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 1.22 2004/05/20 21:40:50 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);
94 static vm_page_t _vm_page_list_find2(int basequeue, int index);
97 * Associated with page of user-allocatable memory is a
101 static struct vm_page **vm_page_buckets; /* Array of buckets */
102 static int vm_page_bucket_count; /* How big is array? */
103 static int vm_page_hash_mask; /* Mask for hash function */
104 static volatile int vm_page_bucket_generation;
106 struct vpgqueues vm_page_queues[PQ_COUNT];
109 vm_page_queue_init(void) {
112 for(i=0;i<PQ_L2_SIZE;i++) {
113 vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count;
115 vm_page_queues[PQ_INACTIVE].cnt = &vmstats.v_inactive_count;
117 vm_page_queues[PQ_ACTIVE].cnt = &vmstats.v_active_count;
118 vm_page_queues[PQ_HOLD].cnt = &vmstats.v_active_count;
119 for(i=0;i<PQ_L2_SIZE;i++) {
120 vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count;
122 for(i=0;i<PQ_COUNT;i++) {
123 TAILQ_INIT(&vm_page_queues[i].pl);
127 vm_page_t vm_page_array = 0;
128 int vm_page_array_size = 0;
130 int vm_page_zero_count = 0;
132 static __inline int vm_page_hash (vm_object_t object, vm_pindex_t pindex);
133 static void vm_page_free_wakeup (void);
138 * Sets the page size, perhaps based upon the memory
139 * size. Must be called before any use of page-size
140 * dependent functions.
143 vm_set_page_size(void)
145 if (vmstats.v_page_size == 0)
146 vmstats.v_page_size = PAGE_SIZE;
147 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0)
148 panic("vm_set_page_size: page size not a power of two");
154 * Add a new page to the freelist for use by the system. New pages
155 * are added to both the head and tail of the associated free page
156 * queue in a bottom-up fashion, so both zero'd and non-zero'd page
157 * requests pull 'recent' adds (higher physical addresses) first.
159 * Must be called at splhigh().
162 vm_add_new_page(vm_paddr_t pa)
165 struct vpgqueues *vpq;
167 ++vmstats.v_page_count;
168 ++vmstats.v_free_count;
169 m = PHYS_TO_VM_PAGE(pa);
172 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK;
173 m->queue = m->pc + PQ_FREE;
174 vpq = &vm_page_queues[m->queue];
176 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
178 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
179 vpq->flipflop = 1 - vpq->flipflop;
180 vm_page_queues[m->queue].lcnt++;
187 * Initializes the resident memory module.
189 * Allocates memory for the page cells, and
190 * for the object/offset-to-page hash table headers.
191 * Each page cell is initialized and placed on the free list.
195 vm_page_startup(vm_offset_t starta, vm_offset_t enda, vm_offset_t vaddr)
198 struct vm_page **bucket;
200 vm_paddr_t page_range;
207 /* the biggest memory array is the second group of pages */
209 vm_paddr_t biggestone, biggestsize;
217 vaddr = round_page(vaddr);
219 for (i = 0; phys_avail[i + 1]; i += 2) {
220 phys_avail[i] = round_page(phys_avail[i]);
221 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
224 for (i = 0; phys_avail[i + 1]; i += 2) {
225 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
227 if (size > biggestsize) {
235 end = phys_avail[biggestone+1];
238 * Initialize the queue headers for the free queue, the active queue
239 * and the inactive queue.
242 vm_page_queue_init();
245 * Allocate (and initialize) the hash table buckets.
247 * The number of buckets MUST BE a power of 2, and the actual value is
248 * the next power of 2 greater than the number of physical pages in
251 * We make the hash table approximately 2x the number of pages to
252 * reduce the chain length. This is about the same size using the
253 * singly-linked list as the 1x hash table we were using before
254 * using TAILQ but the chain length will be smaller.
256 * Note: This computation can be tweaked if desired.
258 vm_page_buckets = (struct vm_page **)vaddr;
259 bucket = vm_page_buckets;
260 if (vm_page_bucket_count == 0) {
261 vm_page_bucket_count = 1;
262 while (vm_page_bucket_count < atop(total))
263 vm_page_bucket_count <<= 1;
265 vm_page_bucket_count <<= 1;
266 vm_page_hash_mask = vm_page_bucket_count - 1;
269 * Validate these addresses.
