((m->flags & PG_MARKER) == 0, ("marker page %p was dequeued", m)); vm_page_change_lock(m, &mtx); recheck: /* * The page may have been disassociated from the queue * or even freed while locks were dropped. We thus must be * careful whenever modifying page state. Once the object lock * has been acquired, we have a stable reference to the page. */ if (vm_page_queue(m) != PQ_INACTIVE) { addl_page_shortage++; continue; } /* * The page was re-enqueued after the page queue lock was * dropped, or a requeue was requested. This page gets a second * chance. */ if ((m->aflags & (PGA_ENQUEUED | PGA_REQUEUE | PGA_REQUEUE_HEAD)) != 0) goto reinsert; /* * Wired pages may not be freed. Complete their removal * from the queue now to avoid needless revisits during * future scans. This check is racy and must be reverified once * we hold the object lock and have verified that the page * is not busy. */ if (vm_page_wired(m)) { vm_page_dequeue_deferred(m); continue; } if (object != m->object) { if (object != NULL) VM_OBJECT_WUNLOCK(object); /* * A page's object pointer may be set to NULL before * the object lock is acquired. */ object = (vm_object_t)atomic_load_ptr(&m->object); if (object != NULL && !VM_OBJECT_TRYWLOCK(object)) { mtx_unlock(mtx); /* Depends on type-stability. */ VM_OBJECT_WLOCK(object); mtx_lock(mtx); goto recheck; } } if (__predict_false(m->object == NULL)) /* * The page has been removed from its object. */ continue; KASSERT(m->object == object, ("page %p does not belong to %p", m, object)); if (vm_page_tryxbusy(m) == 0) { /* * Don't mess with busy pages. Leave them at * the front of the queue. Most likely, they * are being paged out and will leave the * queue shortly after the scan finishes. So, * they ought to be discounted from the * inactive count. */ addl_page_shortage++; goto reinsert; } /* * Re-check for wirings now that we hold the object lock and * have verified that the page is unbusied. If the page is * mapped, it may still be wired by pmap lookups. The call to * vm_page_try_remove_all() below atomically checks for such * wirings and removes mappings. If the page is unmapped, the * wire count is guaranteed not to increase. */ if (__predict_false(vm_page_wired(m))) { vm_page_xunbusy(m); vm_page_dequeue_deferred(m); continue; } /* * Invalid pages can be easily freed. They cannot be * mapped, vm_page_free() asserts this. */ - if (m->valid == 0) + if (vm_page_none_valid(m)) goto free_page; /* * If the page has been referenced and the object is not dead, * reactivate or requeue the page depending on whether the * object is mapped. * * Test PGA_REFERENCED after calling pmap_ts_referenced() so * that a reference from a concurrently destroyed mapping is * observed here and now. */ if (object->ref_count != 0) act_delta = pmap_ts_referenced(m); else { KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m)); act_delta = 0; } if ((m->aflags & PGA_REFERENCED) != 0) { vm_page_aflag_clear(m, PGA_REFERENCED); act_delta++; } if (act_delta != 0) { if (object->ref_count != 0) { vm_page_xunbusy(m); VM_CNT_INC(v_reactivated); vm_page_activate(m); /* * Increase the activation count if the page * was referenced while in the inactive queue. * This makes it less likely that the page will * be returned prematurely to the inactive * queue. */ m->act_count += act_delta + ACT_ADVANCE; continue; } else if ((object->flags & OBJ_DEAD) == 0) { vm_page_xunbusy(m); vm_page_aflag_set(m, PGA_REQUEUE); goto reinsert; } } /* * If the page appears to be clean at the machine-independent * layer, then remove all of its mappings from the pmap in * anticipation of freeing it. If, however, any of the page's * mappings allow write access, then the page may still be * modified until the last of those mappings are removed. */ if (object->ref_count != 0) { vm_page_test_dirty(m); if (m->dirty == 0 && !vm_page_try_remove_all(m)) { vm_page_xunbusy(m); vm_page_dequeue_deferred(m); continue; } } /* * Clean pages can be freed, but dirty pages must be sent back * to the laundry, unless they belong to a dead object. * Requeueing dirty pages from dead objects is pointless, as * they are being paged out and freed by the thread that * destroyed the object. */ if (m->dirty == 0) { free_page: /* * Because we dequeued the page and have already * checked for concurrent dequeue and enqueue * requests, we can safely disassociate the page * from the inactive queue. */ KASSERT((m->aflags & PGA_QUEUE_STATE_MASK) == 0, ("page %p has queue state", m)); m->queue = PQ_NONE; vm_page_free(m); page_shortage--; continue; } vm_page_xunbusy(m); if ((object->flags & OBJ_DEAD) == 0) vm_page_launder(m); continue; reinsert: vm_pageout_reinsert_inactive(&ss, &rq, m); } if (mtx != NULL) mtx_unlock(mtx); if (object != NULL) VM_OBJECT_WUNLOCK(object); vm_pageout_reinsert_inactive(&ss, &rq, NULL); vm_pageout_reinsert_inactive(&ss, &ss.bq, NULL); vm_pagequeue_lock(pq); vm_pageout_end_scan(&ss); vm_pagequeue_unlock(pq); VM_CNT_ADD(v_dfree, starting_page_shortage - page_shortage); /* * Wake up the laundry thread so that it can perform any needed * laundering. If we didn't meet our target, we're in shortfall and * need to launder more aggressively. If PQ_LAUNDRY is empty and no * swap devices are configured, the laundry thread has no work to do, so * don't bother waking it up. * * The laundry thread uses the number of inactive queue scans elapsed * since the last laundering to determine whether to launder again, so * keep count. */ if (starting_page_shortage > 0) { pq = &vmd->vmd_pagequeues[PQ_LAUNDRY]; vm_pagequeue_lock(pq); if (vmd->vmd_laundry_request == VM_LAUNDRY_IDLE && (pq->pq_cnt > 0 || atomic_load_acq_int(&swapdev_enabled))) { if (page_shortage > 0) { vmd->vmd_laundry_request = VM_LAUNDRY_SHORTFALL; VM_CNT_INC(v_pdshortfalls); } else if (vmd->vmd_laundry_request != VM_LAUNDRY_SHORTFALL) vmd->vmd_laundry_request = VM_LAUNDRY_BACKGROUND; wakeup(&vmd->vmd_laundry_request); } vmd->vmd_clean_pages_freed += starting_page_shortage - page_shortage; vm_pagequeue_unlock(pq); } /* * Wakeup the swapout daemon if we didn't free the targeted number of * pages. */ if (page_shortage > 0) vm_swapout_run(); /* * If the inactive queue scan fails repeatedly to meet its * target, kill the largest process. */ vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage); /* * Reclaim pages by swapping out idle processes, if configured to do so. */ vm_swapout_run_idle(); /* * See the description of addl_page_shortage above. */ *addl_shortage = addl_page_shortage + deficit; return (page_shortage <= 0); } static int vm_pageout_oom_vote; /* * The pagedaemon threads randlomly select one to perform the * OOM. Trying to kill processes before all pagedaemons * failed to reach free target is premature. */ static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage, int starting_page_shortage) { int old_vote; if (starting_page_shortage <= 0 || starting_page_shortage != page_shortage) vmd->vmd_oom_seq = 0; else vmd->vmd_oom_seq++; if (vmd->vmd_oom_seq < vm_pageout_oom_seq) { if (vmd->vmd_oom) { vmd->vmd_oom = FALSE; atomic_subtract_int(&vm_pageout_oom_vote, 1); } return; } /* * Do not follow the call sequence until OOM condition is * cleared. */ vmd->vmd_oom_seq = 0; if (vmd->vmd_oom) return; vmd->vmd_oom = TRUE; old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1); if (old_vote != vm_ndomains - 1) return; /* * The current pagedaemon thread is the last in the quorum to * start OOM. Initiate the selection and signaling of the * victim. */ vm_pageout_oom(VM_OOM_MEM); /* * After one round of OOM terror, recall our vote. On the * next pass, current pagedaemon would vote again if the low * memory condition is still there, due to vmd_oom being * false. */ vmd->vmd_oom = FALSE; atomic_subtract_int(&vm_pageout_oom_vote, 1); } /* * The OOM killer is the page daemon's action of last resort when * memory allocation requests have been stalled for a prolonged period * of time because it cannot reclaim memory. This function computes * the approximate number of physical pages that could be reclaimed if * the specified address space is destroyed. * * Private, anonymous memory owned by the address space is the * principal resource that we expect to recover after an OOM kill. * Since the physical pages mapped by the address space's COW entries * are typically shared pages, they are unlikely to be released and so * they are not counted. * * To get to the point where the page daemon runs the OOM killer, its * efforts to write-back vnode-backed pages may have stalled. This * could be caused by a memory allocation deadlock in the write path * that might be resolved by an OOM kill. Therefore, physical pages * belonging to vnode-backed objects are counted, because they might * be freed without being written out first if the address space holds * the last reference to an unlinked vnode. * * Similarly, physical pages belonging to OBJT_PHYS objects are * counted because the address space might hold the last reference to * the object. */ static long vm_pageout_oom_pagecount(struct vmspace *vmspace) { vm_map_t map; vm_map_entry_t entry; vm_object_t obj; long res; map = &vmspace->vm_map; KASSERT(!map->system_map, ("system map")); sx_assert(&map->lock, SA_LOCKED); res = 0; VM_MAP_ENTRY_FOREACH(entry, map) { if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) continue; obj = entry->object.vm_object; if (obj == NULL) continue; if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 && obj->ref_count != 1) continue; switch (obj->type) { case OBJT_DEFAULT: case OBJT_SWAP: case OBJT_PHYS: case OBJT_VNODE: res += obj->resident_page_count; break; } } return (res); } static int vm_oom_ratelim_last; static int vm_oom_pf_secs = 10; SYSCTL_INT(_vm, OID_AUTO, oom_pf_secs, CTLFLAG_RWTUN, &vm_oom_pf_secs, 0, ""); static struct mtx vm_oom_ratelim_mtx; void vm_pageout_oom(int shortage) { struct proc *p, *bigproc; vm_offset_t size, bigsize; struct thread *td; struct vmspace *vm; int now; bool breakout; /* * For OOM requests originating from vm_fault(), there is a high * chance that a single large process faults simultaneously in * several threads. Also, on an active system running many * processes of middle-size, like buildworld, all of them * could fault almost simultaneously as well. * * To avoid killing too many processes, rate-limit OOMs * initiated by vm_fault() time-outs on the waits for free * pages. */ mtx_lock(&vm_oom_ratelim_mtx); now = ticks; if (shortage == VM_OOM_MEM_PF && (u_int)(now - vm_oom_ratelim_last) < hz * vm_oom_pf_secs) { mtx_unlock(&vm_oom_ratelim_mtx); return; } vm_oom_ratelim_last = now; mtx_unlock(&vm_oom_ratelim_mtx); /* * We keep the process bigproc locked once we find it to keep anyone * from messing with it; however, there is a possibility of * deadlock if process B is bigproc and one of its child processes * attempts to propagate a signal to B while we are waiting for A's * lock while walking this list. To avoid this, we don't block on * the process lock but just skip a process if it is already locked. */ bigproc = NULL; bigsize = 0; sx_slock(&allproc_lock); FOREACH_PROC_IN_SYSTEM(p) { PROC_LOCK(p); /* * If this is a system, protected or killed process, skip it. */ if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 || p->p_pid == 1 || P_KILLED(p) || (p->p_pid < 48 && swap_pager_avail != 0)) { PROC_UNLOCK(p); continue; } /* * If the process is in a non-running type state, * don't touch it. Check all the threads individually. */ breakout = false; FOREACH_THREAD_IN_PROC(p, td) { thread_lock(td); if (!TD_ON_RUNQ(td) && !TD_IS_RUNNING(td) && !TD_IS_SLEEPING(td) && !TD_IS_SUSPENDED(td) && !TD_IS_SWAPPED(td)) { thread_unlock(td); breakout = true; break; } thread_unlock(td); } if (breakout) { PROC_UNLOCK(p); continue; } /* * get the process size */ vm = vmspace_acquire_ref(p); if (vm == NULL) { PROC_UNLOCK(p); continue; } _PHOLD_LITE(p); PROC_UNLOCK(p); sx_sunlock(&allproc_lock); if (!vm_map_trylock_read(&vm->vm_map)) { vmspace_free(vm); sx_slock(&allproc_lock); PRELE(p); continue; } size = vmspace_swap_count(vm); if (shortage == VM_OOM_MEM || shortage == VM_OOM_MEM_PF) size += vm_pageout_oom_pagecount(vm); vm_map_unlock_read(&vm->vm_map); vmspace_free(vm); sx_slock(&allproc_lock); /* * If this process is bigger than the biggest one, * remember it. */ if (size > bigsize) { if (bigproc != NULL) PRELE(bigproc); bigproc = p; bigsize = size; } else { PRELE(p); } } sx_sunlock(&allproc_lock); if (bigproc != NULL) { if (vm_panic_on_oom != 0) panic("out of swap space"); PROC_LOCK(bigproc); killproc(bigproc, "out of swap space"); sched_nice(bigproc, PRIO_MIN); _PRELE(bigproc); PROC_UNLOCK(bigproc); } } static bool vm_pageout_lowmem(void) { static int lowmem_ticks = 0; int last; last = atomic_load_int(&lowmem_ticks); while ((u_int)(ticks - last) / hz >= lowmem_period) { if (atomic_fcmpset_int(&lowmem_ticks, &last, ticks) == 0) continue; /* * Decrease registered cache sizes. */ SDT_PROBE0(vm, , , vm__lowmem_scan); EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_PAGES); /* * We do this explicitly after the caches have been * drained above. If we have a severe page shortage on * our hands, completely drain all UMA zones. Otherwise, * just prune the caches. */ uma_reclaim(vm_page_count_min() ? UMA_RECLAIM_DRAIN_CPU : UMA_RECLAIM_TRIM); return (true); } return (false); } static void vm_pageout_worker(void *arg) { struct vm_domain *vmd; u_int ofree; int addl_shortage, domain, shortage; bool target_met; domain = (uintptr_t)arg; vmd = VM_DOMAIN(domain); shortage = 0; target_met = true; /* * XXXKIB It could be useful to bind pageout daemon threads to * the cores belonging to the domain, from which vm_page_array * is allocated. */ KASSERT(vmd->vmd_segs != 0, ("domain without segments")); vmd->vmd_last_active_scan = ticks; /* * The pageout daemon worker is never done, so loop forever. */ while (TRUE) { vm_domain_pageout_lock(vmd); /* * We need to clear wanted before we check the limits. This * prevents races with wakers who will check wanted after they * reach the limit. */ atomic_store_int(&vmd->vmd_pageout_wanted, 0); /* * Might the page daemon need to run again? */ if (vm_paging_needed(vmd, vmd->vmd_free_count)) { /* * Yes. If the scan failed to produce enough free * pages, sleep uninterruptibly for some time in the * hope that the laundry thread will clean some pages. */ vm_domain_pageout_unlock(vmd); if (!target_met) pause("pwait", hz / VM_INACT_SCAN_RATE); } else { /* * No, sleep until the next wakeup or until pages * need to have their reference stats updated. */ if (mtx_sleep(&vmd->vmd_pageout_wanted, vm_domain_pageout_lockptr(vmd), PDROP | PVM, "psleep", hz / VM_INACT_SCAN_RATE) == 0) VM_CNT_INC(v_pdwakeups); } /* Prevent spurious wakeups by ensuring that wanted is set. */ atomic_store_int(&vmd->vmd_pageout_wanted, 1); /* * Use the controller to calculate how many pages to free in * this interval, and scan the inactive queue. If the lowmem * handlers appear to have freed up some pages, subtract the * difference from the inactive queue scan target. */ shortage = pidctrl_daemon(&vmd->vmd_pid, vmd->vmd_free_count); if (shortage > 0) { ofree = vmd->vmd_free_count; if (vm_pageout_lowmem() && vmd->vmd_free_count > ofree) shortage -= min(vmd->vmd_free_count - ofree, (u_int)shortage); target_met = vm_pageout_scan_inactive(vmd, shortage, &addl_shortage); } else addl_shortage = 0; /* * Scan the active queue. A positive value for shortage * indicates that we must aggressively deactivate pages to avoid * a shortfall. */ shortage = vm_pageout_active_target(vmd) + addl_shortage; vm_pageout_scan_active(vmd, shortage); } } /* * vm_pageout_init initialises basic pageout daemon settings. */ static void vm_pageout_init_domain(int domain) { struct vm_domain *vmd; struct sysctl_oid *oid; vmd = VM_DOMAIN(domain); vmd->vmd_interrupt_free_min = 2; /* * v_free_reserved needs to include enough for the largest * swap pager structures plus enough for any pv_entry structs * when paging. */ if (vmd->vmd_page_count > 1024) vmd->vmd_free_min = 4 + (vmd->vmd_page_count - 1024) / 200; else vmd->vmd_free_min = 4; vmd->vmd_pageout_free_min = 2 * MAXBSIZE / PAGE_SIZE + vmd->vmd_interrupt_free_min; vmd->vmd_free_reserved = vm_pageout_page_count + vmd->vmd_pageout_free_min + (vmd->vmd_page_count / 768); vmd->vmd_free_severe = vmd->vmd_free_min / 2; vmd->vmd_free_target = 4 * vmd->vmd_free_min + vmd->vmd_free_reserved; vmd->vmd_free_min += vmd->vmd_free_reserved; vmd->vmd_free_severe += vmd->vmd_free_reserved; vmd->vmd_inactive_target = (3 * vmd->vmd_free_target) / 2; if (vmd->vmd_inactive_target > vmd->vmd_free_count / 3) vmd->vmd_inactive_target = vmd->vmd_free_count / 3; /* * Set the default wakeup threshold to be 10% below the paging * target. This keeps the steady state out of shortfall. */ vmd->vmd_pageout_wakeup_thresh = (vmd->vmd_free_target / 10) * 9; /* * Target amount of memory to move out of the laundry queue during a * background laundering. This is proportional to the amount of system * memory. */ vmd->vmd_background_launder_target = (vmd->vmd_free_target - vmd->vmd_free_min) / 10; /* Initialize the pageout daemon pid controller. */ pidctrl_init(&vmd->vmd_pid, hz / VM_INACT_SCAN_RATE, vmd->vmd_free_target, PIDCTRL_BOUND, PIDCTRL_KPD, PIDCTRL_KID, PIDCTRL_KDD); oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(vmd->vmd_oid), OID_AUTO, "pidctrl", CTLFLAG_RD, NULL, ""); pidctrl_init_sysctl(&vmd->vmd_pid, SYSCTL_CHILDREN(oid)); } static void vm_pageout_init(void) { u_int freecount; int i; /* * Initialize some paging parameters. */ if (vm_cnt.v_page_count < 2000) vm_pageout_page_count = 8; freecount = 0; for (i = 0; i < vm_ndomains; i++) { struct vm_domain *vmd; vm_pageout_init_domain(i); vmd = VM_DOMAIN(i); vm_cnt.v_free_reserved += vmd->vmd_free_reserved; vm_cnt.v_free_target += vmd->vmd_free_target; vm_cnt.v_free_min += vmd->vmd_free_min; vm_cnt.v_inactive_target += vmd->vmd_inactive_target; vm_cnt.v_pageout_free_min += vmd->vmd_pageout_free_min; vm_cnt.v_interrupt_free_min += vmd->vmd_interrupt_free_min; vm_cnt.v_free_severe += vmd->vmd_free_severe; freecount += vmd->vmd_free_count; } /* * Set interval in seconds for active scan. We want to visit each * page at least once every ten minutes. This is to prevent worst * case paging behaviors with stale active LRU. */ if (vm_pageout_update_period == 0) vm_pageout_update_period = 600; if (vm_page_max_user_wired == 0) vm_page_max_user_wired = freecount / 3; } /* * vm_pageout is the high level pageout daemon. */ static void vm_pageout(void) { struct proc *p; struct thread *td; int error, first, i; p = curproc; td = curthread; mtx_init(&vm_oom_ratelim_mtx, "vmoomr", NULL, MTX_DEF); swap_pager_swap_init(); for (first = -1, i = 0; i < vm_ndomains; i++) { if (VM_DOMAIN_EMPTY(i)) { if (bootverbose) printf("domain %d empty; skipping pageout\n", i); continue; } if (first == -1) first = i; else { error = kthread_add(vm_pageout_worker, (void *)(uintptr_t)i, p, NULL, 0, 0, "dom%d", i); if (error != 0) panic("starting pageout for domain %d: %d\n", i, error); } error = kthread_add(vm_pageout_laundry_worker, (void *)(uintptr_t)i, p, NULL, 0, 0, "laundry: dom%d", i); if (error != 0) panic("starting laundry for domain %d: %d", i, error); } error = kthread_add(uma_reclaim_worker, NULL, p, NULL, 0, 0, "uma"); if (error != 0) panic("starting uma_reclaim helper, error %d\n", error); snprintf(td->td_name, sizeof(td->td_name), "dom%d", first); vm_pageout_worker((void *)(uintptr_t)first); } /* * Perform an advisory wakeup of the page daemon. */ void pagedaemon_wakeup(int domain) { struct vm_domain *vmd; vmd = VM_DOMAIN(domain); vm_domain_pageout_assert_unlocked(vmd); if (curproc == pageproc) return; if (atomic_fetchadd_int(&vmd->vmd_pageout_wanted, 1) == 0) { vm_domain_pageout_lock(vmd); atomic_store_int(&vmd->vmd_pageout_wanted, 1); wakeup(&vmd->vmd_pageout_wanted); vm_domain_pageout_unlock(vmd); } } Index: head/sys/vm/vm_swapout.c =================================================================== --- head/sys/vm/vm_swapout.c (revision 353538) +++ head/sys/vm/vm_swapout.c (revision 353539) @@ -1,966 +1,966 @@ /*- * SPDX-License-Identifier: (BSD-4-Clause AND MIT-CMU) * * Copyright (c) 1991 Regents of the University of California. * All rights reserved. * Copyright (c) 1994 John S. Dyson * All rights reserved. * Copyright (c) 1994 David Greenman * All rights reserved. * Copyright (c) 2005 Yahoo! Technologies Norway AS * All rights reserved. * * This code is derived from software contributed to Berkeley by * The Mach Operating System project at Carnegie-Mellon University. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. All advertising materials mentioning features or use of this software * must display the following acknowledgement: * This product includes software developed by the University of * California, Berkeley and its contributors. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91 * * * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Authors: Avadis Tevanian, Jr., Michael Wayne Young * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. */ #include __FBSDID("$FreeBSD$"); #include "opt_kstack_pages.h" #include "opt_kstack_max_pages.h" #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* the kernel process "vm_daemon" */ static void vm_daemon(void); static struct proc *vmproc; static struct kproc_desc vm_kp = { "vmdaemon", vm_daemon, &vmproc }; SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp); static int vm_swap_enabled = 1; static int vm_swap_idle_enabled = 0; SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); /* * Swap_idle_threshold1 is the guaranteed swapped in time for a process */ static int swap_idle_threshold1 = 2; SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW, &swap_idle_threshold1, 0, "Guaranteed swapped in time for a process"); /* * Swap_idle_threshold2 is the time that a process can be idle before * it will be swapped out, if idle swapping is enabled. */ static int swap_idle_threshold2 = 10; SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW, &swap_idle_threshold2, 0, "Time before a process will be swapped out"); static int vm_pageout_req_swapout; /* XXX */ static int vm_daemon_needed; static struct mtx vm_daemon_mtx; /* Allow for use by vm_pageout before vm_daemon is initialized. */ MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF); static int swapped_cnt; static int swap_inprogress; /* Pending swap-ins done outside swapper. */ static int last_swapin; static void swapclear(struct proc *); static int swapout(struct proc *); static void vm_swapout_map_deactivate_pages(vm_map_t, long); static void vm_swapout_object_deactivate_pages(pmap_t, vm_object_t, long); static void swapout_procs(int action); static void vm_req_vmdaemon(int req); static void vm_thread_swapout(struct thread *td); /* * vm_swapout_object_deactivate_pages * * Deactivate enough pages to satisfy the inactive target * requirements. * * The object and map must be locked. */ static void vm_swapout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object, long desired) { vm_object_t backing_object, object; vm_page_t p; int act_delta, remove_mode; VM_OBJECT_ASSERT_LOCKED(first_object); if ((first_object->flags & OBJ_FICTITIOUS) != 0) return; for (object = first_object;; object = backing_object) { if (pmap_resident_count(pmap) <= desired) goto unlock_return; VM_OBJECT_ASSERT_LOCKED(object); if ((object->flags & OBJ_UNMANAGED) != 0 || REFCOUNT_COUNT(object->paging_in_progress) > 0) goto unlock_return; remove_mode = 0; if (object->shadow_count > 1) remove_mode = 1; /* * Scan the object's entire memory queue. */ TAILQ_FOREACH(p, &object->memq, listq) { if (pmap_resident_count(pmap) <= desired) goto unlock_return; if (should_yield()) goto unlock_return; if (vm_page_tryxbusy(p) == 0) continue; VM_CNT_INC(v_pdpages); /* * The page may acquire a wiring after this check. * The page daemon handles wired pages, so there is * no harm done if a wiring appears while we are * attempting to deactivate the page. */ if (vm_page_wired(p) || !pmap_page_exists_quick(pmap, p)) { vm_page_xunbusy(p); continue; } act_delta = pmap_ts_referenced(p); vm_page_lock(p); if ((p->aflags & PGA_REFERENCED) != 0) { if (act_delta == 0) act_delta = 1; vm_page_aflag_clear(p, PGA_REFERENCED); } if (!vm_page_active(p) && act_delta != 0) { vm_page_activate(p); p->act_count += act_delta; } else if (vm_page_active(p)) { /* * The page daemon does not requeue pages * after modifying their activation count. */ if (act_delta == 0) { p->act_count -= min(p->act_count, ACT_DECLINE); if (!remove_mode && p->act_count == 0) { (void)vm_page_try_remove_all(p); vm_page_deactivate(p); } } else { vm_page_activate(p); if (p->act_count < ACT_MAX - ACT_ADVANCE) p->act_count += ACT_ADVANCE; } } else if (vm_page_inactive(p)) (void)vm_page_try_remove_all(p); vm_page_unlock(p); vm_page_xunbusy(p); } if ((backing_object = object->backing_object) == NULL) goto unlock_return; VM_OBJECT_RLOCK(backing_object); if (object != first_object) VM_OBJECT_RUNLOCK(object); } unlock_return: if (object != first_object) VM_OBJECT_RUNLOCK(object); } /* * deactivate some number of pages in a map, try to do it fairly, but * that is really hard to do. */ static void vm_swapout_map_deactivate_pages(vm_map_t map, long desired) { vm_map_entry_t tmpe; vm_object_t obj, bigobj; int nothingwired; if (!vm_map_trylock_read(map)) return; bigobj = NULL; nothingwired = TRUE; /* * first, search out the biggest object, and try to free pages from * that. */ VM_MAP_ENTRY_FOREACH(tmpe, map) { if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { obj = tmpe->object.vm_object; if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) { if (obj->shadow_count <= 1 && (bigobj == NULL || bigobj->resident_page_count < obj->resident_page_count)) { if (bigobj != NULL) VM_OBJECT_RUNLOCK(bigobj); bigobj = obj; } else VM_OBJECT_RUNLOCK(obj); } } if (tmpe->wired_count > 0) nothingwired = FALSE; } if (bigobj != NULL) { vm_swapout_object_deactivate_pages(map->pmap, bigobj, desired); VM_OBJECT_RUNLOCK(bigobj); } /* * Next, hunt around for other pages to deactivate. We actually * do this search sort of wrong -- .text first is not the best idea. */ VM_MAP_ENTRY_FOREACH(tmpe, map) { if (pmap_resident_count(vm_map_pmap(map)) <= desired) break; if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { obj = tmpe->object.vm_object; if (obj != NULL) { VM_OBJECT_RLOCK(obj); vm_swapout_object_deactivate_pages(map->pmap, obj, desired); VM_OBJECT_RUNLOCK(obj); } } } /* * Remove all mappings if a process is swapped out, this will free page * table pages. */ if (desired == 0 && nothingwired) { pmap_remove(vm_map_pmap(map), vm_map_min(map), vm_map_max(map)); } vm_map_unlock_read(map); } /* * Swap out requests */ #define VM_SWAP_NORMAL 1 #define VM_SWAP_IDLE 2 void vm_swapout_run(void) { if (vm_swap_enabled) vm_req_vmdaemon(VM_SWAP_NORMAL); } /* * Idle process swapout -- run once per second when pagedaemons are * reclaiming pages. */ void vm_swapout_run_idle(void) { static long lsec; if (!vm_swap_idle_enabled || time_second == lsec) return; vm_req_vmdaemon(VM_SWAP_IDLE); lsec = time_second; } static void vm_req_vmdaemon(int req) { static int lastrun = 0; mtx_lock(&vm_daemon_mtx); vm_pageout_req_swapout |= req; if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { wakeup(&vm_daemon_needed); lastrun = ticks; } mtx_unlock(&vm_daemon_mtx); } static void vm_daemon(void) { struct rlimit rsslim; struct proc *p; struct thread *td; struct vmspace *vm; int breakout, swapout_flags, tryagain, attempts; #ifdef RACCT uint64_t rsize, ravailable; #endif while (TRUE) { mtx_lock(&vm_daemon_mtx); msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", #ifdef RACCT racct_enable ? hz : 0 #else 0 #endif ); swapout_flags = vm_pageout_req_swapout; vm_pageout_req_swapout = 0; mtx_unlock(&vm_daemon_mtx); if (swapout_flags != 0) { /* * Drain the per-CPU page queue batches as a deadlock * avoidance measure. */ if ((swapout_flags & VM_SWAP_NORMAL) != 0) vm_page_pqbatch_drain(); swapout_procs(swapout_flags); } /* * scan the processes for exceeding their rlimits or if * process is swapped out -- deactivate pages */ tryagain = 0; attempts = 0; again: attempts++; sx_slock(&allproc_lock); FOREACH_PROC_IN_SYSTEM(p) { vm_pindex_t limit, size; /* * if this is a system process or if we have already * looked at this process, skip it. */ PROC_LOCK(p); if (p->p_state != PRS_NORMAL || p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) { PROC_UNLOCK(p); continue; } /* * if the process is in a non-running type state, * don't touch it. */ breakout = 0; FOREACH_THREAD_IN_PROC(p, td) { thread_lock(td); if (!TD_ON_RUNQ(td) && !TD_IS_RUNNING(td) && !TD_IS_SLEEPING(td) && !TD_IS_SUSPENDED(td)) { thread_unlock(td); breakout = 1; break; } thread_unlock(td); } if (breakout) { PROC_UNLOCK(p); continue; } /* * get a limit */ lim_rlimit_proc(p, RLIMIT_RSS, &rsslim); limit = OFF_TO_IDX( qmin(rsslim.rlim_cur, rsslim.rlim_max)); /* * let processes that are swapped out really be * swapped out set the limit to nothing (will force a * swap-out.) */ if ((p->p_flag & P_INMEM) == 0) limit = 0; /* XXX */ vm = vmspace_acquire_ref(p); _PHOLD_LITE(p); PROC_UNLOCK(p); if (vm == NULL) { PRELE(p); continue; } sx_sunlock(&allproc_lock); size = vmspace_resident_count(vm); if (size >= limit) { vm_swapout_map_deactivate_pages( &vm->vm_map, limit); size = vmspace_resident_count(vm); } #ifdef RACCT if (racct_enable) { rsize = IDX_TO_OFF(size); PROC_LOCK(p); if (p->p_state == PRS_NORMAL) racct_set(p, RACCT_RSS, rsize); ravailable = racct_get_available(p, RACCT_RSS); PROC_UNLOCK(p); if (rsize > ravailable) { /* * Don't be overly aggressive; this * might be an innocent process, * and the limit could've been exceeded * by some memory hog. Don't try * to deactivate