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__)
297 * Allocate a bitmap to indicate that a random physical page
298 * needs to be included in a minidump.
300 * The amd64 port needs this to indicate which direct map pages
301 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
303 * However, i386 still needs this workspace internally within the
304 * minidump code. In theory, they are not needed on i386, but are
305 * included should the sf_buf code decide to use them.
307 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE;
308 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
309 new_end -= vm_page_dump_size;
310 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
311 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
312 bzero((void *)vm_page_dump, vm_page_dump_size);
315 * Compute the number of pages of memory that will be available for
316 * use (taking into account the overhead of a page structure per
319 first_page = low_water / PAGE_SIZE;
320 #ifdef VM_PHYSSEG_SPARSE
322 for (i = 0; phys_avail[i + 1] != 0; i += 2)
323 page_range += atop(phys_avail[i + 1] - phys_avail[i]);
324 #elif defined(VM_PHYSSEG_DENSE)
325 page_range = high_water / PAGE_SIZE - first_page;
327 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
332 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
337 * Initialize the mem entry structures now, and put them in the free
340 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
341 mapped = pmap_map(&vaddr, new_end, end,
342 VM_PROT_READ | VM_PROT_WRITE);
343 vm_page_array = (vm_page_t) mapped;
344 #if VM_NRESERVLEVEL > 0
346 * Allocate memory for the reservation management system's data
349 new_end = vm_reserv_startup(&vaddr, new_end, high_water);
353 * pmap_map on amd64 comes out of the direct-map, not kvm like i386,
354 * so the pages must be tracked for a crashdump to include this data.
355 * This includes the vm_page_array and the early UMA bootstrap pages.
357 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
360 phys_avail[biggestone + 1] = new_end;
363 * Clear all of the page structures
365 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
366 for (i = 0; i < page_range; i++)
367 vm_page_array[i].order = VM_NFREEORDER;
368 vm_page_array_size = page_range;
371 * Initialize the physical memory allocator.
376 * Add every available physical page that is not blacklisted to
379 cnt.v_page_count = 0;
380 cnt.v_free_count = 0;
381 list = getenv("vm.blacklist");
382 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
384 last_pa = phys_avail[i + 1];
385 while (pa < last_pa) {
387 vm_page_blacklist_lookup(list, pa))
388 printf("Skipping page with pa 0x%jx\n",
391 vm_phys_add_page(pa);
396 #if VM_NRESERVLEVEL > 0
398 * Initialize the reservation management system.
406 vm_page_flag_set(vm_page_t m, unsigned short bits)
409 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
414 vm_page_flag_clear(vm_page_t m, unsigned short bits)
417 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
422 vm_page_busy(vm_page_t m)
425 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
426 KASSERT((m->oflags & VPO_BUSY) == 0,
427 ("vm_page_busy: page already busy!!!"));
428 m->oflags |= VPO_BUSY;
434 * wakeup anyone waiting for the page.
437 vm_page_flash(vm_page_t m)
440 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
441 if (m->oflags & VPO_WANTED) {
442 m->oflags &= ~VPO_WANTED;
450 * clear the VPO_BUSY flag and wakeup anyone waiting for the
455 vm_page_wakeup(vm_page_t m)
458 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
459 KASSERT(m->oflags & VPO_BUSY, ("vm_page_wakeup: page not busy!!!"));
460 m->oflags &= ~VPO_BUSY;
465 vm_page_io_start(vm_page_t m)
468 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
473 vm_page_io_finish(vm_page_t m)
476 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
483 * Keep page from being freed by the page daemon
484 * much of the same effect as wiring, except much lower
485 * overhead and should be used only for *very* temporary
486 * holding ("wiring").
489 vm_page_hold(vm_page_t mem)
492 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
497 vm_page_unhold(vm_page_t mem)
500 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
502 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
503 if (mem->hold_count == 0 && VM_PAGE_INQUEUE2(mem, PQ_HOLD))
504 vm_page_free_toq(mem);
513 vm_page_free(vm_page_t m)
516 m->flags &= ~PG_ZERO;
523 * Free a page to the zerod-pages queue
526 vm_page_free_zero(vm_page_t m)
536 * Sleep and release the page queues lock.
538 * The object containing the given page must be locked.
541 vm_page_sleep(vm_page_t m, const char *msg)
544 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
545 if (!mtx_owned(&vm_page_queue_mtx))
546 vm_page_lock_queues();
547 vm_page_flag_set(m, PG_REFERENCED);
548 vm_page_unlock_queues();
551 * It's possible that while we sleep, the page will get
552 * unbusied and freed. If we are holding the object
553 * lock, we will assume we hold a reference to the object
554 * such that even if m->object changes, we can re-lock
557 m->oflags |= VPO_WANTED;
558 msleep(m, VM_OBJECT_MTX(m->object), PVM, msg, 0);
564 * make page all dirty
567 vm_page_dirty(vm_page_t m)
570 KASSERT((m->flags & PG_CACHED) == 0,
571 ("vm_page_dirty: page in cache!"));
572 KASSERT(!VM_PAGE_IS_FREE(m),
573 ("vm_page_dirty: page is free!"));
574 KASSERT(m->valid == VM_PAGE_BITS_ALL,
575 ("vm_page_dirty: page is invalid!"));
576 m->dirty = VM_PAGE_BITS_ALL;
582 * Implements Sleator and Tarjan's top-down splay algorithm. Returns
583 * the vm_page containing the given pindex. If, however, that
584 * pindex is not found in the vm_object, returns a vm_page that is
585 * adjacent to the pindex, coming before or after it.