271 new_end = end - vm_page_bucket_count * sizeof(struct vm_page *);
272 new_end = trunc_page(new_end);
273 mapped = round_page(vaddr);
274 vaddr = pmap_map(mapped, new_end, end,
275 VM_PROT_READ | VM_PROT_WRITE);
276 vaddr = round_page(vaddr);
277 bzero((caddr_t) mapped, vaddr - mapped);
279 for (i = 0; i < vm_page_bucket_count; i++) {
285 * Compute the number of pages of memory that will be available for
286 * use (taking into account the overhead of a page structure per
290 first_page = phys_avail[0] / PAGE_SIZE;
292 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
293 npages = (total - (page_range * sizeof(struct vm_page)) -
294 (end - new_end)) / PAGE_SIZE;
298 * Initialize the mem entry structures now, and put them in the free
301 vm_page_array = (vm_page_t) vaddr;
305 * Validate these addresses.
308 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
309 mapped = pmap_map(mapped, new_end, end,
310 VM_PROT_READ | VM_PROT_WRITE);
313 * Clear all of the page structures
315 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
316 vm_page_array_size = page_range;
319 * Construct the free queue(s) in ascending order (by physical
320 * address) so that the first 16MB of physical memory is allocated
321 * last rather than first. On large-memory machines, this avoids
322 * the exhaustion of low physical memory before isa_dmainit has run.
324 vmstats.v_page_count = 0;
325 vmstats.v_free_count = 0;
326 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
331 last_pa = phys_avail[i + 1];
332 while (pa < last_pa && npages-- > 0) {
343 * Distributes the object/offset key pair among hash buckets.
345 * NOTE: This macro depends on vm_page_bucket_count being a power of 2.
346 * This routine may not block.
348 * We try to randomize the hash based on the object to spread the pages
349 * out in the hash table without it costing us too much.
352 vm_page_hash(vm_object_t object, vm_pindex_t pindex)
354 int i = ((uintptr_t)object + pindex) ^ object->hash_rand;
356 return(i & vm_page_hash_mask);
360 vm_page_unhold(vm_page_t mem)
363 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
364 if (mem->hold_count == 0 && mem->queue == PQ_HOLD)
365 vm_page_free_toq(mem);
369 * vm_page_insert: [ internal use only ]
371 * Inserts the given mem entry into the object and object list.
373 * The pagetables are not updated but will presumably fault the page
374 * in if necessary, or if a kernel page the caller will at some point
375 * enter the page into the kernel's pmap. We are not allowed to block
376 * here so we *can't* do this anyway.
378 * This routine may not block.
379 * This routine must be called at splvm().
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
396 * Insert it into the object_object/offset hash table
398 bucket = &vm_page_buckets[vm_page_hash(object, pindex)];
401 vm_page_bucket_generation++;
404 * Now link into the object's list of backed pages.
407 TAILQ_INSERT_TAIL(&object->memq, m, listq);
408 object->generation++;
411 * show that the object has one more resident page.
414 object->resident_page_count++;
417 * Since we are inserting a new and possibly dirty page,
418 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
420 if (m->flags & PG_WRITEABLE)
421 vm_object_set_writeable_dirty(object);
426 * NOTE: used by device pager as well -wfj
428 * Removes the given mem entry from the object/offset-page
429 * table and the object page list, but do not invalidate/terminate
432 * This routine must be called at splvm()
433 * The underlying pmap entry (if any) is NOT removed here.
434 * This routine may not block.
437 vm_page_remove(vm_page_t m)
441 if (m->object == NULL)
444 if ((m->flags & PG_BUSY) == 0) {
445 panic("vm_page_remove: page not busy");
449 * Basically destroy the page.
457 * Remove from the object_object/offset hash table. The object
458 * must be on the hash queue, we will panic if it isn't
460 * Note: we must NULL-out m->hnext to prevent loops in detached
461 * buffers with vm_page_lookup().
464 struct vm_page **bucket;
466 bucket = &vm_page_buckets[vm_page_hash(m->object, m->pindex)];
467 while (*bucket != m) {
469 panic("vm_page_remove(): page not found in hash");
470 bucket = &(*bucket)->hnext;
474 vm_page_bucket_generation++;
478 * Now remove from the object's list of backed pages.
480 TAILQ_REMOVE(&object->memq, m, listq);
483 * And show that the object has one fewer resident page.
