2 * Copyright (c) 1991 Regents of the University of California.
5 * This code is derived from software contributed to Berkeley by
6 * The Mach Operating System project at Carnegie-Mellon University.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 3. All advertising materials mentioning features or use of this software
17 * must display the following acknowledgement:
18 * This product includes software developed by the University of
19 * California, Berkeley and its contributors.
20 * 4. Neither the name of the University nor the names of its contributors
21 * may be used to endorse or promote products derived from this software
22 * without specific prior written permission.
24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91
37 * $FreeBSD: src/sys/vm/vm_page.c,v 1.147.2.18 2002/03/10 05:03:19 alc Exp $
38 * $DragonFly: src/sys/vm/vm_page.c,v 1.15 2003/11/03 17:11:23 dillon Exp $
42 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
43 * All rights reserved.
45 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
47 * Permission to use, copy, modify and distribute this software and
48 * its documentation is hereby granted, provided that both the copyright
49 * notice and this permission notice appear in all copies of the
50 * software, derivative works or modified versions, and any portions
51 * thereof, and that both notices appear in supporting documentation.
53 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
54 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
55 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
57 * Carnegie Mellon requests users of this software to return to
59 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
60 * School of Computer Science
61 * Carnegie Mellon University
62 * Pittsburgh PA 15213-3890
64 * any improvements or extensions that they make and grant Carnegie the
65 * rights to redistribute these changes.
69 * Resident memory management module.
72 #include <sys/param.h>
73 #include <sys/systm.h>
74 #include <sys/malloc.h>
76 #include <sys/vmmeter.h>
77 #include <sys/vnode.h>
80 #include <vm/vm_param.h>
82 #include <vm/vm_kern.h>
84 #include <vm/vm_map.h>
85 #include <vm/vm_object.h>
86 #include <vm/vm_page.h>
87 #include <vm/vm_pageout.h>
88 #include <vm/vm_pager.h>
89 #include <vm/vm_extern.h>
90 #include <vm/vm_page2.h>
92 static void vm_page_queue_init (void);
93 static vm_page_t vm_page_select_cache (vm_object_t, vm_pindex_t);
96 * Associated with page of user-allocatable memory is a
100 static struct vm_page **vm_page_buckets; /* Array of buckets */
101 static int vm_page_bucket_count; /* How big is array? */
102 static int vm_page_hash_mask; /* Mask for hash function */
103 static volatile int vm_page_bucket_generation;
105 struct vpgqueues vm_page_queues[PQ_COUNT];
108 vm_page_queue_init(void) {
111 for(i=0;i<PQ_L2_SIZE;i++) {
112 vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count;
114 vm_page_queues[PQ_INACTIVE].cnt = &vmstats.v_inactive_count;
116 vm_page_queues[PQ_ACTIVE].cnt = &vmstats.v_active_count;
117 vm_page_queues[PQ_HOLD].cnt = &vmstats.v_active_count;
118 for(i=0;i<PQ_L2_SIZE;i++) {
119 vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count;
121 for(i=0;i<PQ_COUNT;i++) {
122 TAILQ_INIT(&vm_page_queues[i].pl);
126 vm_page_t vm_page_array = 0;
127 int vm_page_array_size = 0;
129 int vm_page_zero_count = 0;
131 static __inline int vm_page_hash (vm_object_t object, vm_pindex_t pindex);
132 static void vm_page_free_wakeup (void);
137 * Sets the page size, perhaps based upon the memory
138 * size. Must be called before any use of page-size
139 * dependent functions.
142 vm_set_page_size(void)
144 if (vmstats.v_page_size == 0)
145 vmstats.v_page_size = PAGE_SIZE;
146 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0)
147 panic("vm_set_page_size: page size not a power of two");
153 * Add a new page to the freelist for use by the system. New pages
154 * are added to both the head and tail of the associated free page
155 * queue in a bottom-up fashion, so both zero'd and non-zero'd page
156 * requests pull 'recent' adds (higher physical addresses) first.
158 * Must be called at splhigh().
161 vm_add_new_page(vm_paddr_t pa)
164 struct vpgqueues *vpq;
166 ++vmstats.v_page_count;
167 ++vmstats.v_free_count;
168 m = PHYS_TO_VM_PAGE(pa);
171 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK;
172 m->queue = m->pc + PQ_FREE;
173 vpq = &vm_page_queues[m->queue];
175 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
177 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
178 vpq->flipflop = 1 - vpq->flipflop;
179 vm_page_queues[m->queue].lcnt++;
186 * Initializes the resident memory module.
