2 * Copyright (c) 1991 Regents of the University of California.
4 * Copyright (c) 1998 Matthew Dillon. All Rights Reserved.
6 * This code is derived from software contributed to Berkeley by
7 * The Mach Operating System project at Carnegie-Mellon University.
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
17 * 4. Neither the name of the University nor the names of its contributors
18 * may be used to endorse or promote products derived from this software
19 * without specific prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91
37 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38 * All rights reserved.
40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
42 * Permission to use, copy, modify and distribute this software and
43 * its documentation is hereby granted, provided that both the copyright
44 * notice and this permission notice appear in all copies of the
45 * software, derivative works or modified versions, and any portions
46 * thereof, and that both notices appear in supporting documentation.
48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
52 * Carnegie Mellon requests users of this software to return to
54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
55 * School of Computer Science
56 * Carnegie Mellon University
57 * Pittsburgh PA 15213-3890
59 * any improvements or extensions that they make and grant Carnegie the
60 * rights to redistribute these changes.
64 * GENERAL RULES ON VM_PAGE MANIPULATION
66 * - a pageq mutex is required when adding or removing a page from a
67 * page queue (vm_page_queue[]), regardless of other mutexes or the
68 * busy state of a page.
70 * - a hash chain mutex is required when associating or disassociating
71 * a page from the VM PAGE CACHE hash table (vm_page_buckets),
72 * regardless of other mutexes or the busy state of a page.
74 * - either a hash chain mutex OR a busied page is required in order
75 * to modify the page flags. A hash chain mutex must be obtained in
76 * order to busy a page. A page's flags cannot be modified by a
77 * hash chain mutex if the page is marked busy.
79 * - The object memq mutex is held when inserting or removing
80 * pages from an object (vm_page_insert() or vm_page_remove()). This
81 * is different from the object's main mutex.
83 * Generally speaking, you have to be aware of side effects when running
84 * vm_page ops. A vm_page_lookup() will return with the hash chain
85 * locked, whether it was able to lookup the page or not. vm_page_free(),
86 * vm_page_cache(), vm_page_activate(), and a number of other routines
87 * will release the hash chain mutex for you. Intermediate manipulation
88 * routines such as vm_page_flag_set() expect the hash chain to be held
89 * on entry and the hash chain will remain held on return.
91 * pageq scanning can only occur with the pageq in question locked.
92 * We have a known bottleneck with the active queue, but the cache
93 * and free queues are actually arrays already.
97 * Resident memory management module.
100 #include <sys/cdefs.h>
101 __FBSDID("$FreeBSD$");
105 #include <sys/param.h>
106 #include <sys/systm.h>
107 #include <sys/lock.h>
108 #include <sys/kernel.h>
109 #include <sys/limits.h>
110 #include <sys/malloc.h>
111 #include <sys/mutex.h>
112 #include <sys/proc.h>
113 #include <sys/sysctl.h>
114 #include <sys/vmmeter.h>
115 #include <sys/vnode.h>
118 #include <vm/vm_param.h>
119 #include <vm/vm_kern.h>
120 #include <vm/vm_object.h>
121 #include <vm/vm_page.h>
122 #include <vm/vm_pageout.h>
123 #include <vm/vm_pager.h>
124 #include <vm/vm_phys.h>
125 #include <vm/vm_reserv.h>
126 #include <vm/vm_extern.h>
128 #include <vm/uma_int.h>
130 #include <machine/md_var.h>
133 * Associated with page of user-allocatable memory is a
137 struct vpgqueues vm_page_queues[PQ_COUNT];
138 struct mtx vm_page_queue_mtx;
139 struct mtx vm_page_queue_free_mtx;
141 vm_page_t vm_page_array = 0;
142 int vm_page_array_size = 0;
144 int vm_page_zero_count = 0;
146 static int boot_pages = UMA_BOOT_PAGES;
147 TUNABLE_INT("vm.boot_pages", &boot_pages);
148 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RD, &boot_pages, 0,
149 "number of pages allocated for bootstrapping the VM system");
151 static void vm_page_enqueue(int queue, vm_page_t m);
153 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
154 #if PAGE_SIZE == 32768
156 CTASSERT(sizeof(u_long) >= 8);
163 * Sets the page size, perhaps based upon the memory
164 * size. Must be called before any use of page-size
165 * dependent functions.
168 vm_set_page_size(void)
170 if (cnt.v_page_size == 0)
171 cnt.v_page_size = PAGE_SIZE;
172 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0)
173 panic("vm_set_page_size: page size not a power of two");
177 * vm_page_blacklist_lookup:
179 * See if a physical address in this page has been listed
180 * in the blacklist tunable. Entries in the tunable are
181 * separated by spaces or commas. If an invalid integer is
182 * encountered then the rest of the string is skipped.
185 vm_page_blacklist_lookup(char *list, vm_paddr_t pa)
190 for (pos = list; *pos != '\0'; pos = cp) {
191 bad = strtoq(pos, &cp, 0);
193 if (*cp == ' ' || *cp == ',') {
200 if (pa == trunc_page(bad))
209 * Initializes the resident memory module.
211 * Allocates memory for the page cells, and
212 * for the object/offset-to-page hash table headers.
213 * Each page cell is initialized and placed on the free list.
216 vm_page_startup(vm_offset_t vaddr)
219 vm_paddr_t page_range;
227 /* the biggest memory array is the second group of pages */
229 vm_paddr_t biggestsize;
230 vm_paddr_t low_water, high_water;
236 vaddr = round_page(vaddr);
238 for (i = 0; phys_avail[i + 1]; i += 2) {
239 phys_avail[i] = round_page(phys_avail[i]);
240 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
243 low_water = phys_avail[0];
244 high_water = phys_avail[1];
246 for (i = 0; phys_avail[i + 1]; i += 2) {
247 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
249 if (size > biggestsize) {
253 if (phys_avail[i] < low_water)
254 low_water = phys_avail[i];
255 if (phys_avail[i + 1] > high_water)
256 high_water = phys_avail[i + 1];
264 end = phys_avail[biggestone+1];
267 * Initialize the locks.
269 mtx_init(&vm_page_queue_mtx, "vm page queue mutex", NULL, MTX_DEF |
271 mtx_init(&vm_page_queue_free_mtx, "vm page queue free mutex", NULL,
275 * Initialize the queue headers for the hold queue, the active queue,
276 * and the inactive queue.
