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 1.26 2004/09/17 10:02:12 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.
68 * Resident memory management module. The module manipulates 'VM pages'.
69 * A VM page is the core building block for memory management.
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 #include <sys/thread2.h>
94 static void vm_page_queue_init(void);
95 static void vm_page_free_wakeup(void);
96 static vm_page_t vm_page_select_cache(vm_object_t, vm_pindex_t);
97 static vm_page_t _vm_page_list_find2(int basequeue, int index);
99 static int vm_page_bucket_count; /* How big is array? */
100 static int vm_page_hash_mask; /* Mask for hash function */
101 static struct vm_page **vm_page_buckets; /* Array of buckets */
102 static volatile int vm_page_bucket_generation;
103 struct vpgqueues vm_page_queues[PQ_COUNT]; /* Array of tailq lists */
106 vm_page_queue_init(void)
110 for (i = 0; i < PQ_L2_SIZE; i++)
111 vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count;
112 for (i = 0; i < PQ_L2_SIZE; i++)
113 vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count;
115 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 /* PQ_NONE has no queue */
120 for (i = 0; i < PQ_COUNT; i++)
121 TAILQ_INIT(&vm_page_queues[i].pl);
125 * note: place in initialized data section? Is this necessary?
128 int vm_page_array_size = 0;
129 int vm_page_zero_count = 0;
130 vm_page_t vm_page_array = 0;
135 * Sets the page size, perhaps based upon the memory size.
136 * Must be called before any use of page-size dependent functions.
139 vm_set_page_size(void)
141 if (vmstats.v_page_size == 0)
142 vmstats.v_page_size = PAGE_SIZE;
143 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0)
144 panic("vm_set_page_size: page size not a power of two");
150 * Add a new page to the freelist for use by the system. New pages
151 * are added to both the head and tail of the associated free page
152 * queue in a bottom-up fashion, so both zero'd and non-zero'd page
153 * requests pull 'recent' adds (higher physical addresses) first.
155 * Must be called at splhigh().
158 vm_add_new_page(vm_paddr_t pa)
160 struct vpgqueues *vpq;
163 ++vmstats.v_page_count;
164 ++vmstats.v_free_count;
165 m = PHYS_TO_VM_PAGE(pa);
168 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK;
169 m->queue = m->pc + PQ_FREE;
171 vpq = &vm_page_queues[m->queue];
173 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
175 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
176 vpq->flipflop = 1 - vpq->flipflop;
178 vm_page_queues[m->queue].lcnt++;
185 * Initializes the resident memory module.
187 * Allocates memory for the page cells, and for the object/offset-to-page
188 * hash table headers. Each page cell is initialized and placed on the
192 vm_page_startup(vm_offset_t starta, vm_offset_t enda, vm_offset_t vaddr)
195 struct vm_page **bucket;
197 vm_paddr_t page_range;
204 /* the biggest memory array is the second group of pages */
206 vm_paddr_t biggestone, biggestsize;
214 vaddr = round_page(vaddr);
216 for (i = 0; phys_avail[i + 1]; i += 2) {
217 phys_avail[i] = round_page(phys_avail[i]);
218 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
221 for (i = 0; phys_avail[i + 1]; i += 2) {
222 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
224 if (size > biggestsize) {
232 end = phys_avail[biggestone+1];
235 * Initialize the queue headers for the free queue, the active queue
236 * and the inactive queue.
239 vm_page_queue_init();
242 * Allocate (and initialize) the hash table buckets.
244 * The number of buckets MUST BE a power of 2, and the actual value is
245 * the next power of 2 greater than the number of physical pages in
248 * We make the hash table approximately 2x the number of pages to
249 * reduce the chain length. This is about the same size using the
250 * singly-linked list as the 1x hash table we were using before
251 * using TAILQ but the chain length will be smaller.
253 * Note: This computation can be tweaked if desired.
255 vm_page_buckets = (struct vm_page **)vaddr;
256 bucket = vm_page_buckets;
257 if (vm_page_bucket_count == 0) {
258 vm_page_bucket_count = 1;
259 while (vm_page_bucket_count < atop(total))
260 vm_page_bucket_count <<= 1;
262 vm_page_bucket_count <<= 1;
263 vm_page_hash_mask = vm_page_bucket_count - 1;
266 * Validate these addresses.
