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 1.24 2004/05/20 22:42:25 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 static void vm_page_queue_init(void);
93 static void vm_page_free_wakeup(void);
94 static vm_page_t vm_page_select_cache(vm_object_t, vm_pindex_t);
95 static vm_page_t _vm_page_list_find2(int basequeue, int index);
97 static int vm_page_bucket_count; /* How big is array? */
98 static int vm_page_hash_mask; /* Mask for hash function */
99 static struct vm_page **vm_page_buckets; /* Array of buckets */
100 static volatile int vm_page_bucket_generation;
101 struct vpgqueues vm_page_queues[PQ_COUNT]; /* Array of tailq lists */
104 vm_page_queue_init(void)
108 for (i = 0; i < PQ_L2_SIZE; i++)
109 vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count;
110 for (i = 0; i < PQ_L2_SIZE; i++)
111 vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count;
113 vm_page_queues[PQ_INACTIVE].cnt = &vmstats.v_inactive_count;
114 vm_page_queues[PQ_ACTIVE].cnt = &vmstats.v_active_count;
115 vm_page_queues[PQ_HOLD].cnt = &vmstats.v_active_count;
116 /* PQ_NONE has no queue */
118 for (i = 0; i < PQ_COUNT; i++)
119 TAILQ_INIT(&vm_page_queues[i].pl);
123 * note: place in initialized data section? Is this necessary?
126 int vm_page_array_size = 0;
127 int vm_page_zero_count = 0;
128 vm_page_t vm_page_array = 0;
133 * Sets the page size, perhaps based upon the memory size.
134 * Must be called before any use of page-size dependent functions.
137 vm_set_page_size(void)
139 if (vmstats.v_page_size == 0)
140 vmstats.v_page_size = PAGE_SIZE;
141 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0)
142 panic("vm_set_page_size: page size not a power of two");
148 * Add a new page to the freelist for use by the system. New pages
149 * are added to both the head and tail of the associated free page
150 * queue in a bottom-up fashion, so both zero'd and non-zero'd page
151 * requests pull 'recent' adds (higher physical addresses) first.
153 * Must be called at splhigh().
156 vm_add_new_page(vm_paddr_t pa)
158 struct vpgqueues *vpq;
161 ++vmstats.v_page_count;
162 ++vmstats.v_free_count;
163 m = PHYS_TO_VM_PAGE(pa);
166 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK;
167 m->queue = m->pc + PQ_FREE;
169 vpq = &vm_page_queues[m->queue];
171 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
173 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
174 vpq->flipflop = 1 - vpq->flipflop;
176 vm_page_queues[m->queue].lcnt++;
183 * Initializes the resident memory module.
185 * Allocates memory for the page cells, and for the object/offset-to-page
186 * hash table headers. Each page cell is initialized and placed on the
190 vm_page_startup(vm_offset_t starta, vm_offset_t enda, vm_offset_t vaddr)
193 struct vm_page **bucket;
195 vm_paddr_t page_range;
202 /* the biggest memory array is the second group of pages */
204 vm_paddr_t biggestone, biggestsize;
212 vaddr = round_page(vaddr);
214 for (i = 0; phys_avail[i + 1]; i += 2) {
215 phys_avail[i] = round_page(phys_avail[i]);
216 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
219 for (i = 0; phys_avail[i + 1]; i += 2) {
220 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
222 if (size > biggestsize) {
230 end = phys_avail[biggestone+1];
233 * Initialize the queue headers for the free queue, the active queue
234 * and the inactive queue.
237 vm_page_queue_init();
240 * Allocate (and initialize) the hash table buckets.
242 * The number of buckets MUST BE a power of 2, and the actual value is
243 * the next power of 2 greater than the number of physical pages in
246 * We make the hash table approximately 2x the number of pages to
247 * reduce the chain length. This is about the same size using the
248 * singly-linked list as the 1x hash table we were using before
249 * using TAILQ but the chain length will be smaller.
251 * Note: This computation can be tweaked if desired.