588 vm_page_splay(vm_pindex_t pindex, vm_page_t root)
590 struct vm_page dummy;
591 vm_page_t lefttreemax, righttreemin, y;
595 lefttreemax = righttreemin = &dummy;
597 if (pindex < root->pindex) {
598 if ((y = root->left) == NULL)
600 if (pindex < y->pindex) {
602 root->left = y->right;
605 if ((y = root->left) == NULL)
608 /* Link into the new root's right tree. */
609 righttreemin->left = root;
611 } else if (pindex > root->pindex) {
612 if ((y = root->right) == NULL)
614 if (pindex > y->pindex) {
616 root->right = y->left;
619 if ((y = root->right) == NULL)
622 /* Link into the new root's left tree. */
623 lefttreemax->right = root;
628 /* Assemble the new root. */
629 lefttreemax->right = root->left;
630 righttreemin->left = root->right;
631 root->left = dummy.right;
632 root->right = dummy.left;
637 * vm_page_insert: [ internal use only ]
639 * Inserts the given mem entry into the object and object list.
641 * The pagetables are not updated but will presumably fault the page
642 * in if necessary, or if a kernel page the caller will at some point
643 * enter the page into the kernel's pmap. We are not allowed to block
644 * here so we *can't* do this anyway.
646 * The object and page must be locked.
647 * This routine may not block.
650 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
654 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
655 if (m->object != NULL)
656 panic("vm_page_insert: page already inserted");
659 * Record the object/offset pair in this page
665 * Now link into the object's ordered list of backed pages.
671 TAILQ_INSERT_TAIL(&object->memq, m, listq);
673 root = vm_page_splay(pindex, root);
674 if (pindex < root->pindex) {
675 m->left = root->left;
678 TAILQ_INSERT_BEFORE(root, m, listq);
679 } else if (pindex == root->pindex)
680 panic("vm_page_insert: offset already allocated");
682 m->right = root->right;
685 TAILQ_INSERT_AFTER(&object->memq, root, m, listq);
689 object->generation++;
692 * show that the object has one more resident page.
694 object->resident_page_count++;
696 * Hold the vnode until the last page is released.
698 if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
699 vhold((struct vnode *)object->handle);
702 * Since we are inserting a new and possibly dirty page,
703 * update the object's OBJ_MIGHTBEDIRTY flag.
705 if (m->flags & PG_WRITEABLE)
706 vm_object_set_writeable_dirty(object);
711 * NOTE: used by device pager as well -wfj
713 * Removes the given mem entry from the object/offset-page
714 * table and the object page list, but do not invalidate/terminate
717 * The object and page must be locked.
718 * The underlying pmap entry (if any) is NOT removed here.
719 * This routine may not block.
722 vm_page_remove(vm_page_t m)
727 if ((object = m->object) == NULL)
729 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
730 if (m->oflags & VPO_BUSY) {
731 m->oflags &= ~VPO_BUSY;
734 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
737 * Now remove from the object's list of backed pages.
739 if (m != object->root)
740 vm_page_splay(m->pindex, object->root);
744 root = vm_page_splay(m->pindex, m->left);
745 root->right = m->right;
748 TAILQ_REMOVE(&object->memq, m, listq);
751 * And show that the object has one fewer resident page.
753 object->resident_page_count--;
754 object->generation++;
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)
791 * Move the given memory entry from its
792 * current object to the specified target object/offset.
794 * The object must be locked.
795 * This routine may not block.
797 * Note: swap associated with the page must be invalidated by the move. We
798 * have to do this for several reasons: (1) we aren't freeing the
799 * page, (2) we are dirtying the page, (3) the VM system is probably
800 * moving the page from object A to B, and will then later move
801 * the backing store from A to B and we can't have a conflict.
803 * Note: we *always* dirty the page. It is necessary both for the
804 * fact that we moved it, and because we may be invalidating
805 * swap. If the page is on the cache, we have to deactivate it
806 * or vm_page_dirty() will panic. Dirty pages are not allowed
810 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
814 vm_page_insert(m, new_object, new_pindex);
819 * Convert all of the given object's cached pages that have a
820 * pindex within the given range into free pages. If the value
821 * zero is given for "end", then the range's upper bound is
822 * infinity. If the given object is backed by a vnode and it
823 * transitions from having one or more cached pages to none, the
824 * vnode's hold count is reduced.
827 vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
832 mtx_lock(&vm_page_queue_free_mtx);
833 if (__predict_false(object->cache == NULL)) {
834 mtx_unlock(&vm_page_queue_free_mtx);
837 m = object->cache = vm_page_splay(start, object->cache);
838 if (m->pindex < start) {
839 if (m->right == NULL)
842 m_next = vm_page_splay(start, m->right);
845 m = object->cache = m_next;
850 * At this point, "m" is either (1) a reference to the page
851 * with the least pindex that is greater than or equal to
852 * "start" or (2) NULL.
