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/inttypes.h>
91 #include <machine/md_var.h>
93 #include <vm/vm_page2.h>
94 #include <sys/spinlock2.h>
96 #define VMACTION_HSIZE 256
97 #define VMACTION_HMASK (VMACTION_HSIZE - 1)
99 static void vm_page_queue_init(void);
100 static void vm_page_free_wakeup(void);
101 static vm_page_t vm_page_select_cache(u_short pg_color);
102 static vm_page_t _vm_page_list_find2(int basequeue, int index);
103 static void _vm_page_deactivate_locked(vm_page_t m, int athead);
106 * Array of tailq lists
108 __cachealign struct vpgqueues vm_page_queues[PQ_COUNT];
110 LIST_HEAD(vm_page_action_list, vm_page_action);
111 struct vm_page_action_list action_list[VMACTION_HSIZE];
112 static volatile int vm_pages_waiting;
115 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare,
116 vm_pindex_t, pindex);
119 vm_page_queue_init(void)
123 for (i = 0; i < PQ_L2_SIZE; i++)
124 vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count;
125 for (i = 0; i < PQ_L2_SIZE; i++)
126 vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count;
127 for (i = 0; i < PQ_L2_SIZE; i++)
128 vm_page_queues[PQ_INACTIVE+i].cnt = &vmstats.v_inactive_count;
129 for (i = 0; i < PQ_L2_SIZE; i++)
130 vm_page_queues[PQ_ACTIVE+i].cnt = &vmstats.v_active_count;
131 for (i = 0; i < PQ_L2_SIZE; i++)
132 vm_page_queues[PQ_HOLD+i].cnt = &vmstats.v_active_count;
133 /* PQ_NONE has no queue */
135 for (i = 0; i < PQ_COUNT; i++) {
136 TAILQ_INIT(&vm_page_queues[i].pl);
137 spin_init(&vm_page_queues[i].spin);
140 for (i = 0; i < VMACTION_HSIZE; i++)
141 LIST_INIT(&action_list[i]);
145 * note: place in initialized data section? Is this necessary?
148 int vm_page_array_size = 0;
149 int vm_page_zero_count = 0;
150 vm_page_t vm_page_array = 0;
155 * Sets the page size, perhaps based upon the memory size.
156 * Must be called before any use of page-size dependent functions.
159 vm_set_page_size(void)
161 if (vmstats.v_page_size == 0)
162 vmstats.v_page_size = PAGE_SIZE;
163 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0)
164 panic("vm_set_page_size: page size not a power of two");
170 * Add a new page to the freelist for use by the system. New pages
171 * are added to both the head and tail of the associated free page
172 * queue in a bottom-up fashion, so both zero'd and non-zero'd page
173 * requests pull 'recent' adds (higher physical addresses) first.
175 * Must be called in a critical section.
178 vm_add_new_page(vm_paddr_t pa)
180 struct vpgqueues *vpq;
183 m = PHYS_TO_VM_PAGE(pa);
186 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK;
189 * Twist for cpu localization instead of page coloring.
191 m->pc ^= ((pa >> PAGE_SHIFT) / PQ_L2_SIZE) & PQ_L2_MASK;
192 m->pc ^= ((pa >> PAGE_SHIFT) / (PQ_L2_SIZE * PQ_L2_SIZE)) & PQ_L2_MASK;
194 m->queue = m->pc + PQ_FREE;
195 KKASSERT(m->dirty == 0);
197 atomic_add_int(&vmstats.v_page_count, 1);
198 atomic_add_int(&vmstats.v_free_count, 1);
199 vpq = &vm_page_queues[m->queue];
201 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
203 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
204 vpq->flipflop = 1 - vpq->flipflop;
213 * Initializes the resident memory module.
215 * Preallocates memory for critical VM structures and arrays prior to
216 * kernel_map becoming available.
218 * Memory is allocated from (virtual2_start, virtual2_end) if available,
219 * otherwise memory is allocated from (virtual_start, virtual_end).
221 * On x86-64 (virtual_start, virtual_end) is only 2GB and may not be
222 * large enough to hold vm_page_array & other structures for machines with
223 * large amounts of ram, so we want to use virtual2* when available.
226 vm_page_startup(void)
228 vm_offset_t vaddr = virtual2_start ? virtual2_start : virtual_start;
231 vm_paddr_t page_range;
238 vm_paddr_t biggestone, biggestsize;
245 vaddr = round_page(vaddr);
247 for (i = 0; phys_avail[i + 1]; i += 2) {
248 phys_avail[i] = round_page64(phys_avail[i]);
249 phys_avail[i + 1] = trunc_page64(phys_avail[i + 1]);
252 for (i = 0; phys_avail[i + 1]; i += 2) {
253 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
255 if (size > biggestsize) {
263 end = phys_avail[biggestone+1];
264 end = trunc_page(end);
267 * Initialize the queue headers for the free queue, the active queue
268 * and the inactive queue.
271 vm_page_queue_init();
273 #if !defined(_KERNEL_VIRTUAL)
275 * VKERNELs don't support minidumps and as such don't need
278 * Allocate a bitmap to indicate that a random physical page
279 * needs to be included in a minidump.
281 * The amd64 port needs this to indicate which direct map pages
282 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
284 * However, i386 still needs this workspace internally within the
285 * minidump code. In theory, they are not needed on i386, but are
286 * included should the sf_buf code decide to use them.
288 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE;
289 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
290 end -= vm_page_dump_size;
291 vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size,
292 VM_PROT_READ | VM_PROT_WRITE);
293 bzero((void *)vm_page_dump, vm_page_dump_size);
297 * Compute the number of pages of memory that will be available for
298 * use (taking into account the overhead of a page structure per
301 first_page = phys_avail[0] / PAGE_SIZE;
302 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
303 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE;
306 * Initialize the mem entry structures now, and put them in the free
309 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
310 mapped = pmap_map(&vaddr, new_end, end,
311 VM_PROT_READ | VM_PROT_WRITE);
312 vm_page_array = (vm_page_t)mapped;
314 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL)
316 * since pmap_map on amd64 returns stuff out of a direct-map region,
317 * we have to manually add these pages to the minidump tracking so
318 * that they can be dumped, including the vm_page_array.
320 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
325 * Clear all of the page structures
327 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
328 vm_page_array_size = page_range;
331 * Construct the free queue(s) in ascending order (by physical
332 * address) so that the first 16MB of physical memory is allocated
333 * last rather than first. On large-memory machines, this avoids
334 * the exhaustion of low physical memory before isa_dmainit has run.
336 vmstats.v_page_count = 0;
337 vmstats.v_free_count = 0;
338 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
343 last_pa = phys_avail[i + 1];
344 while (pa < last_pa && npages-- > 0) {
350 virtual2_start = vaddr;
352 virtual_start = vaddr;
356 * Scan comparison function for Red-Black tree scans. An inclusive
357 * (start,end) is expected. Other fields are not used.
360 rb_vm_page_scancmp(struct vm_page *p, void *data)
362 struct rb_vm_page_scan_info *info = data;
364 if (p->pindex < info->start_pindex)
366 if (p->pindex > info->end_pindex)
372 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2)
374 if (p1->pindex < p2->pindex)
376 if (p1->pindex > p2->pindex)
382 * Each page queue has its own spin lock, which is fairly optimal for
383 * allocating and freeing pages at least.
385 * The caller must hold the vm_page_spin_lock() before locking a vm_page's
386 * queue spinlock via this function. Also note that m->queue cannot change
387 * unless both the page and queue are locked.
