4 * Copyright (c) 1991 Regents of the University of California.
7 * This code is derived from software contributed to Berkeley by
8 * The Mach Operating System project at Carnegie-Mellon University.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 4. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91
35 * $FreeBSD: src/sys/vm/vm_page.c,v 1.147.2.18 2002/03/10 05:03:19 alc Exp $
39 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40 * All rights reserved.
42 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
44 * Permission to use, copy, modify and distribute this software and
45 * its documentation is hereby granted, provided that both the copyright
46 * notice and this permission notice appear in all copies of the
47 * software, derivative works or modified versions, and any portions
48 * thereof, and that both notices appear in supporting documentation.
50 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
54 * Carnegie Mellon requests users of this software to return to
56 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
57 * School of Computer Science
58 * Carnegie Mellon University
59 * Pittsburgh PA 15213-3890
61 * any improvements or extensions that they make and grant Carnegie the
62 * rights to redistribute these changes.
65 * Resident memory management module. The module manipulates 'VM pages'.
66 * A VM page is the core building block for memory management.
69 #include <sys/param.h>
70 #include <sys/systm.h>
71 #include <sys/malloc.h>
73 #include <sys/vmmeter.h>
74 #include <sys/vnode.h>
75 #include <sys/kernel.h>
78 #include <vm/vm_param.h>
80 #include <vm/vm_kern.h>
82 #include <vm/vm_map.h>
83 #include <vm/vm_object.h>
84 #include <vm/vm_page.h>
85 #include <vm/vm_pageout.h>
86 #include <vm/vm_pager.h>
87 #include <vm/vm_extern.h>
88 #include <vm/swap_pager.h>
90 #include <machine/inttypes.h>
91 #include <machine/md_var.h>
93 #include <vm/vm_page2.h>
94 #include <sys/spinlock2.h>
96 #define VMACTION_HSIZE 256
97 #define VMACTION_HMASK (VMACTION_HSIZE - 1)
99 static void vm_page_queue_init(void);
100 static void vm_page_free_wakeup(void);
101 static vm_page_t vm_page_select_cache(u_short pg_color);
102 static vm_page_t _vm_page_list_find2(int basequeue, int index);
103 static void _vm_page_deactivate_locked(vm_page_t m, int athead);
106 * Array of tailq lists
108 __cachealign struct vpgqueues vm_page_queues[PQ_COUNT];
110 LIST_HEAD(vm_page_action_list, vm_page_action);
111 struct vm_page_action_list action_list[VMACTION_HSIZE];
112 static volatile int vm_pages_waiting;
115 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare,
116 vm_pindex_t, pindex);
119 vm_page_queue_init(void)
123 for (i = 0; i < PQ_L2_SIZE; i++)
124 vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count;
125 for (i = 0; i < PQ_L2_SIZE; i++)
126 vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count;
128 vm_page_queues[PQ_INACTIVE].cnt = &vmstats.v_inactive_count;
129 vm_page_queues[PQ_ACTIVE].cnt = &vmstats.v_active_count;
130 vm_page_queues[PQ_HOLD].cnt = &vmstats.v_active_count;
131 /* PQ_NONE has no queue */
133 for (i = 0; i < PQ_COUNT; i++) {
134 TAILQ_INIT(&vm_page_queues[i].pl);
135 spin_init(&vm_page_queues[i].spin);
138 for (i = 0; i < VMACTION_HSIZE; i++)
139 LIST_INIT(&action_list[i]);
143 * note: place in initialized data section? Is this necessary?
146 int vm_page_array_size = 0;
147 int vm_page_zero_count = 0;
148 vm_page_t vm_page_array = 0;
153 * Sets the page size, perhaps based upon the memory size.
154 * Must be called before any use of page-size dependent functions.
157 vm_set_page_size(void)
159 if (vmstats.v_page_size == 0)
160 vmstats.v_page_size = PAGE_SIZE;
161 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0)
162 panic("vm_set_page_size: page size not a power of two");
168 * Add a new page to the freelist for use by the system. New pages
169 * are added to both the head and tail of the associated free page
170 * queue in a bottom-up fashion, so both zero'd and non-zero'd page
171 * requests pull 'recent' adds (higher physical addresses) first.
173 * Must be called in a critical section.
176 vm_add_new_page(vm_paddr_t pa)
178 struct vpgqueues *vpq;
181 m = PHYS_TO_VM_PAGE(pa);
184 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK;
187 * Twist for cpu localization instead of page coloring.
