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/swap_pager.h>
92 #include <machine/md_var.h>
94 #include <vm/vm_page2.h>
95 #include <sys/mplock2.h>
97 static void vm_page_queue_init(void);
98 static void vm_page_free_wakeup(void);
99 static vm_page_t vm_page_select_cache(vm_object_t, vm_pindex_t);
100 static vm_page_t _vm_page_list_find2(int basequeue, int index);
102 struct vpgqueues vm_page_queues[PQ_COUNT]; /* Array of tailq lists */
104 #define ASSERT_IN_CRIT_SECTION() KKASSERT(crit_test(curthread));
106 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare,
107 vm_pindex_t, pindex);
110 vm_page_queue_init(void)
114 for (i = 0; i < PQ_L2_SIZE; i++)
115 vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count;
116 for (i = 0; i < PQ_L2_SIZE; i++)
117 vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count;
119 vm_page_queues[PQ_INACTIVE].cnt = &vmstats.v_inactive_count;
120 vm_page_queues[PQ_ACTIVE].cnt = &vmstats.v_active_count;
121 vm_page_queues[PQ_HOLD].cnt = &vmstats.v_active_count;
122 /* PQ_NONE has no queue */
124 for (i = 0; i < PQ_COUNT; i++)
125 TAILQ_INIT(&vm_page_queues[i].pl);
129 * note: place in initialized data section? Is this necessary?
132 int vm_page_array_size = 0;
133 int vm_page_zero_count = 0;
134 vm_page_t vm_page_array = 0;
139 * Sets the page size, perhaps based upon the memory size.
140 * Must be called before any use of page-size dependent functions.
143 vm_set_page_size(void)
145 if (vmstats.v_page_size == 0)
146 vmstats.v_page_size = PAGE_SIZE;
147 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0)
148 panic("vm_set_page_size: page size not a power of two");
154 * Add a new page to the freelist for use by the system. New pages
155 * are added to both the head and tail of the associated free page
156 * queue in a bottom-up fashion, so both zero'd and non-zero'd page
157 * requests pull 'recent' adds (higher physical addresses) first.
159 * Must be called in a critical section.
162 vm_add_new_page(vm_paddr_t pa)
164 struct vpgqueues *vpq;
167 ++vmstats.v_page_count;
168 ++vmstats.v_free_count;
169 m = PHYS_TO_VM_PAGE(pa);
172 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK;
173 m->queue = m->pc + PQ_FREE;
174 KKASSERT(m->dirty == 0);
176 vpq = &vm_page_queues[m->queue];
178 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
180 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
181 vpq->flipflop = 1 - vpq->flipflop;
183 vm_page_queues[m->queue].lcnt++;
190 * Initializes the resident memory module.
192 * Allocates memory for the page cells, and for the object/offset-to-page
193 * hash table headers. Each page cell is initialized and placed on the
196 * starta/enda represents the range of physical memory addresses available
197 * for use (skipping memory already used by the kernel), subject to
198 * phys_avail[]. Note that phys_avail[] has already mapped out memory
199 * already in use by the kernel.
202 vm_page_startup(vm_offset_t vaddr)
206 vm_paddr_t page_range;
213 vm_paddr_t biggestone, biggestsize;
220 vaddr = round_page(vaddr);
222 for (i = 0; phys_avail[i + 1]; i += 2) {
223 phys_avail[i] = round_page64(phys_avail[i]);
224 phys_avail[i + 1] = trunc_page64(phys_avail[i + 1]);
227 for (i = 0; phys_avail[i + 1]; i += 2) {
228 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
230 if (size > biggestsize) {
238 end = phys_avail[biggestone+1];
239 end = trunc_page(end);
242 * Initialize the queue headers for the free queue, the active queue
243 * and the inactive queue.
246 vm_page_queue_init();
248 /* VKERNELs don't support minidumps and as such don't need vm_page_dump */
249 #if !defined(_KERNEL_VIRTUAL)
251 * Allocate a bitmap to indicate that a random physical page
252 * needs to be included in a minidump.
254 * The amd64 port needs this to indicate which direct map pages
255 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
257 * However, i386 still needs this workspace internally within the
258 * minidump code. In theory, they are not needed on i386, but are
259 * included should the sf_buf code decide to use them.
261 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE;
262 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
263 end -= vm_page_dump_size;
264 vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size,
265 VM_PROT_READ | VM_PROT_WRITE);
266 bzero((void *)vm_page_dump, vm_page_dump_size);
270 * Compute the number of pages of memory that will be available for
271 * use (taking into account the overhead of a page structure per
274 first_page = phys_avail[0] / PAGE_SIZE;
275 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
276 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE;
279 * Initialize the mem entry structures now, and put them in the free
282 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
283 mapped = pmap_map(&vaddr, new_end, end,
284 VM_PROT_READ | VM_PROT_WRITE);
285 vm_page_array = (vm_page_t)mapped;
287 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL)
289 * since pmap_map on amd64 returns stuff out of a direct-map region,
290 * we have to manually add these pages to the minidump tracking so
291 * that they can be dumped, including the vm_page_array.