391 _vm_page_queue_spin_lock(vm_page_t m)
396 if (queue != PQ_NONE) {
397 spin_lock(&vm_page_queues[queue].spin);
398 KKASSERT(queue == m->queue);
404 _vm_page_queue_spin_unlock(vm_page_t m)
410 if (queue != PQ_NONE)
411 spin_unlock(&vm_page_queues[queue].spin);
416 _vm_page_queues_spin_lock(u_short queue)
419 if (queue != PQ_NONE)
420 spin_lock(&vm_page_queues[queue].spin);
426 _vm_page_queues_spin_unlock(u_short queue)
429 if (queue != PQ_NONE)
430 spin_unlock(&vm_page_queues[queue].spin);
434 vm_page_queue_spin_lock(vm_page_t m)
436 _vm_page_queue_spin_lock(m);
440 vm_page_queues_spin_lock(u_short queue)
442 _vm_page_queues_spin_lock(queue);
446 vm_page_queue_spin_unlock(vm_page_t m)
448 _vm_page_queue_spin_unlock(m);
452 vm_page_queues_spin_unlock(u_short queue)
454 _vm_page_queues_spin_unlock(queue);
458 * This locks the specified vm_page and its queue in the proper order
459 * (page first, then queue). The queue may change so the caller must
464 _vm_page_and_queue_spin_lock(vm_page_t m)
466 vm_page_spin_lock(m);
467 _vm_page_queue_spin_lock(m);
472 _vm_page_and_queue_spin_unlock(vm_page_t m)
474 _vm_page_queues_spin_unlock(m->queue);
475 vm_page_spin_unlock(m);
479 vm_page_and_queue_spin_unlock(vm_page_t m)
481 _vm_page_and_queue_spin_unlock(m);
485 vm_page_and_queue_spin_lock(vm_page_t m)
487 _vm_page_and_queue_spin_lock(m);
491 * Helper function removes vm_page from its current queue.
492 * Returns the base queue the page used to be on.
494 * The vm_page and the queue must be spinlocked.
495 * This function will unlock the queue but leave the page spinlocked.
497 static __inline u_short
498 _vm_page_rem_queue_spinlocked(vm_page_t m)
500 struct vpgqueues *pq;
504 if (queue != PQ_NONE) {
505 pq = &vm_page_queues[queue];
506 TAILQ_REMOVE(&pq->pl, m, pageq);
507 atomic_add_int(pq->cnt, -1);
510 vm_page_queues_spin_unlock(queue);
511 if ((queue - m->pc) == PQ_FREE && (m->flags & PG_ZERO))
512 atomic_subtract_int(&vm_page_zero_count, 1);
513 if ((queue - m->pc) == PQ_CACHE || (queue - m->pc) == PQ_FREE)
514 return (queue - m->pc);
520 * Helper function places the vm_page on the specified queue.
522 * The vm_page must be spinlocked.
523 * This function will return with both the page and the queue locked.
526 _vm_page_add_queue_spinlocked(vm_page_t m, u_short queue, int athead)
528 struct vpgqueues *pq;
530 KKASSERT(m->queue == PQ_NONE);
532 if (queue != PQ_NONE) {
533 vm_page_queues_spin_lock(queue);
534 pq = &vm_page_queues[queue];
536 atomic_add_int(pq->cnt, 1);
540 * Put zero'd pages on the end ( where we look for zero'd pages
541 * first ) and non-zerod pages at the head.
543 if (queue - m->pc == PQ_FREE) {
544 if (m->flags & PG_ZERO) {
545 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
546 atomic_add_int(&vm_page_zero_count, 1);
548 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
551 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
553 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
555 /* leave the queue spinlocked */
560 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE)
561 * m->busy is zero. Returns TRUE if it had to sleep, FALSE if we
562 * did not. Only one sleep call will be made before returning.
564 * This function does NOT busy the page and on return the page is not
565 * guaranteed to be available.
568 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg)
576 if ((flags & PG_BUSY) == 0 &&
577 (also_m_busy == 0 || (flags & PG_SBUSY) == 0)) {
580 tsleep_interlock(m, 0);
581 if (atomic_cmpset_int(&m->flags, flags,
582 flags | PG_WANTED | PG_REFERENCED)) {
583 tsleep(m, PINTERLOCKED, msg, 0);
590 * Wait until PG_BUSY can be set, then set it. If also_m_busy is TRUE we
591 * also wait for m->busy to become 0 before setting PG_BUSY.
594 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m,
595 int also_m_busy, const char *msg
603 if (flags & PG_BUSY) {
604 tsleep_interlock(m, 0);
605 if (atomic_cmpset_int(&m->flags, flags,
606 flags | PG_WANTED | PG_REFERENCED)) {
607 tsleep(m, PINTERLOCKED, msg, 0);
609 } else if (also_m_busy && (flags & PG_SBUSY)) {
610 tsleep_interlock(m, 0);
611 if (atomic_cmpset_int(&m->flags, flags,
612 flags | PG_WANTED | PG_REFERENCED)) {
613 tsleep(m, PINTERLOCKED, msg, 0);
616 if (atomic_cmpset_int(&m->flags, flags,
620 m->busy_line = lineno;
629 * Attempt to set PG_BUSY. If also_m_busy is TRUE we only succeed if m->busy
632 * Returns non-zero on failure.
635 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy
645 if (also_m_busy && (flags & PG_SBUSY))
647 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
650 m->busy_line = lineno;
658 * Clear the PG_BUSY flag and return non-zero to indicate to the caller
659 * that a wakeup() should be performed.
661 * The vm_page must be spinlocked and will remain spinlocked on return.
662 * The related queue must NOT be spinlocked (which could deadlock us).
668 _vm_page_wakeup(vm_page_t m)
675 if (atomic_cmpset_int(&m->flags, flags,
676 flags & ~(PG_BUSY | PG_WANTED))) {
680 return(flags & PG_WANTED);
684 * Clear the PG_BUSY flag and wakeup anyone waiting for the page. This
685 * is typically the last call you make on a page before moving onto
689 vm_page_wakeup(vm_page_t m)
691 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!"));
692 vm_page_spin_lock(m);
693 if (_vm_page_wakeup(m)) {
694 vm_page_spin_unlock(m);
697 vm_page_spin_unlock(m);
702 * Holding a page keeps it from being reused. Other parts of the system
703 * can still disassociate the page from its current object and free it, or
704 * perform read or write I/O on it and/or otherwise manipulate the page,
705 * but if the page is held the VM system will leave the page and its data
706 * intact and not reuse the page for other purposes until the last hold
707 * reference is released. (see vm_page_wire() if you want to prevent the
708 * page from being disassociated from its object too).
710 * The caller must still validate the contents of the page and, if necessary,
711 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete
712 * before manipulating the page.
714 * XXX get vm_page_spin_lock() here and move FREE->HOLD if necessary
717 vm_page_hold(vm_page_t m)
719 vm_page_spin_lock(m);
720 atomic_add_int(&m->hold_count, 1);
721 if (m->queue - m->pc == PQ_FREE) {
722 _vm_page_queue_spin_lock(m);
723 _vm_page_rem_queue_spinlocked(m);
724 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
725 _vm_page_queue_spin_unlock(m);
727 vm_page_spin_unlock(m);
731 * The opposite of vm_page_hold(). A page can be freed while being held,
732 * which places it on the PQ_HOLD queue. If we are able to busy the page
733 * after the hold count drops to zero we will move the page to the
734 * appropriate PQ_FREE queue by calling vm_page_free_toq().
737 vm_page_unhold(vm_page_t m)
739 vm_page_spin_lock(m);
740 atomic_add_int(&m->hold_count, -1);
741 if (m->hold_count == 0 && m->queue - m->pc == PQ_HOLD) {
742 _vm_page_queue_spin_lock(m);
743 _vm_page_rem_queue_spinlocked(m);
744 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0);
745 _vm_page_queue_spin_unlock(m);
747 vm_page_spin_unlock(m);
751 * Inserts the given vm_page into the object and object list.