278 for (i = 0; i < PQ_COUNT; i++)
279 TAILQ_INIT(&vm_page_queues[i].pl);
280 vm_page_queues[PQ_INACTIVE].cnt = &cnt.v_inactive_count;
281 vm_page_queues[PQ_ACTIVE].cnt = &cnt.v_active_count;
282 vm_page_queues[PQ_HOLD].cnt = &cnt.v_active_count;
285 * Allocate memory for use when boot strapping the kernel memory
288 new_end = end - (boot_pages * UMA_SLAB_SIZE);
289 new_end = trunc_page(new_end);
290 mapped = pmap_map(&vaddr, new_end, end,
291 VM_PROT_READ | VM_PROT_WRITE);
292 bzero((void *)mapped, end - new_end);
293 uma_startup((void *)mapped, boot_pages);
295 #if defined(__amd64__) || defined(__i386__) || defined(__arm__) || \
298 * Allocate a bitmap to indicate that a random physical page
299 * needs to be included in a minidump.
301 * The amd64 port needs this to indicate which direct map pages
302 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
304 * However, i386 still needs this workspace internally within the
305 * minidump code. In theory, they are not needed on i386, but are
306 * included should the sf_buf code decide to use them.
308 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE;
309 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
310 new_end -= vm_page_dump_size;
311 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
312 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
313 bzero((void *)vm_page_dump, vm_page_dump_size);
316 * Compute the number of pages of memory that will be available for
317 * use (taking into account the overhead of a page structure per
320 first_page = low_water / PAGE_SIZE;
321 #ifdef VM_PHYSSEG_SPARSE
323 for (i = 0; phys_avail[i + 1] != 0; i += 2)
324 page_range += atop(phys_avail[i + 1] - phys_avail[i]);
325 #elif defined(VM_PHYSSEG_DENSE)
326 page_range = high_water / PAGE_SIZE - first_page;
328 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
333 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
338 * Initialize the mem entry structures now, and put them in the free
341 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
342 mapped = pmap_map(&vaddr, new_end, end,
343 VM_PROT_READ | VM_PROT_WRITE);
344 vm_page_array = (vm_page_t) mapped;
345 #if VM_NRESERVLEVEL > 0
347 * Allocate memory for the reservation management system's data
350 new_end = vm_reserv_startup(&vaddr, new_end, high_water);
354 * pmap_map on amd64 comes out of the direct-map, not kvm like i386,
355 * so the pages must be tracked for a crashdump to include this data.
356 * This includes the vm_page_array and the early UMA bootstrap pages.
358 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
361 phys_avail[biggestone + 1] = new_end;
364 * Clear all of the page structures
366 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
367 for (i = 0; i < page_range; i++)
368 vm_page_array[i].order = VM_NFREEORDER;
369 vm_page_array_size = page_range;
372 * Initialize the physical memory allocator.
377 * Add every available physical page that is not blacklisted to
380 cnt.v_page_count = 0;
381 cnt.v_free_count = 0;
382 list = getenv("vm.blacklist");
383 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
385 last_pa = phys_avail[i + 1];
386 while (pa < last_pa) {
388 vm_page_blacklist_lookup(list, pa))
389 printf("Skipping page with pa 0x%jx\n",
392 vm_phys_add_page(pa);
397 #if VM_NRESERVLEVEL > 0
399 * Initialize the reservation management system.
407 vm_page_flag_set(vm_page_t m, unsigned short bits)
410 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
415 vm_page_flag_clear(vm_page_t m, unsigned short bits)
418 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
423 vm_page_busy(vm_page_t m)
426 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
427 KASSERT((m->oflags & VPO_BUSY) == 0,
428 ("vm_page_busy: page already busy!!!"));
429 m->oflags |= VPO_BUSY;
435 * wakeup anyone waiting for the page.
438 vm_page_flash(vm_page_t m)
441 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
442 if (m->oflags & VPO_WANTED) {
443 m->oflags &= ~VPO_WANTED;
451 * clear the VPO_BUSY flag and wakeup anyone waiting for the
456 vm_page_wakeup(vm_page_t m)
459 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
460 KASSERT(m->oflags & VPO_BUSY, ("vm_page_wakeup: page not busy!!!"));
461 m->oflags &= ~VPO_BUSY;
466 vm_page_io_start(vm_page_t m)
469 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
474 vm_page_io_finish(vm_page_t m)
477 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
484 * Keep page from being freed by the page daemon
485 * much of the same effect as wiring, except much lower
486 * overhead and should be used only for *very* temporary
487 * holding ("wiring").
490 vm_page_hold(vm_page_t mem)
493 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
498 vm_page_unhold(vm_page_t mem)
501 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
503 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
504 if (mem->hold_count == 0 && VM_PAGE_INQUEUE2(mem, PQ_HOLD))
505 vm_page_free_toq(mem);
514 vm_page_free(vm_page_t m)
517 m->flags &= ~PG_ZERO;
524 * Free a page to the zerod-pages queue
527 vm_page_free_zero(vm_page_t m)
537 * Sleep and release the page queues lock.
539 * The object containing the given page must be locked.
542 vm_page_sleep(vm_page_t m, const char *msg)
545 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
546 if (!mtx_owned(&vm_page_queue_mtx))
547 vm_page_lock_queues();
548 vm_page_flag_set(m, PG_REFERENCED);
549 vm_page_unlock_queues();
552 * It's possible that while we sleep, the page will get
553 * unbusied and freed. If we are holding the object
554 * lock, we will assume we hold a reference to the object
555 * such that even if m->object changes, we can re-lock
558 m->oflags |= VPO_WANTED;
559 msleep(m, VM_OBJECT_MTX(m->object), PVM, msg, 0);
565 * make page all dirty
568 vm_page_dirty(vm_page_t m)
571 KASSERT((m->flags & PG_CACHED) == 0,
572 ("vm_page_dirty: page in cache!"));
573 KASSERT(!VM_PAGE_IS_FREE(m),
574 ("vm_page_dirty: page is free!"));
575 KASSERT(m->valid == VM_PAGE_BITS_ALL,
576 ("vm_page_dirty: page is invalid!"));
577 m->dirty = VM_PAGE_BITS_ALL;
583 * Implements Sleator and Tarjan's top-down splay algorithm. Returns
584 * the vm_page containing the given pindex. If, however, that
585 * pindex is not found in the vm_object, returns a vm_page that is
586 * adjacent to the pindex, coming before or after it.
589 vm_page_splay(vm_pindex_t pindex, vm_page_t root)
591 struct vm_page dummy;
592 vm_page_t lefttreemax, righttreemin, y;
596 lefttreemax = righttreemin = &dummy;
598 if (pindex < root->pindex) {
599 if ((y = root->left) == NULL)
601 if (pindex < y->pindex) {
603 root->left = y->right;
606 if ((y = root->left) == NULL)
609 /* Link into the new root's right tree. */
610 righttreemin->left = root;
612 } else if (pindex > root->pindex) {
613 if ((y = root->right) == NULL)
615 if (pindex > y->pindex) {
617 root->right = y->left;
620 if ((y = root->right) == NULL)
623 /* Link into the new root's left tree. */
624 lefttreemax->right = root;
629 /* Assemble the new root. */
630 lefttreemax->right = root->left;
631 righttreemin->left = root->right;
632 root->left = dummy.right;
633 root->right = dummy.left;
638 * vm_page_insert: [ internal use only ]
640 * Inserts the given mem entry into the object and object list.