189 m->pc ^= ((pa >> PAGE_SHIFT) / PQ_L2_SIZE) & PQ_L2_MASK;
190 m->pc ^= ((pa >> PAGE_SHIFT) / (PQ_L2_SIZE * PQ_L2_SIZE)) & PQ_L2_MASK;
192 m->queue = m->pc + PQ_FREE;
193 KKASSERT(m->dirty == 0);
195 atomic_add_int(&vmstats.v_page_count, 1);
196 atomic_add_int(&vmstats.v_free_count, 1);
197 vpq = &vm_page_queues[m->queue];
199 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
201 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
202 vpq->flipflop = 1 - vpq->flipflop;
211 * Initializes the resident memory module.
213 * Preallocates memory for critical VM structures and arrays prior to
214 * kernel_map becoming available.
216 * Memory is allocated from (virtual2_start, virtual2_end) if available,
217 * otherwise memory is allocated from (virtual_start, virtual_end).
219 * On x86-64 (virtual_start, virtual_end) is only 2GB and may not be
220 * large enough to hold vm_page_array & other structures for machines with
221 * large amounts of ram, so we want to use virtual2* when available.
224 vm_page_startup(void)
226 vm_offset_t vaddr = virtual2_start ? virtual2_start : virtual_start;
229 vm_paddr_t page_range;
236 vm_paddr_t biggestone, biggestsize;
243 vaddr = round_page(vaddr);
245 for (i = 0; phys_avail[i + 1]; i += 2) {
246 phys_avail[i] = round_page64(phys_avail[i]);
247 phys_avail[i + 1] = trunc_page64(phys_avail[i + 1]);
250 for (i = 0; phys_avail[i + 1]; i += 2) {
251 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
253 if (size > biggestsize) {
261 end = phys_avail[biggestone+1];
262 end = trunc_page(end);
265 * Initialize the queue headers for the free queue, the active queue
266 * and the inactive queue.
269 vm_page_queue_init();
271 #if !defined(_KERNEL_VIRTUAL)
273 * VKERNELs don't support minidumps and as such don't need
276 * Allocate a bitmap to indicate that a random physical page
277 * needs to be included in a minidump.
279 * The amd64 port needs this to indicate which direct map pages
280 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
282 * However, i386 still needs this workspace internally within the
283 * minidump code. In theory, they are not needed on i386, but are
284 * included should the sf_buf code decide to use them.
286 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE;
287 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
288 end -= vm_page_dump_size;
289 vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size,
290 VM_PROT_READ | VM_PROT_WRITE);
291 bzero((void *)vm_page_dump, vm_page_dump_size);
295 * Compute the number of pages of memory that will be available for
296 * use (taking into account the overhead of a page structure per
299 first_page = phys_avail[0] / PAGE_SIZE;
300 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
301 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE;
304 * Initialize the mem entry structures now, and put them in the free
307 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
308 mapped = pmap_map(&vaddr, new_end, end,
309 VM_PROT_READ | VM_PROT_WRITE);
310 vm_page_array = (vm_page_t)mapped;
312 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL)
314 * since pmap_map on amd64 returns stuff out of a direct-map region,
315 * we have to manually add these pages to the minidump tracking so
316 * that they can be dumped, including the vm_page_array.
318 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
323 * Clear all of the page structures
325 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
326 vm_page_array_size = page_range;
329 * Construct the free queue(s) in ascending order (by physical
330 * address) so that the first 16MB of physical memory is allocated
331 * last rather than first. On large-memory machines, this avoids
332 * the exhaustion of low physical memory before isa_dmainit has run.
334 vmstats.v_page_count = 0;
335 vmstats.v_free_count = 0;
336 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
341 last_pa = phys_avail[i + 1];
342 while (pa < last_pa && npages-- > 0) {
348 virtual2_start = vaddr;
350 virtual_start = vaddr;
354 * Scan comparison function for Red-Black tree scans. An inclusive
355 * (start,end) is expected. Other fields are not used.
358 rb_vm_page_scancmp(struct vm_page *p, void *data)
360 struct rb_vm_page_scan_info *info = data;
362 if (p->pindex < info->start_pindex)
364 if (p->pindex > info->end_pindex)
370 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2)
372 if (p1->pindex < p2->pindex)
374 if (p1->pindex > p2->pindex)
380 * Each page queue has its own spin lock, which is fairly optimal for
381 * allocating and freeing pages at least.
383 * The caller must hold the vm_page_spin_lock() before locking a vm_page's
384 * queue spinlock via this function. Also note that m->queue cannot change
385 * unless both the page and queue are locked.
