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 1.40 2008/08/25 17:01:42 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 <machine/md_var.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 struct vpgqueues vm_page_queues[PQ_COUNT]; /* Array of tailq lists */
101 #define ASSERT_IN_CRIT_SECTION() KKASSERT(crit_test(curthread));
103 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare,
104 vm_pindex_t, pindex);
107 vm_page_queue_init(void)
111 for (i = 0; i < PQ_L2_SIZE; i++)
112 vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count;
113 for (i = 0; i < PQ_L2_SIZE; i++)
114 vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count;
116 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 /* PQ_NONE has no queue */
121 for (i = 0; i < PQ_COUNT; i++)
122 TAILQ_INIT(&vm_page_queues[i].pl);
126 * note: place in initialized data section? Is this necessary?
129 int vm_page_array_size = 0;
130 int vm_page_zero_count = 0;
131 vm_page_t vm_page_array = 0;
136 * Sets the page size, perhaps based upon the memory size.
137 * Must be called before any use of page-size dependent functions.
140 vm_set_page_size(void)
142 if (vmstats.v_page_size == 0)
143 vmstats.v_page_size = PAGE_SIZE;
144 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0)
145 panic("vm_set_page_size: page size not a power of two");
151 * Add a new page to the freelist for use by the system. New pages
152 * are added to both the head and tail of the associated free page
153 * queue in a bottom-up fashion, so both zero'd and non-zero'd page
154 * requests pull 'recent' adds (higher physical addresses) first.
156 * Must be called in a critical section.
159 vm_add_new_page(vm_paddr_t pa)
161 struct vpgqueues *vpq;
164 ++vmstats.v_page_count;
165 ++vmstats.v_free_count;
166 m = PHYS_TO_VM_PAGE(pa);
169 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK;
170 m->queue = m->pc + PQ_FREE;
171 KKASSERT(m->dirty == 0);
173 vpq = &vm_page_queues[m->queue];
175 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
177 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
178 vpq->flipflop = 1 - vpq->flipflop;
180 vm_page_queues[m->queue].lcnt++;
187 * Initializes the resident memory module.
189 * Allocates memory for the page cells, and for the object/offset-to-page
190 * hash table headers. Each page cell is initialized and placed on the
193 * starta/enda represents the range of physical memory addresses available
194 * for use (skipping memory already used by the kernel), subject to
195 * phys_avail[]. Note that phys_avail[] has already mapped out memory
196 * already in use by the kernel.
199 vm_page_startup(vm_offset_t vaddr)
203 vm_paddr_t page_range;
210 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];
236 end = trunc_page(end);
239 * Initialize the queue headers for the free queue, the active queue
240 * and the inactive queue.
243 vm_page_queue_init();
245 /* VKERNELs don't support minidumps and as such don't need vm_page_dump */
246 #if !defined(_KERNEL_VIRTUAL)
248 * Allocate a bitmap to indicate that a random physical page
249 * needs to be included in a minidump.
251 * The amd64 port needs this to indicate which direct map pages
252 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
254 * However, i386 still needs this workspace internally within the
255 * minidump code. In theory, they are not needed on i386, but are
256 * included should the sf_buf code decide to use them.
258 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE;
259 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
260 end -= vm_page_dump_size;
261 vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size,
262 VM_PROT_READ | VM_PROT_WRITE);
263 bzero((void *)vm_page_dump, vm_page_dump_size);
267 * Compute the number of pages of memory that will be available for
268 * use (taking into account the overhead of a page structure per
271 first_page = phys_avail[0] / PAGE_SIZE;
272 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
273 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE;
276 * Initialize the mem entry structures now, and put them in the free
279 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
280 mapped = pmap_map(&vaddr, new_end, end,
281 VM_PROT_READ | VM_PROT_WRITE);
282 vm_page_array = (vm_page_t)mapped;
286 * since pmap_map on amd64 returns stuff out of a direct-map region,
287 * we have to manually add these pages to the minidump tracking so
288 * that they can be dumped, including the vm_page_array.
290 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
295 * Clear all of the page structures
297 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
298 vm_page_array_size = page_range;
301 * Construct the free queue(s) in ascending order (by physical
302 * address) so that the first 16MB of physical memory is allocated
303 * last rather than first. On large-memory machines, this avoids
304 * the exhaustion of low physical memory before isa_dmainit has run.