253 vm_page_buckets = (struct vm_page **)vaddr;
254 bucket = vm_page_buckets;
255 if (vm_page_bucket_count == 0) {
256 vm_page_bucket_count = 1;
257 while (vm_page_bucket_count < atop(total))
258 vm_page_bucket_count <<= 1;
260 vm_page_bucket_count <<= 1;
261 vm_page_hash_mask = vm_page_bucket_count - 1;
264 * Validate these addresses.
266 new_end = end - vm_page_bucket_count * sizeof(struct vm_page *);
267 new_end = trunc_page(new_end);
268 mapped = round_page(vaddr);
269 vaddr = pmap_map(mapped, new_end, end,
270 VM_PROT_READ | VM_PROT_WRITE);
271 vaddr = round_page(vaddr);
272 bzero((caddr_t) mapped, vaddr - mapped);
274 for (i = 0; i < vm_page_bucket_count; i++) {
280 * Compute the number of pages of memory that will be available for
281 * use (taking into account the overhead of a page structure per
284 first_page = phys_avail[0] / PAGE_SIZE;
285 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
286 npages = (total - (page_range * sizeof(struct vm_page)) -
287 (end - new_end)) / PAGE_SIZE;
292 * Initialize the mem entry structures now, and put them in the free
295 vm_page_array = (vm_page_t) vaddr;
299 * Validate these addresses.
301 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
302 mapped = pmap_map(mapped, new_end, end,
303 VM_PROT_READ | VM_PROT_WRITE);
306 * Clear all of the page structures
308 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
309 vm_page_array_size = page_range;
312 * Construct the free queue(s) in ascending order (by physical
313 * address) so that the first 16MB of physical memory is allocated
314 * last rather than first. On large-memory machines, this avoids
315 * the exhaustion of low physical memory before isa_dmainit has run.
317 vmstats.v_page_count = 0;
318 vmstats.v_free_count = 0;
319 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
324 last_pa = phys_avail[i + 1];
325 while (pa < last_pa && npages-- > 0) {
334 * Distributes the object/offset key pair among hash buckets.
336 * NOTE: This macro depends on vm_page_bucket_count being a power of 2.
337 * This routine may not block.
339 * We try to randomize the hash based on the object to spread the pages
340 * out in the hash table without it costing us too much.
343 vm_page_hash(vm_object_t object, vm_pindex_t pindex)
345 int i = ((uintptr_t)object + pindex) ^ object->hash_rand;
347 return(i & vm_page_hash_mask);
351 * The opposite of vm_page_hold(). A page can be freed while being held,
352 * which places it on the PQ_HOLD queue. We must call vm_page_free_toq()
353 * in this case to actually free it once the hold count drops to 0.
355 * This routine must be called at splvm().
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 * Inserts the given mem entry into the object and object list.
369 * The pagetables are not updated but will presumably fault the page
370 * in if necessary, or if a kernel page the caller will at some point
371 * enter the page into the kernel's pmap. We are not allowed to block
372 * here so we *can't* do this anyway.
374 * This routine may not block.
375 * This routine must be called at splvm().
378 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
380 struct vm_page **bucket;
382 if (m->object != NULL)
383 panic("vm_page_insert: already inserted");
386 * Record the object/offset pair in this page
392 * Insert it into the object_object/offset hash table
394 bucket = &vm_page_buckets[vm_page_hash(object, pindex)];
397 vm_page_bucket_generation++;
400 * Now link into the object's list of backed pages.
402 TAILQ_INSERT_TAIL(&object->memq, m, listq);
403 object->generation++;
406 * show that the object has one more resident page.
408 object->resident_page_count++;
411 * Since we are inserting a new and possibly dirty page,
412 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
414 if (m->flags & PG_WRITEABLE)
415 vm_object_set_writeable_dirty(object);
419 * Removes the given mem entry from the object/offset-page table and
420 * the object page list, but do not invalidate/terminate the backing store.
422 * This routine must be called at splvm().
423 * The underlying pmap entry (if any) is NOT removed here.
424 * This routine may not block.
425 * The page must be BUSY.
