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(u_short pg_color);
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;
186 * Twist for cpu localization instead of page coloring.
188 m->pc ^= ((pa >> PAGE_SHIFT) / PQ_L2_SIZE) & PQ_L2_MASK;
189 m->pc ^= ((pa >> PAGE_SHIFT) / (PQ_L2_SIZE * PQ_L2_SIZE)) & PQ_L2_MASK;
191 m->queue = m->pc + PQ_FREE;
192 KKASSERT(m->dirty == 0);
194 atomic_add_int(&vmstats.v_page_count, 1);
195 atomic_add_int(&vmstats.v_free_count, 1);
196 vpq = &vm_page_queues[m->queue];
198 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
200 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
201 vpq->flipflop = 1 - vpq->flipflop;
210 * Initializes the resident memory module.
212 * Preallocates memory for critical VM structures and arrays prior to
213 * kernel_map becoming available.
215 * Memory is allocated from (virtual2_start, virtual2_end) if available,
216 * otherwise memory is allocated from (virtual_start, virtual_end).
218 * On x86-64 (virtual_start, virtual_end) is only 2GB and may not be
219 * large enough to hold vm_page_array & other structures for machines with
220 * large amounts of ram, so we want to use virtual2* when available.
223 vm_page_startup(void)
225 vm_offset_t vaddr = virtual2_start ? virtual2_start : virtual_start;
228 vm_paddr_t page_range;
235 vm_paddr_t biggestone, biggestsize;
242 vaddr = round_page(vaddr);
244 for (i = 0; phys_avail[i + 1]; i += 2) {
245 phys_avail[i] = round_page64(phys_avail[i]);
246 phys_avail[i + 1] = trunc_page64(phys_avail[i + 1]);
249 for (i = 0; phys_avail[i + 1]; i += 2) {
250 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
252 if (size > biggestsize) {
260 end = phys_avail[biggestone+1];
261 end = trunc_page(end);
264 * Initialize the queue headers for the free queue, the active queue
265 * and the inactive queue.
268 vm_page_queue_init();
270 #if !defined(_KERNEL_VIRTUAL)
272 * VKERNELs don't support minidumps and as such don't need
275 * Allocate a bitmap to indicate that a random physical page
276 * needs to be included in a minidump.
278 * The amd64 port needs this to indicate which direct map pages
279 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
281 * However, i386 still needs this workspace internally within the
282 * minidump code. In theory, they are not needed on i386, but are
283 * included should the sf_buf code decide to use them.
285 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE;
286 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
287 end -= vm_page_dump_size;
288 vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size,
289 VM_PROT_READ | VM_PROT_WRITE);
290 bzero((void *)vm_page_dump, vm_page_dump_size);
294 * Compute the number of pages of memory that will be available for
295 * use (taking into account the overhead of a page structure per
298 first_page = phys_avail[0] / PAGE_SIZE;
299 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
300 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE;
303 * Initialize the mem entry structures now, and put them in the free
306 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
307 mapped = pmap_map(&vaddr, new_end, end,
308 VM_PROT_READ | VM_PROT_WRITE);
309 vm_page_array = (vm_page_t)mapped;
311 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL)
313 * since pmap_map on amd64 returns stuff out of a direct-map region,
314 * we have to manually add these pages to the minidump tracking so
315 * that they can be dumped, including the vm_page_array.
317 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
322 * Clear all of the page structures
324 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
325 vm_page_array_size = page_range;
328 * Construct the free queue(s) in ascending order (by physical
329 * address) so that the first 16MB of physical memory is allocated
330 * last rather than first. On large-memory machines, this avoids
331 * the exhaustion of low physical memory before isa_dmainit has run.
333 vmstats.v_page_count = 0;
334 vmstats.v_free_count = 0;
335 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
340 last_pa = phys_avail[i + 1];
341 while (pa < last_pa && npages-- > 0) {
347 virtual2_start = vaddr;
349 virtual_start = vaddr;
353 * Scan comparison function for Red-Black tree scans. An inclusive
354 * (start,end) is expected. Other fields are not used.
357 rb_vm_page_scancmp(struct vm_page *p, void *data)
359 struct rb_vm_page_scan_info *info = data;
361 if (p->pindex < info->start_pindex)
363 if (p->pindex > info->end_pindex)
369 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2)
371 if (p1->pindex < p2->pindex)
373 if (p1->pindex > p2->pindex)
379 * Each page queue has its own spin lock, which is fairly optimal for
380 * allocating and freeing pages at least.
382 * The caller must hold the vm_page_spin_lock() before locking a vm_page's
383 * queue spinlock via this function. Also note that m->queue cannot change
384 * unless both the page and queue are locked.
388 _vm_page_queue_spin_lock(vm_page_t m)
393 if (queue != PQ_NONE) {
394 spin_lock(&vm_page_queues[queue].spin);
395 KKASSERT(queue == m->queue);
401 _vm_page_queue_spin_unlock(vm_page_t m)
407 if (queue != PQ_NONE)
408 spin_unlock(&vm_page_queues[queue].spin);
413 _vm_page_queues_spin_lock(u_short queue)
416 if (queue != PQ_NONE)
417 spin_lock(&vm_page_queues[queue].spin);
423 _vm_page_queues_spin_unlock(u_short queue)
426 if (queue != PQ_NONE)
427 spin_unlock(&vm_page_queues[queue].spin);
431 vm_page_queue_spin_lock(vm_page_t m)
433 _vm_page_queue_spin_lock(m);
437 vm_page_queues_spin_lock(u_short queue)
439 _vm_page_queues_spin_lock(queue);
443 vm_page_queue_spin_unlock(vm_page_t m)
445 _vm_page_queue_spin_unlock(m);
449 vm_page_queues_spin_unlock(u_short queue)
451 _vm_page_queues_spin_unlock(queue);
455 * This locks the specified vm_page and its queue in the proper order
456 * (page first, then queue). The queue may change so the caller must
461 _vm_page_and_queue_spin_lock(vm_page_t m)
463 vm_page_spin_lock(m);
464 _vm_page_queue_spin_lock(m);
469 _vm_page_and_queue_spin_unlock(vm_page_t m)
471 _vm_page_queues_spin_unlock(m->queue);
472 vm_page_spin_unlock(m);
476 vm_page_and_queue_spin_unlock(vm_page_t m)
478 _vm_page_and_queue_spin_unlock(m);
482 vm_page_and_queue_spin_lock(vm_page_t m)
484 _vm_page_and_queue_spin_lock(m);
488 * Helper function removes vm_page from its current queue.
489 * Returns the base queue the page used to be on.
491 * The vm_page and the queue must be spinlocked.
492 * This function will unlock the queue but leave the page spinlocked.
494 static __inline u_short
495 _vm_page_rem_queue_spinlocked(vm_page_t m)
497 struct vpgqueues *pq;
501 if (queue != PQ_NONE) {
502 pq = &vm_page_queues[queue];
503 TAILQ_REMOVE(&pq->pl, m, pageq);
504 atomic_add_int(pq->cnt, -1);
507 vm_page_queues_spin_unlock(queue);
508 if ((queue - m->pc) == PQ_FREE && (m->flags & PG_ZERO))
509 atomic_subtract_int(&vm_page_zero_count, 1);
510 if ((queue - m->pc) == PQ_CACHE || (queue - m->pc) == PQ_FREE)
511 return (queue - m->pc);
517 * Helper function places the vm_page on the specified queue.