293 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
298 * Clear all of the page structures
300 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
301 vm_page_array_size = page_range;
304 * Construct the free queue(s) in ascending order (by physical
305 * address) so that the first 16MB of physical memory is allocated
306 * last rather than first. On large-memory machines, this avoids
307 * the exhaustion of low physical memory before isa_dmainit has run.
309 vmstats.v_page_count = 0;
310 vmstats.v_free_count = 0;
311 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
316 last_pa = phys_avail[i + 1];
317 while (pa < last_pa && npages-- > 0) {
326 * Scan comparison function for Red-Black tree scans. An inclusive
327 * (start,end) is expected. Other fields are not used.
330 rb_vm_page_scancmp(struct vm_page *p, void *data)
332 struct rb_vm_page_scan_info *info = data;
334 if (p->pindex < info->start_pindex)
336 if (p->pindex > info->end_pindex)
342 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2)
344 if (p1->pindex < p2->pindex)
346 if (p1->pindex > p2->pindex)
352 * The opposite of vm_page_hold(). A page can be freed while being held,
353 * which places it on the PQ_HOLD queue. We must call vm_page_free_toq()
354 * in this case to actually free it once the hold count drops to 0.
356 * This routine must be called at splvm().
359 vm_page_unhold(vm_page_t mem)
362 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
363 if (mem->hold_count == 0 && mem->queue == PQ_HOLD) {
365 vm_page_free_toq(mem);
370 * Inserts the given mem entry into the object and object list.
372 * The pagetables are not updated but will presumably fault the page
373 * in if necessary, or if a kernel page the caller will at some point
374 * enter the page into the kernel's pmap. We are not allowed to block
375 * here so we *can't* do this anyway.
377 * This routine may not block.
378 * This routine must be called with a critical section held.
381 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
383 ASSERT_IN_CRIT_SECTION();
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.
396 ASSERT_MP_LOCK_HELD(curthread);
397 vm_page_rb_tree_RB_INSERT(&object->rb_memq, m);
398 object->generation++;
401 * show that the object has one more resident page.
403 object->resident_page_count++;
406 * Since we are inserting a new and possibly dirty page,
407 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
409 if ((m->valid & m->dirty) || (m->flags & PG_WRITEABLE))
410 vm_object_set_writeable_dirty(object);
413 * Checks for a swap assignment and sets PG_SWAPPED if appropriate.
415 swap_pager_page_inserted(m);
419 * Removes the given vm_page_t from the global (object,index) hash table
420 * and from the object's memq.
422 * The underlying pmap entry (if any) is NOT removed here.
423 * This routine may not block.
425 * The page must be BUSY and will remain BUSY on return. No spl needs to be
426 * held on call to this routine.
428 * note: FreeBSD side effect was to unbusy the page on return. We leave
432 vm_page_remove(vm_page_t m)
437 if (m->object == NULL) {
442 if ((m->flags & PG_BUSY) == 0)
443 panic("vm_page_remove: page not busy");
448 * Remove the page from the object and update the object.
450 ASSERT_MP_LOCK_HELD(curthread);
451 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m);
452 object->resident_page_count--;
453 object->generation++;
460 * Locate and return the page at (object, pindex), or NULL if the
461 * page could not be found.
463 * This routine will operate properly without spl protection, but
464 * the returned page could be in flux if it is busy. Because an
465 * interrupt can race a caller's busy check (unbusying and freeing the
466 * page we return before the caller is able to check the busy bit),
467 * the caller should generally call this routine with a critical
470 * Callers may call this routine without spl protection if they know
471 * 'for sure' that the page will not be ripped out from under them
475 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
480 * Search the hash table for this object/offset pair
482 ASSERT_MP_LOCK_HELD(curthread);
484 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
486 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex));
493 * Move the given memory entry from its current object to the specified
494 * target object/offset.
496 * The object must be locked.
497 * This routine may not block.
499 * Note: This routine will raise itself to splvm(), the caller need not.
501 * Note: Swap associated with the page must be invalidated by the move. We
502 * have to do this for several reasons: (1) we aren't freeing the
503 * page, (2) we are dirtying the page, (3) the VM system is probably
504 * moving the page from object A to B, and will then later move
505 * the backing store from A to B and we can't have a conflict.
