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 1.14 2003/10/15 16:48:04 hmp 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_offset_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;
198 vm_size_t npages, page_range;
205 /* the biggest memory array is the second group of pages */
207 vm_offset_t biggestone, biggestsize;
215 vaddr = round_page(vaddr);
217 for (i = 0; phys_avail[i + 1]; i += 2) {
218 phys_avail[i] = round_page(phys_avail[i]);
219 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
222 for (i = 0; phys_avail[i + 1]; i += 2) {
223 int size = phys_avail[i + 1] - phys_avail[i];
225 if (size > biggestsize) {
233 end = phys_avail[biggestone+1];
236 * Initialize the queue headers for the free queue, the active queue
237 * and the inactive queue.
240 vm_page_queue_init();
243 * Allocate (and initialize) the hash table buckets.
245 * The number of buckets MUST BE a power of 2, and the actual value is
246 * the next power of 2 greater than the number of physical pages in
249 * We make the hash table approximately 2x the number of pages to
250 * reduce the chain length. This is about the same size using the
251 * singly-linked list as the 1x hash table we were using before
252 * using TAILQ but the chain length will be smaller.
254 * Note: This computation can be tweaked if desired.
256 vm_page_buckets = (struct vm_page **)vaddr;
257 bucket = vm_page_buckets;
258 if (vm_page_bucket_count == 0) {
259 vm_page_bucket_count = 1;
260 while (vm_page_bucket_count < atop(total))
261 vm_page_bucket_count <<= 1;
263 vm_page_bucket_count <<= 1;
264 vm_page_hash_mask = vm_page_bucket_count - 1;
267 * Validate these addresses.
269 new_end = end - vm_page_bucket_count * sizeof(struct vm_page *);
270 new_end = trunc_page(new_end);
271 mapped = round_page(vaddr);
272 vaddr = pmap_map(mapped, new_end, end,
273 VM_PROT_READ | VM_PROT_WRITE);
274 vaddr = round_page(vaddr);
275 bzero((caddr_t) mapped, vaddr - mapped);
277 for (i = 0; i < vm_page_bucket_count; i++) {
283 * Compute the number of pages of memory that will be available for
284 * use (taking into account the overhead of a page structure per
288 first_page = phys_avail[0] / PAGE_SIZE;
290 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
291 npages = (total - (page_range * sizeof(struct vm_page)) -
292 (end - new_end)) / PAGE_SIZE;
296 * Initialize the mem entry structures now, and put them in the free
299 vm_page_array = (vm_page_t) vaddr;
303 * Validate these addresses.
306 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
307 mapped = pmap_map(mapped, new_end, end,
308 VM_PROT_READ | VM_PROT_WRITE);
311 * Clear all of the page structures
313 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
314 vm_page_array_size = page_range;
317 * Construct the free queue(s) in ascending order (by physical
318 * address) so that the first 16MB of physical memory is allocated
319 * last rather than first. On large-memory machines, this avoids
320 * the exhaustion of low physical memory before isa_dmainit has run.
322 vmstats.v_page_count = 0;
323 vmstats.v_free_count = 0;
324 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
329 last_pa = phys_avail[i + 1];
330 while (pa < last_pa && npages-- > 0) {
341 * Distributes the object/offset key pair among hash buckets.
343 * NOTE: This macro depends on vm_page_bucket_count being a power of 2.
344 * This routine may not block.
346 * We try to randomize the hash based on the object to spread the pages
347 * out in the hash table without it costing us too much.
350 vm_page_hash(vm_object_t object, vm_pindex_t pindex)
352 int i = ((uintptr_t)object + pindex) ^ object->hash_rand;
354 return(i & vm_page_hash_mask);
358 vm_page_unhold(vm_page_t mem)
361 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
362 if (mem->hold_count == 0 && mem->queue == PQ_HOLD)
363 vm_page_free_toq(mem);
367 * vm_page_insert: [ internal use only ]
369 * Inserts the given mem entry into the object and object list.
