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.
759 * This routine returns TRUE if the page was inserted into the object
760 * successfully, and FALSE if the page already exists in the object.
763 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
765 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
766 if (m->object != NULL)
767 panic("vm_page_insert: already inserted");
769 object->generation++;
772 * Record the object/offset pair in this page and add the
773 * pv_list_count of the page to the object.
775 * The vm_page spin lock is required for interactions with the pmap.
777 vm_page_spin_lock(m);
780 if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) {
783 vm_page_spin_unlock(m);
786 object->resident_page_count++;
787 /* atomic_add_int(&object->agg_pv_list_count, m->md.pv_list_count); */
788 vm_page_spin_unlock(m);
791 * Since we are inserting a new and possibly dirty page,
792 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
794 if ((m->valid & m->dirty) || (m->flags & PG_WRITEABLE))
795 vm_object_set_writeable_dirty(object);
798 * Checks for a swap assignment and sets PG_SWAPPED if appropriate.
800 swap_pager_page_inserted(m);
805 * Removes the given vm_page_t from the (object,index) table
807 * The underlying pmap entry (if any) is NOT removed here.
808 * This routine may not block.
810 * The page must be BUSY and will remain BUSY on return.
811 * No other requirements.
813 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave
817 vm_page_remove(vm_page_t m)
821 if (m->object == NULL) {
825 if ((m->flags & PG_BUSY) == 0)
826 panic("vm_page_remove: page not busy");
830 vm_object_hold(object);
833 * Remove the page from the object and update the object.
835 * The vm_page spin lock is required for interactions with the pmap.
837 vm_page_spin_lock(m);
838 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m);
839 object->resident_page_count--;
840 /* atomic_add_int(&object->agg_pv_list_count, -m->md.pv_list_count); */
842 vm_page_spin_unlock(m);
844 object->generation++;
846 vm_object_drop(object);
850 * Locate and return the page at (object, pindex), or NULL if the
851 * page could not be found.
853 * The caller must hold the vm_object token.
856 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
861 * Search the hash table for this object/offset pair
863 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
864 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
865 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex));
870 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object,
872 int also_m_busy, const char *msg
878 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
879 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
881 KKASSERT(m->object == object && m->pindex == pindex);
884 if (flags & PG_BUSY) {
885 tsleep_interlock(m, 0);
886 if (atomic_cmpset_int(&m->flags, flags,
887 flags | PG_WANTED | PG_REFERENCED)) {
888 tsleep(m, PINTERLOCKED, msg, 0);
889 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
892 } else if (also_m_busy && (flags & PG_SBUSY)) {
893 tsleep_interlock(m, 0);
894 if (atomic_cmpset_int(&m->flags, flags,
895 flags | PG_WANTED | PG_REFERENCED)) {
896 tsleep(m, PINTERLOCKED, msg, 0);
897 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
900 } else if (atomic_cmpset_int(&m->flags, flags,
904 m->busy_line = lineno;
913 * Attempt to lookup and busy a page.
915 * Returns NULL if the page could not be found
917 * Returns a vm_page and error == TRUE if the page exists but could not
920 * Returns a vm_page and error == FALSE on success.
923 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object,
925 int also_m_busy, int *errorp
931 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
932 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
935 KKASSERT(m->object == object && m->pindex == pindex);
938 if (flags & PG_BUSY) {
942 if (also_m_busy && (flags & PG_SBUSY)) {
946 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
949 m->busy_line = lineno;
958 * Caller must hold the related vm_object
961 vm_page_next(vm_page_t m)
965 next = vm_page_rb_tree_RB_NEXT(m);
966 if (next && next->pindex != m->pindex + 1)
974 * Move the given vm_page from its current object to the specified
975 * target object/offset. The page must be busy and will remain so
978 * new_object must be held.
979 * This routine might block. XXX ?
981 * NOTE: Swap associated with the page must be invalidated by the move. We
982 * have to do this for several reasons: (1) we aren't freeing the
983 * page, (2) we are dirtying the page, (3) the VM system is probably
984 * moving the page from object A to B, and will then later move
985 * the backing store from A to B and we can't have a conflict.
987 * NOTE: We *always* dirty the page. It is necessary both for the
988 * fact that we moved it, and because we may be invalidating
989 * swap. If the page is on the cache, we have to deactivate it
990 * or vm_page_dirty() will panic. Dirty pages are not allowed
994 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
996 KKASSERT(m->flags & PG_BUSY);
997 ASSERT_LWKT_TOKEN_HELD(vm_object_token(new_object));
999 ASSERT_LWKT_TOKEN_HELD(vm_object_token(m->object));
1002 if (vm_page_insert(m, new_object, new_pindex) == FALSE) {
1003 panic("vm_page_rename: target exists (%p,%ld)",
1004 new_object, new_pindex);
1006 if (m->queue - m->pc == PQ_CACHE)
1007 vm_page_deactivate(m);
1012 * vm_page_unqueue() without any wakeup. This routine is used when a page
1013 * is being moved between queues or otherwise is to remain BUSYied by the
1016 * This routine may not block.
1019 vm_page_unqueue_nowakeup(vm_page_t m)
1021 vm_page_and_queue_spin_lock(m);
1022 (void)_vm_page_rem_queue_spinlocked(m);
1023 vm_page_spin_unlock(m);
1027 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
1030 * This routine may not block.
1033 vm_page_unqueue(vm_page_t m)
1037 vm_page_and_queue_spin_lock(m);
1038 queue = _vm_page_rem_queue_spinlocked(m);
1039 if (queue == PQ_FREE || queue == PQ_CACHE) {
1040 vm_page_spin_unlock(m);
1041 pagedaemon_wakeup();
1043 vm_page_spin_unlock(m);
1048 * vm_page_list_find()
1050 * Find a page on the specified queue with color optimization.