854 for (; m != NULL && (m->pindex < end || end == 0); m = m_next) {
856 * Find "m"'s successor and remove "m" from the
859 if (m->right == NULL) {
860 object->cache = m->left;
863 m_next = vm_page_splay(start, m->right);
864 m_next->left = m->left;
865 object->cache = m_next;
867 /* Convert "m" to a free page. */
870 /* Clear PG_CACHED and set PG_FREE. */
871 m->flags ^= PG_CACHED | PG_FREE;
872 KASSERT((m->flags & (PG_CACHED | PG_FREE)) == PG_FREE,
873 ("vm_page_cache_free: page %p has inconsistent flags", m));
877 empty = object->cache == NULL;
878 mtx_unlock(&vm_page_queue_free_mtx);
879 if (object->type == OBJT_VNODE && empty)
880 vdrop(object->handle);
884 * Returns the cached page that is associated with the given
885 * object and offset. If, however, none exists, returns NULL.
887 * The free page queue must be locked.
889 static inline vm_page_t
890 vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex)
894 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
895 if ((m = object->cache) != NULL && m->pindex != pindex) {
896 m = vm_page_splay(pindex, m);
897 if ((object->cache = m)->pindex != pindex)
904 * Remove the given cached page from its containing object's
905 * collection of cached pages.
907 * The free page queue must be locked.
910 vm_page_cache_remove(vm_page_t m)
915 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
916 KASSERT((m->flags & PG_CACHED) != 0,
917 ("vm_page_cache_remove: page %p is not cached", m));
919 if (m != object->cache) {
920 root = vm_page_splay(m->pindex, object->cache);
922 ("vm_page_cache_remove: page %p is not cached in object %p",
927 else if (m->right == NULL)
930 root = vm_page_splay(m->pindex, m->left);
931 root->right = m->right;
933 object->cache = root;
939 * Transfer all of the cached pages with offset greater than or
940 * equal to 'offidxstart' from the original object's cache to the
941 * new object's cache. However, any cached pages with offset
942 * greater than or equal to the new object's size are kept in the
943 * original object. Initially, the new object's cache must be
944 * empty. Offset 'offidxstart' in the original object must
945 * correspond to offset zero in the new object.
947 * The new object must be locked.
950 vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart,
951 vm_object_t new_object)
956 * Insertion into an object's collection of cached pages
957 * requires the object to be locked. In contrast, removal does
960 VM_OBJECT_LOCK_ASSERT(new_object, MA_OWNED);
961 KASSERT(new_object->cache == NULL,
962 ("vm_page_cache_transfer: object %p has cached pages",
964 mtx_lock(&vm_page_queue_free_mtx);
965 if ((m = orig_object->cache) != NULL) {
967 * Transfer all of the pages with offset greater than or
968 * equal to 'offidxstart' from the original object's
969 * cache to the new object's cache.
971 m = vm_page_splay(offidxstart, m);
972 if (m->pindex < offidxstart) {
973 orig_object->cache = m;
974 new_object->cache = m->right;
977 orig_object->cache = m->left;
978 new_object->cache = m;
981 while ((m = new_object->cache) != NULL) {
982 if ((m->pindex - offidxstart) >= new_object->size) {
984 * Return all of the cached pages with
985 * offset greater than or equal to the
986 * new object's size to the original
989 new_object->cache = m->left;
990 m->left = orig_object->cache;
991 orig_object->cache = m;
994 m_next = vm_page_splay(m->pindex, m->right);
995 /* Update the page's object and offset. */
996 m->object = new_object;
997 m->pindex -= offidxstart;
1002 new_object->cache = m_next;
1004 KASSERT(new_object->cache == NULL ||
1005 new_object->type == OBJT_SWAP,
1006 ("vm_page_cache_transfer: object %p's type is incompatible"
1007 " with cached pages", new_object));
1009 mtx_unlock(&vm_page_queue_free_mtx);
1015 * Allocate and return a memory cell associated
1016 * with this VM object/offset pair.
1019 * VM_ALLOC_NORMAL normal process request
1020 * VM_ALLOC_SYSTEM system *really* needs a page
1021 * VM_ALLOC_INTERRUPT interrupt time request
1022 * VM_ALLOC_ZERO zero page
1024 * This routine may not block.
1027 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1029 struct vnode *vp = NULL;
1030 vm_object_t m_object;
1032 int flags, page_req;
1034 page_req = req & VM_ALLOC_CLASS_MASK;
1035 KASSERT(curthread->td_intr_nesting_level == 0 ||
1036 page_req == VM_ALLOC_INTERRUPT,
1037 ("vm_page_alloc(NORMAL|SYSTEM) in interrupt context"));
1039 if ((req & VM_ALLOC_NOOBJ) == 0) {
1040 KASSERT(object != NULL,
1041 ("vm_page_alloc: NULL object."));
1042 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1046 * The pager is allowed to eat deeper into the free page list.
1048 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) {
1049 page_req = VM_ALLOC_SYSTEM;
1052 mtx_lock(&vm_page_queue_free_mtx);
1053 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved ||
1054 (page_req == VM_ALLOC_SYSTEM &&
1055 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) ||
1056 (page_req == VM_ALLOC_INTERRUPT &&
1057 cnt.v_free_count + cnt.v_cache_count > 0)) {
1059 * Allocate from the free queue if the number of free pages
1060 * exceeds the minimum for the request class.