753 * The pagetables are not updated but will presumably fault the page
754 * in if necessary, or if a kernel page the caller will at some point
755 * enter the page into the kernel's pmap. We are not allowed to block
756 * here so we *can't* do this anyway.
758 * This routine may not block.
759 * This routine must be called with the vm_object held.
760 * This routine must be called with a critical section held.
762 * This routine returns TRUE if the page was inserted into the object
763 * successfully, and FALSE if the page already exists in the object.
766 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
768 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
769 if (m->object != NULL)
770 panic("vm_page_insert: already inserted");
772 object->generation++;
775 * Record the object/offset pair in this page and add the
776 * pv_list_count of the page to the object.
778 * The vm_page spin lock is required for interactions with the pmap.
780 vm_page_spin_lock(m);
783 if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) {
786 vm_page_spin_unlock(m);
789 object->resident_page_count++;
790 /* atomic_add_int(&object->agg_pv_list_count, m->md.pv_list_count); */
791 vm_page_spin_unlock(m);
794 * Since we are inserting a new and possibly dirty page,
795 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
797 if ((m->valid & m->dirty) || (m->flags & PG_WRITEABLE))
798 vm_object_set_writeable_dirty(object);
801 * Checks for a swap assignment and sets PG_SWAPPED if appropriate.
803 swap_pager_page_inserted(m);
808 * Removes the given vm_page_t from the (object,index) table
810 * The underlying pmap entry (if any) is NOT removed here.
811 * This routine may not block.
813 * The page must be BUSY and will remain BUSY on return.
814 * No other requirements.
816 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave
820 vm_page_remove(vm_page_t m)
824 if (m->object == NULL) {
828 if ((m->flags & PG_BUSY) == 0)
829 panic("vm_page_remove: page not busy");
833 vm_object_hold(object);
836 * Remove the page from the object and update the object.
838 * The vm_page spin lock is required for interactions with the pmap.
840 vm_page_spin_lock(m);
841 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m);
842 object->resident_page_count--;
843 /* atomic_add_int(&object->agg_pv_list_count, -m->md.pv_list_count); */
845 vm_page_spin_unlock(m);
847 object->generation++;
849 vm_object_drop(object);
853 * Locate and return the page at (object, pindex), or NULL if the
854 * page could not be found.
856 * The caller must hold the vm_object token.
859 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
864 * Search the hash table for this object/offset pair
866 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
867 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
868 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex));
873 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object,
875 int also_m_busy, const char *msg
881 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
882 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
884 KKASSERT(m->object == object && m->pindex == pindex);
887 if (flags & PG_BUSY) {
888 tsleep_interlock(m, 0);
889 if (atomic_cmpset_int(&m->flags, flags,
890 flags | PG_WANTED | PG_REFERENCED)) {
891 tsleep(m, PINTERLOCKED, msg, 0);
892 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
895 } else if (also_m_busy && (flags & PG_SBUSY)) {
896 tsleep_interlock(m, 0);
897 if (atomic_cmpset_int(&m->flags, flags,
898 flags | PG_WANTED | PG_REFERENCED)) {
899 tsleep(m, PINTERLOCKED, msg, 0);
900 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
903 } else if (atomic_cmpset_int(&m->flags, flags,
907 m->busy_line = lineno;
916 * Attempt to lookup and busy a page.
918 * Returns NULL if the page could not be found
920 * Returns a vm_page and error == TRUE if the page exists but could not
923 * Returns a vm_page and error == FALSE on success.
926 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object,
928 int also_m_busy, int *errorp
934 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
935 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
938 KKASSERT(m->object == object && m->pindex == pindex);
941 if (flags & PG_BUSY) {
945 if (also_m_busy && (flags & PG_SBUSY)) {
949 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
952 m->busy_line = lineno;
961 * Caller must hold the related vm_object
964 vm_page_next(vm_page_t m)
968 next = vm_page_rb_tree_RB_NEXT(m);
969 if (next && next->pindex != m->pindex + 1)
977 * Move the given vm_page from its current object to the specified
978 * target object/offset. The page must be busy and will remain so
981 * new_object must be held.
982 * This routine might block. XXX ?
984 * NOTE: Swap associated with the page must be invalidated by the move. We
985 * have to do this for several reasons: (1) we aren't freeing the
986 * page, (2) we are dirtying the page, (3) the VM system is probably
987 * moving the page from object A to B, and will then later move
988 * the backing store from A to B and we can't have a conflict.
990 * NOTE: We *always* dirty the page. It is necessary both for the
991 * fact that we moved it, and because we may be invalidating
992 * swap. If the page is on the cache, we have to deactivate it
993 * or vm_page_dirty() will panic. Dirty pages are not allowed
997 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
999 KKASSERT(m->flags & PG_BUSY);
1000 ASSERT_LWKT_TOKEN_HELD(vm_object_token(new_object));
1002 ASSERT_LWKT_TOKEN_HELD(vm_object_token(m->object));
1005 if (vm_page_insert(m, new_object, new_pindex) == FALSE) {
1006 panic("vm_page_rename: target exists (%p,%"PRIu64")",
1007 new_object, new_pindex);
1009 if (m->queue - m->pc == PQ_CACHE)
1010 vm_page_deactivate(m);
1015 * vm_page_unqueue() without any wakeup. This routine is used when a page
1016 * is being moved between queues or otherwise is to remain BUSYied by the
1019 * This routine may not block.
1022 vm_page_unqueue_nowakeup(vm_page_t m)
1024 vm_page_and_queue_spin_lock(m);
1025 (void)_vm_page_rem_queue_spinlocked(m);
1026 vm_page_spin_unlock(m);
1030 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
1033 * This routine may not block.
1036 vm_page_unqueue(vm_page_t m)
1040 vm_page_and_queue_spin_lock(m);
1041 queue = _vm_page_rem_queue_spinlocked(m);
1042 if (queue == PQ_FREE || queue == PQ_CACHE) {
1043 vm_page_spin_unlock(m);
1044 pagedaemon_wakeup();
1046 vm_page_spin_unlock(m);
1051 * vm_page_list_find()
1053 * Find a page on the specified queue with color optimization.
1055 * The page coloring optimization attempts to locate a page that does
1056 * not overload other nearby pages in the object in the cpu's L1 or L2
1057 * caches. We need this optimization because cpu caches tend to be
1058 * physical caches, while object spaces tend to be virtual.
1060 * On MP systems each PQ_FREE and PQ_CACHE color queue has its own spinlock
1061 * and the algorithm is adjusted to localize allocations on a per-core basis.
1062 * This is done by 'twisting' the colors.
1064 * The page is returned spinlocked and removed from its queue (it will
1065 * be on PQ_NONE), or NULL. The page is not PG_BUSY'd. The caller
1066 * is responsible for dealing with the busy-page case (usually by
1067 * deactivating the page and looping).
1069 * NOTE: This routine is carefully inlined. A non-inlined version
1070 * is available for outside callers but the only critical path is
1071 * from within this source file.
1073 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE
1074 * represent stable storage, allowing us to order our locks vm_page
1075 * first, then queue.
1079 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
1085 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist);
1087 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl);
1089 m = _vm_page_list_find2(basequeue, index);
1092 vm_page_and_queue_spin_lock(m);
1093 if (m->queue == basequeue + index) {
1094 _vm_page_rem_queue_spinlocked(m);
1095 /* vm_page_t spin held, no queue spin */
1098 vm_page_and_queue_spin_unlock(m);
1104 _vm_page_list_find2(int basequeue, int index)
1108 struct vpgqueues *pq;
1110 pq = &vm_page_queues[basequeue];
1113 * Note that for the first loop, index+i and index-i wind up at the
1114 * same place. Even though this is not totally optimal, we've already
1115 * blown it by missing the cache case so we do not care.