485 object->resident_page_count--;
486 object->generation++;
494 * Locate and return the page at (object, pindex), or NULL if the
495 * page could not be found.
497 * This routine will operate properly without spl protection, but
498 * the returned page could be in flux if it is busy. Because an
499 * interrupt can race a caller's busy check (unbusying and freeing the
500 * page we return before the caller is able to check the busy bit),
501 * the caller should generally call this routine at splvm().
503 * Callers may call this routine without spl protection if they know
504 * 'for sure' that the page will not be ripped out from under them
509 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
512 struct vm_page **bucket;
516 * Search the hash table for this object/offset pair
520 generation = vm_page_bucket_generation;
521 bucket = &vm_page_buckets[vm_page_hash(object, pindex)];
522 for (m = *bucket; m != NULL; m = m->hnext) {
523 if ((m->object == object) && (m->pindex == pindex)) {
524 if (vm_page_bucket_generation != generation)
529 if (vm_page_bucket_generation != generation)
537 * Move the given memory entry from its
538 * current object to the specified target object/offset.
540 * The object must be locked.
541 * This routine may not block.
543 * Note: this routine will raise itself to splvm(), the caller need not.
545 * Note: swap associated with the page must be invalidated by the move. We
546 * have to do this for several reasons: (1) we aren't freeing the
547 * page, (2) we are dirtying the page, (3) the VM system is probably
548 * moving the page from object A to B, and will then later move
549 * the backing store from A to B and we can't have a conflict.
551 * Note: we *always* dirty the page. It is necessary both for the
552 * fact that we moved it, and because we may be invalidating
553 * swap. If the page is on the cache, we have to deactivate it
554 * or vm_page_dirty() will panic. Dirty pages are not allowed
559 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
565 vm_page_insert(m, new_object, new_pindex);
566 if (m->queue - m->pc == PQ_CACHE)
567 vm_page_deactivate(m);
573 * vm_page_unqueue_nowakeup:
575 * vm_page_unqueue() without any wakeup
577 * This routine must be called at splhigh().
578 * This routine may not block.
582 vm_page_unqueue_nowakeup(vm_page_t m)
584 int queue = m->queue;
585 struct vpgqueues *pq;
586 if (queue != PQ_NONE) {
587 pq = &vm_page_queues[queue];
589 TAILQ_REMOVE(&pq->pl, m, pageq);
598 * Remove a page from its queue.
600 * This routine must be called at splhigh().
601 * This routine may not block.
605 vm_page_unqueue(vm_page_t m)
607 int queue = m->queue;
608 struct vpgqueues *pq;
609 if (queue != PQ_NONE) {
611 pq = &vm_page_queues[queue];
612 TAILQ_REMOVE(&pq->pl, m, pageq);
615 if ((queue - m->pc) == PQ_CACHE) {
616 if (vm_paging_needed())
627 * Find a page on the specified queue with color optimization.
629 * The page coloring optimization attempts to locate a page
630 * that does not overload other nearby pages in the object in
631 * the cpu's L1 or L2 caches. We need this optimization because
632 * cpu caches tend to be physical caches, while object spaces tend
635 * This routine must be called at splvm().
636 * This routine may not block.
638 * Note that this routine is carefully inlined. A non-inlined version
639 * is available for outside callers but the only critical path is
640 * from within this source file.
644 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
649 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist);
651 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl);
653 m = _vm_page_list_find2(basequeue, index);
658 _vm_page_list_find2(int basequeue, int index)
662 struct vpgqueues *pq;
664 pq = &vm_page_queues[basequeue];
667 * Note that for the first loop, index+i and index-i wind up at the
668 * same place. Even though this is not totally optimal, we've already
669 * blown it by missing the cache case so we do not care.
672 for(i = PQ_L2_SIZE / 2; i > 0; --i) {
673 if ((m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl)) != NULL)
676 if ((m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl)) != NULL)
683 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
685 return(_vm_page_list_find(basequeue, index, prefer_zero));
691 * vm_page_select_cache:
693 * Find a page on the cache queue with color optimization. As pages
694 * might be found, but not applicable, they are deactivated. This
695 * keeps us from using potentially busy cached pages.
697 * This routine must be called at splvm().
698 * This routine may not block.
701 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex)
706 m = vm_page_list_find(
708 (pindex + object->pg_color) & PQ_L2_MASK,
711 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
712 m->hold_count || m->wire_count)) {
713 vm_page_deactivate(m);
721 * vm_page_select_free:
723 * Find a free or zero page, with specified preference. We attempt to
724 * inline the nominal case and fall back to _vm_page_select_free()
727 * This routine must be called at splvm().