1052 * The page coloring optimization attempts to locate a page that does
1053 * not overload other nearby pages in the object in the cpu's L1 or L2
1054 * caches. We need this optimization because cpu caches tend to be
1055 * physical caches, while object spaces tend to be virtual.
1057 * On MP systems each PQ_FREE and PQ_CACHE color queue has its own spinlock
1058 * and the algorithm is adjusted to localize allocations on a per-core basis.
1059 * This is done by 'twisting' the colors.
1061 * The page is returned spinlocked and removed from its queue (it will
1062 * be on PQ_NONE), or NULL. The page is not PG_BUSY'd. The caller
1063 * is responsible for dealing with the busy-page case (usually by
1064 * deactivating the page and looping).
1066 * NOTE: This routine is carefully inlined. A non-inlined version
1067 * is available for outside callers but the only critical path is
1068 * from within this source file.
1070 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE
1071 * represent stable storage, allowing us to order our locks vm_page
1072 * first, then queue.
1076 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
1082 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist);
1084 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl);
1086 m = _vm_page_list_find2(basequeue, index);
1089 vm_page_and_queue_spin_lock(m);
1090 if (m->queue == basequeue + index) {
1091 _vm_page_rem_queue_spinlocked(m);
1092 /* vm_page_t spin held, no queue spin */
1095 vm_page_and_queue_spin_unlock(m);
1101 _vm_page_list_find2(int basequeue, int index)
1105 struct vpgqueues *pq;
1107 pq = &vm_page_queues[basequeue];
1110 * Note that for the first loop, index+i and index-i wind up at the
1111 * same place. Even though this is not totally optimal, we've already
1112 * blown it by missing the cache case so we do not care.
1114 for (i = PQ_L2_SIZE / 2; i > 0; --i) {
1116 m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl);
1118 _vm_page_and_queue_spin_lock(m);
1120 basequeue + ((index + i) & PQ_L2_MASK)) {
1121 _vm_page_rem_queue_spinlocked(m);
1124 _vm_page_and_queue_spin_unlock(m);
1127 m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl);
1129 _vm_page_and_queue_spin_lock(m);
1131 basequeue + ((index - i) & PQ_L2_MASK)) {
1132 _vm_page_rem_queue_spinlocked(m);
1135 _vm_page_and_queue_spin_unlock(m);
1145 * Returns a vm_page candidate for allocation. The page is not busied so
1146 * it can move around. The caller must busy the page (and typically
1147 * deactivate it if it cannot be busied!)
1149 * Returns a spinlocked vm_page that has been removed from its queue.
1152 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
1154 return(_vm_page_list_find(basequeue, index, prefer_zero));
1158 * Find a page on the cache queue with color optimization, remove it
1159 * from the queue, and busy it. The returned page will not be spinlocked.
1161 * A candidate failure will be deactivated. Candidates can fail due to
1162 * being busied by someone else, in which case they will be deactivated.
1164 * This routine may not block.
1168 vm_page_select_cache(u_short pg_color)
1173 m = _vm_page_list_find(PQ_CACHE, pg_color & PQ_L2_MASK, FALSE);
1177 * (m) has been removed from its queue and spinlocked
1179 if (vm_page_busy_try(m, TRUE)) {
1180 _vm_page_deactivate_locked(m, 0);
1181 vm_page_spin_unlock(m);
1183 kprintf("Warning: busy page %p found in cache\n", m);
1187 * We successfully busied the page
1189 if ((m->flags & PG_UNMANAGED) == 0 &&
1190 m->hold_count == 0 &&
1191 m->wire_count == 0) {
1192 vm_page_spin_unlock(m);
1193 pagedaemon_wakeup();
1196 _vm_page_deactivate_locked(m, 0);
1197 if (_vm_page_wakeup(m)) {
1198 vm_page_spin_unlock(m);
1201 vm_page_spin_unlock(m);
1209 * Find a free or zero page, with specified preference. We attempt to
1210 * inline the nominal case and fall back to _vm_page_select_free()
1211 * otherwise. A busied page is removed from the queue and returned.
1213 * This routine may not block.
1215 static __inline vm_page_t
1216 vm_page_select_free(u_short pg_color, boolean_t prefer_zero)
1221 m = _vm_page_list_find(PQ_FREE, pg_color & PQ_L2_MASK,
1225 if (vm_page_busy_try(m, TRUE)) {
1226 _vm_page_deactivate_locked(m, 0);
1227 vm_page_spin_unlock(m);
1229 kprintf("Warning: busy page %p found in cache\n", m);
1232 KKASSERT((m->flags & PG_UNMANAGED) == 0);
1233 KKASSERT(m->hold_count == 0);
1234 KKASSERT(m->wire_count == 0);
1235 vm_page_spin_unlock(m);
1236 pagedaemon_wakeup();
1238 /* return busied and removed page */
1248 * Allocate and return a memory cell associated with this VM object/offset
1249 * pair. If object is NULL an unassociated page will be allocated.
1251 * The returned page will be busied and removed from its queues. This
1252 * routine can block and may return NULL if a race occurs and the page
1253 * is found to already exist at the specified (object, pindex).