1062 if (object != NULL &&
1063 (m = vm_page_cache_lookup(object, pindex)) != NULL) {
1064 if ((req & VM_ALLOC_IFNOTCACHED) != 0) {
1065 mtx_unlock(&vm_page_queue_free_mtx);
1068 if (vm_phys_unfree_page(m))
1069 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0);
1070 #if VM_NRESERVLEVEL > 0
1071 else if (!vm_reserv_reactivate_page(m))
1075 panic("vm_page_alloc: cache page %p is missing"
1076 " from the free queue", m);
1077 } else if ((req & VM_ALLOC_IFCACHED) != 0) {
1078 mtx_unlock(&vm_page_queue_free_mtx);
1080 #if VM_NRESERVLEVEL > 0
1081 } else if (object == NULL || object->type == OBJT_DEVICE ||
1082 (object->flags & OBJ_COLORED) == 0 ||
1083 (m = vm_reserv_alloc_page(object, pindex)) == NULL) {
1087 m = vm_phys_alloc_pages(object != NULL ?
1088 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1089 #if VM_NRESERVLEVEL > 0
1090 if (m == NULL && vm_reserv_reclaim_inactive()) {
1091 m = vm_phys_alloc_pages(object != NULL ?
1092 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1099 * Not allocatable, give up.
1101 mtx_unlock(&vm_page_queue_free_mtx);
1102 atomic_add_int(&vm_pageout_deficit, 1);
1103 pagedaemon_wakeup();
1108 * At this point we had better have found a good page.
1111 KASSERT(m != NULL, ("vm_page_alloc: missing page"));
1112 KASSERT(m->queue == PQ_NONE,
1113 ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue));
1114 KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m));
1115 KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m));
1116 KASSERT(m->busy == 0, ("vm_page_alloc: page %p is busy", m));
1117 KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m));
1118 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1119 ("vm_page_alloc: page %p has unexpected memattr %d", m,
1120 pmap_page_get_memattr(m)));
1121 if ((m->flags & PG_CACHED) != 0) {
1122 KASSERT(m->valid != 0,
1123 ("vm_page_alloc: cached page %p is invalid", m));
1124 if (m->object == object && m->pindex == pindex)
1125 cnt.v_reactivated++;
1128 m_object = m->object;
1129 vm_page_cache_remove(m);
1130 if (m_object->type == OBJT_VNODE && m_object->cache == NULL)
1131 vp = m_object->handle;
1133 KASSERT(VM_PAGE_IS_FREE(m),
1134 ("vm_page_alloc: page %p is not free", m));
1135 KASSERT(m->valid == 0,
1136 ("vm_page_alloc: free page %p is valid", m));
1141 * Initialize structure. Only the PG_ZERO flag is inherited.
1144 if (m->flags & PG_ZERO) {
1145 vm_page_zero_count--;
1146 if (req & VM_ALLOC_ZERO)
1149 if (object == NULL || object->type == OBJT_PHYS)
1150 flags |= PG_UNMANAGED;
1152 if (req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ))
1155 m->oflags = VPO_BUSY;
1156 if (req & VM_ALLOC_WIRED) {
1157 atomic_add_int(&cnt.v_wire_count, 1);
1161 mtx_unlock(&vm_page_queue_free_mtx);
1163 if (object != NULL) {
1164 /* Ignore device objects; the pager sets "memattr" for them. */
1165 if (object->memattr != VM_MEMATTR_DEFAULT &&
1166 object->type != OBJT_DEVICE && object->type != OBJT_SG)
1167 pmap_page_set_memattr(m, object->memattr);
1168 vm_page_insert(m, object, pindex);
1173 * The following call to vdrop() must come after the above call
1174 * to vm_page_insert() in case both affect the same object and
1175 * vnode. Otherwise, the affected vnode's hold count could
1176 * temporarily become zero.
1182 * Don't wakeup too often - wakeup the pageout daemon when
1183 * we would be nearly out of memory.
1185 if (vm_paging_needed())
1186 pagedaemon_wakeup();
1192 * vm_wait: (also see VM_WAIT macro)
1194 * Block until free pages are available for allocation
1195 * - Called in various places before memory allocations.
1201 mtx_lock(&vm_page_queue_free_mtx);
1202 if (curproc == pageproc) {
1203 vm_pageout_pages_needed = 1;
1204 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx,
1205 PDROP | PSWP, "VMWait", 0);
1207 if (!vm_pages_needed) {
1208 vm_pages_needed = 1;
1209 wakeup(&vm_pages_needed);
1211 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM,
1217 * vm_waitpfault: (also see VM_WAITPFAULT macro)
1219 * Block until free pages are available for allocation
1220 * - Called only in vm_fault so that processes page faulting
1221 * can be easily tracked.
1222 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
1223 * processes will be able to grab memory first. Do not change
1224 * this balance without careful testing first.
1230 mtx_lock(&vm_page_queue_free_mtx);
1231 if (!vm_pages_needed) {
1232 vm_pages_needed = 1;
1233 wakeup(&vm_pages_needed);
1235 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER,
1242 * If the given page is contained within a page queue, move it to the tail
1245 * The page queues must be locked.
1248 vm_page_requeue(vm_page_t m)
1250 int queue = VM_PAGE_GETQUEUE(m);
1251 struct vpgqueues *vpq;
1253 if (queue != PQ_NONE) {
1254 vpq = &vm_page_queues[queue];
1255 TAILQ_REMOVE(&vpq->pl, m, pageq);
1256 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
1263 * Remove a page from its queue.
1265 * The queue containing the given page must be locked.
1266 * This routine may not block.