268 new_end = end - vm_page_bucket_count * sizeof(struct vm_page *);
269 new_end = trunc_page(new_end);
270 mapped = round_page(vaddr);
271 vaddr = pmap_map(mapped, new_end, end,
272 VM_PROT_READ | VM_PROT_WRITE);
273 vaddr = round_page(vaddr);
274 bzero((caddr_t) mapped, vaddr - mapped);
276 for (i = 0; i < vm_page_bucket_count; i++) {
282 * Compute the number of pages of memory that will be available for
283 * use (taking into account the overhead of a page structure per
286 first_page = phys_avail[0] / PAGE_SIZE;
287 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
288 npages = (total - (page_range * sizeof(struct vm_page)) -
289 (end - new_end)) / PAGE_SIZE;
294 * Initialize the mem entry structures now, and put them in the free
297 vm_page_array = (vm_page_t) vaddr;
301 * Validate these addresses.
303 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
304 mapped = pmap_map(mapped, new_end, end,
305 VM_PROT_READ | VM_PROT_WRITE);
308 * Clear all of the page structures
310 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
311 vm_page_array_size = page_range;
314 * Construct the free queue(s) in ascending order (by physical
315 * address) so that the first 16MB of physical memory is allocated
316 * last rather than first. On large-memory machines, this avoids
317 * the exhaustion of low physical memory before isa_dmainit has run.
319 vmstats.v_page_count = 0;
320 vmstats.v_free_count = 0;
321 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
326 last_pa = phys_avail[i + 1];
327 while (pa < last_pa && npages-- > 0) {
336 * Distributes the object/offset key pair among hash buckets.
338 * NOTE: This macro depends on vm_page_bucket_count being a power of 2.
339 * This routine may not block.
341 * We try to randomize the hash based on the object to spread the pages
342 * out in the hash table without it costing us too much.
345 vm_page_hash(vm_object_t object, vm_pindex_t pindex)
347 int i = ((uintptr_t)object + pindex) ^ object->hash_rand;
349 return(i & vm_page_hash_mask);
353 * The opposite of vm_page_hold(). A page can be freed while being held,
354 * which places it on the PQ_HOLD queue. We must call vm_page_free_toq()
355 * in this case to actually free it once the hold count drops to 0.
357 * This routine must be called at splvm().
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 * 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 * This routine may not block.
377 * This routine must be called at splvm().
380 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
382 struct vm_page **bucket;
384 if (m->object != NULL)
385 panic("vm_page_insert: already inserted");
388 * Record the object/offset pair in this page
394 * Insert it into the object_object/offset hash table
396 bucket = &vm_page_buckets[vm_page_hash(object, pindex)];
399 vm_page_bucket_generation++;
402 * Now link into the object's list of backed pages.
404 TAILQ_INSERT_TAIL(&object->memq, m, listq);
405 object->generation++;
408 * show that the object has one more resident page.
410 object->resident_page_count++;
413 * Since we are inserting a new and possibly dirty page,
414 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
416 if (m->flags & PG_WRITEABLE)
417 vm_object_set_writeable_dirty(object);
421 * Removes the given vm_page_t from the global (object,index) hash table
422 * and from the object's memq.
424 * The underlying pmap entry (if any) is NOT removed here.
425 * This routine may not block.
427 * The page must be BUSY and will remain BUSY on return. No spl needs to be
428 * held on call to this routine.
430 * note: FreeBSD side effect was to unbusy the page on return. We leave
434 vm_page_remove(vm_page_t m)
437 struct vm_page **bucket;
439 if (m->object == NULL)
442 if ((m->flags & PG_BUSY) == 0)
443 panic("vm_page_remove: page not busy");
448 * Remove from the object_object/offset hash table. The object
449 * must be on the hash queue, we will panic if it isn't
451 * Note: we must NULL-out m->hnext to prevent loops in detached
452 * buffers with vm_page_lookup().
455 bucket = &vm_page_buckets[vm_page_hash(m->object, m->pindex)];
456 while (*bucket != m) {
458 panic("vm_page_remove(): page not found in hash");
459 bucket = &(*bucket)->hnext;
463 vm_page_bucket_generation++;
466 * Now remove from the object's list of backed pages.
468 TAILQ_REMOVE(&object->memq, m, listq);
471 * And show that the object has one fewer resident page.