728 * This routine may not block.
731 static __inline vm_page_t
732 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero)
736 m = vm_page_list_find(
738 (pindex + object->pg_color) & PQ_L2_MASK,
747 * Allocate and return a memory cell associated
748 * with this VM object/offset pair.
751 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain
752 * VM_ALLOC_SYSTEM greater free drain
753 * VM_ALLOC_INTERRUPT allow free list to be completely drained
754 * VM_ALLOC_ZERO advisory request for pre-zero'd page
756 * Object must be locked.
757 * This routine may not block.
759 * Additional special handling is required when called from an
760 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with
761 * the page cache in this case.
765 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
770 KASSERT(!vm_page_lookup(object, pindex),
771 ("vm_page_alloc: page already allocated"));
773 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
776 * The pager is allowed to eat deeper into the free page list.
778 if (curthread == pagethread)
779 page_req |= VM_ALLOC_SYSTEM;
783 if (vmstats.v_free_count > vmstats.v_free_reserved ||
784 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) ||
785 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 &&
786 vmstats.v_free_count > vmstats.v_interrupt_free_min)
789 * The free queue has sufficient free pages to take one out.
791 if (page_req & VM_ALLOC_ZERO)
792 m = vm_page_select_free(object, pindex, TRUE);
794 m = vm_page_select_free(object, pindex, FALSE);
795 } else if (page_req & VM_ALLOC_NORMAL) {
797 * Allocatable from the cache (non-interrupt only). On
798 * success, we must free the page and try again, thus
799 * ensuring that vmstats.v_*_free_min counters are replenished.
802 if (curthread->td_preempted) {
803 printf("vm_page_alloc(): warning, attempt to allocate"
804 " cache page from preempting interrupt\n");
807 m = vm_page_select_cache(object, pindex);
810 m = vm_page_select_cache(object, pindex);
813 * On succuess move the page into the free queue and loop.
816 KASSERT(m->dirty == 0,
817 ("Found dirty cache page %p", m));
819 vm_page_protect(m, VM_PROT_NONE);
825 * On failure return NULL
828 #if defined(DIAGNOSTIC)
829 if (vmstats.v_cache_count > 0)
830 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count);
832 vm_pageout_deficit++;
837 * No pages available, wakeup the pageout daemon and give up.
840 vm_pageout_deficit++;
848 KASSERT(m != NULL, ("vm_page_alloc(): missing page on free queue\n"));
851 * Remove from free queue
853 vm_page_unqueue_nowakeup(m);
856 * Initialize structure. Only the PG_ZERO flag is inherited.
858 if (m->flags & PG_ZERO) {
859 vm_page_zero_count--;
860 m->flags = PG_ZERO | PG_BUSY;
869 KASSERT(m->dirty == 0,
870 ("vm_page_alloc: free/cache page %p was dirty", m));
873 * vm_page_insert() is safe prior to the splx(). Note also that
874 * inserting a page here does not insert it into the pmap (which
875 * could cause us to block allocating memory). We cannot block
878 vm_page_insert(m, object, pindex);
881 * Don't wakeup too often - wakeup the pageout daemon when
882 * we would be nearly out of memory.
884 if (vm_paging_needed())
892 * vm_wait: (also see VM_WAIT macro)
894 * Block until free pages are available for allocation
895 * - Called in various places before memory allocations.
904 if (curthread == pagethread) {
905 vm_pageout_pages_needed = 1;
906 tsleep(&vm_pageout_pages_needed, 0, "VMWait", 0);
908 if (!vm_pages_needed) {
910 wakeup(&vm_pages_needed);
912 tsleep(&vmstats.v_free_count, 0, "vmwait", 0);
918 * vm_waitpfault: (also see VM_WAITPFAULT macro)
920 * Block until free pages are available for allocation
921 * - Called only in vm_fault so that processes page faulting
922 * can be easily tracked.
923 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
924 * processes will be able to grab memory first. Do not change
925 * this balance without careful testing first.
934 if (!vm_pages_needed) {
936 wakeup(&vm_pages_needed);
938 tsleep(&vmstats.v_free_count, 0, "pfault", 0);
945 * Put the specified page on the active list (if appropriate).