188 * Allocates memory for the page cells, and
189 * for the object/offset-to-page hash table headers.
190 * Each page cell is initialized and placed on the free list.
194 vm_page_startup(vm_offset_t starta, vm_offset_t enda, vm_offset_t vaddr)
197 struct vm_page **bucket;
199 vm_paddr_t page_range;
206 /* the biggest memory array is the second group of pages */
208 vm_paddr_t biggestone, biggestsize;
216 vaddr = round_page(vaddr);
218 for (i = 0; phys_avail[i + 1]; i += 2) {
219 phys_avail[i] = round_page(phys_avail[i]);
220 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
223 for (i = 0; phys_avail[i + 1]; i += 2) {
224 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
226 if (size > biggestsize) {
234 end = phys_avail[biggestone+1];
237 * Initialize the queue headers for the free queue, the active queue
238 * and the inactive queue.
241 vm_page_queue_init();
244 * Allocate (and initialize) the hash table buckets.
246 * The number of buckets MUST BE a power of 2, and the actual value is
247 * the next power of 2 greater than the number of physical pages in
250 * We make the hash table approximately 2x the number of pages to
251 * reduce the chain length. This is about the same size using the
252 * singly-linked list as the 1x hash table we were using before
253 * using TAILQ but the chain length will be smaller.
255 * Note: This computation can be tweaked if desired.
257 vm_page_buckets = (struct vm_page **)vaddr;
258 bucket = vm_page_buckets;
259 if (vm_page_bucket_count == 0) {
260 vm_page_bucket_count = 1;
261 while (vm_page_bucket_count < atop(total))
262 vm_page_bucket_count <<= 1;
264 vm_page_bucket_count <<= 1;
265 vm_page_hash_mask = vm_page_bucket_count - 1;
268 * Validate these addresses.
270 new_end = end - vm_page_bucket_count * sizeof(struct vm_page *);
271 new_end = trunc_page(new_end);
272 mapped = round_page(vaddr);
273 vaddr = pmap_map(mapped, new_end, end,
274 VM_PROT_READ | VM_PROT_WRITE);
275 vaddr = round_page(vaddr);
276 bzero((caddr_t) mapped, vaddr - mapped);
278 for (i = 0; i < vm_page_bucket_count; i++) {
284 * Compute the number of pages of memory that will be available for
285 * use (taking into account the overhead of a page structure per
289 first_page = phys_avail[0] / PAGE_SIZE;
291 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
292 npages = (total - (page_range * sizeof(struct vm_page)) -
293 (end - new_end)) / PAGE_SIZE;
297 * Initialize the mem entry structures now, and put them in the free
300 vm_page_array = (vm_page_t) vaddr;
304 * Validate these addresses.
307 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
308 mapped = pmap_map(mapped, new_end, end,
309 VM_PROT_READ | VM_PROT_WRITE);
312 * Clear all of the page structures
314 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
315 vm_page_array_size = page_range;
318 * Construct the free queue(s) in ascending order (by physical
319 * address) so that the first 16MB of physical memory is allocated
320 * last rather than first. On large-memory machines, this avoids
321 * the exhaustion of low physical memory before isa_dmainit has run.
323 vmstats.v_page_count = 0;
324 vmstats.v_free_count = 0;
325 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
330 last_pa = phys_avail[i + 1];
331 while (pa < last_pa && npages-- > 0) {
342 * Distributes the object/offset key pair among hash buckets.
344 * NOTE: This macro depends on vm_page_bucket_count being a power of 2.
345 * This routine may not block.
347 * We try to randomize the hash based on the object to spread the pages
348 * out in the hash table without it costing us too much.
351 vm_page_hash(vm_object_t object, vm_pindex_t pindex)
353 int i = ((uintptr_t)object + pindex) ^ object->hash_rand;
355 return(i & vm_page_hash_mask);
359 vm_page_unhold(vm_page_t mem)
362 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
363 if (mem->hold_count == 0 && mem->queue == PQ_HOLD)
364 vm_page_free_toq(mem);
368 * vm_page_insert: [ internal use only ]
370 * Inserts the given mem entry into the object and object list.
372 * The pagetables are not updated but will presumably fault the page
373 * in if necessary, or if a kernel page the caller will at some point
374 * enter the page into the kernel's pmap. We are not allowed to block
375 * here so we *can't* do this anyway.
377 * The object and page must be locked, and must be splhigh.
378 * This routine may not block.