473 object->resident_page_count--;
474 object->generation++;
481 * Locate and return the page at (object, pindex), or NULL if the
482 * page could not be found.
484 * This routine will operate properly without spl protection, but
485 * the returned page could be in flux if it is busy. Because an
486 * interrupt can race a caller's busy check (unbusying and freeing the
487 * page we return before the caller is able to check the busy bit),
488 * the caller should generally call this routine at splvm().
490 * Callers may call this routine without spl protection if they know
491 * 'for sure' that the page will not be ripped out from under them
495 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
498 struct vm_page **bucket;
502 * Search the hash table for this object/offset pair
505 generation = vm_page_bucket_generation;
506 bucket = &vm_page_buckets[vm_page_hash(object, pindex)];
507 for (m = *bucket; m != NULL; m = m->hnext) {
508 if ((m->object == object) && (m->pindex == pindex)) {
509 if (vm_page_bucket_generation != generation)
514 if (vm_page_bucket_generation != generation)
522 * Move the given memory entry from its current object to the specified
523 * target object/offset.
525 * The object must be locked.
526 * This routine may not block.
528 * Note: This routine will raise itself to splvm(), the caller need not.
530 * Note: Swap associated with the page must be invalidated by the move. We
531 * have to do this for several reasons: (1) we aren't freeing the
532 * page, (2) we are dirtying the page, (3) the VM system is probably
533 * moving the page from object A to B, and will then later move
534 * the backing store from A to B and we can't have a conflict.
536 * Note: We *always* dirty the page. It is necessary both for the
537 * fact that we moved it, and because we may be invalidating
538 * swap. If the page is on the cache, we have to deactivate it
539 * or vm_page_dirty() will panic. Dirty pages are not allowed
543 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
547 vm_page_insert(m, new_object, new_pindex);
548 if (m->queue - m->pc == PQ_CACHE)
549 vm_page_deactivate(m);
556 * vm_page_unqueue() without any wakeup. This routine is used when a page
557 * is being moved between queues or otherwise is to remain BUSYied by the
560 * This routine must be called at splhigh().
561 * This routine may not block.
564 vm_page_unqueue_nowakeup(vm_page_t m)
566 int queue = m->queue;
567 struct vpgqueues *pq;
569 if (queue != PQ_NONE) {
570 pq = &vm_page_queues[queue];
572 TAILQ_REMOVE(&pq->pl, m, pageq);
579 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
582 * This routine must be called at splhigh().
583 * This routine may not block.
586 vm_page_unqueue(vm_page_t m)
588 int queue = m->queue;
589 struct vpgqueues *pq;
591 if (queue != PQ_NONE) {
593 pq = &vm_page_queues[queue];
594 TAILQ_REMOVE(&pq->pl, m, pageq);
597 if ((queue - m->pc) == PQ_CACHE) {
598 if (vm_paging_needed())
605 * vm_page_list_find()
607 * Find a page on the specified queue with color optimization.
609 * The page coloring optimization attempts to locate a page that does
610 * not overload other nearby pages in the object in the cpu's L1 or L2
611 * caches. We need this optimization because cpu caches tend to be
612 * physical caches, while object spaces tend to be virtual.
614 * This routine must be called at splvm().
615 * This routine may not block.
617 * Note that this routine is carefully inlined. A non-inlined version
618 * is available for outside callers but the only critical path is
619 * from within this source file.
623 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
628 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist);
630 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl);
632 m = _vm_page_list_find2(basequeue, index);
637 _vm_page_list_find2(int basequeue, int index)
641 struct vpgqueues *pq;
643 pq = &vm_page_queues[basequeue];
646 * Note that for the first loop, index+i and index-i wind up at the
647 * same place. Even though this is not totally optimal, we've already
648 * blown it by missing the cache case so we do not care.
651 for(i = PQ_L2_SIZE / 2; i > 0; --i) {
652 if ((m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl)) != NULL)
655 if ((m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl)) != NULL)
662 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
664 return(_vm_page_list_find(basequeue, index, prefer_zero));
668 * Find a page on the cache queue with color optimization. As pages
669 * might be found, but not applicable, they are deactivated. This
670 * keeps us from using potentially busy cached pages.
672 * This routine must be called at splvm().
673 * This routine may not block.