428 vm_page_remove(vm_page_t m)
432 if (m->object == NULL)
435 if ((m->flags & PG_BUSY) == 0)
436 panic("vm_page_remove: page not busy");
439 * Basically destroy the page.
446 * Remove from the object_object/offset hash table. The object
447 * must be on the hash queue, we will panic if it isn't
449 * Note: we must NULL-out m->hnext to prevent loops in detached
450 * buffers with vm_page_lookup().
453 struct vm_page **bucket;
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++;
467 * Now remove from the object's list of backed pages.
469 TAILQ_REMOVE(&object->memq, m, listq);
472 * And show that the object has one fewer resident page.
474 object->resident_page_count--;
475 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)
549 vm_page_insert(m, new_object, new_pindex);
550 if (m->queue - m->pc == PQ_CACHE)
551 vm_page_deactivate(m);
557 * vm_page_unqueue() without any wakeup. This routine is used when a page
558 * is being moved between queues or otherwise is to remain BUSYied by the
561 * This routine must be called at splhigh().
562 * This routine may not block.
565 vm_page_unqueue_nowakeup(vm_page_t m)
567 int queue = m->queue;
568 struct vpgqueues *pq;
570 if (queue != PQ_NONE) {
571 pq = &vm_page_queues[queue];
573 TAILQ_REMOVE(&pq->pl, m, pageq);
580 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
583 * This routine must be called at splhigh().
584 * This routine may not block.
587 vm_page_unqueue(vm_page_t m)
589 int queue = m->queue;
590 struct vpgqueues *pq;
592 if (queue != PQ_NONE) {
594 pq = &vm_page_queues[queue];
595 TAILQ_REMOVE(&pq->pl, m, pageq);
598 if ((queue - m->pc) == PQ_CACHE) {
599 if (vm_paging_needed())
606 * vm_page_list_find()
608 * Find a page on the specified queue with color optimization.
610 * The page coloring optimization attempts to locate a page that does
611 * not overload other nearby pages in the object in the cpu's L1 or L2
612 * caches. We need this optimization because cpu caches tend to be
613 * physical caches, while object spaces tend to be virtual.
615 * This routine must be called at splvm().
616 * This routine may not block.
618 * Note that this routine is carefully inlined. A non-inlined version
619 * is available for outside callers but the only critical path is
620 * from within this source file.
624 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
629 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist);
631 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl);
633 m = _vm_page_list_find2(basequeue, index);
638 _vm_page_list_find2(int basequeue, int index)
642 struct vpgqueues *pq;
644 pq = &vm_page_queues[basequeue];
647 * Note that for the first loop, index+i and index-i wind up at the
648 * same place. Even though this is not totally optimal, we've already
649 * blown it by missing the cache case so we do not care.
652 for(i = PQ_L2_SIZE / 2; i > 0; --i) {
653 if ((m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl)) != NULL)
656 if ((m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl)) != NULL)
663 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
665 return(_vm_page_list_find(basequeue, index, prefer_zero));
669 * Find a page on the cache queue with color optimization. As pages
670 * might be found, but not applicable, they are deactivated. This
671 * keeps us from using potentially busy cached pages.
673 * This routine must be called at splvm().
674 * This routine may not block.
677 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex)
682 m = _vm_page_list_find(
684 (pindex + object->pg_color) & PQ_L2_MASK,
687 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
688 m->hold_count || m->wire_count)) {
689 vm_page_deactivate(m);
698 * Find a free or zero page, with specified preference. We attempt to
699 * inline the nominal case and fall back to _vm_page_select_free()
702 * This routine must be called at splvm().
703 * This routine may not block.
705 static __inline vm_page_t
706 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero)
710 m = _vm_page_list_find(
712 (pindex + object->pg_color) & PQ_L2_MASK,
721 * Allocate and return a memory cell associated with this VM object/offset
726 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain
727 * VM_ALLOC_SYSTEM greater free drain
728 * VM_ALLOC_INTERRUPT allow free list to be completely drained
729 * VM_ALLOC_ZERO advisory request for pre-zero'd page
731 * The object must be locked.