306 vmstats.v_page_count = 0;
307 vmstats.v_free_count = 0;
308 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
313 last_pa = phys_avail[i + 1];
314 while (pa < last_pa && npages-- > 0) {
323 * Scan comparison function for Red-Black tree scans. An inclusive
324 * (start,end) is expected. Other fields are not used.
327 rb_vm_page_scancmp(struct vm_page *p, void *data)
329 struct rb_vm_page_scan_info *info = data;
331 if (p->pindex < info->start_pindex)
333 if (p->pindex > info->end_pindex)
339 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2)
341 if (p1->pindex < p2->pindex)
343 if (p1->pindex > p2->pindex)
349 * The opposite of vm_page_hold(). A page can be freed while being held,
350 * which places it on the PQ_HOLD queue. We must call vm_page_free_toq()
351 * in this case to actually free it once the hold count drops to 0.
353 * This routine must be called at splvm().
356 vm_page_unhold(vm_page_t mem)
359 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
360 if (mem->hold_count == 0 && mem->queue == PQ_HOLD) {
362 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 with a critical section held.
378 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
380 ASSERT_IN_CRIT_SECTION();
381 if (m->object != NULL)
382 panic("vm_page_insert: already inserted");
385 * Record the object/offset pair in this page
391 * Insert it into the object.
393 vm_page_rb_tree_RB_INSERT(&object->rb_memq, m);
394 object->generation++;
397 * show that the object has one more resident page.
399 object->resident_page_count++;
402 * Since we are inserting a new and possibly dirty page,
403 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
405 if ((m->valid & m->dirty) || (m->flags & PG_WRITEABLE))
406 vm_object_set_writeable_dirty(object);
410 * Removes the given vm_page_t from the global (object,index) hash table
411 * and from the object's memq.
413 * The underlying pmap entry (if any) is NOT removed here.
414 * This routine may not block.
416 * The page must be BUSY and will remain BUSY on return. No spl needs to be
417 * held on call to this routine.
419 * note: FreeBSD side effect was to unbusy the page on return. We leave
423 vm_page_remove(vm_page_t m)
428 if (m->object == NULL) {
433 if ((m->flags & PG_BUSY) == 0)
434 panic("vm_page_remove: page not busy");
439 * Remove the page from the object and update the object.
441 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m);
442 object->resident_page_count--;
443 object->generation++;
450 * Locate and return the page at (object, pindex), or NULL if the
451 * page could not be found.
453 * This routine will operate properly without spl protection, but
454 * the returned page could be in flux if it is busy. Because an
455 * interrupt can race a caller's busy check (unbusying and freeing the
456 * page we return before the caller is able to check the busy bit),
457 * the caller should generally call this routine with a critical
460 * Callers may call this routine without spl protection if they know
461 * 'for sure' that the page will not be ripped out from under them
465 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
470 * Search the hash table for this object/offset pair
473 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
475 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex));
482 * Move the given memory entry from its current object to the specified
483 * target object/offset.
485 * The object must be locked.
486 * This routine may not block.
488 * Note: This routine will raise itself to splvm(), the caller need not.
490 * Note: Swap associated with the page must be invalidated by the move. We
491 * have to do this for several reasons: (1) we aren't freeing the
492 * page, (2) we are dirtying the page, (3) the VM system is probably
493 * moving the page from object A to B, and will then later move
494 * the backing store from A to B and we can't have a conflict.
496 * Note: We *always* dirty the page. It is necessary both for the
497 * fact that we moved it, and because we may be invalidating
498 * swap. If the page is on the cache, we have to deactivate it
499 * or vm_page_dirty() will panic. Dirty pages are not allowed
503 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
507 vm_page_insert(m, new_object, new_pindex);
508 if (m->queue - m->pc == PQ_CACHE)
509 vm_page_deactivate(m);
516 * vm_page_unqueue() without any wakeup. This routine is used when a page
517 * is being moved between queues or otherwise is to remain BUSYied by the
520 * This routine must be called at splhigh().
521 * This routine may not block.
524 vm_page_unqueue_nowakeup(vm_page_t m)
526 int queue = m->queue;
527 struct vpgqueues *pq;
529 if (queue != PQ_NONE) {
530 pq = &vm_page_queues[queue];
532 TAILQ_REMOVE(&pq->pl, m, pageq);
539 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
542 * This routine must be called at splhigh().