1269 vm_pageq_remove(vm_page_t m)
1271 int queue = VM_PAGE_GETQUEUE(m);
1272 struct vpgqueues *pq;
1274 if (queue != PQ_NONE) {
1275 VM_PAGE_SETQUEUE2(m, PQ_NONE);
1276 pq = &vm_page_queues[queue];
1277 TAILQ_REMOVE(&pq->pl, m, pageq);
1285 * Add the given page to the specified queue.
1287 * The page queues must be locked.
1290 vm_page_enqueue(int queue, vm_page_t m)
1292 struct vpgqueues *vpq;
1294 vpq = &vm_page_queues[queue];
1295 VM_PAGE_SETQUEUE2(m, queue);
1296 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
1303 * Put the specified page on the active list (if appropriate).
1304 * Ensure that act_count is at least ACT_INIT but do not otherwise
1307 * The page queues must be locked.
1308 * This routine may not block.
1311 vm_page_activate(vm_page_t m)
1314 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1315 if (VM_PAGE_GETKNOWNQUEUE2(m) != PQ_ACTIVE) {
1317 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1318 if (m->act_count < ACT_INIT)
1319 m->act_count = ACT_INIT;
1320 vm_page_enqueue(PQ_ACTIVE, m);
1323 if (m->act_count < ACT_INIT)
1324 m->act_count = ACT_INIT;
1329 * vm_page_free_wakeup:
1331 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
1332 * routine is called when a page has been added to the cache or free
1335 * The page queues must be locked.
1336 * This routine may not block.
1339 vm_page_free_wakeup(void)
1342 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1344 * if pageout daemon needs pages, then tell it that there are
1347 if (vm_pageout_pages_needed &&
1348 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) {
1349 wakeup(&vm_pageout_pages_needed);
1350 vm_pageout_pages_needed = 0;
1353 * wakeup processes that are waiting on memory if we hit a
1354 * high water mark. And wakeup scheduler process if we have
1355 * lots of memory. this process will swapin processes.
1357 if (vm_pages_needed && !vm_page_count_min()) {
1358 vm_pages_needed = 0;
1359 wakeup(&cnt.v_free_count);
1366 * Returns the given page to the free list,
1367 * disassociating it with any VM object.
1369 * Object and page must be locked prior to entry.
1370 * This routine may not block.
1374 vm_page_free_toq(vm_page_t m)
1377 if (VM_PAGE_GETQUEUE(m) != PQ_NONE)
1378 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1379 KASSERT(!pmap_page_is_mapped(m),
1380 ("vm_page_free_toq: freeing mapped page %p", m));
1381 PCPU_INC(cnt.v_tfree);
1383 if (m->busy || VM_PAGE_IS_FREE(m)) {
1385 "vm_page_free: pindex(%lu), busy(%d), VPO_BUSY(%d), hold(%d)\n",
1386 (u_long)m->pindex, m->busy, (m->oflags & VPO_BUSY) ? 1 : 0,
1388 if (VM_PAGE_IS_FREE(m))
1389 panic("vm_page_free: freeing free page");
1391 panic("vm_page_free: freeing busy page");
1395 * unqueue, then remove page. Note that we cannot destroy
1396 * the page here because we do not want to call the pager's
1397 * callback routine until after we've put the page on the
1398 * appropriate free queue.
1404 * If fictitious remove object association and
1405 * return, otherwise delay object association removal.
1407 if ((m->flags & PG_FICTITIOUS) != 0) {
1414 if (m->wire_count != 0) {
1415 if (m->wire_count > 1) {
1416 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1417 m->wire_count, (long)m->pindex);
1419 panic("vm_page_free: freeing wired page");
1421 if (m->hold_count != 0) {
1422 m->flags &= ~PG_ZERO;
1423 vm_page_enqueue(PQ_HOLD, m);
1426 * Restore the default memory attribute to the page.
1428 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
1429 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
1432 * Insert the page into the physical memory allocator's
1433 * cache/free page queues.
1435 mtx_lock(&vm_page_queue_free_mtx);
1436 m->flags |= PG_FREE;
1438 #if VM_NRESERVLEVEL > 0
1439 if (!vm_reserv_free_page(m))
1443 vm_phys_free_pages(m, 0);
1444 if ((m->flags & PG_ZERO) != 0)
1445 ++vm_page_zero_count;
1447 vm_page_zero_idle_wakeup();
1448 vm_page_free_wakeup();
1449 mtx_unlock(&vm_page_queue_free_mtx);
1456 * Mark this page as wired down by yet
1457 * another map, removing it from paging queues
1460 * The page queues must be locked.
1461 * This routine may not block.
1464 vm_page_wire(vm_page_t m)
1468 * Only bump the wire statistics if the page is not already wired,
1469 * and only unqueue the page if it is on some queue (if it is unmanaged
1470 * it is already off the queues).
1472 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1473 if (m->flags & PG_FICTITIOUS)
1475 if (m->wire_count == 0) {
1476 if ((m->flags & PG_UNMANAGED) == 0)
1478 atomic_add_int(&cnt.v_wire_count, 1);
1481 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
1487 * Release one wiring of this page, potentially
1488 * enabling it to be paged again.
1490 * Many pages placed on the inactive queue should actually go
1491 * into the cache, but it is difficult to figure out which. What
1492 * we do instead, if the inactive target is well met, is to put
1493 * clean pages at the head of the inactive queue instead of the tail.