946 * Ensure that act_count is at least ACT_INIT but do not otherwise
949 * The page queues must be locked.
950 * This routine may not block.
953 vm_page_activate(vm_page_t m)
958 if (m->queue != PQ_ACTIVE) {
959 if ((m->queue - m->pc) == PQ_CACHE)
960 mycpu->gd_cnt.v_reactivated++;
964 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
965 m->queue = PQ_ACTIVE;
966 vm_page_queues[PQ_ACTIVE].lcnt++;
967 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
968 if (m->act_count < ACT_INIT)
969 m->act_count = ACT_INIT;
970 vmstats.v_active_count++;
973 if (m->act_count < ACT_INIT)
974 m->act_count = ACT_INIT;
981 * vm_page_free_wakeup:
983 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
984 * routine is called when a page has been added to the cache or free
987 * This routine may not block.
988 * This routine must be called at splvm()
991 vm_page_free_wakeup(void)
994 * if pageout daemon needs pages, then tell it that there are
997 if (vm_pageout_pages_needed &&
998 vmstats.v_cache_count + vmstats.v_free_count >= vmstats.v_pageout_free_min) {
999 wakeup(&vm_pageout_pages_needed);
1000 vm_pageout_pages_needed = 0;
1003 * wakeup processes that are waiting on memory if we hit a
1004 * high water mark. And wakeup scheduler process if we have
1005 * lots of memory. this process will swapin processes.
1007 if (vm_pages_needed && !vm_page_count_min()) {
1008 vm_pages_needed = 0;
1009 wakeup(&vmstats.v_free_count);
1016 * Returns the given page to the PQ_FREE list,
1017 * disassociating it with any VM object.
1019 * Object and page must be locked prior to entry.
1020 * This routine may not block.
1024 vm_page_free_toq(vm_page_t m)
1027 struct vpgqueues *pq;
1029 vm_object_t object = m->object;
1034 mycpu->gd_cnt.v_tfree++;
1036 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
1038 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
1039 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
1041 if ((m->queue - m->pc) == PQ_FREE)
1042 panic("vm_page_free: freeing free page");
1044 panic("vm_page_free: freeing busy page");
1048 * unqueue, then remove page. Note that we cannot destroy
1049 * the page here because we do not want to call the pager's
1050 * callback routine until after we've put the page on the
1051 * appropriate free queue.
1054 vm_page_unqueue_nowakeup(m);
1058 * If fictitious remove object association and
1059 * return, otherwise delay object association removal.
1062 if ((m->flags & PG_FICTITIOUS) != 0) {
1070 if (m->wire_count != 0) {
1071 if (m->wire_count > 1) {
1072 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1073 m->wire_count, (long)m->pindex);
1075 panic("vm_page_free: freeing wired page");
1079 * We used to free the underlying vnode if the object was empty,
1080 * but we no longer do that because it can block. Instead, the
1081 * sync code is made responsible for the cleanup.
1085 (object->type == OBJT_VNODE) &&
1086 ((object->flags & OBJ_DEAD) == 0) &&
1087 object->handle != NULL
1089 struct vnode *vp = (struct vnode *)object->handle;
1091 if (vp && VSHOULDFREE(vp))
1097 * Clear the UNMANAGED flag when freeing an unmanaged page.
1100 if (m->flags & PG_UNMANAGED) {
1101 m->flags &= ~PG_UNMANAGED;
1104 pmap_page_is_free(m);
1108 if (m->hold_count != 0) {
1109 m->flags &= ~PG_ZERO;
1112 m->queue = PQ_FREE + m->pc;
1113 pq = &vm_page_queues[m->queue];
1118 * Put zero'd pages on the end ( where we look for zero'd pages
1119 * first ) and non-zerod pages at the head.
1122 if (m->flags & PG_ZERO) {
1123 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
1124 ++vm_page_zero_count;
1126 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1129 vm_page_free_wakeup();
1137 * Prevent PV management from being done on the page. The page is
1138 * removed from the paging queues as if it were wired, and as a
1139 * consequence of no longer being managed the pageout daemon will not
1140 * touch it (since there is no way to locate the pte mappings for the
1141 * page). madvise() calls that mess with the pmap will also no longer
1142 * operate on the page.
1144 * Beyond that the page is still reasonably 'normal'. Freeing the page
1145 * will clear the flag.