382 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
384 struct vm_page **bucket;
386 if (m->object != NULL)
387 panic("vm_page_insert: already inserted");
390 * Record the object/offset pair in this page
397 * Insert it into the object_object/offset hash table
400 bucket = &vm_page_buckets[vm_page_hash(object, pindex)];
403 vm_page_bucket_generation++;
406 * Now link into the object's list of backed pages.
409 TAILQ_INSERT_TAIL(&object->memq, m, listq);
410 object->generation++;
413 * show that the object has one more resident page.
416 object->resident_page_count++;
419 * Since we are inserting a new and possibly dirty page,
420 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
422 if (m->flags & PG_WRITEABLE)
423 vm_object_set_writeable_dirty(object);
428 * NOTE: used by device pager as well -wfj
430 * Removes the given mem entry from the object/offset-page
431 * table and the object page list, but do not invalidate/terminate
434 * The object and page must be locked, and at splhigh.
435 * The underlying pmap entry (if any) is NOT removed here.
436 * This routine may not block.
440 vm_page_remove(vm_page_t m)
444 if (m->object == NULL)
447 if ((m->flags & PG_BUSY) == 0) {
448 panic("vm_page_remove: page not busy");
452 * Basically destroy the page.
460 * Remove from the object_object/offset hash table. The object
461 * must be on the hash queue, we will panic if it isn't
463 * Note: we must NULL-out m->hnext to prevent loops in detached
464 * buffers with vm_page_lookup().
468 struct vm_page **bucket;
470 bucket = &vm_page_buckets[vm_page_hash(m->object, m->pindex)];
471 while (*bucket != m) {
473 panic("vm_page_remove(): page not found in hash");
474 bucket = &(*bucket)->hnext;
478 vm_page_bucket_generation++;
482 * Now remove from the object's list of backed pages.
485 TAILQ_REMOVE(&object->memq, m, listq);
488 * And show that the object has one fewer resident page.
491 object->resident_page_count--;
492 object->generation++;
500 * Returns the page associated with the object/offset
501 * pair specified; if none is found, NULL is returned.
503 * NOTE: the code below does not lock. It will operate properly if
504 * an interrupt makes a change, but the generation algorithm will not
505 * operate properly in an SMP environment where both cpu's are able to run
506 * kernel code simultaneously.
508 * The object must be locked. No side effects.
509 * This routine may not block.
510 * This is a critical path routine
514 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
517 struct vm_page **bucket;
521 * Search the hash table for this object/offset pair
525 generation = vm_page_bucket_generation;
526 bucket = &vm_page_buckets[vm_page_hash(object, pindex)];
527 for (m = *bucket; m != NULL; m = m->hnext) {
528 if ((m->object == object) && (m->pindex == pindex)) {
529 if (vm_page_bucket_generation != generation)
534 if (vm_page_bucket_generation != generation)
542 * Move the given memory entry from its
543 * current object to the specified target object/offset.
545 * The object must be locked.
546 * This routine may not block.
548 * Note: this routine will raise itself to splvm(), the caller need not.
550 * Note: swap associated with the page must be invalidated by the move. We
551 * have to do this for several reasons: (1) we aren't freeing the
552 * page, (2) we are dirtying the page, (3) the VM system is probably
553 * moving the page from object A to B, and will then later move
554 * the backing store from A to B and we can't have a conflict.
556 * Note: we *always* dirty the page. It is necessary both for the
557 * fact that we moved it, and because we may be invalidating
558 * swap. If the page is on the cache, we have to deactivate it
559 * or vm_page_dirty() will panic. Dirty pages are not allowed
564 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
570 vm_page_insert(m, new_object, new_pindex);
571 if (m->queue - m->pc == PQ_CACHE)
572 vm_page_deactivate(m);
578 * vm_page_unqueue_nowakeup:
580 * vm_page_unqueue() without any wakeup
582 * This routine must be called at splhigh().
583 * This routine may not block.
587 vm_page_unqueue_nowakeup(vm_page_t m)
589 int queue = m->queue;
590 struct vpgqueues *pq;
591 if (queue != PQ_NONE) {
592 pq = &vm_page_queues[queue];
594 TAILQ_REMOVE(&pq->pl, m, pageq);
603 * Remove a page from its queue.
605 * This routine must be called at splhigh().
606 * This routine may not block.
610 vm_page_unqueue(vm_page_t m)
612 int queue = m->queue;
613 struct vpgqueues *pq;
614 if (queue != PQ_NONE) {
616 pq = &vm_page_queues[queue];
617 TAILQ_REMOVE(&pq->pl, m, pageq);
620 if ((queue - m->pc) == PQ_CACHE) {
621 if (vm_paging_needed())
632 * Find a page on the specified queue with color optimization.