371 * The pagetables are not updated but will presumably fault the page
372 * in if necessary, or if a kernel page the caller will at some point
373 * enter the page into the kernel's pmap. We are not allowed to block
374 * here so we *can't* do this anyway.
376 * The object and page must be locked, and must be splhigh.
377 * This routine may not block.
381 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
383 struct vm_page **bucket;
385 if (m->object != NULL)
386 panic("vm_page_insert: already inserted");
389 * Record the object/offset pair in this page
396 * Insert it into the object_object/offset hash table
399 bucket = &vm_page_buckets[vm_page_hash(object, pindex)];
402 vm_page_bucket_generation++;
405 * Now link into the object's list of backed pages.
408 TAILQ_INSERT_TAIL(&object->memq, m, listq);
409 object->generation++;
412 * show that the object has one more resident page.
415 object->resident_page_count++;
418 * Since we are inserting a new and possibly dirty page,
419 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
421 if (m->flags & PG_WRITEABLE)
422 vm_object_set_writeable_dirty(object);
427 * NOTE: used by device pager as well -wfj
429 * Removes the given mem entry from the object/offset-page
430 * table and the object page list, but do not invalidate/terminate
433 * The object and page must be locked, and at splhigh.
434 * The underlying pmap entry (if any) is NOT removed here.
435 * This routine may not block.
439 vm_page_remove(vm_page_t m)
443 if (m->object == NULL)
446 if ((m->flags & PG_BUSY) == 0) {
447 panic("vm_page_remove: page not busy");
451 * Basically destroy the page.
459 * Remove from the object_object/offset hash table. The object
460 * must be on the hash queue, we will panic if it isn't
462 * Note: we must NULL-out m->hnext to prevent loops in detached
463 * buffers with vm_page_lookup().
467 struct vm_page **bucket;
469 bucket = &vm_page_buckets[vm_page_hash(m->object, m->pindex)];
470 while (*bucket != m) {
472 panic("vm_page_remove(): page not found in hash");
473 bucket = &(*bucket)->hnext;
477 vm_page_bucket_generation++;
481 * Now remove from the object's list of backed pages.
484 TAILQ_REMOVE(&object->memq, m, listq);
487 * And show that the object has one fewer resident page.
490 object->resident_page_count--;
491 object->generation++;
499 * Returns the page associated with the object/offset
500 * pair specified; if none is found, NULL is returned.
502 * NOTE: the code below does not lock. It will operate properly if
503 * an interrupt makes a change, but the generation algorithm will not
504 * operate properly in an SMP environment where both cpu's are able to run
505 * kernel code simultaneously.
507 * The object must be locked. No side effects.
508 * This routine may not block.
509 * This is a critical path routine
513 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
516 struct vm_page **bucket;
520 * Search the hash table for this object/offset pair
524 generation = vm_page_bucket_generation;
525 bucket = &vm_page_buckets[vm_page_hash(object, pindex)];
526 for (m = *bucket; m != NULL; m = m->hnext) {
527 if ((m->object == object) && (m->pindex == pindex)) {
528 if (vm_page_bucket_generation != generation)
533 if (vm_page_bucket_generation != generation)
541 * Move the given memory entry from its
542 * current object to the specified target object/offset.
544 * The object must be locked.
545 * This routine may not block.
547 * Note: this routine will raise itself to splvm(), the caller need not.
549 * Note: swap associated with the page must be invalidated by the move. We
550 * have to do this for several reasons: (1) we aren't freeing the
551 * page, (2) we are dirtying the page, (3) the VM system is probably
552 * moving the page from object A to B, and will then later move
553 * the backing store from A to B and we can't have a conflict.
