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 1.39 2008/07/01 02:02:56 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 struct vpgqueues vm_page_queues[PQ_COUNT]; /* Array of tailq lists */
99 #define ASSERT_IN_CRIT_SECTION() KKASSERT(crit_test(curthread));
101 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare,
102 vm_pindex_t, pindex);
105 vm_page_queue_init(void)
109 for (i = 0; i < PQ_L2_SIZE; i++)
110 vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count;
111 for (i = 0; i < PQ_L2_SIZE; i++)
112 vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count;
114 vm_page_queues[PQ_INACTIVE].cnt = &vmstats.v_inactive_count;
115 vm_page_queues[PQ_ACTIVE].cnt = &vmstats.v_active_count;
116 vm_page_queues[PQ_HOLD].cnt = &vmstats.v_active_count;
117 /* PQ_NONE has no queue */
119 for (i = 0; i < PQ_COUNT; i++)
120 TAILQ_INIT(&vm_page_queues[i].pl);
124 * note: place in initialized data section? Is this necessary?
127 int vm_page_array_size = 0;
128 int vm_page_zero_count = 0;
129 vm_page_t vm_page_array = 0;
134 * Sets the page size, perhaps based upon the memory size.
135 * Must be called before any use of page-size dependent functions.
138 vm_set_page_size(void)
140 if (vmstats.v_page_size == 0)
141 vmstats.v_page_size = PAGE_SIZE;
142 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0)
143 panic("vm_set_page_size: page size not a power of two");
149 * Add a new page to the freelist for use by the system. New pages
150 * are added to both the head and tail of the associated free page
151 * queue in a bottom-up fashion, so both zero'd and non-zero'd page
152 * requests pull 'recent' adds (higher physical addresses) first.
154 * Must be called in a critical section.
157 vm_add_new_page(vm_paddr_t pa)
159 struct vpgqueues *vpq;
162 ++vmstats.v_page_count;
163 ++vmstats.v_free_count;
164 m = PHYS_TO_VM_PAGE(pa);
167 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK;
168 m->queue = m->pc + PQ_FREE;
169 KKASSERT(m->dirty == 0);
171 vpq = &vm_page_queues[m->queue];
173 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
175 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
176 vpq->flipflop = 1 - vpq->flipflop;
178 vm_page_queues[m->queue].lcnt++;
185 * Initializes the resident memory module.
187 * Allocates memory for the page cells, and for the object/offset-to-page
188 * hash table headers. Each page cell is initialized and placed on the
191 * starta/enda represents the range of physical memory addresses available
192 * for use (skipping memory already used by the kernel), subject to
193 * phys_avail[]. Note that phys_avail[] has already mapped out memory
194 * already in use by the kernel.
197 vm_page_startup(vm_offset_t vaddr)
201 vm_paddr_t page_range;
208 vm_paddr_t biggestone, biggestsize;
215 vaddr = round_page(vaddr);
217 for (i = 0; phys_avail[i + 1]; i += 2) {
218 phys_avail[i] = round_page(phys_avail[i]);
219 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
222 for (i = 0; phys_avail[i + 1]; i += 2) {
223 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
225 if (size > biggestsize) {
233 end = phys_avail[biggestone+1];
234 end = trunc_page(end);
237 * Initialize the queue headers for the free queue, the active queue
238 * and the inactive queue.
241 vm_page_queue_init();
244 * Compute the number of pages of memory that will be available for
245 * use (taking into account the overhead of a page structure per
248 first_page = phys_avail[0] / PAGE_SIZE;
249 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
250 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE;
253 * Initialize the mem entry structures now, and put them in the free
256 vm_page_array = (vm_page_t) vaddr;
260 * Validate these addresses.
262 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
263 mapped = pmap_map(mapped, new_end, end,
264 VM_PROT_READ | VM_PROT_WRITE);
267 * Clear all of the page structures
269 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
270 vm_page_array_size = page_range;
273 * Construct the free queue(s) in ascending order (by physical
274 * address) so that the first 16MB of physical memory is allocated
275 * last rather than first. On large-memory machines, this avoids
276 * the exhaustion of low physical memory before isa_dmainit has run.
