diff --git a/sys/vm/vm_fault.c b/sys/vm/vm_fault.c index 6e0415b30600..8c7fe9e37af1 100644 --- a/sys/vm/vm_fault.c +++ b/sys/vm/vm_fault.c @@ -1,2302 +1,2302 @@ /*- * 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. * * * 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 #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 #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_DONTNEED_MIN 1048576 struct faultstate { /* Fault parameters. */ vm_offset_t vaddr; vm_page_t *m_hold; vm_prot_t fault_type; vm_prot_t prot; int fault_flags; boolean_t wired; /* Control state. */ struct timeval oom_start_time; bool oom_started; int nera; bool can_read_lock; /* Page reference for cow. */ vm_page_t m_cow; /* Current object. */ vm_object_t object; vm_pindex_t pindex; vm_page_t m; /* Top-level map object. */ vm_object_t first_object; vm_pindex_t first_pindex; vm_page_t first_m; /* Map state. */ vm_map_t map; vm_map_entry_t entry; int map_generation; bool lookup_still_valid; /* Vnode if locked. */ struct vnode *vp; }; /* * Return codes for internal fault routines. */ enum fault_status { FAULT_SUCCESS = 10000, /* Return success to user. */ FAULT_FAILURE, /* Return failure to user. */ FAULT_CONTINUE, /* Continue faulting. */ FAULT_RESTART, /* Restart fault. */ FAULT_OUT_OF_BOUNDS, /* Invalid address for pager. */ FAULT_HARD, /* Performed I/O. */ FAULT_SOFT, /* Found valid page. */ FAULT_PROTECTION_FAILURE, /* Invalid access. */ }; enum fault_next_status { FAULT_NEXT_GOTOBJ = 1, FAULT_NEXT_NOOBJ, FAULT_NEXT_RESTART, }; 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 vm_fault_page_release(vm_page_t *mp) { vm_page_t m; m = *mp; if (m != NULL) { /* * We are likely to loop around again and attempt to busy * this page. Deactivating it leaves it available for * pageout while optimizing fault restarts. */ vm_page_deactivate(m); vm_page_xunbusy(m); *mp = NULL; } } static inline void vm_fault_page_free(vm_page_t *mp) { vm_page_t m; m = *mp; if (m != NULL) { VM_OBJECT_ASSERT_WLOCKED(m->object); if (!vm_page_wired(m)) vm_page_free(m); else vm_page_xunbusy(m); *mp = NULL; } } /* * Return true if a vm_pager_get_pages() call is needed in order to check * whether the pager might have a particular page, false if it can be determined * immediately that the pager can not have a copy. For swap objects, this can * be checked quickly. */ static inline bool vm_fault_object_needs_getpages(vm_object_t object) { VM_OBJECT_ASSERT_LOCKED(object); return ((object->flags & OBJ_SWAP) == 0 || !pctrie_is_empty(&object->un_pager.swp.swp_blks)); } static inline void vm_fault_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 vm_fault_unlock_vp(struct faultstate *fs) { if (fs->vp != NULL) { vput(fs->vp); fs->vp = NULL; } } static void vm_fault_deallocate(struct faultstate *fs) { vm_fault_page_release(&fs->m_cow); vm_fault_page_release(&fs->m); vm_object_pip_wakeup(fs->object); if (fs->object != fs->first_object) { VM_OBJECT_WLOCK(fs->first_object); vm_fault_page_free(&fs->first_m); VM_OBJECT_WUNLOCK(fs->first_object); vm_object_pip_wakeup(fs->first_object); } vm_object_deallocate(fs->first_object); vm_fault_unlock_map(fs); vm_fault_unlock_vp(fs); } static void vm_fault_unlock_and_deallocate(struct faultstate *fs) { VM_OBJECT_UNLOCK(fs->object); vm_fault_deallocate(fs); } static void vm_fault_dirty(struct faultstate *fs, vm_page_t m) { bool need_dirty; if (((fs->prot & VM_PROT_WRITE) == 0 && (fs->fault_flags & VM_FAULT_DIRTY) == 0) || (m->oflags & VPO_UNMANAGED) != 0) return; VM_PAGE_OBJECT_BUSY_ASSERT(m); need_dirty = ((fs->fault_type & VM_PROT_WRITE) != 0 && (fs->fault_flags & VM_FAULT_WIRE) == 0) || (fs->fault_flags & VM_FAULT_DIRTY) != 0; vm_object_set_writeable_dirty(m->object); /* * 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. */ if (need_dirty && vm_page_set_dirty(m) == 0) { /* * If this is a NOSYNC mmap we do not want to set PGA_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 ((fs->entry->eflags & MAP_ENTRY_NOSYNC) != 0) vm_page_aflag_set(m, PGA_NOSYNC); else vm_page_aflag_clear(m, PGA_NOSYNC); } } /* * Unlocks fs.first_object and fs.map on success. */ static enum fault_status vm_fault_soft_fast(struct faultstate *fs) { vm_page_t m, m_map; #if VM_NRESERVLEVEL > 0 vm_page_t m_super; int flags; #endif int psind; vm_offset_t vaddr; MPASS(fs->vp == NULL); /* * If we fail, vast majority of the time it is because the page is not * there to begin with. Opportunistically perform the lookup and * subsequent checks without the object lock, revalidate later. * * Note: a busy page can be mapped for read|execute access. */ m = vm_page_lookup_unlocked(fs->first_object, fs->first_pindex); if (m == NULL || !vm_page_all_valid(m) || ((fs->prot & VM_PROT_WRITE) != 0 && vm_page_busied(m))) { VM_OBJECT_WLOCK(fs->first_object); return (FAULT_FAILURE); } vaddr = fs->vaddr; VM_OBJECT_RLOCK(fs->first_object); /* * Now that we stabilized the state, revalidate the page is in the shape * we encountered above. */ if (m->object != fs->first_object || m->pindex != fs->first_pindex) goto fail; vm_object_busy(fs->first_object); if (!vm_page_all_valid(m) || ((fs->prot & VM_PROT_WRITE) != 0 && vm_page_busied(m))) goto fail_busy; m_map = m; psind = 0; #if VM_NRESERVLEVEL > 0 if ((m->flags & PG_FICTITIOUS) == 0 && (m_super = vm_reserv_to_superpage(m)) != NULL) { psind = m_super->psind; KASSERT(psind > 0, ("psind %d of m_super %p < 1", psind, m_super)); flags = PS_ALL_VALID; if ((fs->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; } while (rounddown2(vaddr, pagesizes[psind]) < fs->entry->start || roundup2(vaddr + 1, pagesizes[psind]) > fs->entry->end || (vaddr & (pagesizes[psind] - 1)) != (VM_PAGE_TO_PHYS(m) & (pagesizes[psind] - 1)) || !vm_page_ps_test(m_super, psind, flags, m) || !pmap_ps_enabled(fs->map->pmap)) { psind--; if (psind == 0) break; m_super += rounddown2(m - m_super, atop(pagesizes[psind])); KASSERT(m_super->psind >= psind, ("psind %d of m_super %p < %d", m_super->psind, m_super, psind)); } if (psind > 0) { m_map = m_super; vaddr = rounddown2(vaddr, pagesizes[psind]); /* Preset the modified bit for dirty superpages. */ if ((flags & PS_ALL_DIRTY) != 0) fs->fault_type |= VM_PROT_WRITE; } } #endif if (pmap_enter(fs->map->pmap, vaddr, m_map, fs->prot, fs->fault_type | PMAP_ENTER_NOSLEEP | (fs->wired ? PMAP_ENTER_WIRED : 0), psind) != KERN_SUCCESS) goto fail_busy; if (fs->m_hold != NULL) { (*fs->m_hold) = m; vm_page_wire(m); } if (psind == 0 && !fs->wired) vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true); VM_OBJECT_RUNLOCK(fs->first_object); vm_fault_dirty(fs, m); vm_object_unbusy(fs->first_object); vm_map_lookup_done(fs->map, fs->entry); curthread->td_ru.ru_minflt++; return (FAULT_SUCCESS); fail_busy: vm_object_unbusy(fs->first_object); fail: if (!VM_OBJECT_TRYUPGRADE(fs->first_object)) { VM_OBJECT_RUNLOCK(fs->first_object); VM_OBJECT_WLOCK(fs->first_object); } return (FAULT_FAILURE); } static void vm_fault_restore_map_lock(struct faultstate *fs) { VM_OBJECT_ASSERT_WLOCKED(fs->first_object); MPASS(blockcount_read(&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(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 = TAILQ_NEXT(m, listq)) { KASSERT(m != NULL && m->pindex == pidx, ("%s: pindex mismatch", __func__)); vm_fault_populate_check_page(m); vm_page_deactivate(m); vm_page_xunbusy(m); } } static enum fault_status vm_fault_populate(struct faultstate *fs) { vm_offset_t vaddr; vm_page_t m; vm_pindex_t map_first, map_last, pager_first, pager_last, pidx; int bdry_idx, i, npages, psind, rv; enum fault_status res; MPASS(fs->object == fs->first_object); VM_OBJECT_ASSERT_WLOCKED(fs->first_object); MPASS(blockcount_read(&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; vm_fault_unlock_map(fs); vm_fault_unlock_vp(fs); res = FAULT_SUCCESS; /* * 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, fs->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 (FAULT_RESTART); return (FAULT_CONTINUE); } if (rv != VM_PAGER_OK) return (FAULT_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); bdry_idx = MAP_ENTRY_SPLIT_BOUNDARY_INDEX(fs->entry); if (fs->map->timestamp != fs->map_generation) { if (bdry_idx == 0) { vm_fault_populate_cleanup(fs->first_object, pager_first, pager_last); } else { m = vm_page_lookup(fs->first_object, pager_first); if (m != fs->m) vm_page_xunbusy(m); } return (FAULT_RESTART); } /* * The map is unchanged after our last unlock. Process the fault. * * First, the special case of largepage mappings, where * populate only busies the first page in superpage run. */ if (bdry_idx != 0) { KASSERT(PMAP_HAS_LARGEPAGES, ("missing pmap support for large pages")); m = vm_page_lookup(fs->first_object, pager_first); vm_fault_populate_check_page(m); VM_OBJECT_WUNLOCK(fs->first_object); vaddr = fs->entry->start + IDX_TO_OFF(pager_first) - fs->entry->offset; /* assert alignment for entry */ KASSERT((vaddr & (pagesizes[bdry_idx] - 1)) == 0, ("unaligned superpage start %#jx pager_first %#jx offset %#jx vaddr %#jx", (uintmax_t)fs->entry->start, (uintmax_t)pager_first, (uintmax_t)fs->entry->offset, (uintmax_t)vaddr)); KASSERT((VM_PAGE_TO_PHYS(m) & (pagesizes[bdry_idx] - 1)) == 0, ("unaligned superpage m %p %#jx", m, (uintmax_t)VM_PAGE_TO_PHYS(m))); rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot, fs->fault_type | (fs->wired ? PMAP_ENTER_WIRED : 0) | PMAP_ENTER_LARGEPAGE, bdry_idx); VM_OBJECT_WLOCK(fs->first_object); vm_page_xunbusy(m); if (rv != KERN_SUCCESS) { res = FAULT_FAILURE; goto out; } if ((fs->fault_flags & VM_FAULT_WIRE) != 0) { for (i = 0; i < atop(pagesizes[bdry_idx]); i++) vm_page_wire(m + i); } if (fs->m_hold != NULL) { *fs->m_hold = m + (fs->first_pindex - pager_first); vm_page_wire(*fs->m_hold); } goto out; } /* * 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 = TAILQ_NEXT(&m[npages - 1], listq)) { vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset; KASSERT(m != NULL && m->pindex == pidx, ("%s: pindex mismatch", __func__)); psind = m->psind; while (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 || pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last || !pmap_ps_enabled(fs->map->pmap))) psind--; npages = atop(pagesizes[psind]); for (i = 0; i < npages; i++) { vm_fault_populate_check_page(&m[i]); vm_fault_dirty(fs, &m[i]); } VM_OBJECT_WUNLOCK(fs->first_object); rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot, fs->fault_type | (fs->wired ? PMAP_ENTER_WIRED : 0), psind); /* * pmap_enter() may fail for a superpage mapping if additional * protection policies prevent the full mapping. * For example, this will happen on amd64 if the entire * address range does not share the same userspace protection * key. Revert to single-page mappings if this happens. */ MPASS(rv == KERN_SUCCESS || (psind > 0 && rv == KERN_PROTECTION_FAILURE)); if (__predict_false(psind > 0 && rv == KERN_PROTECTION_FAILURE)) { MPASS(!fs->wired); for (i = 0; i < npages; i++) { rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i), &m[i], fs->prot, fs->fault_type, 0); MPASS(rv == KERN_SUCCESS); } } VM_OBJECT_WLOCK(fs->first_object); for (i = 0; i < npages; i++) { if ((fs->fault_flags & VM_FAULT_WIRE) != 0 && m[i].pindex == fs->first_pindex) vm_page_wire(&m[i]); else vm_page_activate(&m[i]); if (fs->m_hold != NULL && m[i].pindex == fs->first_pindex) { (*fs->m_hold) = &m[i]; vm_page_wire(&m[i]); } vm_page_xunbusy(&m[i]); } } out: curthread->td_ru.ru_majflt++; return (res); } 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); } static bool vm_fault_object_ensure_wlocked(struct faultstate *fs) { if (fs->object == fs->first_object) VM_OBJECT_ASSERT_WLOCKED(fs->object); if (!fs->can_read_lock) { VM_OBJECT_ASSERT_WLOCKED(fs->object); return (true); } if (VM_OBJECT_WOWNED(fs->object)) return (true); if (VM_OBJECT_TRYUPGRADE(fs->object)) return (true); return (false); } static enum fault_status vm_fault_lock_vnode(struct faultstate *fs, bool objlocked) { struct vnode *vp; int error, locked; if (fs->object->type != OBJT_VNODE) return (FAULT_CONTINUE); vp = fs->object->handle; if (vp == fs->vp) { ASSERT_VOP_LOCKED(vp, "saved vnode is not locked"); return (FAULT_CONTINUE); } /* * Perform an unlock in case the desired vnode changed while * the map was unlocked during a retry. */ vm_fault_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); if (error == 0) { fs->vp = vp; return (FAULT_CONTINUE); } vhold(vp); if (objlocked) vm_fault_unlock_and_deallocate(fs); else vm_fault_deallocate(fs); error = vget(vp, locked | LK_RETRY | LK_CANRECURSE); vdrop(vp); fs->vp = vp; KASSERT(error == 0, ("vm_fault: vget failed %d", error)); return (FAULT_RESTART); } /* * Calculate the desired readahead. Handle drop-behind. * * Returns the number of readahead blocks to pass to the pager. */ static int vm_fault_readahead(struct faultstate *fs) { int era, nera; u_char behavior; 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 (fs->vaddr == fs->entry->next_read) vm_fault_dontneed(fs, fs->vaddr, nera); } else if (fs->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, 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; } return (nera); } static int vm_fault_lookup(struct faultstate *fs) { int result; KASSERT(!fs->lookup_still_valid, ("vm_fault_lookup: Map already locked.")); result = vm_map_lookup(&fs->map, fs->vaddr, fs->fault_type | VM_PROT_FAULT_LOOKUP, &fs->entry, &fs->first_object, &fs->first_pindex, &fs->prot, &fs->wired); if (result != KERN_SUCCESS) { vm_fault_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)fs->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, fs->vaddr, &fs->entry) && (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION)) { vm_fault_unlock_vp(fs); fs->entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; vm_map_unlock_and_wait(fs->map, 0); } else vm_map_unlock(fs->map); return (KERN_RESOURCE_SHORTAGE); } MPASS((fs->entry->eflags & MAP_ENTRY_GUARD) == 0); if (fs->wired) fs->fault_type = fs->prot | (fs->fault_type & VM_PROT_COPY); else KASSERT((fs->fault_flags & VM_FAULT_WIRE) == 0, ("!fs->wired && VM_FAULT_WIRE")); fs->lookup_still_valid = true; return (KERN_SUCCESS); } static int vm_fault_relookup(struct faultstate *fs) { vm_object_t retry_object; vm_pindex_t retry_pindex; vm_prot_t retry_prot; int result; if (!vm_map_trylock_read(fs->map)) return (KERN_RESTART); fs->lookup_still_valid = true; if (fs->map->timestamp == fs->map_generation) return (KERN_SUCCESS); result = vm_map_lookup_locked(&fs->map, fs->vaddr, fs->fault_type, &fs->entry, &retry_object, &retry_pindex, &retry_prot, &fs->wired); if (result != KERN_SUCCESS) { /* * If retry of map lookup would have blocked then * retry fault from start. */ if (result == KERN_FAILURE) return (KERN_RESTART); return (result); } if (retry_object != fs->first_object || retry_pindex != fs->first_pindex) return (KERN_RESTART); /* * 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. */ fs->prot &= retry_prot; fs->fault_type &= retry_prot; if (fs->prot == 0) return (KERN_RESTART); /* Reassert because wired may have changed. */ KASSERT(fs->wired || (fs->fault_flags & VM_FAULT_WIRE) == 0, ("!wired && VM_FAULT_WIRE")); return (KERN_SUCCESS); } static void vm_fault_cow(struct faultstate *fs) { bool is_first_object_locked; KASSERT(fs->object != fs->first_object, ("source and target COW objects are identical")); /* * 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 and no other refs. */ fs->object->shadow_count == 1 && fs->object->ref_count == 1 && /* * No other ways to look the object up */ fs->object->handle == NULL && (fs->object->flags & OBJ_ANON) != 0 && /* * We don't chase down the shadow chain and we can acquire locks. */ (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs->first_object)) && fs->object == fs->first_object->backing_object && VM_OBJECT_TRYWLOCK(fs->object)) { /* * Remove but keep xbusy for replace. fs->m is moved into * fs->first_object and left busy while fs->first_m is * conditionally freed. */ vm_page_remove_xbusy(fs->m); vm_page_replace(fs->m, fs->first_object, fs->first_pindex, 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 VM_OBJECT_WUNLOCK(fs->object); VM_OBJECT_WUNLOCK(fs->first_object); fs->first_m = fs->m; fs->m = NULL; VM_CNT_INC(v_cow_optim); } else { if (is_first_object_locked) VM_OBJECT_WUNLOCK(fs->first_object); /* * Oh, well, lets copy it. */ pmap_copy_page(fs->m, fs->first_m); vm_page_valid(fs->first_m); if (fs->wired && (fs->fault_flags & VM_FAULT_WIRE) == 0) { vm_page_wire(fs->first_m); vm_page_unwire(fs->m, PQ_INACTIVE); } /* * Save the cow page to be released after * pmap_enter is complete. */ fs->m_cow = fs->m; fs->m = NULL; /* * Typically, the shadow object is either private to this * address space (OBJ_ONEMAPPING) or its pages are read only. * In the highly unusual case where the pages of a shadow object * are read/write shared between this and other address spaces, * we need to ensure that any pmap-level mappings to the * original, copy-on-write page from the backing object are * removed from those other address spaces. * * The flag check is racy, but this is tolerable: if * OBJ_ONEMAPPING is cleared after the check, the busy state * ensures that new mappings of m_cow can't be created. * pmap_enter() will replace an existing mapping