642 * The pagetables are not updated but will presumably fault the page
643 * in if necessary, or if a kernel page the caller will at some point
644 * enter the page into the kernel's pmap. We are not allowed to block
645 * here so we *can't* do this anyway.
647 * The object and page must be locked.
648 * This routine may not block.
651 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
655 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
656 if (m->object != NULL)
657 panic("vm_page_insert: page already inserted");
660 * Record the object/offset pair in this page
666 * Now link into the object's ordered list of backed pages.
672 TAILQ_INSERT_TAIL(&object->memq, m, listq);
674 root = vm_page_splay(pindex, root);
675 if (pindex < root->pindex) {
676 m->left = root->left;
679 TAILQ_INSERT_BEFORE(root, m, listq);
680 } else if (pindex == root->pindex)
681 panic("vm_page_insert: offset already allocated");
683 m->right = root->right;
686 TAILQ_INSERT_AFTER(&object->memq, root, m, listq);
690 object->generation++;
693 * show that the object has one more resident page.
695 object->resident_page_count++;
697 * Hold the vnode until the last page is released.
699 if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
700 vhold((struct vnode *)object->handle);
703 * Since we are inserting a new and possibly dirty page,
704 * update the object's OBJ_MIGHTBEDIRTY flag.
706 if (m->flags & PG_WRITEABLE)
707 vm_object_set_writeable_dirty(object);
712 * NOTE: used by device pager as well -wfj
714 * Removes the given mem entry from the object/offset-page
715 * table and the object page list, but do not invalidate/terminate
718 * The object and page must be locked.
719 * The underlying pmap entry (if any) is NOT removed here.
720 * This routine may not block.
723 vm_page_remove(vm_page_t m)
728 if ((object = m->object) == NULL)
730 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
731 if (m->oflags & VPO_BUSY) {
732 m->oflags &= ~VPO_BUSY;
735 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
738 * Now remove from the object's list of backed pages.
740 if (m != object->root)
741 vm_page_splay(m->pindex, object->root);
745 root = vm_page_splay(m->pindex, m->left);
746 root->right = m->right;
749 TAILQ_REMOVE(&object->memq, m, listq);
752 * And show that the object has one fewer resident page.
754 object->resident_page_count--;
756 * The vnode may now be recycled.
758 if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
759 vdrop((struct vnode *)object->handle);
767 * Returns the page associated with the object/offset
768 * pair specified; if none is found, NULL is returned.
770 * The object must be locked.
771 * This routine may not block.
772 * This is a critical path routine
775 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
779 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
780 if ((m = object->root) != NULL && m->pindex != pindex) {
781 m = vm_page_splay(pindex, m);
782 if ((object->root = m)->pindex != pindex)
789 * vm_page_find_least:
791 * Returns the page associated with the object with least pindex
792 * greater than or equal to the parameter pindex, or NULL.
794 * The object must be locked.
795 * The routine may not block.
798 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
802 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
803 if ((m = TAILQ_FIRST(&object->memq)) != NULL) {
804 if (m->pindex < pindex) {
805 m = vm_page_splay(pindex, object->root);
806 if ((object->root = m)->pindex < pindex)
807 m = TAILQ_NEXT(m, listq);
814 * Returns the given page's successor (by pindex) within the object if it is
815 * resident; if none is found, NULL is returned.
817 * The object must be locked.
820 vm_page_next(vm_page_t m)
824 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
825 if ((next = TAILQ_NEXT(m, listq)) != NULL &&
826 next->pindex != m->pindex + 1)
832 * Returns the given page's predecessor (by pindex) within the object if it is
833 * resident; if none is found, NULL is returned.
835 * The object must be locked.
838 vm_page_prev(vm_page_t m)
842 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
843 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
844 prev->pindex != m->pindex - 1)
852 * Move the given memory entry from its
853 * current object to the specified target object/offset.
855 * The object must be locked.
856 * This routine may not block.
858 * Note: swap associated with the page must be invalidated by the move. We
859 * have to do this for several reasons: (1) we aren't freeing the
860 * page, (2) we are dirtying the page, (3) the VM system is probably
861 * moving the page from object A to B, and will then later move
862 * the backing store from A to B and we can't have a conflict.
864 * Note: we *always* dirty the page. It is necessary both for the
865 * fact that we moved it, and because we may be invalidating
866 * swap. If the page is on the cache, we have to deactivate it
867 * or vm_page_dirty() will panic. Dirty pages are not allowed
871 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
875 vm_page_insert(m, new_object, new_pindex);
880 * Convert all of the given object's cached pages that have a
881 * pindex within the given range into free pages. If the value
882 * zero is given for "end", then the range's upper bound is
883 * infinity. If the given object is backed by a vnode and it
884 * transitions from having one or more cached pages to none, the
885 * vnode's hold count is reduced.
888 vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
893 mtx_lock(&vm_page_queue_free_mtx);
894 if (__predict_false(object->cache == NULL)) {
895 mtx_unlock(&vm_page_queue_free_mtx);
898 m = object->cache = vm_page_splay(start, object->cache);
899 if (m->pindex < start) {
900 if (m->right == NULL)
903 m_next = vm_page_splay(start, m->right);
906 m = object->cache = m_next;
911 * At this point, "m" is either (1) a reference to the page
912 * with the least pindex that is greater than or equal to
913 * "start" or (2) NULL.
915 for (; m != NULL && (m->pindex < end || end == 0); m = m_next) {
917 * Find "m"'s successor and remove "m" from the
920 if (m->right == NULL) {
921 object->cache = m->left;
924 m_next = vm_page_splay(start, m->right);
925 m_next->left = m->left;
926 object->cache = m_next;
928 /* Convert "m" to a free page. */
931 /* Clear PG_CACHED and set PG_FREE. */
932 m->flags ^= PG_CACHED | PG_FREE;
933 KASSERT((m->flags & (PG_CACHED | PG_FREE)) == PG_FREE,
934 ("vm_page_cache_free: page %p has inconsistent flags", m));
938 empty = object->cache == NULL;
939 mtx_unlock(&vm_page_queue_free_mtx);
940 if (object->type == OBJT_VNODE && empty)
941 vdrop(object->handle);
945 * Returns the cached page that is associated with the given
946 * object and offset. If, however, none exists, returns NULL.
948 * The free page queue must be locked.
950 static inline vm_page_t
951 vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex)
955 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
956 if ((m = object->cache) != NULL && m->pindex != pindex) {
957 m = vm_page_splay(pindex, m);
958 if ((object->cache = m)->pindex != pindex)
965 * Remove the given cached page from its containing object's
966 * collection of cached pages.
968 * The free page queue must be locked.