1494 * This will cause them to be moved to the cache more quickly and
1495 * if not actively re-referenced, freed more quickly. If we just
1496 * stick these pages at the end of the inactive queue, heavy filesystem
1497 * meta-data accesses can cause an unnecessary paging load on memory bound
1498 * processes. This optimization causes one-time-use metadata to be
1499 * reused more quickly.
1501 * BUT, if we are in a low-memory situation we have no choice but to
1502 * put clean pages on the cache queue.
1504 * A number of routines use vm_page_unwire() to guarantee that the page
1505 * will go into either the inactive or active queues, and will NEVER
1506 * be placed in the cache - for example, just after dirtying a page.
1507 * dirty pages in the cache are not allowed.
1509 * The page queues must be locked.
1510 * This routine may not block.
1513 vm_page_unwire(vm_page_t m, int activate)
1516 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1517 if (m->flags & PG_FICTITIOUS)
1519 if (m->wire_count > 0) {
1521 if (m->wire_count == 0) {
1522 atomic_subtract_int(&cnt.v_wire_count, 1);
1523 if (m->flags & PG_UNMANAGED) {
1525 } else if (activate)
1526 vm_page_enqueue(PQ_ACTIVE, m);
1528 vm_page_flag_clear(m, PG_WINATCFLS);
1529 vm_page_enqueue(PQ_INACTIVE, m);
1533 panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
1539 * Move the specified page to the inactive queue. If the page has
1540 * any associated swap, the swap is deallocated.
1542 * Normally athead is 0 resulting in LRU operation. athead is set
1543 * to 1 if we want this page to be 'as if it were placed in the cache',
1544 * except without unmapping it from the process address space.
1546 * This routine may not block.
1549 _vm_page_deactivate(vm_page_t m, int athead)
1552 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1555 * Ignore if already inactive.
1557 if (VM_PAGE_INQUEUE2(m, PQ_INACTIVE))
1559 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1560 vm_page_flag_clear(m, PG_WINATCFLS);
1563 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1565 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1566 VM_PAGE_SETQUEUE2(m, PQ_INACTIVE);
1567 cnt.v_inactive_count++;
1572 vm_page_deactivate(vm_page_t m)
1574 _vm_page_deactivate(m, 0);
1578 * vm_page_try_to_cache:
1580 * Returns 0 on failure, 1 on success
1583 vm_page_try_to_cache(vm_page_t m)
1586 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1587 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1588 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1589 (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED)) {
1600 * vm_page_try_to_free()
1602 * Attempt to free the page. If we cannot free it, we do nothing.
1603 * 1 is returned on success, 0 on failure.
1606 vm_page_try_to_free(vm_page_t m)
1609 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1610 if (m->object != NULL)
1611 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1612 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1613 (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED)) {
1626 * Put the specified page onto the page cache queue (if appropriate).
1628 * This routine may not block.
1631 vm_page_cache(vm_page_t m)
1636 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1638 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1639 if ((m->flags & PG_UNMANAGED) || (m->oflags & VPO_BUSY) || m->busy ||
1640 m->hold_count || m->wire_count) {
1641 panic("vm_page_cache: attempting to cache busy page");
1645 panic("vm_page_cache: page %p is dirty", m);
1646 if (m->valid == 0 || object->type == OBJT_DEFAULT ||
1647 (object->type == OBJT_SWAP &&
1648 !vm_pager_has_page(object, m->pindex, NULL, NULL))) {
1650 * Hypothesis: A cache-elgible page belonging to a
1651 * default object or swap object but without a backing
1652 * store must be zero filled.
1657 KASSERT((m->flags & PG_CACHED) == 0,
1658 ("vm_page_cache: page %p is already cached", m));
1662 * Remove the page from the paging queues.
1667 * Remove the page from the object's collection of resident
1670 if (m != object->root)
1671 vm_page_splay(m->pindex, object->root);
1672 if (m->left == NULL)
1675 root = vm_page_splay(m->pindex, m->left);
1676 root->right = m->right;
1678 object->root = root;
1679 TAILQ_REMOVE(&object->memq, m, listq);
1680 object->resident_page_count--;
1681 object->generation++;
1684 * Restore the default memory attribute to the page.
1686 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
1687 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
1690 * Insert the page into the object's collection of cached pages
1691 * and the physical memory allocator's cache/free page queues.
1693 vm_page_flag_clear(m, PG_ZERO);
1694 mtx_lock(&vm_page_queue_free_mtx);
1695 m->flags |= PG_CACHED;
1696 cnt.v_cache_count++;
1697 root = object->cache;
1702 root = vm_page_splay(m->pindex, root);
1703 if (m->pindex < root->pindex) {
1704 m->left = root->left;
1707 } else if (__predict_false(m->pindex == root->pindex))
1708 panic("vm_page_cache: offset already cached");
1710 m->right = root->right;
1716 #if VM_NRESERVLEVEL > 0
1717 if (!vm_reserv_free_page(m)) {
1721 vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0);
1722 vm_phys_free_pages(m, 0);
1724 vm_page_free_wakeup();
1725 mtx_unlock(&vm_page_queue_free_mtx);
1728 * Increment the vnode's hold count if this is the object's only
1729 * cached page. Decrement the vnode's hold count if this was
1730 * the object's only resident page.
