Index: head/sys/dev/xen/gntdev/gntdev.c =================================================================== --- head/sys/dev/xen/gntdev/gntdev.c (revision 353534) +++ head/sys/dev/xen/gntdev/gntdev.c (revision 353535) @@ -1,1291 +1,1291 @@ /*- * Copyright (c) 2016 Akshay Jaggi * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * gntdev.c * * Interface to /dev/xen/gntdev. * */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include MALLOC_DEFINE(M_GNTDEV, "gntdev", "Xen grant-table user-space device"); #define MAX_OFFSET_COUNT ((0xffffffffffffffffull >> PAGE_SHIFT) + 1) static d_open_t gntdev_open; static d_ioctl_t gntdev_ioctl; static d_mmap_single_t gntdev_mmap_single; static struct cdevsw gntdev_devsw = { .d_version = D_VERSION, .d_open = gntdev_open, .d_ioctl = gntdev_ioctl, .d_mmap_single = gntdev_mmap_single, .d_name = "gntdev", }; static device_t gntdev_dev = NULL; struct gntdev_gref; struct gntdev_gmap; STAILQ_HEAD(gref_list_head, gntdev_gref); STAILQ_HEAD(gmap_list_head, gntdev_gmap); RB_HEAD(gref_tree_head, gntdev_gref); RB_HEAD(gmap_tree_head, gntdev_gmap); struct file_offset_struct { RB_ENTRY(file_offset_struct) next; uint64_t file_offset; uint64_t count; }; static int offset_cmp(struct file_offset_struct *f1, struct file_offset_struct *f2) { return (f1->file_offset - f2->file_offset); } RB_HEAD(file_offset_head, file_offset_struct); RB_GENERATE_STATIC(file_offset_head, file_offset_struct, next, offset_cmp); struct per_user_data { struct mtx user_data_lock; struct gref_tree_head gref_tree; struct gmap_tree_head gmap_tree; struct file_offset_head file_offset; }; /* * Get offset into the file which will be used while mmapping the * appropriate pages by the userspace program. */ static int get_file_offset(struct per_user_data *priv_user, uint32_t count, uint64_t *file_offset) { struct file_offset_struct *offset, *offset_tmp; if (count == 0) return (EINVAL); mtx_lock(&priv_user->user_data_lock); RB_FOREACH_SAFE(offset, file_offset_head, &priv_user->file_offset, offset_tmp) { if (offset->count >= count) { offset->count -= count; *file_offset = offset->file_offset + offset->count * PAGE_SIZE; if (offset->count == 0) { RB_REMOVE(file_offset_head, &priv_user->file_offset, offset); free(offset, M_GNTDEV); } mtx_unlock(&priv_user->user_data_lock); return (0); } } mtx_unlock(&priv_user->user_data_lock); return (ENOSPC); } static void put_file_offset(struct per_user_data *priv_user, uint32_t count, uint64_t file_offset) { struct file_offset_struct *offset, *offset_nxt, *offset_prv; offset = malloc(sizeof(*offset), M_GNTDEV, M_WAITOK | M_ZERO); offset->file_offset = file_offset; offset->count = count; mtx_lock(&priv_user->user_data_lock); RB_INSERT(file_offset_head, &priv_user->file_offset, offset); offset_nxt = RB_NEXT(file_offset_head, &priv_user->file_offset, offset); offset_prv = RB_PREV(file_offset_head, &priv_user->file_offset, offset); if (offset_nxt != NULL && offset_nxt->file_offset == offset->file_offset + offset->count * PAGE_SIZE) { offset->count += offset_nxt->count; RB_REMOVE(file_offset_head, &priv_user->file_offset, offset_nxt); free(offset_nxt, M_GNTDEV); } if (offset_prv != NULL && offset->file_offset == offset_prv->file_offset + offset_prv->count * PAGE_SIZE) { offset_prv->count += offset->count; RB_REMOVE(file_offset_head, &priv_user->file_offset, offset); free(offset, M_GNTDEV); } mtx_unlock(&priv_user->user_data_lock); } static int gntdev_gmap_pg_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t foff, struct ucred *cred, u_short *color); static void gntdev_gmap_pg_dtor(void *handle); static int gntdev_gmap_pg_fault(vm_object_t object, vm_ooffset_t offset, int prot, vm_page_t *mres); static struct cdev_pager_ops gntdev_gmap_pg_ops = { .cdev_pg_fault = gntdev_gmap_pg_fault, .cdev_pg_ctor = gntdev_gmap_pg_ctor, .cdev_pg_dtor = gntdev_gmap_pg_dtor, }; struct cleanup_data_struct { struct mtx to_kill_grefs_mtx; struct mtx to_kill_gmaps_mtx; struct gref_list_head to_kill_grefs; struct gmap_list_head to_kill_gmaps; }; static struct cleanup_data_struct cleanup_data = { .to_kill_grefs = STAILQ_HEAD_INITIALIZER(cleanup_data.to_kill_grefs), .to_kill_gmaps = STAILQ_HEAD_INITIALIZER(cleanup_data.to_kill_gmaps), }; MTX_SYSINIT(to_kill_grefs_mtx, &cleanup_data.to_kill_grefs_mtx, "gntdev to_kill_grefs mutex", MTX_DEF); MTX_SYSINIT(to_kill_gmaps_mtx, &cleanup_data.to_kill_gmaps_mtx, "gntdev to_kill_gmaps mutex", MTX_DEF); static void cleanup_function(void *arg, __unused int pending); static struct task cleanup_task = TASK_INITIALIZER(0, cleanup_function, &cleanup_data); struct notify_data { uint64_t index; uint32_t action; uint32_t event_channel_port; xen_intr_handle_t notify_evtchn_handle; }; static void notify(struct notify_data *notify, vm_page_t page); /*-------------------- Grant Allocation Methods -----------------------------*/ struct gntdev_gref { union gref_next_union { STAILQ_ENTRY(gntdev_gref) list; RB_ENTRY(gntdev_gref) tree; } gref_next; uint64_t file_index; grant_ref_t gref_id; vm_page_t page; struct notify_data *notify; }; static int gref_cmp(struct gntdev_gref *g1, struct gntdev_gref *g2) { return (g1->file_index - g2->file_index); } RB_GENERATE_STATIC(gref_tree_head, gntdev_gref, gref_next.tree, gref_cmp); /* * Traverse over the device-list of to-be-deleted grants allocated, and * if all accesses, both local mmaps and foreign maps, to them have ended, * destroy them. */ static void gref_list_dtor(struct cleanup_data_struct *cleanup_data) { struct gref_list_head tmp_grefs; struct gntdev_gref *gref, *gref_tmp, *gref_previous; STAILQ_INIT(&tmp_grefs); mtx_lock(&cleanup_data->to_kill_grefs_mtx); STAILQ_SWAP(&cleanup_data->to_kill_grefs, &tmp_grefs, gntdev_gref); mtx_unlock(&cleanup_data->to_kill_grefs_mtx); gref_previous = NULL; STAILQ_FOREACH_SAFE(gref, &tmp_grefs, gref_next.list, gref_tmp) { if (gref->page && gref->page->object == NULL) { if (gref->notify) { notify(gref->notify, gref->page); } if (gref->gref_id != GRANT_REF_INVALID) { if (gnttab_query_foreign_access(gref->gref_id)) continue; if (gnttab_end_foreign_access_ref(gref->gref_id) == 0) continue; gnttab_free_grant_reference(gref->gref_id); } vm_page_unwire_noq(gref->page); vm_page_free(gref->page); gref->page = NULL; } if (gref->page == NULL) { if (gref_previous == NULL) STAILQ_REMOVE_HEAD(&tmp_grefs, gref_next.list); else STAILQ_REMOVE_AFTER(&tmp_grefs, gref_previous, gref_next.list); if (gref->notify) free(gref->notify, M_GNTDEV); free(gref, M_GNTDEV); } else gref_previous = gref; } if (!STAILQ_EMPTY(&tmp_grefs)) { mtx_lock(&cleanup_data->to_kill_grefs_mtx); STAILQ_CONCAT(&cleanup_data->to_kill_grefs, &tmp_grefs); mtx_unlock(&cleanup_data->to_kill_grefs_mtx); } } /* * Find count number of contiguous allocated grants for a given userspace * program by file-offset (index). */ static struct gntdev_gref* gntdev_find_grefs(struct per_user_data *priv_user, uint64_t index, uint32_t count) { struct gntdev_gref find_gref, *gref, *gref_start = NULL; find_gref.file_index = index; mtx_lock(&priv_user->user_data_lock); gref_start = RB_FIND(gref_tree_head, &priv_user->gref_tree, &find_gref); for (gref = gref_start; gref != NULL && count > 0; gref = RB_NEXT(gref_tree_head, &priv_user->gref_tree, gref)) { if (index != gref->file_index) break; index += PAGE_SIZE; count--; } mtx_unlock(&priv_user->user_data_lock); if (count) return (NULL); return (gref_start); } /* * IOCTL_GNTDEV_ALLOC_GREF * Allocate required number of wired pages for the request, grant foreign * access to the physical frames for these pages, and add details about * this allocation to the per user private data, so that these pages can * be mmapped by the userspace program. */ static int gntdev_alloc_gref(struct ioctl_gntdev_alloc_gref *arg) { uint32_t i; int error, readonly; uint64_t file_offset; struct gntdev_gref *grefs; struct per_user_data *priv_user; readonly = !(arg->flags & GNTDEV_ALLOC_FLAG_WRITABLE); error = devfs_get_cdevpriv((void**) &priv_user); if (error != 0) return (EINVAL); /* Cleanup grefs and free pages. */ taskqueue_enqueue(taskqueue_thread, &cleanup_task); /* Get file offset for this request. */ error = get_file_offset(priv_user, arg->count, &file_offset); if (error != 0) return (error); /* Allocate grefs. */ grefs = malloc(sizeof(*grefs) * arg->count, M_GNTDEV, M_WAITOK); for (i = 0; i < arg->count; i++) { grefs[i].file_index = file_offset + i * PAGE_SIZE; grefs[i].gref_id = GRANT_REF_INVALID; grefs[i].notify = NULL; grefs[i].page = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO); if (grefs[i].page == NULL) { log(LOG_ERR, "Page allocation failed."); error = ENOMEM; break; } if ((grefs[i].page->flags & PG_ZERO) == 0) { /* * Zero the allocated page, as we don't want to * leak our memory to other domains. */ pmap_zero_page(grefs[i].page); } grefs[i].page->valid = VM_PAGE_BITS_ALL; error = gnttab_grant_foreign_access(arg->domid, (VM_PAGE_TO_PHYS(grefs[i].page) >> PAGE_SHIFT), readonly, &grefs[i].gref_id); if (error != 0) { log(LOG_ERR, "Grant Table Hypercall failed."); break; } } if (error != 0) { /* * If target domain maps the gref (by guessing the gref-id), * then we can't clean it up yet and we have to leave the * page in place so as to not leak our memory to that domain. * Add it to a global list to be cleaned up later. */ mtx_lock(&cleanup_data.to_kill_grefs_mtx); for (i = 0; i < arg->count; i++) STAILQ_INSERT_TAIL(&cleanup_data.to_kill_grefs, &grefs[i], gref_next.list); mtx_unlock(&cleanup_data.to_kill_grefs_mtx); taskqueue_enqueue(taskqueue_thread, &cleanup_task); return (error); } /* Copy the output values. */ arg->index = file_offset; for (i = 0; i < arg->count; i++) suword32(&arg->gref_ids[i], grefs[i].gref_id); /* Modify the per user private data. */ mtx_lock(&priv_user->user_data_lock); for (i = 0; i < arg->count; i++) RB_INSERT(gref_tree_head, &priv_user->gref_tree, &grefs[i]); mtx_unlock(&priv_user->user_data_lock); return (error); } /* * IOCTL_GNTDEV_DEALLOC_GREF * Remove grant allocation information from the per user private data, so * that it can't be mmapped anymore by the userspace program, and add it * to the to-be-deleted grants global device-list. */ static int gntdev_dealloc_gref(struct ioctl_gntdev_dealloc_gref *arg) { int error; uint32_t count; struct gntdev_gref *gref, *gref_tmp; struct per_user_data *priv_user; error = devfs_get_cdevpriv((void**) &priv_user); if (error != 0) return (EINVAL); gref = gntdev_find_grefs(priv_user, arg->index, arg->count); if (gref == NULL) { log(LOG_ERR, "Can't find requested grant-refs."); return (EINVAL); } /* Remove the grefs from user private data. */ count = arg->count; mtx_lock(&priv_user->user_data_lock); mtx_lock(&cleanup_data.to_kill_grefs_mtx); for (; gref != NULL && count > 0; gref = gref_tmp) { gref_tmp = RB_NEXT(gref_tree_head, &priv_user->gref_tree, gref); RB_REMOVE(gref_tree_head, &priv_user->gref_tree, gref); STAILQ_INSERT_TAIL(&cleanup_data.to_kill_grefs, gref, gref_next.list); count--; } mtx_unlock(&cleanup_data.to_kill_grefs_mtx); mtx_unlock(&priv_user->user_data_lock); taskqueue_enqueue(taskqueue_thread, &cleanup_task); put_file_offset(priv_user, arg->count, arg->index); return (0); } /*-------------------- Grant Mapping Methods --------------------------------*/ struct gntdev_gmap_map { vm_object_t mem; struct resource *pseudo_phys_res; int pseudo_phys_res_id; vm_paddr_t phys_base_addr; }; struct gntdev_gmap { union gmap_next_union { STAILQ_ENTRY(gntdev_gmap) list; RB_ENTRY(gntdev_gmap) tree; } gmap_next; uint64_t file_index; uint32_t count; struct gnttab_map_grant_ref *grant_map_ops; struct gntdev_gmap_map *map; struct notify_data *notify; }; static int gmap_cmp(struct gntdev_gmap *g1, struct gntdev_gmap *g2) { return (g1->file_index - g2->file_index); } RB_GENERATE_STATIC(gmap_tree_head, gntdev_gmap, gmap_next.tree, gmap_cmp); /* * Traverse over the device-list of to-be-deleted grant mappings, and if * the region is no longer mmapped by anyone, free the memory used to * store information about the mapping. */ static void gmap_list_dtor(struct cleanup_data_struct *cleanup_data) { struct gmap_list_head tmp_gmaps; struct gntdev_gmap *gmap, *gmap_tmp, *gmap_previous; STAILQ_INIT(&tmp_gmaps); mtx_lock(&cleanup_data->to_kill_gmaps_mtx); STAILQ_SWAP(&cleanup_data->to_kill_gmaps, &tmp_gmaps, gntdev_gmap); mtx_unlock(&cleanup_data->to_kill_gmaps_mtx); gmap_previous = NULL; STAILQ_FOREACH_SAFE(gmap, &tmp_gmaps, gmap_next.list, gmap_tmp) { if (gmap->map == NULL) { if (gmap_previous == NULL) STAILQ_REMOVE_HEAD(&tmp_gmaps, gmap_next.list); else STAILQ_REMOVE_AFTER(&tmp_gmaps, gmap_previous, gmap_next.list); if (gmap->notify) free(gmap->notify, M_GNTDEV); free(gmap->grant_map_ops, M_GNTDEV); free(gmap, M_GNTDEV); } else gmap_previous = gmap; } if (!STAILQ_EMPTY(&tmp_gmaps)) { mtx_lock(&cleanup_data->to_kill_gmaps_mtx); STAILQ_CONCAT(&cleanup_data->to_kill_gmaps, &tmp_gmaps); mtx_unlock(&cleanup_data->to_kill_gmaps_mtx); } } /* * Find mapped grants for a given userspace program, by file-offset (index) * and count, as supplied during the map-ioctl. */ static struct gntdev_gmap* gntdev_find_gmap(struct per_user_data *priv_user, uint64_t index, uint32_t count) { struct gntdev_gmap find_gmap, *gmap; find_gmap.file_index = index; mtx_lock(&priv_user->user_data_lock); gmap = RB_FIND(gmap_tree_head, &priv_user->gmap_tree, &find_gmap); mtx_unlock(&priv_user->user_data_lock); if (gmap != NULL && gmap->count == count) return (gmap); return (NULL); } /* * Remove the pages from the mgtdevice pager, call the unmap hypercall, * free the xenmem resource. This function is called during the * destruction of the mgtdevice pager, which happens when all mmaps to * it have been removed, and the unmap-ioctl has been performed. */ static int notify_unmap_cleanup(struct gntdev_gmap *gmap) { uint32_t i; int error, count; vm_page_t m; struct gnttab_unmap_grant_ref *unmap_ops; unmap_ops = malloc(sizeof(struct gnttab_unmap_grant_ref) * gmap->count, M_GNTDEV, M_WAITOK); /* Enumerate freeable maps. */ count = 0; for (i = 0; i < gmap->count; i++) { if (gmap->grant_map_ops[i].handle != -1) { unmap_ops[count].handle = gmap->grant_map_ops[i].handle; unmap_ops[count].host_addr = gmap->grant_map_ops[i].host_addr; unmap_ops[count].dev_bus_addr = 0; count++; } } /* Perform notification. */ if (count > 0 && gmap->notify) { vm_page_t page; uint64_t page_offset; page_offset = gmap->notify->index - gmap->file_index; page = PHYS_TO_VM_PAGE(gmap->map->phys_base_addr + page_offset); notify(gmap->notify, page); } /* Free the pages. */ VM_OBJECT_WLOCK(gmap->map->mem); retry: for (i = 0; i < gmap->count; i++) { m = vm_page_lookup(gmap->map->mem, i); if (m == NULL) continue; if (vm_page_busy_acquire(m, VM_ALLOC_WAITFAIL) == 0) goto retry; cdev_pager_free_page(gmap->map->mem, m); } VM_OBJECT_WUNLOCK(gmap->map->mem); /* Perform unmap hypercall. */ error = HYPERVISOR_grant_table_op(GNTTABOP_unmap_grant_ref, unmap_ops, count); for (i = 0; i < gmap->count; i++) { gmap->grant_map_ops[i].handle = -1; gmap->grant_map_ops[i].host_addr = 0; } if (gmap->map) { error = xenmem_free(gntdev_dev, gmap->map->pseudo_phys_res_id, gmap->map->pseudo_phys_res); KASSERT(error == 0, ("Unable to release memory resource: %d", error)); free(gmap->map, M_GNTDEV); gmap->map = NULL; } free(unmap_ops, M_GNTDEV); return (error); } /* * IOCTL_GNTDEV_MAP_GRANT_REF * Populate structures for mapping the grant reference in the per user * private data. Actual resource allocation and map hypercall is performed * during the mmap. */ static int gntdev_map_grant_ref(struct ioctl_gntdev_map_grant_ref *arg) { uint32_t i; int error; struct gntdev_gmap *gmap; struct per_user_data *priv_user; error = devfs_get_cdevpriv((void**) &priv_user); if (error != 0) return (EINVAL); gmap = malloc(sizeof(*gmap), M_GNTDEV, M_WAITOK | M_ZERO); gmap->count = arg->count; gmap->grant_map_ops = malloc(sizeof(struct gnttab_map_grant_ref) * arg->count, M_GNTDEV, M_WAITOK | M_ZERO); for (i = 0; i < arg->count; i++) { struct ioctl_gntdev_grant_ref ref; error = copyin(&arg->refs[i], &ref, sizeof(ref)); if (error != 0) { free(gmap->grant_map_ops, M_GNTDEV); free(gmap, M_GNTDEV); return (error); } gmap->grant_map_ops[i].dom = ref.domid; gmap->grant_map_ops[i].ref = ref.ref; gmap->grant_map_ops[i].handle = -1; gmap->grant_map_ops[i].flags = GNTMAP_host_map; } error = get_file_offset(priv_user, arg->count, &gmap->file_index); if (error != 0) { free(gmap->grant_map_ops, M_GNTDEV); free(gmap, M_GNTDEV); return (error); } mtx_lock(&priv_user->user_data_lock); RB_INSERT(gmap_tree_head, &priv_user->gmap_tree, gmap); mtx_unlock(&priv_user->user_data_lock); arg->index = gmap->file_index; return (error); } /* * IOCTL_GNTDEV_UNMAP_GRANT_REF * Remove the map information from the per user private data and add it * to the global device-list of mappings to be deleted. A reference to * the mgtdevice pager is also decreased, the reason for which is * explained in mmap_gmap(). */ static int gntdev_unmap_grant_ref(struct ioctl_gntdev_unmap_grant_ref *arg) { int error; struct gntdev_gmap *gmap; struct per_user_data *priv_user; error = devfs_get_cdevpriv((void**) &priv_user); if (error != 0) return (EINVAL); gmap = gntdev_find_gmap(priv_user, arg->index, arg->count); if (gmap == NULL) { log(LOG_ERR, "Can't find requested grant-map."); return (EINVAL); } mtx_lock(&priv_user->user_data_lock); mtx_lock(&cleanup_data.to_kill_gmaps_mtx); RB_REMOVE(gmap_tree_head, &priv_user->gmap_tree, gmap); STAILQ_INSERT_TAIL(&cleanup_data.to_kill_gmaps, gmap, gmap_next.list); mtx_unlock(&cleanup_data.to_kill_gmaps_mtx); mtx_unlock(&priv_user->user_data_lock); if (gmap->map) vm_object_deallocate(gmap->map->mem); taskqueue_enqueue(taskqueue_thread, &cleanup_task); put_file_offset(priv_user, arg->count, arg->index); return (0); } /* * IOCTL_GNTDEV_GET_OFFSET_FOR_VADDR * Get file-offset and count for a given mapping, from the virtual address * where the mapping is mmapped. * Please note, this only works for grants mapped by this domain, and not * grants allocated. Count doesn't make much sense in reference to grants * allocated. Also, because this function is present in the linux gntdev * device, but not in the linux gntalloc one, most userspace code only use * it for mapped grants. */ static int gntdev_get_offset_for_vaddr(struct ioctl_gntdev_get_offset_for_vaddr *arg, struct thread *td) { int error; vm_map_t map; vm_map_entry_t entry; vm_object_t mem; vm_pindex_t pindex; vm_prot_t prot; boolean_t wired; struct gntdev_gmap *gmap; int rc; map = &td->td_proc->p_vmspace->vm_map; error = vm_map_lookup(&map, arg->vaddr, VM_PROT_NONE, &entry, &mem, &pindex, &prot, &wired); if (error != KERN_SUCCESS) return (EINVAL); if ((mem->type != OBJT_MGTDEVICE) || (mem->un_pager.devp.ops != &gntdev_gmap_pg_ops)) { rc = EINVAL; goto out; } gmap = mem->handle; if (gmap == NULL || (entry->end - entry->start) != (gmap->count * PAGE_SIZE)) { rc = EINVAL; goto out; } arg->count = gmap->count; arg->offset = gmap->file_index; rc = 0; out: vm_map_lookup_done(map, entry); return (rc); } /*-------------------- Grant Mapping Pager ----------------------------------*/ static int gntdev_gmap_pg_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t foff, struct ucred *cred, u_short *color) { return (0); } static void gntdev_gmap_pg_dtor(void *handle) { notify_unmap_cleanup((struct gntdev_gmap *)handle); } static int gntdev_gmap_pg_fault(vm_object_t object, vm_ooffset_t offset, int prot, vm_page_t *mres) { struct gntdev_gmap *gmap = object->handle; vm_pindex_t pidx, ridx; vm_page_t page, oldm; vm_ooffset_t relative_offset; if (gmap->map == NULL) return (VM_PAGER_FAIL); relative_offset = offset - gmap->file_index; pidx = OFF_TO_IDX(offset); ridx = OFF_TO_IDX(relative_offset); if (ridx >= gmap->count || gmap->grant_map_ops[ridx].status != GNTST_okay) return (VM_PAGER_FAIL); page = PHYS_TO_VM_PAGE(gmap->map->phys_base_addr + relative_offset); if (page == NULL) return (VM_PAGER_FAIL); KASSERT((page->flags & PG_FICTITIOUS) != 0, ("not fictitious %p", page)); KASSERT(vm_page_wired(page), ("page %p is not wired", page)); KASSERT(!vm_page_busied(page), ("page %p is busy", page)); if (*mres != NULL) { oldm = *mres; vm_page_free(oldm); *mres = NULL; } + vm_page_busy_acquire(page, 0); vm_page_insert(page, object, pidx); page->valid = VM_PAGE_BITS_ALL; - vm_page_xbusy(page); *mres = page; return (VM_PAGER_OK); } /*------------------ Grant Table Methods ------------------------------------*/ static void notify(struct notify_data *notify, vm_page_t page) { if (notify->action & UNMAP_NOTIFY_CLEAR_BYTE) { uint8_t *mem; uint64_t offset; offset = notify->index & PAGE_MASK; mem = (uint8_t *)pmap_quick_enter_page(page); mem[offset] = 0; pmap_quick_remove_page((vm_offset_t)mem); } if (notify->action & UNMAP_NOTIFY_SEND_EVENT) { xen_intr_signal(notify->notify_evtchn_handle); xen_intr_unbind(¬ify->notify_evtchn_handle); } notify->action = 0; } /* * Helper to copy new arguments from the notify ioctl into * the existing notify data. */ static int copy_notify_helper(struct notify_data *destination, struct ioctl_gntdev_unmap_notify *source) { xen_intr_handle_t handlep = NULL; /* * "Get" before "Put"ting previous reference, as we might be * holding the last reference to the event channel port. */ if (source->action & UNMAP_NOTIFY_SEND_EVENT) if (xen_intr_get_evtchn_from_port(source->event_channel_port, &handlep) != 0) return (EINVAL); if (destination->action & UNMAP_NOTIFY_SEND_EVENT) xen_intr_unbind(&destination->notify_evtchn_handle); destination->action = source->action; destination->event_channel_port = source->event_channel_port; destination->index = source->index; destination->notify_evtchn_handle = handlep; return (0); } /* * IOCTL_GNTDEV_SET_UNMAP_NOTIFY * Set unmap notification inside the appropriate grant. It sends a * notification when the grant is completely munmapped by this domain * and ready for destruction. */ static int gntdev_set_unmap_notify(struct ioctl_gntdev_unmap_notify *arg) { int error; uint64_t index; struct per_user_data *priv_user; struct gntdev_gref *gref = NULL; struct gntdev_gmap *gmap; error = devfs_get_cdevpriv((void**) &priv_user); if (error != 0) return (EINVAL); if (arg->action & ~(UNMAP_NOTIFY_CLEAR_BYTE|UNMAP_NOTIFY_SEND_EVENT)) return (EINVAL); index = arg->index & ~PAGE_MASK; gref = gntdev_find_grefs(priv_user, index, 1); if (gref) { if (gref->notify == NULL) gref->notify = malloc(sizeof(*arg), M_GNTDEV, M_WAITOK | M_ZERO); return (copy_notify_helper(gref->notify, arg)); } error = EINVAL; mtx_lock(&priv_user->user_data_lock); RB_FOREACH(gmap, gmap_tree_head, &priv_user->gmap_tree) { if (arg->index >= gmap->file_index && arg->index < gmap->file_index + gmap->count * PAGE_SIZE) { if (gmap->notify == NULL) gmap->notify = malloc(sizeof(*arg), M_GNTDEV, M_WAITOK | M_ZERO); error = copy_notify_helper(gmap->notify, arg); break; } } mtx_unlock(&priv_user->user_data_lock); return (error); } /*------------------ Gntdev Char Device Methods -----------------------------*/ static void cleanup_function(void *arg, __unused int pending) { gref_list_dtor((struct cleanup_data_struct *) arg); gmap_list_dtor((struct cleanup_data_struct *) arg); } static void per_user_data_dtor(void *arg) { struct gntdev_gref *gref, *gref_tmp; struct gntdev_gmap *gmap, *gmap_tmp; struct file_offset_struct *offset, *offset_tmp; struct per_user_data *priv_user; priv_user = (struct per_user_data *) arg; mtx_lock(&priv_user->user_data_lock); mtx_lock(&cleanup_data.to_kill_grefs_mtx); RB_FOREACH_SAFE(gref, gref_tree_head, &priv_user->gref_tree, gref_tmp) { RB_REMOVE(gref_tree_head, &priv_user->gref_tree, gref); STAILQ_INSERT_TAIL(&cleanup_data.to_kill_grefs, gref, gref_next.list); } mtx_unlock(&cleanup_data.to_kill_grefs_mtx); mtx_lock(&cleanup_data.to_kill_gmaps_mtx); RB_FOREACH_SAFE(gmap, gmap_tree_head, &priv_user->gmap_tree, gmap_tmp) { RB_REMOVE(gmap_tree_head, &priv_user->gmap_tree, gmap); STAILQ_INSERT_TAIL(&cleanup_data.to_kill_gmaps, gmap, gmap_next.list); if (gmap->map) vm_object_deallocate(gmap->map->mem); } mtx_unlock(&cleanup_data.to_kill_gmaps_mtx); RB_FOREACH_SAFE(offset, file_offset_head, &priv_user->file_offset, offset_tmp) { RB_REMOVE(file_offset_head, &priv_user->file_offset, offset); free(offset, M_GNTDEV); } mtx_unlock(&priv_user->user_data_lock); taskqueue_enqueue(taskqueue_thread, &cleanup_task); mtx_destroy(&priv_user->user_data_lock); free(priv_user, M_GNTDEV); } static int gntdev_open(struct cdev *dev, int flag, int otyp, struct thread *td) { int error; struct per_user_data *priv_user; struct file_offset_struct *offset; priv_user = malloc(sizeof(*priv_user), M_GNTDEV, M_WAITOK | M_ZERO); RB_INIT(&priv_user->gref_tree); RB_INIT(&priv_user->gmap_tree); RB_INIT(&priv_user->file_offset); offset = malloc(sizeof(*offset), M_GNTDEV, M_WAITOK | M_ZERO); offset->file_offset = 0; offset->count = MAX_OFFSET_COUNT; RB_INSERT(file_offset_head, &priv_user->file_offset, offset); mtx_init(&priv_user->user_data_lock, "per user data mutex", NULL, MTX_DEF); error = devfs_set_cdevpriv(priv_user, per_user_data_dtor); if (error != 0) per_user_data_dtor(priv_user); return (error); } static int gntdev_ioctl(struct cdev *dev, u_long cmd, caddr_t data, int fflag, struct thread *td) { int error; switch (cmd) { case IOCTL_GNTDEV_SET_UNMAP_NOTIFY: error = gntdev_set_unmap_notify( (struct ioctl_gntdev_unmap_notify*) data); break; case IOCTL_GNTDEV_ALLOC_GREF: error = gntdev_alloc_gref( (struct ioctl_gntdev_alloc_gref*) data); break; case IOCTL_GNTDEV_DEALLOC_GREF: error = gntdev_dealloc_gref( (struct ioctl_gntdev_dealloc_gref*) data); break; case IOCTL_GNTDEV_MAP_GRANT_REF: error = gntdev_map_grant_ref( (struct ioctl_gntdev_map_grant_ref*) data); break; case IOCTL_GNTDEV_UNMAP_GRANT_REF: error = gntdev_unmap_grant_ref( (struct ioctl_gntdev_unmap_grant_ref*) data); break; case IOCTL_GNTDEV_GET_OFFSET_FOR_VADDR: error = gntdev_get_offset_for_vaddr( (struct ioctl_gntdev_get_offset_for_vaddr*) data, td); break; default: error = ENOSYS; break; } return (error); } /* * MMAP an allocated grant into user memory. * Please note, that the grants must not already be mmapped, otherwise * this function will fail. */ static int mmap_gref(struct per_user_data *priv_user, struct gntdev_gref *gref_start, uint32_t count, vm_size_t size, struct vm_object **object) { vm_object_t mem_obj; struct gntdev_gref *gref; mem_obj = vm_object_allocate(OBJT_PHYS, size); if (mem_obj == NULL) return (ENOMEM); mtx_lock(&priv_user->user_data_lock); VM_OBJECT_WLOCK(mem_obj); for (gref = gref_start; gref != NULL && count > 0; gref = RB_NEXT(gref_tree_head, &priv_user->gref_tree, gref)) { if (gref->page->object) break; vm_page_insert(gref->page, mem_obj, OFF_TO_IDX(gref->file_index)); count--; } VM_OBJECT_WUNLOCK(mem_obj); mtx_unlock(&priv_user->user_data_lock); if (count) { vm_object_deallocate(mem_obj); return (EINVAL); } *object = mem_obj; return (0); } /* * MMAP a mapped grant into user memory. */ static int mmap_gmap(struct per_user_data *priv_user, struct gntdev_gmap *gmap_start, vm_ooffset_t *offset, vm_size_t size, struct vm_object **object, int nprot) { uint32_t i; int error; /* * The grant map hypercall might already be done. * If that is the case, increase a reference to the * vm object and return the already allocated object. */ if (gmap_start->map) { vm_object_reference(gmap_start->map->mem); *object = gmap_start->map->mem; return (0); } gmap_start->map = malloc(sizeof(*(gmap_start->map)), M_GNTDEV, M_WAITOK | M_ZERO); /* Allocate the xen pseudo physical memory resource. */ gmap_start->map->pseudo_phys_res_id = 0; gmap_start->map->pseudo_phys_res = xenmem_alloc(gntdev_dev, &gmap_start->map->pseudo_phys_res_id, size); if (gmap_start->map->pseudo_phys_res == NULL) { free(gmap_start->map, M_GNTDEV); gmap_start->map = NULL; return (ENOMEM); } gmap_start->map->phys_base_addr = rman_get_start(gmap_start->map->pseudo_phys_res); /* Allocate the mgtdevice pager. */ gmap_start->map->mem = cdev_pager_allocate(gmap_start, OBJT_MGTDEVICE, &gntdev_gmap_pg_ops, size, nprot, *offset, NULL); if (gmap_start->map->mem == NULL) { xenmem_free(gntdev_dev, gmap_start->map->pseudo_phys_res_id, gmap_start->map->pseudo_phys_res); free(gmap_start->map, M_GNTDEV); gmap_start->map = NULL; return (ENOMEM); } for (i = 0; i < gmap_start->count; i++) { gmap_start->grant_map_ops[i].host_addr = gmap_start->map->phys_base_addr + i * PAGE_SIZE; if ((nprot & PROT_WRITE) == 0) gmap_start->grant_map_ops[i].flags |= GNTMAP_readonly; } /* Make the MAP hypercall. */ error = HYPERVISOR_grant_table_op(GNTTABOP_map_grant_ref, gmap_start->grant_map_ops, gmap_start->count); if (error != 0) { /* * Deallocate pager. * Pager deallocation will automatically take care of * xenmem deallocation, etc. */ vm_object_deallocate(gmap_start->map->mem); return (EINVAL); } /* Retry EAGAIN maps. */ for (i = 0; i < gmap_start->count; i++) { int delay = 1; while (delay < 256 && gmap_start->grant_map_ops[i].status == GNTST_eagain) { HYPERVISOR_grant_table_op( GNTTABOP_map_grant_ref, &gmap_start->grant_map_ops[i], 1); pause(("gntmap"), delay * SBT_1MS); delay++; } if (gmap_start->grant_map_ops[i].status == GNTST_eagain) gmap_start->grant_map_ops[i].status = GNTST_bad_page; if (gmap_start->grant_map_ops[i].status != GNTST_okay) { /* * Deallocate pager. * Pager deallocation will automatically take care of * xenmem deallocation, notification, unmap hypercall, * etc. */ vm_object_deallocate(gmap_start->map->mem); return (EINVAL); } } /* * Add a reference to the vm object. We do not want * the vm object to be deleted when all the mmaps are * unmapped, because it may be re-mmapped. Instead, * we want the object to be deleted, when along with * munmaps, we have also processed the unmap-ioctl. */ vm_object_reference(gmap_start->map->mem); *object = gmap_start->map->mem; return (0); } static int gntdev_mmap_single(struct cdev *cdev, vm_ooffset_t *offset, vm_size_t size, struct vm_object **object, int nprot) { int error; uint32_t count; struct gntdev_gref *gref_start; struct gntdev_gmap *gmap_start; struct per_user_data *priv_user; error = devfs_get_cdevpriv((void**) &priv_user); if (error != 0) return (EINVAL); count = OFF_TO_IDX(size); gref_start = gntdev_find_grefs(priv_user, *offset, count); if (gref_start) { error = mmap_gref(priv_user, gref_start, count, size, object); return (error); } gmap_start = gntdev_find_gmap(priv_user, *offset, count); if (gmap_start) { error = mmap_gmap(priv_user, gmap_start, offset, size, object, nprot); return (error); } return (EINVAL); } /*------------------ Private Device Attachment Functions --------------------*/ static void gntdev_identify(driver_t *driver, device_t parent) { KASSERT((xen_domain()), ("Trying to attach gntdev device on non Xen domain")); if (BUS_ADD_CHILD(parent, 0, "gntdev", 0) == NULL) panic("unable to attach gntdev user-space device"); } static int gntdev_probe(device_t dev) { gntdev_dev = dev; device_set_desc(dev, "Xen grant-table user-space device"); return (BUS_PROBE_NOWILDCARD); } static int gntdev_attach(device_t dev) { make_dev_credf(MAKEDEV_ETERNAL, &gntdev_devsw, 0, NULL, UID_ROOT, GID_WHEEL, 0600, "xen/gntdev"); return (0); } /*-------------------- Private Device Attachment Data -----------------------*/ static device_method_t gntdev_methods[] = { DEVMETHOD(device_identify, gntdev_identify), DEVMETHOD(device_probe, gntdev_probe), DEVMETHOD(device_attach, gntdev_attach), DEVMETHOD_END }; static driver_t gntdev_driver = { "gntdev", gntdev_methods, 0, }; devclass_t gntdev_devclass; DRIVER_MODULE(gntdev, xenpv, gntdev_driver, gntdev_devclass, 0, 0); MODULE_DEPEND(gntdev, xenpv, 1, 1, 1); Index: head/sys/dev/xen/privcmd/privcmd.c =================================================================== --- head/sys/dev/xen/privcmd/privcmd.c (revision 353534) +++ head/sys/dev/xen/privcmd/privcmd.c (revision 353535) @@ -1,429 +1,429 @@ /* * Copyright (c) 2014 Roger Pau MonnĂ© * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include MALLOC_DEFINE(M_PRIVCMD, "privcmd_dev", "Xen privcmd user-space device"); struct privcmd_map { vm_object_t mem; vm_size_t size; struct resource *pseudo_phys_res; int pseudo_phys_res_id; vm_paddr_t phys_base_addr; boolean_t mapped; BITSET_DEFINE_VAR() *err; }; static d_ioctl_t privcmd_ioctl; static d_mmap_single_t privcmd_mmap_single; static struct cdevsw privcmd_devsw = { .d_version = D_VERSION, .d_ioctl = privcmd_ioctl, .d_mmap_single = privcmd_mmap_single, .d_name = "privcmd", }; static int privcmd_pg_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t foff, struct ucred *cred, u_short *color); static void privcmd_pg_dtor(void *handle); static int privcmd_pg_fault(vm_object_t object, vm_ooffset_t offset, int prot, vm_page_t *mres); static struct cdev_pager_ops privcmd_pg_ops = { .cdev_pg_fault = privcmd_pg_fault, .cdev_pg_ctor = privcmd_pg_ctor, .cdev_pg_dtor = privcmd_pg_dtor, }; static device_t privcmd_dev = NULL; /*------------------------- Privcmd Pager functions --------------------------*/ static int privcmd_pg_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t foff, struct ucred *cred, u_short *color) { return (0); } static void privcmd_pg_dtor(void *handle) { struct xen_remove_from_physmap rm = { .domid = DOMID_SELF }; struct privcmd_map *map = handle; int error; vm_size_t i; vm_page_t m; /* * Remove the mappings from the used pages. This will remove the * underlying p2m bindings in Xen second stage translation. */ if (map->mapped == true) { VM_OBJECT_WLOCK(map->mem); retry: for (i = 0; i < map->size; i++) { m = vm_page_lookup(map->mem, i); if (m == NULL) continue; if (vm_page_busy_acquire(m, VM_ALLOC_WAITFAIL) == 0) goto retry; cdev_pager_free_page(map->mem, m); } VM_OBJECT_WUNLOCK(map->mem); for (i = 0; i < map->size; i++) { rm.gpfn = atop(map->phys_base_addr) + i; HYPERVISOR_memory_op(XENMEM_remove_from_physmap, &rm); } free(map->err, M_PRIVCMD); } error = xenmem_free(privcmd_dev, map->pseudo_phys_res_id, map->pseudo_phys_res); KASSERT(error == 0, ("Unable to release memory resource: %d", error)); free(map, M_PRIVCMD); } static int privcmd_pg_fault(vm_object_t object, vm_ooffset_t offset, int prot, vm_page_t *mres) { struct privcmd_map *map = object->handle; vm_pindex_t pidx; vm_page_t page, oldm; if (map->mapped != true) return (VM_PAGER_FAIL); pidx = OFF_TO_IDX(offset); if (pidx >= map->size || BIT_ISSET(map->size, pidx, map->err)) return (VM_PAGER_FAIL); page = PHYS_TO_VM_PAGE(map->phys_base_addr + offset); if (page == NULL) return (VM_PAGER_FAIL); KASSERT((page->flags & PG_FICTITIOUS) != 0, ("not fictitious %p", page)); KASSERT(vm_page_wired(page), ("page %p not wired", page)); KASSERT(!vm_page_busied(page), ("page %p is busy", page)); if (*mres != NULL) { oldm = *mres; vm_page_free(oldm); *mres = NULL; } + vm_page_busy_acquire(page, 0); vm_page_insert(page, object, pidx); page->valid = VM_PAGE_BITS_ALL; - vm_page_xbusy(page); *mres = page; return (VM_PAGER_OK); } /*----------------------- Privcmd char device methods ------------------------*/ static int privcmd_mmap_single(struct cdev *cdev, vm_ooffset_t *offset, vm_size_t size, vm_object_t *object, int nprot) { struct privcmd_map *map; map = malloc(sizeof(*map), M_PRIVCMD, M_WAITOK | M_ZERO); map->size = OFF_TO_IDX(size); map->pseudo_phys_res_id = 0; map->pseudo_phys_res = xenmem_alloc(privcmd_dev, &map->pseudo_phys_res_id, size); if (map->pseudo_phys_res == NULL) { free(map, M_PRIVCMD); return (ENOMEM); } map->phys_base_addr = rman_get_start(map->pseudo_phys_res); map->mem = cdev_pager_allocate(map, OBJT_MGTDEVICE, &privcmd_pg_ops, size, nprot, *offset, NULL); if (map->mem == NULL) { xenmem_free(privcmd_dev, map->pseudo_phys_res_id, map->pseudo_phys_res); free(map, M_PRIVCMD); return (ENOMEM); } *object = map->mem; return (0); } static int privcmd_ioctl(struct cdev *dev, unsigned long cmd, caddr_t arg, int mode, struct thread *td) { int error, i; switch (cmd) { case IOCTL_PRIVCMD_HYPERCALL: { struct ioctl_privcmd_hypercall *hcall; hcall = (struct ioctl_privcmd_hypercall *)arg; #ifdef __amd64__ /* * The hypervisor page table walker will refuse to access * user-space pages if SMAP is enabled, so temporary disable it * while performing the hypercall. */ if (cpu_stdext_feature & CPUID_STDEXT_SMAP) stac(); #endif error = privcmd_hypercall(hcall->op, hcall->arg[0], hcall->arg[1], hcall->arg[2], hcall->arg[3], hcall->arg[4]); #ifdef __amd64__ if (cpu_stdext_feature & CPUID_STDEXT_SMAP) clac(); #endif if (error >= 0) { hcall->retval = error; error = 0; } else { error = xen_translate_error(error); hcall->retval = 0; } break; } case IOCTL_PRIVCMD_MMAPBATCH: { struct ioctl_privcmd_mmapbatch *mmap; vm_map_t map; vm_map_entry_t entry; vm_object_t mem; vm_pindex_t pindex; vm_prot_t prot; boolean_t wired; struct xen_add_to_physmap_range add; xen_ulong_t *idxs; xen_pfn_t *gpfns; int *errs, index; struct privcmd_map *umap; uint16_t num; mmap = (struct ioctl_privcmd_mmapbatch *)arg; if ((mmap->num == 0) || ((mmap->addr & PAGE_MASK) != 0)) { error = EINVAL; break; } map = &td->td_proc->p_vmspace->vm_map; error = vm_map_lookup(&map, mmap->addr, VM_PROT_NONE, &entry, &mem, &pindex, &prot, &wired); if (error != KERN_SUCCESS) { error = EINVAL; break; } if ((entry->start != mmap->addr) || (entry->end != mmap->addr + (mmap->num * PAGE_SIZE))) { vm_map_lookup_done(map, entry); error = EINVAL; break; } vm_map_lookup_done(map, entry); if ((mem->type != OBJT_MGTDEVICE) || (mem->un_pager.devp.ops != &privcmd_pg_ops)) { error = EINVAL; break; } umap = mem->handle; add.domid = DOMID_SELF; add.space = XENMAPSPACE_gmfn_foreign; add.foreign_domid = mmap->dom; /* * The 'size' field in the xen_add_to_physmap_range only * allows for UINT16_MAX mappings in a single hypercall. */ num = MIN(mmap->num, UINT16_MAX); idxs = malloc(sizeof(*idxs) * num, M_PRIVCMD, M_WAITOK); gpfns = malloc(sizeof(*gpfns) * num, M_PRIVCMD, M_WAITOK); errs = malloc(sizeof(*errs) * num, M_PRIVCMD, M_WAITOK); set_xen_guest_handle(add.idxs, idxs); set_xen_guest_handle(add.gpfns, gpfns); set_xen_guest_handle(add.errs, errs); /* Allocate a bitset to store broken page mappings. */ umap->err = BITSET_ALLOC(mmap->num, M_PRIVCMD, M_WAITOK | M_ZERO); for (index = 0; index < mmap->num; index += num) { num = MIN(mmap->num - index, UINT16_MAX); add.size = num; error = copyin(&mmap->arr[index], idxs, sizeof(idxs[0]) * num); if (error != 0) goto mmap_out; for (i = 0; i < num; i++) gpfns[i] = atop(umap->phys_base_addr + (i + index) * PAGE_SIZE); bzero(errs, sizeof(*errs) * num); error = HYPERVISOR_memory_op( XENMEM_add_to_physmap_range, &add); if (error != 0) { error = xen_translate_error(error); goto mmap_out; } for (i = 0; i < num; i++) { if (errs[i] != 0) { errs[i] = xen_translate_error(errs[i]); /* Mark the page as invalid. */ BIT_SET(mmap->num, index + i, umap->err); } } error = copyout(errs, &mmap->err[index], sizeof(errs[0]) * num); if (error != 0) goto mmap_out; } umap->mapped = true; mmap_out: free(idxs, M_PRIVCMD); free(gpfns, M_PRIVCMD); free(errs, M_PRIVCMD); if (!umap->mapped) free(umap->err, M_PRIVCMD); break; } default: error = ENOSYS; break; } return (error); } /*------------------ Private Device Attachment Functions --------------------*/ static void privcmd_identify(driver_t *driver, device_t parent) { KASSERT(xen_domain(), ("Trying to attach privcmd device on non Xen domain")); if (BUS_ADD_CHILD(parent, 0, "privcmd", 0) == NULL) panic("unable to attach privcmd user-space device"); } static int privcmd_probe(device_t dev) { privcmd_dev = dev; device_set_desc(dev, "Xen privileged interface user-space device"); return (BUS_PROBE_NOWILDCARD); } static int privcmd_attach(device_t dev) { make_dev_credf(MAKEDEV_ETERNAL, &privcmd_devsw, 0, NULL, UID_ROOT, GID_WHEEL, 0600, "xen/privcmd"); return (0); } /*-------------------- Private Device Attachment Data -----------------------*/ static device_method_t privcmd_methods[] = { DEVMETHOD(device_identify, privcmd_identify), DEVMETHOD(device_probe, privcmd_probe), DEVMETHOD(device_attach, privcmd_attach), DEVMETHOD_END }; static driver_t privcmd_driver = { "privcmd", privcmd_methods, 0, }; devclass_t privcmd_devclass; DRIVER_MODULE(privcmd, xenpv, privcmd_driver, privcmd_devclass, 0, 0); MODULE_DEPEND(privcmd, xenpv, 1, 1, 1); Index: head/sys/fs/tmpfs/tmpfs_subr.c =================================================================== --- head/sys/fs/tmpfs/tmpfs_subr.c (revision 353534) +++ head/sys/fs/tmpfs/tmpfs_subr.c (revision 353535) @@ -1,1875 +1,1873 @@ /* $NetBSD: tmpfs_subr.c,v 1.35 2007/07/09 21:10:50 ad Exp $ */ /*- * SPDX-License-Identifier: BSD-2-Clause-NetBSD * * Copyright (c) 2005 The NetBSD Foundation, Inc. * All rights reserved. * * This code is derived from software contributed to The NetBSD Foundation * by Julio M. Merino Vidal, developed as part of Google's Summer of Code * 2005 program. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE * POSSIBILITY OF SUCH DAMAGE. */ /* * Efficient memory file system supporting functions. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include SYSCTL_NODE(_vfs, OID_AUTO, tmpfs, CTLFLAG_RW, 0, "tmpfs file system"); static long tmpfs_pages_reserved = TMPFS_PAGES_MINRESERVED; static int sysctl_mem_reserved(SYSCTL_HANDLER_ARGS) { int error; long pages, bytes; pages = *(long *)arg1; bytes = pages * PAGE_SIZE; error = sysctl_handle_long(oidp, &bytes, 0, req); if (error || !req->newptr) return (error); pages = bytes / PAGE_SIZE; if (pages < TMPFS_PAGES_MINRESERVED) return (EINVAL); *(long *)arg1 = pages; return (0); } SYSCTL_PROC(_vfs_tmpfs, OID_AUTO, memory_reserved, CTLTYPE_LONG|CTLFLAG_RW, &tmpfs_pages_reserved, 0, sysctl_mem_reserved, "L", "Amount of available memory and swap below which tmpfs growth stops"); static __inline int tmpfs_dirtree_cmp(struct tmpfs_dirent *a, struct tmpfs_dirent *b); RB_PROTOTYPE_STATIC(tmpfs_dir, tmpfs_dirent, uh.td_entries, tmpfs_dirtree_cmp); size_t tmpfs_mem_avail(void) { vm_ooffset_t avail; avail = swap_pager_avail + vm_free_count() - tmpfs_pages_reserved; if (__predict_false(avail < 0)) avail = 0; return (avail); } size_t tmpfs_pages_used(struct tmpfs_mount *tmp) { const size_t node_size = sizeof(struct tmpfs_node) + sizeof(struct tmpfs_dirent); size_t meta_pages; meta_pages = howmany((uintmax_t)tmp->tm_nodes_inuse * node_size, PAGE_SIZE); return (meta_pages + tmp->tm_pages_used); } static size_t tmpfs_pages_check_avail(struct tmpfs_mount *tmp, size_t req_pages) { if (tmpfs_mem_avail() < req_pages) return (0); if (tmp->tm_pages_max != ULONG_MAX && tmp->tm_pages_max < req_pages + tmpfs_pages_used(tmp)) return (0); return (1); } void tmpfs_ref_node(struct tmpfs_node *node) { TMPFS_NODE_LOCK(node); tmpfs_ref_node_locked(node); TMPFS_NODE_UNLOCK(node); } void tmpfs_ref_node_locked(struct tmpfs_node *node) { TMPFS_NODE_ASSERT_LOCKED(node); KASSERT(node->tn_refcount > 0, ("node %p zero refcount", node)); KASSERT(node->tn_refcount < UINT_MAX, ("node %p refcount %u", node, node->tn_refcount)); node->tn_refcount++; } /* * Allocates a new node of type 'type' inside the 'tmp' mount point, with * its owner set to 'uid', its group to 'gid' and its mode set to 'mode', * using the credentials of the process 'p'. * * If the node type is set to 'VDIR', then the parent parameter must point * to the parent directory of the node being created. It may only be NULL * while allocating the root node. * * If the node type is set to 'VBLK' or 'VCHR', then the rdev parameter * specifies the device the node represents. * * If the node type is set to 'VLNK', then the parameter target specifies * the file name of the target file for the symbolic link that is being * created. * * Note that new nodes are retrieved from the available list if it has * items or, if it is empty, from the node pool as long as there is enough * space to create them. * * Returns zero on success or an appropriate error code on failure. */ int tmpfs_alloc_node(struct mount *mp, struct tmpfs_mount *tmp, enum vtype type, uid_t uid, gid_t gid, mode_t mode, struct tmpfs_node *parent, const char *target, dev_t rdev, struct tmpfs_node **node) { struct tmpfs_node *nnode; vm_object_t obj; /* If the root directory of the 'tmp' file system is not yet * allocated, this must be the request to do it. */ MPASS(IMPLIES(tmp->tm_root == NULL, parent == NULL && type == VDIR)); MPASS(IFF(type == VLNK, target != NULL)); MPASS(IFF(type == VBLK || type == VCHR, rdev != VNOVAL)); if (tmp->tm_nodes_inuse >= tmp->tm_nodes_max) return (ENOSPC); if (tmpfs_pages_check_avail(tmp, 1) == 0) return (ENOSPC); if ((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) { /* * When a new tmpfs node is created for fully * constructed mount point, there must be a parent * node, which vnode is locked exclusively. As * consequence, if the unmount is executing in * parallel, vflush() cannot reclaim the parent vnode. * Due to this, the check for MNTK_UNMOUNT flag is not * racy: if we did not see MNTK_UNMOUNT flag, then tmp * cannot be destroyed until node construction is * finished and the parent vnode unlocked. * * Tmpfs does not need to instantiate new nodes during * unmount. */ return (EBUSY); } if ((mp->mnt_kern_flag & MNT_RDONLY) != 0) return (EROFS); nnode = (struct tmpfs_node *)uma_zalloc_arg(tmp->tm_node_pool, tmp, M_WAITOK); /* Generic initialization. */ nnode->tn_type = type; vfs_timestamp(&nnode->tn_atime); nnode->tn_birthtime = nnode->tn_ctime = nnode->tn_mtime = nnode->tn_atime; nnode->tn_uid = uid; nnode->tn_gid = gid; nnode->tn_mode = mode; nnode->tn_id = alloc_unr64(&tmp->tm_ino_unr); nnode->tn_refcount = 1; /* Type-specific initialization. */ switch (nnode->tn_type) { case VBLK: case VCHR: nnode->tn_rdev = rdev; break; case VDIR: RB_INIT(&nnode->tn_dir.tn_dirhead); LIST_INIT(&nnode->tn_dir.tn_dupindex); MPASS(parent != nnode); MPASS(IMPLIES(parent == NULL, tmp->tm_root == NULL)); nnode->tn_dir.tn_parent = (parent == NULL) ? nnode : parent; nnode->tn_dir.tn_readdir_lastn = 0; nnode->tn_dir.tn_readdir_lastp = NULL; nnode->tn_links++; TMPFS_NODE_LOCK(nnode->tn_dir.tn_parent); nnode->tn_dir.tn_parent->tn_links++; TMPFS_NODE_UNLOCK(nnode->tn_dir.tn_parent); break; case VFIFO: /* FALLTHROUGH */ case VSOCK: break; case VLNK: MPASS(strlen(target) < MAXPATHLEN); nnode->tn_size = strlen(target); nnode->tn_link = malloc(nnode->tn_size, M_TMPFSNAME, M_WAITOK); memcpy(nnode->tn_link, target, nnode->tn_size); break; case VREG: obj = nnode->tn_reg.tn_aobj = vm_pager_allocate(OBJT_SWAP, NULL, 0, VM_PROT_DEFAULT, 0, NULL /* XXXKIB - tmpfs needs swap reservation */); VM_OBJECT_WLOCK(obj); /* OBJ_TMPFS is set together with the setting of vp->v_object */ vm_object_set_flag(obj, OBJ_NOSPLIT | OBJ_TMPFS_NODE); vm_object_clear_flag(obj, OBJ_ONEMAPPING); VM_OBJECT_WUNLOCK(obj); break; default: panic("tmpfs_alloc_node: type %p %d", nnode, (int)nnode->tn_type); } TMPFS_LOCK(tmp); LIST_INSERT_HEAD(&tmp->tm_nodes_used, nnode, tn_entries); nnode->tn_attached = true; tmp->tm_nodes_inuse++; tmp->tm_refcount++; TMPFS_UNLOCK(tmp); *node = nnode; return (0); } /* * Destroys the node pointed to by node from the file system 'tmp'. * If the node references a directory, no entries are allowed. */ void tmpfs_free_node(struct tmpfs_mount *tmp, struct tmpfs_node *node) { TMPFS_LOCK(tmp); TMPFS_NODE_LOCK(node); if (!tmpfs_free_node_locked(tmp, node, false)) { TMPFS_NODE_UNLOCK(node); TMPFS_UNLOCK(tmp); } } bool tmpfs_free_node_locked(struct tmpfs_mount *tmp, struct tmpfs_node *node, bool detach) { vm_object_t uobj; TMPFS_MP_ASSERT_LOCKED(tmp); TMPFS_NODE_ASSERT_LOCKED(node); KASSERT(node->tn_refcount > 0, ("node %p refcount zero", node)); node->tn_refcount--; if (node->tn_attached && (detach || node->tn_refcount == 0)) { MPASS(tmp->tm_nodes_inuse > 0); tmp->tm_nodes_inuse--; LIST_REMOVE(node, tn_entries); node->tn_attached = false; } if (node->tn_refcount > 0) return (false); #ifdef INVARIANTS MPASS(node->tn_vnode == NULL); MPASS((node->tn_vpstate & TMPFS_VNODE_ALLOCATING) == 0); #endif TMPFS_NODE_UNLOCK(node); TMPFS_UNLOCK(tmp); switch (node->tn_type) { case VBLK: /* FALLTHROUGH */ case VCHR: /* FALLTHROUGH */ case VDIR: /* FALLTHROUGH */ case VFIFO: /* FALLTHROUGH */ case VSOCK: break; case VLNK: free(node->tn_link, M_TMPFSNAME); break; case VREG: uobj = node->tn_reg.tn_aobj; if (uobj != NULL) { if (uobj->size != 0) atomic_subtract_long(&tmp->tm_pages_used, uobj->size); KASSERT((uobj->flags & OBJ_TMPFS) == 0, ("leaked OBJ_TMPFS node %p vm_obj %p", node, uobj)); vm_object_deallocate(uobj); } break; default: panic("tmpfs_free_node: type %p %d", node, (int)node->tn_type); } uma_zfree(tmp->tm_node_pool, node); TMPFS_LOCK(tmp); tmpfs_free_tmp(tmp); return (true); } static __inline uint32_t tmpfs_dirent_hash(const char *name, u_int len) { uint32_t hash; hash = fnv_32_buf(name, len, FNV1_32_INIT + len) & TMPFS_DIRCOOKIE_MASK; #ifdef TMPFS_DEBUG_DIRCOOKIE_DUP hash &= 0xf; #endif if (hash < TMPFS_DIRCOOKIE_MIN) hash += TMPFS_DIRCOOKIE_MIN; return (hash); } static __inline off_t tmpfs_dirent_cookie(struct tmpfs_dirent *de) { if (de == NULL) return (TMPFS_DIRCOOKIE_EOF); MPASS(de->td_cookie >= TMPFS_DIRCOOKIE_MIN); return (de->td_cookie); } static __inline boolean_t tmpfs_dirent_dup(struct tmpfs_dirent *de) { return ((de->td_cookie & TMPFS_DIRCOOKIE_DUP) != 0); } static __inline boolean_t tmpfs_dirent_duphead(struct tmpfs_dirent *de) { return ((de->td_cookie & TMPFS_DIRCOOKIE_DUPHEAD) != 0); } void tmpfs_dirent_init(struct tmpfs_dirent *de, const char *name, u_int namelen) { de->td_hash = de->td_cookie = tmpfs_dirent_hash(name, namelen); memcpy(de->ud.td_name, name, namelen); de->td_namelen = namelen; } /* * Allocates a new directory entry for the node node with a name of name. * The new directory entry is returned in *de. * * The link count of node is increased by one to reflect the new object * referencing it. * * Returns zero on success or an appropriate error code on failure. */ int tmpfs_alloc_dirent(struct tmpfs_mount *tmp, struct tmpfs_node *node, const char *name, u_int len, struct tmpfs_dirent **de) { struct tmpfs_dirent *nde; nde = uma_zalloc(tmp->tm_dirent_pool, M_WAITOK); nde->td_node = node; if (name != NULL) { nde->ud.td_name = malloc(len, M_TMPFSNAME, M_WAITOK); tmpfs_dirent_init(nde, name, len); } else nde->td_namelen = 0; if (node != NULL) node->tn_links++; *de = nde; return 0; } /* * Frees a directory entry. It is the caller's responsibility to destroy * the node referenced by it if needed. * * The link count of node is decreased by one to reflect the removal of an * object that referenced it. This only happens if 'node_exists' is true; * otherwise the function will not access the node referred to by the * directory entry, as it may already have been released from the outside. */ void tmpfs_free_dirent(struct tmpfs_mount *tmp, struct tmpfs_dirent *de) { struct tmpfs_node *node; node = de->td_node; if (node != NULL) { MPASS(node->tn_links > 0); node->tn_links--; } if (!tmpfs_dirent_duphead(de) && de->ud.td_name != NULL) free(de->ud.td_name, M_TMPFSNAME); uma_zfree(tmp->tm_dirent_pool, de); } void tmpfs_destroy_vobject(struct vnode *vp, vm_object_t obj) { ASSERT_VOP_ELOCKED(vp, "tmpfs_destroy_vobject"); if (vp->v_type != VREG || obj == NULL) return; VM_OBJECT_WLOCK(obj); VI_LOCK(vp); vm_object_clear_flag(obj, OBJ_TMPFS); obj->un_pager.swp.swp_tmpfs = NULL; if (vp->v_writecount < 0) vp->v_writecount = 0; VI_UNLOCK(vp); VM_OBJECT_WUNLOCK(obj); } /* * Need to clear v_object for insmntque failure. */ static void tmpfs_insmntque_dtr(struct vnode *vp, void *dtr_arg) { tmpfs_destroy_vobject(vp, vp->v_object); vp->v_object = NULL; vp->v_data = NULL; vp->v_op = &dead_vnodeops; vgone(vp); vput(vp); } /* * Allocates a new vnode for the node node or returns a new reference to * an existing one if the node had already a vnode referencing it. The * resulting locked vnode is returned in *vpp. * * Returns zero on success or an appropriate error code on failure. */ int tmpfs_alloc_vp(struct mount *mp, struct tmpfs_node *node, int lkflag, struct vnode **vpp) { struct vnode *vp; struct tmpfs_mount *tm; vm_object_t object; int error; error = 0; tm = VFS_TO_TMPFS(mp); TMPFS_NODE_LOCK(node); tmpfs_ref_node_locked(node); loop: TMPFS_NODE_ASSERT_LOCKED(node); if ((vp = node->tn_vnode) != NULL) { MPASS((node->tn_vpstate & TMPFS_VNODE_DOOMED) == 0); VI_LOCK(vp); if ((node->tn_type == VDIR && node->tn_dir.tn_parent == NULL) || ((vp->v_iflag & VI_DOOMED) != 0 && (lkflag & LK_NOWAIT) != 0)) { VI_UNLOCK(vp); TMPFS_NODE_UNLOCK(node); error = ENOENT; vp = NULL; goto out; } if ((vp->v_iflag & VI_DOOMED) != 0) { VI_UNLOCK(vp); node->tn_vpstate |= TMPFS_VNODE_WRECLAIM; while ((node->tn_vpstate & TMPFS_VNODE_WRECLAIM) != 0) { msleep(&node->tn_vnode, TMPFS_NODE_MTX(node), 0, "tmpfsE", 0); } goto loop; } TMPFS_NODE_UNLOCK(node); error = vget(vp, lkflag | LK_INTERLOCK, curthread); if (error == ENOENT) { TMPFS_NODE_LOCK(node); goto loop; } if (error != 0) { vp = NULL; goto out; } /* * Make sure the vnode is still there after * getting the interlock to avoid racing a free. */ if (node->tn_vnode == NULL || node->tn_vnode != vp) { vput(vp); TMPFS_NODE_LOCK(node); goto loop; } goto out; } if ((node->tn_vpstate & TMPFS_VNODE_DOOMED) || (node->tn_type == VDIR && node->tn_dir.tn_parent == NULL)) { TMPFS_NODE_UNLOCK(node); error = ENOENT; vp = NULL; goto out; } /* * otherwise lock the vp list while we call getnewvnode * since that can block. */ if (node->tn_vpstate & TMPFS_VNODE_ALLOCATING) { node->tn_vpstate |= TMPFS_VNODE_WANT; error = msleep((caddr_t) &node->tn_vpstate, TMPFS_NODE_MTX(node), 0, "tmpfs_alloc_vp", 0); if (error != 0) goto out; goto loop; } else node->tn_vpstate |= TMPFS_VNODE_ALLOCATING; TMPFS_NODE_UNLOCK(node); /* Get a new vnode and associate it with our node. */ error = getnewvnode("tmpfs", mp, VFS_TO_TMPFS(mp)->tm_nonc ? &tmpfs_vnodeop_nonc_entries : &tmpfs_vnodeop_entries, &vp); if (error != 0) goto unlock; MPASS(vp != NULL); /* lkflag is ignored, the lock is exclusive */ (void) vn_lock(vp, lkflag | LK_RETRY); vp->v_data = node; vp->v_type = node->tn_type; /* Type-specific initialization. */ switch (node->tn_type) { case VBLK: /* FALLTHROUGH */ case VCHR: /* FALLTHROUGH */ case VLNK: /* FALLTHROUGH */ case VSOCK: break; case VFIFO: vp->v_op = &tmpfs_fifoop_entries; break; case VREG: object = node->tn_reg.tn_aobj; VM_OBJECT_WLOCK(object); VI_LOCK(vp); KASSERT(vp->v_object == NULL, ("Not NULL v_object in tmpfs")); vp->v_object = object; object->un_pager.swp.swp_tmpfs = vp; vm_object_set_flag(object, OBJ_TMPFS); VI_UNLOCK(vp); VM_OBJECT_WUNLOCK(object); break; case VDIR: MPASS(node->tn_dir.tn_parent != NULL); if (node->tn_dir.tn_parent == node) vp->v_vflag |= VV_ROOT; break; default: panic("tmpfs_alloc_vp: type %p %d", node, (int)node->tn_type); } if (vp->v_type != VFIFO) VN_LOCK_ASHARE(vp); error = insmntque1(vp, mp, tmpfs_insmntque_dtr, NULL); if (error != 0) vp = NULL; unlock: TMPFS_NODE_LOCK(node); MPASS(node->tn_vpstate & TMPFS_VNODE_ALLOCATING); node->tn_vpstate &= ~TMPFS_VNODE_ALLOCATING; node->tn_vnode = vp; if (node->tn_vpstate & TMPFS_VNODE_WANT) { node->tn_vpstate &= ~TMPFS_VNODE_WANT; TMPFS_NODE_UNLOCK(node); wakeup((caddr_t) &node->tn_vpstate); } else TMPFS_NODE_UNLOCK(node); out: if (error == 0) { *vpp = vp; #ifdef INVARIANTS MPASS(*vpp != NULL && VOP_ISLOCKED(*vpp)); TMPFS_NODE_LOCK(node); MPASS(*vpp == node->tn_vnode); TMPFS_NODE_UNLOCK(node); #endif } tmpfs_free_node(tm, node); return (error); } /* * Destroys the association between the vnode vp and the node it * references. */ void tmpfs_free_vp(struct vnode *vp) { struct tmpfs_node *node; node = VP_TO_TMPFS_NODE(vp); TMPFS_NODE_ASSERT_LOCKED(node); node->tn_vnode = NULL; if ((node->tn_vpstate & TMPFS_VNODE_WRECLAIM) != 0) wakeup(&node->tn_vnode); node->tn_vpstate &= ~TMPFS_VNODE_WRECLAIM; vp->v_data = NULL; } /* * Allocates a new file of type 'type' and adds it to the parent directory * 'dvp'; this addition is done using the component name given in 'cnp'. * The ownership of the new file is automatically assigned based on the * credentials of the caller (through 'cnp'), the group is set based on * the parent directory and the mode is determined from the 'vap' argument. * If successful, *vpp holds a vnode to the newly created file and zero * is returned. Otherwise *vpp is NULL and the function returns an * appropriate error code. */ int tmpfs_alloc_file(struct vnode *dvp, struct vnode **vpp, struct vattr *vap, struct componentname *cnp, const char *target) { int error; struct tmpfs_dirent *de; struct tmpfs_mount *tmp; struct tmpfs_node *dnode; struct tmpfs_node *node; struct tmpfs_node *parent; ASSERT_VOP_ELOCKED(dvp, "tmpfs_alloc_file"); MPASS(cnp->cn_flags & HASBUF); tmp = VFS_TO_TMPFS(dvp->v_mount); dnode = VP_TO_TMPFS_DIR(dvp); *vpp = NULL; /* If the entry we are creating is a directory, we cannot overflow * the number of links of its parent, because it will get a new * link. */ if (vap->va_type == VDIR) { /* Ensure that we do not overflow the maximum number of links * imposed by the system. */ MPASS(dnode->tn_links <= TMPFS_LINK_MAX); if (dnode->tn_links == TMPFS_LINK_MAX) { return (EMLINK); } parent = dnode; MPASS(parent != NULL); } else parent = NULL; /* Allocate a node that represents the new file. */ error = tmpfs_alloc_node(dvp->v_mount, tmp, vap->va_type, cnp->cn_cred->cr_uid, dnode->tn_gid, vap->va_mode, parent, target, vap->va_rdev, &node); if (error != 0) return (error); /* Allocate a directory entry that points to the new file. */ error = tmpfs_alloc_dirent(tmp, node, cnp->cn_nameptr, cnp->cn_namelen, &de); if (error != 0) { tmpfs_free_node(tmp, node); return (error); } /* Allocate a vnode for the new file. */ error = tmpfs_alloc_vp(dvp->v_mount, node, LK_EXCLUSIVE, vpp); if (error != 0) { tmpfs_free_dirent(tmp, de); tmpfs_free_node(tmp, node); return (error); } /* Now that all required items are allocated, we can proceed to * insert the new node into the directory, an operation that * cannot fail. */ if (cnp->cn_flags & ISWHITEOUT) tmpfs_dir_whiteout_remove(dvp, cnp); tmpfs_dir_attach(dvp, de); return (0); } struct tmpfs_dirent * tmpfs_dir_first(struct tmpfs_node *dnode, struct tmpfs_dir_cursor *dc) { struct tmpfs_dirent *de; de = RB_MIN(tmpfs_dir, &dnode->tn_dir.tn_dirhead); dc->tdc_tree = de; if (de != NULL && tmpfs_dirent_duphead(de)) de = LIST_FIRST(&de->ud.td_duphead); dc->tdc_current = de; return (dc->tdc_current); } struct tmpfs_dirent * tmpfs_dir_next(struct tmpfs_node *dnode, struct tmpfs_dir_cursor *dc) { struct tmpfs_dirent *de; MPASS(dc->tdc_tree != NULL); if (tmpfs_dirent_dup(dc->tdc_current)) { dc->tdc_current = LIST_NEXT(dc->tdc_current, uh.td_dup.entries); if (dc->tdc_current != NULL) return (dc->tdc_current); } dc->tdc_tree = dc->tdc_current = RB_NEXT(tmpfs_dir, &dnode->tn_dir.tn_dirhead, dc->tdc_tree); if ((de = dc->tdc_current) != NULL && tmpfs_dirent_duphead(de)) { dc->tdc_current = LIST_FIRST(&de->ud.td_duphead); MPASS(dc->tdc_current != NULL); } return (dc->tdc_current); } /* Lookup directory entry in RB-Tree. Function may return duphead entry. */ static struct tmpfs_dirent * tmpfs_dir_xlookup_hash(struct tmpfs_node *dnode, uint32_t hash) { struct tmpfs_dirent *de, dekey; dekey.td_hash = hash; de = RB_FIND(tmpfs_dir, &dnode->tn_dir.tn_dirhead, &dekey); return (de); } /* Lookup directory entry by cookie, initialize directory cursor accordingly. */ static struct tmpfs_dirent * tmpfs_dir_lookup_cookie(struct tmpfs_node *node, off_t cookie, struct tmpfs_dir_cursor *dc) { struct tmpfs_dir *dirhead = &node->tn_dir.tn_dirhead; struct tmpfs_dirent *de, dekey; MPASS(cookie >= TMPFS_DIRCOOKIE_MIN); if (cookie == node->tn_dir.tn_readdir_lastn && (de = node->tn_dir.tn_readdir_lastp) != NULL) { /* Protect against possible race, tn_readdir_last[pn] * may be updated with only shared vnode lock held. */ if (cookie == tmpfs_dirent_cookie(de)) goto out; } if ((cookie & TMPFS_DIRCOOKIE_DUP) != 0) { LIST_FOREACH(de, &node->tn_dir.tn_dupindex, uh.td_dup.index_entries) { MPASS(tmpfs_dirent_dup(de)); if (de->td_cookie == cookie) goto out; /* dupindex list is sorted. */ if (de->td_cookie < cookie) { de = NULL; goto out; } } MPASS(de == NULL); goto out; } if ((cookie & TMPFS_DIRCOOKIE_MASK) != cookie) { de = NULL; } else { dekey.td_hash = cookie; /* Recover if direntry for cookie was removed */ de = RB_NFIND(tmpfs_dir, dirhead, &dekey); } dc->tdc_tree = de; dc->tdc_current = de; if (de != NULL && tmpfs_dirent_duphead(de)) { dc->tdc_current = LIST_FIRST(&de->ud.td_duphead); MPASS(dc->tdc_current != NULL); } return (dc->tdc_current); out: dc->tdc_tree = de; dc->tdc_current = de; if (de != NULL && tmpfs_dirent_dup(de)) dc->tdc_tree = tmpfs_dir_xlookup_hash(node, de->td_hash); return (dc->tdc_current); } /* * Looks for a directory entry in the directory represented by node. * 'cnp' describes the name of the entry to look for. Note that the . * and .. components are not allowed as they do not physically exist * within directories. * * Returns a pointer to the entry when found, otherwise NULL. */ struct tmpfs_dirent * tmpfs_dir_lookup(struct tmpfs_node *node, struct tmpfs_node *f, struct componentname *cnp) { struct tmpfs_dir_duphead *duphead; struct tmpfs_dirent *de; uint32_t hash; MPASS(IMPLIES(cnp->cn_namelen == 1, cnp->cn_nameptr[0] != '.')); MPASS(IMPLIES(cnp->cn_namelen == 2, !(cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.'))); TMPFS_VALIDATE_DIR(node); hash = tmpfs_dirent_hash(cnp->cn_nameptr, cnp->cn_namelen); de = tmpfs_dir_xlookup_hash(node, hash); if (de != NULL && tmpfs_dirent_duphead(de)) { duphead = &de->ud.td_duphead; LIST_FOREACH(de, duphead, uh.td_dup.entries) { if (TMPFS_DIRENT_MATCHES(de, cnp->cn_nameptr, cnp->cn_namelen)) break; } } else if (de != NULL) { if (!TMPFS_DIRENT_MATCHES(de, cnp->cn_nameptr, cnp->cn_namelen)) de = NULL; } if (de != NULL && f != NULL && de->td_node != f) de = NULL; return (de); } /* * Attach duplicate-cookie directory entry nde to dnode and insert to dupindex * list, allocate new cookie value. */ static void tmpfs_dir_attach_dup(struct tmpfs_node *dnode, struct tmpfs_dir_duphead *duphead, struct tmpfs_dirent *nde) { struct tmpfs_dir_duphead *dupindex; struct tmpfs_dirent *de, *pde; dupindex = &dnode->tn_dir.tn_dupindex; de = LIST_FIRST(dupindex); if (de == NULL || de->td_cookie < TMPFS_DIRCOOKIE_DUP_MAX) { if (de == NULL) nde->td_cookie = TMPFS_DIRCOOKIE_DUP_MIN; else nde->td_cookie = de->td_cookie + 1; MPASS(tmpfs_dirent_dup(nde)); LIST_INSERT_HEAD(dupindex, nde, uh.td_dup.index_entries); LIST_INSERT_HEAD(duphead, nde, uh.td_dup.entries); return; } /* * Cookie numbers are near exhaustion. Scan dupindex list for unused * numbers. dupindex list is sorted in descending order. Keep it so * after inserting nde. */ while (1) { pde = de; de = LIST_NEXT(de, uh.td_dup.index_entries); if (de == NULL && pde->td_cookie != TMPFS_DIRCOOKIE_DUP_MIN) { /* * Last element of the index doesn't have minimal cookie * value, use it. */ nde->td_cookie = TMPFS_DIRCOOKIE_DUP_MIN; LIST_INSERT_AFTER(pde, nde, uh.td_dup.index_entries); LIST_INSERT_HEAD(duphead, nde, uh.td_dup.entries); return; } else if (de == NULL) { /* * We are so lucky have 2^30 hash duplicates in single * directory :) Return largest possible cookie value. * It should be fine except possible issues with * VOP_READDIR restart. */ nde->td_cookie = TMPFS_DIRCOOKIE_DUP_MAX; LIST_INSERT_HEAD(dupindex, nde, uh.td_dup.index_entries); LIST_INSERT_HEAD(duphead, nde, uh.td_dup.entries); return; } if (de->td_cookie + 1 == pde->td_cookie || de->td_cookie >= TMPFS_DIRCOOKIE_DUP_MAX) continue; /* No hole or invalid cookie. */ nde->td_cookie = de->td_cookie + 1; MPASS(tmpfs_dirent_dup(nde)); MPASS(pde->td_cookie > nde->td_cookie); MPASS(nde->td_cookie > de->td_cookie); LIST_INSERT_BEFORE(de, nde, uh.td_dup.index_entries); LIST_INSERT_HEAD(duphead, nde, uh.td_dup.entries); return; } } /* * Attaches the directory entry de to the directory represented by vp. * Note that this does not change the link count of the node pointed by * the directory entry, as this is done by tmpfs_alloc_dirent. */ void tmpfs_dir_attach(struct vnode *vp, struct tmpfs_dirent *de) { struct tmpfs_node *dnode; struct tmpfs_dirent *xde, *nde; ASSERT_VOP_ELOCKED(vp, __func__); MPASS(de->td_namelen > 0); MPASS(de->td_hash >= TMPFS_DIRCOOKIE_MIN); MPASS(de->td_cookie == de->td_hash); dnode = VP_TO_TMPFS_DIR(vp); dnode->tn_dir.tn_readdir_lastn = 0; dnode->tn_dir.tn_readdir_lastp = NULL; MPASS(!tmpfs_dirent_dup(de)); xde = RB_INSERT(tmpfs_dir, &dnode->tn_dir.tn_dirhead, de); if (xde != NULL && tmpfs_dirent_duphead(xde)) tmpfs_dir_attach_dup(dnode, &xde->ud.td_duphead, de); else if (xde != NULL) { /* * Allocate new duphead. Swap xde with duphead to avoid * adding/removing elements with the same hash. */ MPASS(!tmpfs_dirent_dup(xde)); tmpfs_alloc_dirent(VFS_TO_TMPFS(vp->v_mount), NULL, NULL, 0, &nde); /* *nde = *xde; XXX gcc 4.2.1 may generate invalid code. */ memcpy(nde, xde, sizeof(*xde)); xde->td_cookie |= TMPFS_DIRCOOKIE_DUPHEAD; LIST_INIT(&xde->ud.td_duphead); xde->td_namelen = 0; xde->td_node = NULL; tmpfs_dir_attach_dup(dnode, &xde->ud.td_duphead, nde); tmpfs_dir_attach_dup(dnode, &xde->ud.td_duphead, de); } dnode->tn_size += sizeof(struct tmpfs_dirent); dnode->tn_status |= TMPFS_NODE_ACCESSED | TMPFS_NODE_CHANGED | \ TMPFS_NODE_MODIFIED; tmpfs_update(vp); } /* * Detaches the directory entry de from the directory represented by vp. * Note that this does not change the link count of the node pointed by * the directory entry, as this is done by tmpfs_free_dirent. */ void tmpfs_dir_detach(struct vnode *vp, struct tmpfs_dirent *de) { struct tmpfs_mount *tmp; struct tmpfs_dir *head; struct tmpfs_node *dnode; struct tmpfs_dirent *xde; ASSERT_VOP_ELOCKED(vp, __func__); dnode = VP_TO_TMPFS_DIR(vp); head = &dnode->tn_dir.tn_dirhead; dnode->tn_dir.tn_readdir_lastn = 0; dnode->tn_dir.tn_readdir_lastp = NULL; if (tmpfs_dirent_dup(de)) { /* Remove duphead if de was last entry. */ if (LIST_NEXT(de, uh.td_dup.entries) == NULL) { xde = tmpfs_dir_xlookup_hash(dnode, de->td_hash); MPASS(tmpfs_dirent_duphead(xde)); } else xde = NULL; LIST_REMOVE(de, uh.td_dup.entries); LIST_REMOVE(de, uh.td_dup.index_entries); if (xde != NULL) { if (LIST_EMPTY(&xde->ud.td_duphead)) { RB_REMOVE(tmpfs_dir, head, xde); tmp = VFS_TO_TMPFS(vp->v_mount); MPASS(xde->td_node == NULL); tmpfs_free_dirent(tmp, xde); } } de->td_cookie = de->td_hash; } else RB_REMOVE(tmpfs_dir, head, de); dnode->tn_size -= sizeof(struct tmpfs_dirent); dnode->tn_status |= TMPFS_NODE_ACCESSED | TMPFS_NODE_CHANGED | \ TMPFS_NODE_MODIFIED; tmpfs_update(vp); } void tmpfs_dir_destroy(struct tmpfs_mount *tmp, struct tmpfs_node *dnode) { struct tmpfs_dirent *de, *dde, *nde; RB_FOREACH_SAFE(de, tmpfs_dir, &dnode->tn_dir.tn_dirhead, nde) { RB_REMOVE(tmpfs_dir, &dnode->tn_dir.tn_dirhead, de); /* Node may already be destroyed. */ de->td_node = NULL; if (tmpfs_dirent_duphead(de)) { while ((dde = LIST_FIRST(&de->ud.td_duphead)) != NULL) { LIST_REMOVE(dde, uh.td_dup.entries); dde->td_node = NULL; tmpfs_free_dirent(tmp, dde); } } tmpfs_free_dirent(tmp, de); } } /* * Helper function for tmpfs_readdir. Creates a '.' entry for the given * directory and returns it in the uio space. The function returns 0 * on success, -1 if there was not enough space in the uio structure to * hold the directory entry or an appropriate error code if another * error happens. */ static int tmpfs_dir_getdotdent(struct tmpfs_mount *tm, struct tmpfs_node *node, struct uio *uio) { int error; struct dirent dent; TMPFS_VALIDATE_DIR(node); MPASS(uio->uio_offset == TMPFS_DIRCOOKIE_DOT); dent.d_fileno = node->tn_id; dent.d_type = DT_DIR; dent.d_namlen = 1; dent.d_name[0] = '.'; dent.d_reclen = GENERIC_DIRSIZ(&dent); dirent_terminate(&dent); if (dent.d_reclen > uio->uio_resid) error = EJUSTRETURN; else error = uiomove(&dent, dent.d_reclen, uio); tmpfs_set_status(tm, node, TMPFS_NODE_ACCESSED); return (error); } /* * Helper function for tmpfs_readdir. Creates a '..' entry for the given * directory and returns it in the uio space. The function returns 0 * on success, -1 if there was not enough space in the uio structure to * hold the directory entry or an appropriate error code if another * error happens. */ static int tmpfs_dir_getdotdotdent(struct tmpfs_mount *tm, struct tmpfs_node *node, struct uio *uio) { struct tmpfs_node *parent; struct dirent dent; int error; TMPFS_VALIDATE_DIR(node); MPASS(uio->uio_offset == TMPFS_DIRCOOKIE_DOTDOT); /* * Return ENOENT if the current node is already removed. */ TMPFS_ASSERT_LOCKED(node); parent = node->tn_dir.tn_parent; if (parent == NULL) return (ENOENT); TMPFS_NODE_LOCK(parent); dent.d_fileno = parent->tn_id; TMPFS_NODE_UNLOCK(parent); dent.d_type = DT_DIR; dent.d_namlen = 2; dent.d_name[0] = '.'; dent.d_name[1] = '.'; dent.d_reclen = GENERIC_DIRSIZ(&dent); dirent_terminate(&dent); if (dent.d_reclen > uio->uio_resid) error = EJUSTRETURN; else error = uiomove(&dent, dent.d_reclen, uio); tmpfs_set_status(tm, node, TMPFS_NODE_ACCESSED); return (error); } /* * Helper function for tmpfs_readdir. Returns as much directory entries * as can fit in the uio space. The read starts at uio->uio_offset. * The function returns 0 on success, -1 if there was not enough space * in the uio structure to hold the directory entry or an appropriate * error code if another error happens. */ int tmpfs_dir_getdents(struct tmpfs_mount *tm, struct tmpfs_node *node, struct uio *uio, int maxcookies, u_long *cookies, int *ncookies) { struct tmpfs_dir_cursor dc; struct tmpfs_dirent *de; off_t off; int error; TMPFS_VALIDATE_DIR(node); off = 0; /* * Lookup the node from the current offset. The starting offset of * 0 will lookup both '.' and '..', and then the first real entry, * or EOF if there are none. Then find all entries for the dir that * fit into the buffer. Once no more entries are found (de == NULL), * the offset is set to TMPFS_DIRCOOKIE_EOF, which will cause the next * call to return 0. */ switch (uio->uio_offset) { case TMPFS_DIRCOOKIE_DOT: error = tmpfs_dir_getdotdent(tm, node, uio); if (error != 0) return (error); uio->uio_offset = TMPFS_DIRCOOKIE_DOTDOT; if (cookies != NULL) cookies[(*ncookies)++] = off = uio->uio_offset; /* FALLTHROUGH */ case TMPFS_DIRCOOKIE_DOTDOT: error = tmpfs_dir_getdotdotdent(tm, node, uio); if (error != 0) return (error); de = tmpfs_dir_first(node, &dc); uio->uio_offset = tmpfs_dirent_cookie(de); if (cookies != NULL) cookies[(*ncookies)++] = off = uio->uio_offset; /* EOF. */ if (de == NULL) return (0); break; case TMPFS_DIRCOOKIE_EOF: return (0); default: de = tmpfs_dir_lookup_cookie(node, uio->uio_offset, &dc); if (de == NULL) return (EINVAL); if (cookies != NULL) off = tmpfs_dirent_cookie(de); } /* Read as much entries as possible; i.e., until we reach the end of * the directory or we exhaust uio space. */ do { struct dirent d; /* Create a dirent structure representing the current * tmpfs_node and fill it. */ if (de->td_node == NULL) { d.d_fileno = 1; d.d_type = DT_WHT; } else { d.d_fileno = de->td_node->tn_id; switch (de->td_node->tn_type) { case VBLK: d.d_type = DT_BLK; break; case VCHR: d.d_type = DT_CHR; break; case VDIR: d.d_type = DT_DIR; break; case VFIFO: d.d_type = DT_FIFO; break; case VLNK: d.d_type = DT_LNK; break; case VREG: d.d_type = DT_REG; break; case VSOCK: d.d_type = DT_SOCK; break; default: panic("tmpfs_dir_getdents: type %p %d", de->td_node, (int)de->td_node->tn_type); } } d.d_namlen = de->td_namelen; MPASS(de->td_namelen < sizeof(d.d_name)); (void)memcpy(d.d_name, de->ud.td_name, de->td_namelen); d.d_reclen = GENERIC_DIRSIZ(&d); dirent_terminate(&d); /* Stop reading if the directory entry we are treating is * bigger than the amount of data that can be returned. */ if (d.d_reclen > uio->uio_resid) { error = EJUSTRETURN; break; } /* Copy the new dirent structure into the output buffer and * advance pointers. */ error = uiomove(&d, d.d_reclen, uio); if (error == 0) { de = tmpfs_dir_next(node, &dc); if (cookies != NULL) { off = tmpfs_dirent_cookie(de); MPASS(*ncookies < maxcookies); cookies[(*ncookies)++] = off; } } } while (error == 0 && uio->uio_resid > 0 && de != NULL); /* Skip setting off when using cookies as it is already done above. */ if (cookies == NULL) off = tmpfs_dirent_cookie(de); /* Update the offset and cache. */ uio->uio_offset = off; node->tn_dir.tn_readdir_lastn = off; node->tn_dir.tn_readdir_lastp = de; tmpfs_set_status(tm, node, TMPFS_NODE_ACCESSED); return error; } int tmpfs_dir_whiteout_add(struct vnode *dvp, struct componentname *cnp) { struct tmpfs_dirent *de; int error; error = tmpfs_alloc_dirent(VFS_TO_TMPFS(dvp->v_mount), NULL, cnp->cn_nameptr, cnp->cn_namelen, &de); if (error != 0) return (error); tmpfs_dir_attach(dvp, de); return (0); } void tmpfs_dir_whiteout_remove(struct vnode *dvp, struct componentname *cnp) { struct tmpfs_dirent *de; de = tmpfs_dir_lookup(VP_TO_TMPFS_DIR(dvp), NULL, cnp); MPASS(de != NULL && de->td_node == NULL); tmpfs_dir_detach(dvp, de); tmpfs_free_dirent(VFS_TO_TMPFS(dvp->v_mount), de); } /* * Resizes the aobj associated with the regular file pointed to by 'vp' to the * size 'newsize'. 'vp' must point to a vnode that represents a regular file. * 'newsize' must be positive. * * Returns zero on success or an appropriate error code on failure. */ int tmpfs_reg_resize(struct vnode *vp, off_t newsize, boolean_t ignerr) { struct tmpfs_mount *tmp; struct tmpfs_node *node; vm_object_t uobj; vm_page_t m; vm_pindex_t idx, newpages, oldpages; off_t oldsize; int base, rv; MPASS(vp->v_type == VREG); MPASS(newsize >= 0); node = VP_TO_TMPFS_NODE(vp); uobj = node->tn_reg.tn_aobj; tmp = VFS_TO_TMPFS(vp->v_mount); /* * Convert the old and new sizes to the number of pages needed to * store them. It may happen that we do not need to do anything * because the last allocated page can accommodate the change on * its own. */ oldsize = node->tn_size; oldpages = OFF_TO_IDX(oldsize + PAGE_MASK); MPASS(oldpages == uobj->size); newpages = OFF_TO_IDX(newsize + PAGE_MASK); if (__predict_true(newpages == oldpages && newsize >= oldsize)) { node->tn_size = newsize; return (0); } if (newpages > oldpages && tmpfs_pages_check_avail(tmp, newpages - oldpages) == 0) return (ENOSPC); VM_OBJECT_WLOCK(uobj); if (newsize < oldsize) { /* * Zero the truncated part of the last page. */ base = newsize & PAGE_MASK; if (base != 0) { idx = OFF_TO_IDX(newsize); retry: - m = vm_page_lookup(uobj, idx); + m = vm_page_grab(uobj, idx, VM_ALLOC_NOCREAT); if (m != NULL) { - if (vm_page_sleep_if_busy(m, "tmfssz")) - goto retry; MPASS(m->valid == VM_PAGE_BITS_ALL); } else if (vm_pager_has_page(uobj, idx, NULL, NULL)) { m = vm_page_alloc(uobj, idx, VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL); if (m == NULL) goto retry; rv = vm_pager_get_pages(uobj, &m, 1, NULL, NULL); if (rv == VM_PAGER_OK) { /* * Since the page was not resident, * and therefore not recently * accessed, immediately enqueue it * for asynchronous laundering. The * current operation is not regarded * as an access. */ vm_page_lock(m); vm_page_launder(m); vm_page_unlock(m); - vm_page_xunbusy(m); } else { vm_page_free(m); if (ignerr) m = NULL; else { VM_OBJECT_WUNLOCK(uobj); return (EIO); } } } if (m != NULL) { pmap_zero_page_area(m, base, PAGE_SIZE - base); vm_page_dirty(m); + vm_page_xunbusy(m); vm_pager_page_unswapped(m); } } /* * Release any swap space and free any whole pages. */ if (newpages < oldpages) { swap_pager_freespace(uobj, newpages, oldpages - newpages); vm_object_page_remove(uobj, newpages, 0, 0); } } uobj->size = newpages; VM_OBJECT_WUNLOCK(uobj); atomic_add_long(&tmp->tm_pages_used, newpages - oldpages); node->tn_size = newsize; return (0); } void tmpfs_check_mtime(struct vnode *vp) { struct tmpfs_node *node; struct vm_object *obj; ASSERT_VOP_ELOCKED(vp, "check_mtime"); if (vp->v_type != VREG) return; obj = vp->v_object; KASSERT((obj->flags & (OBJ_TMPFS_NODE | OBJ_TMPFS)) == (OBJ_TMPFS_NODE | OBJ_TMPFS), ("non-tmpfs obj")); /* unlocked read */ if ((obj->flags & OBJ_TMPFS_DIRTY) != 0) { VM_OBJECT_WLOCK(obj); if ((obj->flags & OBJ_TMPFS_DIRTY) != 0) { obj->flags &= ~OBJ_TMPFS_DIRTY; node = VP_TO_TMPFS_NODE(vp); node->tn_status |= TMPFS_NODE_MODIFIED | TMPFS_NODE_CHANGED; } VM_OBJECT_WUNLOCK(obj); } } /* * Change flags of the given vnode. * Caller should execute tmpfs_update on vp after a successful execution. * The vnode must be locked on entry and remain locked on exit. */ int tmpfs_chflags(struct vnode *vp, u_long flags, struct ucred *cred, struct thread *p) { int error; struct tmpfs_node *node; ASSERT_VOP_ELOCKED(vp, "chflags"); node = VP_TO_TMPFS_NODE(vp); if ((flags & ~(SF_APPEND | SF_ARCHIVED | SF_IMMUTABLE | SF_NOUNLINK | UF_APPEND | UF_ARCHIVE | UF_HIDDEN | UF_IMMUTABLE | UF_NODUMP | UF_NOUNLINK | UF_OFFLINE | UF_OPAQUE | UF_READONLY | UF_REPARSE | UF_SPARSE | UF_SYSTEM)) != 0) return (EOPNOTSUPP); /* Disallow this operation if the file system is mounted read-only. */ if (vp->v_mount->mnt_flag & MNT_RDONLY) return EROFS; /* * Callers may only modify the file flags on objects they * have VADMIN rights for. */ if ((error = VOP_ACCESS(vp, VADMIN, cred, p))) return (error); /* * Unprivileged processes are not permitted to unset system * flags, or modify flags if any system flags are set. */ if (!priv_check_cred(cred, PRIV_VFS_SYSFLAGS)) { if (node->tn_flags & (SF_NOUNLINK | SF_IMMUTABLE | SF_APPEND)) { error = securelevel_gt(cred, 0); if (error) return (error); } } else { if (node->tn_flags & (SF_NOUNLINK | SF_IMMUTABLE | SF_APPEND) || ((flags ^ node->tn_flags) & SF_SETTABLE)) return (EPERM); } node->tn_flags = flags; node->tn_status |= TMPFS_NODE_CHANGED; ASSERT_VOP_ELOCKED(vp, "chflags2"); return (0); } /* * Change access mode on the given vnode. * Caller should execute tmpfs_update on vp after a successful execution. * The vnode must be locked on entry and remain locked on exit. */ int tmpfs_chmod(struct vnode *vp, mode_t mode, struct ucred *cred, struct thread *p) { int error; struct tmpfs_node *node; ASSERT_VOP_ELOCKED(vp, "chmod"); node = VP_TO_TMPFS_NODE(vp); /* Disallow this operation if the file system is mounted read-only. */ if (vp->v_mount->mnt_flag & MNT_RDONLY) return EROFS; /* Immutable or append-only files cannot be modified, either. */ if (node->tn_flags & (IMMUTABLE | APPEND)) return EPERM; /* * To modify the permissions on a file, must possess VADMIN * for that file. */ if ((error = VOP_ACCESS(vp, VADMIN, cred, p))) return (error); /* * Privileged processes may set the sticky bit on non-directories, * as well as set the setgid bit on a file with a group that the * process is not a member of. */ if (vp->v_type != VDIR && (mode & S_ISTXT)) { if (priv_check_cred(cred, PRIV_VFS_STICKYFILE)) return (EFTYPE); } if (!groupmember(node->tn_gid, cred) && (mode & S_ISGID)) { error = priv_check_cred(cred, PRIV_VFS_SETGID); if (error) return (error); } node->tn_mode &= ~ALLPERMS; node->tn_mode |= mode & ALLPERMS; node->tn_status |= TMPFS_NODE_CHANGED; ASSERT_VOP_ELOCKED(vp, "chmod2"); return (0); } /* * Change ownership of the given vnode. At least one of uid or gid must * be different than VNOVAL. If one is set to that value, the attribute * is unchanged. * Caller should execute tmpfs_update on vp after a successful execution. * The vnode must be locked on entry and remain locked on exit. */ int tmpfs_chown(struct vnode *vp, uid_t uid, gid_t gid, struct ucred *cred, struct thread *p) { int error; struct tmpfs_node *node; uid_t ouid; gid_t ogid; ASSERT_VOP_ELOCKED(vp, "chown"); node = VP_TO_TMPFS_NODE(vp); /* Assign default values if they are unknown. */ MPASS(uid != VNOVAL || gid != VNOVAL); if (uid == VNOVAL) uid = node->tn_uid; if (gid == VNOVAL) gid = node->tn_gid; MPASS(uid != VNOVAL && gid != VNOVAL); /* Disallow this operation if the file system is mounted read-only. */ if (vp->v_mount->mnt_flag & MNT_RDONLY) return EROFS; /* Immutable or append-only files cannot be modified, either. */ if (node->tn_flags & (IMMUTABLE | APPEND)) return EPERM; /* * To modify the ownership of a file, must possess VADMIN for that * file. */ if ((error = VOP_ACCESS(vp, VADMIN, cred, p))) return (error); /* * To change the owner of a file, or change the group of a file to a * group of which we are not a member, the caller must have * privilege. */ if ((uid != node->tn_uid || (gid != node->tn_gid && !groupmember(gid, cred))) && (error = priv_check_cred(cred, PRIV_VFS_CHOWN))) return (error); ogid = node->tn_gid; ouid = node->tn_uid; node->tn_uid = uid; node->tn_gid = gid; node->tn_status |= TMPFS_NODE_CHANGED; if ((node->tn_mode & (S_ISUID | S_ISGID)) && (ouid != uid || ogid != gid)) { if (priv_check_cred(cred, PRIV_VFS_RETAINSUGID)) node->tn_mode &= ~(S_ISUID | S_ISGID); } ASSERT_VOP_ELOCKED(vp, "chown2"); return (0); } /* * Change size of the given vnode. * Caller should execute tmpfs_update on vp after a successful execution. * The vnode must be locked on entry and remain locked on exit. */ int tmpfs_chsize(struct vnode *vp, u_quad_t size, struct ucred *cred, struct thread *p) { int error; struct tmpfs_node *node; ASSERT_VOP_ELOCKED(vp, "chsize"); node = VP_TO_TMPFS_NODE(vp); /* Decide whether this is a valid operation based on the file type. */ error = 0; switch (vp->v_type) { case VDIR: return EISDIR; case VREG: if (vp->v_mount->mnt_flag & MNT_RDONLY) return EROFS; break; case VBLK: /* FALLTHROUGH */ case VCHR: /* FALLTHROUGH */ case VFIFO: /* Allow modifications of special files even if in the file * system is mounted read-only (we are not modifying the * files themselves, but the objects they represent). */ return 0; default: /* Anything else is unsupported. */ return EOPNOTSUPP; } /* Immutable or append-only files cannot be modified, either. */ if (node->tn_flags & (IMMUTABLE | APPEND)) return EPERM; error = tmpfs_truncate(vp, size); /* tmpfs_truncate will raise the NOTE_EXTEND and NOTE_ATTRIB kevents * for us, as will update tn_status; no need to do that here. */ ASSERT_VOP_ELOCKED(vp, "chsize2"); return (error); } /* * Change access and modification times of the given vnode. * Caller should execute tmpfs_update on vp after a successful execution. * The vnode must be locked on entry and remain locked on exit. */ int tmpfs_chtimes(struct vnode *vp, struct vattr *vap, struct ucred *cred, struct thread *l) { int error; struct tmpfs_node *node; ASSERT_VOP_ELOCKED(vp, "chtimes"); node = VP_TO_TMPFS_NODE(vp); /* Disallow this operation if the file system is mounted read-only. */ if (vp->v_mount->mnt_flag & MNT_RDONLY) return EROFS; /* Immutable or append-only files cannot be modified, either. */ if (node->tn_flags & (IMMUTABLE | APPEND)) return EPERM; error = vn_utimes_perm(vp, vap, cred, l); if (error != 0) return (error); if (vap->va_atime.tv_sec != VNOVAL) node->tn_status |= TMPFS_NODE_ACCESSED; if (vap->va_mtime.tv_sec != VNOVAL) node->tn_status |= TMPFS_NODE_MODIFIED; if (vap->va_birthtime.tv_sec != VNOVAL) node->tn_status |= TMPFS_NODE_MODIFIED; tmpfs_itimes(vp, &vap->va_atime, &vap->va_mtime); if (vap->va_birthtime.tv_sec != VNOVAL) node->tn_birthtime = vap->va_birthtime; ASSERT_VOP_ELOCKED(vp, "chtimes2"); return (0); } void tmpfs_set_status(struct tmpfs_mount *tm, struct tmpfs_node *node, int status) { if ((node->tn_status & status) == status || tm->tm_ronly) return; TMPFS_NODE_LOCK(node); node->tn_status |= status; TMPFS_NODE_UNLOCK(node); } /* Sync timestamps */ void tmpfs_itimes(struct vnode *vp, const struct timespec *acc, const struct timespec *mod) { struct tmpfs_node *node; struct timespec now; ASSERT_VOP_LOCKED(vp, "tmpfs_itimes"); node = VP_TO_TMPFS_NODE(vp); if ((node->tn_status & (TMPFS_NODE_ACCESSED | TMPFS_NODE_MODIFIED | TMPFS_NODE_CHANGED)) == 0) return; vfs_timestamp(&now); TMPFS_NODE_LOCK(node); if (node->tn_status & TMPFS_NODE_ACCESSED) { if (acc == NULL) acc = &now; node->tn_atime = *acc; } if (node->tn_status & TMPFS_NODE_MODIFIED) { if (mod == NULL) mod = &now; node->tn_mtime = *mod; } if (node->tn_status & TMPFS_NODE_CHANGED) node->tn_ctime = now; node->tn_status &= ~(TMPFS_NODE_ACCESSED | TMPFS_NODE_MODIFIED | TMPFS_NODE_CHANGED); TMPFS_NODE_UNLOCK(node); /* XXX: FIX? The entropy here is desirable, but the harvesting may be expensive */ random_harvest_queue(node, sizeof(*node), RANDOM_FS_ATIME); } void tmpfs_update(struct vnode *vp) { tmpfs_itimes(vp, NULL, NULL); } int tmpfs_truncate(struct vnode *vp, off_t length) { int error; struct tmpfs_node *node; node = VP_TO_TMPFS_NODE(vp); if (length < 0) { error = EINVAL; goto out; } if (node->tn_size == length) { error = 0; goto out; } if (length > VFS_TO_TMPFS(vp->v_mount)->tm_maxfilesize) return (EFBIG); error = tmpfs_reg_resize(vp, length, FALSE); if (error == 0) node->tn_status |= TMPFS_NODE_CHANGED | TMPFS_NODE_MODIFIED; out: tmpfs_update(vp); return (error); } static __inline int tmpfs_dirtree_cmp(struct tmpfs_dirent *a, struct tmpfs_dirent *b) { if (a->td_hash > b->td_hash) return (1); else if (a->td_hash < b->td_hash) return (-1); return (0); } RB_GENERATE_STATIC(tmpfs_dir, tmpfs_dirent, uh.td_entries, tmpfs_dirtree_cmp); Index: head/sys/kern/kern_exec.c =================================================================== --- head/sys/kern/kern_exec.c (revision 353534) +++ head/sys/kern/kern_exec.c (revision 353535) @@ -1,1827 +1,1831 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 1993, David Greenman * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include "opt_capsicum.h" #include "opt_hwpmc_hooks.h" #include "opt_ktrace.h" #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef KTRACE #include #endif #include #include #include #include #include #include #include #include #include #ifdef HWPMC_HOOKS #include #endif #include #include #include #ifdef KDTRACE_HOOKS #include dtrace_execexit_func_t dtrace_fasttrap_exec; #endif SDT_PROVIDER_DECLARE(proc); SDT_PROBE_DEFINE1(proc, , , exec, "char *"); SDT_PROBE_DEFINE1(proc, , , exec__failure, "int"); SDT_PROBE_DEFINE1(proc, , , exec__success, "char *"); MALLOC_DEFINE(M_PARGS, "proc-args", "Process arguments"); int coredump_pack_fileinfo = 1; SYSCTL_INT(_kern, OID_AUTO, coredump_pack_fileinfo, CTLFLAG_RWTUN, &coredump_pack_fileinfo, 0, "Enable file path packing in 