971 vm_page_cache_remove(vm_page_t m)
976 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
977 KASSERT((m->flags & PG_CACHED) != 0,
978 ("vm_page_cache_remove: page %p is not cached", m));
980 if (m != object->cache) {
981 root = vm_page_splay(m->pindex, object->cache);
983 ("vm_page_cache_remove: page %p is not cached in object %p",
988 else if (m->right == NULL)
991 root = vm_page_splay(m->pindex, m->left);
992 root->right = m->right;
994 object->cache = root;
1000 * Transfer all of the cached pages with offset greater than or
1001 * equal to 'offidxstart' from the original object's cache to the
1002 * new object's cache. However, any cached pages with offset
1003 * greater than or equal to the new object's size are kept in the
1004 * original object. Initially, the new object's cache must be
1005 * empty. Offset 'offidxstart' in the original object must
1006 * correspond to offset zero in the new object.
1008 * The new object must be locked.
1011 vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart,
1012 vm_object_t new_object)
1014 vm_page_t m, m_next;
1017 * Insertion into an object's collection of cached pages
1018 * requires the object to be locked. In contrast, removal does
1021 VM_OBJECT_LOCK_ASSERT(new_object, MA_OWNED);
1022 KASSERT(new_object->cache == NULL,
1023 ("vm_page_cache_transfer: object %p has cached pages",
1025 mtx_lock(&vm_page_queue_free_mtx);
1026 if ((m = orig_object->cache) != NULL) {
1028 * Transfer all of the pages with offset greater than or
1029 * equal to 'offidxstart' from the original object's
1030 * cache to the new object's cache.
1032 m = vm_page_splay(offidxstart, m);
1033 if (m->pindex < offidxstart) {
1034 orig_object->cache = m;
1035 new_object->cache = m->right;
1038 orig_object->cache = m->left;
1039 new_object->cache = m;
1042 while ((m = new_object->cache) != NULL) {
1043 if ((m->pindex - offidxstart) >= new_object->size) {
1045 * Return all of the cached pages with
1046 * offset greater than or equal to the
1047 * new object's size to the original
1050 new_object->cache = m->left;
1051 m->left = orig_object->cache;
1052 orig_object->cache = m;
1055 m_next = vm_page_splay(m->pindex, m->right);
1056 /* Update the page's object and offset. */
1057 m->object = new_object;
1058 m->pindex -= offidxstart;
1063 new_object->cache = m_next;
1065 KASSERT(new_object->cache == NULL ||
1066 new_object->type == OBJT_SWAP,
1067 ("vm_page_cache_transfer: object %p's type is incompatible"
1068 " with cached pages", new_object));
1070 mtx_unlock(&vm_page_queue_free_mtx);
1076 * Allocate and return a memory cell associated
1077 * with this VM object/offset pair.
1079 * The caller must always specify an allocation class.
1081 * allocation classes:
1082 * VM_ALLOC_NORMAL normal process request
1083 * VM_ALLOC_SYSTEM system *really* needs a page
1084 * VM_ALLOC_INTERRUPT interrupt time request
1086 * optional allocation flags:
1087 * VM_ALLOC_ZERO prefer a zeroed page
1088 * VM_ALLOC_WIRED wire the allocated page
1089 * VM_ALLOC_NOOBJ page is not associated with a vm object
1090 * VM_ALLOC_NOBUSY do not set the page busy
1091 * VM_ALLOC_IFCACHED return page only if it is cached
1092 * VM_ALLOC_IFNOTCACHED return NULL, do not reactivate if the page
1095 * This routine may not sleep.
1098 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1100 struct vnode *vp = NULL;
1101 vm_object_t m_object;
1103 int flags, page_req;
1105 page_req = req & VM_ALLOC_CLASS_MASK;
1106 KASSERT(curthread->td_intr_nesting_level == 0 ||
1107 page_req == VM_ALLOC_INTERRUPT,
1108 ("vm_page_alloc(NORMAL|SYSTEM) in interrupt context"));
1110 if ((req & VM_ALLOC_NOOBJ) == 0) {
1111 KASSERT(object != NULL,
1112 ("vm_page_alloc: NULL object."));
1113 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1117 * The pager is allowed to eat deeper into the free page list.
1119 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) {
1120 page_req = VM_ALLOC_SYSTEM;
1123 mtx_lock(&vm_page_queue_free_mtx);
1124 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved ||
1125 (page_req == VM_ALLOC_SYSTEM &&
1126 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) ||
1127 (page_req == VM_ALLOC_INTERRUPT &&
1128 cnt.v_free_count + cnt.v_cache_count > 0)) {
1130 * Allocate from the free queue if the number of free pages
1131 * exceeds the minimum for the request class.
1133 if (object != NULL &&
1134 (m = vm_page_cache_lookup(object, pindex)) != NULL) {
1135 if ((req & VM_ALLOC_IFNOTCACHED) != 0) {
1136 mtx_unlock(&vm_page_queue_free_mtx);
1139 if (vm_phys_unfree_page(m))
1140 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0);
1141 #if VM_NRESERVLEVEL > 0
1142 else if (!vm_reserv_reactivate_page(m))
1146 panic("vm_page_alloc: cache page %p is missing"
1147 " from the free queue", m);
1148 } else if ((req & VM_ALLOC_IFCACHED) != 0) {
1149 mtx_unlock(&vm_page_queue_free_mtx);
1151 #if VM_NRESERVLEVEL > 0
1152 } else if (object == NULL || object->type == OBJT_DEVICE ||
1153 object->type == OBJT_SG ||
1154 (object->flags & OBJ_COLORED) == 0 ||
1155 (m = vm_reserv_alloc_page(object, pindex)) == NULL) {
1159 m = vm_phys_alloc_pages(object != NULL ?
1160 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1161 #if VM_NRESERVLEVEL > 0
1162 if (m == NULL && vm_reserv_reclaim_inactive()) {
1163 m = vm_phys_alloc_pages(object != NULL ?
1164 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1171 * Not allocatable, give up.
1173 mtx_unlock(&vm_page_queue_free_mtx);
1174 atomic_add_int(&vm_pageout_deficit, 1);
1175 pagedaemon_wakeup();
1180 * At this point we had better have found a good page.
1183 KASSERT(m != NULL, ("vm_page_alloc: missing page"));
1184 KASSERT(m->queue == PQ_NONE,
1185 ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue));
1186 KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m));
1187 KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m));
1188 KASSERT(m->busy == 0, ("vm_page_alloc: page %p is busy", m));
1189 KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m));
1190 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1191 ("vm_page_alloc: page %p has unexpected memattr %d", m,
1192 pmap_page_get_memattr(m)));
1193 if ((m->flags & PG_CACHED) != 0) {
1194 KASSERT(m->valid != 0,
1195 ("vm_page_alloc: cached page %p is invalid", m));
1196 if (m->object == object && m->pindex == pindex)
1197 cnt.v_reactivated++;
1200 m_object = m->object;
1201 vm_page_cache_remove(m);
1202 if (m_object->type == OBJT_VNODE && m_object->cache == NULL)
1203 vp = m_object->handle;
1205 KASSERT(VM_PAGE_IS_FREE(m),
1206 ("vm_page_alloc: page %p is not free", m));
1207 KASSERT(m->valid == 0,
1208 ("vm_page_alloc: free page %p is valid", m));
1213 * Initialize structure. Only the PG_ZERO flag is inherited.