543 * This routine may not block.
546 vm_page_unqueue(vm_page_t m)
548 int queue = m->queue;
549 struct vpgqueues *pq;
551 if (queue != PQ_NONE) {
553 pq = &vm_page_queues[queue];
554 TAILQ_REMOVE(&pq->pl, m, pageq);
557 if ((queue - m->pc) == PQ_CACHE || (queue - m->pc) == PQ_FREE)
563 * vm_page_list_find()
565 * Find a page on the specified queue with color optimization.
567 * The page coloring optimization attempts to locate a page that does
568 * not overload other nearby pages in the object in the cpu's L1 or L2
569 * caches. We need this optimization because cpu caches tend to be
570 * physical caches, while object spaces tend to be virtual.
572 * This routine must be called at splvm().
573 * This routine may not block.
575 * Note that this routine is carefully inlined. A non-inlined version
576 * is available for outside callers but the only critical path is
577 * from within this source file.
581 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
586 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist);
588 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl);
590 m = _vm_page_list_find2(basequeue, index);
595 _vm_page_list_find2(int basequeue, int index)
599 struct vpgqueues *pq;
601 pq = &vm_page_queues[basequeue];
604 * Note that for the first loop, index+i and index-i wind up at the
605 * same place. Even though this is not totally optimal, we've already
606 * blown it by missing the cache case so we do not care.
609 for(i = PQ_L2_SIZE / 2; i > 0; --i) {
610 if ((m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl)) != NULL)
613 if ((m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl)) != NULL)
620 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
622 return(_vm_page_list_find(basequeue, index, prefer_zero));
626 * Find a page on the cache queue with color optimization. As pages
627 * might be found, but not applicable, they are deactivated. This
628 * keeps us from using potentially busy cached pages.
630 * This routine must be called with a critical section held.
631 * This routine may not block.
634 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex)
639 m = _vm_page_list_find(
641 (pindex + object->pg_color) & PQ_L2_MASK,
644 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
645 m->hold_count || m->wire_count)) {
646 vm_page_deactivate(m);
655 * Find a free or zero page, with specified preference. We attempt to
656 * inline the nominal case and fall back to _vm_page_select_free()
659 * This routine must be called with a critical section held.
660 * This routine may not block.
662 static __inline vm_page_t
663 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero)
667 m = _vm_page_list_find(
669 (pindex + object->pg_color) & PQ_L2_MASK,
678 * Allocate and return a memory cell associated with this VM object/offset
683 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain
684 * VM_ALLOC_QUICK like normal but cannot use cache
685 * VM_ALLOC_SYSTEM greater free drain
686 * VM_ALLOC_INTERRUPT allow free list to be completely drained
687 * VM_ALLOC_ZERO advisory request for pre-zero'd page
689 * The object must be locked.
690 * This routine may not block.
691 * The returned page will be marked PG_BUSY
693 * Additional special handling is required when called from an interrupt
694 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache
698 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
702 KKASSERT(object != NULL);
703 KASSERT(!vm_page_lookup(object, pindex),
704 ("vm_page_alloc: page already allocated"));
706 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK|
707 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
710 * Certain system threads (pageout daemon, buf_daemon's) are
711 * allowed to eat deeper into the free page list.
713 if (curthread->td_flags & TDF_SYSTHREAD)
714 page_req |= VM_ALLOC_SYSTEM;
718 if (vmstats.v_free_count > vmstats.v_free_reserved ||
719 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) ||
720 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 &&
721 vmstats.v_free_count > vmstats.v_interrupt_free_min)
724 * The free queue has sufficient free pages to take one out.
726 if (page_req & VM_ALLOC_ZERO)
727 m = vm_page_select_free(object, pindex, TRUE);
729 m = vm_page_select_free(object, pindex, FALSE);
730 } else if (page_req & VM_ALLOC_NORMAL) {
732 * Allocatable from the cache (non-interrupt only). On
733 * success, we must free the page and try again, thus
734 * ensuring that vmstats.v_*_free_min counters are replenished.
737 if (curthread->td_preempted) {
738 kprintf("vm_page_alloc(): warning, attempt to allocate"
739 " cache page from preempting interrupt\n");
742 m = vm_page_select_cache(object, pindex);
745 m = vm_page_select_cache(object, pindex);
748 * On success move the page into the free queue and loop.