278 vmstats.v_page_count = 0;
279 vmstats.v_free_count = 0;
280 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
285 last_pa = phys_avail[i + 1];
286 while (pa < last_pa && npages-- > 0) {
295 * Scan comparison function for Red-Black tree scans. An inclusive
296 * (start,end) is expected. Other fields are not used.
299 rb_vm_page_scancmp(struct vm_page *p, void *data)
301 struct rb_vm_page_scan_info *info = data;
303 if (p->pindex < info->start_pindex)
305 if (p->pindex > info->end_pindex)
311 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2)
313 if (p1->pindex < p2->pindex)
315 if (p1->pindex > p2->pindex)
321 * The opposite of vm_page_hold(). A page can be freed while being held,
322 * which places it on the PQ_HOLD queue. We must call vm_page_free_toq()
323 * in this case to actually free it once the hold count drops to 0.
325 * This routine must be called at splvm().
328 vm_page_unhold(vm_page_t mem)
331 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
332 if (mem->hold_count == 0 && mem->queue == PQ_HOLD) {
334 vm_page_free_toq(mem);
339 * Inserts the given mem entry into the object and object list.
341 * The pagetables are not updated but will presumably fault the page
342 * in if necessary, or if a kernel page the caller will at some point
343 * enter the page into the kernel's pmap. We are not allowed to block
344 * here so we *can't* do this anyway.
346 * This routine may not block.
347 * This routine must be called with a critical section held.
350 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
352 ASSERT_IN_CRIT_SECTION();
353 if (m->object != NULL)
354 panic("vm_page_insert: already inserted");
357 * Record the object/offset pair in this page
363 * Insert it into the object.
365 vm_page_rb_tree_RB_INSERT(&object->rb_memq, m);
366 object->generation++;
369 * show that the object has one more resident page.
371 object->resident_page_count++;
374 * Since we are inserting a new and possibly dirty page,
375 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
377 if ((m->valid & m->dirty) || (m->flags & PG_WRITEABLE))
378 vm_object_set_writeable_dirty(object);
382 * Removes the given vm_page_t from the global (object,index) hash table
383 * and from the object's memq.
385 * The underlying pmap entry (if any) is NOT removed here.
386 * This routine may not block.
388 * The page must be BUSY and will remain BUSY on return. No spl needs to be
389 * held on call to this routine.
391 * note: FreeBSD side effect was to unbusy the page on return. We leave
395 vm_page_remove(vm_page_t m)
400 if (m->object == NULL) {
405 if ((m->flags & PG_BUSY) == 0)
406 panic("vm_page_remove: page not busy");
411 * Remove the page from the object and update the object.
413 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m);
414 object->resident_page_count--;
415 object->generation++;
422 * Locate and return the page at (object, pindex), or NULL if the
423 * page could not be found.
425 * This routine will operate properly without spl protection, but
426 * the returned page could be in flux if it is busy. Because an
427 * interrupt can race a caller's busy check (unbusying and freeing the
428 * page we return before the caller is able to check the busy bit),
429 * the caller should generally call this routine with a critical
432 * Callers may call this routine without spl protection if they know
433 * 'for sure' that the page will not be ripped out from under them
437 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
442 * Search the hash table for this object/offset pair
445 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
447 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex));
454 * Move the given memory entry from its current object to the specified
455 * target object/offset.
457 * The object must be locked.
458 * This routine may not block.
460 * Note: This routine will raise itself to splvm(), the caller need not.
462 * Note: Swap associated with the page must be invalidated by the move. We
463 * have to do this for several reasons: (1) we aren't freeing the
464 * page, (2) we are dirtying the page, (3) the VM system is probably
465 * moving the page from object A to B, and will then later move
466 * the backing store from A to B and we can't have a conflict.
468 * Note: We *always* dirty the page. It is necessary both for the
469 * fact that we moved it, and because we may be invalidating
470 * swap. If the page is on the cache, we have to deactivate it
471 * or vm_page_dirty() will panic. Dirty pages are not allowed
475 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
479 vm_page_insert(m, new_object, new_pindex);
480 if (m->queue - m->pc == PQ_CACHE)
481 vm_page_deactivate(m);
488 * vm_page_unqueue() without any wakeup. This routine is used when a page
489 * is being moved between queues or otherwise is to remain BUSYied by the
492 * This routine must be called at splhigh().
493 * This routine may not block.