555 * Note: we *always* dirty the page. It is necessary both for the
556 * fact that we moved it, and because we may be invalidating
557 * swap. If the page is on the cache, we have to deactivate it
558 * or vm_page_dirty() will panic. Dirty pages are not allowed
563 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
569 vm_page_insert(m, new_object, new_pindex);
570 if (m->queue - m->pc == PQ_CACHE)
571 vm_page_deactivate(m);
577 * vm_page_unqueue_nowakeup:
579 * vm_page_unqueue() without any wakeup
581 * This routine must be called at splhigh().
582 * This routine may not block.
586 vm_page_unqueue_nowakeup(vm_page_t m)
588 int queue = m->queue;
589 struct vpgqueues *pq;
590 if (queue != PQ_NONE) {
591 pq = &vm_page_queues[queue];
593 TAILQ_REMOVE(&pq->pl, m, pageq);
602 * Remove a page from its queue.
604 * This routine must be called at splhigh().
605 * This routine may not block.
609 vm_page_unqueue(vm_page_t m)
611 int queue = m->queue;
612 struct vpgqueues *pq;
613 if (queue != PQ_NONE) {
615 pq = &vm_page_queues[queue];
616 TAILQ_REMOVE(&pq->pl, m, pageq);
619 if ((queue - m->pc) == PQ_CACHE) {
620 if (vm_paging_needed())
631 * Find a page on the specified queue with color optimization.
633 * The page coloring optimization attempts to locate a page
634 * that does not overload other nearby pages in the object in
635 * the cpu's L1 or L2 caches. We need this optimization because
636 * cpu caches tend to be physical caches, while object spaces tend
639 * This routine must be called at splvm().
640 * This routine may not block.
642 * This routine may only be called from the vm_page_list_find() macro
646 _vm_page_list_find(int basequeue, int index)
650 struct vpgqueues *pq;
652 pq = &vm_page_queues[basequeue];
655 * Note that for the first loop, index+i and index-i wind up at the
656 * same place. Even though this is not totally optimal, we've already
657 * blown it by missing the cache case so we do not care.
660 for(i = PQ_L2_SIZE / 2; i > 0; --i) {
661 if ((m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl)) != NULL)
664 if ((m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl)) != NULL)
673 * vm_page_select_cache:
675 * Find a page on the cache queue with color optimization. As pages
676 * might be found, but not applicable, they are deactivated. This
677 * keeps us from using potentially busy cached pages.
679 * This routine must be called at splvm().
680 * This routine may not block.
683 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex)
688 m = vm_page_list_find(
690 (pindex + object->pg_color) & PQ_L2_MASK,
693 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
694 m->hold_count || m->wire_count)) {
695 vm_page_deactivate(m);
703 * vm_page_select_free:
705 * Find a free or zero page, with specified preference. We attempt to
706 * inline the nominal case and fall back to _vm_page_select_free()
709 * This routine must be called at splvm().
710 * This routine may not block.
713 static __inline vm_page_t
714 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero)
718 m = vm_page_list_find(
720 (pindex + object->pg_color) & PQ_L2_MASK,
729 * Allocate and return a memory cell associated
730 * with this VM object/offset pair.
733 * VM_ALLOC_NORMAL normal process request
734 * VM_ALLOC_SYSTEM system *really* needs a page
735 * VM_ALLOC_INTERRUPT interrupt time request
736 * VM_ALLOC_ZERO zero page
738 * Object must be locked.
739 * This routine may not block.
741 * Additional special handling is required when called from an
742 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with
743 * the page cache in this case.
747 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
752 KASSERT(!vm_page_lookup(object, pindex),
753 ("vm_page_alloc: page already allocated"));
756 * The pager is allowed to eat deeper into the free page list.
759 if ((curthread == pagethread) && (page_req != VM_ALLOC_INTERRUPT)) {
760 page_req = VM_ALLOC_SYSTEM;
766 if (vmstats.v_free_count > vmstats.v_free_reserved) {
768 * Allocate from the free queue if there are plenty of pages
771 if (page_req == VM_ALLOC_ZERO)
772 m = vm_page_select_free(object, pindex, TRUE);
774 m = vm_page_select_free(object, pindex, FALSE);
776 (page_req == VM_ALLOC_SYSTEM &&
777 vmstats.v_cache_count == 0 &&
778 vmstats.v_free_count > vmstats.v_interrupt_free_min) ||
779 (page_req == VM_ALLOC_INTERRUPT && vmstats.v_free_count > 0)
782 * Interrupt or system, dig deeper into the free list.