1255 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain
1256 * VM_ALLOC_QUICK like normal but cannot use cache
1257 * VM_ALLOC_SYSTEM greater free drain
1258 * VM_ALLOC_INTERRUPT allow free list to be completely drained
1259 * VM_ALLOC_ZERO advisory request for pre-zero'd page only
1260 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only
1261 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision
1262 * (see vm_page_grab())
1263 * The object must be held if not NULL
1264 * This routine may not block
1266 * Additional special handling is required when called from an interrupt
1267 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache
1271 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
1278 * Cpu twist - cpu localization algorithm
1281 pg_color = mycpu->gd_cpuid + (pindex & ~ncpus_fit_mask) +
1282 (object->pg_color & ~ncpus_fit_mask);
1284 pg_color = mycpu->gd_cpuid + (pindex & ~ncpus_fit_mask);
1288 * Normal page coloring algorithm
1291 pg_color = object->pg_color + pindex;
1297 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK|
1298 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
1301 * Certain system threads (pageout daemon, buf_daemon's) are
1302 * allowed to eat deeper into the free page list.
1304 if (curthread->td_flags & TDF_SYSTHREAD)
1305 page_req |= VM_ALLOC_SYSTEM;
1308 if (vmstats.v_free_count > vmstats.v_free_reserved ||
1309 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) ||
1310 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 &&
1311 vmstats.v_free_count > vmstats.v_interrupt_free_min)
1314 * The free queue has sufficient free pages to take one out.
1316 if (page_req & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO))
1317 m = vm_page_select_free(pg_color, TRUE);
1319 m = vm_page_select_free(pg_color, FALSE);
1320 } else if (page_req & VM_ALLOC_NORMAL) {
1322 * Allocatable from the cache (non-interrupt only). On
1323 * success, we must free the page and try again, thus
1324 * ensuring that vmstats.v_*_free_min counters are replenished.
1327 if (curthread->td_preempted) {
1328 kprintf("vm_page_alloc(): warning, attempt to allocate"
1329 " cache page from preempting interrupt\n");
1332 m = vm_page_select_cache(pg_color);
1335 m = vm_page_select_cache(pg_color);
1338 * On success move the page into the free queue and loop.
1341 KASSERT(m->dirty == 0,
1342 ("Found dirty cache page %p", m));
1343 vm_page_protect(m, VM_PROT_NONE);
1349 * On failure return NULL
1351 #if defined(DIAGNOSTIC)
1352 if (vmstats.v_cache_count > 0)
1353 kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count);
1355 vm_pageout_deficit++;
1356 pagedaemon_wakeup();
1360 * No pages available, wakeup the pageout daemon and give up.
1362 vm_pageout_deficit++;
1363 pagedaemon_wakeup();
1368 * v_free_count can race so loop if we don't find the expected
1375 * Good page found. The page has already been busied for us and
1376 * removed from its queues.
1378 KASSERT(m->dirty == 0,
1379 ("vm_page_alloc: free/cache page %p was dirty", m));
1380 KKASSERT(m->queue == PQ_NONE);
1383 * Initialize the structure, inheriting some flags but clearing
1384 * all the rest. The page has already been busied for us.
1386 vm_page_flag_clear(m, ~(PG_ZERO | PG_BUSY | PG_SBUSY));
1387 KKASSERT(m->wire_count == 0);
1388 KKASSERT(m->busy == 0);
1393 * Caller must be holding the object lock (asserted by
1394 * vm_page_insert()).
1396 * NOTE: Inserting a page here does not insert it into any pmaps
1397 * (which could cause us to block allocating memory).
1399 * NOTE: If no object an unassociated page is allocated, m->pindex
1400 * can be used by the caller for any purpose.
1403 if (vm_page_insert(m, object, pindex) == FALSE) {
1404 kprintf("PAGE RACE (%p:%d,%ld)\n",
1405 object, object->type, pindex);
1408 if ((page_req & VM_ALLOC_NULL_OK) == 0)
1416 * Don't wakeup too often - wakeup the pageout daemon when
1417 * we would be nearly out of memory.
1419 pagedaemon_wakeup();
1422 * A PG_BUSY page is returned.
1428 * Wait for sufficient free memory for nominal heavy memory use kernel
1432 vm_wait_nominal(void)
1434 while (vm_page_count_min(0))
1439 * Test if vm_wait_nominal() would block.
1442 vm_test_nominal(void)
1444 if (vm_page_count_min(0))
1450 * Block until free pages are available for allocation, called in various
1451 * places before memory allocations.
1453 * The caller may loop if vm_page_count_min() == FALSE so we cannot be
1454 * more generous then that.
1460 * never wait forever
1464 lwkt_gettoken(&vm_token);
1466 if (curthread == pagethread) {
1468 * The pageout daemon itself needs pages, this is bad.
1470 if (vm_page_count_min(0)) {
1471 vm_pageout_pages_needed = 1;
1472 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo);
1476 * Wakeup the pageout daemon if necessary and wait.
1478 if (vm_page_count_target()) {
1479 if (vm_pages_needed == 0) {
1480 vm_pages_needed = 1;
1481 wakeup(&vm_pages_needed);
1483 ++vm_pages_waiting; /* SMP race ok */
1484 tsleep(&vmstats.v_free_count, 0, "vmwait", timo);
1487 lwkt_reltoken(&vm_token);
1491 * Block until free pages are available for allocation
1493 * Called only from vm_fault so that processes page faulting can be
1500 * Wakeup the pageout daemon if necessary and wait.
1502 if (vm_page_count_target()) {
1503 lwkt_gettoken(&vm_token);
1504 if (vm_page_count_target()) {
1505 if (vm_pages_needed == 0) {
1506 vm_pages_needed = 1;
1507 wakeup(&vm_pages_needed);
1509 ++vm_pages_waiting; /* SMP race ok */
1510 tsleep(&vmstats.v_free_count, 0, "pfault", hz);
1512 lwkt_reltoken(&vm_token);
1517 * Put the specified page on the active list (if appropriate). Ensure
1518 * that act_count is at least ACT_INIT but do not otherwise mess with it.