'procstat -f' coredump notes"); int coredump_pack_vmmapinfo = 1; SYSCTL_INT(_kern, OID_AUTO, coredump_pack_vmmapinfo, CTLFLAG_RWTUN, &coredump_pack_vmmapinfo, 0, "Enable file path packing in 'procstat -v' coredump notes"); static int sysctl_kern_ps_strings(SYSCTL_HANDLER_ARGS); static int sysctl_kern_usrstack(SYSCTL_HANDLER_ARGS); static int sysctl_kern_stackprot(SYSCTL_HANDLER_ARGS); static int do_execve(struct thread *td, struct image_args *args, struct mac *mac_p); /* XXX This should be vm_size_t. */ SYSCTL_PROC(_kern, KERN_PS_STRINGS, ps_strings, CTLTYPE_ULONG|CTLFLAG_RD| CTLFLAG_CAPRD|CTLFLAG_MPSAFE, NULL, 0, sysctl_kern_ps_strings, "LU", ""); /* XXX This should be vm_size_t. */ SYSCTL_PROC(_kern, KERN_USRSTACK, usrstack, CTLTYPE_ULONG|CTLFLAG_RD| CTLFLAG_CAPRD|CTLFLAG_MPSAFE, NULL, 0, sysctl_kern_usrstack, "LU", ""); SYSCTL_PROC(_kern, OID_AUTO, stackprot, CTLTYPE_INT|CTLFLAG_RD|CTLFLAG_MPSAFE, NULL, 0, sysctl_kern_stackprot, "I", ""); u_long ps_arg_cache_limit = PAGE_SIZE / 16; SYSCTL_ULONG(_kern, OID_AUTO, ps_arg_cache_limit, CTLFLAG_RW, &ps_arg_cache_limit, 0, ""); static int disallow_high_osrel; SYSCTL_INT(_kern, OID_AUTO, disallow_high_osrel, CTLFLAG_RW, &disallow_high_osrel, 0, "Disallow execution of binaries built for higher version of the world"); static int map_at_zero = 0; SYSCTL_INT(_security_bsd, OID_AUTO, map_at_zero, CTLFLAG_RWTUN, &map_at_zero, 0, "Permit processes to map an object at virtual address 0."); static int sysctl_kern_ps_strings(SYSCTL_HANDLER_ARGS) { struct proc *p; int error; p = curproc; #ifdef SCTL_MASK32 if (req->flags & SCTL_MASK32) { unsigned int val; val = (unsigned int)p->p_sysent->sv_psstrings; error = SYSCTL_OUT(req, &val, sizeof(val)); } else #endif error = SYSCTL_OUT(req, &p->p_sysent->sv_psstrings, sizeof(p->p_sysent->sv_psstrings)); return error; } static int sysctl_kern_usrstack(SYSCTL_HANDLER_ARGS) { struct proc *p; int error; p = curproc; #ifdef SCTL_MASK32 if (req->flags & SCTL_MASK32) { unsigned int val; val = (unsigned int)p->p_sysent->sv_usrstack; error = SYSCTL_OUT(req, &val, sizeof(val)); } else #endif error = SYSCTL_OUT(req, &p->p_sysent->sv_usrstack, sizeof(p->p_sysent->sv_usrstack)); return error; } static int sysctl_kern_stackprot(SYSCTL_HANDLER_ARGS) { struct proc *p; p = curproc; return (SYSCTL_OUT(req, &p->p_sysent->sv_stackprot, sizeof(p->p_sysent->sv_stackprot))); } /* * Each of the items is a pointer to a `const struct execsw', hence the * double pointer here. */ static const struct execsw **execsw; #ifndef _SYS_SYSPROTO_H_ struct execve_args { char *fname; char **argv; char **envv; }; #endif int sys_execve(struct thread *td, struct execve_args *uap) { struct image_args args; struct vmspace *oldvmspace; int error; error = pre_execve(td, &oldvmspace); if (error != 0) return (error); error = exec_copyin_args(&args, uap->fname, UIO_USERSPACE, uap->argv, uap->envv); if (error == 0) error = kern_execve(td, &args, NULL); post_execve(td, error, oldvmspace); return (error); } #ifndef _SYS_SYSPROTO_H_ struct fexecve_args { int fd; char **argv; char **envv; } #endif int sys_fexecve(struct thread *td, struct fexecve_args *uap) { struct image_args args; struct vmspace *oldvmspace; int error; error = pre_execve(td, &oldvmspace); if (error != 0) return (error); error = exec_copyin_args(&args, NULL, UIO_SYSSPACE, uap->argv, uap->envv); if (error == 0) { args.fd = uap->fd; error = kern_execve(td, &args, NULL); } post_execve(td, error, oldvmspace); return (error); } #ifndef _SYS_SYSPROTO_H_ struct __mac_execve_args { char *fname; char **argv; char **envv; struct mac *mac_p; }; #endif int sys___mac_execve(struct thread *td, struct __mac_execve_args *uap) { #ifdef MAC struct image_args args; struct vmspace *oldvmspace; int error; error = pre_execve(td, &oldvmspace); if (error != 0) return (error); error = exec_copyin_args(&args, uap->fname, UIO_USERSPACE, uap->argv, uap->envv); if (error == 0) error = kern_execve(td, &args, uap->mac_p); post_execve(td, error, oldvmspace); return (error); #else return (ENOSYS); #endif } int pre_execve(struct thread *td, struct vmspace **oldvmspace) { struct proc *p; int error; KASSERT(td == curthread, ("non-current thread %p", td)); error = 0; p = td->td_proc; if ((p->p_flag & P_HADTHREADS) != 0) { PROC_LOCK(p); if (thread_single(p, SINGLE_BOUNDARY) != 0) error = ERESTART; PROC_UNLOCK(p); } KASSERT(error != 0 || (td->td_pflags & TDP_EXECVMSPC) == 0, ("nested execve")); *oldvmspace = p->p_vmspace; return (error); } void post_execve(struct thread *td, int error, struct vmspace *oldvmspace) { struct proc *p; KASSERT(td == curthread, ("non-current thread %p", td)); p = td->td_proc; if ((p->p_flag & P_HADTHREADS) != 0) { PROC_LOCK(p); /* * If success, we upgrade to SINGLE_EXIT state to * force other threads to suicide. */ if (error == EJUSTRETURN) thread_single(p, SINGLE_EXIT); else thread_single_end(p, SINGLE_BOUNDARY); PROC_UNLOCK(p); } if ((td->td_pflags & TDP_EXECVMSPC) != 0) { KASSERT(p->p_vmspace != oldvmspace, ("oldvmspace still used")); vmspace_free(oldvmspace); td->td_pflags &= ~TDP_EXECVMSPC; } } /* * XXX: kern_execve has the astonishing property of not always returning to * the caller. If sufficiently bad things happen during the call to * do_execve(), it can end up calling exit1(); as a result, callers must * avoid doing anything which they might need to undo (e.g., allocating * memory). */ int kern_execve(struct thread *td, struct image_args *args, struct mac *mac_p) { AUDIT_ARG_ARGV(args->begin_argv, args->argc, exec_args_get_begin_envv(args) - args->begin_argv); AUDIT_ARG_ENVV(exec_args_get_begin_envv(args), args->envc, args->endp - exec_args_get_begin_envv(args)); return (do_execve(td, args, mac_p)); } /* * In-kernel implementation of execve(). All arguments are assumed to be * userspace pointers from the passed thread. */ static int do_execve(struct thread *td, struct image_args *args, struct mac *mac_p) { struct proc *p = td->td_proc; struct nameidata nd; struct ucred *oldcred; struct uidinfo *euip = NULL; register_t *stack_base; int error, i; struct image_params image_params, *imgp; struct vattr attr; int (*img_first)(struct image_params *); struct pargs *oldargs = NULL, *newargs = NULL; struct sigacts *oldsigacts = NULL, *newsigacts = NULL; #ifdef KTRACE struct vnode *tracevp = NULL; struct ucred *tracecred = NULL; #endif struct vnode *oldtextvp = NULL, *newtextvp; int credential_changing; #ifdef MAC struct label *interpvplabel = NULL; int will_transition; #endif #ifdef HWPMC_HOOKS struct pmckern_procexec pe; #endif static const char fexecv_proc_title[] = "(fexecv)"; imgp = &image_params; /* * Lock the process and set the P_INEXEC flag to indicate that * it should be left alone until we're done here. This is * necessary to avoid race conditions - e.g. in ptrace() - * that might allow a local user to illicitly obtain elevated * privileges. */ PROC_LOCK(p); KASSERT((p->p_flag & P_INEXEC) == 0, ("%s(): process already has P_INEXEC flag", __func__)); p->p_flag |= P_INEXEC; PROC_UNLOCK(p); /* * Initialize part of the common data */ bzero(imgp, sizeof(*imgp)); imgp->proc = p; imgp->attr = &attr; imgp->args = args; oldcred = p->p_ucred; #ifdef MAC error = mac_execve_enter(imgp, mac_p); if (error) goto exec_fail; #endif /* * Translate the file name. namei() returns a vnode pointer * in ni_vp among other things. * * XXXAUDIT: It would be desirable to also audit the name of the * interpreter if this is an interpreted binary. */ if (args->fname != NULL) { NDINIT(&nd, LOOKUP, ISOPEN | LOCKLEAF | LOCKSHARED | FOLLOW | SAVENAME | AUDITVNODE1, UIO_SYSSPACE, args->fname, td); } SDT_PROBE1(proc, , , exec, args->fname); interpret: if (args->fname != NULL) { #ifdef CAPABILITY_MODE /* * While capability mode can't reach this point via direct * path arguments to execve(), we also don't allow * interpreters to be used in capability mode (for now). * Catch indirect lookups and return a permissions error. */ if (IN_CAPABILITY_MODE(td)) { error = ECAPMODE; goto exec_fail; } #endif error = namei(&nd); if (error) goto exec_fail; newtextvp = nd.ni_vp; imgp->vp = newtextvp; } else { AUDIT_ARG_FD(args->fd); /* * Descriptors opened only with O_EXEC or O_RDONLY are allowed. */ error = fgetvp_exec(td, args->fd, &cap_fexecve_rights, &newtextvp); if (error) goto exec_fail; vn_lock(newtextvp, LK_SHARED | LK_RETRY); AUDIT_ARG_VNODE1(newtextvp); imgp->vp = newtextvp; } /* * Check file permissions. Also 'opens' file and sets its vnode to * text mode. */ error = exec_check_permissions(imgp); if (error) goto exec_fail_dealloc; imgp->object = imgp->vp->v_object; if (imgp->object != NULL) vm_object_reference(imgp->object); error = exec_map_first_page(imgp); if (error) goto exec_fail_dealloc; imgp->proc->p_osrel = 0; imgp->proc->p_fctl0 = 0; /* * Implement image setuid/setgid. * * Determine new credentials before attempting image activators * so that it can be used by process_exec handlers to determine * credential/setid changes. * * Don't honor setuid/setgid if the filesystem prohibits it or if * the process is being traced. * * We disable setuid/setgid/etc in capability mode on the basis * that most setugid applications are not written with that * environment in mind, and will therefore almost certainly operate * incorrectly. In principle there's no reason that setugid * applications might not be useful in capability mode, so we may want * to reconsider this conservative design choice in the future. * * XXXMAC: For the time being, use NOSUID to also prohibit * transitions on the file system. */ credential_changing = 0; credential_changing |= (attr.va_mode & S_ISUID) && oldcred->cr_uid != attr.va_uid; credential_changing |= (attr.va_mode & S_ISGID) && oldcred->cr_gid != attr.va_gid; #ifdef MAC will_transition = mac_vnode_execve_will_transition(oldcred, imgp->vp, interpvplabel, imgp); credential_changing |= will_transition; #endif /* Don't inherit PROC_PDEATHSIG_CTL value if setuid/setgid. */ if (credential_changing) imgp->proc->p_pdeathsig = 0; if (credential_changing && #ifdef CAPABILITY_MODE ((oldcred->cr_flags & CRED_FLAG_CAPMODE) == 0) && #endif (imgp->vp->v_mount->mnt_flag & MNT_NOSUID) == 0 && (p->p_flag & P_TRACED) == 0) { imgp->credential_setid = true; VOP_UNLOCK(imgp->vp, 0); imgp->newcred = crdup(oldcred); if (attr.va_mode & S_ISUID) { euip = uifind(attr.va_uid); change_euid(imgp->newcred, euip); } vn_lock(imgp->vp, LK_SHARED | LK_RETRY); if (attr.va_mode & S_ISGID) change_egid(imgp->newcred, attr.va_gid); /* * Implement correct POSIX saved-id behavior. * * XXXMAC: Note that the current logic will save the * uid and gid if a MAC domain transition occurs, even * though maybe it shouldn't. */ change_svuid(imgp->newcred, imgp->newcred->cr_uid); change_svgid(imgp->newcred, imgp->newcred->cr_gid); } else { /* * Implement correct POSIX saved-id behavior. * * XXX: It's not clear that the existing behavior is * POSIX-compliant. A number of sources indicate that the * saved uid/gid should only be updated if the new ruid is * not equal to the old ruid, or the new euid is not equal * to the old euid and the new euid is not equal to the old * ruid. The FreeBSD code always updates the saved uid/gid. * Also, this code uses the new (replaced) euid and egid as * the source, which may or may not be the right ones to use. */ if (oldcred->cr_svuid != oldcred->cr_uid || oldcred->cr_svgid != oldcred->cr_gid) { VOP_UNLOCK(imgp->vp, 0); imgp->newcred = crdup(oldcred); vn_lock(imgp->vp, LK_SHARED | LK_RETRY); change_svuid(imgp->newcred, imgp->newcred->cr_uid); change_svgid(imgp->newcred, imgp->newcred->cr_gid); } } /* The new credentials are installed into the process later. */ /* * Do the best to calculate the full path to the image file. */ if (args->fname != NULL && args->fname[0] == '/') imgp->execpath = args->fname; else { VOP_UNLOCK(imgp->vp, 0); if (vn_fullpath(td, imgp->vp, &imgp->execpath, &imgp->freepath) != 0) imgp->execpath = args->fname; vn_lock(imgp->vp, LK_SHARED | LK_RETRY); } /* * If the current process has a special image activator it * wants to try first, call it. For example, emulating shell * scripts differently. */ error = -1; if ((img_first = imgp->proc->p_sysent->sv_imgact_try) != NULL) error = img_first(imgp); /* * Loop through the list of image activators, calling each one. * An activator returns -1 if there is no match, 0 on success, * and an error otherwise. */ for (i = 0; error == -1 && execsw[i]; ++i) { if (execsw[i]->ex_imgact == NULL || execsw[i]->ex_imgact == img_first) { continue; } error = (*execsw[i]->ex_imgact)(imgp); } if (error) { if (error == -1) error = ENOEXEC; goto exec_fail_dealloc; } /* * Special interpreter operation, cleanup and loop up to try to * activate the interpreter. */ if (imgp->interpreted) { exec_unmap_first_page(imgp); /* * The text reference needs to be removed for scripts. * There is a short period before we determine that * something is a script where text reference is active. * The vnode lock is held over this entire period * so nothing should illegitimately be blocked. */ MPASS(imgp->textset); VOP_UNSET_TEXT_CHECKED(newtextvp); imgp->textset = false; /* free name buffer and old vnode */ if (args->fname != NULL) NDFREE(&nd, NDF_ONLY_PNBUF); #ifdef MAC mac_execve_interpreter_enter(newtextvp, &interpvplabel); #endif if (imgp->opened) { VOP_CLOSE(newtextvp, FREAD, td->td_ucred, td); imgp->opened = 0; } vput(newtextvp); vm_object_deallocate(imgp->object); imgp->object = NULL; imgp->credential_setid = false; if (imgp->newcred != NULL) { crfree(imgp->newcred); imgp->newcred = NULL; } imgp->execpath = NULL; free(imgp->freepath, M_TEMP); imgp->freepath = NULL; /* set new name to that of the interpreter */ NDINIT(&nd, LOOKUP, ISOPEN | LOCKLEAF | FOLLOW | SAVENAME, UIO_SYSSPACE, imgp->interpreter_name, td); args->fname = imgp->interpreter_name; goto interpret; } /* * NB: We unlock the vnode here because it is believed that none * of the sv_copyout_strings/sv_fixup operations require the vnode. */ VOP_UNLOCK(imgp->vp, 0); if (disallow_high_osrel && P_OSREL_MAJOR(p->p_osrel) > P_OSREL_MAJOR(__FreeBSD_version)) { error = ENOEXEC; uprintf("Osrel %d for image %s too high\n", p->p_osrel, imgp->execpath != NULL ? imgp->execpath : ""); vn_lock(imgp->vp, LK_SHARED | LK_RETRY); goto exec_fail_dealloc; } /* ABI enforces the use of Capsicum. Switch into capabilities mode. */ if (SV_PROC_FLAG(p, SV_CAPSICUM)) sys_cap_enter(td, NULL); /* * Copy out strings (args and env) and initialize stack base. */ stack_base = (*p->p_sysent->sv_copyout_strings)(imgp); /* * Stack setup. */ error = (*p->p_sysent->sv_fixup)(&stack_base, imgp); if (error != 0) { vn_lock(imgp->vp, LK_SHARED | LK_RETRY); goto exec_fail_dealloc; } if (args->fdp != NULL) { /* Install a brand new file descriptor table. */ fdinstall_remapped(td, args->fdp); args->fdp = NULL; } else { /* * Keep on using the existing file descriptor table. For * security and other reasons, the file descriptor table * cannot be shared after an exec. */ fdunshare(td); /* close files on exec */ fdcloseexec(td); } /* * Malloc things before we need locks. */ i = exec_args_get_begin_envv(imgp->args) - imgp->args->begin_argv; /* Cache arguments if they fit inside our allowance */ if (ps_arg_cache_limit >= i + sizeof(struct pargs)) { newargs = pargs_alloc(i); bcopy(imgp->args->begin_argv, newargs->ar_args, i); } /* * For security and other reasons, signal handlers cannot * be shared after an exec. The new process gets a copy of the old * handlers. In execsigs(), the new process will have its signals * reset. */ if (sigacts_shared(p->p_sigacts)) { oldsigacts = p->p_sigacts; newsigacts = sigacts_alloc(); sigacts_copy(newsigacts, oldsigacts); } vn_lock(imgp->vp, LK_SHARED | LK_RETRY); PROC_LOCK(p); if (oldsigacts) p->p_sigacts = newsigacts; /* Stop profiling */ stopprofclock(p); /* reset caught signals */ execsigs(p); /* name this process - nameiexec(p, ndp) */ bzero(p->p_comm, sizeof(p->p_comm)); if (args->fname) bcopy(nd.ni_cnd.cn_nameptr, p->p_comm, min(nd.ni_cnd.cn_namelen, MAXCOMLEN)); else if (vn_commname(newtextvp, p->p_comm, sizeof(p->p_comm)) != 0) bcopy(fexecv_proc_title, p->p_comm, sizeof(fexecv_proc_title)); bcopy(p->p_comm, td->td_name, sizeof(td->td_name)); #ifdef KTR sched_clear_tdname(td); #endif /* * mark as execed, wakeup the process that vforked (if any) and tell * it that it now has its own resources back */ p->p_flag |= P_EXEC; if ((p->p_flag2 & P2_NOTRACE_EXEC) == 0) p->p_flag2 &= ~P2_NOTRACE; if ((p->p_flag2 & P2_STKGAP_DISABLE_EXEC) == 0) p->p_flag2 &= ~P2_STKGAP_DISABLE; if (p->p_flag & P_PPWAIT) { p->p_flag &= ~(P_PPWAIT | P_PPTRACE); cv_broadcast(&p->p_pwait); /* STOPs are no longer ignored, arrange for AST */ signotify(td); } /* * Implement image setuid/setgid installation. */ if (imgp->credential_setid) { /* * Turn off syscall tracing for set-id programs, except for * root. Record any set-id flags first to make sure that * we do not regain any tracing during a possible block. */ setsugid(p); #ifdef KTRACE if (p->p_tracecred != NULL && priv_check_cred(p->p_tracecred, PRIV_DEBUG_DIFFCRED)) ktrprocexec(p, &tracecred, &tracevp); #endif /* * Close any file descriptors 0..2 that reference procfs, * then make sure file descriptors 0..2 are in use. * * Both fdsetugidsafety() and fdcheckstd() may call functions * taking sleepable locks, so temporarily drop our locks. */ PROC_UNLOCK(p); VOP_UNLOCK(imgp->vp, 0); fdsetugidsafety(td); error = fdcheckstd(td); vn_lock(imgp->vp, LK_SHARED | LK_RETRY); if (error != 0) goto exec_fail_dealloc; PROC_LOCK(p); #ifdef MAC if (will_transition) { mac_vnode_execve_transition(oldcred, imgp->newcred, imgp->vp, interpvplabel, imgp); } #endif } else { if (oldcred->cr_uid == oldcred->cr_ruid && oldcred->cr_gid == oldcred->cr_rgid) p->p_flag &= ~P_SUGID; } /* * Set the new credentials. */ if (imgp->newcred != NULL) { proc_set_cred(p, imgp->newcred); crfree(oldcred); oldcred = NULL; } /* * Store the vp for use in procfs. This vnode was referenced by namei * or fgetvp_exec. */ oldtextvp = p->p_textvp; p->p_textvp = newtextvp; #ifdef KDTRACE_HOOKS /* * Tell the DTrace fasttrap provider about the exec if it * has declared an interest. */ if (dtrace_fasttrap_exec) dtrace_fasttrap_exec(p); #endif /* * Notify others that we exec'd, and clear the P_INEXEC flag * as we're now a bona fide freshly-execed process. */ KNOTE_LOCKED(p->p_klist, NOTE_EXEC); p->p_flag &= ~P_INEXEC; /* clear "fork but no exec" flag, as we _are_ execing */ p->p_acflag &= ~AFORK; /* * Free any previous argument cache and replace it with * the new argument cache, if any. */ oldargs = p->p_args; p->p_args = newargs; newargs = NULL; PROC_UNLOCK(p); #ifdef HWPMC_HOOKS /* * Check if system-wide sampling is in effect or if the * current process is using PMCs. If so, do exec() time * processing. This processing needs to happen AFTER the * P_INEXEC flag is cleared. */ if (PMC_SYSTEM_SAMPLING_ACTIVE() || PMC_PROC_IS_USING_PMCS(p)) { VOP_UNLOCK(imgp->vp, 0); pe.pm_credentialschanged = credential_changing; pe.pm_entryaddr = imgp->entry_addr; PMC_CALL_HOOK_X(td, PMC_FN_PROCESS_EXEC, (void *) &pe); vn_lock(imgp->vp, LK_SHARED | LK_RETRY); } #endif /* Set values passed into the program in registers. */ (*p->p_sysent->sv_setregs)(td, imgp, (u_long)(uintptr_t)stack_base); vfs_mark_atime(imgp->vp, td->td_ucred); SDT_PROBE1(proc, , , exec__success, args->fname); exec_fail_dealloc: if (imgp->firstpage != NULL) exec_unmap_first_page(imgp); if (imgp->vp != NULL) { if (args->fname) NDFREE(&nd, NDF_ONLY_PNBUF); if (imgp->opened) VOP_CLOSE(imgp->vp, FREAD, td->td_ucred, td); if (imgp->textset) VOP_UNSET_TEXT_CHECKED(imgp->vp); if (error != 0) vput(imgp->vp); else VOP_UNLOCK(imgp->vp, 0); } if (imgp->object != NULL) vm_object_deallocate(imgp->object); free(imgp->freepath, M_TEMP); if (error == 0) { if (p->p_ptevents & PTRACE_EXEC) { PROC_LOCK(p); if (p->p_ptevents & PTRACE_EXEC) td->td_dbgflags |= TDB_EXEC; PROC_UNLOCK(p); } /* * Stop the process here if its stop event mask has * the S_EXEC bit set. */ STOPEVENT(p, S_EXEC, 0); } else { exec_fail: /* we're done here, clear P_INEXEC */ PROC_LOCK(p); p->p_flag &= ~P_INEXEC; PROC_UNLOCK(p); SDT_PROBE1(proc, , , exec__failure, error); } if (imgp->newcred != NULL && oldcred != NULL) crfree(imgp->newcred); #ifdef MAC mac_execve_exit(imgp); mac_execve_interpreter_exit(interpvplabel); #endif exec_free_args(args); /* * Handle deferred decrement of ref counts. */ if (oldtextvp != NULL) vrele(oldtextvp); #ifdef KTRACE if (tracevp != NULL) vrele(tracevp); if (tracecred != NULL) crfree(tracecred); #endif pargs_drop(oldargs); pargs_drop(newargs); if (oldsigacts != NULL) sigacts_free(oldsigacts); if (euip != NULL) uifree(euip); if (error && imgp->vmspace_destroyed) { /* sorry, no more process anymore. exit gracefully */ exit1(td, 0, SIGABRT); /* NOT REACHED */ } #ifdef KTRACE if (error == 0) ktrprocctor(p); #endif /* * We don't want cpu_set_syscall_retval() to overwrite any of * the register values put in place by exec_setregs(). * Implementations of cpu_set_syscall_retval() will leave * registers unmodified when returning EJUSTRETURN. */ return (error == 0 ? EJUSTRETURN : error); } int exec_map_first_page(struct image_params *imgp) { int rv, i, after, initial_pagein; vm_page_t ma[VM_INITIAL_PAGEIN]; vm_object_t object; if (imgp->firstpage != NULL) exec_unmap_first_page(imgp); object = imgp->vp->v_object; if (object == NULL) return (EACCES); VM_OBJECT_WLOCK(object); #if VM_NRESERVLEVEL > 0 vm_object_color(object, 0); #endif +retry: ma[0] = vm_page_grab(object, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY | VM_ALLOC_WIRED); if (ma[0]->valid != VM_PAGE_BITS_ALL) { - vm_page_xbusy(ma[0]); + if (vm_page_busy_acquire(ma[0], VM_ALLOC_WAITFAIL) == 0) { + vm_page_unwire_noq(ma[0]); + goto retry; + } if (!vm_pager_has_page(object, 0, NULL, &after)) { if (vm_page_unwire_noq(ma[0])) vm_page_free(ma[0]); else vm_page_xunbusy(ma[0]); VM_OBJECT_WUNLOCK(object); return (EIO); } initial_pagein = min(after, VM_INITIAL_PAGEIN); KASSERT(initial_pagein <= object->size, ("%s: initial_pagein %d object->size %ju", __func__, initial_pagein, (uintmax_t )object->size)); for (i = 1; i < initial_pagein; i++) { if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) { if (ma[i]->valid) break; if (!vm_page_tryxbusy(ma[i])) break; } else { ma[i] = vm_page_alloc(object, i, VM_ALLOC_NORMAL); if (ma[i] == NULL) break; } } initial_pagein = i; rv = vm_pager_get_pages(object, ma, initial_pagein, NULL, NULL); if (rv != VM_PAGER_OK) { if (vm_page_unwire_noq(ma[0])) vm_page_free(ma[0]); else vm_page_xunbusy(ma[0]); for (i = 1; i < initial_pagein; i++) { if (!vm_page_wired(ma[i])) vm_page_free(ma[i]); else vm_page_xunbusy(ma[i]); } VM_OBJECT_WUNLOCK(object); return (EIO); } vm_page_xunbusy(ma[0]); for (i = 1; i < initial_pagein; i++) vm_page_readahead_finish(ma[i]); } VM_OBJECT_WUNLOCK(object); imgp->firstpage = sf_buf_alloc(ma[0], 0); imgp->image_header = (char *)sf_buf_kva(imgp->firstpage); return (0); } void exec_unmap_first_page(struct image_params *imgp) { vm_page_t m; if (imgp->firstpage != NULL) { m = sf_buf_page(imgp->firstpage); sf_buf_free(imgp->firstpage); imgp->firstpage = NULL; vm_page_unwire(m, PQ_ACTIVE); } } /* * Destroy old address space, and allocate a new stack. * The new stack is only sgrowsiz large because it is grown * automatically on a page fault. */ int exec_new_vmspace(struct image_params *imgp, struct sysentvec *sv) { int error; struct proc *p = imgp->proc; struct vmspace *vmspace = p->p_vmspace; vm_object_t obj; struct rlimit rlim_stack; vm_offset_t sv_minuser, stack_addr; vm_map_t map; u_long ssiz; imgp->vmspace_destroyed = 1; imgp->sysent = sv; /* May be called with Giant held */ EVENTHANDLER_DIRECT_INVOKE(process_exec, p, imgp); /* * Blow away entire process VM, if address space not shared, * otherwise, create a new VM space so that other threads are * not disrupted */ map = &vmspace->vm_map; if (map_at_zero) sv_minuser = sv->sv_minuser; else sv_minuser = MAX(sv->sv_minuser, PAGE_SIZE); if (vmspace->vm_refcnt == 1 && vm_map_min(map) == sv_minuser && vm_map_max(map) == sv->sv_maxuser && cpu_exec_vmspace_reuse(p, map)) { shmexit(vmspace); pmap_remove_pages(vmspace_pmap(vmspace)); vm_map_remove(map, vm_map_min(map), vm_map_max(map)); /* * An exec terminates mlockall(MCL_FUTURE), ASLR state * must be re-evaluated. */ vm_map_lock(map); vm_map_modflags(map, 0, MAP_WIREFUTURE | MAP_ASLR | MAP_ASLR_IGNSTART); vm_map_unlock(map); } else { error = vmspace_exec(p, sv_minuser, sv->sv_maxuser); if (error) return (error); vmspace = p->p_vmspace; map = &vmspace->vm_map; } map->flags |= imgp->map_flags; /* Map a shared page */ obj = sv->sv_shared_page_obj; if (obj != NULL) { vm_object_reference(obj); error = vm_map_fixed(map, obj, 0, sv->sv_shared_page_base, sv->sv_shared_page_len, VM_PROT_READ | VM_PROT_EXECUTE, VM_PROT_READ | VM_PROT_EXECUTE, MAP_INHERIT_SHARE | MAP_ACC_NO_CHARGE); if (error != KERN_SUCCESS) { vm_object_deallocate(obj); return (vm_mmap_to_errno(error)); } } /* Allocate a new stack */ if (imgp->stack_sz != 0) { ssiz = trunc_page(imgp->stack_sz); PROC_LOCK(p); lim_rlimit_proc(p, RLIMIT_STACK, &rlim_stack); PROC_UNLOCK(p); if (ssiz > rlim_stack.rlim_max) ssiz = rlim_stack.rlim_max; if (ssiz > rlim_stack.rlim_cur) { rlim_stack.rlim_cur = ssiz; kern_setrlimit(curthread, RLIMIT_STACK, &rlim_stack); } } else if (sv->sv_maxssiz != NULL) { ssiz = *sv->sv_maxssiz; } else { ssiz = maxssiz; } imgp->eff_stack_sz = lim_cur(curthread, RLIMIT_STACK); if (ssiz < imgp->eff_stack_sz) imgp->eff_stack_sz = ssiz; stack_addr = sv->sv_usrstack - ssiz; error = vm_map_stack(map, stack_addr, (vm_size_t)ssiz, obj != NULL && imgp->stack_prot != 0 ? imgp->stack_prot : sv->sv_stackprot, VM_PROT_ALL, MAP_STACK_GROWS_DOWN); if (error != KERN_SUCCESS) return (vm_mmap_to_errno(error)); /* * vm_ssize and vm_maxsaddr are somewhat antiquated concepts, but they * are still used to enforce the stack rlimit on the process stack. */ vmspace->vm_ssize = sgrowsiz >> PAGE_SHIFT; vmspace->vm_maxsaddr = (char *)stack_addr; return (0); } /* * Copy out argument and environment strings from the old process address * space into the temporary string buffer. */ int exec_copyin_args(struct image_args *args, const char *fname, enum uio_seg segflg, char **argv, char **envv) { u_long arg, env; int error; bzero(args, sizeof(*args)); if (argv == NULL) return (EFAULT); /* * Allocate demand-paged memory for the file name, argument, and * environment strings. */ error = exec_alloc_args(args); if (error != 0) return (error); /* * Copy the file name. */ error = exec_args_add_fname(args, fname, segflg); if (error != 0) goto err_exit; /* * extract arguments first */ for (;;) { error = fueword(argv++, &arg); if (error == -1) { error = EFAULT; goto err_exit; } if (arg == 0) break; error = exec_args_add_arg(args, (char *)(uintptr_t)arg, UIO_USERSPACE); if (error != 0) goto err_exit; } /* * extract environment strings */ if (envv) { for (;;) { error = fueword(envv++, &env); if (error == -1) { error = EFAULT; goto err_exit; } if (env == 0) break; error = exec_args_add_env(args, (char *)(uintptr_t)env, UIO_USERSPACE); if (error != 0) goto err_exit; } } return (0); err_exit: exec_free_args(args); return (error); } int exec_copyin_data_fds(struct thread *td, struct image_args *args, const void *data, size_t datalen, const int *fds, size_t fdslen) { struct filedesc *ofdp; const char *p; int *kfds; int error; memset(args, '\0', sizeof(*args)); ofdp = td->td_proc->p_fd; if (datalen >= ARG_MAX || fdslen > ofdp->fd_lastfile + 1) return (E2BIG); error = exec_alloc_args(args); if (error != 0) return (error); args->begin_argv = args->buf; args->stringspace = ARG_MAX; if (datalen > 0) { /* * Argument buffer has been provided. Copy it into the * kernel as a single string and add a terminating null * byte. */ error = copyin(data, args->begin_argv, datalen); if (error != 0) goto err_exit; args->begin_argv[datalen] = '\0'; args->endp = args->begin_argv + datalen + 1; args->stringspace -= datalen + 1; /* * Traditional argument counting. Count the number of * null bytes. */ for (p = args->begin_argv; p < args->endp; ++p) if (*p == '\0') ++args->argc; } else { /* No argument buffer provided. */ args->endp = args->begin_argv; } /* Create new file descriptor table. */ kfds = malloc(fdslen * sizeof(int), M_TEMP, M_WAITOK); error = copyin(fds, kfds, fdslen * sizeof(int)); if (error != 0) { free(kfds, M_TEMP); goto err_exit; } error = fdcopy_remapped(ofdp, kfds, fdslen, &args->fdp); free(kfds, M_TEMP); if (error != 0) goto err_exit; return (0); err_exit: exec_free_args(args); return (error); } struct exec_args_kva { vm_offset_t addr; u_int gen; SLIST_ENTRY(exec_args_kva) next; }; DPCPU_DEFINE_STATIC(struct exec_args_kva *, exec_args_kva); static SLIST_HEAD(, exec_args_kva) exec_args_kva_freelist; static struct mtx exec_args_kva_mtx; static u_int exec_args_gen; static void exec_prealloc_args_kva(void *arg __unused) { struct exec_args_kva *argkva; u_int i; SLIST_INIT(&exec_args_kva_freelist); mtx_init(&exec_args_kva_mtx, "exec args kva", NULL, MTX_DEF); for (i = 0; i < exec_map_entries; i++) { argkva = malloc(sizeof(*argkva), M_PARGS, M_WAITOK); argkva->addr = kmap_alloc_wait(exec_map, exec_map_entry_size); argkva->gen = exec_args_gen; SLIST_INSERT_HEAD(&exec_args_kva_freelist, argkva, next); } } SYSINIT(exec_args_kva, SI_SUB_EXEC, SI_ORDER_ANY, exec_prealloc_args_kva, NULL); static vm_offset_t exec_alloc_args_kva(void **cookie) { struct exec_args_kva *argkva; argkva = (void *)atomic_readandclear_ptr( (uintptr_t *)DPCPU_PTR(exec_args_kva)); if (argkva == NULL) { mtx_lock(&exec_args_kva_mtx); while ((argkva = SLIST_FIRST(&exec_args_kva_freelist)) == NULL) (void)mtx_sleep(&exec_args_kva_freelist, &exec_args_kva_mtx, 0, "execkva", 0); SLIST_REMOVE_HEAD(&exec_args_kva_freelist, next); mtx_unlock(&exec_args_kva_mtx); } *(struct exec_args_kva **)cookie = argkva; return (argkva->addr); } static void exec_release_args_kva(struct exec_args_kva *argkva, u_int gen) { vm_offset_t base; base = argkva->addr; if (argkva->gen != gen) { (void)vm_map_madvise(exec_map, base, base + exec_map_entry_size, MADV_FREE); argkva->gen = gen; } if (!atomic_cmpset_ptr((uintptr_t *)DPCPU_PTR(exec_args_kva), (uintptr_t)NULL, (uintptr_t)argkva)) { mtx_lock(&exec_args_kva_mtx); SLIST_INSERT_HEAD(&exec_args_kva_freelist, argkva, next); wakeup_one(&exec_args_kva_freelist); mtx_unlock(&exec_args_kva_mtx); } } static void exec_free_args_kva(void *cookie) { exec_release_args_kva(cookie, exec_args_gen); } static void exec_args_kva_lowmem(void *arg __unused) { SLIST_HEAD(, exec_args_kva) head; struct exec_args_kva *argkva; u_int gen; int i; gen = atomic_fetchadd_int(&exec_args_gen, 1) + 1; /* * Force an madvise of each KVA range. Any currently allocated ranges * will have MADV_FREE applied once they are freed. */ SLIST_INIT(&head); mtx_lock(&exec_args_kva_mtx); SLIST_SWAP(&head, &exec_args_kva_freelist, exec_args_kva); mtx_unlock(&exec_args_kva_mtx); while ((argkva = SLIST_FIRST(&head)) != NULL) { SLIST_REMOVE_HEAD(&head, next); exec_release_args_kva(argkva, gen); } CPU_FOREACH(i) { argkva = (void *)atomic_readandclear_ptr( (uintptr_t *)DPCPU_ID_PTR(i, exec_args_kva)); if (argkva != NULL) exec_release_args_kva(argkva, gen); } } EVENTHANDLER_DEFINE(vm_lowmem, exec_args_kva_lowmem, NULL, EVENTHANDLER_PRI_ANY); /* * Allocate temporary demand-paged, zero-filled memory for the file name, * argument, and environment strings. */ int exec_alloc_args(struct image_args *args) { args->buf = (char *)exec_alloc_args_kva(&args->bufkva); return (0); } void exec_free_args(struct image_args *args) { if (args->buf != NULL) { exec_free_args_kva(args->bufkva); args->buf = NULL; } if (args->fname_buf != NULL) { free(args->fname_buf, M_TEMP); args->fname_buf = NULL; } if (args->fdp != NULL) fdescfree_remapped(args->fdp); } /* * A set to functions to fill struct image args. * * NOTE: exec_args_add_fname() must be called (possibly with a NULL * fname) before the other functions. All exec_args_add_arg() calls must * be made before any exec_args_add_env() calls. exec_args_adjust_args() * may be called any time after exec_args_add_fname(). * * exec_args_add_fname() - install path to be executed * exec_args_add_arg() - append an argument string * exec_args_add_env() - append an env string * exec_args_adjust_args() - adjust location of the argument list to * allow new arguments to be prepended */ int exec_args_add_fname(struct image_args *args, const char *fname, enum uio_seg segflg) { int error; size_t length; KASSERT(args->fname == NULL, ("fname already appended")); KASSERT(args->endp == NULL, ("already appending to args")); if (fname != NULL) { args->fname = args->buf; error = segflg == UIO_SYSSPACE ? copystr(fname, args->fname, PATH_MAX, &length) : copyinstr(fname, args->fname, PATH_MAX, &length); if (error != 0) return (error == ENAMETOOLONG ? E2BIG : error); } else length = 0; /* Set up for _arg_*()/_env_*() */ args->endp = args->buf + length; /* begin_argv must be set and kept updated */ args->begin_argv = args->endp; KASSERT(exec_map_entry_size - length >= ARG_MAX, ("too little space remaining for arguments %zu < %zu", exec_map_entry_size - length, (size_t)ARG_MAX)); args->stringspace = ARG_MAX; return (0); } static int exec_args_add_str(struct image_args *args, const char *str, enum uio_seg segflg, int *countp) { int error; size_t length; KASSERT(args->endp != NULL, ("endp not initialized")); KASSERT(args->begin_argv != NULL, ("begin_argp not initialized")); error = (segflg == UIO_SYSSPACE) ? copystr(str, args->endp, args->stringspace, &length) : copyinstr(str, args->endp, args->stringspace, &length); if (error != 0) return (error == ENAMETOOLONG ? E2BIG : error); args->stringspace -= length; args->endp += length; (*countp)++; return (0); } int exec_args_add_arg(struct image_args *args, const char *argp, enum uio_seg segflg) { KASSERT(args->envc == 0, ("appending args after env")); return (exec_args_add_str(args, argp, segflg, &args->argc)); } int exec_args_add_env(struct image_args *args, const char *envp, enum uio_seg segflg) { if (args->envc == 0) args->begin_envv = args->endp; return (exec_args_add_str(args, envp, segflg, &args->envc)); } int exec_args_adjust_args(struct image_args *args, size_t consume, ssize_t extend) { ssize_t offset; KASSERT(args->endp != NULL, ("endp not initialized")); KASSERT(args->begin_argv != NULL, ("begin_argp not initialized")); offset = extend - consume; if (args->stringspace < offset) return (E2BIG); memmove(args->begin_argv + extend, args->begin_argv + consume, args->endp - args->begin_argv + consume); if (args->envc > 0) args->begin_envv += offset; args->endp += offset; args->stringspace -= offset; return (0); } char * exec_args_get_begin_envv(struct image_args *args) { KASSERT(args->endp != NULL, ("endp not initialized")); if (args->envc > 0) return (args->begin_envv); return (args->endp); } /* * Copy strings out to the new process address space, constructing new arg * and env vector tables. Return a pointer to the base so that it can be used * as the initial stack pointer. */ register_t * exec_copyout_strings(struct image_params *imgp) { int argc, envc; char **vectp; char *stringp; uintptr_t destp; register_t *stack_base; struct ps_strings *arginfo; struct proc *p; size_t execpath_len; int szsigcode, szps; char canary[sizeof(long) * 8]; szps = sizeof(pagesizes[0]) * MAXPAGESIZES; /* * Calculate string base and vector table pointers. * Also deal with signal trampoline code for this exec type. */ if (imgp->execpath != NULL && imgp->auxargs != NULL) execpath_len = strlen(imgp->execpath) + 1; else execpath_len = 0; p = imgp->proc; szsigcode = 0; arginfo = (struct ps_strings *)p->p_sysent->sv_psstrings; if (p->p_sysent->sv_sigcode_base == 0) { if (p->p_sysent->sv_szsigcode != NULL) szsigcode = *(p->p_sysent->sv_szsigcode); } destp = (uintptr_t)arginfo; /* * install sigcode */ if (szsigcode != 0) { destp -= szsigcode; destp = rounddown2(destp, sizeof(void *)); copyout(p->p_sysent->sv_sigcode, (void *)destp, szsigcode); } /* * Copy the image path for the rtld. */ if (execpath_len != 0) { destp -= execpath_len; destp = rounddown2(destp, sizeof(void *)); imgp->execpathp = destp; copyout(imgp->execpath, (void *)destp, execpath_len); } /* * Prepare the canary for SSP. */ arc4rand(canary, sizeof(canary), 0); destp -= sizeof(canary); imgp->canary = destp; copyout(canary, (void *)destp, sizeof(canary)); imgp->canarylen = sizeof(canary); /* * Prepare the pagesizes array. */ destp -= szps; destp = rounddown2(destp, sizeof(void *)); imgp->pagesizes = destp; copyout(pagesizes, (void *)destp, szps); imgp->pagesizeslen = szps; destp -= ARG_MAX - imgp->args->stringspace; destp = rounddown2(destp, sizeof(void *)); vectp = (char **)destp; if (imgp->sysent->sv_stackgap != NULL) imgp->sysent->sv_stackgap(imgp, (u_long *)&vectp); if (imgp->auxargs) { /* * Allocate room on the stack for the ELF auxargs * array. It has up to AT_COUNT entries. */ vectp -= howmany(AT_COUNT * sizeof(Elf_Auxinfo), sizeof(*vectp)); } /* * Allocate room for the argv[] and env vectors including the * terminating NULL pointers. */ vectp -= imgp->args->argc + 1 + imgp->args->envc + 1; /* * vectp also becomes our initial stack base */ stack_base = (register_t *)vectp; stringp = imgp->args->begin_argv; argc = imgp->args->argc; envc = imgp->args->envc; /* * Copy out strings - arguments and environment. */ copyout(stringp, (void *)destp, ARG_MAX - imgp->args->stringspace); /* * Fill in "ps_strings" struct for ps, w, etc. */ suword(&arginfo->ps_argvstr, (long)(intptr_t)vectp); suword32(&arginfo->ps_nargvstr, argc); /* * Fill in argument portion of vector table. */ for (; argc > 0; --argc) { suword(vectp++, (long)(intptr_t)destp); while (*stringp++ != 0) destp++; destp++; } /* a null vector table pointer separates the argp's from the envp's */ suword(vectp++, 0); suword(&arginfo->ps_envstr, (long)(intptr_t)vectp); suword32(&arginfo->ps_nenvstr, envc); /* * Fill in environment portion of vector table. */ for (; envc > 0; --envc) { suword(vectp++, (long)(intptr_t)destp); while (*stringp++ != 0) destp++; destp++; } /* end of vector table is a null pointer */ suword(vectp, 0); return (stack_base); } /* * Check permissions of file to execute. * Called with imgp->vp locked. * Return 0 for success or error code on failure. */ int exec_check_permissions(struct image_params *imgp) { struct vnode *vp = imgp->vp; struct vattr *attr = imgp->attr; struct thread *td; int error; td = curthread; /* Get file attributes */ error = VOP_GETATTR(vp, attr, td->td_ucred); if (error) return (error); #ifdef MAC error = mac_vnode_check_exec(td->td_ucred, imgp->vp, imgp); if (error) return (error); #endif /* * 1) Check if file execution is disabled for the filesystem that * this file resides on. * 2) Ensure that at least one execute bit is on. Otherwise, a * privileged user will always succeed, and we don't want this * to happen unless the file really is executable. * 3) Ensure that the file is a regular file. */ if ((vp->v_mount->mnt_flag & MNT_NOEXEC) || (attr->va_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) == 0 || (attr->va_type != VREG)) return (EACCES); /* * Zero length files can't be exec'd */ if (attr->va_size == 0) return (ENOEXEC); /* * Check for execute permission to file based on current credentials. */ error = VOP_ACCESS(vp, VEXEC, td->td_ucred, td); if (error) return (error); /* * Check number of open-for-writes on the file and deny execution * if there are any. * * Add a text reference now so no one can write to the * executable while we're activating it. * * Remember if this was set before and unset it in case this is not * actually an executable image. */ error = VOP_SET_TEXT(vp); if (error != 0) return (error); imgp->textset = true; /* * Call filesystem specific open routine (which does nothing in the * general case). */ error = VOP_OPEN(vp, FREAD, td->td_ucred, td, NULL); if (error == 0) imgp->opened = 1; return (error); } /* * Exec handler registration */ int exec_register(const struct execsw *execsw_arg) { const struct execsw **es, **xs, **newexecsw; u_int count = 2; /* New slot and trailing NULL */ if (execsw) for (es = execsw; *es; es++) count++; newexecsw = malloc(count * sizeof(*es), M_TEMP, M_WAITOK); xs = newexecsw; if (execsw) for (es = execsw; *es; es++) *xs++ = *es; *xs++ = execsw_arg; *xs = NULL; if (execsw) free(execsw, M_TEMP); execsw = newexecsw; return (0); } int exec_unregister(const struct execsw *execsw_arg) { const struct execsw **es, **xs, **newexecsw; int count = 1; if (execsw == NULL) panic("unregister with no handlers left?\n"); for (es = execsw; *es; es++) { if (*es == execsw_arg) break; } if (*es == NULL) return (ENOENT); for (es = execsw; *es; es++) if (*es != execsw_arg) count++; newexecsw = malloc(count * sizeof(*es), M_TEMP, M_WAITOK); xs = newexecsw; for (es = execsw; *es; es++) if (*es != execsw_arg) *xs++ = *es; *xs = NULL; if (execsw) free(execsw, M_TEMP); execsw = newexecsw; return (0); } Index: head/sys/kern/uipc_shm.c =================================================================== --- head/sys/kern/uipc_shm.c (revision 353534) +++ head/sys/kern/uipc_shm.c (revision 353535) @@ -1,1506 +1,1505 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2006, 2011, 2016-2017 Robert N. M. Watson * All rights reserved. * * Portions of this software were developed by BAE Systems, the University of * Cambridge Computer Laboratory, and Memorial University under DARPA/AFRL * contract FA8650-15-C-7558 ("CADETS"), as part of the DARPA Transparent * Computing (TC) research program. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ /* * Support for shared swap-backed anonymous memory objects via * shm_open(2), shm_rename(2), and shm_unlink(2). * While most of the implementation is here, vm_mmap.c contains * mapping logic changes. * * posixshmcontrol(1) allows users to inspect the state of the memory * objects. Per-uid swap resource limit controls total amount of * memory that user can consume for anonymous objects, including * shared. */ #include __FBSDID("$FreeBSD$"); #include "opt_capsicum.h" #include "opt_ktrace.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include struct shm_mapping { char *sm_path; Fnv32_t sm_fnv; struct shmfd *sm_shmfd; LIST_ENTRY(shm_mapping) sm_link; }; static MALLOC_DEFINE(M_SHMFD, "shmfd", "shared memory file descriptor"); static LIST_HEAD(, shm_mapping) *shm_dictionary; static struct sx shm_dict_lock; static struct mtx shm_timestamp_lock; static u_long shm_hash; static struct unrhdr64 shm_ino_unr; static dev_t shm_dev_ino; #define SHM_HASH(fnv) (&shm_dictionary[(fnv) & shm_hash]) static void shm_init(void *arg); static void shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd); static struct shmfd *shm_lookup(char *path, Fnv32_t fnv); static int shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred); static int shm_dotruncate_locked(struct shmfd *shmfd, off_t length, void *rl_cookie); static fo_rdwr_t shm_read; static fo_rdwr_t shm_write; static fo_truncate_t shm_truncate; static fo_ioctl_t shm_ioctl; static fo_stat_t shm_stat; static fo_close_t shm_close; static fo_chmod_t shm_chmod; static fo_chown_t shm_chown; static fo_seek_t shm_seek; static fo_fill_kinfo_t shm_fill_kinfo; static fo_mmap_t shm_mmap; static fo_get_seals_t shm_get_seals; static fo_add_seals_t shm_add_seals; /* File descriptor operations. */ struct fileops shm_ops = { .fo_read = shm_read, .fo_write = shm_write, .fo_truncate = shm_truncate, .fo_ioctl = shm_ioctl, .fo_poll = invfo_poll, .fo_kqfilter = invfo_kqfilter, .fo_stat = shm_stat, .fo_close = shm_close, .fo_chmod = shm_chmod, .fo_chown = shm_chown, .fo_sendfile = vn_sendfile, .fo_seek = shm_seek, .fo_fill_kinfo = shm_fill_kinfo, .fo_mmap = shm_mmap, .fo_get_seals = shm_get_seals, .fo_add_seals = shm_add_seals, .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE }; FEATURE(posix_shm, "POSIX shared memory"); static int uiomove_object_page(vm_object_t obj, size_t len, struct uio *uio) { vm_page_t m; vm_pindex_t idx; size_t tlen; int error, offset, rv; idx = OFF_TO_IDX(uio->uio_offset); offset = uio->uio_offset & PAGE_MASK; tlen = MIN(PAGE_SIZE - offset, len); VM_OBJECT_WLOCK(obj); /* * Read I/O without either a corresponding resident page or swap * page: use zero_region. This is intended to avoid instantiating * pages on read from a sparse region. */ if (uio->uio_rw == UIO_READ && vm_page_lookup(obj, idx) == NULL && !vm_pager_has_page(obj, idx, NULL, NULL)) { VM_OBJECT_WUNLOCK(obj); return (uiomove(__DECONST(void *, zero_region), tlen, uio)); } /* * Parallel reads of the page content from disk are prevented * by exclusive busy. * * Although the tmpfs vnode lock is held here, it is * nonetheless safe to sleep waiting for a free page. The * pageout daemon does not need to acquire the tmpfs vnode * lock to page out tobj's pages because tobj is a OBJT_SWAP * type object. */ rv = vm_page_grab_valid(&m, obj, idx, VM_ALLOC_NORMAL | VM_ALLOC_WIRED | VM_ALLOC_NOBUSY); if (rv != VM_PAGER_OK) { VM_OBJECT_WUNLOCK(obj); printf("uiomove_object: vm_obj %p idx %jd pager error %d\n", obj, idx, rv); return (EIO); } VM_OBJECT_WUNLOCK(obj); error = uiomove_fromphys(&m, offset, tlen, uio); if (uio->uio_rw == UIO_WRITE && error == 0) { VM_OBJECT_WLOCK(obj); vm_page_dirty(m); vm_pager_page_unswapped(m); VM_OBJECT_WUNLOCK(obj); } vm_page_unwire(m, PQ_ACTIVE); return (error); } int uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio) { ssize_t resid; size_t len; int error; error = 0; while ((resid = uio->uio_resid) > 0) { if (obj_size <= uio->uio_offset) break; len = MIN(obj_size - uio->uio_offset, resid); if (len == 0) break; error = uiomove_object_page(obj, len, uio); if (error != 0 || resid == uio->uio_resid) break; } return (error); } static int shm_seek(struct file *fp, off_t offset, int whence, struct thread *td) { struct shmfd *shmfd; off_t foffset; int error; shmfd = fp->f_data; foffset = foffset_lock(fp, 0); error = 0; switch (whence) { case L_INCR: if (foffset < 0 || (offset > 0 && foffset > OFF_MAX - offset)) { error = EOVERFLOW; break; } offset += foffset; break; case L_XTND: if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) { error = EOVERFLOW; break; } offset += shmfd->shm_size; break; case L_SET: break; default: error = EINVAL; } if (error == 0) { if (offset < 0 || offset > shmfd->shm_size) error = EINVAL; else td->td_uretoff.tdu_off = offset; } foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0); return (error); } static int shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, struct thread *td) { struct shmfd *shmfd; void *rl_cookie; int error; shmfd = fp->f_data; #ifdef MAC error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd); if (error) return (error); #endif foffset_lock_uio(fp, uio, flags); rl_cookie = rangelock_rlock(&shmfd->shm_rl, uio->uio_offset, uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx); error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio); rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); foffset_unlock_uio(fp, uio, flags); return (error); } static int shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, struct thread *td) { struct shmfd *shmfd; void *rl_cookie; int error; shmfd = fp->f_data; #ifdef MAC error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd); if (error) return (error); #endif foffset_lock_uio(fp, uio, flags); if ((flags & FOF_OFFSET) == 0) { rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX, &shmfd->shm_mtx); } else { rl_cookie = rangelock_wlock(&shmfd->shm_rl, uio->uio_offset, uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx); } if ((shmfd->shm_seals & F_SEAL_WRITE) != 0) error = EPERM; else error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio); rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); foffset_unlock_uio(fp, uio, flags); return (error); } static int shm_truncate(struct file *fp, off_t length, struct ucred *active_cred, struct thread *td) { struct shmfd *shmfd; #ifdef MAC int error; #endif shmfd = fp->f_data; #ifdef MAC error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd); if (error) return (error); #endif return (shm_dotruncate(shmfd, length)); } int shm_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred, struct thread *td) { switch (com) { case FIONBIO: case FIOASYNC: /* * Allow fcntl(fd, F_SETFL, O_NONBLOCK) to work, * just like it would on an unlinked regular file */ return (0); default: return (ENOTTY); } } static int shm_stat(struct file *fp, struct stat *sb, struct ucred *active_cred, struct thread *td) { struct shmfd *shmfd; #ifdef MAC int error; #endif shmfd = fp->f_data; #ifdef MAC error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd); if (error) return (error); #endif /* * Attempt to return sanish values for fstat() on a memory file * descriptor. */ bzero(sb, sizeof(*sb)); sb->st_blksize = PAGE_SIZE; sb->st_size = shmfd->shm_size; sb->st_blocks = howmany(sb->st_size, sb->st_blksize); mtx_lock(&shm_timestamp_lock); sb->st_atim = shmfd->shm_atime; sb->st_ctim = shmfd->shm_ctime; sb->st_mtim = shmfd->shm_mtime; sb->st_birthtim = shmfd->shm_birthtime; sb->st_mode = S_IFREG | shmfd->shm_mode; /* XXX */ sb->st_uid = shmfd->shm_uid; sb->st_gid = shmfd->shm_gid; mtx_unlock(&shm_timestamp_lock); sb->st_dev = shm_dev_ino; sb->st_ino = shmfd->shm_ino; sb->st_nlink = shmfd->shm_object->ref_count; return (0); } static int shm_close(struct file *fp, struct thread *td) { struct shmfd *shmfd; shmfd = fp->f_data; fp->f_data = NULL; shm_drop(shmfd); return (0); } static int shm_dotruncate_locked(struct shmfd *shmfd, off_t length, void *rl_cookie) { vm_object_t object; vm_page_t m; vm_pindex_t idx, nobjsize; vm_ooffset_t delta; int base, rv; KASSERT(length >= 0, ("shm_dotruncate: length < 0")); object = shmfd->shm_object; VM_OBJECT_ASSERT_WLOCKED(object); rangelock_cookie_assert(rl_cookie, RA_WLOCKED); if (length == shmfd->shm_size) return (0); nobjsize = OFF_TO_IDX(length + PAGE_MASK); /* Are we shrinking? If so, trim the end. */ if (length < shmfd->shm_size) { if ((shmfd->shm_seals & F_SEAL_SHRINK) != 0) return (EPERM); /* * Disallow any requests to shrink the size if this * object is mapped into the kernel. */ if (shmfd->shm_kmappings > 0) return (EBUSY); /* * Zero the truncated part of the last page. */ base = length & PAGE_MASK; if (base != 0) { idx = OFF_TO_IDX(length); retry: - m = vm_page_lookup(object, idx); + m = vm_page_grab(object, idx, VM_ALLOC_NOCREAT); if (m != NULL) { - if (vm_page_sleep_if_busy(m, "shmtrc")) - goto retry; + MPASS(m->valid == VM_PAGE_BITS_ALL); } else if (vm_pager_has_page(object, idx, NULL, NULL)) { m = vm_page_alloc(object, idx, VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL); if (m == NULL) goto retry; rv = vm_pager_get_pages(object, &m, 1, NULL, NULL); if (rv == VM_PAGER_OK) { /* * Since the page was not resident, * and therefore not recently * accessed, immediately enqueue it * for asynchronous laundering. The * current operation is not regarded * as an access. */ vm_page_launder(m); - vm_page_xunbusy(m); } else { vm_page_free(m); VM_OBJECT_WUNLOCK(object); return (EIO); } } if (m != NULL) { pmap_zero_page_area(m, base, PAGE_SIZE - base); KASSERT(m->valid == VM_PAGE_BITS_ALL, ("shm_dotruncate: page %p is invalid", m)); vm_page_dirty(m); + vm_page_xunbusy(m); vm_pager_page_unswapped(m); } } delta = IDX_TO_OFF(object->size - nobjsize); /* Toss in memory pages. */ if (nobjsize < object->size) vm_object_page_remove(object, nobjsize, object->size, 0); /* Toss pages from swap. */ if (object->type == OBJT_SWAP) swap_pager_freespace(object, nobjsize, delta); /* Free the swap accounted for shm */ swap_release_by_cred(delta, object->cred); object->charge -= delta; } else { if ((shmfd->shm_seals & F_SEAL_GROW) != 0) return (EPERM); /* Try to reserve additional swap space. */ delta = IDX_TO_OFF(nobjsize - object->size); if (!swap_reserve_by_cred(delta, object->cred)) return (ENOMEM); object->charge += delta; } shmfd->shm_size = length; mtx_lock(&shm_timestamp_lock); vfs_timestamp(&shmfd->shm_ctime); shmfd->shm_mtime = shmfd->shm_ctime; mtx_unlock(&shm_timestamp_lock); object->size = nobjsize; return (0); } int shm_dotruncate(struct shmfd *shmfd, off_t length) { void *rl_cookie; int error; rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX, &shmfd->shm_mtx); VM_OBJECT_WLOCK(shmfd->shm_object); error = shm_dotruncate_locked(shmfd, length, rl_cookie); VM_OBJECT_WUNLOCK(shmfd->shm_object); rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); return (error); } /* * shmfd object management including creation and reference counting * routines. */ struct shmfd * shm_alloc(struct ucred *ucred, mode_t mode) { struct shmfd *shmfd; shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO); shmfd->shm_size = 0; shmfd->shm_uid = ucred->cr_uid; shmfd->shm_gid = ucred->cr_gid; shmfd->shm_mode = mode; shmfd->shm_object = vm_pager_allocate(OBJT_SWAP, NULL, shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred); KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate")); shmfd->shm_object->pg_color = 0; VM_OBJECT_WLOCK(shmfd->shm_object); vm_object_clear_flag(shmfd->shm_object, OBJ_ONEMAPPING); vm_object_set_flag(shmfd->shm_object, OBJ_COLORED | OBJ_NOSPLIT); VM_OBJECT_WUNLOCK(shmfd->shm_object); vfs_timestamp(&shmfd->shm_birthtime); shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime = shmfd->shm_birthtime; shmfd->shm_ino = alloc_unr64(&shm_ino_unr); refcount_init(&shmfd->shm_refs, 1); mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF); rangelock_init(&shmfd->shm_rl); #ifdef MAC mac_posixshm_init(shmfd); mac_posixshm_create(ucred, shmfd); #endif return (shmfd); } struct shmfd * shm_hold(struct shmfd *shmfd) { refcount_acquire(&shmfd->shm_refs); return (shmfd); } void shm_drop(struct shmfd *shmfd) { if (refcount_release(&shmfd->shm_refs)) { #ifdef MAC mac_posixshm_destroy(shmfd); #endif rangelock_destroy(&shmfd->shm_rl); mtx_destroy(&shmfd->shm_mtx); vm_object_deallocate(shmfd->shm_object); free(shmfd, M_SHMFD); } } /* * Determine if the credentials have sufficient permissions for a * specified combination of FREAD and FWRITE. */ int shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags) { accmode_t accmode; int error; accmode = 0; if (flags & FREAD) accmode |= VREAD; if (flags & FWRITE) accmode |= VWRITE; mtx_lock(&shm_timestamp_lock); error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid, accmode, ucred, NULL); mtx_unlock(&shm_timestamp_lock); return (error); } /* * Dictionary management. We maintain an in-kernel dictionary to map * paths to shmfd objects. We use the FNV hash on the path to store * the mappings in a hash table. */ static void shm_init(void *arg) { mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF); sx_init(&shm_dict_lock, "shm dictionary"); shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash); new_unrhdr64(&shm_ino_unr, 1); shm_dev_ino = devfs_alloc_cdp_inode(); KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized")); } SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL); static struct shmfd * shm_lookup(char *path, Fnv32_t fnv) { struct shm_mapping *map; LIST_FOREACH(map, SHM_HASH(fnv), sm_link) { if (map->sm_fnv != fnv) continue; if (strcmp(map->sm_path, path) == 0) return (map->sm_shmfd); } return (NULL); } static void shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd) { struct shm_mapping *map; map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK); map->sm_path = path; map->sm_fnv = fnv; map->sm_shmfd = shm_hold(shmfd); shmfd->shm_path = path; LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link); } static int shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred) { struct shm_mapping *map; int error; LIST_FOREACH(map, SHM_HASH(fnv), sm_link) { if (map->sm_fnv != fnv) continue; if (strcmp(map->sm_path, path) == 0) { #ifdef MAC error = mac_posixshm_check_unlink(ucred, map->sm_shmfd); if (error) return (error); #endif error = shm_access(map->sm_shmfd, ucred, FREAD | FWRITE); if (error) return (error); map->sm_shmfd->shm_path = NULL; LIST_REMOVE(map, sm_link); shm_drop(map->sm_shmfd); free(map->sm_path, M_SHMFD); free(map, M_SHMFD); return (0); } } return (ENOENT); } int kern_shm_open(struct thread *td, const char *userpath, int flags, mode_t mode, struct filecaps *fcaps, int initial_seals) { struct filedesc *fdp; struct shmfd *shmfd; struct file *fp; char *path; const char *pr_path; void *rl_cookie; size_t pr_pathlen; Fnv32_t fnv; mode_t cmode; int fd, error; #ifdef CAPABILITY_MODE /* * shm_open(2) is only allowed for anonymous objects. */ if (IN_CAPABILITY_MODE(td) && (userpath != SHM_ANON)) return (ECAPMODE); #endif AUDIT_ARG_FFLAGS(flags); AUDIT_ARG_MODE(mode); if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR) return (EINVAL); if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0) return (EINVAL); /* * Currently only F_SEAL_SEAL may be set when creating or opening shmfd. * If the decision is made later to allow additional seals, care must be * taken below to ensure that the seals are properly set if the shmfd * already existed -- this currently assumes that only F_SEAL_SEAL can * be set and doesn't take further precautions to ensure the validity of * the seals being added with respect to current mappings. */ if ((initial_seals & ~F_SEAL_SEAL) != 0) return (EINVAL); fdp = td->td_proc->p_fd; cmode = (mode & ~fdp->fd_cmask) & ACCESSPERMS; /* * shm_open(2) created shm should always have O_CLOEXEC set, as mandated * by POSIX. We allow it to be unset here so that an in-kernel * interface may be written as a thin layer around shm, optionally not * setting CLOEXEC. For shm_open(2), O_CLOEXEC is set unconditionally * in sys_shm_open() to keep this implementation compliant. */ error = falloc_caps(td, &fp, &fd, flags & O_CLOEXEC, fcaps); if (error) return (error); /* A SHM_ANON path pointer creates an anonymous object. */ if (userpath == SHM_ANON) { /* A read-only anonymous object is pointless. */ if ((flags & O_ACCMODE) == O_RDONLY) { fdclose(td, fp, fd); fdrop(fp, td); return (EINVAL); } shmfd = shm_alloc(td->td_ucred, cmode); shmfd->shm_seals = initial_seals; } else { path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK); pr_path = td->td_ucred->cr_prison->pr_path; /* Construct a full pathname for jailed callers. */ pr_pathlen = strcmp(pr_path, "/") == 0 ? 0 : strlcpy(path, pr_path, MAXPATHLEN); error = copyinstr(userpath, path + pr_pathlen, MAXPATHLEN - pr_pathlen, NULL); #ifdef KTRACE if (error == 0 && KTRPOINT(curthread, KTR_NAMEI)) ktrnamei(path); #endif /* Require paths to start with a '/' character. */ if (error == 0 && path[pr_pathlen] != '/') error = EINVAL; if (error) { fdclose(td, fp, fd); fdrop(fp, td); free(path, M_SHMFD); return (error); } AUDIT_ARG_UPATH1_CANON(path); fnv = fnv_32_str(path, FNV1_32_INIT); sx_xlock(&shm_dict_lock); shmfd = shm_lookup(path, fnv); if (shmfd == NULL) { /* Object does not yet exist, create it if requested. */ if (flags & O_CREAT) { #ifdef MAC error = mac_posixshm_check_create(td->td_ucred, path); if (error == 0) { #endif shmfd = shm_alloc(td->td_ucred, cmode); shmfd->shm_seals = initial_seals; shm_insert(path, fnv, shmfd); #ifdef MAC } #endif } else { free(path, M_SHMFD); error = ENOENT; } } else { rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX, &shmfd->shm_mtx); /* * kern_shm_open() likely shouldn't ever error out on * trying to set a seal that already exists, unlike * F_ADD_SEALS. This would break terribly as * shm_open(2) actually sets F_SEAL_SEAL to maintain * historical behavior where the underlying file could * not be sealed. */ initial_seals &= ~shmfd->shm_seals; /* * Object already exists, obtain a new * reference if requested and permitted. */ free(path, M_SHMFD); /* * initial_seals can't set additional seals if we've * already been set F_SEAL_SEAL. If F_SEAL_SEAL is set, * then we've already removed that one from * initial_seals. This is currently redundant as we * only allow setting F_SEAL_SEAL at creation time, but * it's cheap to check and decreases the effort required * to allow additional seals. */ if ((shmfd->shm_seals & F_SEAL_SEAL) != 0 && initial_seals != 0) error = EPERM; else if ((flags & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL)) error = EEXIST; else { #ifdef MAC error = mac_posixshm_check_open(td->td_ucred, shmfd, FFLAGS(flags & O_ACCMODE)); if (error == 0) #endif error = shm_access(shmfd, td->td_ucred, FFLAGS(flags & O_ACCMODE)); } /* * Truncate the file back to zero length if * O_TRUNC was specified and the object was * opened with read/write. */ if (error == 0 && (flags & (O_ACCMODE | O_TRUNC)) == (O_RDWR | O_TRUNC)) { VM_OBJECT_WLOCK(shmfd->shm_object); #ifdef MAC error = mac_posixshm_check_truncate( td->td_ucred, fp->f_cred, shmfd); if (error == 0) #endif error = shm_dotruncate_locked(shmfd, 0, rl_cookie); VM_OBJECT_WUNLOCK(shmfd->shm_object); } if (error == 0) { /* * Currently we only allow F_SEAL_SEAL to be * set initially. As noted above, this would * need to be reworked should that change. */ shmfd->shm_seals |= initial_seals; shm_hold(shmfd); } rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); } sx_xunlock(&shm_dict_lock); if (error) { fdclose(td, fp, fd); fdrop(fp, td); return (error); } } finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops); td->td_retval[0] = fd; fdrop(fp, td); return (0); } /* System calls. */ #ifdef COMPAT_FREEBSD12 int freebsd12_shm_open(struct thread *td, struct freebsd12_shm_open_args *uap) { return (kern_shm_open(td, uap->path, uap->flags | O_CLOEXEC, uap->mode, NULL, F_SEAL_SEAL)); } #endif int sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap) { char *path; const char *pr_path; size_t pr_pathlen; Fnv32_t fnv; int error; path = malloc(MAXPATHLEN, M_TEMP, M_WAITOK); pr_path = td->td_ucred->cr_prison->pr_path; pr_pathlen = strcmp(pr_path, "/") == 0 ? 0 : strlcpy(path, pr_path, MAXPATHLEN); error = copyinstr(uap->path, path + pr_pathlen, MAXPATHLEN - pr_pathlen, NULL); if (error) { free(path, M_TEMP); return (error); } #ifdef KTRACE if (KTRPOINT(curthread, KTR_NAMEI)) ktrnamei(path); #endif AUDIT_ARG_UPATH1_CANON(path); fnv = fnv_32_str(path, FNV1_32_INIT); sx_xlock(&shm_dict_lock); error = shm_remove(path, fnv, td->td_ucred); sx_xunlock(&shm_dict_lock); free(path, M_TEMP); return (error); } int sys_shm_rename(struct thread *td, struct shm_rename_args *uap) { char *path_from = NULL, *path_to = NULL; Fnv32_t fnv_from, fnv_to; struct shmfd *fd_from; struct shmfd *fd_to; int error; int flags; flags = uap->flags; /* * Make sure the user passed only valid flags. * If you add a new flag, please add a new term here. */ if ((flags & ~( SHM_RENAME_NOREPLACE | SHM_RENAME_EXCHANGE )) != 0) { error = EINVAL; goto out; } /* * EXCHANGE and NOREPLACE don't quite make sense together. Let's * force the user to choose one or the other. */ if ((flags & SHM_RENAME_NOREPLACE) != 0 && (flags & SHM_RENAME_EXCHANGE) != 0) { error = EINVAL; goto out; } /* * Malloc zone M_SHMFD, since this path may end up freed later from * M_SHMFD if we end up doing an insert. */ path_from = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK); error = copyinstr(uap->path_from, path_from, MAXPATHLEN, NULL); if (error) goto out; path_to = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK); error = copyinstr(uap->path_to, path_to, MAXPATHLEN, NULL); if (error) goto out; /* Rename with from/to equal is a no-op */ if (strncmp(path_from, path_to, MAXPATHLEN) == 0) goto out; fnv_from = fnv_32_str(path_from, FNV1_32_INIT); fnv_to = fnv_32_str(path_to, FNV1_32_INIT); sx_xlock(&shm_dict_lock); fd_from = shm_lookup(path_from, fnv_from); if (fd_from == NULL) { sx_xunlock(&shm_dict_lock); error = ENOENT; goto out; } fd_to = shm_lookup(path_to, fnv_to); if ((flags & SHM_RENAME_NOREPLACE) != 0 && fd_to != NULL) { sx_xunlock(&shm_dict_lock); error = EEXIST; goto out; } /* * Unconditionally prevents shm_remove from invalidating the 'from' * shm's state. */ shm_hold(fd_from); error = shm_remove(path_from, fnv_from, td->td_ucred); /* * One of my assumptions failed if ENOENT (e.g. locking didn't * protect us) */ KASSERT(error != ENOENT, ("Our shm disappeared during shm_rename: %s", path_from)); if (error) { shm_drop(fd_from); sx_xunlock(&shm_dict_lock); goto out; } /* * If we are exchanging, we need to ensure the shm_remove below * doesn't invalidate the dest shm's state. */ if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) shm_hold(fd_to); /* * NOTE: if path_to is not already in the hash, c'est la vie; * it simply means we have nothing already at path_to to unlink. * That is the ENOENT case. * * If we somehow don't have access to unlink this guy, but * did for the shm at path_from, then relink the shm to path_from * and abort with EACCES. * * All other errors: that is weird; let's relink and abort the * operation. */ error = shm_remove(path_to, fnv_to, td->td_ucred); if (error && error != ENOENT) { shm_insert(path_from, fnv_from, fd_from); shm_drop(fd_from); /* Don't free path_from now, since the hash references it */ path_from = NULL; sx_xunlock(&shm_dict_lock); goto out; } shm_insert(path_to, fnv_to, fd_from); /* Don't free path_to now, since the hash references it */ path_to = NULL; /* We kept a ref when we removed, and incremented again in insert */ shm_drop(fd_from); #ifdef DEBUG KASSERT(fd_from->shm_refs > 0, ("Expected >0 refs; got: %d\n", fd_from->shm_refs)); #endif if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) { shm_insert(path_from, fnv_from, fd_to); path_from = NULL; shm_drop(fd_to); #ifdef DEBUG KASSERT(fd_to->shm_refs > 0, ("Expected >0 refs; got: %d\n", fd_to->shm_refs)); #endif } error = 0; sx_xunlock(&shm_dict_lock); out: if (path_from != NULL) free(path_from, M_SHMFD); if (path_to != NULL) free(path_to, M_SHMFD); return(error); } int shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize, vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff, struct thread *td) { struct shmfd *shmfd; vm_prot_t maxprot; int error; bool writecnt; void *rl_cookie; shmfd = fp->f_data; maxprot = VM_PROT_NONE; rl_cookie = rangelock_rlock(&shmfd->shm_rl, 0, objsize, &shmfd->shm_mtx); /* FREAD should always be set. */ if ((fp->f_flag & FREAD) != 0) maxprot |= VM_PROT_EXECUTE | VM_PROT_READ; if ((fp->f_flag & FWRITE) != 0) maxprot |= VM_PROT_WRITE; writecnt = (flags & MAP_SHARED) != 0 && (prot & VM_PROT_WRITE) != 0; if (writecnt && (shmfd->shm_seals & F_SEAL_WRITE) != 0) { error = EPERM; goto out; } /* Don't permit shared writable mappings on read-only descriptors. */ if (writecnt && (maxprot & VM_PROT_WRITE) == 0) { error = EACCES; goto out; } maxprot &= cap_maxprot; /* See comment in vn_mmap(). */ if ( #ifdef _LP64 objsize > OFF_MAX || #endif foff < 0 || foff > OFF_MAX - objsize) { error = EINVAL; goto out; } #ifdef MAC error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags); if (error != 0) goto out; #endif mtx_lock(&shm_timestamp_lock); vfs_timestamp(&shmfd->shm_atime); mtx_unlock(&shm_timestamp_lock); vm_object_reference(shmfd->shm_object); if (writecnt) vm_pager_update_writecount(shmfd->shm_object, 0, objsize); error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags, shmfd->shm_object, foff, writecnt, td); if (error != 0) { if (writecnt) vm_pager_release_writecount(shmfd->shm_object, 0, objsize); vm_object_deallocate(shmfd->shm_object); } out: rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); return (error); } static int shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td) { struct shmfd *shmfd; int error; error = 0; shmfd = fp->f_data; mtx_lock(&shm_timestamp_lock); /* * SUSv4 says that x bits of permission need not be affected. * Be consistent with our shm_open there. */ #ifdef MAC error = mac_posixshm_check_setmode(active_cred, shmfd, mode); if (error != 0) goto out; #endif error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid, VADMIN, active_cred, NULL); if (error != 0) goto out; shmfd->shm_mode = mode & ACCESSPERMS; out: mtx_unlock(&shm_timestamp_lock); return (error); } static int shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred, struct thread *td) { struct shmfd *shmfd; int error; error = 0; shmfd = fp->f_data; mtx_lock(&shm_timestamp_lock); #ifdef MAC error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid); if (error != 0) goto out; #endif if (uid == (uid_t)-1) uid = shmfd->shm_uid; if (gid == (gid_t)-1) gid = shmfd->shm_gid; if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) || (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) && (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN))) goto out; shmfd->shm_uid = uid; shmfd->shm_gid = gid; out: mtx_unlock(&shm_timestamp_lock); return (error); } /* * Helper routines to allow the backing object of a shared memory file * descriptor to be mapped in the kernel. */ int shm_map(struct file *fp, size_t size, off_t offset, void **memp) { struct shmfd *shmfd; vm_offset_t kva, ofs; vm_object_t obj; int rv; if (fp->f_type != DTYPE_SHM) return (EINVAL); shmfd = fp->f_data; obj = shmfd->shm_object; VM_OBJECT_WLOCK(obj); /* * XXXRW: This validation is probably insufficient, and subject to * sign errors. It should be fixed. */ if (offset >= shmfd->shm_size || offset + size > round_page(shmfd->shm_size)) { VM_OBJECT_WUNLOCK(obj); return (EINVAL); } shmfd->shm_kmappings++; vm_object_reference_locked(obj); VM_OBJECT_WUNLOCK(obj); /* Map the object into the kernel_map and wire it. */ kva = vm_map_min(kernel_map); ofs = offset & PAGE_MASK; offset = trunc_page(offset); size = round_page(size + ofs); rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0, VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE, VM_PROT_READ | VM_PROT_WRITE, 0); if (rv == KERN_SUCCESS) { rv = vm_map_wire(kernel_map, kva, kva + size, VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES); if (rv == KERN_SUCCESS) { *memp = (void *)(kva + ofs); return (0); } vm_map_remove(kernel_map, kva, kva + size); } else vm_object_deallocate(obj); /* On failure, drop our mapping reference. */ VM_OBJECT_WLOCK(obj); shmfd->shm_kmappings--; VM_OBJECT_WUNLOCK(obj); return (vm_mmap_to_errno(rv)); } /* * We require the caller to unmap the entire entry. This allows us to * safely decrement shm_kmappings when a mapping is removed. */ int shm_unmap(struct file *fp, void *mem, size_t size) { struct shmfd *shmfd; vm_map_entry_t entry; vm_offset_t kva, ofs; vm_object_t obj; vm_pindex_t pindex; vm_prot_t prot; boolean_t wired; vm_map_t map; int rv; if (fp->f_type != DTYPE_SHM) return (EINVAL); shmfd = fp->f_data; kva = (vm_offset_t)mem; ofs = kva & PAGE_MASK; kva = trunc_page(kva); size = round_page(size + ofs); map = kernel_map; rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry, &obj, &pindex, &prot, &wired); if (rv != KERN_SUCCESS) return (EINVAL); if (entry->start != kva || entry->end != kva + size) { vm_map_lookup_done(map, entry); return (EINVAL); } vm_map_lookup_done(map, entry); if (obj != shmfd->shm_object) return (EINVAL); vm_map_remove(map, kva, kva + size); VM_OBJECT_WLOCK(obj); KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped")); shmfd->shm_kmappings--; VM_OBJECT_WUNLOCK(obj); return (0); } static int shm_fill_kinfo_locked(struct shmfd *shmfd, struct kinfo_file *kif, bool list) { const char *path, *pr_path; size_t pr_pathlen; bool visible; sx_assert(&shm_dict_lock, SA_LOCKED); kif->kf_type = KF_TYPE_SHM; kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode; kif->kf_un.kf_file.kf_file_size = shmfd->shm_size; if (shmfd->shm_path != NULL) { if (shmfd->shm_path != NULL) { path = shmfd->shm_path; pr_path = curthread->td_ucred->cr_prison->pr_path; if (strcmp(pr_path, "/") != 0) { /* Return the jail-rooted pathname. */ pr_pathlen = strlen(pr_path); visible = strncmp(path, pr_path, pr_pathlen) == 0 && path[pr_pathlen] == '/'; if (list && !visible) return (EPERM); if (visible) path += pr_pathlen; } strlcpy(kif->kf_path, path, sizeof(kif->kf_path)); } } return (0); } static int shm_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp __unused) { int res; sx_slock(&shm_dict_lock); res = shm_fill_kinfo_locked(fp->f_data, kif, false); sx_sunlock(&shm_dict_lock); return (res); } static int shm_add_seals(struct file *fp, int seals) { struct shmfd *shmfd; void *rl_cookie; vm_ooffset_t writemappings; int error, nseals; error = 0; shmfd = fp->f_data; rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX, &shmfd->shm_mtx); /* Even already-set seals should result in EPERM. */ if ((shmfd->shm_seals & F_SEAL_SEAL) != 0) { error = EPERM; goto out; } nseals = seals & ~shmfd->shm_seals; if ((nseals & F_SEAL_WRITE) != 0) { /* * The rangelock above prevents writable mappings from being * added after we've started applying seals. The RLOCK here * is to avoid torn reads on ILP32 arches as unmapping/reducing * writemappings will be done without a rangelock. */ VM_OBJECT_RLOCK(shmfd->shm_object); writemappings = shmfd->shm_object->un_pager.swp.writemappings; VM_OBJECT_RUNLOCK(shmfd->shm_object); /* kmappings are also writable */ if (writemappings > 0) { error = EBUSY; goto out; } } shmfd->shm_seals |= nseals; out: rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); return (error); } static int shm_get_seals(struct file *fp, int *seals) { struct shmfd *shmfd; shmfd = fp->f_data; *seals = shmfd->shm_seals; return (0); } static int sysctl_posix_shm_list(SYSCTL_HANDLER_ARGS) { struct shm_mapping *shmm; struct sbuf sb; struct kinfo_file kif; u_long i; ssize_t curlen; int error, error2; sbuf_new_for_sysctl(&sb, NULL, sizeof(struct kinfo_file) * 5, req); sbuf_clear_flags(&sb, SBUF_INCLUDENUL); curlen = 0; error = 0; sx_slock(&shm_dict_lock); for (i = 0; i < shm_hash + 1; i++) { LIST_FOREACH(shmm, &shm_dictionary[i], sm_link) { error = shm_fill_kinfo_locked(shmm->sm_shmfd, &kif, true); if (error == EPERM) continue; if (error != 0) break; pack_kinfo(&kif); if (req->oldptr != NULL && kif.kf_structsize + curlen > req->oldlen) break; error = sbuf_bcat(&sb, &kif, kif.kf_structsize) == 0 ? 0 : ENOMEM; if (error != 0) break; curlen += kif.kf_structsize; } } sx_sunlock(&shm_dict_lock); error2 = sbuf_finish(&sb); sbuf_delete(&sb); return (error != 0 ? error : error2); } SYSCTL_PROC(_kern_ipc, OID_AUTO, posix_shm_list, CTLFLAG_RD | CTLFLAG_MPSAFE | CTLTYPE_OPAQUE, NULL, 0, sysctl_posix_shm_list, "", "POSIX SHM list"); int kern_shm_open2(struct thread *td, const char *path, int flags, mode_t mode, int shmflags, const char *name __unused) { int initial_seals; if ((shmflags & ~SHM_ALLOW_SEALING) != 0) return (EINVAL); initial_seals = F_SEAL_SEAL; if ((shmflags & SHM_ALLOW_SEALING) != 0) initial_seals &= ~F_SEAL_SEAL; return (kern_shm_open(td, path, flags, mode, NULL, initial_seals)); } /* * This version of the shm_open() interface leaves CLOEXEC behavior up to the * caller, and libc will enforce it for the traditional shm_open() call. This * allows other consumers, like memfd_create(), to opt-in for CLOEXEC. This * interface also includes a 'name' argument that is currently unused, but could * potentially be exported later via some interface for debugging purposes. * From the kernel's perspective, it is optional. Individual consumers like * memfd_create() may require it in order to be compatible with other systems * implementing the same function. */ int sys_shm_open2(struct thread *td, struct shm_open2_args *uap) { return (kern_shm_open2(td, uap->path, uap->flags, uap->mode, uap->shmflags, uap->name)); } Index: head/sys/kern/vfs_bio.c =================================================================== --- head/sys/kern/vfs_bio.c (revision 353534) +++ head/sys/kern/vfs_bio.c (revision 353535) @@ -1,5465 +1,5463 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2004 Poul-Henning Kamp * Copyright (c) 1994,1997 John S. Dyson * Copyright (c) 2013 The FreeBSD Foundation * All rights reserved. * * Portions of this software were developed by Konstantin Belousov * under sponsorship from the FreeBSD Foundation. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ /* * this file contains a new buffer I/O scheme implementing a coherent * VM object and buffer cache scheme. Pains have been taken to make * sure that the performance degradation associated with schemes such * as this is not realized. * * Author: John S. Dyson * Significant help during the development and debugging phases * had been provided by David Greenman, also of the FreeBSD core team. * * see man buf(9) for more info. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer"); struct bio_ops bioops; /* I/O operation notification */ struct buf_ops buf_ops_bio = { .bop_name = "buf_ops_bio", .bop_write = bufwrite, .bop_strategy = bufstrategy, .bop_sync = bufsync, .bop_bdflush = bufbdflush, }; struct bufqueue { struct mtx_padalign bq_lock; TAILQ_HEAD(, buf) bq_queue; uint8_t bq_index; uint16_t bq_subqueue; int bq_len; } __aligned(CACHE_LINE_SIZE); #define BQ_LOCKPTR(bq) (&(bq)->bq_lock) #define BQ_LOCK(bq) mtx_lock(BQ_LOCKPTR((bq))) #define BQ_UNLOCK(bq) mtx_unlock(BQ_LOCKPTR((bq))) #define BQ_ASSERT_LOCKED(bq) mtx_assert(BQ_LOCKPTR((bq)), MA_OWNED) struct bufdomain { struct bufqueue bd_subq[MAXCPU + 1]; /* Per-cpu sub queues + global */ struct bufqueue bd_dirtyq; struct bufqueue *bd_cleanq; struct mtx_padalign bd_run_lock; /* Constants */ long bd_maxbufspace; long bd_hibufspace; long bd_lobufspace; long bd_bufspacethresh; int bd_hifreebuffers; int bd_lofreebuffers; int bd_hidirtybuffers; int bd_lodirtybuffers; int bd_dirtybufthresh; int bd_lim; /* atomics */ int bd_wanted; int __aligned(CACHE_LINE_SIZE) bd_numdirtybuffers; int __aligned(CACHE_LINE_SIZE) bd_running; long __aligned(CACHE_LINE_SIZE) bd_bufspace; int __aligned(CACHE_LINE_SIZE) bd_freebuffers; } __aligned(CACHE_LINE_SIZE); #define BD_LOCKPTR(bd) (&(bd)->bd_cleanq->bq_lock) #define BD_LOCK(bd) mtx_lock(BD_LOCKPTR((bd))) #define BD_UNLOCK(bd) mtx_unlock(BD_LOCKPTR((bd))) #define BD_ASSERT_LOCKED(bd) mtx_assert(BD_LOCKPTR((bd)), MA_OWNED) #define BD_RUN_LOCKPTR(bd) (&(bd)->bd_run_lock) #define BD_RUN_LOCK(bd) mtx_lock(BD_RUN_LOCKPTR((bd))) #define BD_RUN_UNLOCK(bd) mtx_unlock(BD_RUN_LOCKPTR((bd))) #define BD_DOMAIN(bd) (bd - bdomain) static struct buf *buf; /* buffer header pool */ extern struct buf *swbuf; /* Swap buffer header pool. */ caddr_t unmapped_buf; /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */ struct proc *bufdaemonproc; static int inmem(struct vnode *vp, daddr_t blkno); static void vm_hold_free_pages(struct buf *bp, int newbsize); static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to); static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m); static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m); static void vfs_clean_pages_dirty_buf(struct buf *bp); static void vfs_setdirty_locked_object(struct buf *bp); static void vfs_vmio_invalidate(struct buf *bp); static void vfs_vmio_truncate(struct buf *bp, int npages); static void vfs_vmio_extend(struct buf *bp, int npages, int size); static int vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno); static void breada(struct vnode *, daddr_t *, int *, int, struct ucred *, int, void (*)(struct buf *)); static int buf_flush(struct vnode *vp, struct bufdomain *, int); static int flushbufqueues(struct vnode *, struct bufdomain *, int, int); static void buf_daemon(void); static __inline void bd_wakeup(void); static int sysctl_runningspace(SYSCTL_HANDLER_ARGS); static void bufkva_reclaim(vmem_t *, int); static void bufkva_free(struct buf *); static int buf_import(void *, void **, int, int, int); static void buf_release(void *, void **, int); static void maxbcachebuf_adjust(void); static inline struct bufdomain *bufdomain(struct buf *); static void bq_remove(struct bufqueue *bq, struct buf *bp); static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock); static int buf_recycle(struct bufdomain *, bool kva); static void bq_init(struct bufqueue *bq, int qindex, int cpu, const char *lockname); static void bd_init(struct bufdomain *bd); static int bd_flushall(struct bufdomain *bd); static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS); static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS); static int sysctl_bufspace(SYSCTL_HANDLER_ARGS); int vmiodirenable = TRUE; SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, "Use the VM system for directory writes"); long runningbufspace; SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, "Amount of presently outstanding async buffer io"); SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_bufspace, "L", "Physical memory used for buffers"); static counter_u64_t bufkvaspace; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace, "Kernel virtual memory used for buffers"); static long maxbufspace; SYSCTL_PROC(_vfs, OID_AUTO, maxbufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &maxbufspace, __offsetof(struct bufdomain, bd_maxbufspace), sysctl_bufdomain_long, "L", "Maximum allowed value of bufspace (including metadata)"); static long bufmallocspace; SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, "Amount of malloced memory for buffers"); static long maxbufmallocspace; SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, "Maximum amount of malloced memory for buffers"); static long lobufspace; SYSCTL_PROC(_vfs, OID_AUTO, lobufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lobufspace, __offsetof(struct bufdomain, bd_lobufspace), sysctl_bufdomain_long, "L", "Minimum amount of buffers we want to have"); long hibufspace; SYSCTL_PROC(_vfs, OID_AUTO, hibufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hibufspace, __offsetof(struct bufdomain, bd_hibufspace), sysctl_bufdomain_long, "L", "Maximum allowed value of bufspace (excluding metadata)"); long bufspacethresh; SYSCTL_PROC(_vfs, OID_AUTO, bufspacethresh, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &bufspacethresh, __offsetof(struct bufdomain, bd_bufspacethresh), sysctl_bufdomain_long, "L", "Bufspace consumed before waking the daemon to free some"); static counter_u64_t buffreekvacnt; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, "Number of times we have freed the KVA space from some buffer"); static counter_u64_t bufdefragcnt; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, "Number of times we have had to repeat buffer allocation to defragment"); static long lorunningspace; SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE | CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L", "Minimum preferred space used for in-progress I/O"); static long hirunningspace; SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE | CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L", "Maximum amount of space to use for in-progress I/O"); int dirtybufferflushes; SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); int bdwriteskip; SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip, 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk"); int altbufferflushes; SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW | CTLFLAG_STATS, &altbufferflushes, 0, "Number of fsync flushes to limit dirty buffers"); static int recursiveflushes; SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW | CTLFLAG_STATS, &recursiveflushes, 0, "Number of flushes skipped due to being recursive"); static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS); SYSCTL_PROC(_vfs, OID_AUTO, numdirtybuffers, CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_numdirtybuffers, "I", "Number of buffers that are dirty (has unwritten changes) at the moment"); static int lodirtybuffers; SYSCTL_PROC(_vfs, OID_AUTO, lodirtybuffers, CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lodirtybuffers, __offsetof(struct bufdomain, bd_lodirtybuffers), sysctl_bufdomain_int, "I", "How many buffers we want to have free before bufdaemon can sleep"); static int hidirtybuffers; SYSCTL_PROC(_vfs, OID_AUTO, hidirtybuffers, CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hidirtybuffers, __offsetof(struct bufdomain, bd_hidirtybuffers), sysctl_bufdomain_int, "I", "When the number of dirty buffers is considered severe"); int dirtybufthresh; SYSCTL_PROC(_vfs, OID_AUTO, dirtybufthresh, CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &dirtybufthresh, __offsetof(struct bufdomain, bd_dirtybufthresh), sysctl_bufdomain_int, "I", "Number of bdwrite to bawrite conversions to clear dirty buffers"); static int numfreebuffers; SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, "Number of free buffers"); static int lofreebuffers; SYSCTL_PROC(_vfs, OID_AUTO, lofreebuffers, CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lofreebuffers, __offsetof(struct bufdomain, bd_lofreebuffers), sysctl_bufdomain_int, "I", "Target number of free buffers"); static int hifreebuffers; SYSCTL_PROC(_vfs, OID_AUTO, hifreebuffers, CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hifreebuffers, __offsetof(struct bufdomain, bd_hifreebuffers), sysctl_bufdomain_int, "I", "Threshold for clean buffer recycling"); static counter_u64_t getnewbufcalls; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, "Number of calls to getnewbuf"); static counter_u64_t getnewbufrestarts; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts, "Number of times getnewbuf has had to restart a buffer acquisition"); static counter_u64_t mappingrestarts; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RD, &mappingrestarts, "Number of times getblk has had to restart a buffer mapping for " "unmapped buffer"); static counter_u64_t numbufallocfails; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW, &numbufallocfails, "Number of times buffer allocations failed"); static int flushbufqtarget = 100; SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0, "Amount of work to do in flushbufqueues when helping bufdaemon"); static counter_u64_t notbufdflushes; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, ¬bufdflushes, "Number of dirty buffer flushes done by the bufdaemon helpers"); static long barrierwrites; SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW | CTLFLAG_STATS, &barrierwrites, 0, "Number of barrier writes"); SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD, &unmapped_buf_allowed, 0, "Permit the use of the unmapped i/o"); int maxbcachebuf = MAXBCACHEBUF; SYSCTL_INT(_vfs, OID_AUTO, maxbcachebuf, CTLFLAG_RDTUN, &maxbcachebuf, 0, "Maximum size of a buffer cache block"); /* * This lock synchronizes access to bd_request. */ static struct mtx_padalign __exclusive_cache_line bdlock; /* * This lock protects the runningbufreq and synchronizes runningbufwakeup and * waitrunningbufspace(). */ static struct mtx_padalign __exclusive_cache_line rbreqlock; /* * Lock that protects bdirtywait. */ static struct mtx_padalign __exclusive_cache_line bdirtylock; /* * Wakeup point for bufdaemon, as well as indicator of whether it is already * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it * is idling. */ static int bd_request; /* * Request for the buf daemon to write more buffers than is indicated by * lodirtybuf. This may be necessary to push out excess dependencies or * defragment the address space where a simple count of the number of dirty * buffers is insufficient to characterize the demand for flushing them. */ static int bd_speedupreq; /* * Synchronization (sleep/wakeup) variable for active buffer space requests. * Set when wait starts, cleared prior to wakeup(). * Used in runningbufwakeup() and waitrunningbufspace(). */ static int runningbufreq; /* * Synchronization for bwillwrite() waiters. */ static int bdirtywait; /* * Definitions for the buffer free lists. */ #define QUEUE_NONE 0 /* on no queue */ #define QUEUE_EMPTY 1 /* empty buffer headers */ #define QUEUE_DIRTY 2 /* B_DELWRI buffers */ #define QUEUE_CLEAN 3 /* non-B_DELWRI buffers */ #define QUEUE_SENTINEL 4 /* not an queue index, but mark for sentinel */ /* Maximum number of buffer domains. */ #define BUF_DOMAINS 8 struct bufdomainset bdlodirty; /* Domains > lodirty */ struct bufdomainset bdhidirty; /* Domains > hidirty */ /* Configured number of clean queues. */ static int __read_mostly buf_domains; BITSET_DEFINE(bufdomainset, BUF_DOMAINS); struct bufdomain __exclusive_cache_line bdomain[BUF_DOMAINS]; struct bufqueue __exclusive_cache_line bqempty; /* * per-cpu empty buffer cache. */ uma_zone_t buf_zone; /* * Single global constant for BUF_WMESG, to avoid getting multiple references. * buf_wmesg is referred from macros. */ const char *buf_wmesg = BUF_WMESG; static int sysctl_runningspace(SYSCTL_HANDLER_ARGS) { long value; int error; value = *(long *)arg1; error = sysctl_handle_long(oidp, &value, 0, req); if (error != 0 || req->newptr == NULL) return (error); mtx_lock(&rbreqlock); if (arg1 == &hirunningspace) { if (value < lorunningspace) error = EINVAL; else hirunningspace = value; } else { KASSERT(arg1 == &lorunningspace, ("%s: unknown arg1", __func__)); if (value > hirunningspace) error = EINVAL; else lorunningspace = value; } mtx_unlock(&rbreqlock); return (error); } static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS) { int error; int value; int i; value = *(int *)arg1; error = sysctl_handle_int(oidp, &value, 0, req); if (error != 0 || req->newptr == NULL) return (error); *(int *)arg1 = value; for (i = 0; i < buf_domains; i++) *(int *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) = value / buf_domains; return (error); } static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS) { long value; int error; int i; value = *(long *)arg1; error = sysctl_handle_long(oidp, &value, 0, req); if (error != 0 || req->newptr == NULL) return (error); *(long *)arg1 = value; for (i = 0; i < buf_domains; i++) *(long *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) = value / buf_domains; return (error); } #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) static int sysctl_bufspace(SYSCTL_HANDLER_ARGS) { long lvalue; int ivalue; int i; lvalue = 0; for (i = 0; i < buf_domains; i++) lvalue += bdomain[i].bd_bufspace; if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long)) return (sysctl_handle_long(oidp, &lvalue, 0, req)); if (lvalue > INT_MAX) /* On overflow, still write out a long to trigger ENOMEM. */ return (sysctl_handle_long(oidp, &lvalue, 0, req)); ivalue = lvalue; return (sysctl_handle_int(oidp, &ivalue, 0, req)); } #else static int sysctl_bufspace(SYSCTL_HANDLER_ARGS) { long lvalue; int i; lvalue = 0; for (i = 0; i < buf_domains; i++) lvalue += bdomain[i].bd_bufspace; return (sysctl_handle_long(oidp, &lvalue, 0, req)); } #endif static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS) { int value; int i; value = 0; for (i = 0; i < buf_domains; i++) value += bdomain[i].bd_numdirtybuffers; return (sysctl_handle_int(oidp, &value, 0, req)); } /* * bdirtywakeup: * * Wakeup any bwillwrite() waiters. */ static void bdirtywakeup(void) { mtx_lock(&bdirtylock); if (bdirtywait) { bdirtywait = 0; wakeup(&bdirtywait); } mtx_unlock(&bdirtylock); } /* * bd_clear: * * Clear a domain from the appropriate bitsets when dirtybuffers * is decremented. */ static void bd_clear(struct bufdomain *bd) { mtx_lock(&bdirtylock); if (bd->bd_numdirtybuffers <= bd->bd_lodirtybuffers) BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty); if (bd->bd_numdirtybuffers <= bd->bd_hidirtybuffers) BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty); mtx_unlock(&bdirtylock); } /* * bd_set: * * Set a domain in the appropriate bitsets when dirtybuffers * is incremented. */ static void bd_set(struct bufdomain *bd) { mtx_lock(&bdirtylock); if (bd->bd_numdirtybuffers > bd->bd_lodirtybuffers) BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty); if (bd->bd_numdirtybuffers > bd->bd_hidirtybuffers) BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty); mtx_unlock(&bdirtylock); } /* * bdirtysub: * * Decrement the numdirtybuffers count by one and wakeup any * threads blocked in bwillwrite(). */ static void bdirtysub(struct buf *bp) { struct bufdomain *bd; int num; bd = bufdomain(bp); num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, -1); if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2) bdirtywakeup(); if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers) bd_clear(bd); } /* * bdirtyadd: * * Increment the numdirtybuffers count by one and wakeup the buf * daemon if needed. */ static void bdirtyadd(struct buf *bp) { struct bufdomain *bd; int num; /* * Only do the wakeup once as we cross the boundary. The * buf daemon will keep running until the condition clears. */ bd = bufdomain(bp); num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, 1); if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2) bd_wakeup(); if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers) bd_set(bd); } /* * bufspace_daemon_wakeup: * * Wakeup the daemons responsible for freeing clean bufs. */ static void bufspace_daemon_wakeup(struct bufdomain *bd) { /* * avoid the lock if the daemon is running. */ if (atomic_fetchadd_int(&bd->bd_running, 1) == 0) { BD_RUN_LOCK(bd); atomic_store_int(&bd->bd_running, 1); wakeup(&bd->bd_running); BD_RUN_UNLOCK(bd); } } /* * bufspace_daemon_wait: * * Sleep until the domain falls below a limit or one second passes. */ static void bufspace_daemon_wait(struct bufdomain *bd) { /* * Re-check our limits and sleep. bd_running must be * cleared prior to checking the limits to avoid missed * wakeups. The waker will adjust one of bufspace or * freebuffers prior to checking bd_running. */ BD_RUN_LOCK(bd); atomic_store_int(&bd->bd_running, 0); if (bd->bd_bufspace < bd->bd_bufspacethresh && bd->bd_freebuffers > bd->bd_lofreebuffers) { msleep(&bd->bd_running, BD_RUN_LOCKPTR(bd), PRIBIO|PDROP, "-", hz); } else { /* Avoid spurious wakeups while running. */ atomic_store_int(&bd->bd_running, 1); BD_RUN_UNLOCK(bd); } } /* * bufspace_adjust: * * Adjust the reported bufspace for a KVA managed buffer, possibly * waking any waiters. */ static void bufspace_adjust(struct buf *bp, int bufsize) { struct bufdomain *bd; long space; int diff; KASSERT((bp->b_flags & B_MALLOC) == 0, ("bufspace_adjust: malloc buf %p", bp)); bd = bufdomain(bp); diff = bufsize - bp->b_bufsize; if (diff < 0) { atomic_subtract_long(&bd->bd_bufspace, -diff); } else if (diff > 0) { space = atomic_fetchadd_long(&bd->bd_bufspace, diff); /* Wake up the daemon on the transition. */ if (space < bd->bd_bufspacethresh && space + diff >= bd->bd_bufspacethresh) bufspace_daemon_wakeup(bd); } bp->b_bufsize = bufsize; } /* * bufspace_reserve: * * Reserve bufspace before calling allocbuf(). metadata has a * different space limit than data. */ static int bufspace_reserve(struct bufdomain *bd, int size, bool metadata) { long limit, new; long space; if (metadata) limit = bd->bd_maxbufspace; else limit = bd->bd_hibufspace; space = atomic_fetchadd_long(&bd->bd_bufspace, size); new = space + size; if (new > limit) { atomic_subtract_long(&bd->bd_bufspace, size); return (ENOSPC); } /* Wake up the daemon on the transition. */ if (space < bd->bd_bufspacethresh && new >= bd->bd_bufspacethresh) bufspace_daemon_wakeup(bd); return (0); } /* * bufspace_release: * * Release reserved bufspace after bufspace_adjust() has consumed it. */ static void bufspace_release(struct bufdomain *bd, int size) { atomic_subtract_long(&bd->bd_bufspace, size); } /* * bufspace_wait: * * Wait for bufspace, acting as the buf daemon if a locked vnode is * supplied. bd_wanted must be set prior to polling for space. The * operation must be re-tried on return. */ static void bufspace_wait(struct bufdomain *bd, struct vnode *vp, int gbflags, int slpflag, int slptimeo) { struct thread *td; int error, fl, norunbuf; if ((gbflags & GB_NOWAIT_BD) != 0) return; td = curthread; BD_LOCK(bd); while (bd->bd_wanted) { if (vp != NULL && vp->v_type != VCHR && (td->td_pflags & TDP_BUFNEED) == 0) { BD_UNLOCK(bd); /* * getblk() is called with a vnode locked, and * some majority of the dirty buffers may as * well belong to the vnode. Flushing the * buffers there would make a progress that * cannot be achieved by the buf_daemon, that * cannot lock the vnode. */ norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) | (td->td_pflags & TDP_NORUNNINGBUF); /* * Play bufdaemon. The getnewbuf() function * may be called while the thread owns lock * for another dirty buffer for the same * vnode, which makes it impossible to use * VOP_FSYNC() there, due to the buffer lock * recursion. */ td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF; fl = buf_flush(vp, bd, flushbufqtarget); td->td_pflags &= norunbuf; BD_LOCK(bd); if (fl != 0) continue; if (bd->bd_wanted == 0) break; } error = msleep(&bd->bd_wanted, BD_LOCKPTR(bd), (PRIBIO + 4) | slpflag, "newbuf", slptimeo); if (error != 0) break; } BD_UNLOCK(bd); } /* * bufspace_daemon: * * buffer space management daemon. Tries to maintain some marginal * amount of free buffer space so that requesting processes neither * block nor work to reclaim buffers. */ static void bufspace_daemon(void *arg) { struct bufdomain *bd; EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread, SHUTDOWN_PRI_LAST + 100); bd = arg; for (;;) { kthread_suspend_check(); /* * Free buffers from the clean queue until we meet our * targets. * * Theory of operation: The buffer cache is most efficient * when some free buffer headers and space are always * available to getnewbuf(). This daemon attempts to prevent * the excessive blocking and synchronization associated * with shortfall. It goes through three phases according * demand: * * 1) The daemon wakes up voluntarily once per-second * during idle periods when the counters are below * the wakeup thresholds (bufspacethresh, lofreebuffers). * * 2) The daemon wakes up as we cross the thresholds * ahead of any potential blocking. This may bounce * slightly according to the rate of consumption and * release. * * 3) The daemon and consumers are starved for working * clean buffers. This is the 'bufspace' sleep below * which will inefficiently trade bufs with bqrelse * until we return to condition 2. */ while (bd->bd_bufspace > bd->bd_lobufspace || bd->bd_freebuffers < bd->bd_hifreebuffers) { if (buf_recycle(bd, false) != 0) { if (bd_flushall(bd)) continue; /* * Speedup dirty if we've run out of clean * buffers. This is possible in particular * because softdep may held many bufs locked * pending writes to other bufs which are * marked for delayed write, exhausting * clean space until they are written. */ bd_speedup(); BD_LOCK(bd); if (bd->bd_wanted) { msleep(&bd->bd_wanted, BD_LOCKPTR(bd), PRIBIO|PDROP, "bufspace", hz/10); } else BD_UNLOCK(bd); } maybe_yield(); } bufspace_daemon_wait(bd); } } /* * bufmallocadjust: * * Adjust the reported bufspace for a malloc managed buffer, possibly * waking any waiters. */ static void bufmallocadjust(struct buf *bp, int bufsize) { int diff; KASSERT((bp->b_flags & B_MALLOC) != 0, ("bufmallocadjust: non-malloc buf %p", bp)); diff = bufsize - bp->b_bufsize; if (diff < 0) atomic_subtract_long(&bufmallocspace, -diff); else atomic_add_long(&bufmallocspace, diff); bp->b_bufsize = bufsize; } /* * runningwakeup: * * Wake up processes that are waiting on asynchronous writes to fall * below lorunningspace. */ static void runningwakeup(void) { mtx_lock(&rbreqlock); if (runningbufreq) { runningbufreq = 0; wakeup(&runningbufreq); } mtx_unlock(&rbreqlock); } /* * runningbufwakeup: * * Decrement the outstanding write count according. */ void runningbufwakeup(struct buf *bp) { long space, bspace; bspace = bp->b_runningbufspace; if (bspace == 0) return; space = atomic_fetchadd_long(&runningbufspace, -bspace); KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld", space, bspace)); bp->b_runningbufspace = 0; /* * Only acquire the lock and wakeup on the transition from exceeding * the threshold to falling below it. */ if (space < lorunningspace) return; if (space - bspace > lorunningspace) return; runningwakeup(); } /* * waitrunningbufspace() * * runningbufspace is a measure of the amount of I/O currently * running. This routine is used in async-write situations to * prevent creating huge backups of pending writes to a device. * Only asynchronous writes are governed by this function. * * This does NOT turn an async write into a sync write. It waits * for earlier writes to complete and generally returns before the * caller's write has reached the device. */ void waitrunningbufspace(void) { mtx_lock(&rbreqlock); while (runningbufspace > hirunningspace) { runningbufreq = 1; msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); } mtx_unlock(&rbreqlock); } /* * vfs_buf_test_cache: * * Called when a buffer is extended. This function clears the B_CACHE * bit if the newly extended portion of the buffer does not contain * valid data. */ static __inline void vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, vm_page_t m) { VM_OBJECT_ASSERT_LOCKED(m->object); if (bp->b_flags & B_CACHE) { int base = (foff + off) & PAGE_MASK; if (vm_page_is_valid(m, base, size) == 0) bp->b_flags &= ~B_CACHE; } } /* Wake up the buffer daemon if necessary */ static void bd_wakeup(void) { mtx_lock(&bdlock); if (bd_request == 0) { bd_request = 1; wakeup(&bd_request); } mtx_unlock(&bdlock); } /* * Adjust the maxbcachbuf tunable. */ static void maxbcachebuf_adjust(void) { int i; /* * maxbcachebuf must be a power of 2 >= MAXBSIZE. */ i = 2; while (i * 2 <= maxbcachebuf) i *= 2; maxbcachebuf = i; if (maxbcachebuf < MAXBSIZE) maxbcachebuf = MAXBSIZE; if (maxbcachebuf > MAXPHYS) maxbcachebuf = MAXPHYS; if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF) printf("maxbcachebuf=%d\n", maxbcachebuf); } /* * bd_speedup - speedup the buffer cache flushing code */ void bd_speedup(void) { int needwake; mtx_lock(&bdlock); needwake = 0; if (bd_speedupreq == 0 || bd_request == 0) needwake = 1; bd_speedupreq = 1; bd_request = 1; if (needwake) wakeup(&bd_request); mtx_unlock(&bdlock); } #ifdef __i386__ #define TRANSIENT_DENOM 5 #else #define TRANSIENT_DENOM 10 #endif /* * Calculating buffer cache scaling values and reserve space for buffer * headers. This is called during low level kernel initialization and * may be called more then once. We CANNOT write to the memory area * being reserved at this time. */ caddr_t kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) { int tuned_nbuf; long maxbuf, maxbuf_sz, buf_sz, biotmap_sz; /* * physmem_est is in pages. Convert it to kilobytes (assumes * PAGE_SIZE is >= 1K) */ physmem_est = physmem_est * (PAGE_SIZE / 1024); maxbcachebuf_adjust(); /* * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. * For the first 64MB of ram nominally allocate sufficient buffers to * cover 1/4 of our ram. Beyond the first 64MB allocate additional * buffers to cover 1/10 of our ram over 64MB. When auto-sizing * the buffer cache we limit the eventual kva reservation to * maxbcache bytes. * * factor represents the 1/4 x ram conversion. */ if (nbuf == 0) { int factor = 4 * BKVASIZE / 1024; nbuf = 50; if (physmem_est > 4096) nbuf += min((physmem_est - 4096) / factor, 65536 / factor); if (physmem_est > 65536) nbuf += min((physmem_est - 65536) * 2 / (factor * 5), 32 * 1024 * 1024 / (factor * 5)); if (maxbcache && nbuf > maxbcache / BKVASIZE) nbuf = maxbcache / BKVASIZE; tuned_nbuf = 1; } else tuned_nbuf = 0; /* XXX Avoid unsigned long overflows later on with maxbufspace. */ maxbuf = (LONG_MAX / 3) / BKVASIZE; if (nbuf > maxbuf) { if (!tuned_nbuf) printf("Warning: nbufs lowered from %d to %ld\n", nbuf, maxbuf); nbuf = maxbuf; } /* * Ideal allocation size for the transient bio submap is 10% * of the maximal space buffer map. This roughly corresponds * to the amount of the buffer mapped for typical UFS load. * * Clip the buffer map to reserve space for the transient * BIOs, if its extent is bigger than 90% (80% on i386) of the * maximum buffer map extent on the platform. * * The fall-back to the maxbuf in case of maxbcache unset, * allows to not trim the buffer KVA for the architectures * with ample KVA space. */ if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) { maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE; buf_sz = (long)nbuf * BKVASIZE; if (buf_sz < maxbuf_sz / TRANSIENT_DENOM * (TRANSIENT_DENOM - 1)) { /* * There is more KVA than memory. Do not * adjust buffer map size, and assign the rest * of maxbuf to transient map. */ biotmap_sz = maxbuf_sz - buf_sz; } else { /* * Buffer map spans all KVA we could afford on * this platform. Give 10% (20% on i386) of * the buffer map to the transient bio map. */ biotmap_sz = buf_sz / TRANSIENT_DENOM; buf_sz -= biotmap_sz; } if (biotmap_sz / INT_MAX > MAXPHYS) bio_transient_maxcnt = INT_MAX; else bio_transient_maxcnt = biotmap_sz / MAXPHYS; /* * Artificially limit to 1024 simultaneous in-flight I/Os * using the transient mapping. */ if (bio_transient_maxcnt > 1024) bio_transient_maxcnt = 1024; if (tuned_nbuf) nbuf = buf_sz / BKVASIZE; } if (nswbuf == 0) { nswbuf = min(nbuf / 4, 256); if (nswbuf < NSWBUF_MIN) nswbuf = NSWBUF_MIN; } /* * Reserve space for the buffer cache buffers */ buf = (void *)v; v = (caddr_t)(buf + nbuf); return(v); } /* Initialize the buffer subsystem. Called before use of any buffers. */ void bufinit(void) { struct buf *bp; int i; KASSERT(maxbcachebuf >= MAXBSIZE, ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf, MAXBSIZE)); bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock"); mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF); unmapped_buf = (caddr_t)kva_alloc(MAXPHYS); /* finally, initialize each buffer header and stick on empty q */ for (i = 0; i < nbuf; i++) { bp = &buf[i]; bzero(bp, sizeof *bp); bp->b_flags = B_INVAL; bp->b_rcred = NOCRED; bp->b_wcred = NOCRED; bp->b_qindex = QUEUE_NONE; bp->b_domain = -1; bp->b_subqueue = mp_maxid + 1; bp->b_xflags = 0; bp->b_data = bp->b_kvabase = unmapped_buf; LIST_INIT(&bp->b_dep); BUF_LOCKINIT(bp); bq_insert(&bqempty, bp, false); } /* * maxbufspace is the absolute maximum amount of buffer space we are * allowed to reserve in KVM and in real terms. The absolute maximum * is nominally used by metadata. hibufspace is the nominal maximum * used by most other requests. The differential is required to * ensure that metadata deadlocks don't occur. * * maxbufspace is based on BKVASIZE. Allocating buffers larger then * this may result in KVM fragmentation which is not handled optimally * by the system. XXX This is less true with vmem. We could use * PAGE_SIZE. */ maxbufspace = (long)nbuf * BKVASIZE; hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10); lobufspace = (hibufspace / 20) * 19; /* 95% */ bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2; /* * Note: The 16 MiB upper limit for hirunningspace was chosen * arbitrarily and may need further tuning. It corresponds to * 128 outstanding write IO requests (if IO size is 128 KiB), * which fits with many RAID controllers' tagged queuing limits. * The lower 1 MiB limit is the historical upper limit for * hirunningspace. */ hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf), 16 * 1024 * 1024), 1024 * 1024); lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf); /* * Limit the amount of malloc memory since it is wired permanently into * the kernel space. Even though this is accounted for in the buffer * allocation, we don't want the malloced region to grow uncontrolled. * The malloc scheme improves memory utilization significantly on * average (small) directories. */ maxbufmallocspace = hibufspace / 20; /* * Reduce the chance of a deadlock occurring by limiting the number * of delayed-write dirty buffers we allow to stack up. */ hidirtybuffers = nbuf / 4 + 20; dirtybufthresh = hidirtybuffers * 9 / 10; /* * To support extreme low-memory systems, make sure hidirtybuffers * cannot eat up all available buffer space. This occurs when our * minimum cannot be met. We try to size hidirtybuffers to 3/4 our * buffer space assuming BKVASIZE'd buffers. */ while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { hidirtybuffers >>= 1; } lodirtybuffers = hidirtybuffers / 2; /* * lofreebuffers should be sufficient to avoid stalling waiting on * buf headers under heavy utilization. The bufs in per-cpu caches * are counted as free but will be unavailable to threads executing * on other cpus. * * hifreebuffers is the free target for the bufspace daemon. This * should be set appropriately to limit work per-iteration. */ lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus); hifreebuffers = (3 * lofreebuffers) / 2; numfreebuffers = nbuf; /* Setup the kva and free list allocators. */ vmem_set_reclaim(buffer_arena, bufkva_reclaim); buf_zone = uma_zcache_create("buf free cache", sizeof(struct buf), NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0); /* * Size the clean queue according to the amount of buffer space. * One queue per-256mb up to the max. More queues gives better * concurrency but less accurate LRU. */ buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS); for (i = 0 ; i < buf_domains; i++) { struct bufdomain *bd; bd = &bdomain[i]; bd_init(bd); bd->bd_freebuffers = nbuf / buf_domains; bd->bd_hifreebuffers = hifreebuffers / buf_domains; bd->bd_lofreebuffers = lofreebuffers / buf_domains; bd->bd_bufspace = 0; bd->bd_maxbufspace = maxbufspace / buf_domains; bd->bd_hibufspace = hibufspace / buf_domains; bd->bd_lobufspace = lobufspace / buf_domains; bd->bd_bufspacethresh = bufspacethresh / buf_domains; bd->bd_numdirtybuffers = 0; bd->bd_hidirtybuffers = hidirtybuffers / buf_domains; bd->bd_lodirtybuffers = lodirtybuffers / buf_domains; bd->bd_dirtybufthresh = dirtybufthresh / buf_domains; /* Don't allow more than 2% of bufs in the per-cpu caches. */ bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus; } getnewbufcalls = counter_u64_alloc(M_WAITOK); getnewbufrestarts = counter_u64_alloc(M_WAITOK); mappingrestarts = counter_u64_alloc(M_WAITOK); numbufallocfails = counter_u64_alloc(M_WAITOK); notbufdflushes = counter_u64_alloc(M_WAITOK); buffreekvacnt = counter_u64_alloc(M_WAITOK); bufdefragcnt = counter_u64_alloc(M_WAITOK); bufkvaspace = counter_u64_alloc(M_WAITOK); } #ifdef INVARIANTS static inline void vfs_buf_check_mapped(struct buf *bp) { KASSERT(bp->b_kvabase != unmapped_buf, ("mapped buf: b_kvabase was not updated %p", bp)); KASSERT(bp->b_data != unmapped_buf, ("mapped buf: b_data was not updated %p", bp)); KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf + MAXPHYS, ("b_data + b_offset unmapped %p", bp)); } static inline void vfs_buf_check_unmapped(struct buf *bp) { KASSERT(bp->b_data == unmapped_buf, ("unmapped buf: corrupted b_data %p", bp)); } #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp) #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp) #else #define BUF_CHECK_MAPPED(bp) do {} while (0) #define BUF_CHECK_UNMAPPED(bp) do {} while (0) #endif static int isbufbusy(struct buf *bp) { if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) || ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI)) return (1); return (0); } /* * Shutdown the system cleanly to prepare for reboot, halt, or power off. */ void bufshutdown(int show_busybufs) { static int first_buf_printf = 1; struct buf *bp; int iter, nbusy, pbusy; #ifndef PREEMPTION int subiter; #endif /* * Sync filesystems for shutdown */ wdog_kern_pat(WD_LASTVAL); sys_sync(curthread, NULL); /* * With soft updates, some buffers that are * written will be remarked as dirty until other * buffers are written. */ for (iter = pbusy = 0; iter < 20; iter++) { nbusy = 0; for (bp = &buf[nbuf]; --bp >= buf; ) if (isbufbusy(bp)) nbusy++; if (nbusy == 0) { if (first_buf_printf) printf("All buffers synced."); break; } if (first_buf_printf) { printf("Syncing disks, buffers remaining... "); first_buf_printf = 0; } printf("%d ", nbusy); if (nbusy < pbusy) iter = 0; pbusy = nbusy; wdog_kern_pat(WD_LASTVAL); sys_sync(curthread, NULL); #ifdef PREEMPTION /* * Spin for a while to allow interrupt threads to run. */ DELAY(50000 * iter); #else /* * Context switch several times to allow interrupt * threads to run. */ for (subiter = 0; subiter < 50 * iter; subiter++) { thread_lock(curthread); mi_switch(SW_VOL, NULL); thread_unlock(curthread); DELAY(1000); } #endif } printf("\n"); /* * Count only busy local buffers to prevent forcing * a fsck if we're just a client of a wedged NFS server */ nbusy = 0; for (bp = &buf[nbuf]; --bp >= buf; ) { if (isbufbusy(bp)) { #if 0 /* XXX: This is bogus. We should probably have a BO_REMOTE flag instead */ if (bp->b_dev == NULL) { TAILQ_REMOVE(&mountlist, bp->b_vp->v_mount, mnt_list); continue; } #endif nbusy++; if (show_busybufs > 0) { printf( "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:", nbusy, bp, bp->b_vp, bp->b_flags, (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno); BUF_LOCKPRINTINFO(bp); if (show_busybufs > 1) vn_printf(bp->b_vp, "vnode content: "); } } } if (nbusy) { /* * Failed to sync all blocks. Indicate this and don't * unmount filesystems (thus forcing an fsck on reboot). */ printf("Giving up on %d buffers\n", nbusy); DELAY(5000000); /* 5 seconds */ } else { if (!first_buf_printf) printf("Final sync complete\n"); /* * Unmount filesystems */ if (panicstr == NULL) vfs_unmountall(); } swapoff_all(); DELAY(100000); /* wait for console output to finish */ } static void bpmap_qenter(struct buf *bp) { BUF_CHECK_MAPPED(bp); /* * bp->b_data is relative to bp->b_offset, but * bp->b_offset may be offset into the first page. */ bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data); pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages); bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | (vm_offset_t)(bp->b_offset & PAGE_MASK)); } static inline struct bufdomain * bufdomain(struct buf *bp) { return (&bdomain[bp->b_domain]); } static struct bufqueue * bufqueue(struct buf *bp) { switch (bp->b_qindex) { case QUEUE_NONE: /* FALLTHROUGH */ case QUEUE_SENTINEL: return (NULL); case QUEUE_EMPTY: return (&bqempty); case QUEUE_DIRTY: return (&bufdomain(bp)->bd_dirtyq); case QUEUE_CLEAN: return (&bufdomain(bp)->bd_subq[bp->b_subqueue]); default: break; } panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex); } /* * Return the locked bufqueue that bp is a member of. */ static struct bufqueue * bufqueue_acquire(struct buf *bp) { struct bufqueue *bq, *nbq; /* * bp can be pushed from a per-cpu queue to the * cleanq while we're waiting on the lock. Retry * if the queues don't match. */ bq = bufqueue(bp); BQ_LOCK(bq); for (;;) { nbq = bufqueue(bp); if (bq == nbq) break; BQ_UNLOCK(bq); BQ_LOCK(nbq); bq = nbq; } return (bq); } /* * binsfree: * * Insert the buffer into the appropriate free list. Requires a * locked buffer on entry and buffer is unlocked before return. */ static void binsfree(struct buf *bp, int qindex) { struct bufdomain *bd; struct bufqueue *bq; KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY, ("binsfree: Invalid qindex %d", qindex)); BUF_ASSERT_XLOCKED(bp); /* * Handle delayed bremfree() processing. */ if (bp->b_flags & B_REMFREE) { if (bp->b_qindex == qindex) { bp->b_flags |= B_REUSE; bp->b_flags &= ~B_REMFREE; BUF_UNLOCK(bp); return; } bq = bufqueue_acquire(bp); bq_remove(bq, bp); BQ_UNLOCK(bq); } bd = bufdomain(bp); if (qindex == QUEUE_CLEAN) { if (bd->bd_lim != 0) bq = &bd->bd_subq[PCPU_GET(cpuid)]; else bq = bd->bd_cleanq; } else bq = &bd->bd_dirtyq; bq_insert(bq, bp, true); } /* * buf_free: * * Free a buffer to the buf zone once it no longer has valid contents. */ static void buf_free(struct buf *bp) { if (bp->b_flags & B_REMFREE) bremfreef(bp); if (bp->b_vflags & BV_BKGRDINPROG) panic("losing buffer 1"); if (bp->b_rcred != NOCRED) { crfree(bp->b_rcred); bp->b_rcred = NOCRED; } if (bp->b_wcred != NOCRED) { crfree(bp->b_wcred); bp->b_wcred = NOCRED; } if (!LIST_EMPTY(&bp->b_dep)) buf_deallocate(bp); bufkva_free(bp); atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1); BUF_UNLOCK(bp); uma_zfree(buf_zone, bp); } /* * buf_import: * * Import bufs into the uma cache from the buf list. The system still * expects a static array of bufs and much of the synchronization * around bufs assumes type stable storage. As a result, UMA is used * only as a per-cpu cache of bufs still maintained on a global list. */ static int buf_import(void *arg, void **store, int cnt, int domain, int flags) { struct buf *bp; int i; BQ_LOCK(&bqempty); for (i = 0; i < cnt; i++) { bp = TAILQ_FIRST(&bqempty.bq_queue); if (bp == NULL) break; bq_remove(&bqempty, bp); store[i] = bp; } BQ_UNLOCK(&bqempty); return (i); } /* * buf_release: * * Release bufs from the uma cache back to the buffer queues. */ static void buf_release(void *arg, void **store, int cnt) { struct bufqueue *bq; struct buf *bp; int i; bq = &bqempty; BQ_LOCK(bq); for (i = 0; i < cnt; i++) { bp = store[i]; /* Inline bq_insert() to batch locking. */ TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist); bp->b_flags &= ~(B_AGE | B_REUSE); bq->bq_len++; bp->b_qindex = bq->bq_index; } BQ_UNLOCK(bq); } /* * buf_alloc: * * Allocate an empty buffer header. */ static struct buf * buf_alloc(struct bufdomain *bd) { struct buf *bp; int freebufs; /* * We can only run out of bufs in the buf zone if the average buf * is less than BKVASIZE. In this case the actual wait/block will * come from buf_reycle() failing to flush one of these small bufs. */ bp = NULL; freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1); if (freebufs > 0) bp = uma_zalloc(buf_zone, M_NOWAIT); if (bp == NULL) { atomic_add_int(&bd->bd_freebuffers, 1); bufspace_daemon_wakeup(bd); counter_u64_add(numbufallocfails, 1); return (NULL); } /* * Wake-up the bufspace daemon on transition below threshold. */ if (freebufs == bd->bd_lofreebuffers) bufspace_daemon_wakeup(bd); if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) panic("getnewbuf_empty: Locked buf %p on free queue.", bp); KASSERT(bp->b_vp == NULL, ("bp: %p still has vnode %p.", bp, bp->b_vp)); KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0, ("invalid buffer %p flags %#x", bp, bp->b_flags)); KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags)); KASSERT(bp->b_npages == 0, ("bp: %p still has %d vm pages\n", bp, bp->b_npages)); KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp)); KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp)); bp->b_domain = BD_DOMAIN(bd); bp->b_flags = 0; bp->b_ioflags = 0; bp->b_xflags = 0; bp->b_vflags = 0; bp->b_vp = NULL; bp->b_blkno = bp->b_lblkno = 0; bp->b_offset = NOOFFSET; bp->b_iodone = 0; bp->b_error = 0; bp->b_resid = 0; bp->b_bcount = 0; bp->b_npages = 0; bp->b_dirtyoff = bp->b_dirtyend = 0; bp->b_bufobj = NULL; bp->b_data = bp->b_kvabase = unmapped_buf; bp->b_fsprivate1 = NULL; bp->b_fsprivate2 = NULL; bp->b_fsprivate3 = NULL; LIST_INIT(&bp->b_dep); return (bp); } /* * buf_recycle: * * Free a buffer from the given bufqueue. kva controls whether the * freed buf must own some kva resources. This is used for * defragmenting. */ static int buf_recycle(struct bufdomain *bd, bool kva) { struct bufqueue *bq; struct buf *bp, *nbp; if (kva) counter_u64_add(bufdefragcnt, 1); nbp = NULL; bq = bd->bd_cleanq; BQ_LOCK(bq); KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd), ("buf_recycle: Locks don't match")); nbp = TAILQ_FIRST(&bq->bq_queue); /* * Run scan, possibly freeing data and/or kva mappings on the fly * depending. */ while ((bp = nbp) != NULL) { /* * Calculate next bp (we can only use it if we do not * release the bqlock). */ nbp = TAILQ_NEXT(bp, b_freelist); /* * If we are defragging then we need a buffer with * some kva to reclaim. */ if (kva && bp->b_kvasize == 0) continue; if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) continue; /* * Implement a second chance algorithm for frequently * accessed buffers. */ if ((bp->b_flags & B_REUSE) != 0) { TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist); TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist); bp->b_flags &= ~B_REUSE; BUF_UNLOCK(bp); continue; } /* * Skip buffers with background writes in progress. */ if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { BUF_UNLOCK(bp); continue; } KASSERT(bp->b_qindex == QUEUE_CLEAN, ("buf_recycle: inconsistent queue %d bp %p", bp->b_qindex, bp)); KASSERT(bp->b_domain == BD_DOMAIN(bd), ("getnewbuf: queue domain %d doesn't match request %d", bp->b_domain, (int)BD_DOMAIN(bd))); /* * NOTE: nbp is now entirely invalid. We can only restart * the scan from this point on. */ bq_remove(bq, bp); BQ_UNLOCK(bq); /* * Requeue the background write buffer with error and * restart the scan. */ if ((bp->b_vflags & BV_BKGRDERR) != 0) { bqrelse(bp); BQ_LOCK(bq); nbp = TAILQ_FIRST(&bq->bq_queue); continue; } bp->b_flags |= B_INVAL; brelse(bp); return (0); } bd->bd_wanted = 1; BQ_UNLOCK(bq); return (ENOBUFS); } /* * bremfree: * * Mark the buffer for removal from the appropriate free list. * */ void bremfree(struct buf *bp) { CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT((bp->b_flags & B_REMFREE) == 0, ("bremfree: buffer %p already marked for delayed removal.", bp)); KASSERT(bp->b_qindex != QUEUE_NONE, ("bremfree: buffer %p not on a queue.", bp)); BUF_ASSERT_XLOCKED(bp); bp->b_flags |= B_REMFREE; } /* * bremfreef: * * Force an immediate removal from a free list. Used only in nfs when * it abuses the b_freelist pointer. */ void bremfreef(struct buf *bp) { struct bufqueue *bq; bq = bufqueue_acquire(bp); bq_remove(bq, bp); BQ_UNLOCK(bq); } static void bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname) { mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF); TAILQ_INIT(&bq->bq_queue); bq->bq_len = 0; bq->bq_index = qindex; bq->bq_subqueue = subqueue; } static void bd_init(struct bufdomain *bd) { int i; bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1]; bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock"); bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock"); for (i = 0; i <= mp_maxid; i++) bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i, "bufq clean subqueue lock"); mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF); } /* * bq_remove: * * Removes a buffer from the free list, must be called with the * correct qlock held. */ static void bq_remove(struct bufqueue *bq, struct buf *bp) { CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT(bp->b_qindex != QUEUE_NONE, ("bq_remove: buffer %p not on a queue.", bp)); KASSERT(bufqueue(bp) == bq, ("bq_remove: Remove buffer %p from wrong queue.", bp)); BQ_ASSERT_LOCKED(bq); if (bp->b_qindex != QUEUE_EMPTY) { BUF_ASSERT_XLOCKED(bp); } KASSERT(bq->bq_len >= 1, ("queue %d underflow", bp->b_qindex)); TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist); bq->bq_len--; bp->b_qindex = QUEUE_NONE; bp->b_flags &= ~(B_REMFREE | B_REUSE); } static void bd_flush(struct bufdomain *bd, struct bufqueue *bq) { struct buf *bp; BQ_ASSERT_LOCKED(bq); if (bq != bd->bd_cleanq) { BD_LOCK(bd); while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) { TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist); TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp, b_freelist); bp->b_subqueue = bd->bd_cleanq->bq_subqueue; } bd->bd_cleanq->bq_len += bq->bq_len; bq->bq_len = 0; } if (bd->bd_wanted) { bd->bd_wanted = 0; wakeup(&bd->bd_wanted); } if (bq != bd->bd_cleanq) BD_UNLOCK(bd); } static int bd_flushall(struct bufdomain *bd) { struct bufqueue *bq; int flushed; int i; if (bd->bd_lim == 0) return (0); flushed = 0; for (i = 0; i <= mp_maxid; i++) { bq = &bd->bd_subq[i]; if (bq->bq_len == 0) continue; BQ_LOCK(bq); bd_flush(bd, bq); BQ_UNLOCK(bq); flushed++; } return (flushed); } static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock) { struct bufdomain *bd; if (bp->b_qindex != QUEUE_NONE) panic("bq_insert: free buffer %p onto another queue?", bp); bd = bufdomain(bp); if (bp->b_flags & B_AGE) { /* Place this buf directly on the real queue. */ if (bq->bq_index == QUEUE_CLEAN) bq = bd->bd_cleanq; BQ_LOCK(bq); TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist); } else { BQ_LOCK(bq); TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist); } bp->b_flags &= ~(B_AGE | B_REUSE); bq->bq_len++; bp->b_qindex = bq->bq_index; bp->b_subqueue = bq->bq_subqueue; /* * Unlock before we notify so that we don't wakeup a waiter that * fails a trylock on the buf and sleeps again. */ if (unlock) BUF_UNLOCK(bp); if (bp->b_qindex == QUEUE_CLEAN) { /* * Flush the per-cpu queue and notify any waiters. */ if (bd->bd_wanted || (bq != bd->bd_cleanq && bq->bq_len >= bd->bd_lim)) bd_flush(bd, bq); } BQ_UNLOCK(bq); } /* * bufkva_free: * * Free the kva allocation for a buffer. * */ static void bufkva_free(struct buf *bp) { #ifdef INVARIANTS if (bp->b_kvasize == 0) { KASSERT(bp->b_kvabase == unmapped_buf && bp->b_data == unmapped_buf, ("Leaked KVA space on %p", bp)); } else if (buf_mapped(bp)) BUF_CHECK_MAPPED(bp); else BUF_CHECK_UNMAPPED(bp); #endif if (bp->b_kvasize == 0) return; vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize); counter_u64_add(bufkvaspace, -bp->b_kvasize); counter_u64_add(buffreekvacnt, 1); bp->b_data = bp->b_kvabase = unmapped_buf; bp->b_kvasize = 0; } /* * bufkva_alloc: * * Allocate the buffer KVA and set b_kvasize and b_kvabase. */ static int bufkva_alloc(struct buf *bp, int maxsize, int gbflags) { vm_offset_t addr; int error; KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0, ("Invalid gbflags 0x%x in %s", gbflags, __func__)); bufkva_free(bp); addr = 0; error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr); if (error != 0) { /* * Buffer map is too fragmented. Request the caller * to defragment the map. */ return (error); } bp->b_kvabase = (caddr_t)addr; bp->b_kvasize = maxsize; counter_u64_add(bufkvaspace, bp->b_kvasize); if ((gbflags & GB_UNMAPPED) != 0) { bp->b_data = unmapped_buf; BUF_CHECK_UNMAPPED(bp); } else { bp->b_data = bp->b_kvabase; BUF_CHECK_MAPPED(bp); } return (0); } /* * bufkva_reclaim: * * Reclaim buffer kva by freeing buffers holding kva. This is a vmem * callback that fires to avoid returning failure. */ static void bufkva_reclaim(vmem_t *vmem, int flags) { bool done; int q; int i; done = false; for (i = 0; i < 5; i++) { for (q = 0; q < buf_domains; q++) if (buf_recycle(&bdomain[q], true) != 0) done = true; if (done) break; } return; } /* * Attempt to initiate asynchronous I/O on read-ahead blocks. We must * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, * the buffer is valid and we do not have to do anything. */ static void breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt, struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *)) { struct buf *rabp; struct thread *td; int i; td = curthread; for (i = 0; i < cnt; i++, rablkno++, rabsize++) { if (inmem(vp, *rablkno)) continue; rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); if ((rabp->b_flags & B_CACHE) != 0) { brelse(rabp); continue; } #ifdef RACCT if (racct_enable) { PROC_LOCK(curproc); racct_add_buf(curproc, rabp, 0); PROC_UNLOCK(curproc); } #endif /* RACCT */ td->td_ru.ru_inblock++; rabp->b_flags |= B_ASYNC; rabp->b_flags &= ~B_INVAL; if ((flags & GB_CKHASH) != 0) { rabp->b_flags |= B_CKHASH; rabp->b_ckhashcalc = ckhashfunc; } rabp->b_ioflags &= ~BIO_ERROR; rabp->b_iocmd = BIO_READ; if (rabp->b_rcred == NOCRED && cred != NOCRED) rabp->b_rcred = crhold(cred); vfs_busy_pages(rabp, 0); BUF_KERNPROC(rabp); rabp->b_iooffset = dbtob(rabp->b_blkno); bstrategy(rabp); } } /* * Entry point for bread() and breadn() via #defines in sys/buf.h. * * Get a buffer with the specified data. Look in the cache first. We * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE * is set, the buffer is valid and we do not have to do anything, see * getblk(). Also starts asynchronous I/O on read-ahead blocks. * * Always return a NULL buffer pointer (in bpp) when returning an error. */ int breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno, int *rabsize, int cnt, struct ucred *cred, int flags, void (*ckhashfunc)(struct buf *), struct buf **bpp) { struct buf *bp; struct thread *td; int error, readwait, rv; CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); td = curthread; /* * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags * are specified. */ error = getblkx(vp, blkno, size, 0, 0, flags, &bp); if (error != 0) { *bpp = NULL; return (error); } flags &= ~GB_NOSPARSE; *bpp = bp; /* * If not found in cache, do some I/O */ readwait = 0; if ((bp->b_flags & B_CACHE) == 0) { #ifdef RACCT if (racct_enable) { PROC_LOCK(td->td_proc); racct_add_buf(td->td_proc, bp, 0); PROC_UNLOCK(td->td_proc); } #endif /* RACCT */ td->td_ru.ru_inblock++; bp->b_iocmd = BIO_READ; bp->b_flags &= ~B_INVAL; if ((flags & GB_CKHASH) != 0) { bp->b_flags |= B_CKHASH; bp->b_ckhashcalc = ckhashfunc; } bp->b_ioflags &= ~BIO_ERROR; if (bp->b_rcred == NOCRED && cred != NOCRED) bp->b_rcred = crhold(cred); vfs_busy_pages(bp, 0); bp->b_iooffset = dbtob(bp->b_blkno); bstrategy(bp); ++readwait; } /* * Attempt to initiate asynchronous I/O on read-ahead blocks. */ breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc); rv = 0; if (readwait) { rv = bufwait(bp); if (rv != 0) { brelse(bp); *bpp = NULL; } } return (rv); } /* * Write, release buffer on completion. (Done by iodone * if async). Do not bother writing anything if the buffer * is invalid. * * Note that we set B_CACHE here, indicating that buffer is * fully valid and thus cacheable. This is true even of NFS * now so we set it generally. This could be set either here * or in biodone() since the I/O is synchronous. We put it * here. */ int bufwrite(struct buf *bp) { int oldflags; struct vnode *vp; long space; int vp_md; CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) { bp->b_flags |= B_INVAL | B_RELBUF; bp->b_flags &= ~B_CACHE; brelse(bp); return (ENXIO); } if (bp->b_flags & B_INVAL) { brelse(bp); return (0); } if (bp->b_flags & B_BARRIER) atomic_add_long(&barrierwrites, 1); oldflags = bp->b_flags; KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), ("FFS background buffer should not get here %p", bp)); vp = bp->b_vp; if (vp) vp_md = vp->v_vflag & VV_MD; else vp_md = 0; /* * Mark the buffer clean. Increment the bufobj write count * before bundirty() call, to prevent other thread from seeing * empty dirty list and zero counter for writes in progress, * falsely indicating that the bufobj is clean. */ bufobj_wref(bp->b_bufobj); bundirty(bp); bp->b_flags &= ~B_DONE; bp->b_ioflags &= ~BIO_ERROR; bp->b_flags |= B_CACHE; bp->b_iocmd = BIO_WRITE; vfs_busy_pages(bp, 1); /* * Normal bwrites pipeline writes */ bp->b_runningbufspace = bp->b_bufsize; space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace); #ifdef RACCT if (racct_enable) { PROC_LOCK(curproc); racct_add_buf(curproc, bp, 1); PROC_UNLOCK(curproc); } #endif /* RACCT */ curthread->td_ru.ru_oublock++; if (oldflags & B_ASYNC) BUF_KERNPROC(bp); bp->b_iooffset = dbtob(bp->b_blkno); buf_track(bp, __func__); bstrategy(bp); if ((oldflags & B_ASYNC) == 0) { int rtval = bufwait(bp); brelse(bp); return (rtval); } else if (space > hirunningspace) { /* * don't allow the async write to saturate the I/O * system. We will not deadlock here because * we are blocking waiting for I/O that is already in-progress * to complete. We do not block here if it is the update * or syncer daemon trying to clean up as that can lead * to deadlock. */ if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) waitrunningbufspace(); } return (0); } void bufbdflush(struct bufobj *bo, struct buf *bp) { struct buf *nbp; if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); altbufferflushes++; } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { BO_LOCK(bo); /* * Try to find a buffer to flush. */ TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { if ((nbp->b_vflags & BV_BKGRDINPROG) || BUF_LOCK(nbp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) continue; if (bp == nbp) panic("bdwrite: found ourselves"); BO_UNLOCK(bo); /* Don't countdeps with the bo lock held. */ if (buf_countdeps(nbp, 0)) { BO_LOCK(bo); BUF_UNLOCK(nbp); continue; } if (nbp->b_flags & B_CLUSTEROK) { vfs_bio_awrite(nbp); } else { bremfree(nbp); bawrite(nbp); } dirtybufferflushes++; break; } if (nbp == NULL) BO_UNLOCK(bo); } } /* * Delayed write. (Buffer is marked dirty). Do not bother writing * anything if the buffer is marked invalid. * * Note that since the buffer must be completely valid, we can safely * set B_CACHE. In fact, we have to set B_CACHE here rather then in * biodone() in order to prevent getblk from writing the buffer * out synchronously. */ void bdwrite(struct buf *bp) { struct thread *td = curthread; struct vnode *vp; struct bufobj *bo; CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); KASSERT((bp->b_flags & B_BARRIER) == 0, ("Barrier request in delayed write %p", bp)); if (bp->b_flags & B_INVAL) { brelse(bp); return; } /* * If we have too many dirty buffers, don't create any more. * If we are wildly over our limit, then force a complete * cleanup. Otherwise, just keep the situation from getting * out of control. Note that we have to avoid a recursive * disaster and not try to clean up after our own cleanup! */ vp = bp->b_vp; bo = bp->b_bufobj; if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { td->td_pflags |= TDP_INBDFLUSH; BO_BDFLUSH(bo, bp); td->td_pflags &= ~TDP_INBDFLUSH; } else recursiveflushes++; bdirty(bp); /* * Set B_CACHE, indicating that the buffer is fully valid. This is * true even of NFS now. */ bp->b_flags |= B_CACHE; /* * This bmap keeps the system from needing to do the bmap later, * perhaps when the system is attempting to do a sync. Since it * is likely that the indirect block -- or whatever other datastructure * that the filesystem needs is still in memory now, it is a good * thing to do this. Note also, that if the pageout daemon is * requesting a sync -- there might not be enough memory to do * the bmap then... So, this is important to do. */ if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); } buf_track(bp, __func__); /* * Set the *dirty* buffer range based upon the VM system dirty * pages. * * Mark the buffer pages as clean. We need to do this here to * satisfy the vnode_pager and the pageout daemon, so that it * thinks that the pages have been "cleaned". Note that since * the pages are in a delayed write buffer -- the VFS layer * "will" see that the pages get written out on the next sync, * or perhaps the cluster will be completed. */ vfs_clean_pages_dirty_buf(bp); bqrelse(bp); /* * note: we cannot initiate I/O from a bdwrite even if we wanted to, * due to the softdep code. */ } /* * bdirty: * * Turn buffer into delayed write request. We must clear BIO_READ and * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to * itself to properly update it in the dirty/clean lists. We mark it * B_DONE to ensure that any asynchronization of the buffer properly * clears B_DONE ( else a panic will occur later ). * * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() * should only be called if the buffer is known-good. * * Since the buffer is not on a queue, we do not update the numfreebuffers * count. * * The buffer must be on QUEUE_NONE. */ void bdirty(struct buf *bp) { CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); bp->b_flags &= ~(B_RELBUF); bp->b_iocmd = BIO_WRITE; if ((bp->b_flags & B_DELWRI) == 0) { bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; reassignbuf(bp); bdirtyadd(bp); } } /* * bundirty: * * Clear B_DELWRI for buffer. * * Since the buffer is not on a queue, we do not update the numfreebuffers * count. * * The buffer must be on QUEUE_NONE. */ void bundirty(struct buf *bp) { CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); if (bp->b_flags & B_DELWRI) { bp->b_flags &= ~B_DELWRI; reassignbuf(bp); bdirtysub(bp); } /* * Since it is now being written, we can clear its deferred write flag. */ bp->b_flags &= ~B_DEFERRED; } /* * bawrite: * * Asynchronous write. Start output on a buffer, but do not wait for * it to complete. The buffer is released when the output completes. * * bwrite() ( or the VOP routine anyway ) is responsible for handling * B_INVAL buffers. Not us. */ void bawrite(struct buf *bp) { bp->b_flags |= B_ASYNC; (void) bwrite(bp); } /* * babarrierwrite: * * Asynchronous barrier write. Start output on a buffer, but do not * wait for it to complete. Place a write barrier after this write so * that this buffer and all buffers written before it are committed to * the disk before any buffers written after this write are committed * to the disk. The buffer is released when the output completes. */ void babarrierwrite(struct buf *bp) { bp->b_flags |= B_ASYNC | B_BARRIER; (void) bwrite(bp); } /* * bbarrierwrite: * * Synchronous barrier write. Start output on a buffer and wait for * it to complete. Place a write barrier after this write so that * this buffer and all buffers written before it are committed to * the disk before any buffers written after this write are committed * to the disk. The buffer is released when the output completes. */ int bbarrierwrite(struct buf *bp) { bp->b_flags |= B_BARRIER; return (bwrite(bp)); } /* * bwillwrite: * * Called prior to the locking of any vnodes when we are expecting to * write. We do not want to starve the buffer cache with too many * dirty buffers so we block here. By blocking prior to the locking * of any vnodes we attempt to avoid the situation where a locked vnode * prevents the various system daemons from flushing related buffers. */ void bwillwrite(void) { if (buf_dirty_count_severe()) { mtx_lock(&bdirtylock); while (buf_dirty_count_severe()) { bdirtywait = 1; msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4), "flswai", 0); } mtx_unlock(&bdirtylock); } } /* * Return true if we have too many dirty buffers. */ int buf_dirty_count_severe(void) { return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty)); } /* * brelse: * * Release a busy buffer and, if requested, free its resources. The * buffer will be stashed in the appropriate bufqueue[] allowing it * to be accessed later as a cache entity or reused for other purposes. */ void brelse(struct buf *bp) { struct mount *v_mnt; int qindex; /* * Many functions erroneously call brelse with a NULL bp under rare * error conditions. Simply return when called with a NULL bp. */ if (bp == NULL) return; CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0, ("brelse: non-VMIO buffer marked NOREUSE")); if (BUF_LOCKRECURSED(bp)) { /* * Do not process, in particular, do not handle the * B_INVAL/B_RELBUF and do not release to free list. */ BUF_UNLOCK(bp); return; } if (bp->b_flags & B_MANAGED) { bqrelse(bp); return; } if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) { BO_LOCK(bp->b_bufobj); bp->b_vflags &= ~BV_BKGRDERR; BO_UNLOCK(bp->b_bufobj); bdirty(bp); } if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && (bp->b_flags & B_INVALONERR)) { /* * Forced invalidation of dirty buffer contents, to be used * after a failed write in the rare case that the loss of the * contents is acceptable. The buffer is invalidated and * freed. */ bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE; bp->b_flags &= ~(B_ASYNC | B_CACHE); } if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) && !(bp->b_flags & B_INVAL)) { /* * Failed write, redirty. All errors except ENXIO (which * means the device is gone) are treated as being * transient. * * XXX Treating EIO as transient is not correct; the * contract with the local storage device drivers is that * they will only return EIO once the I/O is no longer * retriable. Network I/O also respects this through the * guarantees of TCP and/or the internal retries of NFS. * ENOMEM might be transient, but we also have no way of * knowing when its ok to retry/reschedule. In general, * this entire case should be made obsolete through better * error handling/recovery and resource scheduling. * * Do this also for buffers that failed with ENXIO, but have * non-empty dependencies - the soft updates code might need * to access the buffer to untangle them. * * Must clear BIO_ERROR to prevent pages from being scrapped. */ bp->b_ioflags &= ~BIO_ERROR; bdirty(bp); } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { /* * Either a failed read I/O, or we were asked to free or not * cache the buffer, or we failed to write to a device that's * no longer present. */ bp->b_flags |= B_INVAL; if (!LIST_EMPTY(&bp->b_dep)) buf_deallocate(bp); if (bp->b_flags & B_DELWRI) bdirtysub(bp); bp->b_flags &= ~(B_DELWRI | B_CACHE); if ((bp->b_flags & B_VMIO) == 0) { allocbuf(bp, 0); if (bp->b_vp) brelvp(bp); } } /* * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_truncate() * is called with B_DELWRI set, the underlying pages may wind up * getting freed causing a previous write (bdwrite()) to get 'lost' * because pages associated with a B_DELWRI bp are marked clean. * * We still allow the B_INVAL case to call vfs_vmio_truncate(), even * if B_DELWRI is set. */ if (bp->b_flags & B_DELWRI) bp->b_flags &= ~B_RELBUF; /* * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer * constituted, not even NFS buffers now. Two flags effect this. If * B_INVAL, the struct buf is invalidated but the VM object is kept * around ( i.e. so it is trivial to reconstitute the buffer later ). * * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be * invalidated. BIO_ERROR cannot be set for a failed write unless the * buffer is also B_INVAL because it hits the re-dirtying code above. * * Normally we can do this whether a buffer is B_DELWRI or not. If * the buffer is an NFS buffer, it is tracking piecemeal writes or * the commit state and we cannot afford to lose the buffer. If the * buffer has a background write in progress, we need to keep it * around to prevent it from being reconstituted and starting a second * background write. */ v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL; if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE || (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) && (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 || vn_isdisk(bp->b_vp, NULL) || (bp->b_flags & B_DELWRI) == 0)) { vfs_vmio_invalidate(bp); allocbuf(bp, 0); } if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 || (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) { allocbuf(bp, 0); bp->b_flags &= ~B_NOREUSE; if (bp->b_vp != NULL) brelvp(bp); } /* * If the buffer has junk contents signal it and eventually * clean up B_DELWRI and diassociate the vnode so that gbincore() * doesn't find it. */ if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 || (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0) bp->b_flags |= B_INVAL; if (bp->b_flags & B_INVAL) { if (bp->b_flags & B_DELWRI) bundirty(bp); if (bp->b_vp) brelvp(bp); } buf_track(bp, __func__); /* buffers with no memory */ if (bp->b_bufsize == 0) { buf_free(bp); return; } /* buffers with junk contents */ if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || (bp->b_ioflags & BIO_ERROR)) { bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); if (bp->b_vflags & BV_BKGRDINPROG) panic("losing buffer 2"); qindex = QUEUE_CLEAN; bp->b_flags |= B_AGE; /* remaining buffers */ } else if (bp->b_flags & B_DELWRI) qindex = QUEUE_DIRTY; else qindex = QUEUE_CLEAN; if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) panic("brelse: not dirty"); bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT); /* binsfree unlocks bp. */ binsfree(bp, qindex); } /* * Release a buffer back to the appropriate queue but do not try to free * it. The buffer is expected to be used again soon. * * bqrelse() is used by bdwrite() to requeue a delayed write, and used by * biodone() to requeue an async I/O on completion. It is also used when * known good buffers need to be requeued but we think we may need the data * again soon. * * XXX we should be able to leave the B_RELBUF hint set on completion. */ void bqrelse(struct buf *bp) { int qindex; CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); qindex = QUEUE_NONE; if (BUF_LOCKRECURSED(bp)) { /* do not release to free list */ BUF_UNLOCK(bp); return; } bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); if (bp->b_flags & B_MANAGED) { if (bp->b_flags & B_REMFREE) bremfreef(bp); goto out; } /* buffers with stale but valid contents */ if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) { BO_LOCK(bp->b_bufobj); bp->b_vflags &= ~BV_BKGRDERR; BO_UNLOCK(bp->b_bufobj); qindex = QUEUE_DIRTY; } else { if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) panic("bqrelse: not dirty"); if ((bp->b_flags & B_NOREUSE) != 0) { brelse(bp); return; } qindex = QUEUE_CLEAN; } buf_track(bp, __func__); /* binsfree unlocks bp. */ binsfree(bp, qindex); return; out: buf_track(bp, __func__); /* unlock */ BUF_UNLOCK(bp); } /* * Complete I/O to a VMIO backed page. Validate the pages as appropriate, * restore bogus pages. */ static void vfs_vmio_iodone(struct buf *bp) { vm_ooffset_t foff; vm_page_t m; vm_object_t obj; struct vnode *vp __unused; int i, iosize, resid; bool bogus; obj = bp->b_bufobj->bo_object; KASSERT(REFCOUNT_COUNT(obj->paging_in_progress) >= bp->b_npages, ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)", REFCOUNT_COUNT(obj->paging_in_progress), bp->b_npages)); vp = bp->b_vp; KASSERT(vp->v_holdcnt > 0, ("vfs_vmio_iodone: vnode %p has zero hold count", vp)); KASSERT(vp->v_object != NULL, ("vfs_vmio_iodone: vnode %p has no vm_object", vp)); foff = bp->b_offset; KASSERT(bp->b_offset != NOOFFSET, ("vfs_vmio_iodone: bp %p has no buffer offset", bp)); bogus = false; iosize = bp->b_bcount - bp->b_resid; VM_OBJECT_WLOCK(obj); for (i = 0; i < bp->b_npages; i++) { resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; if (resid > iosize) resid = iosize; /* * cleanup bogus pages, restoring the originals */ m = bp->b_pages[i]; if (m == bogus_page) { bogus = true; m = vm_page_lookup(obj, OFF_TO_IDX(foff)); if (m == NULL) panic("biodone: page disappeared!"); bp->b_pages[i] = m; } else if ((bp->b_iocmd == BIO_READ) && resid > 0) { /* * In the write case, the valid and clean bits are * already changed correctly ( see bdwrite() ), so we * only need to do this here in the read case. */ KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK, resid)) == 0, ("vfs_vmio_iodone: page %p " "has unexpected dirty bits", m)); vfs_page_set_valid(bp, foff, m); } KASSERT(OFF_TO_IDX(foff) == m->pindex, ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch", (intmax_t)foff, (uintmax_t)m->pindex)); vm_page_sunbusy(m); foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; iosize -= resid; } vm_object_pip_wakeupn(obj, bp->b_npages); VM_OBJECT_WUNLOCK(obj); if (bogus && buf_mapped(bp)) { BUF_CHECK_MAPPED(bp); pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); } } /* * Perform page invalidation when a buffer is released. The fully invalid * pages will be reclaimed later in vfs_vmio_truncate(). */ static void vfs_vmio_invalidate(struct buf *bp) { vm_object_t obj; vm_page_t m; int flags, i, resid, poffset, presid; if (buf_mapped(bp)) { BUF_CHECK_MAPPED(bp); pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages); } else BUF_CHECK_UNMAPPED(bp); /* * Get the base offset and length of the buffer. Note that * in the VMIO case if the buffer block size is not * page-aligned then b_data pointer may not be page-aligned. * But our b_pages[] array *IS* page aligned. * * block sizes less then DEV_BSIZE (usually 512) are not * supported due to the page granularity bits (m->valid, * m->dirty, etc...). * * See man buf(9) for more information */ flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0; obj = bp->b_bufobj->bo_object; resid = bp->b_bufsize; poffset = bp->b_offset & PAGE_MASK; VM_OBJECT_WLOCK(obj); for (i = 0; i < bp->b_npages; i++) { m = bp->b_pages[i]; if (m == bogus_page) panic("vfs_vmio_invalidate: Unexpected bogus page."); bp->b_pages[i] = NULL; presid = resid > (PAGE_SIZE - poffset) ? (PAGE_SIZE - poffset) : resid; KASSERT(presid >= 0, ("brelse: extra page")); - while (vm_page_xbusied(m)) - vm_page_sleep_if_xbusy(m, "mbncsh"); + vm_page_busy_acquire(m, VM_ALLOC_SBUSY); if (pmap_page_wired_mappings(m) == 0) vm_page_set_invalid(m, poffset, presid); + vm_page_sunbusy(m); vm_page_release_locked(m, flags); resid -= presid; poffset = 0; } VM_OBJECT_WUNLOCK(obj); bp->b_npages = 0; } /* * Page-granular truncation of an existing VMIO buffer. */ static void vfs_vmio_truncate(struct buf *bp, int desiredpages) { vm_object_t obj; vm_page_t m; int flags, i; if (bp->b_npages == desiredpages) return; if (buf_mapped(bp)) { BUF_CHECK_MAPPED(bp); pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) + (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages); } else BUF_CHECK_UNMAPPED(bp); /* * The object lock is needed only if we will attempt to free pages. */ flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0; if ((bp->b_flags & B_DIRECT) != 0) { flags |= VPR_TRYFREE; obj = bp->b_bufobj->bo_object; VM_OBJECT_WLOCK(obj); } else { obj = NULL; } for (i = desiredpages; i < bp->b_npages; i++) { m = bp->b_pages[i]; KASSERT(m != bogus_page, ("allocbuf: bogus page found")); bp->b_pages[i] = NULL; if (obj != NULL) vm_page_release_locked(m, flags); else vm_page_release(m, flags); } if (obj != NULL) VM_OBJECT_WUNLOCK(obj); bp->b_npages = desiredpages; } /* * Byte granular extension of VMIO buffers. */ static void vfs_vmio_extend(struct buf *bp, int desiredpages, int size) { /* * We are growing the buffer, possibly in a * byte-granular fashion. */ vm_object_t obj; vm_offset_t toff; vm_offset_t tinc; vm_page_t m; /* * Step 1, bring in the VM pages from the object, allocating * them if necessary. We must clear B_CACHE if these pages * are not valid for the range covered by the buffer. */ obj = bp->b_bufobj->bo_object; VM_OBJECT_WLOCK(obj); if (bp->b_npages < desiredpages) { /* * We must allocate system pages since blocking * here could interfere with paging I/O, no * matter which process we are. * * Only exclusive busy can be tested here. * Blocking on shared busy might lead to * deadlocks once allocbuf() is called after * pages are vfs_busy_pages(). */ (void)vm_page_grab_pages(obj, OFF_TO_IDX(bp->b_offset) + bp->b_npages, VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY | VM_ALLOC_NOBUSY | VM_ALLOC_WIRED, &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages); bp->b_npages = desiredpages; } /* * Step 2. We've loaded the pages into the buffer, * we have to figure out if we can still have B_CACHE * set. Note that B_CACHE is set according to the * byte-granular range ( bcount and size ), not the * aligned range ( newbsize ). * * The VM test is against m->valid, which is DEV_BSIZE * aligned. Needless to say, the validity of the data * needs to also be DEV_BSIZE aligned. Note that this * fails with NFS if the server or some other client * extends the file's EOF. If our buffer is resized, * B_CACHE may remain set! XXX */ toff = bp->b_bcount; tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); while ((bp->b_flags & B_CACHE) && toff < size) { vm_pindex_t pi; if (tinc > (size - toff)) tinc = size - toff; pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT; m = bp->b_pages[pi]; vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m); toff += tinc; tinc = PAGE_SIZE; } VM_OBJECT_WUNLOCK(obj); /* * Step 3, fixup the KVA pmap. */ if (buf_mapped(bp)) bpmap_qenter(bp); else BUF_CHECK_UNMAPPED(bp); } /* * Check to see if a block at a particular lbn is available for a clustered * write. */ static int vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) { struct buf *bpa; int match; match = 0; /* If the buf isn't in core skip it */ if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) return (0); /* If the buf is busy we don't want to wait for it */ if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) return (0); /* Only cluster with valid clusterable delayed write buffers */ if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != (B_DELWRI | B_CLUSTEROK)) goto done; if (bpa->b_bufsize != size) goto done; /* * Check to see if it is in the expected place on disk and that the * block has been mapped. */ if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) match = 1; done: BUF_UNLOCK(bpa); return (match); } /* * vfs_bio_awrite: * * Implement clustered async writes for clearing out B_DELWRI buffers. * This is much better then the old way of writing only one buffer at * a time. Note that we may not be presented with the buffers in the * correct order, so we search for the cluster in both directions. */ int vfs_bio_awrite(struct buf *bp) { struct bufobj *bo; int i; int j; daddr_t lblkno = bp->b_lblkno; struct vnode *vp = bp->b_vp; int ncl; int nwritten; int size; int maxcl; int gbflags; bo = &vp->v_bufobj; gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0; /* * right now we support clustered writing only to regular files. If * we find a clusterable block we could be in the middle of a cluster * rather then at the beginning. */ if ((vp->v_type == VREG) && (vp->v_mount != 0) && /* Only on nodes that have the size info */ (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { size = vp->v_mount->mnt_stat.f_iosize; maxcl = MAXPHYS / size; BO_RLOCK(bo); for (i = 1; i < maxcl; i++) if (vfs_bio_clcheck(vp, size, lblkno + i, bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) break; for (j = 1; i + j <= maxcl && j <= lblkno; j++) if (vfs_bio_clcheck(vp, size, lblkno - j, bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) break; BO_RUNLOCK(bo); --j; ncl = i + j; /* * this is a possible cluster write */ if (ncl != 1) { BUF_UNLOCK(bp); nwritten = cluster_wbuild(vp, size, lblkno - j, ncl, gbflags); return (nwritten); } } bremfree(bp); bp->b_flags |= B_ASYNC; /* * default (old) behavior, writing out only one block * * XXX returns b_bufsize instead of b_bcount for nwritten? */ nwritten = bp->b_bufsize; (void) bwrite(bp); return (nwritten); } /* * getnewbuf_kva: * * Allocate KVA for an empty buf header according to gbflags. */ static int getnewbuf_kva(struct buf *bp, int gbflags, int maxsize) { if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) { /* * In order to keep fragmentation sane we only allocate kva * in BKVASIZE chunks. XXX with vmem we can do page size. */ maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; if (maxsize != bp->b_kvasize && bufkva_alloc(bp, maxsize, gbflags)) return (ENOSPC); } return (0); } /* * getnewbuf: * * Find and initialize a new buffer header, freeing up existing buffers * in the bufqueues as necessary. The new buffer is returned locked. * * We block if: * We have insufficient buffer headers * We have insufficient buffer space * buffer_arena is too fragmented ( space reservation fails ) * If we have to flush dirty buffers ( but we try to avoid this ) * * The caller is responsible for releasing the reserved bufspace after * allocbuf() is called. */ static struct buf * getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags) { struct bufdomain *bd; struct buf *bp; bool metadata, reserved; bp = NULL; KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); if (!unmapped_buf_allowed) gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC); if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 || vp->v_type == VCHR) metadata = true; else metadata = false; if (vp == NULL) bd = &bdomain[0]; else bd = &bdomain[vp->v_bufobj.bo_domain]; counter_u64_add(getnewbufcalls, 1); reserved = false; do { if (reserved == false && bufspace_reserve(bd, maxsize, metadata) != 0) { counter_u64_add(getnewbufrestarts, 1); continue; } reserved = true; if ((bp = buf_alloc(bd)) == NULL) { counter_u64_add(getnewbufrestarts, 1); continue; } if (getnewbuf_kva(bp, gbflags, maxsize) == 0) return (bp); break; } while (buf_recycle(bd, false) == 0); if (reserved) bufspace_release(bd, maxsize); if (bp != NULL) { bp->b_flags |= B_INVAL; brelse(bp); } bufspace_wait(bd, vp, gbflags, slpflag, slptimeo); return (NULL); } /* * buf_daemon: * * buffer flushing daemon. Buffers are normally flushed by the * update daemon but if it cannot keep up this process starts to * take the load in an attempt to prevent getnewbuf() from blocking. */ static struct kproc_desc buf_kp = { "bufdaemon", buf_daemon, &bufdaemonproc }; SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); static int buf_flush(struct vnode *vp, struct bufdomain *bd, int target) { int flushed; flushed = flushbufqueues(vp, bd, target, 0); if (flushed == 0) { /* * Could not find any buffers without rollback * dependencies, so just write the first one * in the hopes of eventually making progress. */ if (vp != NULL && target > 2) target /= 2; flushbufqueues(vp, bd, target, 1); } return (flushed); } static void buf_daemon() { struct bufdomain *bd; int speedupreq; int lodirty; int i; /* * This process needs to be suspended prior to shutdown sync. */ EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread, SHUTDOWN_PRI_LAST + 100); /* * Start the buf clean daemons as children threads. */ for (i = 0 ; i < buf_domains; i++) { int error; error = kthread_add((void (*)(void *))bufspace_daemon, &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i); if (error) panic("error %d spawning bufspace daemon", error); } /* * This process is allowed to take the buffer cache to the limit */ curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; mtx_lock(&bdlock); for (;;) { bd_request = 0; mtx_unlock(&bdlock); kthread_suspend_check(); /* * Save speedupreq for this pass and reset to capture new * requests. */ speedupreq = bd_speedupreq; bd_speedupreq = 0; /* * Flush each domain sequentially according to its level and * the speedup request. */ for (i = 0; i < buf_domains; i++) { bd = &bdomain[i]; if (speedupreq) lodirty = bd->bd_numdirtybuffers / 2; else lodirty = bd->bd_lodirtybuffers; while (bd->bd_numdirtybuffers > lodirty) { if (buf_flush(NULL, bd, bd->bd_numdirtybuffers - lodirty) == 0) break; kern_yield(PRI_USER); } } /* * Only clear bd_request if we have reached our low water * mark. The buf_daemon normally waits 1 second and * then incrementally flushes any dirty buffers that have * built up, within reason. * * If we were unable to hit our low water mark and couldn't * find any flushable buffers, we sleep for a short period * to avoid endless loops on unlockable buffers. */ mtx_lock(&bdlock); if (!BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) { /* * We reached our low water mark, reset the * request and sleep until we are needed again. * The sleep is just so the suspend code works. */ bd_request = 0; /* * Do an extra wakeup in case dirty threshold * changed via sysctl and the explicit transition * out of shortfall was missed. */ bdirtywakeup(); if (runningbufspace <= lorunningspace) runningwakeup(); msleep(&bd_request, &bdlock, PVM, "psleep", hz); } else { /* * We couldn't find any flushable dirty buffers but * still have too many dirty buffers, we * have to sleep and try again. (rare) */ msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); } } } /* * flushbufqueues: * * Try to flush a buffer in the dirty queue. We must be careful to * free up B_INVAL buffers instead of write them, which NFS is * particularly sensitive to. */ static int flushwithdeps = 0; SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW | CTLFLAG_STATS, &flushwithdeps, 0, "Number of buffers flushed with dependecies that require rollbacks"); static int flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target, int flushdeps) { struct bufqueue *bq; struct buf *sentinel; struct vnode *vp; struct mount *mp; struct buf *bp; int hasdeps; int flushed; int error; bool unlock; flushed = 0; bq = &bd->bd_dirtyq; bp = NULL; sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); sentinel->b_qindex = QUEUE_SENTINEL; BQ_LOCK(bq); TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist); BQ_UNLOCK(bq); while (flushed != target) { maybe_yield(); BQ_LOCK(bq); bp = TAILQ_NEXT(sentinel, b_freelist); if (bp != NULL) { TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist); TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel, b_freelist); } else { BQ_UNLOCK(bq); break; } /* * Skip sentinels inserted by other invocations of the * flushbufqueues(), taking care to not reorder them. * * Only flush the buffers that belong to the * vnode locked by the curthread. */ if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL && bp->b_vp != lvp)) { BQ_UNLOCK(bq); continue; } error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL); BQ_UNLOCK(bq); if (error != 0) continue; /* * BKGRDINPROG can only be set with the buf and bufobj * locks both held. We tolerate a race to clear it here. */ if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || (bp->b_flags & B_DELWRI) == 0) { BUF_UNLOCK(bp); continue; } if (bp->b_flags & B_INVAL) { bremfreef(bp); brelse(bp); flushed++; continue; } if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { if (flushdeps == 0) { BUF_UNLOCK(bp); continue; } hasdeps = 1; } else hasdeps = 0; /* * We must hold the lock on a vnode before writing * one of its buffers. Otherwise we may confuse, or * in the case of a snapshot vnode, deadlock the * system. * * The lock order here is the reverse of the normal * of vnode followed by buf lock. This is ok because * the NOWAIT will prevent deadlock. */ vp = bp->b_vp; if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { BUF_UNLOCK(bp); continue; } if (lvp == NULL) { unlock = true; error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT); } else { ASSERT_VOP_LOCKED(vp, "getbuf"); unlock = false; error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 : vn_lock(vp, LK_TRYUPGRADE); } if (error == 0) { CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); if (curproc == bufdaemonproc) { vfs_bio_awrite(bp); } else { bremfree(bp); bwrite(bp); counter_u64_add(notbufdflushes, 1); } vn_finished_write(mp); if (unlock) VOP_UNLOCK(vp, 0); flushwithdeps += hasdeps; flushed++; /* * Sleeping on runningbufspace while holding * vnode lock leads to deadlock. */ if (curproc == bufdaemonproc && runningbufspace > hirunningspace) waitrunningbufspace(); continue; } vn_finished_write(mp); BUF_UNLOCK(bp); } BQ_LOCK(bq); TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist); BQ_UNLOCK(bq); free(sentinel, M_TEMP); return (flushed); } /* * Check to see if a block is currently memory resident. */ struct buf * incore(struct bufobj *bo, daddr_t blkno) { struct buf *bp; BO_RLOCK(bo); bp = gbincore(bo, blkno); BO_RUNLOCK(bo); return (bp); } /* * Returns true if no I/O is needed to access the * associated VM object. This is like incore except * it also hunts around in the VM system for the data. */ static int inmem(struct vnode * vp, daddr_t blkno) { vm_object_t obj; vm_offset_t toff, tinc, size; vm_page_t m; vm_ooffset_t off; ASSERT_VOP_LOCKED(vp, "inmem"); if (incore(&vp->v_bufobj, blkno)) return 1; if (vp->v_mount == NULL) return 0; obj = vp->v_object; if (obj == NULL) return (0); size = PAGE_SIZE; if (size > vp->v_mount->mnt_stat.f_iosize) size = vp->v_mount->mnt_stat.f_iosize; off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; VM_OBJECT_RLOCK(obj); for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); if (!m) goto notinmem; tinc = size; if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); if (vm_page_is_valid(m, (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) goto notinmem; } VM_OBJECT_RUNLOCK(obj); return 1; notinmem: VM_OBJECT_RUNLOCK(obj); return (0); } /* * Set the dirty range for a buffer based on the status of the dirty * bits in the pages comprising the buffer. The range is limited * to the size of the buffer. * * Tell the VM system that the pages associated with this buffer * are clean. This is used for delayed writes where the data is * going to go to disk eventually without additional VM intevention. * * Note that while we only really need to clean through to b_bcount, we * just go ahead and clean through to b_bufsize. */ static void vfs_clean_pages_dirty_buf(struct buf *bp) { vm_ooffset_t foff, noff, eoff; vm_page_t m; int i; if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0) return; foff = bp->b_offset; KASSERT(bp->b_offset != NOOFFSET, ("vfs_clean_pages_dirty_buf: no buffer offset")); VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); - vfs_drain_busy_pages(bp); + vfs_busy_pages_acquire(bp); vfs_setdirty_locked_object(bp); for (i = 0; i < bp->b_npages; i++) { noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; eoff = noff; if (eoff > bp->b_offset + bp->b_bufsize) eoff = bp->b_offset + bp->b_bufsize; m = bp->b_pages[i]; vfs_page_set_validclean(bp, foff, m); /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ foff = noff; } + vfs_busy_pages_release(bp); VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); } static void vfs_setdirty_locked_object(struct buf *bp) { vm_object_t object; int i; object = bp->b_bufobj->bo_object; VM_OBJECT_ASSERT_WLOCKED(object); /* * We qualify the scan for modified pages on whether the * object has been flushed yet. */ if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) { vm_offset_t boffset; vm_offset_t eoffset; /* * test the pages to see if they have been modified directly * by users through the VM system. */ for (i = 0; i < bp->b_npages; i++) vm_page_test_dirty(bp->b_pages[i]); /* * Calculate the encompassing dirty range, boffset and eoffset, * (eoffset - boffset) bytes. */ for (i = 0; i < bp->b_npages; i++) { if (bp->b_pages[i]->dirty) break; } boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); for (i = bp->b_npages - 1; i >= 0; --i) { if (bp->b_pages[i]->dirty) { break; } } eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); /* * Fit it to the buffer. */ if (eoffset > bp->b_bcount) eoffset = bp->b_bcount; /* * If we have a good dirty range, merge with the existing * dirty range. */ if (boffset < eoffset) { if (bp->b_dirtyoff > boffset) bp->b_dirtyoff = boffset; if (bp->b_dirtyend < eoffset) bp->b_dirtyend = eoffset; } } } /* * Allocate the KVA mapping for an existing buffer. * If an unmapped buffer is provided but a mapped buffer is requested, take * also care to properly setup mappings between pages and KVA. */ static void bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags) { int bsize, maxsize, need_mapping, need_kva; off_t offset; need_mapping = bp->b_data == unmapped_buf && (gbflags & GB_UNMAPPED) == 0; need_kva = bp->b_kvabase == unmapped_buf && bp->b_data == unmapped_buf && (gbflags & GB_KVAALLOC) != 0; if (!need_mapping && !need_kva) return; BUF_CHECK_UNMAPPED(bp); if (need_mapping && bp->b_kvabase != unmapped_buf) { /* * Buffer is not mapped, but the KVA was already * reserved at the time of the instantiation. Use the * allocated space. */ goto has_addr; } /* * Calculate the amount of the address space we would reserve * if the buffer was mapped. */ bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize; KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize")); offset = blkno * bsize; maxsize = size + (offset & PAGE_MASK); maxsize = imax(maxsize, bsize); while (bufkva_alloc(bp, maxsize, gbflags) != 0) { if ((gbflags & GB_NOWAIT_BD) != 0) { /* * XXXKIB: defragmentation cannot * succeed, not sure what else to do. */ panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp); } counter_u64_add(mappingrestarts, 1); bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0); } has_addr: if (need_mapping) { /* b_offset is handled by bpmap_qenter. */ bp->b_data = bp->b_kvabase; BUF_CHECK_MAPPED(bp); bpmap_qenter(bp); } } struct buf * getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo, int flags) { struct buf *bp; int error; error = getblkx(vp, blkno, size, slpflag, slptimeo, flags, &bp); if (error != 0) return (NULL); return (bp); } /* * getblkx: * * Get a block given a specified block and offset into a file/device. * The buffers B_DONE bit will be cleared on return, making it almost * ready for an I/O initiation. B_INVAL may or may not be set on * return. The caller should clear B_INVAL prior to initiating a * READ. * * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for * an existing buffer. * * For a VMIO buffer, B_CACHE is modified according to the backing VM. * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set * and then cleared based on the backing VM. If the previous buffer is * non-0-sized but invalid, B_CACHE will be cleared. * * If getblk() must create a new buffer, the new buffer is returned with * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which * case it is returned with B_INVAL clear and B_CACHE set based on the * backing VM. * * getblk() also forces a bwrite() for any B_DELWRI buffer whos * B_CACHE bit is clear. * * What this means, basically, is that the caller should use B_CACHE to * determine whether the buffer is fully valid or not and should clear * B_INVAL prior to issuing a read. If the caller intends to validate * the buffer by loading its data area with something, the caller needs * to clear B_INVAL. If the caller does this without issuing an I/O, * the caller should set B_CACHE ( as an optimization ), else the caller * should issue the I/O and biodone() will set B_CACHE if the I/O was * a write attempt or if it was a successful read. If the caller * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR * prior to issuing the READ. biodone() will *not* clear B_INVAL. */ int getblkx(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo, int flags, struct buf **bpp) { struct buf *bp; struct bufobj *bo; daddr_t d_blkno; int bsize, error, maxsize, vmio; off_t offset; CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); ASSERT_VOP_LOCKED(vp, "getblk"); if (size > maxbcachebuf) panic("getblk: size(%d) > maxbcachebuf(%d)\n", size, maxbcachebuf); if (!unmapped_buf_allowed) flags &= ~(GB_UNMAPPED | GB_KVAALLOC); bo = &vp->v_bufobj; d_blkno = blkno; loop: BO_RLOCK(bo); bp = gbincore(bo, blkno); if (bp != NULL) { int lockflags; /* * Buffer is in-core. If the buffer is not busy nor managed, * it must be on a queue. */ lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; if ((flags & GB_LOCK_NOWAIT) != 0) lockflags |= LK_NOWAIT; error = BUF_TIMELOCK(bp, lockflags, BO_LOCKPTR(bo), "getblk", slpflag, slptimeo); /* * If we slept and got the lock we have to restart in case * the buffer changed identities. */ if (error == ENOLCK) goto loop; /* We timed out or were interrupted. */ else if (error != 0) return (error); /* If recursed, assume caller knows the rules. */ else if (BUF_LOCKRECURSED(bp)) goto end; /* * The buffer is locked. B_CACHE is cleared if the buffer is * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set * and for a VMIO buffer B_CACHE is adjusted according to the * backing VM cache. */ if (bp->b_flags & B_INVAL) bp->b_flags &= ~B_CACHE; else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) bp->b_flags |= B_CACHE; if (bp->b_flags & B_MANAGED) MPASS(bp->b_qindex == QUEUE_NONE); else bremfree(bp); /* * check for size inconsistencies for non-VMIO case. */ if (bp->b_bcount != size) { if ((bp->b_flags & B_VMIO) == 0 || (size > bp->b_kvasize)) { if (bp->b_flags & B_DELWRI) { bp->b_flags |= B_NOCACHE; bwrite(bp); } else { if (LIST_EMPTY(&bp->b_dep)) { bp->b_flags |= B_RELBUF; brelse(bp); } else { bp->b_flags |= B_NOCACHE; bwrite(bp); } } goto loop; } } /* * Handle the case of unmapped buffer which should * become mapped, or the buffer for which KVA * reservation is requested. */ bp_unmapped_get_kva(bp, blkno, size, flags); /* * If the size is inconsistent in the VMIO case, we can resize * the buffer. This might lead to B_CACHE getting set or * cleared. If the size has not changed, B_CACHE remains * unchanged from its previous state. */ allocbuf(bp, size); KASSERT(bp->b_offset != NOOFFSET, ("getblk: no buffer offset")); /* * A buffer with B_DELWRI set and B_CACHE clear must * be committed before we can return the buffer in * order to prevent the caller from issuing a read * ( due to B_CACHE not being set ) and overwriting * it. * * Most callers, including NFS and FFS, need this to * operate properly either because they assume they * can issue a read if B_CACHE is not set, or because * ( for example ) an uncached B_DELWRI might loop due * to softupdates re-dirtying the buffer. In the latter * case, B_CACHE is set after the first write completes, * preventing further loops. * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE * above while extending the buffer, we cannot allow the * buffer to remain with B_CACHE set after the write * completes or it will represent a corrupt state. To * deal with this we set B_NOCACHE to scrap the buffer * after the write. * * We might be able to do something fancy, like setting * B_CACHE in bwrite() except if B_DELWRI is already set, * so the below call doesn't set B_CACHE, but that gets real * confusing. This is much easier. */ if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { bp->b_flags |= B_NOCACHE; bwrite(bp); goto loop; } bp->b_flags &= ~B_DONE; } else { /* * Buffer is not in-core, create new buffer. The buffer * returned by getnewbuf() is locked. Note that the returned * buffer is also considered valid (not marked B_INVAL). */ BO_RUNLOCK(bo); /* * If the user does not want us to create the buffer, bail out * here. */ if (flags & GB_NOCREAT) return (EEXIST); bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize; KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize")); offset = blkno * bsize; vmio = vp->v_object != NULL; if (vmio) { maxsize = size + (offset & PAGE_MASK); } else { maxsize = size; /* Do not allow non-VMIO notmapped buffers. */ flags &= ~(GB_UNMAPPED | GB_KVAALLOC); } maxsize = imax(maxsize, bsize); if ((flags & GB_NOSPARSE) != 0 && vmio && !vn_isdisk(vp, NULL)) { error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0); KASSERT(error != EOPNOTSUPP, ("GB_NOSPARSE from fs not supporting bmap, vp %p", vp)); if (error != 0) return (error); if (d_blkno == -1) return (EJUSTRETURN); } bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags); if (bp == NULL) { if (slpflag || slptimeo) return (ETIMEDOUT); /* * XXX This is here until the sleep path is diagnosed * enough to work under very low memory conditions. * * There's an issue on low memory, 4BSD+non-preempt * systems (eg MIPS routers with 32MB RAM) where buffer * exhaustion occurs without sleeping for buffer * reclaimation. This just sticks in a loop and * constantly attempts to allocate a buffer, which * hits exhaustion and tries to wakeup bufdaemon. * This never happens because we never yield. * * The real solution is to identify and fix these cases * so we aren't effectively busy-waiting in a loop * until the reclaimation path has cycles to run. */ kern_yield(PRI_USER); goto loop; } /* * This code is used to make sure that a buffer is not * created while the getnewbuf routine is blocked. * This can be a problem whether the vnode is locked or not. * If the buffer is created out from under us, we have to * throw away the one we just created. * * Note: this must occur before we associate the buffer * with the vp especially considering limitations in * the splay tree implementation when dealing with duplicate * lblkno's. */ BO_LOCK(bo); if (gbincore(bo, blkno)) { BO_UNLOCK(bo); bp->b_flags |= B_INVAL; bufspace_release(bufdomain(bp), maxsize); brelse(bp); goto loop; } /* * Insert the buffer into the hash, so that it can * be found by incore. */ bp->b_lblkno = blkno; bp->b_blkno = d_blkno; bp->b_offset = offset; bgetvp(vp, bp); BO_UNLOCK(bo); /* * set B_VMIO bit. allocbuf() the buffer bigger. Since the * buffer size starts out as 0, B_CACHE will be set by * allocbuf() for the VMIO case prior to it testing the * backing store for validity. */ if (vmio) { bp->b_flags |= B_VMIO; KASSERT(vp->v_object == bp->b_bufobj->bo_object, ("ARGH! different b_bufobj->bo_object %p %p %p\n", bp, vp->v_object, bp->b_bufobj->bo_object)); } else { bp->b_flags &= ~B_VMIO; KASSERT(bp->b_bufobj->bo_object == NULL, ("ARGH! has b_bufobj->bo_object %p %p\n", bp, bp->b_bufobj->bo_object)); BUF_CHECK_MAPPED(bp); } allocbuf(bp, size); bufspace_release(bufdomain(bp), maxsize); bp->b_flags &= ~B_DONE; } CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); end: buf_track(bp, __func__); KASSERT(bp->b_bufobj == bo, ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); *bpp = bp; return (0); } /* * Get an empty, disassociated buffer of given size. The buffer is initially * set to B_INVAL. */ struct buf * geteblk(int size, int flags) { struct buf *bp; int maxsize; maxsize = (size + BKVAMASK) & ~BKVAMASK; while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) { if ((flags & GB_NOWAIT_BD) && (curthread->td_pflags & TDP_BUFNEED) != 0) return (NULL); } allocbuf(bp, size); bufspace_release(bufdomain(bp), maxsize); bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ return (bp); } /* * Truncate the backing store for a non-vmio buffer. */ static void vfs_nonvmio_truncate(struct buf *bp, int newbsize) { if (bp->b_flags & B_MALLOC) { /* * malloced buffers are not shrunk */ if (newbsize == 0) { bufmallocadjust(bp, 0); free(bp->b_data, M_BIOBUF); bp->b_data = bp->b_kvabase; bp->b_flags &= ~B_MALLOC; } return; } vm_hold_free_pages(bp, newbsize); bufspace_adjust(bp, newbsize); } /* * Extend the backing for a non-VMIO buffer. */ static void vfs_nonvmio_extend(struct buf *bp, int newbsize) { caddr_t origbuf; int origbufsize; /* * We only use malloced memory on the first allocation. * and revert to page-allocated memory when the buffer * grows. * * There is a potential smp race here that could lead * to bufmallocspace slightly passing the max. It * is probably extremely rare and not worth worrying * over. */ if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 && bufmallocspace < maxbufmallocspace) { bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK); bp->b_flags |= B_MALLOC; bufmallocadjust(bp, newbsize); return; } /* * If the buffer is growing on its other-than-first * allocation then we revert to the page-allocation * scheme. */ origbuf = NULL; origbufsize = 0; if (bp->b_flags & B_MALLOC) { origbuf = bp->b_data; origbufsize = bp->b_bufsize; bp->b_data = bp->b_kvabase; bufmallocadjust(bp, 0); bp->b_flags &= ~B_MALLOC; newbsize = round_page(newbsize); } vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize, (vm_offset_t) bp->b_data + newbsize); if (origbuf != NULL) { bcopy(origbuf, bp->b_data, origbufsize); free(origbuf, M_BIOBUF); } bufspace_adjust(bp, newbsize); } /* * This code constitutes the buffer memory from either anonymous system * memory (in the case of non-VMIO operations) or from an associated * VM object (in the case of VMIO operations). This code is able to * resize a buffer up or down. * * Note that this code is tricky, and has many complications to resolve * deadlock or inconsistent data situations. Tread lightly!!! * There are B_CACHE and B_DELWRI interactions that must be dealt with by * the caller. Calling this code willy nilly can result in the loss of data. * * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with * B_CACHE for the non-VMIO case. */ int allocbuf(struct buf *bp, int size) { int newbsize; if (bp->b_bcount == size) return (1); if (bp->b_kvasize != 0 && bp->b_kvasize < size) panic("allocbuf: buffer too small"); newbsize = roundup2(size, DEV_BSIZE); if ((bp->b_flags & B_VMIO) == 0) { if ((bp->b_flags & B_MALLOC) == 0) newbsize = round_page(newbsize); /* * Just get anonymous memory from the kernel. Don't * mess with B_CACHE. */ if (newbsize < bp->b_bufsize) vfs_nonvmio_truncate(bp, newbsize); else if (newbsize > bp->b_bufsize) vfs_nonvmio_extend(bp, newbsize); } else { int desiredpages; desiredpages = (size == 0) ? 