732 * This routine may not block.
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)
744 KASSERT(!vm_page_lookup(object, pindex),
745 ("vm_page_alloc: page already allocated"));
747 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
750 * The pager is allowed to eat deeper into the free page list.
752 if (curthread == pagethread)
753 page_req |= VM_ALLOC_SYSTEM;
757 if (vmstats.v_free_count > vmstats.v_free_reserved ||
758 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) ||
759 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 &&
760 vmstats.v_free_count > vmstats.v_interrupt_free_min)
763 * The free queue has sufficient free pages to take one out.
765 if (page_req & VM_ALLOC_ZERO)
766 m = vm_page_select_free(object, pindex, TRUE);
768 m = vm_page_select_free(object, pindex, FALSE);
769 } else if (page_req & VM_ALLOC_NORMAL) {
771 * Allocatable from the cache (non-interrupt only). On
772 * success, we must free the page and try again, thus
773 * ensuring that vmstats.v_*_free_min counters are replenished.
776 if (curthread->td_preempted) {
777 printf("vm_page_alloc(): warning, attempt to allocate"
778 " cache page from preempting interrupt\n");
781 m = vm_page_select_cache(object, pindex);
784 m = vm_page_select_cache(object, pindex);
787 * On succuess move the page into the free queue and loop.
790 KASSERT(m->dirty == 0,
791 ("Found dirty cache page %p", m));
793 vm_page_protect(m, VM_PROT_NONE);
799 * On failure return NULL
802 #if defined(DIAGNOSTIC)
803 if (vmstats.v_cache_count > 0)
804 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count);
806 vm_pageout_deficit++;
811 * No pages available, wakeup the pageout daemon and give up.
814 vm_pageout_deficit++;
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.
832 if (m->flags & PG_ZERO) {
833 vm_page_zero_count--;
834 m->flags = PG_ZERO | PG_BUSY;
843 KASSERT(m->dirty == 0,
844 ("vm_page_alloc: free/cache page %p was dirty", m));
847 * vm_page_insert() is safe prior to the splx(). Note also that
848 * inserting a page here does not insert it into the pmap (which
849 * could cause us to block allocating memory). We cannot block
852 vm_page_insert(m, object, pindex);
855 * Don't wakeup too often - wakeup the pageout daemon when
856 * we would be nearly out of memory.
858 if (vm_paging_needed())
866 * Block until free pages are available for allocation, called in various
867 * places before memory allocations.
875 if (curthread == pagethread) {
876 vm_pageout_pages_needed = 1;
877 tsleep(&vm_pageout_pages_needed, 0, "VMWait", 0);
879 if (!vm_pages_needed) {
881 wakeup(&vm_pages_needed);
883 tsleep(&vmstats.v_free_count, 0, "vmwait", 0);
889 * Block until free pages are available for allocation
891 * Called only in vm_fault so that processes page faulting can be
894 * Sleeps at a lower priority than vm_wait() so that vm_wait()ing
895 * processes will be able to grab memory first. Do not change
896 * this balance without careful testing first.
904 if (!vm_pages_needed) {
906 wakeup(&vm_pages_needed);
908 tsleep(&vmstats.v_free_count, 0, "pfault", 0);
913 * Put the specified page on the active list (if appropriate). Ensure
914 * that act_count is at least ACT_INIT but do not otherwise mess with it.
916 * The page queues must be locked.
917 * This routine may not block.
920 vm_page_activate(vm_page_t m)
925 if (m->queue != PQ_ACTIVE) {
926 if ((m->queue - m->pc) == PQ_CACHE)
927 mycpu->gd_cnt.v_reactivated++;
931 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
932 m->queue = PQ_ACTIVE;
933 vm_page_queues[PQ_ACTIVE].lcnt++;
934 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl,
936 if (m->act_count < ACT_INIT)
937 m->act_count = ACT_INIT;
938 vmstats.v_active_count++;
941 if (m->act_count < ACT_INIT)
942 m->act_count = ACT_INIT;
949 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
950 * routine is called when a page has been added to the cache or free
953 * This routine may not block.