507 * Note: We *always* dirty the page. It is necessary both for the
508 * fact that we moved it, and because we may be invalidating
509 * swap. If the page is on the cache, we have to deactivate it
510 * or vm_page_dirty() will panic. Dirty pages are not allowed
514 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
518 vm_page_insert(m, new_object, new_pindex);
519 if (m->queue - m->pc == PQ_CACHE)
520 vm_page_deactivate(m);
527 * vm_page_unqueue() without any wakeup. This routine is used when a page
528 * is being moved between queues or otherwise is to remain BUSYied by the
531 * This routine must be called at splhigh().
532 * This routine may not block.
535 vm_page_unqueue_nowakeup(vm_page_t m)
537 int queue = m->queue;
538 struct vpgqueues *pq;
540 if (queue != PQ_NONE) {
541 pq = &vm_page_queues[queue];
543 TAILQ_REMOVE(&pq->pl, m, pageq);
550 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
553 * This routine must be called at splhigh().
554 * This routine may not block.
557 vm_page_unqueue(vm_page_t m)
559 int queue = m->queue;
560 struct vpgqueues *pq;
562 if (queue != PQ_NONE) {
564 pq = &vm_page_queues[queue];
565 TAILQ_REMOVE(&pq->pl, m, pageq);
568 if ((queue - m->pc) == PQ_CACHE || (queue - m->pc) == PQ_FREE)
574 * vm_page_list_find()
576 * Find a page on the specified queue with color optimization.
578 * The page coloring optimization attempts to locate a page that does
579 * not overload other nearby pages in the object in the cpu's L1 or L2
580 * caches. We need this optimization because cpu caches tend to be
581 * physical caches, while object spaces tend to be virtual.
583 * This routine must be called at splvm().
584 * This routine may not block.
586 * Note that this routine is carefully inlined. A non-inlined version
587 * is available for outside callers but the only critical path is
588 * from within this source file.
592 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
597 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist);
599 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl);
601 m = _vm_page_list_find2(basequeue, index);
606 _vm_page_list_find2(int basequeue, int index)
610 struct vpgqueues *pq;
612 pq = &vm_page_queues[basequeue];
615 * Note that for the first loop, index+i and index-i wind up at the
616 * same place. Even though this is not totally optimal, we've already
617 * blown it by missing the cache case so we do not care.
620 for(i = PQ_L2_SIZE / 2; i > 0; --i) {
621 if ((m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl)) != NULL)
624 if ((m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl)) != NULL)
631 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
633 return(_vm_page_list_find(basequeue, index, prefer_zero));
637 * Find a page on the cache queue with color optimization. As pages
638 * might be found, but not applicable, they are deactivated. This
639 * keeps us from using potentially busy cached pages.
641 * This routine must be called with a critical section held.
642 * This routine may not block.
645 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex)
650 m = _vm_page_list_find(
652 (pindex + object->pg_color) & PQ_L2_MASK,
655 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
656 m->hold_count || m->wire_count)) {
657 vm_page_deactivate(m);
666 * Find a free or zero page, with specified preference. We attempt to
667 * inline the nominal case and fall back to _vm_page_select_free()
670 * This routine must be called with a critical section held.
671 * This routine may not block.
673 static __inline vm_page_t
674 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero)
678 m = _vm_page_list_find(
680 (pindex + object->pg_color) & PQ_L2_MASK,
689 * Allocate and return a memory cell associated with this VM object/offset
694 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain
695 * VM_ALLOC_QUICK like normal but cannot use cache
696 * VM_ALLOC_SYSTEM greater free drain
697 * VM_ALLOC_INTERRUPT allow free list to be completely drained
698 * VM_ALLOC_ZERO advisory request for pre-zero'd page
700 * The object must be locked.
701 * This routine may not block.
702 * The returned page will be marked PG_BUSY
704 * Additional special handling is required when called from an interrupt
705 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache
709 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
713 KKASSERT(object != NULL);
714 KASSERT(!vm_page_lookup(object, pindex),
715 ("vm_page_alloc: page already allocated"));
717 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK|
718 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
721 * Certain system threads (pageout daemon, buf_daemon's) are
722 * allowed to eat deeper into the free page list.
724 if (curthread->td_flags & TDF_SYSTHREAD)
725 page_req |= VM_ALLOC_SYSTEM;
729 if (vmstats.v_free_count > vmstats.v_free_reserved ||
730 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) ||
731 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 &&
732 vmstats.v_free_count > vmstats.v_interrupt_free_min)
735 * The free queue has sufficient free pages to take one out.