0 : num_pages((bp->b_offset & PAGE_MASK) + newbsize); if (bp->b_flags & B_MALLOC) panic("allocbuf: VMIO buffer can't be malloced"); /* * Set B_CACHE initially if buffer is 0 length or will become * 0-length. */ if (size == 0 || bp->b_bufsize == 0) bp->b_flags |= B_CACHE; if (newbsize < bp->b_bufsize) vfs_vmio_truncate(bp, desiredpages); /* XXX This looks as if it should be newbsize > b_bufsize */ else if (size > bp->b_bcount) vfs_vmio_extend(bp, desiredpages, size); bufspace_adjust(bp, newbsize); } bp->b_bcount = size; /* requested buffer size. */ return (1); } extern int inflight_transient_maps; static struct bio_queue nondump_bios; void biodone(struct bio *bp) { struct mtx *mtxp; void (*done)(struct bio *); vm_offset_t start, end; biotrack(bp, __func__); /* * Avoid completing I/O when dumping after a panic since that may * result in a deadlock in the filesystem or pager code. Note that * this doesn't affect dumps that were started manually since we aim * to keep the system usable after it has been resumed. */ if (__predict_false(dumping && SCHEDULER_STOPPED())) { TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue); return; } if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) { bp->bio_flags &= ~BIO_TRANSIENT_MAPPING; bp->bio_flags |= BIO_UNMAPPED; start = trunc_page((vm_offset_t)bp->bio_data); end = round_page((vm_offset_t)bp->bio_data + bp->bio_length); bp->bio_data = unmapped_buf; pmap_qremove(start, atop(end - start)); vmem_free(transient_arena, start, end - start); atomic_add_int(&inflight_transient_maps, -1); } done = bp->bio_done; if (done == NULL) { mtxp = mtx_pool_find(mtxpool_sleep, bp); mtx_lock(mtxp); bp->bio_flags |= BIO_DONE; wakeup(bp); mtx_unlock(mtxp); } else done(bp); } /* * Wait for a BIO to finish. */ int biowait(struct bio *bp, const char *wchan) { struct mtx *mtxp; mtxp = mtx_pool_find(mtxpool_sleep, bp); mtx_lock(mtxp); while ((bp->bio_flags & BIO_DONE) == 0) msleep(bp, mtxp, PRIBIO, wchan, 0); mtx_unlock(mtxp); if (bp->bio_error != 0) return (bp->bio_error); if (!(bp->bio_flags & BIO_ERROR)) return (0); return (EIO); } void biofinish(struct bio *bp, struct devstat *stat, int error) { if (error) { bp->bio_error = error; bp->bio_flags |= BIO_ERROR; } if (stat != NULL) devstat_end_transaction_bio(stat, bp); biodone(bp); } #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING) void biotrack_buf(struct bio *bp, const char *location) { buf_track(bp->bio_track_bp, location); } #endif /* * bufwait: * * Wait for buffer I/O completion, returning error status. The buffer * is left locked and B_DONE on return. B_EINTR is converted into an EINTR * error and cleared. */ int bufwait(struct buf *bp) { if (bp->b_iocmd == BIO_READ) bwait(bp, PRIBIO, "biord"); else bwait(bp, PRIBIO, "biowr"); if (bp->b_flags & B_EINTR) { bp->b_flags &= ~B_EINTR; return (EINTR); } if (bp->b_ioflags & BIO_ERROR) { return (bp->b_error ? bp->b_error : EIO); } else { return (0); } } /* * bufdone: * * Finish I/O on a buffer, optionally calling a completion function. * This is usually called from an interrupt so process blocking is * not allowed. * * biodone is also responsible for setting B_CACHE in a B_VMIO bp. * In a non-VMIO bp, B_CACHE will be set on the next getblk() * assuming B_INVAL is clear. * * For the VMIO case, we set B_CACHE if the op was a read and no * read error occurred, or if the op was a write. B_CACHE is never * set if the buffer is invalid or otherwise uncacheable. * * bufdone does not mess with B_INVAL, allowing the I/O routine or the * initiator to leave B_INVAL set to brelse the buffer out of existence * in the biodone routine. */ void bufdone(struct buf *bp) { struct bufobj *dropobj; void (*biodone)(struct buf *); buf_track(bp, __func__); CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); dropobj = NULL; KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); runningbufwakeup(bp); if (bp->b_iocmd == BIO_WRITE) dropobj = bp->b_bufobj; /* call optional completion function if requested */ if (bp->b_iodone != NULL) { biodone = bp->b_iodone; bp->b_iodone = NULL; (*biodone) (bp); if (dropobj) bufobj_wdrop(dropobj); return; } if (bp->b_flags & B_VMIO) { /* * Set B_CACHE if the op was a normal read and no error * occurred. B_CACHE is set for writes in the b*write() * routines. */ if (bp->b_iocmd == BIO_READ && !(bp->b_flags & (B_INVAL|B_NOCACHE)) && !(bp->b_ioflags & BIO_ERROR)) bp->b_flags |= B_CACHE; vfs_vmio_iodone(bp); } if (!LIST_EMPTY(&bp->b_dep)) buf_complete(bp); if ((bp->b_flags & B_CKHASH) != 0) { KASSERT(bp->b_iocmd == BIO_READ, ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd)); KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp)); (*bp->b_ckhashcalc)(bp); } /* * For asynchronous completions, release the buffer now. The brelse * will do a wakeup there if necessary - so no need to do a wakeup * here in the async case. The sync case always needs to do a wakeup. */ if (bp->b_flags & B_ASYNC) { if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) brelse(bp); else bqrelse(bp); } else bdone(bp); if (dropobj) bufobj_wdrop(dropobj); } /* * This routine is called in lieu of iodone in the case of * incomplete I/O. This keeps the busy status for pages * consistent. */ void vfs_unbusy_pages(struct buf *bp) { int i; vm_object_t obj; vm_page_t m; runningbufwakeup(bp); if (!(bp->b_flags & B_VMIO)) return; obj = bp->b_bufobj->bo_object; VM_OBJECT_WLOCK(obj); for (i = 0; i < bp->b_npages; i++) { m = bp->b_pages[i]; if (m == bogus_page) { m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); if (!m) panic("vfs_unbusy_pages: page missing\n"); bp->b_pages[i] = m; if (buf_mapped(bp)) { BUF_CHECK_MAPPED(bp); pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); } else BUF_CHECK_UNMAPPED(bp); } vm_page_sunbusy(m); } vm_object_pip_wakeupn(obj, bp->b_npages); VM_OBJECT_WUNLOCK(obj); } /* * vfs_page_set_valid: * * Set the valid bits in a page based on the supplied offset. The * range is restricted to the buffer's size. * * This routine is typically called after a read completes. */ static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) { vm_ooffset_t eoff; /* * Compute the end offset, eoff, such that [off, eoff) does not span a * page boundary and eoff is not greater than the end of the buffer. * The end of the buffer, in this case, is our file EOF, not the * allocation size of the buffer. */ eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; if (eoff > bp->b_offset + bp->b_bcount) eoff = bp->b_offset + bp->b_bcount; /* * Set valid range. This is typically the entire buffer and thus the * entire page. */ if (eoff > off) vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off); } /* * vfs_page_set_validclean: * * Set the valid bits and clear the dirty bits in a page based on the * supplied offset. The range is restricted to the buffer's size. */ static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) { vm_ooffset_t soff, eoff; /* * Start and end offsets in buffer. eoff - soff may not cross a * page boundary or cross the end of the buffer. The end of the * buffer, in this case, is our file EOF, not the allocation size * of the buffer. */ soff = off; eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; if (eoff > bp->b_offset + bp->b_bcount) eoff = bp->b_offset + bp->b_bcount; /* * Set valid range. This is typically the entire buffer and thus the * entire page. */ if (eoff > soff) { vm_page_set_validclean( m, (vm_offset_t) (soff & PAGE_MASK), (vm_offset_t) (eoff - soff) ); } } /* - * Ensure that all buffer pages are not exclusive busied. If any page is - * exclusive busy, drain it. + * Acquire a shared busy on all pages in the buf. */ void -vfs_drain_busy_pages(struct buf *bp) +vfs_busy_pages_acquire(struct buf *bp) { - vm_page_t m; - int i, last_busied; + int i; VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object); - last_busied = 0; - for (i = 0; i < bp->b_npages; i++) { - m = bp->b_pages[i]; - if (vm_page_xbusied(m)) { - for (; last_busied < i; last_busied++) - vm_page_sbusy(bp->b_pages[last_busied]); - while (vm_page_xbusied(m)) { - vm_page_sleep_if_xbusy(m, "vbpage"); - } - } - } - for (i = 0; i < last_busied; i++) + for (i = 0; i < bp->b_npages; i++) + vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY); +} + +void +vfs_busy_pages_release(struct buf *bp) +{ + int i; + + VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object); + for (i = 0; i < bp->b_npages; i++) vm_page_sunbusy(bp->b_pages[i]); } /* * This routine is called before a device strategy routine. * It is used to tell the VM system that paging I/O is in * progress, and treat the pages associated with the buffer * almost as being exclusive busy. Also the object paging_in_progress * flag is handled to make sure that the object doesn't become * inconsistent. * * Since I/O has not been initiated yet, certain buffer flags * such as BIO_ERROR or B_INVAL may be in an inconsistent state * and should be ignored. */ void vfs_busy_pages(struct buf *bp, int clear_modify) { vm_object_t obj; vm_ooffset_t foff; vm_page_t m; int i; bool bogus; if (!(bp->b_flags & B_VMIO)) return; obj = bp->b_bufobj->bo_object; foff = bp->b_offset; KASSERT(bp->b_offset != NOOFFSET, ("vfs_busy_pages: no buffer offset")); VM_OBJECT_WLOCK(obj); - vfs_drain_busy_pages(bp); + if ((bp->b_flags & B_CLUSTER) == 0) { + vm_object_pip_add(obj, bp->b_npages); + vfs_busy_pages_acquire(bp); + } if (bp->b_bufsize != 0) vfs_setdirty_locked_object(bp); bogus = false; for (i = 0; i < bp->b_npages; i++) { m = bp->b_pages[i]; + vm_page_assert_sbusied(m); - if ((bp->b_flags & B_CLUSTER) == 0) { - vm_object_pip_add(obj, 1); - vm_page_sbusy(m); - } /* * When readying a buffer for a read ( i.e * clear_modify == 0 ), it is important to do * bogus_page replacement for valid pages in * partially instantiated buffers. Partially * instantiated buffers can, in turn, occur when * reconstituting a buffer from its VM backing store * base. We only have to do this if B_CACHE is * clear ( which causes the I/O to occur in the * first place ). The replacement prevents the read * I/O from overwriting potentially dirty VM-backed * pages. XXX bogus page replacement is, uh, bogus. * It may not work properly with small-block devices. * We need to find a better way. */ if (clear_modify) { pmap_remove_write(m); vfs_page_set_validclean(bp, foff, m); } else if (m->valid == VM_PAGE_BITS_ALL && (bp->b_flags & B_CACHE) == 0) { bp->b_pages[i] = bogus_page; bogus = true; } foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; } VM_OBJECT_WUNLOCK(obj); if (bogus && buf_mapped(bp)) { BUF_CHECK_MAPPED(bp); pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); } } /* * vfs_bio_set_valid: * * Set the range within the buffer to valid. The range is * relative to the beginning of the buffer, b_offset. Note that * b_offset itself may be offset from the beginning of the first * page. */ void vfs_bio_set_valid(struct buf *bp, int base, int size) { int i, n; vm_page_t m; if (!(bp->b_flags & B_VMIO)) return; /* * Fixup base to be relative to beginning of first page. * Set initial n to be the maximum number of bytes in the * first page that can be validated. */ base += (bp->b_offset & PAGE_MASK); n = PAGE_SIZE - (base & PAGE_MASK); VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { m = bp->b_pages[i]; if (n > size) n = size; vm_page_set_valid_range(m, base & PAGE_MASK, n); base += n; size -= n; n = PAGE_SIZE; } VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); } /* * vfs_bio_clrbuf: * * If the specified buffer is a non-VMIO buffer, clear the entire * buffer. If the specified buffer is a VMIO buffer, clear and * validate only the previously invalid portions of the buffer. * This routine essentially fakes an I/O, so we need to clear * BIO_ERROR and B_INVAL. * * Note that while we only theoretically need to clear through b_bcount, * we go ahead and clear through b_bufsize. */ void vfs_bio_clrbuf(struct buf *bp) { int i, j, mask, sa, ea, slide; if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { clrbuf(bp); return; } bp->b_flags &= ~B_INVAL; bp->b_ioflags &= ~BIO_ERROR; VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && (bp->b_offset & PAGE_MASK) == 0) { if (bp->b_pages[0] == bogus_page) goto unlock; mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object); if ((bp->b_pages[0]->valid & mask) == mask) goto unlock; if ((bp->b_pages[0]->valid & mask) == 0) { pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize); bp->b_pages[0]->valid |= mask; goto unlock; } } sa = bp->b_offset & PAGE_MASK; slide = 0; for (i = 0; i < bp->b_npages; i++, sa = 0) { slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize); ea = slide & PAGE_MASK; if (ea == 0) ea = PAGE_SIZE; if (bp->b_pages[i] == bogus_page) continue; j = sa / DEV_BSIZE; mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object); if ((bp->b_pages[i]->valid & mask) == mask) continue; if ((bp->b_pages[i]->valid & mask) == 0) pmap_zero_page_area(bp->b_pages[i], sa, ea - sa); else { for (; sa < ea; sa += DEV_BSIZE, j++) { if ((bp->b_pages[i]->valid & (1 << j)) == 0) { pmap_zero_page_area(bp->b_pages[i], sa, DEV_BSIZE); } } } bp->b_pages[i]->valid |= mask; } unlock: VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); bp->b_resid = 0; } void vfs_bio_bzero_buf(struct buf *bp, int base, int size) { vm_page_t m; int i, n; if (buf_mapped(bp)) { BUF_CHECK_MAPPED(bp); bzero(bp->b_data + base, size); } else { BUF_CHECK_UNMAPPED(bp); n = PAGE_SIZE - (base & PAGE_MASK); for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { m = bp->b_pages[i]; if (n > size) n = size; pmap_zero_page_area(m, base & PAGE_MASK, n); base += n; size -= n; n = PAGE_SIZE; } } } /* * Update buffer flags based on I/O request parameters, optionally releasing the * buffer. If it's VMIO or direct I/O, the buffer pages are released to the VM, * where they may be placed on a page queue (VMIO) or freed immediately (direct * I/O). Otherwise the buffer is released to the cache. */ static void b_io_dismiss(struct buf *bp, int ioflag, bool release) { KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0, ("buf %p non-VMIO noreuse", bp)); if ((ioflag & IO_DIRECT) != 0) bp->b_flags |= B_DIRECT; if ((ioflag & IO_EXT) != 0) bp->b_xflags |= BX_ALTDATA; if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) { bp->b_flags |= B_RELBUF; if ((ioflag & IO_NOREUSE) != 0) bp->b_flags |= B_NOREUSE; if (release) brelse(bp); } else if (release) bqrelse(bp); } void vfs_bio_brelse(struct buf *bp, int ioflag) { b_io_dismiss(bp, ioflag, true); } void vfs_bio_set_flags(struct buf *bp, int ioflag) { b_io_dismiss(bp, ioflag, false); } /* * vm_hold_load_pages and vm_hold_free_pages get pages into * a buffers address space. The pages are anonymous and are * not associated with a file object. */ static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) { vm_offset_t pg; vm_page_t p; int index; BUF_CHECK_MAPPED(bp); to = round_page(to); from = round_page(from); index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; for (pg = from; pg < to; pg += PAGE_SIZE, index++) { /* * note: must allocate system pages since blocking here * could interfere with paging I/O, no matter which * process we are. */ p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) | VM_ALLOC_WAITOK); pmap_qenter(pg, &p, 1); bp->b_pages[index] = p; } bp->b_npages = index; } /* Return pages associated with this buf to the vm system */ static void vm_hold_free_pages(struct buf *bp, int newbsize) { vm_offset_t from; vm_page_t p; int index, newnpages; BUF_CHECK_MAPPED(bp); from = round_page((vm_offset_t)bp->b_data + newbsize); newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; if (bp->b_npages > newnpages) pmap_qremove(from, bp->b_npages - newnpages); for (index = newnpages; index < bp->b_npages; index++) { p = bp->b_pages[index]; bp->b_pages[index] = NULL; vm_page_unwire_noq(p); vm_page_free(p); } bp->b_npages = newnpages; } /* * Map an IO request into kernel virtual address space. * * All requests are (re)mapped into kernel VA space. * Notice that we use b_bufsize for the size of the buffer * to be mapped. b_bcount might be modified by the driver. * * Note that even if the caller determines that the address space should * be valid, a race or a smaller-file mapped into a larger space may * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST * check the return value. * * This function only works with pager buffers. */ int vmapbuf(struct buf *bp, int mapbuf) { vm_prot_t prot; int pidx; if (bp->b_bufsize < 0) return (-1); prot = VM_PROT_READ; if (bp->b_iocmd == BIO_READ) prot |= VM_PROT_WRITE; /* Less backwards than it looks */ if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages, btoc(MAXPHYS))) < 0) return (-1); bp->b_npages = pidx; bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK; if (mapbuf || !unmapped_buf_allowed) { pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx); bp->b_data = bp->b_kvabase + bp->b_offset; } else bp->b_data = unmapped_buf; return(0); } /* * Free the io map PTEs associated with this IO operation. * We also invalidate the TLB entries and restore the original b_addr. * * This function only works with pager buffers. */ void vunmapbuf(struct buf *bp) { int npages; npages = bp->b_npages; if (buf_mapped(bp)) pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); vm_page_unhold_pages(bp->b_pages, npages); bp->b_data = unmapped_buf; } void bdone(struct buf *bp) { struct mtx *mtxp; mtxp = mtx_pool_find(mtxpool_sleep, bp); mtx_lock(mtxp); bp->b_flags |= B_DONE; wakeup(bp); mtx_unlock(mtxp); } void bwait(struct buf *bp, u_char pri, const char *wchan) { struct mtx *mtxp; mtxp = mtx_pool_find(mtxpool_sleep, bp); mtx_lock(mtxp); while ((bp->b_flags & B_DONE) == 0) msleep(bp, mtxp, pri, wchan, 0); mtx_unlock(mtxp); } int bufsync(struct bufobj *bo, int waitfor) { return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread)); } void bufstrategy(struct bufobj *bo, struct buf *bp) { int i __unused; struct vnode *vp; vp = bp->b_vp; KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); i = VOP_STRATEGY(vp, bp); KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); } /* * Initialize a struct bufobj before use. Memory is assumed zero filled. */ void bufobj_init(struct bufobj *bo, void *private) { static volatile int bufobj_cleanq; bo->bo_domain = atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains; rw_init(BO_LOCKPTR(bo), "bufobj interlock"); bo->bo_private = private; TAILQ_INIT(&bo->bo_clean.bv_hd); TAILQ_INIT(&bo->bo_dirty.bv_hd); } void bufobj_wrefl(struct bufobj *bo) { KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); ASSERT_BO_WLOCKED(bo); bo->bo_numoutput++; } void bufobj_wref(struct bufobj *bo) { KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); BO_LOCK(bo); bo->bo_numoutput++; BO_UNLOCK(bo); } void bufobj_wdrop(struct bufobj *bo) { KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); BO_LOCK(bo); KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { bo->bo_flag &= ~BO_WWAIT; wakeup(&bo->bo_numoutput); } BO_UNLOCK(bo); } int bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) { int error; KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); ASSERT_BO_WLOCKED(bo); error = 0; while (bo->bo_numoutput) { bo->bo_flag |= BO_WWAIT; error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo), slpflag | (PRIBIO + 1), "bo_wwait", timeo); if (error) break; } return (error); } /* * Set bio_data or bio_ma for struct bio from the struct buf. */ void bdata2bio(struct buf *bp, struct bio *bip) { if (!buf_mapped(bp)) { KASSERT(unmapped_buf_allowed, ("unmapped")); bip->bio_ma = bp->b_pages; bip->bio_ma_n = bp->b_npages; bip->bio_data = unmapped_buf; bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; bip->bio_flags |= BIO_UNMAPPED; KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) / PAGE_SIZE == bp->b_npages, ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset, (long long)bip->bio_length, bip->bio_ma_n)); } else { bip->bio_data = bp->b_data; bip->bio_ma = NULL; } } /* * The MIPS pmap code currently doesn't handle aliased pages. * The VIPT caches may not handle page aliasing themselves, leading * to data corruption. * * As such, this code makes a system extremely unhappy if said * system doesn't support unaliasing the above situation in hardware. * Some "recent" systems (eg some mips24k/mips74k cores) don't enable * this feature at build time, so it has to be handled in software. * * Once the MIPS pmap/cache code grows to support this function on * earlier chips, it should be flipped back off. */ #ifdef __mips__ static int buf_pager_relbuf = 1; #else static int buf_pager_relbuf = 0; #endif SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN, &buf_pager_relbuf, 0, "Make buffer pager release buffers after reading"); /* * The buffer pager. It uses buffer reads to validate pages. * * In contrast to the generic local pager from vm/vnode_pager.c, this * pager correctly and easily handles volumes where the underlying * device block size is greater than the machine page size. The * buffer cache transparently extends the requested page run to be * aligned at the block boundary, and does the necessary bogus page * replacements in the addends to avoid obliterating already valid * pages. * * The only non-trivial issue is that the exclusive busy state for * pages, which is assumed by the vm_pager_getpages() interface, is * incompatible with the VMIO buffer cache's desire to share-busy the * pages. This function performs a trivial downgrade of the pages' * state before reading buffers, and a less trivial upgrade from the * shared-busy to excl-busy state after the read. */ int vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count, int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno, vbg_get_blksize_t get_blksize) { vm_page_t m; vm_object_t object; struct buf *bp; struct mount *mp; daddr_t lbn, lbnp; vm_ooffset_t la, lb, poff, poffe; long bsize; int bo_bs, br_flags, error, i, pgsin, pgsin_a, pgsin_b; bool redo, lpart; object = vp->v_object; mp = vp->v_mount; error = 0; la = IDX_TO_OFF(ma[count - 1]->pindex); if (la >= object->un_pager.vnp.vnp_size) return (VM_PAGER_BAD); /* * Change the meaning of la from where the last requested page starts * to where it ends, because that's the end of the requested region * and the start of the potential read-ahead region. */ la += PAGE_SIZE; lpart = la > object->un_pager.vnp.vnp_size; bo_bs = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex))); /* * Calculate read-ahead, behind and total pages. */ pgsin = count; lb = IDX_TO_OFF(ma[0]->pindex); pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs)); pgsin += pgsin_b; if (rbehind != NULL) *rbehind = pgsin_b; pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la); if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size) pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size, PAGE_SIZE) - la); pgsin += pgsin_a; if (rahead != NULL) *rahead = pgsin_a; VM_CNT_INC(v_vnodein); VM_CNT_ADD(v_vnodepgsin, pgsin); br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS) != 0) ? GB_UNMAPPED : 0; VM_OBJECT_WLOCK(object); again: for (i = 0; i < count; i++) vm_page_busy_downgrade(ma[i]); VM_OBJECT_WUNLOCK(object); lbnp = -1; for (i = 0; i < count; i++) { m = ma[i]; /* * Pages are shared busy and the object lock is not * owned, which together allow for the pages' * invalidation. The racy test for validity avoids * useless creation of the buffer for the most typical * case when invalidation is not used in redo or for * parallel read. The shared->excl upgrade loop at * the end of the function catches the race in a * reliable way (protected by the object lock). */ if (m->valid == VM_PAGE_BITS_ALL) continue; poff = IDX_TO_OFF(m->pindex); poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size); for (; poff < poffe; poff += bsize) { lbn = get_lblkno(vp, poff); if (lbn == lbnp) goto next_page; lbnp = lbn; bsize = get_blksize(vp, lbn); error = bread_gb(vp, lbn, bsize, curthread->td_ucred, br_flags, &bp); if (error != 0) goto end_pages; if (LIST_EMPTY(&bp->b_dep)) { /* * Invalidation clears m->valid, but * may leave B_CACHE flag if the * buffer existed at the invalidation * time. In this case, recycle the * buffer to do real read on next * bread() after redo. * * Otherwise B_RELBUF is not strictly * necessary, enable to reduce buf * cache pressure. */ if (buf_pager_relbuf || m->valid != VM_PAGE_BITS_ALL) bp->b_flags |= B_RELBUF; bp->b_flags &= ~B_NOCACHE; brelse(bp); } else { bqrelse(bp); } } KASSERT(1 /* racy, enable for debugging */ || m->valid == VM_PAGE_BITS_ALL || i == count - 1, ("buf %d %p invalid", i, m)); if (i == count - 1 && lpart) { VM_OBJECT_WLOCK(object); if (m->valid != 0 && m->valid != VM_PAGE_BITS_ALL) vm_page_zero_invalid(m, TRUE); VM_OBJECT_WUNLOCK(object); } next_page:; } end_pages: VM_OBJECT_WLOCK(object); redo = false; for (i = 0; i < count; i++) { vm_page_sunbusy(ma[i]); ma[i] = vm_page_grab(object, ma[i]->pindex, VM_ALLOC_NORMAL); /* * Since the pages were only sbusy while neither the * buffer nor the object lock was held by us, or * reallocated while vm_page_grab() slept for busy * relinguish, they could have been invalidated. * Recheck the valid bits and re-read as needed. * * Note that the last page is made fully valid in the * read loop, and partial validity for the page at * index count - 1 could mean that the page was * invalidated or removed, so we must restart for * safety as well. */ if (ma[i]->valid != VM_PAGE_BITS_ALL) redo = true; } if (redo && error == 0) goto again; VM_OBJECT_WUNLOCK(object); return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK); } #include "opt_ddb.h" #ifdef DDB #include /* DDB command to show buffer data */ DB_SHOW_COMMAND(buffer, db_show_buffer) { /* get args */ struct buf *bp = (struct buf *)addr; #ifdef FULL_BUF_TRACKING uint32_t i, j; #endif if (!have_addr) { db_printf("usage: show buffer \n"); return; } db_printf("buf at %p\n", bp); db_printf("b_flags = 0x%b, b_xflags=0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags, PRINT_BUF_XFLAGS); db_printf("b_vflags=0x%b b_ioflags0x%b\n", (u_int)bp->b_vflags, PRINT_BUF_VFLAGS, (u_int)bp->b_ioflags, PRINT_BIO_FLAGS); db_printf( "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, " "b_vp = %p, b_dep = %p\n", bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first); db_printf("b_kvabase = %p, b_kvasize = %d\n", bp->b_kvabase, bp->b_kvasize); if (bp->b_npages) { int i; db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); for (i = 0; i < bp->b_npages; i++) { vm_page_t m; m = bp->b_pages[i]; if (m != NULL) db_printf("(%p, 0x%lx, 0x%lx)", m->object, (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); else db_printf("( ??? )"); if ((i + 1) < bp->b_npages) db_printf(","); } db_printf("\n"); } BUF_LOCKPRINTINFO(bp); #if defined(FULL_BUF_TRACKING) db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt); i = bp->b_io_tcnt % BUF_TRACKING_SIZE; for (j = 1; j <= BUF_TRACKING_SIZE; j++) { if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL) continue; db_printf(" %2u: %s\n", j, bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]); } #elif defined(BUF_TRACKING) db_printf("b_io_tracking: %s\n", bp->b_io_tracking); #endif db_printf(" "); } DB_SHOW_COMMAND(bufqueues, bufqueues) { struct bufdomain *bd; struct buf *bp; long total; int i, j, cnt; db_printf("bqempty: %d\n", bqempty.bq_len); for (i = 0; i < buf_domains; i++) { bd = &bdomain[i]; db_printf("Buf domain %d\n", i); db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers); db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers); db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers); db_printf("\n"); db_printf("\tbufspace\t%ld\n", bd->bd_bufspace); db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace); db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace); db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace); db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh); db_printf("\n"); db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers); db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers); db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers); db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh); db_printf("\n"); total = 0; TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist) total += bp->b_bufsize; db_printf("\tcleanq count\t%d (%ld)\n", bd->bd_cleanq->bq_len, total); total = 0; TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist) total += bp->b_bufsize; db_printf("\tdirtyq count\t%d (%ld)\n", bd->bd_dirtyq.bq_len, total); db_printf("\twakeup\t\t%d\n", bd->bd_wanted); db_printf("\tlim\t\t%d\n", bd->bd_lim); db_printf("\tCPU "); for (j = 0; j <= mp_maxid; j++) db_printf("%d, ", bd->bd_subq[j].bq_len); db_printf("\n"); cnt = 0; total = 0; for (j = 0; j < nbuf; j++) if (buf[j].b_domain == i && BUF_ISLOCKED(&buf[j])) { cnt++; total += buf[j].b_bufsize; } db_printf("\tLocked buffers: %d space %ld\n", cnt, total); cnt = 0; total = 0; for (j = 0; j < nbuf; j++) if (buf[j].b_domain == i) { cnt++; total += buf[j].b_bufsize; } db_printf("\tTotal buffers: %d space %ld\n", cnt, total); } } DB_SHOW_COMMAND(lockedbufs, lockedbufs) { struct buf *bp; int i; for (i = 0; i < nbuf; i++) { bp = &buf[i]; if (BUF_ISLOCKED(bp)) { db_show_buffer((uintptr_t)bp, 1, 0, NULL); db_printf("\n"); if (db_pager_quit) break; } } } DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) { struct vnode *vp; struct buf *bp; if (!have_addr) { db_printf("usage: show vnodebufs \n"); return; } vp = (struct vnode *)addr; db_printf("Clean buffers:\n"); TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { db_show_buffer((uintptr_t)bp, 1, 0, NULL); db_printf("\n"); } db_printf("Dirty buffers:\n"); TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { db_show_buffer((uintptr_t)bp, 1, 0, NULL); db_printf("\n"); } } DB_COMMAND(countfreebufs, db_coundfreebufs) { struct buf *bp; int i, used = 0, nfree = 0; if (have_addr) { db_printf("usage: countfreebufs\n"); return; } for (i = 0; i < nbuf; i++) { bp = &buf[i]; if (bp->b_qindex == QUEUE_EMPTY) nfree++; else used++; } db_printf("Counted %d free, %d used (%d tot)\n", nfree, used, nfree + used); db_printf("numfreebuffers is %d\n", numfreebuffers); } #endif /* DDB */ Index: head/sys/kern/vfs_cluster.c =================================================================== --- head/sys/kern/vfs_cluster.c (revision 353534) +++ head/sys/kern/vfs_cluster.c (revision 353535) @@ -1,1097 +1,1097 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1993 * The Regents of the University of California. All rights reserved. * Modifications/enhancements: * Copyright (c) 1995 John S. Dyson. All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)vfs_cluster.c 8.7 (Berkeley) 2/13/94 */ #include __FBSDID("$FreeBSD$"); #include "opt_debug_cluster.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #if defined(CLUSTERDEBUG) static int rcluster= 0; SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0, "Debug VFS clustering code"); #endif static MALLOC_DEFINE(M_SEGMENT, "cl_savebuf", "cluster_save buffer"); static uma_zone_t cluster_pbuf_zone; static void cluster_init(void *); static struct cluster_save *cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int gbflags); static struct buf *cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn, daddr_t blkno, long size, int run, int gbflags, struct buf *fbp); static void cluster_callback(struct buf *); static int write_behind = 1; SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0, "Cluster write-behind; 0: disable, 1: enable, 2: backed off"); static int read_max = 64; SYSCTL_INT(_vfs, OID_AUTO, read_max, CTLFLAG_RW, &read_max, 0, "Cluster read-ahead max block count"); static int read_min = 1; SYSCTL_INT(_vfs, OID_AUTO, read_min, CTLFLAG_RW, &read_min, 0, "Cluster read min block count"); SYSINIT(cluster, SI_SUB_CPU, SI_ORDER_ANY, cluster_init, NULL); static void cluster_init(void *dummy) { cluster_pbuf_zone = pbuf_zsecond_create("clpbuf", nswbuf / 2); } /* * Read data to a buf, including read-ahead if we find this to be beneficial. * cluster_read replaces bread. */ int cluster_read(struct vnode *vp, u_quad_t filesize, daddr_t lblkno, long size, struct ucred *cred, long totread, int seqcount, int gbflags, struct buf **bpp) { struct buf *bp, *rbp, *reqbp; struct bufobj *bo; struct thread *td; daddr_t blkno, origblkno; int maxra, racluster; int error, ncontig; int i; error = 0; td = curthread; bo = &vp->v_bufobj; if (!unmapped_buf_allowed) gbflags &= ~GB_UNMAPPED; /* * Try to limit the amount of read-ahead by a few * ad-hoc parameters. This needs work!!! */ racluster = vp->v_mount->mnt_iosize_max / size; maxra = seqcount; maxra = min(read_max, maxra); maxra = min(nbuf/8, maxra); if (((u_quad_t)(lblkno + maxra + 1) * size) > filesize) maxra = (filesize / size) - lblkno; /* * get the requested block */ error = getblkx(vp, lblkno, size, 0, 0, gbflags, &bp); if (error != 0) { *bpp = NULL; return (error); } gbflags &= ~GB_NOSPARSE; origblkno = lblkno; *bpp = reqbp = bp; /* * if it is in the cache, then check to see if the reads have been * sequential. If they have, then try some read-ahead, otherwise * back-off on prospective read-aheads. */ if (bp->b_flags & B_CACHE) { if (!seqcount) { return 0; } else if ((bp->b_flags & B_RAM) == 0) { return 0; } else { bp->b_flags &= ~B_RAM; BO_RLOCK(bo); for (i = 1; i < maxra; i++) { /* * Stop if the buffer does not exist or it * is invalid (about to go away?) */ rbp = gbincore(&vp->v_bufobj, lblkno+i); if (rbp == NULL || (rbp->b_flags & B_INVAL)) break; /* * Set another read-ahead mark so we know * to check again. (If we can lock the * buffer without waiting) */ if ((((i % racluster) == (racluster - 1)) || (i == (maxra - 1))) && (0 == BUF_LOCK(rbp, LK_EXCLUSIVE | LK_NOWAIT, NULL))) { rbp->b_flags |= B_RAM; BUF_UNLOCK(rbp); } } BO_RUNLOCK(bo); if (i >= maxra) { return 0; } lblkno += i; } reqbp = bp = NULL; /* * If it isn't in the cache, then get a chunk from * disk if sequential, otherwise just get the block. */ } else { off_t firstread = bp->b_offset; int nblks; long minread; KASSERT(bp->b_offset != NOOFFSET, ("cluster_read: no buffer offset")); ncontig = 0; /* * Adjust totread if needed */ minread = read_min * size; if (minread > totread) totread = minread; /* * Compute the total number of blocks that we should read * synchronously. */ if (firstread + totread > filesize) totread = filesize - firstread; nblks = howmany(totread, size); if (nblks > racluster) nblks = racluster; /* * Now compute the number of contiguous blocks. */ if (nblks > 1) { error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL); /* * If this failed to map just do the original block. */ if (error || blkno == -1) ncontig = 0; } /* * If we have contiguous data available do a cluster * otherwise just read the requested block. */ if (ncontig) { /* Account for our first block. */ ncontig = min(ncontig + 1, nblks); if (ncontig < nblks) nblks = ncontig; bp = cluster_rbuild(vp, filesize, lblkno, blkno, size, nblks, gbflags, bp); lblkno += (bp->b_bufsize / size); } else { bp->b_flags |= B_RAM; bp->b_iocmd = BIO_READ; lblkno += 1; } } /* * handle the synchronous read so that it is available ASAP. */ if (bp) { if ((bp->b_flags & B_CLUSTER) == 0) { vfs_busy_pages(bp, 0); } bp->b_flags &= ~B_INVAL; bp->b_ioflags &= ~BIO_ERROR; if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL) BUF_KERNPROC(bp); bp->b_iooffset = dbtob(bp->b_blkno); bstrategy(bp); #ifdef RACCT if (racct_enable) { PROC_LOCK(td->td_proc); racct_add_buf(td->td_proc, bp, 0); PROC_UNLOCK(td->td_proc); } #endif /* RACCT */ td->td_ru.ru_inblock++; } /* * If we have been doing sequential I/O, then do some read-ahead. */ while (lblkno < (origblkno + maxra)) { error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL); if (error) break; if (blkno == -1) break; /* * We could throttle ncontig here by maxra but we might as * well read the data if it is contiguous. We're throttled * by racluster anyway. */ if (ncontig) { ncontig = min(ncontig + 1, racluster); rbp = cluster_rbuild(vp, filesize, lblkno, blkno, size, ncontig, gbflags, NULL); lblkno += (rbp->b_bufsize / size); if (rbp->b_flags & B_DELWRI) { bqrelse(rbp); continue; } } else { rbp = getblk(vp, lblkno, size, 0, 0, gbflags); lblkno += 1; if (rbp->b_flags & B_DELWRI) { bqrelse(rbp); continue; } rbp->b_flags |= B_ASYNC | B_RAM; rbp->b_iocmd = BIO_READ; rbp->b_blkno = blkno; } if (rbp->b_flags & B_CACHE) { rbp->b_flags &= ~B_ASYNC; bqrelse(rbp); continue; } if ((rbp->b_flags & B_CLUSTER) == 0) { vfs_busy_pages(rbp, 0); } rbp->b_flags &= ~B_INVAL; rbp->b_ioflags &= ~BIO_ERROR; if ((rbp->b_flags & B_ASYNC) || rbp->b_iodone != NULL) BUF_KERNPROC(rbp); rbp->b_iooffset = dbtob(rbp->b_blkno); bstrategy(rbp); #ifdef RACCT if (racct_enable) { PROC_LOCK(td->td_proc); racct_add_buf(td->td_proc, rbp, 0); PROC_UNLOCK(td->td_proc); } #endif /* RACCT */ td->td_ru.ru_inblock++; } if (reqbp) { /* * Like bread, always brelse() the buffer when * returning an error. */ error = bufwait(reqbp); if (error != 0) { brelse(reqbp); *bpp = NULL; } } return (error); } /* * If blocks are contiguous on disk, use this to provide clustered * read ahead. We will read as many blocks as possible sequentially * and then parcel them up into logical blocks in the buffer hash table. */ static struct buf * cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn, daddr_t blkno, long size, int run, int gbflags, struct buf *fbp) { struct buf *bp, *tbp; daddr_t bn; off_t off; long tinc, tsize; int i, inc, j, k, toff; KASSERT(size == vp->v_mount->mnt_stat.f_iosize, ("cluster_rbuild: size %ld != f_iosize %jd\n", size, (intmax_t)vp->v_mount->mnt_stat.f_iosize)); /* * avoid a division */ while ((u_quad_t) size * (lbn + run) > filesize) { --run; } if (fbp) { tbp = fbp; tbp->b_iocmd = BIO_READ; } else { tbp = getblk(vp, lbn, size, 0, 0, gbflags); if (tbp->b_flags & B_CACHE) return tbp; tbp->b_flags |= B_ASYNC | B_RAM; tbp->b_iocmd = BIO_READ; } tbp->b_blkno = blkno; if( (tbp->b_flags & B_MALLOC) || ((tbp->b_flags & B_VMIO) == 0) || (run <= 1) ) return tbp; bp = uma_zalloc(cluster_pbuf_zone, M_NOWAIT); if (bp == NULL) return tbp; /* * We are synthesizing a buffer out of vm_page_t's, but * if the block size is not page aligned then the starting * address may not be either. Inherit the b_data offset * from the original buffer. */ bp->b_flags = B_ASYNC | B_CLUSTER | B_VMIO; if ((gbflags & GB_UNMAPPED) != 0) { bp->b_data = unmapped_buf; } else { bp->b_data = (char *)((vm_offset_t)bp->b_data | ((vm_offset_t)tbp->b_data & PAGE_MASK)); } bp->b_iocmd = BIO_READ; bp->b_iodone = cluster_callback; bp->b_blkno = blkno; bp->b_lblkno = lbn; bp->b_offset = tbp->b_offset; KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset")); pbgetvp(vp, bp); TAILQ_INIT(&bp->b_cluster.cluster_head); bp->b_bcount = 0; bp->b_bufsize = 0; bp->b_npages = 0; inc = btodb(size); for (bn = blkno, i = 0; i < run; ++i, bn += inc) { if (i == 0) { VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object); - vfs_drain_busy_pages(tbp); vm_object_pip_add(tbp->b_bufobj->bo_object, tbp->b_npages); - for (k = 0; k < tbp->b_npages; k++) - vm_page_sbusy(tbp->b_pages[k]); + vfs_busy_pages_acquire(tbp); VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object); } else { if ((bp->b_npages * PAGE_SIZE) + round_page(size) > vp->v_mount->mnt_iosize_max) { break; } tbp = getblk(vp, lbn + i, size, 0, 0, GB_LOCK_NOWAIT | (gbflags & GB_UNMAPPED)); /* Don't wait around for locked bufs. */ if (tbp == NULL) break; /* * Stop scanning if the buffer is fully valid * (marked B_CACHE), or locked (may be doing a * background write), or if the buffer is not * VMIO backed. The clustering code can only deal * with VMIO-backed buffers. The bo lock is not * required for the BKGRDINPROG check since it * can not be set without the buf lock. */ if ((tbp->b_vflags & BV_BKGRDINPROG) || (tbp->b_flags & B_CACHE) || (tbp->b_flags & B_VMIO) == 0) { bqrelse(tbp); break; } /* * The buffer must be completely invalid in order to * take part in the cluster. If it is partially valid * then we stop. */ off = tbp->b_offset; tsize = size; VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object); for (j = 0; tsize > 0; j++) { toff = off & PAGE_MASK; tinc = tsize; if (toff + tinc > PAGE_SIZE) tinc = PAGE_SIZE - toff; VM_OBJECT_ASSERT_WLOCKED(tbp->b_pages[j]->object); if ((tbp->b_pages[j]->valid & vm_page_bits(toff, tinc)) != 0) break; - if (vm_page_xbusied(tbp->b_pages[j])) + if (vm_page_trysbusy(tbp->b_pages[j]) == 0) break; vm_object_pip_add(tbp->b_bufobj->bo_object, 1); - vm_page_sbusy(tbp->b_pages[j]); off += tinc; tsize -= tinc; } if (tsize > 0) { clean_sbusy: vm_object_pip_wakeupn(tbp->b_bufobj->bo_object, j); for (k = 0; k < j; k++) vm_page_sunbusy(tbp->b_pages[k]); VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object); bqrelse(tbp); break; } VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object); /* * Set a read-ahead mark as appropriate */ if ((fbp && (i == 1)) || (i == (run - 1))) tbp->b_flags |= B_RAM; /* * Set the buffer up for an async read (XXX should * we do this only if we do not wind up brelse()ing?). * Set the block number if it isn't set, otherwise * if it is make sure it matches the block number we * expect. */ tbp->b_flags |= B_ASYNC; tbp->b_iocmd = BIO_READ; if (tbp->b_blkno == tbp->b_lblkno) { tbp->b_blkno = bn; } else if (tbp->b_blkno != bn) { VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object); goto clean_sbusy; } } /* * XXX fbp from caller may not be B_ASYNC, but we are going * to biodone() it in cluster_callback() anyway */ BUF_KERNPROC(tbp); TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head, tbp, b_cluster.cluster_entry); VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object); for (j = 0; j < tbp->b_npages; j += 1) { vm_page_t m; m = tbp->b_pages[j]; if ((bp->b_npages == 0) || (bp->b_pages[bp->b_npages-1] != m)) { bp->b_pages[bp->b_npages] = m; bp->b_npages++; } if (m->valid == VM_PAGE_BITS_ALL) tbp->b_pages[j] = bogus_page; } VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object); /* * Don't inherit tbp->b_bufsize as it may be larger due to * a non-page-aligned size. Instead just aggregate using * 'size'. */ if (tbp->b_bcount != size) printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size); if (tbp->b_bufsize != size) printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size); bp->b_bcount += size; bp->b_bufsize += size; } /* * Fully valid pages in the cluster are already good and do not need * to be re-read from disk. Replace the page with bogus_page */ VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); for (j = 0; j < bp->b_npages; j++) { VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[j]->object); if (bp->b_pages[j]->valid == VM_PAGE_BITS_ALL) bp->b_pages[j] = bogus_page; } VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); if (bp->b_bufsize > bp->b_kvasize) panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n", bp->b_bufsize, bp->b_kvasize); if (buf_mapped(bp)) { pmap_qenter(trunc_page((vm_offset_t) bp->b_data), (vm_page_t *)bp->b_pages, bp->b_npages); } return (bp); } /* * Cleanup after a clustered read or write. * This is complicated by the fact that any of the buffers might have * extra memory (if there were no empty buffer headers at allocbuf time) * that we will need to shift around. */ static void cluster_callback(struct buf *bp) { struct buf *nbp, *tbp; int error = 0; /* * Must propagate errors to all the components. */ if (bp->b_ioflags & BIO_ERROR) error = bp->b_error; if (buf_mapped(bp)) { pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); } /* * Move memory from the large cluster buffer into the component * buffers and mark IO as done on these. */ for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head); tbp; tbp = nbp) { nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry); if (error) { tbp->b_ioflags |= BIO_ERROR; tbp->b_error = error; } else { tbp->b_dirtyoff = tbp->b_dirtyend = 0; tbp->b_flags &= ~B_INVAL; tbp->b_ioflags &= ~BIO_ERROR; /* * XXX the bdwrite()/bqrelse() issued during * cluster building clears B_RELBUF (see bqrelse() * comment). If direct I/O was specified, we have * to restore it here to allow the buffer and VM * to be freed. */ if (tbp->b_flags & B_DIRECT) tbp->b_flags |= B_RELBUF; } bufdone(tbp); } pbrelvp(bp); uma_zfree(cluster_pbuf_zone, bp); } /* * cluster_wbuild_wb: * * Implement modified write build for cluster. * * write_behind = 0 write behind disabled * write_behind = 1 write behind normal (default) * write_behind = 2 write behind backed-off */ static __inline int cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len, int gbflags) { int r = 0; switch (write_behind) { case 2: if (start_lbn < len) break; start_lbn -= len; /* FALLTHROUGH */ case 1: r = cluster_wbuild(vp, size, start_lbn, len, gbflags); /* FALLTHROUGH */ default: /* FALLTHROUGH */ break; } return(r); } /* * Do clustered write for FFS. * * Three cases: * 1. Write is not sequential (write asynchronously) * Write is sequential: * 2. beginning of cluster - begin cluster * 3. middle of a cluster - add to cluster * 4. end of a cluster - asynchronously write cluster */ void cluster_write(struct vnode *vp, struct buf *bp, u_quad_t filesize, int seqcount, int gbflags) { daddr_t lbn; int maxclen, cursize; int lblocksize; int async; if (!unmapped_buf_allowed) gbflags &= ~GB_UNMAPPED; if (vp->v_type == VREG) { async = DOINGASYNC(vp); lblocksize = vp->v_mount->mnt_stat.f_iosize; } else { async = 0; lblocksize = bp->b_bufsize; } lbn = bp->b_lblkno; KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset")); /* Initialize vnode to beginning of file. */ if (lbn == 0) vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0; if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 || (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) { maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1; if (vp->v_clen != 0) { /* * Next block is not sequential. * * If we are not writing at end of file, the process * seeked to another point in the file since its last * write, or we have reached our maximum cluster size, * then push the previous cluster. Otherwise try * reallocating to make it sequential. * * Change to algorithm: only push previous cluster if * it was sequential from the point of view of the * seqcount heuristic, otherwise leave the buffer * intact so we can potentially optimize the I/O * later on in the buf_daemon or update daemon * flush. */ cursize = vp->v_lastw - vp->v_cstart + 1; if (((u_quad_t) bp->b_offset + lblocksize) != filesize || lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) { if (!async && seqcount > 0) { cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, cursize, gbflags); } } else { struct buf **bpp, **endbp; struct cluster_save *buflist; buflist = cluster_collectbufs(vp, bp, gbflags); if (buflist == NULL) { /* * Cluster build failed so just write * it now. */ bawrite(bp); return; } endbp = &buflist->bs_children [buflist->bs_nchildren - 1]; if (VOP_REALLOCBLKS(vp, buflist)) { /* * Failed, push the previous cluster * if *really* writing sequentially * in the logical file (seqcount > 1), * otherwise delay it in the hopes that * the low level disk driver can * optimize the write ordering. */ for (bpp = buflist->bs_children; bpp < endbp; bpp++) brelse(*bpp); free(buflist, M_SEGMENT); if (seqcount > 1) { cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, cursize, gbflags); } } else { /* * Succeeded, keep building cluster. */ for (bpp = buflist->bs_children; bpp <= endbp; bpp++) bdwrite(*bpp); free(buflist, M_SEGMENT); vp->v_lastw = lbn; vp->v_lasta = bp->b_blkno; return; } } } /* * Consider beginning a cluster. If at end of file, make * cluster as large as possible, otherwise find size of * existing cluster. */ if ((vp->v_type == VREG) && ((u_quad_t) bp->b_offset + lblocksize) != filesize && (bp->b_blkno == bp->b_lblkno) && (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) || bp->b_blkno == -1)) { bawrite(bp); vp->v_clen = 0; vp->v_lasta = bp->b_blkno; vp->v_cstart = lbn + 1; vp->v_lastw = lbn; return; } vp->v_clen = maxclen; if (!async && maxclen == 0) { /* I/O not contiguous */ vp->v_cstart = lbn + 1; bawrite(bp); } else { /* Wait for rest of cluster */ vp->v_cstart = lbn; bdwrite(bp); } } else if (lbn == vp->v_cstart + vp->v_clen) { /* * At end of cluster, write it out if seqcount tells us we * are operating sequentially, otherwise let the buf or * update daemon handle it. */ bdwrite(bp); if (seqcount > 1) { cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, vp->v_clen + 1, gbflags); } vp->v_clen = 0; vp->v_cstart = lbn + 1; } else if (vm_page_count_severe()) { /* * We are low on memory, get it going NOW */ bawrite(bp); } else { /* * In the middle of a cluster, so just delay the I/O for now. */ bdwrite(bp); } vp->v_lastw = lbn; vp->v_lasta = bp->b_blkno; } /* * This is an awful lot like cluster_rbuild...wish they could be combined. * The last lbn argument is the current block on which I/O is being * performed. Check to see that it doesn't fall in the middle of * the current block (if last_bp == NULL). */ int cluster_wbuild(struct vnode *vp, long size, daddr_t start_lbn, int len, int gbflags) { struct buf *bp, *tbp; struct bufobj *bo; int i, j; int totalwritten = 0; int dbsize = btodb(size); if (!unmapped_buf_allowed) gbflags &= ~GB_UNMAPPED; bo = &vp->v_bufobj; while (len > 0) { /* * If the buffer is not delayed-write (i.e. dirty), or it * is delayed-write but either locked or inval, it cannot * partake in the clustered write. */ BO_LOCK(bo); if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL || (tbp->b_vflags & BV_BKGRDINPROG)) { BO_UNLOCK(bo); ++start_lbn; --len; continue; } if (BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, BO_LOCKPTR(bo))) { ++start_lbn; --len; continue; } if ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) { BUF_UNLOCK(tbp); ++start_lbn; --len; continue; } bremfree(tbp); tbp->b_flags &= ~B_DONE; /* * Extra memory in the buffer, punt on this buffer. * XXX we could handle this in most cases, but we would * have to push the extra memory down to after our max * possible cluster size and then potentially pull it back * up if the cluster was terminated prematurely--too much * hassle. */ if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) != (B_CLUSTEROK | B_VMIO)) || (tbp->b_bcount != tbp->b_bufsize) || (tbp->b_bcount != size) || (len == 1) || ((bp = uma_zalloc(cluster_pbuf_zone, (vp->v_vflag & VV_MD) != 0 ? M_NOWAIT : M_WAITOK)) == NULL)) { totalwritten += tbp->b_bufsize; bawrite(tbp); ++start_lbn; --len; continue; } /* * We got a pbuf to make the cluster in. * so initialise it. */ TAILQ_INIT(&bp->b_cluster.cluster_head); bp->b_bcount = 0; bp->b_bufsize = 0; bp->b_npages = 0; if (tbp->b_wcred != NOCRED) bp->b_wcred = crhold(tbp->b_wcred); bp->b_blkno = tbp->b_blkno; bp->b_lblkno = tbp->b_lblkno; bp->b_offset = tbp->b_offset; /* * We are synthesizing a buffer out of vm_page_t's, but * if the block size is not page aligned then the starting * address may not be either. Inherit the b_data offset * from the original buffer. */ if ((gbflags & GB_UNMAPPED) == 0 || (tbp->b_flags & B_VMIO) == 0) { bp->b_data = (char *)((vm_offset_t)bp->b_data | ((vm_offset_t)tbp->b_data & PAGE_MASK)); } else { bp->b_data = unmapped_buf; } bp->b_flags |= B_CLUSTER | (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT)); bp->b_iodone = cluster_callback; pbgetvp(vp, bp); /* * From this location in the file, scan forward to see * if there are buffers with adjacent data that need to * be written as well. */ for (i = 0; i < len; ++i, ++start_lbn) { if (i != 0) { /* If not the first buffer */ /* * If the adjacent data is not even in core it * can't need to be written. */ BO_LOCK(bo); if ((tbp = gbincore(bo, start_lbn)) == NULL || (tbp->b_vflags & BV_BKGRDINPROG)) { BO_UNLOCK(bo); break; } /* * If it IS in core, but has different * characteristics, or is locked (which * means it could be undergoing a background * I/O or be in a weird state), then don't * cluster with it. */ if (BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, BO_LOCKPTR(bo))) break; if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK | B_INVAL | B_DELWRI | B_NEEDCOMMIT)) != (B_DELWRI | B_CLUSTEROK | (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) || tbp->b_wcred != bp->b_wcred) { BUF_UNLOCK(tbp); break; } /* * Check that the combined cluster * would make sense with regard to pages * and would not be too large */ if ((tbp->b_bcount != size) || ((bp->b_blkno + (dbsize * i)) != tbp->b_blkno) || ((tbp->b_npages + bp->b_npages) > (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) { BUF_UNLOCK(tbp); break; } /* * Ok, it's passed all the tests, * so remove it from the free list * and mark it busy. We will use it. */ bremfree(tbp); tbp->b_flags &= ~B_DONE; } /* end of code for non-first buffers only */ /* * If the IO is via the VM then we do some * special VM hackery (yuck). Since the buffer's * block size may not be page-aligned it is possible * for a page to be shared between two buffers. We * have to get rid of the duplication when building * the cluster. */ if (tbp->b_flags & B_VMIO) { vm_page_t m; VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object); if (i == 0) { - vfs_drain_busy_pages(tbp); + vfs_busy_pages_acquire(tbp); } else { /* if not first buffer */ for (j = 0; j < tbp->b_npages; j += 1) { m = tbp->b_pages[j]; - if (vm_page_xbusied(m)) { + if (vm_page_trysbusy(m) == 0) { + for (j--; j >= 0; j--) + vm_page_sunbusy( + tbp->b_pages[j]); VM_OBJECT_WUNLOCK( tbp->b_object); bqrelse(tbp); goto finishcluster; } } } + vm_object_pip_add(tbp->b_bufobj->bo_object, + tbp->b_npages); for (j = 0; j < tbp->b_npages; j += 1) { m = tbp->b_pages[j]; - vm_page_sbusy(m); - vm_object_pip_add(m->object, 1); if ((bp->b_npages == 0) || (bp->b_pages[bp->b_npages - 1] != m)) { bp->b_pages[bp->b_npages] = m; bp->b_npages++; } } VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object); } bp->b_bcount += size; bp->b_bufsize += size; /* * If any of the clustered buffers have their * B_BARRIER flag set, transfer that request to * the cluster. */ bp->b_flags |= (tbp->b_flags & B_BARRIER); tbp->b_flags &= ~(B_DONE | B_BARRIER); tbp->b_flags |= B_ASYNC; tbp->b_ioflags &= ~BIO_ERROR; tbp->b_iocmd = BIO_WRITE; bundirty(tbp); reassignbuf(tbp); /* put on clean list */ bufobj_wref(tbp->b_bufobj); BUF_KERNPROC(tbp); buf_track(tbp, __func__); TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head, tbp, b_cluster.cluster_entry); } finishcluster: if (buf_mapped(bp)) { pmap_qenter(trunc_page((vm_offset_t) bp->b_data), (vm_page_t *)bp->b_pages, bp->b_npages); } if (bp->b_bufsize > bp->b_kvasize) panic( "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n", bp->b_bufsize, bp->b_kvasize); totalwritten += bp->b_bufsize; bp->b_dirtyoff = 0; bp->b_dirtyend = bp->b_bufsize; bawrite(bp); len -= i; } return totalwritten; } /* * Collect together all the buffers in a cluster. * Plus add one additional buffer. */ static struct cluster_save * cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int gbflags) { struct cluster_save *buflist; struct buf *bp; daddr_t lbn; int i, j, len, error; len = vp->v_lastw - vp->v_cstart + 1; buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist), M_SEGMENT, M_WAITOK); buflist->bs_nchildren = 0; buflist->bs_children = (struct buf **) (buflist + 1); for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) { error = bread_gb(vp, lbn, last_bp->b_bcount, NOCRED, gbflags, &bp); if (error != 0) { /* * If read fails, release collected buffers * and return failure. */ for (j = 0; j < i; j++) brelse(buflist->bs_children[j]); free(buflist, M_SEGMENT); return (NULL); } buflist->bs_children[i] = bp; if (bp->b_blkno == bp->b_lblkno) VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); } buflist->bs_children[i] = bp = last_bp; if (bp->b_blkno == bp->b_lblkno) VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); buflist->bs_nchildren = i + 1; return (buflist); } Index: head/sys/sys/buf.h =================================================================== --- head/sys/sys/buf.h (revision 353534) +++ head/sys/sys/buf.h (revision 353535) @@ -1,589 +1,590 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1989, 1993 * The Regents of the University of California. All rights reserved. * (c) UNIX System Laboratories, Inc. * All or some portions of this file are derived from material licensed * to the University of California by American Telephone and Telegraph * Co. or Unix System Laboratories, Inc. and are reproduced herein with * the permission of UNIX System Laboratories, Inc. * * 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. * * @(#)buf.h 8.9 (Berkeley) 3/30/95 * $FreeBSD$ */ #ifndef _SYS_BUF_H_ #define _SYS_BUF_H_ #include #include #include #include #include struct bio; struct buf; struct bufobj; struct mount; struct vnode; struct uio; /* * To avoid including */ LIST_HEAD(workhead, worklist); /* * These are currently used only by the soft dependency code, hence * are stored once in a global variable. If other subsystems wanted * to use these hooks, a pointer to a set of bio_ops could be added * to each buffer. */ extern struct bio_ops { void (*io_start)(struct buf *); void (*io_complete)(struct buf *); void (*io_deallocate)(struct buf *); int (*io_countdeps)(struct buf *, int); } bioops; struct vm_object; struct vm_page; typedef uint32_t b_xflags_t; /* * The buffer header describes an I/O operation in the kernel. * * NOTES: * b_bufsize, b_bcount. b_bufsize is the allocation size of the * buffer, either DEV_BSIZE or PAGE_SIZE aligned. b_bcount is the * originally requested buffer size and can serve as a bounds check * against EOF. For most, but not all uses, b_bcount == b_bufsize. * * b_dirtyoff, b_dirtyend. Buffers support piecemeal, unaligned * ranges of dirty data that need to be written to backing store. * The range is typically clipped at b_bcount ( not b_bufsize ). * * b_resid. Number of bytes remaining in I/O. After an I/O operation * completes, b_resid is usually 0 indicating 100% success. * * All fields are protected by the buffer lock except those marked: * V - Protected by owning bufobj lock * Q - Protected by the buf queue lock * D - Protected by an dependency implementation specific lock */ struct buf { struct bufobj *b_bufobj; long b_bcount; void *b_caller1; caddr_t b_data; int b_error; uint16_t b_iocmd; /* BIO_* bio_cmd from bio.h */ uint16_t b_ioflags; /* BIO_* bio_flags from bio.h */ off_t b_iooffset; long b_resid; void (*b_iodone)(struct buf *); void (*b_ckhashcalc)(struct buf *); uint64_t b_ckhash; /* B_CKHASH requested check-hash */ daddr_t b_blkno; /* Underlying physical block number. */ off_t b_offset; /* Offset into file. */ TAILQ_ENTRY(buf) b_bobufs; /* (V) Buffer's associated vnode. */ uint32_t b_vflags; /* (V) BV_* flags */ uint8_t b_qindex; /* (Q) buffer queue index */ uint8_t b_domain; /* (Q) buf domain this resides in */ uint16_t b_subqueue; /* (Q) per-cpu q if any */ uint32_t b_flags; /* B_* flags. */ b_xflags_t b_xflags; /* extra flags */ struct lock b_lock; /* Buffer lock */ long b_bufsize; /* Allocated buffer size. */ int b_runningbufspace; /* when I/O is running, pipelining */ int b_kvasize; /* size of kva for buffer */ int b_dirtyoff; /* Offset in buffer of dirty region. */ int b_dirtyend; /* Offset of end of dirty region. */ caddr_t b_kvabase; /* base kva for buffer */ daddr_t b_lblkno; /* Logical block number. */ struct vnode *b_vp; /* Device vnode. */ struct ucred *b_rcred; /* Read credentials reference. */ struct ucred *b_wcred; /* Write credentials reference. */ union { TAILQ_ENTRY(buf) b_freelist; /* (Q) */ struct { void (*b_pgiodone)(void *, vm_page_t *, int, int); int b_pgbefore; int b_pgafter; }; }; union cluster_info { TAILQ_HEAD(cluster_list_head, buf) cluster_head; TAILQ_ENTRY(buf) cluster_entry; } b_cluster; struct vm_page *b_pages[btoc(MAXPHYS)]; int b_npages; struct workhead b_dep; /* (D) List of filesystem dependencies. */ void *b_fsprivate1; void *b_fsprivate2; void *b_fsprivate3; #if defined(FULL_BUF_TRACKING) #define BUF_TRACKING_SIZE 32 #define BUF_TRACKING_ENTRY(x) ((x) & (BUF_TRACKING_SIZE - 1)) const char *b_io_tracking[BUF_TRACKING_SIZE]; uint32_t b_io_tcnt; #elif defined(BUF_TRACKING) const char *b_io_tracking; #endif }; #define b_object b_bufobj->bo_object /* * These flags are kept in b_flags. * * Notes: * * B_ASYNC VOP calls on bp's are usually async whether or not * B_ASYNC is set, but some subsystems, such as NFS, like * to know what is best for the caller so they can * optimize the I/O. * * B_PAGING Indicates that bp is being used by the paging system or * some paging system and that the bp is not linked into * the b_vp's clean/dirty linked lists or ref counts. * Buffer vp reassignments are illegal in this case. * * B_CACHE This may only be set if the buffer is entirely valid. * The situation where B_DELWRI is set and B_CACHE is * clear MUST be committed to disk by getblk() so * B_DELWRI can also be cleared. See the comments for * getblk() in kern/vfs_bio.c. If B_CACHE is clear, * the caller is expected to clear BIO_ERROR and B_INVAL, * set BIO_READ, and initiate an I/O. * * The 'entire buffer' is defined to be the range from * 0 through b_bcount. * * B_MALLOC Request that the buffer be allocated from the malloc * pool, DEV_BSIZE aligned instead of PAGE_SIZE aligned. * * B_CLUSTEROK This flag is typically set for B_DELWRI buffers * by filesystems that allow clustering when the buffer * is fully dirty and indicates that it may be clustered * with other adjacent dirty buffers. Note the clustering * may not be used with the stage 1 data write under NFS * but may be used for the commit rpc portion. * * B_INVALONERR This flag is set on dirty buffers. It specifies that a * write error should forcibly invalidate the buffer * contents. This flag should be used with caution, as it * discards data. It is incompatible with B_ASYNC. * * B_VMIO Indicates that the buffer is tied into an VM object. * The buffer's data is always PAGE_SIZE aligned even * if b_bufsize and b_bcount are not. ( b_bufsize is * always at least DEV_BSIZE aligned, though ). * * B_DIRECT Hint that we should attempt to completely free * the pages underlying the buffer. B_DIRECT is * sticky until the buffer is released and typically * only has an effect when B_RELBUF is also set. * */ #define B_AGE 0x00000001 /* Move to age queue when I/O done. */ #define B_NEEDCOMMIT 0x00000002 /* Append-write in progress. */ #define B_ASYNC 0x00000004 /* Start I/O, do not wait. */ #define B_DIRECT 0x00000008 /* direct I/O flag (pls free vmio) */ #define B_DEFERRED 0x00000010 /* Skipped over for cleaning */ #define B_CACHE 0x00000020 /* Bread found us in the cache. */ #define B_VALIDSUSPWRT 0x00000040 /* Valid write during suspension. */ #define B_DELWRI 0x00000080 /* Delay I/O until buffer reused. */ #define B_CKHASH 0x00000100 /* checksum hash calculated on read */ #define B_DONE 0x00000200 /* I/O completed. */ #define B_EINTR 0x00000400 /* I/O was interrupted */ #define B_NOREUSE 0x00000800 /* Contents not reused once released. */ #define B_REUSE 0x00001000 /* Contents reused, second chance. */ #define B_INVAL 0x00002000 /* Does not contain valid info. */ #define B_BARRIER 0x00004000 /* Write this and all preceding first. */ #define B_NOCACHE 0x00008000 /* Do not cache block after use. */ #define B_MALLOC 0x00010000 /* malloced b_data */ #define B_CLUSTEROK 0x00020000 /* Pagein op, so swap() can count it. */ #define B_INVALONERR 0x00040000 /* Invalidate on write error. */ #define B_00080000 0x00080000 /* Available flag. */ #define B_00100000 0x00100000 /* Available flag. */ #define B_00200000 0x00200000 /* Available flag. */ #define B_RELBUF 0x00400000 /* Release VMIO buffer. */ #define B_FS_FLAG1 0x00800000 /* Available flag for FS use. */ #define B_NOCOPY 0x01000000 /* Don't copy-on-write this buf. */ #define B_INFREECNT 0x02000000 /* buf is counted in numfreebufs */ #define B_PAGING 0x04000000 /* volatile paging I/O -- bypass VMIO */ #define B_MANAGED 0x08000000 /* Managed by FS. */ #define B_RAM 0x10000000 /* Read ahead mark (flag) */ #define B_VMIO 0x20000000 /* VMIO flag */ #define B_CLUSTER 0x40000000 /* pagein op, so swap() can count it */ #define B_REMFREE 0x80000000 /* Delayed bremfree */ #define PRINT_BUF_FLAGS "\20\40remfree\37cluster\36vmio\35ram\34managed" \ "\33paging\32infreecnt\31nocopy\30b23\27relbuf\26b21\25b20" \ "\24b19\23invalonerr\22clusterok\21malloc\20nocache\17b14\16inval" \ "\15reuse\14noreuse\13eintr\12done\11b8\10delwri" \ "\7validsuspwrt\6cache\5deferred\4direct\3async\2needcommit\1age" /* * These flags are kept in b_xflags. * * BX_FSPRIV reserves a set of eight flags that may be used by individual * filesystems for their own purpose. Their specific definitions are * found in the header files for each filesystem that uses them. */ #define BX_VNDIRTY 0x00000001 /* On vnode dirty list */ #define BX_VNCLEAN 0x00000002 /* On vnode clean list */ #define BX_BKGRDWRITE 0x00000010 /* Do writes in background */ #define BX_BKGRDMARKER 0x00000020 /* Mark buffer for splay tree */ #define BX_ALTDATA 0x00000040 /* Holds extended data */ #define BX_FSPRIV 0x00FF0000 /* filesystem-specific flags mask */ #define PRINT_BUF_XFLAGS "\20\7altdata\6bkgrdmarker\5bkgrdwrite\2clean\1dirty" #define NOOFFSET (-1LL) /* No buffer offset calculated yet */ /* * These flags are kept in b_vflags. */ #define BV_SCANNED 0x00000001 /* VOP_FSYNC funcs mark written bufs */ #define BV_BKGRDINPROG 0x00000002 /* Background write in progress */ #define BV_BKGRDWAIT 0x00000004 /* Background write waiting */ #define BV_BKGRDERR 0x00000008 /* Error from background write */ #define PRINT_BUF_VFLAGS "\20\4bkgrderr\3bkgrdwait\2bkgrdinprog\1scanned" #ifdef _KERNEL #ifndef NSWBUF_MIN #define NSWBUF_MIN 16 #endif /* * Buffer locking */ extern const char *buf_wmesg; /* Default buffer lock message */ #define BUF_WMESG "bufwait" #include /* XXX for curthread */ #include /* * Initialize a lock. */ #define BUF_LOCKINIT(bp) \ lockinit(&(bp)->b_lock, PRIBIO + 4, buf_wmesg, 0, LK_NEW) /* * * Get a lock sleeping non-interruptably until it becomes available. */ #define BUF_LOCK(bp, locktype, interlock) \ _lockmgr_args_rw(&(bp)->b_lock, (locktype), (interlock), \ LK_WMESG_DEFAULT, LK_PRIO_DEFAULT, LK_TIMO_DEFAULT, \ LOCK_FILE, LOCK_LINE) /* * Get a lock sleeping with specified interruptably and timeout. */ #define BUF_TIMELOCK(bp, locktype, interlock, wmesg, catch, timo) \ _lockmgr_args_rw(&(bp)->b_lock, (locktype) | LK_TIMELOCK, \ (interlock), (wmesg), (PRIBIO + 4) | (catch), (timo), \ LOCK_FILE, LOCK_LINE) /* * Release a lock. Only the acquiring process may free the lock unless * it has been handed off to biodone. */ #define BUF_UNLOCK(bp) do { \ KASSERT(((bp)->b_flags & B_REMFREE) == 0, \ ("BUF_UNLOCK %p while B_REMFREE is still set.", (bp))); \ \ (void)_lockmgr_args(&(bp)->b_lock, LK_RELEASE, NULL, \ LK_WMESG_DEFAULT, LK_PRIO_DEFAULT, LK_TIMO_DEFAULT, \ LOCK_FILE, LOCK_LINE); \ } while (0) /* * Check if a buffer lock is recursed. */ #define BUF_LOCKRECURSED(bp) \ lockmgr_recursed(&(bp)->b_lock) /* * Check if a buffer lock is currently held. */ #define BUF_ISLOCKED(bp) \ lockstatus(&(bp)->b_lock) /* * Free a buffer lock. */ #define BUF_LOCKFREE(bp) \ lockdestroy(&(bp)->b_lock) /* * Print informations on a buffer lock. */ #define BUF_LOCKPRINTINFO(bp) \ lockmgr_printinfo(&(bp)->b_lock) /* * Buffer lock assertions. */ #if defined(INVARIANTS) && defined(INVARIANT_SUPPORT) #define BUF_ASSERT_LOCKED(bp) \ _lockmgr_assert(&(bp)->b_lock, KA_LOCKED, LOCK_FILE, LOCK_LINE) #define BUF_ASSERT_SLOCKED(bp) \ _lockmgr_assert(&(bp)->b_lock, KA_SLOCKED, LOCK_FILE, LOCK_LINE) #define BUF_ASSERT_XLOCKED(bp) \ _lockmgr_assert(&(bp)->b_lock, KA_XLOCKED, LOCK_FILE, LOCK_LINE) #define BUF_ASSERT_UNLOCKED(bp) \ _lockmgr_assert(&(bp)->b_lock, KA_UNLOCKED, LOCK_FILE, LOCK_LINE) #else #define BUF_ASSERT_LOCKED(bp) #define BUF_ASSERT_SLOCKED(bp) #define BUF_ASSERT_XLOCKED(bp) #define BUF_ASSERT_UNLOCKED(bp) #endif #ifdef _SYS_PROC_H_ /* Avoid #include pollution */ /* * When initiating asynchronous I/O, change ownership of the lock to the * kernel. Once done, the lock may legally released by biodone. The * original owning process can no longer acquire it recursively, but must * wait until the I/O is completed and the lock has been freed by biodone. */ #define BUF_KERNPROC(bp) \ _lockmgr_disown(&(bp)->b_lock, LOCK_FILE, LOCK_LINE) #endif #endif /* _KERNEL */ struct buf_queue_head { TAILQ_HEAD(buf_queue, buf) queue; daddr_t last_pblkno; struct buf *insert_point; struct buf *switch_point; }; /* * This structure describes a clustered I/O. */ struct cluster_save { long bs_bcount; /* Saved b_bcount. */ long bs_bufsize; /* Saved b_bufsize. */ int bs_nchildren; /* Number of associated buffers. */ struct buf **bs_children; /* List of associated buffers. */ }; #ifdef _KERNEL static __inline int bwrite(struct buf *bp) { KASSERT(bp->b_bufobj != NULL, ("bwrite: no bufobj bp=%p", bp)); KASSERT(bp->b_bufobj->bo_ops != NULL, ("bwrite: no bo_ops bp=%p", bp)); KASSERT(bp->b_bufobj->bo_ops->bop_write != NULL, ("bwrite: no bop_write bp=%p", bp)); return (BO_WRITE(bp->b_bufobj, bp)); } static __inline void bstrategy(struct buf *bp) { KASSERT(bp->b_bufobj != NULL, ("bstrategy: no bufobj bp=%p", bp)); KASSERT(bp->b_bufobj->bo_ops != NULL, ("bstrategy: no bo_ops bp=%p", bp)); KASSERT(bp->b_bufobj->bo_ops->bop_strategy != NULL, ("bstrategy: no bop_strategy bp=%p", bp)); BO_STRATEGY(bp->b_bufobj, bp); } static __inline void buf_start(struct buf *bp) { if (bioops.io_start) (*bioops.io_start)(bp); } static __inline void buf_complete(struct buf *bp) { if (bioops.io_complete) (*bioops.io_complete)(bp); } static __inline void buf_deallocate(struct buf *bp) { if (bioops.io_deallocate) (*bioops.io_deallocate)(bp); } static __inline int buf_countdeps(struct buf *bp, int i) { if (bioops.io_countdeps) return ((*bioops.io_countdeps)(bp, i)); else return (0); } static __inline void buf_track(struct buf *bp, const char *location) { #if defined(FULL_BUF_TRACKING) bp->b_io_tracking[BUF_TRACKING_ENTRY(bp->b_io_tcnt++)] = location; #elif defined(BUF_TRACKING) bp->b_io_tracking = location; #endif } #endif /* _KERNEL */ /* * Zero out the buffer's data area. */ #define clrbuf(bp) { \ bzero((bp)->b_data, (u_int)(bp)->b_bcount); \ (bp)->b_resid = 0; \ } /* * Flags for getblk's last parameter. */ #define GB_LOCK_NOWAIT 0x0001 /* Fail if we block on a buf lock. */ #define GB_NOCREAT 0x0002 /* Don't create a buf if not found. */ #define GB_NOWAIT_BD 0x0004 /* Do not wait for bufdaemon. */ #define GB_UNMAPPED 0x0008 /* Do not mmap buffer pages. */ #define GB_KVAALLOC 0x0010 /* But allocate KVA. */ #define GB_CKHASH 0x0020 /* If reading, calc checksum hash */ #define GB_NOSPARSE 0x0040 /* Do not instantiate holes */ #ifdef _KERNEL extern int nbuf; /* The number of buffer headers */ extern long maxswzone; /* Max KVA for swap structures */ extern long maxbcache; /* Max KVA for buffer cache */ extern int maxbcachebuf; /* Max buffer cache block size */ extern long runningbufspace; extern long hibufspace; extern int dirtybufthresh; extern int bdwriteskip; extern int dirtybufferflushes; extern int altbufferflushes; extern int nswbuf; /* Number of swap I/O buffer headers. */ extern caddr_t unmapped_buf; /* Data address for unmapped buffers. */ static inline int buf_mapped(struct buf *bp) { return (bp->b_data != unmapped_buf); } void runningbufwakeup(struct buf *); void waitrunningbufspace(void); caddr_t kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est); void bufinit(void); void bufshutdown(int); void bdata2bio(struct buf *bp, struct bio *bip); void bwillwrite(void); int buf_dirty_count_severe(void); void bremfree(struct buf *); void bremfreef(struct buf *); /* XXX Force bremfree, only for nfs. */ #define bread(vp, blkno, size, cred, bpp) \ breadn_flags(vp, blkno, size, NULL, NULL, 0, cred, 0, NULL, bpp) #define bread_gb(vp, blkno, size, cred, gbflags, bpp) \ breadn_flags(vp, blkno, size, NULL, NULL, 0, cred, \ gbflags, NULL, bpp) #define breadn(vp, blkno, size, rablkno, rabsize, cnt, cred, bpp) \ breadn_flags(vp, blkno, size, rablkno, rabsize, cnt, cred, \ 0, NULL, bpp) int breadn_flags(struct vnode *, daddr_t, int, daddr_t *, int *, int, struct ucred *, int, void (*)(struct buf *), struct buf **); void bdwrite(struct buf *); void bawrite(struct buf *); void babarrierwrite(struct buf *); int bbarrierwrite(struct buf *); void bdirty(struct buf *); void bundirty(struct buf *); void bufstrategy(struct bufobj *, struct buf *); void brelse(struct buf *); void bqrelse(struct buf *); int vfs_bio_awrite(struct buf *); -void vfs_drain_busy_pages(struct buf *bp); +void vfs_busy_pages_acquire(struct buf *bp); +void vfs_busy_pages_release(struct buf *bp); struct buf *incore(struct bufobj *, daddr_t); struct buf *gbincore(struct bufobj *, daddr_t); struct buf *getblk(struct vnode *, daddr_t, int, int, int, int); int getblkx(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo, int flags, struct buf **bpp); struct buf *geteblk(int, int); int bufwait(struct buf *); int bufwrite(struct buf *); void bufdone(struct buf *); void bd_speedup(void); extern uma_zone_t pbuf_zone; uma_zone_t pbuf_zsecond_create(char *name, int max); int cluster_read(struct vnode *, u_quad_t, daddr_t, long, struct ucred *, long, int, int, struct buf **); int cluster_wbuild(struct vnode *, long, daddr_t, int, int); void cluster_write(struct vnode *, struct buf *, u_quad_t, int, int); void vfs_bio_brelse(struct buf *bp, int ioflags); void vfs_bio_bzero_buf(struct buf *bp, int base, int size); void vfs_bio_clrbuf(struct buf *); void vfs_bio_set_flags(struct buf *bp, int ioflags); void vfs_bio_set_valid(struct buf *, int base, int size); void vfs_busy_pages(struct buf *, int clear_modify); void vfs_unbusy_pages(struct buf *); int vmapbuf(struct buf *, int); void vunmapbuf(struct buf *); void brelvp(struct buf *); void bgetvp(struct vnode *, struct buf *); void pbgetbo(struct bufobj *bo, struct buf *bp); void pbgetvp(struct vnode *, struct buf *); void pbrelbo(struct buf *); void pbrelvp(struct buf *); int allocbuf(struct buf *bp, int size); void reassignbuf(struct buf *); void bwait(struct buf *, u_char, const char *); void bdone(struct buf *); typedef daddr_t (vbg_get_lblkno_t)(struct vnode *, vm_ooffset_t); typedef int (vbg_get_blksize_t)(struct vnode *, daddr_t); int vfs_bio_getpages(struct vnode *vp, struct vm_page **ma, int count, int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno, vbg_get_blksize_t get_blksize); #endif /* _KERNEL */ #endif /* !_SYS_BUF_H_ */ Index: head/sys/vm/phys_pager.c =================================================================== --- head/sys/vm/phys_pager.c (revision 353534) +++ head/sys/vm/phys_pager.c (revision 353535) @@ -1,269 +1,253 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2000 Peter Wemm * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* list of phys pager objects */ static struct pagerlst phys_pager_object_list; /* protect access to phys_pager_object_list */ static struct mtx phys_pager_mtx; static void phys_pager_init(void) { TAILQ_INIT(&phys_pager_object_list); mtx_init(&phys_pager_mtx, "phys_pager list", NULL, MTX_DEF); } static vm_object_t phys_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t foff, struct ucred *cred) { vm_object_t object, object1; vm_pindex_t pindex; /* * Offset should be page aligned. */ if (foff & PAGE_MASK) return (NULL); pindex = OFF_TO_IDX(foff + PAGE_MASK + size); if (handle != NULL) { mtx_lock(&phys_pager_mtx); /* * Look up pager, creating as necessary. */ object1 = NULL; object = vm_pager_object_lookup(&phys_pager_object_list, handle); if (object == NULL) { /* * Allocate object and associate it with the pager. */ mtx_unlock(&phys_pager_mtx); object1 = vm_object_allocate(OBJT_PHYS, pindex); mtx_lock(&phys_pager_mtx); object = vm_pager_object_lookup(&phys_pager_object_list, handle); if (object != NULL) { /* * We raced with other thread while * allocating object. */ if (pindex > object->size) object->size = pindex; } else { object = object1; object1 = NULL; object->handle = handle; vm_object_set_flag(object, OBJ_POPULATE); TAILQ_INSERT_TAIL(&phys_pager_object_list, object, pager_object_list); } } else { if (pindex > object->size) object->size = pindex; } mtx_unlock(&phys_pager_mtx); vm_object_deallocate(object1); } else { object = vm_object_allocate(OBJT_PHYS, pindex); vm_object_set_flag(object, OBJ_POPULATE); } return (object); } static void phys_pager_dealloc(vm_object_t object) { if (object->handle != NULL) { VM_OBJECT_WUNLOCK(object); mtx_lock(&phys_pager_mtx); TAILQ_REMOVE(&phys_pager_object_list, object, pager_object_list); mtx_unlock(&phys_pager_mtx); VM_OBJECT_WLOCK(object); } object->handle = NULL; object->type = OBJT_DEAD; } /* * Fill as many pages as vm_fault has allocated for us. */ static int phys_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind, int *rahead) { int i; VM_OBJECT_ASSERT_WLOCKED(object); for (i = 0; i < count; i++) { if (m[i]->valid == 0) { if ((m[i]->flags & PG_ZERO) == 0) pmap_zero_page(m[i]); m[i]->valid = VM_PAGE_BITS_ALL; } KASSERT(m[i]->valid == VM_PAGE_BITS_ALL, ("phys_pager_getpages: partially valid page %p", m[i])); KASSERT(m[i]->dirty == 0, ("phys_pager_getpages: dirty page %p", m[i])); } if (rbehind) *rbehind = 0; if (rahead) *rahead = 0; return (VM_PAGER_OK); } /* * Implement a pretty aggressive clustered getpages strategy. Hint that * everything in an entire 4MB window should be prefaulted at once. * * 4MB (1024 slots per page table page) is convenient for x86, * but may not be for other arches. */ #ifndef PHYSCLUSTER #define PHYSCLUSTER 1024 #endif static int phys_pager_cluster = PHYSCLUSTER; SYSCTL_INT(_vm, OID_AUTO, phys_pager_cluster, CTLFLAG_RWTUN, &phys_pager_cluster, 0, "prefault window size for phys pager"); /* * Max hint to vm_page_alloc() about the further allocation needs * inside the phys_pager_populate() loop. The number of bits used to * implement VM_ALLOC_COUNT() determines the hard limit on this value. * That limit is currently 65535. */ #define PHYSALLOC 16 static int phys_pager_populate(vm_object_t object, vm_pindex_t pidx, int fault_type __unused, vm_prot_t max_prot __unused, vm_pindex_t *first, vm_pindex_t *last) { vm_page_t m; vm_pindex_t base, end, i; int ahead; base = rounddown(pidx, phys_pager_cluster); end = base + phys_pager_cluster - 1; if (end >= object->size) end = object->size - 1; if (*first > base) base = *first; if (end > *last) end = *last; *first = base; *last = end; for (i = base; i <= end; i++) { -retry: - m = vm_page_lookup(object, i); - if (m == NULL) { - ahead = MIN(end - i, PHYSALLOC); - m = vm_page_alloc(object, i, VM_ALLOC_NORMAL | - VM_ALLOC_ZERO | VM_ALLOC_WAITFAIL | - VM_ALLOC_COUNT(ahead)); - if (m == NULL) - goto retry; - if ((m->flags & PG_ZERO) == 0) - pmap_zero_page(m); + ahead = MIN(end - i, PHYSALLOC); + m = vm_page_grab(object, i, + VM_ALLOC_NORMAL | VM_ALLOC_COUNT(ahead)); + if (m->valid != VM_PAGE_BITS_ALL) { + vm_page_zero_invalid(m, TRUE); m->valid = VM_PAGE_BITS_ALL; - } else if (vm_page_xbusied(m)) { - vm_page_sleep_if_xbusy(m, "physb"); - goto retry; - } else { - vm_page_xbusy(m); - if (m->valid != VM_PAGE_BITS_ALL) - vm_page_zero_invalid(m, TRUE); } - - KASSERT(m->valid == VM_PAGE_BITS_ALL, - ("phys_pager_populate: partially valid page %p", m)); KASSERT(m->dirty == 0, ("phys_pager_populate: dirty page %p", m)); } return (VM_PAGER_OK); } static void phys_pager_putpages(vm_object_t object, vm_page_t *m, int count, boolean_t sync, int *rtvals) { panic("phys_pager_putpage called"); } static boolean_t phys_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after) { vm_pindex_t base, end; base = rounddown(pindex, phys_pager_cluster); end = base + phys_pager_cluster - 1; if (before != NULL) *before = pindex - base; if (after != NULL) *after = end - pindex; return (TRUE); } struct pagerops physpagerops = { .pgo_init = phys_pager_init, .pgo_alloc = phys_pager_alloc, .pgo_dealloc = phys_pager_dealloc, .pgo_getpages = phys_pager_getpages, .pgo_putpages = phys_pager_putpages, .pgo_haspage = phys_pager_haspage, .pgo_populate = phys_pager_populate, }; Index: head/sys/vm/vm_fault.c =================================================================== --- head/sys/vm/vm_fault.c (revision 353534) +++ head/sys/vm/vm_fault.c (revision 353535) @@ -1,1904 +1,1904 @@ /*- * SPDX-License-Identifier: (BSD-4-Clause AND MIT-CMU) * * Copyright (c) 1991, 1993 * The Regents of the University of California. All rights reserved. * Copyright (c) 1994 John S. Dyson * All rights reserved. * Copyright (c) 1994 David Greenman * All rights reserved. * * * This code is derived from software contributed to Berkeley by * The Mach Operating System project at Carnegie-Mellon University. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. All advertising materials mentioning features or use of this software * must display the following acknowledgement: * This product includes software developed by the University of * California, Berkeley and its contributors. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)vm_fault.c 8.4 (Berkeley) 1/12/94 * * * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Authors: Avadis Tevanian, Jr., Michael Wayne Young * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. */ /* * Page fault handling module. */ #include __FBSDID("$FreeBSD$"); #include "opt_ktrace.h" #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef KTRACE #include #endif #include #include #include #include #include #include #include #include #include #include #include #define PFBAK 4 #define PFFOR 4 #define VM_FAULT_READ_DEFAULT (1 + VM_FAULT_READ_AHEAD_INIT) #define VM_FAULT_READ_MAX (1 + VM_FAULT_READ_AHEAD_MAX) #define VM_FAULT_DONTNEED_MIN 1048576 struct faultstate { vm_page_t m; vm_object_t object; vm_pindex_t pindex; vm_page_t first_m; vm_object_t first_object; vm_pindex_t first_pindex; vm_map_t map; vm_map_entry_t entry; int map_generation; bool lookup_still_valid; struct vnode *vp; }; static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead); static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra, int backward, int forward, bool obj_locked); static int vm_pfault_oom_attempts = 3; SYSCTL_INT(_vm, OID_AUTO, pfault_oom_attempts, CTLFLAG_RWTUN, &vm_pfault_oom_attempts, 0, "Number of page allocation attempts in page fault handler before it " "triggers OOM handling"); static int vm_pfault_oom_wait = 10; SYSCTL_INT(_vm, OID_AUTO, pfault_oom_wait, CTLFLAG_RWTUN, &vm_pfault_oom_wait, 0, "Number of seconds to wait for free pages before retrying " "the page fault handler"); static inline void release_page(struct faultstate *fs) { vm_page_xunbusy(fs->m); vm_page_lock(fs->m); vm_page_deactivate(fs->m); vm_page_unlock(fs->m); fs->m = NULL; } static inline void unlock_map(struct faultstate *fs) { if (fs->lookup_still_valid) { vm_map_lookup_done(fs->map, fs->entry); fs->lookup_still_valid = false; } } static void unlock_vp(struct faultstate *fs) { if (fs->vp != NULL) { vput(fs->vp); fs->vp = NULL; } } static void unlock_and_deallocate(struct faultstate *fs) { vm_object_pip_wakeup(fs->object); VM_OBJECT_WUNLOCK(fs->object); if (fs->object != fs->first_object) { VM_OBJECT_WLOCK(fs->first_object); vm_page_free(fs->first_m); vm_object_pip_wakeup(fs->first_object); VM_OBJECT_WUNLOCK(fs->first_object); fs->first_m = NULL; } vm_object_deallocate(fs->first_object); unlock_map(fs); unlock_vp(fs); } static void vm_fault_dirty(vm_map_entry_t entry, vm_page_t m, vm_prot_t prot, vm_prot_t fault_type, int fault_flags, bool set_wd) { bool need_dirty; if (((prot & VM_PROT_WRITE) == 0 && (fault_flags & VM_FAULT_DIRTY) == 0) || (m->oflags & VPO_UNMANAGED) != 0) return; VM_OBJECT_ASSERT_LOCKED(m->object); need_dirty = ((fault_type & VM_PROT_WRITE) != 0 && (fault_flags & VM_FAULT_WIRE) == 0) || (fault_flags & VM_FAULT_DIRTY) != 0; if (set_wd) vm_object_set_writeable_dirty(m->object); else /* * If two callers of vm_fault_dirty() with set_wd == * FALSE, one for the map entry with MAP_ENTRY_NOSYNC * flag set, other with flag clear, race, it is * possible for the no-NOSYNC thread to see m->dirty * != 0 and not clear VPO_NOSYNC. Take vm_page lock * around manipulation of VPO_NOSYNC and * vm_page_dirty() call, to avoid the race and keep * m->oflags consistent. */ vm_page_lock(m); /* * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC * if the page is already dirty to prevent data written with * the expectation of being synced from not being synced. * Likewise if this entry does not request NOSYNC then make * sure the page isn't marked NOSYNC. Applications sharing * data should use the same flags to avoid ping ponging. */ if ((entry->eflags & MAP_ENTRY_NOSYNC) != 0) { if (m->dirty == 0) { m->oflags |= VPO_NOSYNC; } } else { m->oflags &= ~VPO_NOSYNC; } /* * If the fault is a write, we know that this page is being * written NOW so dirty it explicitly to save on * pmap_is_modified() calls later. * * Also, since the page is now dirty, we can possibly tell * the pager to release any swap backing the page. Calling * the pager requires a write lock on the object. */ if (need_dirty) vm_page_dirty(m); if (!set_wd) vm_page_unlock(m); else if (need_dirty) vm_pager_page_unswapped(m); } /* * Unlocks fs.first_object and fs.map on success. */ static int vm_fault_soft_fast(struct faultstate *fs, vm_offset_t vaddr, vm_prot_t prot, int fault_type, int fault_flags, boolean_t wired, vm_page_t *m_hold) { vm_page_t m, m_map; #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \ __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \ VM_NRESERVLEVEL > 0 vm_page_t m_super; int flags; #endif int psind, rv; MPASS(fs->vp == NULL); m = vm_page_lookup(fs->first_object, fs->first_pindex); /* A busy page can be mapped for read|execute access. */ if (m == NULL || ((prot & VM_PROT_WRITE) != 0 && vm_page_busied(m)) || m->valid != VM_PAGE_BITS_ALL) return (KERN_FAILURE); m_map = m; psind = 0; #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \ __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \ VM_NRESERVLEVEL > 0 if ((m->flags & PG_FICTITIOUS) == 0 && (m_super = vm_reserv_to_superpage(m)) != NULL && rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start && roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end && (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) & (pagesizes[m_super->psind] - 1)) && !wired && pmap_ps_enabled(fs->map->pmap)) { flags = PS_ALL_VALID; if ((prot & VM_PROT_WRITE) != 0) { /* * Create a superpage mapping allowing write access * only if none of the constituent pages are busy and * all of them are already dirty (except possibly for * the page that was faulted on). */ flags |= PS_NONE_BUSY; if ((fs->first_object->flags & OBJ_UNMANAGED) == 0) flags |= PS_ALL_DIRTY; } if (vm_page_ps_test(m_super, flags, m)) { m_map = m_super; psind = m_super->psind; vaddr = rounddown2(vaddr, pagesizes[psind]); /* Preset the modified bit for dirty superpages. */ if ((flags & PS_ALL_DIRTY) != 0) fault_type |= VM_PROT_WRITE; } } #endif rv = pmap_enter(fs->map->pmap, vaddr, m_map, prot, fault_type | PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED : 0), psind); if (rv != KERN_SUCCESS) return (rv); if (m_hold != NULL) { *m_hold = m; vm_page_wire(m); } vm_fault_dirty(fs->entry, m, prot, fault_type, fault_flags, false); if (psind == 0 && !wired) vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true); VM_OBJECT_RUNLOCK(fs->first_object); vm_map_lookup_done(fs->map, fs->entry); curthread->td_ru.ru_minflt++; return (KERN_SUCCESS); } static void vm_fault_restore_map_lock(struct faultstate *fs) { VM_OBJECT_ASSERT_WLOCKED(fs->first_object); MPASS(REFCOUNT_COUNT(fs->first_object->paging_in_progress) > 0); if (!vm_map_trylock_read(fs->map)) { VM_OBJECT_WUNLOCK(fs->first_object); vm_map_lock_read(fs->map); VM_OBJECT_WLOCK(fs->first_object); } fs->lookup_still_valid = true; } static void vm_fault_populate_check_page(vm_page_t m) { /* * Check each page to ensure that the pager is obeying the * interface: the page must be installed in the object, fully * valid, and exclusively busied. */ MPASS(m != NULL); MPASS(m->valid == VM_PAGE_BITS_ALL); MPASS(vm_page_xbusied(m)); } static void vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first, vm_pindex_t last) { vm_page_t m; vm_pindex_t pidx; VM_OBJECT_ASSERT_WLOCKED(object); MPASS(first <= last); for (pidx = first, m = vm_page_lookup(object, pidx); pidx <= last; pidx++, m = vm_page_next(m)) { vm_fault_populate_check_page(m); vm_page_lock(m); vm_page_deactivate(m); vm_page_unlock(m); vm_page_xunbusy(m); } } static int vm_fault_populate(struct faultstate *fs, vm_prot_t prot, int fault_type, int fault_flags, boolean_t wired, vm_page_t *m_hold) { struct mtx *m_mtx; vm_offset_t vaddr; vm_page_t m; vm_pindex_t map_first, map_last, pager_first, pager_last, pidx; int i, npages, psind, rv; MPASS(fs->object == fs->first_object); VM_OBJECT_ASSERT_WLOCKED(fs->first_object); MPASS(REFCOUNT_COUNT(fs->first_object->paging_in_progress) > 0); MPASS(fs->first_object->backing_object == NULL); MPASS(fs->lookup_still_valid); pager_first = OFF_TO_IDX(fs->entry->offset); pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1; unlock_map(fs); unlock_vp(fs); /* * Call the pager (driver) populate() method. * * There is no guarantee that the method will be called again * if the current fault is for read, and a future fault is * for write. Report the entry's maximum allowed protection * to the driver. */ rv = vm_pager_populate(fs->first_object, fs->first_pindex, fault_type, fs->entry->max_protection, &pager_first, &pager_last); VM_OBJECT_ASSERT_WLOCKED(fs->first_object); if (rv == VM_PAGER_BAD) { /* * VM_PAGER_BAD is the backdoor for a pager to request * normal fault handling. */ vm_fault_restore_map_lock(fs); if (fs->map->timestamp != fs->map_generation) return (KERN_RESOURCE_SHORTAGE); /* RetryFault */ return (KERN_NOT_RECEIVER); } if (rv != VM_PAGER_OK) return (KERN_FAILURE); /* AKA SIGSEGV */ /* Ensure that the driver is obeying the interface. */ MPASS(pager_first <= pager_last); MPASS(fs->first_pindex <= pager_last); MPASS(fs->first_pindex >= pager_first); MPASS(pager_last < fs->first_object->size); vm_fault_restore_map_lock(fs); if (fs->map->timestamp != fs->map_generation) { vm_fault_populate_cleanup(fs->first_object, pager_first, pager_last); return (KERN_RESOURCE_SHORTAGE); /* RetryFault */ } /* * The map is unchanged after our last unlock. Process the fault. * * The range [pager_first, pager_last] that is given to the * pager is only a hint. The pager may populate any range * within the object that includes the requested page index. * In case the pager expanded the range, clip it to fit into * the map entry. */ map_first = OFF_TO_IDX(fs->entry->offset); if (map_first > pager_first) { vm_fault_populate_cleanup(fs->first_object, pager_first, map_first - 1); pager_first = map_first; } map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1; if (map_last < pager_last) { vm_fault_populate_cleanup(fs->first_object, map_last + 1, pager_last); pager_last = map_last; } for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx); pidx <= pager_last; pidx += npages, m = vm_page_next(&m[npages - 1])) { vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset; #if defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \ __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv) psind = m->psind; if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 || pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last || !pmap_ps_enabled(fs->map->pmap) || wired)) psind = 0; #else psind = 0; #endif npages = atop(pagesizes[psind]); for (i = 0; i < npages; i++) { vm_fault_populate_check_page(&m[i]); vm_fault_dirty(fs->entry, &m[i], prot, fault_type, fault_flags, true); } VM_OBJECT_WUNLOCK(fs->first_object); rv = pmap_enter(fs->map->pmap, vaddr, m, prot, fault_type | (wired ? PMAP_ENTER_WIRED : 0), psind); #if defined(__amd64__) if (psind > 0 && rv == KERN_FAILURE) { for (i = 0; i < npages; i++) { rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i), &m[i], prot, fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0); MPASS(rv == KERN_SUCCESS); } } #else MPASS(rv == KERN_SUCCESS); #endif VM_OBJECT_WLOCK(fs->first_object); m_mtx = NULL; for (i = 0; i < npages; i++) { if ((fault_flags & VM_FAULT_WIRE) != 0) { vm_page_wire(&m[i]); } else { vm_page_change_lock(&m[i], &m_mtx); vm_page_activate(&m[i]); } if (m_hold != NULL && m[i].pindex == fs->first_pindex) { *m_hold = &m[i]; vm_page_wire(&m[i]); } vm_page_xunbusy(&m[i]); } if (m_mtx != NULL) mtx_unlock(m_mtx); } curthread->td_ru.ru_majflt++; return (KERN_SUCCESS); } static int prot_fault_translation; SYSCTL_INT(_machdep, OID_AUTO, prot_fault_translation, CTLFLAG_RWTUN, &prot_fault_translation, 0, "Control signal to deliver on protection fault"); /* compat definition to keep common code for signal translation */ #define UCODE_PAGEFLT 12 #ifdef T_PAGEFLT _Static_assert(UCODE_PAGEFLT == T_PAGEFLT, "T_PAGEFLT"); #endif /* * vm_fault_trap: * * Handle a page fault occurring at the given address, * requiring the given permissions, in the map specified. * If successful, the page is inserted into the * associated physical map. * * NOTE: the given address should be truncated to the * proper page address. * * KERN_SUCCESS is returned if the page fault is handled; otherwise, * a standard error specifying why the fault is fatal is returned. * * The map in question must be referenced, and remains so. * Caller may hold no locks. */ int vm_fault_trap(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags, int *signo, int *ucode) { int result; MPASS(signo == NULL || ucode != NULL); #ifdef KTRACE if (map != kernel_map && KTRPOINT(curthread, KTR_FAULT)) ktrfault(vaddr, fault_type); #endif result = vm_fault(map, trunc_page(vaddr), fault_type, fault_flags, NULL); KASSERT(result == KERN_SUCCESS || result == KERN_FAILURE || result == KERN_INVALID_ADDRESS || result == KERN_RESOURCE_SHORTAGE || result == KERN_PROTECTION_FAILURE || result == KERN_OUT_OF_BOUNDS, ("Unexpected Mach error %d from vm_fault()", result)); #ifdef KTRACE if (map != kernel_map && KTRPOINT(curthread, KTR_FAULTEND)) ktrfaultend(result); #endif if (result != KERN_SUCCESS && signo != NULL) { switch (result) { case KERN_FAILURE: case KERN_INVALID_ADDRESS: *signo = SIGSEGV; *ucode = SEGV_MAPERR; break; case KERN_RESOURCE_SHORTAGE: *signo = SIGBUS; *ucode = BUS_OOMERR; break; case KERN_OUT_OF_BOUNDS: *signo = SIGBUS; *ucode = BUS_OBJERR; break; case KERN_PROTECTION_FAILURE: if (prot_fault_translation == 0) { /* * Autodetect. This check also covers * the images without the ABI-tag ELF * note. */ if (SV_CURPROC_ABI() == SV_ABI_FREEBSD && curproc->p_osrel >= P_OSREL_SIGSEGV) { *signo = SIGSEGV; *ucode = SEGV_ACCERR; } else { *signo = SIGBUS; *ucode = UCODE_PAGEFLT; } } else if (prot_fault_translation == 1) { /* Always compat mode. */ *signo = SIGBUS; *ucode = UCODE_PAGEFLT; } else { /* Always SIGSEGV mode. */ *signo = SIGSEGV; *ucode = SEGV_ACCERR; } break; default: KASSERT(0, ("Unexpected Mach error %d from vm_fault()", result)); break; } } return (result); } int vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags, vm_page_t *m_hold) { struct faultstate fs; struct vnode *vp; struct domainset *dset; vm_object_t next_object, retry_object; vm_offset_t e_end, e_start; vm_pindex_t retry_pindex; vm_prot_t prot, retry_prot; int ahead, alloc_req, behind, cluster_offset, error, era, faultcount; int locked, nera, oom, result, rv; u_char behavior; boolean_t wired; /* Passed by reference. */ bool dead, hardfault, is_first_object_locked; VM_CNT_INC(v_vm_faults); if ((curthread->td_pflags & TDP_NOFAULTING) != 0) return (KERN_PROTECTION_FAILURE); fs.vp = NULL; faultcount = 0; nera = -1; hardfault = false; RetryFault: oom = 0; RetryFault_oom: /* * Find the backing store object and offset into it to begin the * search. */ fs.map = map; result = vm_map_lookup(&fs.map, vaddr, fault_type | VM_PROT_FAULT_LOOKUP, &fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired); if (result != KERN_SUCCESS) { unlock_vp(&fs); return (result); } fs.map_generation = fs.map->timestamp; if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { panic("%s: fault on nofault entry, addr: %#lx", __func__, (u_long)vaddr); } if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION && fs.entry->wiring_thread != curthread) { vm_map_unlock_read(fs.map); vm_map_lock(fs.map); if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) && (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) { unlock_vp(&fs); fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; vm_map_unlock_and_wait(fs.map, 0); } else vm_map_unlock(fs.map); goto RetryFault; } MPASS((fs.entry->eflags & MAP_ENTRY_GUARD) == 0); if (wired) fault_type = prot | (fault_type & VM_PROT_COPY); else KASSERT((fault_flags & VM_FAULT_WIRE) == 0, ("!wired && VM_FAULT_WIRE")); /* * Try to avoid lock contention on the top-level object through * special-case handling of some types of page faults, specifically, * those that are both (1) mapping an existing page from the top- * level object and (2) not having to mark that object as containing * dirty pages. Under these conditions, a read lock on the top-level * object suffices, allowing multiple page faults of a similar type to * run in parallel on the same top-level object. */ if (fs.vp == NULL /* avoid locked vnode leak */ && (fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0 && /* avoid calling vm_object_set_writeable_dirty() */ ((prot & VM_PROT_WRITE) == 0 || (fs.first_object->type != OBJT_VNODE && (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) || (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0)) { VM_OBJECT_RLOCK(fs.first_object); if ((prot & VM_PROT_WRITE) == 0 || (fs.first_object->type != OBJT_VNODE && (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) || (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0) { rv = vm_fault_soft_fast(&fs, vaddr, prot, fault_type, fault_flags, wired, m_hold); if (rv == KERN_SUCCESS) return (rv); } if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) { VM_OBJECT_RUNLOCK(fs.first_object); VM_OBJECT_WLOCK(fs.first_object); } } else { VM_OBJECT_WLOCK(fs.first_object); } /* * Make a reference to this object to prevent its disposal while we * are messing with it. Once we have the reference, the map is free * to be diddled. Since objects reference their shadows (and copies), * they will stay around as well. * * Bump the paging-in-progress count to prevent size changes (e.g. * truncation operations) during I/O. */ vm_object_reference_locked(fs.first_object); vm_object_pip_add(fs.first_object, 1); fs.lookup_still_valid = true; fs.first_m = NULL; /* * Search for the page at object/offset. */ fs.object = fs.first_object; fs.pindex = fs.first_pindex; while (TRUE) { /* * If the object is marked for imminent termination, * we retry here, since the collapse pass has raced * with us. Otherwise, if we see terminally dead * object, return fail. */ if ((fs.object->flags & OBJ_DEAD) != 0) { dead = fs.object->type == OBJT_DEAD; unlock_and_deallocate(&fs); if (dead) return (KERN_PROTECTION_FAILURE); pause("vmf_de", 1); goto RetryFault; } /* * See if page is resident */ fs.m = vm_page_lookup(fs.object, fs.pindex); if (fs.m != NULL) { /* * Wait/Retry if the page is busy. We have to do this * if the page is either exclusive or shared busy * because the vm_pager may be using read busy for * pageouts (and even pageins if it is the vnode * pager), and we could end up trying to pagein and * pageout the same page simultaneously. * * We can theoretically allow the busy case on a read * fault if the page is marked valid, but since such * pages are typically already pmap'd, putting that * special case in might be more effort then it is * worth. We cannot under any circumstances mess * around with a shared busied page except, perhaps, * to pmap it. */ - if (vm_page_busied(fs.m)) { + if (vm_page_tryxbusy(fs.m) == 0) { /* * Reference the page before unlocking and * sleeping so that the page daemon is less * likely to reclaim it. */ vm_page_aflag_set(fs.m, PGA_REFERENCED); if (fs.object != fs.first_object) { if (!VM_OBJECT_TRYWLOCK( fs.first_object)) { VM_OBJECT_WUNLOCK(fs.object); VM_OBJECT_WLOCK(fs.first_object); VM_OBJECT_WLOCK(fs.object); } vm_page_free(fs.first_m); vm_object_pip_wakeup(fs.first_object); VM_OBJECT_WUNLOCK(fs.first_object); fs.first_m = NULL; } unlock_map(&fs); if (fs.m == vm_page_lookup(fs.object, fs.pindex)) { vm_page_sleep_if_busy(fs.m, "vmpfw"); } vm_object_pip_wakeup(fs.object); VM_OBJECT_WUNLOCK(fs.object); VM_CNT_INC(v_intrans); vm_object_deallocate(fs.first_object); goto RetryFault; } /* - * Mark page busy for other processes, and the + * The page is marked busy for other processes and the * pagedaemon. If it still isn't completely valid * (readable), jump to readrest, else break-out ( we * found the page ). */ - vm_page_xbusy(fs.m); if (fs.m->valid != VM_PAGE_BITS_ALL) goto readrest; break; /* break to PAGE HAS BEEN FOUND */ } KASSERT(fs.m == NULL, ("fs.m should be NULL, not %p", fs.m)); /* * Page is not resident. If the pager might contain the page * or this is the beginning of the search, allocate a new * page. (Default objects are zero-fill, so there is no real * pager for them.) */ if (fs.object->type != OBJT_DEFAULT || fs.object == fs.first_object) { if (fs.pindex >= fs.object->size) { unlock_and_deallocate(&fs); return (KERN_OUT_OF_BOUNDS); } if (fs.object == fs.first_object && (fs.first_object->flags & OBJ_POPULATE) != 0 && fs.first_object->shadow_count == 0) { rv = vm_fault_populate(&fs, prot, fault_type, fault_flags, wired, m_hold); switch (rv) { case KERN_SUCCESS: case KERN_FAILURE: unlock_and_deallocate(&fs); return (rv); case KERN_RESOURCE_SHORTAGE: unlock_and_deallocate(&fs); goto RetryFault; case KERN_NOT_RECEIVER: /* * Pager's populate() method * returned VM_PAGER_BAD. */ break; default: panic("inconsistent return codes"); } } /* * Allocate a new page for this object/offset pair. * * Unlocked read of the p_flag is harmless. At * worst, the P_KILLED might be not observed * there, and allocation can fail, causing * restart and new reading of the p_flag. */ dset = fs.object->domain.dr_policy; if (dset == NULL) dset = curthread->td_domain.dr_policy; if (!vm_page_count_severe_set(&dset->ds_mask) || P_KILLED(curproc)) { #if VM_NRESERVLEVEL > 0 vm_object_color(fs.object, atop(vaddr) - fs.pindex); #endif alloc_req = P_KILLED(curproc) ? VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL; if (fs.object->type != OBJT_VNODE && fs.object->backing_object == NULL) alloc_req |= VM_ALLOC_ZERO; fs.m = vm_page_alloc(fs.object, fs.pindex, alloc_req); } if (fs.m == NULL) { unlock_and_deallocate(&fs); if (vm_pfault_oom_attempts < 0 || oom < vm_pfault_oom_attempts) { oom++; vm_waitpfault(dset, vm_pfault_oom_wait * hz); goto RetryFault_oom; } if (bootverbose) printf( "proc %d (%s) failed to alloc page on fault, starting OOM\n", curproc->p_pid, curproc->p_comm); vm_pageout_oom(VM_OOM_MEM_PF); goto RetryFault; } } readrest: /* * At this point, we have either allocated a new page or found * an existing page that is only partially valid. * * We hold a reference on the current object and the page is * exclusive busied. */ /* * If the pager for the current object might have the page, * then determine the number of additional pages to read and * potentially reprioritize previously read pages for earlier * reclamation. These operations should only be performed * once per page fault. Even if the current pager doesn't * have the page, the number of additional pages to read will * apply to subsequent objects in the shadow chain. */ if (fs.object->type != OBJT_DEFAULT && nera == -1 && !P_KILLED(curproc)) { KASSERT(fs.lookup_still_valid, ("map unlocked")); era = fs.entry->read_ahead; behavior = vm_map_entry_behavior(fs.entry); if (behavior == MAP_ENTRY_BEHAV_RANDOM) { nera = 0; } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) { nera = VM_FAULT_READ_AHEAD_MAX; if (vaddr == fs.entry->next_read) vm_fault_dontneed(&fs, vaddr, nera); } else if (vaddr == fs.entry->next_read) { /* * This is a sequential fault. Arithmetically * increase the requested number of pages in * the read-ahead window. The requested * number of pages is "# of sequential faults * x (read ahead min + 1) + read ahead min" */ nera = VM_FAULT_READ_AHEAD_MIN; if (era > 0) { nera += era + 1; if (nera > VM_FAULT_READ_AHEAD_MAX) nera = VM_FAULT_READ_AHEAD_MAX; } if (era == VM_FAULT_READ_AHEAD_MAX) vm_fault_dontneed(&fs, vaddr, nera); } else { /* * This is a non-sequential fault. */ nera = 0; } if (era != nera) { /* * A read lock on the map suffices to update * the read ahead count safely. */ fs.entry->read_ahead = nera; } /* * Prepare for unlocking the map. Save the map * entry's start and end addresses, which are used to * optimize the size of the pager operation below. * Even if the map entry's addresses change after * unlocking the map, using the saved addresses is * safe. */ e_start = fs.entry->start; e_end = fs.entry->end; } /* * Call the pager to retrieve the page if there is a chance * that the pager has it, and potentially retrieve additional * pages at the same time. */ if (fs.object->type != OBJT_DEFAULT) { /* * Release the map lock before locking the vnode or * sleeping in the pager. (If the current object has * a shadow, then an earlier iteration of this loop * may have already unlocked the map.) */ unlock_map(&fs); if (fs.object->type == OBJT_VNODE && (vp = fs.object->handle) != fs.vp) { /* * Perform an unlock in case the desired vnode * changed while the map was unlocked during a * retry. */ unlock_vp(&fs); locked = VOP_ISLOCKED(vp); if (locked != LK_EXCLUSIVE) locked = LK_SHARED; /* * We must not sleep acquiring the vnode lock * while we have the page exclusive busied or * the object's paging-in-progress count * incremented. Otherwise, we could deadlock. */ error = vget(vp, locked | LK_CANRECURSE | LK_NOWAIT, curthread); if (error != 0) { vhold(vp); release_page(&fs); unlock_and_deallocate(&fs); error = vget(vp, locked | LK_RETRY | LK_CANRECURSE, curthread); vdrop(vp); fs.vp = vp; KASSERT(error == 0, ("vm_fault: vget failed")); goto RetryFault; } fs.vp = vp; } KASSERT(fs.vp == NULL || !fs.map->system_map, ("vm_fault: vnode-backed object mapped by system map")); /* * Page in the requested page and hint the pager, * that it may bring up surrounding pages. */ if (nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM || P_KILLED(curproc)) { behind = 0; ahead = 0; } else { /* Is this a sequential fault? */ if (nera > 0) { behind = 0; ahead = nera; } else { /* * Request a cluster of pages that is * aligned to a VM_FAULT_READ_DEFAULT * page offset boundary within the * object. Alignment to a page offset * boundary is more likely to coincide * with the underlying file system * block than alignment to a virtual * address boundary. */ cluster_offset = fs.pindex % VM_FAULT_READ_DEFAULT; behind = ulmin(cluster_offset, atop(vaddr - e_start)); ahead = VM_FAULT_READ_DEFAULT - 1 - cluster_offset; } ahead = ulmin(ahead, atop(e_end - vaddr) - 1); } rv = vm_pager_get_pages(fs.object, &fs.m, 1, &behind, &ahead); if (rv == VM_PAGER_OK) { faultcount = behind + 1 + ahead; hardfault = true; break; /* break to PAGE HAS BEEN FOUND */ } if (rv == VM_PAGER_ERROR) printf("vm_fault: pager read error, pid %d (%s)\n", curproc->p_pid, curproc->p_comm); /* * If an I/O error occurred or the requested page was * outside the range of the pager, clean up and return * an error. */ if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) { if (!vm_page_wired(fs.m)) vm_page_free(fs.m); else vm_page_xunbusy(fs.m); fs.m = NULL; unlock_and_deallocate(&fs); return (KERN_OUT_OF_BOUNDS); } /* * The requested page does not exist at this object/ * offset. Remove the invalid page from the object, * waking up anyone waiting for it, and continue on to * the next object. However, if this is the top-level * object, we must leave the busy page in place to * prevent another process from rushing past us, and * inserting the page in that object at the same time * that we are. */ if (fs.object != fs.first_object) { if (!vm_page_wired(fs.m)) vm_page_free(fs.m); else vm_page_xunbusy(fs.m); fs.m = NULL; } } /* * We get here if the object has default pager (or unwiring) * or the pager doesn't have the page. */ if (fs.object == fs.first_object) fs.first_m = fs.m; /* * Move on to the next object. Lock the next object before * unlocking the current one. */ next_object = fs.object->backing_object; if (next_object == NULL) { /* * If there's no object left, fill the page in the top * object with zeros. */ if (fs.object != fs.first_object) { vm_object_pip_wakeup(fs.object); VM_OBJECT_WUNLOCK(fs.object); fs.object = fs.first_object; fs.pindex = fs.first_pindex; fs.m = fs.first_m; VM_OBJECT_WLOCK(fs.object); } fs.first_m = NULL; /* * Zero the page if necessary and mark it valid. */ if ((fs.m->flags & PG_ZERO) == 0) { pmap_zero_page(fs.m); } else { VM_CNT_INC(v_ozfod); } VM_CNT_INC(v_zfod); fs.m->valid = VM_PAGE_BITS_ALL; /* Don't try to prefault neighboring pages. */ faultcount = 1; break; /* break to PAGE HAS BEEN FOUND */ } else { KASSERT(fs.object != next_object, ("object loop %p", next_object)); VM_OBJECT_WLOCK(next_object); vm_object_pip_add(next_object, 1); if (fs.object != fs.first_object) vm_object_pip_wakeup(fs.object); fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset); VM_OBJECT_WUNLOCK(fs.object); fs.object = next_object; } } vm_page_assert_xbusied(fs.m); /* * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock * is held.] */ /* * If the page is being written, but isn't already owned by the * top-level object, we have to copy it into a new page owned by the * top-level object. */ if (fs.object != fs.first_object) { /* * We only really need to copy if we want to write it. */ if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) { /* * This allows pages to be virtually copied from a * backing_object into the first_object, where the * backing object has no other refs to it, and cannot * gain any more refs. Instead of a bcopy, we just * move the page from the backing object to the * first object. Note that we must mark the page * dirty in the first object so that it will go out * to swap when needed. */ is_first_object_locked = false; if ( /* * Only one shadow object */ (fs.object->shadow_count == 1) && /* * No COW refs, except us */ (fs.object->ref_count == 1) && /* * No one else can look this object up */ (fs.object->handle == NULL) && /* * No other ways to look the object up */ ((fs.object->type == OBJT_DEFAULT) || (fs.object->type == OBJT_SWAP)) && (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) && /* * We don't chase down the shadow chain */ fs.object == fs.first_object->backing_object) { (void)vm_page_remove(fs.m); vm_page_replace_checked(fs.m, fs.first_object, fs.first_pindex, fs.first_m); vm_page_free(fs.first_m); vm_page_dirty(fs.m); #if VM_NRESERVLEVEL > 0 /* * Rename the reservation. */ vm_reserv_rename(fs.m, fs.first_object, fs.object, OFF_TO_IDX( fs.first_object->backing_object_offset)); #endif /* * Removing the page from the backing object * unbusied it. */ vm_page_xbusy(fs.m); fs.first_m = fs.m; fs.m = NULL; VM_CNT_INC(v_cow_optim); } else { /* * Oh, well, lets copy it. */ pmap_copy_page(fs.m, fs.first_m); fs.first_m->valid = VM_PAGE_BITS_ALL; if (wired && (fault_flags & VM_FAULT_WIRE) == 0) { vm_page_wire(fs.first_m); vm_page_unwire(fs.m, PQ_INACTIVE); } /* * We no longer need the old page or object. */ release_page(&fs); } /* * fs.object != fs.first_object due to above * conditional */ vm_object_pip_wakeup(fs.object); VM_OBJECT_WUNLOCK(fs.object); /* * We only try to prefault read-only mappings to the * neighboring pages when this copy-on-write fault is * a hard fault. In other cases, trying to prefault * is typically wasted effort. */ if (faultcount == 0) faultcount = 1; /* * Only use the new page below... */ fs.object = fs.first_object; fs.pindex = fs.first_pindex; fs.m = fs.first_m; if (!is_first_object_locked) VM_OBJECT_WLOCK(fs.object); VM_CNT_INC(v_cow_faults); curthread->td_cow++; } else { prot &= ~VM_PROT_WRITE; } } /* * We must verify that the maps have not changed since our last * lookup. */ if (!fs.lookup_still_valid) { if (!vm_map_trylock_read(fs.map)) { release_page(&fs); unlock_and_deallocate(&fs); goto RetryFault; } fs.lookup_still_valid = true; if (fs.map->timestamp != fs.map_generation) { result = vm_map_lookup_locked(&fs.map, vaddr, fault_type, &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired); /* * If we don't need the page any longer, put it on the inactive * list (the easiest thing to do here). If no one needs it, * pageout will grab it eventually. */ if (result != KERN_SUCCESS) { release_page(&fs); unlock_and_deallocate(&fs); /* * If retry of map lookup would have blocked then * retry fault from start. */ if (result == KERN_FAILURE) goto RetryFault; return (result); } if ((retry_object != fs.first_object) || (retry_pindex != fs.first_pindex)) { release_page(&fs); unlock_and_deallocate(&fs); goto RetryFault; } /* * Check whether the protection has changed or the object has * been copied while we left the map unlocked. Changing from * read to write permission is OK - we leave the page * write-protected, and catch the write fault. Changing from * write to read permission means that we can't mark the page * write-enabled after all. */ prot &= retry_prot; fault_type &= retry_prot; if (prot == 0) { release_page(&fs); unlock_and_deallocate(&fs); goto RetryFault; } /* Reassert because wired may have changed. */ KASSERT(wired || (fault_flags & VM_FAULT_WIRE) == 0, ("!wired && VM_FAULT_WIRE")); } } /* * If the page was filled by a pager, save the virtual address that * should be faulted on next under a sequential access pattern to the * map entry. A read lock on the map suffices to update this address * safely. */ if (hardfault) fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE; vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags, true); vm_page_assert_xbusied(fs.m); /* * Page must be completely valid or it is not fit to * map into user space. vm_pager_get_pages() ensures this. */ KASSERT(fs.m->valid == VM_PAGE_BITS_ALL, ("vm_fault: page %p partially invalid", fs.m)); VM_OBJECT_WUNLOCK(fs.object); /* * Put this page into the physical map. We had to do the unlock above * because pmap_enter() may sleep. We don't put the page * back on the active queue until later so that the pageout daemon * won't find it (yet). */ pmap_enter(fs.map->pmap, vaddr, fs.m, prot, fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0); if (faultcount != 1 && (fault_flags & VM_FAULT_WIRE) == 0 && wired == 0) vm_fault_prefault(&fs, vaddr, faultcount > 0 ? behind : PFBAK, faultcount > 0 ? ahead : PFFOR, false); VM_OBJECT_WLOCK(fs.object); /* * If the page is not wired down, then put it where the pageout daemon * can find it. */ if ((fault_flags & VM_FAULT_WIRE) != 0) { vm_page_wire(fs.m); } else { vm_page_lock(fs.m); vm_page_activate(fs.m); vm_page_unlock(fs.m); } if (m_hold != NULL) { *m_hold = fs.m; vm_page_wire(fs.m); } vm_page_xunbusy(fs.m); /* * Unlock everything, and return */ unlock_and_deallocate(&fs); if (hardfault) { VM_CNT_INC(v_io_faults); curthread->td_ru.ru_majflt++; #ifdef RACCT if (racct_enable && fs.object->type == OBJT_VNODE) { PROC_LOCK(curproc); if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) { racct_add_force(curproc, RACCT_WRITEBPS, PAGE_SIZE + behind * PAGE_SIZE); racct_add_force(curproc, RACCT_WRITEIOPS, 1); } else { racct_add_force(curproc, RACCT_READBPS, PAGE_SIZE + ahead * PAGE_SIZE); racct_add_force(curproc, RACCT_READIOPS, 1); } PROC_UNLOCK(curproc); } #endif } else curthread->td_ru.ru_minflt++; return (KERN_SUCCESS); } /* * Speed up the reclamation of pages that precede the faulting pindex within * the first object of the shadow chain. Essentially, perform the equivalent * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes * the faulting pindex by the cluster size when the pages read by vm_fault() * cross a cluster-size boundary. The cluster size is the greater of the * smallest superpage size and VM_FAULT_DONTNEED_MIN. * * When "fs->first_object" is a shadow object, the pages in the backing object * that precede the faulting pindex are deactivated by vm_fault(). So, this * function must only be concerned with pages in the first object. */ static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead) { vm_map_entry_t entry; vm_object_t first_object, object; vm_offset_t end, start; vm_page_t m, m_next; vm_pindex_t pend, pstart; vm_size_t size; object = fs->object; VM_OBJECT_ASSERT_WLOCKED(object); first_object = fs->first_object; if (first_object != object) { if (!VM_OBJECT_TRYWLOCK(first_object)) { VM_OBJECT_WUNLOCK(object); VM_OBJECT_WLOCK(first_object); VM_OBJECT_WLOCK(object); } } /* Neither fictitious nor unmanaged pages can be reclaimed. */ if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) { size = VM_FAULT_DONTNEED_MIN; if (MAXPAGESIZES > 1 && size < pagesizes[1]) size = pagesizes[1]; end = rounddown2(vaddr, size); if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) && (entry = fs->entry)->start < end) { if (end - entry->start < size) start = entry->start; else start = end - size; pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED); pstart = OFF_TO_IDX(entry->offset) + atop(start - entry->start); m_next = vm_page_find_least(first_object, pstart); pend = OFF_TO_IDX(entry->offset) + atop(end - entry->start); while ((m = m_next) != NULL && m->pindex < pend) { m_next = TAILQ_NEXT(m, listq); if (m->valid != VM_PAGE_BITS_ALL || vm_page_busied(m)) continue; /* * Don't clear PGA_REFERENCED, since it would * likely represent a reference by a different * process. * * Typically, at this point, prefetched pages * are still in the inactive queue. Only * pages that triggered page faults are in the * active queue. */ vm_page_lock(m); if (!vm_page_inactive(m)) vm_page_deactivate(m); vm_page_unlock(m); } } } if (first_object != object) VM_OBJECT_WUNLOCK(first_object); } /* * vm_fault_prefault provides a quick way of clustering * pagefaults into a processes address space. It is a "cousin" * of vm_map_pmap_enter, except it runs at page fault time instead * of mmap time. */ static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra, int backward, int forward, bool obj_locked) { pmap_t pmap; vm_map_entry_t entry; vm_object_t backing_object, lobject; vm_offset_t addr, starta; vm_pindex_t pindex; vm_page_t m; int i; pmap = fs->map->pmap; if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)) return; entry = fs->entry; if (addra < backward * PAGE_SIZE) { starta = entry->start; } else { starta = addra - backward * PAGE_SIZE; if (starta < entry->start) starta = entry->start; } /* * Generate the sequence of virtual addresses that are candidates for * prefaulting in an outward spiral from the faulting virtual address, * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ... * If the candidate address doesn't have a backing physical page, then * the loop immediately terminates. */ for (i = 0; i < 2 * imax(backward, forward); i++) { addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE : PAGE_SIZE); if (addr > addra + forward * PAGE_SIZE) addr = 0; if (addr < starta || addr >= entry->end) continue; if (!pmap_is_prefaultable(pmap, addr)) continue; pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; lobject = entry->object.vm_object; if (!obj_locked) VM_OBJECT_RLOCK(lobject); while ((m = vm_page_lookup(lobject, pindex)) == NULL && lobject->type == OBJT_DEFAULT && (backing_object = lobject->backing_object) != NULL) { KASSERT((lobject->backing_object_offset & PAGE_MASK) == 0, ("vm_fault_prefault: unaligned object offset")); pindex += lobject->backing_object_offset >> PAGE_SHIFT; VM_OBJECT_RLOCK(backing_object); if (!obj_locked || lobject != entry->object.vm_object) VM_OBJECT_RUNLOCK(lobject); lobject = backing_object; } if (m == NULL) { if (!obj_locked || lobject != entry->object.vm_object) VM_OBJECT_RUNLOCK(lobject); break; } if (m->valid == VM_PAGE_BITS_ALL && (m->flags & PG_FICTITIOUS) == 0) pmap_enter_quick(pmap, addr, m, entry->protection); if (!obj_locked || lobject != entry->object.vm_object) VM_OBJECT_RUNLOCK(lobject); } } /* * Hold each of the physical pages that are mapped by the specified range of * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid * and allow the specified types of access, "prot". If all of the implied * pages are successfully held, then the number of held pages is returned * together with pointers to those pages in the array "ma". However, if any * of the pages cannot be held, -1 is returned. */ int vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len, vm_prot_t prot, vm_page_t *ma, int max_count) { vm_offset_t end, va; vm_page_t *mp; int count; boolean_t pmap_failed; if (len == 0) return (0); end = round_page(addr + len); addr = trunc_page(addr); /* * Check for illegal addresses. */ if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map)) return (-1); if (atop(end - addr) > max_count) panic("vm_fault_quick_hold_pages: count > max_count"); count = atop(end - addr); /* * Most likely, the physical pages are resident in the pmap, so it is * faster to try pmap_extract_and_hold() first. */ pmap_failed = FALSE; for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) { *mp = pmap_extract_and_hold(map->pmap, va, prot); if (*mp == NULL) pmap_failed = TRUE; else if ((prot & VM_PROT_WRITE) != 0 && (*mp)->dirty != VM_PAGE_BITS_ALL) { /* * Explicitly dirty the physical page. Otherwise, the * caller's changes may go unnoticed because they are * performed through an unmanaged mapping or by a DMA * operation. * * The object lock is not held here. * See vm_page_clear_dirty_mask(). */ vm_page_dirty(*mp); } } if (pmap_failed) { /* * One or more pages could not be held by the pmap. Either no * page was mapped at the specified virtual address or that * mapping had insufficient permissions. Attempt to fault in * and hold these pages. * * If vm_fault_disable_pagefaults() was called, * i.e., TDP_NOFAULTING is set, we must not sleep nor * acquire MD VM locks, which means we must not call * vm_fault(). Some (out of tree) callers mark * too wide a code area with vm_fault_disable_pagefaults() * already, use the VM_PROT_QUICK_NOFAULT flag to request * the proper behaviour explicitly. */ if ((prot & VM_PROT_QUICK_NOFAULT) != 0 && (curthread->td_pflags & TDP_NOFAULTING) != 0) goto error; for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) if (*mp == NULL && vm_fault(map, va, prot, VM_FAULT_NORMAL, mp) != KERN_SUCCESS) goto error; } return (count); error: for (mp = ma; mp < ma + count; mp++) if (*mp != NULL) vm_page_unwire(*mp, PQ_INACTIVE); return (-1); } /* * Routine: * vm_fault_copy_entry * Function: * Create new shadow object backing dst_entry with private copy of * all underlying pages. When src_entry is equal to dst_entry, * function implements COW for wired-down map entry. Otherwise, * it forks wired entry into dst_map. * * In/out conditions: * The source and destination maps must be locked for write. * The source map entry must be wired down (or be a sharing map * entry corresponding to a main map entry that is wired down). */ void vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map, vm_map_entry_t dst_entry, vm_map_entry_t src_entry, vm_ooffset_t *fork_charge) { vm_object_t backing_object, dst_object, object, src_object; vm_pindex_t dst_pindex, pindex, src_pindex; vm_prot_t access, prot; vm_offset_t vaddr; vm_page_t dst_m; vm_page_t src_m; boolean_t upgrade; #ifdef lint src_map++; #endif /* lint */ upgrade = src_entry == dst_entry; access = prot = dst_entry->protection; src_object = src_entry->object.vm_object; src_pindex = OFF_TO_IDX(src_entry->offset); if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) { dst_object = src_object; vm_object_reference(dst_object); } else { /* * Create the top-level object for the destination entry. (Doesn't * actually shadow anything - we copy the pages directly.) */ dst_object = vm_object_allocate(OBJT_DEFAULT, atop(dst_entry->end - dst_entry->start)); #if VM_NRESERVLEVEL > 0 dst_object->flags |= OBJ_COLORED; dst_object->pg_color = atop(dst_entry->start); #endif dst_object->domain = src_object->domain; dst_object->charge = dst_entry->end - dst_entry->start; } VM_OBJECT_WLOCK(dst_object); KASSERT(upgrade || dst_entry->object.vm_object == NULL, ("vm_fault_copy_entry: vm_object not NULL")); if (src_object != dst_object) { dst_entry->object.vm_object = dst_object; dst_entry->offset = 0; dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC; } if (fork_charge != NULL) { KASSERT(dst_entry->cred == NULL, ("vm_fault_copy_entry: leaked swp charge")); dst_object->cred = curthread->td_ucred; crhold(dst_object->cred); *fork_charge += dst_object->charge; } else if ((dst_object->type == OBJT_DEFAULT || dst_object->type == OBJT_SWAP) && dst_object->cred == NULL) { KASSERT(dst_entry->cred != NULL, ("no cred for entry %p", dst_entry)); dst_object->cred = dst_entry->cred; dst_entry->cred = NULL; } /* * If not an upgrade, then enter the mappings in the pmap as * read and/or execute accesses. Otherwise, enter them as * write accesses. * * A writeable large page mapping is only created if all of * the constituent small page mappings are modified. Marking * PTEs as modified on inception allows promotion to happen * without taking potentially large number of soft faults. */ if (!upgrade) access &= ~VM_PROT_WRITE; /* * Loop through all of the virtual pages within the entry's * range, copying each page from the source object to the * destination object. Since the source is wired, those pages * must exist. In contrast, the destination is pageable. * Since the destination object doesn't share any backing storage * with the source object, all of its pages must be dirtied, * regardless of whether they can be written. */ for (vaddr = dst_entry->start, dst_pindex = 0; vaddr < dst_entry->end; vaddr += PAGE_SIZE, dst_pindex++) { again: /* * Find the page in the source object, and copy it in. * Because the source is wired down, the page will be * in memory. */ if (src_object != dst_object) VM_OBJECT_RLOCK(src_object); object = src_object; pindex = src_pindex + dst_pindex; while ((src_m = vm_page_lookup(object, pindex)) == NULL && (backing_object = object->backing_object) != NULL) { /* * Unless the source mapping is read-only or * it is presently being upgraded from * read-only, the first object in the shadow * chain should provide all of the pages. In * other words, this loop body should never be * executed when the source mapping is already * read/write. */ KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 || upgrade, ("vm_fault_copy_entry: main object missing page")); VM_OBJECT_RLOCK(backing_object); pindex += OFF_TO_IDX(object->backing_object_offset); if (object != dst_object) VM_OBJECT_RUNLOCK(object); object = backing_object; } KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing")); if (object != dst_object) { /* * Allocate a page in the destination object. */ dst_m = vm_page_alloc(dst_object, (src_object == dst_object ? src_pindex : 0) + dst_pindex, VM_ALLOC_NORMAL); if (dst_m == NULL) { VM_OBJECT_WUNLOCK(dst_object); VM_OBJECT_RUNLOCK(object); vm_wait(dst_object); VM_OBJECT_WLOCK(dst_object); goto again; } pmap_copy_page(src_m, dst_m); VM_OBJECT_RUNLOCK(object); dst_m->dirty = dst_m->valid = src_m->valid; } else { dst_m = src_m; - if (vm_page_sleep_if_busy(dst_m, "fltupg")) + if (vm_page_busy_acquire(dst_m, VM_ALLOC_WAITFAIL) == 0) goto again; - if (dst_m->pindex >= dst_object->size) + if (dst_m->pindex >= dst_object->size) { /* * We are upgrading. Index can occur * out of bounds if the object type is * vnode and the file was truncated. */ + vm_page_xunbusy(dst_m); break; - vm_page_xbusy(dst_m); + } } VM_OBJECT_WUNLOCK(dst_object); /* * Enter it in the pmap. If a wired, copy-on-write * mapping is being replaced by a write-enabled * mapping, then wire that new mapping. * * The page can be invalid if the user called * msync(MS_INVALIDATE) or truncated the backing vnode * or shared memory object. In this case, do not * insert it into pmap, but still do the copy so that * all copies of the wired map entry have similar * backing pages. */ if (dst_m->valid == VM_PAGE_BITS_ALL) { pmap_enter(dst_map->pmap, vaddr, dst_m, prot, access | (upgrade ? PMAP_ENTER_WIRED : 0), 0); } /* * Mark it no longer busy, and put it on the active list. */ VM_OBJECT_WLOCK(dst_object); if (upgrade) { if (src_m != dst_m) { vm_page_unwire(src_m, PQ_INACTIVE); vm_page_wire(dst_m); } else { KASSERT(vm_page_wired(dst_m), ("dst_m %p is not wired", dst_m)); } } else { vm_page_lock(dst_m); vm_page_activate(dst_m); vm_page_unlock(dst_m); } vm_page_xunbusy(dst_m); } VM_OBJECT_WUNLOCK(dst_object); if (upgrade) { dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY); vm_object_deallocate(src_object); } } /* * Block entry into the machine-independent layer's page fault handler by * the calling thread. Subsequent calls to vm_fault() by that thread will * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of * spurious page faults. */ int vm_fault_disable_pagefaults(void) { return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR)); } void vm_fault_enable_pagefaults(int save) { curthread_pflags_restore(save); } Index: head/sys/vm/vm_object.c =================================================================== --- head/sys/vm/vm_object.c (revision 353534) +++ head/sys/vm/vm_object.c (revision 353535) @@ -1,2591 +1,2612 @@ /*- * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU) * * Copyright (c) 1991, 1993 * The Regents of the University of California. All rights reserved. * * This code is derived from software contributed to Berkeley by * The Mach Operating System project at Carnegie-Mellon University. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94 * * * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Authors: Avadis Tevanian, Jr., Michael Wayne Young * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. */ /* * Virtual memory object module. */ #include __FBSDID("$FreeBSD$"); #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include /* for curproc, pageproc */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include static int old_msync; SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0, "Use old (insecure) msync behavior"); static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags, int flags, boolean_t *clearobjflags, boolean_t *eio); static boolean_t vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags); static void vm_object_qcollapse(vm_object_t object); static void vm_object_vndeallocate(vm_object_t object); /* * Virtual memory objects maintain the actual data * associated with allocated virtual memory. A given * page of memory exists within exactly one object. * * An object is only deallocated when all "references" * are given up. Only one "reference" to a given * region of an object should be writeable. * * Associated with each object is a list of all resident * memory pages belonging to that object; this list is * maintained by the "vm_page" module, and locked by the object's * lock. * * Each object also records a "pager" routine which is * used to retrieve (and store) pages to the proper backing * storage. In addition, objects may be backed by other * objects from which they were virtual-copied. * * The only items within the object structure which are * modified after time of creation are: * reference count locked by object's lock * pager routine locked by object's lock * */ struct object_q vm_object_list; struct mtx vm_object_list_mtx; /* lock for object list and count */ struct vm_object kernel_object_store; static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0, "VM object stats"); static counter_u64_t object_collapses = EARLY_COUNTER; SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD, &object_collapses, "VM object collapses"); static counter_u64_t object_bypasses = EARLY_COUNTER; SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD, &object_bypasses, "VM object bypasses"); static void counter_startup(void) { object_collapses = counter_u64_alloc(M_WAITOK); object_bypasses = counter_u64_alloc(M_WAITOK); } SYSINIT(object_counters, SI_SUB_CPU, SI_ORDER_ANY, counter_startup, NULL); static uma_zone_t obj_zone; static int vm_object_zinit(void *mem, int size, int flags); #ifdef INVARIANTS static void vm_object_zdtor(void *mem, int size, void *arg); static void vm_object_zdtor(void *mem, int size, void *arg) { vm_object_t object; object = (vm_object_t)mem; KASSERT(object->ref_count == 0, ("object %p ref_count = %d", object, object->ref_count)); KASSERT(TAILQ_EMPTY(&object->memq), ("object %p has resident pages in its memq", object)); KASSERT(vm_radix_is_empty(&object->rtree), ("object %p has resident pages in its trie", object)); #if VM_NRESERVLEVEL > 0 KASSERT(LIST_EMPTY(&object->rvq), ("object %p has reservations", object)); #endif KASSERT(REFCOUNT_COUNT(object->paging_in_progress) == 0, ("object %p paging_in_progress = %d", object, REFCOUNT_COUNT(object->paging_in_progress))); KASSERT(object->resident_page_count == 0, ("object %p resident_page_count = %d", object, object->resident_page_count)); KASSERT(object->shadow_count == 0, ("object %p shadow_count = %d", object, object->shadow_count)); KASSERT(object->type == OBJT_DEAD, ("object %p has non-dead type %d", object, object->type)); } #endif static int vm_object_zinit(void *mem, int size, int flags) { vm_object_t object; object = (vm_object_t)mem; rw_init_flags(&object->lock, "vm object", RW_DUPOK | RW_NEW); /* These are true for any object that has been freed */ object->type = OBJT_DEAD; object->ref_count = 0; vm_radix_init(&object->rtree); refcount_init(&object->paging_in_progress, 0); object->resident_page_count = 0; object->shadow_count = 0; object->flags = OBJ_DEAD; mtx_lock(&vm_object_list_mtx); TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); mtx_unlock(&vm_object_list_mtx); return (0); } static void _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object) { TAILQ_INIT(&object->memq); LIST_INIT(&object->shadow_head); object->type = type; if (type == OBJT_SWAP) pctrie_init(&object->un_pager.swp.swp_blks); /* * Ensure that swap_pager_swapoff() iteration over object_list * sees up to date type and pctrie head if it observed * non-dead object. */ atomic_thread_fence_rel(); switch (type) { case OBJT_DEAD: panic("_vm_object_allocate: can't create OBJT_DEAD"); case OBJT_DEFAULT: case OBJT_SWAP: object->flags = OBJ_ONEMAPPING; break; case OBJT_DEVICE: case OBJT_SG: object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED; break; case OBJT_MGTDEVICE: object->flags = OBJ_FICTITIOUS; break; case OBJT_PHYS: object->flags = OBJ_UNMANAGED; break; case OBJT_VNODE: object->flags = 0; break; default: panic("_vm_object_allocate: type %d is undefined", type); } object->size = size; object->domain.dr_policy = NULL; object->generation = 1; object->ref_count = 1; object->memattr = VM_MEMATTR_DEFAULT; object->cred = NULL; object->charge = 0; object->handle = NULL; object->backing_object = NULL; object->backing_object_offset = (vm_ooffset_t) 0; #if VM_NRESERVLEVEL > 0 LIST_INIT(&object->rvq); #endif umtx_shm_object_init(object); } /* * vm_object_init: * * Initialize the VM objects module. */ void vm_object_init(void) { TAILQ_INIT(&vm_object_list); mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); rw_init(&kernel_object->lock, "kernel vm object"); _vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), kernel_object); #if VM_NRESERVLEVEL > 0 kernel_object->flags |= OBJ_COLORED; kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); #endif /* * The lock portion of struct vm_object must be type stable due * to vm_pageout_fallback_object_lock locking a vm object * without holding any references to it. */ obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, #ifdef INVARIANTS vm_object_zdtor, #else NULL, #endif vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); vm_radix_zinit(); } void vm_object_clear_flag(vm_object_t object, u_short bits) { VM_OBJECT_ASSERT_WLOCKED(object); object->flags &= ~bits; } /* * Sets the default memory attribute for the specified object. Pages * that are allocated to this object are by default assigned this memory * attribute. * * Presently, this function must be called before any pages are allocated * to the object. In the future, this requirement may be relaxed for * "default" and "swap" objects. */ int vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr) { VM_OBJECT_ASSERT_WLOCKED(object); switch (object->type) { case OBJT_DEFAULT: case OBJT_DEVICE: case OBJT_MGTDEVICE: case OBJT_PHYS: case OBJT_SG: case OBJT_SWAP: case OBJT_VNODE: if (!TAILQ_EMPTY(&object->memq)) return (KERN_FAILURE); break; case OBJT_DEAD: return (KERN_INVALID_ARGUMENT); default: panic("vm_object_set_memattr: object %p is of undefined type", object); } object->memattr = memattr; return (KERN_SUCCESS); } void vm_object_pip_add(vm_object_t object, short i) { refcount_acquiren(&object->paging_in_progress, i); } void vm_object_pip_wakeup(vm_object_t object) { refcount_release(&object->paging_in_progress); } void vm_object_pip_wakeupn(vm_object_t object, short i) { refcount_releasen(&object->paging_in_progress, i); } void vm_object_pip_wait(vm_object_t object, char *waitid) { VM_OBJECT_ASSERT_WLOCKED(object); while (REFCOUNT_COUNT(object->paging_in_progress) > 0) { VM_OBJECT_WUNLOCK(object); refcount_wait(&object->paging_in_progress, waitid, PVM); VM_OBJECT_WLOCK(object); } } void vm_object_pip_wait_unlocked(vm_object_t object, char *waitid) { VM_OBJECT_ASSERT_UNLOCKED(object); while (REFCOUNT_COUNT(object->paging_in_progress) > 0) refcount_wait(&object->paging_in_progress, waitid, PVM); } /* * vm_object_allocate: * * Returns a new object with the given size. */ vm_object_t vm_object_allocate(objtype_t type, vm_pindex_t size) { vm_object_t object; object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK); _vm_object_allocate(type, size, object); return (object); } /* * vm_object_reference: * * Gets another reference to the given object. Note: OBJ_DEAD * objects can be referenced during final cleaning. */ void vm_object_reference(vm_object_t object) { if (object == NULL) return; VM_OBJECT_WLOCK(object); vm_object_reference_locked(object); VM_OBJECT_WUNLOCK(object); } /* * vm_object_reference_locked: * * Gets another reference to the given object. * * The object must be locked. */ void vm_object_reference_locked(vm_object_t object) { struct vnode *vp; VM_OBJECT_ASSERT_WLOCKED(object); object->ref_count++; if (object->type == OBJT_VNODE) { vp = object->handle; vref(vp); } } /* * Handle deallocating an object of type OBJT_VNODE. */ static void vm_object_vndeallocate(vm_object_t object) { struct vnode *vp = (struct vnode *) object->handle; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(object->type == OBJT_VNODE, ("vm_object_vndeallocate: not a vnode object")); KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); #ifdef INVARIANTS if (object->ref_count == 0) { vn_printf(vp, "vm_object_vndeallocate "); panic("vm_object_vndeallocate: bad object reference count"); } #endif if (!umtx_shm_vnobj_persistent && object->ref_count == 1) umtx_shm_object_terminated(object); object->ref_count--; /* vrele may need the vnode lock. */ VM_OBJECT_WUNLOCK(object); vrele(vp); } /* * vm_object_deallocate: * * Release a reference to the specified object, * gained either through a vm_object_allocate * or a vm_object_reference call. When all references * are gone, storage associated with this object * may be relinquished. * * No object may be locked. */ void vm_object_deallocate(vm_object_t object) { vm_object_t temp; while (object != NULL) { VM_OBJECT_WLOCK(object); if (object->type == OBJT_VNODE) { vm_object_vndeallocate(object); return; } KASSERT(object->ref_count != 0, ("vm_object_deallocate: object deallocated too many times: %d", object->type)); /* * If the reference count goes to 0 we start calling * vm_object_terminate() on the object chain. * A ref count of 1 may be a special case depending on the * shadow count being 0 or 1. */ object->ref_count--; if (object->ref_count > 1) { VM_OBJECT_WUNLOCK(object); return; } else if (object->ref_count == 1) { if (object->shadow_count == 0 && object->handle == NULL && (object->type == OBJT_DEFAULT || (object->type == OBJT_SWAP && (object->flags & OBJ_TMPFS_NODE) == 0))) { vm_object_set_flag(object, OBJ_ONEMAPPING); } else if ((object->shadow_count == 1) && (object->handle == NULL) && (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { vm_object_t robject; robject = LIST_FIRST(&object->shadow_head); KASSERT(robject != NULL, ("vm_object_deallocate: ref_count: %d, shadow_count: %d", object->ref_count, object->shadow_count)); KASSERT((robject->flags & OBJ_TMPFS_NODE) == 0, ("shadowed tmpfs v_object %p", object)); if (!VM_OBJECT_TRYWLOCK(robject)) { /* * Avoid a potential deadlock. */ object->ref_count++; VM_OBJECT_WUNLOCK(object); /* * More likely than not the thread * holding robject's lock has lower * priority than the current thread. * Let the lower priority thread run. */ pause("vmo_de", 1); continue; } /* * Collapse object into its shadow unless its * shadow is dead. In that case, object will * be deallocated by the thread that is * deallocating its shadow. */ if ((robject->flags & OBJ_DEAD) == 0 && (robject->handle == NULL) && (robject->type == OBJT_DEFAULT || robject->type == OBJT_SWAP)) { robject->ref_count++; retry: if (REFCOUNT_COUNT(robject->paging_in_progress) > 0) { VM_OBJECT_WUNLOCK(object); vm_object_pip_wait(robject, "objde1"); temp = robject->backing_object; if (object == temp) { VM_OBJECT_WLOCK(object); goto retry; } } else if (REFCOUNT_COUNT(object->paging_in_progress) > 0) { VM_OBJECT_WUNLOCK(robject); VM_OBJECT_WUNLOCK(object); refcount_wait( &object->paging_in_progress, "objde2", PVM); VM_OBJECT_WLOCK(robject); temp = robject->backing_object; if (object == temp) { VM_OBJECT_WLOCK(object); goto retry; } } else VM_OBJECT_WUNLOCK(object); if (robject->ref_count == 1) { robject->ref_count--; object = robject; goto doterm; } object = robject; vm_object_collapse(object); VM_OBJECT_WUNLOCK(object); continue; } VM_OBJECT_WUNLOCK(robject); } VM_OBJECT_WUNLOCK(object); return; } doterm: umtx_shm_object_terminated(object); temp = object->backing_object; if (temp != NULL) { KASSERT((object->flags & OBJ_TMPFS_NODE) == 0, ("shadowed tmpfs v_object 2 %p", object)); VM_OBJECT_WLOCK(temp); LIST_REMOVE(object, shadow_list); temp->shadow_count--; VM_OBJECT_WUNLOCK(temp); object->backing_object = NULL; } /* * Don't double-terminate, we could be in a termination * recursion due to the terminate having to sync data * to disk. */ if ((object->flags & OBJ_DEAD) == 0) { vm_object_set_flag(object, OBJ_DEAD); vm_object_terminate(object); } else VM_OBJECT_WUNLOCK(object); object = temp; } } /* * vm_object_destroy removes the object from the global object list * and frees the space for the object. */ void vm_object_destroy(vm_object_t object) { /* * Release the allocation charge. */ if (object->cred != NULL) { swap_release_by_cred(object->charge, object->cred); object->charge = 0; crfree(object->cred); object->cred = NULL; } /* * Free the space for the object. */ uma_zfree(obj_zone, object); } /* * vm_object_terminate_pages removes any remaining pageable pages * from the object and resets the object to an empty state. */ static void vm_object_terminate_pages(vm_object_t object) { vm_page_t p, p_next; VM_OBJECT_ASSERT_WLOCKED(object); /* * Free any remaining pageable pages. This also removes them from the * paging queues. However, don't free wired pages, just remove them * from the object. Rather than incrementally removing each page from * the object, the page and object are reset to any empty state. */ TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) { vm_page_assert_unbusied(p); KASSERT(p->object == object && (p->ref_count & VPRC_OBJREF) != 0, ("vm_object_terminate_pages: page %p is inconsistent", p)); p->object = NULL; if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) { VM_CNT_INC(v_pfree); vm_page_free(p); } } /* * If the object contained any pages, then reset it to an empty state. * None of the object's fields, including "resident_page_count", were * modified by the preceding loop. */ if (object->resident_page_count != 0) { vm_radix_reclaim_allnodes(&object->rtree); TAILQ_INIT(&object->memq); object->resident_page_count = 0; if (object->type == OBJT_VNODE) vdrop(object->handle); } } /* * vm_object_terminate actually destroys the specified object, freeing * up all previously used resources. * * The object must be locked. * This routine may block. */ void vm_object_terminate(vm_object_t object) { VM_OBJECT_ASSERT_WLOCKED(object); KASSERT((object->flags & OBJ_DEAD) != 0, ("terminating non-dead obj %p", object)); /* * wait for the pageout daemon to be done with the object */ vm_object_pip_wait(object, "objtrm"); KASSERT(!REFCOUNT_COUNT(object->paging_in_progress), ("vm_object_terminate: pageout in progress")); KASSERT(object->ref_count == 0, ("vm_object_terminate: object with references, ref_count=%d", object->ref_count)); if ((object->flags & OBJ_PG_DTOR) == 0) vm_object_terminate_pages(object); #if VM_NRESERVLEVEL > 0 if (__predict_false(!LIST_EMPTY(&object->rvq))) vm_reserv_break_all(object); #endif KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT || object->type == OBJT_SWAP, ("%s: non-swap obj %p has cred", __func__, object)); /* * Let the pager know object is dead. */ vm_pager_deallocate(object); VM_OBJECT_WUNLOCK(object); vm_object_destroy(object); } /* * Make the page read-only so that we can clear the object flags. However, if * this is a nosync mmap then the object is likely to stay dirty so do not * mess with the page and do not clear the object flags. Returns TRUE if the * page should be flushed, and FALSE otherwise. */ static boolean_t vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags) { /* * If we have been asked to skip nosync pages and this is a * nosync page, skip it. Note that the object flags were not * cleared in this case so we do not have to set them. */ if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) { *clearobjflags = FALSE; return (FALSE); } else { pmap_remove_write(p); return (p->dirty != 0); } } /* * vm_object_page_clean * * Clean all dirty pages in the specified range of object. Leaves page * on whatever queue it is currently on. If NOSYNC is set then do not * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC), * leaving the object dirty. * * When stuffing pages asynchronously, allow clustering. XXX we need a * synchronous clustering mode implementation. * * Odd semantics: if start == end, we clean everything. * * The object must be locked. * * Returns FALSE if some page from the range was not written, as * reported by the pager, and TRUE otherwise. */ boolean_t vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end, int flags) { vm_page_t np, p; vm_pindex_t pi, tend, tstart; int curgeneration, n, pagerflags; boolean_t clearobjflags, eio, res; VM_OBJECT_ASSERT_WLOCKED(object); /* * The OBJ_MIGHTBEDIRTY flag is only set for OBJT_VNODE * objects. The check below prevents the function from * operating on non-vnode objects. */ if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 || object->resident_page_count == 0) return (TRUE); pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0; tstart = OFF_TO_IDX(start); tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK); clearobjflags = tstart == 0 && tend >= object->size; res = TRUE; rescan: curgeneration = object->generation; for (p = vm_page_find_least(object, tstart); p != NULL; p = np) { pi = p->pindex; if (pi >= tend) break; np = TAILQ_NEXT(p, listq); if (p->valid == 0) continue; - if (vm_page_sleep_if_busy(p, "vpcwai")) { + if (vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL) == 0) { if (object->generation != curgeneration) { if ((flags & OBJPC_SYNC) != 0) goto rescan; else clearobjflags = FALSE; } np = vm_page_find_least(object, pi); continue; } - if (!vm_object_page_remove_write(p, flags, &clearobjflags)) + if (!vm_object_page_remove_write(p, flags, &clearobjflags)) { + vm_page_xunbusy(p); continue; + } n = vm_object_page_collect_flush(object, p, pagerflags, flags, &clearobjflags, &eio); if (eio) { res = FALSE; clearobjflags = FALSE; } if (object->generation != curgeneration) { if ((flags & OBJPC_SYNC) != 0) goto rescan; else clearobjflags = FALSE; } /* * If the VOP_PUTPAGES() did a truncated write, so * that even the first page of the run is not fully * written, vm_pageout_flush() returns 0 as the run * length. Since the condition that caused truncated * write may be permanent, e.g. exhausted free space, * accepting n == 0 would cause an infinite loop. * * Forwarding the iterator leaves the unwritten page * behind, but there is not much we can do there if * filesystem refuses to write it. */ if (n == 0) { n = 1; clearobjflags = FALSE; } np = vm_page_find_least(object, pi + n); } #if 0 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0); #endif if (clearobjflags) vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY); return (res); } static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags, int flags, boolean_t *clearobjflags, boolean_t *eio) { vm_page_t ma[vm_pageout_page_count], p_first, tp; int count, i, mreq, runlen; vm_page_lock_assert(p, MA_NOTOWNED); + vm_page_assert_xbusied(p); VM_OBJECT_ASSERT_WLOCKED(object); count = 1; mreq = 0; for (tp = p; count < vm_pageout_page_count; count++) { tp = vm_page_next(tp); - if (tp == NULL || vm_page_busied(tp)) + if (tp == NULL || vm_page_tryxbusy(tp) == 0) break; - if (!vm_object_page_remove_write(tp, flags, clearobjflags)) + if (!vm_object_page_remove_write(tp, flags, clearobjflags)) { + vm_page_xunbusy(tp); break; + } } for (p_first = p; count < vm_pageout_page_count; count++) { tp = vm_page_prev(p_first); - if (tp == NULL || vm_page_busied(tp)) + if (tp == NULL || vm_page_tryxbusy(tp) == 0) break; - if (!vm_object_page_remove_write(tp, flags, clearobjflags)) + if (!vm_object_page_remove_write(tp, flags, clearobjflags)) { + vm_page_xunbusy(tp); break; + } p_first = tp; mreq++; } for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++) ma[i] = tp; vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio); return (runlen); } /* * Note that there is absolutely no sense in writing out * anonymous objects, so we track down the vnode object * to write out. * We invalidate (remove) all pages from the address space * for semantic correctness. * * If the backing object is a device object with unmanaged pages, then any * mappings to the specified range of pages must be removed before this * function is called. * * Note: certain anonymous maps, such as MAP_NOSYNC maps, * may start out with a NULL object. */ boolean_t vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size, boolean_t syncio, boolean_t invalidate) { vm_object_t backing_object; struct vnode *vp; struct mount *mp; int error, flags, fsync_after; boolean_t res; if (object == NULL) return (TRUE); res = TRUE; error = 0; VM_OBJECT_WLOCK(object); while ((backing_object = object->backing_object) != NULL) { VM_OBJECT_WLOCK(backing_object); offset += object->backing_object_offset; VM_OBJECT_WUNLOCK(object); object = backing_object; if (object->size < OFF_TO_IDX(offset + size)) size = IDX_TO_OFF(object->size) - offset; } /* * Flush pages if writing is allowed, invalidate them * if invalidation requested. Pages undergoing I/O * will be ignored by vm_object_page_remove(). * * We cannot lock the vnode and then wait for paging * to complete without deadlocking against vm_fault. * Instead we simply call vm_object_page_remove() and * allow it to block internally on a page-by-page * basis when it encounters pages undergoing async * I/O. */ if (object->type == OBJT_VNODE && (object->flags & OBJ_MIGHTBEDIRTY) != 0 && ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) { VM_OBJECT_WUNLOCK(object); (void) vn_start_write(vp, &mp, V_WAIT); vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); if (syncio && !invalidate && offset == 0 && atop(size) == object->size) { /* * If syncing the whole mapping of the file, * it is faster to schedule all the writes in * async mode, also allowing the clustering, * and then wait for i/o to complete. */ flags = 0; fsync_after = TRUE; } else { flags = (syncio || invalidate) ? OBJPC_SYNC : 0; flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0; fsync_after = FALSE; } VM_OBJECT_WLOCK(object); res = vm_object_page_clean(object, offset, offset + size, flags); VM_OBJECT_WUNLOCK(object); if (fsync_after) error = VOP_FSYNC(vp, MNT_WAIT, curthread); VOP_UNLOCK(vp, 0); vn_finished_write(mp); if (error != 0) res = FALSE; VM_OBJECT_WLOCK(object); } if ((object->type == OBJT_VNODE || object->type == OBJT_DEVICE) && invalidate) { if (object->type == OBJT_DEVICE) /* * The option OBJPR_NOTMAPPED must be passed here * because vm_object_page_remove() cannot remove * unmanaged mappings. */ flags = OBJPR_NOTMAPPED; else if (old_msync) flags = 0; else flags = OBJPR_CLEANONLY; vm_object_page_remove(object, OFF_TO_IDX(offset), OFF_TO_IDX(offset + size + PAGE_MASK), flags); } VM_OBJECT_WUNLOCK(object); return (res); } /* * Determine whether the given advice can be applied to the object. Advice is * not applied to unmanaged pages since they never belong to page queues, and * since MADV_FREE is destructive, it can apply only to anonymous pages that * have been mapped at most once. */ static bool vm_object_advice_applies(vm_object_t object, int advice) { if ((object->flags & OBJ_UNMANAGED) != 0) return (false); if (advice != MADV_FREE) return (true); return ((object->type == OBJT_DEFAULT || object->type == OBJT_SWAP) && (object->flags & OBJ_ONEMAPPING) != 0); } static void vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex, vm_size_t size) { if (advice == MADV_FREE && object->type == OBJT_SWAP) swap_pager_freespace(object, pindex, size); } /* * vm_object_madvise: * * Implements the madvise function at the object/page level. * * MADV_WILLNEED (any object) * * Activate the specified pages if they are resident. * * MADV_DONTNEED (any object) * * Deactivate the specified pages if they are resident. * * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, * OBJ_ONEMAPPING only) * * Deactivate and clean the specified pages if they are * resident. This permits the process to reuse the pages * without faulting or the kernel to reclaim the pages * without I/O. */ void vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end, int advice) { vm_pindex_t tpindex; vm_object_t backing_object, tobject; vm_page_t m, tm; if (object == NULL) return; relookup: VM_OBJECT_WLOCK(object); if (!vm_object_advice_applies(object, advice)) { VM_OBJECT_WUNLOCK(object); return; } for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) { tobject = object; /* * If the next page isn't resident in the top-level object, we * need to search the shadow chain. When applying MADV_FREE, we * take care to release any swap space used to store * non-resident pages. */ if (m == NULL || pindex < m->pindex) { /* * Optimize a common case: if the top-level object has * no backing object, we can skip over the non-resident * range in constant time. */ if (object->backing_object == NULL) { tpindex = (m != NULL && m->pindex < end) ? m->pindex : end; vm_object_madvise_freespace(object, advice, pindex, tpindex - pindex); if ((pindex = tpindex) == end) break; goto next_page; } tpindex = pindex; do { vm_object_madvise_freespace(tobject, advice, tpindex, 1); /* * Prepare to search the next object in the * chain. */ backing_object = tobject->backing_object; if (backing_object == NULL) goto next_pindex; VM_OBJECT_WLOCK(backing_object); tpindex += OFF_TO_IDX(tobject->backing_object_offset); if (tobject != object) VM_OBJECT_WUNLOCK(tobject); tobject = backing_object; if (!vm_object_advice_applies(tobject, advice)) goto next_pindex; } while ((tm = vm_page_lookup(tobject, tpindex)) == NULL); } else { next_page: tm = m; m = TAILQ_NEXT(m, listq); } /* * If the page is not in a normal state, skip it. */ if (tm->valid != VM_PAGE_BITS_ALL || vm_page_wired(tm)) goto next_pindex; KASSERT((tm->flags & PG_FICTITIOUS) == 0, ("vm_object_madvise: page %p is fictitious", tm)); KASSERT((tm->oflags & VPO_UNMANAGED) == 0, ("vm_object_madvise: page %p is not managed", tm)); - if (vm_page_busied(tm)) { + if (vm_page_tryxbusy(tm) == 0) { if (object != tobject) VM_OBJECT_WUNLOCK(object); if (advice == MADV_WILLNEED) { /* * Reference the page before unlocking and * sleeping so that the page daemon is less * likely to reclaim it. */ vm_page_aflag_set(tm, PGA_REFERENCED); } vm_page_busy_sleep(tm, "madvpo", false); goto relookup; } vm_page_lock(tm); vm_page_advise(tm, advice); vm_page_unlock(tm); + vm_page_xunbusy(tm); vm_object_madvise_freespace(tobject, advice, tm->pindex, 1); next_pindex: if (tobject != object) VM_OBJECT_WUNLOCK(tobject); } VM_OBJECT_WUNLOCK(object); } /* * vm_object_shadow: * * Create a new object which is backed by the * specified existing object range. The source * object reference is deallocated. * * The new object and offset into that object * are returned in the source parameters. */ void vm_object_shadow( vm_object_t *object, /* IN/OUT */ vm_ooffset_t *offset, /* IN/OUT */ vm_size_t length) { vm_object_t source; vm_object_t result; source = *object; /* * Don't create the new object if the old object isn't shared. */ if (source != NULL) { VM_OBJECT_WLOCK(source); if (source->ref_count == 1 && source->handle == NULL && (source->type == OBJT_DEFAULT || source->type == OBJT_SWAP)) { VM_OBJECT_WUNLOCK(source); return; } VM_OBJECT_WUNLOCK(source); } /* * Allocate a new object with the given length. */ result = vm_object_allocate(OBJT_DEFAULT, atop(length)); /* * The new object shadows the source object, adding a reference to it. * Our caller changes his reference to point to the new object, * removing a reference to the source object. Net result: no change * of reference count. * * Try to optimize the result object's page color when shadowing * in order to maintain page coloring consistency in the combined * shadowed object. */ result->backing_object = source; /* * Store the offset into the source object, and fix up the offset into * the new object. */ result->backing_object_offset = *offset; if (source != NULL) { VM_OBJECT_WLOCK(source); result->domain = source->domain; LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list); source->shadow_count++; #if VM_NRESERVLEVEL > 0 result->flags |= source->flags & OBJ_COLORED; result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER - 1)) - 1); #endif VM_OBJECT_WUNLOCK(source); } /* * Return the new things */ *offset = 0; *object = result; } /* * vm_object_split: * * Split the pages in a map entry into a new object. This affords * easier removal of unused pages, and keeps object inheritance from * being a negative impact on memory usage. */ void vm_object_split(vm_map_entry_t entry) { vm_page_t m, m_next; vm_object_t orig_object, new_object, source; vm_pindex_t idx, offidxstart; vm_size_t size; orig_object = entry->object.vm_object; if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP) return; if (orig_object->ref_count <= 1) return; VM_OBJECT_WUNLOCK(orig_object); offidxstart = OFF_TO_IDX(entry->offset); size = atop(entry->end - entry->start); /* * If swap_pager_copy() is later called, it will convert new_object * into a swap object. */ new_object = vm_object_allocate(OBJT_DEFAULT, size); /* * At this point, the new object is still private, so the order in * which the original and new objects are locked does not matter. */ VM_OBJECT_WLOCK(new_object); VM_OBJECT_WLOCK(orig_object); new_object->domain = orig_object->domain; source = orig_object->backing_object; if (source != NULL) { VM_OBJECT_WLOCK(source); if ((source->flags & OBJ_DEAD) != 0) { VM_OBJECT_WUNLOCK(source); VM_OBJECT_WUNLOCK(orig_object); VM_OBJECT_WUNLOCK(new_object); vm_object_deallocate(new_object); VM_OBJECT_WLOCK(orig_object); return; } LIST_INSERT_HEAD(&source->shadow_head, new_object, shadow_list); source->shadow_count++; vm_object_reference_locked(source); /* for new_object */ vm_object_clear_flag(source, OBJ_ONEMAPPING); VM_OBJECT_WUNLOCK(source); new_object->backing_object_offset = orig_object->backing_object_offset + entry->offset; new_object->backing_object = source; } if (orig_object->cred != NULL) { new_object->cred = orig_object->cred; crhold(orig_object->cred); new_object->charge = ptoa(size); KASSERT(orig_object->charge >= ptoa(size), ("orig_object->charge < 0")); orig_object->charge -= ptoa(size); } retry: m = vm_page_find_least(orig_object, offidxstart); for (; m != NULL && (idx = m->pindex - offidxstart) < size; m = m_next) { m_next = TAILQ_NEXT(m, listq); /* * We must wait for pending I/O to complete before we can * rename the page. * * We do not have to VM_PROT_NONE the page as mappings should * not be changed by this operation. */ - if (vm_page_busied(m)) { + if (vm_page_tryxbusy(m) == 0) { VM_OBJECT_WUNLOCK(new_object); vm_page_sleep_if_busy(m, "spltwt"); VM_OBJECT_WLOCK(new_object); goto retry; } /* vm_page_rename() will dirty the page. */ if (vm_page_rename(m, new_object, idx)) { + vm_page_xunbusy(m); VM_OBJECT_WUNLOCK(new_object); VM_OBJECT_WUNLOCK(orig_object); vm_radix_wait(); VM_OBJECT_WLOCK(orig_object); VM_OBJECT_WLOCK(new_object); goto retry; } + /* Rename released the xbusy lock. */ + #if VM_NRESERVLEVEL > 0 /* * If some of the reservation's allocated pages remain with * the original object, then transferring the reservation to * the new object is neither particularly beneficial nor * particularly harmful as compared to leaving the reservation * with the original object. If, however, all of the * reservation's allocated pages are transferred to the new * object, then transferring the reservation is typically * beneficial. Determining which of these two cases applies * would be more costly than unconditionally renaming the * reservation. */ vm_reserv_rename(m, new_object, orig_object, offidxstart); #endif if (orig_object->type == OBJT_SWAP) vm_page_xbusy(m); } if (orig_object->type == OBJT_SWAP) { /* * swap_pager_copy() can sleep, in which case the orig_object's * and new_object's locks are released and reacquired. */ swap_pager_copy(orig_object, new_object, offidxstart, 0); TAILQ_FOREACH(m, &new_object->memq, listq) vm_page_xunbusy(m); } VM_OBJECT_WUNLOCK(orig_object); VM_OBJECT_WUNLOCK(new_object); entry->object.vm_object = new_object; entry->offset = 0LL; vm_object_deallocate(orig_object); VM_OBJECT_WLOCK(new_object); } #define OBSC_COLLAPSE_NOWAIT 0x0002 #define OBSC_COLLAPSE_WAIT 0x0004 static vm_page_t vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p, vm_page_t next, int op) { vm_object_t backing_object; VM_OBJECT_ASSERT_WLOCKED(object); backing_object = object->backing_object; VM_OBJECT_ASSERT_WLOCKED(backing_object); - KASSERT(p == NULL || vm_page_busied(p), ("unbusy page %p", p)); KASSERT(p == NULL || p->object == object || p->object == backing_object, ("invalid ownership %p %p %p", p, object, backing_object)); if ((op & OBSC_COLLAPSE_NOWAIT) != 0) return (next); /* The page is only NULL when rename fails. */ if (p == NULL) { vm_radix_wait(); } else { if (p->object == object) VM_OBJECT_WUNLOCK(backing_object); else VM_OBJECT_WUNLOCK(object); vm_page_busy_sleep(p, "vmocol", false); } VM_OBJECT_WLOCK(object); VM_OBJECT_WLOCK(backing_object); return (TAILQ_FIRST(&backing_object->memq)); } static bool vm_object_scan_all_shadowed(vm_object_t object) { vm_object_t backing_object; vm_page_t p, pp; vm_pindex_t backing_offset_index, new_pindex, pi, ps; VM_OBJECT_ASSERT_WLOCKED(object); VM_OBJECT_ASSERT_WLOCKED(object->backing_object); backing_object = object->backing_object; if (backing_object->type != OBJT_DEFAULT && backing_object->type != OBJT_SWAP) return (false); pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset); p = vm_page_find_least(backing_object, pi); ps = swap_pager_find_least(backing_object, pi); /* * Only check pages inside the parent object's range and * inside the parent object's mapping of the backing object. */ for (;; pi++) { if (p != NULL && p->pindex < pi) p = TAILQ_NEXT(p, listq); if (ps < pi) ps = swap_pager_find_least(backing_object, pi); if (p == NULL && ps >= backing_object->size) break; else if (p == NULL) pi = ps; else pi = MIN(p->pindex, ps); new_pindex = pi - backing_offset_index; if (new_pindex >= object->size) break; /* * See if the parent has the page or if the parent's object * pager has the page. If the parent has the page but the page * is not valid, the parent's object pager must have the page. * * If this fails, the parent does not completely shadow the * object and we might as well give up now. */ pp = vm_page_lookup(object, new_pindex); if ((pp == NULL || pp->valid == 0) && !vm_pager_has_page(object, new_pindex, NULL, NULL)) return (false); } return (true); } static bool vm_object_collapse_scan(vm_object_t object, int op) { vm_object_t backing_object; vm_page_t next, p, pp; vm_pindex_t backing_offset_index, new_pindex; VM_OBJECT_ASSERT_WLOCKED(object); VM_OBJECT_ASSERT_WLOCKED(object->backing_object); backing_object = object->backing_object; backing_offset_index = OFF_TO_IDX(object->backing_object_offset); /* * Initial conditions */ if ((op & OBSC_COLLAPSE_WAIT) != 0) vm_object_set_flag(backing_object, OBJ_DEAD); /* * Our scan */ for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) { next = TAILQ_NEXT(p, listq); new_pindex = p->pindex - backing_offset_index; /* * Check for busy page */ - if (vm_page_busied(p)) { + if (vm_page_tryxbusy(p) == 0) { next = vm_object_collapse_scan_wait(object, p, next, op); continue; } KASSERT(p->object == backing_object, ("vm_object_collapse_scan: object mismatch")); if (p->pindex < backing_offset_index || new_pindex >= object->size) { if (backing_object->type == OBJT_SWAP) swap_pager_freespace(backing_object, p->pindex, 1); KASSERT(!pmap_page_is_mapped(p), ("freeing mapped page %p", p)); if (vm_page_remove(p)) vm_page_free(p); continue; } pp = vm_page_lookup(object, new_pindex); - if (pp != NULL && vm_page_busied(pp)) { + if (pp != NULL && vm_page_tryxbusy(pp) == 0) { + vm_page_xunbusy(p); /* * The page in the parent is busy and possibly not * (yet) valid. Until its state is finalized by the * busy bit owner, we can't tell whether it shadows the * original page. Therefore, we must either skip it * and the original (backing_object) page or wait for * its state to be finalized. * * This is due to a race with vm_fault() where we must * unbusy the original (backing_obj) page before we can * (re)lock the parent. Hence we can get here. */ next = vm_object_collapse_scan_wait(object, pp, next, op); continue; } KASSERT(pp == NULL || pp->valid != 0, ("unbusy invalid page %p", pp)); if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL, NULL)) { /* * The page already exists in the parent OR swap exists * for this location in the parent. Leave the parent's * page alone. Destroy the original page from the * backing object. */ if (backing_object->type == OBJT_SWAP) swap_pager_freespace(backing_object, p->pindex, 1); KASSERT(!pmap_page_is_mapped(p), ("freeing mapped page %p", p)); if (vm_page_remove(p)) vm_page_free(p); + if (pp != NULL) + vm_page_xunbusy(pp); continue; } /* * Page does not exist in parent, rename the page from the * backing object to the main object. * * If the page was mapped to a process, it can remain mapped * through the rename. vm_page_rename() will dirty the page. */ if (vm_page_rename(p, object, new_pindex)) { + vm_page_xunbusy(p); + if (pp != NULL) + vm_page_xunbusy(pp); next = vm_object_collapse_scan_wait(object, NULL, next, op); continue; } + /* Rename released the xbusy lock. */ /* Use the old pindex to free the right page. */ if (backing_object->type == OBJT_SWAP) swap_pager_freespace(backing_object, new_pindex + backing_offset_index, 1); #if VM_NRESERVLEVEL > 0 /* * Rename the reservation. */ vm_reserv_rename(p, object, backing_object, backing_offset_index); #endif } return (true); } /* * this version of collapse allows the operation to occur earlier and * when paging_in_progress is true for an object... This is not a complete * operation, but should plug 99.9% of the rest of the leaks. */ static void vm_object_qcollapse(vm_object_t object) { vm_object_t backing_object = object->backing_object; VM_OBJECT_ASSERT_WLOCKED(object); VM_OBJECT_ASSERT_WLOCKED(backing_object); if (backing_object->ref_count != 1) return; vm_object_collapse_scan(object, OBSC_COLLAPSE_NOWAIT); } /* * vm_object_collapse: * * Collapse an object with the object backing it. * Pages in the backing object are moved into the * parent, and the backing object is deallocated. */ void vm_object_collapse(vm_object_t object) { vm_object_t backing_object, new_backing_object; VM_OBJECT_ASSERT_WLOCKED(object); while (TRUE) { /* * Verify that the conditions are right for collapse: * * The object exists and the backing object exists. */ if ((backing_object = object->backing_object) == NULL) break; /* * we check the backing object first, because it is most likely * not collapsable. */ VM_OBJECT_WLOCK(backing_object); if (backing_object->handle != NULL || (backing_object->type != OBJT_DEFAULT && backing_object->type != OBJT_SWAP) || (backing_object->flags & (OBJ_DEAD | OBJ_NOSPLIT)) != 0 || object->handle != NULL || (object->type != OBJT_DEFAULT && object->type != OBJT_SWAP) || (object->flags & OBJ_DEAD)) { VM_OBJECT_WUNLOCK(backing_object); break; } if (REFCOUNT_COUNT(object->paging_in_progress) > 0 || REFCOUNT_COUNT(backing_object->paging_in_progress) > 0) { vm_object_qcollapse(object); VM_OBJECT_WUNLOCK(backing_object); break; } /* * We know that we can either collapse the backing object (if * the parent is the only reference to it) or (perhaps) have * the parent bypass the object if the parent happens to shadow * all the resident pages in the entire backing object. * * This is ignoring pager-backed pages such as swap pages. * vm_object_collapse_scan fails the shadowing test in this * case. */ if (backing_object->ref_count == 1) { vm_object_pip_add(object, 1); vm_object_pip_add(backing_object, 1); /* * If there is exactly one reference to the backing * object, we can collapse it into the parent. */ vm_object_collapse_scan(object, OBSC_COLLAPSE_WAIT); #if VM_NRESERVLEVEL > 0 /* * Break any reservations from backing_object. */ if (__predict_false(!LIST_EMPTY(&backing_object->rvq))) vm_reserv_break_all(backing_object); #endif /* * Move the pager from backing_object to object. */ if (backing_object->type == OBJT_SWAP) { /* * swap_pager_copy() can sleep, in which case * the backing_object's and object's locks are * released and reacquired. * Since swap_pager_copy() is being asked to * destroy the source, it will change the * backing_object's type to OBJT_DEFAULT. */ swap_pager_copy( backing_object, object, OFF_TO_IDX(object->backing_object_offset), TRUE); } /* * Object now shadows whatever backing_object did. * Note that the reference to * backing_object->backing_object moves from within * backing_object to within object. */ LIST_REMOVE(object, shadow_list); backing_object->shadow_count--; if (backing_object->backing_object) { VM_OBJECT_WLOCK(backing_object->backing_object); LIST_REMOVE(backing_object, shadow_list); LIST_INSERT_HEAD( &backing_object->backing_object->shadow_head, object, shadow_list); /* * The shadow_count has not changed. */ VM_OBJECT_WUNLOCK(backing_object->backing_object); } object->backing_object = backing_object->backing_object; object->backing_object_offset += backing_object->backing_object_offset; /* * Discard backing_object. * * Since the backing object has no pages, no pager left, * and no object references within it, all that is * necessary is to dispose of it. */ KASSERT(backing_object->ref_count == 1, ( "backing_object %p was somehow re-referenced during collapse!", backing_object)); vm_object_pip_wakeup(backing_object); backing_object->type = OBJT_DEAD; backing_object->ref_count = 0; VM_OBJECT_WUNLOCK(backing_object); vm_object_destroy(backing_object); vm_object_pip_wakeup(object); counter_u64_add(object_collapses, 1); } else { /* * If we do not entirely shadow the backing object, * there is nothing we can do so we give up. */ if (object->resident_page_count != object->size && !vm_object_scan_all_shadowed(object)) { VM_OBJECT_WUNLOCK(backing_object); break; } /* * Make the parent shadow the next object in the * chain. Deallocating backing_object will not remove * it, since its reference count is at least 2. */ LIST_REMOVE(object, shadow_list); backing_object->shadow_count--; new_backing_object = backing_object->backing_object; if ((object->backing_object = new_backing_object) != NULL) { VM_OBJECT_WLOCK(new_backing_object); LIST_INSERT_HEAD( &new_backing_object->shadow_head, object, shadow_list ); new_backing_object->shadow_count++; vm_object_reference_locked(new_backing_object); VM_OBJECT_WUNLOCK(new_backing_object); object->backing_object_offset += backing_object->backing_object_offset; } /* * Drop the reference count on backing_object. Since * its ref_count was at least 2, it will not vanish. */ backing_object->ref_count--; VM_OBJECT_WUNLOCK(backing_object); counter_u64_add(object_bypasses, 1); } /* * Try again with this object's new backing object. */ } } /* * vm_object_page_remove: * * For the given object, either frees or invalidates each of the * specified pages. In general, a page is freed. However, if a page is * wired for any reason other than the existence of a managed, wired * mapping, then it may be invalidated but not removed from the object. * Pages are specified by the given range ["start", "end") and the option * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range * extends from "start" to the end of the object. If the option * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the * specified range are affected. If the option OBJPR_NOTMAPPED is * specified, then the pages within the specified range must have no * mappings. Otherwise, if this option is not specified, any mappings to * the specified pages are removed before the pages are freed or * invalidated. * * In general, this operation should only be performed on objects that * contain managed pages. There are, however, two exceptions. First, it * is performed on the kernel and kmem objects by vm_map_entry_delete(). * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device- * backed pages. In both of these cases, the option OBJPR_CLEANONLY must * not be specified and the option OBJPR_NOTMAPPED must be specified. * * The object must be locked. */ void vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int options) { vm_page_t p, next; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT((object->flags & OBJ_UNMANAGED) == 0 || (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED, ("vm_object_page_remove: illegal options for object %p", object)); if (object->resident_page_count == 0) return; vm_object_pip_add(object, 1); again: p = vm_page_find_least(object, start); /* * Here, the variable "p" is either (1) the page with the least pindex * greater than or equal to the parameter "start" or (2) NULL. */ for (; p != NULL && (p->pindex < end || end == 0); p = next) { next = TAILQ_NEXT(p, listq); /* * If the page is wired for any reason besides the existence * of managed, wired mappings, then it cannot be freed. For * example, fictitious pages, which represent device memory, * are inherently wired and cannot be freed. They can, * however, be invalidated if the option OBJPR_CLEANONLY is * not specified. */ - if (vm_page_busied(p)) { + if (vm_page_tryxbusy(p) == 0) { vm_page_sleep_if_busy(p, "vmopar"); goto again; } if (vm_page_wired(p)) { wired: if ((options & OBJPR_NOTMAPPED) == 0 && object->ref_count != 0) pmap_remove_all(p); if ((options & OBJPR_CLEANONLY) == 0) { p->valid = 0; vm_page_undirty(p); } + vm_page_xunbusy(p); continue; } KASSERT((p->flags & PG_FICTITIOUS) == 0, ("vm_object_page_remove: page %p is fictitious", p)); if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) { if ((options & OBJPR_NOTMAPPED) == 0 && object->ref_count != 0 && !vm_page_try_remove_write(p)) goto wired; - if (p->dirty != 0) + if (p->dirty != 0) { + vm_page_xunbusy(p); continue; + } } if ((options & OBJPR_NOTMAPPED) == 0 && object->ref_count != 0 && !vm_page_try_remove_all(p)) goto wired; vm_page_free(p); } vm_object_pip_wakeup(object); } /* * vm_object_page_noreuse: * * For the given object, attempt to move the specified pages to * the head of the inactive queue. This bypasses regular LRU * operation and allows the pages to be reused quickly under memory * pressure. If a page is wired for any reason, then it will not * be queued. Pages are specified by the range ["start", "end"). * As a special case, if "end" is zero, then the range extends from * "start" to the end of the object. * * This operation should only be performed on objects that * contain non-fictitious, managed pages. * * The object must be locked. */ void vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end) { struct mtx *mtx; vm_page_t p, next; VM_OBJECT_ASSERT_LOCKED(object); KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0, ("vm_object_page_noreuse: illegal object %p", object)); if (object->resident_page_count == 0) return; p = vm_page_find_least(object, start); /* * Here, the variable "p" is either (1) the page with the least pindex * greater than or equal to the parameter "start" or (2) NULL. */ mtx = NULL; for (; p != NULL && (p->pindex < end || end == 0); p = next) { next = TAILQ_NEXT(p, listq); vm_page_change_lock(p, &mtx); vm_page_deactivate_noreuse(p); } if (mtx != NULL) mtx_unlock(mtx); } /* * Populate the specified range of the object with valid pages. Returns * TRUE if the range is successfully populated and FALSE otherwise. * * Note: This function should be optimized to pass a larger array of * pages to vm_pager_get_pages() before it is applied to a non- * OBJT_DEVICE object. * * The object must be locked. */ boolean_t vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end) { vm_page_t m; vm_pindex_t pindex; int rv; VM_OBJECT_ASSERT_WLOCKED(object); for (pindex = start; pindex < end; pindex++) { rv = vm_page_grab_valid(&m, object, pindex, VM_ALLOC_NORMAL); if (rv != VM_PAGER_OK) break; /* * Keep "m" busy because a subsequent iteration may unlock * the object. */ } if (pindex > start) { m = vm_page_lookup(object, start); while (m != NULL && m->pindex < pindex) { vm_page_xunbusy(m); m = TAILQ_NEXT(m, listq); } } return (pindex == end); } /* * Routine: vm_object_coalesce * Function: Coalesces two objects backing up adjoining * regions of memory into a single object. * * returns TRUE if objects were combined. * * NOTE: Only works at the moment if the second object is NULL - * if it's not, which object do we lock first? * * Parameters: * prev_object First object to coalesce * prev_offset Offset into prev_object * prev_size Size of reference to prev_object * next_size Size of reference to the second object * reserved Indicator that extension region has * swap accounted for * * Conditions: * The object must *not* be locked. */ boolean_t vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset, vm_size_t prev_size, vm_size_t next_size, boolean_t reserved) { vm_pindex_t next_pindex; if (prev_object == NULL) return (TRUE); VM_OBJECT_WLOCK(prev_object); if ((prev_object->type != OBJT_DEFAULT && prev_object->type != OBJT_SWAP) || (prev_object->flags & OBJ_NOSPLIT) != 0) { VM_OBJECT_WUNLOCK(prev_object); return (FALSE); } /* * Try to collapse the object first */ vm_object_collapse(prev_object); /* * Can't coalesce if: . more than one reference . paged out . shadows * another object . has a copy elsewhere (any of which mean that the * pages not mapped to prev_entry may be in use anyway) */ if (prev_object->backing_object != NULL) { VM_OBJECT_WUNLOCK(prev_object); return (FALSE); } prev_size >>= PAGE_SHIFT; next_size >>= PAGE_SHIFT; next_pindex = OFF_TO_IDX(prev_offset) + prev_size; if (prev_object->ref_count > 1 && prev_object->size != next_pindex && (prev_object->flags & OBJ_ONEMAPPING) == 0) { VM_OBJECT_WUNLOCK(prev_object); return (FALSE); } /* * Account for the charge. */ if (prev_object->cred != NULL) { /* * If prev_object was charged, then this mapping, * although not charged now, may become writable * later. Non-NULL cred in the object would prevent * swap reservation during enabling of the write * access, so reserve swap now. Failed reservation * cause allocation of the separate object for the map * entry, and swap reservation for this entry is * managed in appropriate time. */ if (!reserved && !swap_reserve_by_cred(ptoa(next_size), prev_object->cred)) { VM_OBJECT_WUNLOCK(prev_object); return (FALSE); } prev_object->charge += ptoa(next_size); } /* * Remove any pages that may still be in the object from a previous * deallocation. */ if (next_pindex < prev_object->size) { vm_object_page_remove(prev_object, next_pindex, next_pindex + next_size, 0); if (prev_object->type == OBJT_SWAP) swap_pager_freespace(prev_object, next_pindex, next_size); #if 0 if (prev_object->cred != NULL) { KASSERT(prev_object->charge >= ptoa(prev_object->size - next_pindex), ("object %p overcharged 1 %jx %jx", prev_object, (uintmax_t)next_pindex, (uintmax_t)next_size)); prev_object->charge -= ptoa(prev_object->size - next_pindex); } #endif } /* * Extend the object if necessary. */ if (next_pindex + next_size > prev_object->size) prev_object->size = next_pindex + next_size; VM_OBJECT_WUNLOCK(prev_object); return (TRUE); } void vm_object_set_writeable_dirty(vm_object_t object) { VM_OBJECT_ASSERT_WLOCKED(object); if (object->type != OBJT_VNODE) { if ((object->flags & OBJ_TMPFS_NODE) != 0) { KASSERT(object->type == OBJT_SWAP, ("non-swap tmpfs")); vm_object_set_flag(object, OBJ_TMPFS_DIRTY); } return; } object->generation++; if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) return; vm_object_set_flag(object, OBJ_MIGHTBEDIRTY); } /* * vm_object_unwire: * * For each page offset within the specified range of the given object, * find the highest-level page in the shadow chain and unwire it. A page * must exist at every page offset, and the highest-level page must be * wired. */ void vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length, uint8_t queue) { vm_object_t tobject, t1object; vm_page_t m, tm; vm_pindex_t end_pindex, pindex, tpindex; int depth, locked_depth; KASSERT((offset & PAGE_MASK) == 0, ("vm_object_unwire: offset is not page aligned")); KASSERT((length & PAGE_MASK) == 0, ("vm_object_unwire: length is not a multiple of PAGE_SIZE")); /* The wired count of a fictitious page never changes. */ if ((object->flags & OBJ_FICTITIOUS) != 0) return; pindex = OFF_TO_IDX(offset); end_pindex = pindex + atop(length); again: locked_depth = 1; VM_OBJECT_RLOCK(object); m = vm_page_find_least(object, pindex); while (pindex < end_pindex) { if (m == NULL || pindex < m->pindex) { /* * The first object in the shadow chain doesn't * contain a page at the current index. Therefore, * the page must exist in a backing object. */ tobject = object; tpindex = pindex; depth = 0; do { tpindex += OFF_TO_IDX(tobject->backing_object_offset); tobject = tobject->backing_object; KASSERT(tobject != NULL, ("vm_object_unwire: missing page")); if ((tobject->flags & OBJ_FICTITIOUS) != 0) goto next_page; depth++; if (depth == locked_depth) { locked_depth++; VM_OBJECT_RLOCK(tobject); } } while ((tm = vm_page_lookup(tobject, tpindex)) == NULL); } else { tm = m; m = TAILQ_NEXT(m, listq); } - if (vm_page_xbusied(tm)) { + if (vm_page_trysbusy(tm) == 0) { for (tobject = object; locked_depth >= 1; locked_depth--) { t1object = tobject->backing_object; if (tm->object != tobject) VM_OBJECT_RUNLOCK(tobject); tobject = t1object; } vm_page_busy_sleep(tm, "unwbo", true); goto again; } vm_page_unwire(tm, queue); + vm_page_sunbusy(tm); next_page: pindex++; } /* Release the accumulated object locks. */ for (tobject = object; locked_depth >= 1; locked_depth--) { t1object = tobject->backing_object; VM_OBJECT_RUNLOCK(tobject); tobject = t1object; } } /* * Return the vnode for the given object, or NULL if none exists. * For tmpfs objects, the function may return NULL if there is * no vnode allocated at the time of the call. */ struct vnode * vm_object_vnode(vm_object_t object) { struct vnode *vp; VM_OBJECT_ASSERT_LOCKED(object); if (object->type == OBJT_VNODE) { vp = object->handle; KASSERT(vp != NULL, ("%s: OBJT_VNODE has no vnode", __func__)); } else if (object->type == OBJT_SWAP && (object->flags & OBJ_TMPFS) != 0) { vp = object->un_pager.swp.swp_tmpfs; KASSERT(vp != NULL, ("%s: OBJT_TMPFS has no vnode", __func__)); } else { vp = NULL; } return (vp); } /* * Return the kvme type of the given object. * If vpp is not NULL, set it to the object's vm_object_vnode() or NULL. */ int vm_object_kvme_type(vm_object_t object, struct vnode **vpp) { VM_OBJECT_ASSERT_LOCKED(object); if (vpp != NULL) *vpp = vm_object_vnode(object); switch (object->type) { case OBJT_DEFAULT: return (KVME_TYPE_DEFAULT); case OBJT_VNODE: return (KVME_TYPE_VNODE); case OBJT_SWAP: if ((object->flags & OBJ_TMPFS_NODE) != 0) return (KVME_TYPE_VNODE); return (KVME_TYPE_SWAP); case OBJT_DEVICE: return (KVME_TYPE_DEVICE); case OBJT_PHYS: return (KVME_TYPE_PHYS); case OBJT_DEAD: return (KVME_TYPE_DEAD); case OBJT_SG: return (KVME_TYPE_SG); case OBJT_MGTDEVICE: return (KVME_TYPE_MGTDEVICE); default: return (KVME_TYPE_UNKNOWN); } } static int sysctl_vm_object_list(SYSCTL_HANDLER_ARGS) { struct kinfo_vmobject *kvo; char *fullpath, *freepath; struct vnode *vp; struct vattr va; vm_object_t obj; vm_page_t m; int count, error; if (req->oldptr == NULL) { /* * If an old buffer has not been provided, generate an * estimate of the space needed for a subsequent call. */ mtx_lock(&vm_object_list_mtx); count = 0; TAILQ_FOREACH(obj, &vm_object_list, object_list) { if (obj->type == OBJT_DEAD) continue; count++; } mtx_unlock(&vm_object_list_mtx); return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) * count * 11 / 10)); } kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK); error = 0; /* * VM objects are type stable and are never removed from the * list once added. This allows us to safely read obj->object_list * after reacquiring the VM object lock. */ mtx_lock(&vm_object_list_mtx); TAILQ_FOREACH(obj, &vm_object_list, object_list) { if (obj->type == OBJT_DEAD) continue; VM_OBJECT_RLOCK(obj); if (obj->type == OBJT_DEAD) { VM_OBJECT_RUNLOCK(obj); continue; } mtx_unlock(&vm_object_list_mtx); kvo->kvo_size = ptoa(obj->size); kvo->kvo_resident = obj->resident_page_count; kvo->kvo_ref_count = obj->ref_count; kvo->kvo_shadow_count = obj->shadow_count; kvo->kvo_memattr = obj->memattr; kvo->kvo_active = 0; kvo->kvo_inactive = 0; TAILQ_FOREACH(m, &obj->memq, listq) { /* * A page may belong to the object but be * dequeued and set to PQ_NONE while the * object lock is not held. This makes the * reads of m->queue below racy, and we do not * count pages set to PQ_NONE. However, this * sysctl is only meant to give an * approximation of the system anyway. */ if (m->queue == PQ_ACTIVE) kvo->kvo_active++; else if (m->queue == PQ_INACTIVE) kvo->kvo_inactive++; } kvo->kvo_vn_fileid = 0; kvo->kvo_vn_fsid = 0; kvo->kvo_vn_fsid_freebsd11 = 0; freepath = NULL; fullpath = ""; kvo->kvo_type = vm_object_kvme_type(obj, &vp); if (vp != NULL) vref(vp); VM_OBJECT_RUNLOCK(obj); if (vp != NULL) { vn_fullpath(curthread, vp, &fullpath, &freepath); vn_lock(vp, LK_SHARED | LK_RETRY); if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) { kvo->kvo_vn_fileid = va.va_fileid; kvo->kvo_vn_fsid = va.va_fsid; kvo->kvo_vn_fsid_freebsd11 = va.va_fsid; /* truncate */ } vput(vp); } strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path)); if (freepath != NULL) free(freepath, M_TEMP); /* Pack record size down */ kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path) + strlen(kvo->kvo_path) + 1; kvo->kvo_structsize = roundup(kvo->kvo_structsize, sizeof(uint64_t)); error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize); mtx_lock(&vm_object_list_mtx); if (error) break; } mtx_unlock(&vm_object_list_mtx); free(kvo, M_TEMP); return (error); } SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject", "List of VM objects"); #include "opt_ddb.h" #ifdef DDB #include #include #include static int _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) { vm_map_t tmpm; vm_map_entry_t tmpe; vm_object_t obj; if (map == 0) return 0; if (entry == 0) { VM_MAP_ENTRY_FOREACH(tmpe, map) { if (_vm_object_in_map(map, object, tmpe)) { return 1; } } } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { tmpm = entry->object.sub_map; VM_MAP_ENTRY_FOREACH(tmpe, tmpm) { if (_vm_object_in_map(tmpm, object, tmpe)) { return 1; } } } else if ((obj = entry->object.vm_object) != NULL) { for (; obj; obj = obj->backing_object) if (obj == object) { return 1; } } return 0; } static int vm_object_in_map(vm_object_t object) { struct proc *p; /* sx_slock(&allproc_lock); */ FOREACH_PROC_IN_SYSTEM(p) { if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) continue; if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { /* sx_sunlock(&allproc_lock); */ return 1; } } /* sx_sunlock(&allproc_lock); */ if (_vm_object_in_map(kernel_map, object, 0)) return 1; return 0; } DB_SHOW_COMMAND(vmochk, vm_object_check) { vm_object_t object; /* * make sure that internal objs are in a map somewhere * and none have zero ref counts. */ TAILQ_FOREACH(object, &vm_object_list, object_list) { if (object->handle == NULL && (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { if (object->ref_count == 0) { db_printf("vmochk: internal obj has zero ref count: %ld\n", (long)object->size); } if (!vm_object_in_map(object)) { db_printf( "vmochk: internal obj is not in a map: " "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", object->ref_count, (u_long)object->size, (u_long)object->size, (void *)object->backing_object); } } } } /* * vm_object_print: [ debug ] */ DB_SHOW_COMMAND(object, vm_object_print_static) { /* XXX convert args. */ vm_object_t object = (vm_object_t)addr; boolean_t full = have_addr; vm_page_t p; /* XXX count is an (unused) arg. Avoid shadowing it. */ #define count was_count int count; if (object == NULL) return; db_iprintf( "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n", object, (int)object->type, (uintmax_t)object->size, object->resident_page_count, object->ref_count, object->flags, object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge); db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n", object->shadow_count, object->backing_object ? object->backing_object->ref_count : 0, object->backing_object, (uintmax_t)object->backing_object_offset); if (!full) return; db_indent += 2; count = 0; TAILQ_FOREACH(p, &object->memq, listq) { if (count == 0) db_iprintf("memory:="); else if (count == 6) { db_printf("\n"); db_iprintf(" ..."); count = 0; } else db_printf(","); count++; db_printf("(off=0x%jx,page=0x%jx)", (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p)); } if (count != 0) db_printf("\n"); db_indent -= 2; } /* XXX. */ #undef count /* XXX need this non-static entry for calling from vm_map_print. */ void vm_object_print( /* db_expr_t */ long addr, boolean_t have_addr, /* db_expr_t */ long count, char *modif) { vm_object_print_static(addr, have_addr, count, modif); } DB_SHOW_COMMAND(vmopag, vm_object_print_pages) { vm_object_t object; vm_pindex_t fidx; vm_paddr_t pa; vm_page_t m, prev_m; int rcount, nl, c; nl = 0; TAILQ_FOREACH(object, &vm_object_list, object_list) { db_printf("new object: %p\n", (void *)object); if (nl > 18) { c = cngetc(); if (c != ' ') return; nl = 0; } nl++; rcount = 0; fidx = 0; pa = -1; TAILQ_FOREACH(m, &object->memq, listq) { if (m->pindex > 128) break; if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL && prev_m->pindex + 1 != m->pindex) { if (rcount) { db_printf(" index(%ld)run(%d)pa(0x%lx)\n", (long)fidx, rcount, (long)pa); if (nl > 18) { c = cngetc(); if (c != ' ') return; nl = 0; } nl++; rcount = 0; } } if (rcount && (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { ++rcount; continue; } if (rcount) { db_printf(" index(%ld)run(%d)pa(0x%lx)\n", (long)fidx, rcount, (long)pa); if (nl > 18) { c = cngetc(); if (c != ' ') return; nl = 0; } nl++; } fidx = m->pindex; pa = VM_PAGE_TO_PHYS(m); rcount = 1; } if (rcount) { db_printf(" index(%ld)run(%d)pa(0x%lx)\n", (long)fidx, rcount, (long)pa); if (nl > 18) { c = cngetc(); if (c != ' ') return; nl = 0; } nl++; } } } #endif /* DDB */ Index: head/sys/vm/vm_page.c =================================================================== --- head/sys/vm/vm_page.c (revision 353534) +++ head/sys/vm/vm_page.c (revision 353535) @@ -1,4931 +1,4977 @@ /*- * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU) * * Copyright (c) 1991 Regents of the University of California. * All rights reserved. * Copyright (c) 1998 Matthew Dillon. All Rights Reserved. * * This code is derived from software contributed to Berkeley by * The Mach Operating System project at Carnegie-Mellon University. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91 */ /*- * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Authors: Avadis Tevanian, Jr., Michael Wayne Young * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. */ /* * Resident memory management module. */ #include __FBSDID("$FreeBSD$"); #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include extern int uma_startup_count(int); extern void uma_startup(void *, int); extern int vmem_startup_count(void); struct vm_domain vm_dom[MAXMEMDOM]; DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]); struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT]; struct mtx_padalign __exclusive_cache_line vm_domainset_lock; /* The following fields are protected by the domainset lock. */ domainset_t __exclusive_cache_line vm_min_domains; domainset_t __exclusive_cache_line vm_severe_domains; static int vm_min_waiters; static int vm_severe_waiters; static int vm_pageproc_waiters; static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD, 0, "VM page statistics"); static counter_u64_t queue_ops = EARLY_COUNTER; SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops, CTLFLAG_RD, &queue_ops, "Number of batched queue operations"); static counter_u64_t queue_nops = EARLY_COUNTER; SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops, CTLFLAG_RD, &queue_nops, "Number of batched queue operations with no effects"); static void counter_startup(void) { queue_ops = counter_u64_alloc(M_WAITOK); queue_nops = counter_u64_alloc(M_WAITOK); } SYSINIT(page_counters, SI_SUB_CPU, SI_ORDER_ANY, counter_startup, NULL); /* * bogus page -- for I/O to/from partially complete buffers, * or for paging into sparsely invalid regions. */ vm_page_t bogus_page; vm_page_t vm_page_array; long vm_page_array_size; long first_page; static int boot_pages; SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &boot_pages, 0, "number of pages allocated for bootstrapping the VM system"); static int pa_tryrelock_restart; SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD, &pa_tryrelock_restart, 0, "Number of tryrelock restarts"); static TAILQ_HEAD(, vm_page) blacklist_head; static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS); SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages"); static uma_zone_t fakepg_zone; static void vm_page_alloc_check(vm_page_t m); static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); static void vm_page_dequeue_complete(vm_page_t m); static void vm_page_enqueue(vm_page_t m, uint8_t queue); static void vm_page_init(void *dummy); static int vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, vm_page_t mpred); static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred); static void vm_page_mvqueue(vm_page_t m, uint8_t queue); static int vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run, vm_paddr_t high); static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req); static int vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags); static void vm_page_zone_release(void *arg, void **store, int cnt); SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL); static void vm_page_init(void *dummy) { fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM); bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | VM_ALLOC_NORMAL | VM_ALLOC_WIRED); } /* * The cache page zone is initialized later since we need to be able to allocate * pages before UMA is fully initialized. */ static void vm_page_init_cache_zones(void *dummy __unused) { struct vm_domain *vmd; struct vm_pgcache *pgcache; int domain, pool; for (domain = 0; domain < vm_ndomains; domain++) { vmd = VM_DOMAIN(domain); /* * Don't allow the page caches to take up more than .25% of * memory. */ if (vmd->vmd_page_count / 400 < 256 * mp_ncpus * VM_NFREEPOOL) continue; for (pool = 0; pool < VM_NFREEPOOL; pool++) { pgcache = &vmd->vmd_pgcache[pool]; pgcache->domain = domain; pgcache->pool = pool; pgcache->zone = uma_zcache_create("vm pgcache", sizeof(struct vm_page), NULL, NULL, NULL, NULL, vm_page_zone_import, vm_page_zone_release, pgcache, UMA_ZONE_MAXBUCKET | UMA_ZONE_VM); (void)uma_zone_set_maxcache(pgcache->zone, 0); } } } SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL); /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ #if PAGE_SIZE == 32768 #ifdef CTASSERT CTASSERT(sizeof(u_long) >= 8); #endif #endif /* * Try to acquire a physical address lock while a pmap is locked. If we * fail to trylock we unlock and lock the pmap directly and cache the * locked pa in *locked. The caller should then restart their loop in case * the virtual to physical mapping has changed. */ int vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked) { vm_paddr_t lockpa; lockpa = *locked; *locked = pa; if (lockpa) { PA_LOCK_ASSERT(lockpa, MA_OWNED); if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa)) return (0); PA_UNLOCK(lockpa); } if (PA_TRYLOCK(pa)) return (0); PMAP_UNLOCK(pmap); atomic_add_int(&pa_tryrelock_restart, 1); PA_LOCK(pa); PMAP_LOCK(pmap); return (EAGAIN); } /* * vm_set_page_size: * * Sets the page size, perhaps based upon the memory * size. Must be called before any use of page-size * dependent functions. */ void vm_set_page_size(void) { if (vm_cnt.v_page_size == 0) vm_cnt.v_page_size = PAGE_SIZE; if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0) panic("vm_set_page_size: page size not a power of two"); } /* * vm_page_blacklist_next: * * Find the next entry in the provided string of blacklist * addresses. Entries are separated by space, comma, or newline. * If an invalid integer is encountered then the rest of the * string is skipped. Updates the list pointer to the next * character, or NULL if the string is exhausted or invalid. */ static vm_paddr_t vm_page_blacklist_next(char **list, char *end) { vm_paddr_t bad; char *cp, *pos; if (list == NULL || *list == NULL) return (0); if (**list =='\0') { *list = NULL; return (0); } /* * If there's no end pointer then the buffer is coming from * the kenv and we know it's null-terminated. */ if (end == NULL) end = *list + strlen(*list); /* Ensure that strtoq() won't walk off the end */ if (*end != '\0') { if (*end == '\n' || *end == ' ' || *end == ',') *end = '\0'; else { printf("Blacklist not terminated, skipping\n"); *list = NULL; return (0); } } for (pos = *list; *pos != '\0'; pos = cp) { bad = strtoq(pos, &cp, 0); if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') { if (bad == 0) { if (++cp < end) continue; else break; } } else break; if (*cp == '\0' || ++cp >= end) *list = NULL; else *list = cp; return (trunc_page(bad)); } printf("Garbage in RAM blacklist, skipping\n"); *list = NULL; return (0); } bool vm_page_blacklist_add(vm_paddr_t pa, bool verbose) { struct vm_domain *vmd; vm_page_t m; int ret; m = vm_phys_paddr_to_vm_page(pa); if (m == NULL) return (true); /* page does not exist, no failure */ vmd = vm_pagequeue_domain(m); vm_domain_free_lock(vmd); ret = vm_phys_unfree_page(m); vm_domain_free_unlock(vmd); if (ret != 0) { vm_domain_freecnt_inc(vmd, -1); TAILQ_INSERT_TAIL(&blacklist_head, m, listq); if (verbose) printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa); } return (ret); } /* * vm_page_blacklist_check: * * Iterate through the provided string of blacklist addresses, pulling * each entry out of the physical allocator free list and putting it * onto a list for reporting via the vm.page_blacklist sysctl. */ static void vm_page_blacklist_check(char *list, char *end) { vm_paddr_t pa; char *next; next = list; while (next != NULL) { if ((pa = vm_page_blacklist_next(&next, end)) == 0) continue; vm_page_blacklist_add(pa, bootverbose); } } /* * vm_page_blacklist_load: * * Search for a special module named "ram_blacklist". It'll be a * plain text file provided by the user via the loader directive * of the same name. */ static void vm_page_blacklist_load(char **list, char **end) { void *mod; u_char *ptr; u_int len; mod = NULL; ptr = NULL; mod = preload_search_by_type("ram_blacklist"); if (mod != NULL) { ptr = preload_fetch_addr(mod); len = preload_fetch_size(mod); } *list = ptr; if (ptr != NULL) *end = ptr + len; else *end = NULL; return; } static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS) { vm_page_t m; struct sbuf sbuf; int error, first; first = 1; error = sysctl_wire_old_buffer(req, 0); if (error != 0) return (error); sbuf_new_for_sysctl(&sbuf, NULL, 128, req); TAILQ_FOREACH(m, &blacklist_head, listq) { sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",", (uintmax_t)m->phys_addr); first = 0; } error = sbuf_finish(&sbuf); sbuf_delete(&sbuf); return (error); } /* * Initialize a dummy page for use in scans of the specified paging queue. * In principle, this function only needs to set the flag PG_MARKER. * Nonetheless, it write busies the page as a safety precaution. */ static void vm_page_init_marker(vm_page_t marker, int queue, uint8_t aflags) { bzero(marker, sizeof(*marker)); marker->flags = PG_MARKER; marker->aflags = aflags; marker->busy_lock = VPB_SINGLE_EXCLUSIVER; marker->queue = queue; } static void vm_page_domain_init(int domain) { struct vm_domain *vmd; struct vm_pagequeue *pq; int i; vmd = VM_DOMAIN(domain); bzero(vmd, sizeof(*vmd)); *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) = "vm inactive pagequeue"; *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) = "vm active pagequeue"; *__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) = "vm laundry pagequeue"; *__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) = "vm unswappable pagequeue"; vmd->vmd_domain = domain; vmd->vmd_page_count = 0; vmd->vmd_free_count = 0; vmd->vmd_segs = 0; vmd->vmd_oom = FALSE; for (i = 0; i < PQ_COUNT; i++) { pq = &vmd->vmd_pagequeues[i]; TAILQ_INIT(&pq->pq_pl); mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue", MTX_DEF | MTX_DUPOK); pq->pq_pdpages = 0; vm_page_init_marker(&vmd->vmd_markers[i], i, 0); } mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF); mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF); snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain); /* * inacthead is used to provide FIFO ordering for LRU-bypassing * insertions. */ vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED); TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl, &vmd->vmd_inacthead, plinks.q); /* * The clock pages are used to implement active queue scanning without * requeues. Scans start at clock[0], which is advanced after the scan * ends. When the two clock hands meet, they are reset and scanning * resumes from the head of the queue. */ vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED); vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED); TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, &vmd->vmd_clock[0], plinks.q); TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, &vmd->vmd_clock[1], plinks.q); } /* * Initialize a physical page in preparation for adding it to the free * lists. */ static void vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind) { m->object = NULL; m->ref_count = 0; m->busy_lock = VPB_UNBUSIED; m->flags = m->aflags = 0; m->phys_addr = pa; m->queue = PQ_NONE; m->psind = 0; m->segind = segind; m->order = VM_NFREEORDER; m->pool = VM_FREEPOOL_DEFAULT; m->valid = m->dirty = 0; pmap_page_init(m); } #ifndef PMAP_HAS_PAGE_ARRAY static vm_paddr_t vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range) { vm_paddr_t new_end; /* * Reserve an unmapped guard page to trap access to vm_page_array[-1]. * However, because this page is allocated from KVM, out-of-bounds * accesses using the direct map will not be trapped. */ *vaddr += PAGE_SIZE; /* * Allocate physical memory for the page structures, and map it. */ new_end = trunc_page(end - page_range * sizeof(struct vm_page)); vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE); vm_page_array_size = page_range; return (new_end); } #endif /* * vm_page_startup: * * Initializes the resident memory module. Allocates physical memory for * bootstrapping UMA and some data structures that are used to manage * physical pages. Initializes these structures, and populates the free * page queues. */ vm_offset_t vm_page_startup(vm_offset_t vaddr) { struct vm_phys_seg *seg; vm_page_t m; char *list, *listend; vm_offset_t mapped; vm_paddr_t end, high_avail, low_avail, new_end, page_range, size; vm_paddr_t last_pa, pa; u_long pagecount; int biggestone, i, segind; #ifdef WITNESS int witness_size; #endif #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) long ii; #endif vaddr = round_page(vaddr); vm_phys_early_startup(); biggestone = vm_phys_avail_largest(); end = phys_avail[biggestone+1]; /* * Initialize the page and queue locks. */ mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF); for (i = 0; i < PA_LOCK_COUNT; i++) mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF); for (i = 0; i < vm_ndomains; i++) vm_page_domain_init(i); /* * Allocate memory for use when boot strapping the kernel memory * allocator. Tell UMA how many zones we are going to create * before going fully functional. UMA will add its zones. * * VM startup zones: vmem, vmem_btag, VM OBJECT, RADIX NODE, MAP, * KMAP ENTRY, MAP ENTRY, VMSPACE. */ boot_pages = uma_startup_count(8); #ifndef UMA_MD_SMALL_ALLOC /* vmem_startup() calls uma_prealloc(). */ boot_pages += vmem_startup_count(); /* vm_map_startup() calls uma_prealloc(). */ boot_pages += howmany(MAX_KMAP, UMA_SLAB_SPACE / sizeof(struct vm_map)); /* * Before going fully functional kmem_init() does allocation * from "KMAP ENTRY" and vmem_create() does allocation from "vmem". */ boot_pages += 2; #endif /* * CTFLAG_RDTUN doesn't work during the early boot process, so we must * manually fetch the value. */ TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages); new_end = end - (boot_pages * UMA_SLAB_SIZE); new_end = trunc_page(new_end); mapped = pmap_map(&vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE); bzero((void *)mapped, end - new_end); uma_startup((void *)mapped, boot_pages); #ifdef WITNESS witness_size = round_page(witness_startup_count()); new_end -= witness_size; mapped = pmap_map(&vaddr, new_end, new_end + witness_size, VM_PROT_READ | VM_PROT_WRITE); bzero((void *)mapped, witness_size); witness_startup((void *)mapped); #endif #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \ defined(__i386__) || defined(__mips__) || defined(__riscv) || \ defined(__powerpc64__) /* * Allocate a bitmap to indicate that a random physical page * needs to be included in a minidump. * * The amd64 port needs this to indicate which direct map pages * need to be dumped, via calls to dump_add_page()/dump_drop_page(). * * However, i386 still needs this workspace internally within the * minidump code. In theory, they are not needed on i386, but are * included should the sf_buf code decide to use them. */ last_pa = 0; for (i = 0; dump_avail[i + 1] != 0; i += 2) if (dump_avail[i + 1] > last_pa) last_pa = dump_avail[i + 1]; page_range = last_pa / PAGE_SIZE; vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); new_end -= vm_page_dump_size; vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); bzero((void *)vm_page_dump, vm_page_dump_size); #else (void)last_pa; #endif #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \ defined(__riscv) || defined(__powerpc64__) /* * Include the UMA bootstrap pages, witness pages and vm_page_dump * in a crash dump. When pmap_map() uses the direct map, they are * not automatically included. */ for (pa = new_end; pa < end; pa += PAGE_SIZE) dump_add_page(pa); #endif phys_avail[biggestone + 1] = new_end; #ifdef __amd64__ /* * Request that the physical pages underlying the message buffer be * included in a crash dump. Since the message buffer is accessed * through the direct map, they are not automatically included. */ pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); last_pa = pa + round_page(msgbufsize); while (pa < last_pa) { dump_add_page(pa); pa += PAGE_SIZE; } #endif /* * Compute the number of pages of memory that will be available for * use, taking into account the overhead of a page structure per page. * In other words, solve * "available physical memory" - round_page(page_range * * sizeof(struct vm_page)) = page_range * PAGE_SIZE * for page_range. */ low_avail = phys_avail[0]; high_avail = phys_avail[1]; for (i = 0; i < vm_phys_nsegs; i++) { if (vm_phys_segs[i].start < low_avail) low_avail = vm_phys_segs[i].start; if (vm_phys_segs[i].end > high_avail) high_avail = vm_phys_segs[i].end; } /* Skip the first chunk. It is already accounted for. */ for (i = 2; phys_avail[i + 1] != 0; i += 2) { if (phys_avail[i] < low_avail) low_avail = phys_avail[i]; if (phys_avail[i + 1] > high_avail) high_avail = phys_avail[i + 1]; } first_page = low_avail / PAGE_SIZE; #ifdef VM_PHYSSEG_SPARSE size = 0; for (i = 0; i < vm_phys_nsegs; i++) size += vm_phys_segs[i].end - vm_phys_segs[i].start; for (i = 0; phys_avail[i + 1] != 0; i += 2) size += phys_avail[i + 1] - phys_avail[i]; #elif defined(VM_PHYSSEG_DENSE) size = high_avail - low_avail; #else #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." #endif #ifdef PMAP_HAS_PAGE_ARRAY pmap_page_array_startup(size / PAGE_SIZE); biggestone = vm_phys_avail_largest(); end = new_end = phys_avail[biggestone + 1]; #else #ifdef VM_PHYSSEG_DENSE /* * In the VM_PHYSSEG_DENSE case, the number of pages can account for * the overhead of a page structure per page only if vm_page_array is * allocated from the last physical memory chunk. Otherwise, we must * allocate page structures representing the physical memory * underlying vm_page_array, even though they will not be used. */ if (new_end != high_avail) page_range = size / PAGE_SIZE; else #endif { page_range = size / (PAGE_SIZE + sizeof(struct vm_page)); /* * If the partial bytes remaining are large enough for * a page (PAGE_SIZE) without a corresponding * 'struct vm_page', then new_end will contain an * extra page after subtracting the length of the VM * page array. Compensate by subtracting an extra * page from new_end. */ if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) { if (new_end == high_avail) high_avail -= PAGE_SIZE; new_end -= PAGE_SIZE; } } end = new_end; new_end = vm_page_array_alloc(&vaddr, end, page_range); #endif #if VM_NRESERVLEVEL > 0 /* * Allocate physical memory for the reservation management system's * data structures, and map it. */ new_end = vm_reserv_startup(&vaddr, new_end); #endif #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \ defined(__riscv) || defined(__powerpc64__) /* * Include vm_page_array and vm_reserv_array in a crash dump. */ for (pa = new_end; pa < end; pa += PAGE_SIZE) dump_add_page(pa); #endif phys_avail[biggestone + 1] = new_end; /* * Add physical memory segments corresponding to the available * physical pages. */ for (i = 0; phys_avail[i + 1] != 0; i += 2) if (vm_phys_avail_size(i) != 0) vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]); /* * Initialize the physical memory allocator. */ vm_phys_init(); /* * Initialize the page structures and add every available page to the * physical memory allocator's free lists. */ #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) for (ii = 0; ii < vm_page_array_size; ii++) { m = &vm_page_array[ii]; vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0); m->flags = PG_FICTITIOUS; } #endif vm_cnt.v_page_count = 0; for (segind = 0; segind < vm_phys_nsegs; segind++) { seg = &vm_phys_segs[segind]; for (m = seg->first_page, pa = seg->start; pa < seg->end; m++, pa += PAGE_SIZE) vm_page_init_page(m, pa, segind); /* * Add the segment to the free lists only if it is covered by * one of the ranges in phys_avail. Because we've added the * ranges to the vm_phys_segs array, we can assume that each * segment is either entirely contained in one of the ranges, * or doesn't overlap any of them. */ for (i = 0; phys_avail[i + 1] != 0; i += 2) { struct vm_domain *vmd; if (seg->start < phys_avail[i] || seg->end > phys_avail[i + 1]) continue; m = seg->first_page; pagecount = (u_long)atop(seg->end - seg->start); vmd = VM_DOMAIN(seg->domain); vm_domain_free_lock(vmd); vm_phys_enqueue_contig(m, pagecount); vm_domain_free_unlock(vmd); vm_domain_freecnt_inc(vmd, pagecount); vm_cnt.v_page_count += (u_int)pagecount; vmd = VM_DOMAIN(seg->domain); vmd->vmd_page_count += (u_int)pagecount; vmd->vmd_segs |= 1UL << m->segind; break; } } /* * Remove blacklisted pages from the physical memory allocator. */ TAILQ_INIT(&blacklist_head); vm_page_blacklist_load(&list, &listend); vm_page_blacklist_check(list, listend); list = kern_getenv("vm.blacklist"); vm_page_blacklist_check(list, NULL); freeenv(list); #if VM_NRESERVLEVEL > 0 /* * Initialize the reservation management system. */ vm_reserv_init(); #endif return (vaddr); } void vm_page_reference(vm_page_t m) { vm_page_aflag_set(m, PGA_REFERENCED); } /* * vm_page_busy_acquire: * * Acquire the busy lock as described by VM_ALLOC_* flags. Will loop * and drop the object lock if necessary. */ int vm_page_busy_acquire(vm_page_t m, int allocflags) { vm_object_t obj; u_int x; bool locked; /* * The page-specific object must be cached because page * identity can change during the sleep, causing the * re-lock of a different object. * It is assumed that a reference to the object is already * held by the callers. */ obj = m->object; for (;;) { if ((allocflags & VM_ALLOC_SBUSY) == 0) { if (vm_page_tryxbusy(m)) return (TRUE); } else { if (vm_page_trysbusy(m)) return (TRUE); } if ((allocflags & VM_ALLOC_NOWAIT) != 0) return (FALSE); if (obj != NULL) { locked = VM_OBJECT_WOWNED(obj); } else { MPASS(vm_page_wired(m)); locked = FALSE; } sleepq_lock(m); x = m->busy_lock; if (x == VPB_UNBUSIED || ((allocflags & VM_ALLOC_SBUSY) != 0 && (x & VPB_BIT_SHARED) != 0) || ((x & VPB_BIT_WAITERS) == 0 && !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) { sleepq_release(m); continue; } if (locked) VM_OBJECT_WUNLOCK(obj); sleepq_add(m, NULL, "vmpba", 0, 0); sleepq_wait(m, PVM); if (locked) VM_OBJECT_WLOCK(obj); MPASS(m->object == obj || m->object == NULL); if ((allocflags & VM_ALLOC_WAITFAIL) != 0) return (FALSE); } } /* * vm_page_busy_downgrade: * * Downgrade an exclusive busy page into a single shared busy page. */ void vm_page_busy_downgrade(vm_page_t m) { u_int x; vm_page_assert_xbusied(m); x = m->busy_lock; for (;;) { if (atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_SHARERS_WORD(1))) break; } if ((x & VPB_BIT_WAITERS) != 0) wakeup(m); } /* + * + * vm_page_busy_tryupgrade: + * + * Attempt to upgrade a single shared busy into an exclusive busy. + */ +int +vm_page_busy_tryupgrade(vm_page_t m) +{ + u_int x; + + vm_page_assert_sbusied(m); + + x = m->busy_lock; + for (;;) { + if (VPB_SHARERS(x) > 1) + return (0); + KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), + ("vm_page_busy_tryupgrade: invalid lock state")); + if (!atomic_fcmpset_acq_int(&m->busy_lock, &x, + VPB_SINGLE_EXCLUSIVER | (x & VPB_BIT_WAITERS))) + continue; + return (1); + } +} + +/* * vm_page_sbusied: * * Return a positive value if the page is shared busied, 0 otherwise. */ int vm_page_sbusied(vm_page_t m) { u_int x; x = m->busy_lock; return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); } /* * vm_page_sunbusy: * * Shared unbusy a page. */ void vm_page_sunbusy(vm_page_t m) { u_int x; vm_page_assert_sbusied(m); x = m->busy_lock; for (;;) { if (VPB_SHARERS(x) > 1) { if (atomic_fcmpset_int(&m->busy_lock, &x, x - VPB_ONE_SHARER)) break; continue; } KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), ("vm_page_sunbusy: invalid lock state")); if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) continue; if ((x & VPB_BIT_WAITERS) == 0) break; wakeup(m); break; } } /* * vm_page_busy_sleep: * * Sleep if the page is busy, using the page pointer as wchan. * This is used to implement the hard-path of busying mechanism. * * If nonshared is true, sleep only if the page is xbusy. * * The object lock must be held on entry and will be released on exit. */ void vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared) { vm_object_t obj; u_int x; obj = m->object; vm_page_lock_assert(m, MA_NOTOWNED); VM_OBJECT_ASSERT_LOCKED(obj); sleepq_lock(m); x = m->busy_lock; if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) || ((x & VPB_BIT_WAITERS) == 0 && !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) { VM_OBJECT_DROP(obj); sleepq_release(m); return; } VM_OBJECT_DROP(obj); sleepq_add(m, NULL, wmesg, 0, 0); sleepq_wait(m, PVM); } /* * vm_page_trysbusy: * * Try to shared busy a page. * If the operation succeeds 1 is returned otherwise 0. * The operation never sleeps. */ int vm_page_trysbusy(vm_page_t m) { u_int x; x = m->busy_lock; for (;;) { if ((x & VPB_BIT_SHARED) == 0) return (0); if (atomic_fcmpset_acq_int(&m->busy_lock, &x, x + VPB_ONE_SHARER)) return (1); } } /* * vm_page_xunbusy_hard: * * Called when unbusy has failed because there is a waiter. */ void vm_page_xunbusy_hard(vm_page_t m) { vm_page_assert_xbusied(m); /* * Wake the waiter. */ atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); wakeup(m); } /* * Avoid releasing and reacquiring the same page lock. */ void vm_page_change_lock(vm_page_t m, struct mtx **mtx) { struct mtx *mtx1; mtx1 = vm_page_lockptr(m); if (*mtx == mtx1) return; if (*mtx != NULL) mtx_unlock(*mtx); *mtx = mtx1; mtx_lock(mtx1); } /* * vm_page_unhold_pages: * * Unhold each of the pages that is referenced by the given array. */ void vm_page_unhold_pages(vm_page_t *ma, int count) { for (; count != 0; count--) { vm_page_unwire(*ma, PQ_ACTIVE); ma++; } } vm_page_t PHYS_TO_VM_PAGE(vm_paddr_t pa) { vm_page_t m; #ifdef VM_PHYSSEG_SPARSE m = vm_phys_paddr_to_vm_page(pa); if (m == NULL) m = vm_phys_fictitious_to_vm_page(pa); return (m); #elif defined(VM_PHYSSEG_DENSE) long pi; pi = atop(pa); if (pi >= first_page && (pi - first_page) < vm_page_array_size) { m = &vm_page_array[pi - first_page]; return (m); } return (vm_phys_fictitious_to_vm_page(pa)); #else #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." #endif } /* * vm_page_getfake: * * Create a fictitious page with the specified physical address and * memory attribute. The memory attribute is the only the machine- * dependent aspect of a fictitious page that must be initialized. */ vm_page_t vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) { vm_page_t m; m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); vm_page_initfake(m, paddr, memattr); return (m); } void vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) { if ((m->flags & PG_FICTITIOUS) != 0) { /* * The page's memattr might have changed since the * previous initialization. Update the pmap to the * new memattr. */ goto memattr; } m->phys_addr = paddr; m->queue = PQ_NONE; /* Fictitious pages don't use "segind". */ m->flags = PG_FICTITIOUS; /* Fictitious pages don't use "order" or "pool". */ m->oflags = VPO_UNMANAGED; m->busy_lock = VPB_SINGLE_EXCLUSIVER; /* Fictitious pages are unevictable. */ m->ref_count = 1; pmap_page_init(m); memattr: pmap_page_set_memattr(m, memattr); } /* * vm_page_putfake: * * Release a fictitious page. */ void vm_page_putfake(vm_page_t m) { KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); KASSERT((m->flags & PG_FICTITIOUS) != 0, ("vm_page_putfake: bad page %p", m)); uma_zfree(fakepg_zone, m); } /* * vm_page_updatefake: * * Update the given fictitious page to the specified physical address and * memory attribute. */ void vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) { KASSERT((m->flags & PG_FICTITIOUS) != 0, ("vm_page_updatefake: bad page %p", m)); m->phys_addr = paddr; pmap_page_set_memattr(m, memattr); } /* * vm_page_free: * * Free a page. */ void vm_page_free(vm_page_t m) { m->flags &= ~PG_ZERO; vm_page_free_toq(m); } /* * vm_page_free_zero: * * Free a page to the zerod-pages queue */ void vm_page_free_zero(vm_page_t m) { m->flags |= PG_ZERO; vm_page_free_toq(m); } /* * Unbusy and handle the page queueing for a page from a getpages request that * was optionally read ahead or behind. */ void vm_page_readahead_finish(vm_page_t m) { /* We shouldn't put invalid pages on queues. */ KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m)); /* * Since the page is not the actually needed one, whether it should * be activated or deactivated is not obvious. Empirical results * have shown that deactivating the page is usually the best choice, * unless the page is wanted by another thread. */ vm_page_lock(m); if ((m->busy_lock & VPB_BIT_WAITERS) != 0) vm_page_activate(m); else vm_page_deactivate(m); vm_page_unlock(m); vm_page_xunbusy(m); } /* * vm_page_sleep_if_busy: * * Sleep and release the object lock if the page is busied. * Returns TRUE if the thread slept. * * The given page must be unlocked and object containing it must * be locked. */ int vm_page_sleep_if_busy(vm_page_t m, const char *msg) { vm_object_t obj; vm_page_lock_assert(m, MA_NOTOWNED); VM_OBJECT_ASSERT_WLOCKED(m->object); if (vm_page_busied(m)) { /* * The page-specific object must be cached because page * identity can change during the sleep, causing the * re-lock of a different object. * It is assumed that a reference to the object is already * held by the callers. */ obj = m->object; vm_page_busy_sleep(m, msg, false); VM_OBJECT_WLOCK(obj); return (TRUE); } return (FALSE); } /* * vm_page_sleep_if_xbusy: * * Sleep and release the object lock if the page is xbusied. * Returns TRUE if the thread slept. * * The given page must be unlocked and object containing it must * be locked. */ int vm_page_sleep_if_xbusy(vm_page_t m, const char *msg) { vm_object_t obj; vm_page_lock_assert(m, MA_NOTOWNED); VM_OBJECT_ASSERT_WLOCKED(m->object); if (vm_page_xbusied(m)) { /* * The page-specific object must be cached because page * identity can change during the sleep, causing the * re-lock of a different object. * It is assumed that a reference to the object is already * held by the callers. */ obj = m->object; vm_page_busy_sleep(m, msg, true); VM_OBJECT_WLOCK(obj); return (TRUE); } return (FALSE); } /* * vm_page_dirty_KBI: [ internal use only ] * * Set all bits in the page's dirty field. * * The object containing the specified page must be locked if the * call is made from the machine-independent layer. * * See vm_page_clear_dirty_mask(). * * This function should only be called by vm_page_dirty(). */ void vm_page_dirty_KBI(vm_page_t m) { /* Refer to this operation by its public name. */ KASSERT(m->valid == VM_PAGE_BITS_ALL, ("vm_page_dirty: page is invalid!")); m->dirty = VM_PAGE_BITS_ALL; } /* * vm_page_insert: [ internal use only ] * * Inserts the given mem entry into the object and object list. * * The object must be locked. */ int vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) { vm_page_t mpred; VM_OBJECT_ASSERT_WLOCKED(object); mpred = vm_radix_lookup_le(&object->rtree, pindex); return (vm_page_insert_after(m, object, pindex, mpred)); } /* * vm_page_insert_after: * * Inserts the page "m" into the specified object at offset "pindex". * * The page "mpred" must immediately precede the offset "pindex" within * the specified object. * * The object must be locked. */ static int vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, vm_page_t mpred) { vm_page_t msucc; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(m->object == NULL, ("vm_page_insert_after: page already inserted")); if (mpred != NULL) { KASSERT(mpred->object == object, ("vm_page_insert_after: object doesn't contain mpred")); KASSERT(mpred->pindex < pindex, ("vm_page_insert_after: mpred doesn't precede pindex")); msucc = TAILQ_NEXT(mpred, listq); } else msucc = TAILQ_FIRST(&object->memq); if (msucc != NULL) KASSERT(msucc->pindex > pindex, ("vm_page_insert_after: msucc doesn't succeed pindex")); /* * Record the object/offset pair in this page. */ m->object = object; m->pindex = pindex; m->ref_count |= VPRC_OBJREF; /* * Now link into the object's ordered list of backed pages. */ if (vm_radix_insert(&object->rtree, m)) { m->object = NULL; m->pindex = 0; m->ref_count &= ~VPRC_OBJREF; return (1); } vm_page_insert_radixdone(m, object, mpred); return (0); } /* * vm_page_insert_radixdone: * * Complete page "m" insertion into the specified object after the * radix trie hooking. * * The page "mpred" must precede the offset "m->pindex" within the * specified object. * * The object must be locked. */ static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) { VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(object != NULL && m->object == object, ("vm_page_insert_radixdone: page %p has inconsistent object", m)); KASSERT((m->ref_count & VPRC_OBJREF) != 0, ("vm_page_insert_radixdone: page %p is missing object ref", m)); if (mpred != NULL) { KASSERT(mpred->object == object, ("vm_page_insert_radixdone: object doesn't contain mpred")); KASSERT(mpred->pindex < m->pindex, ("vm_page_insert_radixdone: mpred doesn't precede pindex")); } if (mpred != NULL) TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); else TAILQ_INSERT_HEAD(&object->memq, m, listq); /* * Show that the object has one more resident page. */ object->resident_page_count++; /* * Hold the vnode until the last page is released. */ if (object->resident_page_count == 1 && object->type == OBJT_VNODE) vhold(object->handle); /* * Since we are inserting a new and possibly dirty page, * update the object's OBJ_MIGHTBEDIRTY flag. */ if (pmap_page_is_write_mapped(m)) vm_object_set_writeable_dirty(object); } /* * Do the work to remove a page from its object. The caller is responsible for * updating the page's fields to reflect this removal. */ static void vm_page_object_remove(vm_page_t m) { vm_object_t object; vm_page_t mrem; object = m->object; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT((m->ref_count & VPRC_OBJREF) != 0, ("page %p is missing its object ref", m)); if (vm_page_xbusied(m)) vm_page_xunbusy(m); mrem = vm_radix_remove(&object->rtree, m->pindex); KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m)); /* * Now remove from the object's list of backed pages. */ TAILQ_REMOVE(&object->memq, m, listq); /* * And show that the object has one fewer resident page. */ object->resident_page_count--; /* * The vnode may now be recycled. */ if (object->resident_page_count == 0 && object->type == OBJT_VNODE) vdrop(object->handle); } /* * vm_page_remove: * * Removes the specified page from its containing object, but does not * invalidate any backing storage. Returns true if the object's reference * was the last reference to the page, and false otherwise. * * The object must be locked. */ bool vm_page_remove(vm_page_t m) { vm_page_object_remove(m); m->object = NULL; return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF); } /* * vm_page_lookup: * * Returns the page associated with the object/offset * pair specified; if none is found, NULL is returned. * * The object must be locked. */ vm_page_t vm_page_lookup(vm_object_t object, vm_pindex_t pindex) { VM_OBJECT_ASSERT_LOCKED(object); return (vm_radix_lookup(&object->rtree, pindex)); } /* * vm_page_find_least: * * Returns the page associated with the object with least pindex * greater than or equal to the parameter pindex, or NULL. * * The object must be locked. */ vm_page_t vm_page_find_least(vm_object_t object, vm_pindex_t pindex) { vm_page_t m; VM_OBJECT_ASSERT_LOCKED(object); if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) m = vm_radix_lookup_ge(&object->rtree, pindex); return (m); } /* * Returns the given page's successor (by pindex) within the object if it is * resident; if none is found, NULL is returned. * * The object must be locked. */ vm_page_t vm_page_next(vm_page_t m) { vm_page_t next; VM_OBJECT_ASSERT_LOCKED(m->object); if ((next = TAILQ_NEXT(m, listq)) != NULL) { MPASS(next->object == m->object); if (next->pindex != m->pindex + 1) next = NULL; } return (next); } /* * Returns the given page's predecessor (by pindex) within the object if it is * resident; if none is found, NULL is returned. * * The object must be locked. */ vm_page_t vm_page_prev(vm_page_t m) { vm_page_t prev; VM_OBJECT_ASSERT_LOCKED(m->object); if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) { MPASS(prev->object == m->object); if (prev->pindex != m->pindex - 1) prev = NULL; } return (prev); } /* * Uses the page mnew as a replacement for an existing page at index * pindex which must be already present in the object. */ vm_page_t vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex) { vm_page_t mold; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0, ("vm_page_replace: page %p already in object", mnew)); /* * This function mostly follows vm_page_insert() and * vm_page_remove() without the radix, object count and vnode * dance. Double check such functions for more comments. */ mnew->object = object; mnew->pindex = pindex; atomic_set_int(&mnew->ref_count, VPRC_OBJREF); mold = vm_radix_replace(&object->rtree, mnew); KASSERT(mold->queue == PQ_NONE, ("vm_page_replace: old page %p is on a paging queue", mold)); /* Keep the resident page list in sorted order. */ TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq); TAILQ_REMOVE(&object->memq, mold, listq); mold->object = NULL; atomic_clear_int(&mold->ref_count, VPRC_OBJREF); vm_page_xunbusy(mold); /* * The object's resident_page_count does not change because we have * swapped one page for another, but OBJ_MIGHTBEDIRTY. */ if (pmap_page_is_write_mapped(mnew)) vm_object_set_writeable_dirty(object); return (mold); } /* * vm_page_rename: * * Move the given memory entry from its * current object to the specified target object/offset. * * Note: swap associated with the page must be invalidated by the move. We * have to do this for several reasons: (1) we aren't freeing the * page, (2) we are dirtying the page, (3) the VM system is probably * moving the page from object A to B, and will then later move * the backing store from A to B and we can't have a conflict. * * Note: we *always* dirty the page. It is necessary both for the * fact that we moved it, and because we may be invalidating * swap. * * The objects must be locked. */ int vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) { vm_page_t mpred; vm_pindex_t opidx; VM_OBJECT_ASSERT_WLOCKED(new_object); KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m)); mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex); KASSERT(mpred == NULL || mpred->pindex != new_pindex, ("vm_page_rename: pindex already renamed")); /* * Create a custom version of vm_page_insert() which does not depend * by m_prev and can cheat on the implementation aspects of the * function. */ opidx = m->pindex; m->pindex = new_pindex; if (vm_radix_insert(&new_object->rtree, m)) { m->pindex = opidx; return (1); } /* * The operation cannot fail anymore. The removal must happen before * the listq iterator is tainted. */ m->pindex = opidx; vm_page_object_remove(m); /* Return back to the new pindex to complete vm_page_insert(). */ m->pindex = new_pindex; m->object = new_object; vm_page_insert_radixdone(m, new_object, mpred); vm_page_dirty(m); return (0); } /* * vm_page_alloc: * * Allocate and return a page that is associated with the specified * object and offset pair. By default, this page is exclusive busied. * * The caller must always specify an allocation class. * * allocation classes: * VM_ALLOC_NORMAL normal process request * VM_ALLOC_SYSTEM system *really* needs a page * VM_ALLOC_INTERRUPT interrupt time request * * optional allocation flags: * VM_ALLOC_COUNT(number) the number of additional pages that the caller * intends to allocate * VM_ALLOC_NOBUSY do not exclusive busy the page * VM_ALLOC_NODUMP do not include the page in a kernel core dump * VM_ALLOC_NOOBJ page is not associated with an object and * should not be exclusive busy * VM_ALLOC_SBUSY shared busy the allocated page * VM_ALLOC_WIRED wire the allocated page * VM_ALLOC_ZERO prefer a zeroed page */ vm_page_t vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) { return (vm_page_alloc_after(object, pindex, req, object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) : NULL)); } vm_page_t vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain, int req) { return (vm_page_alloc_domain_after(object, pindex, domain, req, object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) : NULL)); } /* * Allocate a page in the specified object with the given page index. To * optimize insertion of the page into the object, the caller must also specifiy * the resident page in the object with largest index smaller than the given * page index, or NULL if no such page exists. */ vm_page_t vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex, int req, vm_page_t mpred) { struct vm_domainset_iter di; vm_page_t m; int domain; vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); do { m = vm_page_alloc_domain_after(object, pindex, domain, req, mpred); if (m != NULL) break; } while (vm_domainset_iter_page(&di, object, &domain) == 0); return (m); } /* * Returns true if the number of free pages exceeds the minimum * for the request class and false otherwise. */ int vm_domain_allocate(struct vm_domain *vmd, int req, int npages) { u_int limit, old, new; req = req & VM_ALLOC_CLASS_MASK; /* * The page daemon is allowed to dig deeper into the free page list. */ if (curproc == pageproc && req != VM_ALLOC_INTERRUPT) req = VM_ALLOC_SYSTEM; if (req == VM_ALLOC_INTERRUPT) limit = 0; else if (req == VM_ALLOC_SYSTEM) limit = vmd->vmd_interrupt_free_min; else limit = vmd->vmd_free_reserved; /* * Attempt to reserve the pages. Fail if we're below the limit. */ limit += npages; old = vmd->vmd_free_count; do { if (old < limit) return (0); new = old - npages; } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0); /* Wake the page daemon if we've crossed the threshold. */ if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old)) pagedaemon_wakeup(vmd->vmd_domain); /* Only update bitsets on transitions. */ if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) || (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe)) vm_domain_set(vmd); return (1); } vm_page_t vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain, int req, vm_page_t mpred) { struct vm_domain *vmd; vm_page_t m; int flags, pool; KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), ("inconsistent object(%p)/req(%x)", object, req)); KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0, ("Can't sleep and retry object insertion.")); KASSERT(mpred == NULL || mpred->pindex < pindex, ("mpred %p doesn't precede pindex 0x%jx", mpred, (uintmax_t)pindex)); if (object != NULL) VM_OBJECT_ASSERT_WLOCKED(object); flags = 0; m = NULL; pool = object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT; again: #if VM_NRESERVLEVEL > 0 /* * Can we allocate the page from a reservation? */ if (vm_object_reserv(object) && (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) != NULL) { domain = vm_phys_domain(m); vmd = VM_DOMAIN(domain); goto found; } #endif vmd = VM_DOMAIN(domain); if (vmd->vmd_pgcache[pool].zone != NULL) { m = uma_zalloc(vmd->vmd_pgcache[pool].zone, M_NOWAIT); if (m != NULL) { flags |= PG_PCPU_CACHE; goto found; } } if (vm_domain_allocate(vmd, req, 1)) { /* * If not, allocate it from the free page queues. */ vm_domain_free_lock(vmd); m = vm_phys_alloc_pages(domain, pool, 0); vm_domain_free_unlock(vmd); if (m == NULL) { vm_domain_freecnt_inc(vmd, 1); #if VM_NRESERVLEVEL > 0 if (vm_reserv_reclaim_inactive(domain)) goto again; #endif } } if (m == NULL) { /* * Not allocatable, give up. */ if (vm_domain_alloc_fail(vmd, object, req)) goto again; return (NULL); } /* * At this point we had better have found a good page. */ found: vm_page_dequeue(m); vm_page_alloc_check(m); /* * Initialize the page. Only the PG_ZERO flag is inherited. */ if ((req & VM_ALLOC_ZERO) != 0) flags |= (m->flags & PG_ZERO); if ((req & VM_ALLOC_NODUMP) != 0) flags |= PG_NODUMP; m->flags = flags; m->aflags = 0; m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; m->busy_lock = VPB_UNBUSIED; if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) m->busy_lock = VPB_SINGLE_EXCLUSIVER; if ((req & VM_ALLOC_SBUSY) != 0) m->busy_lock = VPB_SHARERS_WORD(1); if (req & VM_ALLOC_WIRED) { /* * The page lock is not required for wiring a page until that * page is inserted into the object. */ vm_wire_add(1); m->ref_count = 1; } m->act_count = 0; if (object != NULL) { if (vm_page_insert_after(m, object, pindex, mpred)) { if (req & VM_ALLOC_WIRED) { vm_wire_sub(1); m->ref_count = 0; } KASSERT(m->object == NULL, ("page %p has object", m)); m->oflags = VPO_UNMANAGED; m->busy_lock = VPB_UNBUSIED; /* Don't change PG_ZERO. */ vm_page_free_toq(m); if (req & VM_ALLOC_WAITFAIL) { VM_OBJECT_WUNLOCK(object); vm_radix_wait(); VM_OBJECT_WLOCK(object); } return (NULL); } /* Ignore device objects; the pager sets "memattr" for them. */ if (object->memattr != VM_MEMATTR_DEFAULT && (object->flags & OBJ_FICTITIOUS) == 0) pmap_page_set_memattr(m, object->memattr); } else m->pindex = pindex; return (m); } /* * vm_page_alloc_contig: * * Allocate a contiguous set of physical pages of the given size "npages" * from the free lists. All of the physical pages must be at or above * the given physical address "low" and below the given physical address * "high". The given value "alignment" determines the alignment of the * first physical page in the set. If the given value "boundary" is * non-zero, then the set of physical pages cannot cross any physical * address boundary that is a multiple of that value. Both "alignment" * and "boundary" must be a power of two. * * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, * then the memory attribute setting for the physical pages is configured * to the object's memory attribute setting. Otherwise, the memory * attribute setting for the physical pages is configured to "memattr", * overriding the object's memory attribute setting. However, if the * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the * memory attribute setting for the physical pages cannot be configured * to VM_MEMATTR_DEFAULT. * * The specified object may not contain fictitious pages. * * The caller must always specify an allocation class. * * allocation classes: * VM_ALLOC_NORMAL normal process request * VM_ALLOC_SYSTEM system *really* needs a page * VM_ALLOC_INTERRUPT interrupt time request * * optional allocation flags: * VM_ALLOC_NOBUSY do not exclusive busy the page * VM_ALLOC_NODUMP do not include the page in a kernel core dump * VM_ALLOC_NOOBJ page is not associated with an object and * should not be exclusive busy * VM_ALLOC_SBUSY shared busy the allocated page * VM_ALLOC_WIRED wire the allocated page * VM_ALLOC_ZERO prefer a zeroed page */ vm_page_t vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr) { struct vm_domainset_iter di; vm_page_t m; int domain; vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); do { m = vm_page_alloc_contig_domain(object, pindex, domain, req, npages, low, high, alignment, boundary, memattr); if (m != NULL) break; } while (vm_domainset_iter_page(&di, object, &domain) == 0); return (m); } vm_page_t vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr) { struct vm_domain *vmd; vm_page_t m, m_ret, mpred; u_int busy_lock, flags, oflags; mpred = NULL; /* XXX: pacify gcc */ KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object, req)); KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0, ("Can't sleep and retry object insertion.")); if (object != NULL) { VM_OBJECT_ASSERT_WLOCKED(object); KASSERT((object->flags & OBJ_FICTITIOUS) == 0, ("vm_page_alloc_contig: object %p has fictitious pages", object)); } KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); if (object != NULL) { mpred = vm_radix_lookup_le(&object->rtree, pindex); KASSERT(mpred == NULL || mpred->pindex != pindex, ("vm_page_alloc_contig: pindex already allocated")); } /* * Can we allocate the pages without the number of free pages falling * below the lower bound for the allocation class? */ m_ret = NULL; again: #if VM_NRESERVLEVEL > 0 /* * Can we allocate the pages from a reservation? */ if (vm_object_reserv(object) && (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req, mpred, npages, low, high, alignment, boundary)) != NULL) { domain = vm_phys_domain(m_ret); vmd = VM_DOMAIN(domain); goto found; } #endif vmd = VM_DOMAIN(domain); if (vm_domain_allocate(vmd, req, npages)) { /* * allocate them from the free page queues. */ vm_domain_free_lock(vmd); m_ret = vm_phys_alloc_contig(domain, npages, low, high, alignment, boundary); vm_domain_free_unlock(vmd); if (m_ret == NULL) { vm_domain_freecnt_inc(vmd, npages); #if VM_NRESERVLEVEL > 0 if (vm_reserv_reclaim_contig(domain, npages, low, high, alignment, boundary)) goto again; #endif } } if (m_ret == NULL) { if (vm_domain_alloc_fail(vmd, object, req)) goto again; return (NULL); } #if VM_NRESERVLEVEL > 0 found: #endif for (m = m_ret; m < &m_ret[npages]; m++) { vm_page_dequeue(m); vm_page_alloc_check(m); } /* * Initialize the pages. Only the PG_ZERO flag is inherited. */ flags = 0; if ((req & VM_ALLOC_ZERO) != 0) flags = PG_ZERO; if ((req & VM_ALLOC_NODUMP) != 0) flags |= PG_NODUMP; oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; busy_lock = VPB_UNBUSIED; if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) busy_lock = VPB_SINGLE_EXCLUSIVER; if ((req & VM_ALLOC_SBUSY) != 0) busy_lock = VPB_SHARERS_WORD(1); if ((req & VM_ALLOC_WIRED) != 0) vm_wire_add(npages); if (object != NULL) { if (object->memattr != VM_MEMATTR_DEFAULT && memattr == VM_MEMATTR_DEFAULT) memattr = object->memattr; } for (m = m_ret; m < &m_ret[npages]; m++) { m->aflags = 0; m->flags = (m->flags | PG_NODUMP) & flags; m->busy_lock = busy_lock; if ((req & VM_ALLOC_WIRED) != 0) m->ref_count = 1; m->act_count = 0; m->oflags = oflags; if (object != NULL) { if (vm_page_insert_after(m, object, pindex, mpred)) { if ((req & VM_ALLOC_WIRED) != 0) vm_wire_sub(npages); KASSERT(m->object == NULL, ("page %p has object", m)); mpred = m; for (m = m_ret; m < &m_ret[npages]; m++) { if (m <= mpred && (req & VM_ALLOC_WIRED) != 0) m->ref_count = 0; m->oflags = VPO_UNMANAGED; m->busy_lock = VPB_UNBUSIED; /* Don't change PG_ZERO. */ vm_page_free_toq(m); } if (req & VM_ALLOC_WAITFAIL) { VM_OBJECT_WUNLOCK(object); vm_radix_wait(); VM_OBJECT_WLOCK(object); } return (NULL); } mpred = m; } else m->pindex = pindex; if (memattr != VM_MEMATTR_DEFAULT) pmap_page_set_memattr(m, memattr); pindex++; } return (m_ret); } /* * Check a page that has been freshly dequeued from a freelist. */ static void vm_page_alloc_check(vm_page_t m) { KASSERT(m->object == NULL, ("page %p has object", m)); KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0, ("page %p has unexpected queue %d, flags %#x", m, m->queue, (m->aflags & PGA_QUEUE_STATE_MASK))); KASSERT(m->ref_count == 0, ("page %p has references", m)); KASSERT(!vm_page_busied(m), ("page %p is busy", m)); KASSERT(m->dirty == 0, ("page %p is dirty", m)); KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, ("page %p has unexpected memattr %d", m, pmap_page_get_memattr(m))); KASSERT(m->valid == 0, ("free page %p is valid", m)); } /* * vm_page_alloc_freelist: * * Allocate a physical page from the specified free page list. * * The caller must always specify an allocation class. * * allocation classes: * VM_ALLOC_NORMAL normal process request * VM_ALLOC_SYSTEM system *really* needs a page * VM_ALLOC_INTERRUPT interrupt time request * * optional allocation flags: * VM_ALLOC_COUNT(number) the number of additional pages that the caller * intends to allocate * VM_ALLOC_WIRED wire the allocated page * VM_ALLOC_ZERO prefer a zeroed page */ vm_page_t vm_page_alloc_freelist(int freelist, int req) { struct vm_domainset_iter di; vm_page_t m; int domain; vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); do { m = vm_page_alloc_freelist_domain(domain, freelist, req); if (m != NULL) break; } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); return (m); } vm_page_t vm_page_alloc_freelist_domain(int domain, int freelist, int req) { struct vm_domain *vmd; vm_page_t m; u_int flags; m = NULL; vmd = VM_DOMAIN(domain); again: if (vm_domain_allocate(vmd, req, 1)) { vm_domain_free_lock(vmd); m = vm_phys_alloc_freelist_pages(domain, freelist, VM_FREEPOOL_DIRECT, 0); vm_domain_free_unlock(vmd); if (m == NULL) vm_domain_freecnt_inc(vmd, 1); } if (m == NULL) { if (vm_domain_alloc_fail(vmd, NULL, req)) goto again; return (NULL); } vm_page_dequeue(m); vm_page_alloc_check(m); /* * Initialize the page. Only the PG_ZERO flag is inherited. */ m->aflags = 0; flags = 0; if ((req & VM_ALLOC_ZERO) != 0) flags = PG_ZERO; m->flags &= flags; if ((req & VM_ALLOC_WIRED) != 0) { /* * The page lock is not required for wiring a page that does * not belong to an object. */ vm_wire_add(1); m->ref_count = 1; } /* Unmanaged pages don't use "act_count". */ m->oflags = VPO_UNMANAGED; return (m); } static int vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags) { struct vm_domain *vmd; struct vm_pgcache *pgcache; int i; pgcache = arg; vmd = VM_DOMAIN(pgcache->domain); /* Only import if we can bring in a full bucket. */ if (cnt == 1 || !vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt)) return (0); domain = vmd->vmd_domain; vm_domain_free_lock(vmd); i = vm_phys_alloc_npages(domain, pgcache->pool, cnt, (vm_page_t *)store); vm_domain_free_unlock(vmd); if (cnt != i) vm_domain_freecnt_inc(vmd, cnt - i); return (i); } static void vm_page_zone_release(void *arg, void **store, int cnt) { struct vm_domain *vmd; struct vm_pgcache *pgcache; vm_page_t m; int i; pgcache = arg; vmd = VM_DOMAIN(pgcache->domain); vm_domain_free_lock(vmd); for (i = 0; i < cnt; i++) { m = (vm_page_t)store[i]; vm_phys_free_pages(m, 0); } vm_domain_free_unlock(vmd); vm_domain_freecnt_inc(vmd, cnt); } #define VPSC_ANY 0 /* No restrictions. */ #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */ #define VPSC_NOSUPER 2 /* Skip superpages. */ /* * vm_page_scan_contig: * * Scan vm_page_array[] between the specified entries "m_start" and * "m_end" for a run of contiguous physical pages that satisfy the * specified conditions, and return the lowest page in the run. The * specified "alignment" determines the alignment of the lowest physical * page in the run. If the specified "boundary" is non-zero, then the * run of physical pages cannot span a physical address that is a * multiple of "boundary". * * "m_end" is never dereferenced, so it need not point to a vm_page * structure within vm_page_array[]. * * "npages" must be greater than zero. "m_start" and "m_end" must not * span a hole (or discontiguity) in the physical address space. Both * "alignment" and "boundary" must be a power of two. */ vm_page_t vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, u_long alignment, vm_paddr_t boundary, int options) { struct mtx *m_mtx; vm_object_t object; vm_paddr_t pa; vm_page_t m, m_run; #if VM_NRESERVLEVEL > 0 int level; #endif int m_inc, order, run_ext, run_len; KASSERT(npages > 0, ("npages is 0")); KASSERT(powerof2(alignment), ("alignment is not a power of 2")); KASSERT(powerof2(boundary), ("boundary is not a power of 2")); m_run = NULL; run_len = 0; m_mtx = NULL; for (m = m_start; m < m_end && run_len < npages; m += m_inc) { KASSERT((m->flags & PG_MARKER) == 0, ("page %p is PG_MARKER", m)); KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1, ("fictitious page %p has invalid ref count", m)); /* * If the current page would be the start of a run, check its * physical address against the end, alignment, and boundary * conditions. If it doesn't satisfy these conditions, either * terminate the scan or advance to the next page that * satisfies the failed condition. */ if (run_len == 0) { KASSERT(m_run == NULL, ("m_run != NULL")); if (m + npages > m_end) break; pa = VM_PAGE_TO_PHYS(m); if ((pa & (alignment - 1)) != 0) { m_inc = atop(roundup2(pa, alignment) - pa); continue; } if (rounddown2(pa ^ (pa + ptoa(npages) - 1), boundary) != 0) { m_inc = atop(roundup2(pa, boundary) - pa); continue; } } else KASSERT(m_run != NULL, ("m_run == NULL")); vm_page_change_lock(m, &m_mtx); m_inc = 1; retry: if (vm_page_wired(m)) run_ext = 0; #if VM_NRESERVLEVEL > 0 else if ((level = vm_reserv_level(m)) >= 0 && (options & VPSC_NORESERV) != 0) { run_ext = 0; /* Advance to the end of the reservation. */ pa = VM_PAGE_TO_PHYS(m); m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - pa); } #endif else if ((object = m->object) != NULL) { /* * The page is considered eligible for relocation if * and only if it could be laundered or reclaimed by * the page daemon. */ if (!VM_OBJECT_TRYRLOCK(object)) { mtx_unlock(m_mtx); VM_OBJECT_RLOCK(object); mtx_lock(m_mtx); if (m->object != object) { /* * The page may have been freed. */ VM_OBJECT_RUNLOCK(object); goto retry; } } /* Don't care: PG_NODUMP, PG_ZERO. */ if (object->type != OBJT_DEFAULT && object->type != OBJT_SWAP && object->type != OBJT_VNODE) { run_ext = 0; #if VM_NRESERVLEVEL > 0 } else if ((options & VPSC_NOSUPER) != 0 && (level = vm_reserv_level_iffullpop(m)) >= 0) { run_ext = 0; /* Advance to the end of the superpage. */ pa = VM_PAGE_TO_PHYS(m); m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - pa); #endif } else if (object->memattr == VM_MEMATTR_DEFAULT && vm_page_queue(m) != PQ_NONE && !vm_page_busied(m) && !vm_page_wired(m)) { /* * The page is allocated but eligible for * relocation. Extend the current run by one * page. */ KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, ("page %p has an unexpected memattr", m)); KASSERT((m->oflags & (VPO_SWAPINPROG | VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, ("page %p has unexpected oflags", m)); /* Don't care: VPO_NOSYNC. */ run_ext = 1; } else run_ext = 0; VM_OBJECT_RUNLOCK(object); #if VM_NRESERVLEVEL > 0 } else if (level >= 0) { /* * The page is reserved but not yet allocated. In * other words, it is still free. Extend the current * run by one page. */ run_ext = 1; #endif } else if ((order = m->order) < VM_NFREEORDER) { /* * The page is enqueued in the physical memory * allocator's free page queues. Moreover, it is the * first page in a power-of-two-sized run of * contiguous free pages. Add these pages to the end * of the current run, and jump ahead. */ run_ext = 1 << order; m_inc = 1 << order; } else { /* * Skip the page for one of the following reasons: (1) * It is enqueued in the physical memory allocator's * free page queues. However, it is not the first * page in a run of contiguous free pages. (This case * rarely occurs because the scan is performed in * ascending order.) (2) It is not reserved, and it is * transitioning from free to allocated. (Conversely, * the transition from allocated to free for managed * pages is blocked by the page lock.) (3) It is * allocated but not contained by an object and not * wired, e.g., allocated by Xen's balloon driver. */ run_ext = 0; } /* * Extend or reset the current run of pages. */ if (run_ext > 0) { if (run_len == 0) m_run = m; run_len += run_ext; } else { if (run_len > 0) { m_run = NULL; run_len = 0; } } } if (m_mtx != NULL) mtx_unlock(m_mtx); if (run_len >= npages) return (m_run); return (NULL); } /* * vm_page_reclaim_run: * * Try to relocate each of the allocated virtual pages within the * specified run of physical pages to a new physical address. Free the * physical pages underlying the relocated virtual pages. A virtual page * is relocatable if and only if it could be laundered or reclaimed by * the page daemon. Whenever possible, a virtual page is relocated to a * physical address above "high". * * Returns 0 if every physical page within the run was already free or * just freed by a successful relocation. Otherwise, returns a non-zero * value indicating why the last attempt to relocate a virtual page was * unsuccessful. * * "req_class" must be an allocation class. */ static int vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run, vm_paddr_t high) { struct vm_domain *vmd; struct mtx *m_mtx; struct spglist free; vm_object_t object; vm_paddr_t pa; vm_page_t m, m_end, m_new; int error, order, req; KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class, ("req_class is not an allocation class")); SLIST_INIT(&free); error = 0; m = m_run; m_end = m_run + npages; m_mtx = NULL; for (; error == 0 && m < m_end; m++) { KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, ("page %p is PG_FICTITIOUS or PG_MARKER", m)); /* * Avoid releasing and reacquiring the same page lock. */ vm_page_change_lock(m, &m_mtx); retry: /* * Racily check for wirings. Races are handled below. */ if (vm_page_wired(m)) error = EBUSY; else if ((object = m->object) != NULL) { /* * The page is relocated if and only if it could be * laundered or reclaimed by the page daemon. */ if (!VM_OBJECT_TRYWLOCK(object)) { mtx_unlock(m_mtx); VM_OBJECT_WLOCK(object); mtx_lock(m_mtx); if (m->object != object) { /* * The page may have been freed. */ VM_OBJECT_WUNLOCK(object); goto retry; } } /* Don't care: PG_NODUMP, PG_ZERO. */ if (object->type != OBJT_DEFAULT && object->type != OBJT_SWAP && object->type != OBJT_VNODE) error = EINVAL; else if (object->memattr != VM_MEMATTR_DEFAULT) error = EINVAL; else if (vm_page_queue(m) != PQ_NONE && - !vm_page_busied(m) && !vm_page_wired(m)) { + vm_page_tryxbusy(m) != 0) { + if (vm_page_wired(m)) { + vm_page_xunbusy(m); + error = EBUSY; + goto unlock; + } KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, ("page %p has an unexpected memattr", m)); KASSERT((m->oflags & (VPO_SWAPINPROG | VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, ("page %p has unexpected oflags", m)); /* Don't care: VPO_NOSYNC. */ if (m->valid != 0) { /* * First, try to allocate a new page * that is above "high". Failing * that, try to allocate a new page * that is below "m_run". Allocate * the new page between the end of * "m_run" and "high" only as a last * resort. */ req = req_class | VM_ALLOC_NOOBJ; if ((m->flags & PG_NODUMP) != 0) req |= VM_ALLOC_NODUMP; if (trunc_page(high) != ~(vm_paddr_t)PAGE_MASK) { m_new = vm_page_alloc_contig( NULL, 0, req, 1, round_page(high), ~(vm_paddr_t)0, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT); } else m_new = NULL; if (m_new == NULL) { pa = VM_PAGE_TO_PHYS(m_run); m_new = vm_page_alloc_contig( NULL, 0, req, 1, 0, pa - 1, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT); } if (m_new == NULL) { pa += ptoa(npages); m_new = vm_page_alloc_contig( NULL, 0, req, 1, pa, high, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT); } if (m_new == NULL) { + vm_page_xunbusy(m); error = ENOMEM; goto unlock; } /* * Unmap the page and check for new * wirings that may have been acquired * through a pmap lookup. */ if (object->ref_count != 0 && !vm_page_try_remove_all(m)) { vm_page_free(m_new); error = EBUSY; goto unlock; } /* * Replace "m" with the new page. For * vm_page_replace(), "m" must be busy * and dequeued. Finally, change "m" * as if vm_page_free() was called. */ m_new->aflags = m->aflags & ~PGA_QUEUE_STATE_MASK; KASSERT(m_new->oflags == VPO_UNMANAGED, ("page %p is managed", m_new)); m_new->oflags = m->oflags & VPO_NOSYNC; pmap_copy_page(m, m_new); m_new->valid = m->valid; m_new->dirty = m->dirty; m->flags &= ~PG_ZERO; - vm_page_xbusy(m); vm_page_dequeue(m); vm_page_replace_checked(m_new, object, m->pindex, m); if (vm_page_free_prep(m)) SLIST_INSERT_HEAD(&free, m, plinks.s.ss); /* * The new page must be deactivated * before the object is unlocked. */ vm_page_change_lock(m_new, &m_mtx); vm_page_deactivate(m_new); } else { m->flags &= ~PG_ZERO; vm_page_dequeue(m); if (vm_page_free_prep(m)) SLIST_INSERT_HEAD(&free, m, plinks.s.ss); KASSERT(m->dirty == 0, ("page %p is dirty", m)); } } else error = EBUSY; unlock: VM_OBJECT_WUNLOCK(object); } else { MPASS(vm_phys_domain(m) == domain); vmd = VM_DOMAIN(domain); vm_domain_free_lock(vmd); order = m->order; if (order < VM_NFREEORDER) { /* * The page is enqueued in the physical memory * allocator's free page queues. Moreover, it * is the first page in a power-of-two-sized * run of contiguous free pages. Jump ahead * to the last page within that run, and * continue from there. */ m += (1 << order) - 1; } #if VM_NRESERVLEVEL > 0 else if (vm_reserv_is_page_free(m)) order = 0; #endif vm_domain_free_unlock(vmd); if (order == VM_NFREEORDER) error = EINVAL; } } if (m_mtx != NULL) mtx_unlock(m_mtx); if ((m = SLIST_FIRST(&free)) != NULL) { int cnt; vmd = VM_DOMAIN(domain); cnt = 0; vm_domain_free_lock(vmd); do { MPASS(vm_phys_domain(m) == domain); SLIST_REMOVE_HEAD(&free, plinks.s.ss); vm_phys_free_pages(m, 0); cnt++; } while ((m = SLIST_FIRST(&free)) != NULL); vm_domain_free_unlock(vmd); vm_domain_freecnt_inc(vmd, cnt); } return (error); } #define NRUNS 16 CTASSERT(powerof2(NRUNS)); #define RUN_INDEX(count) ((count) & (NRUNS - 1)) #define MIN_RECLAIM 8 /* * vm_page_reclaim_contig: * * Reclaim allocated, contiguous physical memory satisfying the specified * conditions by relocating the virtual pages using that physical memory. * Returns true if reclamation is successful and false otherwise. Since * relocation requires the allocation of physical pages, reclamation may * fail due to a shortage of free pages. When reclamation fails, callers * are expected to perform vm_wait() before retrying a failed allocation * operation, e.g., vm_page_alloc_contig(). * * The caller must always specify an allocation class through "req". * * allocation classes: * VM_ALLOC_NORMAL normal process request * VM_ALLOC_SYSTEM system *really* needs a page * VM_ALLOC_INTERRUPT interrupt time request * * The optional allocation flags are ignored. * * "npages" must be greater than zero. Both "alignment" and "boundary" * must be a power of two. */ bool vm_page_reclaim_contig_domain(int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) { struct vm_domain *vmd; vm_paddr_t curr_low; vm_page_t m_run, m_runs[NRUNS]; u_long count, reclaimed; int error, i, options, req_class; KASSERT(npages > 0, ("npages is 0")); KASSERT(powerof2(alignment), ("alignment is not a power of 2")); KASSERT(powerof2(boundary), ("boundary is not a power of 2")); req_class = req & VM_ALLOC_CLASS_MASK; /* * The page daemon is allowed to dig deeper into the free page list. */ if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) req_class = VM_ALLOC_SYSTEM; /* * Return if the number of free pages cannot satisfy the requested * allocation. */ vmd = VM_DOMAIN(domain); count = vmd->vmd_free_count; if (count < npages + vmd->vmd_free_reserved || (count < npages + vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) || (count < npages && req_class == VM_ALLOC_INTERRUPT)) return (false); /* * Scan up to three times, relaxing the restrictions ("options") on * the reclamation of reservations and superpages each time. */ for (options = VPSC_NORESERV;;) { /* * Find the highest runs that satisfy the given constraints * and restrictions, and record them in "m_runs". */ curr_low = low; count = 0; for (;;) { m_run = vm_phys_scan_contig(domain, npages, curr_low, high, alignment, boundary, options); if (m_run == NULL) break; curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages); m_runs[RUN_INDEX(count)] = m_run; count++; } /* * Reclaim the highest runs in LIFO (descending) order until * the number of reclaimed pages, "reclaimed", is at least * MIN_RECLAIM. Reset "reclaimed" each time because each * reclamation is idempotent, and runs will (likely) recur * from one scan to the next as restrictions are relaxed. */ reclaimed = 0; for (i = 0; count > 0 && i < NRUNS; i++) { count--; m_run = m_runs[RUN_INDEX(count)]; error = vm_page_reclaim_run(req_class, domain, npages, m_run, high); if (error == 0) { reclaimed += npages; if (reclaimed >= MIN_RECLAIM) return (true); } } /* * Either relax the restrictions on the next scan or return if * the last scan had no restrictions. */ if (options == VPSC_NORESERV) options = VPSC_NOSUPER; else if (options == VPSC_NOSUPER) options = VPSC_ANY; else if (options == VPSC_ANY) return (reclaimed != 0); } } bool vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) { struct vm_domainset_iter di; int domain; bool ret; vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); do { ret = vm_page_reclaim_contig_domain(domain, req, npages, low, high, alignment, boundary); if (ret) break; } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); return (ret); } /* * Set the domain in the appropriate page level domainset. */ void vm_domain_set(struct vm_domain *vmd) { mtx_lock(&vm_domainset_lock); if (!vmd->vmd_minset && vm_paging_min(vmd)) { vmd->vmd_minset = 1; DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains); } if (!vmd->vmd_severeset && vm_paging_severe(vmd)) { vmd->vmd_severeset = 1; DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains); } mtx_unlock(&vm_domainset_lock); } /* * Clear the domain from the appropriate page level domainset. */ void vm_domain_clear(struct vm_domain *vmd) { mtx_lock(&vm_domainset_lock); if (vmd->vmd_minset && !vm_paging_min(vmd)) { vmd->vmd_minset = 0; DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains); if (vm_min_waiters != 0) { vm_min_waiters = 0; wakeup(&vm_min_domains); } } if (vmd->vmd_severeset && !vm_paging_severe(vmd)) { vmd->vmd_severeset = 0; DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains); if (vm_severe_waiters != 0) { vm_severe_waiters = 0; wakeup(&vm_severe_domains); } } /* * If pageout daemon needs pages, then tell it that there are * some free. */ if (vmd->vmd_pageout_pages_needed && vmd->vmd_free_count >= vmd->vmd_pageout_free_min) { wakeup(&vmd->vmd_pageout_pages_needed); vmd->vmd_pageout_pages_needed = 0; } /* See comments in vm_wait_doms(). */ if (vm_pageproc_waiters) { vm_pageproc_waiters = 0; wakeup(&vm_pageproc_waiters); } mtx_unlock(&vm_domainset_lock); } /* * Wait for free pages to exceed the min threshold globally. */ void vm_wait_min(void) { mtx_lock(&vm_domainset_lock); while (vm_page_count_min()) { vm_min_waiters++; msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0); } mtx_unlock(&vm_domainset_lock); } /* * Wait for free pages to exceed the severe threshold globally. */ void vm_wait_severe(void) { mtx_lock(&vm_domainset_lock); while (vm_page_count_severe()) { vm_severe_waiters++; msleep(&vm_severe_domains, &vm_domainset_lock, PVM, "vmwait", 0); } mtx_unlock(&vm_domainset_lock); } u_int vm_wait_count(void) { return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters); } void vm_wait_doms(const domainset_t *wdoms) { /* * We use racey wakeup synchronization to avoid expensive global * locking for the pageproc when sleeping with a non-specific vm_wait. * To handle this, we only sleep for one tick in this instance. It * is expected that most allocations for the pageproc will come from * kmem or vm_page_grab* which will use the more specific and * race-free vm_wait_domain(). */ if (curproc == pageproc) { mtx_lock(&vm_domainset_lock); vm_pageproc_waiters++; msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP, "pageprocwait", 1); } else { /* * XXX Ideally we would wait only until the allocation could * be satisfied. This condition can cause new allocators to * consume all freed pages while old allocators wait. */ mtx_lock(&vm_domainset_lock); if (vm_page_count_min_set(wdoms)) { vm_min_waiters++; msleep(&vm_min_domains, &vm_domainset_lock, PVM | PDROP, "vmwait", 0); } else mtx_unlock(&vm_domainset_lock); } } /* * vm_wait_domain: * * Sleep until free pages are available for allocation. * - Called in various places after failed memory allocations. */ void vm_wait_domain(int domain) { struct vm_domain *vmd; domainset_t wdom; vmd = VM_DOMAIN(domain); vm_domain_free_assert_unlocked(vmd); if (curproc == pageproc) { mtx_lock(&vm_domainset_lock); if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) { vmd->vmd_pageout_pages_needed = 1; msleep(&vmd->vmd_pageout_pages_needed, &vm_domainset_lock, PDROP | PSWP, "VMWait", 0); } else mtx_unlock(&vm_domainset_lock); } else { if (pageproc == NULL) panic("vm_wait in early boot"); DOMAINSET_ZERO(&wdom); DOMAINSET_SET(vmd->vmd_domain, &wdom); vm_wait_doms(&wdom); } } /* * vm_wait: * * Sleep until free pages are available for allocation in the * affinity domains of the obj. If obj is NULL, the domain set * for the calling thread is used. * Called in various places after failed memory allocations. */ void vm_wait(vm_object_t obj) { struct domainset *d; d = NULL; /* * Carefully fetch pointers only once: the struct domainset * itself is ummutable but the pointer might change. */ if (obj != NULL) d = obj->domain.dr_policy; if (d == NULL) d = curthread->td_domain.dr_policy; vm_wait_doms(&d->ds_mask); } /* * vm_domain_alloc_fail: * * Called when a page allocation function fails. Informs the * pagedaemon and performs the requested wait. Requires the * domain_free and object lock on entry. Returns with the * object lock held and free lock released. Returns an error when * retry is necessary. * */ static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req) { vm_domain_free_assert_unlocked(vmd); atomic_add_int(&vmd->vmd_pageout_deficit, max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) { if (object != NULL) VM_OBJECT_WUNLOCK(object); vm_wait_domain(vmd->vmd_domain); if (object != NULL) VM_OBJECT_WLOCK(object); if (req & VM_ALLOC_WAITOK) return (EAGAIN); } return (0); } /* * vm_waitpfault: * * Sleep until free pages are available for allocation. * - Called only in vm_fault so that processes page faulting * can be easily tracked. * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing * processes will be able to grab memory first. Do not change * this balance without careful testing first. */ void vm_waitpfault(struct domainset *dset, int timo) { /* * XXX Ideally we would wait only until the allocation could * be satisfied. This condition can cause new allocators to * consume all freed pages while old allocators wait. */ mtx_lock(&vm_domainset_lock); if (vm_page_count_min_set(&dset->ds_mask)) { vm_min_waiters++; msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP, "pfault", timo); } else mtx_unlock(&vm_domainset_lock); } static struct vm_pagequeue * vm_page_pagequeue(vm_page_t m) { uint8_t queue; if ((queue = atomic_load_8(&m->queue)) == PQ_NONE) return (NULL); return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]); } static inline void vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m) { struct vm_domain *vmd; uint8_t qflags; CRITICAL_ASSERT(curthread); vm_pagequeue_assert_locked(pq); /* * The page daemon is allowed to set m->queue = PQ_NONE without * the page queue lock held. In this case it is about to free the page, * which must not have any queue state. */ qflags = atomic_load_8(&m->aflags); KASSERT(pq == vm_page_pagequeue(m) || (qflags & PGA_QUEUE_STATE_MASK) == 0, ("page %p doesn't belong to queue %p but has aflags %#x", m, pq, qflags)); if ((qflags & PGA_DEQUEUE) != 0) { if (__predict_true((qflags & PGA_ENQUEUED) != 0)) vm_pagequeue_remove(pq, m); vm_page_dequeue_complete(m); counter_u64_add(queue_ops, 1); } else if ((qflags & (PGA_REQUEUE | PGA_REQUEUE_HEAD)) != 0) { if ((qflags & PGA_ENQUEUED) != 0) TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); else { vm_pagequeue_cnt_inc(pq); vm_page_aflag_set(m, PGA_ENQUEUED); } /* * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE. * In particular, if both flags are set in close succession, * only PGA_REQUEUE_HEAD will be applied, even if it was set * first. */ if ((qflags & PGA_REQUEUE_HEAD) != 0) { KASSERT(m->queue == PQ_INACTIVE, ("head enqueue not supported for page %p", m)); vmd = vm_pagequeue_domain(m); TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q); } else TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); vm_page_aflag_clear(m, qflags & (PGA_REQUEUE | PGA_REQUEUE_HEAD)); counter_u64_add(queue_ops, 1); } else { counter_u64_add(queue_nops, 1); } } static void vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq, uint8_t queue) { vm_page_t m; int i; for (i = 0; i < bq->bq_cnt; i++) { m = bq->bq_pa[i]; if (__predict_false(m->queue != queue)) continue; vm_pqbatch_process_page(pq, m); } vm_batchqueue_init(bq); } /* * vm_page_pqbatch_submit: [ internal use only ] * * Enqueue a page in the specified page queue's batched work queue. * The caller must have encoded the requested operation in the page * structure's aflags field. */ void vm_page_pqbatch_submit(vm_page_t m, uint8_t queue) { struct vm_batchqueue *bq; struct vm_pagequeue *pq; int domain; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("page %p is unmanaged", m)); KASSERT(mtx_owned(vm_page_lockptr(m)) || m->object == NULL, ("missing synchronization for page %p", m)); KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue)); domain = vm_phys_domain(m); pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue]; critical_enter(); bq = DPCPU_PTR(pqbatch[domain][queue]); if (vm_batchqueue_insert(bq, m)) { critical_exit(); return; } critical_exit(); vm_pagequeue_lock(pq); critical_enter(); bq = DPCPU_PTR(pqbatch[domain][queue]); vm_pqbatch_process(pq, bq, queue); /* * The page may have been logically dequeued before we acquired the * page queue lock. In this case, since we either hold the page lock * or the page is being freed, a different thread cannot be concurrently * enqueuing the page. */ if (__predict_true(m->queue == queue)) vm_pqbatch_process_page(pq, m); else { KASSERT(m->queue == PQ_NONE, ("invalid queue transition for page %p", m)); KASSERT((m->aflags & PGA_ENQUEUED) == 0, ("page %p is enqueued with invalid queue index", m)); } vm_pagequeue_unlock(pq); critical_exit(); } /* * vm_page_pqbatch_drain: [ internal use only ] * * Force all per-CPU page queue batch queues to be drained. This is * intended for use in severe memory shortages, to ensure that pages * do not remain stuck in the batch queues. */ void vm_page_pqbatch_drain(void) { struct thread *td; struct vm_domain *vmd; struct vm_pagequeue *pq; int cpu, domain, queue; td = curthread; CPU_FOREACH(cpu) { thread_lock(td); sched_bind(td, cpu); thread_unlock(td); for (domain = 0; domain < vm_ndomains; domain++) { vmd = VM_DOMAIN(domain); for (queue = 0; queue < PQ_COUNT; queue++) { pq = &vmd->vmd_pagequeues[queue]; vm_pagequeue_lock(pq); critical_enter(); vm_pqbatch_process(pq, DPCPU_PTR(pqbatch[domain][queue]), queue); critical_exit(); vm_pagequeue_unlock(pq); } } } thread_lock(td); sched_unbind(td); thread_unlock(td); } /* * Complete the logical removal of a page from a page queue. We must be * careful to synchronize with the page daemon, which may be concurrently * examining the page with only the page lock held. The page must not be * in a state where it appears to be logically enqueued. */ static void vm_page_dequeue_complete(vm_page_t m) { m->queue = PQ_NONE; atomic_thread_fence_rel(); vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK); } /* * vm_page_dequeue_deferred: [ internal use only ] * * Request removal of the given page from its current page * queue. Physical removal from the queue may be deferred * indefinitely. * * The page must be locked. */ void vm_page_dequeue_deferred(vm_page_t m) { uint8_t queue; vm_page_assert_locked(m); if ((queue = vm_page_queue(m)) == PQ_NONE) return; /* * Set PGA_DEQUEUE if it is not already set to handle a concurrent call * to vm_page_dequeue_deferred_free(). In particular, avoid modifying * the page's queue state once vm_page_dequeue_deferred_free() has been * called. In the event of a race, two batch queue entries for the page * will be created, but the second will have no effect. */ if (vm_page_pqstate_cmpset(m, queue, queue, PGA_DEQUEUE, PGA_DEQUEUE)) vm_page_pqbatch_submit(m, queue); } /* * A variant of vm_page_dequeue_deferred() that does not assert the page * lock and is only to be called from vm_page_free_prep(). Because the * page is being freed, we can assume that nothing other than the page * daemon is scheduling queue operations on this page, so we get for * free the mutual exclusion that is otherwise provided by the page lock. * To handle races, the page daemon must take care to atomically check * for PGA_DEQUEUE when updating queue state. */ static void vm_page_dequeue_deferred_free(vm_page_t m) { uint8_t queue; KASSERT(m->ref_count == 0, ("page %p has references", m)); for (;;) { if ((m->aflags & PGA_DEQUEUE) != 0) return; atomic_thread_fence_acq(); if ((queue = atomic_load_8(&m->queue)) == PQ_NONE) return; if (vm_page_pqstate_cmpset(m, queue, queue, PGA_DEQUEUE, PGA_DEQUEUE)) { vm_page_pqbatch_submit(m, queue); break; } } } /* * vm_page_dequeue: * * Remove the page from whichever page queue it's in, if any. * The page must either be locked or unallocated. This constraint * ensures that the queue state of the page will remain consistent * after this function returns. */ void vm_page_dequeue(vm_page_t m) { struct vm_pagequeue *pq, *pq1; uint8_t aflags; KASSERT(mtx_owned(vm_page_lockptr(m)) || m->object == NULL, ("page %p is allocated and unlocked", m)); for (pq = vm_page_pagequeue(m);; pq = pq1) { if (pq == NULL) { /* * A thread may be concurrently executing * vm_page_dequeue_complete(). Ensure that all queue * state is cleared before we return. */ aflags = atomic_load_8(&m->aflags); if ((aflags & PGA_QUEUE_STATE_MASK) == 0) return; KASSERT((aflags & PGA_DEQUEUE) != 0, ("page %p has unexpected queue state flags %#x", m, aflags)); /* * Busy wait until the thread updating queue state is * finished. Such a thread must be executing in a * critical section. */ cpu_spinwait(); pq1 = vm_page_pagequeue(m); continue; } vm_pagequeue_lock(pq); if ((pq1 = vm_page_pagequeue(m)) == pq) break; vm_pagequeue_unlock(pq); } KASSERT(pq == vm_page_pagequeue(m), ("%s: page %p migrated directly between queues", __func__, m)); KASSERT((m->aflags & PGA_DEQUEUE) != 0 || mtx_owned(vm_page_lockptr(m)), ("%s: queued unlocked page %p", __func__, m)); if ((m->aflags & PGA_ENQUEUED) != 0) vm_pagequeue_remove(pq, m); vm_page_dequeue_complete(m); vm_pagequeue_unlock(pq); } /* * Schedule the given page for insertion into the specified page queue. * Physical insertion of the page may be deferred indefinitely. */ static void vm_page_enqueue(vm_page_t m, uint8_t queue) { vm_page_assert_locked(m); KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0, ("%s: page %p is already enqueued", __func__, m)); m->queue = queue; if ((m->aflags & PGA_REQUEUE) == 0) vm_page_aflag_set(m, PGA_REQUEUE); vm_page_pqbatch_submit(m, queue); } /* * vm_page_requeue: [ internal use only ] * * Schedule a requeue of the given page. * * The page must be locked. */ void vm_page_requeue(vm_page_t m) { vm_page_assert_locked(m); KASSERT(vm_page_queue(m) != PQ_NONE, ("%s: page %p is not logically enqueued", __func__, m)); if ((m->aflags & PGA_REQUEUE) == 0) vm_page_aflag_set(m, PGA_REQUEUE); vm_page_pqbatch_submit(m, atomic_load_8(&m->queue)); } /* * vm_page_swapqueue: [ internal use only ] * * Move the page from one queue to another, or to the tail of its * current queue, in the face of a possible concurrent call to * vm_page_dequeue_deferred_free(). */ void vm_page_swapqueue(vm_page_t m, uint8_t oldq, uint8_t newq) { struct vm_pagequeue *pq; vm_page_t next; bool queued; KASSERT(oldq < PQ_COUNT && newq < PQ_COUNT && oldq != newq, ("vm_page_swapqueue: invalid queues (%d, %d)", oldq, newq)); vm_page_assert_locked(m); pq = &vm_pagequeue_domain(m)->vmd_pagequeues[oldq]; vm_pagequeue_lock(pq); /* * The physical queue state might change at any point before the page * queue lock is acquired, so we must verify that we hold the correct * lock before proceeding. */ if (__predict_false(m->queue != oldq)) { vm_pagequeue_unlock(pq); return; } /* * Once the queue index of the page changes, there is nothing * synchronizing with further updates to the physical queue state. * Therefore we must remove the page from the queue now in anticipation * of a successful commit, and be prepared to roll back. */ if (__predict_true((m->aflags & PGA_ENQUEUED) != 0)) { next = TAILQ_NEXT(m, plinks.q); TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); vm_page_aflag_clear(m, PGA_ENQUEUED); queued = true; } else { queued = false; } /* * Atomically update the queue field and set PGA_REQUEUE while * ensuring that PGA_DEQUEUE has not been set. */ if (__predict_false(!vm_page_pqstate_cmpset(m, oldq, newq, PGA_DEQUEUE, PGA_REQUEUE))) { if (queued) { vm_page_aflag_set(m, PGA_ENQUEUED); if (next != NULL) TAILQ_INSERT_BEFORE(next, m, plinks.q); else TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); } vm_pagequeue_unlock(pq); return; } vm_pagequeue_cnt_dec(pq); vm_pagequeue_unlock(pq); vm_page_pqbatch_submit(m, newq); } /* * vm_page_free_prep: * * Prepares the given page to be put on the free list, * disassociating it from any VM object. The caller may return * the page to the free list only if this function returns true. * * The object must be locked. The page must be locked if it is * managed. */ bool vm_page_free_prep(vm_page_t m) { /* * Synchronize with threads that have dropped a reference to this * page. */ atomic_thread_fence_acq(); #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP) if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) { uint64_t *p; int i; p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++) KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx", m, i, (uintmax_t)*p)); } #endif if ((m->oflags & VPO_UNMANAGED) == 0) { KASSERT(!pmap_page_is_mapped(m), ("vm_page_free_prep: freeing mapped page %p", m)); KASSERT((m->aflags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0, ("vm_page_free_prep: mapping flags set in page %p", m)); } else { KASSERT(m->queue == PQ_NONE, ("vm_page_free_prep: unmanaged page %p is queued", m)); } VM_CNT_INC(v_tfree); if (vm_page_sbusied(m)) panic("vm_page_free_prep: freeing busy page %p", m); if (m->object != NULL) { vm_page_object_remove(m); /* * The object reference can be released without an atomic * operation. */ KASSERT((m->flags & PG_FICTITIOUS) != 0 || m->ref_count == VPRC_OBJREF, ("vm_page_free_prep: page %p has unexpected ref_count %u", m, m->ref_count)); m->object = NULL; m->ref_count -= VPRC_OBJREF; } /* * If fictitious remove object association and * return. */ if ((m->flags & PG_FICTITIOUS) != 0) { KASSERT(m->ref_count == 1, ("fictitious page %p is referenced", m)); KASSERT(m->queue == PQ_NONE, ("fictitious page %p is queued", m)); return (false); } /* * Pages need not be dequeued before they are returned to the physical * memory allocator, but they must at least be marked for a deferred * dequeue. */ if ((m->oflags & VPO_UNMANAGED) == 0) vm_page_dequeue_deferred_free(m); m->valid = 0; vm_page_undirty(m); if (m->ref_count != 0) panic("vm_page_free_prep: page %p has references", m); /* * Restore the default memory attribute to the page. */ if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); #if VM_NRESERVLEVEL > 0 /* * Determine whether the page belongs to a reservation. If the page was * allocated from a per-CPU cache, it cannot belong to a reservation, so * as an optimization, we avoid the check in that case. */ if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m)) return (false); #endif return (true); } /* * vm_page_free_toq: * * Returns the given page to the free list, disassociating it * from any VM object. * * The object must be locked. The page must be locked if it is * managed. */ void vm_page_free_toq(vm_page_t m) { struct vm_domain *vmd; uma_zone_t zone; if (!vm_page_free_prep(m)) return; vmd = vm_pagequeue_domain(m); zone = vmd->vmd_pgcache[m->pool].zone; if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) { uma_zfree(zone, m); return; } vm_domain_free_lock(vmd); vm_phys_free_pages(m, 0); vm_domain_free_unlock(vmd); vm_domain_freecnt_inc(vmd, 1); } /* * vm_page_free_pages_toq: * * Returns a list of pages to the free list, disassociating it * from any VM object. In other words, this is equivalent to * calling vm_page_free_toq() for each page of a list of VM objects. * * The objects must be locked. The pages must be locked if it is * managed. */ void vm_page_free_pages_toq(struct spglist *free, bool update_wire_count) { vm_page_t m; int count; if (SLIST_EMPTY(free)) return; count = 0; while ((m = SLIST_FIRST(free)) != NULL) { count++; SLIST_REMOVE_HEAD(free, plinks.s.ss); vm_page_free_toq(m); } if (update_wire_count) vm_wire_sub(count); } /* * Mark this page as wired down, preventing reclamation by the page daemon * or when the containing object is destroyed. */ void vm_page_wire(vm_page_t m) { u_int old; KASSERT(m->object != NULL, ("vm_page_wire: page %p does not belong to an object", m)); if (!vm_page_busied(m)) VM_OBJECT_ASSERT_LOCKED(m->object); KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(m->ref_count) >= 1, ("vm_page_wire: fictitious page %p has zero wirings", m)); old = atomic_fetchadd_int(&m->ref_count, 1); KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX, ("vm_page_wire: counter overflow for page %p", m)); if (VPRC_WIRE_COUNT(old) == 0) vm_wire_add(1); } /* * Attempt to wire a mapped page following a pmap lookup of that page. * This may fail if a thread is concurrently tearing down mappings of the page. */ bool vm_page_wire_mapped(vm_page_t m) { u_int old; old = m->ref_count; do { KASSERT(old > 0, ("vm_page_wire_mapped: wiring unreferenced page %p", m)); if ((old & VPRC_BLOCKED) != 0) return (false); } while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1)); if (VPRC_WIRE_COUNT(old) == 0) vm_wire_add(1); return (true); } /* * Release one wiring of the specified page, potentially allowing it to be * paged out. * * Only managed pages belonging to an object can be paged out. If the number * of wirings transitions to zero and the page is eligible for page out, then * the page is added to the specified paging queue. If the released wiring * represented the last reference to the page, the page is freed. * * A managed page must be locked. */ void vm_page_unwire(vm_page_t m, uint8_t queue) { u_int old; bool locked; KASSERT(queue < PQ_COUNT, ("vm_page_unwire: invalid queue %u request for page %p", queue, m)); if ((m->oflags & VPO_UNMANAGED) != 0) { if (vm_page_unwire_noq(m) && m->ref_count == 0) vm_page_free(m); return; } /* * Update LRU state before releasing the wiring reference. * We only need to do this once since we hold the page lock. * Use a release store when updating the reference count to * synchronize with vm_page_free_prep(). */ old = m->ref_count; locked = false; do { KASSERT(VPRC_WIRE_COUNT(old) > 0, ("vm_page_unwire: wire count underflow for page %p", m)); if (!locked && VPRC_WIRE_COUNT(old) == 1) { vm_page_lock(m); locked = true; if (queue == PQ_ACTIVE && vm_page_queue(m) == PQ_ACTIVE) vm_page_reference(m); else vm_page_mvqueue(m, queue); } } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1)); /* * Release the lock only after the wiring is released, to ensure that * the page daemon does not encounter and dequeue the page while it is * still wired. */ if (locked) vm_page_unlock(m); if (VPRC_WIRE_COUNT(old) == 1) { vm_wire_sub(1); if (old == 1) vm_page_free(m); } } /* * Unwire a page without (re-)inserting it into a page queue. It is up * to the caller to enqueue, requeue, or free the page as appropriate. * In most cases involving managed pages, vm_page_unwire() should be used * instead. */ bool vm_page_unwire_noq(vm_page_t m) { u_int old; old = vm_page_drop(m, 1); KASSERT(VPRC_WIRE_COUNT(old) != 0, ("vm_page_unref: counter underflow for page %p", m)); KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1, ("vm_page_unref: missing ref on fictitious page %p", m)); if (VPRC_WIRE_COUNT(old) > 1) return (false); vm_wire_sub(1); return (true); } /* * Ensure that the page is in the specified page queue. If the page is * active or being moved to the active queue, ensure that its act_count is * at least ACT_INIT but do not otherwise mess with it. Otherwise, ensure that * the page is at the tail of its page queue. * * The page may be wired. The caller should release its wiring reference * before releasing the page lock, otherwise the page daemon may immediately * dequeue the page. * * A managed page must be locked. */ static __always_inline void vm_page_mvqueue(vm_page_t m, const uint8_t nqueue) { vm_page_assert_locked(m); KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("vm_page_mvqueue: page %p is unmanaged", m)); if (vm_page_queue(m) != nqueue) { vm_page_dequeue(m); vm_page_enqueue(m, nqueue); } else if (nqueue != PQ_ACTIVE) { vm_page_requeue(m); } if (nqueue == PQ_ACTIVE && m->act_count < ACT_INIT) m->act_count = ACT_INIT; } /* * Put the specified page on the active list (if appropriate). */ void vm_page_activate(vm_page_t m) { if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m)) return; vm_page_mvqueue(m, PQ_ACTIVE); } /* * Move the specified page to the tail of the inactive queue, or requeue * the page if it is already in the inactive queue. */ void vm_page_deactivate(vm_page_t m) { if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m)) return; vm_page_mvqueue(m, PQ_INACTIVE); } /* * Move the specified page close to the head of the inactive queue, * bypassing LRU. A marker page is used to maintain FIFO ordering. * As with regular enqueues, we use a per-CPU batch queue to reduce * contention on the page queue lock. */ static void _vm_page_deactivate_noreuse(vm_page_t m) { vm_page_assert_locked(m); if (!vm_page_inactive(m)) { vm_page_dequeue(m); m->queue = PQ_INACTIVE; } if ((m->aflags & PGA_REQUEUE_HEAD) == 0) vm_page_aflag_set(m, PGA_REQUEUE_HEAD); vm_page_pqbatch_submit(m, PQ_INACTIVE); } void vm_page_deactivate_noreuse(vm_page_t m) { KASSERT(m->object != NULL, ("vm_page_deactivate_noreuse: page %p has no object", m)); if ((m->oflags & VPO_UNMANAGED) == 0 && !vm_page_wired(m)) _vm_page_deactivate_noreuse(m); } /* * Put a page in the laundry, or requeue it if it is already there. */ void vm_page_launder(vm_page_t m) { if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m)) return; vm_page_mvqueue(m, PQ_LAUNDRY); } /* * Put a page in the PQ_UNSWAPPABLE holding queue. */ void vm_page_unswappable(vm_page_t m) { vm_page_assert_locked(m); KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0, ("page %p already unswappable", m)); vm_page_dequeue(m); vm_page_enqueue(m, PQ_UNSWAPPABLE); } static void vm_page_release_toq(vm_page_t m, int flags) { vm_page_assert_locked(m); /* * Use a check of the valid bits to determine whether we should * accelerate reclamation of the page. The object lock might not be * held here, in which case the check is racy. At worst we will either * accelerate reclamation of a valid page and violate LRU, or * unnecessarily defer reclamation of an invalid page. * * If we were asked to not cache the page, place it near the head of the * inactive queue so that is reclaimed sooner. */ if ((flags & (VPR_TRYFREE | VPR_NOREUSE)) != 0 || m->valid == 0) _vm_page_deactivate_noreuse(m); else if (vm_page_active(m)) vm_page_reference(m); else vm_page_mvqueue(m, PQ_INACTIVE); } /* * Unwire a page and either attempt to free it or re-add it to the page queues. */ void vm_page_release(vm_page_t m, int flags) { vm_object_t object; u_int old; bool locked; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("vm_page_release: page %p is unmanaged", m)); if ((flags & VPR_TRYFREE) != 0) { for (;;) { object = (vm_object_t)atomic_load_ptr(&m->object); if (object == NULL) break; /* Depends on type-stability. */ if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object)) { object = NULL; break; } if (object == m->object) break; VM_OBJECT_WUNLOCK(object); } if (__predict_true(object != NULL)) { vm_page_release_locked(m, flags); VM_OBJECT_WUNLOCK(object); return; } } /* * Update LRU state before releasing the wiring reference. * Use a release store when updating the reference count to * synchronize with vm_page_free_prep(). */ old = m->ref_count; locked = false; do { KASSERT(VPRC_WIRE_COUNT(old) > 0, ("vm_page_unwire: wire count underflow for page %p", m)); if (!locked && VPRC_WIRE_COUNT(old) == 1) { vm_page_lock(m); locked = true; vm_page_release_toq(m, flags); } } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1)); /* * Release the lock only after the wiring is released, to ensure that * the page daemon does not encounter and dequeue the page while it is * still wired. */ if (locked) vm_page_unlock(m); if (VPRC_WIRE_COUNT(old) == 1) { vm_wire_sub(1); if (old == 1) vm_page_free(m); } } /* See vm_page_release(). */ void vm_page_release_locked(vm_page_t m, int flags) { VM_OBJECT_ASSERT_WLOCKED(m->object); KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("vm_page_release_locked: page %p is unmanaged", m)); if (vm_page_unwire_noq(m)) { if ((flags & VPR_TRYFREE) != 0 && (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) && m->dirty == 0 && !vm_page_busied(m)) { vm_page_free(m); } else { vm_page_lock(m); vm_page_release_toq(m, flags); vm_page_unlock(m); } } } static bool vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t)) { u_int old; KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0, ("vm_page_try_blocked_op: page %p has no object", m)); - KASSERT(!vm_page_busied(m), - ("vm_page_try_blocked_op: page %p is busy", m)); + KASSERT(vm_page_busied(m), + ("vm_page_try_blocked_op: page %p is not busy", m)); VM_OBJECT_ASSERT_LOCKED(m->object); old = m->ref_count; do { KASSERT(old != 0, ("vm_page_try_blocked_op: page %p has no references", m)); if (VPRC_WIRE_COUNT(old) != 0) return (false); } while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED)); (op)(m); /* * If the object is read-locked, new wirings may be created via an * object lookup. */ old = vm_page_drop(m, VPRC_BLOCKED); KASSERT(!VM_OBJECT_WOWNED(m->object) || old == (VPRC_BLOCKED | VPRC_OBJREF), ("vm_page_try_blocked_op: unexpected refcount value %u for %p", old, m)); return (true); } /* * Atomically check for wirings and remove all mappings of the page. */ bool vm_page_try_remove_all(vm_page_t m) { return (vm_page_try_blocked_op(m, pmap_remove_all)); } /* * Atomically check for wirings and remove all writeable mappings of the page. */ bool vm_page_try_remove_write(vm_page_t m) { return (vm_page_try_blocked_op(m, pmap_remove_write)); } /* * vm_page_advise * * Apply the specified advice to the given page. * * The object and page must be locked. */ void vm_page_advise(vm_page_t m, int advice) { vm_page_assert_locked(m); VM_OBJECT_ASSERT_WLOCKED(m->object); if (advice == MADV_FREE) /* * Mark the page clean. This will allow the page to be freed * without first paging it out. MADV_FREE pages are often * quickly reused by malloc(3), so we do not do anything that * would result in a page fault on a later access. */ vm_page_undirty(m); else if (advice != MADV_DONTNEED) { if (advice == MADV_WILLNEED) vm_page_activate(m); return; } /* * Clear any references to the page. Otherwise, the page daemon will * immediately reactivate the page. */ vm_page_aflag_clear(m, PGA_REFERENCED); if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) vm_page_dirty(m); /* * Place clean pages near the head of the inactive queue rather than * the tail, thus defeating the queue's LRU operation and ensuring that * the page will be reused quickly. Dirty pages not already in the * laundry are moved there. */ if (m->dirty == 0) vm_page_deactivate_noreuse(m); else if (!vm_page_in_laundry(m)) vm_page_launder(m); } /* * Grab a page, waiting until we are waken up due to the page * changing state. We keep on waiting, if the page continues * to be in the object. If the page doesn't exist, first allocate it * and then conditionally zero it. * * This routine may sleep. * * The object must be locked on entry. The lock will, however, be released * and reacquired if the routine sleeps. */ vm_page_t vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) { vm_page_t m; int sleep; int pflags; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || (allocflags & VM_ALLOC_IGN_SBUSY) != 0, ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); pflags = allocflags & - ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL); + ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL | + VM_ALLOC_NOBUSY); if ((allocflags & VM_ALLOC_NOWAIT) == 0) pflags |= VM_ALLOC_WAITFAIL; + if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) + pflags |= VM_ALLOC_SBUSY; retrylookup: if ((m = vm_page_lookup(object, pindex)) != NULL) { - sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ? - vm_page_xbusied(m) : vm_page_busied(m); + if ((allocflags & (VM_ALLOC_IGN_SBUSY | VM_ALLOC_SBUSY)) != 0) + sleep = !vm_page_trysbusy(m); + else + sleep = !vm_page_tryxbusy(m); if (sleep) { if ((allocflags & VM_ALLOC_NOWAIT) != 0) return (NULL); /* * Reference the page before unlocking and * sleeping so that the page daemon is less * likely to reclaim it. */ if ((allocflags & VM_ALLOC_NOCREAT) == 0) vm_page_aflag_set(m, PGA_REFERENCED); vm_page_busy_sleep(m, "pgrbwt", (allocflags & VM_ALLOC_IGN_SBUSY) != 0); VM_OBJECT_WLOCK(object); if ((allocflags & VM_ALLOC_WAITFAIL) != 0) return (NULL); goto retrylookup; } else { if ((allocflags & VM_ALLOC_WIRED) != 0) vm_page_wire(m); - if ((allocflags & - (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) - vm_page_xbusy(m); - else if ((allocflags & VM_ALLOC_SBUSY) != 0) - vm_page_sbusy(m); - return (m); + goto out; } } if ((allocflags & VM_ALLOC_NOCREAT) != 0) return (NULL); m = vm_page_alloc(object, pindex, pflags); if (m == NULL) { if ((allocflags & VM_ALLOC_NOWAIT) != 0) return (NULL); goto retrylookup; } if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) pmap_zero_page(m); + +out: + if ((allocflags & VM_ALLOC_NOBUSY) != 0) { + if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) + vm_page_sunbusy(m); + else + vm_page_xunbusy(m); + } return (m); } /* * Grab a page and make it valid, paging in if necessary. Pages missing from * their pager are zero filled and validated. */ int vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags) { vm_page_t m; bool sleep, xbusy; int pflags; int rv; KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || (allocflags & VM_ALLOC_IGN_SBUSY) != 0, ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); KASSERT((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, ("vm_page_grab_valid: Invalid flags 0x%X", allocflags)); VM_OBJECT_ASSERT_WLOCKED(object); pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY); pflags |= VM_ALLOC_WAITFAIL; retrylookup: xbusy = false; if ((m = vm_page_lookup(object, pindex)) != NULL) { /* * If the page is fully valid it can only become invalid * with the object lock held. If it is not valid it can * become valid with the busy lock held. Therefore, we * may unnecessarily lock the exclusive busy here if we * race with I/O completion not using the object lock. * However, we will not end up with an invalid page and a * shared lock. */ if (m->valid != VM_PAGE_BITS_ALL || (allocflags & (VM_ALLOC_IGN_SBUSY | VM_ALLOC_SBUSY)) == 0) { sleep = !vm_page_tryxbusy(m); xbusy = true; } else sleep = !vm_page_trysbusy(m); if (sleep) { /* * Reference the page before unlocking and * sleeping so that the page daemon is less * likely to reclaim it. */ if ((allocflags & VM_ALLOC_NOCREAT) == 0) vm_page_aflag_set(m, PGA_REFERENCED); vm_page_busy_sleep(m, "pgrbwt", (allocflags & VM_ALLOC_IGN_SBUSY) != 0); VM_OBJECT_WLOCK(object); goto retrylookup; } if ((allocflags & VM_ALLOC_NOCREAT) != 0 && m->valid != VM_PAGE_BITS_ALL) { if (xbusy) vm_page_xunbusy(m); else vm_page_sunbusy(m); *mp = NULL; return (VM_PAGER_FAIL); } if ((allocflags & VM_ALLOC_WIRED) != 0) vm_page_wire(m); if (m->valid == VM_PAGE_BITS_ALL) goto out; } else if ((allocflags & VM_ALLOC_NOCREAT) != 0) { *mp = NULL; return (VM_PAGER_FAIL); } else if ((m = vm_page_alloc(object, pindex, pflags)) != NULL) { xbusy = true; } else { goto retrylookup; } vm_page_assert_xbusied(m); MPASS(xbusy); if (vm_pager_has_page(object, pindex, NULL, NULL)) { rv = vm_pager_get_pages(object, &m, 1, NULL, NULL); if (rv != VM_PAGER_OK) { if (allocflags & VM_ALLOC_WIRED) vm_page_unwire_noq(m); vm_page_free(m); *mp = NULL; return (rv); } MPASS(m->valid == VM_PAGE_BITS_ALL); } else { vm_page_zero_invalid(m, TRUE); } out: if ((allocflags & VM_ALLOC_NOBUSY) != 0) { if (xbusy) vm_page_xunbusy(m); else vm_page_sunbusy(m); } if ((allocflags & VM_ALLOC_SBUSY) != 0 && xbusy) vm_page_busy_downgrade(m); *mp = m; return (VM_PAGER_OK); } /* * Return the specified range of pages from the given object. For each * page offset within the range, if a page already exists within the object * at that offset and it is busy, then wait for it to change state. If, * instead, the page doesn't exist, then allocate it. * * The caller must always specify an allocation class. * * allocation classes: * VM_ALLOC_NORMAL normal process request * VM_ALLOC_SYSTEM system *really* needs the pages * * The caller must always specify that the pages are to be busied and/or * wired. * * optional allocation flags: * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages * VM_ALLOC_NOBUSY do not exclusive busy the page * VM_ALLOC_NOWAIT do not sleep * VM_ALLOC_SBUSY set page to sbusy state * VM_ALLOC_WIRED wire the pages * VM_ALLOC_ZERO zero and validate any invalid pages * * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it * may return a partial prefix of the requested range. */ int vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, vm_page_t *ma, int count) { vm_page_t m, mpred; int pflags; int i; bool sleep; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0, ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed")); KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 || (allocflags & VM_ALLOC_WIRED) != 0, ("vm_page_grab_pages: the pages must be busied or wired")); KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || (allocflags & VM_ALLOC_IGN_SBUSY) != 0, ("vm_page_grab_pages: VM_ALLOC_SBUSY/IGN_SBUSY mismatch")); if (count == 0) return (0); - pflags = allocflags & ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | - VM_ALLOC_WAITFAIL | VM_ALLOC_IGN_SBUSY); + pflags = allocflags & + ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL | + VM_ALLOC_NOBUSY); if ((allocflags & VM_ALLOC_NOWAIT) == 0) pflags |= VM_ALLOC_WAITFAIL; + if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) + pflags |= VM_ALLOC_SBUSY; i = 0; retrylookup: m = vm_radix_lookup_le(&object->rtree, pindex + i); if (m == NULL || m->pindex != pindex + i) { mpred = m; m = NULL; } else mpred = TAILQ_PREV(m, pglist, listq); for (; i < count; i++) { if (m != NULL) { - sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ? - vm_page_xbusied(m) : vm_page_busied(m); + if ((allocflags & + (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0) + sleep = !vm_page_trysbusy(m); + else + sleep = !vm_page_tryxbusy(m); if (sleep) { if ((allocflags & VM_ALLOC_NOWAIT) != 0) break; /* * Reference the page before unlocking and * sleeping so that the page daemon is less * likely to reclaim it. */ if ((allocflags & VM_ALLOC_NOCREAT) == 0) vm_page_aflag_set(m, PGA_REFERENCED); vm_page_busy_sleep(m, "grbmaw", (allocflags & VM_ALLOC_IGN_SBUSY) != 0); VM_OBJECT_WLOCK(object); goto retrylookup; } if ((allocflags & VM_ALLOC_WIRED) != 0) vm_page_wire(m); - if ((allocflags & (VM_ALLOC_NOBUSY | - VM_ALLOC_SBUSY)) == 0) - vm_page_xbusy(m); - if ((allocflags & VM_ALLOC_SBUSY) != 0) - vm_page_sbusy(m); } else { if ((allocflags & VM_ALLOC_NOCREAT) != 0) break; m = vm_page_alloc_after(object, pindex + i, pflags | VM_ALLOC_COUNT(count - i), mpred); if (m == NULL) { if ((allocflags & VM_ALLOC_NOWAIT) != 0) break; goto retrylookup; } } if (m->valid == 0 && (allocflags & VM_ALLOC_ZERO) != 0) { if ((m->flags & PG_ZERO) == 0) pmap_zero_page(m); m->valid = VM_PAGE_BITS_ALL; + } + if ((allocflags & VM_ALLOC_NOBUSY) != 0) { + if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) + vm_page_sunbusy(m); + else + vm_page_xunbusy(m); } ma[i] = mpred = m; m = vm_page_next(m); } return (i); } /* * Mapping function for valid or dirty bits in a page. * * Inputs are required to range within a page. */ vm_page_bits_t vm_page_bits(int base, int size) { int first_bit; int last_bit; KASSERT( base + size <= PAGE_SIZE, ("vm_page_bits: illegal base/size %d/%d", base, size) ); if (size == 0) /* handle degenerate case */ return (0); first_bit = base >> DEV_BSHIFT; last_bit = (base + size - 1) >> DEV_BSHIFT; return (((vm_page_bits_t)2 << last_bit) - ((vm_page_bits_t)1 << first_bit)); } /* * vm_page_set_valid_range: * * Sets portions of a page valid. The arguments are expected * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive * of any partial chunks touched by the range. The invalid portion of * such chunks will be zeroed. * * (base + size) must be less then or equal to PAGE_SIZE. */ void vm_page_set_valid_range(vm_page_t m, int base, int size) { int endoff, frag; VM_OBJECT_ASSERT_WLOCKED(m->object); if (size == 0) /* handle degenerate case */ return; /* * If the base is not DEV_BSIZE aligned and the valid * bit is clear, we have to zero out a portion of the * first block. */ if ((frag = rounddown2(base, DEV_BSIZE)) != base && (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) pmap_zero_page_area(m, frag, base - frag); /* * If the ending offset is not DEV_BSIZE aligned and the * valid bit is clear, we have to zero out a portion of * the last block. */ endoff = base + size; if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) pmap_zero_page_area(m, endoff, DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); /* * Assert that no previously invalid block that is now being validated * is already dirty. */ KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, ("vm_page_set_valid_range: page %p is dirty", m)); /* * Set valid bits inclusive of any overlap. */ m->valid |= vm_page_bits(base, size); } /* * Clear the given bits from the specified page's dirty field. */ static __inline void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) { uintptr_t addr; #if PAGE_SIZE < 16384 int shift; #endif /* * If the object is locked and the page is neither exclusive busy nor * write mapped, then the page's dirty field cannot possibly be * set by a concurrent pmap operation. */ VM_OBJECT_ASSERT_WLOCKED(m->object); if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) m->dirty &= ~pagebits; else { /* * The pmap layer can call vm_page_dirty() without * holding a distinguished lock. The combination of * the object's lock and an atomic operation suffice * to guarantee consistency of the page dirty field. * * For PAGE_SIZE == 32768 case, compiler already * properly aligns the dirty field, so no forcible * alignment is needed. Only require existence of * atomic_clear_64 when page size is 32768. */ addr = (uintptr_t)&m->dirty; #if PAGE_SIZE == 32768 atomic_clear_64((uint64_t *)addr, pagebits); #elif PAGE_SIZE == 16384 atomic_clear_32((uint32_t *)addr, pagebits); #else /* PAGE_SIZE <= 8192 */ /* * Use a trick to perform a 32-bit atomic on the * containing aligned word, to not depend on the existence * of atomic_clear_{8, 16}. */ shift = addr & (sizeof(uint32_t) - 1); #if BYTE_ORDER == BIG_ENDIAN shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY; #else shift *= NBBY; #endif addr &= ~(sizeof(uint32_t) - 1); atomic_clear_32((uint32_t *)addr, pagebits << shift); #endif /* PAGE_SIZE */ } } /* * vm_page_set_validclean: * * Sets portions of a page valid and clean. The arguments are expected * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive * of any partial chunks touched by the range. The invalid portion of * such chunks will be zero'd. * * (base + size) must be less then or equal to PAGE_SIZE. */ void vm_page_set_validclean(vm_page_t m, int base, int size) { vm_page_bits_t oldvalid, pagebits; int endoff, frag; VM_OBJECT_ASSERT_WLOCKED(m->object); if (size == 0) /* handle degenerate case */ return; /* * If the base is not DEV_BSIZE aligned and the valid * bit is clear, we have to zero out a portion of the * first block. */ if ((frag = rounddown2(base, DEV_BSIZE)) != base && (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) pmap_zero_page_area(m, frag, base - frag); /* * If the ending offset is not DEV_BSIZE aligned and the * valid bit is clear, we have to zero out a portion of * the last block. */ endoff = base + size; if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) pmap_zero_page_area(m, endoff, DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); /* * Set valid, clear dirty bits. If validating the entire * page we can safely clear the pmap modify bit. We also * use this opportunity to clear the VPO_NOSYNC flag. If a process * takes a write fault on a MAP_NOSYNC memory area the flag will * be set again. * * We set valid bits inclusive of any overlap, but we can only * clear dirty bits for DEV_BSIZE chunks that are fully within * the range. */ oldvalid = m->valid; pagebits = vm_page_bits(base, size); m->valid |= pagebits; #if 0 /* NOT YET */ if ((frag = base & (DEV_BSIZE - 1)) != 0) { frag = DEV_BSIZE - frag; base += frag; size -= frag; if (size < 0) size = 0; } pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); #endif if (base == 0 && size == PAGE_SIZE) { /* * The page can only be modified within the pmap if it is * mapped, and it can only be mapped if it was previously * fully valid. */ if (oldvalid == VM_PAGE_BITS_ALL) /* * Perform the pmap_clear_modify() first. Otherwise, * a concurrent pmap operation, such as * pmap_protect(), could clear a modification in the * pmap and set the dirty field on the page before * pmap_clear_modify() had begun and after the dirty * field was cleared here. */ pmap_clear_modify(m); m->dirty = 0; m->oflags &= ~VPO_NOSYNC; } else if (oldvalid != VM_PAGE_BITS_ALL) m->dirty &= ~pagebits; else vm_page_clear_dirty_mask(m, pagebits); } void vm_page_clear_dirty(vm_page_t m, int base, int size) { vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); } /* * vm_page_set_invalid: * * Invalidates DEV_BSIZE'd chunks within a page. Both the * valid and dirty bits for the effected areas are cleared. */ void vm_page_set_invalid(vm_page_t m, int base, int size) { vm_page_bits_t bits; vm_object_t object; object = m->object; VM_OBJECT_ASSERT_WLOCKED(object); if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + size >= object->un_pager.vnp.vnp_size) bits = VM_PAGE_BITS_ALL; else bits = vm_page_bits(base, size); if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL && bits != 0) pmap_remove_all(m); KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) || !pmap_page_is_mapped(m), ("vm_page_set_invalid: page %p is mapped", m)); m->valid &= ~bits; m->dirty &= ~bits; } /* * vm_page_zero_invalid() * * The kernel assumes that the invalid portions of a page contain * garbage, but such pages can be mapped into memory by user code. * When this occurs, we must zero out the non-valid portions of the * page so user code sees what it expects. * * Pages are most often semi-valid when the end of a file is mapped * into memory and the file's size is not page aligned. */ void vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) { int b; int i; VM_OBJECT_ASSERT_WLOCKED(m->object); /* * Scan the valid bits looking for invalid sections that * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the * valid bit may be set ) have already been zeroed by * vm_page_set_validclean(). */ for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { if (i == (PAGE_SIZE / DEV_BSIZE) || (m->valid & ((vm_page_bits_t)1 << i))) { if (i > b) { pmap_zero_page_area(m, b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); } b = i + 1; } } /* * setvalid is TRUE when we can safely set the zero'd areas * as being valid. We can do this if there are no cache consistancy * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. */ if (setvalid) m->valid = VM_PAGE_BITS_ALL; } /* * vm_page_is_valid: * * Is (partial) page valid? Note that the case where size == 0 * will return FALSE in the degenerate case where the page is * entirely invalid, and TRUE otherwise. */ int vm_page_is_valid(vm_page_t m, int base, int size) { vm_page_bits_t bits; VM_OBJECT_ASSERT_LOCKED(m->object); bits = vm_page_bits(base, size); return (m->valid != 0 && (m->valid & bits) == bits); } /* * Returns true if all of the specified predicates are true for the entire * (super)page and false otherwise. */ bool vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m) { vm_object_t object; int i, npages; object = m->object; if (skip_m != NULL && skip_m->object != object) return (false); VM_OBJECT_ASSERT_LOCKED(object); npages = atop(pagesizes[m->psind]); /* * The physically contiguous pages that make up a superpage, i.e., a * page with a page size index ("psind") greater than zero, will * occupy adjacent entries in vm_page_array[]. */ for (i = 0; i < npages; i++) { /* Always test object consistency, including "skip_m". */ if (m[i].object != object) return (false); if (&m[i] == skip_m) continue; if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i])) return (false); if ((flags & PS_ALL_DIRTY) != 0) { /* * Calling vm_page_test_dirty() or pmap_is_modified() * might stop this case from spuriously returning * "false". However, that would require a write lock * on the object containing "m[i]". */ if (m[i].dirty != VM_PAGE_BITS_ALL) return (false); } if ((flags & PS_ALL_VALID) != 0 && m[i].valid != VM_PAGE_BITS_ALL) return (false); } return (true); } /* * Set the page's dirty bits if the page is modified. */ void vm_page_test_dirty(vm_page_t m) { VM_OBJECT_ASSERT_WLOCKED(m->object); if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) vm_page_dirty(m); } void vm_page_lock_KBI(vm_page_t m, const char *file, int line) { mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); } void vm_page_unlock_KBI(vm_page_t m, const char *file, int line) { mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); } int vm_page_trylock_KBI(vm_page_t m, const char *file, int line) { return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); } #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) void vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) { vm_page_lock_assert_KBI(m, MA_OWNED, file, line); } void vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) { mtx_assert_(vm_page_lockptr(m), a, file, line); } #endif #ifdef INVARIANTS void vm_page_object_lock_assert(vm_page_t m) { /* * Certain of the page's fields may only be modified by the * holder of the containing object's lock or the exclusive busy. * holder. Unfortunately, the holder of the write busy is * not recorded, and thus cannot be checked here. */ if (m->object != NULL && !vm_page_xbusied(m)) VM_OBJECT_ASSERT_WLOCKED(m->object); } void vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits) { if ((bits & PGA_WRITEABLE) == 0) return; /* * The PGA_WRITEABLE flag can only be set if the page is * managed, is exclusively busied or the object is locked. * Currently, this flag is only set by pmap_enter(). */ KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("PGA_WRITEABLE on unmanaged page")); if (!vm_page_xbusied(m)) VM_OBJECT_ASSERT_LOCKED(m->object); } #endif #include "opt_ddb.h" #ifdef DDB #include #include DB_SHOW_COMMAND(page, vm_page_print_page_info) { db_printf("vm_cnt.v_free_count: %d\n", vm_free_count()); db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count()); db_printf("vm_cnt.v_active_count: %d\n", vm_active_count()); db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count()); db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count()); db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); } DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) { int dom; db_printf("pq_free %d\n", vm_free_count()); for (dom = 0; dom < vm_ndomains; dom++) { db_printf( "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n", dom, vm_dom[dom].vmd_page_count, vm_dom[dom].vmd_free_count, vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt, vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt); } } DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) { vm_page_t m; boolean_t phys, virt; if (!have_addr) { db_printf("show pginfo addr\n"); return; } phys = strchr(modif, 'p') != NULL; virt = strchr(modif, 'v') != NULL; if (virt) m = PHYS_TO_VM_PAGE(pmap_kextract(addr)); else if (phys) m = PHYS_TO_VM_PAGE(addr); else m = (vm_page_t)addr; db_printf( "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref %u\n" " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, m->queue, m->ref_count, m->aflags, m->oflags, m->flags, m->act_count, m->busy_lock, m->valid, m->dirty); } #endif /* DDB */ Index: head/sys/vm/vm_page.h =================================================================== --- head/sys/vm/vm_page.h (revision 353534) +++ head/sys/vm/vm_page.h (revision 353535) @@ -1,934 +1,935 @@ /*- * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU) * * Copyright (c) 1991, 1993 * The Regents of the University of California. All rights reserved. * * This code is derived from software contributed to Berkeley by * The Mach Operating System project at Carnegie-Mellon University. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)vm_page.h 8.2 (Berkeley) 12/13/93 * * * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Authors: Avadis Tevanian, Jr., Michael Wayne Young * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. * * $FreeBSD$ */ /* * Resident memory system definitions. */ #ifndef _VM_PAGE_ #define _VM_PAGE_ #include /* * Management of resident (logical) pages. * * A small structure is kept for each resident * page, indexed by page number. Each structure * is an element of several collections: * * A radix tree used to quickly * perform object/offset lookups * * A list of all pages for a given object, * so they can be quickly deactivated at * time of deallocation. * * An ordered list of pages due for pageout. * * In addition, the structure contains the object * and offset to which this page belongs (for pageout), * and sundry status bits. * * In general, operations on this structure's mutable fields are * synchronized using either one of or a combination of the lock on the * object that the page belongs to (O), the page lock (P), * the per-domain lock for the free queues (F), or the page's queue * lock (Q). The physical address of a page is used to select its page * lock from a pool. The queue lock for a page depends on the value of * its queue field and described in detail below. If a field is * annotated below with two of these locks, then holding either lock is * sufficient for read access, but both locks are required for write * access. An annotation of (C) indicates that the field is immutable. * An annotation of (A) indicates that modifications to the field must * be atomic. Accesses to such fields may require additional * synchronization depending on the context. * * In contrast, the synchronization of accesses to the page's * dirty field is machine dependent (M). In the * machine-independent layer, the lock on the object that the * page belongs to must be held in order to operate on the field. * However, the pmap layer is permitted to set all bits within * the field without holding that lock. If the underlying * architecture does not support atomic read-modify-write * operations on the field's type, then the machine-independent * layer uses a 32-bit atomic on the aligned 32-bit word that * contains the dirty field. In the machine-independent layer, * the implementation of read-modify-write operations on the * field is encapsulated in vm_page_clear_dirty_mask(). * * The ref_count field tracks references to the page. References that * prevent the page from being reclaimable are called wirings and are * counted in the low bits of ref_count. The containing object's * reference, if one exists, is counted using the VPRC_OBJREF bit in the * ref_count field. Additionally, the VPRC_BLOCKED bit is used to * atomically check for wirings and prevent new wirings via * pmap_extract_and_hold(). When a page belongs to an object, it may be * wired only when the object is locked, or the page is busy, or by * pmap_extract_and_hold(). As a result, if the object is locked and the * page is not busy (or is exclusively busied by the current thread), and * the page is unmapped, its wire count will not increase. The ref_count * field is updated using atomic operations in most cases, except when it * is known that no other references to the page exist, such as in the page * allocator. A page may be present in the page queues, or even actively * scanned by the page daemon, without an explicitly counted referenced. * The page daemon must therefore handle the possibility of a concurrent * free of the page. * * The busy lock is an embedded reader-writer lock which protects the * page's contents and identity (i.e., its tuple) and * interlocks with the object lock (O). In particular, a page may be * busied or unbusied only with the object write lock held. To avoid * bloating the page structure, the busy lock lacks some of the * features available to the kernel's general-purpose synchronization * primitives. As a result, busy lock ordering rules are not verified, * lock recursion is not detected, and an attempt to xbusy a busy page * or sbusy an xbusy page results will trigger a panic rather than * causing the thread to block. vm_page_sleep_if_busy() can be used to * sleep until the page's busy state changes, after which the caller * must re-lookup the page and re-evaluate its state. * * The queue field is the index of the page queue containing the page, * or PQ_NONE if the page is not enqueued. The queue lock of a page is * the page queue lock corresponding to the page queue index, or the * page lock (P) for the page if it is not enqueued. To modify the * queue field, the queue lock for the old value of the field must be * held. There is one exception to this rule: the page daemon may * transition the queue field from PQ_INACTIVE to PQ_NONE immediately * prior to freeing a page during an inactive queue scan. At that * point the page has already been physically dequeued and no other * references to that vm_page structure exist. * * To avoid contention on page queue locks, page queue operations * (enqueue, dequeue, requeue) are batched using per-CPU queues. A * deferred operation is requested by inserting an entry into a batch * queue; the entry is simply a pointer to the page, and the request * type is encoded in the page's aflags field using the values in * PGA_QUEUE_STATE_MASK. The type-stability of struct vm_pages is * crucial to this scheme since the processing of entries in a given * batch queue may be deferred indefinitely. In particular, a page may * be freed before its pending batch queue entries have been processed. * The page lock (P) must be held to schedule a batched queue * operation, and the page queue lock must be held in order to process * batch queue entries for the page queue. There is one exception to * this rule: the thread freeing a page may schedule a dequeue without * holding the page lock. In this scenario the only other thread which * may hold a reference to the page is the page daemon, which is * careful to avoid modifying the page's queue state once the dequeue * has been requested by setting PGA_DEQUEUE. */ #if PAGE_SIZE == 4096 #define VM_PAGE_BITS_ALL 0xffu typedef uint8_t vm_page_bits_t; #elif PAGE_SIZE == 8192 #define VM_PAGE_BITS_ALL 0xffffu typedef uint16_t vm_page_bits_t; #elif PAGE_SIZE == 16384 #define VM_PAGE_BITS_ALL 0xffffffffu typedef uint32_t vm_page_bits_t; #elif PAGE_SIZE == 32768 #define VM_PAGE_BITS_ALL 0xfffffffffffffffflu typedef uint64_t vm_page_bits_t; #endif struct vm_page { union { TAILQ_ENTRY(vm_page) q; /* page queue or free list (Q) */ struct { SLIST_ENTRY(vm_page) ss; /* private slists */ void *pv; } s; struct { u_long p; u_long v; } memguard; } plinks; TAILQ_ENTRY(vm_page) listq; /* pages in same object (O) */ vm_object_t object; /* which object am I in (O) */ vm_pindex_t pindex; /* offset into object (O,P) */ vm_paddr_t phys_addr; /* physical address of page (C) */ struct md_page md; /* machine dependent stuff */ u_int ref_count; /* page references (A) */ volatile u_int busy_lock; /* busy owners lock */ uint16_t flags; /* page PG_* flags (P) */ uint8_t order; /* index of the buddy queue (F) */ uint8_t pool; /* vm_phys freepool index (F) */ uint8_t aflags; /* access is atomic */ uint8_t oflags; /* page VPO_* flags (O) */ uint8_t queue; /* page queue index (Q) */ int8_t psind; /* pagesizes[] index (O) */ int8_t segind; /* vm_phys segment index (C) */ u_char act_count; /* page usage count (P) */ /* NOTE that these must support one bit per DEV_BSIZE in a page */ /* so, on normal X86 kernels, they must be at least 8 bits wide */ vm_page_bits_t valid; /* map of valid DEV_BSIZE chunks (O) */ vm_page_bits_t dirty; /* map of dirty DEV_BSIZE chunks (M) */ }; /* * Special bits used in the ref_count field. * * ref_count is normally used to count wirings that prevent the page from being * reclaimed, but also supports several special types of references that do not * prevent reclamation. Accesses to the ref_count field must be atomic unless * the page is unallocated. * * VPRC_OBJREF is the reference held by the containing object. It can set or * cleared only when the corresponding object's write lock is held. * * VPRC_BLOCKED is used to atomically block wirings via pmap lookups while * attempting to tear down all mappings of a given page. The page lock and * object write lock must both be held in order to set or clear this bit. */ #define VPRC_BLOCKED 0x40000000u /* mappings are being removed */ #define VPRC_OBJREF 0x80000000u /* object reference, cleared with (O) */ #define VPRC_WIRE_COUNT(c) ((c) & ~(VPRC_BLOCKED | VPRC_OBJREF)) #define VPRC_WIRE_COUNT_MAX (~(VPRC_BLOCKED | VPRC_OBJREF)) /* * Page flags stored in oflags: * * Access to these page flags is synchronized by the lock on the object * containing the page (O). * * Note: VPO_UNMANAGED (used by OBJT_DEVICE, OBJT_PHYS and OBJT_SG) * indicates that the page is not under PV management but * otherwise should be treated as a normal page. Pages not * under PV management cannot be paged out via the * object/vm_page_t because there is no knowledge of their pte * mappings, and such pages are also not on any PQ queue. * */ #define VPO_KMEM_EXEC 0x01 /* kmem mapping allows execution */ #define VPO_SWAPSLEEP 0x02 /* waiting for swap to finish */ #define VPO_UNMANAGED 0x04 /* no PV management for page */ #define VPO_SWAPINPROG 0x08 /* swap I/O in progress on page */ #define VPO_NOSYNC 0x10 /* do not collect for syncer */ /* * Busy page implementation details. * The algorithm is taken mostly by rwlock(9) and sx(9) locks implementation, * even if the support for owner identity is removed because of size * constraints. Checks on lock recursion are then not possible, while the * lock assertions effectiveness is someway reduced. */ #define VPB_BIT_SHARED 0x01 #define VPB_BIT_EXCLUSIVE 0x02 #define VPB_BIT_WAITERS 0x04 #define VPB_BIT_FLAGMASK \ (VPB_BIT_SHARED | VPB_BIT_EXCLUSIVE | VPB_BIT_WAITERS) #define VPB_SHARERS_SHIFT 3 #define VPB_SHARERS(x) \ (((x) & ~VPB_BIT_FLAGMASK) >> VPB_SHARERS_SHIFT) #define VPB_SHARERS_WORD(x) ((x) << VPB_SHARERS_SHIFT | VPB_BIT_SHARED) #define VPB_ONE_SHARER (1 << VPB_SHARERS_SHIFT) #define VPB_SINGLE_EXCLUSIVER VPB_BIT_EXCLUSIVE #define VPB_UNBUSIED VPB_SHARERS_WORD(0) #define PQ_NONE 255 #define PQ_INACTIVE 0 #define PQ_ACTIVE 1 #define PQ_LAUNDRY 2 #define PQ_UNSWAPPABLE 3 #define PQ_COUNT 4 #ifndef VM_PAGE_HAVE_PGLIST TAILQ_HEAD(pglist, vm_page); #define VM_PAGE_HAVE_PGLIST #endif SLIST_HEAD(spglist, vm_page); #ifdef _KERNEL extern vm_page_t bogus_page; #endif /* _KERNEL */ extern struct mtx_padalign pa_lock[]; #if defined(__arm__) #define PDRSHIFT PDR_SHIFT #elif !defined(PDRSHIFT) #define PDRSHIFT 21 #endif #define pa_index(pa) ((pa) >> PDRSHIFT) #define PA_LOCKPTR(pa) ((struct mtx *)(&pa_lock[pa_index(pa) % PA_LOCK_COUNT])) #define PA_LOCKOBJPTR(pa) ((struct lock_object *)PA_LOCKPTR((pa))) #define PA_LOCK(pa) mtx_lock(PA_LOCKPTR(pa)) #define PA_TRYLOCK(pa) mtx_trylock(PA_LOCKPTR(pa)) #define PA_UNLOCK(pa) mtx_unlock(PA_LOCKPTR(pa)) #define PA_UNLOCK_COND(pa) \ do { \ if ((pa) != 0) { \ PA_UNLOCK((pa)); \ (pa) = 0; \ } \ } while (0) #define PA_LOCK_ASSERT(pa, a) mtx_assert(PA_LOCKPTR(pa), (a)) #if defined(KLD_MODULE) && !defined(KLD_TIED) #define vm_page_lock(m) vm_page_lock_KBI((m), LOCK_FILE, LOCK_LINE) #define vm_page_unlock(m) vm_page_unlock_KBI((m), LOCK_FILE, LOCK_LINE) #define vm_page_trylock(m) vm_page_trylock_KBI((m), LOCK_FILE, LOCK_LINE) #else /* !KLD_MODULE */ #define vm_page_lockptr(m) (PA_LOCKPTR(VM_PAGE_TO_PHYS((m)))) #define vm_page_lock(m) mtx_lock(vm_page_lockptr((m))) #define vm_page_unlock(m) mtx_unlock(vm_page_lockptr((m))) #define vm_page_trylock(m) mtx_trylock(vm_page_lockptr((m))) #endif #if defined(INVARIANTS) #define vm_page_assert_locked(m) \ vm_page_assert_locked_KBI((m), __FILE__, __LINE__) #define vm_page_lock_assert(m, a) \ vm_page_lock_assert_KBI((m), (a), __FILE__, __LINE__) #else #define vm_page_assert_locked(m) #define vm_page_lock_assert(m, a) #endif /* * The vm_page's aflags are updated using atomic operations. To set or clear * these flags, the functions vm_page_aflag_set() and vm_page_aflag_clear() * must be used. Neither these flags nor these functions are part of the KBI. * * PGA_REFERENCED may be cleared only if the page is locked. It is set by * both the MI and MD VM layers. However, kernel loadable modules should not * directly set this flag. They should call vm_page_reference() instead. * * PGA_WRITEABLE is set exclusively on managed pages by pmap_enter(). * When it does so, the object must be locked, or the page must be * exclusive busied. The MI VM layer must never access this flag * directly. Instead, it should call pmap_page_is_write_mapped(). * * PGA_EXECUTABLE may be set by pmap routines, and indicates that a page has * at least one executable mapping. It is not consumed by the MI VM layer. * * PGA_ENQUEUED is set and cleared when a page is inserted into or removed * from a page queue, respectively. It determines whether the plinks.q field * of the page is valid. To set or clear this flag, the queue lock for the * page must be held: the page queue lock corresponding to the page's "queue" * field if its value is not PQ_NONE, and the page lock otherwise. * * PGA_DEQUEUE is set when the page is scheduled to be dequeued from a page * queue, and cleared when the dequeue request is processed. A page may * have PGA_DEQUEUE set and PGA_ENQUEUED cleared, for instance if a dequeue * is requested after the page is scheduled to be enqueued but before it is * actually inserted into the page queue. For allocated pages, the page lock * must be held to set this flag, but it may be set by vm_page_free_prep() * without the page lock held. The page queue lock must be held to clear the * PGA_DEQUEUE flag. * * PGA_REQUEUE is set when the page is scheduled to be enqueued or requeued * in its page queue. The page lock must be held to set this flag, and the * queue lock for the page must be held to clear it. * * PGA_REQUEUE_HEAD is a special flag for enqueuing pages near the head of * the inactive queue, thus bypassing LRU. The page lock must be held to * set this flag, and the queue lock for the page must be held to clear it. */ #define PGA_WRITEABLE 0x01 /* page may be mapped writeable */ #define PGA_REFERENCED 0x02 /* page has been referenced */ #define PGA_EXECUTABLE 0x04 /* page may be mapped executable */ #define PGA_ENQUEUED 0x08 /* page is enqueued in a page queue */ #define PGA_DEQUEUE 0x10 /* page is due to be dequeued */ #define PGA_REQUEUE 0x20 /* page is due to be requeued */ #define PGA_REQUEUE_HEAD 0x40 /* page requeue should bypass LRU */ #define PGA_QUEUE_STATE_MASK (PGA_ENQUEUED | PGA_DEQUEUE | PGA_REQUEUE | \ PGA_REQUEUE_HEAD) /* * Page flags. If changed at any other time than page allocation or * freeing, the modification must be protected by the vm_page lock. * * The PG_PCPU_CACHE flag is set at allocation time if the page was * allocated from a per-CPU cache. It is cleared the next time that the * page is allocated from the physical memory allocator. */ #define PG_PCPU_CACHE 0x0001 /* was allocated from per-CPU caches */ #define PG_FICTITIOUS 0x0004 /* physical page doesn't exist */ #define PG_ZERO 0x0008 /* page is zeroed */ #define PG_MARKER 0x0010 /* special queue marker page */ #define PG_NODUMP 0x0080 /* don't include this page in a dump */ /* * Misc constants. */ #define ACT_DECLINE 1 #define ACT_ADVANCE 3 #define ACT_INIT 5 #define ACT_MAX 64 #ifdef _KERNEL #include #include /* * Each pageable resident page falls into one of five lists: * * free * Available for allocation now. * * inactive * Low activity, candidates for reclamation. * This list is approximately LRU ordered. * * laundry * This is the list of pages that should be * paged out next. * * unswappable * Dirty anonymous pages that cannot be paged * out because no swap device is configured. * * active * Pages that are "active", i.e., they have been * recently referenced. * */ extern vm_page_t vm_page_array; /* First resident page in table */ extern long vm_page_array_size; /* number of vm_page_t's */ extern long first_page; /* first physical page number */ #define VM_PAGE_TO_PHYS(entry) ((entry)->phys_addr) /* * PHYS_TO_VM_PAGE() returns the vm_page_t object that represents a memory * page to which the given physical address belongs. The correct vm_page_t * object is returned for addresses that are not page-aligned. */ vm_page_t PHYS_TO_VM_PAGE(vm_paddr_t pa); /* * Page allocation parameters for vm_page for the functions * vm_page_alloc(), vm_page_grab(), vm_page_alloc_contig() and * vm_page_alloc_freelist(). Some functions support only a subset * of the flags, and ignore others, see the flags legend. * * The meaning of VM_ALLOC_ZERO differs slightly between the vm_page_alloc*() * and the vm_page_grab*() functions. See these functions for details. * * Bits 0 - 1 define class. * Bits 2 - 15 dedicated for flags. * Legend: * (a) - vm_page_alloc() supports the flag. * (c) - vm_page_alloc_contig() supports the flag. * (f) - vm_page_alloc_freelist() supports the flag. * (g) - vm_page_grab() supports the flag. * (p) - vm_page_grab_pages() supports the flag. * Bits above 15 define the count of additional pages that the caller * intends to allocate. */ #define VM_ALLOC_NORMAL 0 #define VM_ALLOC_INTERRUPT 1 #define VM_ALLOC_SYSTEM 2 #define VM_ALLOC_CLASS_MASK 3 #define VM_ALLOC_WAITOK 0x0008 /* (acf) Sleep and retry */ #define VM_ALLOC_WAITFAIL 0x0010 /* (acf) Sleep and return error */ #define VM_ALLOC_WIRED 0x0020 /* (acfgp) Allocate a wired page */ #define VM_ALLOC_ZERO 0x0040 /* (acfgp) Allocate a prezeroed page */ #define VM_ALLOC_NOOBJ 0x0100 /* (acg) No associated object */ #define VM_ALLOC_NOBUSY 0x0200 /* (acgp) Do not excl busy the page */ #define VM_ALLOC_NOCREAT 0x0400 /* (gp) Don't create a page */ #define VM_ALLOC_IGN_SBUSY 0x1000 /* (gp) Ignore shared busy flag */ #define VM_ALLOC_NODUMP 0x2000 /* (ag) don't include in dump */ #define VM_ALLOC_SBUSY 0x4000 /* (acgp) Shared busy the page */ #define VM_ALLOC_NOWAIT 0x8000 /* (acfgp) Do not sleep */ #define VM_ALLOC_COUNT_SHIFT 16 #define VM_ALLOC_COUNT(count) ((count) << VM_ALLOC_COUNT_SHIFT) #ifdef M_NOWAIT static inline int malloc2vm_flags(int malloc_flags) { int pflags; KASSERT((malloc_flags & M_USE_RESERVE) == 0 || (malloc_flags & M_NOWAIT) != 0, ("M_USE_RESERVE requires M_NOWAIT")); pflags = (malloc_flags & M_USE_RESERVE) != 0 ? VM_ALLOC_INTERRUPT : VM_ALLOC_SYSTEM; if ((malloc_flags & M_ZERO) != 0) pflags |= VM_ALLOC_ZERO; if ((malloc_flags & M_NODUMP) != 0) pflags |= VM_ALLOC_NODUMP; if ((malloc_flags & M_NOWAIT)) pflags |= VM_ALLOC_NOWAIT; if ((malloc_flags & M_WAITOK)) pflags |= VM_ALLOC_WAITOK; return (pflags); } #endif /* * Predicates supported by vm_page_ps_test(): * * PS_ALL_DIRTY is true only if the entire (super)page is dirty. * However, it can be spuriously false when the (super)page has become * dirty in the pmap but that information has not been propagated to the * machine-independent layer. */ #define PS_ALL_DIRTY 0x1 #define PS_ALL_VALID 0x2 #define PS_NONE_BUSY 0x4 int vm_page_busy_acquire(vm_page_t m, int allocflags); void vm_page_busy_downgrade(vm_page_t m); +int vm_page_busy_tryupgrade(vm_page_t m); void vm_page_busy_sleep(vm_page_t m, const char *msg, bool nonshared); void vm_page_free(vm_page_t m); void vm_page_free_zero(vm_page_t m); void vm_page_activate (vm_page_t); void vm_page_advise(vm_page_t m, int advice); vm_page_t vm_page_alloc(vm_object_t, vm_pindex_t, int); vm_page_t vm_page_alloc_domain(vm_object_t, vm_pindex_t, int, int); vm_page_t vm_page_alloc_after(vm_object_t, vm_pindex_t, int, vm_page_t); vm_page_t vm_page_alloc_domain_after(vm_object_t, vm_pindex_t, int, int, vm_page_t); vm_page_t vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr); vm_page_t vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr); vm_page_t vm_page_alloc_freelist(int, int); vm_page_t vm_page_alloc_freelist_domain(int, int, int); bool vm_page_blacklist_add(vm_paddr_t pa, bool verbose); void vm_page_change_lock(vm_page_t m, struct mtx **mtx); vm_page_t vm_page_grab (vm_object_t, vm_pindex_t, int); int vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, vm_page_t *ma, int count); int vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags); void vm_page_deactivate(vm_page_t); void vm_page_deactivate_noreuse(vm_page_t); void vm_page_dequeue(vm_page_t m); void vm_page_dequeue_deferred(vm_page_t m); vm_page_t vm_page_find_least(vm_object_t, vm_pindex_t); bool vm_page_free_prep(vm_page_t m); vm_page_t vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr); void vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr); int vm_page_insert (vm_page_t, vm_object_t, vm_pindex_t); void vm_page_launder(vm_page_t m); vm_page_t vm_page_lookup (vm_object_t, vm_pindex_t); vm_page_t vm_page_next(vm_page_t m); int vm_page_pa_tryrelock(pmap_t, vm_paddr_t, vm_paddr_t *); void vm_page_pqbatch_drain(void); void vm_page_pqbatch_submit(vm_page_t m, uint8_t queue); vm_page_t vm_page_prev(vm_page_t m); bool vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m); void vm_page_putfake(vm_page_t m); void vm_page_readahead_finish(vm_page_t m); bool vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary); bool vm_page_reclaim_contig_domain(int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary); void vm_page_reference(vm_page_t m); #define VPR_TRYFREE 0x01 #define VPR_NOREUSE 0x02 void vm_page_release(vm_page_t m, int flags); void vm_page_release_locked(vm_page_t m, int flags); bool vm_page_remove(vm_page_t); int vm_page_rename(vm_page_t, vm_object_t, vm_pindex_t); vm_page_t vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex); void vm_page_requeue(vm_page_t m); int vm_page_sbusied(vm_page_t m); vm_page_t vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, u_long alignment, vm_paddr_t boundary, int options); void vm_page_set_valid_range(vm_page_t m, int base, int size); int vm_page_sleep_if_busy(vm_page_t m, const char *msg); int vm_page_sleep_if_xbusy(vm_page_t m, const char *msg); vm_offset_t vm_page_startup(vm_offset_t vaddr); void vm_page_sunbusy(vm_page_t m); void vm_page_swapqueue(vm_page_t m, uint8_t oldq, uint8_t newq); bool vm_page_try_remove_all(vm_page_t m); bool vm_page_try_remove_write(vm_page_t m); int vm_page_trysbusy(vm_page_t m); void vm_page_unhold_pages(vm_page_t *ma, int count); void vm_page_unswappable(vm_page_t m); void vm_page_unwire(vm_page_t m, uint8_t queue); bool vm_page_unwire_noq(vm_page_t m); void vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr); void vm_page_wire(vm_page_t); bool vm_page_wire_mapped(vm_page_t m); void vm_page_xunbusy_hard(vm_page_t m); void vm_page_set_validclean (vm_page_t, int, int); void vm_page_clear_dirty (vm_page_t, int, int); void vm_page_set_invalid (vm_page_t, int, int); int vm_page_is_valid (vm_page_t, int, int); void vm_page_test_dirty (vm_page_t); vm_page_bits_t vm_page_bits(int base, int size); void vm_page_zero_invalid(vm_page_t m, boolean_t setvalid); void vm_page_free_toq(vm_page_t m); void vm_page_free_pages_toq(struct spglist *free, bool update_wire_count); void vm_page_dirty_KBI(vm_page_t m); void vm_page_lock_KBI(vm_page_t m, const char *file, int line); void vm_page_unlock_KBI(vm_page_t m, const char *file, int line); int vm_page_trylock_KBI(vm_page_t m, const char *file, int line); #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) void vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line); void vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line); #endif #define vm_page_assert_sbusied(m) \ KASSERT(vm_page_sbusied(m), \ ("vm_page_assert_sbusied: page %p not shared busy @ %s:%d", \ (m), __FILE__, __LINE__)) #define vm_page_assert_unbusied(m) \ KASSERT(!vm_page_busied(m), \ ("vm_page_assert_unbusied: page %p busy @ %s:%d", \ (m), __FILE__, __LINE__)) #define vm_page_assert_xbusied(m) \ KASSERT(vm_page_xbusied(m), \ ("vm_page_assert_xbusied: page %p not exclusive busy @ %s:%d", \ (m), __FILE__, __LINE__)) #define vm_page_busied(m) \ ((m)->busy_lock != VPB_UNBUSIED) #define vm_page_sbusy(m) do { \ if (!vm_page_trysbusy(m)) \ panic("%s: page %p failed shared busying", __func__, \ (m)); \ } while (0) #define vm_page_tryxbusy(m) \ (atomic_cmpset_acq_int(&(m)->busy_lock, VPB_UNBUSIED, \ VPB_SINGLE_EXCLUSIVER)) #define vm_page_xbusied(m) \ (((m)->busy_lock & VPB_SINGLE_EXCLUSIVER) != 0) #define vm_page_xbusy(m) do { \ if (!vm_page_tryxbusy(m)) \ panic("%s: page %p failed exclusive busying", __func__, \ (m)); \ } while (0) /* Note: page m's lock must not be owned by the caller. */ #define vm_page_xunbusy(m) do { \ if (!atomic_cmpset_rel_int(&(m)->busy_lock, \ VPB_SINGLE_EXCLUSIVER, VPB_UNBUSIED)) \ vm_page_xunbusy_hard(m); \ } while (0) #ifdef INVARIANTS void vm_page_object_lock_assert(vm_page_t m); #define VM_PAGE_OBJECT_LOCK_ASSERT(m) vm_page_object_lock_assert(m) void vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits); #define VM_PAGE_ASSERT_PGA_WRITEABLE(m, bits) \ vm_page_assert_pga_writeable(m, bits) #else #define VM_PAGE_OBJECT_LOCK_ASSERT(m) (void)0 #define VM_PAGE_ASSERT_PGA_WRITEABLE(m, bits) (void)0 #endif /* * We want to use atomic updates for the aflags field, which is 8 bits wide. * However, not all architectures support atomic operations on 8-bit * destinations. In order that we can easily use a 32-bit operation, we * require that the aflags field be 32-bit aligned. */ _Static_assert(offsetof(struct vm_page, aflags) % sizeof(uint32_t) == 0, "aflags field is not 32-bit aligned"); /* * We want to be able to update the aflags and queue fields atomically in * the same operation. */ _Static_assert(offsetof(struct vm_page, aflags) / sizeof(uint32_t) == offsetof(struct vm_page, queue) / sizeof(uint32_t), "aflags and queue fields do not belong to the same 32-bit word"); _Static_assert(offsetof(struct vm_page, queue) % sizeof(uint32_t) == 2, "queue field is at an unexpected offset"); _Static_assert(sizeof(((struct vm_page *)NULL)->queue) == 1, "queue field has an unexpected size"); #if BYTE_ORDER == LITTLE_ENDIAN #define VM_PAGE_AFLAG_SHIFT 0 #define VM_PAGE_QUEUE_SHIFT 16 #else #define VM_PAGE_AFLAG_SHIFT 24 #define VM_PAGE_QUEUE_SHIFT 8 #endif #define VM_PAGE_QUEUE_MASK (0xff << VM_PAGE_QUEUE_SHIFT) /* * Clear the given bits in the specified page. */ static inline void vm_page_aflag_clear(vm_page_t m, uint8_t bits) { uint32_t *addr, val; /* * The PGA_REFERENCED flag can only be cleared if the page is locked. */ if ((bits & PGA_REFERENCED) != 0) vm_page_assert_locked(m); /* * Access the whole 32-bit word containing the aflags field with an * atomic update. Parallel non-atomic updates to the other fields * within this word are handled properly by the atomic update. */ addr = (void *)&m->aflags; val = bits << VM_PAGE_AFLAG_SHIFT; atomic_clear_32(addr, val); } /* * Set the given bits in the specified page. */ static inline void vm_page_aflag_set(vm_page_t m, uint8_t bits) { uint32_t *addr, val; VM_PAGE_ASSERT_PGA_WRITEABLE(m, bits); /* * Access the whole 32-bit word containing the aflags field with an * atomic update. Parallel non-atomic updates to the other fields * within this word are handled properly by the atomic update. */ addr = (void *)&m->aflags; val = bits << VM_PAGE_AFLAG_SHIFT; atomic_set_32(addr, val); } /* * Atomically update the queue state of the page. The operation fails if * any of the queue flags in "fflags" are set or if the "queue" field of * the page does not match the expected value; if the operation is * successful, the flags in "nflags" are set and all other queue state * flags are cleared. */ static inline bool vm_page_pqstate_cmpset(vm_page_t m, uint32_t oldq, uint32_t newq, uint32_t fflags, uint32_t nflags) { uint32_t *addr, nval, oval, qsmask; fflags <<= VM_PAGE_AFLAG_SHIFT; nflags <<= VM_PAGE_AFLAG_SHIFT; newq <<= VM_PAGE_QUEUE_SHIFT; oldq <<= VM_PAGE_QUEUE_SHIFT; qsmask = ((PGA_DEQUEUE | PGA_REQUEUE | PGA_REQUEUE_HEAD) << VM_PAGE_AFLAG_SHIFT) | VM_PAGE_QUEUE_MASK; addr = (void *)&m->aflags; oval = atomic_load_32(addr); do { if ((oval & fflags) != 0) return (false); if ((oval & VM_PAGE_QUEUE_MASK) != oldq) return (false); nval = (oval & ~qsmask) | nflags | newq; } while (!atomic_fcmpset_32(addr, &oval, nval)); return (true); } /* * vm_page_dirty: * * Set all bits in the page's dirty field. * * The object containing the specified page must be locked if the * call is made from the machine-independent layer. * * See vm_page_clear_dirty_mask(). */ static __inline void vm_page_dirty(vm_page_t m) { /* Use vm_page_dirty_KBI() under INVARIANTS to save memory. */ #if (defined(KLD_MODULE) && !defined(KLD_TIED)) || defined(INVARIANTS) vm_page_dirty_KBI(m); #else m->dirty = VM_PAGE_BITS_ALL; #endif } /* * vm_page_undirty: * * Set page to not be dirty. Note: does not clear pmap modify bits */ static __inline void vm_page_undirty(vm_page_t m) { VM_PAGE_OBJECT_LOCK_ASSERT(m); m->dirty = 0; } static inline void vm_page_replace_checked(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, vm_page_t mold) { vm_page_t mret; mret = vm_page_replace(mnew, object, pindex); KASSERT(mret == mold, ("invalid page replacement, mold=%p, mret=%p", mold, mret)); /* Unused if !INVARIANTS. */ (void)mold; (void)mret; } /* * vm_page_queue: * * Return the index of the queue containing m. This index is guaranteed * not to change while the page lock is held. */ static inline uint8_t vm_page_queue(vm_page_t m) { vm_page_assert_locked(m); if ((m->aflags & PGA_DEQUEUE) != 0) return (PQ_NONE); atomic_thread_fence_acq(); return (m->queue); } static inline bool vm_page_active(vm_page_t m) { return (vm_page_queue(m) == PQ_ACTIVE); } static inline bool vm_page_inactive(vm_page_t m) { return (vm_page_queue(m) == PQ_INACTIVE); } static inline bool vm_page_in_laundry(vm_page_t m) { uint8_t queue; queue = vm_page_queue(m); return (queue == PQ_LAUNDRY || queue == PQ_UNSWAPPABLE); } /* * vm_page_drop: * * Release a reference to a page and return the old reference count. */ static inline u_int vm_page_drop(vm_page_t m, u_int val) { u_int old; /* * Synchronize with vm_page_free_prep(): ensure that all updates to the * page structure are visible before it is freed. */ atomic_thread_fence_rel(); old = atomic_fetchadd_int(&m->ref_count, -val); KASSERT(old != VPRC_BLOCKED, ("vm_page_drop: page %p has an invalid refcount value", m)); return (old); } /* * vm_page_wired: * * Perform a racy check to determine whether a reference prevents the page * from being reclaimable. If the page's object is locked, and the page is * unmapped and unbusied or exclusively busied by the current thread, no * new wirings may be created. */ static inline bool vm_page_wired(vm_page_t m) { return (VPRC_WIRE_COUNT(m->ref_count) > 0); } #endif /* _KERNEL */ #endif /* !