954 * This routine must be called at splvm()
957 vm_page_free_wakeup(void)
960 * if pageout daemon needs pages, then tell it that there are
963 if (vm_pageout_pages_needed &&
964 vmstats.v_cache_count + vmstats.v_free_count >=
965 vmstats.v_pageout_free_min
967 wakeup(&vm_pageout_pages_needed);
968 vm_pageout_pages_needed = 0;
972 * wakeup processes that are waiting on memory if we hit a
973 * high water mark. And wakeup scheduler process if we have
974 * lots of memory. this process will swapin processes.
976 if (vm_pages_needed && !vm_page_count_min()) {
978 wakeup(&vmstats.v_free_count);
985 * Returns the given page to the PQ_FREE list,
986 * disassociating it with any VM object.
988 * Object and page must be locked prior to entry.
989 * This routine may not block.
993 vm_page_free_toq(vm_page_t m)
996 struct vpgqueues *pq;
999 mycpu->gd_cnt.v_tfree++;
1001 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
1003 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
1004 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
1006 if ((m->queue - m->pc) == PQ_FREE)
1007 panic("vm_page_free: freeing free page");
1009 panic("vm_page_free: freeing busy page");
1013 * unqueue, then remove page. Note that we cannot destroy
1014 * the page here because we do not want to call the pager's
1015 * callback routine until after we've put the page on the
1016 * appropriate free queue.
1018 vm_page_unqueue_nowakeup(m);
1022 * If fictitious remove object association and
1023 * return, otherwise delay object association removal.
1025 if ((m->flags & PG_FICTITIOUS) != 0) {
1033 if (m->wire_count != 0) {
1034 if (m->wire_count > 1) {
1036 "vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1037 m->wire_count, (long)m->pindex);
1039 panic("vm_page_free: freeing wired page");
1043 * Clear the UNMANAGED flag when freeing an unmanaged page.
1045 if (m->flags & PG_UNMANAGED) {
1046 m->flags &= ~PG_UNMANAGED;
1049 pmap_page_is_free(m);
1053 if (m->hold_count != 0) {
1054 m->flags &= ~PG_ZERO;
1057 m->queue = PQ_FREE + m->pc;
1059 pq = &vm_page_queues[m->queue];
1064 * Put zero'd pages on the end ( where we look for zero'd pages
1065 * first ) and non-zerod pages at the head.
1067 if (m->flags & PG_ZERO) {
1068 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
1069 ++vm_page_zero_count;
1071 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1074 vm_page_free_wakeup();
1079 * vm_page_unmanage()
1081 * Prevent PV management from being done on the page. The page is
1082 * removed from the paging queues as if it were wired, and as a
1083 * consequence of no longer being managed the pageout daemon will not
1084 * touch it (since there is no way to locate the pte mappings for the
1085 * page). madvise() calls that mess with the pmap will also no longer
1086 * operate on the page.
1088 * Beyond that the page is still reasonably 'normal'. Freeing the page
1089 * will clear the flag.
1091 * This routine is used by OBJT_PHYS objects - objects using unswappable
1092 * physical memory as backing store rather then swap-backed memory and
1093 * will eventually be extended to support 4MB unmanaged physical
1097 vm_page_unmanage(vm_page_t m)
1102 if ((m->flags & PG_UNMANAGED) == 0) {
1103 if (m->wire_count == 0)
1106 vm_page_flag_set(m, PG_UNMANAGED);
1111 * Mark this page as wired down by yet another map, removing it from
1112 * paging queues as necessary.
1114 * The page queues must be locked.
1115 * This routine may not block.
1118 vm_page_wire(vm_page_t m)
1123 * Only bump the wire statistics if the page is not already wired,
1124 * and only unqueue the page if it is on some queue (if it is unmanaged
1125 * it is already off the queues).