in the current * address space. If OBJ_ONEMAPPING is set after the check, * removing mappings will at worse trigger some unnecessary page * faults. */ vm_page_assert_xbusied(fs->m_cow); if ((fs->first_object->flags & OBJ_ONEMAPPING) == 0) pmap_remove_all(fs->m_cow); } vm_object_pip_wakeup(fs->object); /* * Only use the new page below... */ fs->object = fs->first_object; fs->pindex = fs->first_pindex; fs->m = fs->first_m; VM_CNT_INC(v_cow_faults); curthread->td_cow++; } static enum fault_next_status vm_fault_next(struct faultstate *fs) { vm_object_t next_object; if (fs->object == fs->first_object || !fs->can_read_lock) VM_OBJECT_ASSERT_WLOCKED(fs->object); else VM_OBJECT_ASSERT_LOCKED(fs->object); /* * 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) { fs->first_m = fs->m; fs->m = NULL; } else if (fs->m != NULL) { if (!vm_fault_object_ensure_wlocked(fs)) { fs->can_read_lock = false; vm_fault_unlock_and_deallocate(fs); return (FAULT_NEXT_RESTART); } vm_fault_page_free(&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) return (FAULT_NEXT_NOOBJ); MPASS(fs->first_m != NULL); KASSERT(fs->object != next_object, ("object loop %p", next_object)); if (fs->can_read_lock) VM_OBJECT_RLOCK(next_object); else 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_UNLOCK(fs->object); fs->object = next_object; return (FAULT_NEXT_GOTOBJ); } static void vm_fault_zerofill(struct faultstate *fs) { /* * 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); fs->object = fs->first_object; fs->pindex = fs->first_pindex; } MPASS(fs->first_m != NULL); MPASS(fs->m == NULL); fs->m = fs->first_m; 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); vm_page_valid(fs->m); } /* * Initiate page fault after timeout. Returns true if caller should * do vm_waitpfault() after the call. */ static bool vm_fault_allocate_oom(struct faultstate *fs) { struct timeval now; vm_fault_unlock_and_deallocate(fs); if (vm_pfault_oom_attempts < 0) return (true); if (!fs->oom_started) { fs->oom_started = true; getmicrotime(&fs->oom_start_time); return (true); } getmicrotime(&now); timevalsub(&now, &fs->oom_start_time); if (now.tv_sec < vm_pfault_oom_attempts * vm_pfault_oom_wait) return (true); 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); fs->oom_started = false; return (false); } /* * Allocate a page directly or via the object populate method. */ static enum fault_status vm_fault_allocate(struct faultstate *fs) { struct domainset *dset; enum fault_status res; if ((fs->object->flags & OBJ_SIZEVNLOCK) != 0) { res = vm_fault_lock_vnode(fs, true); MPASS(res == FAULT_CONTINUE || res == FAULT_RESTART); if (res == FAULT_RESTART) return (res); } if (fs->pindex >= fs->object->size) { vm_fault_unlock_and_deallocate(fs); return (FAULT_OUT_OF_BOUNDS); } if (fs->object == fs->first_object && (fs->first_object->flags & OBJ_POPULATE) != 0 && fs->first_object->shadow_count == 0) { res = vm_fault_populate(fs); switch (res) { case FAULT_SUCCESS: case FAULT_FAILURE: case FAULT_RESTART: vm_fault_unlock_and_deallocate(fs); return (res); case FAULT_CONTINUE: /* * Pager's populate() method * returned VM_PAGER_BAD. */ break; default: panic("inconsistent return codes"); } } /* * Allocate a new page for this object/offset pair. * * If the process has a fatal signal pending, prioritize the allocation * with the expectation that the process will exit shortly and free some * pages. In particular, the signal may have been posted by the page * daemon in an attempt to resolve an out-of-memory condition. * * The unlocked read of the p_flag is harmless. At worst, the P_KILLED * might be not observed here, and allocation fails, causing a 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(fs->vaddr) - fs->pindex); #endif if (!vm_pager_can_alloc_page(fs->object, fs->pindex)) { vm_fault_unlock_and_deallocate(fs); return (FAULT_FAILURE); } fs->m = vm_page_alloc(fs->object, fs->pindex, P_KILLED(curproc) ? VM_ALLOC_SYSTEM : 0); } if (fs->m == NULL) { if (vm_fault_allocate_oom(fs)) vm_waitpfault(dset, vm_pfault_oom_wait * hz); return (FAULT_RESTART); } fs->oom_started = false; return (FAULT_CONTINUE); } /* * 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. */ static enum fault_status vm_fault_getpages(struct faultstate *fs, int *behindp, int *aheadp) { vm_offset_t e_end, e_start; int ahead, behind, cluster_offset, rv; enum fault_status status; u_char behavior; /* * 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; behavior = vm_map_entry_behavior(fs->entry); /* * 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->nera == -1 && !P_KILLED(curproc)) fs->nera = vm_fault_readahead(fs); /* * 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.) */ vm_fault_unlock_map(fs); status = vm_fault_lock_vnode(fs, false); MPASS(status == FAULT_CONTINUE || status == FAULT_RESTART); if (status == FAULT_RESTART) return (status); - KASSERT(fs->vp == NULL || !fs->map->system_map, + KASSERT(fs->vp == NULL || !vm_map_is_system(fs->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 (fs->nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM || P_KILLED(curproc)) { behind = 0; ahead = 0; } else { /* Is this a sequential fault? */ if (fs->nera > 0) { behind = 0; ahead = fs->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(fs->vaddr - e_start)); ahead = VM_FAULT_READ_DEFAULT - 1 - cluster_offset; } ahead = ulmin(ahead, atop(e_end - fs->vaddr) - 1); } *behindp = behind; *aheadp = ahead; rv = vm_pager_get_pages(fs->object, &fs->m, 1, behindp, aheadp); if (rv == VM_PAGER_OK) return (FAULT_HARD); 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) { VM_OBJECT_WLOCK(fs->object); vm_fault_page_free(&fs->m); vm_fault_unlock_and_deallocate(fs); return (FAULT_OUT_OF_BOUNDS); } KASSERT(rv == VM_PAGER_FAIL, ("%s: unexpected pager error %d", __func__, rv)); return (FAULT_CONTINUE); } /* * 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. */ static void vm_fault_busy_sleep(struct faultstate *fs) { /* * 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) { vm_fault_page_release(&fs->first_m); vm_object_pip_wakeup(fs->first_object); } vm_object_pip_wakeup(fs->object); vm_fault_unlock_map(fs); if (fs->m != vm_page_lookup(fs->object, fs->pindex) || !vm_page_busy_sleep(fs->m, "vmpfw", 0)) VM_OBJECT_UNLOCK(fs->object); VM_CNT_INC(v_intrans); vm_object_deallocate(fs->first_object); } /* * Handle page lookup, populate, allocate, page-in for the current * object. * * The object is locked on entry and will remain locked with a return * code of FAULT_CONTINUE so that fault may follow the shadow chain. * Otherwise, the object will be unlocked upon return. */ static enum fault_status vm_fault_object(struct faultstate *fs, int *behindp, int *aheadp) { enum fault_status res; bool dead; if (fs->object == fs->first_object || !fs->can_read_lock) VM_OBJECT_ASSERT_WLOCKED(fs->object); else VM_OBJECT_ASSERT_LOCKED(fs->object); /* * 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; vm_fault_unlock_and_deallocate(fs); if (dead) return (FAULT_PROTECTION_FAILURE); pause("vmf_de", 1); return (FAULT_RESTART); } /* * See if the page is resident. */ fs->m = vm_page_lookup(fs->object, fs->pindex); if (fs->m != NULL) { if (!vm_page_tryxbusy(fs->m)) { vm_fault_busy_sleep(fs); return (FAULT_RESTART); } /* * The page is marked busy for other processes and the * pagedaemon. If it is still completely valid we are * done. */ if (vm_page_all_valid(fs->m)) { VM_OBJECT_UNLOCK(fs->object); return (FAULT_SOFT); } } /* * Page is not resident. If the pager might contain the page * or this is the beginning of the search, allocate a new * page. */ if (fs->m == NULL && (vm_fault_object_needs_getpages(fs->object) || fs->object == fs->first_object)) { if (!vm_fault_object_ensure_wlocked(fs)) { fs->can_read_lock = false; vm_fault_unlock_and_deallocate(fs); return (FAULT_RESTART); } res = vm_fault_allocate(fs); if (res != FAULT_CONTINUE) return (res); } /* * Check to see if the pager can possibly satisfy this fault. * If not, skip to the next object without dropping the lock to * preserve atomicity of shadow faults. */ if (vm_fault_object_needs_getpages(fs->object)) { /* * 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. The exclusive busy * prevents simultaneous faults and collapses while * the object lock is dropped. */ VM_OBJECT_UNLOCK(fs->object); res = vm_fault_getpages(fs, behindp, aheadp); if (res == FAULT_CONTINUE) VM_OBJECT_WLOCK(fs->object); } else { res = FAULT_CONTINUE; } return (res); } 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; int ahead, behind, faultcount, rv; enum fault_status res; enum fault_next_status res_next; bool hardfault; VM_CNT_INC(v_vm_faults); if ((curthread->td_pflags & TDP_NOFAULTING) != 0) return (KERN_PROTECTION_FAILURE); fs.vp = NULL; fs.vaddr = vaddr; fs.m_hold = m_hold; fs.fault_flags = fault_flags; fs.map = map; fs.lookup_still_valid = false; fs.oom_started = false; fs.nera = -1; fs.can_read_lock = true; faultcount = 0; hardfault = false; RetryFault: fs.fault_type = fault_type; /* * Find the backing store object and offset into it to begin the * search. */ rv = vm_fault_lookup(&fs); if (rv != KERN_SUCCESS) { if (rv == KERN_RESOURCE_SHORTAGE) goto RetryFault; return (rv); } /* * Try to avoid lock contention on the top-level object through * special-case handling of some types of page faults, specifically, * those that are mapping an existing page from the top-level object. * Under this condition, a read lock on the object suffices, allowing * multiple page faults of a similar type to run in parallel. */ if (fs.vp == NULL /* avoid locked vnode leak */ && (fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) == 0 && (fs.fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0) { res = vm_fault_soft_fast(&fs); if (res == FAULT_SUCCESS) { VM_OBJECT_ASSERT_UNLOCKED(fs.first_object); return (KERN_SUCCESS); } VM_OBJECT_ASSERT_WLOCKED(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.m_cow = fs.m = fs.first_m = NULL; /* * Search for the page at object/offset. */ fs.object = fs.first_object; fs.pindex = fs.first_pindex; if ((fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) != 0) { res = vm_fault_allocate(&fs); switch (res) { case FAULT_RESTART: goto RetryFault; case FAULT_SUCCESS: return (KERN_SUCCESS); case FAULT_FAILURE: return (KERN_FAILURE); case FAULT_OUT_OF_BOUNDS: return (KERN_OUT_OF_BOUNDS); case FAULT_CONTINUE: break; default: panic("vm_fault: Unhandled status %d", res); } } while (TRUE) { KASSERT(fs.m == NULL, ("page still set %p at loop start", fs.m)); res = vm_fault_object(&fs, &behind, &ahead); switch (res) { case FAULT_SOFT: goto found; case FAULT_HARD: faultcount = behind + 1 + ahead; hardfault = true; goto found; case FAULT_RESTART: goto RetryFault; case FAULT_SUCCESS: return (KERN_SUCCESS); case FAULT_FAILURE: return (KERN_FAILURE); case FAULT_OUT_OF_BOUNDS: return (KERN_OUT_OF_BOUNDS); case FAULT_PROTECTION_FAILURE: return (KERN_PROTECTION_FAILURE); case FAULT_CONTINUE: break; default: panic("vm_fault: Unhandled status %d", res); } /* * The page was not found in the current object. Try to * traverse into a backing object or zero fill if none is * found. */ res_next = vm_fault_next(&fs); if (res_next == FAULT_NEXT_RESTART) goto RetryFault; else if (res_next == FAULT_NEXT_GOTOBJ) continue; MPASS(res_next == FAULT_NEXT_NOOBJ); if ((fs.fault_flags & VM_FAULT_NOFILL) != 0) { if (fs.first_object == fs.object) vm_fault_page_free(&fs.first_m); vm_fault_unlock_and_deallocate(&fs); return (KERN_OUT_OF_BOUNDS); } VM_OBJECT_UNLOCK(fs.object); vm_fault_zerofill(&fs); /* Don't try to prefault neighboring pages. */ faultcount = 1; break; } found: /* * A valid page has been found and exclusively busied. The * object lock must no longer be held. */ vm_page_assert_xbusied(fs.m); VM_OBJECT_ASSERT_UNLOCKED(fs.object); /* * 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 ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) { vm_fault_cow(&fs); /* * 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; } else { fs.prot &= ~VM_PROT_WRITE; } } /* * We must verify that the maps have not changed since our last * lookup. */ if (!fs.lookup_still_valid) { rv = vm_fault_relookup(&fs); if (rv != KERN_SUCCESS) { vm_fault_deallocate(&fs); if (rv == KERN_RESTART) goto RetryFault; return (rv); } } VM_OBJECT_ASSERT_UNLOCKED(fs.object); /* * 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; /* * Page must be completely valid or it is not fit to * map into user space. vm_pager_get_pages() ensures this. */ vm_page_assert_xbusied(fs.m); KASSERT(vm_page_all_valid(fs.m), ("vm_fault: page %p partially invalid", fs.m)); vm_fault_dirty(&fs, fs.m); /* * 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, fs.prot, fs.fault_type | (fs.wired ? PMAP_ENTER_WIRED : 0), 0); if (faultcount != 1 && (fs.fault_flags & VM_FAULT_WIRE) == 0 && fs.wired == 0) vm_fault_prefault(&fs, vaddr, faultcount > 0 ? behind : PFBAK, faultcount > 0 ? ahead : PFFOR, false); /* * If the page is not wired down, then put it where the pageout daemon * can find it. */ if ((fs.fault_flags & VM_FAULT_WIRE) != 0) vm_page_wire(fs.m); else vm_page_activate(fs.m); if (fs.m_hold != NULL) { (*fs.m_hold) = fs.m; vm_page_wire(fs.m); } vm_page_xunbusy(fs.m); fs.m = NULL; /* * Unlock everything, and return */ vm_fault_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 ((fs.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; vm_offset_t end, start; vm_page_t m, m_next; vm_pindex_t pend, pstart; vm_size_t size; VM_OBJECT_ASSERT_UNLOCKED(fs->object); first_object = fs->first_object; /* Neither fictitious nor unmanaged pages can be reclaimed. */ if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) { VM_OBJECT_RLOCK(first_object); 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 (!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. The test for whether the page * is in the inactive queue is racy; in the * worst case we will requeue the page * unnecessarily. */ if (!vm_page_inactive(m)) vm_page_deactivate(m); } } VM_OBJECT_RUNLOCK(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; vm_prot_t prot; 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; } prot = entry->protection; /* * If pmap_enter() has enabled write access on a nearby mapping, then * don't attempt promotion, because it will fail. */ if ((fs->prot & VM_PROT_WRITE) != 0) prot |= VM_PROT_NO_PROMOTE; /* * 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 && !vm_fault_object_needs_getpages(lobject) && (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 (vm_page_all_valid(m) && (m->flags & PG_FICTITIOUS) == 0) pmap_enter_quick(pmap, addr, m, prot); 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); if (!vm_map_range_valid(map, addr, end)) 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 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 __unused, 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; bool upgrade; upgrade = src_entry == dst_entry; KASSERT(upgrade || dst_entry->object.vm_object == NULL, ("vm_fault_copy_entry: vm_object not 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. */ access = prot = dst_entry->protection; if (!upgrade) access &= ~VM_PROT_WRITE; 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_anon(atop(dst_entry->end - dst_entry->start), NULL, NULL, 0); #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; dst_entry->object.vm_object = dst_object; dst_entry->offset = 0; dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC; } VM_OBJECT_WLOCK(dst_object); 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->flags & OBJ_SWAP) != 0 && 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; } /* * 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; } /* * See the comment in vm_fault_cow(). */ if (src_object == dst_object && (object->flags & OBJ_ONEMAPPING) == 0) pmap_remove_all(src_m); pmap_copy_page(src_m, dst_m); /* * The object lock does not guarantee that "src_m" will * transition from invalid to valid, but it does ensure * that "src_m" will not transition from valid to * invalid. */ dst_m->dirty = dst_m->valid = src_m->valid; VM_OBJECT_RUNLOCK(object); } 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; } } /* * 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 (vm_page_all_valid(dst_m)) { VM_OBJECT_WUNLOCK(dst_object); pmap_enter(dst_map->pmap, vaddr, dst_m, prot, access | (upgrade ? PMAP_ENTER_WIRED : 0), 0); VM_OBJECT_WLOCK(dst_object); } /* * Mark it no longer busy, and put it on the active list. */ 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_activate(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); } diff --git a/sys/vm/vm_map.c b/sys/vm/vm_map.c index c3aa499e0dc7..f7847a244586 100644 --- a/sys/vm/vm_map.c +++ b/sys/vm/vm_map.c @@ -1,5433 +1,5432 @@ /*- * 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. * * * 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 #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 vmspace_zone; static int vmspace_zinit(void *mem, int size, 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 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 CONTAINS_BITS(set, bits) ((~(set) & (bits)) == 0) #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; \ } #ifndef UMA_USE_DMAP /* * Allocate a new slab for kernel map entries. The kernel map may be locked or * unlocked, depending on whether the request is coming from the kernel map or a * submap. This function allocates a virtual address range directly from the * kernel map instead of the kmem_* layer to avoid recursion on the kernel map * lock and also to avoid triggering allocator recursion in the vmem boundary * tag allocator. */ static void * kmapent_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, int wait) { vm_offset_t addr; int error, locked; *pflag = UMA_SLAB_PRIV; if (!