1520 * The caller should be holding the page busied ? XXX
1521 * This routine may not block.
1524 vm_page_activate(vm_page_t m)
1528 vm_page_spin_lock(m);
1529 if (m->queue != PQ_ACTIVE) {
1530 _vm_page_queue_spin_lock(m);
1531 oqueue = _vm_page_rem_queue_spinlocked(m);
1532 /* page is left spinlocked, queue is unlocked */
1534 if (oqueue == PQ_CACHE)
1535 mycpu->gd_cnt.v_reactivated++;
1536 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1537 if (m->act_count < ACT_INIT)
1538 m->act_count = ACT_INIT;
1539 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE, 0);
1541 _vm_page_and_queue_spin_unlock(m);
1542 if (oqueue == PQ_CACHE || oqueue == PQ_FREE)
1543 pagedaemon_wakeup();
1545 if (m->act_count < ACT_INIT)
1546 m->act_count = ACT_INIT;
1547 vm_page_spin_unlock(m);
1552 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
1553 * routine is called when a page has been added to the cache or free
1556 * This routine may not block.
1558 static __inline void
1559 vm_page_free_wakeup(void)
1562 * If the pageout daemon itself needs pages, then tell it that
1563 * there are some free.
1565 if (vm_pageout_pages_needed &&
1566 vmstats.v_cache_count + vmstats.v_free_count >=
1567 vmstats.v_pageout_free_min
1569 wakeup(&vm_pageout_pages_needed);
1570 vm_pageout_pages_needed = 0;
1574 * Wakeup processes that are waiting on memory.
1576 * NOTE: vm_paging_target() is the pageout daemon's target, while
1577 * vm_page_count_target() is somewhere inbetween. We want
1578 * to wake processes up prior to the pageout daemon reaching
1579 * its target to provide some hysteresis.
1581 if (vm_pages_waiting) {
1582 if (!vm_page_count_target()) {
1584 * Plenty of pages are free, wakeup everyone.
1586 vm_pages_waiting = 0;
1587 wakeup(&vmstats.v_free_count);
1588 ++mycpu->gd_cnt.v_ppwakeups;
1589 } else if (!vm_page_count_min(0)) {
1591 * Some pages are free, wakeup someone.
1593 int wcount = vm_pages_waiting;
1596 vm_pages_waiting = wcount;
1597 wakeup_one(&vmstats.v_free_count);
1598 ++mycpu->gd_cnt.v_ppwakeups;
1604 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates
1605 * it from its VM object.
1607 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on
1608 * return (the page will have been freed).
1611 vm_page_free_toq(vm_page_t m)
1613 mycpu->gd_cnt.v_tfree++;
1614 KKASSERT((m->flags & PG_MAPPED) == 0);
1615 KKASSERT(m->flags & PG_BUSY);
1617 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
1619 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
1620 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
1622 if ((m->queue - m->pc) == PQ_FREE)
1623 panic("vm_page_free: freeing free page");
1625 panic("vm_page_free: freeing busy page");
1629 * Remove from object, spinlock the page and its queues and
1630 * remove from any queue. No queue spinlock will be held
1631 * after this section (because the page was removed from any
1635 vm_page_and_queue_spin_lock(m);
1636 _vm_page_rem_queue_spinlocked(m);
1639 * No further management of fictitious pages occurs beyond object
1640 * and queue removal.
1642 if ((m->flags & PG_FICTITIOUS) != 0) {
1643 vm_page_spin_unlock(m);
1651 if (m->wire_count != 0) {
1652 if (m->wire_count > 1) {
1654 "vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1655 m->wire_count, (long)m->pindex);
1657 panic("vm_page_free: freeing wired page");
1661 * Clear the UNMANAGED flag when freeing an unmanaged page.
1663 if (m->flags & PG_UNMANAGED) {
1664 vm_page_flag_clear(m, PG_UNMANAGED);
1667 if (m->hold_count != 0) {
1668 vm_page_flag_clear(m, PG_ZERO);
1669 _vm_page_add_queue_spinlocked(m, PQ_HOLD, 0);
1671 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0);
1675 * This sequence allows us to clear PG_BUSY while still holding
1676 * its spin lock, which reduces contention vs allocators. We
1677 * must not leave the queue locked or _vm_page_wakeup() may
1680 _vm_page_queue_spin_unlock(m);
1681 if (_vm_page_wakeup(m)) {
1682 vm_page_spin_unlock(m);
1685 vm_page_spin_unlock(m);
1687 vm_page_free_wakeup();
1691 * vm_page_free_fromq_fast()
1693 * Remove a non-zero page from one of the free queues; the page is removed for
1694 * zeroing, so do not issue a wakeup.
1697 vm_page_free_fromq_fast(void)
1703 for (i = 0; i < PQ_L2_SIZE; ++i) {
1704 m = vm_page_list_find(PQ_FREE, qi, FALSE);
1705 /* page is returned spinlocked and removed from its queue */
1707 if (vm_page_busy_try(m, TRUE)) {
1709 * We were unable to busy the page, deactivate
1712 _vm_page_deactivate_locked(m, 0);
1713 vm_page_spin_unlock(m);
1714 } else if ((m->flags & PG_ZERO) == 0) {
1716 * The page is not PG_ZERO'd so return it.