751 KASSERT(m->dirty == 0,
752 ("Found dirty cache page %p", m));
754 vm_page_protect(m, VM_PROT_NONE);
760 * On failure return NULL
763 #if defined(DIAGNOSTIC)
764 if (vmstats.v_cache_count > 0)
765 kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count);
767 vm_pageout_deficit++;
772 * No pages available, wakeup the pageout daemon and give up.
775 vm_pageout_deficit++;
781 * Good page found. The page has not yet been busied. We are in
782 * a critical section.
784 KASSERT(m != NULL, ("vm_page_alloc(): missing page on free queue\n"));
785 KASSERT(m->dirty == 0,
786 ("vm_page_alloc: free/cache page %p was dirty", m));
789 * Remove from free queue
791 vm_page_unqueue_nowakeup(m);
794 * Initialize structure. Only the PG_ZERO flag is inherited. Set
797 if (m->flags & PG_ZERO) {
798 vm_page_zero_count--;
799 m->flags = PG_ZERO | PG_BUSY;
810 * vm_page_insert() is safe prior to the crit_exit(). Note also that
811 * inserting a page here does not insert it into the pmap (which
812 * could cause us to block allocating memory). We cannot block
815 vm_page_insert(m, object, pindex);
818 * Don't wakeup too often - wakeup the pageout daemon when
819 * we would be nearly out of memory.
826 * A PG_BUSY page is returned.
832 * Block until free pages are available for allocation, called in various
833 * places before memory allocations.
839 if (curthread == pagethread) {
840 vm_pageout_pages_needed = 1;
841 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo);
843 if (vm_pages_needed == 0) {
845 wakeup(&vm_pages_needed);
847 tsleep(&vmstats.v_free_count, 0, "vmwait", timo);
853 * Block until free pages are available for allocation
855 * Called only in vm_fault so that processes page faulting can be
862 if (vm_pages_needed == 0) {
864 wakeup(&vm_pages_needed);
866 tsleep(&vmstats.v_free_count, 0, "pfault", 0);
871 * Put the specified page on the active list (if appropriate). Ensure
872 * that act_count is at least ACT_INIT but do not otherwise mess with it.
874 * The page queues must be locked.
875 * This routine may not block.
878 vm_page_activate(vm_page_t m)
881 if (m->queue != PQ_ACTIVE) {
882 if ((m->queue - m->pc) == PQ_CACHE)
883 mycpu->gd_cnt.v_reactivated++;
887 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
888 m->queue = PQ_ACTIVE;
889 vm_page_queues[PQ_ACTIVE].lcnt++;
890 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl,
892 if (m->act_count < ACT_INIT)
893 m->act_count = ACT_INIT;
894 vmstats.v_active_count++;
897 if (m->act_count < ACT_INIT)
898 m->act_count = ACT_INIT;
904 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
905 * routine is called when a page has been added to the cache or free
908 * This routine may not block.
909 * This routine must be called at splvm()
912 vm_page_free_wakeup(void)
915 * if pageout daemon needs pages, then tell it that there are
918 if (vm_pageout_pages_needed &&
919 vmstats.v_cache_count + vmstats.v_free_count >=
920 vmstats.v_pageout_free_min
922 wakeup(&vm_pageout_pages_needed);
923 vm_pageout_pages_needed = 0;
927 * wakeup processes that are waiting on memory if we hit a
928 * high water mark. And wakeup scheduler process if we have
929 * lots of memory. this process will swapin processes.
931 if (vm_pages_needed && !vm_page_count_min(0)) {
933 wakeup(&vmstats.v_free_count);
940 * Returns the given page to the PQ_FREE list, disassociating it with
943 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on
944 * return (the page will have been freed). No particular spl is required
947 * This routine may not block.
950 vm_page_free_toq(vm_page_t m)
952 struct vpgqueues *pq;
955 mycpu->gd_cnt.v_tfree++;
957 KKASSERT((m->flags & PG_MAPPED) == 0);
959 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
961 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
962 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
964 if ((m->queue - m->pc) == PQ_FREE)
965 panic("vm_page_free: freeing free page");
967 panic("vm_page_free: freeing busy page");
971 * unqueue, then remove page. Note that we cannot destroy
972 * the page here because we do not want to call the pager's
973 * callback routine until after we've put the page on the
974 * appropriate free queue.