784 m = vm_page_select_free(object, pindex, FALSE);
785 } else if (page_req != VM_ALLOC_INTERRUPT) {
787 * Allocatable from cache (non-interrupt only). On success,
788 * we must free the page and try again, thus ensuring that
789 * vmstats.v_*_free_min counters are replenished.
791 m = vm_page_select_cache(object, pindex);
794 #if defined(DIAGNOSTIC)
795 if (vmstats.v_cache_count > 0)
796 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count);
798 vm_pageout_deficit++;
802 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m));
804 vm_page_protect(m, VM_PROT_NONE);
809 * Not allocatable from cache from interrupt, give up.
812 vm_pageout_deficit++;
818 * At this point we had better have found a good page.
823 ("vm_page_alloc(): missing page on free queue\n")
827 * Remove from free queue
830 vm_page_unqueue_nowakeup(m);
833 * Initialize structure. Only the PG_ZERO flag is inherited.
836 if (m->flags & PG_ZERO) {
837 vm_page_zero_count--;
838 m->flags = PG_ZERO | PG_BUSY;
847 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m));
850 * vm_page_insert() is safe prior to the splx(). Note also that
851 * inserting a page here does not insert it into the pmap (which
852 * could cause us to block allocating memory). We cannot block
856 vm_page_insert(m, object, pindex);
859 * Don't wakeup too often - wakeup the pageout daemon when
860 * we would be nearly out of memory.
862 if (vm_paging_needed())
871 * vm_wait: (also see VM_WAIT macro)
873 * Block until free pages are available for allocation
874 * - Called in various places before memory allocations.
883 if (curthread == pagethread) {
884 vm_pageout_pages_needed = 1;
885 tsleep(&vm_pageout_pages_needed, 0, "VMWait", 0);
887 if (!vm_pages_needed) {
889 wakeup(&vm_pages_needed);
891 tsleep(&vmstats.v_free_count, 0, "vmwait", 0);
897 * vm_waitpfault: (also see VM_WAITPFAULT macro)
899 * Block until free pages are available for allocation
900 * - Called only in vm_fault so that processes page faulting
901 * can be easily tracked.
902 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
903 * processes will be able to grab memory first. Do not change
904 * this balance without careful testing first.
913 if (!vm_pages_needed) {
915 wakeup(&vm_pages_needed);
917 tsleep(&vmstats.v_free_count, 0, "pfault", 0);
924 * Put the specified page on the active list (if appropriate).
925 * Ensure that act_count is at least ACT_INIT but do not otherwise
928 * The page queues must be locked.
929 * This routine may not block.
932 vm_page_activate(vm_page_t m)
937 if (m->queue != PQ_ACTIVE) {
938 if ((m->queue - m->pc) == PQ_CACHE)
939 mycpu->gd_cnt.v_reactivated++;
943 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
944 m->queue = PQ_ACTIVE;
945 vm_page_queues[PQ_ACTIVE].lcnt++;
946 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
947 if (m->act_count < ACT_INIT)
948 m->act_count = ACT_INIT;
949 vmstats.v_active_count++;
952 if (m->act_count < ACT_INIT)
953 m->act_count = ACT_INIT;
960 * vm_page_free_wakeup:
962 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
963 * routine is called when a page has been added to the cache or free
966 * This routine may not block.
967 * This routine must be called at splvm()
970 vm_page_free_wakeup(void)
973 * if pageout daemon needs pages, then tell it that there are
976 if (vm_pageout_pages_needed &&
977 vmstats.v_cache_count + vmstats.v_free_count >= vmstats.v_pageout_free_min) {
978 wakeup(&vm_pageout_pages_needed);
979 vm_pageout_pages_needed = 0;
982 * wakeup processes that are waiting on memory if we hit a
983 * high water mark. And wakeup scheduler process if we have
984 * lots of memory. this process will swapin processes.