634 * The page coloring optimization attempts to locate a page
635 * that does not overload other nearby pages in the object in
636 * the cpu's L1 or L2 caches. We need this optimization because
637 * cpu caches tend to be physical caches, while object spaces tend
640 * This routine must be called at splvm().
641 * This routine may not block.
643 * This routine may only be called from the vm_page_list_find() macro
647 _vm_page_list_find(int basequeue, int index)
651 struct vpgqueues *pq;
653 pq = &vm_page_queues[basequeue];
656 * Note that for the first loop, index+i and index-i wind up at the
657 * same place. Even though this is not totally optimal, we've already
658 * blown it by missing the cache case so we do not care.
661 for(i = PQ_L2_SIZE / 2; i > 0; --i) {
662 if ((m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl)) != NULL)
665 if ((m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl)) != NULL)
674 * vm_page_select_cache:
676 * Find a page on the cache queue with color optimization. As pages
677 * might be found, but not applicable, they are deactivated. This
678 * keeps us from using potentially busy cached pages.
680 * This routine must be called at splvm().
681 * This routine may not block.
684 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex)
689 m = vm_page_list_find(
691 (pindex + object->pg_color) & PQ_L2_MASK,
694 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
695 m->hold_count || m->wire_count)) {
696 vm_page_deactivate(m);
704 * vm_page_select_free:
706 * Find a free or zero page, with specified preference. We attempt to
707 * inline the nominal case and fall back to _vm_page_select_free()
710 * This routine must be called at splvm().
711 * This routine may not block.
714 static __inline vm_page_t
715 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero)
719 m = vm_page_list_find(
721 (pindex + object->pg_color) & PQ_L2_MASK,
730 * Allocate and return a memory cell associated
731 * with this VM object/offset pair.
734 * VM_ALLOC_NORMAL normal process request
735 * VM_ALLOC_SYSTEM system *really* needs a page
736 * VM_ALLOC_INTERRUPT interrupt time request
737 * VM_ALLOC_ZERO zero page
739 * Object must be locked.
740 * This routine may not block.
742 * Additional special handling is required when called from an
743 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with
744 * the page cache in this case.
748 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
753 KASSERT(!vm_page_lookup(object, pindex),
754 ("vm_page_alloc: page already allocated"));
757 * The pager is allowed to eat deeper into the free page list.
760 if ((curthread == pagethread) && (page_req != VM_ALLOC_INTERRUPT)) {
761 page_req = VM_ALLOC_SYSTEM;
767 if (vmstats.v_free_count > vmstats.v_free_reserved) {
769 * Allocate from the free queue if there are plenty of pages
772 if (page_req == VM_ALLOC_ZERO)
773 m = vm_page_select_free(object, pindex, TRUE);
775 m = vm_page_select_free(object, pindex, FALSE);
777 (page_req == VM_ALLOC_SYSTEM &&
778 vmstats.v_cache_count == 0 &&
779 vmstats.v_free_count > vmstats.v_interrupt_free_min) ||
780 (page_req == VM_ALLOC_INTERRUPT && vmstats.v_free_count > 0)
783 * Interrupt or system, dig deeper into the free list.
785 m = vm_page_select_free(object, pindex, FALSE);
786 } else if (page_req != VM_ALLOC_INTERRUPT) {
788 * Allocatable from cache (non-interrupt only). On success,
789 * we must free the page and try again, thus ensuring that
790 * vmstats.v_*_free_min counters are replenished.
792 m = vm_page_select_cache(object, pindex);
795 #if defined(DIAGNOSTIC)
796 if (vmstats.v_cache_count > 0)
797 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count);
799 vm_pageout_deficit++;
803 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m));
805 vm_page_protect(m, VM_PROT_NONE);
810 * Not allocatable from cache from interrupt, give up.
813 vm_pageout_deficit++;
819 * At this point we had better have found a good page.
824 ("vm_page_alloc(): missing page on free queue\n")
828 * Remove from free queue
831 vm_page_unqueue_nowakeup(m);
834 * Initialize structure. Only the PG_ZERO flag is inherited.
837 if (m->flags & PG_ZERO) {
838 vm_page_zero_count--;
839 m->flags = PG_ZERO | PG_BUSY;
848 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m));
851 * vm_page_insert() is safe prior to the splx(). Note also that
852 * inserting a page here does not insert it into the pmap (which
853 * could cause us to block allocating memory). We cannot block
857 vm_page_insert(m, object, pindex);
860 * Don't wakeup too often - wakeup the pageout daemon when
861 * we would be nearly out of memory.