519 * The vm_page must be spinlocked.
520 * This function will return with both the page and the queue locked.
523 _vm_page_add_queue_spinlocked(vm_page_t m, u_short queue, int athead)
525 struct vpgqueues *pq;
527 KKASSERT(m->queue == PQ_NONE);
529 if (queue != PQ_NONE) {
530 vm_page_queues_spin_lock(queue);
531 pq = &vm_page_queues[queue];
533 atomic_add_int(pq->cnt, 1);
537 * Put zero'd pages on the end ( where we look for zero'd pages
538 * first ) and non-zerod pages at the head.
540 if (queue - m->pc == PQ_FREE) {
541 if (m->flags & PG_ZERO) {
542 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
543 atomic_add_int(&vm_page_zero_count, 1);
545 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
548 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
550 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
552 /* leave the queue spinlocked */
557 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE)
558 * m->busy is zero. Returns TRUE if it had to sleep, FALSE if we
559 * did not. Only one sleep call will be made before returning.
561 * This function does NOT busy the page and on return the page is not
562 * guaranteed to be available.
565 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg)
573 if ((flags & PG_BUSY) == 0 &&
574 (also_m_busy == 0 || (flags & PG_SBUSY) == 0)) {
577 tsleep_interlock(m, 0);
578 if (atomic_cmpset_int(&m->flags, flags,
579 flags | PG_WANTED | PG_REFERENCED)) {
580 tsleep(m, PINTERLOCKED, msg, 0);
587 * Wait until PG_BUSY can be set, then set it. If also_m_busy is TRUE we
588 * also wait for m->busy to become 0 before setting PG_BUSY.
591 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m,
592 int also_m_busy, const char *msg
600 if (flags & PG_BUSY) {
601 tsleep_interlock(m, 0);
602 if (atomic_cmpset_int(&m->flags, flags,
603 flags | PG_WANTED | PG_REFERENCED)) {
604 tsleep(m, PINTERLOCKED, msg, 0);
606 } else if (also_m_busy && (flags & PG_SBUSY)) {
607 tsleep_interlock(m, 0);
608 if (atomic_cmpset_int(&m->flags, flags,
609 flags | PG_WANTED | PG_REFERENCED)) {
610 tsleep(m, PINTERLOCKED, msg, 0);
613 if (atomic_cmpset_int(&m->flags, flags,
617 m->busy_line = lineno;
626 * Attempt to set PG_BUSY. If also_m_busy is TRUE we only succeed if m->busy
629 * Returns non-zero on failure.
632 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy
642 if (also_m_busy && (flags & PG_SBUSY))
644 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
647 m->busy_line = lineno;
655 * Clear the PG_BUSY flag and return non-zero to indicate to the caller
656 * that a wakeup() should be performed.
658 * The vm_page must be spinlocked and will remain spinlocked on return.
659 * The related queue must NOT be spinlocked (which could deadlock us).
665 _vm_page_wakeup(vm_page_t m)
672 if (atomic_cmpset_int(&m->flags, flags,
673 flags & ~(PG_BUSY | PG_WANTED))) {
677 return(flags & PG_WANTED);
681 * Clear the PG_BUSY flag and wakeup anyone waiting for the page. This
682 * is typically the last call you make on a page before moving onto
686 vm_page_wakeup(vm_page_t m)
688 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!"));
689 vm_page_spin_lock(m);
690 if (_vm_page_wakeup(m)) {
691 vm_page_spin_unlock(m);
694 vm_page_spin_unlock(m);
699 * Holding a page keeps it from being reused. Other parts of the system
700 * can still disassociate the page from its current object and free it, or
701 * perform read or write I/O on it and/or otherwise manipulate the page,
702 * but if the page is held the VM system will leave the page and its data
703 * intact and not reuse the page for other purposes until the last hold
704 * reference is released. (see vm_page_wire() if you want to prevent the
705 * page from being disassociated from its object too).
707 * The caller must still validate the contents of the page and, if necessary,
708 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete
709 * before manipulating the page.
711 * XXX get vm_page_spin_lock() here and move FREE->HOLD if necessary
714 vm_page_hold(vm_page_t m)
716 vm_page_spin_lock(m);
717 atomic_add_int(&m->hold_count, 1);
718 if (m->queue - m->pc == PQ_FREE) {
719 _vm_page_queue_spin_lock(m);
720 _vm_page_rem_queue_spinlocked(m);
721 _vm_page_add_queue_spinlocked(m, PQ_HOLD, 0);
722 _vm_page_queue_spin_unlock(m);
724 vm_page_spin_unlock(m);
728 * The opposite of vm_page_hold(). A page can be freed while being held,
729 * which places it on the PQ_HOLD queue. If we are able to busy the page
730 * after the hold count drops to zero we will move the page to the
731 * appropriate PQ_FREE queue by calling vm_page_free_toq().
734 vm_page_unhold(vm_page_t m)
736 vm_page_spin_lock(m);
737 atomic_add_int(&m->hold_count, -1);
738 if (m->hold_count == 0 && m->queue == PQ_HOLD) {
739 _vm_page_queue_spin_lock(m);
740 _vm_page_rem_queue_spinlocked(m);
741 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0);
742 _vm_page_queue_spin_unlock(m);
744 vm_page_spin_unlock(m);
748 * Inserts the given vm_page into the object and object list.
750 * The pagetables are not updated but will presumably fault the page
751 * in if necessary, or if a kernel page the caller will at some point
752 * enter the page into the kernel's pmap. We are not allowed to block
753 * here so we *can't* do this anyway.
755 * This routine may not block.
756 * This routine must be called with the vm_object held.
757 * This routine must be called with a critical section held.
760 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
762 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
763 if (m->object != NULL)
764 panic("vm_page_insert: already inserted");
766 object->generation++;
767 object->resident_page_count++;
770 * Record the object/offset pair in this page and add the
771 * pv_list_count of the page to the object.
773 * The vm_page spin lock is required for interactions with the pmap.
775 vm_page_spin_lock(m);
778 vm_page_rb_tree_RB_INSERT(&object->rb_memq, m);
779 /* atomic_add_int(&object->agg_pv_list_count, m->md.pv_list_count); */
780 vm_page_spin_unlock(m);
783 * Since we are inserting a new and possibly dirty page,
784 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
786 if ((m->valid & m->dirty) || (m->flags & PG_WRITEABLE))
787 vm_object_set_writeable_dirty(object);
790 * Checks for a swap assignment and sets PG_SWAPPED if appropriate.
792 swap_pager_page_inserted(m);
796 * Removes the given vm_page_t from the (object,index) table
798 * The underlying pmap entry (if any) is NOT removed here.
799 * This routine may not block.
801 * The page must be BUSY and will remain BUSY on return.
802 * No other requirements.
804 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave
808 vm_page_remove(vm_page_t m)
812 if (m->object == NULL) {
816 if ((m->flags & PG_BUSY) == 0)
817 panic("vm_page_remove: page not busy");
821 vm_object_hold(object);
824 * Remove the page from the object and update the object.