1216 if (m->flags & PG_ZERO) {
1217 vm_page_zero_count--;
1218 if (req & VM_ALLOC_ZERO)
1221 if (object == NULL || object->type == OBJT_PHYS)
1222 flags |= PG_UNMANAGED;
1224 if (req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ))
1227 m->oflags = VPO_BUSY;
1228 if (req & VM_ALLOC_WIRED) {
1229 atomic_add_int(&cnt.v_wire_count, 1);
1233 mtx_unlock(&vm_page_queue_free_mtx);
1235 if (object != NULL) {
1236 /* Ignore device objects; the pager sets "memattr" for them. */
1237 if (object->memattr != VM_MEMATTR_DEFAULT &&
1238 object->type != OBJT_DEVICE && object->type != OBJT_SG)
1239 pmap_page_set_memattr(m, object->memattr);
1240 vm_page_insert(m, object, pindex);
1245 * The following call to vdrop() must come after the above call
1246 * to vm_page_insert() in case both affect the same object and
1247 * vnode. Otherwise, the affected vnode's hold count could
1248 * temporarily become zero.
1254 * Don't wakeup too often - wakeup the pageout daemon when
1255 * we would be nearly out of memory.
1257 if (vm_paging_needed())
1258 pagedaemon_wakeup();
1264 * Initialize a page that has been freshly dequeued from a freelist.
1265 * The caller has to drop the vnode returned, if it is not NULL.
1267 * To be called with vm_page_queue_free_mtx held.
1270 vm_page_alloc_init(vm_page_t m)
1273 vm_object_t m_object;
1275 KASSERT(m->queue == PQ_NONE,
1276 ("vm_page_alloc_init: page %p has unexpected queue %d",
1278 KASSERT(m->wire_count == 0,
1279 ("vm_page_alloc_init: page %p is wired", m));
1280 KASSERT(m->hold_count == 0,
1281 ("vm_page_alloc_init: page %p is held", m));
1282 KASSERT(m->busy == 0,
1283 ("vm_page_alloc_init: page %p is busy", m));
1284 KASSERT(m->dirty == 0,
1285 ("vm_page_alloc_init: page %p is dirty", m));
1286 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1287 ("vm_page_alloc_init: page %p has unexpected memattr %d",
1288 m, pmap_page_get_memattr(m)));
1289 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1291 if ((m->flags & PG_CACHED) != 0) {
1293 m_object = m->object;
1294 vm_page_cache_remove(m);
1295 if (m_object->type == OBJT_VNODE &&
1296 m_object->cache == NULL)
1297 drop = m_object->handle;
1299 KASSERT(VM_PAGE_IS_FREE(m),
1300 ("vm_page_alloc_init: page %p is not free", m));
1301 KASSERT(m->valid == 0,
1302 ("vm_page_alloc_init: free page %p is valid", m));
1305 if (m->flags & PG_ZERO)
1306 vm_page_zero_count--;
1307 /* Don't clear the PG_ZERO flag; we'll need it later. */
1308 m->flags = PG_UNMANAGED | (m->flags & PG_ZERO);
1310 /* Unmanaged pages don't use "act_count". */
1315 * vm_page_alloc_freelist:
1317 * Allocate a page from the specified freelist.
1318 * Only the ALLOC_CLASS values in req are honored, other request flags
1322 vm_page_alloc_freelist(int flind, int req)
1329 page_req = req & VM_ALLOC_CLASS_MASK;
1330 mtx_lock(&vm_page_queue_free_mtx);
1332 * Do not allocate reserved pages unless the req has asked for it.
1334 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved ||
1335 (page_req == VM_ALLOC_SYSTEM &&
1336 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) ||
1337 (page_req == VM_ALLOC_INTERRUPT &&
1338 cnt.v_free_count + cnt.v_cache_count > 0)) {
1339 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0);
1342 mtx_unlock(&vm_page_queue_free_mtx);
1345 drop = vm_page_alloc_init(m);
1346 mtx_unlock(&vm_page_queue_free_mtx);
1353 * vm_wait: (also see VM_WAIT macro)
1355 * Block until free pages are available for allocation
1356 * - Called in various places before memory allocations.
1362 mtx_lock(&vm_page_queue_free_mtx);
1363 if (curproc == pageproc) {
1364 vm_pageout_pages_needed = 1;
1365 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx,
1366 PDROP | PSWP, "VMWait", 0);
1368 if (!vm_pages_needed) {
1369 vm_pages_needed = 1;
1370 wakeup(&vm_pages_needed);
1372 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM,
1378 * vm_waitpfault: (also see VM_WAITPFAULT macro)
1380 * Block until free pages are available for allocation
1381 * - Called only in vm_fault so that processes page faulting
1382 * can be easily tracked.
1383 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
1384 * processes will be able to grab memory first. Do not change
1385 * this balance without careful testing first.
1391 mtx_lock(&vm_page_queue_free_mtx);
1392 if (!vm_pages_needed) {
1393 vm_pages_needed = 1;
1394 wakeup(&vm_pages_needed);
1396 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER,
1403 * If the given page is contained within a page queue, move it to the tail
1406 * The page queues must be locked.
1409 vm_page_requeue(vm_page_t m)
1411 int queue = VM_PAGE_GETQUEUE(m);
1412 struct vpgqueues *vpq;
1414 if (queue != PQ_NONE) {
1415 vpq = &vm_page_queues[queue];
1416 TAILQ_REMOVE(&vpq->pl, m, pageq);
1417 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
1424 * Remove a page from its queue.
1426 * The queue containing the given page must be locked.
1427 * This routine may not block.
1430 vm_pageq_remove(vm_page_t m)
1432 int queue = VM_PAGE_GETQUEUE(m);
1433 struct vpgqueues *pq;
1435 if (queue != PQ_NONE) {
1436 VM_PAGE_SETQUEUE2(m, PQ_NONE);
1437 pq = &vm_page_queues[queue];
1438 TAILQ_REMOVE(&pq->pl, m, pageq);
1446 * Add the given page to the specified queue.
1448 * The page queues must be locked.
1451 vm_page_enqueue(int queue, vm_page_t m)
1453 struct vpgqueues *vpq;
1455 vpq = &vm_page_queues[queue];
1456 VM_PAGE_SETQUEUE2(m, queue);
1457 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
1464 * Put the specified page on the active list (if appropriate).
1465 * Ensure that act_count is at least ACT_INIT but do not otherwise
1468 * The page queues must be locked.
