4 * Copyright (c) 1991 Regents of the University of California.
7 * This code is derived from software contributed to Berkeley by
8 * The Mach Operating System project at Carnegie-Mellon University.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 4. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91
35 * $FreeBSD: src/sys/vm/vm_page.c,v 1.147.2.18 2002/03/10 05:03:19 alc Exp $
39 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40 * All rights reserved.
42 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
44 * Permission to use, copy, modify and distribute this software and
45 * its documentation is hereby granted, provided that both the copyright
46 * notice and this permission notice appear in all copies of the
47 * software, derivative works or modified versions, and any portions
48 * thereof, and that both notices appear in supporting documentation.
50 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
54 * Carnegie Mellon requests users of this software to return to
56 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
57 * School of Computer Science
58 * Carnegie Mellon University
59 * Pittsburgh PA 15213-3890
61 * any improvements or extensions that they make and grant Carnegie the
62 * rights to redistribute these changes.
65 * Resident memory management module. The module manipulates 'VM pages'.
66 * A VM page is the core building block for memory management.
69 #include <sys/param.h>
70 #include <sys/systm.h>
71 #include <sys/malloc.h>
73 #include <sys/vmmeter.h>
74 #include <sys/vnode.h>
75 #include <sys/kernel.h>
78 #include <vm/vm_param.h>
80 #include <vm/vm_kern.h>
82 #include <vm/vm_map.h>
83 #include <vm/vm_object.h>
84 #include <vm/vm_page.h>
85 #include <vm/vm_pageout.h>
86 #include <vm/vm_pager.h>
87 #include <vm/vm_extern.h>
88 #include <vm/swap_pager.h>
90 #include <machine/md_var.h>
92 #include <vm/vm_page2.h>
93 #include <sys/spinlock2.h>
95 #define VMACTION_HSIZE 256
96 #define VMACTION_HMASK (VMACTION_HSIZE - 1)
98 static void vm_page_queue_init(void);
99 static void vm_page_free_wakeup(void);
100 static vm_page_t vm_page_select_cache(vm_object_t, vm_pindex_t);
101 static vm_page_t _vm_page_list_find2(int basequeue, int index);
102 static void _vm_page_deactivate_locked(vm_page_t m, int athead);
105 * Array of tailq lists
107 __cachealign struct vpgqueues vm_page_queues[PQ_COUNT];
109 LIST_HEAD(vm_page_action_list, vm_page_action);
110 struct vm_page_action_list action_list[VMACTION_HSIZE];
111 static volatile int vm_pages_waiting;
114 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare,
115 vm_pindex_t, pindex);
118 vm_page_queue_init(void)
122 for (i = 0; i < PQ_L2_SIZE; i++)
123 vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count;
124 for (i = 0; i < PQ_L2_SIZE; i++)
125 vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count;
127 vm_page_queues[PQ_INACTIVE].cnt = &vmstats.v_inactive_count;
128 vm_page_queues[PQ_ACTIVE].cnt = &vmstats.v_active_count;
129 vm_page_queues[PQ_HOLD].cnt = &vmstats.v_active_count;
130 /* PQ_NONE has no queue */
132 for (i = 0; i < PQ_COUNT; i++) {
133 TAILQ_INIT(&vm_page_queues[i].pl);
134 spin_init(&vm_page_queues[i].spin);
137 for (i = 0; i < VMACTION_HSIZE; i++)
138 LIST_INIT(&action_list[i]);
142 * note: place in initialized data section? Is this necessary?
145 int vm_page_array_size = 0;
146 int vm_page_zero_count = 0;
147 vm_page_t vm_page_array = 0;
152 * Sets the page size, perhaps based upon the memory size.
153 * Must be called before any use of page-size dependent functions.
156 vm_set_page_size(void)
158 if (vmstats.v_page_size == 0)
159 vmstats.v_page_size = PAGE_SIZE;
160 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0)
161 panic("vm_set_page_size: page size not a power of two");
167 * Add a new page to the freelist for use by the system. New pages
168 * are added to both the head and tail of the associated free page
169 * queue in a bottom-up fashion, so both zero'd and non-zero'd page
170 * requests pull 'recent' adds (higher physical addresses) first.
172 * Must be called in a critical section.
175 vm_add_new_page(vm_paddr_t pa)
177 struct vpgqueues *vpq;
180 m = PHYS_TO_VM_PAGE(pa);
183 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK;
184 m->queue = m->pc + PQ_FREE;
185 KKASSERT(m->dirty == 0);
187 atomic_add_int(&vmstats.v_page_count, 1);
188 atomic_add_int(&vmstats.v_free_count, 1);
189 vpq = &vm_page_queues[m->queue];
191 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
193 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
194 vpq->flipflop = 1 - vpq->flipflop;
203 * Initializes the resident memory module.
205 * Preallocates memory for critical VM structures and arrays prior to
206 * kernel_map becoming available.
208 * Memory is allocated from (virtual2_start, virtual2_end) if available,
209 * otherwise memory is allocated from (virtual_start, virtual_end).
211 * On x86-64 (virtual_start, virtual_end) is only 2GB and may not be
212 * large enough to hold vm_page_array & other structures for machines with
213 * large amounts of ram, so we want to use virtual2* when available.
216 vm_page_startup(void)
218 vm_offset_t vaddr = virtual2_start ? virtual2_start : virtual_start;
221 vm_paddr_t page_range;
228 vm_paddr_t biggestone, biggestsize;
235 vaddr = round_page(vaddr);
237 for (i = 0; phys_avail[i + 1]; i += 2) {
238 phys_avail[i] = round_page64(phys_avail[i]);
239 phys_avail[i + 1] = trunc_page64(phys_avail[i + 1]);
242 for (i = 0; phys_avail[i + 1]; i += 2) {
243 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
245 if (size > biggestsize) {
253 end = phys_avail[biggestone+1];
254 end = trunc_page(end);
257 * Initialize the queue headers for the free queue, the active queue
258 * and the inactive queue.
261 vm_page_queue_init();
263 #if !defined(_KERNEL_VIRTUAL)
265 * VKERNELs don't support minidumps and as such don't need
268 * Allocate a bitmap to indicate that a random physical page
269 * needs to be included in a minidump.
271 * The amd64 port needs this to indicate which direct map pages
272 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
274 * However, i386 still needs this workspace internally within the
275 * minidump code. In theory, they are not needed on i386, but are
276 * included should the sf_buf code decide to use them.
278 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE;
279 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
280 end -= vm_page_dump_size;
281 vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size,
282 VM_PROT_READ | VM_PROT_WRITE);
283 bzero((void *)vm_page_dump, vm_page_dump_size);
287 * Compute the number of pages of memory that will be available for
288 * use (taking into account the overhead of a page structure per
291 first_page = phys_avail[0] / PAGE_SIZE;
292 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
293 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE;
296 * Initialize the mem entry structures now, and put them in the free
299 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
300 mapped = pmap_map(&vaddr, new_end, end,
301 VM_PROT_READ | VM_PROT_WRITE);
302 vm_page_array = (vm_page_t)mapped;
304 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL)
306 * since pmap_map on amd64 returns stuff out of a direct-map region,
307 * we have to manually add these pages to the minidump tracking so
308 * that they can be dumped, including the vm_page_array.
310 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
315 * Clear all of the page structures
317 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
318 vm_page_array_size = page_range;
321 * Construct the free queue(s) in ascending order (by physical
322 * address) so that the first 16MB of physical memory is allocated
323 * last rather than first. On large-memory machines, this avoids
324 * the exhaustion of low physical memory before isa_dmainit has run.