_VM_PAGE_ */ Index: head/sys/vm/vm_pageout.c =================================================================== --- head/sys/vm/vm_pageout.c (revision 353534) +++ head/sys/vm/vm_pageout.c (revision 353535) @@ -1,2216 +1,2245 @@ /*- * SPDX-License-Identifier: (BSD-4-Clause AND MIT-CMU) * * Copyright (c) 1991 Regents of the University of California. * All rights reserved. * Copyright (c) 1994 John S. Dyson * All rights reserved. * Copyright (c) 1994 David Greenman * All rights reserved. * Copyright (c) 2005 Yahoo! Technologies Norway AS * All rights reserved. * * This code is derived from software contributed to Berkeley by * The Mach Operating System project at Carnegie-Mellon University. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. All advertising materials mentioning features or use of this software * must display the following acknowledgement: * This product includes software developed by the University of * California, Berkeley and its contributors. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91 * * * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Authors: Avadis Tevanian, Jr., Michael Wayne Young * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. */ /* * The proverbial page-out daemon. */ #include __FBSDID("$FreeBSD$"); #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* * System initialization */ /* the kernel process "vm_pageout"*/ static void vm_pageout(void); static void vm_pageout_init(void); static int vm_pageout_clean(vm_page_t m, int *numpagedout); static int vm_pageout_cluster(vm_page_t m); static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage, int starting_page_shortage); SYSINIT(pagedaemon_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, vm_pageout_init, NULL); struct proc *pageproc; static struct kproc_desc page_kp = { "pagedaemon", vm_pageout, &pageproc }; SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start, &page_kp); SDT_PROVIDER_DEFINE(vm); SDT_PROBE_DEFINE(vm, , , vm__lowmem_scan); /* Pagedaemon activity rates, in subdivisions of one second. */ #define VM_LAUNDER_RATE 10 #define VM_INACT_SCAN_RATE 10 static int vm_pageout_oom_seq = 12; static int vm_pageout_update_period; static int disable_swap_pageouts; static int lowmem_period = 10; static int swapdev_enabled; static int vm_panic_on_oom = 0; SYSCTL_INT(_vm, OID_AUTO, panic_on_oom, CTLFLAG_RWTUN, &vm_panic_on_oom, 0, "panic on out of memory instead of killing the largest process"); SYSCTL_INT(_vm, OID_AUTO, pageout_update_period, CTLFLAG_RWTUN, &vm_pageout_update_period, 0, "Maximum active LRU update period"); SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RWTUN, &lowmem_period, 0, "Low memory callback period"); SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, CTLFLAG_RWTUN, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); static int pageout_lock_miss; SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); SYSCTL_INT(_vm, OID_AUTO, pageout_oom_seq, CTLFLAG_RWTUN, &vm_pageout_oom_seq, 0, "back-to-back calls to oom detector to start OOM"); static int act_scan_laundry_weight = 3; SYSCTL_INT(_vm, OID_AUTO, act_scan_laundry_weight, CTLFLAG_RWTUN, &act_scan_laundry_weight, 0, "weight given to clean vs. dirty pages in active queue scans"); static u_int vm_background_launder_rate = 4096; SYSCTL_UINT(_vm, OID_AUTO, background_launder_rate, CTLFLAG_RWTUN, &vm_background_launder_rate, 0, "background laundering rate, in kilobytes per second"); static u_int vm_background_launder_max = 20 * 1024; SYSCTL_UINT(_vm, OID_AUTO, background_launder_max, CTLFLAG_RWTUN, &vm_background_launder_max, 0, "background laundering cap, in kilobytes"); int vm_pageout_page_count = 32; u_long vm_page_max_user_wired; SYSCTL_ULONG(_vm, OID_AUTO, max_user_wired, CTLFLAG_RW, &vm_page_max_user_wired, 0, "system-wide limit to user-wired page count"); static u_int isqrt(u_int num); static int vm_pageout_launder(struct vm_domain *vmd, int launder, bool in_shortfall); static void vm_pageout_laundry_worker(void *arg); struct scan_state { struct vm_batchqueue bq; struct vm_pagequeue *pq; vm_page_t marker; int maxscan; int scanned; }; static void vm_pageout_init_scan(struct scan_state *ss, struct vm_pagequeue *pq, vm_page_t marker, vm_page_t after, int maxscan) { vm_pagequeue_assert_locked(pq); KASSERT((marker->aflags & PGA_ENQUEUED) == 0, ("marker %p already enqueued", marker)); if (after == NULL) TAILQ_INSERT_HEAD(&pq->pq_pl, marker, plinks.q); else TAILQ_INSERT_AFTER(&pq->pq_pl, after, marker, plinks.q); vm_page_aflag_set(marker, PGA_ENQUEUED); vm_batchqueue_init(&ss->bq); ss->pq = pq; ss->marker = marker; ss->maxscan = maxscan; ss->scanned = 0; vm_pagequeue_unlock(pq); } static void vm_pageout_end_scan(struct scan_state *ss) { struct vm_pagequeue *pq; pq = ss->pq; vm_pagequeue_assert_locked(pq); KASSERT((ss->marker->aflags & PGA_ENQUEUED) != 0, ("marker %p not enqueued", ss->marker)); TAILQ_REMOVE(&pq->pq_pl, ss->marker, plinks.q); vm_page_aflag_clear(ss->marker, PGA_ENQUEUED); pq->pq_pdpages += ss->scanned; } /* * Add a small number of queued pages to a batch queue for later processing * without the corresponding queue lock held. The caller must have enqueued a * marker page at the desired start point for the scan. Pages will be * physically dequeued if the caller so requests. Otherwise, the returned * batch may contain marker pages, and it is up to the caller to handle them. * * When processing the batch queue, vm_page_queue() must be used to * determine whether the page has been logically dequeued by another thread. * Once this check is performed, the page lock guarantees that the page will * not be disassociated from the queue. */ static __always_inline void vm_pageout_collect_batch(struct scan_state *ss, const bool dequeue) { struct vm_pagequeue *pq; vm_page_t m, marker, n; marker = ss->marker; pq = ss->pq; KASSERT((marker->aflags & PGA_ENQUEUED) != 0, ("marker %p not enqueued", ss->marker)); vm_pagequeue_lock(pq); for (m = TAILQ_NEXT(marker, plinks.q); m != NULL && ss->scanned < ss->maxscan && ss->bq.bq_cnt < VM_BATCHQUEUE_SIZE; m = n, ss->scanned++) { n = TAILQ_NEXT(m, plinks.q); if ((m->flags & PG_MARKER) == 0) { KASSERT((m->aflags & PGA_ENQUEUED) != 0, ("page %p not enqueued", m)); KASSERT((m->flags & PG_FICTITIOUS) == 0, ("Fictitious page %p cannot be in page queue", m)); KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("Unmanaged page %p cannot be in page queue", m)); } else if (dequeue) continue; (void)vm_batchqueue_insert(&ss->bq, m); if (dequeue) { TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); vm_page_aflag_clear(m, PGA_ENQUEUED); } } TAILQ_REMOVE(&pq->pq_pl, marker, plinks.q); if (__predict_true(m != NULL)) TAILQ_INSERT_BEFORE(m, marker, plinks.q); else TAILQ_INSERT_TAIL(&pq->pq_pl, marker, plinks.q); if (dequeue) vm_pagequeue_cnt_add(pq, -ss->bq.bq_cnt); vm_pagequeue_unlock(pq); } /* * Return the next page to be scanned, or NULL if the scan is complete. */ static __always_inline vm_page_t vm_pageout_next(struct scan_state *ss, const bool dequeue) { if (ss->bq.bq_cnt == 0) vm_pageout_collect_batch(ss, dequeue); return (vm_batchqueue_pop(&ss->bq)); } /* * Scan for pages at adjacent offsets within the given page's object that are * eligible for laundering, form a cluster of these pages and the given page, * and launder that cluster. */ static int vm_pageout_cluster(vm_page_t m) { vm_object_t object; vm_page_t mc[2 * vm_pageout_page_count], p, pb, ps; vm_pindex_t pindex; int ib, is, page_base, pageout_count; object = m->object; VM_OBJECT_ASSERT_WLOCKED(object); pindex = m->pindex; - vm_page_assert_unbusied(m); + vm_page_assert_xbusied(m); mc[vm_pageout_page_count] = pb = ps = m; pageout_count = 1; page_base = vm_pageout_page_count; ib = 1; is = 1; /* * We can cluster only if the page is not clean, busy, or held, and * the page is in the laundry queue. * * During heavy mmap/modification loads the pageout * daemon can really fragment the underlying file * due to flushing pages out of order and not trying to * align the clusters (which leaves sporadic out-of-order * holes). To solve this problem we do the reverse scan * first and attempt to align our cluster, then do a * forward scan if room remains. */ more: while (ib != 0 && pageout_count < vm_pageout_page_count) { if (ib > pindex) { ib = 0; break; } - if ((p = vm_page_prev(pb)) == NULL || vm_page_busied(p) || - vm_page_wired(p)) { + if ((p = vm_page_prev(pb)) == NULL || + vm_page_tryxbusy(p) == 0) { ib = 0; break; } + if (vm_page_wired(p)) { + ib = 0; + vm_page_xunbusy(p); + break; + } vm_page_test_dirty(p); if (p->dirty == 0) { ib = 0; + vm_page_xunbusy(p); break; } vm_page_lock(p); if (!vm_page_in_laundry(p) || !vm_page_try_remove_write(p)) { vm_page_unlock(p); + vm_page_xunbusy(p); ib = 0; break; } vm_page_unlock(p); mc[--page_base] = pb = p; ++pageout_count; ++ib; /* * We are at an alignment boundary. Stop here, and switch * directions. Do not clear ib. */ if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) break; } while (pageout_count < vm_pageout_page_count && pindex + is < object->size) { - if ((p = vm_page_next(ps)) == NULL || vm_page_busied(p) || - vm_page_wired(p)) + if ((p = vm_page_next(ps)) == NULL || + vm_page_tryxbusy(p) == 0) break; + if (vm_page_wired(p)) { + vm_page_xunbusy(p); + break; + } vm_page_test_dirty(p); - if (p->dirty == 0) + if (p->dirty == 0) { + vm_page_xunbusy(p); break; + } vm_page_lock(p); if (!vm_page_in_laundry(p) || !vm_page_try_remove_write(p)) { vm_page_unlock(p); + vm_page_xunbusy(p); break; } vm_page_unlock(p); mc[page_base + pageout_count] = ps = p; ++pageout_count; ++is; } /* * If we exhausted our forward scan, continue with the reverse scan * when possible, even past an alignment boundary. This catches * boundary conditions. */ if (ib != 0 && pageout_count < vm_pageout_page_count) goto more; return (vm_pageout_flush(&mc[page_base], pageout_count, VM_PAGER_PUT_NOREUSE, 0, NULL, NULL)); } /* * vm_pageout_flush() - launder the given pages * * The given pages are laundered. Note that we setup for the start of * I/O ( i.e. busy the page ), mark it read-only, and bump the object * reference count all in here rather then in the parent. If we want * the parent to do more sophisticated things we may have to change * the ordering. * * Returned runlen is the count of pages between mreq and first * page after mreq with status VM_PAGER_AGAIN. * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL * for any page in runlen set. */ int vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen, boolean_t *eio) { vm_object_t object = mc[0]->object; int pageout_status[count]; int numpagedout = 0; int i, runlen; VM_OBJECT_ASSERT_WLOCKED(object); /* - * Initiate I/O. Mark the pages busy and verify that they're valid - * and read-only. + * Initiate I/O. Mark the pages shared busy and verify that they're + * valid and read-only. * * We do not have to fixup the clean/dirty bits here... we can * allow the pager to do it after the I/O completes. * * NOTE! mc[i]->dirty may be partial or fragmented due to an * edge case with file fragments. */ for (i = 0; i < count; i++) { KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, ("vm_pageout_flush: partially invalid page %p index %d/%d", mc[i], i, count)); KASSERT((mc[i]->aflags & PGA_WRITEABLE) == 0, ("vm_pageout_flush: writeable page %p", mc[i])); - vm_page_sbusy(mc[i]); + vm_page_busy_downgrade(mc[i]); } vm_object_pip_add(object, count); vm_pager_put_pages(object, mc, count, flags, pageout_status); runlen = count - mreq; if (eio != NULL) *eio = FALSE; for (i = 0; i < count; i++) { vm_page_t mt = mc[i]; KASSERT(pageout_status[i] == VM_PAGER_PEND || !pmap_page_is_write_mapped(mt), ("vm_pageout_flush: page %p is not write protected", mt)); switch (pageout_status[i]) { case VM_PAGER_OK: vm_page_lock(mt); if (vm_page_in_laundry(mt)) vm_page_deactivate_noreuse(mt); vm_page_unlock(mt); /* FALLTHROUGH */ case VM_PAGER_PEND: numpagedout++; break; case VM_PAGER_BAD: /* * The page is outside the object's range. We pretend * that the page out worked and clean the page, so the * changes will be lost if the page is reclaimed by * the page daemon. */ vm_page_undirty(mt); vm_page_lock(mt); if (vm_page_in_laundry(mt)) vm_page_deactivate_noreuse(mt); vm_page_unlock(mt); break; case VM_PAGER_ERROR: case VM_PAGER_FAIL: /* * If the page couldn't be paged out to swap because the * pager wasn't able to find space, place the page in * the PQ_UNSWAPPABLE holding queue. This is an * optimization that prevents the page daemon from * wasting CPU cycles on pages that cannot be reclaimed * becase no swap device is configured. * * Otherwise, reactivate the page so that it doesn't * clog the laundry and inactive queues. (We will try * paging it out again later.) */ vm_page_lock(mt); if (object->type == OBJT_SWAP && pageout_status[i] == VM_PAGER_FAIL) { vm_page_unswappable(mt); numpagedout++; } else vm_page_activate(mt); vm_page_unlock(mt); if (eio != NULL && i >= mreq && i - mreq < runlen) *eio = TRUE; break; case VM_PAGER_AGAIN: if (i >= mreq && i - mreq < runlen) runlen = i - mreq; break; } /* * If the operation is still going, leave the page busy to * block all other accesses. Also, leave the paging in * progress indicator set so that we don't attempt an object * collapse. */ if (pageout_status[i] != VM_PAGER_PEND) { vm_object_pip_wakeup(object); vm_page_sunbusy(mt); } } if (prunlen != NULL) *prunlen = runlen; return (numpagedout); } static void vm_pageout_swapon(void *arg __unused, struct swdevt *sp __unused) { atomic_store_rel_int(&swapdev_enabled, 1); } static void vm_pageout_swapoff(void *arg __unused, struct swdevt *sp __unused) { if (swap_pager_nswapdev() == 1) atomic_store_rel_int(&swapdev_enabled, 0); } /* * Attempt to acquire all of the necessary locks to launder a page and * then call through the clustering layer to PUTPAGES. Wait a short * time for a vnode lock. * * Requires the page and object lock on entry, releases both before return. * Returns 0 on success and an errno otherwise. */ static int vm_pageout_clean(vm_page_t m, int *numpagedout) { struct vnode *vp; struct mount *mp; vm_object_t object; vm_pindex_t pindex; int error, lockmode; vm_page_assert_locked(m); object = m->object; VM_OBJECT_ASSERT_WLOCKED(object); error = 0; vp = NULL; mp = NULL; /* * The object is already known NOT to be dead. It * is possible for the vget() to block the whole * pageout daemon, but the new low-memory handling * code should prevent it. * * We can't wait forever for the vnode lock, we might * deadlock due to a vn_read() getting stuck in * vm_wait while holding this vnode. We skip the * vnode if we can't get it in a reasonable amount * of time. */ if (object->type == OBJT_VNODE) { vm_page_unlock(m); + vm_page_xunbusy(m); vp = object->handle; if (vp->v_type == VREG && vn_start_write(vp, &mp, V_NOWAIT) != 0) { mp = NULL; error = EDEADLK; goto unlock_all; } KASSERT(mp != NULL, ("vp %p with NULL v_mount", vp)); vm_object_reference_locked(object); pindex = m->pindex; VM_OBJECT_WUNLOCK(object); lockmode = MNT_SHARED_WRITES(vp->v_mount) ? LK_SHARED : LK_EXCLUSIVE; if (vget(vp, lockmode | LK_TIMELOCK, curthread)) { vp = NULL; error = EDEADLK; goto unlock_mp; } VM_OBJECT_WLOCK(object); /* * Ensure that the object and vnode were not disassociated * while locks were dropped. */ if (vp->v_object != object) { error = ENOENT; goto unlock_all; } vm_page_lock(m); /* * While the object and page were unlocked, the page * may have been: * (1) moved to a different queue, * (2) reallocated to a different object, * (3) reallocated to a different offset, or * (4) cleaned. */ if (!vm_page_in_laundry(m) || m->object != object || m->pindex != pindex || m->dirty == 0) { vm_page_unlock(m); error = ENXIO; goto unlock_all; } /* * The page may have been busied while the object and page * locks were released. */ - if (vm_page_busied(m)) { + if (vm_page_tryxbusy(m) == 0) { vm_page_unlock(m); error = EBUSY; goto unlock_all; } } /* * Remove all writeable mappings, failing if the page is wired. */ if (!vm_page_try_remove_write(m)) { + vm_page_xunbusy(m); vm_page_unlock(m); error = EBUSY; goto unlock_all; } vm_page_unlock(m); /* * If a page is dirty, then it is either being washed * (but not yet cleaned) or it is still in the * laundry. If it is still in the laundry, then we * start the cleaning operation. */ if ((*numpagedout = vm_pageout_cluster(m)) == 0) error = EIO; unlock_all: VM_OBJECT_WUNLOCK(object); unlock_mp: vm_page_lock_assert(m, MA_NOTOWNED); if (mp != NULL) { if (vp != NULL) vput(vp); vm_object_deallocate(object); vn_finished_write(mp); } return (error); } /* * Attempt to launder the specified number of pages. * * Returns the number of pages successfully laundered. */ static int vm_pageout_launder(struct vm_domain *vmd, int launder, bool in_shortfall) { struct scan_state ss; struct vm_pagequeue *pq; struct mtx *mtx; vm_object_t object; vm_page_t m, marker; int act_delta, error, numpagedout, queue, starting_target; int vnodes_skipped; bool pageout_ok; mtx = NULL; object = NULL; starting_target = launder; vnodes_skipped = 0; /* * Scan the laundry queues for pages eligible to be laundered. We stop * once the target number of dirty pages have been laundered, or once * we've reached the end of the queue. A single iteration of this loop * may cause more than one page to be laundered because of clustering. * * As an optimization, we avoid laundering from PQ_UNSWAPPABLE when no * swap devices are configured. */ if (atomic_load_acq_int(&swapdev_enabled)) queue = PQ_UNSWAPPABLE; else queue = PQ_LAUNDRY; scan: marker = &vmd->vmd_markers[queue]; pq = &vmd->vmd_pagequeues[queue]; vm_pagequeue_lock(pq); vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt); while (launder > 0 && (m = vm_pageout_next(&ss, false)) != NULL) { if (__predict_false((m->flags & PG_MARKER) != 0)) continue; vm_page_change_lock(m, &mtx); recheck: /* * The page may have been disassociated from the queue * or even freed while locks were dropped. We thus must be * careful whenever modifying page state. Once the object lock * has been acquired, we have a stable reference to the page. */ if (vm_page_queue(m) != queue) continue; /* * A requeue was requested, so this page gets a second * chance. */ if ((m->aflags & PGA_REQUEUE) != 0) { vm_page_pqbatch_submit(m, queue); continue; } /* * Wired pages may not be freed. Complete their removal * from the queue now to avoid needless revisits during * future scans. This check is racy and must be reverified once * we hold the object lock and have verified that the page * is not busy. */ if (vm_page_wired(m)) { vm_page_dequeue_deferred(m); continue; } if (object != m->object) { if (object != NULL) VM_OBJECT_WUNLOCK(object); /* * A page's object pointer may be set to NULL before * the object lock is acquired. */ object = (vm_object_t)atomic_load_ptr(&m->object); if (object != NULL && !VM_OBJECT_TRYWLOCK(object)) { mtx_unlock(mtx); /* Depends on type-stability. */ VM_OBJECT_WLOCK(object); mtx_lock(mtx); goto recheck; } } if (__predict_false(m->object == NULL)) /* * The page has been removed from its object. */ continue; KASSERT(m->object == object, ("page %p does not belong to %p", m, object)); - if (vm_page_busied(m)) + if (vm_page_tryxbusy(m) == 0) continue; /* * Re-check for wirings now that we hold the object lock and * have verified that the page is unbusied. If the page is * mapped, it may still be wired by pmap lookups. The call to * vm_page_try_remove_all() below atomically checks for such * wirings and removes mappings. If the page is unmapped, the * wire count is guaranteed not to increase. */ if (__predict_false(vm_page_wired(m))) { + vm_page_xunbusy(m); vm_page_dequeue_deferred(m); continue; } /* * Invalid pages can be easily freed. They cannot be * mapped; vm_page_free() asserts this. */ if (m->valid == 0) goto free_page; /* * If the page has been referenced and the object is not dead, * reactivate or requeue the page depending on whether the * object is mapped. * * Test PGA_REFERENCED after calling pmap_ts_referenced() so * that a reference from a concurrently destroyed mapping is * observed here and now. */ if (object->ref_count != 0) act_delta = pmap_ts_referenced(m); else { KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m)); act_delta = 0; } if ((m->aflags & PGA_REFERENCED) != 0) { vm_page_aflag_clear(m, PGA_REFERENCED); act_delta++; } if (act_delta != 0) { if (object->ref_count != 0) { + vm_page_xunbusy(m); VM_CNT_INC(v_reactivated); vm_page_activate(m); /* * Increase the activation count if the page * was referenced while in the laundry queue. * This makes it less likely that the page will * be returned prematurely to the inactive * queue. */ m->act_count += act_delta + ACT_ADVANCE; /* * If this was a background laundering, count * activated pages towards our target. The * purpose of background laundering is to ensure * that pages are eventually cycled through the * laundry queue, and an activation is a valid * way out. */ if (!in_shortfall) launder--; continue; } else if ((object->flags & OBJ_DEAD) == 0) { + vm_page_xunbusy(m); vm_page_requeue(m); continue; } } /* * If the page appears to be clean at the machine-independent * layer, then remove all of its mappings from the pmap in * anticipation of freeing it. If, however, any of the page's * mappings allow write access, then the page may still be * modified until the last of those mappings are removed. */ if (object->ref_count != 0) { vm_page_test_dirty(m); if (m->dirty == 0 && !vm_page_try_remove_all(m)) { + vm_page_xunbusy(m); vm_page_dequeue_deferred(m); continue; } } /* * Clean pages are freed, and dirty pages are paged out unless * they belong to a dead object. Requeueing dirty pages from * dead objects is pointless, as they are being paged out and * freed by the thread that destroyed the object. */ if (m->dirty == 0) { free_page: vm_page_free(m); VM_CNT_INC(v_dfree); } else if ((object->flags & OBJ_DEAD) == 0) { if (object->type != OBJT_SWAP && object->type != OBJT_DEFAULT) pageout_ok = true; else if (disable_swap_pageouts) pageout_ok = false; else pageout_ok = true; if (!pageout_ok) { + vm_page_xunbusy(m); vm_page_requeue(m); continue; } /* * Form a cluster with adjacent, dirty pages from the * same object, and page out that entire cluster. * * The adjacent, dirty pages must also be in the * laundry. However, their mappings are not checked * for new references. Consequently, a recently * referenced page may be paged out. However, that * page will not be prematurely reclaimed. After page * out, the page will be placed in the inactive queue, * where any new references will be detected and the * page reactivated. */ error = vm_pageout_clean(m, &numpagedout); if (error == 0) { launder -= numpagedout; ss.scanned += numpagedout; } else if (error == EDEADLK) { pageout_lock_miss++; vnodes_skipped++; } mtx = NULL; object = NULL; - } + } else + vm_page_xunbusy(m); } if (mtx != NULL) { mtx_unlock(mtx); mtx = NULL; } if (object != NULL) { VM_OBJECT_WUNLOCK(object); object = NULL; } vm_pagequeue_lock(pq); vm_pageout_end_scan(&ss); vm_pagequeue_unlock(pq); if (launder > 0 && queue == PQ_UNSWAPPABLE) { queue = PQ_LAUNDRY; goto scan; } /* * Wakeup the sync daemon if we skipped a vnode in a writeable object * and we didn't launder enough pages. */ if (vnodes_skipped > 0 && launder > 0) (void)speedup_syncer(); return (starting_target - launder); } /* * Compute the integer square root. */ static u_int isqrt(u_int num) { u_int bit, root, tmp; bit = num != 0 ? (1u << ((fls(num) - 1) & ~1)) : 0; root = 0; while (bit != 0) { tmp = root + bit; root >>= 1; if (num >= tmp) { num -= tmp; root += bit; } bit >>= 2; } return (root); } /* * Perform the work of the laundry thread: periodically wake up and determine * whether any pages need to be laundered. If so, determine the number of pages * that need to be laundered, and launder them. */ static void vm_pageout_laundry_worker(void *arg) { struct vm_domain *vmd; struct vm_pagequeue *pq; uint64_t nclean, ndirty, nfreed; int domain, last_target, launder, shortfall, shortfall_cycle, target; bool in_shortfall; domain = (uintptr_t)arg; vmd = VM_DOMAIN(domain); pq = &vmd->vmd_pagequeues[PQ_LAUNDRY]; KASSERT(vmd->vmd_segs != 0, ("domain without segments")); shortfall = 0; in_shortfall = false; shortfall_cycle = 0; last_target = target = 0; nfreed = 0; /* * Calls to these handlers are serialized by the swap syscall lock. */ (void)EVENTHANDLER_REGISTER(swapon, vm_pageout_swapon, vmd, EVENTHANDLER_PRI_ANY); (void)EVENTHANDLER_REGISTER(swapoff, vm_pageout_swapoff, vmd, EVENTHANDLER_PRI_ANY); /* * The pageout laundry worker is never done, so loop forever. */ for (;;) { KASSERT(target >= 0, ("negative target %d", target)); KASSERT(shortfall_cycle >= 0, ("negative cycle %d", shortfall_cycle)); launder = 0; /* * First determine whether we need to launder pages to meet a * shortage of free pages. */ if (shortfall > 0) { in_shortfall = true; shortfall_cycle = VM_LAUNDER_RATE / VM_INACT_SCAN_RATE; target = shortfall; } else if (!in_shortfall) goto trybackground; else if (shortfall_cycle == 0 || vm_laundry_target(vmd) <= 0) { /* * We recently entered shortfall and began laundering * pages. If we have completed that laundering run * (and we are no longer in shortfall) or we have met * our laundry target through other activity, then we * can stop laundering pages. */ in_shortfall = false; target = 0; goto trybackground; } launder = target / shortfall_cycle--; goto dolaundry; /* * There's no immediate need to launder any pages; see if we * meet the conditions to perform background laundering: * * 1. The ratio of dirty to clean inactive pages exceeds the * background laundering threshold, or * 2. we haven't yet reached the target of the current * background laundering run. * * The background laundering threshold is not a constant. * Instead, it is a slowly growing function of the number of * clean pages freed by the page daemon since the last * background laundering. Thus, as the ratio of dirty to * clean inactive pages grows, the amount of memory pressure * required to trigger laundering decreases. We ensure * that the threshold is non-zero after an inactive queue * scan, even if that scan failed to free a single clean page. */ trybackground: nclean = vmd->vmd_free_count + vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt; ndirty = vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt; if (target == 0 && ndirty * isqrt(howmany(nfreed + 1, vmd->vmd_free_target - vmd->vmd_free_min)) >= nclean) { target = vmd->vmd_background_launder_target; } /* * We have a non-zero background laundering target. If we've * laundered up to our maximum without observing a page daemon * request, just stop. This is a safety belt that ensures we * don't launder an excessive amount if memory pressure is low * and the ratio of dirty to clean pages is large. Otherwise, * proceed at the background laundering rate. */ if (target > 0) { if (nfreed > 0) { nfreed = 0; last_target = target; } else if (last_target - target >= vm_background_launder_max * PAGE_SIZE / 1024) { target = 0; } launder = vm_background_launder_rate * PAGE_SIZE / 1024; launder /= VM_LAUNDER_RATE; if (launder > target) launder = target; } dolaundry: if (launder > 0) { /* * Because of I/O clustering, the number of laundered * pages could exceed "target" by the maximum size of * a cluster minus one. */ target -= min(vm_pageout_launder(vmd, launder, in_shortfall), target); pause("laundp", hz / VM_LAUNDER_RATE); } /* * If we're not currently laundering pages and the page daemon * hasn't posted a new request, sleep until the page daemon * kicks us. */ vm_pagequeue_lock(pq); if (target == 0 && vmd->vmd_laundry_request == VM_LAUNDRY_IDLE) (void)mtx_sleep(&vmd->vmd_laundry_request, vm_pagequeue_lockptr(pq), PVM, "launds", 0); /* * If the pagedaemon has indicated that it's in shortfall, start * a shortfall laundering unless we're already in the middle of * one. This may preempt a background laundering. */ if (vmd->vmd_laundry_request == VM_LAUNDRY_SHORTFALL && (!in_shortfall || shortfall_cycle == 0)) { shortfall = vm_laundry_target(vmd) + vmd->vmd_pageout_deficit; target = 0; } else shortfall = 0; if (target == 0) vmd->vmd_laundry_request = VM_LAUNDRY_IDLE; nfreed += vmd->vmd_clean_pages_freed; vmd->vmd_clean_pages_freed = 0; vm_pagequeue_unlock(pq); } } /* * Compute the number of pages we want to try to move from the * active queue to either the inactive or laundry queue. * * When scanning active pages during a shortage, we make clean pages * count more heavily towards the page shortage than dirty pages. * This is because dirty pages must be laundered before they can be * reused and thus have less utility when attempting to quickly * alleviate a free page shortage. However, this weighting also * causes the scan to deactivate dirty pages more aggressively, * improving the effectiveness of clustering. */ static int vm_pageout_active_target(struct vm_domain *vmd) { int shortage; shortage = vmd->vmd_inactive_target + vm_paging_target(vmd) - (vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt + vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt / act_scan_laundry_weight); shortage *= act_scan_laundry_weight; return (shortage); } /* * Scan the active queue. If there is no shortage of inactive pages, scan a * small portion of the queue in order to maintain quasi-LRU. */ static void vm_pageout_scan_active(struct vm_domain *vmd, int page_shortage) { struct scan_state ss; struct mtx *mtx; vm_object_t object; vm_page_t m, marker; struct vm_pagequeue *pq; long min_scan; int act_delta, max_scan, scan_tick; marker = &vmd->vmd_markers[PQ_ACTIVE]; pq = &vmd->vmd_pagequeues[PQ_ACTIVE]; vm_pagequeue_lock(pq); /* * If we're just idle polling attempt to visit every * active page within 'update_period' seconds. */ scan_tick = ticks; if (vm_pageout_update_period != 0) { min_scan = pq->pq_cnt; min_scan *= scan_tick - vmd->vmd_last_active_scan; min_scan /= hz * vm_pageout_update_period; } else min_scan = 0; if (min_scan > 0 || (page_shortage > 0 && pq->pq_cnt > 0)) vmd->vmd_last_active_scan = scan_tick; /* * Scan the active queue for pages that can be deactivated. Update * the per-page activity counter and use it to identify deactivation * candidates. Held pages may be deactivated. * * To avoid requeuing each page that remains in the active queue, we * implement the CLOCK algorithm. To keep the implementation of the * enqueue operation consistent for all page queues, we use two hands, * represented by marker pages. Scans begin at the first hand, which * precedes the second hand in the queue. When the two hands meet, * they are moved back to the head and tail of the queue, respectively, * and scanning resumes. */ max_scan = page_shortage > 0 ? pq->pq_cnt : min_scan; mtx = NULL; act_scan: vm_pageout_init_scan(&ss, pq, marker, &vmd->vmd_clock[0], max_scan); while ((m = vm_pageout_next(&ss, false)) != NULL) { if (__predict_false(m == &vmd->vmd_clock[1])) { vm_pagequeue_lock(pq); TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q); TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[1], plinks.q); TAILQ_INSERT_HEAD(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q); TAILQ_INSERT_TAIL(&pq->pq_pl, &vmd->vmd_clock[1], plinks.q); max_scan -= ss.scanned; vm_pageout_end_scan(&ss); goto act_scan; } if (__predict_false((m->flags & PG_MARKER) != 0)) continue; vm_page_change_lock(m, &mtx); /* * The page may have been disassociated from the queue * or even freed while locks were dropped. We thus must be * careful whenever modifying page state. Once the object lock * has been acquired, we have a stable reference to the page. */ if (vm_page_queue(m) != PQ_ACTIVE) continue; /* * Wired pages are dequeued lazily. */ if (vm_page_wired(m)) { vm_page_dequeue_deferred(m); continue; } /* * A page's object pointer may be set to NULL before * the object lock is acquired. */ object = (vm_object_t)atomic_load_ptr(&m->object); if (__predict_false(object == NULL)) /* * The page has been removed from its object. */ continue; /* * Check to see "how much" the page has been used. * * Test PGA_REFERENCED after calling pmap_ts_referenced() so * that a reference from a concurrently destroyed mapping is * observed here and now. * * Perform an unsynchronized object ref count check. While * the page lock ensures that the page is not reallocated to * another object, in particular, one with unmanaged mappings * that cannot support pmap_ts_referenced(), two races are, * nonetheless, possible: * 1) The count was transitioning to zero, but we saw a non- * zero value. pmap_ts_referenced() will return zero * because the page is not mapped. * 2) The count was transitioning to one, but we saw zero. * This race delays the detection of a new reference. At * worst, we will deactivate and reactivate the page. */ if (object->ref_count != 0) act_delta = pmap_ts_referenced(m); else act_delta = 0; if ((m->aflags & PGA_REFERENCED) != 0) { vm_page_aflag_clear(m, PGA_REFERENCED); act_delta++; } /* * Advance or decay the act_count based on recent usage. */ if (act_delta != 0) { m->act_count += ACT_ADVANCE + act_delta; if (m->act_count > ACT_MAX) m->act_count = ACT_MAX; } else m->act_count -= min(m->act_count, ACT_DECLINE); if (m->act_count == 0) { /* * When not short for inactive pages, let dirty pages go * through the inactive queue before moving to the * laundry queues. This gives them some extra time to * be reactivated, potentially avoiding an expensive * pageout. However, during a page shortage, the * inactive queue is necessarily small, and so dirty * pages would only spend a trivial amount of time in * the inactive queue. Therefore, we might as well * place them directly in the laundry queue to reduce * queuing overhead. */ if (page_shortage <= 0) { vm_page_swapqueue(m, PQ_ACTIVE, PQ_INACTIVE); } else { /* * Calling vm_page_test_dirty() here would * require acquisition of the object's write * lock. However, during a page shortage, * directing dirty pages into the laundry * queue is only an optimization and not a * requirement. Therefore, we simply rely on * the opportunistic updates to the page's * dirty field by the pmap. */ if (m->dirty == 0) { vm_page_swapqueue(m, PQ_ACTIVE, PQ_INACTIVE); page_shortage -= act_scan_laundry_weight; } else { vm_page_swapqueue(m, PQ_ACTIVE, PQ_LAUNDRY); page_shortage--; } } } } if (mtx != NULL) { mtx_unlock(mtx); mtx = NULL; } vm_pagequeue_lock(pq); TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q); TAILQ_INSERT_AFTER(&pq->pq_pl, marker, &vmd->vmd_clock[0], plinks.q); vm_pageout_end_scan(&ss); vm_pagequeue_unlock(pq); } static int vm_pageout_reinsert_inactive_page(struct scan_state *ss, vm_page_t m) { struct vm_domain *vmd; if (m->queue != PQ_INACTIVE || (m->aflags & PGA_ENQUEUED) != 0) return (0); vm_page_aflag_set(m, PGA_ENQUEUED); if ((m->aflags & PGA_REQUEUE_HEAD) != 0) { vmd = vm_pagequeue_domain(m); TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q); vm_page_aflag_clear(m, PGA_REQUEUE | PGA_REQUEUE_HEAD); } else if ((m->aflags & PGA_REQUEUE) != 0) { TAILQ_INSERT_TAIL(&ss->pq->pq_pl, m, plinks.q); vm_page_aflag_clear(m, PGA_REQUEUE | PGA_REQUEUE_HEAD); } else TAILQ_INSERT_BEFORE(ss->marker, m, plinks.q); return (1); } /* * Re-add stuck pages to the inactive queue. We will examine them again * during the next scan. If the queue state of a page has changed since * it was physically removed from the page queue in * vm_pageout_collect_batch(), don't do anything with that page. */ static void vm_pageout_reinsert_inactive(struct scan_state *ss, struct vm_batchqueue *bq, vm_page_t m) { struct vm_pagequeue *pq; int delta; delta = 0; pq = ss->pq; if (m != NULL) { if (vm_batchqueue_insert(bq, m)) return; vm_pagequeue_lock(pq); delta += vm_pageout_reinsert_inactive_page(ss, m); } else vm_pagequeue_lock(pq); while ((m = vm_batchqueue_pop(bq)) != NULL) delta += vm_pageout_reinsert_inactive_page(ss, m); vm_pagequeue_cnt_add(pq, delta); vm_pagequeue_unlock(pq); vm_batchqueue_init(bq); } /* * Attempt to reclaim the requested number of pages from the inactive queue. * Returns true if the shortage was addressed. */ static int vm_pageout_scan_inactive(struct vm_domain *vmd, int shortage, int *addl_shortage) { struct scan_state ss; struct vm_batchqueue rq; struct mtx *mtx; vm_page_t m, marker; struct vm_pagequeue *pq; vm_object_t object; int act_delta, addl_page_shortage, deficit, page_shortage; int starting_page_shortage; /* * The addl_page_shortage is an estimate of the number of temporarily * stuck pages in the inactive queue. In other words, the * number of pages from the inactive count that should be * discounted in setting the target for the active queue scan. */ addl_page_shortage = 0; /* * vmd_pageout_deficit counts the number of pages requested in * allocations that failed because of a free page shortage. We assume * that the allocations will be reattempted and thus include the deficit * in our scan target. */ deficit = atomic_readandclear_int(&vmd->vmd_pageout_deficit); starting_page_shortage = page_shortage = shortage + deficit; mtx = NULL; object = NULL; vm_batchqueue_init(&rq); /* * Start scanning the inactive queue for pages that we can free. The * scan will stop when we reach the target or we have scanned the * entire queue. (Note that m->act_count is not used to make * decisions for the inactive queue, only for the active queue.) */ marker = &vmd->vmd_markers[PQ_INACTIVE]; pq = &vmd->vmd_pagequeues[PQ_INACTIVE]; vm_pagequeue_lock(pq); vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt); while (page_shortage > 0 && (m = vm_pageout_next(&ss, true)) != NULL) { KASSERT((m->flags & PG_MARKER) == 0, ("marker page %p was dequeued", m)); vm_page_change_lock(m, &mtx); recheck: /* * The page may have been disassociated from the queue * or even freed while locks were dropped. We thus must be * careful whenever modifying page state. Once the object lock * has been acquired, we have a stable reference to the page. */ if (vm_page_queue(m) != PQ_INACTIVE) { addl_page_shortage++; continue; } /* * The page was re-enqueued after the page queue lock was * dropped, or a requeue was requested. This page gets a second * chance. */ if ((m->aflags & (PGA_ENQUEUED | PGA_REQUEUE | PGA_REQUEUE_HEAD)) != 0) goto reinsert; /* * Wired pages may not be freed. Complete their removal * from the queue now to avoid needless revisits during * future scans. This check is racy and must be reverified once * we hold the object lock and have verified that the page * is not busy. */ if (vm_page_wired(m)) { vm_page_dequeue_deferred(m); continue; } if (object != m->object) { if (object != NULL) VM_OBJECT_WUNLOCK(object); /* * A page's object pointer may be set to NULL before * the object lock is acquired. */ object = (vm_object_t)atomic_load_ptr(&m->object); if (object != NULL && !VM_OBJECT_TRYWLOCK(object)) { mtx_unlock(mtx); /* Depends on type-stability. */ VM_OBJECT_WLOCK(object); mtx_lock(mtx); goto recheck; } } if (__predict_false(m->object == NULL)) /* * The page has been removed from its object. */ continue; KASSERT(m->object == object, ("page %p does not belong to %p", m, object)); - if (vm_page_busied(m)) { + if (vm_page_tryxbusy(m) == 0) { /* * Don't mess with busy pages. Leave them at * the front of the queue. Most likely, they * are being paged out and will leave the * queue shortly after the scan finishes. So, * they ought to be discounted from the * inactive count. */ addl_page_shortage++; goto reinsert; } /* * Re-check for wirings now that we hold the object lock and * have verified that the page is unbusied. If the page is * mapped, it may still be wired by pmap lookups. The call to * vm_page_try_remove_all() below atomically checks for such * wirings and removes mappings. If the page is unmapped, the * wire count is guaranteed not to increase. */ if (__predict_false(vm_page_wired(m))) { + vm_page_xunbusy(m); vm_page_dequeue_deferred(m); continue; } /* * Invalid pages can be easily freed. They cannot be * mapped, vm_page_free() asserts this. */ if (m->valid == 0) goto free_page; /* * If the page has been referenced and the object is not dead, * reactivate or requeue the page depending on whether the * object is mapped. * * Test PGA_REFERENCED after calling pmap_ts_referenced() so * that a reference from a concurrently destroyed mapping is * observed here and now. */ if (object->ref_count != 0) act_delta = pmap_ts_referenced(m); else { KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m)); act_delta = 0; } if ((m->aflags & PGA_REFERENCED) != 0) { vm_page_aflag_clear(m, PGA_REFERENCED); act_delta++; } if (act_delta != 0) { if (object->ref_count != 0) { + vm_page_xunbusy(m); VM_CNT_INC(v_reactivated); vm_page_activate(m); /* * Increase the activation count if the page * was referenced while in the inactive queue. * This makes it less likely that the page will * be returned prematurely to the inactive * queue. */ m->act_count += act_delta + ACT_ADVANCE; continue; } else if ((object->flags & OBJ_DEAD) == 0) { + vm_page_xunbusy(m); vm_page_aflag_set(m, PGA_REQUEUE); goto reinsert; } } /* * If the page appears to be clean at the machine-independent * layer, then remove all of its mappings from the pmap in * anticipation of freeing it. If, however, any of the page's * mappings allow write access, then the page may still be * modified until the last of those mappings are removed. */ if (object->ref_count != 0) { vm_page_test_dirty(m); if (m->dirty == 0 && !vm_page_try_remove_all(m)) { + vm_page_xunbusy(m); vm_page_dequeue_deferred(m); continue; } } /* * Clean pages can be freed, but dirty pages must be sent back * to the laundry, unless they belong to a dead object. * Requeueing dirty pages from dead objects is pointless, as * they are being paged out and freed by the thread that * destroyed the object. */ if (m->dirty == 0) { free_page: /* * Because we dequeued the page and have already * checked for concurrent dequeue and enqueue * requests, we can safely disassociate the page * from the inactive queue. */ KASSERT((m->aflags & PGA_QUEUE_STATE_MASK) == 0, ("page %p has queue state", m)); m->queue = PQ_NONE; vm_page_free(m); page_shortage--; - } else if ((object->flags & OBJ_DEAD) == 0) + continue; + } + vm_page_xunbusy(m); + if ((object->flags & OBJ_DEAD) == 0) vm_page_launder(m); continue; reinsert: vm_pageout_reinsert_inactive(&ss, &rq, m); } if (mtx != NULL) mtx_unlock(mtx); if (object != NULL) VM_OBJECT_WUNLOCK(object); vm_pageout_reinsert_inactive(&ss, &rq, NULL); vm_pageout_reinsert_inactive(&ss, &ss.bq, NULL); vm_pagequeue_lock(pq); vm_pageout_end_scan(&ss); vm_pagequeue_unlock(pq); VM_CNT_ADD(v_dfree, starting_page_shortage - page_shortage); /* * Wake up the laundry thread so that it can perform any needed * laundering. If we didn't meet our target, we're in shortfall and * need to launder more aggressively. If PQ_LAUNDRY is empty and no * swap devices are configured, the laundry thread has no work to do, so * don't bother waking it up. * * The laundry thread uses the number of inactive queue scans elapsed * since the last laundering to determine whether to launder again, so * keep count. */ if (starting_page_shortage > 0) { pq = &vmd->vmd_pagequeues[PQ_LAUNDRY]; vm_pagequeue_lock(pq); if (vmd->vmd_laundry_request == VM_LAUNDRY_IDLE && (pq->pq_cnt > 0 || atomic_load_acq_int(&swapdev_enabled))) { if (page_shortage > 0) { vmd->vmd_laundry_request = VM_LAUNDRY_SHORTFALL; VM_CNT_INC(v_pdshortfalls); } else if (vmd->vmd_laundry_request != VM_LAUNDRY_SHORTFALL) vmd->vmd_laundry_request = VM_LAUNDRY_BACKGROUND; wakeup(&vmd->vmd_laundry_request); } vmd->vmd_clean_pages_freed += starting_page_shortage - page_shortage; vm_pagequeue_unlock(pq); } /* * Wakeup the swapout daemon if we didn't free the targeted number of * pages. */ if (page_shortage > 0) vm_swapout_run(); /* * If the inactive queue scan fails repeatedly to meet its * target, kill the largest process. */ vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage); /* * Reclaim pages by swapping out idle processes, if configured to do so. */ vm_swapout_run_idle(); /* * See the description of addl_page_shortage above. */ *addl_shortage = addl_page_shortage + deficit; return (page_shortage <= 0); } static int vm_pageout_oom_vote; /* * The pagedaemon threads randlomly select one to perform the * OOM. Trying to kill processes before all pagedaemons * failed to reach free target is premature. */ static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage, int starting_page_shortage) { int old_vote; if (starting_page_shortage <= 0 || starting_page_shortage != page_shortage) vmd->vmd_oom_seq = 0; else vmd->vmd_oom_seq++; if (vmd->vmd_oom_seq < vm_pageout_oom_seq) { if (vmd->vmd_oom) { vmd->vmd_oom = FALSE; atomic_subtract_int(&vm_pageout_oom_vote, 1); } return; } /* * Do not follow the call sequence until OOM condition is * cleared. */ vmd->vmd_oom_seq = 0; if (vmd->vmd_oom) return; vmd->vmd_oom = TRUE; old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1); if (old_vote != vm_ndomains - 1) return; /* * The current pagedaemon thread is the last in the quorum to * start OOM. Initiate the selection and signaling of the * victim. */ vm_pageout_oom(VM_OOM_MEM); /* * After one round of OOM terror, recall our vote. On the * next pass, current pagedaemon would vote again if the low * memory condition is still there, due to vmd_oom being * false. */ vmd->vmd_oom = FALSE; atomic_subtract_int(&vm_pageout_oom_vote, 1); } /* * The OOM killer is the page daemon's action of last resort when * memory allocation requests have been stalled for a prolonged period * of time because it cannot reclaim memory. This function computes * the approximate number of physical pages that could be reclaimed if * the specified address space is destroyed. * * Private, anonymous memory owned by the address space is the * principal resource that we expect to recover after an OOM kill. * Since the physical pages mapped by the address space's COW entries * are typically shared pages, they are unlikely to be released and so * they are not counted. * * To get to the point where the page daemon runs the OOM killer, its * efforts to write-back vnode-backed pages may have stalled. This * could be caused by a memory allocation deadlock in the write path * that might be resolved by an OOM kill. Therefore, physical pages * belonging to vnode-backed objects are counted, because they might * be freed without being written out first if the address space holds * the last reference to an unlinked vnode. * * Similarly, physical pages belonging to OBJT_PHYS objects are * counted because the address space might hold the last reference to * the object. */ static long vm_pageout_oom_pagecount(struct vmspace *vmspace) { vm_map_t map; vm_map_entry_t entry; vm_object_t obj; long res; map = &vmspace->vm_map; KASSERT(!map->system_map, ("system map")); sx_assert(&map->lock, SA_LOCKED); res = 0; VM_MAP_ENTRY_FOREACH(entry, map) { if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) continue; obj = entry->object.vm_object; if (obj == NULL) continue; if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 && obj->ref_count != 1) continue; switch (obj->type) { case OBJT_DEFAULT: case OBJT_SWAP: case OBJT_PHYS: case OBJT_VNODE: res += obj->resident_page_count; break; } } return (res); } static int vm_oom_ratelim_last; static int vm_oom_pf_secs = 10; SYSCTL_INT(_vm, OID_AUTO, oom_pf_secs, CTLFLAG_RWTUN, &vm_oom_pf_secs, 0, ""); static struct mtx vm_oom_ratelim_mtx; void vm_pageout_oom(int shortage) { struct proc *p, *bigproc; vm_offset_t size, bigsize; struct thread *td; struct vmspace *vm; int now; bool breakout; /* * For OOM requests originating from vm_fault(), there is a high * chance that a single large process faults simultaneously in * several threads. Also, on an active system running many * processes of middle-size, like buildworld, all of them * could fault almost simultaneously as well. * * To avoid killing too many processes, rate-limit OOMs * initiated by vm_fault() time-outs on the waits for free * pages. */ mtx_lock(&vm_oom_ratelim_mtx); now = ticks; if (shortage == VM_OOM_MEM_PF && (u_int)(now - vm_oom_ratelim_last) < hz * vm_oom_pf_secs) { mtx_unlock(&vm_oom_ratelim_mtx); return; } vm_oom_ratelim_last = now; mtx_unlock(&vm_oom_ratelim_mtx); /* * We keep the process bigproc locked once we find it to keep anyone * from messing with it; however, there is a possibility of * deadlock if process B is bigproc and one of its child processes * attempts to propagate a signal to B while we are waiting for A's * lock while walking this list. To avoid this, we don't block on * the process lock but just skip a process if it is already locked. */ bigproc = NULL; bigsize = 0; sx_slock(&allproc_lock); FOREACH_PROC_IN_SYSTEM(p) { PROC_LOCK(p); /* * If this is a system, protected or killed process, skip it. */ if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 || p->p_pid == 1 || P_KILLED(p) || (p->p_pid < 48 && swap_pager_avail != 0)) { PROC_UNLOCK(p); continue; } /* * If the process is in a non-running type state, * don't touch it. Check all the threads individually. */ breakout = false; FOREACH_THREAD_IN_PROC(p, td) { thread_lock(td); if (!TD_ON_RUNQ(td) && !TD_IS_RUNNING(td) && !TD_IS_SLEEPING(td) && !TD_IS_SUSPENDED(td) && !TD_IS_SWAPPED(td)) { thread_unlock(td); breakout = true; break; } thread_unlock(td); } if (breakout) { PROC_UNLOCK(p); continue; } /* * get the process size */ vm = vmspace_acquire_ref(p); if (vm == NULL) { PROC_UNLOCK(p); continue; } _PHOLD_LITE(p); PROC_UNLOCK(p); sx_sunlock(&allproc_lock); if (!vm_map_trylock_read(&vm->vm_map)) { vmspace_free(vm); sx_slock(&allproc_lock); PRELE(p); continue; } size = vmspace_swap_count(vm); if (shortage == VM_OOM_MEM || shortage == VM_OOM_MEM_PF) size += vm_pageout_oom_pagecount(vm); vm_map_unlock_read(&vm->vm_map); vmspace_free(vm); sx_slock(&allproc_lock); /* * If this process is bigger than the biggest one, * remember it. */ if (size > bigsize) { if (bigproc != NULL) PRELE(bigproc); bigproc = p; bigsize = size; } else { PRELE(p); } } sx_sunlock(&allproc_lock); if (bigproc != NULL) { if (vm_panic_on_oom != 0) panic("out of swap space"); PROC_LOCK(bigproc); killproc(bigproc, "out of swap space"); sched_nice(bigproc, PRIO_MIN); _PRELE(bigproc); PROC_UNLOCK(bigproc); } } static bool vm_pageout_lowmem(void) { static int lowmem_ticks = 0; int last; last = atomic_load_int(&lowmem_ticks); while ((u_int)(ticks - last) / hz >= lowmem_period) { if (atomic_fcmpset_int(&lowmem_ticks, &last, ticks) == 0) continue; /* * Decrease registered cache sizes. */ SDT_PROBE0(vm, , , vm__lowmem_scan); EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_PAGES); /* * We do this explicitly after the caches have been * drained above. If we have a severe page shortage on * our hands, completely drain all UMA zones. Otherwise, * just prune the caches. */ uma_reclaim(vm_page_count_min() ? UMA_RECLAIM_DRAIN_CPU : UMA_RECLAIM_TRIM); return (true); } return (false); } static void vm_pageout_worker(void *arg) { struct vm_domain *vmd; u_int ofree; int addl_shortage, domain, shortage; bool target_met; domain = (uintptr_t)arg; vmd = VM_DOMAIN(domain); shortage = 0; target_met = true; /* * XXXKIB It could be useful to bind pageout daemon threads to * the cores belonging to the domain, from which vm_page_array * is allocated. */ KASSERT(vmd->vmd_segs != 0, ("domain without segments")); vmd->vmd_last_active_scan = ticks; /* * The pageout daemon worker is never done, so loop forever. */ while (TRUE) { vm_domain_pageout_lock(vmd); /* * We need to clear wanted before we check the limits. This * prevents races with wakers who will check wanted after they * reach the limit. */ atomic_store_int(&vmd->vmd_pageout_wanted, 0); /* * Might the page daemon need to run again? */ if (vm_paging_needed(vmd, vmd->vmd_free_count)) { /* * Yes. If the scan failed to produce enough free * pages, sleep uninterruptibly for some time in the * hope that the laundry thread will clean some pages. */ vm_domain_pageout_unlock(vmd); if (!target_met) pause("pwait", hz / VM_INACT_SCAN_RATE); } else { /* * No, sleep until the next wakeup or until pages * need to have their reference stats updated. */ if (mtx_sleep(&vmd->vmd_pageout_wanted, vm_domain_pageout_lockptr(vmd), PDROP | PVM, "psleep", hz / VM_INACT_SCAN_RATE) == 0) VM_CNT_INC(v_pdwakeups); } /* Prevent spurious wakeups by ensuring that wanted is set. */ atomic_store_int(&vmd->vmd_pageout_wanted, 1); /* * Use the controller to calculate how many pages to free in * this interval, and scan the inactive queue. If the lowmem * handlers appear to have freed up some pages, subtract the * difference from the inactive queue scan target. */ shortage = pidctrl_daemon(&vmd->vmd_pid, vmd->vmd_free_count); if (shortage > 0) { ofree = vmd->vmd_free_count; if (vm_pageout_lowmem() && vmd->vmd_free_count > ofree) shortage -= min(vmd->vmd_free_count - ofree, (u_int)shortage); target_met = vm_pageout_scan_inactive(vmd, shortage, &addl_shortage); } else addl_shortage = 0; /* * Scan the active queue. A positive value for shortage * indicates that we must aggressively deactivate pages to avoid * a shortfall. */ shortage = vm_pageout_active_target(vmd) + addl_shortage; vm_pageout_scan_active(vmd, shortage); } } /* * vm_pageout_init initialises basic pageout daemon settings. */ static void vm_pageout_init_domain(int domain) { struct vm_domain *vmd; struct sysctl_oid *oid; vmd = VM_DOMAIN(domain); vmd->vmd_interrupt_free_min = 2; /* * v_free_reserved needs to include enough for the largest * swap pager structures plus enough for any pv_entry structs * when paging. */ if (vmd->vmd_page_count > 1024) vmd->vmd_free_min = 4 + (vmd->vmd_page_count - 1024) / 200; else vmd->vmd_free_min = 4; vmd->vmd_pageout_free_min = 2 * MAXBSIZE / PAGE_SIZE + vmd->vmd_interrupt_free_min; vmd->vmd_free_reserved = vm_pageout_page_count + vmd->vmd_pageout_free_min + (vmd->vmd_page_count / 768); vmd->vmd_free_severe = vmd->vmd_free_min / 2; vmd->vmd_free_target = 4 * vmd->vmd_free_min + vmd->vmd_free_reserved; vmd->vmd_free_min += vmd->vmd_free_reserved; vmd->vmd_free_severe += vmd->vmd_free_reserved; vmd->vmd_inactive_target = (3 * vmd->vmd_free_target) / 2; if (vmd->vmd_inactive_target > vmd->vmd_free_count / 3) vmd->vmd_inactive_target = vmd->vmd_free_count / 3; /* * Set the default wakeup threshold to be 10% below the paging * target. This keeps the steady state out of shortfall. */ vmd->vmd_pageout_wakeup_thresh = (vmd->vmd_free_target / 10) * 9; /* * Target amount of memory to move out of the laundry queue during a * background laundering. This is proportional to the amount of system * memory. */ vmd->vmd_background_launder_target = (vmd->vmd_free_target - vmd->vmd_free_min) / 10; /* Initialize the pageout daemon pid controller. */ pidctrl_init(&vmd->vmd_pid, hz / VM_INACT_SCAN_RATE, vmd->vmd_free_target, PIDCTRL_BOUND, PIDCTRL_KPD, PIDCTRL_KID, PIDCTRL_KDD); oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(vmd->vmd_oid), OID_AUTO, "pidctrl", CTLFLAG_RD, NULL, ""); pidctrl_init_sysctl(&vmd->vmd_pid, SYSCTL_CHILDREN(oid)); } static void vm_pageout_init(void) { u_int freecount; int i; /* * Initialize some paging parameters. */ if (vm_cnt.v_page_count < 2000) vm_pageout_page_count = 8; freecount = 0; for (i = 0; i < vm_ndomains; i++) { struct vm_domain *vmd; vm_pageout_init_domain(i); vmd = VM_DOMAIN(i); vm_cnt.v_free_reserved += vmd->vmd_free_reserved; vm_cnt.v_free_target += vmd->vmd_free_target; vm_cnt.v_free_min += vmd->vmd_free_min; vm_cnt.v_inactive_target += vmd->vmd_inactive_target; vm_cnt.v_pageout_free_min += vmd->vmd_pageout_free_min; vm_cnt.v_interrupt_free_min += vmd->vmd_interrupt_free_min; vm_cnt.v_free_severe += vmd->vmd_free_severe; freecount += vmd->vmd_free_count; } /* * Set interval in seconds for active scan. We want to visit each * page at least once every ten minutes. This is to prevent worst * case paging behaviors with stale active LRU. */ if (vm_pageout_update_period == 0) vm_pageout_update_period = 600; if (vm_page_max_user_wired == 0) vm_page_max_user_wired = freecount / 3; } /* * vm_pageout is the high level pageout daemon. */ static void vm_pageout(void) { struct proc *p; struct thread *td; int error, first, i; p = curproc; td = curthread; mtx_init(&vm_oom_ratelim_mtx, "vmoomr", NULL, MTX_DEF); swap_pager_swap_init(); for (first = -1, i = 0; i < vm_ndomains; i++) { if (VM_DOMAIN_EMPTY(i)) { if (bootverbose) printf("domain %d empty; skipping pageout\n", i); continue; } if (first == -1) first = i; else { error = kthread_add(vm_pageout_worker, (void *)(uintptr_t)i, p, NULL, 0, 0, "dom%d", i); if (error != 0) panic("starting pageout for domain %d: %d\n", i, error); } error = kthread_add(vm_pageout_laundry_worker, (void *)(uintptr_t)i, p, NULL, 0, 0, "laundry: dom%d", i); if (error != 0) panic("starting laundry for domain %d: %d", i, error); } error = kthread_add(uma_reclaim_worker, NULL, p, NULL, 0, 0, "uma"); if (error != 0) panic("starting uma_reclaim helper, error %d\n", error); snprintf(td->td_name, sizeof(td->td_name), "dom%d", first); vm_pageout_worker((void *)(uintptr_t)first); } /* * Perform an advisory wakeup of the page daemon. */ void pagedaemon_wakeup(int domain) { struct vm_domain *vmd; vmd = VM_DOMAIN(domain); vm_domain_pageout_assert_unlocked(vmd); if (curproc == pageproc) return; if (atomic_fetchadd_int(&vmd->vmd_pageout_wanted, 1) == 0) { vm_domain_pageout_lock(vmd); atomic_store_int(&vmd->vmd_pageout_wanted, 1); wakeup(&vmd->vmd_pageout_wanted); vm_domain_pageout_unlock(vmd); } } Index: head/sys/vm/vm_swapout.c =================================================================== --- head/sys/vm/vm_swapout.c (revision 353534) +++ head/sys/vm/vm_swapout.c (revision 353535) @@ -1,963 +1,966 @@ /*- * SPDX-License-Identifier: (BSD-4-Clause AND MIT-CMU) * * Copyright (c) 1991 Regents of the University of California. * All rights reserved. * Copyright (c) 1994 John S. Dyson * All rights reserved. * Copyright (c) 1994 David Greenman * All rights reserved. * Copyright (c) 2005 Yahoo! Technologies Norway AS * All rights reserved. * * This code is derived from software contributed to Berkeley by * The Mach Operating System project at Carnegie-Mellon University. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. All advertising materials mentioning features or use of this software * must display the following acknowledgement: * This product includes software developed by the University of * California, Berkeley and its contributors. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91 * * * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Authors: Avadis Tevanian, Jr., Michael Wayne Young * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. */ #include __FBSDID("$FreeBSD$"); #include "opt_kstack_pages.h" #include "opt_kstack_max_pages.h" #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* the kernel process "vm_daemon" */ static void vm_daemon(void); static struct proc *vmproc; static struct kproc_desc vm_kp = { "vmdaemon", vm_daemon, &vmproc }; SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp); static int vm_swap_enabled = 1; static int vm_swap_idle_enabled = 0; SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); /* * Swap_idle_threshold1 is the guaranteed swapped in time for a process */ static int swap_idle_threshold1 = 2; SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW, &swap_idle_threshold1, 0, "Guaranteed swapped in time for a process"); /* * Swap_idle_threshold2 is the time that a process can be idle before * it will be swapped out, if idle swapping is enabled. */ static int swap_idle_threshold2 = 10; SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW, &swap_idle_threshold2, 0, "Time before a process will be swapped out"); static int vm_pageout_req_swapout; /* XXX */ static int vm_daemon_needed; static struct mtx vm_daemon_mtx; /* Allow for use by vm_pageout before vm_daemon is initialized. */ MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF); static int swapped_cnt; static int swap_inprogress; /* Pending swap-ins done outside swapper. */ static int last_swapin; static void swapclear(struct proc *); static int swapout(struct proc *); static void vm_swapout_map_deactivate_pages(vm_map_t, long); static void vm_swapout_object_deactivate_pages(pmap_t, vm_object_t, long); static void swapout_procs(int action); static void vm_req_vmdaemon(int req); static void vm_thread_swapout(struct thread *td); /* * vm_swapout_object_deactivate_pages * * Deactivate enough pages to satisfy the inactive target * requirements. * * The object and map must be locked. */ static void vm_swapout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object, long desired) { vm_object_t backing_object, object; vm_page_t p; int act_delta, remove_mode; VM_OBJECT_ASSERT_LOCKED(first_object); if ((first_object->flags & OBJ_FICTITIOUS) != 0) return; for (object = first_object;; object = backing_object) { if (pmap_resident_count(pmap) <= desired) goto unlock_return; VM_OBJECT_ASSERT_LOCKED(object); if ((object->flags & OBJ_UNMANAGED) != 0 || REFCOUNT_COUNT(object->paging_in_progress) > 0) goto unlock_return; remove_mode = 0; if (object->shadow_count > 1) remove_mode = 1; /* * Scan the object's entire memory queue. */ TAILQ_FOREACH(p, &object->memq, listq) { if (pmap_resident_count(pmap) <= desired) goto unlock_return; if (should_yield()) goto unlock_return; + if (vm_page_tryxbusy(p) == 0) + continue; + VM_CNT_INC(v_pdpages); /* * The page may acquire a wiring after this check. * The page daemon handles wired pages, so there is * no harm done if a wiring appears while we are * attempting to deactivate the page. */ - if (vm_page_busied(p) || vm_page_wired(p)) + if (vm_page_wired(p) || !pmap_page_exists_quick(pmap, p)) { + vm_page_xunbusy(p); continue; - VM_CNT_INC(v_pdpages); - if (!pmap_page_exists_quick(pmap, p)) - continue; + } act_delta = pmap_ts_referenced(p); vm_page_lock(p); if ((p->aflags & PGA_REFERENCED) != 0) { if (act_delta == 0) act_delta = 1; vm_page_aflag_clear(p, PGA_REFERENCED); } if (!vm_page_active(p) && act_delta != 0) { vm_page_activate(p); p->act_count += act_delta; } else if (vm_page_active(p)) { /* * The page daemon does not requeue pages * after modifying their activation count. */ if (act_delta == 0) { p->act_count -= min(p->act_count, ACT_DECLINE); if (!remove_mode && p->act_count == 0) { (void)vm_page_try_remove_all(p); vm_page_deactivate(p); } } else { vm_page_activate(p); if (p->act_count < ACT_MAX - ACT_ADVANCE) p->act_count += ACT_ADVANCE; } } else if (vm_page_inactive(p)) (void)vm_page_try_remove_all(p); vm_page_unlock(p); + vm_page_xunbusy(p); } if ((backing_object = object->backing_object) == NULL) goto unlock_return; VM_OBJECT_RLOCK(backing_object); if (object != first_object) VM_OBJECT_RUNLOCK(object); } unlock_return: if (object != first_object) VM_OBJECT_RUNLOCK(object); } /* * deactivate some number of pages in a map, try to do it fairly, but * that is really hard to do. */ static void vm_swapout_map_deactivate_pages(vm_map_t map, long desired) { vm_map_entry_t tmpe; vm_object_t obj, bigobj; int nothingwired; if (!vm_map_trylock_read(map)) return; bigobj = NULL; nothingwired = TRUE; /* * first, search out the biggest object, and try to free pages from * that. */ VM_MAP_ENTRY_FOREACH(tmpe, map) { if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { obj = tmpe->object.vm_object; if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) { if (obj->shadow_count <= 1 && (bigobj == NULL || bigobj->resident_page_count < obj->resident_page_count)) { if (bigobj != NULL) VM_OBJECT_RUNLOCK(bigobj); bigobj = obj; } else VM_OBJECT_RUNLOCK(obj); } } if (tmpe->wired_count > 0) nothingwired = FALSE; } if (bigobj != NULL) { vm_swapout_object_deactivate_pages(map->pmap, bigobj, desired); VM_OBJECT_RUNLOCK(bigobj); } /* * Next, hunt around for other pages to deactivate. We actually * do this search sort of wrong -- .text first is not the best idea. */ VM_MAP_ENTRY_FOREACH(tmpe, map) { if (pmap_resident_count(vm_map_pmap(map)) <= desired) break; if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { obj = tmpe->object.vm_object; if (obj != NULL) { VM_OBJECT_RLOCK(obj); vm_swapout_object_deactivate_pages(map->pmap, obj, desired); VM_OBJECT_RUNLOCK(obj); } } } /* * Remove all mappings if a process is swapped out, this will free page * table pages. */ if (desired == 0 && nothingwired) { pmap_remove(vm_map_pmap(map), vm_map_min(map), vm_map_max(map)); } vm_map_unlock_read(map); } /* * Swap out requests */ #define VM_SWAP_NORMAL 1 #define VM_SWAP_IDLE 2 void vm_swapout_run(void) { if (vm_swap_enabled) vm_req_vmdaemon(VM_SWAP_NORMAL); } /* * Idle process swapout -- run once per second when pagedaemons are * reclaiming pages. */ void vm_swapout_run_idle(void) { static long lsec; if (!vm_swap_idle_enabled || time_second == lsec) return; vm_req_vmdaemon(VM_SWAP_IDLE); lsec = time_second; } static void vm_req_vmdaemon(int req) { static int lastrun = 0; mtx_lock(&vm_daemon_mtx); vm_pageout_req_swapout |= req; if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { wakeup(&vm_daemon_needed); lastrun = ticks; } mtx_unlock(&vm_daemon_mtx); } static void vm_daemon(void) { struct rlimit rsslim; struct proc *p; struct thread *td; struct vmspace *vm; int breakout, swapout_flags, tryagain, attempts; #ifdef RACCT uint64_t rsize, ravailable; #endif while (TRUE) { mtx_lock(&vm_daemon_mtx); msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", #ifdef RACCT racct_enable ? hz : 0 #else 0 #endif ); swapout_flags = vm_pageout_req_swapout; vm_pageout_req_swapout = 0; mtx_unlock(&vm_daemon_mtx); if (swapout_flags != 0) { /* * Drain the per-CPU page queue batches as a deadlock * avoidance measure. */ if ((swapout_flags & VM_SWAP_NORMAL) != 0) vm_page_pqbatch_drain(); swapout_procs(swapout_flags); } /* * scan the processes for exceeding their rlimits or if * process is swapped out -- deactivate pages */ tryagain = 0; attempts = 0; again: attempts++; sx_slock(&allproc_lock); FOREACH_PROC_IN_SYSTEM(p) { vm_pindex_t limit, size; /* * if this is a system process or if we have already * looked at this process, skip it. */ PROC_LOCK(p); if (p->p_state != PRS_NORMAL || p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) { PROC_UNLOCK(p); continue; } /* * if the process is in a non-running type state, * don't touch it. */ breakout = 0; FOREACH_THREAD_IN_PROC(p, td) { thread_lock(td); if (!TD_ON_RUNQ(td) && !TD_IS_RUNNING(td) && !TD_IS_SLEEPING(td) && !TD_IS_SUSPENDED(td)) { thread_unlock(td); breakout = 1; break; } thread_unlock(td); } if (breakout) { PROC_UNLOCK(p); continue; } /* * get a limit */ lim_rlimit_proc(p, RLIMIT_RSS, &rsslim); limit = OFF_TO_IDX( qmin(rsslim.rlim_cur, rsslim.rlim_max)); /* * let processes that are swapped out really be * swapped out set the limit to nothing (will force a * swap-out.) */ if ((p->p_flag & P_INMEM) == 0) limit = 0; /* XXX */ vm = vmspace_acquire_ref(p); _PHOLD_LITE(p); PROC_UNLOCK(p); if (vm == NULL) { PRELE(p); continue; } sx_sunlock(&allproc_lock); size = vmspace_resident_count(vm); if (size >= limit) { vm_swapout_map_deactivate_pages( &vm->vm_map, limit); size = vmspace_resident_count(vm); } #ifdef RACCT if (racct_enable) { rsize = IDX_TO_OFF(size); PROC_LOCK(p); if (p->p_state == PRS_NORMAL) racct_set(p, RACCT_RSS, rsize); ravailable = racct_get_available(p, RACCT_RSS); PROC_UNLOCK(p); if (rsize > ravailable) { /* * Don't be overly aggressive; this * might be an innocent process, * and the limit could've been exceeded * by some memory hog. Don't try * to deactivate more than 1/4th * of process' resident set size. */ if (attempts <= 8) { if (ravailable < rsize - (rsize / 4)) { ravailable = rsize - (rsize / 4); } } vm_swapout_map_deactivate_pages( &vm->vm_map, OFF_TO_IDX(ravailable)); /* Update RSS usage after paging out. */ size = vmspace_resident_count(vm); rsize = IDX_TO_OFF(size); PROC_LOCK(p); if (p->p_state == PRS_NORMAL) racct_set(p, RACCT_RSS, rsize); PROC_UNLOCK(p); if (rsize > ravailable) tryagain = 1; } } #endif vmspace_free(vm); sx_slock(&allproc_lock); PRELE(p); } sx_sunlock(&allproc_lock); if (tryagain != 0 && attempts <= 10) { maybe_yield(); goto again; } } } /* * Allow a thread's kernel stack to be paged out. */ static void vm_thread_swapout(struct thread *td) { vm_object_t ksobj; vm_page_t m; int i, pages; cpu_thread_swapout(td); pages = td->td_kstack_pages; ksobj = td->td_kstack_obj; pmap_qremove(td->td_kstack, pages); VM_OBJECT_WLOCK(ksobj); for (i = 0; i < pages; i++) { m = vm_page_lookup(ksobj, i); if (m == NULL) panic("vm_thread_swapout: kstack already missing?"); vm_page_dirty(m); vm_page_unwire(m, PQ_LAUNDRY); } VM_OBJECT_WUNLOCK(ksobj); } /* * Bring the kernel stack for a specified thread back in. */ static void vm_thread_swapin(struct thread *td, int oom_alloc) { vm_object_t ksobj; vm_page_t ma[KSTACK_MAX_PAGES]; int a, count, i, j, pages, rv; pages = td->td_kstack_pages; ksobj = td->td_kstack_obj; VM_OBJECT_WLOCK(ksobj); (void)vm_page_grab_pages(ksobj, 0, oom_alloc | VM_ALLOC_WIRED, ma, pages); for (i = 0; i < pages;) { vm_page_assert_xbusied(ma[i]); if (ma[i]->valid == VM_PAGE_BITS_ALL) { vm_page_xunbusy(ma[i]); i++; continue; } vm_object_pip_add(ksobj, 1); for (j = i + 1; j < pages; j++) if (ma[j]->valid == VM_PAGE_BITS_ALL) break; rv = vm_pager_has_page(ksobj, ma[i]->pindex, NULL, &a); KASSERT(rv == 1, ("%s: missing page %p", __func__, ma[i])); count = min(a + 1, j - i); rv = vm_pager_get_pages(ksobj, ma + i, count, NULL, NULL); KASSERT(rv == VM_PAGER_OK, ("%s: cannot get kstack for proc %d", __func__, td->td_proc->p_pid)); vm_object_pip_wakeup(ksobj); for (j = i; j < i + count; j++) vm_page_xunbusy(ma[j]); i += count; } VM_OBJECT_WUNLOCK(ksobj); pmap_qenter(td->td_kstack, ma, pages); cpu_thread_swapin(td); } void faultin(struct proc *p) { struct thread *td; int oom_alloc; PROC_LOCK_ASSERT(p, MA_OWNED); /* * If another process is swapping in this process, * just wait until it finishes. */ if (p->p_flag & P_SWAPPINGIN) { while (p->p_flag & P_SWAPPINGIN) msleep(&p->p_flag, &p->p_mtx, PVM, "faultin", 0); return; } if ((p->p_flag & P_INMEM) == 0) { oom_alloc = (p->p_flag & P_WKILLED) != 0 ? VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL; /* * Don't let another thread swap process p out while we are * busy swapping it in. */ ++p->p_lock; p->p_flag |= P_SWAPPINGIN; PROC_UNLOCK(p); sx_xlock(&allproc_lock); MPASS(swapped_cnt > 0); swapped_cnt--; if (curthread != &thread0) swap_inprogress++; sx_xunlock(&allproc_lock); /* * We hold no lock here because the list of threads * can not change while all threads in the process are * swapped out. */ FOREACH_THREAD_IN_PROC(p, td) vm_thread_swapin(td, oom_alloc); if (curthread != &thread0) { sx_xlock(&allproc_lock); MPASS(swap_inprogress > 0); swap_inprogress--; last_swapin = ticks; sx_xunlock(&allproc_lock); } PROC_LOCK(p); swapclear(p); p->p_swtick = ticks; /* Allow other threads to swap p out now. */ wakeup(&p->p_flag); --p->p_lock; } } /* * This swapin algorithm attempts to swap-in processes only if there * is enough space for them. Of course, if a process waits for a long * time, it will be swapped in anyway. */ static struct proc * swapper_selector(bool wkilled_only) { struct proc *p, *res; struct thread *td; int ppri, pri, slptime, swtime; sx_assert(&allproc_lock, SA_SLOCKED); if (swapped_cnt == 0) return (NULL); res = NULL; ppri = INT_MIN; FOREACH_PROC_IN_SYSTEM(p) { PROC_LOCK(p); if (p->p_state == PRS_NEW || (p->p_flag & (P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) != 0) { PROC_UNLOCK(p); continue; } if (p->p_state == PRS_NORMAL && (p->p_flag & P_WKILLED) != 0) { /* * A swapped-out process might have mapped a * large portion of the system's pages as * anonymous memory. There is no other way to * release the memory other than to kill the * process, for which we need to swap it in. */ return (p); } if (wkilled_only) { PROC_UNLOCK(p); continue; } swtime = (ticks - p->p_swtick) / hz; FOREACH_THREAD_IN_PROC(p, td) { /* * An otherwise runnable thread of a process * swapped out has only the TDI_SWAPPED bit set. */ thread_lock(td); if (td->td_inhibitors == TDI_SWAPPED) { slptime = (ticks - td->td_slptick) / hz; pri = swtime + slptime; if ((td->td_flags & TDF_SWAPINREQ) == 0) pri -= p->p_nice * 8; /* * if this thread is higher priority * and there is enough space, then select * this process instead of the previous * selection. */ if (pri > ppri) { res = p; ppri = pri; } } thread_unlock(td); } PROC_UNLOCK(p); } if (res != NULL) PROC_LOCK(res); return (res); } #define SWAPIN_INTERVAL (MAXSLP * hz / 2) /* * Limit swapper to swap in one non-WKILLED process in MAXSLP/2 * interval, assuming that there is: * - at least one domain that is not suffering from a shortage of free memory; * - no parallel swap-ins; * - no other swap-ins in the current SWAPIN_INTERVAL. */ static bool swapper_wkilled_only(void) { return (vm_page_count_min_set(&all_domains) || swap_inprogress > 0 || (u_int)(ticks - last_swapin) < SWAPIN_INTERVAL); } void swapper(void) { struct proc *p; for (;;) { sx_slock(&allproc_lock); p = swapper_selector(swapper_wkilled_only()); sx_sunlock(&allproc_lock); if (p == NULL) { tsleep(&proc0, PVM, "swapin", SWAPIN_INTERVAL); } else { PROC_LOCK_ASSERT(p, MA_OWNED); /* * Another process may be bringing or may have * already brought this process in while we * traverse all threads. Or, this process may * have exited or even being swapped out * again. */ if (p->p_state == PRS_NORMAL && (p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) == 0) { faultin(p); } PROC_UNLOCK(p); } } } /* * First, if any processes have been sleeping or stopped for at least * "swap_idle_threshold1" seconds, they are swapped out. If, however, * no such processes exist, then the longest-sleeping or stopped * process is swapped out. Finally, and only as a last resort, if * there are no sleeping or stopped processes, the longest-resident * process is swapped out. */ static void swapout_procs(int action) { struct proc *p; struct thread *td; int slptime; bool didswap, doswap; MPASS((action & (VM_SWAP_NORMAL | VM_SWAP_IDLE)) != 0); didswap = false; sx_slock(&allproc_lock); FOREACH_PROC_IN_SYSTEM(p) { /* * Filter out not yet fully constructed processes. Do * not swap out held processes. Avoid processes which * are system, exiting, execing, traced, already swapped * out or are in the process of being swapped in or out. */ PROC_LOCK(p); if (p->p_state != PRS_NORMAL || p->p_lock != 0 || (p->p_flag & (P_SYSTEM | P_WEXIT | P_INEXEC | P_STOPPED_SINGLE | P_TRACED | P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) != P_INMEM) { PROC_UNLOCK(p); continue; } /* * Further consideration of this process for swap out * requires iterating over its threads. We release * allproc_lock here so that process creation and * destruction are not blocked while we iterate. * * To later reacquire allproc_lock and resume * iteration over the allproc list, we will first have * to release the lock on the process. We place a * hold on the process so that it remains in the * allproc list while it is unlocked. */ _PHOLD_LITE(p); sx_sunlock(&allproc_lock); /* * Do not swapout a realtime process. * Guarantee swap_idle_threshold1 time in memory. * If the system is under memory stress, or if we are * swapping idle processes >= swap_idle_threshold2, * then swap the process out. */ doswap = true; FOREACH_THREAD_IN_PROC(p, td) { thread_lock(td); slptime = (ticks - td->td_slptick) / hz; if (PRI_IS_REALTIME(td->td_pri_class) || slptime < swap_idle_threshold1 || !thread_safetoswapout(td) || ((action & VM_SWAP_NORMAL) == 0 && slptime < swap_idle_threshold2)) doswap = false; thread_unlock(td); if (!doswap) break; } if (doswap && swapout(p) == 0) didswap = true; PROC_UNLOCK(p); if (didswap) { sx_xlock(&allproc_lock); swapped_cnt++; sx_downgrade(&allproc_lock); } else sx_slock(&allproc_lock); PRELE(p); } sx_sunlock(&allproc_lock); /* * If we swapped something out, and another process needed memory, * then wakeup the sched process. */ if (didswap) wakeup(&proc0); } static void swapclear(struct proc *p) { struct thread *td; PROC_LOCK_ASSERT(p, MA_OWNED); FOREACH_THREAD_IN_PROC(p, td) { thread_lock(td); td->td_flags |= TDF_INMEM; td->td_flags &= ~TDF_SWAPINREQ; TD_CLR_SWAPPED(td); if (TD_CAN_RUN(td)) if (setrunnable(td)) { #ifdef INVARIANTS /* * XXX: We just cleared TDI_SWAPPED * above and set TDF_INMEM, so this * should never happen. */ panic("not waking up swapper"); #endif } thread_unlock(td); } p->p_flag &= ~(P_SWAPPINGIN | P_SWAPPINGOUT); p->p_flag |= P_INMEM; } static int swapout(struct proc *p) { struct thread *td; PROC_LOCK_ASSERT(p, MA_OWNED); /* * The states of this process and its threads may have changed * by now. Assuming that there is only one pageout daemon thread, * this process should still be in memory. */ KASSERT((p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) == P_INMEM, ("swapout: lost a swapout race?")); /* * Remember the resident count. */ p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace); /* * Check and mark all threads before we proceed. */ p->p_flag &= ~P_INMEM; p->p_flag |= P_SWAPPINGOUT; FOREACH_THREAD_IN_PROC(p, td) { thread_lock(td); if (!thread_safetoswapout(td)) { thread_unlock(td); swapclear(p); return (EBUSY); } td->td_flags &= ~TDF_INMEM; TD_SET_SWAPPED(td); thread_unlock(td); } td = FIRST_THREAD_IN_PROC(p); ++td->td_ru.ru_nswap; PROC_UNLOCK(p); /* * This list is stable because all threads are now prevented from * running. The list is only modified in the context of a running * thread in this process. */ FOREACH_THREAD_IN_PROC(p, td) vm_thread_swapout(td); PROC_LOCK(p); p->p_flag &= ~P_SWAPPINGOUT; p->p_swtick = ticks; return (0); }