1469 * This routine may not block.
1472 vm_page_activate(vm_page_t m)
1475 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1476 if (VM_PAGE_GETKNOWNQUEUE2(m) != PQ_ACTIVE) {
1478 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1479 if (m->act_count < ACT_INIT)
1480 m->act_count = ACT_INIT;
1481 vm_page_enqueue(PQ_ACTIVE, m);
1484 if (m->act_count < ACT_INIT)
1485 m->act_count = ACT_INIT;
1490 * vm_page_free_wakeup:
1492 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
1493 * routine is called when a page has been added to the cache or free
1496 * The page queues must be locked.
1497 * This routine may not block.
1500 vm_page_free_wakeup(void)
1503 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1505 * if pageout daemon needs pages, then tell it that there are
1508 if (vm_pageout_pages_needed &&
1509 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) {
1510 wakeup(&vm_pageout_pages_needed);
1511 vm_pageout_pages_needed = 0;
1514 * wakeup processes that are waiting on memory if we hit a
1515 * high water mark. And wakeup scheduler process if we have
1516 * lots of memory. this process will swapin processes.
1518 if (vm_pages_needed && !vm_page_count_min()) {
1519 vm_pages_needed = 0;
1520 wakeup(&cnt.v_free_count);
1527 * Returns the given page to the free list,
1528 * disassociating it with any VM object.
1530 * Object and page must be locked prior to entry.
1531 * This routine may not block.
1535 vm_page_free_toq(vm_page_t m)
1538 if (VM_PAGE_GETQUEUE(m) != PQ_NONE)
1539 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1540 KASSERT(!pmap_page_is_mapped(m),
1541 ("vm_page_free_toq: freeing mapped page %p", m));
1542 PCPU_INC(cnt.v_tfree);
1544 if (m->busy || VM_PAGE_IS_FREE(m)) {
1546 "vm_page_free: pindex(%lu), busy(%d), VPO_BUSY(%d), hold(%d)\n",
1547 (u_long)m->pindex, m->busy, (m->oflags & VPO_BUSY) ? 1 : 0,
1549 if (VM_PAGE_IS_FREE(m))
1550 panic("vm_page_free: freeing free page");
1552 panic("vm_page_free: freeing busy page");
1556 * unqueue, then remove page. Note that we cannot destroy
1557 * the page here because we do not want to call the pager's
1558 * callback routine until after we've put the page on the
1559 * appropriate free queue.
1565 * If fictitious remove object association and
1566 * return, otherwise delay object association removal.
1568 if ((m->flags & PG_FICTITIOUS) != 0) {
1575 if (m->wire_count != 0) {
1576 if (m->wire_count > 1) {
1577 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1578 m->wire_count, (long)m->pindex);
1580 panic("vm_page_free: freeing wired page");
1582 if (m->hold_count != 0) {
1583 m->flags &= ~PG_ZERO;
1584 vm_page_enqueue(PQ_HOLD, m);
1587 * Restore the default memory attribute to the page.
1589 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
1590 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
1593 * Insert the page into the physical memory allocator's
1594 * cache/free page queues.
1596 mtx_lock(&vm_page_queue_free_mtx);
1597 m->flags |= PG_FREE;
1599 #if VM_NRESERVLEVEL > 0
1600 if (!vm_reserv_free_page(m))
1604 vm_phys_free_pages(m, 0);
1605 if ((m->flags & PG_ZERO) != 0)
1606 ++vm_page_zero_count;
1608 vm_page_zero_idle_wakeup();
1609 vm_page_free_wakeup();
1610 mtx_unlock(&vm_page_queue_free_mtx);
1617 * Mark this page as wired down by yet
1618 * another map, removing it from paging queues
1621 * The page queues must be locked.
1622 * This routine may not block.
1625 vm_page_wire(vm_page_t m)
1629 * Only bump the wire statistics if the page is not already wired,
1630 * and only unqueue the page if it is on some queue (if it is unmanaged
1631 * it is already off the queues).
1633 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1634 if (m->flags & PG_FICTITIOUS)
1636 if (m->wire_count == 0) {
1637 if ((m->flags & PG_UNMANAGED) == 0)
1639 atomic_add_int(&cnt.v_wire_count, 1);
1642 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
1648 * Release one wiring of this page, potentially
1649 * enabling it to be paged again.
1651 * Many pages placed on the inactive queue should actually go
1652 * into the cache, but it is difficult to figure out which. What
1653 * we do instead, if the inactive target is well met, is to put
1654 * clean pages at the head of the inactive queue instead of the tail.
1655 * This will cause them to be moved to the cache more quickly and
1656 * if not actively re-referenced, freed more quickly. If we just
1657 * stick these pages at the end of the inactive queue, heavy filesystem
1658 * meta-data accesses can cause an unnecessary paging load on memory bound
1659 * processes. This optimization causes one-time-use metadata to be
1660 * reused more quickly.
1662 * BUT, if we are in a low-memory situation we have no choice but to
1663 * put clean pages on the cache queue.
1665 * A number of routines use vm_page_unwire() to guarantee that the page
1666 * will go into either the inactive or active queues, and will NEVER
1667 * be placed in the cache - for example, just after dirtying a page.
1668 * dirty pages in the cache are not allowed.
1670 * The page queues must be locked.
1671 * This routine may not block.
1674 vm_page_unwire(vm_page_t m, int activate)
1677 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1678 if (m->flags & PG_FICTITIOUS)
1680 if (m->wire_count > 0) {
1682 if (m->wire_count == 0) {
1683 atomic_subtract_int(&cnt.v_wire_count, 1);
1684 if (m->flags & PG_UNMANAGED) {
1686 } else if (activate)
1687 vm_page_enqueue(PQ_ACTIVE, m);
1689 vm_page_flag_clear(m, PG_WINATCFLS);
1690 vm_page_enqueue(PQ_INACTIVE, m);
1694 panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
1700 * Move the specified page to the inactive queue. If the page has
1701 * any associated swap, the swap is deallocated.
1703 * Normally athead is 0 resulting in LRU operation. athead is set
1704 * to 1 if we want this page to be 'as if it were placed in the cache',
1705 * except without unmapping it from the process address space.
1707 * This routine may not block.
1710 _vm_page_deactivate(vm_page_t m, int athead)
1713 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1716 * Ignore if already inactive.
1718 if (VM_PAGE_INQUEUE2(m, PQ_INACTIVE))
1720 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1721 vm_page_flag_clear(m, PG_WINATCFLS);
1724 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1726 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1727 VM_PAGE_SETQUEUE2(m, PQ_INACTIVE);
1728 cnt.v_inactive_count++;
1733 vm_page_deactivate(vm_page_t m)
1735 _vm_page_deactivate(m, 0);
1739 * vm_page_try_to_cache:
1741 * Returns 0 on failure, 1 on success
1744 vm_page_try_to_cache(vm_page_t m)
1747 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1748 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1749 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1750 (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED)) {
1761 * vm_page_try_to_free()
1763 * Attempt to free the page. If we cannot free it, we do nothing.