826 * The vm_page spin lock is required for interactions with the pmap.
828 vm_page_spin_lock(m);
829 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m);
830 object->resident_page_count--;
831 /* atomic_add_int(&object->agg_pv_list_count, -m->md.pv_list_count); */
833 vm_page_spin_unlock(m);
835 object->generation++;
837 vm_object_drop(object);
841 * Locate and return the page at (object, pindex), or NULL if the
842 * page could not be found.
844 * The caller must hold the vm_object token.
847 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
852 * Search the hash table for this object/offset pair
854 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
855 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
856 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex));
861 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object,
863 int also_m_busy, const char *msg
869 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
870 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
872 KKASSERT(m->object == object && m->pindex == pindex);
875 if (flags & PG_BUSY) {
876 tsleep_interlock(m, 0);
877 if (atomic_cmpset_int(&m->flags, flags,
878 flags | PG_WANTED | PG_REFERENCED)) {
879 tsleep(m, PINTERLOCKED, msg, 0);
880 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
883 } else if (also_m_busy && (flags & PG_SBUSY)) {
884 tsleep_interlock(m, 0);
885 if (atomic_cmpset_int(&m->flags, flags,
886 flags | PG_WANTED | PG_REFERENCED)) {
887 tsleep(m, PINTERLOCKED, msg, 0);
888 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
891 } else if (atomic_cmpset_int(&m->flags, flags,
895 m->busy_line = lineno;
904 * Attempt to lookup and busy a page.
906 * Returns NULL if the page could not be found
908 * Returns a vm_page and error == TRUE if the page exists but could not
911 * Returns a vm_page and error == FALSE on success.
914 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object,
916 int also_m_busy, int *errorp
922 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
923 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
926 KKASSERT(m->object == object && m->pindex == pindex);
929 if (flags & PG_BUSY) {
933 if (also_m_busy && (flags & PG_SBUSY)) {
937 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
940 m->busy_line = lineno;
949 * Caller must hold the related vm_object
952 vm_page_next(vm_page_t m)
956 next = vm_page_rb_tree_RB_NEXT(m);
957 if (next && next->pindex != m->pindex + 1)
965 * Move the given vm_page from its current object to the specified
966 * target object/offset. The page must be busy and will remain so
969 * new_object must be held.
970 * This routine might block. XXX ?
972 * NOTE: Swap associated with the page must be invalidated by the move. We
973 * have to do this for several reasons: (1) we aren't freeing the
974 * page, (2) we are dirtying the page, (3) the VM system is probably
975 * moving the page from object A to B, and will then later move
976 * the backing store from A to B and we can't have a conflict.
978 * NOTE: We *always* dirty the page. It is necessary both for the
979 * fact that we moved it, and because we may be invalidating
980 * swap. If the page is on the cache, we have to deactivate it
981 * or vm_page_dirty() will panic. Dirty pages are not allowed
985 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
987 KKASSERT(m->flags & PG_BUSY);
988 ASSERT_LWKT_TOKEN_HELD(vm_object_token(new_object));
990 ASSERT_LWKT_TOKEN_HELD(vm_object_token(m->object));
993 vm_page_insert(m, new_object, new_pindex);
994 if (m->queue - m->pc == PQ_CACHE)
995 vm_page_deactivate(m);
1000 * vm_page_unqueue() without any wakeup. This routine is used when a page
1001 * is being moved between queues or otherwise is to remain BUSYied by the
1004 * This routine may not block.
1007 vm_page_unqueue_nowakeup(vm_page_t m)
1009 vm_page_and_queue_spin_lock(m);
1010 (void)_vm_page_rem_queue_spinlocked(m);
1011 vm_page_spin_unlock(m);
1015 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
1018 * This routine may not block.
1021 vm_page_unqueue(vm_page_t m)
1025 vm_page_and_queue_spin_lock(m);
1026 queue = _vm_page_rem_queue_spinlocked(m);
1027 if (queue == PQ_FREE || queue == PQ_CACHE) {
1028 vm_page_spin_unlock(m);
1029 pagedaemon_wakeup();
1031 vm_page_spin_unlock(m);
1036 * vm_page_list_find()
1038 * Find a page on the specified queue with color optimization.
1040 * The page coloring optimization attempts to locate a page that does
1041 * not overload other nearby pages in the object in the cpu's L1 or L2
1042 * caches. We need this optimization because cpu caches tend to be
1043 * physical caches, while object spaces tend to be virtual.
1045 * On MP systems each PQ_FREE and PQ_CACHE color queue has its own spinlock
1046 * and the algorithm is adjusted to localize allocations on a per-core basis.
1047 * This is done by 'twisting' the colors.
1049 * The page is returned spinlocked and removed from its queue (it will
1050 * be on PQ_NONE), or NULL. The page is not PG_BUSY'd. The caller
1051 * is responsible for dealing with the busy-page case (usually by
1052 * deactivating the page and looping).
1054 * NOTE: This routine is carefully inlined. A non-inlined version
1055 * is available for outside callers but the only critical path is
1056 * from within this source file.
1058 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE
1059 * represent stable storage, allowing us to order our locks vm_page
1060 * first, then queue.
1064 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
1070 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist);
1072 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl);
1074 m = _vm_page_list_find2(basequeue, index);
1077 vm_page_and_queue_spin_lock(m);
1078 if (m->queue == basequeue + index) {
1079 _vm_page_rem_queue_spinlocked(m);
1080 /* vm_page_t spin held, no queue spin */
1083 vm_page_and_queue_spin_unlock(m);
1089 _vm_page_list_find2(int basequeue, int index)
1093 struct vpgqueues *pq;
1095 pq = &vm_page_queues[basequeue];
1098 * Note that for the first loop, index+i and index-i wind up at the
1099 * same place. Even though this is not totally optimal, we've already
1100 * blown it by missing the cache case so we do not care.
1102 for (i = PQ_L2_SIZE / 2; i > 0; --i) {
1104 m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl);
1106 _vm_page_and_queue_spin_lock(m);
1108 basequeue + ((index + i) & PQ_L2_MASK)) {
1109 _vm_page_rem_queue_spinlocked(m);
1112 _vm_page_and_queue_spin_unlock(m);
1115 m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl);
1117 _vm_page_and_queue_spin_lock(m);
1119 basequeue + ((index - i) & PQ_L2_MASK)) {
1120 _vm_page_rem_queue_spinlocked(m);
1123 _vm_page_and_queue_spin_unlock(m);
1133 * Returns a vm_page candidate for allocation. The page is not busied so
1134 * it can move around. The caller must busy the page (and typically
1135 * deactivate it if it cannot be busied!)
1137 * Returns a spinlocked vm_page that has been removed from its queue.
1140 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
1142 return(_vm_page_list_find(basequeue, index, prefer_zero));
1146 * Find a page on the cache queue with color optimization, remove it
1147 * from the queue, and busy it. The returned page will not be spinlocked.
1149 * A candidate failure will be deactivated. Candidates can fail due to
1150 * being busied by someone else, in which case they will be deactivated.
1152 * This routine may not block.