(locked = vm_map_locked(kernel_map))) vm_map_lock(kernel_map); addr = vm_map_findspace(kernel_map, vm_map_min(kernel_map), bytes); if (addr + bytes < addr || addr + bytes > vm_map_max(kernel_map)) panic("%s: kernel map is exhausted", __func__); error = vm_map_insert(kernel_map, NULL, 0, addr, addr + bytes, VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT); if (error != KERN_SUCCESS) panic("%s: vm_map_insert() failed: %d", __func__, error); if (!locked) vm_map_unlock(kernel_map); error = kmem_back_domain(domain, kernel_object, addr, bytes, M_NOWAIT | M_USE_RESERVE | (wait & M_ZERO)); if (error == KERN_SUCCESS) { return ((void *)addr); } else { if (!locked) vm_map_lock(kernel_map); vm_map_delete(kernel_map, addr, bytes); if (!locked) vm_map_unlock(kernel_map); return (NULL); } } static void kmapent_free(void *item, vm_size_t size, uint8_t pflag) { vm_offset_t addr; int error __diagused; if ((pflag & UMA_SLAB_PRIV) == 0) /* XXX leaked */ return; addr = (vm_offset_t)item; kmem_unback(kernel_object, addr, size); error = vm_map_remove(kernel_map, addr, addr + size); KASSERT(error == KERN_SUCCESS, ("%s: vm_map_remove failed: %d", __func__, error)); } /* * The worst-case upper bound on the number of kernel map entries that may be * created before the zone must be replenished in _vm_map_unlock(). */ #define KMAPENT_RESERVE 1 #endif /* !UMD_MD_SMALL_ALLOC */ /* * vm_map_startup: * * Initialize the vm_map module. Must be called before any other vm_map * routines. * * User map and entry structures are allocated from the general purpose * memory pool. Kernel maps are statically defined. Kernel map entries * require special handling to avoid recursion; see the comments above * kmapent_alloc() and in vm_map_entry_create(). */ void vm_map_startup(void) { mtx_init(&map_sleep_mtx, "vm map sleep mutex", NULL, MTX_DEF); /* * Disable the use of per-CPU buckets: map entry allocation is * serialized by the kernel map lock. */ kmapentzone = uma_zcreate("KMAP ENTRY", sizeof(struct vm_map_entry), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM | UMA_ZONE_NOBUCKET); #ifndef UMA_USE_DMAP /* Reserve an extra map entry for use when replenishing the reserve. */ uma_zone_reserve(kmapentzone, KMAPENT_RESERVE + 1); uma_prealloc(kmapentzone, KMAPENT_RESERVE + 1); uma_zone_set_allocf(kmapentzone, kmapent_alloc); uma_zone_set_freef(kmapentzone, kmapent_free); #endif 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_map_t map; vm = (struct vmspace *)mem; map = &vm->vm_map; memset(map, 0, sizeof(*map)); /* set map->system_map = false */ sx_init(&map->lock, "vm map (user)"); PMAP_LOCK_INIT(vmspace_pmap(vm)); return (0); } #ifdef INVARIANTS static void vmspace_zdtor(void *mem, int size, void *arg) { struct vmspace *vm; vm = (struct vmspace *)mem; KASSERT(vm->vm_map.nentries == 0, ("vmspace %p nentries == %d on free", vm, vm->vm_map.nentries)); KASSERT(vm->vm_map.size == 0, ("vmspace %p size == %ju on free", vm, (uintmax_t)vm->vm_map.size)); } #endif /* INVARIANTS */ /* * Allocate a vmspace structure, including a vm_map and pmap, * and initialize those structures. The refcnt is set to 1. */ 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); refcount_init(&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 (refcount_release(&vm->vm_refcnt)) 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) { struct vmspace *vm; struct proc *p; bool released; p = td->td_proc; vm = p->p_vmspace; /* * Prepare to release the vmspace reference. The thread that releases * the last reference is responsible for tearing down the vmspace. * However, threads not releasing the final reference must switch to the * kernel's vmspace0 before the decrement so that the subsequent pmap * deactivation does not modify a freed vmspace. */ refcount_acquire(&vmspace0.vm_refcnt); if (!(released = refcount_release_if_last(&vm->vm_refcnt))) { if (p->p_vmspace != &vmspace0) { PROC_VMSPACE_LOCK(p); p->p_vmspace = &vmspace0; PROC_VMSPACE_UNLOCK(p); pmap_activate(td); } released = refcount_release(&vm->vm_refcnt); } if (released) { /* * pmap_remove_pages() expects the pmap to be active, so switch * back first if necessary. */ if (p->p_vmspace != vm) { PROC_VMSPACE_LOCK(p); p->p_vmspace = vm; PROC_VMSPACE_UNLOCK(p); pmap_activate(td); } pmap_remove_pages(vmspace_pmap(vm)); 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; PROC_VMSPACE_LOCK(p); vm = p->p_vmspace; if (vm == NULL || !refcount_acquire_if_not_zero(&vm->vm_refcnt)) { PROC_VMSPACE_UNLOCK(p); return (NULL); } 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(refcount_load(&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; refcount_acquire(&newvm->vm_refcnt); /* 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) + if (vm_map_is_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; struct vnode *vp; bool vp_held; 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)); if ((object->flags & OBJ_ANON) != 0) object = object->handle; else KASSERT(object->backing_object == NULL, ("non-anon object %p shadows", object)); KASSERT(object != NULL, ("No content object for text, entry %p obj %p", entry, entry->object.vm_object)); /* * Mostly, we do not lock the backing object. It is * referenced by the entry we are processing, so it cannot go * away. */ vm_pager_getvp(object, &vp, &vp_held); if (vp != NULL) { if (add) { VOP_SET_TEXT_CHECKED(vp); } else { vn_lock(vp, LK_SHARED | LK_RETRY); VOP_UNSET_TEXT_CHECKED(vp); VOP_UNLOCK(vp); } if (vp_held) vdrop(vp); } } /* * Use a different name for this vm_map_entry field when it's use * is not consistent with its use as part of an ordered search tree. */ #define defer_next right 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->defer_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; } } #ifdef INVARIANTS static void _vm_map_assert_locked(vm_map_t map, const char *file, int line) { - if (map->system_map) + if (vm_map_is_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) enum { VMMAP_CHECK_NONE, VMMAP_CHECK_UNLOCK, VMMAP_CHECK_ALL }; #ifdef DIAGNOSTIC static int enable_vmmap_check = VMMAP_CHECK_UNLOCK; #else static int enable_vmmap_check = VMMAP_CHECK_NONE; #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, int check); #define VM_MAP_ASSERT_CONSISTENT(map) \ _vm_map_assert_consistent(map, VMMAP_CHECK_ALL) #ifdef DIAGNOSTIC #define VM_MAP_UNLOCK_CONSISTENT(map) do { \ if (map->nupdates > map->nentries) { \ _vm_map_assert_consistent(map, VMMAP_CHECK_UNLOCK); \ map->nupdates = 0; \ } \ } while (0) #else #define VM_MAP_UNLOCK_CONSISTENT(map) #endif #else #define VM_MAP_ASSERT_LOCKED(map) #define VM_MAP_ASSERT_CONSISTENT(map) #define VM_MAP_UNLOCK_CONSISTENT(map) #endif /* INVARIANTS */ void _vm_map_unlock(vm_map_t map, const char *file, int line) { VM_MAP_UNLOCK_CONSISTENT(map); - if (map->system_map) { + if (vm_map_is_system(map)) { #ifndef UMA_USE_DMAP if (map == kernel_map && (map->flags & MAP_REPLENISH) != 0) { uma_prealloc(kmapentzone, 1); map->flags &= ~MAP_REPLENISH; } #endif 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) + if (vm_map_is_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) { + if (vm_map_is_system(map)) { KASSERT((map->flags & MAP_REPLENISH) == 0, ("%s: MAP_REPLENISH leaked", __func__)); 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 ? + error = vm_map_is_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 ? + error = vm_map_is_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) { + if (vm_map_is_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) { + if (vm_map_is_system(map)) { KASSERT((map->flags & MAP_REPLENISH) == 0, ("%s: MAP_REPLENISH leaked", __func__)); mtx_assert_(&map->system_mtx, MA_OWNED, file, line); } else { VM_MAP_UNLOCK_CONSISTENT(map); 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) + if (vm_map_is_system(map)) return (mtx_owned(&map->system_mtx)); - else - return (sx_xlocked(&map->lock)); + return (sx_xlocked(&map->lock)); } /* * _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) { VM_MAP_UNLOCK_CONSISTENT(map); mtx_lock(&map_sleep_mtx); - if (map->system_map) { + if (vm_map_is_system(map)) { KASSERT((map->flags & MAP_REPLENISH) == 0, ("%s: MAP_REPLENISH leaked", __func__)); 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) + if (vm_map_is_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)); } /* * 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.eflags = MAP_ENTRY_HEADER; map->needs_wakeup = FALSE; map->pmap = pmap; map->header.end = min; map->header.start = max; map->flags = 0; map->header.left = map->header.right = &map->header; map->root = NULL; map->timestamp = 0; map->busy = 0; map->anon_loc = 0; #ifdef DIAGNOSTIC map->nupdates = 0; #endif } 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); map->system_map = false; sx_init(&map->lock, "vm map (user)"); } void vm_map_init_system(vm_map_t map, pmap_t pmap, vm_offset_t min, vm_offset_t max) { _vm_map_init(map, pmap, min, max); map->system_map = true; mtx_init(&map->system_mtx, "vm map (system)", NULL, MTX_DEF | MTX_DUPOK); } /* * 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); + uma_zfree(vm_map_is_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; #ifndef UMA_USE_DMAP if (map == kernel_map) { VM_MAP_ASSERT_LOCKED(map); /* * A new slab of kernel map entries cannot be allocated at this * point because the kernel map has not yet been updated to * reflect the caller's request. Therefore, we allocate a new * map entry, dipping into the reserve if necessary, and set a * flag indicating that the reserve must be replenished before * the map is unlocked. */ new_entry = uma_zalloc(kmapentzone, M_NOWAIT | M_NOVM); if (new_entry == NULL) { new_entry = uma_zalloc(kmapentzone, M_NOWAIT | M_NOVM | M_USE_RESERVE); kernel_map->flags |= MAP_REPLENISH; } } else #endif - if (map->system_map) { + if (vm_map_is_system(map)) { new_entry = uma_zalloc(kmapentzone, M_NOWAIT); } else { new_entry = uma_zalloc(mapentzone, M_WAITOK); } KASSERT(new_entry != NULL, ("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 != left_ancestor ? 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 != right_ancestor ? root->right->max_free : right_ancestor->start - root->end); } /* * vm_map_entry_{pred,succ}: * * Find the {predecessor, successor} of the entry by taking one step * in the appropriate direction and backtracking as much as necessary. * vm_map_entry_succ is defined in vm_map.h. */ static inline vm_map_entry_t vm_map_entry_pred(vm_map_entry_t entry) { vm_map_entry_t prior; prior = entry->left; if (prior->right->start < entry->start) { do prior = prior->right; while (prior->right != entry); } return (prior); } static inline vm_size_t vm_size_max(vm_size_t a, vm_size_t b) { return (a > b ? a : b); } #define SPLAY_LEFT_STEP(root, y, llist, rlist, test) do { \ vm_map_entry_t z; \ 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_size_max( \ vm_map_entry_max_free_left(root, llist), \ vm_map_entry_max_free_right(root, rlist)), \ ("%s: max_free invariant fails", __func__)); \ if (max_free - 1 < vm_map_entry_max_free_left(root, llist)) \ max_free = vm_map_entry_max_free_right(root, rlist); \ if (y != llist && (test)) { \ /* Rotate right and make y root. */ \ z = y->right; \ if (z != root) { \ root->left = z; \ y->right = root; \ if (max_free < y->max_free) \ root->max_free = max_free = \ vm_size_max(max_free, z->max_free); \ } else if (max_free < y->max_free) \ root->max_free = max_free = \ vm_size_max(max_free, root->start - y->end);\ 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 != llist ? y : NULL; \ } while (0) #define SPLAY_RIGHT_STEP(root, y, llist, rlist, test) do { \ vm_map_entry_t z; \ 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_size_max( \ vm_map_entry_max_free_left(root, llist), \ vm_map_entry_max_free_right(root, rlist)), \ ("%s: max_free invariant fails", __func__)); \ if (max_free - 1 < vm_map_entry_max_free_right(root, rlist)) \ max_free = vm_map_entry_max_free_left(root, llist); \ if (y != rlist && (test)) { \ /* Rotate left and make y root. */ \ z = y->left; \ if (z != root) { \ root->right = z; \ y->left = root; \ if (max_free < y->max_free) \ root->max_free = max_free = \ vm_size_max(max_free, z->max_free); \ } else if (max_free < y->max_free) \ root->max_free = max_free = \ vm_size_max(max_free, y->start - root->end);\ 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 != rlist ? y : NULL; \ } while (0) /* * Walk down the tree until we find addr or a gap where addr would go, breaking * off left and right subtrees of nodes less than, or greater than addr. Treat * subtrees with root->max_free < length as empty trees. 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_{left,right,pred,succ}, rely on the start and end address * values in &map->header. */ static __always_inline vm_map_entry_t vm_map_splay_split(vm_map_t map, vm_offset_t addr, vm_size_t length, vm_map_entry_t *llist, vm_map_entry_t *rlist) { vm_map_entry_t left, right, root, y; left = right = &map->header; root = map->root; while (root != NULL && root->max_free >= length) { KASSERT(left->end <= root->start && root->end <= right->start, ("%s: root not within tree bounds", __func__)); if (addr < root->start) { SPLAY_LEFT_STEP(root, y, left, right, y->max_free >= length && addr < y->start); } else if (addr >= root->end) { SPLAY_RIGHT_STEP(root, y, left, right, y->max_free >= length && addr >= y->end); } else break; } *llist = left; *rlist = right; return (root); } static __always_inline void vm_map_splay_findnext(vm_map_entry_t root, vm_map_entry_t *rlist) { vm_map_entry_t hi, right, y; right = *rlist; hi = root->right == right ? NULL : root->right; if (hi == NULL) return; do SPLAY_LEFT_STEP(hi, y, root, right, true); while (hi != NULL); *rlist = right; } static __always_inline void vm_map_splay_findprev(vm_map_entry_t root, vm_map_entry_t *llist) { vm_map_entry_t left, lo, y; left = *llist; lo = root->left == left ? NULL : root->left; if (lo == NULL) return; do SPLAY_RIGHT_STEP(lo, y, left, root, true); while (lo != NULL); *llist = left; } 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 vm_size_t vm_map_splay_merge_left_walk(vm_map_entry_t header, vm_map_entry_t root, vm_map_entry_t tail, vm_size_t max_free, vm_map_entry_t llist) { do { /* * The max_free values of the children of llist are in * llist->max_free and max_free. Update with the * max value. */ llist->max_free = max_free = vm_size_max(llist->max_free, max_free); vm_map_entry_swap(&llist->right, &tail); vm_map_entry_swap(&tail, &llist); } while (llist != header); root->left = tail; return (max_free); } /* * When llist is known to be the predecessor of root. */ static inline vm_size_t vm_map_splay_merge_pred(vm_map_entry_t header, vm_map_entry_t root, vm_map_entry_t llist) { vm_size_t max_free; max_free = root->start - llist->end; if (llist != header) { max_free = vm_map_splay_merge_left_walk(header, root, root, max_free, llist); } else { root->left = header; header->right = root; } return (max_free); } /* * When llist may or may not be the predecessor of root. */ static inline vm_size_t vm_map_splay_merge_left(vm_map_entry_t header, vm_map_entry_t root, vm_map_entry_t llist) { vm_size_t max_free; max_free = vm_map_entry_max_free_left(root, llist); if (llist != header) { max_free = vm_map_splay_merge_left_walk(header, root, root->left == llist ? root : root->left, max_free, llist); } return (max_free); } static vm_size_t vm_map_splay_merge_right_walk(vm_map_entry_t header, vm_map_entry_t root, vm_map_entry_t tail, vm_size_t max_free, vm_map_entry_t rlist) { do { /* * The max_free values of the children of rlist are in * rlist->max_free and max_free. Update with the * max value. */ rlist->max_free = max_free = vm_size_max(rlist->max_free, max_free); vm_map_entry_swap(&rlist->left, &tail); vm_map_entry_swap(&tail, &rlist); } while (rlist != header); root->right = tail; return (max_free); } /* * When rlist is known to be the succecessor of root. */ static inline vm_size_t vm_map_splay_merge_succ(vm_map_entry_t header, vm_map_entry_t root, vm_map_entry_t rlist) { vm_size_t max_free; max_free = rlist->start - root->end; if (rlist != header) { max_free = vm_map_splay_merge_right_walk(header, root, root, max_free, rlist); } else { root->right = header; header->left = root; } return (max_free); } /* * When rlist may or may not be the succecessor of root. */ static inline vm_size_t vm_map_splay_merge_right(vm_map_entry_t header, vm_map_entry_t root, vm_map_entry_t rlist) { vm_size_t max_free; max_free = vm_map_entry_max_free_right(root, rlist); if (rlist != header) { max_free = vm_map_splay_merge_right_walk(header, root, root->right == rlist ? root : root->right, max_free, rlist); } return (max_free); } /* * 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 tree is threaded, which means that there are no null pointers. * When a node has no left child, its left pointer points to its * predecessor, which the last ancestor on the search path from the root * where the search branched right. Likewise, when a node has no right * child, its right pointer points to its successor. The map header node * is the predecessor of the first map entry, and the successor of the * last. * * 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 header, llist, rlist, root; vm_size_t max_free_left, max_free_right; header = &map->header; root = vm_map_splay_split(map, addr, 0, &llist, &rlist); if (root != NULL) { max_free_left = vm_map_splay_merge_left(header, root, llist); max_free_right = vm_map_splay_merge_right(header, root, rlist); } else if (llist != header) { /* * Recover the greatest node in the left * subtree and make it the root. */ root = llist; llist = root->right; max_free_left = vm_map_splay_merge_left(header, root, llist); max_free_right = vm_map_splay_merge_succ(header, root, rlist); } else if (rlist != header) { /* * Recover the least node in the right * subtree and make it the root. */ root = rlist; rlist = root->left; max_free_left = vm_map_splay_merge_pred(header, root, llist); max_free_right = vm_map_splay_merge_right(header, root, rlist); } else { /* There is no root. */ return (NULL); } root->max_free = vm_size_max(max_free_left, max_free_right); map->root = root; VM_MAP_ASSERT_CONSISTENT(map); return (root); } /* * vm_map_entry_{un,}link: * * Insert/remove entries from maps. On linking, if new entry clips * existing entry, trim existing entry to avoid overlap, and manage * offsets. On unlinking, merge disappearing entry with neighbor, if * called for, and manage offsets. Callers should not modify fields in * entries already mapped. */ static void vm_map_entry_link(vm_map_t map, vm_map_entry_t entry) { vm_map_entry_t header, llist, rlist, root; vm_size_t max_free_left, max_free_right; CTR3(KTR_VM, "vm_map_entry_link: map %p, nentries %d, entry %p", map, map->nentries, entry); VM_MAP_ASSERT_LOCKED(map); map->nentries++; header = &map->header; root = vm_map_splay_split(map, entry->start, 0, &llist, &rlist); if (root == NULL) { /* * The new entry does not overlap any existing entry in the * map, so it becomes the new root of the map tree. */ max_free_left = vm_map_splay_merge_pred(header, entry, llist); max_free_right = vm_map_splay_merge_succ(header, entry, rlist); } else if (entry->start == root->start) { /* * The new entry is a clone of root, with only the end field * changed. The root entry will be shrunk to abut the new * entry, and will be the right child of the new root entry in * the modified map. */ KASSERT(entry->end < root->end, ("%s: clip_start not within entry", __func__)); vm_map_splay_findprev(root, &llist); if ((root->eflags & MAP_ENTRY_STACK_GAP) == 0) root->offset += entry->end - root->start; root->start = entry->end; max_free_left = vm_map_splay_merge_pred(header, entry, llist); max_free_right = root->max_free = vm_size_max( vm_map_splay_merge_pred(entry, root, entry), vm_map_splay_merge_right(header, root, rlist)); } else { /* * The new entry is a clone of root, with only the start field * changed. The root entry will be shrunk to abut the new * entry, and will be the left child of the new root entry in * the modified map. */ KASSERT(entry->end == root->end, ("%s: clip_start not within entry", __func__)); vm_map_splay_findnext(root, &rlist); if ((entry->eflags & MAP_ENTRY_STACK_GAP) == 0) entry->offset += entry->start - root->start; root->end = entry->start; max_free_left = root->max_free = vm_size_max( vm_map_splay_merge_left(header, root, llist), vm_map_splay_merge_succ(entry, root, entry)); max_free_right = vm_map_splay_merge_succ(header, entry, rlist); } entry->max_free = vm_size_max(max_free_left, max_free_right); map->root = entry; 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 header, llist, rlist, root; vm_size_t max_free_left, max_free_right; VM_MAP_ASSERT_LOCKED(map); header = &map->header; 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_findprev(root, &llist); vm_map_splay_findnext(root, &rlist); if (op == UNLINK_MERGE_NEXT) { rlist->start = root->start; MPASS((rlist->eflags & MAP_ENTRY_STACK_GAP) == 0); rlist->offset = root->offset; } if (llist != header) { root = llist; llist = root->right; max_free_left = vm_map_splay_merge_left(header, root, llist); max_free_right = vm_map_splay_merge_succ(header, root, rlist); } else if (rlist != header) { root = rlist; rlist = root->left; max_free_left = vm_map_splay_merge_pred(header, root, llist); max_free_right = vm_map_splay_merge_right(header, root, rlist); } else { header->left = header->right = header; root = NULL; } if (root != NULL) root->max_free = vm_size_max(max_free_left, max_free_right); map->root = root; 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 header, llist, rlist, root; VM_MAP_ASSERT_LOCKED(map); header = &map->header; 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); entry->end += grow_amount; root->max_free = vm_size_max( vm_map_splay_merge_left(header, root, llist), vm_map_splay_merge_succ(header, root, rlist)); map->root = root; 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, header, lbound, ubound; boolean_t locked; /* * If the map is empty, then the map entry immediately preceding * "address" is the map's header. */ header = &map->header; cur = map->root; if (cur == NULL) { *entry = 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) { VM_MAP_UNLOCK_CONSISTENT(map); 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 = 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 = ubound = header; for (;;) { if (address < cur->start) { ubound = cur; cur = cur->left; if (cur == lbound) break; } else if (cur->end <= address) { lbound = cur; cur = cur->right; if (cur == ubound) break; } else { *entry = cur; return (TRUE); } } *entry = lbound; return (FALSE); } /* * vm_map_insert1() is identical to vm_map_insert() except that it * returns the newly inserted map entry in '*res'. In case the new * entry is coalesced with a neighbor or an existing entry was * resized, that entry is returned. In any case, the returned entry * covers the specified address range. */ static int vm_map_insert1(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 *res) { vm_map_entry_t new_entry, next_entry, prev_entry; struct ucred *cred; vm_eflags_t protoeflags; vm_inherit_t inheritance; u_long bdry; u_int bidx; 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 || (cow & MAP_SPLIT_BOUNDARY_MASK) != 0, ("vm_map_insert: paradoxical MAP_NOFAULT request, obj %p cow %#x", object, cow)); 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 == end || !vm_map_range_valid(map, start, end)) return (KERN_INVALID_ADDRESS); if ((map->flags & MAP_WXORX) != 0 && (prot & (VM_PROT_WRITE | VM_PROT_EXECUTE)) == (VM_PROT_WRITE | VM_PROT_EXECUTE)) return (KERN_PROTECTION_FAILURE); /* * 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. */ next_entry = vm_map_entry_succ(prev_entry); if (next_entry->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_AREA) protoeflags |= MAP_ENTRY_GROWS_DOWN; 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) != 0) protoeflags |= MAP_ENTRY_STACK_GAP; if (cow & MAP_INHERIT_SHARE) inheritance = VM_INHERIT_SHARE; else inheritance = VM_INHERIT_DEFAULT; if ((cow & MAP_SPLIT_BOUNDARY_MASK) != 0) { /* This magically ignores index 0, for usual page size. */ bidx = (cow & MAP_SPLIT_BOUNDARY_MASK) >> MAP_SPLIT_BOUNDARY_SHIFT; if (bidx >= MAXPAGESIZES) return (KERN_INVALID_ARGUMENT); bdry = pagesizes[bidx] - 1; if ((start & bdry) != 0 || (end & bdry) != 0) return (KERN_INVALID_ARGUMENT); protoeflags |= bidx << MAP_ENTRY_SPLIT_BOUNDARY_SHIFT; } 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.) */ if ((object->flags & OBJ_ANON) != 0) { 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_AREA | 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); *res = vm_map_try_merge_entries(map, prev_entry, next_entry); 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); *res = vm_map_try_merge_entries(map, new_entry, next_entry); 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_insert: * * Inserts the given VM object into the target map at the * specified 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 res; return (vm_map_insert1(map, object, offset, start, end, prot, max, cow, &res)); } /* * 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 header, llist, rlist, root, y; vm_size_t left_length, max_free_left, max_free_right; vm_offset_t gap_end; VM_MAP_ASSERT_LOCKED(map); /* * 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. */ header = &map->header; root = vm_map_splay_split(map, start, length, &llist, &rlist); gap_end = rlist->start; if (root != NULL) { start = root->end; if (root->right != rlist) gap_end = start; max_free_left = vm_map_splay_merge_left(header, root, llist); max_free_right = vm_map_splay_merge_right(header, root, rlist); } else if (rlist != header) { root = rlist; rlist = root->left; max_free_left = vm_map_splay_merge_pred(header, root, llist); max_free_right = vm_map_splay_merge_right(header, root, rlist); } else { root = llist; llist = root->right; max_free_left = vm_map_splay_merge_left(header, root, llist); max_free_right = vm_map_splay_merge_succ(header, root, rlist); } root->max_free = vm_size_max(max_free_left, max_free_right); map->root = root; VM_MAP_ASSERT_CONSISTENT(map); if (length <= gap_end - start) return (start); /* With max_free, can immediately tell if no solution. */ if (root->right == header || 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 = header; for (left_length = 0;; left_length = vm_map_entry_max_free_left(root, llist)) { if (length <= left_length) SPLAY_LEFT_STEP(root, y, llist, rlist, length <= vm_map_entry_max_free_left(y, llist)); else SPLAY_RIGHT_STEP(root, y, llist, rlist, length > vm_map_entry_max_free_left(y, root)); if (root == NULL) break; } root = llist; llist = root->right; max_free_left = vm_map_splay_merge_left(header, root, llist); if (rlist == header) { root->max_free = vm_size_max(max_free_left, vm_map_splay_merge_succ(header, root, rlist)); } else { y = rlist; rlist = y->left; y->max_free = vm_size_max( vm_map_splay_merge_pred(root, y, root), vm_map_splay_merge_right(header, y, rlist)); root->max_free = vm_size_max(max_free_left, y->max_free); } map->root = root; 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_AREA) == 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) { result = vm_map_delete(map, start, end); if (result != KERN_SUCCESS) goto out; } if ((cow & MAP_STACK_AREA) != 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); } out: vm_map_unlock(map); return (result); } #if VM_NRESERVLEVEL <= 1 static const int aslr_pages_rnd_64[2] = {0x1000, 0x10}; static const int aslr_pages_rnd_32[2] = {0x100, 0x4}; #elif VM_NRESERVLEVEL == 2 static const int aslr_pages_rnd_64[3] = {0x1000, 0x1000, 0x10}; static const int aslr_pages_rnd_32[3] = {0x100, 0x100, 0x4}; #else #error "Unsupported VM_NRESERVLEVEL" #endif 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, int cow) { switch (cluster_anon) { case 0: return (false); case 1: return (addr == 0 || (cow & MAP_NO_HINT) != 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"); /* * 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 *addr = roundup2(*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); } } } int vm_map_find_aligned(vm_map_t map, vm_offset_t *addr, vm_size_t length, vm_offset_t max_addr, vm_offset_t alignment) { /* XXXKIB ASLR eh ? */ *addr = vm_map_findspace(map, *addr, length); if (*addr + length > vm_map_max(map) || (max_addr != 0 && *addr + length > max_addr)) return (KERN_NO_SPACE); return (vm_map_alignspace(map, NULL, 0, addr, length, max_addr, alignment)); } /* * 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) { int rv; vm_map_lock(map); rv = vm_map_find_locked(map, object, offset, addr, length, max_addr, find_space, prot, max, cow); vm_map_unlock(map); return (rv); } int vm_map_find_locked(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_AREA) == 0 || object == NULL, ("non-NULL backing object for stack")); MPASS((cow & MAP_REMAP) == 0 || (find_space == VMFS_NO_SPACE && (cow & MAP_STACK_AREA) == 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, cow) && (map->flags & MAP_IS_SUB_MAP) == 0 && max_addr == 0 && find_space != VMFS_NO_SPACE && object == NULL && (cow & (MAP_INHERIT_SHARE | MAP_STACK_AREA)) == 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; 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 = 0; #if VM_NRESERVLEVEL > 0 if ((find_space == VMFS_SUPER_SPACE || find_space == VMFS_OPTIMAL_SPACE) && pagesizes[VM_NRESERVLEVEL] != 0) { /* * Do not pointlessly increase the space that * is requested from vm_map_findspace(). * pmap_align_superpage() will only change a * mapping's alignment if that mapping is at * least a superpage in size. */ pidx = VM_NRESERVLEVEL; while (pidx > 0 && length < pagesizes[pidx]) pidx--; } #endif 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; } return (KERN_NO_SPACE); } } 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; } return (rv); } } else if ((cow & MAP_REMAP) != 0) { if (!vm_map_range_valid(map, *addr, *addr + length)) return (KERN_INVALID_ADDRESS); rv = vm_map_delete(map, *addr, *addr + length); if (rv != KERN_SUCCESS) return (rv); } if ((cow & MAP_STACK_AREA) != 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); } /* * Update the starting address for clustered anonymous memory mappings * if a starting address was not previously defined or an ASLR restart * placed an anonymous memory mapping at a lower address. */ if (update_anon && rv == KERN_SUCCESS && (map->anon_loc == 0 || *addr < map->anon_loc)) map->anon_loc = *addr; return (rv); } /* * vm_map_find_min() is a variant of vm_map_find() that takes an * additional parameter ("default_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 "default_addr", it performs a second pass, replacing * the hint with "default_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 default_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; if (hint == 0) { cow |= MAP_NO_HINT; *addr = hint = default_addr; } for (;;) { rv = vm_map_find(map, object, offset, addr, length, max_addr, find_space, prot, max, cow); if (rv == KERN_SUCCESS || default_addr >= hint) return (rv); *addr = hint = default_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_IN_TRANSITION | MAP_ENTRY_IS_SUB_MAP | MAP_ENTRY_VN_EXEC | \ MAP_ENTRY_STACK_GAP) 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 two map entries that represent consecutive ranges. If * the entries can be merged, expand the range of the second to * cover the range of the first and delete the first. Then return * the map entry that includes the first range. * * The map must be locked. */ vm_map_entry_t vm_map_try_merge_entries(vm_map_t map, vm_map_entry_t prev_entry, vm_map_entry_t entry) { VM_MAP_ASSERT_LOCKED(map); if ((entry->eflags & MAP_ENTRY_NOMERGE_MASK) == 0 && vm_map_mergeable_neighbors(prev_entry, entry)) { vm_map_entry_unlink(map, prev_entry, UNLINK_MERGE_NEXT); vm_map_merged_neighbor_dispose(map, prev_entry); return (entry); } return (prev_entry); } /* * 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_anon(atop(entry->end - entry->start), NULL, entry->cred, entry->end - entry->start); entry->object.vm_object = object; entry->offset = 0; 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 && + if (entry->object.vm_object == NULL && !vm_map_is_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_entry_clone * * Create a duplicate map entry for clipping. */ static vm_map_entry_t vm_map_entry_clone(vm_map_t map, vm_map_entry_t entry) { vm_map_entry_t new_entry; VM_MAP_ASSERT_LOCKED(map); /* * 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; if (new_entry->cred != NULL) crhold(entry->cred); 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. */ } return (new_entry); } /* * 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. */ static int vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t startaddr) { vm_map_entry_t new_entry; int bdry_idx; - if (!map->system_map) + if (!vm_map_is_system(map)) WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "%s: map %p entry %p start 0x%jx", __func__, map, entry, (uintmax_t)startaddr); if (startaddr <= entry->start) return (KERN_SUCCESS); VM_MAP_ASSERT_LOCKED(map); KASSERT(entry->end > startaddr && entry->start < startaddr, ("%s: invalid clip of entry %p", __func__, entry)); bdry_idx = MAP_ENTRY_SPLIT_BOUNDARY_INDEX(entry); if (bdry_idx != 0) { if ((startaddr & (pagesizes[bdry_idx] - 1)) != 0) return (KERN_INVALID_ARGUMENT); } new_entry = vm_map_entry_clone(map, 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 = startaddr; vm_map_entry_link(map, new_entry); return (KERN_SUCCESS); } /* * vm_map_lookup_clip_start: * * Find the entry at or just after 'start', and clip it if 'start' is in * the interior of the entry. Return entry after 'start', and in * prev_entry set