976 vm_page_unqueue_nowakeup(m);
980 * No further management of fictitious pages occurs beyond object
983 if ((m->flags & PG_FICTITIOUS) != 0) {
992 if (m->wire_count != 0) {
993 if (m->wire_count > 1) {
995 "vm_page_free: invalid wire count (%d), pindex: 0x%lx",
996 m->wire_count, (long)m->pindex);
998 panic("vm_page_free: freeing wired page");
1002 * Clear the UNMANAGED flag when freeing an unmanaged page.
1004 if (m->flags & PG_UNMANAGED) {
1005 m->flags &= ~PG_UNMANAGED;
1008 if (m->hold_count != 0) {
1009 m->flags &= ~PG_ZERO;
1012 m->queue = PQ_FREE + m->pc;
1014 pq = &vm_page_queues[m->queue];
1019 * Put zero'd pages on the end ( where we look for zero'd pages
1020 * first ) and non-zerod pages at the head.
1022 if (m->flags & PG_ZERO) {
1023 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
1024 ++vm_page_zero_count;
1026 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1029 vm_page_free_wakeup();
1034 * vm_page_unmanage()
1036 * Prevent PV management from being done on the page. The page is
1037 * removed from the paging queues as if it were wired, and as a
1038 * consequence of no longer being managed the pageout daemon will not
1039 * touch it (since there is no way to locate the pte mappings for the
1040 * page). madvise() calls that mess with the pmap will also no longer
1041 * operate on the page.
1043 * Beyond that the page is still reasonably 'normal'. Freeing the page
1044 * will clear the flag.
1046 * This routine is used by OBJT_PHYS objects - objects using unswappable
1047 * physical memory as backing store rather then swap-backed memory and
1048 * will eventually be extended to support 4MB unmanaged physical
1051 * Must be called with a critical section held.
1054 vm_page_unmanage(vm_page_t m)
1056 ASSERT_IN_CRIT_SECTION();
1057 if ((m->flags & PG_UNMANAGED) == 0) {
1058 if (m->wire_count == 0)
1061 vm_page_flag_set(m, PG_UNMANAGED);
1065 * Mark this page as wired down by yet another map, removing it from
1066 * paging queues as necessary.
1068 * The page queues must be locked.
1069 * This routine may not block.
1072 vm_page_wire(vm_page_t m)
1075 * Only bump the wire statistics if the page is not already wired,
1076 * and only unqueue the page if it is on some queue (if it is unmanaged
1077 * it is already off the queues). Don't do anything with fictitious
1078 * pages because they are always wired.
1081 if ((m->flags & PG_FICTITIOUS) == 0) {
1082 if (m->wire_count == 0) {
1083 if ((m->flags & PG_UNMANAGED) == 0)
1085 vmstats.v_wire_count++;
1088 KASSERT(m->wire_count != 0,
1089 ("vm_page_wire: wire_count overflow m=%p", m));
1095 * Release one wiring of this page, potentially enabling it to be paged again.
1097 * Many pages placed on the inactive queue should actually go
1098 * into the cache, but it is difficult to figure out which. What
1099 * we do instead, if the inactive target is well met, is to put
1100 * clean pages at the head of the inactive queue instead of the tail.
1101 * This will cause them to be moved to the cache more quickly and
1102 * if not actively re-referenced, freed more quickly. If we just
1103 * stick these pages at the end of the inactive queue, heavy filesystem
1104 * meta-data accesses can cause an unnecessary paging load on memory bound
1105 * processes. This optimization causes one-time-use metadata to be
1106 * reused more quickly.
1108 * BUT, if we are in a low-memory situation we have no choice but to
1109 * put clean pages on the cache queue.
1111 * A number of routines use vm_page_unwire() to guarantee that the page
1112 * will go into either the inactive or active queues, and will NEVER
1113 * be placed in the cache - for example, just after dirtying a page.
1114 * dirty pages in the cache are not allowed.
1116 * The page queues must be locked.
1117 * This routine may not block.