more than 1/4th * of process' resident set size. */ if (attempts <= 8) { if (ravailable < rsize - (rsize / 4)) { ravailable = rsize - (rsize / 4); } } vm_swapout_map_deactivate_pages( &vm->vm_map, OFF_TO_IDX(ravailable)); /* Update RSS usage after paging out. */ size = vmspace_resident_count(vm); rsize = IDX_TO_OFF(size); PROC_LOCK(p); if (p->p_state == PRS_NORMAL) racct_set(p, RACCT_RSS, rsize); PROC_UNLOCK(p); if (rsize > ravailable) tryagain = 1; } } #endif vmspace_free(vm); sx_slock(&allproc_lock); PRELE(p); } sx_sunlock(&allproc_lock); if (tryagain != 0 && attempts <= 10) { maybe_yield(); goto again; } } } /* * Allow a thread's kernel stack to be paged out. */ static void vm_thread_swapout(struct thread *td) { vm_object_t ksobj; vm_page_t m; int i, pages; cpu_thread_swapout(td); pages = td->td_kstack_pages; ksobj = td->td_kstack_obj; pmap_qremove(td->td_kstack, pages); VM_OBJECT_WLOCK(ksobj); for (i = 0; i < pages; i++) { m = vm_page_lookup(ksobj, i); if (m == NULL) panic("vm_thread_swapout: kstack already missing?"); vm_page_dirty(m); vm_page_unwire(m, PQ_LAUNDRY); } VM_OBJECT_WUNLOCK(ksobj); } /* * Bring the kernel stack for a specified thread back in. */ static void vm_thread_swapin(struct thread *td, int oom_alloc) { vm_object_t ksobj; vm_page_t ma[KSTACK_MAX_PAGES]; int a, count, i, j, pages, rv; pages = td->td_kstack_pages; ksobj = td->td_kstack_obj; VM_OBJECT_WLOCK(ksobj); (void)vm_page_grab_pages(ksobj, 0, oom_alloc | VM_ALLOC_WIRED, ma, pages); for (i = 0; i < pages;) { vm_page_assert_xbusied(ma[i]); - if (ma[i]->valid == VM_PAGE_BITS_ALL) { + if (vm_page_all_valid(ma[i])) { vm_page_xunbusy(ma[i]); i++; continue; } vm_object_pip_add(ksobj, 1); for (j = i + 1; j < pages; j++) - if (ma[j]->valid == VM_PAGE_BITS_ALL) + if (vm_page_all_valid(ma[j])) break; rv = vm_pager_has_page(ksobj, ma[i]->pindex, NULL, &a); KASSERT(rv == 1, ("%s: missing page %p", __func__, ma[i])); count = min(a + 1, j - i); rv = vm_pager_get_pages(ksobj, ma + i, count, NULL, NULL); KASSERT(rv == VM_PAGER_OK, ("%s: cannot get kstack for proc %d", __func__, td->td_proc->p_pid)); vm_object_pip_wakeup(ksobj); for (j = i; j < i + count; j++) vm_page_xunbusy(ma[j]); i += count; } VM_OBJECT_WUNLOCK(ksobj); pmap_qenter(td->td_kstack, ma, pages); cpu_thread_swapin(td); } void faultin(struct proc *p) { struct thread *td; int oom_alloc; PROC_LOCK_ASSERT(p, MA_OWNED); /* * If another process is swapping in this process, * just wait until it finishes. */ if (p->p_flag & P_SWAPPINGIN) { while (p->p_flag & P_SWAPPINGIN) msleep(&p->p_flag, &p->p_mtx, PVM, "faultin", 0); return; } if ((p->p_flag & P_INMEM) == 0) { oom_alloc = (p->p_flag & P_WKILLED) != 0 ? VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL; /* * Don't let another thread swap process p out while we are * busy swapping it in. */ ++p->p_lock; p->p_flag |= P_SWAPPINGIN; PROC_UNLOCK(p); sx_xlock(&allproc_lock); MPASS(swapped_cnt > 0); swapped_cnt--; if (curthread != &thread0) swap_inprogress++; sx_xunlock(&allproc_lock); /* * We hold no lock here because the list of threads * can not change while all threads in the process are * swapped out. */ FOREACH_THREAD_IN_PROC(p, td) vm_thread_swapin(td, oom_alloc); if (curthread != &thread0) { sx_xlock(&allproc_lock); MPASS(swap_inprogress > 0); swap_inprogress--; last_swapin = ticks; sx_xunlock(&allproc_lock); } PROC_LOCK(p); swapclear(p); p->p_swtick = ticks; /* Allow other threads to swap p out now. */ wakeup(&p->p_flag); --p->p_lock; } } /* * This swapin algorithm attempts to swap-in processes only if there * is enough space for them. Of course, if a process waits for a long * time, it will be swapped in anyway. */ static struct proc * swapper_selector(bool wkilled_only) { struct proc *p, *res; struct thread *td; int ppri, pri, slptime, swtime; sx_assert(&allproc_lock, SA_SLOCKED); if (swapped_cnt == 0) return (NULL); res = NULL; ppri = INT_MIN; FOREACH_PROC_IN_SYSTEM(p) { PROC_LOCK(p); if (p->p_state == PRS_NEW || (p->p_flag & (P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) != 0) { PROC_UNLOCK(p); continue; } if (p->p_state == PRS_NORMAL && (p->p_flag & P_WKILLED) != 0) { /* * A swapped-out process might have mapped a * large portion of the system's pages as * anonymous memory. There is no other way to * release the memory other than to kill the * process, for which we need to swap it in. */ return (p); } if (wkilled_only) { PROC_UNLOCK(p); continue; } swtime = (ticks - p->p_swtick) / hz; FOREACH_THREAD_IN_PROC(p, td) { /* * An otherwise runnable thread of a process * swapped out has only the TDI_SWAPPED bit set. */ thread_lock(td); if (td->td_inhibitors == TDI_SWAPPED) { slptime = (ticks - td->td_slptick) / hz; pri = swtime + slptime; if ((td->td_flags & TDF_SWAPINREQ) == 0) pri -= p->p_nice * 8; /* * if this thread is higher priority * and there is enough space, then select * this process instead of the previous * selection. */ if (pri > ppri) { res = p; ppri = pri; } } thread_unlock(td); } PROC_UNLOCK(p); } if (res != NULL) PROC_LOCK(res); return (res); } #define SWAPIN_INTERVAL (MAXSLP * hz / 2) /* * Limit swapper to swap in one non-WKILLED process in MAXSLP/2 * interval, assuming that there is: * - at least one domain that is not suffering from a shortage of free memory; * - no parallel swap-ins; * - no other swap-ins in the current SWAPIN_INTERVAL. */ static bool swapper_wkilled_only(void) { return (vm_page_count_min_set(&all_domains) || swap_inprogress > 0 || (u_int)(ticks - last_swapin) < SWAPIN_INTERVAL); } void swapper(void) { struct proc *p; for (;;) { sx_slock(&allproc_lock); p = swapper_selector(swapper_wkilled_only()); sx_sunlock(&allproc_lock); if (p == NULL) { tsleep(&proc0, PVM, "swapin", SWAPIN_INTERVAL); } else { PROC_LOCK_ASSERT(p, MA_OWNED); /* * Another process may be bringing or may have * already brought this process in while we * traverse all threads. Or, this process may * have exited or even being swapped out * again. */ if (p->p_state == PRS_NORMAL && (p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) == 0) { faultin(p); } PROC_UNLOCK(p); } } } /* * First, if any processes have been sleeping or stopped for at least * "swap_idle_threshold1" seconds, they are swapped out. If, however, * no such processes exist, then the longest-sleeping or stopped * process is swapped out. Finally, and only as a last resort, if * there are no sleeping or stopped processes, the longest-resident * process is swapped out. */ static void swapout_procs(int action) { struct proc *p; struct thread *td; int slptime; bool didswap, doswap; MPASS((action & (VM_SWAP_NORMAL | VM_SWAP_IDLE)) != 0); didswap = false; sx_slock(&allproc_lock); FOREACH_PROC_IN_SYSTEM(p) { /* * Filter out not yet fully constructed processes. Do * not swap out held processes. Avoid processes which * are system, exiting, execing, traced, already swapped * out or are in the process of being swapped in or out. */ PROC_LOCK(p); if (p->p_state != PRS_NORMAL || p->p_lock != 0 || (p->p_flag & (P_SYSTEM | P_WEXIT | P_INEXEC | P_STOPPED_SINGLE | P_TRACED | P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) != P_INMEM) { PROC_UNLOCK(p); continue; } /* * Further consideration of this process for swap out * requires iterating over its threads. We release * allproc_lock here so that process creation and * destruction are not blocked while we iterate. * * To later reacquire allproc_lock and resume * iteration over the allproc list, we will first have * to release the lock on the process. We place a * hold on the process so that it remains in the * allproc list while it is unlocked. */ _PHOLD_LITE(p); sx_sunlock(&allproc_lock); /* * Do not swapout a realtime process. * Guarantee swap_idle_threshold1 time in memory. * If the system is under memory stress, or if we are * swapping idle processes >= swap_idle_threshold2, * then swap the process out. */ doswap = true; FOREACH_THREAD_IN_PROC(p, td) { thread_lock(td); slptime = (ticks - td->td_slptick) / hz; if (PRI_IS_REALTIME(td->td_pri_class) || slptime < swap_idle_threshold1 || !thread_safetoswapout(td) || ((action & VM_SWAP_NORMAL) == 0 && slptime < swap_idle_threshold2)) doswap = false; thread_unlock(td); if (!doswap) break; } if (doswap && swapout(p) == 0) didswap = true; PROC_UNLOCK(p); if (didswap) { sx_xlock(&allproc_lock); swapped_cnt++; sx_downgrade(&allproc_lock); } else sx_slock(&allproc_lock); PRELE(p); } sx_sunlock(&allproc_lock); /* * If we swapped something out, and another process needed memory, * then wakeup the sched process. */ if (didswap) wakeup(&proc0); } static void swapclear(struct proc *p) { struct thread *td; PROC_LOCK_ASSERT(p, MA_OWNED); FOREACH_THREAD_IN_PROC(p, td) { thread_lock(td); td->td_flags |= TDF_INMEM; td->td_flags &= ~TDF_SWAPINREQ; TD_CLR_SWAPPED(td); if (TD_CAN_RUN(td)) if (setrunnable(td)) { #ifdef INVARIANTS /* * XXX: We just cleared TDI_SWAPPED * above and set TDF_INMEM, so this * should never happen. */ panic("not waking up swapper"); #endif } thread_unlock(td); } p->p_flag &= ~(P_SWAPPINGIN | P_SWAPPINGOUT); p->p_flag |= P_INMEM; } static int swapout(struct proc *p) { struct thread *td; PROC_LOCK_ASSERT(p, MA_OWNED); /* * The states of this process and its threads may have changed * by now. Assuming that there is only one pageout daemon thread, * this process should still be in memory. */ KASSERT((p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) == P_INMEM, ("swapout: lost a swapout race?")); /* * Remember the resident count. */ p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace); /* * Check and mark all threads before we proceed. */ p->p_flag &= ~P_INMEM; p->p_flag |= P_SWAPPINGOUT; FOREACH_THREAD_IN_PROC(p, td) { thread_lock(td); if (!thread_safetoswapout(td)) { thread_unlock(td); swapclear(p); return (EBUSY); } td->td_flags &= ~TDF_INMEM; TD_SET_SWAPPED(td); thread_unlock(td); } td = FIRST_THREAD_IN_PROC(p); ++td->td_ru.ru_nswap; PROC_UNLOCK(p); /* * This list is stable because all threads are now prevented from * running. The list is only modified in the context of a running * thread in this process. */ FOREACH_THREAD_IN_PROC(p, td) vm_thread_swapout(td); PROC_LOCK(p); p->p_flag &= ~P_SWAPPINGOUT; p->p_swtick = ticks; return (0); } Index: head/sys/vm/vnode_pager.c =================================================================== --- head/sys/vm/vnode_pager.c (revision 353538) +++ head/sys/vm/vnode_pager.c (revision 353539) @@ -1,1586 +1,1591 @@ /*- * SPDX-License-Identifier: BSD-4-Clause * * Copyright (c) 1990 University of Utah. * Copyright (c) 1991 The Regents of the University of California. * All rights reserved. * Copyright (c) 1993, 1994 John S. Dyson * Copyright (c) 1995, David Greenman * * This code is derived from software contributed to Berkeley by * the Systems Programming Group of the University of Utah Computer * Science Department. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. All advertising materials mentioning features or use of this software * must display the following acknowledgement: * This product includes software developed by the University of * California, Berkeley and its contributors. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)vnode_pager.c 7.5 (Berkeley) 4/20/91 */ /* * Page to/from files (vnodes). */ /* * TODO: * Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will * greatly re-simplify the vnode_pager. */ #include __FBSDID("$FreeBSD$"); #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include static int vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress, int *run); static int vnode_pager_input_smlfs(vm_object_t object, vm_page_t m); static int vnode_pager_input_old(vm_object_t object, vm_page_t m); static void vnode_pager_dealloc(vm_object_t); static int vnode_pager_getpages(vm_object_t, vm_page_t *, int, int *, int *); static int vnode_pager_getpages_async(vm_object_t, vm_page_t *, int, int *, int *, vop_getpages_iodone_t, void *); static void vnode_pager_putpages(vm_object_t, vm_page_t *, int, int, int *); static boolean_t vnode_pager_haspage(vm_object_t, vm_pindex_t, int *, int *); static vm_object_t vnode_pager_alloc(void *, vm_ooffset_t, vm_prot_t, vm_ooffset_t, struct ucred *cred); static int vnode_pager_generic_getpages_done(struct buf *); static void vnode_pager_generic_getpages_done_async(struct buf *); static void vnode_pager_update_writecount(vm_object_t, vm_offset_t, vm_offset_t); static void vnode_pager_release_writecount(vm_object_t, vm_offset_t, vm_offset_t); struct pagerops vnodepagerops = { .pgo_alloc = vnode_pager_alloc, .pgo_dealloc = vnode_pager_dealloc, .pgo_getpages = vnode_pager_getpages, .pgo_getpages_async = vnode_pager_getpages_async, .pgo_putpages = vnode_pager_putpages, .pgo_haspage = vnode_pager_haspage, .pgo_update_writecount = vnode_pager_update_writecount, .pgo_release_writecount = vnode_pager_release_writecount, }; static struct domainset *vnode_domainset = NULL; SYSCTL_PROC(_debug, OID_AUTO, vnode_domainset, CTLTYPE_STRING | CTLFLAG_RW, &vnode_domainset, 0, sysctl_handle_domainset, "A", "Default vnode NUMA policy"); static int nvnpbufs; SYSCTL_INT(_vm, OID_AUTO, vnode_pbufs, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &nvnpbufs, 0, "number of physical buffers allocated for vnode pager"); static uma_zone_t vnode_pbuf_zone; static void vnode_pager_init(void *dummy) { #ifdef __LP64__ nvnpbufs = nswbuf * 2; #else nvnpbufs = nswbuf / 2; #endif TUNABLE_INT_FETCH("vm.vnode_pbufs", &nvnpbufs); vnode_pbuf_zone = pbuf_zsecond_create("vnpbuf", nvnpbufs); } SYSINIT(vnode_pager, SI_SUB_CPU, SI_ORDER_ANY, vnode_pager_init, NULL); /* Create the VM system backing object for this vnode */ int vnode_create_vobject(struct vnode *vp, off_t isize, struct thread *td) { vm_object_t object; vm_ooffset_t size = isize; struct vattr va; if (!vn_isdisk(vp, NULL) && vn_canvmio(vp) == FALSE) return (0); object = vp->v_object; if (object != NULL) return (0); if (size == 0) { if (vn_isdisk(vp, NULL)) { size = IDX_TO_OFF(INT_MAX); } else { if (VOP_GETATTR(vp, &va, td->td_ucred)) return (0); size = va.va_size; } } object = vnode_pager_alloc(vp, size, 0, 0, td->td_ucred); /* * Dereference the reference we just created. This assumes * that the object is associated with the vp. */ VM_OBJECT_WLOCK(object); object->ref_count--; VM_OBJECT_WUNLOCK(object); vrele(vp); KASSERT(vp->v_object != NULL, ("vnode_create_vobject: NULL object")); return (0); } void vnode_destroy_vobject(struct vnode *vp) { struct vm_object *obj; obj = vp->v_object; if (obj == NULL || obj->handle != vp) return; ASSERT_VOP_ELOCKED(vp, "vnode_destroy_vobject"); VM_OBJECT_WLOCK(obj); MPASS(obj->type == OBJT_VNODE); umtx_shm_object_terminated(obj); if (obj->ref_count == 0) { /* * don't double-terminate the object */ if ((obj->flags & OBJ_DEAD) == 0) { vm_object_set_flag(obj, OBJ_DEAD); /* * Clean pages and flush buffers. */ vm_object_page_clean(obj, 0, 0, OBJPC_SYNC); VM_OBJECT_WUNLOCK(obj); vinvalbuf(vp, V_SAVE, 0, 0); BO_LOCK(&vp->v_bufobj); vp->v_bufobj.bo_flag |= BO_DEAD; BO_UNLOCK(&vp->v_bufobj); VM_OBJECT_WLOCK(obj); vm_object_terminate(obj); } else { /* * Waiters were already handled during object * termination. The exclusive vnode lock hopefully * prevented new waiters from referencing the dying * object. */ vp->v_object = NULL; VM_OBJECT_WUNLOCK(obj); } } else { /* * Woe to the process that tries to page now :-). */ vm_pager_deallocate(obj); VM_OBJECT_WUNLOCK(obj); } KASSERT(vp->v_object == NULL, ("vp %p obj %p", vp, vp->v_object)); } /* * Allocate (or lookup) pager for a vnode. * Handle is a vnode pointer. */ vm_object_t vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t offset, struct ucred *cred) { vm_object_t object; struct vnode *vp; /* * Pageout to vnode, no can do yet. */ if (handle == NULL) return (NULL); vp = (struct vnode *)handle; ASSERT_VOP_LOCKED(vp, "vnode_pager_alloc"); KASSERT(vp->v_usecount != 0, ("vnode_pager_alloc: no vnode reference")); retry: object = vp->v_object; if (object == NULL) { /* * Add an object of the appropriate size */ object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size))); object->un_pager.vnp.vnp_size = size; object->un_pager.vnp.writemappings = 0; object->domain.dr_policy = vnode_domainset; object->handle = handle; VI_LOCK(vp); if (vp->v_object != NULL) { /* * Object has been created while we were allocating. */ VI_UNLOCK(vp); VM_OBJECT_WLOCK(object); KASSERT(object->ref_count == 1, ("leaked ref %p %d", object, object->ref_count)); object->type = OBJT_DEAD; object->ref_count = 0; VM_OBJECT_WUNLOCK(object); vm_object_destroy(object); goto retry; } vp->v_object = object; VI_UNLOCK(vp); } else { VM_OBJECT_WLOCK(object); object->ref_count++; #if VM_NRESERVLEVEL > 0 vm_object_color(object, 0); #endif VM_OBJECT_WUNLOCK(object); } vrefact(vp); return (object); } /* * The object must be locked. */ static void vnode_pager_dealloc(vm_object_t object) { struct vnode *vp; int refs; vp = object->handle; if (vp == NULL) panic("vnode_pager_dealloc: pager already dealloced"); VM_OBJECT_ASSERT_WLOCKED(object); vm_object_pip_wait(object, "vnpdea"); refs = object->ref_count; object->handle = NULL; object->type = OBJT_DEAD; ASSERT_VOP_ELOCKED(vp, "vnode_pager_dealloc"); if (object->un_pager.vnp.writemappings > 0) { object->un_pager.vnp.writemappings = 0; VOP_ADD_WRITECOUNT_CHECKED(vp, -1); CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d", __func__, vp, vp->v_writecount); } vp->v_object = NULL; VI_LOCK(vp); /* * vm_map_entry_set_vnode_text() cannot reach this vnode by * following object->handle. Clear all text references now. * This also clears the transient references from * kern_execve(), which is fine because dead_vnodeops uses nop * for VOP_UNSET_TEXT(). */ if (vp->v_writecount < 0) vp->v_writecount = 0; VI_UNLOCK(vp); VM_OBJECT_WUNLOCK(object); while (refs-- > 0) vunref(vp); VM_OBJECT_WLOCK(object); } static boolean_t vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after) { struct vnode *vp = object->handle; daddr_t bn; uintptr_t lockstate; int err; daddr_t reqblock; int poff; int bsize; int pagesperblock, blocksperpage; VM_OBJECT_ASSERT_LOCKED(object); /* * If no vp or vp is doomed or marked transparent to VM, we do not * have the page. */ if (vp == NULL || vp->v_iflag & VI_DOOMED) return FALSE; /* * If the offset is beyond end of file we do * not have the page. */ if (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size) return FALSE; bsize = vp->v_mount->mnt_stat.f_iosize; pagesperblock = bsize / PAGE_SIZE; blocksperpage = 0; if (pagesperblock > 0) { reqblock = pindex / pagesperblock; } else { blocksperpage = (PAGE_SIZE / bsize); reqblock = pindex * blocksperpage; } lockstate = VM_OBJECT_DROP(object); err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before); VM_OBJECT_PICKUP(object, lockstate); if (err) return TRUE; if (bn == -1) return FALSE; if (pagesperblock > 0) { poff = pindex - (reqblock * pagesperblock); if (before) { *before *= pagesperblock; *before += poff; } if (after) { /* * The BMAP vop can report a partial block in the * 'after', but must not report blocks after EOF. * Assert the latter, and truncate 'after' in case * of the former. */ KASSERT((reqblock + *after) * pagesperblock < roundup2(object->size, pagesperblock), ("%s: reqblock %jd after %d size %ju", __func__, (intmax_t )reqblock, *after, (uintmax_t )object->size)); *after *= pagesperblock; *after += pagesperblock - (poff + 1); if (pindex + *after >= object->size) *after = object->size - 1 - pindex; } } else { if (before) { *before /= blocksperpage; } if (after) { *after /= blocksperpage; } } return TRUE; } /* * Lets the VM system know about a change in size for a file. * We adjust our own internal size and flush any cached pages in * the associated object that are affected by the size change. * * Note: this routine may be invoked as a result of a pager put * operation (possibly at object termination time), so we must be careful. */ void vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize) { vm_object_t object; vm_page_t m; vm_pindex_t nobjsize; if ((object = vp->v_object) == NULL) return; /* ASSERT_VOP_ELOCKED(vp, "vnode_pager_setsize and not locked vnode"); */ VM_OBJECT_WLOCK(object); if (object->type == OBJT_DEAD) { VM_OBJECT_WUNLOCK(object); return; } KASSERT(object->type == OBJT_VNODE, ("not vnode-backed object %p", object)); if (nsize == object->un_pager.vnp.vnp_size) { /* * Hasn't changed size */ VM_OBJECT_WUNLOCK(object); return; } nobjsize = OFF_TO_IDX(nsize + PAGE_MASK); if (nsize < object->un_pager.vnp.vnp_size) { /* * File has shrunk. Toss any cached pages beyond the new EOF. */ if (nobjsize < object->size) vm_object_page_remove(object, nobjsize, object->size, 0); /* * this gets rid of garbage at the end of a page that is now * only partially backed by the vnode. * * XXX for some reason (I don't know yet), if we take a * completely invalid page and mark it partially valid * it can screw up NFS reads, so we don't allow the case. */ - if ((nsize & PAGE_MASK) && - (m = vm_page_lookup(object, OFF_TO_IDX(nsize))) != NULL && - m->valid != 0) { + if (!