1156 vm_page_select_cache(u_short pg_color)
1161 m = _vm_page_list_find(PQ_CACHE, pg_color & PQ_L2_MASK, FALSE);
1165 * (m) has been removed from its queue and spinlocked
1167 if (vm_page_busy_try(m, TRUE)) {
1168 _vm_page_deactivate_locked(m, 0);
1169 vm_page_spin_unlock(m);
1171 kprintf("Warning: busy page %p found in cache\n", m);
1175 * We successfully busied the page
1177 if ((m->flags & PG_UNMANAGED) == 0 &&
1178 m->hold_count == 0 &&
1179 m->wire_count == 0) {
1180 vm_page_spin_unlock(m);
1181 pagedaemon_wakeup();
1184 _vm_page_deactivate_locked(m, 0);
1185 if (_vm_page_wakeup(m)) {
1186 vm_page_spin_unlock(m);
1189 vm_page_spin_unlock(m);
1197 * Find a free or zero page, with specified preference. We attempt to
1198 * inline the nominal case and fall back to _vm_page_select_free()
1199 * otherwise. A busied page is removed from the queue and returned.
1201 * This routine may not block.
1203 static __inline vm_page_t
1204 vm_page_select_free(u_short pg_color, boolean_t prefer_zero)
1209 m = _vm_page_list_find(PQ_FREE, pg_color & PQ_L2_MASK,
1213 if (vm_page_busy_try(m, TRUE)) {
1214 _vm_page_deactivate_locked(m, 0);
1215 vm_page_spin_unlock(m);
1217 kprintf("Warning: busy page %p found in cache\n", m);
1220 KKASSERT((m->flags & PG_UNMANAGED) == 0);
1221 KKASSERT(m->hold_count == 0);
1222 KKASSERT(m->wire_count == 0);
1223 vm_page_spin_unlock(m);
1224 pagedaemon_wakeup();
1226 /* return busied and removed page */
1236 * Allocate and return a memory cell associated with this VM object/offset
1237 * pair. If object is NULL an unassociated page will be allocated.
1241 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain
1242 * VM_ALLOC_QUICK like normal but cannot use cache
1243 * VM_ALLOC_SYSTEM greater free drain
1244 * VM_ALLOC_INTERRUPT allow free list to be completely drained
1245 * VM_ALLOC_ZERO advisory request for pre-zero'd page
1247 * The object must be locked if not NULL
1248 * This routine may not block
1249 * The returned page will be marked PG_BUSY
1251 * Additional special handling is required when called from an interrupt
1252 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache
1256 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
1263 * Cpu twist - cpu localization algorithm
1266 pg_color = mycpu->gd_cpuid + (pindex & ~ncpus_fit_mask) +
1267 (object->pg_color & ~ncpus_fit_mask);
1268 KASSERT(vm_page_lookup(object, pindex) == NULL,
1269 ("vm_page_alloc: page already allocated"));
1271 pg_color = mycpu->gd_cpuid + (pindex & ~ncpus_fit_mask);
1275 * Normal page coloring algorithm
1278 pg_color = object->pg_color + pindex;
1279 KASSERT(vm_page_lookup(object, pindex) == NULL,
1280 ("vm_page_alloc: page already allocated"));
1286 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK|
1287 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
1290 * Certain system threads (pageout daemon, buf_daemon's) are
1291 * allowed to eat deeper into the free page list.
1293 if (curthread->td_flags & TDF_SYSTHREAD)
1294 page_req |= VM_ALLOC_SYSTEM;
1297 if (vmstats.v_free_count > vmstats.v_free_reserved ||
1298 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) ||
1299 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 &&
1300 vmstats.v_free_count > vmstats.v_interrupt_free_min)
1303 * The free queue has sufficient free pages to take one out.
1305 if (page_req & VM_ALLOC_ZERO)
1306 m = vm_page_select_free(pg_color, TRUE);
1308 m = vm_page_select_free(pg_color, FALSE);
1309 } else if (page_req & VM_ALLOC_NORMAL) {
1311 * Allocatable from the cache (non-interrupt only). On
1312 * success, we must free the page and try again, thus
1313 * ensuring that vmstats.v_*_free_min counters are replenished.
1316 if (curthread->td_preempted) {
1317 kprintf("vm_page_alloc(): warning, attempt to allocate"
1318 " cache page from preempting interrupt\n");
1321 m = vm_page_select_cache(pg_color);
1324 m = vm_page_select_cache(pg_color);
1327 * On success move the page into the free queue and loop.
1330 KASSERT(m->dirty == 0,
1331 ("Found dirty cache page %p", m));
1332 vm_page_protect(m, VM_PROT_NONE);
1338 * On failure return NULL
1340 #if defined(DIAGNOSTIC)
1341 if (vmstats.v_cache_count > 0)
1342 kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count);
1344 vm_pageout_deficit++;
1345 pagedaemon_wakeup();
1349 * No pages available, wakeup the pageout daemon and give up.
1351 vm_pageout_deficit++;
1352 pagedaemon_wakeup();
1357 * Good page found. The page has already been busied for us.
1359 * v_free_count can race so loop if we don't find the expected
1364 KASSERT(m->dirty == 0,
1365 ("vm_page_alloc: free/cache page %p was dirty", m));
1368 * NOTE: page has already been removed from its queue and busied.
1370 KKASSERT(m->queue == PQ_NONE);
1373 * Initialize structure. Only the PG_ZERO flag is inherited. Set
1376 vm_page_flag_clear(m, ~(PG_ZERO | PG_BUSY));
1377 KKASSERT(m->wire_count == 0);
1378 KKASSERT(m->busy == 0);
1383 * Caller must be holding the object lock (asserted by
1384 * vm_page_insert()).
1386 * NOTE: Inserting a page here does not insert it into any pmaps
1387 * (which could cause us to block allocating memory).
1389 * NOTE: If no object an unassociated page is allocated, m->pindex
1390 * can be used by the caller for any purpose.
1393 vm_page_insert(m, object, pindex);
1398 * Don't wakeup too often - wakeup the pageout daemon when
1399 * we would be nearly out of memory.
1401 pagedaemon_wakeup();
1404 * A PG_BUSY page is returned.
1410 * Wait for sufficient free memory for nominal heavy memory use kernel
1414 vm_wait_nominal(void)
1416 while (vm_page_count_min(0))
1421 * Test if vm_wait_nominal() would block.
1424 vm_test_nominal(void)
1426 if (vm_page_count_min(0))
1432 * Block until free pages are available for allocation, called in various
1433 * places before memory allocations.
1435 * The caller may loop if vm_page_count_min() == FALSE so we cannot be
1436 * more generous then that.
1442 * never wait forever
1446 lwkt_gettoken(&vm_token);
1448 if (curthread == pagethread) {
1450 * The pageout daemon itself needs pages, this is bad.
1452 if (vm_page_count_min(0)) {
1453 vm_pageout_pages_needed = 1;
1454 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo);
1458 * Wakeup the pageout daemon if necessary and wait.