496 vm_page_unqueue_nowakeup(vm_page_t m)
498 int queue = m->queue;
499 struct vpgqueues *pq;
501 if (queue != PQ_NONE) {
502 pq = &vm_page_queues[queue];
504 TAILQ_REMOVE(&pq->pl, m, pageq);
511 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
514 * This routine must be called at splhigh().
515 * This routine may not block.
518 vm_page_unqueue(vm_page_t m)
520 int queue = m->queue;
521 struct vpgqueues *pq;
523 if (queue != PQ_NONE) {
525 pq = &vm_page_queues[queue];
526 TAILQ_REMOVE(&pq->pl, m, pageq);
529 if ((queue - m->pc) == PQ_CACHE) {
530 if (vm_paging_needed())
537 * vm_page_list_find()
539 * Find a page on the specified queue with color optimization.
541 * The page coloring optimization attempts to locate a page that does
542 * not overload other nearby pages in the object in the cpu's L1 or L2
543 * caches. We need this optimization because cpu caches tend to be
544 * physical caches, while object spaces tend to be virtual.
546 * This routine must be called at splvm().
547 * This routine may not block.
549 * Note that this routine is carefully inlined. A non-inlined version
550 * is available for outside callers but the only critical path is
551 * from within this source file.
555 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
560 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist);
562 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl);
564 m = _vm_page_list_find2(basequeue, index);
569 _vm_page_list_find2(int basequeue, int index)
573 struct vpgqueues *pq;
575 pq = &vm_page_queues[basequeue];
578 * Note that for the first loop, index+i and index-i wind up at the
579 * same place. Even though this is not totally optimal, we've already
580 * blown it by missing the cache case so we do not care.
583 for(i = PQ_L2_SIZE / 2; i > 0; --i) {
584 if ((m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl)) != NULL)
587 if ((m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl)) != NULL)
594 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
596 return(_vm_page_list_find(basequeue, index, prefer_zero));
600 * Find a page on the cache queue with color optimization. As pages
601 * might be found, but not applicable, they are deactivated. This
602 * keeps us from using potentially busy cached pages.
604 * This routine must be called with a critical section held.
605 * This routine may not block.
608 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex)
613 m = _vm_page_list_find(
615 (pindex + object->pg_color) & PQ_L2_MASK,
618 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
619 m->hold_count || m->wire_count)) {
620 vm_page_deactivate(m);
629 * Find a free or zero page, with specified preference. We attempt to
630 * inline the nominal case and fall back to _vm_page_select_free()
633 * This routine must be called with a critical section held.
634 * This routine may not block.
636 static __inline vm_page_t
637 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero)
641 m = _vm_page_list_find(
643 (pindex + object->pg_color) & PQ_L2_MASK,
652 * Allocate and return a memory cell associated with this VM object/offset
657 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain
658 * VM_ALLOC_SYSTEM greater free drain
659 * VM_ALLOC_INTERRUPT allow free list to be completely drained
660 * VM_ALLOC_ZERO advisory request for pre-zero'd page
662 * The object must be locked.
663 * This routine may not block.
664 * The returned page will be marked PG_BUSY
666 * Additional special handling is required when called from an interrupt
667 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache
671 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
675 KASSERT(!vm_page_lookup(object, pindex),
676 ("vm_page_alloc: page already allocated"));
678 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
681 * Certain system threads (pageout daemon, buf_daemon's) are
682 * allowed to eat deeper into the free page list.
684 if (curthread->td_flags & TDF_SYSTHREAD)
685 page_req |= VM_ALLOC_SYSTEM;
689 if (vmstats.v_free_count > vmstats.v_free_reserved ||
690 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) ||
691 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 &&
692 vmstats.v_free_count > vmstats.v_interrupt_free_min)
695 * The free queue has sufficient free pages to take one out.
697 if (page_req & VM_ALLOC_ZERO)
698 m = vm_page_select_free(object, pindex, TRUE);
700 m = vm_page_select_free(object, pindex, FALSE);
701 } else if (page_req & VM_ALLOC_NORMAL) {
703 * Allocatable from the cache (non-interrupt only). On
704 * success, we must free the page and try again, thus
705 * ensuring that vmstats.v_*_free_min counters are replenished.