676 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex)
681 m = _vm_page_list_find(
683 (pindex + object->pg_color) & PQ_L2_MASK,
686 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
687 m->hold_count || m->wire_count)) {
688 vm_page_deactivate(m);
697 * Find a free or zero page, with specified preference. We attempt to
698 * inline the nominal case and fall back to _vm_page_select_free()
701 * This routine must be called at splvm().
702 * This routine may not block.
704 static __inline vm_page_t
705 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero)
709 m = _vm_page_list_find(
711 (pindex + object->pg_color) & PQ_L2_MASK,
720 * Allocate and return a memory cell associated with this VM object/offset
725 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain
726 * VM_ALLOC_SYSTEM greater free drain
727 * VM_ALLOC_INTERRUPT allow free list to be completely drained
728 * VM_ALLOC_ZERO advisory request for pre-zero'd page
730 * The object must be locked.
731 * This routine may not block.
732 * The returned page will be marked PG_BUSY
734 * Additional special handling is required when called from an interrupt
735 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache
739 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
743 KASSERT(!vm_page_lookup(object, pindex),
744 ("vm_page_alloc: page already allocated"));
746 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
749 * The pager is allowed to eat deeper into the free page list.
751 if (curthread == pagethread)
752 page_req |= VM_ALLOC_SYSTEM;
756 if (vmstats.v_free_count > vmstats.v_free_reserved ||
757 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) ||
758 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 &&
759 vmstats.v_free_count > vmstats.v_interrupt_free_min)
762 * The free queue has sufficient free pages to take one out.
764 if (page_req & VM_ALLOC_ZERO)
765 m = vm_page_select_free(object, pindex, TRUE);
767 m = vm_page_select_free(object, pindex, FALSE);
768 } else if (page_req & VM_ALLOC_NORMAL) {
770 * Allocatable from the cache (non-interrupt only). On
771 * success, we must free the page and try again, thus
772 * ensuring that vmstats.v_*_free_min counters are replenished.
775 if (curthread->td_preempted) {
776 printf("vm_page_alloc(): warning, attempt to allocate"
777 " cache page from preempting interrupt\n");
780 m = vm_page_select_cache(object, pindex);
783 m = vm_page_select_cache(object, pindex);
786 * On success move the page into the free queue and loop.
789 KASSERT(m->dirty == 0,
790 ("Found dirty cache page %p", m));
792 vm_page_protect(m, VM_PROT_NONE);
798 * On failure return NULL
801 #if defined(DIAGNOSTIC)
802 if (vmstats.v_cache_count > 0)
803 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count);
805 vm_pageout_deficit++;
810 * No pages available, wakeup the pageout daemon and give up.
813 vm_pageout_deficit++;
819 * Good page found. The page has not yet been busied. We are in
820 * a critical section.
822 KASSERT(m != NULL, ("vm_page_alloc(): missing page on free queue\n"));
825 * Remove from free queue
827 vm_page_unqueue_nowakeup(m);
830 * Initialize structure. Only the PG_ZERO flag is inherited. Set
833 if (m->flags & PG_ZERO) {
834 vm_page_zero_count--;
835 m->flags = PG_ZERO | PG_BUSY;
844 KASSERT(m->dirty == 0,
845 ("vm_page_alloc: free/cache page %p was dirty", m));
848 * vm_page_insert() is safe prior to the crit_exit(). Note also that
849 * inserting a page here does not insert it into the pmap (which
850 * could cause us to block allocating memory). We cannot block
853 vm_page_insert(m, object, pindex);
856 * Don't wakeup too often - wakeup the pageout daemon when
857 * we would be nearly out of memory.
859 if (vm_paging_needed())
865 * A PG_BUSY page is returned.
871 * Block until free pages are available for allocation, called in various
872 * places before memory allocations.
880 if (curthread == pagethread) {
881 vm_pageout_pages_needed = 1;
882 tsleep(&vm_pageout_pages_needed, 0, "VMWait", 0);
884 if (!vm_pages_needed) {
886 wakeup(&vm_pages_needed);
888 tsleep(&vmstats.v_free_count, 0, "vmwait", 0);
894 * Block until free pages are available for allocation
896 * Called only in vm_fault so that processes page faulting can be
899 * Sleeps at a lower priority than vm_wait() so that vm_wait()ing
900 * processes will be able to grab memory first. Do not change
901 * this balance without careful testing first.