1460 if (vm_page_count_target()) {
1461 if (vm_pages_needed == 0) {
1462 vm_pages_needed = 1;
1463 wakeup(&vm_pages_needed);
1465 ++vm_pages_waiting; /* SMP race ok */
1466 tsleep(&vmstats.v_free_count, 0, "vmwait", timo);
1469 lwkt_reltoken(&vm_token);
1473 * Block until free pages are available for allocation
1475 * Called only from vm_fault so that processes page faulting can be
1482 * Wakeup the pageout daemon if necessary and wait.
1484 if (vm_page_count_target()) {
1485 lwkt_gettoken(&vm_token);
1486 if (vm_page_count_target()) {
1487 if (vm_pages_needed == 0) {
1488 vm_pages_needed = 1;
1489 wakeup(&vm_pages_needed);
1491 ++vm_pages_waiting; /* SMP race ok */
1492 tsleep(&vmstats.v_free_count, 0, "pfault", hz);
1494 lwkt_reltoken(&vm_token);
1499 * Put the specified page on the active list (if appropriate). Ensure
1500 * that act_count is at least ACT_INIT but do not otherwise mess with it.
1502 * The caller should be holding the page busied ? XXX
1503 * This routine may not block.
1506 vm_page_activate(vm_page_t m)
1510 vm_page_spin_lock(m);
1511 if (m->queue != PQ_ACTIVE) {
1512 _vm_page_queue_spin_lock(m);
1513 oqueue = _vm_page_rem_queue_spinlocked(m);
1514 /* page is left spinlocked, queue is unlocked */
1516 if (oqueue == PQ_CACHE)
1517 mycpu->gd_cnt.v_reactivated++;
1518 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1519 if (m->act_count < ACT_INIT)
1520 m->act_count = ACT_INIT;
1521 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE, 0);
1523 _vm_page_and_queue_spin_unlock(m);
1524 if (oqueue == PQ_CACHE || oqueue == PQ_FREE)
1525 pagedaemon_wakeup();
1527 if (m->act_count < ACT_INIT)
1528 m->act_count = ACT_INIT;
1529 vm_page_spin_unlock(m);
1534 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
1535 * routine is called when a page has been added to the cache or free
1538 * This routine may not block.
1540 static __inline void
1541 vm_page_free_wakeup(void)
1544 * If the pageout daemon itself needs pages, then tell it that
1545 * there are some free.
1547 if (vm_pageout_pages_needed &&
1548 vmstats.v_cache_count + vmstats.v_free_count >=
1549 vmstats.v_pageout_free_min
1551 wakeup(&vm_pageout_pages_needed);
1552 vm_pageout_pages_needed = 0;
1556 * Wakeup processes that are waiting on memory.
1558 * NOTE: vm_paging_target() is the pageout daemon's target, while
1559 * vm_page_count_target() is somewhere inbetween. We want
1560 * to wake processes up prior to the pageout daemon reaching
1561 * its target to provide some hysteresis.
1563 if (vm_pages_waiting) {
1564 if (!vm_page_count_target()) {
1566 * Plenty of pages are free, wakeup everyone.
1568 vm_pages_waiting = 0;
1569 wakeup(&vmstats.v_free_count);
1570 ++mycpu->gd_cnt.v_ppwakeups;
1571 } else if (!vm_page_count_min(0)) {
1573 * Some pages are free, wakeup someone.
1575 int wcount = vm_pages_waiting;
1578 vm_pages_waiting = wcount;
1579 wakeup_one(&vmstats.v_free_count);
1580 ++mycpu->gd_cnt.v_ppwakeups;
1586 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates
1587 * it from its VM object.
1589 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on
1590 * return (the page will have been freed).
1593 vm_page_free_toq(vm_page_t m)
1595 mycpu->gd_cnt.v_tfree++;
1596 KKASSERT((m->flags & PG_MAPPED) == 0);
1597 KKASSERT(m->flags & PG_BUSY);
1599 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
1601 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
1602 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
1604 if ((m->queue - m->pc) == PQ_FREE)
1605 panic("vm_page_free: freeing free page");
1607 panic("vm_page_free: freeing busy page");
1611 * Remove from object, spinlock the page and its queues and
1612 * remove from any queue. No queue spinlock will be held
1613 * after this section (because the page was removed from any
1617 vm_page_and_queue_spin_lock(m);
1618 _vm_page_rem_queue_spinlocked(m);
1621 * No further management of fictitious pages occurs beyond object
1622 * and queue removal.
1624 if ((m->flags & PG_FICTITIOUS) != 0) {
1625 vm_page_spin_unlock(m);
1633 if (m->wire_count != 0) {
1634 if (m->wire_count > 1) {
1636 "vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1637 m->wire_count, (long)m->pindex);
1639 panic("vm_page_free: freeing wired page");
1643 * Clear the UNMANAGED flag when freeing an unmanaged page.
1645 if (m->flags & PG_UNMANAGED) {
1646 vm_page_flag_clear(m, PG_UNMANAGED);
1649 if (m->hold_count != 0) {
1650 vm_page_flag_clear(m, PG_ZERO);
1651 _vm_page_add_queue_spinlocked(m, PQ_HOLD, 0);
1653 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0);
1657 * This sequence allows us to clear PG_BUSY while still holding
1658 * its spin lock, which reduces contention vs allocators. We
1659 * must not leave the queue locked or _vm_page_wakeup() may
1662 _vm_page_queue_spin_unlock(m);
1663 if (_vm_page_wakeup(m)) {
1664 vm_page_spin_unlock(m);
1667 vm_page_spin_unlock(m);
1669 vm_page_free_wakeup();
1673 * vm_page_free_fromq_fast()
1675 * Remove a non-zero page from one of the free queues; the page is removed for
1676 * zeroing, so do not issue a wakeup.
1679 vm_page_free_fromq_fast(void)
1685 for (i = 0; i < PQ_L2_SIZE; ++i) {
1686 m = vm_page_list_find(PQ_FREE, qi, FALSE);
1687 /* page is returned spinlocked and removed from its queue */
1689 if (vm_page_busy_try(m, TRUE)) {
1691 * We were unable to busy the page, deactivate
1694 _vm_page_deactivate_locked(m, 0);
1695 vm_page_spin_unlock(m);
1696 } else if ((m->flags & PG_ZERO) == 0) {
1698 * The page is not PG_ZERO'd so return it.
1700 vm_page_spin_unlock(m);
1704 * The page is PG_ZERO, requeue it and loop
1706 _vm_page_add_queue_spinlocked(m,
1709 vm_page_queue_spin_unlock(m);
1710 if (_vm_page_wakeup(m)) {
1711 vm_page_spin_unlock(m);
1714 vm_page_spin_unlock(m);
1719 qi = (qi + PQ_PRIME2) & PQ_L2_MASK;
1725 * vm_page_unmanage()
1727 * Prevent PV management from being done on the page. The page is
1728 * removed from the paging queues as if it were wired, and as a
1729 * consequence of no longer being managed the pageout daemon will not
1730 * touch it (since there is no way to locate the pte mappings for the
1731 * page). madvise() calls that mess with the pmap will also no longer
1732 * operate on the page.
1734 * Beyond that the page is still reasonably 'normal'. Freeing the page
1735 * will clear the flag.