389 _vm_page_queue_spin_lock(vm_page_t m)
394 if (queue != PQ_NONE) {
395 spin_lock(&vm_page_queues[queue].spin);
396 KKASSERT(queue == m->queue);
402 _vm_page_queue_spin_unlock(vm_page_t m)
408 if (queue != PQ_NONE)
409 spin_unlock(&vm_page_queues[queue].spin);
414 _vm_page_queues_spin_lock(u_short queue)
417 if (queue != PQ_NONE)
418 spin_lock(&vm_page_queues[queue].spin);
424 _vm_page_queues_spin_unlock(u_short queue)
427 if (queue != PQ_NONE)
428 spin_unlock(&vm_page_queues[queue].spin);
432 vm_page_queue_spin_lock(vm_page_t m)
434 _vm_page_queue_spin_lock(m);
438 vm_page_queues_spin_lock(u_short queue)
440 _vm_page_queues_spin_lock(queue);
444 vm_page_queue_spin_unlock(vm_page_t m)
446 _vm_page_queue_spin_unlock(m);
450 vm_page_queues_spin_unlock(u_short queue)
452 _vm_page_queues_spin_unlock(queue);
456 * This locks the specified vm_page and its queue in the proper order
457 * (page first, then queue). The queue may change so the caller must
462 _vm_page_and_queue_spin_lock(vm_page_t m)
464 vm_page_spin_lock(m);
465 _vm_page_queue_spin_lock(m);
470 _vm_page_and_queue_spin_unlock(vm_page_t m)
472 _vm_page_queues_spin_unlock(m->queue);
473 vm_page_spin_unlock(m);
477 vm_page_and_queue_spin_unlock(vm_page_t m)
479 _vm_page_and_queue_spin_unlock(m);
483 vm_page_and_queue_spin_lock(vm_page_t m)
485 _vm_page_and_queue_spin_lock(m);
489 * Helper function removes vm_page from its current queue.
490 * Returns the base queue the page used to be on.
492 * The vm_page and the queue must be spinlocked.
493 * This function will unlock the queue but leave the page spinlocked.
495 static __inline u_short
496 _vm_page_rem_queue_spinlocked(vm_page_t m)
498 struct vpgqueues *pq;
502 if (queue != PQ_NONE) {
503 pq = &vm_page_queues[queue];
504 TAILQ_REMOVE(&pq->pl, m, pageq);
505 atomic_add_int(pq->cnt, -1);
508 vm_page_queues_spin_unlock(queue);
509 if ((queue - m->pc) == PQ_FREE && (m->flags & PG_ZERO))
510 atomic_subtract_int(&vm_page_zero_count, 1);
511 if ((queue - m->pc) == PQ_CACHE || (queue - m->pc) == PQ_FREE)
512 return (queue - m->pc);
518 * Helper function places the vm_page on the specified queue.
520 * The vm_page must be spinlocked.
521 * This function will return with both the page and the queue locked.
524 _vm_page_add_queue_spinlocked(vm_page_t m, u_short queue, int athead)
526 struct vpgqueues *pq;
528 KKASSERT(m->queue == PQ_NONE);
530 if (queue != PQ_NONE) {
531 vm_page_queues_spin_lock(queue);
532 pq = &vm_page_queues[queue];
534 atomic_add_int(pq->cnt, 1);
538 * Put zero'd pages on the end ( where we look for zero'd pages
539 * first ) and non-zerod pages at the head.
541 if (queue - m->pc == PQ_FREE) {
542 if (m->flags & PG_ZERO) {
543 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
544 atomic_add_int(&vm_page_zero_count, 1);
546 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
549 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
551 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
553 /* leave the queue spinlocked */
558 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE)
559 * m->busy is zero. Returns TRUE if it had to sleep, FALSE if we
560 * did not. Only one sleep call will be made before returning.
562 * This function does NOT busy the page and on return the page is not
563 * guaranteed to be available.
566 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg)
574 if ((flags & PG_BUSY) == 0 &&
575 (also_m_busy == 0 || (flags & PG_SBUSY) == 0)) {
578 tsleep_interlock(m, 0);
579 if (atomic_cmpset_int(&m->flags, flags,
580 flags | PG_WANTED | PG_REFERENCED)) {
581 tsleep(m, PINTERLOCKED, msg, 0);
588 * Wait until PG_BUSY can be set, then set it. If also_m_busy is TRUE we
589 * also wait for m->busy to become 0 before setting PG_BUSY.