1718 vm_page_spin_unlock(m);
1722 * The page is PG_ZERO, requeue it and loop
1724 _vm_page_add_queue_spinlocked(m,
1727 vm_page_queue_spin_unlock(m);
1728 if (_vm_page_wakeup(m)) {
1729 vm_page_spin_unlock(m);
1732 vm_page_spin_unlock(m);
1737 qi = (qi + PQ_PRIME2) & PQ_L2_MASK;
1743 * vm_page_unmanage()
1745 * Prevent PV management from being done on the page. The page is
1746 * removed from the paging queues as if it were wired, and as a
1747 * consequence of no longer being managed the pageout daemon will not
1748 * touch it (since there is no way to locate the pte mappings for the
1749 * page). madvise() calls that mess with the pmap will also no longer
1750 * operate on the page.
1752 * Beyond that the page is still reasonably 'normal'. Freeing the page
1753 * will clear the flag.
1755 * This routine is used by OBJT_PHYS objects - objects using unswappable
1756 * physical memory as backing store rather then swap-backed memory and
1757 * will eventually be extended to support 4MB unmanaged physical
1760 * Caller must be holding the page busy.
1763 vm_page_unmanage(vm_page_t m)
1765 KKASSERT(m->flags & PG_BUSY);
1766 if ((m->flags & PG_UNMANAGED) == 0) {
1767 if (m->wire_count == 0)
1770 vm_page_flag_set(m, PG_UNMANAGED);
1774 * Mark this page as wired down by yet another map, removing it from
1775 * paging queues as necessary.
1777 * Caller must be holding the page busy.
1780 vm_page_wire(vm_page_t m)
1783 * Only bump the wire statistics if the page is not already wired,
1784 * and only unqueue the page if it is on some queue (if it is unmanaged
1785 * it is already off the queues). Don't do anything with fictitious
1786 * pages because they are always wired.
1788 KKASSERT(m->flags & PG_BUSY);
1789 if ((m->flags & PG_FICTITIOUS) == 0) {
1790 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) {
1791 if ((m->flags & PG_UNMANAGED) == 0)
1793 atomic_add_int(&vmstats.v_wire_count, 1);
1795 KASSERT(m->wire_count != 0,
1796 ("vm_page_wire: wire_count overflow m=%p", m));
1801 * Release one wiring of this page, potentially enabling it to be paged again.
1803 * Many pages placed on the inactive queue should actually go
1804 * into the cache, but it is difficult to figure out which. What
1805 * we do instead, if the inactive target is well met, is to put
1806 * clean pages at the head of the inactive queue instead of the tail.
1807 * This will cause them to be moved to the cache more quickly and
1808 * if not actively re-referenced, freed more quickly. If we just
1809 * stick these pages at the end of the inactive queue, heavy filesystem
1810 * meta-data accesses can cause an unnecessary paging load on memory bound
1811 * processes. This optimization causes one-time-use metadata to be
1812 * reused more quickly.
1814 * BUT, if we are in a low-memory situation we have no choice but to
1815 * put clean pages on the cache queue.
1817 * A number of routines use vm_page_unwire() to guarantee that the page
1818 * will go into either the inactive or active queues, and will NEVER
1819 * be placed in the cache - for example, just after dirtying a page.
1820 * dirty pages in the cache are not allowed.
1822 * The page queues must be locked.
1823 * This routine may not block.
1826 vm_page_unwire(vm_page_t m, int activate)
1828 KKASSERT(m->flags & PG_BUSY);
1829 if (m->flags & PG_FICTITIOUS) {
1831 } else if (m->wire_count <= 0) {
1832 panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
1834 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) {
1835 atomic_add_int(&vmstats.v_wire_count, -1);
1836 if (m->flags & PG_UNMANAGED) {
1838 } else if (activate) {
1839 vm_page_spin_lock(m);
1840 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE, 0);
1841 _vm_page_and_queue_spin_unlock(m);
1843 vm_page_spin_lock(m);
1844 vm_page_flag_clear(m, PG_WINATCFLS);
1845 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE,
1847 ++vm_swapcache_inactive_heuristic;
1848 _vm_page_and_queue_spin_unlock(m);
1855 * Move the specified page to the inactive queue. If the page has
1856 * any associated swap, the swap is deallocated.
1858 * Normally athead is 0 resulting in LRU operation. athead is set
1859 * to 1 if we want this page to be 'as if it were placed in the cache',
1860 * except without unmapping it from the process address space.
1862 * vm_page's spinlock must be held on entry and will remain held on return.
1863 * This routine may not block.
1866 _vm_page_deactivate_locked(vm_page_t m, int athead)
1871 * Ignore if already inactive.
1873 if (m->queue == PQ_INACTIVE)
1875 _vm_page_queue_spin_lock(m);
1876 oqueue = _vm_page_rem_queue_spinlocked(m);
1878 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1879 if (oqueue == PQ_CACHE)
1880 mycpu->gd_cnt.v_reactivated++;
1881 vm_page_flag_clear(m, PG_WINATCFLS);
1882 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE, athead);
1884 ++vm_swapcache_inactive_heuristic;
1886 _vm_page_queue_spin_unlock(m);
1887 /* leaves vm_page spinlocked */
1891 * Attempt to deactivate a page.
1896 vm_page_deactivate(vm_page_t m)
1898 vm_page_spin_lock(m);
1899 _vm_page_deactivate_locked(m, 0);
1900 vm_page_spin_unlock(m);
1904 vm_page_deactivate_locked(vm_page_t m)
1906 _vm_page_deactivate_locked(m, 0);
1910 * Attempt to move a page to PQ_CACHE.
1912 * Returns 0 on failure, 1 on success
1914 * The page should NOT be busied by the caller. This function will validate
1915 * whether the page can be safely moved to the cache.