1737 * This routine is used by OBJT_PHYS objects - objects using unswappable
1738 * physical memory as backing store rather then swap-backed memory and
1739 * will eventually be extended to support 4MB unmanaged physical
1742 * Caller must be holding the page busy.
1745 vm_page_unmanage(vm_page_t m)
1747 KKASSERT(m->flags & PG_BUSY);
1748 if ((m->flags & PG_UNMANAGED) == 0) {
1749 if (m->wire_count == 0)
1752 vm_page_flag_set(m, PG_UNMANAGED);
1756 * Mark this page as wired down by yet another map, removing it from
1757 * paging queues as necessary.
1759 * Caller must be holding the page busy.
1762 vm_page_wire(vm_page_t m)
1765 * Only bump the wire statistics if the page is not already wired,
1766 * and only unqueue the page if it is on some queue (if it is unmanaged
1767 * it is already off the queues). Don't do anything with fictitious
1768 * pages because they are always wired.
1770 KKASSERT(m->flags & PG_BUSY);
1771 if ((m->flags & PG_FICTITIOUS) == 0) {
1772 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) {
1773 if ((m->flags & PG_UNMANAGED) == 0)
1775 atomic_add_int(&vmstats.v_wire_count, 1);
1777 KASSERT(m->wire_count != 0,
1778 ("vm_page_wire: wire_count overflow m=%p", m));
1783 * Release one wiring of this page, potentially enabling it to be paged again.
1785 * Many pages placed on the inactive queue should actually go
1786 * into the cache, but it is difficult to figure out which. What
1787 * we do instead, if the inactive target is well met, is to put
1788 * clean pages at the head of the inactive queue instead of the tail.
1789 * This will cause them to be moved to the cache more quickly and
1790 * if not actively re-referenced, freed more quickly. If we just
1791 * stick these pages at the end of the inactive queue, heavy filesystem
1792 * meta-data accesses can cause an unnecessary paging load on memory bound
1793 * processes. This optimization causes one-time-use metadata to be
1794 * reused more quickly.
1796 * BUT, if we are in a low-memory situation we have no choice but to
1797 * put clean pages on the cache queue.
1799 * A number of routines use vm_page_unwire() to guarantee that the page
1800 * will go into either the inactive or active queues, and will NEVER
1801 * be placed in the cache - for example, just after dirtying a page.
1802 * dirty pages in the cache are not allowed.
1804 * The page queues must be locked.
1805 * This routine may not block.
1808 vm_page_unwire(vm_page_t m, int activate)
1810 KKASSERT(m->flags & PG_BUSY);
1811 if (m->flags & PG_FICTITIOUS) {
1813 } else if (m->wire_count <= 0) {
1814 panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
1816 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) {
1817 atomic_add_int(&vmstats.v_wire_count, -1);
1818 if (m->flags & PG_UNMANAGED) {
1820 } else if (activate) {
1821 vm_page_spin_lock(m);
1822 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE, 0);
1823 _vm_page_and_queue_spin_unlock(m);
1825 vm_page_spin_lock(m);
1826 vm_page_flag_clear(m, PG_WINATCFLS);
1827 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE,
1829 ++vm_swapcache_inactive_heuristic;
1830 _vm_page_and_queue_spin_unlock(m);
1837 * Move the specified page to the inactive queue. If the page has
1838 * any associated swap, the swap is deallocated.
1840 * Normally athead is 0 resulting in LRU operation. athead is set
1841 * to 1 if we want this page to be 'as if it were placed in the cache',
1842 * except without unmapping it from the process address space.
1844 * vm_page's spinlock must be held on entry and will remain held on return.
1845 * This routine may not block.
1848 _vm_page_deactivate_locked(vm_page_t m, int athead)
1853 * Ignore if already inactive.
1855 if (m->queue == PQ_INACTIVE)
1857 _vm_page_queue_spin_lock(m);
1858 oqueue = _vm_page_rem_queue_spinlocked(m);
1860 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1861 if (oqueue == PQ_CACHE)
1862 mycpu->gd_cnt.v_reactivated++;
1863 vm_page_flag_clear(m, PG_WINATCFLS);
1864 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE, athead);
1866 ++vm_swapcache_inactive_heuristic;
1868 _vm_page_queue_spin_unlock(m);
1869 /* leaves vm_page spinlocked */
1873 * Attempt to deactivate a page.
1878 vm_page_deactivate(vm_page_t m)
1880 vm_page_spin_lock(m);
1881 _vm_page_deactivate_locked(m, 0);
1882 vm_page_spin_unlock(m);
1886 vm_page_deactivate_locked(vm_page_t m)
1888 _vm_page_deactivate_locked(m, 0);
1892 * Attempt to move a page to PQ_CACHE.
1894 * Returns 0 on failure, 1 on success
1896 * The page should NOT be busied by the caller. This function will validate
1897 * whether the page can be safely moved to the cache.
1900 vm_page_try_to_cache(vm_page_t m)
1902 vm_page_spin_lock(m);
1903 if (vm_page_busy_try(m, TRUE)) {
1904 vm_page_spin_unlock(m);
1907 if (m->dirty || m->hold_count || m->wire_count ||
1908 (m->flags & PG_UNMANAGED)) {
1909 if (_vm_page_wakeup(m)) {
1910 vm_page_spin_unlock(m);
1913 vm_page_spin_unlock(m);
1917 vm_page_spin_unlock(m);
1920 * Page busied by us and no longer spinlocked. Dirty pages cannot
1921 * be moved to the cache.
1923 vm_page_test_dirty(m);
1933 * Attempt to free the page. If we cannot free it, we do nothing.
1934 * 1 is returned on success, 0 on failure.
1939 vm_page_try_to_free(vm_page_t m)
1941 vm_page_spin_lock(m);
1942 if (vm_page_busy_try(m, TRUE)) {
1943 vm_page_spin_unlock(m);
1946 if (m->dirty || m->hold_count || m->wire_count ||
1947 (m->flags & PG_UNMANAGED)) {
1948 if (_vm_page_wakeup(m)) {
1949 vm_page_spin_unlock(m);
1952 vm_page_spin_unlock(m);
1956 vm_page_spin_unlock(m);
1959 * Page busied by us and no longer spinlocked. Dirty pages will
1960 * not be freed by this function. We have to re-test the
1961 * dirty bit after cleaning out the pmaps.
1963 vm_page_test_dirty(m);
1968 vm_page_protect(m, VM_PROT_NONE);
1980 * Put the specified page onto the page cache queue (if appropriate).
1982 * The page must be busy, and this routine will release the busy and
1983 * possibly even free the page.
1986 vm_page_cache(vm_page_t m)
1988 if ((m->flags & PG_UNMANAGED) || m->busy ||
1989 m->wire_count || m->hold_count) {
1990 kprintf("vm_page_cache: attempting to cache busy/held page\n");
1996 * Already in the cache (and thus not mapped)
1998 if ((m->queue - m->pc) == PQ_CACHE) {
1999 KKASSERT((m->flags & PG_MAPPED) == 0);
2005 * Caller is required to test m->dirty, but note that the act of
2006 * removing the page from its maps can cause it to become dirty
2007 * on an SMP system due to another cpu running in usermode.