1117 for (i = PQ_L2_SIZE / 2; i > 0; --i) {
1119 m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl);
1121 _vm_page_and_queue_spin_lock(m);
1123 basequeue + ((index + i) & PQ_L2_MASK)) {
1124 _vm_page_rem_queue_spinlocked(m);
1127 _vm_page_and_queue_spin_unlock(m);
1130 m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl);
1132 _vm_page_and_queue_spin_lock(m);
1134 basequeue + ((index - i) & PQ_L2_MASK)) {
1135 _vm_page_rem_queue_spinlocked(m);
1138 _vm_page_and_queue_spin_unlock(m);
1148 * Returns a vm_page candidate for allocation. The page is not busied so
1149 * it can move around. The caller must busy the page (and typically
1150 * deactivate it if it cannot be busied!)
1152 * Returns a spinlocked vm_page that has been removed from its queue.
1155 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
1157 return(_vm_page_list_find(basequeue, index, prefer_zero));
1161 * Find a page on the cache queue with color optimization, remove it
1162 * from the queue, and busy it. The returned page will not be spinlocked.
1164 * A candidate failure will be deactivated. Candidates can fail due to
1165 * being busied by someone else, in which case they will be deactivated.
1167 * This routine may not block.
1171 vm_page_select_cache(u_short pg_color)
1176 m = _vm_page_list_find(PQ_CACHE, pg_color & PQ_L2_MASK, FALSE);
1180 * (m) has been removed from its queue and spinlocked
1182 if (vm_page_busy_try(m, TRUE)) {
1183 _vm_page_deactivate_locked(m, 0);
1184 vm_page_spin_unlock(m);
1186 kprintf("Warning: busy page %p found in cache\n", m);
1190 * We successfully busied the page
1192 if ((m->flags & PG_UNMANAGED) == 0 &&
1193 m->hold_count == 0 &&
1194 m->wire_count == 0) {
1195 vm_page_spin_unlock(m);
1196 pagedaemon_wakeup();
1199 _vm_page_deactivate_locked(m, 0);
1200 if (_vm_page_wakeup(m)) {
1201 vm_page_spin_unlock(m);
1204 vm_page_spin_unlock(m);
1212 * Find a free or zero page, with specified preference. We attempt to
1213 * inline the nominal case and fall back to _vm_page_select_free()
1214 * otherwise. A busied page is removed from the queue and returned.
1216 * This routine may not block.
1218 static __inline vm_page_t
1219 vm_page_select_free(u_short pg_color, boolean_t prefer_zero)
1224 m = _vm_page_list_find(PQ_FREE, pg_color & PQ_L2_MASK,
1228 if (vm_page_busy_try(m, TRUE)) {
1229 _vm_page_deactivate_locked(m, 0);
1230 vm_page_spin_unlock(m);
1232 kprintf("Warning: busy page %p found in cache\n", m);
1235 KKASSERT((m->flags & PG_UNMANAGED) == 0);
1236 KKASSERT(m->hold_count == 0);
1237 KKASSERT(m->wire_count == 0);
1238 vm_page_spin_unlock(m);
1239 pagedaemon_wakeup();
1241 /* return busied and removed page */
1251 * Allocate and return a memory cell associated with this VM object/offset
1252 * pair. If object is NULL an unassociated page will be allocated.
1254 * The returned page will be busied and removed from its queues. This
1255 * routine can block and may return NULL if a race occurs and the page
1256 * is found to already exist at the specified (object, pindex).
1258 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain
1259 * VM_ALLOC_QUICK like normal but cannot use cache
1260 * VM_ALLOC_SYSTEM greater free drain
1261 * VM_ALLOC_INTERRUPT allow free list to be completely drained
1262 * VM_ALLOC_ZERO advisory request for pre-zero'd page only
1263 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only
1264 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision
1265 * (see vm_page_grab())
1266 * The object must be held if not NULL
1267 * This routine may not block
1269 * Additional special handling is required when called from an interrupt
1270 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache
1274 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
1281 * Cpu twist - cpu localization algorithm
1284 pg_color = mycpu->gd_cpuid + (pindex & ~ncpus_fit_mask) +
1285 (object->pg_color & ~ncpus_fit_mask);
1287 pg_color = mycpu->gd_cpuid + (pindex & ~ncpus_fit_mask);
1291 * Normal page coloring algorithm
1294 pg_color = object->pg_color + pindex;
1300 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK|
1301 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
1304 * Certain system threads (pageout daemon, buf_daemon's) are
1305 * allowed to eat deeper into the free page list.
1307 if (curthread->td_flags & TDF_SYSTHREAD)
1308 page_req |= VM_ALLOC_SYSTEM;
1311 if (vmstats.v_free_count > vmstats.v_free_reserved ||
1312 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) ||
1313 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 &&
1314 vmstats.v_free_count > vmstats.v_interrupt_free_min)
1317 * The free queue has sufficient free pages to take one out.
1319 if (page_req & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO))
1320 m = vm_page_select_free(pg_color, TRUE);
1322 m = vm_page_select_free(pg_color, FALSE);
1323 } else if (page_req & VM_ALLOC_NORMAL) {
1325 * Allocatable from the cache (non-interrupt only). On
1326 * success, we must free the page and try again, thus
1327 * ensuring that vmstats.v_*_free_min counters are replenished.
1330 if (curthread->td_preempted) {
1331 kprintf("vm_page_alloc(): warning, attempt to allocate"
1332 " cache page from preempting interrupt\n");
1335 m = vm_page_select_cache(pg_color);
1338 m = vm_page_select_cache(pg_color);
1341 * On success move the page into the free queue and loop.
1344 KASSERT(m->dirty == 0,
1345 ("Found dirty cache page %p", m));
1346 vm_page_protect(m, VM_PROT_NONE);
1352 * On failure return NULL
1354 #if defined(DIAGNOSTIC)
1355 if (vmstats.v_cache_count > 0)
1356 kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count);
1358 vm_pageout_deficit++;
1359 pagedaemon_wakeup();
1363 * No pages available, wakeup the pageout daemon and give up.
1365 vm_pageout_deficit++;
1366 pagedaemon_wakeup();
1371 * v_free_count can race so loop if we don't find the expected
1378 * Good page found. The page has already been busied for us and
1379 * removed from its queues.
1381 KASSERT(m->dirty == 0,
1382 ("vm_page_alloc: free/cache page %p was dirty", m));
1383 KKASSERT(m->queue == PQ_NONE);
1386 * Initialize the structure, inheriting some flags but clearing
1387 * all the rest. The page has already been busied for us.
1389 vm_page_flag_clear(m, ~(PG_ZERO | PG_BUSY | PG_SBUSY));
1390 KKASSERT(m->wire_count == 0);
1391 KKASSERT(m->busy == 0);
1396 * Caller must be holding the object lock (asserted by
1397 * vm_page_insert()).
1399 * NOTE: Inserting a page here does not insert it into any pmaps
1400 * (which could cause us to block allocating memory).
1402 * NOTE: If no object an unassociated page is allocated, m->pindex
1403 * can be used by the caller for any purpose.
1406 if (vm_page_insert(m, object, pindex) == FALSE) {
1407 kprintf("PAGE RACE (%p:%d,%"PRIu64")\n",
1408 object, object->type, pindex);
1411 if ((page_req & VM_ALLOC_NULL_OK) == 0)
1419 * Don't wakeup too often - wakeup the pageout daemon when
1420 * we would be nearly out of memory.