708 if (curthread->td_preempted) {
709 kprintf("vm_page_alloc(): warning, attempt to allocate"
710 " cache page from preempting interrupt\n");
713 m = vm_page_select_cache(object, pindex);
716 m = vm_page_select_cache(object, pindex);
719 * On success move the page into the free queue and loop.
722 KASSERT(m->dirty == 0,
723 ("Found dirty cache page %p", m));
725 vm_page_protect(m, VM_PROT_NONE);
731 * On failure return NULL
734 #if defined(DIAGNOSTIC)
735 if (vmstats.v_cache_count > 0)
736 kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count);
738 vm_pageout_deficit++;
743 * No pages available, wakeup the pageout daemon and give up.
746 vm_pageout_deficit++;
752 * Good page found. The page has not yet been busied. We are in
753 * a critical section.
755 KASSERT(m != NULL, ("vm_page_alloc(): missing page on free queue\n"));
756 KASSERT(m->dirty == 0,
757 ("vm_page_alloc: free/cache page %p was dirty", m));
760 * Remove from free queue
762 vm_page_unqueue_nowakeup(m);
765 * Initialize structure. Only the PG_ZERO flag is inherited. Set
768 if (m->flags & PG_ZERO) {
769 vm_page_zero_count--;
770 m->flags = PG_ZERO | PG_BUSY;
781 * vm_page_insert() is safe prior to the crit_exit(). Note also that
782 * inserting a page here does not insert it into the pmap (which
783 * could cause us to block allocating memory). We cannot block
786 vm_page_insert(m, object, pindex);
789 * Don't wakeup too often - wakeup the pageout daemon when
790 * we would be nearly out of memory.
792 if (vm_paging_needed())
798 * A PG_BUSY page is returned.
804 * Block until free pages are available for allocation, called in various
805 * places before memory allocations.
811 if (curthread == pagethread) {
812 vm_pageout_pages_needed = 1;
813 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo);
815 if (!vm_pages_needed) {
817 wakeup(&vm_pages_needed);
819 tsleep(&vmstats.v_free_count, 0, "vmwait", timo);
825 * Block until free pages are available for allocation
827 * Called only in vm_fault so that processes page faulting can be
830 * Sleeps at a lower priority than vm_wait() so that vm_wait()ing
831 * processes will be able to grab memory first. Do not change
832 * this balance without careful testing first.
838 if (!vm_pages_needed) {
840 wakeup(&vm_pages_needed);
842 tsleep(&vmstats.v_free_count, 0, "pfault", 0);
847 * Put the specified page on the active list (if appropriate). Ensure
848 * that act_count is at least ACT_INIT but do not otherwise mess with it.
850 * The page queues must be locked.
851 * This routine may not block.
854 vm_page_activate(vm_page_t m)
857 if (m->queue != PQ_ACTIVE) {
858 if ((m->queue - m->pc) == PQ_CACHE)
859 mycpu->gd_cnt.v_reactivated++;
863 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
864 m->queue = PQ_ACTIVE;
865 vm_page_queues[PQ_ACTIVE].lcnt++;
866 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl,
868 if (m->act_count < ACT_INIT)
869 m->act_count = ACT_INIT;
870 vmstats.v_active_count++;
873 if (m->act_count < ACT_INIT)
874 m->act_count = ACT_INIT;
880 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
881 * routine is called when a page has been added to the cache or free
884 * This routine may not block.
885 * This routine must be called at splvm()
888 vm_page_free_wakeup(void)
891 * if pageout daemon needs pages, then tell it that there are
894 if (vm_pageout_pages_needed &&
895 vmstats.v_cache_count + vmstats.v_free_count >=
896 vmstats.v_pageout_free_min
898 wakeup(&vm_pageout_pages_needed);
899 vm_pageout_pages_needed = 0;
903 * wakeup processes that are waiting on memory if we hit a
904 * high water mark. And wakeup scheduler process if we have
905 * lots of memory. this process will swapin processes.
907 if (vm_pages_needed && !vm_page_count_min()) {
909 wakeup(&vmstats.v_free_count);
916 * Returns the given page to the PQ_FREE list, disassociating it with
919 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on
920 * return (the page will have been freed). No particular spl is required
923 * This routine may not block.