1120 vm_page_unwire(vm_page_t m, int activate)
1123 if (m->flags & PG_FICTITIOUS) {
1125 } else if (m->wire_count <= 0) {
1126 panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
1128 if (--m->wire_count == 0) {
1129 --vmstats.v_wire_count;
1130 if (m->flags & PG_UNMANAGED) {
1132 } else if (activate) {
1134 &vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1135 m->queue = PQ_ACTIVE;
1136 vm_page_queues[PQ_ACTIVE].lcnt++;
1137 vmstats.v_active_count++;
1139 vm_page_flag_clear(m, PG_WINATCFLS);
1141 &vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1142 m->queue = PQ_INACTIVE;
1143 vm_page_queues[PQ_INACTIVE].lcnt++;
1144 vmstats.v_inactive_count++;
1153 * Move the specified page to the inactive queue. If the page has
1154 * any associated swap, the swap is deallocated.
1156 * Normally athead is 0 resulting in LRU operation. athead is set
1157 * to 1 if we want this page to be 'as if it were placed in the cache',
1158 * except without unmapping it from the process address space.
1160 * This routine may not block.
1162 static __inline void
1163 _vm_page_deactivate(vm_page_t m, int athead)
1166 * Ignore if already inactive.
1168 if (m->queue == PQ_INACTIVE)
1171 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1172 if ((m->queue - m->pc) == PQ_CACHE)
1173 mycpu->gd_cnt.v_reactivated++;
1174 vm_page_flag_clear(m, PG_WINATCFLS);
1177 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1179 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1180 m->queue = PQ_INACTIVE;
1181 vm_page_queues[PQ_INACTIVE].lcnt++;
1182 vmstats.v_inactive_count++;
1187 vm_page_deactivate(vm_page_t m)
1190 _vm_page_deactivate(m, 0);
1195 * vm_page_try_to_cache:
1197 * Returns 0 on failure, 1 on success
1200 vm_page_try_to_cache(vm_page_t m)
1203 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1204 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1208 vm_page_test_dirty(m);
1219 * Attempt to free the page. If we cannot free it, we do nothing.
1220 * 1 is returned on success, 0 on failure.
1223 vm_page_try_to_free(vm_page_t m)
1226 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1227 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1231 vm_page_test_dirty(m);
1237 vm_page_protect(m, VM_PROT_NONE);
1246 * Put the specified page onto the page cache queue (if appropriate).
1248 * This routine may not block.
1251 vm_page_cache(vm_page_t m)
1253 ASSERT_IN_CRIT_SECTION();
1255 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
1256 m->wire_count || m->hold_count) {
1257 kprintf("vm_page_cache: attempting to cache busy/held page\n");
1262 * Already in the cache (and thus not mapped)
1264 if ((m->queue - m->pc) == PQ_CACHE) {
1265 KKASSERT((m->flags & PG_MAPPED) == 0);
1270 * Caller is required to test m->dirty, but note that the act of
1271 * removing the page from its maps can cause it to become dirty
1272 * on an SMP system due to another cpu running in usermode.
1275 panic("vm_page_cache: caching a dirty page, pindex: %ld",
1280 * Remove all pmaps and indicate that the page is not
1281 * writeable or mapped. Our vm_page_protect() call may
1282 * have blocked (especially w/ VM_PROT_NONE), so recheck
1286 vm_page_protect(m, VM_PROT_NONE);
1288 if ((m->flags & (PG_BUSY|PG_UNMANAGED|PG_MAPPED)) || m->busy ||
1289 m->wire_count || m->hold_count) {
1291 } else if (m->dirty) {
1292 vm_page_deactivate(m);
1294 vm_page_unqueue_nowakeup(m);
1295 m->queue = PQ_CACHE + m->pc;
1296 vm_page_queues[m->queue].lcnt++;
1297 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq);
1298 vmstats.v_cache_count++;
1299 vm_page_free_wakeup();
1304 * vm_page_dontneed()
1306 * Cache, deactivate, or do nothing as appropriate. This routine
1307 * is typically used by madvise() MADV_DONTNEED.
1309 * Generally speaking we want to move the page into the cache so
1310 * it gets reused quickly. However, this can result in a silly syndrome
1311 * due to the page recycling too quickly. Small objects will not be
1312 * fully cached. On the otherhand, if we move the page to the inactive
1313 * queue we wind up with a problem whereby very large objects
1314 * unnecessarily blow away our inactive and cache queues.
1316 * The solution is to move the pages based on a fixed weighting. We
1317 * either leave them alone, deactivate them, or move them to the cache,
1318 * where moving them to the cache has the highest weighting.