1732 if (object->type == OBJT_VNODE) {
1733 if (root == NULL && object->resident_page_count != 0)
1734 vhold(object->handle);
1735 else if (root != NULL && object->resident_page_count == 0)
1736 vdrop(object->handle);
1743 * Cache, deactivate, or do nothing as appropriate. This routine
1744 * is typically used by madvise() MADV_DONTNEED.
1746 * Generally speaking we want to move the page into the cache so
1747 * it gets reused quickly. However, this can result in a silly syndrome
1748 * due to the page recycling too quickly. Small objects will not be
1749 * fully cached. On the otherhand, if we move the page to the inactive
1750 * queue we wind up with a problem whereby very large objects
1751 * unnecessarily blow away our inactive and cache queues.
1753 * The solution is to move the pages based on a fixed weighting. We
1754 * either leave them alone, deactivate them, or move them to the cache,
1755 * where moving them to the cache has the highest weighting.
1756 * By forcing some pages into other queues we eventually force the
1757 * system to balance the queues, potentially recovering other unrelated
1758 * space from active. The idea is to not force this to happen too
1762 vm_page_dontneed(vm_page_t m)
1764 static int dnweight;
1768 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1772 * occassionally leave the page alone
1774 if ((dnw & 0x01F0) == 0 ||
1775 VM_PAGE_INQUEUE2(m, PQ_INACTIVE)) {
1776 if (m->act_count >= ACT_INIT)
1782 * Clear any references to the page. Otherwise, the page daemon will
1783 * immediately reactivate the page.
1785 vm_page_flag_clear(m, PG_REFERENCED);
1786 pmap_clear_reference(m);
1788 if (m->dirty == 0 && pmap_is_modified(m))
1791 if (m->dirty || (dnw & 0x0070) == 0) {
1793 * Deactivate the page 3 times out of 32.
1798 * Cache the page 28 times out of every 32. Note that
1799 * the page is deactivated instead of cached, but placed
1800 * at the head of the queue instead of the tail.
1804 _vm_page_deactivate(m, head);
1808 * Grab a page, waiting until we are waken up due to the page
1809 * changing state. We keep on waiting, if the page continues
1810 * to be in the object. If the page doesn't exist, first allocate it
1811 * and then conditionally zero it.
1813 * This routine may block.
1816 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
1820 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1822 if ((m = vm_page_lookup(object, pindex)) != NULL) {
1823 if (vm_page_sleep_if_busy(m, TRUE, "pgrbwt")) {
1824 if ((allocflags & VM_ALLOC_RETRY) == 0)
1828 if ((allocflags & VM_ALLOC_WIRED) != 0) {
1829 vm_page_lock_queues();
1831 vm_page_unlock_queues();
1833 if ((allocflags & VM_ALLOC_NOBUSY) == 0)
1838 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1840 VM_OBJECT_UNLOCK(object);
1842 VM_OBJECT_LOCK(object);
1843 if ((allocflags & VM_ALLOC_RETRY) == 0)
1846 } else if (m->valid != 0)
1848 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
1854 * Mapping function for valid bits or for dirty bits in
1855 * a page. May not block.
1857 * Inputs are required to range within a page.
1860 vm_page_bits(int base, int size)
1866 base + size <= PAGE_SIZE,
1867 ("vm_page_bits: illegal base/size %d/%d", base, size)
1870 if (size == 0) /* handle degenerate case */
1873 first_bit = base >> DEV_BSHIFT;
1874 last_bit = (base + size - 1) >> DEV_BSHIFT;
1876 return ((2 << last_bit) - (1 << first_bit));
1880 * vm_page_set_valid:
1882 * Sets portions of a page valid. The arguments are expected
1883 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1884 * of any partial chunks touched by the range. The invalid portion of
1885 * such chunks will be zeroed.
1887 * (base + size) must be less then or equal to PAGE_SIZE.
1890 vm_page_set_valid(vm_page_t m, int base, int size)
1894 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1895 if (size == 0) /* handle degenerate case */
1899 * If the base is not DEV_BSIZE aligned and the valid
1900 * bit is clear, we have to zero out a portion of the
1903 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
1904 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
1905 pmap_zero_page_area(m, frag, base - frag);
1908 * If the ending offset is not DEV_BSIZE aligned and the
1909 * valid bit is clear, we have to zero out a portion of
1912 endoff = base + size;
1913 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
1914 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
1915 pmap_zero_page_area(m, endoff,
1916 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
1919 * Assert that no previously invalid block that is now being validated
1922 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
1923 ("vm_page_set_valid: page %p is dirty", m));
1926 * Set valid bits inclusive of any overlap.
1928 m->valid |= vm_page_bits(base, size);
1932 * vm_page_set_validclean:
1934 * Sets portions of a page valid and clean. The arguments are expected
1935 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1936 * of any partial chunks touched by the range. The invalid portion of
1937 * such chunks will be zero'd.
1939 * This routine may not block.
1941 * (base + size) must be less then or equal to PAGE_SIZE.