926 vm_page_free_toq(vm_page_t m)
928 struct vpgqueues *pq;
931 mycpu->gd_cnt.v_tfree++;
933 KKASSERT((m->flags & PG_MAPPED) == 0);
935 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
937 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
938 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
940 if ((m->queue - m->pc) == PQ_FREE)
941 panic("vm_page_free: freeing free page");
943 panic("vm_page_free: freeing busy page");
947 * unqueue, then remove page. Note that we cannot destroy
948 * the page here because we do not want to call the pager's
949 * callback routine until after we've put the page on the
950 * appropriate free queue.
952 vm_page_unqueue_nowakeup(m);
956 * No further management of fictitious pages occurs beyond object
959 if ((m->flags & PG_FICTITIOUS) != 0) {
968 if (m->wire_count != 0) {
969 if (m->wire_count > 1) {
971 "vm_page_free: invalid wire count (%d), pindex: 0x%lx",
972 m->wire_count, (long)m->pindex);
974 panic("vm_page_free: freeing wired page");
978 * Clear the UNMANAGED flag when freeing an unmanaged page.
980 if (m->flags & PG_UNMANAGED) {
981 m->flags &= ~PG_UNMANAGED;
984 if (m->hold_count != 0) {
985 m->flags &= ~PG_ZERO;
988 m->queue = PQ_FREE + m->pc;
990 pq = &vm_page_queues[m->queue];
995 * Put zero'd pages on the end ( where we look for zero'd pages
996 * first ) and non-zerod pages at the head.
998 if (m->flags & PG_ZERO) {
999 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
1000 ++vm_page_zero_count;
1002 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1005 vm_page_free_wakeup();
1010 * vm_page_unmanage()
1012 * Prevent PV management from being done on the page. The page is
1013 * removed from the paging queues as if it were wired, and as a
1014 * consequence of no longer being managed the pageout daemon will not
1015 * touch it (since there is no way to locate the pte mappings for the
1016 * page). madvise() calls that mess with the pmap will also no longer
1017 * operate on the page.
1019 * Beyond that the page is still reasonably 'normal'. Freeing the page
1020 * will clear the flag.
1022 * This routine is used by OBJT_PHYS objects - objects using unswappable
1023 * physical memory as backing store rather then swap-backed memory and
1024 * will eventually be extended to support 4MB unmanaged physical
1027 * Must be called with a critical section held.
1030 vm_page_unmanage(vm_page_t m)
1032 ASSERT_IN_CRIT_SECTION();
1033 if ((m->flags & PG_UNMANAGED) == 0) {
1034 if (m->wire_count == 0)
1037 vm_page_flag_set(m, PG_UNMANAGED);
1041 * Mark this page as wired down by yet another map, removing it from
1042 * paging queues as necessary.
1044 * The page queues must be locked.
1045 * This routine may not block.
1048 vm_page_wire(vm_page_t m)
1051 * Only bump the wire statistics if the page is not already wired,
1052 * and only unqueue the page if it is on some queue (if it is unmanaged
1053 * it is already off the queues). Don't do anything with fictitious
1054 * pages because they are always wired.
1057 if ((m->flags & PG_FICTITIOUS) == 0) {
1058 if (m->wire_count == 0) {
1059 if ((m->flags & PG_UNMANAGED) == 0)
1061 vmstats.v_wire_count++;
1064 KASSERT(m->wire_count != 0,
1065 ("vm_page_wire: wire_count overflow m=%p", m));
1071 * Release one wiring of this page, potentially enabling it to be paged again.
1073 * Many pages placed on the inactive queue should actually go
1074 * into the cache, but it is difficult to figure out which. What
1075 * we do instead, if the inactive target is well met, is to put
1076 * clean pages at the head of the inactive queue instead of the tail.
1077 * This will cause them to be moved to the cache more quickly and
1078 * if not actively re-referenced, freed more quickly. If we just
1079 * stick these pages at the end of the inactive queue, heavy filesystem
1080 * meta-data accesses can cause an unnecessary paging load on memory bound
1081 * processes. This optimization causes one-time-use metadata to be
1082 * reused more quickly.
1084 * BUT, if we are in a low-memory situation we have no choice but to
1085 * put clean pages on the cache queue.