1147 * This routine is used by OBJT_PHYS objects - objects using unswappable
1148 * physical memory as backing store rather then swap-backed memory and
1149 * will eventually be extended to support 4MB unmanaged physical
1154 vm_page_unmanage(vm_page_t m)
1159 if ((m->flags & PG_UNMANAGED) == 0) {
1160 if (m->wire_count == 0)
1163 vm_page_flag_set(m, PG_UNMANAGED);
1170 * Mark this page as wired down by yet
1171 * another map, removing it from paging queues
1174 * The page queues must be locked.
1175 * This routine may not block.
1178 vm_page_wire(vm_page_t m)
1183 * Only bump the wire statistics if the page is not already wired,
1184 * and only unqueue the page if it is on some queue (if it is unmanaged
1185 * it is already off the queues).
1188 if (m->wire_count == 0) {
1189 if ((m->flags & PG_UNMANAGED) == 0)
1191 vmstats.v_wire_count++;
1194 KASSERT(m->wire_count != 0,
1195 ("vm_page_wire: wire_count overflow m=%p", m));
1198 vm_page_flag_set(m, PG_MAPPED);
1204 * Release one wiring of this page, potentially
1205 * enabling it to be paged again.
1207 * Many pages placed on the inactive queue should actually go
1208 * into the cache, but it is difficult to figure out which. What
1209 * we do instead, if the inactive target is well met, is to put
1210 * clean pages at the head of the inactive queue instead of the tail.
1211 * This will cause them to be moved to the cache more quickly and
1212 * if not actively re-referenced, freed more quickly. If we just
1213 * stick these pages at the end of the inactive queue, heavy filesystem
1214 * meta-data accesses can cause an unnecessary paging load on memory bound
1215 * processes. This optimization causes one-time-use metadata to be
1216 * reused more quickly.
1218 * BUT, if we are in a low-memory situation we have no choice but to
1219 * put clean pages on the cache queue.
1221 * A number of routines use vm_page_unwire() to guarantee that the page
1222 * will go into either the inactive or active queues, and will NEVER
1223 * be placed in the cache - for example, just after dirtying a page.
1224 * dirty pages in the cache are not allowed.
1226 * The page queues must be locked.
1227 * This routine may not block.
1230 vm_page_unwire(vm_page_t m, int activate)
1236 if (m->wire_count > 0) {
1238 if (m->wire_count == 0) {
1239 vmstats.v_wire_count--;
1240 if (m->flags & PG_UNMANAGED) {
1242 } else if (activate) {
1243 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1244 m->queue = PQ_ACTIVE;
1245 vm_page_queues[PQ_ACTIVE].lcnt++;
1246 vmstats.v_active_count++;
1248 vm_page_flag_clear(m, PG_WINATCFLS);
1249 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1250 m->queue = PQ_INACTIVE;
1251 vm_page_queues[PQ_INACTIVE].lcnt++;
1252 vmstats.v_inactive_count++;
1256 panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
1263 * Move the specified page to the inactive queue. If the page has
1264 * any associated swap, the swap is deallocated.
1266 * Normally athead is 0 resulting in LRU operation. athead is set
1267 * to 1 if we want this page to be 'as if it were placed in the cache',
1268 * except without unmapping it from the process address space.
1270 * This routine may not block.
1272 static __inline void
1273 _vm_page_deactivate(vm_page_t m, int athead)
1278 * Ignore if already inactive.
1280 if (m->queue == PQ_INACTIVE)
1284 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1285 if ((m->queue - m->pc) == PQ_CACHE)
1286 mycpu->gd_cnt.v_reactivated++;
1287 vm_page_flag_clear(m, PG_WINATCFLS);
1290 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1292 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1293 m->queue = PQ_INACTIVE;
1294 vm_page_queues[PQ_INACTIVE].lcnt++;
1295 vmstats.v_inactive_count++;
1301 vm_page_deactivate(vm_page_t m)
1303 _vm_page_deactivate(m, 0);
1307 * vm_page_try_to_cache:
1309 * Returns 0 on failure, 1 on success
1312 vm_page_try_to_cache(vm_page_t m)
1314 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1315 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1318 vm_page_test_dirty(m);
1326 * vm_page_try_to_free()
1328 * Attempt to free the page. If we cannot free it, we do nothing.
1329 * 1 is returned on success, 0 on failure.
1333 vm_page_try_to_free(vm_page_t m)
1335 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1336 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1339 vm_page_test_dirty(m);
1343 vm_page_protect(m, VM_PROT_NONE);
1352 * Put the specified page onto the page cache queue (if appropriate).