(nsize & PAGE_MASK)) + goto out; + m = vm_page_grab(object, OFF_TO_IDX(nsize), VM_ALLOC_NOCREAT); + if (m == NULL) + goto out; + if (!vm_page_none_valid(m)) { int base = (int)nsize & PAGE_MASK; int size = PAGE_SIZE - base; /* * Clear out partial-page garbage in case * the page has been mapped. */ pmap_zero_page_area(m, base, size); /* * Update the valid bits to reflect the blocks that * have been zeroed. Some of these valid bits may * have already been set. */ vm_page_set_valid_range(m, base, size); /* * Round "base" to the next block boundary so that the * dirty bit for a partially zeroed block is not * cleared. */ base = roundup2(base, DEV_BSIZE); /* * Clear out partial-page dirty bits. * * note that we do not clear out the valid * bits. This would prevent bogus_page * replacement from working properly. */ vm_page_clear_dirty(m, base, PAGE_SIZE - base); } + vm_page_xunbusy(m); } +out: object->un_pager.vnp.vnp_size = nsize; object->size = nobjsize; VM_OBJECT_WUNLOCK(object); } /* * calculate the linear (byte) disk address of specified virtual * file address */ static int vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress, int *run) { int bsize; int err; daddr_t vblock; daddr_t voffset; if (address < 0) return -1; if (vp->v_iflag & VI_DOOMED) return -1; bsize = vp->v_mount->mnt_stat.f_iosize; vblock = address / bsize; voffset = address % bsize; err = VOP_BMAP(vp, vblock, NULL, rtaddress, run, NULL); if (err == 0) { if (*rtaddress != -1) *rtaddress += voffset / DEV_BSIZE; if (run) { *run += 1; *run *= bsize / PAGE_SIZE; *run -= voffset / PAGE_SIZE; } } return (err); } /* * small block filesystem vnode pager input */ static int vnode_pager_input_smlfs(vm_object_t object, vm_page_t m) { struct vnode *vp; struct bufobj *bo; struct buf *bp; struct sf_buf *sf; daddr_t fileaddr; vm_offset_t bsize; vm_page_bits_t bits; int error, i; error = 0; vp = object->handle; if (vp->v_iflag & VI_DOOMED) return VM_PAGER_BAD; bsize = vp->v_mount->mnt_stat.f_iosize; VOP_BMAP(vp, 0, &bo, 0, NULL, NULL); sf = sf_buf_alloc(m, 0); for (i = 0; i < PAGE_SIZE / bsize; i++) { vm_ooffset_t address; bits = vm_page_bits(i * bsize, bsize); if (m->valid & bits) continue; address = IDX_TO_OFF(m->pindex) + i * bsize; if (address >= object->un_pager.vnp.vnp_size) { fileaddr = -1; } else { error = vnode_pager_addr(vp, address, &fileaddr, NULL); if (error) break; } if (fileaddr != -1) { bp = uma_zalloc(vnode_pbuf_zone, M_WAITOK); /* build a minimal buffer header */ bp->b_iocmd = BIO_READ; bp->b_iodone = bdone; KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred")); KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred")); bp->b_rcred = crhold(curthread->td_ucred); bp->b_wcred = crhold(curthread->td_ucred); bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize; bp->b_blkno = fileaddr; pbgetbo(bo, bp); bp->b_vp = vp; bp->b_bcount = bsize; bp->b_bufsize = bsize; bp->b_runningbufspace = bp->b_bufsize; atomic_add_long(&runningbufspace, bp->b_runningbufspace); /* do the input */ bp->b_iooffset = dbtob(bp->b_blkno); bstrategy(bp); bwait(bp, PVM, "vnsrd"); if ((bp->b_ioflags & BIO_ERROR) != 0) error = EIO; /* * free the buffer header back to the swap buffer pool */ bp->b_vp = NULL; pbrelbo(bp); uma_zfree(vnode_pbuf_zone, bp); if (error) break; } else bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize); KASSERT((m->dirty & bits) == 0, ("vnode_pager_input_smlfs: page %p is dirty", m)); VM_OBJECT_WLOCK(object); m->valid |= bits; VM_OBJECT_WUNLOCK(object); } sf_buf_free(sf); if (error) { return VM_PAGER_ERROR; } return VM_PAGER_OK; } /* * old style vnode pager input routine */ static int vnode_pager_input_old(vm_object_t object, vm_page_t m) { struct uio auio; struct iovec aiov; int error; int size; struct sf_buf *sf; struct vnode *vp; VM_OBJECT_ASSERT_WLOCKED(object); error = 0; /* * Return failure if beyond current EOF */ if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) { return VM_PAGER_BAD; } else { size = PAGE_SIZE; if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size) size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex); vp = object->handle; VM_OBJECT_WUNLOCK(object); /* * Allocate a kernel virtual address and initialize so that * we can use VOP_READ/WRITE routines. */ sf = sf_buf_alloc(m, 0); aiov.iov_base = (caddr_t)sf_buf_kva(sf); aiov.iov_len = size; auio.uio_iov = &aiov; auio.uio_iovcnt = 1; auio.uio_offset = IDX_TO_OFF(m->pindex); auio.uio_segflg = UIO_SYSSPACE; auio.uio_rw = UIO_READ; auio.uio_resid = size; auio.uio_td = curthread; error = VOP_READ(vp, &auio, 0, curthread->td_ucred); if (!error) { int count = size - auio.uio_resid; if (count == 0) error = EINVAL; else if (count != PAGE_SIZE) bzero((caddr_t)sf_buf_kva(sf) + count, PAGE_SIZE - count); } sf_buf_free(sf); VM_OBJECT_WLOCK(object); } KASSERT(m->dirty == 0, ("vnode_pager_input_old: page %p is dirty", m)); if (!error) - m->valid = VM_PAGE_BITS_ALL; + vm_page_valid(m); return error ? VM_PAGER_ERROR : VM_PAGER_OK; } /* * generic vnode pager input routine */ /* * Local media VFS's that do not implement their own VOP_GETPAGES * should have their VOP_GETPAGES call to vnode_pager_generic_getpages() * to implement the previous behaviour. * * All other FS's should use the bypass to get to the local media * backing vp's VOP_GETPAGES. */ static int vnode_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind, int *rahead) { struct vnode *vp; int rtval; vp = object->handle; VM_OBJECT_WUNLOCK(object); rtval = VOP_GETPAGES(vp, m, count, rbehind, rahead); KASSERT(rtval != EOPNOTSUPP, ("vnode_pager: FS getpages not implemented\n")); VM_OBJECT_WLOCK(object); return rtval; } static int vnode_pager_getpages_async(vm_object_t object, vm_page_t *m, int count, int *rbehind, int *rahead, vop_getpages_iodone_t iodone, void *arg) { struct vnode *vp; int rtval; vp = object->handle; VM_OBJECT_WUNLOCK(object); rtval = VOP_GETPAGES_ASYNC(vp, m, count, rbehind, rahead, iodone, arg); KASSERT(rtval != EOPNOTSUPP, ("vnode_pager: FS getpages_async not implemented\n")); VM_OBJECT_WLOCK(object); return (rtval); } /* * The implementation of VOP_GETPAGES() and VOP_GETPAGES_ASYNC() for * local filesystems, where partially valid pages can only occur at * the end of file. */ int vnode_pager_local_getpages(struct vop_getpages_args *ap) { return (vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count, ap->a_rbehind, ap->a_rahead, NULL, NULL)); } int vnode_pager_local_getpages_async(struct vop_getpages_async_args *ap) { return (vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count, ap->a_rbehind, ap->a_rahead, ap->a_iodone, ap->a_arg)); } /* * This is now called from local media FS's to operate against their * own vnodes if they fail to implement VOP_GETPAGES. */ int vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *m, int count, int *a_rbehind, int *a_rahead, vop_getpages_iodone_t iodone, void *arg) { vm_object_t object; struct bufobj *bo; struct buf *bp; off_t foff; #ifdef INVARIANTS off_t blkno0; #endif int bsize, pagesperblock; int error, before, after, rbehind, rahead, poff, i; int bytecount, secmask; KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, ("%s does not support devices", __func__)); if (vp->v_iflag & VI_DOOMED) return (VM_PAGER_BAD); object = vp->v_object; foff = IDX_TO_OFF(m[0]->pindex); bsize = vp->v_mount->mnt_stat.f_iosize; pagesperblock = bsize / PAGE_SIZE; KASSERT(foff < object->un_pager.vnp.vnp_size, ("%s: page %p offset beyond vp %p size", __func__, m[0], vp)); KASSERT(count <= nitems(bp->b_pages), ("%s: requested %d pages", __func__, count)); /* * The last page has valid blocks. Invalid part can only * exist at the end of file, and the page is made fully valid * by zeroing in vm_pager_get_pages(). */ - if (m[count - 1]->valid != 0 && --count == 0) { + if (!vm_page_none_valid(m[count - 1]) && --count == 0) { if (iodone != NULL) iodone(arg, m, 1, 0); return (VM_PAGER_OK); } bp = uma_zalloc(vnode_pbuf_zone, M_WAITOK); /* * Get the underlying device blocks for the file with VOP_BMAP(). * If the file system doesn't support VOP_BMAP, use old way of * getting pages via VOP_READ. */ error = VOP_BMAP(vp, foff / bsize, &bo, &bp->b_blkno, &after, &before); if (error == EOPNOTSUPP) { uma_zfree(vnode_pbuf_zone, bp); VM_OBJECT_WLOCK(object); for (i = 0; i < count; i++) { VM_CNT_INC(v_vnodein); VM_CNT_INC(v_vnodepgsin); error = vnode_pager_input_old(object, m[i]); if (error) break; } VM_OBJECT_WUNLOCK(object); return (error); } else if (error != 0) { uma_zfree(vnode_pbuf_zone, bp); return (VM_PAGER_ERROR); } /* * If the file system supports BMAP, but blocksize is smaller * than a page size, then use special small filesystem code. */ if (pagesperblock == 0) { uma_zfree(vnode_pbuf_zone, bp); for (i = 0; i < count; i++) { VM_CNT_INC(v_vnodein); VM_CNT_INC(v_vnodepgsin); error = vnode_pager_input_smlfs(object, m[i]); if (error) break; } return (error); } /* * A sparse file can be encountered only for a single page request, * which may not be preceded by call to vm_pager_haspage(). */ if (bp->b_blkno == -1) { KASSERT(count == 1, ("%s: array[%d] request to a sparse file %p", __func__, count, vp)); uma_zfree(vnode_pbuf_zone, bp); pmap_zero_page(m[0]); KASSERT(m[0]->dirty == 0, ("%s: page %p is dirty", __func__, m[0])); VM_OBJECT_WLOCK(object); - m[0]->valid = VM_PAGE_BITS_ALL; + vm_page_valid(m[0]); VM_OBJECT_WUNLOCK(object); return (VM_PAGER_OK); } #ifdef INVARIANTS blkno0 = bp->b_blkno; #endif bp->b_blkno += (foff % bsize) / DEV_BSIZE; /* Recalculate blocks available after/before to pages. */ poff = (foff % bsize) / PAGE_SIZE; before *= pagesperblock; before += poff; after *= pagesperblock; after += pagesperblock - (poff + 1); if (m[0]->pindex + after >= object->size) after = object->size - 1 - m[0]->pindex; KASSERT(count <= after + 1, ("%s: %d pages asked, can do only %d", __func__, count, after + 1)); after -= count - 1; /* Trim requested rbehind/rahead to possible values. */ rbehind = a_rbehind ? *a_rbehind : 0; rahead = a_rahead ? *a_rahead : 0; rbehind = min(rbehind, before); rbehind = min(rbehind, m[0]->pindex); rahead = min(rahead, after); rahead = min(rahead, object->size - m[count - 1]->pindex); /* * Check that total amount of pages fit into buf. Trim rbehind and * rahead evenly if not. */ if (rbehind + rahead + count > nitems(bp->b_pages)) { int trim, sum; trim = rbehind + rahead + count - nitems(bp->b_pages) + 1; sum = rbehind + rahead; if (rbehind == before) { /* Roundup rbehind trim to block size. */ rbehind -= roundup(trim * rbehind / sum, pagesperblock); if (rbehind < 0) rbehind = 0; } else rbehind -= trim * rbehind / sum; rahead -= trim * rahead / sum; } KASSERT(rbehind + rahead + count <= nitems(bp->b_pages), ("%s: behind %d ahead %d count %d", __func__, rbehind, rahead, count)); /* * Fill in the bp->b_pages[] array with requested and optional * read behind or read ahead pages. Read behind pages are looked * up in a backward direction, down to a first cached page. Same * for read ahead pages, but there is no need to shift the array * in case of encountering a cached page. */ i = bp->b_npages = 0; if (rbehind) { vm_pindex_t startpindex, tpindex; vm_page_t p; VM_OBJECT_WLOCK(object); startpindex = m[0]->pindex - rbehind; if ((p = TAILQ_PREV(m[0], pglist, listq)) != NULL && p->pindex >= startpindex) startpindex = p->pindex + 1; /* tpindex is unsigned; beware of numeric underflow. */ for (tpindex = m[0]->pindex - 1; tpindex >= startpindex && tpindex < m[0]->pindex; tpindex--, i++) { p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); if (p == NULL) { /* Shift the array. */ for (int j = 0; j < i; j++) bp->b_pages[j] = bp->b_pages[j + tpindex + 1 - startpindex]; break; } bp->b_pages[tpindex - startpindex] = p; } bp->b_pgbefore = i; bp->b_npages += i; bp->b_blkno -= IDX_TO_OFF(i) / DEV_BSIZE; } else bp->b_pgbefore = 0; /* Requested pages. */ for (int j = 0; j < count; j++, i++) bp->b_pages[i] = m[j]; bp->b_npages += count; if (rahead) { vm_pindex_t endpindex, tpindex; vm_page_t p; if (!VM_OBJECT_WOWNED(object)) VM_OBJECT_WLOCK(object); endpindex = m[count - 1]->pindex + rahead + 1; if ((p = TAILQ_NEXT(m[count - 1], listq)) != NULL && p->pindex < endpindex) endpindex = p->pindex; if (endpindex > object->size) endpindex = object->size; for (tpindex = m[count - 1]->pindex + 1; tpindex < endpindex; i++, tpindex++) { p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); if (p == NULL) break; bp->b_pages[i] = p; } bp->b_pgafter = i - bp->b_npages; bp->b_npages = i; } else bp->b_pgafter = 0; if (VM_OBJECT_WOWNED(object)) VM_OBJECT_WUNLOCK(object); /* Report back actual behind/ahead read. */ if (a_rbehind) *a_rbehind = bp->b_pgbefore; if (a_rahead) *a_rahead = bp->b_pgafter; #ifdef INVARIANTS KASSERT(bp->b_npages <= nitems(bp->b_pages), ("%s: buf %p overflowed", __func__, bp)); for (int j = 1, prev = 0; j < bp->b_npages; j++) { if (bp->b_pages[j] == bogus_page) continue; KASSERT(bp->b_pages[j]->pindex - bp->b_pages[prev]->pindex == j - prev, ("%s: pages array not consecutive, bp %p", __func__, bp)); prev = j; } #endif /* * Recalculate first offset and bytecount with regards to read behind. * Truncate bytecount to vnode real size and round up physical size * for real devices. */ foff = IDX_TO_OFF(bp->b_pages[0]->pindex); bytecount = bp->b_npages << PAGE_SHIFT; if ((foff + bytecount) > object->un_pager.vnp.vnp_size) bytecount = object->un_pager.vnp.vnp_size - foff; secmask = bo->bo_bsize - 1; KASSERT(secmask < PAGE_SIZE && secmask > 0, ("%s: sector size %d too large", __func__, secmask + 1)); bytecount = (bytecount + secmask) & ~secmask; /* * And map the pages to be read into the kva, if the filesystem * requires mapped buffers. */ if ((vp->v_mount->mnt_kern_flag & MNTK_UNMAPPED_BUFS) != 0 && unmapped_buf_allowed) { bp->b_data = unmapped_buf; bp->b_offset = 0; } else { bp->b_data = bp->b_kvabase; pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages); } /* Build a minimal buffer header. */ bp->b_iocmd = BIO_READ; KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred")); KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred")); bp->b_rcred = crhold(curthread->td_ucred); bp->b_wcred = crhold(curthread->td_ucred); pbgetbo(bo, bp); bp->b_vp = vp; bp->b_bcount = bp->b_bufsize = bp->b_runningbufspace = bytecount; bp->b_iooffset = dbtob(bp->b_blkno); KASSERT(IDX_TO_OFF(m[0]->pindex - bp->b_pages[0]->pindex) == (blkno0 - bp->b_blkno) * DEV_BSIZE + IDX_TO_OFF(m[0]->pindex) % bsize, ("wrong offsets bsize %d m[0] %ju b_pages[0] %ju " "blkno0 %ju b_blkno %ju", bsize, (uintmax_t)m[0]->pindex, (uintmax_t)bp->b_pages[0]->pindex, (uintmax_t)blkno0, (uintmax_t)bp->b_blkno)); atomic_add_long(&runningbufspace, bp->b_runningbufspace); VM_CNT_INC(v_vnodein); VM_CNT_ADD(v_vnodepgsin, bp->b_npages); if (iodone != NULL) { /* async */ bp->b_pgiodone = iodone; bp->b_caller1 = arg; bp->b_iodone = vnode_pager_generic_getpages_done_async; bp->b_flags |= B_ASYNC; BUF_KERNPROC(bp); bstrategy(bp); return (VM_PAGER_OK); } else { bp->b_iodone = bdone; bstrategy(bp); bwait(bp, PVM, "vnread"); error = vnode_pager_generic_getpages_done(bp); for (i = 0; i < bp->b_npages; i++) bp->b_pages[i] = NULL; bp->b_vp = NULL; pbrelbo(bp); uma_zfree(vnode_pbuf_zone, bp); return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK); } } static void vnode_pager_generic_getpages_done_async(struct buf *bp) { int error; error = vnode_pager_generic_getpages_done(bp); /* Run the iodone upon the requested range. */ bp->b_pgiodone(bp->b_caller1, bp->b_pages + bp->b_pgbefore, bp->b_npages - bp->b_pgbefore - bp->b_pgafter, error); for (int i = 0; i < bp->b_npages; i++) bp->b_pages[i] = NULL; bp->b_vp = NULL; pbrelbo(bp); uma_zfree(vnode_pbuf_zone, bp); } static int vnode_pager_generic_getpages_done(struct buf *bp) { vm_object_t object; off_t tfoff, nextoff; int i, error; error = (bp->b_ioflags & BIO_ERROR) != 0 ? EIO : 0; object = bp->b_vp->v_object; if (error == 0 && bp->b_bcount != bp->b_npages * PAGE_SIZE) { if (!buf_mapped(bp)) { bp->b_data = bp->b_kvabase; pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages); } bzero(bp->b_data + bp->b_bcount, PAGE_SIZE * bp->b_npages - bp->b_bcount); } if (buf_mapped(bp)) { pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages); bp->b_data = unmapped_buf; } VM_OBJECT_WLOCK(object); for (i = 0, tfoff = IDX_TO_OFF(bp->b_pages[0]->pindex); i < bp->b_npages; i++, tfoff = nextoff) { vm_page_t mt; nextoff = tfoff + PAGE_SIZE; mt = bp->b_pages[i]; if (nextoff <= object->un_pager.vnp.vnp_size) { /* * Read filled up entire page. */ - mt->valid = VM_PAGE_BITS_ALL; + vm_page_valid(mt); KASSERT(mt->dirty == 0, ("%s: page %p is dirty", __func__, mt)); KASSERT(!pmap_page_is_mapped(mt), ("%s: page %p is mapped", __func__, mt)); } else { /* * Read did not fill up entire page. * * Currently we do not set the entire page valid, * we just try to clear the piece that we couldn't * read. */ vm_page_set_valid_range(mt, 0, object->un_pager.vnp.vnp_size - tfoff); KASSERT((mt->dirty & vm_page_bits(0, object->un_pager.vnp.vnp_size - tfoff)) == 0, ("%s: page %p is dirty", __func__, mt)); } if (i < bp->b_pgbefore || i >= bp->b_npages - bp->b_pgafter) vm_page_readahead_finish(mt); } VM_OBJECT_WUNLOCK(object); if (error != 0) printf("%s: I/O read error %d\n", __func__, error); return (error); } /* * EOPNOTSUPP is no longer legal. For local media VFS's that do not * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to * vnode_pager_generic_putpages() to implement the previous behaviour. * * All other FS's should use the bypass to get to the local media * backing vp's VOP_PUTPAGES. */ static void vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count, int flags, int *rtvals) { int rtval; struct vnode *vp; int bytes = count * PAGE_SIZE; /* * Force synchronous operation if we are extremely low on memory * to prevent a low-memory deadlock. VOP operations often need to * allocate more memory to initiate the I/O ( i.e. do a BMAP * operation ). The swapper handles the case by limiting the amount * of asynchronous I/O, but that sort of solution doesn't scale well * for the vnode pager without a lot of work. * * Also, the backing vnode's iodone routine may not wake the pageout * daemon up. This should be probably be addressed XXX. */ if (vm_page_count_min()) flags |= VM_PAGER_PUT_SYNC; /* * Call device-specific putpages function */ vp = object->handle; VM_OBJECT_WUNLOCK(object); rtval = VOP_PUTPAGES(vp, m, bytes, flags, rtvals); KASSERT(rtval != EOPNOTSUPP, ("vnode_pager: stale FS putpages\n")); VM_OBJECT_WLOCK(object); } static int vn_off2bidx(vm_ooffset_t offset) { return ((offset & PAGE_MASK) / DEV_BSIZE); } static bool vn_dirty_blk(vm_page_t m, vm_ooffset_t