1087 * A number of routines use vm_page_unwire() to guarantee that the page
1088 * will go into either the inactive or active queues, and will NEVER
1089 * be placed in the cache - for example, just after dirtying a page.
1090 * dirty pages in the cache are not allowed.
1092 * The page queues must be locked.
1093 * This routine may not block.
1096 vm_page_unwire(vm_page_t m, int activate)
1099 if (m->flags & PG_FICTITIOUS) {
1101 } else if (m->wire_count <= 0) {
1102 panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
1104 if (--m->wire_count == 0) {
1105 --vmstats.v_wire_count;
1106 if (m->flags & PG_UNMANAGED) {
1108 } else if (activate) {
1110 &vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1111 m->queue = PQ_ACTIVE;
1112 vm_page_queues[PQ_ACTIVE].lcnt++;
1113 vmstats.v_active_count++;
1115 vm_page_flag_clear(m, PG_WINATCFLS);
1117 &vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1118 m->queue = PQ_INACTIVE;
1119 vm_page_queues[PQ_INACTIVE].lcnt++;
1120 vmstats.v_inactive_count++;
1129 * Move the specified page to the inactive queue. If the page has
1130 * any associated swap, the swap is deallocated.
1132 * Normally athead is 0 resulting in LRU operation. athead is set
1133 * to 1 if we want this page to be 'as if it were placed in the cache',
1134 * except without unmapping it from the process address space.
1136 * This routine may not block.
1138 static __inline void
1139 _vm_page_deactivate(vm_page_t m, int athead)
1142 * Ignore if already inactive.
1144 if (m->queue == PQ_INACTIVE)
1147 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1148 if ((m->queue - m->pc) == PQ_CACHE)
1149 mycpu->gd_cnt.v_reactivated++;
1150 vm_page_flag_clear(m, PG_WINATCFLS);
1153 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1155 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1156 m->queue = PQ_INACTIVE;
1157 vm_page_queues[PQ_INACTIVE].lcnt++;
1158 vmstats.v_inactive_count++;
1163 vm_page_deactivate(vm_page_t m)
1166 _vm_page_deactivate(m, 0);
1171 * vm_page_try_to_cache:
1173 * Returns 0 on failure, 1 on success
1176 vm_page_try_to_cache(vm_page_t m)
1179 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1180 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1184 vm_page_test_dirty(m);
1195 * Attempt to free the page. If we cannot free it, we do nothing.
1196 * 1 is returned on success, 0 on failure.
1199 vm_page_try_to_free(vm_page_t m)
1202 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1203 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1207 vm_page_test_dirty(m);
1213 vm_page_protect(m, VM_PROT_NONE);
1222 * Put the specified page onto the page cache queue (if appropriate).
1224 * This routine may not block.
1227 vm_page_cache(vm_page_t m)
1229 ASSERT_IN_CRIT_SECTION();
1231 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
1232 m->wire_count || m->hold_count) {
1233 kprintf("vm_page_cache: attempting to cache busy/held page\n");
1238 * Already in the cache (and thus not mapped)
1240 if ((m->queue - m->pc) == PQ_CACHE) {
1241 KKASSERT((m->flags & PG_MAPPED) == 0);
1246 * Caller is required to test m->dirty, but note that the act of
1247 * removing the page from its maps can cause it to become dirty
1248 * on an SMP system due to another cpu running in usermode.
1251 panic("vm_page_cache: caching a dirty page, pindex: %ld",
1256 * Remove all pmaps and indicate that the page is not
1257 * writeable or mapped. Our vm_page_protect() call may
1258 * have blocked (especially w/ VM_PROT_NONE), so recheck
1262 vm_page_protect(m, VM_PROT_NONE);
1264 if ((m->flags & (PG_BUSY|PG_UNMANAGED|PG_MAPPED)) || m->busy ||
1265 m->wire_count || m->hold_count) {
1267 } else if (m->dirty) {
1268 vm_page_deactivate(m);
1270 vm_page_unqueue_nowakeup(m);
1271 m->queue = PQ_CACHE + m->pc;
1272 vm_page_queues[m->queue].lcnt++;
1273 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq);
1274 vmstats.v_cache_count++;
1275 vm_page_free_wakeup();
1280 * vm_page_dontneed()
1282 * Cache, deactivate, or do nothing as appropriate. This routine
1283 * is typically used by madvise() MADV_DONTNEED.