737 if (page_req & VM_ALLOC_ZERO)
738 m = vm_page_select_free(object, pindex, TRUE);
740 m = vm_page_select_free(object, pindex, FALSE);
741 } else if (page_req & VM_ALLOC_NORMAL) {
743 * Allocatable from the cache (non-interrupt only). On
744 * success, we must free the page and try again, thus
745 * ensuring that vmstats.v_*_free_min counters are replenished.
748 if (curthread->td_preempted) {
749 kprintf("vm_page_alloc(): warning, attempt to allocate"
750 " cache page from preempting interrupt\n");
753 m = vm_page_select_cache(object, pindex);
756 m = vm_page_select_cache(object, pindex);
759 * On success move the page into the free queue and loop.
762 KASSERT(m->dirty == 0,
763 ("Found dirty cache page %p", m));
765 vm_page_protect(m, VM_PROT_NONE);
771 * On failure return NULL
774 #if defined(DIAGNOSTIC)
775 if (vmstats.v_cache_count > 0)
776 kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count);
778 vm_pageout_deficit++;
783 * No pages available, wakeup the pageout daemon and give up.
786 vm_pageout_deficit++;
792 * Good page found. The page has not yet been busied. We are in
793 * a critical section.
795 KASSERT(m != NULL, ("vm_page_alloc(): missing page on free queue\n"));
796 KASSERT(m->dirty == 0,
797 ("vm_page_alloc: free/cache page %p was dirty", m));
800 * Remove from free queue
802 vm_page_unqueue_nowakeup(m);
805 * Initialize structure. Only the PG_ZERO flag is inherited. Set
808 if (m->flags & PG_ZERO) {
809 vm_page_zero_count--;
810 m->flags = PG_ZERO | PG_BUSY;
821 * vm_page_insert() is safe prior to the crit_exit(). Note also that
822 * inserting a page here does not insert it into the pmap (which
823 * could cause us to block allocating memory). We cannot block
826 vm_page_insert(m, object, pindex);
829 * Don't wakeup too often - wakeup the pageout daemon when
830 * we would be nearly out of memory.
837 * A PG_BUSY page is returned.
843 * Wait for sufficient free memory for nominal heavy memory use kernel
847 vm_wait_nominal(void)
849 while (vm_page_count_min(0))
854 * Test if vm_wait_nominal() would block.
857 vm_test_nominal(void)
859 if (vm_page_count_min(0))
865 * Block until free pages are available for allocation, called in various
866 * places before memory allocations.
872 if (curthread == pagethread) {
873 vm_pageout_pages_needed = 1;
874 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo);
876 if (vm_pages_needed == 0) {
878 wakeup(&vm_pages_needed);
880 tsleep(&vmstats.v_free_count, 0, "vmwait", timo);
886 * Block until free pages are available for allocation
888 * Called only in vm_fault so that processes page faulting can be
895 if (vm_pages_needed == 0) {
897 wakeup(&vm_pages_needed);
899 tsleep(&vmstats.v_free_count, 0, "pfault", 0);
904 * Put the specified page on the active list (if appropriate). Ensure
905 * that act_count is at least ACT_INIT but do not otherwise mess with it.
907 * The page queues must be locked.
908 * This routine may not block.
911 vm_page_activate(vm_page_t m)
914 if (m->queue != PQ_ACTIVE) {
915 if ((m->queue - m->pc) == PQ_CACHE)
916 mycpu->gd_cnt.v_reactivated++;
920 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
921 m->queue = PQ_ACTIVE;
922 vm_page_queues[PQ_ACTIVE].lcnt++;
923 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl,
925 if (m->act_count < ACT_INIT)
926 m->act_count = ACT_INIT;
927 vmstats.v_active_count++;
930 if (m->act_count < ACT_INIT)
931 m->act_count = ACT_INIT;
937 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
938 * routine is called when a page has been added to the cache or free
941 * This routine may not block.
942 * This routine must be called at splvm()
945 vm_page_free_wakeup(void)
948 * if pageout daemon needs pages, then tell it that there are
951 if (vm_pageout_pages_needed &&
952 vmstats.v_cache_count + vmstats.v_free_count >=
953 vmstats.v_pageout_free_min
955 wakeup(&vm_pageout_pages_needed);
956 vm_pageout_pages_needed = 0;
960 * wakeup processes that are waiting on memory if we hit a
961 * high water mark. And wakeup scheduler process if we have
962 * lots of memory. this process will swapin processes.
964 if (vm_pages_needed && !vm_page_count_min(0)) {
966 wakeup(&vmstats.v_free_count);
973 * Returns the given page to the PQ_FREE list, disassociating it with
976 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on
977 * return (the page will have been freed). No particular spl is required
980 * This routine may not block.