592 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m,
593 int also_m_busy, const char *msg
601 if (flags & PG_BUSY) {
602 tsleep_interlock(m, 0);
603 if (atomic_cmpset_int(&m->flags, flags,
604 flags | PG_WANTED | PG_REFERENCED)) {
605 tsleep(m, PINTERLOCKED, msg, 0);
607 } else if (also_m_busy && (flags & PG_SBUSY)) {
608 tsleep_interlock(m, 0);
609 if (atomic_cmpset_int(&m->flags, flags,
610 flags | PG_WANTED | PG_REFERENCED)) {
611 tsleep(m, PINTERLOCKED, msg, 0);
614 if (atomic_cmpset_int(&m->flags, flags,
618 m->busy_line = lineno;
627 * Attempt to set PG_BUSY. If also_m_busy is TRUE we only succeed if m->busy
630 * Returns non-zero on failure.
633 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy
643 if (also_m_busy && (flags & PG_SBUSY))
645 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
648 m->busy_line = lineno;
656 * Clear the PG_BUSY flag and return non-zero to indicate to the caller
657 * that a wakeup() should be performed.
659 * The vm_page must be spinlocked and will remain spinlocked on return.
660 * The related queue must NOT be spinlocked (which could deadlock us).
666 _vm_page_wakeup(vm_page_t m)
673 if (atomic_cmpset_int(&m->flags, flags,
674 flags & ~(PG_BUSY | PG_WANTED))) {
678 return(flags & PG_WANTED);
682 * Clear the PG_BUSY flag and wakeup anyone waiting for the page. This
683 * is typically the last call you make on a page before moving onto
687 vm_page_wakeup(vm_page_t m)
689 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!"));
690 vm_page_spin_lock(m);
691 if (_vm_page_wakeup(m)) {
692 vm_page_spin_unlock(m);
695 vm_page_spin_unlock(m);
700 * Holding a page keeps it from being reused. Other parts of the system
701 * can still disassociate the page from its current object and free it, or
702 * perform read or write I/O on it and/or otherwise manipulate the page,
703 * but if the page is held the VM system will leave the page and its data
704 * intact and not reuse the page for other purposes until the last hold
705 * reference is released. (see vm_page_wire() if you want to prevent the
706 * page from being disassociated from its object too).
708 * The caller must still validate the contents of the page and, if necessary,
709 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete
710 * before manipulating the page.
712 * XXX get vm_page_spin_lock() here and move FREE->HOLD if necessary
715 vm_page_hold(vm_page_t m)
717 vm_page_spin_lock(m);
718 atomic_add_int(&m->hold_count, 1);
719 if (m->queue - m->pc == PQ_FREE) {
720 _vm_page_queue_spin_lock(m);
721 _vm_page_rem_queue_spinlocked(m);
722 _vm_page_add_queue_spinlocked(m, PQ_HOLD, 0);
723 _vm_page_queue_spin_unlock(m);
725 vm_page_spin_unlock(m);
729 * The opposite of vm_page_hold(). A page can be freed while being held,
730 * which places it on the PQ_HOLD queue. If we are able to busy the page
731 * after the hold count drops to zero we will move the page to the
732 * appropriate PQ_FREE queue by calling vm_page_free_toq().
735 vm_page_unhold(vm_page_t m)
737 vm_page_spin_lock(m);
738 atomic_add_int(&m->hold_count, -1);
739 if (m->hold_count == 0 && m->queue == PQ_HOLD) {
740 _vm_page_queue_spin_lock(m);
741 _vm_page_rem_queue_spinlocked(m);
742 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0);
743 _vm_page_queue_spin_unlock(m);
745 vm_page_spin_unlock(m);
749 * Inserts the given vm_page into the object and object list.
751 * The pagetables are not updated but will presumably fault the page
752 * in if necessary, or if a kernel page the caller will at some point
753 * enter the page into the kernel's pmap. We are not allowed to block
754 * here so we *can't* do this anyway.
756 * This routine may not block.
757 * This routine must be called with the vm_object held.
758 * This routine must be called with a critical section held.
760 * This routine returns TRUE if the page was inserted into the object
761 * successfully, and FALSE if the page already exists in the object.
764 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
766 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
767 if (m->object != NULL)
768 panic("vm_page_insert: already inserted");
770 object->generation++;
773 * Record the object/offset pair in this page and add the
774 * pv_list_count of the page to the object.
776 * The vm_page spin lock is required for interactions with the pmap.