1422 pagedaemon_wakeup();
1425 * A PG_BUSY page is returned.
1431 * Wait for sufficient free memory for nominal heavy memory use kernel
1435 vm_wait_nominal(void)
1437 while (vm_page_count_min(0))
1442 * Test if vm_wait_nominal() would block.
1445 vm_test_nominal(void)
1447 if (vm_page_count_min(0))
1453 * Block until free pages are available for allocation, called in various
1454 * places before memory allocations.
1456 * The caller may loop if vm_page_count_min() == FALSE so we cannot be
1457 * more generous then that.
1463 * never wait forever
1467 lwkt_gettoken(&vm_token);
1469 if (curthread == pagethread) {
1471 * The pageout daemon itself needs pages, this is bad.
1473 if (vm_page_count_min(0)) {
1474 vm_pageout_pages_needed = 1;
1475 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo);
1479 * Wakeup the pageout daemon if necessary and wait.
1481 if (vm_page_count_target()) {
1482 if (vm_pages_needed == 0) {
1483 vm_pages_needed = 1;
1484 wakeup(&vm_pages_needed);
1486 ++vm_pages_waiting; /* SMP race ok */
1487 tsleep(&vmstats.v_free_count, 0, "vmwait", timo);
1490 lwkt_reltoken(&vm_token);
1494 * Block until free pages are available for allocation
1496 * Called only from vm_fault so that processes page faulting can be
1503 * Wakeup the pageout daemon if necessary and wait.
1505 if (vm_page_count_target()) {
1506 lwkt_gettoken(&vm_token);
1507 if (vm_page_count_target()) {
1508 if (vm_pages_needed == 0) {
1509 vm_pages_needed = 1;
1510 wakeup(&vm_pages_needed);
1512 ++vm_pages_waiting; /* SMP race ok */
1513 tsleep(&vmstats.v_free_count, 0, "pfault", hz);
1515 lwkt_reltoken(&vm_token);
1520 * Put the specified page on the active list (if appropriate). Ensure
1521 * that act_count is at least ACT_INIT but do not otherwise mess with it.
1523 * The caller should be holding the page busied ? XXX
1524 * This routine may not block.
1527 vm_page_activate(vm_page_t m)
1531 vm_page_spin_lock(m);
1532 if (m->queue - m->pc != PQ_ACTIVE) {
1533 _vm_page_queue_spin_lock(m);
1534 oqueue = _vm_page_rem_queue_spinlocked(m);
1535 /* page is left spinlocked, queue is unlocked */
1537 if (oqueue == PQ_CACHE)
1538 mycpu->gd_cnt.v_reactivated++;
1539 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1540 if (m->act_count < ACT_INIT)
1541 m->act_count = ACT_INIT;
1542 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE + m->pc, 0);
1544 _vm_page_and_queue_spin_unlock(m);
1545 if (oqueue == PQ_CACHE || oqueue == PQ_FREE)
1546 pagedaemon_wakeup();
1548 if (m->act_count < ACT_INIT)
1549 m->act_count = ACT_INIT;
1550 vm_page_spin_unlock(m);
1555 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
1556 * routine is called when a page has been added to the cache or free
1559 * This routine may not block.
1561 static __inline void
1562 vm_page_free_wakeup(void)
1565 * If the pageout daemon itself needs pages, then tell it that
1566 * there are some free.
1568 if (vm_pageout_pages_needed &&
1569 vmstats.v_cache_count + vmstats.v_free_count >=
1570 vmstats.v_pageout_free_min
1572 wakeup(&vm_pageout_pages_needed);
1573 vm_pageout_pages_needed = 0;
1577 * Wakeup processes that are waiting on memory.
1579 * NOTE: vm_paging_target() is the pageout daemon's target, while
1580 * vm_page_count_target() is somewhere inbetween. We want
1581 * to wake processes up prior to the pageout daemon reaching
1582 * its target to provide some hysteresis.
1584 if (vm_pages_waiting) {
1585 if (!vm_page_count_target()) {
1587 * Plenty of pages are free, wakeup everyone.
1589 vm_pages_waiting = 0;
1590 wakeup(&vmstats.v_free_count);
1591 ++mycpu->gd_cnt.v_ppwakeups;
1592 } else if (!vm_page_count_min(0)) {
1594 * Some pages are free, wakeup someone.
1596 int wcount = vm_pages_waiting;
1599 vm_pages_waiting = wcount;
1600 wakeup_one(&vmstats.v_free_count);
1601 ++mycpu->gd_cnt.v_ppwakeups;
1607 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates
1608 * it from its VM object.
1610 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on
1611 * return (the page will have been freed).
1614 vm_page_free_toq(vm_page_t m)
1616 mycpu->gd_cnt.v_tfree++;
1617 KKASSERT((m->flags & PG_MAPPED) == 0);
1618 KKASSERT(m->flags & PG_BUSY);
1620 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
1622 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
1623 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
1625 if ((m->queue - m->pc) == PQ_FREE)
1626 panic("vm_page_free: freeing free page");
1628 panic("vm_page_free: freeing busy page");
1632 * Remove from object, spinlock the page and its queues and
1633 * remove from any queue. No queue spinlock will be held
1634 * after this section (because the page was removed from any
1638 vm_page_and_queue_spin_lock(m);
1639 _vm_page_rem_queue_spinlocked(m);
1642 * No further management of fictitious pages occurs beyond object
1643 * and queue removal.
1645 if ((m->flags & PG_FICTITIOUS) != 0) {
1646 vm_page_spin_unlock(m);
1654 if (m->wire_count != 0) {
1655 if (m->wire_count > 1) {
1657 "vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1658 m->wire_count, (long)m->pindex);
1660 panic("vm_page_free: freeing wired page");
1664 * Clear the UNMANAGED flag when freeing an unmanaged page.
1666 if (m->flags & PG_UNMANAGED) {
1667 vm_page_flag_clear(m, PG_UNMANAGED);
1670 if (m->hold_count != 0) {
1671 vm_page_flag_clear(m, PG_ZERO);
1672 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
1674 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0);
1678 * This sequence allows us to clear PG_BUSY while still holding
1679 * its spin lock, which reduces contention vs allocators. We
1680 * must not leave the queue locked or _vm_page_wakeup() may
1683 _vm_page_queue_spin_unlock(m);
1684 if (_vm_page_wakeup(m)) {
1685 vm_page_spin_unlock(m);
1688 vm_page_spin_unlock(m);
1690 vm_page_free_wakeup();
1694 * vm_page_free_fromq_fast()
1696 * Remove a non-zero page from one of the free queues; the page is removed for
1697 * zeroing, so do not issue a wakeup.
1700 vm_page_free_fromq_fast(void)
1706 for (i = 0; i < PQ_L2_SIZE; ++i) {
1707 m = vm_page_list_find(PQ_FREE, qi, FALSE);
1708 /* page is returned spinlocked and removed from its queue */
1710 if (vm_page_busy_try(m, TRUE)) {
1712 * We were unable to busy the page, deactivate
1715 _vm_page_deactivate_locked(m, 0);
1716 vm_page_spin_unlock(m);
1717 } else if ((m->flags & PG_ZERO) == 0) {
1719 * The page is not PG_ZERO'd so return it.