1128 if (m->wire_count == 0) {
1129 if ((m->flags & PG_UNMANAGED) == 0)
1131 vmstats.v_wire_count++;
1134 KASSERT(m->wire_count != 0,
1135 ("vm_page_wire: wire_count overflow m=%p", m));
1138 vm_page_flag_set(m, PG_MAPPED);
1142 * Release one wiring of this page, potentially enabling it to be paged again.
1144 * Many pages placed on the inactive queue should actually go
1145 * into the cache, but it is difficult to figure out which. What
1146 * we do instead, if the inactive target is well met, is to put
1147 * clean pages at the head of the inactive queue instead of the tail.
1148 * This will cause them to be moved to the cache more quickly and
1149 * if not actively re-referenced, freed more quickly. If we just
1150 * stick these pages at the end of the inactive queue, heavy filesystem
1151 * meta-data accesses can cause an unnecessary paging load on memory bound
1152 * processes. This optimization causes one-time-use metadata to be
1153 * reused more quickly.
1155 * BUT, if we are in a low-memory situation we have no choice but to
1156 * put clean pages on the cache queue.
1158 * A number of routines use vm_page_unwire() to guarantee that the page
1159 * will go into either the inactive or active queues, and will NEVER
1160 * be placed in the cache - for example, just after dirtying a page.
1161 * dirty pages in the cache are not allowed.
1163 * The page queues must be locked.
1164 * This routine may not block.
1167 vm_page_unwire(vm_page_t m, int activate)
1173 if (m->wire_count > 0) {
1175 if (m->wire_count == 0) {
1176 vmstats.v_wire_count--;
1177 if (m->flags & PG_UNMANAGED) {
1179 } else if (activate) {
1180 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1181 m->queue = PQ_ACTIVE;
1182 vm_page_queues[PQ_ACTIVE].lcnt++;
1183 vmstats.v_active_count++;
1185 vm_page_flag_clear(m, PG_WINATCFLS);
1186 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1187 m->queue = PQ_INACTIVE;
1188 vm_page_queues[PQ_INACTIVE].lcnt++;
1189 vmstats.v_inactive_count++;
1193 panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
1200 * Move the specified page to the inactive queue. If the page has
1201 * any associated swap, the swap is deallocated.
1203 * Normally athead is 0 resulting in LRU operation. athead is set
1204 * to 1 if we want this page to be 'as if it were placed in the cache',
1205 * except without unmapping it from the process address space.
1207 * This routine may not block.
1209 static __inline void
1210 _vm_page_deactivate(vm_page_t m, int athead)
1215 * Ignore if already inactive.
1217 if (m->queue == PQ_INACTIVE)
1221 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1222 if ((m->queue - m->pc) == PQ_CACHE)
1223 mycpu->gd_cnt.v_reactivated++;
1224 vm_page_flag_clear(m, PG_WINATCFLS);
1227 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1229 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1230 m->queue = PQ_INACTIVE;
1231 vm_page_queues[PQ_INACTIVE].lcnt++;
1232 vmstats.v_inactive_count++;
1238 vm_page_deactivate(vm_page_t m)
1240 _vm_page_deactivate(m, 0);
1244 * vm_page_try_to_cache:
1246 * Returns 0 on failure, 1 on success
1249 vm_page_try_to_cache(vm_page_t m)
1251 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1252 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1255 vm_page_test_dirty(m);
1263 * Attempt to free the page. If we cannot free it, we do nothing.
1264 * 1 is returned on success, 0 on failure.
1267 vm_page_try_to_free(vm_page_t m)
1269 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1270 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1273 vm_page_test_dirty(m);
1277 vm_page_protect(m, VM_PROT_NONE);
1285 * Put the specified page onto the page cache queue (if appropriate).
1287 * This routine may not block.
1290 vm_page_cache(vm_page_t m)
1294 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
1295 m->wire_count || m->hold_count) {
1296 printf("vm_page_cache: attempting to cache busy/held page\n");
1299 if ((m->queue - m->pc) == PQ_CACHE)
1303 * Remove all pmaps and indicate that the page is not
1304 * writeable or mapped.