326 vmstats.v_page_count = 0;
327 vmstats.v_free_count = 0;
328 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
333 last_pa = phys_avail[i + 1];
334 while (pa < last_pa && npages-- > 0) {
340 virtual2_start = vaddr;
342 virtual_start = vaddr;
346 * Scan comparison function for Red-Black tree scans. An inclusive
347 * (start,end) is expected. Other fields are not used.
350 rb_vm_page_scancmp(struct vm_page *p, void *data)
352 struct rb_vm_page_scan_info *info = data;
354 if (p->pindex < info->start_pindex)
356 if (p->pindex > info->end_pindex)
362 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2)
364 if (p1->pindex < p2->pindex)
366 if (p1->pindex > p2->pindex)
372 * Each page queue has its own spin lock, which is fairly optimal for
373 * allocating and freeing pages at least.
375 * The caller must hold the vm_page_spin_lock() before locking a vm_page's
376 * queue spinlock via this function. Also note that m->queue cannot change
377 * unless both the page and queue are locked.
381 _vm_page_queue_spin_lock(vm_page_t m)
386 if (queue != PQ_NONE) {
387 spin_lock(&vm_page_queues[queue].spin);
388 KKASSERT(queue == m->queue);
394 _vm_page_queue_spin_unlock(vm_page_t m)
400 if (queue != PQ_NONE)
401 spin_unlock(&vm_page_queues[queue].spin);
406 _vm_page_queues_spin_lock(u_short queue)
409 if (queue != PQ_NONE)
410 spin_lock(&vm_page_queues[queue].spin);
416 _vm_page_queues_spin_unlock(u_short queue)
419 if (queue != PQ_NONE)
420 spin_unlock(&vm_page_queues[queue].spin);
424 vm_page_queue_spin_lock(vm_page_t m)
426 _vm_page_queue_spin_lock(m);
430 vm_page_queues_spin_lock(u_short queue)
432 _vm_page_queues_spin_lock(queue);
436 vm_page_queue_spin_unlock(vm_page_t m)
438 _vm_page_queue_spin_unlock(m);
442 vm_page_queues_spin_unlock(u_short queue)
444 _vm_page_queues_spin_unlock(queue);
448 * This locks the specified vm_page and its queue in the proper order
449 * (page first, then queue). The queue may change so the caller must
454 _vm_page_and_queue_spin_lock(vm_page_t m)
456 vm_page_spin_lock(m);
457 _vm_page_queue_spin_lock(m);
462 _vm_page_and_queue_spin_unlock(vm_page_t m)
464 _vm_page_queues_spin_unlock(m->queue);
465 vm_page_spin_unlock(m);
469 vm_page_and_queue_spin_unlock(vm_page_t m)
471 _vm_page_and_queue_spin_unlock(m);
475 vm_page_and_queue_spin_lock(vm_page_t m)
477 _vm_page_and_queue_spin_lock(m);
481 * Helper function removes vm_page from its current queue.
482 * Returns the base queue the page used to be on.
484 * The vm_page and the queue must be spinlocked.
485 * This function will unlock the queue but leave the page spinlocked.
487 static __inline u_short
488 _vm_page_rem_queue_spinlocked(vm_page_t m)
490 struct vpgqueues *pq;
494 if (queue != PQ_NONE) {
495 pq = &vm_page_queues[queue];
496 TAILQ_REMOVE(&pq->pl, m, pageq);
497 atomic_add_int(pq->cnt, -1);
500 if ((queue - m->pc) == PQ_FREE && (m->flags & PG_ZERO))
501 atomic_subtract_int(&vm_page_zero_count, 1);
502 vm_page_queues_spin_unlock(queue);
503 if ((queue - m->pc) == PQ_CACHE || (queue - m->pc) == PQ_FREE)
504 return (queue - m->pc);
510 * Helper function places the vm_page on the specified queue.
512 * The vm_page must be spinlocked.
513 * This function will return with both the page and the queue locked.
516 _vm_page_add_queue_spinlocked(vm_page_t m, u_short queue, int athead)
518 struct vpgqueues *pq;
520 KKASSERT(m->queue == PQ_NONE);
522 if (queue != PQ_NONE) {
523 vm_page_queues_spin_lock(queue);
524 pq = &vm_page_queues[queue];
526 atomic_add_int(pq->cnt, 1);
530 * Put zero'd pages on the end ( where we look for zero'd pages
531 * first ) and non-zerod pages at the head.
533 if (queue - m->pc == PQ_FREE) {
534 if (m->flags & PG_ZERO) {
535 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
536 atomic_add_int(&vm_page_zero_count, 1);
538 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
541 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
543 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
545 /* leave the queue spinlocked */
550 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE)
551 * m->busy is zero. Returns TRUE if it had to sleep, FALSE if we
552 * did not. Only one sleep call will be made before returning.
554 * This function does NOT busy the page and on return the page is not
555 * guaranteed to be available.
558 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg)
566 if ((flags & PG_BUSY) == 0 &&
567 (also_m_busy == 0 || (flags & PG_SBUSY) == 0)) {
570 tsleep_interlock(m, 0);
571 if (atomic_cmpset_int(&m->flags, flags,
572 flags | PG_WANTED | PG_REFERENCED)) {
573 tsleep(m, PINTERLOCKED, msg, 0);
580 * Wait until PG_BUSY can be set, then set it. If also_m_busy is TRUE we
581 * also wait for m->busy to become 0 before setting PG_BUSY.
584 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m,
585 int also_m_busy, const char *msg
593 if (flags & PG_BUSY) {
594 tsleep_interlock(m, 0);
595 if (atomic_cmpset_int(&m->flags, flags,
596 flags | PG_WANTED | PG_REFERENCED)) {
597 tsleep(m, PINTERLOCKED, msg, 0);
599 } else if (also_m_busy && (flags & PG_SBUSY)) {
600 tsleep_interlock(m, 0);
601 if (atomic_cmpset_int(&m->flags, flags,
602 flags | PG_WANTED | PG_REFERENCED)) {
603 tsleep(m, PINTERLOCKED, msg, 0);
606 if (atomic_cmpset_int(&m->flags, flags,
610 m->busy_line = lineno;
619 * Attempt to set PG_BUSY. If also_m_busy is TRUE we only succeed if m->busy
622 * Returns non-zero on failure.
625 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy
635 if (also_m_busy && (flags & PG_SBUSY))
637 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
640 m->busy_line = lineno;
648 * Clear the PG_BUSY flag and return non-zero to indicate to the caller
649 * that a wakeup() should be performed.
651 * The vm_page must be spinlocked and will remain spinlocked on return.
652 * The related queue must NOT be spinlocked (which could deadlock us).
658 _vm_page_wakeup(vm_page_t m)
665 if (atomic_cmpset_int(&m->flags, flags,
666 flags & ~(PG_BUSY | PG_WANTED))) {
670 return(flags & PG_WANTED);
674 * Clear the PG_BUSY flag and wakeup anyone waiting for the page. This
675 * is typically the last call you make on a page before moving onto
679 vm_page_wakeup(vm_page_t m)
681 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!"));
682 vm_page_spin_lock(m);
683 if (_vm_page_wakeup(m)) {
684 vm_page_spin_unlock(m);
687 vm_page_spin_unlock(m);
692 * Holding a page keeps it from being reused. Other parts of the system
693 * can still disassociate the page from its current object and free it, or
694 * perform read or write I/O on it and/or otherwise manipulate the page,
695 * but if the page is held the VM system will leave the page and its data
696 * intact and not reuse the page for other purposes until the last hold
697 * reference is released. (see vm_page_wire() if you want to prevent the
698 * page from being disassociated from its object too).