1285 * Generally speaking we want to move the page into the cache so
1286 * it gets reused quickly. However, this can result in a silly syndrome
1287 * due to the page recycling too quickly. Small objects will not be
1288 * fully cached. On the otherhand, if we move the page to the inactive
1289 * queue we wind up with a problem whereby very large objects
1290 * unnecessarily blow away our inactive and cache queues.
1292 * The solution is to move the pages based on a fixed weighting. We
1293 * either leave them alone, deactivate them, or move them to the cache,
1294 * where moving them to the cache has the highest weighting.
1295 * By forcing some pages into other queues we eventually force the
1296 * system to balance the queues, potentially recovering other unrelated
1297 * space from active. The idea is to not force this to happen too
1301 vm_page_dontneed(vm_page_t m)
1303 static int dnweight;
1310 * occassionally leave the page alone
1313 if ((dnw & 0x01F0) == 0 ||
1314 m->queue == PQ_INACTIVE ||
1315 m->queue - m->pc == PQ_CACHE
1317 if (m->act_count >= ACT_INIT)
1324 vm_page_test_dirty(m);
1326 if (m->dirty || (dnw & 0x0070) == 0) {
1328 * Deactivate the page 3 times out of 32.
1333 * Cache the page 28 times out of every 32. Note that
1334 * the page is deactivated instead of cached, but placed
1335 * at the head of the queue instead of the tail.
1339 _vm_page_deactivate(m, head);
1344 * Grab a page, blocking if it is busy and allocating a page if necessary.
1345 * A busy page is returned or NULL.
1347 * If VM_ALLOC_RETRY is specified VM_ALLOC_NORMAL must also be specified.
1348 * If VM_ALLOC_RETRY is not specified
1350 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
1351 * always returned if we had blocked.
1352 * This routine will never return NULL if VM_ALLOC_RETRY is set.
1353 * This routine may not be called from an interrupt.
1354 * The returned page may not be entirely valid.
1356 * This routine may be called from mainline code without spl protection and
1357 * be guarenteed a busied page associated with the object at the specified
1361 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
1366 KKASSERT(allocflags &
1367 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
1370 if ((m = vm_page_lookup(object, pindex)) != NULL) {
1371 if (m->busy || (m->flags & PG_BUSY)) {
1372 generation = object->generation;
1374 while ((object->generation == generation) &&
1375 (m->busy || (m->flags & PG_BUSY))) {
1376 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
1377 tsleep(m, 0, "pgrbwt", 0);
1378 if ((allocflags & VM_ALLOC_RETRY) == 0) {
1389 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1392 if ((allocflags & VM_ALLOC_RETRY) == 0)
1402 * Mapping function for valid bits or for dirty bits in
1403 * a page. May not block.
1405 * Inputs are required to range within a page.
1408 vm_page_bits(int base, int size)
1414 base + size <= PAGE_SIZE,
1415 ("vm_page_bits: illegal base/size %d/%d", base, size)
1418 if (size == 0) /* handle degenerate case */
1421 first_bit = base >> DEV_BSHIFT;
1422 last_bit = (base + size - 1) >> DEV_BSHIFT;
1424 return ((2 << last_bit) - (1 << first_bit));
1428 * Sets portions of a page valid and clean. The arguments are expected
1429 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1430 * of any partial chunks touched by the range. The invalid portion of
1431 * such chunks will be zero'd.
1433 * This routine may not block.
1435 * (base + size) must be less then or equal to PAGE_SIZE.