2010 panic("vm_page_cache: caching a dirty page, pindex: %ld",
2015 * Remove all pmaps and indicate that the page is not
2016 * writeable or mapped. Our vm_page_protect() call may
2017 * have blocked (especially w/ VM_PROT_NONE), so recheck
2020 vm_page_protect(m, VM_PROT_NONE);
2021 if ((m->flags & (PG_UNMANAGED|PG_MAPPED)) || m->busy ||
2022 m->wire_count || m->hold_count) {
2024 } else if (m->dirty) {
2025 vm_page_deactivate(m);
2028 _vm_page_and_queue_spin_lock(m);
2029 _vm_page_rem_queue_spinlocked(m);
2030 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0);
2031 _vm_page_queue_spin_unlock(m);
2032 if (_vm_page_wakeup(m)) {
2033 vm_page_spin_unlock(m);
2036 vm_page_spin_unlock(m);
2038 vm_page_free_wakeup();
2043 * vm_page_dontneed()
2045 * Cache, deactivate, or do nothing as appropriate. This routine
2046 * is typically used by madvise() MADV_DONTNEED.
2048 * Generally speaking we want to move the page into the cache so
2049 * it gets reused quickly. However, this can result in a silly syndrome
2050 * due to the page recycling too quickly. Small objects will not be
2051 * fully cached. On the otherhand, if we move the page to the inactive
2052 * queue we wind up with a problem whereby very large objects
2053 * unnecessarily blow away our inactive and cache queues.
2055 * The solution is to move the pages based on a fixed weighting. We
2056 * either leave them alone, deactivate them, or move them to the cache,
2057 * where moving them to the cache has the highest weighting.
2058 * By forcing some pages into other queues we eventually force the
2059 * system to balance the queues, potentially recovering other unrelated
2060 * space from active. The idea is to not force this to happen too
2063 * The page must be busied.
2066 vm_page_dontneed(vm_page_t m)
2068 static int dnweight;
2075 * occassionally leave the page alone
2077 if ((dnw & 0x01F0) == 0 ||
2078 m->queue == PQ_INACTIVE ||
2079 m->queue - m->pc == PQ_CACHE
2081 if (m->act_count >= ACT_INIT)
2087 * If vm_page_dontneed() is inactivating a page, it must clear
2088 * the referenced flag; otherwise the pagedaemon will see references
2089 * on the page in the inactive queue and reactivate it. Until the
2090 * page can move to the cache queue, madvise's job is not done.
2092 vm_page_flag_clear(m, PG_REFERENCED);
2093 pmap_clear_reference(m);
2096 vm_page_test_dirty(m);
2098 if (m->dirty || (dnw & 0x0070) == 0) {
2100 * Deactivate the page 3 times out of 32.
2105 * Cache the page 28 times out of every 32. Note that
2106 * the page is deactivated instead of cached, but placed
2107 * at the head of the queue instead of the tail.
2111 vm_page_spin_lock(m);
2112 _vm_page_deactivate_locked(m, head);
2113 vm_page_spin_unlock(m);
2117 * These routines manipulate the 'soft busy' count for a page. A soft busy
2118 * is almost like PG_BUSY except that it allows certain compatible operations
2119 * to occur on the page while it is busy. For example, a page undergoing a
2120 * write can still be mapped read-only.
2122 * Because vm_pages can overlap buffers m->busy can be > 1. m->busy is only
2123 * adjusted while the vm_page is PG_BUSY so the flash will occur when the
2124 * busy bit is cleared.
2127 vm_page_io_start(vm_page_t m)
2129 KASSERT(m->flags & PG_BUSY, ("vm_page_io_start: page not busy!!!"));
2130 atomic_add_char(&m->busy, 1);
2131 vm_page_flag_set(m, PG_SBUSY);
2135 vm_page_io_finish(vm_page_t m)
2137 KASSERT(m->flags & PG_BUSY, ("vm_page_io_finish: page not busy!!!"));
2138 atomic_subtract_char(&m->busy, 1);
2140 vm_page_flag_clear(m, PG_SBUSY);
2144 * Grab a page, blocking if it is busy and allocating a page if necessary.
2145 * A busy page is returned or NULL.
2147 * The page is not removed from its queues. XXX?
2149 * If VM_ALLOC_RETRY is specified VM_ALLOC_NORMAL must also be specified.
2150 * If VM_ALLOC_RETRY is not specified
2152 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
2153 * always returned if we had blocked.
2154 * This routine will never return NULL if VM_ALLOC_RETRY is set.
2155 * This routine may not be called from an interrupt.
2156 * The returned page may not be entirely valid.
2158 * This routine may be called from mainline code without spl protection and
2159 * be guarenteed a busied page associated with the object at the specified
2165 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2170 KKASSERT(allocflags &
2171 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
2172 vm_object_hold(object);
2174 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error);
2176 vm_page_sleep_busy(m, TRUE, "pgrbwt");
2177 if ((allocflags & VM_ALLOC_RETRY) == 0) {
2181 } else if (m == NULL) {
2182 m = vm_page_alloc(object, pindex,
2183 allocflags & ~VM_ALLOC_RETRY);
2187 if ((allocflags & VM_ALLOC_RETRY) == 0)
2194 vm_object_drop(object);
2199 * Mapping function for valid bits or for dirty bits in
2200 * a page. May not block.
2202 * Inputs are required to range within a page.
2208 vm_page_bits(int base, int size)
2214 base + size <= PAGE_SIZE,
2215 ("vm_page_bits: illegal base/size %d/%d", base, size)
2218 if (size == 0) /* handle degenerate case */
2221 first_bit = base >> DEV_BSHIFT;
2222 last_bit = (base + size - 1) >> DEV_BSHIFT;
2224 return ((2 << last_bit) - (1 << first_bit));
2228 * Sets portions of a page valid and clean. The arguments are expected
2229 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2230 * of any partial chunks touched by the range. The invalid portion of
2231 * such chunks will be zero'd.
2233 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically
2234 * align base to DEV_BSIZE so as not to mark clean a partially
2235 * truncated device block. Otherwise the dirty page status might be
2238 * This routine may not block.
2240 * (base + size) must be less then or equal to PAGE_SIZE.
2243 _vm_page_zero_valid(vm_page_t m, int base, int size)
2248 if (size == 0) /* handle degenerate case */
2252 * If the base is not DEV_BSIZE aligned and the valid
2253 * bit is clear, we have to zero out a portion of the
2257 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2258 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
2260 pmap_zero_page_area(
2268 * If the ending offset is not DEV_BSIZE aligned and the
2269 * valid bit is clear, we have to zero out a portion of
2273 endoff = base + size;
2275 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2276 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
2278 pmap_zero_page_area(
2281 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
2287 * Set valid, clear dirty bits. If validating the entire
2288 * page we can safely clear the pmap modify bit. We also
2289 * use this opportunity to clear the PG_NOSYNC flag. If a process
2290 * takes a write fault on a MAP_NOSYNC memory area the flag will
2293 * We set valid bits inclusive of any overlap, but we can only
2294 * clear dirty bits for DEV_BSIZE chunks that are fully within
2297 * Page must be busied?
2298 * No other requirements.