863 if (vm_paging_needed())
872 * vm_wait: (also see VM_WAIT macro)
874 * Block until free pages are available for allocation
875 * - Called in various places before memory allocations.
884 if (curthread == pagethread) {
885 vm_pageout_pages_needed = 1;
886 tsleep(&vm_pageout_pages_needed, 0, "VMWait", 0);
888 if (!vm_pages_needed) {
890 wakeup(&vm_pages_needed);
892 tsleep(&vmstats.v_free_count, 0, "vmwait", 0);
898 * vm_waitpfault: (also see VM_WAITPFAULT macro)
900 * Block until free pages are available for allocation
901 * - Called only in vm_fault so that processes page faulting
902 * can be easily tracked.
903 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
904 * processes will be able to grab memory first. Do not change
905 * this balance without careful testing first.
914 if (!vm_pages_needed) {
916 wakeup(&vm_pages_needed);
918 tsleep(&vmstats.v_free_count, 0, "pfault", 0);
925 * Put the specified page on the active list (if appropriate).
926 * Ensure that act_count is at least ACT_INIT but do not otherwise
929 * The page queues must be locked.
930 * This routine may not block.
933 vm_page_activate(vm_page_t m)
938 if (m->queue != PQ_ACTIVE) {
939 if ((m->queue - m->pc) == PQ_CACHE)
940 mycpu->gd_cnt.v_reactivated++;
944 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
945 m->queue = PQ_ACTIVE;
946 vm_page_queues[PQ_ACTIVE].lcnt++;
947 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
948 if (m->act_count < ACT_INIT)
949 m->act_count = ACT_INIT;
950 vmstats.v_active_count++;
953 if (m->act_count < ACT_INIT)
954 m->act_count = ACT_INIT;
961 * vm_page_free_wakeup:
963 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
964 * routine is called when a page has been added to the cache or free
967 * This routine may not block.
968 * This routine must be called at splvm()
971 vm_page_free_wakeup(void)
974 * if pageout daemon needs pages, then tell it that there are
977 if (vm_pageout_pages_needed &&
978 vmstats.v_cache_count + vmstats.v_free_count >= vmstats.v_pageout_free_min) {
979 wakeup(&vm_pageout_pages_needed);
980 vm_pageout_pages_needed = 0;
983 * wakeup processes that are waiting on memory if we hit a
984 * high water mark. And wakeup scheduler process if we have
985 * lots of memory. this process will swapin processes.
987 if (vm_pages_needed && !vm_page_count_min()) {
989 wakeup(&vmstats.v_free_count);
996 * Returns the given page to the PQ_FREE list,
997 * disassociating it with any VM object.
999 * Object and page must be locked prior to entry.
1000 * This routine may not block.
1004 vm_page_free_toq(vm_page_t m)
1007 struct vpgqueues *pq;
1008 vm_object_t object = m->object;
1012 mycpu->gd_cnt.v_tfree++;
1014 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
1016 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
1017 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
1019 if ((m->queue - m->pc) == PQ_FREE)
1020 panic("vm_page_free: freeing free page");
1022 panic("vm_page_free: freeing busy page");
1026 * unqueue, then remove page. Note that we cannot destroy
1027 * the page here because we do not want to call the pager's
1028 * callback routine until after we've put the page on the
1029 * appropriate free queue.
1032 vm_page_unqueue_nowakeup(m);
1036 * If fictitious remove object association and
1037 * return, otherwise delay object association removal.
1040 if ((m->flags & PG_FICTITIOUS) != 0) {
1048 if (m->wire_count != 0) {
1049 if (m->wire_count > 1) {
1050 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1051 m->wire_count, (long)m->pindex);
1053 panic("vm_page_free: freeing wired page\n");
1057 * If we've exhausted the object's resident pages we want to free
1062 (object->type == OBJT_VNODE) &&
1063 ((object->flags & OBJ_DEAD) == 0)
1065 struct vnode *vp = (struct vnode *)object->handle;
1067 if (vp && VSHOULDFREE(vp))
1072 * Clear the UNMANAGED flag when freeing an unmanaged page.
1075 if (m->flags & PG_UNMANAGED) {
1076 m->flags &= ~PG_UNMANAGED;
1079 pmap_page_is_free(m);
1083 if (m->hold_count != 0) {
1084 m->flags &= ~PG_ZERO;
1087 m->queue = PQ_FREE + m->pc;
1088 pq = &vm_page_queues[m->queue];
1093 * Put zero'd pages on the end ( where we look for zero'd pages
1094 * first ) and non-zerod pages at the head.