700 * The caller must still validate the contents of the page and, if necessary,
701 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete
702 * before manipulating the page.
704 * XXX get vm_page_spin_lock() here and move FREE->HOLD if necessary
707 vm_page_hold(vm_page_t m)
709 vm_page_spin_lock(m);
710 atomic_add_int(&m->hold_count, 1);
711 if (m->queue - m->pc == PQ_FREE) {
712 _vm_page_queue_spin_lock(m);
713 _vm_page_rem_queue_spinlocked(m);
714 _vm_page_add_queue_spinlocked(m, PQ_HOLD, 0);
715 _vm_page_queue_spin_unlock(m);
717 vm_page_spin_unlock(m);
721 * The opposite of vm_page_hold(). A page can be freed while being held,
722 * which places it on the PQ_HOLD queue. If we are able to busy the page
723 * after the hold count drops to zero we will move the page to the
724 * appropriate PQ_FREE queue by calling vm_page_free_toq().
727 vm_page_unhold(vm_page_t m)
729 vm_page_spin_lock(m);
730 atomic_add_int(&m->hold_count, -1);
731 if (m->hold_count == 0 && m->queue == PQ_HOLD) {
732 _vm_page_queue_spin_lock(m);
733 _vm_page_rem_queue_spinlocked(m);
734 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0);
735 _vm_page_queue_spin_unlock(m);
737 vm_page_spin_unlock(m);
741 * Inserts the given vm_page into the object and object list.
743 * The pagetables are not updated but will presumably fault the page
744 * in if necessary, or if a kernel page the caller will at some point
745 * enter the page into the kernel's pmap. We are not allowed to block
746 * here so we *can't* do this anyway.
748 * This routine may not block.
749 * This routine must be called with the vm_object held.
750 * This routine must be called with a critical section held.
753 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
755 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
756 if (m->object != NULL)
757 panic("vm_page_insert: already inserted");
759 object->generation++;
760 object->resident_page_count++;
763 * Record the object/offset pair in this page and add the
764 * pv_list_count of the page to the object.
766 * The vm_page spin lock is required for interactions with the pmap.
768 vm_page_spin_lock(m);
771 vm_page_rb_tree_RB_INSERT(&object->rb_memq, m);
772 atomic_add_int(&object->agg_pv_list_count, m->md.pv_list_count);
773 vm_page_spin_unlock(m);
776 * Since we are inserting a new and possibly dirty page,
777 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
779 if ((m->valid & m->dirty) || (m->flags & PG_WRITEABLE))
780 vm_object_set_writeable_dirty(object);
783 * Checks for a swap assignment and sets PG_SWAPPED if appropriate.
785 swap_pager_page_inserted(m);
789 * Removes the given vm_page_t from the (object,index) table
791 * The underlying pmap entry (if any) is NOT removed here.
792 * This routine may not block.
794 * The page must be BUSY and will remain BUSY on return.
795 * No other requirements.
797 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave
801 vm_page_remove(vm_page_t m)
805 if (m->object == NULL) {
809 if ((m->flags & PG_BUSY) == 0)
810 panic("vm_page_remove: page not busy");
814 vm_object_hold(object);
817 * Remove the page from the object and update the object.
819 * The vm_page spin lock is required for interactions with the pmap.
821 vm_page_spin_lock(m);
822 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m);
823 object->resident_page_count--;
824 atomic_add_int(&object->agg_pv_list_count, -m->md.pv_list_count);
826 vm_page_spin_unlock(m);
828 object->generation++;
830 vm_object_drop(object);
834 * Locate and return the page at (object, pindex), or NULL if the
835 * page could not be found.
837 * The caller must hold the vm_object token.
840 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
845 * Search the hash table for this object/offset pair
847 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
848 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
849 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex));
854 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object,
856 int also_m_busy, const char *msg
862 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
863 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
865 KKASSERT(m->object == object && m->pindex == pindex);
868 if (flags & PG_BUSY) {
869 tsleep_interlock(m, 0);
870 if (atomic_cmpset_int(&m->flags, flags,
871 flags | PG_WANTED | PG_REFERENCED)) {
872 tsleep(m, PINTERLOCKED, msg, 0);
873 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
876 } else if (also_m_busy && (flags & PG_SBUSY)) {
877 tsleep_interlock(m, 0);
878 if (atomic_cmpset_int(&m->flags, flags,
879 flags | PG_WANTED | PG_REFERENCED)) {
880 tsleep(m, PINTERLOCKED, msg, 0);
881 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
884 } else if (atomic_cmpset_int(&m->flags, flags,
888 m->busy_line = lineno;
897 * Attempt to lookup and busy a page.
899 * Returns NULL if the page could not be found
901 * Returns a vm_page and error == TRUE if the page exists but could not
904 * Returns a vm_page and error == FALSE on success.
907 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object,
909 int also_m_busy, int *errorp
915 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
916 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
919 KKASSERT(m->object == object && m->pindex == pindex);
922 if (flags & PG_BUSY) {
926 if (also_m_busy && (flags & PG_SBUSY)) {
930 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
933 m->busy_line = lineno;
942 * Caller must hold the related vm_object
945 vm_page_next(vm_page_t m)
949 next = vm_page_rb_tree_RB_NEXT(m);
950 if (next && next->pindex != m->pindex + 1)
958 * Move the given vm_page from its current object to the specified
959 * target object/offset. The page must be busy and will remain so
962 * new_object must be held.
963 * This routine might block. XXX ?
965 * NOTE: Swap associated with the page must be invalidated by the move. We
966 * have to do this for several reasons: (1) we aren't freeing the
967 * page, (2) we are dirtying the page, (3) the VM system is probably
968 * moving the page from object A to B, and will then later move
969 * the backing store from A to B and we can't have a conflict.
971 * NOTE: We *always* dirty the page. It is necessary both for the
972 * fact that we moved it, and because we may be invalidating
973 * swap. If the page is on the cache, we have to deactivate it
974 * or vm_page_dirty() will panic. Dirty pages are not allowed
978 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
980 KKASSERT(m->flags & PG_BUSY);
981 ASSERT_LWKT_TOKEN_HELD(vm_object_token(new_object));
983 ASSERT_LWKT_TOKEN_HELD(vm_object_token(m->object));
986 vm_page_insert(m, new_object, new_pindex);
987 if (m->queue - m->pc == PQ_CACHE)
988 vm_page_deactivate(m);
993 * vm_page_unqueue() without any wakeup. This routine is used when a page
994 * is being moved between queues or otherwise is to remain BUSYied by the
997 * This routine may not block.
1000 vm_page_unqueue_nowakeup(vm_page_t m)
1002 vm_page_and_queue_spin_lock(m);
1003 (void)_vm_page_rem_queue_spinlocked(m);
1004 vm_page_spin_unlock(m);
1008 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
1011 * This routine may not block.
1014 vm_page_unqueue(vm_page_t m)
1018 vm_page_and_queue_spin_lock(m);
1019 queue = _vm_page_rem_queue_spinlocked(m);
1020 if (queue == PQ_FREE || queue == PQ_CACHE) {
1021 vm_page_spin_unlock(m);
1022 pagedaemon_wakeup();
1024 vm_page_spin_unlock(m);
1029 * vm_page_list_find()
1031 * Find a page on the specified queue with color optimization.