909 if (!vm_pages_needed) {
911 wakeup(&vm_pages_needed);
913 tsleep(&vmstats.v_free_count, 0, "pfault", 0);
918 * Put the specified page on the active list (if appropriate). Ensure
919 * that act_count is at least ACT_INIT but do not otherwise mess with it.
921 * The page queues must be locked.
922 * This routine may not block.
925 vm_page_activate(vm_page_t m)
928 if (m->queue != PQ_ACTIVE) {
929 if ((m->queue - m->pc) == PQ_CACHE)
930 mycpu->gd_cnt.v_reactivated++;
934 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
935 m->queue = PQ_ACTIVE;
936 vm_page_queues[PQ_ACTIVE].lcnt++;
937 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl,
939 if (m->act_count < ACT_INIT)
940 m->act_count = ACT_INIT;
941 vmstats.v_active_count++;
944 if (m->act_count < ACT_INIT)
945 m->act_count = ACT_INIT;
951 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
952 * routine is called when a page has been added to the cache or free
955 * This routine may not block.
956 * This routine must be called at splvm()
959 vm_page_free_wakeup(void)
962 * if pageout daemon needs pages, then tell it that there are
965 if (vm_pageout_pages_needed &&
966 vmstats.v_cache_count + vmstats.v_free_count >=
967 vmstats.v_pageout_free_min
969 wakeup(&vm_pageout_pages_needed);
970 vm_pageout_pages_needed = 0;
974 * wakeup processes that are waiting on memory if we hit a
975 * high water mark. And wakeup scheduler process if we have
976 * lots of memory. this process will swapin processes.
978 if (vm_pages_needed && !vm_page_count_min()) {
980 wakeup(&vmstats.v_free_count);
987 * Returns the given page to the PQ_FREE list, disassociating it with
990 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on
991 * return (the page will have been freed). No particular spl is required
994 * This routine may not block.
997 vm_page_free_toq(vm_page_t m)
999 struct vpgqueues *pq;
1002 mycpu->gd_cnt.v_tfree++;
1004 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
1006 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
1007 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
1009 if ((m->queue - m->pc) == PQ_FREE)
1010 panic("vm_page_free: freeing free page");
1012 panic("vm_page_free: freeing busy page");
1016 * unqueue, then remove page. Note that we cannot destroy
1017 * the page here because we do not want to call the pager's
1018 * callback routine until after we've put the page on the
1019 * appropriate free queue.
1021 vm_page_unqueue_nowakeup(m);
1025 * No further management of fictitious pages occurs beyond object
1026 * and queue removal.
1028 if ((m->flags & PG_FICTITIOUS) != 0) {
1037 if (m->wire_count != 0) {
1038 if (m->wire_count > 1) {
1040 "vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1041 m->wire_count, (long)m->pindex);
1043 panic("vm_page_free: freeing wired page");
1047 * Clear the UNMANAGED flag when freeing an unmanaged page.
1049 if (m->flags & PG_UNMANAGED) {
1050 m->flags &= ~PG_UNMANAGED;
1053 pmap_page_is_free(m);
1057 if (m->hold_count != 0) {
1058 m->flags &= ~PG_ZERO;
1061 m->queue = PQ_FREE + m->pc;
1063 pq = &vm_page_queues[m->queue];
1068 * Put zero'd pages on the end ( where we look for zero'd pages
1069 * first ) and non-zerod pages at the head.
1071 if (m->flags & PG_ZERO) {
1072 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
1073 ++vm_page_zero_count;
1075 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1078 vm_page_free_wakeup();
1083 * vm_page_unmanage()
1085 * Prevent PV management from being done on the page. The page is
1086 * removed from the paging queues as if it were wired, and as a
1087 * consequence of no longer being managed the pageout daemon will not
1088 * touch it (since there is no way to locate the pte mappings for the
1089 * page). madvise() calls that mess with the pmap will also no longer
1090 * operate on the page.
1092 * Beyond that the page is still reasonably 'normal'. Freeing the page
1093 * will clear the flag.
1095 * This routine is used by OBJT_PHYS objects - objects using unswappable
1096 * physical memory as backing store rather then swap-backed memory and
1097 * will eventually be extended to support 4MB unmanaged physical
1101 vm_page_unmanage(vm_page_t m)
1106 if ((m->flags & PG_UNMANAGED) == 0) {
1107 if (m->wire_count == 0)
1110 vm_page_flag_set(m, PG_UNMANAGED);
1115 * Mark this page as wired down by yet another map, removing it from
1116 * paging queues as necessary.