1918 vm_page_try_to_cache(vm_page_t m)
1920 vm_page_spin_lock(m);
1921 if (vm_page_busy_try(m, TRUE)) {
1922 vm_page_spin_unlock(m);
1925 if (m->dirty || m->hold_count || m->wire_count ||
1926 (m->flags & PG_UNMANAGED)) {
1927 if (_vm_page_wakeup(m)) {
1928 vm_page_spin_unlock(m);
1931 vm_page_spin_unlock(m);
1935 vm_page_spin_unlock(m);
1938 * Page busied by us and no longer spinlocked. Dirty pages cannot
1939 * be moved to the cache.
1941 vm_page_test_dirty(m);
1951 * Attempt to free the page. If we cannot free it, we do nothing.
1952 * 1 is returned on success, 0 on failure.
1957 vm_page_try_to_free(vm_page_t m)
1959 vm_page_spin_lock(m);
1960 if (vm_page_busy_try(m, TRUE)) {
1961 vm_page_spin_unlock(m);
1964 if (m->dirty || m->hold_count || m->wire_count ||
1965 (m->flags & PG_UNMANAGED)) {
1966 if (_vm_page_wakeup(m)) {
1967 vm_page_spin_unlock(m);
1970 vm_page_spin_unlock(m);
1974 vm_page_spin_unlock(m);
1977 * Page busied by us and no longer spinlocked. Dirty pages will
1978 * not be freed by this function. We have to re-test the
1979 * dirty bit after cleaning out the pmaps.
1981 vm_page_test_dirty(m);
1986 vm_page_protect(m, VM_PROT_NONE);
1998 * Put the specified page onto the page cache queue (if appropriate).
2000 * The page must be busy, and this routine will release the busy and
2001 * possibly even free the page.
2004 vm_page_cache(vm_page_t m)
2006 if ((m->flags & PG_UNMANAGED) || m->busy ||
2007 m->wire_count || m->hold_count) {
2008 kprintf("vm_page_cache: attempting to cache busy/held page\n");
2014 * Already in the cache (and thus not mapped)
2016 if ((m->queue - m->pc) == PQ_CACHE) {
2017 KKASSERT((m->flags & PG_MAPPED) == 0);
2023 * Caller is required to test m->dirty, but note that the act of
2024 * removing the page from its maps can cause it to become dirty
2025 * on an SMP system due to another cpu running in usermode.
2028 panic("vm_page_cache: caching a dirty page, pindex: %ld",
2033 * Remove all pmaps and indicate that the page is not
2034 * writeable or mapped. Our vm_page_protect() call may
2035 * have blocked (especially w/ VM_PROT_NONE), so recheck
2038 vm_page_protect(m, VM_PROT_NONE);
2039 if ((m->flags & (PG_UNMANAGED|PG_MAPPED)) || m->busy ||
2040 m->wire_count || m->hold_count) {
2042 } else if (m->dirty) {
2043 vm_page_deactivate(m);
2046 _vm_page_and_queue_spin_lock(m);
2047 _vm_page_rem_queue_spinlocked(m);
2048 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0);
2049 _vm_page_queue_spin_unlock(m);
2050 if (_vm_page_wakeup(m)) {
2051 vm_page_spin_unlock(m);
2054 vm_page_spin_unlock(m);
2056 vm_page_free_wakeup();
2061 * vm_page_dontneed()
2063 * Cache, deactivate, or do nothing as appropriate. This routine
2064 * is typically used by madvise() MADV_DONTNEED.
2066 * Generally speaking we want to move the page into the cache so
2067 * it gets reused quickly. However, this can result in a silly syndrome
2068 * due to the page recycling too quickly. Small objects will not be
2069 * fully cached. On the otherhand, if we move the page to the inactive
2070 * queue we wind up with a problem whereby very large objects
2071 * unnecessarily blow away our inactive and cache queues.
2073 * The solution is to move the pages based on a fixed weighting. We
2074 * either leave them alone, deactivate them, or move them to the cache,
2075 * where moving them to the cache has the highest weighting.
2076 * By forcing some pages into other queues we eventually force the
2077 * system to balance the queues, potentially recovering other unrelated
2078 * space from active. The idea is to not force this to happen too
2081 * The page must be busied.
2084 vm_page_dontneed(vm_page_t m)
2086 static int dnweight;
2093 * occassionally leave the page alone
2095 if ((dnw & 0x01F0) == 0 ||
2096 m->queue == PQ_INACTIVE ||
2097 m->queue - m->pc == PQ_CACHE
2099 if (m->act_count >= ACT_INIT)
2105 * If vm_page_dontneed() is inactivating a page, it must clear
2106 * the referenced flag; otherwise the pagedaemon will see references
2107 * on the page in the inactive queue and reactivate it. Until the
2108 * page can move to the cache queue, madvise's job is not done.
2110 vm_page_flag_clear(m, PG_REFERENCED);
2111 pmap_clear_reference(m);
2114 vm_page_test_dirty(m);
2116 if (m->dirty || (dnw & 0x0070) == 0) {
2118 * Deactivate the page 3 times out of 32.
2123 * Cache the page 28 times out of every 32. Note that
2124 * the page is deactivated instead of cached, but placed
2125 * at the head of the queue instead of the tail.
2129 vm_page_spin_lock(m);
2130 _vm_page_deactivate_locked(m, head);
2131 vm_page_spin_unlock(m);
2135 * These routines manipulate the 'soft busy' count for a page. A soft busy
2136 * is almost like PG_BUSY except that it allows certain compatible operations
2137 * to occur on the page while it is busy. For example, a page undergoing a
2138 * write can still be mapped read-only.
2140 * Because vm_pages can overlap buffers m->busy can be > 1. m->busy is only
2141 * adjusted while the vm_page is PG_BUSY so the flash will occur when the
2142 * busy bit is cleared.