1721 vm_page_spin_unlock(m);
1725 * The page is PG_ZERO, requeue it and loop
1727 _vm_page_add_queue_spinlocked(m,
1730 vm_page_queue_spin_unlock(m);
1731 if (_vm_page_wakeup(m)) {
1732 vm_page_spin_unlock(m);
1735 vm_page_spin_unlock(m);
1740 qi = (qi + PQ_PRIME2) & PQ_L2_MASK;
1746 * vm_page_unmanage()
1748 * Prevent PV management from being done on the page. The page is
1749 * removed from the paging queues as if it were wired, and as a
1750 * consequence of no longer being managed the pageout daemon will not
1751 * touch it (since there is no way to locate the pte mappings for the
1752 * page). madvise() calls that mess with the pmap will also no longer
1753 * operate on the page.
1755 * Beyond that the page is still reasonably 'normal'. Freeing the page
1756 * will clear the flag.
1758 * This routine is used by OBJT_PHYS objects - objects using unswappable
1759 * physical memory as backing store rather then swap-backed memory and
1760 * will eventually be extended to support 4MB unmanaged physical
1763 * Caller must be holding the page busy.
1766 vm_page_unmanage(vm_page_t m)
1768 KKASSERT(m->flags & PG_BUSY);
1769 if ((m->flags & PG_UNMANAGED) == 0) {
1770 if (m->wire_count == 0)
1773 vm_page_flag_set(m, PG_UNMANAGED);
1777 * Mark this page as wired down by yet another map, removing it from
1778 * paging queues as necessary.
1780 * Caller must be holding the page busy.
1783 vm_page_wire(vm_page_t m)
1786 * Only bump the wire statistics if the page is not already wired,
1787 * and only unqueue the page if it is on some queue (if it is unmanaged
1788 * it is already off the queues). Don't do anything with fictitious
1789 * pages because they are always wired.
1791 KKASSERT(m->flags & PG_BUSY);
1792 if ((m->flags & PG_FICTITIOUS) == 0) {
1793 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) {
1794 if ((m->flags & PG_UNMANAGED) == 0)
1796 atomic_add_int(&vmstats.v_wire_count, 1);
1798 KASSERT(m->wire_count != 0,
1799 ("vm_page_wire: wire_count overflow m=%p", m));
1804 * Release one wiring of this page, potentially enabling it to be paged again.
1806 * Many pages placed on the inactive queue should actually go
1807 * into the cache, but it is difficult to figure out which. What
1808 * we do instead, if the inactive target is well met, is to put
1809 * clean pages at the head of the inactive queue instead of the tail.
1810 * This will cause them to be moved to the cache more quickly and
1811 * if not actively re-referenced, freed more quickly. If we just
1812 * stick these pages at the end of the inactive queue, heavy filesystem
1813 * meta-data accesses can cause an unnecessary paging load on memory bound
1814 * processes. This optimization causes one-time-use metadata to be
1815 * reused more quickly.
1817 * BUT, if we are in a low-memory situation we have no choice but to
1818 * put clean pages on the cache queue.
1820 * A number of routines use vm_page_unwire() to guarantee that the page
1821 * will go into either the inactive or active queues, and will NEVER
1822 * be placed in the cache - for example, just after dirtying a page.
1823 * dirty pages in the cache are not allowed.
1825 * The page queues must be locked.
1826 * This routine may not block.
1829 vm_page_unwire(vm_page_t m, int activate)
1831 KKASSERT(m->flags & PG_BUSY);
1832 if (m->flags & PG_FICTITIOUS) {
1834 } else if (m->wire_count <= 0) {
1835 panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
1837 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) {
1838 atomic_add_int(&vmstats.v_wire_count, -1);
1839 if (m->flags & PG_UNMANAGED) {
1841 } else if (activate) {
1842 vm_page_spin_lock(m);
1843 _vm_page_add_queue_spinlocked(m,
1844 PQ_ACTIVE + m->pc, 0);
1845 _vm_page_and_queue_spin_unlock(m);
1847 vm_page_spin_lock(m);
1848 vm_page_flag_clear(m, PG_WINATCFLS);
1849 _vm_page_add_queue_spinlocked(m,
1850 PQ_INACTIVE + m->pc, 0);
1851 ++vm_swapcache_inactive_heuristic;
1852 _vm_page_and_queue_spin_unlock(m);
1859 * Move the specified page to the inactive queue. If the page has
1860 * any associated swap, the swap is deallocated.
1862 * Normally athead is 0 resulting in LRU operation. athead is set
1863 * to 1 if we want this page to be 'as if it were placed in the cache',
1864 * except without unmapping it from the process address space.
1866 * vm_page's spinlock must be held on entry and will remain held on return.
1867 * This routine may not block.
1870 _vm_page_deactivate_locked(vm_page_t m, int athead)
1875 * Ignore if already inactive.
1877 if (m->queue - m->pc == PQ_INACTIVE)
1879 _vm_page_queue_spin_lock(m);
1880 oqueue = _vm_page_rem_queue_spinlocked(m);
1882 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1883 if (oqueue == PQ_CACHE)
1884 mycpu->gd_cnt.v_reactivated++;
1885 vm_page_flag_clear(m, PG_WINATCFLS);
1886 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead);
1888 ++vm_swapcache_inactive_heuristic;
1890 _vm_page_queue_spin_unlock(m);
1891 /* leaves vm_page spinlocked */
1895 * Attempt to deactivate a page.
1900 vm_page_deactivate(vm_page_t m)
1902 vm_page_spin_lock(m);
1903 _vm_page_deactivate_locked(m, 0);
1904 vm_page_spin_unlock(m);
1908 vm_page_deactivate_locked(vm_page_t m)
1910 _vm_page_deactivate_locked(m, 0);
1914 * Attempt to move a page to PQ_CACHE.
1916 * Returns 0 on failure, 1 on success
1918 * The page should NOT be busied by the caller. This function will validate
1919 * whether the page can be safely moved to the cache.
1922 vm_page_try_to_cache(vm_page_t m)
1924 vm_page_spin_lock(m);
1925 if (vm_page_busy_try(m, TRUE)) {
1926 vm_page_spin_unlock(m);
1929 if (m->dirty || m->hold_count || m->wire_count ||
1930 (m->flags & PG_UNMANAGED)) {
1931 if (_vm_page_wakeup(m)) {
1932 vm_page_spin_unlock(m);
1935 vm_page_spin_unlock(m);
1939 vm_page_spin_unlock(m);
1942 * Page busied by us and no longer spinlocked. Dirty pages cannot
1943 * be moved to the cache.
1945 vm_page_test_dirty(m);
1955 * Attempt to free the page. If we cannot free it, we do nothing.
1956 * 1 is returned on success, 0 on failure.
1961 vm_page_try_to_free(vm_page_t m)
1963 vm_page_spin_lock(m);
1964 if (vm_page_busy_try(m, TRUE)) {
1965 vm_page_spin_unlock(m);
1968 if (m->dirty || m->hold_count || m->wire_count ||
1969 (m->flags & PG_UNMANAGED)) {
1970 if (_vm_page_wakeup(m)) {
1971 vm_page_spin_unlock(m);
1974 vm_page_spin_unlock(m);
1978 vm_page_spin_unlock(m);
1981 * Page busied by us and no longer spinlocked. Dirty pages will
1982 * not be freed by this function. We have to re-test the
1983 * dirty bit after cleaning out the pmaps.
1985 vm_page_test_dirty(m);
1990 vm_page_protect(m, VM_PROT_NONE);
2002 * Put the specified page onto the page cache queue (if appropriate).
2004 * The page must be busy, and this routine will release the busy and
2005 * possibly even free the page.
2008 vm_page_cache(vm_page_t m)
2010 if ((m->flags & PG_UNMANAGED) || m->busy ||
2011 m->wire_count || m->hold_count) {
2012 kprintf("vm_page_cache: attempting to cache busy/held page\n");
2018 * Already in the cache (and thus not mapped)
2020 if ((m->queue - m->pc) == PQ_CACHE) {
2021 KKASSERT((m->flags & PG_MAPPED) == 0);
2027 * Caller is required to test m->dirty, but note that the act of
2028 * removing the page from its maps can cause it to become dirty
2029 * on an SMP system due to another cpu running in usermode.