1033 * The page coloring optimization attempts to locate a page that does
1034 * not overload other nearby pages in the object in the cpu's L1 or L2
1035 * caches. We need this optimization because cpu caches tend to be
1036 * physical caches, while object spaces tend to be virtual. This optimization
1037 * also gives us multiple queues and spinlocks to worth with on SMP systems.
1039 * The page is returned spinlocked and removed from its queue (it will
1040 * be on PQ_NONE), or NULL. The page is not PG_BUSY'd. The caller
1041 * is responsible for dealing with the busy-page case (usually by
1042 * deactivating the page and looping).
1044 * NOTE: This routine is carefully inlined. A non-inlined version
1045 * is available for outside callers but the only critical path is
1046 * from within this source file.
1048 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE
1049 * represent stable storage, allowing us to order our locks vm_page
1050 * first, then queue.
1054 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
1060 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist);
1062 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl);
1064 m = _vm_page_list_find2(basequeue, index);
1067 vm_page_and_queue_spin_lock(m);
1068 if (m->queue == basequeue + index) {
1069 _vm_page_rem_queue_spinlocked(m);
1070 /* vm_page_t spin held, no queue spin */
1073 vm_page_and_queue_spin_unlock(m);
1079 _vm_page_list_find2(int basequeue, int index)
1083 struct vpgqueues *pq;
1085 pq = &vm_page_queues[basequeue];
1088 * Note that for the first loop, index+i and index-i wind up at the
1089 * same place. Even though this is not totally optimal, we've already
1090 * blown it by missing the cache case so we do not care.
1092 for (i = PQ_L2_SIZE / 2; i > 0; --i) {
1094 m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl);
1096 _vm_page_and_queue_spin_lock(m);
1098 basequeue + ((index + i) & PQ_L2_MASK)) {
1099 _vm_page_rem_queue_spinlocked(m);
1102 _vm_page_and_queue_spin_unlock(m);
1105 m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl);
1107 _vm_page_and_queue_spin_lock(m);
1109 basequeue + ((index - i) & PQ_L2_MASK)) {
1110 _vm_page_rem_queue_spinlocked(m);
1113 _vm_page_and_queue_spin_unlock(m);
1123 * Returns a vm_page candidate for allocation. The page is not busied so
1124 * it can move around. The caller must busy the page (and typically
1125 * deactivate it if it cannot be busied!)
1127 * Returns a spinlocked vm_page that has been removed from its queue.
1130 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
1132 return(_vm_page_list_find(basequeue, index, prefer_zero));
1136 * Find a page on the cache queue with color optimization, remove it
1137 * from the queue, and busy it. The returned page will not be spinlocked.
1139 * A candidate failure will be deactivated. Candidates can fail due to
1140 * being busied by someone else, in which case they will be deactivated.
1142 * This routine may not block.
1146 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex)
1151 m = _vm_page_list_find(PQ_CACHE,
1152 (pindex + object->pg_color) & PQ_L2_MASK,
1157 * (m) has been removed from its queue and spinlocked
1159 if (vm_page_busy_try(m, TRUE)) {
1160 _vm_page_deactivate_locked(m, 0);
1161 vm_page_spin_unlock(m);
1163 kprintf("Warning: busy page %p found in cache\n", m);
1167 * We successfully busied the page
1169 if ((m->flags & PG_UNMANAGED) == 0 &&
1170 m->hold_count == 0 &&
1171 m->wire_count == 0) {
1172 vm_page_spin_unlock(m);
1173 pagedaemon_wakeup();
1176 _vm_page_deactivate_locked(m, 0);
1177 if (_vm_page_wakeup(m)) {
1178 vm_page_spin_unlock(m);
1181 vm_page_spin_unlock(m);
1189 * Find a free or zero page, with specified preference. We attempt to
1190 * inline the nominal case and fall back to _vm_page_select_free()
1191 * otherwise. A busied page is removed from the queue and returned.
1193 * This routine may not block.
1195 static __inline vm_page_t
1196 vm_page_select_free(vm_object_t object, vm_pindex_t pindex,
1197 boolean_t prefer_zero)
1202 m = _vm_page_list_find(PQ_FREE,
1203 (pindex + object->pg_color) & PQ_L2_MASK,
1207 if (vm_page_busy_try(m, TRUE)) {
1208 _vm_page_deactivate_locked(m, 0);
1209 vm_page_spin_unlock(m);
1211 kprintf("Warning: busy page %p found in cache\n", m);
1214 KKASSERT((m->flags & PG_UNMANAGED) == 0);
1215 KKASSERT(m->hold_count == 0);
1216 KKASSERT(m->wire_count == 0);
1217 vm_page_spin_unlock(m);
1218 pagedaemon_wakeup();
1220 /* return busied and removed page */
1230 * Allocate and return a memory cell associated with this VM object/offset
1235 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain
1236 * VM_ALLOC_QUICK like normal but cannot use cache
1237 * VM_ALLOC_SYSTEM greater free drain
1238 * VM_ALLOC_INTERRUPT allow free list to be completely drained
1239 * VM_ALLOC_ZERO advisory request for pre-zero'd page
1241 * The object must be locked.
1242 * This routine may not block.
1243 * The returned page will be marked PG_BUSY
1245 * Additional special handling is required when called from an interrupt
1246 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache
1250 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
1254 KKASSERT(object != NULL);
1255 KASSERT(!vm_page_lookup(object, pindex),
1256 ("vm_page_alloc: page already allocated"));
1258 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK|
1259 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
1262 * Certain system threads (pageout daemon, buf_daemon's) are
1263 * allowed to eat deeper into the free page list.
1265 if (curthread->td_flags & TDF_SYSTHREAD)
1266 page_req |= VM_ALLOC_SYSTEM;
1269 if (vmstats.v_free_count > vmstats.v_free_reserved ||
1270 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) ||
1271 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 &&
1272 vmstats.v_free_count > vmstats.v_interrupt_free_min)
1275 * The free queue has sufficient free pages to take one out.
1277 if (page_req & VM_ALLOC_ZERO)
1278 m = vm_page_select_free(object, pindex, TRUE);
1280 m = vm_page_select_free(object, pindex, FALSE);
1281 } else if (page_req & VM_ALLOC_NORMAL) {
1283 * Allocatable from the cache (non-interrupt only). On
1284 * success, we must free the page and try again, thus
1285 * ensuring that vmstats.v_*_free_min counters are replenished.
1288 if (curthread->td_preempted) {
1289 kprintf("vm_page_alloc(): warning, attempt to allocate"
1290 " cache page from preempting interrupt\n");
1293 m = vm_page_select_cache(object, pindex);
1296 m = vm_page_select_cache(object, pindex);
1299 * On success move the page into the free queue and loop.
1302 KASSERT(m->dirty == 0,
1303 ("Found dirty cache page %p", m));
1304 vm_page_protect(m, VM_PROT_NONE);
1310 * On failure return NULL
1312 #if defined(DIAGNOSTIC)
1313 if (vmstats.v_cache_count > 0)
1314 kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count);
1316 vm_pageout_deficit++;
1317 pagedaemon_wakeup();
1321 * No pages available, wakeup the pageout daemon and give up.
1323 vm_pageout_deficit++;
1324 pagedaemon_wakeup();
1329 * Good page found. The page has already been busied for us.
1331 * v_free_count can race so loop if we don't find the expected
1336 KASSERT(m->dirty == 0,
1337 ("vm_page_alloc: free/cache page %p was dirty", m));
1340 * NOTE: page has already been removed from its queue and busied.