1097 if (m->flags & PG_ZERO) {
1098 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
1099 ++vm_page_zero_count;
1101 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1104 vm_page_free_wakeup();
1112 * Prevent PV management from being done on the page. The page is
1113 * removed from the paging queues as if it were wired, and as a
1114 * consequence of no longer being managed the pageout daemon will not
1115 * touch it (since there is no way to locate the pte mappings for the
1116 * page). madvise() calls that mess with the pmap will also no longer
1117 * operate on the page.
1119 * Beyond that the page is still reasonably 'normal'. Freeing the page
1120 * will clear the flag.
1122 * This routine is used by OBJT_PHYS objects - objects using unswappable
1123 * physical memory as backing store rather then swap-backed memory and
1124 * will eventually be extended to support 4MB unmanaged physical
1129 vm_page_unmanage(vm_page_t m)
1134 if ((m->flags & PG_UNMANAGED) == 0) {
1135 if (m->wire_count == 0)
1138 vm_page_flag_set(m, PG_UNMANAGED);
1145 * Mark this page as wired down by yet
1146 * another map, removing it from paging queues
1149 * The page queues must be locked.
1150 * This routine may not block.
1153 vm_page_wire(vm_page_t m)
1158 * Only bump the wire statistics if the page is not already wired,
1159 * and only unqueue the page if it is on some queue (if it is unmanaged
1160 * it is already off the queues).
1163 if (m->wire_count == 0) {
1164 if ((m->flags & PG_UNMANAGED) == 0)
1166 vmstats.v_wire_count++;
1169 KASSERT(m->wire_count != 0,
1170 ("vm_page_wire: wire_count overflow m=%p", m));
1173 vm_page_flag_set(m, PG_MAPPED);
1179 * Release one wiring of this page, potentially
1180 * enabling it to be paged again.
1182 * Many pages placed on the inactive queue should actually go
1183 * into the cache, but it is difficult to figure out which. What
1184 * we do instead, if the inactive target is well met, is to put
1185 * clean pages at the head of the inactive queue instead of the tail.
1186 * This will cause them to be moved to the cache more quickly and
1187 * if not actively re-referenced, freed more quickly. If we just
1188 * stick these pages at the end of the inactive queue, heavy filesystem
1189 * meta-data accesses can cause an unnecessary paging load on memory bound
1190 * processes. This optimization causes one-time-use metadata to be
1191 * reused more quickly.
1193 * BUT, if we are in a low-memory situation we have no choice but to
1194 * put clean pages on the cache queue.
1196 * A number of routines use vm_page_unwire() to guarantee that the page
1197 * will go into either the inactive or active queues, and will NEVER
1198 * be placed in the cache - for example, just after dirtying a page.
1199 * dirty pages in the cache are not allowed.
1201 * The page queues must be locked.
1202 * This routine may not block.
1205 vm_page_unwire(vm_page_t m, int activate)
1211 if (m->wire_count > 0) {
1213 if (m->wire_count == 0) {
1214 vmstats.v_wire_count--;
1215 if (m->flags & PG_UNMANAGED) {
1217 } else if (activate) {
1218 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1219 m->queue = PQ_ACTIVE;
1220 vm_page_queues[PQ_ACTIVE].lcnt++;
1221 vmstats.v_active_count++;
1223 vm_page_flag_clear(m, PG_WINATCFLS);
1224 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1225 m->queue = PQ_INACTIVE;
1226 vm_page_queues[PQ_INACTIVE].lcnt++;
1227 vmstats.v_inactive_count++;
1231 panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count);
1238 * Move the specified page to the inactive queue. If the page has
1239 * any associated swap, the swap is deallocated.
1241 * Normally athead is 0 resulting in LRU operation. athead is set
1242 * to 1 if we want this page to be 'as if it were placed in the cache',
1243 * except without unmapping it from the process address space.
1245 * This routine may not block.
1247 static __inline void
1248 _vm_page_deactivate(vm_page_t m, int athead)
1253 * Ignore if already inactive.