986 if (vm_pages_needed && !vm_page_count_min()) {
988 wakeup(&vmstats.v_free_count);
995 * Returns the given page to the PQ_FREE list,
996 * disassociating it with any VM object.
998 * Object and page must be locked prior to entry.
999 * This routine may not block.
1003 vm_page_free_toq(vm_page_t m)
1006 struct vpgqueues *pq;
1007 vm_object_t object = m->object;
1011 mycpu->gd_cnt.v_tfree++;
1013 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
1015 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
1016 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
1018 if ((m->queue - m->pc) == PQ_FREE)
1019 panic("vm_page_free: freeing free page");
1021 panic("vm_page_free: freeing busy page");
1025 * unqueue, then remove page. Note that we cannot destroy
1026 * the page here because we do not want to call the pager's
1027 * callback routine until after we've put the page on the
1028 * appropriate free queue.
1031 vm_page_unqueue_nowakeup(m);
1035 * If fictitious remove object association and
1036 * return, otherwise delay object association removal.
1039 if ((m->flags & PG_FICTITIOUS) != 0) {
1047 if (m->wire_count != 0) {
1048 if (m->wire_count > 1) {
1049 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1050 m->wire_count, (long)m->pindex);
1052 panic("vm_page_free: freeing wired page\n");
1056 * If we've exhausted the object's resident pages we want to free
1061 (object->type == OBJT_VNODE) &&
1062 ((object->flags & OBJ_DEAD) == 0)
1064 struct vnode *vp = (struct vnode *)object->handle;
1066 if (vp && VSHOULDFREE(vp))
1071 * Clear the UNMANAGED flag when freeing an unmanaged page.
1074 if (m->flags & PG_UNMANAGED) {
1075 m->flags &= ~PG_UNMANAGED;
1078 pmap_page_is_free(m);
1082 if (m->hold_count != 0) {
1083 m->flags &= ~PG_ZERO;
1086 m->queue = PQ_FREE + m->pc;
1087 pq = &vm_page_queues[m->queue];
1092 * Put zero'd pages on the end ( where we look for zero'd pages
1093 * first ) and non-zerod pages at the head.
1096 if (m->flags & PG_ZERO) {
1097 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
1098 ++vm_page_zero_count;
1100 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1103 vm_page_free_wakeup();
1111 * Prevent PV management from being done on the page. The page is
1112 * removed from the paging queues as if it were wired, and as a
1113 * consequence of no longer being managed the pageout daemon will not
1114 * touch it (since there is no way to locate the pte mappings for the
1115 * page). madvise() calls that mess with the pmap will also no longer
1116 * operate on the page.
1118 * Beyond that the page is still reasonably 'normal'. Freeing the page
1119 * will clear the flag.
1121 * This routine is used by OBJT_PHYS objects - objects using unswappable
1122 * physical memory as backing store rather then swap-backed memory and
1123 * will eventually be extended to support 4MB unmanaged physical
1128 vm_page_unmanage(vm_page_t m)
1133 if ((m->flags & PG_UNMANAGED) == 0) {
1134 if (m->wire_count == 0)
1137 vm_page_flag_set(m, PG_UNMANAGED);
1144 * Mark this page as wired down by yet
1145 * another map, removing it from paging queues
1148 * The page queues must be locked.
1149 * This routine may not block.
1152 vm_page_wire(vm_page_t m)
1157 * Only bump the wire statistics if the page is not already wired,
1158 * and only unqueue the page if it is on some queue (if it is unmanaged
1159 * it is already off the queues).