983 vm_page_free_toq(vm_page_t m)
985 struct vpgqueues *pq;
988 mycpu->gd_cnt.v_tfree++;
990 KKASSERT((m->flags & PG_MAPPED) == 0);
992 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
994 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
995 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
997 if ((m->queue - m->pc) == PQ_FREE)
998 panic("vm_page_free: freeing free page");
1000 panic("vm_page_free: freeing busy page");
1004 * unqueue, then remove page. Note that we cannot destroy
1005 * the page here because we do not want to call the pager's
1006 * callback routine until after we've put the page on the
1007 * appropriate free queue.
1009 vm_page_unqueue_nowakeup(m);
1013 * No further management of fictitious pages occurs beyond object
1014 * and queue removal.
1016 if ((m->flags & PG_FICTITIOUS) != 0) {
1025 if (m->wire_count != 0) {
1026 if (m->wire_count > 1) {
1028 "vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1029 m->wire_count, (long)m->pindex);
1031 panic("vm_page_free: freeing wired page");
1035 * Clear the UNMANAGED flag when freeing an unmanaged page.
1037 if (m->flags & PG_UNMANAGED) {
1038 m->flags &= ~PG_UNMANAGED;
1041 if (m->hold_count != 0) {
1042 m->flags &= ~PG_ZERO;
1045 m->queue = PQ_FREE + m->pc;
1047 pq = &vm_page_queues[m->queue];
1052 * Put zero'd pages on the end ( where we look for zero'd pages
1053 * first ) and non-zerod pages at the head.
1055 if (m->flags & PG_ZERO) {
1056 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
1057 ++vm_page_zero_count;
1059 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1062 vm_page_free_wakeup();
1067 * vm_page_free_fromq_fast()
1069 * Remove a non-zero page from one of the free queues; the page is removed for
1070 * zeroing, so do not issue a wakeup.
1075 vm_page_free_fromq_fast(void)
1082 for (i = 0; i < PQ_L2_SIZE; ++i) {
1083 m = vm_page_list_find(PQ_FREE, qi, FALSE);
1084 qi = (qi + PQ_PRIME2) & PQ_L2_MASK;
1085 if (m && (m->flags & PG_ZERO) == 0) {
1086 vm_page_unqueue_nowakeup(m);
1097 * vm_page_unmanage()
1099 * Prevent PV management from being done on the page. The page is
1100 * removed from the paging queues as if it were wired, and as a
1101 * consequence of no longer being managed the pageout daemon will not
1102 * touch it (since there is no way to locate the pte mappings for the
1103 * page). madvise() calls that mess with the pmap will also no longer
1104 * operate on the page.
1106 * Beyond that the page is still reasonably 'normal'. Freeing the page
1107 * will clear the flag.
1109 * This routine is used by OBJT_PHYS objects - objects using unswappable
1110 * physical memory as backing store rather then swap-backed memory and
1111 * will eventually be extended to support 4MB unmanaged physical
1114 * Must be called with a critical section held.
1117 vm_page_unmanage(vm_page_t m)
1119 ASSERT_IN_CRIT_SECTION();
1120 if ((m->flags & PG_UNMANAGED) == 0) {
1121 if (m->wire_count == 0)
1124 vm_page_flag_set(m, PG_UNMANAGED);
1128 * Mark this page as wired down by yet another map, removing it from
1129 * paging queues as necessary.
1131 * The page queues must be locked.
1132 * This routine may not block.
1135 vm_page_wire(vm_page_t m)
1138 * Only bump the wire statistics if the page is not already wired,
1139 * and only unqueue the page if it is on some queue (if it is unmanaged
1140 * it is already off the queues). Don't do anything with fictitious
1141 * pages because they are always wired.
1144 if ((m->flags & PG_FICTITIOUS) == 0) {
1145 if (m->wire_count == 0) {
1146 if ((m->flags & PG_UNMANAGED) == 0)
1148 vmstats.v_wire_count++;
1151 KASSERT(m->wire_count != 0,
1152 ("vm_page_wire: wire_count overflow m=%p", m));
1158 * Release one wiring of this page, potentially enabling it to be paged again.
1160 * Many pages placed on the inactive queue should actually go
1161 * into the cache, but it is difficult to figure out which. What
1162 * we do instead, if the inactive target is well met, is to put
1163 * clean pages at the head of the inactive queue instead of the tail.
1164 * This will cause them to be moved to the cache more quickly and
1165 * if not actively re-referenced, freed more quickly. If we just
1166 * stick these pages at the end of the inactive queue, heavy filesystem
1167 * meta-data accesses can cause an unnecessary paging load on memory bound
1168 * processes. This optimization causes one-time-use metadata to be
1169 * reused more quickly.
1171 * BUT, if we are in a low-memory situation we have no choice but to
1172 * put clean pages on the cache queue.