1255 if (m->queue == PQ_INACTIVE)
1259 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1260 if ((m->queue - m->pc) == PQ_CACHE)
1261 mycpu->gd_cnt.v_reactivated++;
1262 vm_page_flag_clear(m, PG_WINATCFLS);
1265 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1267 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1268 m->queue = PQ_INACTIVE;
1269 vm_page_queues[PQ_INACTIVE].lcnt++;
1270 vmstats.v_inactive_count++;
1276 vm_page_deactivate(vm_page_t m)
1278 _vm_page_deactivate(m, 0);
1282 * vm_page_try_to_cache:
1284 * Returns 0 on failure, 1 on success
1287 vm_page_try_to_cache(vm_page_t m)
1289 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1290 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1293 vm_page_test_dirty(m);
1301 * vm_page_try_to_free()
1303 * Attempt to free the page. If we cannot free it, we do nothing.
1304 * 1 is returned on success, 0 on failure.
1308 vm_page_try_to_free(vm_page_t m)
1310 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1311 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1314 vm_page_test_dirty(m);
1318 vm_page_protect(m, VM_PROT_NONE);
1327 * Put the specified page onto the page cache queue (if appropriate).
1329 * This routine may not block.
1332 vm_page_cache(vm_page_t m)
1336 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || m->wire_count) {
1337 printf("vm_page_cache: attempting to cache busy page\n");
1340 if ((m->queue - m->pc) == PQ_CACHE)
1344 * Remove all pmaps and indicate that the page is not
1345 * writeable or mapped.
1348 vm_page_protect(m, VM_PROT_NONE);
1349 if (m->dirty != 0) {
1350 panic("vm_page_cache: caching a dirty page, pindex: %ld",
1354 vm_page_unqueue_nowakeup(m);
1355 m->queue = PQ_CACHE + m->pc;
1356 vm_page_queues[m->queue].lcnt++;
1357 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq);
1358 vmstats.v_cache_count++;
1359 vm_page_free_wakeup();
1366 * Cache, deactivate, or do nothing as appropriate. This routine
1367 * is typically used by madvise() MADV_DONTNEED.
1369 * Generally speaking we want to move the page into the cache so
1370 * it gets reused quickly. However, this can result in a silly syndrome
1371 * due to the page recycling too quickly. Small objects will not be
1372 * fully cached. On the otherhand, if we move the page to the inactive
1373 * queue we wind up with a problem whereby very large objects
1374 * unnecessarily blow away our inactive and cache queues.
1376 * The solution is to move the pages based on a fixed weighting. We
1377 * either leave them alone, deactivate them, or move them to the cache,
1378 * where moving them to the cache has the highest weighting.
1379 * By forcing some pages into other queues we eventually force the
1380 * system to balance the queues, potentially recovering other unrelated
1381 * space from active. The idea is to not force this to happen too
1386 vm_page_dontneed(vm_page_t m)
1388 static int dnweight;
1395 * occassionally leave the page alone
1398 if ((dnw & 0x01F0) == 0 ||
1399 m->queue == PQ_INACTIVE ||
1400 m->queue - m->pc == PQ_CACHE
1402 if (m->act_count >= ACT_INIT)
1408 vm_page_test_dirty(m);
1410 if (m->dirty || (dnw & 0x0070) == 0) {
1412 * Deactivate the page 3 times out of 32.
1417 * Cache the page 28 times out of every 32. Note that
1418 * the page is deactivated instead of cached, but placed
1419 * at the head of the queue instead of the tail.
1423 _vm_page_deactivate(m, head);
1427 * Grab a page, waiting until we are waken up due to the page
1428 * changing state. We keep on waiting, if the page continues
1429 * to be in the object. If the page doesn't exist, allocate it.
1431 * This routine may block.
1434 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
1441 if ((m = vm_page_lookup(object, pindex)) != NULL) {
1442 if (m->busy || (m->flags & PG_BUSY)) {
1443 generation = object->generation;
1446 while ((object->generation == generation) &&
1447 (m->busy || (m->flags & PG_BUSY))) {
1448 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
1449 tsleep(m, 0, "pgrbwt", 0);
1450 if ((allocflags & VM_ALLOC_RETRY) == 0) {
1463 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1466 if ((allocflags & VM_ALLOC_RETRY) == 0)
1475 * Mapping function for valid bits or for dirty bits in
1476 * a page. May not block.
1478 * Inputs are required to range within a page.
1482 vm_page_bits(int base, int size)
1488 base + size <= PAGE_SIZE,
1489 ("vm_page_bits: illegal base/size %d/%d", base, size)
1492 if (size == 0) /* handle degenerate case */
1495 first_bit = base >> DEV_BSHIFT;
1496 last_bit = (base + size - 1) >> DEV_BSHIFT;
1498 return ((2 << last_bit) - (1 << first_bit));
1502 * vm_page_set_validclean:
1504 * Sets portions of a page valid and clean. The arguments are expected
1505 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1506 * of any partial chunks touched by the range. The invalid portion of
1507 * such chunks will be zero'd.