1764 * 1 is returned on success, 0 on failure.
1767 vm_page_try_to_free(vm_page_t m)
1770 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1771 if (m->object != NULL)
1772 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1773 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1774 (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED)) {
1787 * Put the specified page onto the page cache queue (if appropriate).
1789 * This routine may not block.
1792 vm_page_cache(vm_page_t m)
1797 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1799 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1800 if ((m->flags & PG_UNMANAGED) || (m->oflags & VPO_BUSY) || m->busy ||
1801 m->hold_count || m->wire_count) {
1802 panic("vm_page_cache: attempting to cache busy page");
1806 panic("vm_page_cache: page %p is dirty", m);
1807 if (m->valid == 0 || object->type == OBJT_DEFAULT ||
1808 (object->type == OBJT_SWAP &&
1809 !vm_pager_has_page(object, m->pindex, NULL, NULL))) {
1811 * Hypothesis: A cache-elgible page belonging to a
1812 * default object or swap object but without a backing
1813 * store must be zero filled.
1818 KASSERT((m->flags & PG_CACHED) == 0,
1819 ("vm_page_cache: page %p is already cached", m));
1823 * Remove the page from the paging queues.
1828 * Remove the page from the object's collection of resident
1831 if (m != object->root)
1832 vm_page_splay(m->pindex, object->root);
1833 if (m->left == NULL)
1836 root = vm_page_splay(m->pindex, m->left);
1837 root->right = m->right;
1839 object->root = root;
1840 TAILQ_REMOVE(&object->memq, m, listq);
1841 object->resident_page_count--;
1844 * Restore the default memory attribute to the page.
1846 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
1847 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
1850 * Insert the page into the object's collection of cached pages
1851 * and the physical memory allocator's cache/free page queues.
1853 vm_page_flag_clear(m, PG_ZERO);
1854 mtx_lock(&vm_page_queue_free_mtx);
1855 m->flags |= PG_CACHED;
1856 cnt.v_cache_count++;
1857 root = object->cache;
1862 root = vm_page_splay(m->pindex, root);
1863 if (m->pindex < root->pindex) {
1864 m->left = root->left;
1867 } else if (__predict_false(m->pindex == root->pindex))
1868 panic("vm_page_cache: offset already cached");
1870 m->right = root->right;
1876 #if VM_NRESERVLEVEL > 0
1877 if (!vm_reserv_free_page(m)) {
1881 vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0);
1882 vm_phys_free_pages(m, 0);
1884 vm_page_free_wakeup();
1885 mtx_unlock(&vm_page_queue_free_mtx);
1888 * Increment the vnode's hold count if this is the object's only
1889 * cached page. Decrement the vnode's hold count if this was
1890 * the object's only resident page.
1892 if (object->type == OBJT_VNODE) {
1893 if (root == NULL && object->resident_page_count != 0)
1894 vhold(object->handle);
1895 else if (root != NULL && object->resident_page_count == 0)
1896 vdrop(object->handle);
1903 * Cache, deactivate, or do nothing as appropriate. This routine
1904 * is typically used by madvise() MADV_DONTNEED.
1906 * Generally speaking we want to move the page into the cache so
1907 * it gets reused quickly. However, this can result in a silly syndrome
1908 * due to the page recycling too quickly. Small objects will not be
1909 * fully cached. On the otherhand, if we move the page to the inactive
1910 * queue we wind up with a problem whereby very large objects
1911 * unnecessarily blow away our inactive and cache queues.
1913 * The solution is to move the pages based on a fixed weighting. We
1914 * either leave them alone, deactivate them, or move them to the cache,
1915 * where moving them to the cache has the highest weighting.
1916 * By forcing some pages into other queues we eventually force the
1917 * system to balance the queues, potentially recovering other unrelated
1918 * space from active. The idea is to not force this to happen too
1922 vm_page_dontneed(vm_page_t m)
1924 static int dnweight;
1928 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1932 * occassionally leave the page alone
1934 if ((dnw & 0x01F0) == 0 ||
1935 VM_PAGE_INQUEUE2(m, PQ_INACTIVE)) {
1936 if (m->act_count >= ACT_INIT)
1942 * Clear any references to the page. Otherwise, the page daemon will
1943 * immediately reactivate the page.
1945 vm_page_flag_clear(m, PG_REFERENCED);
1946 pmap_clear_reference(m);
1948 if (m->dirty == 0 && pmap_is_modified(m))
1951 if (m->dirty || (dnw & 0x0070) == 0) {
1953 * Deactivate the page 3 times out of 32.
1958 * Cache the page 28 times out of every 32. Note that
1959 * the page is deactivated instead of cached, but placed
1960 * at the head of the queue instead of the tail.
1964 _vm_page_deactivate(m, head);
1968 * Grab a page, waiting until we are waken up due to the page
1969 * changing state. We keep on waiting, if the page continues
1970 * to be in the object. If the page doesn't exist, first allocate it
1971 * and then conditionally zero it.
1973 * This routine may block.
1976 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
1980 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1982 if ((m = vm_page_lookup(object, pindex)) != NULL) {
1983 if (vm_page_sleep_if_busy(m, TRUE, "pgrbwt")) {
1984 if ((allocflags & VM_ALLOC_RETRY) == 0)
1988 if ((allocflags & VM_ALLOC_WIRED) != 0) {
1989 vm_page_lock_queues();
1991 vm_page_unlock_queues();
1993 if ((allocflags & VM_ALLOC_NOBUSY) == 0)
1998 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
2000 VM_OBJECT_UNLOCK(object);
2002 VM_OBJECT_LOCK(object);
2003 if ((allocflags & VM_ALLOC_RETRY) == 0)
2006 } else if (m->valid != 0)
2008 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
2014 * Mapping function for valid bits or for dirty bits in
2015 * a page. May not block.
2017 * Inputs are required to range within a page.
2020 vm_page_bits(int base, int size)
2026 base + size <= PAGE_SIZE,
2027 ("vm_page_bits: illegal base/size %d/%d", base, size)
2030 if (size == 0) /* handle degenerate case */
2033 first_bit = base >> DEV_BSHIFT;
2034 last_bit = (base + size - 1) >> DEV_BSHIFT;
2036 return ((2 << last_bit) - (1 << first_bit));
2040 * vm_page_set_valid:
2042 * Sets portions of a page valid. The arguments are expected
2043 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2044 * of any partial chunks touched by the range. The invalid portion of
2045 * such chunks will be zeroed.
2047 * (base + size) must be less then or equal to PAGE_SIZE.