1174 * A number of routines use vm_page_unwire() to guarantee that the page
1175 * will go into either the inactive or active queues, and will NEVER
1176 * be placed in the cache - for example, just after dirtying a page.
1177 * dirty pages in the cache are not allowed.
1179 * The page queues must be locked.
1180 * This routine may not block.
1183 vm_page_unwire(vm_page_t m, int activate)
1186 if (m->flags & PG_FICTITIOUS) {
1188 } else if (m->wire_count <= 0) {
1189 panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
1191 if (--m->wire_count == 0) {
1192 --vmstats.v_wire_count;
1193 if (m->flags & PG_UNMANAGED) {
1195 } else if (activate) {
1197 &vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1198 m->queue = PQ_ACTIVE;
1199 vm_page_queues[PQ_ACTIVE].lcnt++;
1200 vmstats.v_active_count++;
1202 vm_page_flag_clear(m, PG_WINATCFLS);
1204 &vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1205 m->queue = PQ_INACTIVE;
1206 vm_page_queues[PQ_INACTIVE].lcnt++;
1207 vmstats.v_inactive_count++;
1208 ++vm_swapcache_inactive_heuristic;
1217 * Move the specified page to the inactive queue. If the page has
1218 * any associated swap, the swap is deallocated.
1220 * Normally athead is 0 resulting in LRU operation. athead is set
1221 * to 1 if we want this page to be 'as if it were placed in the cache',
1222 * except without unmapping it from the process address space.
1224 * This routine may not block.
1226 static __inline void
1227 _vm_page_deactivate(vm_page_t m, int athead)
1230 * Ignore if already inactive.
1232 if (m->queue == PQ_INACTIVE)
1235 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1236 if ((m->queue - m->pc) == PQ_CACHE)
1237 mycpu->gd_cnt.v_reactivated++;
1238 vm_page_flag_clear(m, PG_WINATCFLS);
1241 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl,
1244 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl,
1246 ++vm_swapcache_inactive_heuristic;
1248 m->queue = PQ_INACTIVE;
1249 vm_page_queues[PQ_INACTIVE].lcnt++;
1250 vmstats.v_inactive_count++;
1255 vm_page_deactivate(vm_page_t m)
1258 _vm_page_deactivate(m, 0);
1263 * vm_page_try_to_cache:
1265 * Returns 0 on failure, 1 on success
1268 vm_page_try_to_cache(vm_page_t m)
1271 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1272 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1276 vm_page_test_dirty(m);
1287 * Attempt to free the page. If we cannot free it, we do nothing.
1288 * 1 is returned on success, 0 on failure.
1291 vm_page_try_to_free(vm_page_t m)
1294 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1295 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1299 vm_page_test_dirty(m);
1305 vm_page_protect(m, VM_PROT_NONE);
1314 * Put the specified page onto the page cache queue (if appropriate).
1316 * This routine may not block.
1319 vm_page_cache(vm_page_t m)
1321 ASSERT_IN_CRIT_SECTION();
1323 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
1324 m->wire_count || m->hold_count) {
1325 kprintf("vm_page_cache: attempting to cache busy/held page\n");
1330 * Already in the cache (and thus not mapped)
1332 if ((m->queue - m->pc) == PQ_CACHE) {
1333 KKASSERT((m->flags & PG_MAPPED) == 0);
1338 * Caller is required to test m->dirty, but note that the act of
1339 * removing the page from its maps can cause it to become dirty
1340 * on an SMP system due to another cpu running in usermode.
1343 panic("vm_page_cache: caching a dirty page, pindex: %ld",
1348 * Remove all pmaps and indicate that the page is not
1349 * writeable or mapped. Our vm_page_protect() call may
1350 * have blocked (especially w/ VM_PROT_NONE), so recheck
1354 vm_page_protect(m, VM_PROT_NONE);
1356 if ((m->flags & (PG_BUSY|PG_UNMANAGED|PG_MAPPED)) || m->busy ||
1357 m->wire_count || m->hold_count) {
1359 } else if (m->dirty) {
1360 vm_page_deactivate(m);
1362 vm_page_unqueue_nowakeup(m);
1363 m->queue = PQ_CACHE + m->pc;
1364 vm_page_queues[m->queue].lcnt++;
1365 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq);
1366 vmstats.v_cache_count++;
1367 vm_page_free_wakeup();
1372 * vm_page_dontneed()
1374 * Cache, deactivate, or do nothing as appropriate. This routine
1375 * is typically used by madvise() MADV_DONTNEED.
1377 * Generally speaking we want to move the page into the cache so
1378 * it gets reused quickly. However, this can result in a silly syndrome
1379 * due to the page recycling too quickly. Small objects will not be
1380 * fully cached. On the otherhand, if we move the page to the inactive
1381 * queue we wind up with a problem whereby very large objects
1382 * unnecessarily blow away our inactive and cache queues.