1118 * The page queues must be locked.
1119 * This routine may not block.
1122 vm_page_wire(vm_page_t m)
1127 * Only bump the wire statistics if the page is not already wired,
1128 * and only unqueue the page if it is on some queue (if it is unmanaged
1129 * it is already off the queues). Don't do anything with fictitious
1130 * pages because they are always wired.
1133 if ((m->flags & PG_FICTITIOUS) == 0) {
1134 if (m->wire_count == 0) {
1135 if ((m->flags & PG_UNMANAGED) == 0)
1137 vmstats.v_wire_count++;
1140 KASSERT(m->wire_count != 0,
1141 ("vm_page_wire: wire_count overflow m=%p", m));
1144 vm_page_flag_set(m, PG_MAPPED);
1148 * Release one wiring of this page, potentially enabling it to be paged again.
1150 * Many pages placed on the inactive queue should actually go
1151 * into the cache, but it is difficult to figure out which. What
1152 * we do instead, if the inactive target is well met, is to put
1153 * clean pages at the head of the inactive queue instead of the tail.
1154 * This will cause them to be moved to the cache more quickly and
1155 * if not actively re-referenced, freed more quickly. If we just
1156 * stick these pages at the end of the inactive queue, heavy filesystem
1157 * meta-data accesses can cause an unnecessary paging load on memory bound
1158 * processes. This optimization causes one-time-use metadata to be
1159 * reused more quickly.
1161 * BUT, if we are in a low-memory situation we have no choice but to
1162 * put clean pages on the cache queue.
1164 * A number of routines use vm_page_unwire() to guarantee that the page
1165 * will go into either the inactive or active queues, and will NEVER
1166 * be placed in the cache - for example, just after dirtying a page.
1167 * dirty pages in the cache are not allowed.
1169 * The page queues must be locked.
1170 * This routine may not block.
1173 vm_page_unwire(vm_page_t m, int activate)
1178 if (m->flags & PG_FICTITIOUS) {
1180 } else if (m->wire_count <= 0) {
1181 panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
1183 if (--m->wire_count == 0) {
1184 --vmstats.v_wire_count;
1185 if (m->flags & PG_UNMANAGED) {
1187 } else if (activate) {
1189 &vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1190 m->queue = PQ_ACTIVE;
1191 vm_page_queues[PQ_ACTIVE].lcnt++;
1192 vmstats.v_active_count++;
1194 vm_page_flag_clear(m, PG_WINATCFLS);
1196 &vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1197 m->queue = PQ_INACTIVE;
1198 vm_page_queues[PQ_INACTIVE].lcnt++;
1199 vmstats.v_inactive_count++;
1208 * Move the specified page to the inactive queue. If the page has
1209 * any associated swap, the swap is deallocated.
1211 * Normally athead is 0 resulting in LRU operation. athead is set
1212 * to 1 if we want this page to be 'as if it were placed in the cache',
1213 * except without unmapping it from the process address space.
1215 * This routine may not block.
1217 static __inline void
1218 _vm_page_deactivate(vm_page_t m, int athead)
1223 * Ignore if already inactive.
1225 if (m->queue == PQ_INACTIVE)
1229 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1230 if ((m->queue - m->pc) == PQ_CACHE)
1231 mycpu->gd_cnt.v_reactivated++;
1232 vm_page_flag_clear(m, PG_WINATCFLS);
1235 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1237 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1238 m->queue = PQ_INACTIVE;
1239 vm_page_queues[PQ_INACTIVE].lcnt++;
1240 vmstats.v_inactive_count++;
1246 vm_page_deactivate(vm_page_t m)
1248 _vm_page_deactivate(m, 0);
1252 * vm_page_try_to_cache:
1254 * Returns 0 on failure, 1 on success
1257 vm_page_try_to_cache(vm_page_t m)
1259 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1260 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1263 vm_page_test_dirty(m);
1271 * Attempt to free the page. If we cannot free it, we do nothing.
1272 * 1 is returned on success, 0 on failure.