1319 * By forcing some pages into other queues we eventually force the
1320 * system to balance the queues, potentially recovering other unrelated
1321 * space from active. The idea is to not force this to happen too
1325 vm_page_dontneed(vm_page_t m)
1327 static int dnweight;
1334 * occassionally leave the page alone
1337 if ((dnw & 0x01F0) == 0 ||
1338 m->queue == PQ_INACTIVE ||
1339 m->queue - m->pc == PQ_CACHE
1341 if (m->act_count >= ACT_INIT)
1348 vm_page_test_dirty(m);
1350 if (m->dirty || (dnw & 0x0070) == 0) {
1352 * Deactivate the page 3 times out of 32.
1357 * Cache the page 28 times out of every 32. Note that
1358 * the page is deactivated instead of cached, but placed
1359 * at the head of the queue instead of the tail.
1363 _vm_page_deactivate(m, head);
1368 * Grab a page, blocking if it is busy and allocating a page if necessary.
1369 * A busy page is returned or NULL.
1371 * If VM_ALLOC_RETRY is specified VM_ALLOC_NORMAL must also be specified.
1372 * If VM_ALLOC_RETRY is not specified
1374 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
1375 * always returned if we had blocked.
1376 * This routine will never return NULL if VM_ALLOC_RETRY is set.
1377 * This routine may not be called from an interrupt.
1378 * The returned page may not be entirely valid.
1380 * This routine may be called from mainline code without spl protection and
1381 * be guarenteed a busied page associated with the object at the specified
1385 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
1390 KKASSERT(allocflags &
1391 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
1394 if ((m = vm_page_lookup(object, pindex)) != NULL) {
1395 if (m->busy || (m->flags & PG_BUSY)) {
1396 generation = object->generation;
1398 while ((object->generation == generation) &&
1399 (m->busy || (m->flags & PG_BUSY))) {
1400 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
1401 tsleep(m, 0, "pgrbwt", 0);
1402 if ((allocflags & VM_ALLOC_RETRY) == 0) {
1413 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1416 if ((allocflags & VM_ALLOC_RETRY) == 0)
1426 * Mapping function for valid bits or for dirty bits in
1427 * a page. May not block.
1429 * Inputs are required to range within a page.
1432 vm_page_bits(int base, int size)
1438 base + size <= PAGE_SIZE,
1439 ("vm_page_bits: illegal base/size %d/%d", base, size)
1442 if (size == 0) /* handle degenerate case */
1445 first_bit = base >> DEV_BSHIFT;
1446 last_bit = (base + size - 1) >> DEV_BSHIFT;
1448 return ((2 << last_bit) - (1 << first_bit));
1452 * Sets portions of a page valid and clean. The arguments are expected
1453 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1454 * of any partial chunks touched by the range. The invalid portion of
1455 * such chunks will be zero'd.
1457 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically
1458 * align base to DEV_BSIZE so as not to mark clean a partially
1459 * truncated device block. Otherwise the dirty page status might be
1462 * This routine may not block.
1464 * (base + size) must be less then or equal to PAGE_SIZE.
1467 _vm_page_zero_valid(vm_page_t m, int base, int size)
1472 if (size == 0) /* handle degenerate case */
1476 * If the base is not DEV_BSIZE aligned and the valid
1477 * bit is clear, we have to zero out a portion of the
1481 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
1482 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
1484 pmap_zero_page_area(
1492 * If the ending offset is not DEV_BSIZE aligned and the
1493 * valid bit is clear, we have to zero out a portion of
1497 endoff = base + size;
1499 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
1500 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
1502 pmap_zero_page_area(
1505 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
1511 * Set valid, clear dirty bits. If validating the entire
1512 * page we can safely clear the pmap modify bit. We also
1513 * use this opportunity to clear the PG_NOSYNC flag. If a process
1514 * takes a write fault on a MAP_NOSYNC memory area the flag will
1517 * We set valid bits inclusive of any overlap, but we can only
1518 * clear dirty bits for DEV_BSIZE chunks that are fully within
1522 vm_page_set_valid(vm_page_t m, int base, int size)
1524 _vm_page_zero_valid(m, base, size);
1525 m->valid |= vm_page_bits(base, size);
1530 * Set valid bits and clear dirty bits.
1532 * NOTE: This function does not clear the pmap modified bit.