1384 * The solution is to move the pages based on a fixed weighting. We
1385 * either leave them alone, deactivate them, or move them to the cache,
1386 * where moving them to the cache has the highest weighting.
1387 * By forcing some pages into other queues we eventually force the
1388 * system to balance the queues, potentially recovering other unrelated
1389 * space from active. The idea is to not force this to happen too
1393 vm_page_dontneed(vm_page_t m)
1395 static int dnweight;
1402 * occassionally leave the page alone
1405 if ((dnw & 0x01F0) == 0 ||
1406 m->queue == PQ_INACTIVE ||
1407 m->queue - m->pc == PQ_CACHE
1409 if (m->act_count >= ACT_INIT)
1416 vm_page_test_dirty(m);
1418 if (m->dirty || (dnw & 0x0070) == 0) {
1420 * Deactivate the page 3 times out of 32.
1425 * Cache the page 28 times out of every 32. Note that
1426 * the page is deactivated instead of cached, but placed
1427 * at the head of the queue instead of the tail.
1431 _vm_page_deactivate(m, head);
1436 * Grab a page, blocking if it is busy and allocating a page if necessary.
1437 * A busy page is returned or NULL.
1439 * If VM_ALLOC_RETRY is specified VM_ALLOC_NORMAL must also be specified.
1440 * If VM_ALLOC_RETRY is not specified
1442 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
1443 * always returned if we had blocked.
1444 * This routine will never return NULL if VM_ALLOC_RETRY is set.
1445 * This routine may not be called from an interrupt.
1446 * The returned page may not be entirely valid.
1448 * This routine may be called from mainline code without spl protection and
1449 * be guarenteed a busied page associated with the object at the specified
1453 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
1458 KKASSERT(allocflags &
1459 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
1462 if ((m = vm_page_lookup(object, pindex)) != NULL) {
1463 if (m->busy || (m->flags & PG_BUSY)) {
1464 generation = object->generation;
1466 while ((object->generation == generation) &&
1467 (m->busy || (m->flags & PG_BUSY))) {
1468 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
1469 tsleep(m, 0, "pgrbwt", 0);
1470 if ((allocflags & VM_ALLOC_RETRY) == 0) {
1481 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1484 if ((allocflags & VM_ALLOC_RETRY) == 0)
1494 * Mapping function for valid bits or for dirty bits in
1495 * a page. May not block.
1497 * Inputs are required to range within a page.
1500 vm_page_bits(int base, int size)
1506 base + size <= PAGE_SIZE,
1507 ("vm_page_bits: illegal base/size %d/%d", base, size)
1510 if (size == 0) /* handle degenerate case */
1513 first_bit = base >> DEV_BSHIFT;
1514 last_bit = (base + size - 1) >> DEV_BSHIFT;
1516 return ((2 << last_bit) - (1 << first_bit));
1520 * Sets portions of a page valid and clean. The arguments are expected
1521 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1522 * of any partial chunks touched by the range. The invalid portion of
1523 * such chunks will be zero'd.
1525 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically
1526 * align base to DEV_BSIZE so as not to mark clean a partially
1527 * truncated device block. Otherwise the dirty page status might be
1530 * This routine may not block.
1532 * (base + size) must be less then or equal to PAGE_SIZE.
1535 _vm_page_zero_valid(vm_page_t m, int base, int size)
1540 if (size == 0) /* handle degenerate case */
1544 * If the base is not DEV_BSIZE aligned and the valid
1545 * bit is clear, we have to zero out a portion of the
1549 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
1550 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
1552 pmap_zero_page_area(
1560 * If the ending offset is not DEV_BSIZE aligned and the
1561 * valid bit is clear, we have to zero out a portion of
1565 endoff = base + size;
1567 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
1568 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
1570 pmap_zero_page_area(
1573 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
1579 * Set valid, clear dirty bits. If validating the entire
1580 * page we can safely clear the pmap modify bit. We also
1581 * use this opportunity to clear the PG_NOSYNC flag. If a process
1582 * takes a write fault on a MAP_NOSYNC memory area the flag will
1585 * We set valid bits inclusive of any overlap, but we can only
1586 * clear dirty bits for DEV_BSIZE chunks that are fully within
1590 vm_page_set_valid(vm_page_t m, int base, int size)
1592 _vm_page_zero_valid(m, base, size);
1593 m->valid |= vm_page_bits(base, size);
1598 * Set valid bits and clear dirty bits.
1600 * NOTE: This function does not clear the pmap modified bit.