1354 * This routine may not block.
1357 vm_page_cache(vm_page_t m)
1361 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
1362 m->wire_count || m->hold_count) {
1363 printf("vm_page_cache: attempting to cache busy/held page\n");
1366 if ((m->queue - m->pc) == PQ_CACHE)
1370 * Remove all pmaps and indicate that the page is not
1371 * writeable or mapped.
1374 vm_page_protect(m, VM_PROT_NONE);
1375 if (m->dirty != 0) {
1376 panic("vm_page_cache: caching a dirty page, pindex: %ld",
1380 vm_page_unqueue_nowakeup(m);
1381 m->queue = PQ_CACHE + m->pc;
1382 vm_page_queues[m->queue].lcnt++;
1383 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq);
1384 vmstats.v_cache_count++;
1385 vm_page_free_wakeup();
1392 * Cache, deactivate, or do nothing as appropriate. This routine
1393 * is typically used by madvise() MADV_DONTNEED.
1395 * Generally speaking we want to move the page into the cache so
1396 * it gets reused quickly. However, this can result in a silly syndrome
1397 * due to the page recycling too quickly. Small objects will not be
1398 * fully cached. On the otherhand, if we move the page to the inactive
1399 * queue we wind up with a problem whereby very large objects
1400 * unnecessarily blow away our inactive and cache queues.
1402 * The solution is to move the pages based on a fixed weighting. We
1403 * either leave them alone, deactivate them, or move them to the cache,
1404 * where moving them to the cache has the highest weighting.
1405 * By forcing some pages into other queues we eventually force the
1406 * system to balance the queues, potentially recovering other unrelated
1407 * space from active. The idea is to not force this to happen too
1412 vm_page_dontneed(vm_page_t m)
1414 static int dnweight;
1421 * occassionally leave the page alone
1424 if ((dnw & 0x01F0) == 0 ||
1425 m->queue == PQ_INACTIVE ||
1426 m->queue - m->pc == PQ_CACHE
1428 if (m->act_count >= ACT_INIT)
1434 vm_page_test_dirty(m);
1436 if (m->dirty || (dnw & 0x0070) == 0) {
1438 * Deactivate the page 3 times out of 32.
1443 * Cache the page 28 times out of every 32. Note that
1444 * the page is deactivated instead of cached, but placed
1445 * at the head of the queue instead of the tail.
1449 _vm_page_deactivate(m, head);
1453 * Grab a page, blocking if it is busy and allocating a page if necessary.
1454 * A busy page is returned or NULL.
1456 * If VM_ALLOC_RETRY is specified VM_ALLOC_NORMAL must also be specified.
1457 * If VM_ALLOC_RETRY is not specified
1459 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
1460 * always returned if we had blocked.
1461 * This routine will never return NULL if VM_ALLOC_RETRY is set.
1462 * This routine may not be called from an interrupt.
1463 * The returned page may not be entirely valid.
1465 * This routine may be called from mainline code without spl protection and
1466 * be guarenteed a busied page associated with the object at the specified
1470 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
1475 KKASSERT(allocflags &
1476 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
1479 if ((m = vm_page_lookup(object, pindex)) != NULL) {
1480 if (m->busy || (m->flags & PG_BUSY)) {
1481 generation = object->generation;
1483 while ((object->generation == generation) &&
1484 (m->busy || (m->flags & PG_BUSY))) {
1485 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
1486 tsleep(m, 0, "pgrbwt", 0);
1487 if ((allocflags & VM_ALLOC_RETRY) == 0) {
1498 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1501 if ((allocflags & VM_ALLOC_RETRY) == 0)
1511 * Mapping function for valid bits or for dirty bits in
1512 * a page. May not block.
1514 * Inputs are required to range within a page.
1518 vm_page_bits(int base, int size)
1524 base + size <= PAGE_SIZE,
1525 ("vm_page_bits: illegal base/size %d/%d", base, size)
1528 if (size == 0) /* handle degenerate case */
1531 first_bit = base >> DEV_BSHIFT;
1532 last_bit = (base + size - 1) >> DEV_BSHIFT;
1534 return ((2 << last_bit) - (1 << first_bit));
1538 * vm_page_set_validclean:
1540 * Sets portions of a page valid and clean. The arguments are expected
1541 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1542 * of any partial chunks touched by the range. The invalid portion of
1543 * such chunks will be zero'd.