1162 if (m->wire_count == 0) {
1163 if ((m->flags & PG_UNMANAGED) == 0)
1165 vmstats.v_wire_count++;
1168 KASSERT(m->wire_count != 0,
1169 ("vm_page_wire: wire_count overflow m=%p", m));
1172 vm_page_flag_set(m, PG_MAPPED);
1178 * Release one wiring of this page, potentially
1179 * enabling it to be paged again.
1181 * Many pages placed on the inactive queue should actually go
1182 * into the cache, but it is difficult to figure out which. What
1183 * we do instead, if the inactive target is well met, is to put
1184 * clean pages at the head of the inactive queue instead of the tail.
1185 * This will cause them to be moved to the cache more quickly and
1186 * if not actively re-referenced, freed more quickly. If we just
1187 * stick these pages at the end of the inactive queue, heavy filesystem
1188 * meta-data accesses can cause an unnecessary paging load on memory bound
1189 * processes. This optimization causes one-time-use metadata to be
1190 * reused more quickly.
1192 * BUT, if we are in a low-memory situation we have no choice but to
1193 * put clean pages on the cache queue.
1195 * A number of routines use vm_page_unwire() to guarantee that the page
1196 * will go into either the inactive or active queues, and will NEVER
1197 * be placed in the cache - for example, just after dirtying a page.
1198 * dirty pages in the cache are not allowed.
1200 * The page queues must be locked.
1201 * This routine may not block.
1204 vm_page_unwire(vm_page_t m, int activate)
1210 if (m->wire_count > 0) {
1212 if (m->wire_count == 0) {
1213 vmstats.v_wire_count--;
1214 if (m->flags & PG_UNMANAGED) {
1216 } else if (activate) {
1217 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1218 m->queue = PQ_ACTIVE;
1219 vm_page_queues[PQ_ACTIVE].lcnt++;
1220 vmstats.v_active_count++;
1222 vm_page_flag_clear(m, PG_WINATCFLS);
1223 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1224 m->queue = PQ_INACTIVE;
1225 vm_page_queues[PQ_INACTIVE].lcnt++;
1226 vmstats.v_inactive_count++;
1230 panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count);
1237 * Move the specified page to the inactive queue. If the page has
1238 * any associated swap, the swap is deallocated.
1240 * Normally athead is 0 resulting in LRU operation. athead is set
1241 * to 1 if we want this page to be 'as if it were placed in the cache',
1242 * except without unmapping it from the process address space.
1244 * This routine may not block.
1246 static __inline void
1247 _vm_page_deactivate(vm_page_t m, int athead)
1252 * Ignore if already inactive.
1254 if (m->queue == PQ_INACTIVE)
1258 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1259 if ((m->queue - m->pc) == PQ_CACHE)
1260 mycpu->gd_cnt.v_reactivated++;
1261 vm_page_flag_clear(m, PG_WINATCFLS);
1264 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1266 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1267 m->queue = PQ_INACTIVE;
1268 vm_page_queues[PQ_INACTIVE].lcnt++;
1269 vmstats.v_inactive_count++;
1275 vm_page_deactivate(vm_page_t m)
1277 _vm_page_deactivate(m, 0);
1281 * vm_page_try_to_cache:
1283 * Returns 0 on failure, 1 on success
1286 vm_page_try_to_cache(vm_page_t m)
1288 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1289 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1292 vm_page_test_dirty(m);
1300 * vm_page_try_to_free()
1302 * Attempt to free the page. If we cannot free it, we do nothing.
1303 * 1 is returned on success, 0 on failure.
1307 vm_page_try_to_free(vm_page_t m)
1309 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1310 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1313 vm_page_test_dirty(m);
1317 vm_page_protect(m, VM_PROT_NONE);
1326 * Put the specified page onto the page cache queue (if appropriate).
1328 * This routine may not block.
1331 vm_page_cache(vm_page_t m)
1335 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || m->wire_count) {
1336 printf("vm_page_cache: attempting to cache busy page\n");
1339 if ((m->queue - m->pc) == PQ_CACHE)
1343 * Remove all pmaps and indicate that the page is not
1344 * writeable or mapped.