2301 vm_page_set_valid(vm_page_t m, int base, int size)
2303 _vm_page_zero_valid(m, base, size);
2304 m->valid |= vm_page_bits(base, size);
2309 * Set valid bits and clear dirty bits.
2311 * NOTE: This function does not clear the pmap modified bit.
2312 * Also note that e.g. NFS may use a byte-granular base
2315 * WARNING: Page must be busied? But vfs_clean_one_page() will call
2316 * this without necessarily busying the page (via bdwrite()).
2317 * So for now vm_token must also be held.
2319 * No other requirements.
2322 vm_page_set_validclean(vm_page_t m, int base, int size)
2326 _vm_page_zero_valid(m, base, size);
2327 pagebits = vm_page_bits(base, size);
2328 m->valid |= pagebits;
2329 m->dirty &= ~pagebits;
2330 if (base == 0 && size == PAGE_SIZE) {
2331 /*pmap_clear_modify(m);*/
2332 vm_page_flag_clear(m, PG_NOSYNC);
2337 * Set valid & dirty. Used by buwrite()
2339 * WARNING: Page must be busied? But vfs_dirty_one_page() will
2340 * call this function in buwrite() so for now vm_token must
2343 * No other requirements.
2346 vm_page_set_validdirty(vm_page_t m, int base, int size)
2350 pagebits = vm_page_bits(base, size);
2351 m->valid |= pagebits;
2352 m->dirty |= pagebits;
2354 vm_object_set_writeable_dirty(m->object);
2360 * NOTE: This function does not clear the pmap modified bit.
2361 * Also note that e.g. NFS may use a byte-granular base
2364 * Page must be busied?
2365 * No other requirements.
2368 vm_page_clear_dirty(vm_page_t m, int base, int size)
2370 m->dirty &= ~vm_page_bits(base, size);
2371 if (base == 0 && size == PAGE_SIZE) {
2372 /*pmap_clear_modify(m);*/
2373 vm_page_flag_clear(m, PG_NOSYNC);
2378 * Make the page all-dirty.
2380 * Also make sure the related object and vnode reflect the fact that the
2381 * object may now contain a dirty page.
2383 * Page must be busied?
2384 * No other requirements.
2387 vm_page_dirty(vm_page_t m)
2390 int pqtype = m->queue - m->pc;
2392 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE,
2393 ("vm_page_dirty: page in free/cache queue!"));
2394 if (m->dirty != VM_PAGE_BITS_ALL) {
2395 m->dirty = VM_PAGE_BITS_ALL;
2397 vm_object_set_writeable_dirty(m->object);
2402 * Invalidates DEV_BSIZE'd chunks within a page. Both the
2403 * valid and dirty bits for the effected areas are cleared.
2405 * Page must be busied?
2407 * No other requirements.
2410 vm_page_set_invalid(vm_page_t m, int base, int size)
2414 bits = vm_page_bits(base, size);
2417 m->object->generation++;
2421 * The kernel assumes that the invalid portions of a page contain
2422 * garbage, but such pages can be mapped into memory by user code.
2423 * When this occurs, we must zero out the non-valid portions of the
2424 * page so user code sees what it expects.
2426 * Pages are most often semi-valid when the end of a file is mapped
2427 * into memory and the file's size is not page aligned.
2429 * Page must be busied?
2430 * No other requirements.
2433 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
2439 * Scan the valid bits looking for invalid sections that
2440 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
2441 * valid bit may be set ) have already been zerod by
2442 * vm_page_set_validclean().
2444 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
2445 if (i == (PAGE_SIZE / DEV_BSIZE) ||
2446 (m->valid & (1 << i))
2449 pmap_zero_page_area(
2452 (i - b) << DEV_BSHIFT
2460 * setvalid is TRUE when we can safely set the zero'd areas
2461 * as being valid. We can do this if there are no cache consistency
2462 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
2465 m->valid = VM_PAGE_BITS_ALL;
2469 * Is a (partial) page valid? Note that the case where size == 0
2470 * will return FALSE in the degenerate case where the page is entirely
2471 * invalid, and TRUE otherwise.
2474 * No other requirements.
2477 vm_page_is_valid(vm_page_t m, int base, int size)
2479 int bits = vm_page_bits(base, size);
2481 if (m->valid && ((m->valid & bits) == bits))
2488 * update dirty bits from pmap/mmu. May not block.
2490 * Caller must hold the page busy
2493 vm_page_test_dirty(vm_page_t m)
2495 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
2501 * Register an action, associating it with its vm_page
2504 vm_page_register_action(vm_page_action_t action, vm_page_event_t event)
2506 struct vm_page_action_list *list;
2509 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
2510 list = &action_list[hv];
2512 lwkt_gettoken(&vm_token);
2513 vm_page_flag_set(action->m, PG_ACTIONLIST);
2514 action->event = event;
2515 LIST_INSERT_HEAD(list, action, entry);
2516 lwkt_reltoken(&vm_token);
2520 * Unregister an action, disassociating it from its related vm_page
2523 vm_page_unregister_action(vm_page_action_t action)
2525 struct vm_page_action_list *list;
2528 lwkt_gettoken(&vm_token);
2529 if (action->event != VMEVENT_NONE) {
2530 action->event = VMEVENT_NONE;
2531 LIST_REMOVE(action, entry);
2533 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
2534 list = &action_list[hv];
2535 if (LIST_EMPTY(list))
2536 vm_page_flag_clear(action->m, PG_ACTIONLIST);
2538 lwkt_reltoken(&vm_token);
2542 * Issue an event on a VM page. Corresponding action structures are
2543 * removed from the page's list and called.
2545 * If the vm_page has no more pending action events we clear its
2546 * PG_ACTIONLIST flag.
2549 vm_page_event_internal(vm_page_t m, vm_page_event_t event)
2551 struct vm_page_action_list *list;
2552 struct vm_page_action *scan;
2553 struct vm_page_action *next;
2557 hv = (int)((intptr_t)m >> 8) & VMACTION_HMASK;
2558 list = &action_list[hv];
2561 lwkt_gettoken(&vm_token);
2562 LIST_FOREACH_MUTABLE(scan, list, entry, next) {
2564 if (scan->event == event) {
2565 scan->event = VMEVENT_NONE;
2566 LIST_REMOVE(scan, entry);
2567 scan->func(m, scan);
2575 vm_page_flag_clear(m, PG_ACTIONLIST);
2576 lwkt_reltoken(&vm_token);
2579 #include "opt_ddb.h"
2581 #include <sys/kernel.h>
2583 #include <ddb/ddb.h>
2585 DB_SHOW_COMMAND(page, vm_page_print_page_info)
2587 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count);
2588 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count);
2589 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count);
2590 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count);
2591 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count);
2592 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved);
2593 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min);
2594 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target);
2595 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min);
2596 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target);
2599 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
2602 db_printf("PQ_FREE:");
2603 for(i=0;i<PQ_L2_SIZE;i++) {
2604 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
2608 db_printf("PQ_CACHE:");
2609 for(i=0;i<PQ_L2_SIZE;i++) {
2610 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
2614 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
2615 vm_page_queues[PQ_ACTIVE].lcnt,
2616 vm_page_queues[PQ_INACTIVE].lcnt);