1275 vm_page_try_to_free(vm_page_t m)
1277 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1278 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1281 vm_page_test_dirty(m);
1285 vm_page_protect(m, VM_PROT_NONE);
1293 * Put the specified page onto the page cache queue (if appropriate).
1295 * This routine may not block.
1298 vm_page_cache(vm_page_t m)
1302 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
1303 m->wire_count || m->hold_count) {
1304 printf("vm_page_cache: attempting to cache busy/held page\n");
1307 if ((m->queue - m->pc) == PQ_CACHE)
1311 * Remove all pmaps and indicate that the page is not
1312 * writeable or mapped.
1315 vm_page_protect(m, VM_PROT_NONE);
1316 if (m->dirty != 0) {
1317 panic("vm_page_cache: caching a dirty page, pindex: %ld",
1321 vm_page_unqueue_nowakeup(m);
1322 m->queue = PQ_CACHE + m->pc;
1323 vm_page_queues[m->queue].lcnt++;
1324 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq);
1325 vmstats.v_cache_count++;
1326 vm_page_free_wakeup();
1331 * vm_page_dontneed()
1333 * Cache, deactivate, or do nothing as appropriate. This routine
1334 * is typically used by madvise() MADV_DONTNEED.
1336 * Generally speaking we want to move the page into the cache so
1337 * it gets reused quickly. However, this can result in a silly syndrome
1338 * due to the page recycling too quickly. Small objects will not be
1339 * fully cached. On the otherhand, if we move the page to the inactive
1340 * queue we wind up with a problem whereby very large objects
1341 * unnecessarily blow away our inactive and cache queues.
1343 * The solution is to move the pages based on a fixed weighting. We
1344 * either leave them alone, deactivate them, or move them to the cache,
1345 * where moving them to the cache has the highest weighting.
1346 * By forcing some pages into other queues we eventually force the
1347 * system to balance the queues, potentially recovering other unrelated
1348 * space from active. The idea is to not force this to happen too
1352 vm_page_dontneed(vm_page_t m)
1354 static int dnweight;
1361 * occassionally leave the page alone
1364 if ((dnw & 0x01F0) == 0 ||
1365 m->queue == PQ_INACTIVE ||
1366 m->queue - m->pc == PQ_CACHE
1368 if (m->act_count >= ACT_INIT)
1374 vm_page_test_dirty(m);
1376 if (m->dirty || (dnw & 0x0070) == 0) {
1378 * Deactivate the page 3 times out of 32.
1383 * Cache the page 28 times out of every 32. Note that
1384 * the page is deactivated instead of cached, but placed
1385 * at the head of the queue instead of the tail.
1389 _vm_page_deactivate(m, head);
1393 * Grab a page, blocking if it is busy and allocating a page if necessary.
1394 * A busy page is returned or NULL.
1396 * If VM_ALLOC_RETRY is specified VM_ALLOC_NORMAL must also be specified.
1397 * If VM_ALLOC_RETRY is not specified
1399 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
1400 * always returned if we had blocked.
1401 * This routine will never return NULL if VM_ALLOC_RETRY is set.
1402 * This routine may not be called from an interrupt.
1403 * The returned page may not be entirely valid.
1405 * This routine may be called from mainline code without spl protection and
1406 * be guarenteed a busied page associated with the object at the specified
1410 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
1415 KKASSERT(allocflags &
1416 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
1419 if ((m = vm_page_lookup(object, pindex)) != NULL) {
1420 if (m->busy || (m->flags & PG_BUSY)) {
1421 generation = object->generation;
1423 while ((object->generation == generation) &&
1424 (m->busy || (m->flags & PG_BUSY))) {
1425 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
1426 tsleep(m, 0, "pgrbwt", 0);
1427 if ((allocflags & VM_ALLOC_RETRY) == 0) {
1438 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1441 if ((allocflags & VM_ALLOC_RETRY) == 0)
1451 * Mapping function for valid bits or for dirty bits in
1452 * a page. May not block.
1454 * Inputs are required to range within a page.
1457 vm_page_bits(int base, int size)
1463 base + size <= PAGE_SIZE,
1464 ("vm_page_bits: illegal base/size %d/%d", base, size)
1467 if (size == 0) /* handle degenerate case */
1470 first_bit = base >> DEV_BSHIFT;
1471 last_bit = (base + size - 1) >> DEV_BSHIFT;
1473 return ((2 << last_bit) - (1 << first_bit));
1477 * Sets portions of a page valid and clean. The arguments are expected
1478 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1479 * of any partial chunks touched by the range. The invalid portion of
1480 * such chunks will be zero'd.