1342 KKASSERT(m->queue == PQ_NONE);
1345 * Initialize structure. Only the PG_ZERO flag is inherited. Set
1348 vm_page_flag_clear(m, ~(PG_ZERO | PG_BUSY));
1349 KKASSERT(m->wire_count == 0);
1350 KKASSERT(m->busy == 0);
1355 * Caller must be holding the object lock (asserted by
1356 * vm_page_insert()).
1358 * NOTE: Inserting a page here does not insert it into any pmaps
1359 * (which could cause us to block allocating memory).
1361 vm_page_insert(m, object, pindex);
1364 * Don't wakeup too often - wakeup the pageout daemon when
1365 * we would be nearly out of memory.
1367 pagedaemon_wakeup();
1370 * A PG_BUSY page is returned.
1376 * Wait for sufficient free memory for nominal heavy memory use kernel
1380 vm_wait_nominal(void)
1382 while (vm_page_count_min(0))
1387 * Test if vm_wait_nominal() would block.
1390 vm_test_nominal(void)
1392 if (vm_page_count_min(0))
1398 * Block until free pages are available for allocation, called in various
1399 * places before memory allocations.
1401 * The caller may loop if vm_page_count_min() == FALSE so we cannot be
1402 * more generous then that.
1408 * never wait forever
1412 lwkt_gettoken(&vm_token);
1414 if (curthread == pagethread) {
1416 * The pageout daemon itself needs pages, this is bad.
1418 if (vm_page_count_min(0)) {
1419 vm_pageout_pages_needed = 1;
1420 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo);
1424 * Wakeup the pageout daemon if necessary and wait.
1426 if (vm_page_count_target()) {
1427 if (vm_pages_needed == 0) {
1428 vm_pages_needed = 1;
1429 wakeup(&vm_pages_needed);
1431 ++vm_pages_waiting; /* SMP race ok */
1432 tsleep(&vmstats.v_free_count, 0, "vmwait", timo);
1435 lwkt_reltoken(&vm_token);
1439 * Block until free pages are available for allocation
1441 * Called only from vm_fault so that processes page faulting can be
1448 * Wakeup the pageout daemon if necessary and wait.
1450 if (vm_page_count_target()) {
1451 lwkt_gettoken(&vm_token);
1452 if (vm_page_count_target()) {
1453 if (vm_pages_needed == 0) {
1454 vm_pages_needed = 1;
1455 wakeup(&vm_pages_needed);
1457 ++vm_pages_waiting; /* SMP race ok */
1458 tsleep(&vmstats.v_free_count, 0, "pfault", hz);
1460 lwkt_reltoken(&vm_token);
1465 * Put the specified page on the active list (if appropriate). Ensure
1466 * that act_count is at least ACT_INIT but do not otherwise mess with it.
1468 * The caller should be holding the page busied ? XXX
1469 * This routine may not block.
1472 vm_page_activate(vm_page_t m)
1476 vm_page_spin_lock(m);
1477 if (m->queue != PQ_ACTIVE) {
1478 _vm_page_queue_spin_lock(m);
1479 oqueue = _vm_page_rem_queue_spinlocked(m);
1480 /* page is left spinlocked, queue is unlocked */
1482 if (oqueue == PQ_CACHE)
1483 mycpu->gd_cnt.v_reactivated++;
1484 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1485 if (m->act_count < ACT_INIT)
1486 m->act_count = ACT_INIT;
1487 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE, 0);
1489 _vm_page_and_queue_spin_unlock(m);
1490 if (oqueue == PQ_CACHE || oqueue == PQ_FREE)
1491 pagedaemon_wakeup();
1493 if (m->act_count < ACT_INIT)
1494 m->act_count = ACT_INIT;
1495 vm_page_spin_unlock(m);
1500 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
1501 * routine is called when a page has been added to the cache or free
1504 * This routine may not block.
1506 static __inline void
1507 vm_page_free_wakeup(void)
1510 * If the pageout daemon itself needs pages, then tell it that
1511 * there are some free.
1513 if (vm_pageout_pages_needed &&
1514 vmstats.v_cache_count + vmstats.v_free_count >=
1515 vmstats.v_pageout_free_min
1517 wakeup(&vm_pageout_pages_needed);
1518 vm_pageout_pages_needed = 0;
1522 * Wakeup processes that are waiting on memory.
1524 * NOTE: vm_paging_target() is the pageout daemon's target, while
1525 * vm_page_count_target() is somewhere inbetween. We want
1526 * to wake processes up prior to the pageout daemon reaching
1527 * its target to provide some hysteresis.
1529 if (vm_pages_waiting) {
1530 if (!vm_page_count_target()) {
1532 * Plenty of pages are free, wakeup everyone.
1534 vm_pages_waiting = 0;
1535 wakeup(&vmstats.v_free_count);
1536 ++mycpu->gd_cnt.v_ppwakeups;
1537 } else if (!vm_page_count_min(0)) {
1539 * Some pages are free, wakeup someone.
1541 int wcount = vm_pages_waiting;
1544 vm_pages_waiting = wcount;
1545 wakeup_one(&vmstats.v_free_count);
1546 ++mycpu->gd_cnt.v_ppwakeups;
1552 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates
1553 * it from its VM object.
1555 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on
1556 * return (the page will have been freed).
1559 vm_page_free_toq(vm_page_t m)
1561 mycpu->gd_cnt.v_tfree++;
1562 KKASSERT((m->flags & PG_MAPPED) == 0);
1563 KKASSERT(m->flags & PG_BUSY);
1565 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
1567 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
1568 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
1570 if ((m->queue - m->pc) == PQ_FREE)
1571 panic("vm_page_free: freeing free page");
1573 panic("vm_page_free: freeing busy page");
1577 * Remove from object, spinlock the page and its queues and
1578 * remove from any queue. No queue spinlock will be held
1579 * after this section (because the page was removed from any
1583 vm_page_and_queue_spin_lock(m);
1584 _vm_page_rem_queue_spinlocked(m);
1587 * No further management of fictitious pages occurs beyond object
1588 * and queue removal.
1590 if ((m->flags & PG_FICTITIOUS) != 0) {
1591 vm_page_spin_unlock(m);
1599 if (m->wire_count != 0) {
1600 if (m->wire_count > 1) {
1602 "vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1603 m->wire_count, (long)m->pindex);
1605 panic("vm_page_free: freeing wired page");
1609 * Clear the UNMANAGED flag when freeing an unmanaged page.
1611 if (m->flags & PG_UNMANAGED) {
1612 vm_page_flag_clear(m, PG_UNMANAGED);
1615 if (m->hold_count != 0) {
1616 vm_page_flag_clear(m, PG_ZERO);
1617 _vm_page_add_queue_spinlocked(m, PQ_HOLD, 0);
1619 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0);
1623 * This sequence allows us to clear PG_BUSY while still holding
1624 * its spin lock, which reduces contention vs allocators. We
1625 * must not leave the queue locked or _vm_page_wakeup() may
1628 _vm_page_queue_spin_unlock(m);
1629 if (_vm_page_wakeup(m)) {
1630 vm_page_spin_unlock(m);
1633 vm_page_spin_unlock(m);
1635 vm_page_free_wakeup();
1639 * vm_page_free_fromq_fast()
1641 * Remove a non-zero page from one of the free queues; the page is removed for
1642 * zeroing, so do not issue a wakeup.
1645 vm_page_free_fromq_fast(void)
1651 for (i = 0; i < PQ_L2_SIZE; ++i) {
1652 m = vm_page_list_find(PQ_FREE, qi, FALSE);
1653 /* page is returned spinlocked and removed from its queue */
1655 if (vm_page_busy_try(m, TRUE)) {
1657 * We were unable to busy the page, deactivate
1660 _vm_page_deactivate_locked(m, 0);
1661 vm_page_spin_unlock(m);
1662 } else if ((m->flags & PG_ZERO) == 0) {
1664 * The page is not PG_ZERO'd so return it.