2145 vm_page_io_start(vm_page_t m)
2147 KASSERT(m->flags & PG_BUSY, ("vm_page_io_start: page not busy!!!"));
2148 atomic_add_char(&m->busy, 1);
2149 vm_page_flag_set(m, PG_SBUSY);
2153 vm_page_io_finish(vm_page_t m)
2155 KASSERT(m->flags & PG_BUSY, ("vm_page_io_finish: page not busy!!!"));
2156 atomic_subtract_char(&m->busy, 1);
2158 vm_page_flag_clear(m, PG_SBUSY);
2162 * Grab a page, blocking if it is busy and allocating a page if necessary.
2163 * A busy page is returned or NULL. The page may or may not be valid and
2164 * might not be on a queue (the caller is responsible for the disposition of
2167 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the
2168 * page will be zero'd and marked valid.
2170 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked
2171 * valid even if it already exists.
2173 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also
2174 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified.
2176 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
2177 * always returned if we had blocked.
2179 * This routine may not be called from an interrupt.
2181 * PG_ZERO is *ALWAYS* cleared by this routine.
2183 * No other requirements.
2186 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2191 KKASSERT(allocflags &
2192 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
2193 vm_object_hold(object);
2195 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error);
2197 vm_page_sleep_busy(m, TRUE, "pgrbwt");
2198 if ((allocflags & VM_ALLOC_RETRY) == 0) {
2203 } else if (m == NULL) {
2204 m = vm_page_alloc(object, pindex,
2205 allocflags & ~VM_ALLOC_RETRY);
2209 if ((allocflags & VM_ALLOC_RETRY) == 0)
2218 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid.
2220 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set
2221 * valid even if already valid.
2223 if (m->valid == 0) {
2224 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) {
2225 if ((m->flags & PG_ZERO) == 0)
2226 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2227 m->valid = VM_PAGE_BITS_ALL;
2229 } else if (allocflags & VM_ALLOC_FORCE_ZERO) {
2230 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2231 m->valid = VM_PAGE_BITS_ALL;
2233 vm_page_flag_clear(m, PG_ZERO);
2235 vm_object_drop(object);
2240 * Mapping function for valid bits or for dirty bits in
2241 * a page. May not block.
2243 * Inputs are required to range within a page.
2249 vm_page_bits(int base, int size)
2255 base + size <= PAGE_SIZE,
2256 ("vm_page_bits: illegal base/size %d/%d", base, size)
2259 if (size == 0) /* handle degenerate case */
2262 first_bit = base >> DEV_BSHIFT;
2263 last_bit = (base + size - 1) >> DEV_BSHIFT;
2265 return ((2 << last_bit) - (1 << first_bit));
2269 * Sets portions of a page valid and clean. The arguments are expected
2270 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2271 * of any partial chunks touched by the range. The invalid portion of
2272 * such chunks will be zero'd.
2274 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically
2275 * align base to DEV_BSIZE so as not to mark clean a partially
2276 * truncated device block. Otherwise the dirty page status might be
2279 * This routine may not block.
2281 * (base + size) must be less then or equal to PAGE_SIZE.
2284 _vm_page_zero_valid(vm_page_t m, int base, int size)
2289 if (size == 0) /* handle degenerate case */
2293 * If the base is not DEV_BSIZE aligned and the valid
2294 * bit is clear, we have to zero out a portion of the
2298 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2299 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
2301 pmap_zero_page_area(
2309 * If the ending offset is not DEV_BSIZE aligned and the
2310 * valid bit is clear, we have to zero out a portion of
2314 endoff = base + size;
2316 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2317 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
2319 pmap_zero_page_area(
2322 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
2328 * Set valid, clear dirty bits. If validating the entire
2329 * page we can safely clear the pmap modify bit. We also
2330 * use this opportunity to clear the PG_NOSYNC flag. If a process
2331 * takes a write fault on a MAP_NOSYNC memory area the flag will
2334 * We set valid bits inclusive of any overlap, but we can only
2335 * clear dirty bits for DEV_BSIZE chunks that are fully within
2338 * Page must be busied?
2339 * No other requirements.
2342 vm_page_set_valid(vm_page_t m, int base, int size)
2344 _vm_page_zero_valid(m, base, size);
2345 m->valid |= vm_page_bits(base, size);
2350 * Set valid bits and clear dirty bits.
2352 * NOTE: This function does not clear the pmap modified bit.
2353 * Also note that e.g. NFS may use a byte-granular base
2356 * WARNING: Page must be busied? But vfs_clean_one_page() will call
2357 * this without necessarily busying the page (via bdwrite()).
2358 * So for now vm_token must also be held.
2360 * No other requirements.
2363 vm_page_set_validclean(vm_page_t m, int base, int size)
2367 _vm_page_zero_valid(m, base, size);
2368 pagebits = vm_page_bits(base, size);
2369 m->valid |= pagebits;
2370 m->dirty &= ~pagebits;
2371 if (base == 0 && size == PAGE_SIZE) {
2372 /*pmap_clear_modify(m);*/
2373 vm_page_flag_clear(m, PG_NOSYNC);
2378 * Set valid & dirty. Used by buwrite()
2380 * WARNING: Page must be busied? But vfs_dirty_one_page() will
2381 * call this function in buwrite() so for now vm_token must
2384 * No other requirements.
2387 vm_page_set_validdirty(vm_page_t m, int base, int size)
2391 pagebits = vm_page_bits(base, size);
2392 m->valid |= pagebits;
2393 m->dirty |= pagebits;
2395 vm_object_set_writeable_dirty(m->object);
2401 * NOTE: This function does not clear the pmap modified bit.
2402 * Also note that e.g. NFS may use a byte-granular base
2405 * Page must be busied?