1307 vm_page_protect(m, VM_PROT_NONE);
1308 if (m->dirty != 0) {
1309 panic("vm_page_cache: caching a dirty page, pindex: %ld",
1313 vm_page_unqueue_nowakeup(m);
1314 m->queue = PQ_CACHE + m->pc;
1315 vm_page_queues[m->queue].lcnt++;
1316 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq);
1317 vmstats.v_cache_count++;
1318 vm_page_free_wakeup();
1323 * vm_page_dontneed()
1325 * Cache, deactivate, or do nothing as appropriate. This routine
1326 * is typically used by madvise() MADV_DONTNEED.
1328 * Generally speaking we want to move the page into the cache so
1329 * it gets reused quickly. However, this can result in a silly syndrome
1330 * due to the page recycling too quickly. Small objects will not be
1331 * fully cached. On the otherhand, if we move the page to the inactive
1332 * queue we wind up with a problem whereby very large objects
1333 * unnecessarily blow away our inactive and cache queues.
1335 * The solution is to move the pages based on a fixed weighting. We
1336 * either leave them alone, deactivate them, or move them to the cache,
1337 * where moving them to the cache has the highest weighting.
1338 * By forcing some pages into other queues we eventually force the
1339 * system to balance the queues, potentially recovering other unrelated
1340 * space from active. The idea is to not force this to happen too
1344 vm_page_dontneed(vm_page_t m)
1346 static int dnweight;
1353 * occassionally leave the page alone
1356 if ((dnw & 0x01F0) == 0 ||
1357 m->queue == PQ_INACTIVE ||
1358 m->queue - m->pc == PQ_CACHE
1360 if (m->act_count >= ACT_INIT)
1366 vm_page_test_dirty(m);
1368 if (m->dirty || (dnw & 0x0070) == 0) {
1370 * Deactivate the page 3 times out of 32.
1375 * Cache the page 28 times out of every 32. Note that
1376 * the page is deactivated instead of cached, but placed
1377 * at the head of the queue instead of the tail.
1381 _vm_page_deactivate(m, head);
1385 * Grab a page, blocking if it is busy and allocating a page if necessary.
1386 * A busy page is returned or NULL.
1388 * If VM_ALLOC_RETRY is specified VM_ALLOC_NORMAL must also be specified.
1389 * If VM_ALLOC_RETRY is not specified
1391 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
1392 * always returned if we had blocked.
1393 * This routine will never return NULL if VM_ALLOC_RETRY is set.
1394 * This routine may not be called from an interrupt.
1395 * The returned page may not be entirely valid.
1397 * This routine may be called from mainline code without spl protection and
1398 * be guarenteed a busied page associated with the object at the specified
1402 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
1407 KKASSERT(allocflags &
1408 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
1411 if ((m = vm_page_lookup(object, pindex)) != NULL) {
1412 if (m->busy || (m->flags & PG_BUSY)) {
1413 generation = object->generation;
1415 while ((object->generation == generation) &&
1416 (m->busy || (m->flags & PG_BUSY))) {
1417 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
1418 tsleep(m, 0, "pgrbwt", 0);
1419 if ((allocflags & VM_ALLOC_RETRY) == 0) {
1430 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1433 if ((allocflags & VM_ALLOC_RETRY) == 0)
1443 * Mapping function for valid bits or for dirty bits in
1444 * a page. May not block.
1446 * Inputs are required to range within a page.
1449 vm_page_bits(int base, int size)
1455 base + size <= PAGE_SIZE,
1456 ("vm_page_bits: illegal base/size %d/%d", base, size)
1459 if (size == 0) /* handle degenerate case */
1462 first_bit = base >> DEV_BSHIFT;
1463 last_bit = (base + size - 1) >> DEV_BSHIFT;
1465 return ((2 << last_bit) - (1 << first_bit));
1469 * Sets portions of a page valid and clean. The arguments are expected
1470 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1471 * of any partial chunks touched by the range. The invalid portion of
1472 * such chunks will be zero'd.
1474 * This routine may not block.