the entry before 'start'. */ static int vm_map_lookup_clip_start(vm_map_t map, vm_offset_t start, vm_map_entry_t *res_entry, vm_map_entry_t *prev_entry) { vm_map_entry_t entry; int rv; - if (!map->system_map) + if (!vm_map_is_system(map)) WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "%s: map %p start 0x%jx prev %p", __func__, map, (uintmax_t)start, prev_entry); if (vm_map_lookup_entry(map, start, prev_entry)) { entry = *prev_entry; rv = vm_map_clip_start(map, entry, start); if (rv != KERN_SUCCESS) return (rv); *prev_entry = vm_map_entry_pred(entry); } else entry = vm_map_entry_succ(*prev_entry); *res_entry = entry; return (KERN_SUCCESS); } /* * 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. */ static int vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t endaddr) { vm_map_entry_t new_entry; int bdry_idx; - if (!map->system_map) + if (!vm_map_is_system(map)) WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "%s: map %p entry %p end 0x%jx", __func__, map, entry, (uintmax_t)endaddr); if (endaddr >= entry->end) return (KERN_SUCCESS); VM_MAP_ASSERT_LOCKED(map); KASSERT(entry->start < endaddr && entry->end > endaddr, ("%s: invalid clip of entry %p", __func__, entry)); bdry_idx = MAP_ENTRY_SPLIT_BOUNDARY_INDEX(entry); if (bdry_idx != 0) { if ((endaddr & (pagesizes[bdry_idx] - 1)) != 0) return (KERN_INVALID_ARGUMENT); } new_entry = vm_map_entry_clone(map, 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 = endaddr; vm_map_entry_link(map, new_entry); return (KERN_SUCCESS); } /* * 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) && entry->end >= end && (entry->eflags & MAP_ENTRY_COW) == 0 && entry->object.vm_object == NULL) { result = vm_map_clip_start(map, entry, start); if (result != KERN_SUCCESS) goto unlock; result = vm_map_clip_end(map, entry, end); if (result != KERN_SUCCESS) goto unlock; entry->object.sub_map = submap; entry->eflags |= MAP_ENTRY_IS_SUB_MAP; result = KERN_SUCCESS; } unlock: 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; int psind; if ((prot & (VM_PROT_READ | VM_PROT_EXECUTE)) == 0 || object == NULL) return; if (object->type == OBJT_DEVICE || object->type == OBJT_SG) { 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); } else VM_OBJECT_RLOCK(object); psize = atop(size); if (psize + pindex > object->size) { if (pindex >= object->size) { 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 (vm_page_all_valid(p)) { if (p_start == NULL) { start = addr + ptoa(tmpidx); p_start = p; } /* Jump ahead if a superpage mapping is possible. */ for (psind = p->psind; psind > 0; psind--) { if (((addr + ptoa(tmpidx)) & (pagesizes[psind] - 1)) == 0) { mask = atop(pagesizes[psind]) - 1; if (tmpidx + mask < psize && vm_page_ps_test(p, psind, PS_ALL_VALID, NULL)) { p += mask; threshold += mask; break; } } } } 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); } static void vm_map_protect_guard(vm_map_entry_t entry, vm_prot_t new_prot, vm_prot_t new_maxprot, int flags) { vm_prot_t old_prot; MPASS((entry->eflags & MAP_ENTRY_GUARD) != 0); if ((entry->eflags & MAP_ENTRY_STACK_GAP) == 0) return; old_prot = PROT_EXTRACT(entry->offset); if ((flags & VM_MAP_PROTECT_SET_MAXPROT) != 0) { entry->offset = PROT_MAX(new_maxprot) | (new_maxprot & old_prot); } if ((flags & VM_MAP_PROTECT_SET_PROT) != 0) { entry->offset = new_prot | PROT_MAX( PROT_MAX_EXTRACT(entry->offset)); } } /* * vm_map_protect: * * Sets the protection and/or the maximum protection of the * specified address region in the target map. */ int vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end, vm_prot_t new_prot, vm_prot_t new_maxprot, int flags) { vm_map_entry_t entry, first_entry, in_tran, prev_entry; vm_object_t obj; struct ucred *cred; vm_offset_t orig_start; vm_prot_t check_prot, max_prot, old_prot; int rv; if (start == end) return (KERN_SUCCESS); if (CONTAINS_BITS(flags, VM_MAP_PROTECT_SET_PROT | VM_MAP_PROTECT_SET_MAXPROT) && !CONTAINS_BITS(new_maxprot, new_prot)) return (KERN_OUT_OF_BOUNDS); orig_start = start; again: in_tran = NULL; start = orig_start; vm_map_lock(map); if ((map->flags & MAP_WXORX) != 0 && (flags & VM_MAP_PROTECT_SET_PROT) != 0 && CONTAINS_BITS(new_prot, VM_PROT_WRITE | VM_PROT_EXECUTE)) { vm_map_unlock(map); return (KERN_PROTECTION_FAILURE); } /* * 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, &first_entry)) first_entry = vm_map_entry_succ(first_entry); if ((flags & VM_MAP_PROTECT_GROWSDOWN) != 0 && (first_entry->eflags & MAP_ENTRY_GROWS_DOWN) != 0) { /* * Handle Linux's PROT_GROWSDOWN flag. * It means that protection is applied down to the * whole stack, including the specified range of the * mapped region, and the grow down region (AKA * guard). */ while (!CONTAINS_BITS(first_entry->eflags, MAP_ENTRY_GUARD | MAP_ENTRY_STACK_GAP) && first_entry != vm_map_entry_first(map)) first_entry = vm_map_entry_pred(first_entry); start = first_entry->start; } /* * Make a first pass to check for protection violations. */ check_prot = 0; if ((flags & VM_MAP_PROTECT_SET_PROT) != 0) check_prot |= new_prot; if ((flags & VM_MAP_PROTECT_SET_MAXPROT) != 0) check_prot |= new_maxprot; for (entry = first_entry; entry->start < end; entry = vm_map_entry_succ(entry)) { if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) { vm_map_unlock(map); return (KERN_INVALID_ARGUMENT); } if ((entry->eflags & (MAP_ENTRY_GUARD | MAP_ENTRY_STACK_GAP)) == MAP_ENTRY_GUARD) continue; max_prot = (entry->eflags & MAP_ENTRY_STACK_GAP) != 0 ? PROT_MAX_EXTRACT(entry->offset) : entry->max_protection; if (!CONTAINS_BITS(max_prot, check_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 have * stabilized. An in-transition entry might already have its pages * wired and wired_count incremented, but not yet have its * MAP_ENTRY_USER_WIRED flag set. In which case, we would fail to call * vm_fault_copy_entry() in the final loop below. */ 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 = vm_map_clip_start(map, first_entry, start); if (rv != KERN_SUCCESS) { vm_map_unlock(map); return (rv); } for (entry = first_entry; entry->start < end; entry = vm_map_entry_succ(entry)) { rv = vm_map_clip_end(map, entry, end); if (rv != KERN_SUCCESS) { vm_map_unlock(map); return (rv); } if ((flags & VM_MAP_PROTECT_SET_PROT) == 0 || ((new_prot & ~entry->protection) & VM_PROT_WRITE) == 0 || ENTRY_CHARGED(entry) || (entry->eflags & MAP_ENTRY_GUARD) != 0) continue; cred = curthread->td_ucred; obj = entry->object.vm_object; if (obj == NULL || (entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0) { if (!swap_reserve(entry->end - entry->start)) { rv = KERN_RESOURCE_SHORTAGE; end = entry->end; break; } crhold(cred); entry->cred = cred; continue; } VM_OBJECT_WLOCK(obj); if ((obj->flags & OBJ_SWAP) == 0) { 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, entry)); if (!swap_reserve(ptoa(obj->size))) { VM_OBJECT_WUNLOCK(obj); rv = KERN_RESOURCE_SHORTAGE; end = entry->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 (prev_entry = vm_map_entry_pred(first_entry), entry = first_entry; entry->start < end; vm_map_try_merge_entries(map, prev_entry, entry), prev_entry = entry, entry = vm_map_entry_succ(entry)) { if (rv != KERN_SUCCESS) continue; if ((entry->eflags & MAP_ENTRY_GUARD) != 0) { vm_map_protect_guard(entry, new_prot, new_maxprot, flags); continue; } old_prot = entry->protection; if ((flags & VM_MAP_PROTECT_SET_MAXPROT) != 0) { entry->max_protection = new_maxprot; entry->protection = new_maxprot & old_prot; } if ((flags & VM_MAP_PROTECT_SET_PROT) != 0) entry->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 ((entry->eflags & MAP_ENTRY_USER_WIRED) != 0 && (entry->protection & VM_PROT_WRITE) != 0 && (old_prot & VM_PROT_WRITE) == 0) vm_fault_copy_entry(map, map, entry, entry, NULL); /* * When restricting access, update the physical map. Worry * about copy-on-write here. */ if ((old_prot & ~entry->protection) != 0) { #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \ VM_PROT_ALL) pmap_protect(map->pmap, entry->start, entry->end, entry->protection & MASK(entry)); #undef MASK } } vm_map_try_merge_entries(map, prev_entry, entry); 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 entry, prev_entry; int rv; 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 (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. */ rv = vm_map_lookup_clip_start(map, start, &entry, &prev_entry); if (rv != KERN_SUCCESS) { vm_map_unlock(map); return (vm_mmap_to_errno(rv)); } for (; entry->start < end; prev_entry = entry, entry = vm_map_entry_succ(entry)) { if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) continue; rv = vm_map_clip_end(map, entry, end); if (rv != KERN_SUCCESS) { vm_map_unlock(map); return (vm_mmap_to_errno(rv)); } switch (behav) { case MADV_NORMAL: vm_map_entry_set_behavior(entry, MAP_ENTRY_BEHAV_NORMAL); break; case MADV_SEQUENTIAL: vm_map_entry_set_behavior(entry, MAP_ENTRY_BEHAV_SEQUENTIAL); break; case MADV_RANDOM: vm_map_entry_set_behavior(entry, MAP_ENTRY_BEHAV_RANDOM); break; case MADV_NOSYNC: entry->eflags |= MAP_ENTRY_NOSYNC; break; case MADV_AUTOSYNC: entry->eflags &= ~MAP_ENTRY_NOSYNC; break; case MADV_NOCORE: entry->eflags |= MAP_ENTRY_NOCOREDUMP; break; case MADV_CORE: entry->eflags &= ~MAP_ENTRY_NOCOREDUMP; break; default: break; } vm_map_try_merge_entries(map, prev_entry, entry); } vm_map_try_merge_entries(map, prev_entry, entry); 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. */ if (!vm_map_lookup_entry(map, start, &entry)) entry = vm_map_entry_succ(entry); for (; entry->start < end; entry = vm_map_entry_succ(entry)) { vm_offset_t useEnd, useStart; if ((entry->eflags & (MAP_ENTRY_IS_SUB_MAP | MAP_ENTRY_GUARD)) != 0) 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 && entry->object.vm_object != NULL && entry->object.vm_object->backing_object != NULL) continue; pstart = OFF_TO_IDX(entry->offset); pend = pstart + atop(entry->end - entry->start); useStart = entry->start; useEnd = entry->end; if (entry->start < start) { pstart += atop(start - entry->start); useStart = start; } if (entry->end > end) { pend -= atop(entry->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(entry->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 && entry->wired_count == 0) { vm_map_pmap_enter(map, useStart, entry->protection, entry->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, lentry, prev_entry, start_entry; int rv; 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); rv = vm_map_lookup_clip_start(map, start, &start_entry, &prev_entry); if (rv != KERN_SUCCESS) goto unlock; if (vm_map_lookup_entry(map, end - 1, &lentry)) { rv = vm_map_clip_end(map, lentry, end); if (rv != KERN_SUCCESS) goto unlock; } if (new_inheritance == VM_INHERIT_COPY) { for (entry = start_entry; entry->start < end; prev_entry = entry, entry = vm_map_entry_succ(entry)) { if ((entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) != 0) { rv = KERN_INVALID_ARGUMENT; goto unlock; } } } for (entry = start_entry; entry->start < end; prev_entry = entry, entry = vm_map_entry_succ(entry)) { KASSERT(entry->end <= end, ("non-clipped entry %p end %jx %jx", entry, (uintmax_t)entry->end, (uintmax_t)end)); if ((entry->eflags & MAP_ENTRY_GUARD) == 0 || new_inheritance != VM_INHERIT_ZERO) entry->inheritance = new_inheritance; vm_map_try_merge_entries(map, prev_entry, entry); } vm_map_try_merge_entries(map, prev_entry, entry); unlock: vm_map_unlock(map); return (rv); } /* * 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 = vm_map_entry_succ(entry); } 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, next_entry, prev_entry; int rv; bool 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 = vm_map_entry_succ(first_entry); else { vm_map_unlock(map); return (KERN_INVALID_ADDRESS); } } rv = KERN_SUCCESS; for (entry = first_entry; entry->start < end; entry = next_entry) { if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { /* * We have not yet clipped the entry. */ next_entry = vm_map_entry_in_transition(map, start, &end, holes_ok, entry); if (next_entry == NULL) { if (entry == first_entry) { vm_map_unlock(map); return (KERN_INVALID_ADDRESS); } rv = KERN_INVALID_ADDRESS; break; } first_entry = (entry == first_entry) ? next_entry : NULL; continue; } rv = vm_map_clip_start(map, entry, start); if (rv != KERN_SUCCESS) break; rv = vm_map_clip_end(map, entry, end); if (rv != KERN_SUCCESS) break; /* * 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; next_entry = vm_map_entry_succ(entry); /* * Check the map for holes in the specified region. * If holes_ok, skip this check. */ if (!holes_ok && entry->end < end && next_entry->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; } } need_wakeup = false; if (first_entry == NULL && !vm_map_lookup_entry(map, start, &first_entry)) { KASSERT(holes_ok, ("vm_map_unwire: lookup failed")); prev_entry = first_entry; entry = vm_map_entry_succ(first_entry); } else { prev_entry = vm_map_entry_pred(first_entry); entry = first_entry; } for (; entry->start < end; prev_entry = entry, entry = vm_map_entry_succ(entry)) { /* * 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, prev_entry, entry); } vm_map_try_merge_entries(map, prev_entry, 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, next_entry, prev_entry; vm_offset_t faddr, saved_end, saved_start; u_long incr, npages; u_int bidx, last_timestamp; int rv; bool 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 = vm_map_entry_succ(first_entry); else return (KERN_INVALID_ADDRESS); } for (entry = first_entry; entry->start < end; entry = next_entry) { if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { /* * We have not yet clipped the entry. */ next_entry = vm_map_entry_in_transition(map, start, &end, holes_ok, entry); if (next_entry == NULL) { if (entry == first_entry) return (KERN_INVALID_ADDRESS); rv = KERN_INVALID_ADDRESS; goto done; } first_entry = (entry == first_entry) ? next_entry : NULL; continue; } rv = vm_map_clip_start(map, entry, start); if (rv != KERN_SUCCESS) goto done; rv = vm_map_clip_end(map, entry, end); if (rv != KERN_SUCCESS) goto done; /* * 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; bidx = MAP_ENTRY_SPLIT_BOUNDARY_INDEX(entry); incr = pagesizes[bidx]; vm_map_busy(map); vm_map_unlock(map); for (faddr = saved_start; faddr < saved_end; faddr += incr) { /* * Simulate a fault to get the page and enter * it into the physical map. */ rv = vm_fault(map, faddr, VM_PROT_NONE, VM_FAULT_WIRE, NULL); if (rv != KERN_SUCCESS) break; } 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, &next_entry)) KASSERT(false, ("vm_map_wire: lookup failed")); first_entry = (entry == first_entry) ? next_entry : NULL; for (entry = next_entry; entry->end < saved_end; entry = vm_map_entry_succ(entry)) { /* * 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); } } 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. */ next_entry = vm_map_entry_succ(entry); if (!holes_ok && entry->end < end && next_entry->start > entry->end) { end = entry->end; rv = KERN_INVALID_ADDRESS; goto done; } } 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")); prev_entry = first_entry; entry = vm_map_entry_succ(first_entry); } else { prev_entry = vm_map_entry_pred(first_entry); entry = first_entry; } for (; entry->start < end; prev_entry = entry, entry = vm_map_entry_succ(entry)) { /* * 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. * * Another way to get an entry not marked with * MAP_ENTRY_IN_TRANSITION is after failed clipping, * which set rv to KERN_INVALID_ARGUMENT. */ if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 || entry->wiring_thread != curthread) { KASSERT(holes_ok || rv == KERN_INVALID_ARGUMENT, ("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, prev_entry, entry); } vm_map_try_merge_entries(map, prev_entry, 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 entry, first_entry, next_entry; vm_size_t size; vm_object_t object; vm_ooffset_t offset; unsigned int last_timestamp; int bdry_idx; boolean_t failed; vm_map_lock_read(map); VM_MAP_RANGE_CHECK(map, start, end); if (!vm_map_lookup_entry(map, start, &first_entry)) { vm_map_unlock_read(map); return (KERN_INVALID_ADDRESS); } else if (start == end) { start = first_entry->start; end = first_entry->end; } /* * Make a first pass to check for user-wired memory, holes, * and partial invalidation of largepage mappings. */ for (entry = first_entry; entry->start < end; entry = next_entry) { if (invalidate) { if ((entry->eflags & MAP_ENTRY_USER_WIRED) != 0) { vm_map_unlock_read(map); return (KERN_INVALID_ARGUMENT); } bdry_idx = MAP_ENTRY_SPLIT_BOUNDARY_INDEX(entry); if (bdry_idx != 0 && ((start & (pagesizes[bdry_idx] - 1)) != 0 || (end & (pagesizes[bdry_idx] - 1)) != 0)) { vm_map_unlock_read(map); return (KERN_INVALID_ARGUMENT); } } next_entry = vm_map_entry_succ(entry); if (end > entry->end && entry->end != next_entry->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 (entry = first_entry; entry->start < end;) { offset = entry->offset + (start - entry->start); size = (end <= entry->end ? end : entry->end) - start; if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) { vm_map_t smap; vm_map_entry_t tentry; vm_size_t tsize; smap = entry->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 = entry->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, &entry)) entry = vm_map_entry_succ(entry); } 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, 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); + vm_map_entry_deallocate(entry, vm_map_is_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) { entry->object.vm_object = NULL; } else if ((object->flags & OBJ_ANON) != 0 || object == kernel_object) { KASSERT(entry->cred == NULL || object->cred == NULL || (entry->eflags & MAP_ENTRY_NEEDS_COPY), ("OVERCOMMIT vm_map_entry_delete: both cred %p", entry)); offidxstart = OFF_TO_IDX(entry->offset); offidxend = offidxstart + atop(size); VM_OBJECT_WLOCK(object); if (object->ref_count != 1 && ((object->flags & OBJ_ONEMAPPING) != 0 || 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 (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); } - if (map->system_map) + if (vm_map_is_system(map)) vm_map_entry_deallocate(entry, TRUE); else { entry->defer_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, next_entry, scratch_entry; int rv; VM_MAP_ASSERT_LOCKED(map); if (start == end) return (KERN_SUCCESS); /* * Find the start of the region, and clip it. * Step through all entries in this region. */ rv = vm_map_lookup_clip_start(map, start, &entry, &scratch_entry); if (rv != KERN_SUCCESS) return (rv); for (; entry->start < end; entry = next_entry) { /* * 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; 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. */ rv = vm_map_lookup_clip_start(map, saved_start, &next_entry, &scratch_entry); if (rv != KERN_SUCCESS) break; } else next_entry = entry; continue; } /* XXXKIB or delete to the upper superpage boundary ? */ rv = vm_map_clip_end(map, entry, end); if (rv != KERN_SUCCESS) break; next_entry = vm_map_entry_succ(entry); /* * 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_map_delete(map->pmap, entry->start, entry->end); /* * 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); } return (rv); } /* * 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 = vm_map_entry_succ(entry); } return (TRUE); } /* * * vm_map_copy_swap_object: * * Copies a swap-backed object from an existing map entry to a * new one. Carries forward the swap charge. May change the * src object on return. */ static void vm_map_copy_swap_object(vm_map_entry_t src_entry, vm_map_entry_t dst_entry, vm_offset_t size, vm_ooffset_t *fork_charge) { vm_object_t src_object; struct ucred *cred; int charged; src_object = src_entry->object.vm_object; charged = ENTRY_CHARGED(src_entry); if ((src_object->flags & OBJ_ANON) != 0) { VM_OBJECT_WLOCK(src_object); vm_object_collapse(src_object); if ((src_object->flags & OBJ_ONEMAPPING) != 0) { 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); VM_OBJECT_WUNLOCK(src_object); } else vm_object_reference(src_object); if (src_entry->cred != NULL && !(src_entry->eflags & MAP_ENTRY_NEEDS_COPY)) { KASSERT(src_object->cred == NULL, ("OVERCOMMIT: vm_map_copy_anon_entry: cred %p", src_object)); src_object->cred = src_entry->cred; src_object->charge = size; } 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; } } } /* * 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; 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) { if ((src_object->flags & OBJ_SWAP) != 0) { vm_map_copy_swap_object(src_entry, dst_entry, size, fork_charge); /* May have split/collapsed, reload obj. */ src_object = src_entry->object.vm_object; } else { vm_object_reference(src_object); dst_entry->object.vm_object = src_object; } 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->defer_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; if ((dst_entry->eflags & MAP_ENTRY_GUARD) == 0) 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) != 0) { 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 __diagused; 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; vm2->vm_stacktop = vm1->vm_stacktop; vm2->vm_shp_base = vm1->vm_shp_base; 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; new_map->flags |= old_map->flags & (MAP_ASLR | MAP_ASLR_IGNSTART | MAP_ASLR_STACK | MAP_WXORX); VM_MAP_ENTRY_FOREACH(old_entry, old_map) { if ((old_entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 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->cred, /* Transfer the second reference too. */ true); old_entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; old_entry->cred = NULL; /* * 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; } else { 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 | MAP_ENTRY_SPLIT_BOUNDARY_MASK); 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; } } /* * 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 gap_entry, new_entry, prev_entry; vm_offset_t bot, gap_bot, gap_top, top; vm_size_t init_ssize, sgp; int rv; KASSERT((cow & MAP_STACK_AREA) != 0, ("New mapping is not a stack")); if (max_ssize == 0 || !vm_map_range_valid(map, addrbos, addrbos + max_ssize)) return (KERN_INVALID_ADDRESS); sgp = ((curproc->p_flag2 & P2_STKGAP_DISABLE) != 0 || (curproc->p_fctl0 & NT_FREEBSD_FCTL_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 (vm_map_entry_succ(prev_entry)->start < addrbos + max_ssize) return (KERN_NO_SPACE); /* * We initially map a stack of only init_ssize, at the top of * the range. We will grow as needed later. * * 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. */ bot = addrbos + max_ssize - init_ssize; top = bot + init_ssize; gap_bot = addrbos; gap_top = bot; rv = vm_map_insert1(map, NULL, 0, bot, top, prot, max, cow, &new_entry); if (rv != KERN_SUCCESS) return (rv); KASSERT(new_entry->end == top || new_entry->start == bot, ("Bad entry start/end for new stack entry")); KASSERT((new_entry->eflags & MAP_ENTRY_GROWS_DOWN) != 0, ("new entry lacks MAP_ENTRY_GROWS_DOWN")); if (gap_bot == gap_top) return (KERN_SUCCESS); rv = vm_map_insert1(map, NULL, 0, gap_bot, gap_top, VM_PROT_NONE, VM_PROT_NONE, MAP_CREATE_GUARD | MAP_CREATE_STACK_GAP, &gap_entry); if (rv == KERN_SUCCESS) { KASSERT((gap_entry->eflags & MAP_ENTRY_GUARD) != 0, ("entry %p not gap %#x", gap_entry, gap_entry->eflags)); KASSERT((gap_entry->eflags & MAP_ENTRY_STACK_GAP) != 0, ("entry %p not stack gap %#x", gap_entry, gap_entry->eflags)); /* * 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(). * Similarly, since a gap cannot have a backing object, * store the original stack protections in the * object offset. */ gap_entry->next_read = sgp; gap_entry->offset = prot | PROT_MAX(max); } 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; vm_offset_t gap_end, gap_start, grow_start; vm_size_t grow_amount, guard, max_grow, sgp; vm_prot_t prot, max; rlim_t lmemlim, stacklim, vmemlim; int rv, rv1 __diagused; bool gap_deleted, is_procstack; #ifdef notyet uint64_t limit; #endif #ifdef RACCT int error __diagused; #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); + MPASS(!vm_map_is_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) != 0) { stack_entry = vm_map_entry_succ(gap_entry); 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); } else { return (KERN_FAILURE); } guard = ((curproc->p_flag2 & P2_STKGAP_DISABLE) != 0 || (curproc->p_fctl0 & NT_FREEBSD_FCTL_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)vm->vm_stacktop; 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; } /* * The gap_entry "offset" field is overloaded. See * vm_map_stack_locked(). */ prot = PROT_EXTRACT(gap_entry->offset); max = PROT_MAX_EXTRACT(gap_entry->offset); sgp = gap_entry->next_read; 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, prot, max, MAP_STACK_AREA); if (rv != KERN_SUCCESS) { if (gap_deleted) { rv1 = vm_map_insert1(map, NULL, 0, gap_start, gap_end, VM_PROT_NONE, VM_PROT_NONE, MAP_CREATE_GUARD | MAP_CREATE_STACK_GAP, &gap_entry); MPASS(rv1 == KERN_SUCCESS); gap_entry->next_read = sgp; gap_entry->offset = prot | PROT_MAX(max); } else { vm_map_entry_resize(map, gap_entry, grow_amount); } } 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; /* * The caller is responsible for ensuring that the reference count * cannot concurrently transition 1 -> 2. */ if (refcount_load(&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; 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) != 0 && vm_map_growstack(map, vaddr, entry) == KERN_SUCCESS) goto RetryLookupLocked; } fault_type = fault_typea & VM_PROT_ALL; 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; } eobject = entry->object.vm_object; vm_object_shadow(&entry->object.vm_object, &entry->offset, size, entry->cred, false); if (eobject == entry->object.vm_object) { /* * The object was not shadowed. */ swap_release_by_cred(size, entry->cred); crfree(entry->cred); } entry->cred = NULL; entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; 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 (entry->object.vm_object == NULL && !vm_map_is_system(map)) { if (vm_map_lock_upgrade(map)) goto RetryLookup; entry->object.vm_object = vm_object_allocate_anon(atop(size), NULL, entry->cred, size); entry->offset = 0; 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) + if (entry->object.vm_object == NULL && !vm_map_is_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); } bool vm_map_range_valid_KBI(vm_map_t map, vm_offset_t start, vm_offset_t end) { return (vm_map_range_valid(map, start, end)); } #ifdef INVARIANTS static void _vm_map_assert_consistent(vm_map_t map, int check) { vm_map_entry_t entry, prev; vm_map_entry_t cur, header, lbound, ubound; vm_size_t max_left, max_right; #ifdef DIAGNOSTIC ++map->nupdates; #endif if (enable_vmmap_check != check) return; header = 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->left == header || 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 == header || entry->start < entry->right->start, ("map %p start = %jx, right->start = %jx", map, (uintmax_t)entry->start, (uintmax_t)entry->right->start)); cur = map->root; lbound = ubound = header; for (;;) { if (entry->start < cur->start) { ubound = cur; cur = cur->left; KASSERT(cur != lbound, ("map %p cannot find %jx", map, (uintmax_t)entry->start)); } else if (cur->end <= entry->start) { lbound = cur; cur = cur->right; KASSERT(cur != ubound, ("map %p cannot find %jx", map, (uintmax_t)entry->start)); } else { KASSERT(cur == entry, ("map %p cannot find %jx", map, (uintmax_t)entry->start)); break; } } max_left = vm_map_entry_max_free_left(entry, lbound); max_right = vm_map_entry_max_free_right(entry, ubound); KASSERT(entry->max_free == vm_size_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 const char * const 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 */ diff --git a/sys/vm/vm_map.h b/sys/vm/vm_map.h index 5ecec0531e1c..ed1106734951 100644 --- a/sys/vm/vm_map.h +++ b/sys/vm/vm_map.h @@ -1,546 +1,552 @@ /*- * 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. * * * 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 map module definitions. */ #ifndef _VM_MAP_ #define _VM_MAP_ #include #include #include /* * Types defined: * * vm_map_t the high-level address map data structure. * vm_map_entry_t an entry in an address map. */ typedef u_char vm_flags_t; typedef u_int vm_eflags_t; /* * Objects which live in maps may be either VM objects, or * another map (called a "sharing map") which denotes read-write * sharing with other maps. */ union vm_map_object { struct vm_object *vm_object; /* object object */ struct vm_map *sub_map; /* belongs to another map */ }; /* * Address map entries consist of start and end addresses, * a VM object (or sharing map) and offset into that object, * and user-exported inheritance and protection information. * Also included is control information for virtual copy operations. * * For stack gap map entries (MAP_ENTRY_GUARD | MAP_ENTRY_STACK_GAP), * the next_read member is reused as the stack_guard_page storage, and * offset is the stack protection. */ struct vm_map_entry { struct vm_map_entry *left; /* left child or previous entry */ struct vm_map_entry *right; /* right child or next entry */ vm_offset_t start; /* start address */ vm_offset_t end; /* end address */ vm_offset_t next_read; /* vaddr of the next sequential read */ vm_size_t max_free; /* max free space in subtree */ union vm_map_object object; /* object I point to */ vm_ooffset_t offset; /* offset into object */ vm_eflags_t eflags; /* map entry flags */ vm_prot_t protection; /* protection code */ vm_prot_t max_protection; /* maximum protection */ vm_inherit_t inheritance; /* inheritance */ uint8_t read_ahead; /* pages in the read-ahead window */ int wired_count; /* can be paged if = 0 */ struct ucred *cred; /* tmp storage for creator ref */ struct thread *wiring_thread; }; #define MAP_ENTRY_NOSYNC 0x00000001 #define MAP_ENTRY_IS_SUB_MAP 0x00000002 #define MAP_ENTRY_COW 0x00000004 #define MAP_ENTRY_NEEDS_COPY 0x00000008 #define MAP_ENTRY_NOFAULT 0x00000010 #define MAP_ENTRY_USER_WIRED 0x00000020 #define MAP_ENTRY_BEHAV_NORMAL 0x00000000 /* default behavior */ #define MAP_ENTRY_BEHAV_SEQUENTIAL 0x00000040 /* expect sequential access */ #define MAP_ENTRY_BEHAV_RANDOM 0x00000080 /* expect random access */ #define MAP_ENTRY_BEHAV_RESERVED 0x000000c0 /* future use */ #define MAP_ENTRY_BEHAV_MASK 0x000000c0 #define MAP_ENTRY_IN_TRANSITION 0x00000100 /* entry being changed */ #define MAP_ENTRY_NEEDS_WAKEUP 0x00000200 /* waiters in transition */ #define MAP_ENTRY_NOCOREDUMP 0x00000400 /* don't include in a core */ #define MAP_ENTRY_VN_EXEC 0x00000800 /* text vnode mapping */ #define MAP_ENTRY_GROWS_DOWN 0x00001000 /* top-down stacks */ #define MAP_ENTRY_UNUSED0 0x00002000 #define MAP_ENTRY_WIRE_SKIPPED 0x00004000 #define MAP_ENTRY_WRITECNT 0x00008000 /* tracked writeable mapping */ #define MAP_ENTRY_GUARD 0x00010000 #define MAP_ENTRY_STACK_GAP 0x00020000 #define MAP_ENTRY_UNUSED1 0x00040000 #define MAP_ENTRY_HEADER 0x00080000 #define MAP_ENTRY_SPLIT_BOUNDARY_MASK 0x00300000 #define MAP_ENTRY_SPLIT_BOUNDARY_SHIFT 20 #define MAP_ENTRY_SPLIT_BOUNDARY_INDEX(entry) \ (((entry)->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) >> \ MAP_ENTRY_SPLIT_BOUNDARY_SHIFT) #ifdef _KERNEL static __inline u_char vm_map_entry_behavior(vm_map_entry_t entry) { return (entry->eflags & MAP_ENTRY_BEHAV_MASK); } static __inline int vm_map_entry_user_wired_count(vm_map_entry_t entry) { if (entry->eflags & MAP_ENTRY_USER_WIRED) return (1); return (0); } static __inline int vm_map_entry_system_wired_count(vm_map_entry_t entry) { return (entry->wired_count - vm_map_entry_user_wired_count(entry)); } #endif /* _KERNEL */ /* * A map is a set of map entries. These map entries are * organized as a threaded binary search tree. The tree is * ordered based upon the start and end addresses contained * within each map entry. The largest gap between an entry in a * subtree and one of its neighbors is saved in the max_free * field, and that field is updated when the tree is restructured. * * Sleator and Tarjan's top-down splay algorithm is employed to * control height imbalance in the binary search tree. * * The map's min offset value is stored in map->header.end, and * its max offset value is stored in map->header.start. These * values act as sentinels for any forward or backward address * scan of the list. The right and left fields of the map * header point to the first and list map entries. The map * header has a special value for the eflags field, * MAP_ENTRY_HEADER, that is set initially, is never changed, * and prevents an eflags match of the header with any other map * entry. * * List of locks * (c) const until freed */ struct vm_map { struct vm_map_entry header; /* List of entries */ union { struct sx lock; /* Lock for map data */ struct mtx system_mtx; }; int nentries; /* Number of entries */ vm_size_t size; /* virtual size */ u_int timestamp; /* Version number */ u_char needs_wakeup; u_char system_map; /* (c) Am I a system map? */ vm_flags_t flags; /* flags for this vm_map */ vm_map_entry_t root; /* Root of a binary search tree */ pmap_t pmap; /* (c) Physical map */ vm_offset_t anon_loc; int busy; #ifdef DIAGNOSTIC int nupdates; #endif }; /* * vm_flags_t values */ #define MAP_WIREFUTURE 0x01 /* wire all future pages */ #define MAP_BUSY_WAKEUP 0x02 /* thread(s) waiting on busy state */ #define MAP_IS_SUB_MAP 0x04 /* has parent */ #define MAP_ASLR 0x08 /* enabled ASLR */ #define MAP_ASLR_IGNSTART 0x10 /* ASLR ignores data segment */ #define MAP_REPLENISH 0x20 /* kmapent zone needs to be refilled */ #define MAP_WXORX 0x40 /* enforce W^X */ #define MAP_ASLR_STACK 0x80 /* stack location is randomized */ #ifdef _KERNEL #if defined(KLD_MODULE) && !defined(KLD_TIED) #define vm_map_max(map) vm_map_max_KBI((map)) #define vm_map_min(map) vm_map_min_KBI((map)) #define vm_map_pmap(map) vm_map_pmap_KBI((map)) #define vm_map_range_valid(map, start, end) \ vm_map_range_valid_KBI((map), (start), (end)) #else static __inline vm_offset_t vm_map_max(const struct vm_map *map) { return (map->header.start); } static __inline vm_offset_t vm_map_min(const struct vm_map *map) { return (map->header.end); } static __inline pmap_t vm_map_pmap(vm_map_t map) { return (map->pmap); } static __inline void vm_map_modflags(vm_map_t map, vm_flags_t set, vm_flags_t clear) { map->flags = (map->flags | set) & ~clear; } static