1666 vm_page_spin_unlock(m);
1670 * The page is PG_ZERO, requeue it and loop
1672 _vm_page_add_queue_spinlocked(m,
1675 vm_page_queue_spin_unlock(m);
1676 if (_vm_page_wakeup(m)) {
1677 vm_page_spin_unlock(m);
1680 vm_page_spin_unlock(m);
1685 qi = (qi + PQ_PRIME2) & PQ_L2_MASK;
1691 * vm_page_unmanage()
1693 * Prevent PV management from being done on the page. The page is
1694 * removed from the paging queues as if it were wired, and as a
1695 * consequence of no longer being managed the pageout daemon will not
1696 * touch it (since there is no way to locate the pte mappings for the
1697 * page). madvise() calls that mess with the pmap will also no longer
1698 * operate on the page.
1700 * Beyond that the page is still reasonably 'normal'. Freeing the page
1701 * will clear the flag.
1703 * This routine is used by OBJT_PHYS objects - objects using unswappable
1704 * physical memory as backing store rather then swap-backed memory and
1705 * will eventually be extended to support 4MB unmanaged physical
1708 * Caller must be holding the page busy.
1711 vm_page_unmanage(vm_page_t m)
1713 KKASSERT(m->flags & PG_BUSY);
1714 if ((m->flags & PG_UNMANAGED) == 0) {
1715 if (m->wire_count == 0)
1718 vm_page_flag_set(m, PG_UNMANAGED);
1722 * Mark this page as wired down by yet another map, removing it from
1723 * paging queues as necessary.
1725 * Caller must be holding the page busy.
1728 vm_page_wire(vm_page_t m)
1731 * Only bump the wire statistics if the page is not already wired,
1732 * and only unqueue the page if it is on some queue (if it is unmanaged
1733 * it is already off the queues). Don't do anything with fictitious
1734 * pages because they are always wired.
1736 KKASSERT(m->flags & PG_BUSY);
1737 if ((m->flags & PG_FICTITIOUS) == 0) {
1738 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) {
1739 if ((m->flags & PG_UNMANAGED) == 0)
1741 atomic_add_int(&vmstats.v_wire_count, 1);
1743 KASSERT(m->wire_count != 0,
1744 ("vm_page_wire: wire_count overflow m=%p", m));
1749 * Release one wiring of this page, potentially enabling it to be paged again.
1751 * Many pages placed on the inactive queue should actually go
1752 * into the cache, but it is difficult to figure out which. What
1753 * we do instead, if the inactive target is well met, is to put
1754 * clean pages at the head of the inactive queue instead of the tail.
1755 * This will cause them to be moved to the cache more quickly and
1756 * if not actively re-referenced, freed more quickly. If we just
1757 * stick these pages at the end of the inactive queue, heavy filesystem
1758 * meta-data accesses can cause an unnecessary paging load on memory bound
1759 * processes. This optimization causes one-time-use metadata to be
1760 * reused more quickly.
1762 * BUT, if we are in a low-memory situation we have no choice but to
1763 * put clean pages on the cache queue.
1765 * A number of routines use vm_page_unwire() to guarantee that the page
1766 * will go into either the inactive or active queues, and will NEVER
1767 * be placed in the cache - for example, just after dirtying a page.
1768 * dirty pages in the cache are not allowed.
1770 * The page queues must be locked.
1771 * This routine may not block.
1774 vm_page_unwire(vm_page_t m, int activate)
1776 KKASSERT(m->flags & PG_BUSY);
1777 if (m->flags & PG_FICTITIOUS) {
1779 } else if (m->wire_count <= 0) {
1780 panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
1782 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) {
1783 atomic_add_int(&vmstats.v_wire_count, -1);
1784 if (m->flags & PG_UNMANAGED) {
1786 } else if (activate) {
1787 vm_page_spin_lock(m);
1788 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE, 0);
1789 _vm_page_and_queue_spin_unlock(m);
1791 vm_page_spin_lock(m);
1792 vm_page_flag_clear(m, PG_WINATCFLS);
1793 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE,
1795 ++vm_swapcache_inactive_heuristic;
1796 _vm_page_and_queue_spin_unlock(m);
1803 * Move the specified page to the inactive queue. If the page has
1804 * any associated swap, the swap is deallocated.
1806 * Normally athead is 0 resulting in LRU operation. athead is set
1807 * to 1 if we want this page to be 'as if it were placed in the cache',
1808 * except without unmapping it from the process address space.
1810 * vm_page's spinlock must be held on entry and will remain held on return.
1811 * This routine may not block.
1814 _vm_page_deactivate_locked(vm_page_t m, int athead)
1819 * Ignore if already inactive.
1821 if (m->queue == PQ_INACTIVE)
1823 _vm_page_queue_spin_lock(m);
1824 oqueue = _vm_page_rem_queue_spinlocked(m);
1826 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1827 if (oqueue == PQ_CACHE)
1828 mycpu->gd_cnt.v_reactivated++;
1829 vm_page_flag_clear(m, PG_WINATCFLS);
1830 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE, athead);
1832 ++vm_swapcache_inactive_heuristic;
1834 _vm_page_queue_spin_unlock(m);
1835 /* leaves vm_page spinlocked */
1839 * Attempt to deactivate a page.
1844 vm_page_deactivate(vm_page_t m)
1846 vm_page_spin_lock(m);
1847 _vm_page_deactivate_locked(m, 0);
1848 vm_page_spin_unlock(m);
1852 vm_page_deactivate_locked(vm_page_t m)
1854 _vm_page_deactivate_locked(m, 0);
1858 * Attempt to move a page to PQ_CACHE.
1860 * Returns 0 on failure, 1 on success
1862 * The page should NOT be busied by the caller. This function will validate
1863 * whether the page can be safely moved to the cache.
1866 vm_page_try_to_cache(vm_page_t m)
1868 vm_page_spin_lock(m);
1869 if (vm_page_busy_try(m, TRUE)) {
1870 vm_page_spin_unlock(m);
1873 if (m->dirty || m->hold_count || m->wire_count ||
1874 (m->flags & PG_UNMANAGED)) {
1875 if (_vm_page_wakeup(m)) {
1876 vm_page_spin_unlock(m);
1879 vm_page_spin_unlock(m);
1883 vm_page_spin_unlock(m);
1886 * Page busied by us and no longer spinlocked. Dirty pages cannot
1887 * be moved to the cache.
1889 vm_page_test_dirty(m);
1899 * Attempt to free the page. If we cannot free it, we do nothing.
1900 * 1 is returned on success, 0 on failure.
1905 vm_page_try_to_free(vm_page_t m)
1907 vm_page_spin_lock(m);
1908 if (vm_page_busy_try(m, TRUE)) {
1909 vm_page_spin_unlock(m);
1912 if (m->dirty || m->hold_count || m->wire_count ||
1913 (m->flags & PG_UNMANAGED)) {
1914 if (_vm_page_wakeup(m)) {
1915 vm_page_spin_unlock(m);
1918 vm_page_spin_unlock(m);
1922 vm_page_spin_unlock(m);
1925 * Page busied by us and no longer spinlocked. Dirty pages will
1926 * not be freed by this function. We have to re-test the
1927 * dirty bit after cleaning out the pmaps.
1929 vm_page_test_dirty(m);
1934 vm_page_protect(m, VM_PROT_NONE);
1946 * Put the specified page onto the page cache queue (if appropriate).