2032 panic("vm_page_cache: caching a dirty page, pindex: %ld",
2037 * Remove all pmaps and indicate that the page is not
2038 * writeable or mapped. Our vm_page_protect() call may
2039 * have blocked (especially w/ VM_PROT_NONE), so recheck
2042 vm_page_protect(m, VM_PROT_NONE);
2043 if ((m->flags & (PG_UNMANAGED|PG_MAPPED)) || m->busy ||
2044 m->wire_count || m->hold_count) {
2046 } else if (m->dirty) {
2047 vm_page_deactivate(m);
2050 _vm_page_and_queue_spin_lock(m);
2051 _vm_page_rem_queue_spinlocked(m);
2052 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0);
2053 _vm_page_queue_spin_unlock(m);
2054 if (_vm_page_wakeup(m)) {
2055 vm_page_spin_unlock(m);
2058 vm_page_spin_unlock(m);
2060 vm_page_free_wakeup();
2065 * vm_page_dontneed()
2067 * Cache, deactivate, or do nothing as appropriate. This routine
2068 * is typically used by madvise() MADV_DONTNEED.
2070 * Generally speaking we want to move the page into the cache so
2071 * it gets reused quickly. However, this can result in a silly syndrome
2072 * due to the page recycling too quickly. Small objects will not be
2073 * fully cached. On the otherhand, if we move the page to the inactive
2074 * queue we wind up with a problem whereby very large objects
2075 * unnecessarily blow away our inactive and cache queues.
2077 * The solution is to move the pages based on a fixed weighting. We
2078 * either leave them alone, deactivate them, or move them to the cache,
2079 * where moving them to the cache has the highest weighting.
2080 * By forcing some pages into other queues we eventually force the
2081 * system to balance the queues, potentially recovering other unrelated
2082 * space from active. The idea is to not force this to happen too
2085 * The page must be busied.
2088 vm_page_dontneed(vm_page_t m)
2090 static int dnweight;
2097 * occassionally leave the page alone
2099 if ((dnw & 0x01F0) == 0 ||
2100 m->queue - m->pc == PQ_INACTIVE ||
2101 m->queue - m->pc == PQ_CACHE
2103 if (m->act_count >= ACT_INIT)
2109 * If vm_page_dontneed() is inactivating a page, it must clear
2110 * the referenced flag; otherwise the pagedaemon will see references
2111 * on the page in the inactive queue and reactivate it. Until the
2112 * page can move to the cache queue, madvise's job is not done.
2114 vm_page_flag_clear(m, PG_REFERENCED);
2115 pmap_clear_reference(m);
2118 vm_page_test_dirty(m);
2120 if (m->dirty || (dnw & 0x0070) == 0) {
2122 * Deactivate the page 3 times out of 32.
2127 * Cache the page 28 times out of every 32. Note that
2128 * the page is deactivated instead of cached, but placed
2129 * at the head of the queue instead of the tail.
2133 vm_page_spin_lock(m);
2134 _vm_page_deactivate_locked(m, head);
2135 vm_page_spin_unlock(m);
2139 * These routines manipulate the 'soft busy' count for a page. A soft busy
2140 * is almost like PG_BUSY except that it allows certain compatible operations
2141 * to occur on the page while it is busy. For example, a page undergoing a
2142 * write can still be mapped read-only.
2144 * Because vm_pages can overlap buffers m->busy can be > 1. m->busy is only
2145 * adjusted while the vm_page is PG_BUSY so the flash will occur when the
2146 * busy bit is cleared.
2149 vm_page_io_start(vm_page_t m)
2151 KASSERT(m->flags & PG_BUSY, ("vm_page_io_start: page not busy!!!"));
2152 atomic_add_char(&m->busy, 1);
2153 vm_page_flag_set(m, PG_SBUSY);
2157 vm_page_io_finish(vm_page_t m)
2159 KASSERT(m->flags & PG_BUSY, ("vm_page_io_finish: page not busy!!!"));
2160 atomic_subtract_char(&m->busy, 1);
2162 vm_page_flag_clear(m, PG_SBUSY);
2166 * Grab a page, blocking if it is busy and allocating a page if necessary.
2167 * A busy page is returned or NULL. The page may or may not be valid and
2168 * might not be on a queue (the caller is responsible for the disposition of
2171 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the
2172 * page will be zero'd and marked valid.
2174 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked
2175 * valid even if it already exists.
2177 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also
2178 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified.
2180 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
2181 * always returned if we had blocked.
2183 * This routine may not be called from an interrupt.
2185 * PG_ZERO is *ALWAYS* cleared by this routine.
2187 * No other requirements.
2190 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2195 KKASSERT(allocflags &
2196 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
2197 vm_object_hold(object);
2199 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error);
2201 vm_page_sleep_busy(m, TRUE, "pgrbwt");
2202 if ((allocflags & VM_ALLOC_RETRY) == 0) {
2207 } else if (m == NULL) {
2208 m = vm_page_alloc(object, pindex,
2209 allocflags & ~VM_ALLOC_RETRY);
2213 if ((allocflags & VM_ALLOC_RETRY) == 0)
2222 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid.
2224 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set
2225 * valid even if already valid.
2227 if (m->valid == 0) {
2228 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) {
2229 if ((m->flags & PG_ZERO) == 0)
2230 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2231 m->valid = VM_PAGE_BITS_ALL;
2233 } else if (allocflags & VM_ALLOC_FORCE_ZERO) {
2234 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2235 m->valid = VM_PAGE_BITS_ALL;
2237 vm_page_flag_clear(m, PG_ZERO);
2239 vm_object_drop(object);
2244 * Mapping function for valid bits or for dirty bits in
2245 * a page. May not block.
2247 * Inputs are required to range within a page.
2253 vm_page_bits(int base, int size)
2259 base + size <= PAGE_SIZE,
2260 ("vm_page_bits: illegal base/size %d/%d", base, size)
2263 if (size == 0) /* handle degenerate case */
2266 first_bit = base >> DEV_BSHIFT;
2267 last_bit = (base + size - 1) >> DEV_BSHIFT;
2269 return ((2 << last_bit) - (1 << first_bit));
2273 * Sets portions of a page valid and clean. The arguments are expected
2274 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2275 * of any partial chunks touched by the range. The invalid portion of
2276 * such chunks will be zero'd.
2278 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically
2279 * align base to DEV_BSIZE so as not to mark clean a partially
2280 * truncated device block. Otherwise the dirty page status might be
2283 * This routine may not block.
2285 * (base + size) must be less then or equal to PAGE_SIZE.
2288 _vm_page_zero_valid(vm_page_t m, int base, int size)
2293 if (size == 0) /* handle degenerate case */
2297 * If the base is not DEV_BSIZE aligned and the valid
2298 * bit is clear, we have to zero out a portion of the
2302 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2303 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
2305 pmap_zero_page_area(
2313 * If the ending offset is not DEV_BSIZE aligned and the
2314 * valid bit is clear, we have to zero out a portion of
2318 endoff = base + size;
2320 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2321 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
2323 pmap_zero_page_area(
2326 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
2332 * Set valid, clear dirty bits. If validating the entire
2333 * page we can safely clear the pmap modify bit. We also
2334 * use this opportunity to clear the PG_NOSYNC flag. If a process
2335 * takes a write fault on a MAP_NOSYNC memory area the flag will
2338 * We set valid bits inclusive of any overlap, but we can only
2339 * clear dirty bits for DEV_BSIZE chunks that are fully within
2342 * Page must be busied?