2050 vm_page_set_valid(vm_page_t m, int base, int size)
2054 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2055 if (size == 0) /* handle degenerate case */
2059 * If the base is not DEV_BSIZE aligned and the valid
2060 * bit is clear, we have to zero out a portion of the
2063 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2064 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
2065 pmap_zero_page_area(m, frag, base - frag);
2068 * If the ending offset is not DEV_BSIZE aligned and the
2069 * valid bit is clear, we have to zero out a portion of
2072 endoff = base + size;
2073 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2074 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
2075 pmap_zero_page_area(m, endoff,
2076 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
2079 * Assert that no previously invalid block that is now being validated
2082 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
2083 ("vm_page_set_valid: page %p is dirty", m));
2086 * Set valid bits inclusive of any overlap.
2088 m->valid |= vm_page_bits(base, size);
2092 * vm_page_set_validclean:
2094 * Sets portions of a page valid and clean. The arguments are expected
2095 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2096 * of any partial chunks touched by the range. The invalid portion of
2097 * such chunks will be zero'd.
2099 * This routine may not block.
2101 * (base + size) must be less then or equal to PAGE_SIZE.
2104 vm_page_set_validclean(vm_page_t m, int base, int size)
2110 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2111 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2112 if (size == 0) /* handle degenerate case */
2116 * If the base is not DEV_BSIZE aligned and the valid
2117 * bit is clear, we have to zero out a portion of the
2120 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2121 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
2122 pmap_zero_page_area(m, frag, base - frag);
2125 * If the ending offset is not DEV_BSIZE aligned and the
2126 * valid bit is clear, we have to zero out a portion of
2129 endoff = base + size;
2130 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2131 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
2132 pmap_zero_page_area(m, endoff,
2133 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
2136 * Set valid, clear dirty bits. If validating the entire
2137 * page we can safely clear the pmap modify bit. We also
2138 * use this opportunity to clear the VPO_NOSYNC flag. If a process
2139 * takes a write fault on a MAP_NOSYNC memory area the flag will
2142 * We set valid bits inclusive of any overlap, but we can only
2143 * clear dirty bits for DEV_BSIZE chunks that are fully within
2146 pagebits = vm_page_bits(base, size);
2147 m->valid |= pagebits;
2149 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
2150 frag = DEV_BSIZE - frag;
2156 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
2158 m->dirty &= ~pagebits;
2159 if (base == 0 && size == PAGE_SIZE) {
2160 pmap_clear_modify(m);
2161 m->oflags &= ~VPO_NOSYNC;
2166 vm_page_clear_dirty(vm_page_t m, int base, int size)
2169 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2170 m->dirty &= ~vm_page_bits(base, size);
2174 * vm_page_set_invalid:
2176 * Invalidates DEV_BSIZE'd chunks within a page. Both the
2177 * valid and dirty bits for the effected areas are cleared.
2182 vm_page_set_invalid(vm_page_t m, int base, int size)
2186 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2187 bits = vm_page_bits(base, size);
2188 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2189 if (m->valid == VM_PAGE_BITS_ALL && bits != 0)
2196 * vm_page_zero_invalid()
2198 * The kernel assumes that the invalid portions of a page contain
2199 * garbage, but such pages can be mapped into memory by user code.
2200 * When this occurs, we must zero out the non-valid portions of the
2201 * page so user code sees what it expects.
2203 * Pages are most often semi-valid when the end of a file is mapped
2204 * into memory and the file's size is not page aligned.
2207 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
2212 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2214 * Scan the valid bits looking for invalid sections that
2215 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
2216 * valid bit may be set ) have already been zerod by
2217 * vm_page_set_validclean().
2219 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
2220 if (i == (PAGE_SIZE / DEV_BSIZE) ||
2221 (m->valid & (1 << i))
2224 pmap_zero_page_area(m,
2225 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
2232 * setvalid is TRUE when we can safely set the zero'd areas
2233 * as being valid. We can do this if there are no cache consistancy
2234 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
2237 m->valid = VM_PAGE_BITS_ALL;
2243 * Is (partial) page valid? Note that the case where size == 0
2244 * will return FALSE in the degenerate case where the page is
2245 * entirely invalid, and TRUE otherwise.
2250 vm_page_is_valid(vm_page_t m, int base, int size)
2252 int bits = vm_page_bits(base, size);
2254 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2255 if (m->valid && ((m->valid & bits) == bits))
2262 * update dirty bits from pmap/mmu. May not block.
2265 vm_page_test_dirty(vm_page_t m)
2267 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
2272 int so_zerocp_fullpage = 0;
2275 * Replace the given page with a copy. The copied page assumes
2276 * the portion of the given page's "wire_count" that is not the
2277 * responsibility of this copy-on-write mechanism.
2279 * The object containing the given page must have a non-zero
2280 * paging-in-progress count and be locked.
2283 vm_page_cowfault(vm_page_t m)
2290 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2291 KASSERT(object->paging_in_progress != 0,
2292 ("vm_page_cowfault: object %p's paging-in-progress count is zero.",
2299 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY);
2301 vm_page_insert(m, object, pindex);
2302 vm_page_unlock_queues();
2303 VM_OBJECT_UNLOCK(object);
2305 VM_OBJECT_LOCK(object);
2306 if (m == vm_page_lookup(object, pindex)) {
2307 vm_page_lock_queues();
2311 * Page disappeared during the wait.
2313 vm_page_lock_queues();
2320 * check to see if we raced with an xmit complete when
2321 * waiting to allocate a page. If so, put things back
2325 vm_page_insert(m, object, pindex);
2326 } else { /* clear COW & copy page */
2327 if (!so_zerocp_fullpage)
2328 pmap_copy_page(m, mnew);
2329 mnew->valid = VM_PAGE_BITS_ALL;
2330 vm_page_dirty(mnew);
2331 mnew->wire_count = m->wire_count - m->cow;
2332 m->wire_count = m->cow;
2337 vm_page_cowclear(vm_page_t m)
2340 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2344 * let vm_fault add back write permission lazily
2348 * sf_buf_free() will free the page, so we needn't do it here
2353 vm_page_cowsetup(vm_page_t m)
2356 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2357 if (m->cow == USHRT_MAX - 1)
2360 pmap_remove_write(m);
2364 #include "opt_ddb.h"
2366 #include <sys/kernel.h>
2368 #include <ddb/ddb.h>
2370 DB_SHOW_COMMAND(page, vm_page_print_page_info)
2372 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count);
2373 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count);
2374 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count);
2375 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count);
2376 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count);
2377 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved);
2378 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min);
2379 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target);
2380 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min);
2381 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target);
2384 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
2387 db_printf("PQ_FREE:");
2388 db_printf(" %d", cnt.v_free_count);
2391 db_printf("PQ_CACHE:");
2392 db_printf(" %d", cnt.v_cache_count);
2395 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
2396 *vm_page_queues[PQ_ACTIVE].cnt,
2397 *vm_page_queues[PQ_INACTIVE].cnt);