1545 * This routine may not block.
1547 * (base + size) must be less then or equal to PAGE_SIZE.
1550 vm_page_set_validclean(vm_page_t m, int base, int size)
1556 if (size == 0) /* handle degenerate case */
1560 * If the base is not DEV_BSIZE aligned and the valid
1561 * bit is clear, we have to zero out a portion of the
1565 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
1566 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
1568 pmap_zero_page_area(
1576 * If the ending offset is not DEV_BSIZE aligned and the
1577 * valid bit is clear, we have to zero out a portion of
1581 endoff = base + size;
1583 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
1584 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
1586 pmap_zero_page_area(
1589 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
1594 * Set valid, clear dirty bits. If validating the entire
1595 * page we can safely clear the pmap modify bit. We also
1596 * use this opportunity to clear the PG_NOSYNC flag. If a process
1597 * takes a write fault on a MAP_NOSYNC memory area the flag will
1600 * We set valid bits inclusive of any overlap, but we can only
1601 * clear dirty bits for DEV_BSIZE chunks that are fully within
1605 pagebits = vm_page_bits(base, size);
1606 m->valid |= pagebits;
1608 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
1609 frag = DEV_BSIZE - frag;
1615 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
1617 m->dirty &= ~pagebits;
1618 if (base == 0 && size == PAGE_SIZE) {
1619 pmap_clear_modify(m);
1620 vm_page_flag_clear(m, PG_NOSYNC);
1627 vm_page_set_dirty(vm_page_t m, int base, int size)
1629 m->dirty |= vm_page_bits(base, size);
1635 vm_page_clear_dirty(vm_page_t m, int base, int size)
1637 m->dirty &= ~vm_page_bits(base, size);
1641 * vm_page_set_invalid:
1643 * Invalidates DEV_BSIZE'd chunks within a page. Both the
1644 * valid and dirty bits for the effected areas are cleared.
1649 vm_page_set_invalid(vm_page_t m, int base, int size)
1653 bits = vm_page_bits(base, size);
1656 m->object->generation++;
1660 * vm_page_zero_invalid()
1662 * The kernel assumes that the invalid portions of a page contain
1663 * garbage, but such pages can be mapped into memory by user code.
1664 * When this occurs, we must zero out the non-valid portions of the
1665 * page so user code sees what it expects.
1667 * Pages are most often semi-valid when the end of a file is mapped
1668 * into memory and the file's size is not page aligned.
1672 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
1678 * Scan the valid bits looking for invalid sections that
1679 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
1680 * valid bit may be set ) have already been zerod by
1681 * vm_page_set_validclean().
1684 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
1685 if (i == (PAGE_SIZE / DEV_BSIZE) ||
1686 (m->valid & (1 << i))
1689 pmap_zero_page_area(
1692 (i - b) << DEV_BSHIFT
1700 * setvalid is TRUE when we can safely set the zero'd areas
1701 * as being valid. We can do this if there are no cache consistency
1702 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
1706 m->valid = VM_PAGE_BITS_ALL;
1712 * Is (partial) page valid? Note that the case where size == 0
1713 * will return FALSE in the degenerate case where the page is
1714 * entirely invalid, and TRUE otherwise.
1720 vm_page_is_valid(vm_page_t m, int base, int size)
1722 int bits = vm_page_bits(base, size);
1724 if (m->valid && ((m->valid & bits) == bits))
1731 * update dirty bits from pmap/mmu. May not block.
1735 vm_page_test_dirty(vm_page_t m)
1737 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
1742 #include "opt_ddb.h"
1744 #include <sys/kernel.h>
1746 #include <ddb/ddb.h>
1748 DB_SHOW_COMMAND(page, vm_page_print_page_info)
1750 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count);
1751 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count);
1752 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count);
1753 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count);
1754 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count);
1755 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved);
1756 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min);
1757 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target);
1758 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min);
1759 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target);
1762 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
1765 db_printf("PQ_FREE:");
1766 for(i=0;i<PQ_L2_SIZE;i++) {
1767 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
1771 db_printf("PQ_CACHE:");
1772 for(i=0;i<PQ_L2_SIZE;i++) {
1773 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
1777 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
1778 vm_page_queues[PQ_ACTIVE].lcnt,
1779 vm_page_queues[PQ_INACTIVE].lcnt);