offset) { KASSERT(IDX_TO_OFF(m->pindex) <= offset && offset < IDX_TO_OFF(m->pindex + 1), ("page %p pidx %ju offset %ju", m, (uintmax_t)m->pindex, (uintmax_t)offset)); return ((m->dirty & ((vm_page_bits_t)1 << vn_off2bidx(offset))) != 0); } /* * This is now called from local media FS's to operate against their * own vnodes if they fail to implement VOP_PUTPAGES. * * This is typically called indirectly via the pageout daemon and * clustering has already typically occurred, so in general we ask the * underlying filesystem to write the data out asynchronously rather * then delayed. */ int vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *ma, int bytecount, int flags, int *rtvals) { vm_object_t object; vm_page_t m; vm_ooffset_t maxblksz, next_offset, poffset, prev_offset; struct uio auio; struct iovec aiov; off_t prev_resid, wrsz; int count, error, i, maxsize, ncount, pgoff, ppscheck; bool in_hole; static struct timeval lastfail; static int curfail; object = vp->v_object; count = bytecount / PAGE_SIZE; for (i = 0; i < count; i++) rtvals[i] = VM_PAGER_ERROR; if ((int64_t)ma[0]->pindex < 0) { printf("vnode_pager_generic_putpages: " "attempt to write meta-data 0x%jx(%lx)\n", (uintmax_t)ma[0]->pindex, (u_long)ma[0]->dirty); rtvals[0] = VM_PAGER_BAD; return (VM_PAGER_BAD); } maxsize = count * PAGE_SIZE; ncount = count; poffset = IDX_TO_OFF(ma[0]->pindex); /* * If the page-aligned write is larger then the actual file we * have to invalidate pages occurring beyond the file EOF. However, * there is an edge case where a file may not be page-aligned where * the last page is partially invalid. In this case the filesystem * may not properly clear the dirty bits for the entire page (which * could be VM_PAGE_BITS_ALL due to the page having been mmap()d). * With the page locked we are free to fix-up the dirty bits here. * * We do not under any circumstances truncate the valid bits, as * this will screw up bogus page replacement. */ VM_OBJECT_RLOCK(object); if (maxsize + poffset > object->un_pager.vnp.vnp_size) { if (!VM_OBJECT_TRYUPGRADE(object)) { VM_OBJECT_RUNLOCK(object); VM_OBJECT_WLOCK(object); if (maxsize + poffset <= object->un_pager.vnp.vnp_size) goto downgrade; } if (object->un_pager.vnp.vnp_size > poffset) { maxsize = object->un_pager.vnp.vnp_size - poffset; ncount = btoc(maxsize); if ((pgoff = (int)maxsize & PAGE_MASK) != 0) { pgoff = roundup2(pgoff, DEV_BSIZE); /* * If the object is locked and the following * conditions hold, then the page's dirty * field cannot be concurrently changed by a * pmap operation. */ m = ma[ncount - 1]; vm_page_assert_sbusied(m); KASSERT(!pmap_page_is_write_mapped(m), ("vnode_pager_generic_putpages: page %p is not read-only", m)); MPASS(m->dirty != 0); vm_page_clear_dirty(m, pgoff, PAGE_SIZE - pgoff); } } else { maxsize = 0; ncount = 0; } for (i = ncount; i < count; i++) rtvals[i] = VM_PAGER_BAD; downgrade: VM_OBJECT_LOCK_DOWNGRADE(object); } auio.uio_iov = &aiov; auio.uio_segflg = UIO_NOCOPY; auio.uio_rw = UIO_WRITE; auio.uio_td = NULL; maxblksz = roundup2(poffset + maxsize, DEV_BSIZE); for (prev_offset = poffset; prev_offset < maxblksz;) { /* Skip clean blocks. */ for (in_hole = true; in_hole && prev_offset < maxblksz;) { m = ma[OFF_TO_IDX(prev_offset - poffset)]; for (i = vn_off2bidx(prev_offset); i < sizeof(vm_page_bits_t) * NBBY && prev_offset < maxblksz; i++) { if (vn_dirty_blk(m, prev_offset)) { in_hole = false; break; } prev_offset += DEV_BSIZE; } } if (in_hole) goto write_done; /* Find longest run of dirty blocks. */ for (next_offset = prev_offset; next_offset < maxblksz;) { m = ma[OFF_TO_IDX(next_offset - poffset)]; for (i = vn_off2bidx(next_offset); i < sizeof(vm_page_bits_t) * NBBY && next_offset < maxblksz; i++) { if (!vn_dirty_blk(m, next_offset)) goto start_write; next_offset += DEV_BSIZE; } } start_write: if (next_offset > poffset + maxsize) next_offset = poffset + maxsize; /* * Getting here requires finding a dirty block in the * 'skip clean blocks' loop. */ MPASS(prev_offset < next_offset); VM_OBJECT_RUNLOCK(object); aiov.iov_base = NULL; auio.uio_iovcnt = 1; auio.uio_offset = prev_offset; prev_resid = auio.uio_resid = aiov.iov_len = next_offset - prev_offset; error = VOP_WRITE(vp, &auio, vnode_pager_putpages_ioflags(flags), curthread->td_ucred); wrsz = prev_resid - auio.uio_resid; if (wrsz == 0) { if (ppsratecheck(&lastfail, &curfail, 1) != 0) { vn_printf(vp, "vnode_pager_putpages: " "zero-length write at %ju resid %zd\n", auio.uio_offset, auio.uio_resid); } VM_OBJECT_RLOCK(object); break; } /* Adjust the starting offset for next iteration. */ prev_offset += wrsz; MPASS(auio.uio_offset == prev_offset); ppscheck = 0; if (error != 0 && (ppscheck = ppsratecheck(&lastfail, &curfail, 1)) != 0) vn_printf(vp, "vnode_pager_putpages: I/O error %d\n", error); if (auio.uio_resid != 0 && (ppscheck != 0 || ppsratecheck(&lastfail, &curfail, 1) != 0)) vn_printf(vp, "vnode_pager_putpages: residual I/O %zd " "at %ju\n", auio.uio_resid, (uintmax_t)ma[0]->pindex); VM_OBJECT_RLOCK(object); if (error != 0 || auio.uio_resid != 0) break; } write_done: /* Mark completely processed pages. */ for (i = 0; i < OFF_TO_IDX(prev_offset - poffset); i++) rtvals[i] = VM_PAGER_OK; /* Mark partial EOF page. */ if (prev_offset == poffset + maxsize && (prev_offset & PAGE_MASK) != 0) rtvals[i++] = VM_PAGER_OK; /* Unwritten pages in range, free bonus if the page is clean. */ for (; i < ncount; i++) rtvals[i] = ma[i]->dirty == 0 ? VM_PAGER_OK : VM_PAGER_ERROR; VM_OBJECT_RUNLOCK(object); VM_CNT_ADD(v_vnodepgsout, i); VM_CNT_INC(v_vnodeout); return (rtvals[0]); } int vnode_pager_putpages_ioflags(int pager_flags) { int ioflags; /* * Pageouts are already clustered, use IO_ASYNC to force a * bawrite() rather then a bdwrite() to prevent paging I/O * from saturating the buffer cache. Dummy-up the sequential * heuristic to cause large ranges to cluster. If neither * IO_SYNC or IO_ASYNC is set, the system decides how to * cluster. */ ioflags = IO_VMIO; if ((pager_flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) != 0) ioflags |= IO_SYNC; else if ((pager_flags & VM_PAGER_CLUSTER_OK) == 0) ioflags |= IO_ASYNC; ioflags |= (pager_flags & VM_PAGER_PUT_INVAL) != 0 ? IO_INVAL: 0; ioflags |= (pager_flags & VM_PAGER_PUT_NOREUSE) != 0 ? IO_NOREUSE : 0; ioflags |= IO_SEQMAX << IO_SEQSHIFT; return (ioflags); } /* * vnode_pager_undirty_pages(). * * A helper to mark pages as clean after pageout that was possibly * done with a short write. The lpos argument specifies the page run * length in bytes, and the written argument specifies how many bytes * were actually written. eof is the offset past the last valid byte * in the vnode using the absolute file position of the first byte in * the run as the base from which it is computed. */ void vnode_pager_undirty_pages(vm_page_t *ma, int *rtvals, int written, off_t eof, int lpos) { vm_object_t obj; int i, pos, pos_devb; if (written == 0 && eof >= lpos) return; obj = ma[0]->object; VM_OBJECT_WLOCK(obj); for (i = 0, pos = 0; pos < written; i++, pos += PAGE_SIZE) { if (pos < trunc_page(written)) { rtvals[i] = VM_PAGER_OK; vm_page_undirty(ma[i]); } else { /* Partially written page. */ rtvals[i] = VM_PAGER_AGAIN; vm_page_clear_dirty(ma[i], 0, written & PAGE_MASK); } } if (eof >= lpos) /* avoid truncation */ goto done; for (pos = eof, i = OFF_TO_IDX(trunc_page(pos)); pos < lpos; i++) { if (pos != trunc_page(pos)) { /* * The page contains the last valid byte in * the vnode, mark the rest of the page as * clean, potentially making the whole page * clean. */ pos_devb = roundup2(pos & PAGE_MASK, DEV_BSIZE); vm_page_clear_dirty(ma[i], pos_devb, PAGE_SIZE - pos_devb); /* * If the page was cleaned, report the pageout * on it as successful. msync() no longer * needs to write out the page, endlessly * creating write requests and dirty buffers. */ if (ma[i]->dirty == 0) rtvals[i] = VM_PAGER_OK; pos = round_page(pos); } else { /* vm_pageout_flush() clears dirty */ rtvals[i] = VM_PAGER_BAD; pos += PAGE_SIZE; } } done: VM_OBJECT_WUNLOCK(obj); } static void vnode_pager_update_writecount(vm_object_t object, vm_offset_t start, vm_offset_t end) { struct vnode *vp; vm_ooffset_t old_wm; VM_OBJECT_WLOCK(object); if (object->type != OBJT_VNODE) { VM_OBJECT_WUNLOCK(object); return; } old_wm = object->un_pager.vnp.writemappings; object->un_pager.vnp.writemappings += (vm_ooffset_t)end - start; vp = object->handle; if (old_wm == 0 && object->un_pager.vnp.writemappings != 0) { ASSERT_VOP_LOCKED(vp, "v_writecount inc"); VOP_ADD_WRITECOUNT_CHECKED(vp, 1); CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d", __func__, vp, vp->v_writecount); } else if (old_wm != 0 && object->un_pager.vnp.writemappings == 0) { ASSERT_VOP_LOCKED(vp, "v_writecount dec"); VOP_ADD_WRITECOUNT_CHECKED(vp, -1); CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d", __func__, vp, vp->v_writecount); } VM_OBJECT_WUNLOCK(object); } static void vnode_pager_release_writecount(vm_object_t object, vm_offset_t start, vm_offset_t end) { struct vnode *vp; struct mount *mp; vm_offset_t inc; VM_OBJECT_WLOCK(object); /* * First, recheck the object type to account for the race when * the vnode is reclaimed. */ if (object->type != OBJT_VNODE) { VM_OBJECT_WUNLOCK(object); return; } /* * Optimize for the case when writemappings is not going to * zero. */ inc = end - start; if (object->un_pager.vnp.writemappings != inc) { object->un_pager.vnp.writemappings -= inc; VM_OBJECT_WUNLOCK(object); return; } vp = object->handle; vhold(vp); VM_OBJECT_WUNLOCK(object); mp = NULL; vn_start_write(vp, &mp, V_WAIT); vn_lock(vp, LK_SHARED | LK_RETRY); /* * Decrement the object's writemappings, by swapping the start * and end arguments for vnode_pager_update_writecount(). If * there was not a race with vnode reclaimation, then the * vnode's v_writecount is decremented. */ vnode_pager_update_writecount(object, end, start); VOP_UNLOCK(vp, 0); vdrop(vp); if (mp != NULL) vn_finished_write(mp); }