2343 * No other requirements.
2346 vm_page_set_valid(vm_page_t m, int base, int size)
2348 _vm_page_zero_valid(m, base, size);
2349 m->valid |= vm_page_bits(base, size);
2354 * Set valid bits and clear dirty bits.
2356 * NOTE: This function does not clear the pmap modified bit.
2357 * Also note that e.g. NFS may use a byte-granular base
2360 * WARNING: Page must be busied? But vfs_clean_one_page() will call
2361 * this without necessarily busying the page (via bdwrite()).
2362 * So for now vm_token must also be held.
2364 * No other requirements.
2367 vm_page_set_validclean(vm_page_t m, int base, int size)
2371 _vm_page_zero_valid(m, base, size);
2372 pagebits = vm_page_bits(base, size);
2373 m->valid |= pagebits;
2374 m->dirty &= ~pagebits;
2375 if (base == 0 && size == PAGE_SIZE) {
2376 /*pmap_clear_modify(m);*/
2377 vm_page_flag_clear(m, PG_NOSYNC);
2382 * Set valid & dirty. Used by buwrite()
2384 * WARNING: Page must be busied? But vfs_dirty_one_page() will
2385 * call this function in buwrite() so for now vm_token must
2388 * No other requirements.
2391 vm_page_set_validdirty(vm_page_t m, int base, int size)
2395 pagebits = vm_page_bits(base, size);
2396 m->valid |= pagebits;
2397 m->dirty |= pagebits;
2399 vm_object_set_writeable_dirty(m->object);
2405 * NOTE: This function does not clear the pmap modified bit.
2406 * Also note that e.g. NFS may use a byte-granular base
2409 * Page must be busied?
2410 * No other requirements.
2413 vm_page_clear_dirty(vm_page_t m, int base, int size)
2415 m->dirty &= ~vm_page_bits(base, size);
2416 if (base == 0 && size == PAGE_SIZE) {
2417 /*pmap_clear_modify(m);*/
2418 vm_page_flag_clear(m, PG_NOSYNC);
2423 * Make the page all-dirty.
2425 * Also make sure the related object and vnode reflect the fact that the
2426 * object may now contain a dirty page.
2428 * Page must be busied?
2429 * No other requirements.
2432 vm_page_dirty(vm_page_t m)
2435 int pqtype = m->queue - m->pc;
2437 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE,
2438 ("vm_page_dirty: page in free/cache queue!"));
2439 if (m->dirty != VM_PAGE_BITS_ALL) {
2440 m->dirty = VM_PAGE_BITS_ALL;
2442 vm_object_set_writeable_dirty(m->object);
2447 * Invalidates DEV_BSIZE'd chunks within a page. Both the
2448 * valid and dirty bits for the effected areas are cleared.
2450 * Page must be busied?
2452 * No other requirements.
2455 vm_page_set_invalid(vm_page_t m, int base, int size)
2459 bits = vm_page_bits(base, size);
2462 m->object->generation++;
2466 * The kernel assumes that the invalid portions of a page contain
2467 * garbage, but such pages can be mapped into memory by user code.
2468 * When this occurs, we must zero out the non-valid portions of the
2469 * page so user code sees what it expects.
2471 * Pages are most often semi-valid when the end of a file is mapped
2472 * into memory and the file's size is not page aligned.
2474 * Page must be busied?
2475 * No other requirements.
2478 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
2484 * Scan the valid bits looking for invalid sections that
2485 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
2486 * valid bit may be set ) have already been zerod by
2487 * vm_page_set_validclean().
2489 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
2490 if (i == (PAGE_SIZE / DEV_BSIZE) ||
2491 (m->valid & (1 << i))
2494 pmap_zero_page_area(
2497 (i - b) << DEV_BSHIFT
2505 * setvalid is TRUE when we can safely set the zero'd areas
2506 * as being valid. We can do this if there are no cache consistency
2507 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
2510 m->valid = VM_PAGE_BITS_ALL;
2514 * Is a (partial) page valid? Note that the case where size == 0
2515 * will return FALSE in the degenerate case where the page is entirely
2516 * invalid, and TRUE otherwise.
2519 * No other requirements.
2522 vm_page_is_valid(vm_page_t m, int base, int size)
2524 int bits = vm_page_bits(base, size);
2526 if (m->valid && ((m->valid & bits) == bits))
2533 * update dirty bits from pmap/mmu. May not block.
2535 * Caller must hold the page busy
2538 vm_page_test_dirty(vm_page_t m)
2540 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
2546 * Register an action, associating it with its vm_page
2549 vm_page_register_action(vm_page_action_t action, vm_page_event_t event)
2551 struct vm_page_action_list *list;
2554 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
2555 list = &action_list[hv];
2557 lwkt_gettoken(&vm_token);
2558 vm_page_flag_set(action->m, PG_ACTIONLIST);
2559 action->event = event;
2560 LIST_INSERT_HEAD(list, action, entry);
2561 lwkt_reltoken(&vm_token);
2565 * Unregister an action, disassociating it from its related vm_page
2568 vm_page_unregister_action(vm_page_action_t action)
2570 struct vm_page_action_list *list;
2573 lwkt_gettoken(&vm_token);
2574 if (action->event != VMEVENT_NONE) {
2575 action->event = VMEVENT_NONE;
2576 LIST_REMOVE(action, entry);
2578 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
2579 list = &action_list[hv];
2580 if (LIST_EMPTY(list))
2581 vm_page_flag_clear(action->m, PG_ACTIONLIST);
2583 lwkt_reltoken(&vm_token);
2587 * Issue an event on a VM page. Corresponding action structures are
2588 * removed from the page's list and called.
2590 * If the vm_page has no more pending action events we clear its
2591 * PG_ACTIONLIST flag.
2594 vm_page_event_internal(vm_page_t m, vm_page_event_t event)
2596 struct vm_page_action_list *list;
2597 struct vm_page_action *scan;
2598 struct vm_page_action *next;
2602 hv = (int)((intptr_t)m >> 8) & VMACTION_HMASK;
2603 list = &action_list[hv];
2606 lwkt_gettoken(&vm_token);
2607 LIST_FOREACH_MUTABLE(scan, list, entry, next) {
2609 if (scan->event == event) {
2610 scan->event = VMEVENT_NONE;
2611 LIST_REMOVE(scan, entry);
2612 scan->func(m, scan);
2620 vm_page_flag_clear(m, PG_ACTIONLIST);
2621 lwkt_reltoken(&vm_token);
2624 #include "opt_ddb.h"
2626 #include <sys/kernel.h>
2628 #include <ddb/ddb.h>
2630 DB_SHOW_COMMAND(page, vm_page_print_page_info)
2632 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count);
2633 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count);
2634 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count);
2635 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count);
2636 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count);
2637 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved);
2638 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min);
2639 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target);
2640 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min);
2641 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target);
2644 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
2647 db_printf("PQ_FREE:");
2648 for(i=0;i<PQ_L2_SIZE;i++) {
2649 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
2653 db_printf("PQ_CACHE:");
2654 for(i=0;i<PQ_L2_SIZE;i++) {
2655 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
2659 db_printf("PQ_ACTIVE:");
2660 for(i=0;i<PQ_L2_SIZE;i++) {
2661 db_printf(" %d", vm_page_queues[PQ_ACTIVE + i].lcnt);
2665 db_printf("PQ_INACTIVE:");
2666 for(i=0;i<PQ_L2_SIZE;i++) {
2667 db_printf(" %d", vm_page_queues[PQ_INACTIVE + i].lcnt);