2406 * No other requirements.
2409 vm_page_clear_dirty(vm_page_t m, int base, int size)
2411 m->dirty &= ~vm_page_bits(base, size);
2412 if (base == 0 && size == PAGE_SIZE) {
2413 /*pmap_clear_modify(m);*/
2414 vm_page_flag_clear(m, PG_NOSYNC);
2419 * Make the page all-dirty.
2421 * Also make sure the related object and vnode reflect the fact that the
2422 * object may now contain a dirty page.
2424 * Page must be busied?
2425 * No other requirements.
2428 vm_page_dirty(vm_page_t m)
2431 int pqtype = m->queue - m->pc;
2433 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE,
2434 ("vm_page_dirty: page in free/cache queue!"));
2435 if (m->dirty != VM_PAGE_BITS_ALL) {
2436 m->dirty = VM_PAGE_BITS_ALL;
2438 vm_object_set_writeable_dirty(m->object);
2443 * Invalidates DEV_BSIZE'd chunks within a page. Both the
2444 * valid and dirty bits for the effected areas are cleared.
2446 * Page must be busied?
2448 * No other requirements.
2451 vm_page_set_invalid(vm_page_t m, int base, int size)
2455 bits = vm_page_bits(base, size);
2458 m->object->generation++;
2462 * The kernel assumes that the invalid portions of a page contain
2463 * garbage, but such pages can be mapped into memory by user code.
2464 * When this occurs, we must zero out the non-valid portions of the
2465 * page so user code sees what it expects.
2467 * Pages are most often semi-valid when the end of a file is mapped
2468 * into memory and the file's size is not page aligned.
2470 * Page must be busied?
2471 * No other requirements.
2474 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
2480 * Scan the valid bits looking for invalid sections that
2481 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
2482 * valid bit may be set ) have already been zerod by
2483 * vm_page_set_validclean().
2485 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
2486 if (i == (PAGE_SIZE / DEV_BSIZE) ||
2487 (m->valid & (1 << i))
2490 pmap_zero_page_area(
2493 (i - b) << DEV_BSHIFT
2501 * setvalid is TRUE when we can safely set the zero'd areas
2502 * as being valid. We can do this if there are no cache consistency
2503 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
2506 m->valid = VM_PAGE_BITS_ALL;
2510 * Is a (partial) page valid? Note that the case where size == 0
2511 * will return FALSE in the degenerate case where the page is entirely
2512 * invalid, and TRUE otherwise.
2515 * No other requirements.
2518 vm_page_is_valid(vm_page_t m, int base, int size)
2520 int bits = vm_page_bits(base, size);
2522 if (m->valid && ((m->valid & bits) == bits))
2529 * update dirty bits from pmap/mmu. May not block.
2531 * Caller must hold the page busy
2534 vm_page_test_dirty(vm_page_t m)
2536 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
2542 * Register an action, associating it with its vm_page
2545 vm_page_register_action(vm_page_action_t action, vm_page_event_t event)
2547 struct vm_page_action_list *list;
2550 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
2551 list = &action_list[hv];
2553 lwkt_gettoken(&vm_token);
2554 vm_page_flag_set(action->m, PG_ACTIONLIST);
2555 action->event = event;
2556 LIST_INSERT_HEAD(list, action, entry);
2557 lwkt_reltoken(&vm_token);
2561 * Unregister an action, disassociating it from its related vm_page
2564 vm_page_unregister_action(vm_page_action_t action)
2566 struct vm_page_action_list *list;
2569 lwkt_gettoken(&vm_token);
2570 if (action->event != VMEVENT_NONE) {
2571 action->event = VMEVENT_NONE;
2572 LIST_REMOVE(action, entry);
2574 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
2575 list = &action_list[hv];
2576 if (LIST_EMPTY(list))
2577 vm_page_flag_clear(action->m, PG_ACTIONLIST);
2579 lwkt_reltoken(&vm_token);
2583 * Issue an event on a VM page. Corresponding action structures are
2584 * removed from the page's list and called.
2586 * If the vm_page has no more pending action events we clear its
2587 * PG_ACTIONLIST flag.
2590 vm_page_event_internal(vm_page_t m, vm_page_event_t event)
2592 struct vm_page_action_list *list;
2593 struct vm_page_action *scan;
2594 struct vm_page_action *next;
2598 hv = (int)((intptr_t)m >> 8) & VMACTION_HMASK;
2599 list = &action_list[hv];
2602 lwkt_gettoken(&vm_token);
2603 LIST_FOREACH_MUTABLE(scan, list, entry, next) {
2605 if (scan->event == event) {
2606 scan->event = VMEVENT_NONE;
2607 LIST_REMOVE(scan, entry);
2608 scan->func(m, scan);
2616 vm_page_flag_clear(m, PG_ACTIONLIST);
2617 lwkt_reltoken(&vm_token);
2620 #include "opt_ddb.h"
2622 #include <sys/kernel.h>
2624 #include <ddb/ddb.h>
2626 DB_SHOW_COMMAND(page, vm_page_print_page_info)
2628 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count);
2629 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count);
2630 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count);
2631 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count);
2632 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count);
2633 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved);
2634 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min);
2635 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target);
2636 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min);
2637 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target);
2640 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
2643 db_printf("PQ_FREE:");
2644 for(i=0;i<PQ_L2_SIZE;i++) {
2645 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
2649 db_printf("PQ_CACHE:");
2650 for(i=0;i<PQ_L2_SIZE;i++) {
2651 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
2655 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
2656 vm_page_queues[PQ_ACTIVE].lcnt,
2657 vm_page_queues[PQ_INACTIVE].lcnt);