1347 vm_page_protect(m, VM_PROT_NONE);
1348 if (m->dirty != 0) {
1349 panic("vm_page_cache: caching a dirty page, pindex: %ld",
1353 vm_page_unqueue_nowakeup(m);
1354 m->queue = PQ_CACHE + m->pc;
1355 vm_page_queues[m->queue].lcnt++;
1356 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq);
1357 vmstats.v_cache_count++;
1358 vm_page_free_wakeup();
1365 * Cache, deactivate, or do nothing as appropriate. This routine
1366 * is typically used by madvise() MADV_DONTNEED.
1368 * Generally speaking we want to move the page into the cache so
1369 * it gets reused quickly. However, this can result in a silly syndrome
1370 * due to the page recycling too quickly. Small objects will not be
1371 * fully cached. On the otherhand, if we move the page to the inactive
1372 * queue we wind up with a problem whereby very large objects
1373 * unnecessarily blow away our inactive and cache queues.
1375 * The solution is to move the pages based on a fixed weighting. We
1376 * either leave them alone, deactivate them, or move them to the cache,
1377 * where moving them to the cache has the highest weighting.
1378 * By forcing some pages into other queues we eventually force the
1379 * system to balance the queues, potentially recovering other unrelated
1380 * space from active. The idea is to not force this to happen too
1385 vm_page_dontneed(vm_page_t m)
1387 static int dnweight;
1394 * occassionally leave the page alone
1397 if ((dnw & 0x01F0) == 0 ||
1398 m->queue == PQ_INACTIVE ||
1399 m->queue - m->pc == PQ_CACHE
1401 if (m->act_count >= ACT_INIT)
1407 vm_page_test_dirty(m);
1409 if (m->dirty || (dnw & 0x0070) == 0) {
1411 * Deactivate the page 3 times out of 32.
1416 * Cache the page 28 times out of every 32. Note that
1417 * the page is deactivated instead of cached, but placed
1418 * at the head of the queue instead of the tail.
1422 _vm_page_deactivate(m, head);
1426 * Grab a page, waiting until we are waken up due to the page
1427 * changing state. We keep on waiting, if the page continues
1428 * to be in the object. If the page doesn't exist, allocate it.
1430 * This routine may block.
1433 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
1440 if ((m = vm_page_lookup(object, pindex)) != NULL) {
1441 if (m->busy || (m->flags & PG_BUSY)) {
1442 generation = object->generation;
1445 while ((object->generation == generation) &&
1446 (m->busy || (m->flags & PG_BUSY))) {
1447 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
1448 tsleep(m, 0, "pgrbwt", 0);
1449 if ((allocflags & VM_ALLOC_RETRY) == 0) {
1462 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1465 if ((allocflags & VM_ALLOC_RETRY) == 0)
1474 * Mapping function for valid bits or for dirty bits in
1475 * a page. May not block.
1477 * Inputs are required to range within a page.
1481 vm_page_bits(int base, int size)
1487 base + size <= PAGE_SIZE,
1488 ("vm_page_bits: illegal base/size %d/%d", base, size)
1491 if (size == 0) /* handle degenerate case */
1494 first_bit = base >> DEV_BSHIFT;
1495 last_bit = (base + size - 1) >> DEV_BSHIFT;
1497 return ((2 << last_bit) - (1 << first_bit));
1501 * vm_page_set_validclean:
1503 * Sets portions of a page valid and clean. The arguments are expected
1504 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1505 * of any partial chunks touched by the range. The invalid portion of
1506 * such chunks will be zero'd.
1508 * This routine may not block.
1510 * (base + size) must be less then or equal to PAGE_SIZE.