1476 * (base + size) must be less then or equal to PAGE_SIZE.
1479 vm_page_set_validclean(vm_page_t m, int base, int size)
1485 if (size == 0) /* handle degenerate case */
1489 * If the base is not DEV_BSIZE aligned and the valid
1490 * bit is clear, we have to zero out a portion of the
1494 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
1495 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
1497 pmap_zero_page_area(
1505 * If the ending offset is not DEV_BSIZE aligned and the
1506 * valid bit is clear, we have to zero out a portion of
1510 endoff = base + size;
1512 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
1513 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
1515 pmap_zero_page_area(
1518 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
1523 * Set valid, clear dirty bits. If validating the entire
1524 * page we can safely clear the pmap modify bit. We also
1525 * use this opportunity to clear the PG_NOSYNC flag. If a process
1526 * takes a write fault on a MAP_NOSYNC memory area the flag will
1529 * We set valid bits inclusive of any overlap, but we can only
1530 * clear dirty bits for DEV_BSIZE chunks that are fully within
1534 pagebits = vm_page_bits(base, size);
1535 m->valid |= pagebits;
1537 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
1538 frag = DEV_BSIZE - frag;
1544 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
1546 m->dirty &= ~pagebits;
1547 if (base == 0 && size == PAGE_SIZE) {
1548 pmap_clear_modify(m);
1549 vm_page_flag_clear(m, PG_NOSYNC);
1554 vm_page_clear_dirty(vm_page_t m, int base, int size)
1556 m->dirty &= ~vm_page_bits(base, size);
1560 * Invalidates DEV_BSIZE'd chunks within a page. Both the
1561 * valid and dirty bits for the effected areas are cleared.
1566 vm_page_set_invalid(vm_page_t m, int base, int size)
1570 bits = vm_page_bits(base, size);
1573 m->object->generation++;
1577 * The kernel assumes that the invalid portions of a page contain
1578 * garbage, but such pages can be mapped into memory by user code.
1579 * When this occurs, we must zero out the non-valid portions of the
1580 * page so user code sees what it expects.
1582 * Pages are most often semi-valid when the end of a file is mapped
1583 * into memory and the file's size is not page aligned.
1586 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
1592 * Scan the valid bits looking for invalid sections that
1593 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
1594 * valid bit may be set ) have already been zerod by
1595 * vm_page_set_validclean().
1597 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
1598 if (i == (PAGE_SIZE / DEV_BSIZE) ||
1599 (m->valid & (1 << i))
1602 pmap_zero_page_area(
1605 (i - b) << DEV_BSHIFT
1613 * setvalid is TRUE when we can safely set the zero'd areas
1614 * as being valid. We can do this if there are no cache consistency
1615 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
1618 m->valid = VM_PAGE_BITS_ALL;
1622 * Is a (partial) page valid? Note that the case where size == 0
1623 * will return FALSE in the degenerate case where the page is entirely
1624 * invalid, and TRUE otherwise.
1629 vm_page_is_valid(vm_page_t m, int base, int size)
1631 int bits = vm_page_bits(base, size);
1633 if (m->valid && ((m->valid & bits) == bits))
1640 * update dirty bits from pmap/mmu. May not block.
1643 vm_page_test_dirty(vm_page_t m)
1645 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
1650 #include "opt_ddb.h"
1652 #include <sys/kernel.h>
1654 #include <ddb/ddb.h>
1656 DB_SHOW_COMMAND(page, vm_page_print_page_info)
1658 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count);
1659 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count);
1660 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count);
1661 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count);
1662 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count);
1663 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved);
1664 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min);
1665 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target);
1666 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min);
1667 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target);
1670 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
1673 db_printf("PQ_FREE:");
1674 for(i=0;i<PQ_L2_SIZE;i++) {
1675 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
1679 db_printf("PQ_CACHE:");
1680 for(i=0;i<PQ_L2_SIZE;i++) {
1681 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
1685 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
1686 vm_page_queues[PQ_ACTIVE].lcnt,
1687 vm_page_queues[PQ_INACTIVE].lcnt);