1482 * This routine may not block.
1484 * (base + size) must be less then or equal to PAGE_SIZE.
1487 vm_page_set_validclean(vm_page_t m, int base, int size)
1493 if (size == 0) /* handle degenerate case */
1497 * If the base is not DEV_BSIZE aligned and the valid
1498 * bit is clear, we have to zero out a portion of the
1502 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
1503 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
1505 pmap_zero_page_area(
1513 * If the ending offset is not DEV_BSIZE aligned and the
1514 * valid bit is clear, we have to zero out a portion of
1518 endoff = base + size;
1520 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
1521 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
1523 pmap_zero_page_area(
1526 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
1531 * Set valid, clear dirty bits. If validating the entire
1532 * page we can safely clear the pmap modify bit. We also
1533 * use this opportunity to clear the PG_NOSYNC flag. If a process
1534 * takes a write fault on a MAP_NOSYNC memory area the flag will
1537 * We set valid bits inclusive of any overlap, but we can only
1538 * clear dirty bits for DEV_BSIZE chunks that are fully within
1542 pagebits = vm_page_bits(base, size);
1543 m->valid |= pagebits;
1545 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
1546 frag = DEV_BSIZE - frag;
1552 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
1554 m->dirty &= ~pagebits;
1555 if (base == 0 && size == PAGE_SIZE) {
1556 pmap_clear_modify(m);
1557 vm_page_flag_clear(m, PG_NOSYNC);
1562 vm_page_clear_dirty(vm_page_t m, int base, int size)
1564 m->dirty &= ~vm_page_bits(base, size);
1568 * Invalidates DEV_BSIZE'd chunks within a page. Both the
1569 * valid and dirty bits for the effected areas are cleared.
1574 vm_page_set_invalid(vm_page_t m, int base, int size)
1578 bits = vm_page_bits(base, size);
1581 m->object->generation++;
1585 * The kernel assumes that the invalid portions of a page contain
1586 * garbage, but such pages can be mapped into memory by user code.
1587 * When this occurs, we must zero out the non-valid portions of the
1588 * page so user code sees what it expects.
1590 * Pages are most often semi-valid when the end of a file is mapped
1591 * into memory and the file's size is not page aligned.
1594 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
1600 * Scan the valid bits looking for invalid sections that
1601 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
1602 * valid bit may be set ) have already been zerod by
1603 * vm_page_set_validclean().
1605 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
1606 if (i == (PAGE_SIZE / DEV_BSIZE) ||
1607 (m->valid & (1 << i))
1610 pmap_zero_page_area(
1613 (i - b) << DEV_BSHIFT
1621 * setvalid is TRUE when we can safely set the zero'd areas
1622 * as being valid. We can do this if there are no cache consistency
1623 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
1626 m->valid = VM_PAGE_BITS_ALL;
1630 * Is a (partial) page valid? Note that the case where size == 0
1631 * will return FALSE in the degenerate case where the page is entirely
1632 * invalid, and TRUE otherwise.
1637 vm_page_is_valid(vm_page_t m, int base, int size)
1639 int bits = vm_page_bits(base, size);
1641 if (m->valid && ((m->valid & bits) == bits))
1648 * update dirty bits from pmap/mmu. May not block.
1651 vm_page_test_dirty(vm_page_t m)
1653 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
1658 #include "opt_ddb.h"
1660 #include <sys/kernel.h>
1662 #include <ddb/ddb.h>
1664 DB_SHOW_COMMAND(page, vm_page_print_page_info)
1666 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count);
1667 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count);
1668 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count);
1669 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count);
1670 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count);
1671 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved);
1672 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min);
1673 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target);
1674 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min);
1675 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target);
1678 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
1681 db_printf("PQ_FREE:");
1682 for(i=0;i<PQ_L2_SIZE;i++) {
1683 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
1687 db_printf("PQ_CACHE:");
1688 for(i=0;i<PQ_L2_SIZE;i++) {
1689 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
1693 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
1694 vm_page_queues[PQ_ACTIVE].lcnt,
1695 vm_page_queues[PQ_INACTIVE].lcnt);