1513 vm_page_set_validclean(vm_page_t m, int base, int size)
1519 if (size == 0) /* handle degenerate case */
1523 * If the base is not DEV_BSIZE aligned and the valid
1524 * bit is clear, we have to zero out a portion of the
1528 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
1529 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
1531 pmap_zero_page_area(
1539 * If the ending offset is not DEV_BSIZE aligned and the
1540 * valid bit is clear, we have to zero out a portion of
1544 endoff = base + size;
1546 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
1547 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
1549 pmap_zero_page_area(
1552 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
1557 * Set valid, clear dirty bits. If validating the entire
1558 * page we can safely clear the pmap modify bit. We also
1559 * use this opportunity to clear the PG_NOSYNC flag. If a process
1560 * takes a write fault on a MAP_NOSYNC memory area the flag will
1563 * We set valid bits inclusive of any overlap, but we can only
1564 * clear dirty bits for DEV_BSIZE chunks that are fully within
1568 pagebits = vm_page_bits(base, size);
1569 m->valid |= pagebits;
1571 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
1572 frag = DEV_BSIZE - frag;
1578 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
1580 m->dirty &= ~pagebits;
1581 if (base == 0 && size == PAGE_SIZE) {
1582 pmap_clear_modify(m);
1583 vm_page_flag_clear(m, PG_NOSYNC);
1590 vm_page_set_dirty(vm_page_t m, int base, int size)
1592 m->dirty |= vm_page_bits(base, size);
1598 vm_page_clear_dirty(vm_page_t m, int base, int size)
1600 m->dirty &= ~vm_page_bits(base, size);
1604 * vm_page_set_invalid:
1606 * Invalidates DEV_BSIZE'd chunks within a page. Both the
1607 * valid and dirty bits for the effected areas are cleared.
1612 vm_page_set_invalid(vm_page_t m, int base, int size)
1616 bits = vm_page_bits(base, size);
1619 m->object->generation++;
1623 * vm_page_zero_invalid()
1625 * The kernel assumes that the invalid portions of a page contain
1626 * garbage, but such pages can be mapped into memory by user code.
1627 * When this occurs, we must zero out the non-valid portions of the
1628 * page so user code sees what it expects.
1630 * Pages are most often semi-valid when the end of a file is mapped
1631 * into memory and the file's size is not page aligned.
1635 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
1641 * Scan the valid bits looking for invalid sections that
1642 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
1643 * valid bit may be set ) have already been zerod by
1644 * vm_page_set_validclean().
1647 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
1648 if (i == (PAGE_SIZE / DEV_BSIZE) ||
1649 (m->valid & (1 << i))
1652 pmap_zero_page_area(
1655 (i - b) << DEV_BSHIFT
1663 * setvalid is TRUE when we can safely set the zero'd areas
1664 * as being valid. We can do this if there are no cache consistency
1665 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
1669 m->valid = VM_PAGE_BITS_ALL;
1675 * Is (partial) page valid? Note that the case where size == 0
1676 * will return FALSE in the degenerate case where the page is
1677 * entirely invalid, and TRUE otherwise.
1683 vm_page_is_valid(vm_page_t m, int base, int size)
1685 int bits = vm_page_bits(base, size);
1687 if (m->valid && ((m->valid & bits) == bits))
1694 * update dirty bits from pmap/mmu. May not block.
1698 vm_page_test_dirty(vm_page_t m)
1700 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
1705 #include "opt_ddb.h"
1707 #include <sys/kernel.h>
1709 #include <ddb/ddb.h>
1711 DB_SHOW_COMMAND(page, vm_page_print_page_info)
1713 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count);
1714 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count);
1715 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count);
1716 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count);
1717 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count);
1718 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved);
1719 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min);
1720 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target);
1721 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min);
1722 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target);
1725 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
1728 db_printf("PQ_FREE:");
1729 for(i=0;i<PQ_L2_SIZE;i++) {
1730 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
1734 db_printf("PQ_CACHE:");
1735 for(i=0;i<PQ_L2_SIZE;i++) {
1736 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
1740 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
1741 vm_page_queues[PQ_ACTIVE].lcnt,
1742 vm_page_queues[PQ_INACTIVE].lcnt);