1948 * The page must be busy, and this routine will release the busy and
1949 * possibly even free the page.
1952 vm_page_cache(vm_page_t m)
1954 if ((m->flags & PG_UNMANAGED) || m->busy ||
1955 m->wire_count || m->hold_count) {
1956 kprintf("vm_page_cache: attempting to cache busy/held page\n");
1962 * Already in the cache (and thus not mapped)
1964 if ((m->queue - m->pc) == PQ_CACHE) {
1965 KKASSERT((m->flags & PG_MAPPED) == 0);
1971 * Caller is required to test m->dirty, but note that the act of
1972 * removing the page from its maps can cause it to become dirty
1973 * on an SMP system due to another cpu running in usermode.
1976 panic("vm_page_cache: caching a dirty page, pindex: %ld",
1981 * Remove all pmaps and indicate that the page is not
1982 * writeable or mapped. Our vm_page_protect() call may
1983 * have blocked (especially w/ VM_PROT_NONE), so recheck
1986 vm_page_protect(m, VM_PROT_NONE);
1987 if ((m->flags & (PG_UNMANAGED|PG_MAPPED)) || m->busy ||
1988 m->wire_count || m->hold_count) {
1990 } else if (m->dirty) {
1991 vm_page_deactivate(m);
1994 _vm_page_and_queue_spin_lock(m);
1995 _vm_page_rem_queue_spinlocked(m);
1996 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0);
1997 _vm_page_queue_spin_unlock(m);
1998 if (_vm_page_wakeup(m)) {
1999 vm_page_spin_unlock(m);
2002 vm_page_spin_unlock(m);
2004 vm_page_free_wakeup();
2009 * vm_page_dontneed()
2011 * Cache, deactivate, or do nothing as appropriate. This routine
2012 * is typically used by madvise() MADV_DONTNEED.
2014 * Generally speaking we want to move the page into the cache so
2015 * it gets reused quickly. However, this can result in a silly syndrome
2016 * due to the page recycling too quickly. Small objects will not be
2017 * fully cached. On the otherhand, if we move the page to the inactive
2018 * queue we wind up with a problem whereby very large objects
2019 * unnecessarily blow away our inactive and cache queues.
2021 * The solution is to move the pages based on a fixed weighting. We
2022 * either leave them alone, deactivate them, or move them to the cache,
2023 * where moving them to the cache has the highest weighting.
2024 * By forcing some pages into other queues we eventually force the
2025 * system to balance the queues, potentially recovering other unrelated
2026 * space from active. The idea is to not force this to happen too
2029 * The page must be busied.
2032 vm_page_dontneed(vm_page_t m)
2034 static int dnweight;
2041 * occassionally leave the page alone
2043 if ((dnw & 0x01F0) == 0 ||
2044 m->queue == PQ_INACTIVE ||
2045 m->queue - m->pc == PQ_CACHE
2047 if (m->act_count >= ACT_INIT)
2053 * If vm_page_dontneed() is inactivating a page, it must clear
2054 * the referenced flag; otherwise the pagedaemon will see references
2055 * on the page in the inactive queue and reactivate it. Until the
2056 * page can move to the cache queue, madvise's job is not done.
2058 vm_page_flag_clear(m, PG_REFERENCED);
2059 pmap_clear_reference(m);
2062 vm_page_test_dirty(m);
2064 if (m->dirty || (dnw & 0x0070) == 0) {
2066 * Deactivate the page 3 times out of 32.
2071 * Cache the page 28 times out of every 32. Note that
2072 * the page is deactivated instead of cached, but placed
2073 * at the head of the queue instead of the tail.
2077 vm_page_spin_lock(m);
2078 _vm_page_deactivate_locked(m, head);
2079 vm_page_spin_unlock(m);
2083 * These routines manipulate the 'soft busy' count for a page. A soft busy
2084 * is almost like PG_BUSY except that it allows certain compatible operations
2085 * to occur on the page while it is busy. For example, a page undergoing a
2086 * write can still be mapped read-only.
2088 * Because vm_pages can overlap buffers m->busy can be > 1. m->busy is only
2089 * adjusted while the vm_page is PG_BUSY so the flash will occur when the
2090 * busy bit is cleared.
2093 vm_page_io_start(vm_page_t m)
2095 KASSERT(m->flags & PG_BUSY, ("vm_page_io_start: page not busy!!!"));
2096 atomic_add_char(&m->busy, 1);
2097 vm_page_flag_set(m, PG_SBUSY);
2101 vm_page_io_finish(vm_page_t m)
2103 KASSERT(m->flags & PG_BUSY, ("vm_page_io_finish: page not busy!!!"));
2104 atomic_subtract_char(&m->busy, 1);
2106 vm_page_flag_clear(m, PG_SBUSY);
2110 * Grab a page, blocking if it is busy and allocating a page if necessary.
2111 * A busy page is returned or NULL.
2113 * The page is not removed from its queues. XXX?
2115 * If VM_ALLOC_RETRY is specified VM_ALLOC_NORMAL must also be specified.
2116 * If VM_ALLOC_RETRY is not specified
2118 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
2119 * always returned if we had blocked.
2120 * This routine will never return NULL if VM_ALLOC_RETRY is set.
2121 * This routine may not be called from an interrupt.
2122 * The returned page may not be entirely valid.
2124 * This routine may be called from mainline code without spl protection and
2125 * be guarenteed a busied page associated with the object at the specified
2131 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2136 KKASSERT(allocflags &
2137 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
2138 vm_object_hold(object);
2140 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error);
2142 vm_page_sleep_busy(m, TRUE, "pgrbwt");
2143 if ((allocflags & VM_ALLOC_RETRY) == 0) {
2147 } else if (m == NULL) {
2148 m = vm_page_alloc(object, pindex,
2149 allocflags & ~VM_ALLOC_RETRY);
2153 if ((allocflags & VM_ALLOC_RETRY) == 0)
2160 vm_object_drop(object);
2165 * Mapping function for valid bits or for dirty bits in
2166 * a page. May not block.
2168 * Inputs are required to range within a page.
2174 vm_page_bits(int base, int size)
2180 base + size <= PAGE_SIZE,
2181 ("vm_page_bits: illegal base/size %d/%d", base, size)
2184 if (size == 0) /* handle degenerate case */
2187 first_bit = base >> DEV_BSHIFT;
2188 last_bit = (base + size - 1) >> DEV_BSHIFT;
2190 return ((2 << last_bit) - (1 << first_bit));
2194 * Sets portions of a page valid and clean. The arguments are expected
2195 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2196 * of any partial chunks touched by the range. The invalid portion of
2197 * such chunks will be zero'd.
2199 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically
2200 * align base to DEV_BSIZE so as not to mark clean a partially
2201 * truncated device block. Otherwise the dirty page status might be
2204 * This routine may not block.
2206 * (base + size) must be less then or equal to PAGE_SIZE.
2209 _vm_page_zero_valid(vm_page_t m, int base, int size)
2214 if (size == 0) /* handle degenerate case */
2218 * If the base is not DEV_BSIZE aligned and the valid
2219 * bit is clear, we have to zero out a portion of the
2223 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2224 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
2226 pmap_zero_page_area(
2234 * If the ending offset is not DEV_BSIZE aligned and the
2235 * valid bit is clear, we have to zero out a portion of
2239 endoff = base + size;
2241 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2242 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
2244 pmap_zero_page_area(
2247 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
2253 * Set valid, clear dirty bits. If validating the entire
2254 * page we can safely clear the pmap modify bit. We also
2255 * use this opportunity to clear the PG_NOSYNC flag. If a process
2256 * takes a write fault on a MAP_NOSYNC memory area the flag will
2259 * We set valid bits inclusive of any overlap, but we can only
2260 * clear dirty bits for DEV_BSIZE chunks that are fully within
2263 * Page must be busied?