1944 vm_page_set_validclean(vm_page_t m, int base, int size)
1950 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1951 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1952 if (size == 0) /* handle degenerate case */
1956 * If the base is not DEV_BSIZE aligned and the valid
1957 * bit is clear, we have to zero out a portion of the
1960 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
1961 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
1962 pmap_zero_page_area(m, frag, base - frag);
1965 * If the ending offset is not DEV_BSIZE aligned and the
1966 * valid bit is clear, we have to zero out a portion of
1969 endoff = base + size;
1970 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
1971 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
1972 pmap_zero_page_area(m, endoff,
1973 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
1976 * Set valid, clear dirty bits. If validating the entire
1977 * page we can safely clear the pmap modify bit. We also
1978 * use this opportunity to clear the VPO_NOSYNC flag. If a process
1979 * takes a write fault on a MAP_NOSYNC memory area the flag will
1982 * We set valid bits inclusive of any overlap, but we can only
1983 * clear dirty bits for DEV_BSIZE chunks that are fully within
1986 pagebits = vm_page_bits(base, size);
1987 m->valid |= pagebits;
1989 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
1990 frag = DEV_BSIZE - frag;
1996 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
1998 m->dirty &= ~pagebits;
1999 if (base == 0 && size == PAGE_SIZE) {
2000 pmap_clear_modify(m);
2001 m->oflags &= ~VPO_NOSYNC;
2006 vm_page_clear_dirty(vm_page_t m, int base, int size)
2009 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2010 m->dirty &= ~vm_page_bits(base, size);
2014 * vm_page_set_invalid:
2016 * Invalidates DEV_BSIZE'd chunks within a page. Both the
2017 * valid and dirty bits for the effected areas are cleared.
2022 vm_page_set_invalid(vm_page_t m, int base, int size)
2026 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2027 bits = vm_page_bits(base, size);
2028 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2029 if (m->valid == VM_PAGE_BITS_ALL && bits != 0)
2033 m->object->generation++;
2037 * vm_page_zero_invalid()
2039 * The kernel assumes that the invalid portions of a page contain
2040 * garbage, but such pages can be mapped into memory by user code.
2041 * When this occurs, we must zero out the non-valid portions of the
2042 * page so user code sees what it expects.
2044 * Pages are most often semi-valid when the end of a file is mapped
2045 * into memory and the file's size is not page aligned.
2048 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
2053 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2055 * Scan the valid bits looking for invalid sections that
2056 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
2057 * valid bit may be set ) have already been zerod by
2058 * vm_page_set_validclean().
2060 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
2061 if (i == (PAGE_SIZE / DEV_BSIZE) ||
2062 (m->valid & (1 << i))
2065 pmap_zero_page_area(m,
2066 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
2073 * setvalid is TRUE when we can safely set the zero'd areas
2074 * as being valid. We can do this if there are no cache consistancy
2075 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
2078 m->valid = VM_PAGE_BITS_ALL;
2084 * Is (partial) page valid? Note that the case where size == 0
2085 * will return FALSE in the degenerate case where the page is
2086 * entirely invalid, and TRUE otherwise.
2091 vm_page_is_valid(vm_page_t m, int base, int size)
2093 int bits = vm_page_bits(base, size);
2095 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2096 if (m->valid && ((m->valid & bits) == bits))
2103 * update dirty bits from pmap/mmu. May not block.
2106 vm_page_test_dirty(vm_page_t m)
2108 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
2113 int so_zerocp_fullpage = 0;
2116 * Replace the given page with a copy. The copied page assumes
2117 * the portion of the given page's "wire_count" that is not the
2118 * responsibility of this copy-on-write mechanism.
2120 * The object containing the given page must have a non-zero
2121 * paging-in-progress count and be locked.
2124 vm_page_cowfault(vm_page_t m)
2131 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2132 KASSERT(object->paging_in_progress != 0,
2133 ("vm_page_cowfault: object %p's paging-in-progress count is zero.",
2140 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY);
2142 vm_page_insert(m, object, pindex);
2143 vm_page_unlock_queues();
2144 VM_OBJECT_UNLOCK(object);
2146 VM_OBJECT_LOCK(object);
2147 if (m == vm_page_lookup(object, pindex)) {
2148 vm_page_lock_queues();
2152 * Page disappeared during the wait.
2154 vm_page_lock_queues();
2161 * check to see if we raced with an xmit complete when
2162 * waiting to allocate a page. If so, put things back
2166 vm_page_insert(m, object, pindex);
2167 } else { /* clear COW & copy page */
2168 if (!so_zerocp_fullpage)
2169 pmap_copy_page(m, mnew);
2170 mnew->valid = VM_PAGE_BITS_ALL;
2171 vm_page_dirty(mnew);
2172 mnew->wire_count = m->wire_count - m->cow;
2173 m->wire_count = m->cow;
2178 vm_page_cowclear(vm_page_t m)
2181 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2185 * let vm_fault add back write permission lazily
2189 * sf_buf_free() will free the page, so we needn't do it here
2194 vm_page_cowsetup(vm_page_t m)
2197 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2198 if (m->cow == USHRT_MAX - 1)
2201 pmap_remove_write(m);
2205 #include "opt_ddb.h"
2207 #include <sys/kernel.h>
2209 #include <ddb/ddb.h>
2211 DB_SHOW_COMMAND(page, vm_page_print_page_info)
2213 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count);
2214 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count);
2215 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count);
2216 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count);
2217 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count);
2218 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved);
2219 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min);
2220 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target);
2221 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min);
2222 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target);
2225 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
2228 db_printf("PQ_FREE:");
2229 db_printf(" %d", cnt.v_free_count);
2232 db_printf("PQ_CACHE:");
2233 db_printf(" %d", cnt.v_cache_count);
2236 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
2237 *vm_page_queues[PQ_ACTIVE].cnt,
2238 *vm_page_queues[PQ_INACTIVE].cnt);