1438 vm_page_set_validclean(vm_page_t m, int base, int size)
1444 if (size == 0) /* handle degenerate case */
1448 * If the base is not DEV_BSIZE aligned and the valid
1449 * bit is clear, we have to zero out a portion of the
1453 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
1454 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
1456 pmap_zero_page_area(
1464 * If the ending offset is not DEV_BSIZE aligned and the
1465 * valid bit is clear, we have to zero out a portion of
1469 endoff = base + size;
1471 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
1472 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
1474 pmap_zero_page_area(
1477 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
1482 * Set valid, clear dirty bits. If validating the entire
1483 * page we can safely clear the pmap modify bit. We also
1484 * use this opportunity to clear the PG_NOSYNC flag. If a process
1485 * takes a write fault on a MAP_NOSYNC memory area the flag will
1488 * We set valid bits inclusive of any overlap, but we can only
1489 * clear dirty bits for DEV_BSIZE chunks that are fully within
1493 pagebits = vm_page_bits(base, size);
1494 m->valid |= pagebits;
1496 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
1497 frag = DEV_BSIZE - frag;
1503 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
1505 m->dirty &= ~pagebits;
1506 if (base == 0 && size == PAGE_SIZE) {
1507 pmap_clear_modify(m);
1508 vm_page_flag_clear(m, PG_NOSYNC);
1513 vm_page_clear_dirty(vm_page_t m, int base, int size)
1515 m->dirty &= ~vm_page_bits(base, size);
1519 * Make the page all-dirty.
1521 * Also make sure the related object and vnode reflect the fact that the
1522 * object may now contain a dirty page.
1525 vm_page_dirty(vm_page_t m)
1528 int pqtype = m->queue - m->pc;
1530 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE,
1531 ("vm_page_dirty: page in free/cache queue!"));
1532 if (m->dirty != VM_PAGE_BITS_ALL) {
1533 m->dirty = VM_PAGE_BITS_ALL;
1535 vm_object_set_writeable_dirty(m->object);
1540 * Invalidates DEV_BSIZE'd chunks within a page. Both the
1541 * valid and dirty bits for the effected areas are cleared.
1546 vm_page_set_invalid(vm_page_t m, int base, int size)
1550 bits = vm_page_bits(base, size);
1553 m->object->generation++;
1557 * The kernel assumes that the invalid portions of a page contain
1558 * garbage, but such pages can be mapped into memory by user code.
1559 * When this occurs, we must zero out the non-valid portions of the
1560 * page so user code sees what it expects.
1562 * Pages are most often semi-valid when the end of a file is mapped
1563 * into memory and the file's size is not page aligned.
1566 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
1572 * Scan the valid bits looking for invalid sections that
1573 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
1574 * valid bit may be set ) have already been zerod by
1575 * vm_page_set_validclean().
1577 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
1578 if (i == (PAGE_SIZE / DEV_BSIZE) ||
1579 (m->valid & (1 << i))
1582 pmap_zero_page_area(
1585 (i - b) << DEV_BSHIFT
1593 * setvalid is TRUE when we can safely set the zero'd areas
1594 * as being valid. We can do this if there are no cache consistency
1595 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
1598 m->valid = VM_PAGE_BITS_ALL;
1602 * Is a (partial) page valid? Note that the case where size == 0
1603 * will return FALSE in the degenerate case where the page is entirely
1604 * invalid, and TRUE otherwise.
1609 vm_page_is_valid(vm_page_t m, int base, int size)
1611 int bits = vm_page_bits(base, size);
1613 if (m->valid && ((m->valid & bits) == bits))
1620 * update dirty bits from pmap/mmu. May not block.
1623 vm_page_test_dirty(vm_page_t m)
1625 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
1631 * Issue an event on a VM page. Corresponding action structures are
1632 * removed from the page's list and called.
1635 vm_page_event_internal(vm_page_t m, vm_page_event_t event)
1637 struct vm_page_action *scan, *next;
1639 LIST_FOREACH_MUTABLE(scan, &m->action_list, entry, next) {
1640 if (scan->event == event) {
1641 scan->event = VMEVENT_NONE;
1642 LIST_REMOVE(scan, entry);
1643 scan->func(m, scan);
1648 #include "opt_ddb.h"
1650 #include <sys/kernel.h>
1652 #include <ddb/ddb.h>
1654 DB_SHOW_COMMAND(page, vm_page_print_page_info)
1656 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count);
1657 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count);
1658 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count);
1659 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count);
1660 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count);
1661 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved);
1662 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min);
1663 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target);
1664 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min);
1665 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target);
1668 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
1671 db_printf("PQ_FREE:");
1672 for(i=0;i<PQ_L2_SIZE;i++) {
1673 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
1677 db_printf("PQ_CACHE:");
1678 for(i=0;i<PQ_L2_SIZE;i++) {
1679 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
1683 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
1684 vm_page_queues[PQ_ACTIVE].lcnt,
1685 vm_page_queues[PQ_INACTIVE].lcnt);