1509 * This routine may not block.
1511 * (base + size) must be less then or equal to PAGE_SIZE.
1514 vm_page_set_validclean(vm_page_t m, int base, int size)
1520 if (size == 0) /* handle degenerate case */
1524 * If the base is not DEV_BSIZE aligned and the valid
1525 * bit is clear, we have to zero out a portion of the
1529 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
1530 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
1532 pmap_zero_page_area(
1540 * If the ending offset is not DEV_BSIZE aligned and the
1541 * valid bit is clear, we have to zero out a portion of
1545 endoff = base + size;
1547 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
1548 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
1550 pmap_zero_page_area(
1553 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
1558 * Set valid, clear dirty bits. If validating the entire
1559 * page we can safely clear the pmap modify bit. We also
1560 * use this opportunity to clear the PG_NOSYNC flag. If a process
1561 * takes a write fault on a MAP_NOSYNC memory area the flag will
1564 * We set valid bits inclusive of any overlap, but we can only
1565 * clear dirty bits for DEV_BSIZE chunks that are fully within
1569 pagebits = vm_page_bits(base, size);
1570 m->valid |= pagebits;
1572 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
1573 frag = DEV_BSIZE - frag;
1579 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
1581 m->dirty &= ~pagebits;
1582 if (base == 0 && size == PAGE_SIZE) {
1583 pmap_clear_modify(m);
1584 vm_page_flag_clear(m, PG_NOSYNC);
1591 vm_page_set_dirty(vm_page_t m, int base, int size)
1593 m->dirty |= vm_page_bits(base, size);
1599 vm_page_clear_dirty(vm_page_t m, int base, int size)
1601 m->dirty &= ~vm_page_bits(base, size);
1605 * vm_page_set_invalid:
1607 * Invalidates DEV_BSIZE'd chunks within a page. Both the
1608 * valid and dirty bits for the effected areas are cleared.
1613 vm_page_set_invalid(vm_page_t m, int base, int size)
1617 bits = vm_page_bits(base, size);
1620 m->object->generation++;
1624 * vm_page_zero_invalid()
1626 * The kernel assumes that the invalid portions of a page contain
1627 * garbage, but such pages can be mapped into memory by user code.
1628 * When this occurs, we must zero out the non-valid portions of the
1629 * page so user code sees what it expects.
1631 * Pages are most often semi-valid when the end of a file is mapped
1632 * into memory and the file's size is not page aligned.
1636 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
1642 * Scan the valid bits looking for invalid sections that
1643 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
1644 * valid bit may be set ) have already been zerod by
1645 * vm_page_set_validclean().
1648 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
1649 if (i == (PAGE_SIZE / DEV_BSIZE) ||
1650 (m->valid & (1 << i))
1653 pmap_zero_page_area(
1656 (i - b) << DEV_BSHIFT
1664 * setvalid is TRUE when we can safely set the zero'd areas
1665 * as being valid. We can do this if there are no cache consistency
1666 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
1670 m->valid = VM_PAGE_BITS_ALL;
1676 * Is (partial) page valid? Note that the case where size == 0
1677 * will return FALSE in the degenerate case where the page is
1678 * entirely invalid, and TRUE otherwise.
1684 vm_page_is_valid(vm_page_t m, int base, int size)
1686 int bits = vm_page_bits(base, size);
1688 if (m->valid && ((m->valid & bits) == bits))
1695 * update dirty bits from pmap/mmu. May not block.
1699 vm_page_test_dirty(vm_page_t m)
1701 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
1706 #include "opt_ddb.h"
1708 #include <sys/kernel.h>
1710 #include <ddb/ddb.h>
1712 DB_SHOW_COMMAND(page, vm_page_print_page_info)
1714 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count);
1715 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count);
1716 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count);
1717 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count);
1718 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count);
1719 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved);
1720 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min);
1721 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target);
1722 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min);
1723 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target);
1726 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
1729 db_printf("PQ_FREE:");
1730 for(i=0;i<PQ_L2_SIZE;i++) {
1731 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
1735 db_printf("PQ_CACHE:");
1736 for(i=0;i<PQ_L2_SIZE;i++) {
1737 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
1741 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
1742 vm_page_queues[PQ_ACTIVE].lcnt,
1743 vm_page_queues[PQ_INACTIVE].lcnt);