inline bool vm_map_range_valid(vm_map_t map, vm_offset_t start, vm_offset_t end) { if (end < start) return (false); if (start < vm_map_min(map) || end > vm_map_max(map)) return (false); return (true); } +static inline bool +vm_map_is_system(vm_map_t map) +{ + return ((map->system_map)); +} + #endif /* KLD_MODULE */ #endif /* _KERNEL */ /* * Shareable process virtual address space. * * List of locks * (c) const until freed */ struct vmspace { struct vm_map vm_map; /* VM address map */ struct shmmap_state *vm_shm; /* SYS5 shared memory private data XXX */ segsz_t vm_swrss; /* resident set size before last swap */ segsz_t vm_tsize; /* text size (pages) XXX */ segsz_t vm_dsize; /* data size (pages) XXX */ segsz_t vm_ssize; /* stack size (pages) */ caddr_t vm_taddr; /* (c) user virtual address of text */ caddr_t vm_daddr; /* (c) user virtual address of data */ caddr_t vm_maxsaddr; /* user VA at max stack growth */ vm_offset_t vm_stacktop; /* top of the stack, may not be page-aligned */ vm_offset_t vm_shp_base; /* shared page address */ u_int vm_refcnt; /* number of references */ /* * Keep the PMAP last, so that CPU-specific variations of that * structure on a single architecture don't result in offset * variations of the machine-independent fields in the vmspace. */ struct pmap vm_pmap; /* private physical map */ }; #ifdef _KERNEL static __inline pmap_t vmspace_pmap(struct vmspace *vmspace) { return &vmspace->vm_pmap; } #endif /* _KERNEL */ #ifdef _KERNEL /* * Macros: vm_map_lock, etc. * Function: * Perform locking on the data portion of a map. Note that * these macros mimic procedure calls returning void. The * semicolon is supplied by the user of these macros, not * by the macros themselves. The macros can safely be used * as unbraced elements in a higher level statement. */ void _vm_map_lock(vm_map_t map, const char *file, int line); void _vm_map_unlock(vm_map_t map, const char *file, int line); int _vm_map_unlock_and_wait(vm_map_t map, int timo, const char *file, int line); void _vm_map_lock_read(vm_map_t map, const char *file, int line); void _vm_map_unlock_read(vm_map_t map, const char *file, int line); int _vm_map_trylock(vm_map_t map, const char *file, int line); int _vm_map_trylock_read(vm_map_t map, const char *file, int line); int _vm_map_lock_upgrade(vm_map_t map, const char *file, int line); void _vm_map_lock_downgrade(vm_map_t map, const char *file, int line); int vm_map_locked(vm_map_t map); void vm_map_wakeup(vm_map_t map); void vm_map_busy(vm_map_t map); void vm_map_unbusy(vm_map_t map); void vm_map_wait_busy(vm_map_t map); vm_offset_t vm_map_max_KBI(const struct vm_map *map); vm_offset_t vm_map_min_KBI(const struct vm_map *map); pmap_t vm_map_pmap_KBI(vm_map_t map); bool vm_map_range_valid_KBI(vm_map_t map, vm_offset_t start, vm_offset_t end); #define vm_map_lock(map) _vm_map_lock(map, LOCK_FILE, LOCK_LINE) #define vm_map_unlock(map) _vm_map_unlock(map, LOCK_FILE, LOCK_LINE) #define vm_map_unlock_and_wait(map, timo) \ _vm_map_unlock_and_wait(map, timo, LOCK_FILE, LOCK_LINE) #define vm_map_lock_read(map) _vm_map_lock_read(map, LOCK_FILE, LOCK_LINE) #define vm_map_unlock_read(map) _vm_map_unlock_read(map, LOCK_FILE, LOCK_LINE) #define vm_map_trylock(map) _vm_map_trylock(map, LOCK_FILE, LOCK_LINE) #define vm_map_trylock_read(map) \ _vm_map_trylock_read(map, LOCK_FILE, LOCK_LINE) #define vm_map_lock_upgrade(map) \ _vm_map_lock_upgrade(map, LOCK_FILE, LOCK_LINE) #define vm_map_lock_downgrade(map) \ _vm_map_lock_downgrade(map, LOCK_FILE, LOCK_LINE) long vmspace_resident_count(struct vmspace *vmspace); #endif /* _KERNEL */ /* * Copy-on-write flags for vm_map operations */ #define MAP_INHERIT_SHARE 0x00000001 #define MAP_COPY_ON_WRITE 0x00000002 #define MAP_NOFAULT 0x00000004 #define MAP_PREFAULT 0x00000008 #define MAP_PREFAULT_PARTIAL 0x00000010 #define MAP_DISABLE_SYNCER 0x00000020 #define MAP_CHECK_EXCL 0x00000040 #define MAP_CREATE_GUARD 0x00000080 #define MAP_DISABLE_COREDUMP 0x00000100 #define MAP_PREFAULT_MADVISE 0x00000200 /* from (user) madvise request */ #define MAP_WRITECOUNT 0x00000400 #define MAP_REMAP 0x00000800 #define MAP_STACK_AREA 0x00001000 #define MAP_COW_UNUSED0 0x00002000 #define MAP_ACC_CHARGED 0x00004000 #define MAP_ACC_NO_CHARGE 0x00008000 #define MAP_COW_UNUSED1 0x00010000 #define MAP_CREATE_STACK_GAP 0x00020000 #define MAP_VN_EXEC 0x00040000 #define MAP_SPLIT_BOUNDARY_MASK 0x00180000 #define MAP_NO_HINT 0x00200000 #define MAP_SPLIT_BOUNDARY_SHIFT 19 /* * vm_fault option flags */ #define VM_FAULT_NORMAL 0x00 /* Nothing special */ #define VM_FAULT_WIRE 0x01 /* Wire the mapped page */ #define VM_FAULT_DIRTY 0x02 /* Dirty the page; use w/VM_PROT_COPY */ #define VM_FAULT_NOFILL 0x04 /* Fail if the pager doesn't have a copy */ /* * Initially, mappings are slightly sequential. The maximum window size must * account for the map entry's "read_ahead" field being defined as an uint8_t. */ #define VM_FAULT_READ_AHEAD_MIN 7 #define VM_FAULT_READ_AHEAD_INIT 15 #define VM_FAULT_READ_AHEAD_MAX min(atop(maxphys) - 1, UINT8_MAX) /* * The following "find_space" options are supported by vm_map_find(). * * For VMFS_ALIGNED_SPACE, the desired alignment is specified to * the macro argument as log base 2 of the desired alignment. */ #define VMFS_NO_SPACE 0 /* don't find; use the given range */ #define VMFS_ANY_SPACE 1 /* find a range with any alignment */ #define VMFS_OPTIMAL_SPACE 2 /* find a range with optimal alignment*/ #define VMFS_SUPER_SPACE 3 /* find a superpage-aligned range */ #define VMFS_ALIGNED_SPACE(x) ((x) << 8) /* find a range with fixed alignment */ /* * vm_map_wire and vm_map_unwire option flags */ #define VM_MAP_WIRE_SYSTEM 0 /* wiring in a kernel map */ #define VM_MAP_WIRE_USER 1 /* wiring in a user map */ #define VM_MAP_WIRE_NOHOLES 0 /* region must not have holes */ #define VM_MAP_WIRE_HOLESOK 2 /* region may have holes */ #define VM_MAP_WIRE_WRITE 4 /* Validate writable. */ typedef int vm_map_entry_reader(void *token, vm_map_entry_t addr, vm_map_entry_t dest); #ifndef _KERNEL /* * Find the successor of a map_entry, using a reader to dereference pointers. * '*clone' is a copy of a vm_map entry. 'reader' is used to copy a map entry * at some address into '*clone'. Change *clone to a copy of the next map * entry, and return the address of that entry, or NULL if copying has failed. * * This function is made available to user-space code that needs to traverse * map entries. */ static inline vm_map_entry_t vm_map_entry_read_succ(void *token, struct vm_map_entry *const clone, vm_map_entry_reader reader) { vm_map_entry_t after, backup; vm_offset_t start; after = clone->right; start = clone->start; if (!reader(token, after, clone)) return (NULL); backup = clone->left; if (!reader(token, backup, clone)) return (NULL); if (clone->start > start) { do { after = backup; backup = clone->left; if (!reader(token, backup, clone)) return (NULL); } while (clone->start != start); } if (!reader(token, after, clone)) return (NULL); return (after); } #endif /* ! _KERNEL */ #ifdef _KERNEL boolean_t vm_map_check_protection (vm_map_t, vm_offset_t, vm_offset_t, vm_prot_t); int vm_map_delete(vm_map_t, vm_offset_t, vm_offset_t); int vm_map_find(vm_map_t, vm_object_t, vm_ooffset_t, vm_offset_t *, vm_size_t, vm_offset_t, int, vm_prot_t, vm_prot_t, int); int vm_map_find_locked(vm_map_t, vm_object_t, vm_ooffset_t, vm_offset_t *, vm_size_t, vm_offset_t, int, vm_prot_t, vm_prot_t, int); int vm_map_find_min(vm_map_t, vm_object_t, vm_ooffset_t, vm_offset_t *, vm_size_t, vm_offset_t, vm_offset_t, int, vm_prot_t, vm_prot_t, int); int vm_map_find_aligned(vm_map_t map, vm_offset_t *addr, vm_size_t length, vm_offset_t max_addr, vm_offset_t alignment); int vm_map_fixed(vm_map_t, vm_object_t, vm_ooffset_t, vm_offset_t, vm_size_t, vm_prot_t, vm_prot_t, int); vm_offset_t vm_map_findspace(vm_map_t, vm_offset_t, vm_size_t); int vm_map_inherit (vm_map_t, vm_offset_t, vm_offset_t, vm_inherit_t); void vm_map_init(vm_map_t, pmap_t, vm_offset_t, vm_offset_t); void vm_map_init_system(vm_map_t, pmap_t, vm_offset_t, vm_offset_t); int vm_map_insert (vm_map_t, vm_object_t, vm_ooffset_t, vm_offset_t, vm_offset_t, vm_prot_t, vm_prot_t, int); int vm_map_lookup (vm_map_t *, vm_offset_t, vm_prot_t, vm_map_entry_t *, vm_object_t *, vm_pindex_t *, vm_prot_t *, boolean_t *); int vm_map_lookup_locked(vm_map_t *, vm_offset_t, vm_prot_t, vm_map_entry_t *, vm_object_t *, vm_pindex_t *, vm_prot_t *, boolean_t *); void vm_map_lookup_done (vm_map_t, vm_map_entry_t); boolean_t vm_map_lookup_entry (vm_map_t, vm_offset_t, vm_map_entry_t *); static inline vm_map_entry_t vm_map_entry_first(vm_map_t map) { return (map->header.right); } static inline vm_map_entry_t vm_map_entry_succ(vm_map_entry_t entry) { vm_map_entry_t after; after = entry->right; if (after->left->start > entry->start) { do after = after->left; while (after->left != entry); } return (after); } #define VM_MAP_ENTRY_FOREACH(it, map) \ for ((it) = vm_map_entry_first(map); \ (it) != &(map)->header; \ (it) = vm_map_entry_succ(it)) #define VM_MAP_PROTECT_SET_PROT 0x0001 #define VM_MAP_PROTECT_SET_MAXPROT 0x0002 #define VM_MAP_PROTECT_GROWSDOWN 0x0004 int vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end, vm_prot_t new_prot, vm_prot_t new_maxprot, int flags); int vm_map_remove (vm_map_t, vm_offset_t, vm_offset_t); vm_map_entry_t vm_map_try_merge_entries(vm_map_t map, vm_map_entry_t prev, vm_map_entry_t entry); void vm_map_startup (void); int vm_map_submap (vm_map_t, vm_offset_t, vm_offset_t, vm_map_t); int vm_map_sync(vm_map_t, vm_offset_t, vm_offset_t, boolean_t, boolean_t); int vm_map_madvise (vm_map_t, vm_offset_t, vm_offset_t, int); int vm_map_stack (vm_map_t, vm_offset_t, vm_size_t, vm_prot_t, vm_prot_t, int); int vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end, int flags); int vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t end, int flags); int vm_map_wire_locked(vm_map_t map, vm_offset_t start, vm_offset_t end, int flags); long vmspace_swap_count(struct vmspace *vmspace); void vm_map_entry_set_vnode_text(vm_map_entry_t entry, bool add); #endif /* _KERNEL */ #endif /* _VM_MAP_ */ diff --git a/sys/vm/vm_pageout.c b/sys/vm/vm_pageout.c index f6c5e6291692..28a54a83fd49 100644 --- a/sys/vm/vm_pageout.c +++ b/sys/vm/vm_pageout.c @@ -1,2388 +1,2388 @@ /*- * 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. * * * 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 #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 /* * 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 swapdev_enabled; int vm_pageout_page_count = 32; static int vm_panic_on_oom = 0; SYSCTL_INT(_vm, OID_AUTO, panic_on_oom, CTLFLAG_RWTUN, &vm_panic_on_oom, 0, "Panic on the given number of out-of-memory errors instead of " "killing the largest process"); static int vm_pageout_update_period; SYSCTL_INT(_vm, OID_AUTO, pageout_update_period, CTLFLAG_RWTUN, &vm_pageout_update_period, 0, "Maximum active LRU update period"); static int pageout_cpus_per_thread = 16; SYSCTL_INT(_vm, OID_AUTO, pageout_cpus_per_thread, CTLFLAG_RDTUN, &pageout_cpus_per_thread, 0, "Number of CPUs per pagedaemon worker thread"); static int lowmem_period = 10; SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RWTUN, &lowmem_period, 0, "Low memory callback period"); static int disable_swap_pageouts; 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"); static int vm_pageout_oom_seq = 12; 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"); 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->a.flags & 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->a.flags & 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_pageout_defer() must be used to * determine whether the page has been logically dequeued since the batch was * collected. */ 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->a.flags & 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->a.flags & 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)); } /* * Determine whether processing of a page should be deferred and ensure that any * outstanding queue operations are processed. */ static __always_inline bool vm_pageout_defer(vm_page_t m, const uint8_t queue, const bool enqueued) { vm_page_astate_t as; as = vm_page_astate_load(m); if (__predict_false(as.queue != queue || ((as.flags & PGA_ENQUEUED) != 0) != enqueued)) return (true); if ((as.flags & PGA_QUEUE_OP_MASK) != 0) { vm_page_pqbatch_submit(m, queue); return (true); } return (false); } /* * We can cluster only if the page is not clean, busy, or held, and the page is * in the laundry queue. */ static bool vm_pageout_flushable(vm_page_t m) { if (vm_page_tryxbusy(m) == 0) return (false); if (!vm_page_wired(m)) { vm_page_test_dirty(m); if (m->dirty != 0 && vm_page_in_laundry(m) && vm_page_try_remove_write(m)) return (true); } vm_page_xunbusy(m); return (false); } /* * 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_page_t mc[2 * vm_pageout_page_count - 1]; int alignment, num_ends, page_base, pageout_count; VM_OBJECT_ASSERT_WLOCKED(m->object); vm_page_assert_xbusied(m); alignment = m->pindex % vm_pageout_page_count; num_ends = 0; page_base = nitems(mc) / 2; pageout_count = 1; mc[page_base] = m; /* * 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: m = mc[page_base]; while (pageout_count < vm_pageout_page_count) { /* * If we are at an alignment boundary, and haven't reached the * last flushable page forward, stop here, and switch * directions. */ if (alignment == pageout_count - 1 && num_ends == 0) break; m = vm_page_prev(m); if (m == NULL || !vm_pageout_flushable(m)) { num_ends++; break; } mc[--page_base] = m; ++pageout_count; } m = mc[page_base + pageout_count - 1]; while (num_ends != 2 && pageout_count < vm_pageout_page_count) { m = vm_page_next(m); if (m == NULL || !vm_pageout_flushable(m)) { if (num_ends++ == 0) /* Resume the reverse scan. */ goto more; break; } mc[page_base + pageout_count] = m; ++pageout_count; } 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(vm_page_all_valid(mc[i]), ("vm_pageout_flush: partially invalid page %p index %d/%d", mc[i], i, count)); KASSERT((mc[i]->a.flags & 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: /* * The page may have moved since laundering started, in * which case it should be left alone. */ if (vm_page_in_laundry(mt)) vm_page_deactivate_noreuse(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); if (vm_page_in_laundry(mt)) vm_page_deactivate_noreuse(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 * because 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.) */ if ((object->flags & OBJ_SWAP) != 0 && pageout_status[i] == VM_PAGER_FAIL) { vm_page_unswappable(mt); numpagedout++; } else vm_page_activate(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; 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_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); if (vget(vp, vn_lktype_write(NULL, vp) | LK_TIMELOCK) != 0) { 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; } /* * While the object was 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) { error = ENXIO; goto unlock_all; } /* * The page may have been busied while the object lock was * released. */ if (vm_page_tryxbusy(m) == 0) { 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); error = EBUSY; goto unlock_all; } /* * 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: 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; vm_object_t object; vm_page_t m, marker; vm_page_astate_t new, old; int act_delta, error, numpagedout, queue, refs, starting_target; int vnodes_skipped; bool pageout_ok; 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; /* * Don't touch a page that was removed from the queue after the * page queue lock was released. Otherwise, ensure that any * pending queue operations, such as dequeues for wired pages, * are handled. */ if (vm_pageout_defer(m, queue, true)) continue; /* * Lock the page's object. */ if (object == NULL || object != m->object) { if (object != NULL) VM_OBJECT_WUNLOCK(object); object = atomic_load_ptr(&m->object); if (__predict_false(object == NULL)) /* The page is being freed by another thread. */ continue; /* Depends on type-stability. */ VM_OBJECT_WLOCK(object); if (__predict_false(m->object != object)) { VM_OBJECT_WUNLOCK(object); object = NULL; continue; } } if (vm_page_tryxbusy(m) == 0) continue; /* * Check for wirings now that we hold the object lock and have * exclusively busied the page. 