1601 * Also note that e.g. NFS may use a byte-granular base
1605 vm_page_set_validclean(vm_page_t m, int base, int size)
1609 _vm_page_zero_valid(m, base, size);
1610 pagebits = vm_page_bits(base, size);
1611 m->valid |= pagebits;
1612 m->dirty &= ~pagebits;
1613 if (base == 0 && size == PAGE_SIZE) {
1614 /*pmap_clear_modify(m);*/
1615 vm_page_flag_clear(m, PG_NOSYNC);
1620 * Set valid & dirty. Used by buwrite()
1623 vm_page_set_validdirty(vm_page_t m, int base, int size)
1627 pagebits = vm_page_bits(base, size);
1628 m->valid |= pagebits;
1629 m->dirty |= pagebits;
1631 vm_object_set_writeable_dirty(m->object);
1637 * NOTE: This function does not clear the pmap modified bit.
1638 * Also note that e.g. NFS may use a byte-granular base
1642 vm_page_clear_dirty(vm_page_t m, int base, int size)
1644 m->dirty &= ~vm_page_bits(base, size);
1645 if (base == 0 && size == PAGE_SIZE) {
1646 /*pmap_clear_modify(m);*/
1647 vm_page_flag_clear(m, PG_NOSYNC);
1652 * Make the page all-dirty.
1654 * Also make sure the related object and vnode reflect the fact that the
1655 * object may now contain a dirty page.
1658 vm_page_dirty(vm_page_t m)
1661 int pqtype = m->queue - m->pc;
1663 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE,
1664 ("vm_page_dirty: page in free/cache queue!"));
1665 if (m->dirty != VM_PAGE_BITS_ALL) {
1666 m->dirty = VM_PAGE_BITS_ALL;
1668 vm_object_set_writeable_dirty(m->object);
1673 * Invalidates DEV_BSIZE'd chunks within a page. Both the
1674 * valid and dirty bits for the effected areas are cleared.
1679 vm_page_set_invalid(vm_page_t m, int base, int size)
1683 bits = vm_page_bits(base, size);
1686 m->object->generation++;
1690 * The kernel assumes that the invalid portions of a page contain
1691 * garbage, but such pages can be mapped into memory by user code.
1692 * When this occurs, we must zero out the non-valid portions of the
1693 * page so user code sees what it expects.
1695 * Pages are most often semi-valid when the end of a file is mapped
1696 * into memory and the file's size is not page aligned.
1699 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
1705 * Scan the valid bits looking for invalid sections that
1706 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
1707 * valid bit may be set ) have already been zerod by
1708 * vm_page_set_validclean().
1710 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
1711 if (i == (PAGE_SIZE / DEV_BSIZE) ||
1712 (m->valid & (1 << i))
1715 pmap_zero_page_area(
1718 (i - b) << DEV_BSHIFT
1726 * setvalid is TRUE when we can safely set the zero'd areas
1727 * as being valid. We can do this if there are no cache consistency
1728 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
1731 m->valid = VM_PAGE_BITS_ALL;
1735 * Is a (partial) page valid? Note that the case where size == 0
1736 * will return FALSE in the degenerate case where the page is entirely
1737 * invalid, and TRUE otherwise.
1742 vm_page_is_valid(vm_page_t m, int base, int size)
1744 int bits = vm_page_bits(base, size);
1746 if (m->valid && ((m->valid & bits) == bits))
1753 * update dirty bits from pmap/mmu. May not block.
1756 vm_page_test_dirty(vm_page_t m)
1758 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
1764 * Issue an event on a VM page. Corresponding action structures are
1765 * removed from the page's list and called.
1768 vm_page_event_internal(vm_page_t m, vm_page_event_t event)
1770 struct vm_page_action *scan, *next;
1772 LIST_FOREACH_MUTABLE(scan, &m->action_list, entry, next) {
1773 if (scan->event == event) {
1774 scan->event = VMEVENT_NONE;
1775 LIST_REMOVE(scan, entry);
1776 scan->func(m, scan);
1782 #include "opt_ddb.h"
1784 #include <sys/kernel.h>
1786 #include <ddb/ddb.h>
1788 DB_SHOW_COMMAND(page, vm_page_print_page_info)
1790 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count);
1791 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count);
1792 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count);
1793 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count);
1794 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count);
1795 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved);
1796 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min);
1797 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target);
1798 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min);
1799 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target);
1802 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
1805 db_printf("PQ_FREE:");
1806 for(i=0;i<PQ_L2_SIZE;i++) {
1807 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
1811 db_printf("PQ_CACHE:");
1812 for(i=0;i<PQ_L2_SIZE;i++) {
1813 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
1817 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
1818 vm_page_queues[PQ_ACTIVE].lcnt,
1819 vm_page_queues[PQ_INACTIVE].lcnt);