2264 * No other requirements.
2267 vm_page_set_valid(vm_page_t m, int base, int size)
2269 _vm_page_zero_valid(m, base, size);
2270 m->valid |= vm_page_bits(base, size);
2275 * Set valid bits and clear dirty bits.
2277 * NOTE: This function does not clear the pmap modified bit.
2278 * Also note that e.g. NFS may use a byte-granular base
2281 * WARNING: Page must be busied? But vfs_clean_one_page() will call
2282 * this without necessarily busying the page (via bdwrite()).
2283 * So for now vm_token must also be held.
2285 * No other requirements.
2288 vm_page_set_validclean(vm_page_t m, int base, int size)
2292 _vm_page_zero_valid(m, base, size);
2293 pagebits = vm_page_bits(base, size);
2294 m->valid |= pagebits;
2295 m->dirty &= ~pagebits;
2296 if (base == 0 && size == PAGE_SIZE) {
2297 /*pmap_clear_modify(m);*/
2298 vm_page_flag_clear(m, PG_NOSYNC);
2303 * Set valid & dirty. Used by buwrite()
2305 * WARNING: Page must be busied? But vfs_dirty_one_page() will
2306 * call this function in buwrite() so for now vm_token must
2309 * No other requirements.
2312 vm_page_set_validdirty(vm_page_t m, int base, int size)
2316 pagebits = vm_page_bits(base, size);
2317 m->valid |= pagebits;
2318 m->dirty |= pagebits;
2320 vm_object_set_writeable_dirty(m->object);
2326 * NOTE: This function does not clear the pmap modified bit.
2327 * Also note that e.g. NFS may use a byte-granular base
2330 * Page must be busied?
2331 * No other requirements.
2334 vm_page_clear_dirty(vm_page_t m, int base, int size)
2336 m->dirty &= ~vm_page_bits(base, size);
2337 if (base == 0 && size == PAGE_SIZE) {
2338 /*pmap_clear_modify(m);*/
2339 vm_page_flag_clear(m, PG_NOSYNC);
2344 * Make the page all-dirty.
2346 * Also make sure the related object and vnode reflect the fact that the
2347 * object may now contain a dirty page.
2349 * Page must be busied?
2350 * No other requirements.
2353 vm_page_dirty(vm_page_t m)
2356 int pqtype = m->queue - m->pc;
2358 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE,
2359 ("vm_page_dirty: page in free/cache queue!"));
2360 if (m->dirty != VM_PAGE_BITS_ALL) {
2361 m->dirty = VM_PAGE_BITS_ALL;
2363 vm_object_set_writeable_dirty(m->object);
2368 * Invalidates DEV_BSIZE'd chunks within a page. Both the
2369 * valid and dirty bits for the effected areas are cleared.
2371 * Page must be busied?
2373 * No other requirements.
2376 vm_page_set_invalid(vm_page_t m, int base, int size)
2380 bits = vm_page_bits(base, size);
2383 m->object->generation++;
2387 * The kernel assumes that the invalid portions of a page contain
2388 * garbage, but such pages can be mapped into memory by user code.
2389 * When this occurs, we must zero out the non-valid portions of the
2390 * page so user code sees what it expects.
2392 * Pages are most often semi-valid when the end of a file is mapped
2393 * into memory and the file's size is not page aligned.
2395 * Page must be busied?
2396 * No other requirements.
2399 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
2405 * Scan the valid bits looking for invalid sections that
2406 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
2407 * valid bit may be set ) have already been zerod by
2408 * vm_page_set_validclean().
2410 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
2411 if (i == (PAGE_SIZE / DEV_BSIZE) ||
2412 (m->valid & (1 << i))
2415 pmap_zero_page_area(
2418 (i - b) << DEV_BSHIFT
2426 * setvalid is TRUE when we can safely set the zero'd areas
2427 * as being valid. We can do this if there are no cache consistency
2428 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
2431 m->valid = VM_PAGE_BITS_ALL;
2435 * Is a (partial) page valid? Note that the case where size == 0
2436 * will return FALSE in the degenerate case where the page is entirely
2437 * invalid, and TRUE otherwise.
2440 * No other requirements.
2443 vm_page_is_valid(vm_page_t m, int base, int size)
2445 int bits = vm_page_bits(base, size);
2447 if (m->valid && ((m->valid & bits) == bits))
2454 * update dirty bits from pmap/mmu. May not block.
2456 * Caller must hold the page busy
2459 vm_page_test_dirty(vm_page_t m)
2461 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
2467 * Register an action, associating it with its vm_page
2470 vm_page_register_action(vm_page_action_t action, vm_page_event_t event)
2472 struct vm_page_action_list *list;
2475 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
2476 list = &action_list[hv];
2478 lwkt_gettoken(&vm_token);
2479 vm_page_flag_set(action->m, PG_ACTIONLIST);
2480 action->event = event;
2481 LIST_INSERT_HEAD(list, action, entry);
2482 lwkt_reltoken(&vm_token);
2486 * Unregister an action, disassociating it from its related vm_page
2489 vm_page_unregister_action(vm_page_action_t action)
2491 struct vm_page_action_list *list;
2494 lwkt_gettoken(&vm_token);
2495 if (action->event != VMEVENT_NONE) {
2496 action->event = VMEVENT_NONE;
2497 LIST_REMOVE(action, entry);
2499 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
2500 list = &action_list[hv];
2501 if (LIST_EMPTY(list))
2502 vm_page_flag_clear(action->m, PG_ACTIONLIST);
2504 lwkt_reltoken(&vm_token);
2508 * Issue an event on a VM page. Corresponding action structures are
2509 * removed from the page's list and called.
2511 * If the vm_page has no more pending action events we clear its
2512 * PG_ACTIONLIST flag.
2515 vm_page_event_internal(vm_page_t m, vm_page_event_t event)
2517 struct vm_page_action_list *list;
2518 struct vm_page_action *scan;
2519 struct vm_page_action *next;
2523 hv = (int)((intptr_t)m >> 8) & VMACTION_HMASK;
2524 list = &action_list[hv];
2527 lwkt_gettoken(&vm_token);
2528 LIST_FOREACH_MUTABLE(scan, list, entry, next) {
2530 if (scan->event == event) {
2531 scan->event = VMEVENT_NONE;
2532 LIST_REMOVE(scan, entry);
2533 scan->func(m, scan);
2541 vm_page_flag_clear(m, PG_ACTIONLIST);
2542 lwkt_reltoken(&vm_token);
2545 #include "opt_ddb.h"
2547 #include <sys/kernel.h>
2549 #include <ddb/ddb.h>
2551 DB_SHOW_COMMAND(page, vm_page_print_page_info)
2553 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count);
2554 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count);
2555 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count);
2556 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count);
2557 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count);
2558 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved);
2559 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min);
2560 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target);
2561 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min);
2562 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target);
2565 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
2568 db_printf("PQ_FREE:");
2569 for(i=0;i<PQ_L2_SIZE;i++) {
2570 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
2574 db_printf("PQ_CACHE:");
2575 for(i=0;i<PQ_L2_SIZE;i++) {
2576 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
2580 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
2581 vm_page_queues[PQ_ACTIVE].lcnt,
2582 vm_page_queues[PQ_INACTIVE].lcnt);