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 after this check. */ if (__predict_false(vm_page_wired(m))) goto skip_page; /* * Invalid pages can be easily freed. They cannot be * mapped; vm_page_free() asserts this. */ if (vm_page_none_valid(m)) goto free_page; refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0; for (old = vm_page_astate_load(m);;) { /* * Check to see if the page has been removed from the * queue since the first such check. Leave it alone if * so, discarding any references collected by * pmap_ts_referenced(). */ if (__predict_false(_vm_page_queue(old) == PQ_NONE)) goto skip_page; new = old; act_delta = refs; if ((old.flags & PGA_REFERENCED) != 0) { new.flags &= ~PGA_REFERENCED; act_delta++; } if (act_delta == 0) { ; } else if (object->ref_count != 0) { /* * 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 laundry queue. */ new.act_count += ACT_ADVANCE + act_delta; if (new.act_count > ACT_MAX) new.act_count = ACT_MAX; new.flags &= ~PGA_QUEUE_OP_MASK; new.flags |= PGA_REQUEUE; new.queue = PQ_ACTIVE; if (!vm_page_pqstate_commit(m, &old, new)) continue; /* * 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--; VM_CNT_INC(v_reactivated); goto skip_page; } else if ((object->flags & OBJ_DEAD) == 0) { new.flags |= PGA_REQUEUE; if (!vm_page_pqstate_commit(m, &old, new)) continue; goto skip_page; } break; } /* * 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)) goto skip_page; } /* * 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: /* * Now we are guaranteed that no other threads are * manipulating the page, check for a last-second * reference. */ if (vm_pageout_defer(m, queue, true)) goto skip_page; vm_page_free(m); VM_CNT_INC(v_dfree); } else if ((object->flags & OBJ_DEAD) == 0) { if ((object->flags & OBJ_SWAP) != 0) pageout_ok = disable_swap_pageouts == 0; else pageout_ok = true; if (!pageout_ok) { vm_page_launder(m); goto skip_page; } /* * 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++; } object = NULL; } else { skip_page: vm_page_xunbusy(m); } } 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; vm_object_t object; vm_page_t m, marker; struct vm_pagequeue *pq; vm_page_astate_t old, new; long min_scan; int act_delta, max_scan, ps_delta, refs, scan_tick; uint8_t nqueue; 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; 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; /* * Don't touch a page that was removed from the queue after the * page queue lock was released. Otherwise, ensure that any * pending queue operations, such as dequeues for wired pages, * are handled. */ if (vm_pageout_defer(m, PQ_ACTIVE, true)) continue; /* * A page's object pointer may be set to NULL before * the object lock is acquired. */ object = atomic_load_ptr(&m->object); if (__predict_false(object == NULL)) /* * The page has been removed from its object. */ continue; /* Deferred free of swap space. */ if ((m->a.flags & PGA_SWAP_FREE) != 0 && VM_OBJECT_TRYWLOCK(object)) { if (m->object == object) vm_pager_page_unswapped(m); VM_OBJECT_WUNLOCK(object); } /* * 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. */ refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0; old = vm_page_astate_load(m); do { /* * Check to see if the page has been removed from the * queue since the first such check. Leave it alone if * so, discarding any references collected by * pmap_ts_referenced(). */ if (__predict_false(_vm_page_queue(old) == PQ_NONE)) { ps_delta = 0; break; } /* * Advance or decay the act_count based on recent usage. */ new = old; act_delta = refs; if ((old.flags & PGA_REFERENCED) != 0) { new.flags &= ~PGA_REFERENCED; act_delta++; } if (act_delta != 0) { new.act_count += ACT_ADVANCE + act_delta; if (new.act_count > ACT_MAX) new.act_count = ACT_MAX; } else { new.act_count -= min(new.act_count, ACT_DECLINE); } if (new.act_count > 0) { /* * Adjust the activation count and keep the page * in the active queue. The count might be left * unchanged if it is saturated. The page may * have been moved to a different queue since we * started the scan, in which case we move it * back. */ ps_delta = 0; if (old.queue != PQ_ACTIVE) { new.flags &= ~PGA_QUEUE_OP_MASK; new.flags |= PGA_REQUEUE; new.queue = PQ_ACTIVE; } } else { /* * When not short for inactive pages, let dirty * pages go through the inactive queue before * moving to the laundry queue. 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. * * 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 (page_shortage <= 0) { nqueue = PQ_INACTIVE; ps_delta = 0; } else if (m->dirty == 0) { nqueue = PQ_INACTIVE; ps_delta = act_scan_laundry_weight; } else { nqueue = PQ_LAUNDRY; ps_delta = 1; } new.flags &= ~PGA_QUEUE_OP_MASK; new.flags |= PGA_REQUEUE; new.queue = nqueue; } } while (!vm_page_pqstate_commit(m, &old, new)); page_shortage -= ps_delta; } 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 vm_pagequeue *pq, vm_page_t marker, vm_page_t m) { vm_page_astate_t as; vm_pagequeue_assert_locked(pq); as = vm_page_astate_load(m); if (as.queue != PQ_INACTIVE || (as.flags & PGA_ENQUEUED) != 0) return (0); vm_page_aflag_set(m, PGA_ENQUEUED); TAILQ_INSERT_BEFORE(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; vm_page_t marker; int delta; delta = 0; marker = ss->marker; pq = ss->pq; if (m != NULL) { if (vm_batchqueue_insert(bq, m) != 0) return; vm_pagequeue_lock(pq); delta += vm_pageout_reinsert_inactive_page(pq, marker, m); } else vm_pagequeue_lock(pq); while ((m = vm_batchqueue_pop(bq)) != NULL) delta += vm_pageout_reinsert_inactive_page(pq, marker, m); vm_pagequeue_cnt_add(pq, delta); vm_pagequeue_unlock(pq); vm_batchqueue_init(bq); } static void vm_pageout_scan_inactive(struct vm_domain *vmd, int page_shortage) { struct timeval start, end; struct scan_state ss; struct vm_batchqueue rq; struct vm_page marker_page; vm_page_t m, marker; struct vm_pagequeue *pq; vm_object_t object; vm_page_astate_t old, new; int act_delta, addl_page_shortage, starting_page_shortage, refs; object = NULL; vm_batchqueue_init(&rq); getmicrouptime(&start); /* * 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; /* * 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->a.act_count is not used to make * decisions for the inactive queue, only for the active queue.) */ starting_page_shortage = page_shortage; marker = &marker_page; vm_page_init_marker(marker, PQ_INACTIVE, 0); 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) { /* * If we need to refill the scan batch queue, release any * optimistically held object lock. This gives someone else a * chance to grab the lock, and also avoids holding it while we * do unrelated work. */ if (object != NULL && vm_batchqueue_empty(&ss.bq)) { VM_OBJECT_WUNLOCK(object); object = NULL; } m = vm_pageout_next(&ss, true); if (m == NULL) break; KASSERT((m->flags & PG_MARKER) == 0, ("marker page %p was dequeued", m)); /* * Don't touch a page that was removed from the queue after the * page queue lock was released. Otherwise, ensure that any * pending queue operations, such as dequeues for wired pages, * are handled. */ if (vm_pageout_defer(m, PQ_INACTIVE, false)) continue; /* * Lock the page's object. */ if (object == NULL || object != m->object) { if (object != NULL) VM_OBJECT_WUNLOCK(object); object = atomic_load_ptr(&m->object); if (__predict_false(object == NULL)) /* The page is being freed by another thread. */ continue; /* Depends on type-stability. */ VM_OBJECT_WLOCK(object); if (__predict_false(m->object != object)) { VM_OBJECT_WUNLOCK(object); object = NULL; goto reinsert; } } 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; } /* Deferred free of swap space. */ if ((m->a.flags & PGA_SWAP_FREE) != 0) vm_pager_page_unswapped(m); /* * Check for wirings now that we hold the object lock and have * exclusively busied the page. 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 after this check. */ if (__predict_false(vm_page_wired(m))) goto skip_page; /* * Invalid pages can be easily freed. They cannot be * mapped, vm_page_free() asserts this. */ if (vm_page_none_valid(m)) goto free_page; refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0; for (old = vm_page_astate_load(m);;) { /* * Check to see if the page has been removed from the * queue since the first such check. Leave it alone if * so, discarding any references collected by * pmap_ts_referenced(). */ if (__predict_false(_vm_page_queue(old) == PQ_NONE)) goto skip_page; new = old; act_delta = refs; if ((old.flags & PGA_REFERENCED) != 0) { new.flags &= ~PGA_REFERENCED; act_delta++; } if (act_delta == 0) { ; } else if (object->ref_count != 0) { /* * 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. */ new.act_count += ACT_ADVANCE + act_delta; if (new.act_count > ACT_MAX) new.act_count = ACT_MAX; new.flags &= ~PGA_QUEUE_OP_MASK; new.flags |= PGA_REQUEUE; new.queue = PQ_ACTIVE; if (!vm_page_pqstate_commit(m, &old, new)) continue; VM_CNT_INC(v_reactivated); goto skip_page; } else if ((object->flags & OBJ_DEAD) == 0) { new.queue = PQ_INACTIVE; new.flags |= PGA_REQUEUE; if (!vm_page_pqstate_commit(m, &old, new)) continue; goto skip_page; } break; } /* * 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)) goto skip_page; } /* * 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: /* * Now we are guaranteed that no other threads are * manipulating the page, check for a last-second * reference that would save it from doom. */ if (vm_pageout_defer(m, PQ_INACTIVE, false)) goto skip_page; /* * Because we dequeued the page and have already checked * for pending dequeue and enqueue requests, we can * safely disassociate the page from the inactive queue * without holding the queue lock. */ m->a.queue = PQ_NONE; vm_page_free(m); page_shortage--; continue; } if ((object->flags & OBJ_DEAD) == 0) vm_page_launder(m); skip_page: vm_page_xunbusy(m); continue; reinsert: vm_pageout_reinsert_inactive(&ss, &rq, m); } 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); /* * Record the remaining shortage and the progress and rate it was made. */ atomic_add_int(&vmd->vmd_addl_shortage, addl_page_shortage); getmicrouptime(&end); timevalsub(&end, &start); atomic_add_int(&vmd->vmd_inactive_us, end.tv_sec * 1000000 + end.tv_usec); atomic_add_int(&vmd->vmd_inactive_freed, starting_page_shortage - page_shortage); } /* * Dispatch a number of inactive threads according to load and collect the * results to present a coherent view of paging activity on this domain. */ static int vm_pageout_inactive_dispatch(struct vm_domain *vmd, int shortage) { u_int freed, pps, slop, threads, us; vmd->vmd_inactive_shortage = shortage; slop = 0; /* * If we have more work than we can do in a quarter of our interval, we * fire off multiple threads to process it. */ threads = vmd->vmd_inactive_threads; if (threads > 1 && vmd->vmd_inactive_pps != 0 && shortage > vmd->vmd_inactive_pps / VM_INACT_SCAN_RATE / 4) { vmd->vmd_inactive_shortage /= threads; slop = shortage % threads; vm_domain_pageout_lock(vmd); blockcount_acquire(&vmd->vmd_inactive_starting, threads - 1); blockcount_acquire(&vmd->vmd_inactive_running, threads - 1); wakeup(&vmd->vmd_inactive_shortage); vm_domain_pageout_unlock(vmd); } /* Run the local thread scan. */ vm_pageout_scan_inactive(vmd, vmd->vmd_inactive_shortage + slop); /* * Block until helper threads report results and then accumulate * totals. */ blockcount_wait(&vmd->vmd_inactive_running, NULL, "vmpoid", PVM); freed = atomic_readandclear_int(&vmd->vmd_inactive_freed); VM_CNT_ADD(v_dfree, freed); /* * Calculate the per-thread paging rate with an exponential decay of * prior results. Careful to avoid integer rounding errors with large * us values. */ us = max(atomic_readandclear_int(&vmd->vmd_inactive_us), 1); if (us > 1000000) /* Keep rounding to tenths */ pps = (freed * 10) / ((us * 10) / 1000000); else pps = (1000000 / us) * freed; vmd->vmd_inactive_pps = (vmd->vmd_inactive_pps / 2) + (pps / 2); return (shortage - freed); } /* * Attempt to reclaim the requested number of pages from the inactive queue. * Returns true if the shortage was addressed. */ static int vm_pageout_inactive(struct vm_domain *vmd, int shortage, int *addl_shortage) { struct vm_pagequeue *pq; u_int addl_page_shortage, deficit, page_shortage; u_int starting_page_shortage; /* * 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 = shortage + deficit; /* * Run the inactive scan on as many threads as is necessary. */ page_shortage = vm_pageout_inactive_dispatch(vmd, starting_page_shortage); addl_page_shortage = atomic_readandclear_int(&vmd->vmd_addl_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); } /* * 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); /* * 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")); + KASSERT(!vm_map_is_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; if (obj->type == OBJT_PHYS || obj->type == OBJT_VNODE || (obj->flags & OBJ_SWAP) != 0) res += obj->resident_page_count; } 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) { const char *reason; 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)) { 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(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) { switch (shortage) { case VM_OOM_MEM: reason = "failed to reclaim memory"; break; case VM_OOM_MEM_PF: reason = "a thread waited too long to allocate a page"; break; case VM_OOM_SWAPZ: reason = "out of swap space"; break; default: panic("unknown OOM reason %d", shortage); } if (vm_panic_on_oom != 0 && --vm_panic_on_oom == 0) panic("%s", reason); PROC_LOCK(bigproc); killproc(bigproc, reason); sched_nice(bigproc, PRIO_MIN); _PRELE(bigproc); PROC_UNLOCK(bigproc); } } /* * Signal a free page shortage to subsystems that have registered an event * handler. Reclaim memory from UMA in the event of a severe shortage. * Return true if the free page count should be re-evaluated. */ static bool vm_pageout_lowmem(void) { static int lowmem_ticks = 0; int last; bool ret; ret = false; 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. */ uma_reclaim(UMA_RECLAIM_TRIM); ret = true; break; } /* * Kick off an asynchronous reclaim of cached memory if one of the * page daemons is failing to keep up with demand. Use the "severe" * threshold instead of "min" to ensure that we do not blow away the * caches if a subset of the NUMA domains are depleted by kernel memory * allocations; the domainset iterators automatically skip domains * below the "min" threshold on the first pass. * * UMA reclaim worker has its own rate-limiting mechanism, so don't * worry about kicking it too often. */ if (vm_page_count_severe()) uma_reclaim_wakeup(); return (ret); } 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_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_helper runs additional pageout daemons in times of high paging * activity. */ static void vm_pageout_helper(void *arg) { struct vm_domain *vmd; int domain; domain = (uintptr_t)arg; vmd = VM_DOMAIN(domain); vm_domain_pageout_lock(vmd); for (;;) { msleep(&vmd->vmd_inactive_shortage, vm_domain_pageout_lockptr(vmd), PVM, "psleep", 0); blockcount_release(&vmd->vmd_inactive_starting, 1); vm_domain_pageout_unlock(vmd); vm_pageout_scan_inactive(vmd, vmd->vmd_inactive_shortage); vm_domain_pageout_lock(vmd); /* * Release the running count while the pageout lock is held to * prevent wakeup races. */ blockcount_release(&vmd->vmd_inactive_running, 1); } } static int get_pageout_threads_per_domain(const struct vm_domain *vmd) { unsigned total_pageout_threads, eligible_cpus, domain_cpus; if (VM_DOMAIN_EMPTY(vmd->vmd_domain)) return (0); /* * Semi-arbitrarily constrain pagedaemon threads to less than half the * total number of CPUs in the system as an upper limit. */ if (pageout_cpus_per_thread < 2) pageout_cpus_per_thread = 2; else if (pageout_cpus_per_thread > mp_ncpus) pageout_cpus_per_thread = mp_ncpus; total_pageout_threads = howmany(mp_ncpus, pageout_cpus_per_thread); domain_cpus = CPU_COUNT(&cpuset_domain[vmd->vmd_domain]); /* Pagedaemons are not run in empty domains. */ eligible_cpus = mp_ncpus; for (unsigned i = 0; i < vm_ndomains; i++) if (VM_DOMAIN_EMPTY(i)) eligible_cpus -= CPU_COUNT(&cpuset_domain[i]); /* * Assign a portion of the total pageout threads to this domain * corresponding to the fraction of pagedaemon-eligible CPUs in the * domain. In asymmetric NUMA systems, domains with more CPUs may be * allocated more threads than domains with fewer CPUs. */ return (howmany(total_pageout_threads * domain_cpus, eligible_cpus)); } /* * Initialize basic pageout daemon settings. See the comment above the * definition of vm_domain for some explanation of how these thresholds are * used. */ 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. */ 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_min = vmd->vmd_page_count / 200; 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 | CTLFLAG_MPSAFE, NULL, ""); pidctrl_init_sysctl(&vmd->vmd_pid, SYSCTL_CHILDREN(oid)); vmd->vmd_inactive_threads = get_pageout_threads_per_domain(vmd); } static void vm_pageout_init(void) { u_long freecount; int i; /* * Initialize some paging parameters. */ 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; /* * Set the maximum number of user-wired virtual pages. Historically the * main source of such pages was mlock(2) and mlockall(2). Hypervisors * may also request user-wired memory. */ if (vm_page_max_user_wired == 0) vm_page_max_user_wired = 4 * freecount / 5; } /* * vm_pageout is the high level pageout daemon. */ static void vm_pageout(void) { struct proc *p; struct thread *td; int error, first, i, j, pageout_threads; 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); } pageout_threads = VM_DOMAIN(i)->vmd_inactive_threads; for (j = 0; j < pageout_threads - 1; j++) { error = kthread_add(vm_pageout_helper, (void *)(uintptr_t)i, p, NULL, 0, 0, "dom%d helper%d", i, j); if (error != 0) panic("starting pageout helper %d for domain " "%d: %d\n", j, 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); } }