1533 * Also note that e.g. NFS may use a byte-granular base
1537 vm_page_set_validclean(vm_page_t m, int base, int size)
1541 _vm_page_zero_valid(m, base, size);
1542 pagebits = vm_page_bits(base, size);
1543 m->valid |= pagebits;
1544 m->dirty &= ~pagebits;
1545 if (base == 0 && size == PAGE_SIZE) {
1546 /*pmap_clear_modify(m);*/
1547 vm_page_flag_clear(m, PG_NOSYNC);
1554 * NOTE: This function does not clear the pmap modified bit.
1555 * Also note that e.g. NFS may use a byte-granular base
1559 vm_page_clear_dirty(vm_page_t m, int base, int size)
1561 m->dirty &= ~vm_page_bits(base, size);
1562 if (base == 0 && size == PAGE_SIZE) {
1563 /*pmap_clear_modify(m);*/
1564 vm_page_flag_clear(m, PG_NOSYNC);
1569 * Make the page all-dirty.
1571 * Also make sure the related object and vnode reflect the fact that the
1572 * object may now contain a dirty page.
1575 vm_page_dirty(vm_page_t m)
1578 int pqtype = m->queue - m->pc;
1580 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE,
1581 ("vm_page_dirty: page in free/cache queue!"));
1582 if (m->dirty != VM_PAGE_BITS_ALL) {
1583 m->dirty = VM_PAGE_BITS_ALL;
1585 vm_object_set_writeable_dirty(m->object);
1590 * Invalidates DEV_BSIZE'd chunks within a page. Both the
1591 * valid and dirty bits for the effected areas are cleared.
1596 vm_page_set_invalid(vm_page_t m, int base, int size)
1600 bits = vm_page_bits(base, size);
1603 m->object->generation++;
1607 * The kernel assumes that the invalid portions of a page contain
1608 * garbage, but such pages can be mapped into memory by user code.
1609 * When this occurs, we must zero out the non-valid portions of the
1610 * page so user code sees what it expects.
1612 * Pages are most often semi-valid when the end of a file is mapped
1613 * into memory and the file's size is not page aligned.
1616 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
1622 * Scan the valid bits looking for invalid sections that
1623 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
1624 * valid bit may be set ) have already been zerod by
1625 * vm_page_set_validclean().
1627 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
1628 if (i == (PAGE_SIZE / DEV_BSIZE) ||
1629 (m->valid & (1 << i))
1632 pmap_zero_page_area(
1635 (i - b) << DEV_BSHIFT
1643 * setvalid is TRUE when we can safely set the zero'd areas
1644 * as being valid. We can do this if there are no cache consistency
1645 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
1648 m->valid = VM_PAGE_BITS_ALL;
1652 * Is a (partial) page valid? Note that the case where size == 0
1653 * will return FALSE in the degenerate case where the page is entirely
1654 * invalid, and TRUE otherwise.
1659 vm_page_is_valid(vm_page_t m, int base, int size)
1661 int bits = vm_page_bits(base, size);
1663 if (m->valid && ((m->valid & bits) == bits))
1670 * update dirty bits from pmap/mmu. May not block.
1673 vm_page_test_dirty(vm_page_t m)
1675 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
1681 * Issue an event on a VM page. Corresponding action structures are
1682 * removed from the page's list and called.
1685 vm_page_event_internal(vm_page_t m, vm_page_event_t event)
1687 struct vm_page_action *scan, *next;
1689 LIST_FOREACH_MUTABLE(scan, &m->action_list, entry, next) {
1690 if (scan->event == event) {
1691 scan->event = VMEVENT_NONE;
1692 LIST_REMOVE(scan, entry);
1693 scan->func(m, scan);
1698 #include "opt_ddb.h"
1700 #include <sys/kernel.h>
1702 #include <ddb/ddb.h>
1704 DB_SHOW_COMMAND(page, vm_page_print_page_info)
1706 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count);
1707 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count);
1708 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count);
1709 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count);
1710 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count);
1711 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved);
1712 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min);
1713 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target);
1714 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min);
1715 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target);
1718 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
1721 db_printf("PQ_FREE:");
1722 for(i=0;i<PQ_L2_SIZE;i++) {
1723 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
1727 db_printf("PQ_CACHE:");
1728 for(i=0;i<PQ_L2_SIZE;i++) {
1729 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
1733 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
1734 vm_page_queues[PQ_ACTIVE].lcnt,
1735 vm_page_queues[PQ_INACTIVE].lcnt);