778 vm_page_spin_lock(m);
781 if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) {
784 vm_page_spin_unlock(m);
787 object->resident_page_count++;
788 /* atomic_add_int(&object->agg_pv_list_count, m->md.pv_list_count); */
789 vm_page_spin_unlock(m);
792 * Since we are inserting a new and possibly dirty page,
793 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
795 if ((m->valid & m->dirty) || (m->flags & PG_WRITEABLE))
796 vm_object_set_writeable_dirty(object);
799 * Checks for a swap assignment and sets PG_SWAPPED if appropriate.
801 swap_pager_page_inserted(m);
806 * Removes the given vm_page_t from the (object,index) table
808 * The underlying pmap entry (if any) is NOT removed here.
809 * This routine may not block.
811 * The page must be BUSY and will remain BUSY on return.
812 * No other requirements.
814 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave
818 vm_page_remove(vm_page_t m)
822 if (m->object == NULL) {
826 if ((m->flags & PG_BUSY) == 0)
827 panic("vm_page_remove: page not busy");
831 vm_object_hold(object);
834 * Remove the page from the object and update the object.
836 * The vm_page spin lock is required for interactions with the pmap.
838 vm_page_spin_lock(m);
839 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m);
840 object->resident_page_count--;
841 /* atomic_add_int(&object->agg_pv_list_count, -m->md.pv_list_count); */
843 vm_page_spin_unlock(m);
845 object->generation++;
847 vm_object_drop(object);
851 * Locate and return the page at (object, pindex), or NULL if the
852 * page could not be found.
854 * The caller must hold the vm_object token.
857 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
862 * Search the hash table for this object/offset pair
864 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
865 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
866 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex));
871 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object,
873 int also_m_busy, const char *msg
879 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
880 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
882 KKASSERT(m->object == object && m->pindex == pindex);
885 if (flags & PG_BUSY) {
886 tsleep_interlock(m, 0);
887 if (atomic_cmpset_int(&m->flags, flags,
888 flags | PG_WANTED | PG_REFERENCED)) {
889 tsleep(m, PINTERLOCKED, msg, 0);
890 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
893 } else if (also_m_busy && (flags & PG_SBUSY)) {
894 tsleep_interlock(m, 0);
895 if (atomic_cmpset_int(&m->flags, flags,
896 flags | PG_WANTED | PG_REFERENCED)) {
897 tsleep(m, PINTERLOCKED, msg, 0);
898 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
901 } else if (atomic_cmpset_int(&m->flags, flags,
905 m->busy_line = lineno;
914 * Attempt to lookup and busy a page.
916 * Returns NULL if the page could not be found
918 * Returns a vm_page and error == TRUE if the page exists but could not
921 * Returns a vm_page and error == FALSE on success.
924 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object,
926 int also_m_busy, int *errorp
932 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
933 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
936 KKASSERT(m->object == object && m->pindex == pindex);
939 if (flags & PG_BUSY) {
943 if (also_m_busy && (flags & PG_SBUSY)) {
947 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
950 m->busy_line = lineno;
959 * Caller must hold the related vm_object
962 vm_page_next(vm_page_t m)
966 next = vm_page_rb_tree_RB_NEXT(m);
967 if (next && next->pindex != m->pindex + 1)
975 * Move the given vm_page from its current object to the specified
976 * target object/offset. The page must be busy and will remain so
979 * new_object must be held.
980 * This routine might block. XXX ?
982 * NOTE: Swap associated with the page must be invalidated by the move. We
983 * have to do this for several reasons: (1) we aren't freeing the
984 * page, (2) we are dirtying the page, (3) the VM system is probably
985 * moving the page from object A to B, and will then later move
986 * the backing store from A to B and we can't have a conflict.
988 * NOTE: We *always* dirty the page. It is necessary both for the
989 * fact that we moved it, and because we may be invalidating
990 * swap. If the page is on the cache, we have to deactivate it
991 * or vm_page_dirty() will panic. Dirty pages are not allowed
995 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
997 KKASSERT(m->flags & PG_BUSY);
998 ASSERT_LWKT_TOKEN_HELD(vm_object_token(new_object));
1000 ASSERT_LWKT_TOKEN_HELD(vm_object_token(m->object));
1003 if (vm_page_insert(m, new_object, new_pindex) == FALSE) {
1004 panic("vm_page_rename: target exists (%p,%"PRIu64")",
1005 new_object, new_pindex);
1007 if (m->queue - m->pc == PQ_CACHE)
1008 vm_page_deactivate(m);
1013 * vm_page_unqueue() without any wakeup. This routine is used when a page
1014 * is being moved between queues or otherwise is to remain BUSYied by the
1017 * This routine may not block.
1020 vm_page_unqueue_nowakeup(vm_page_t m)
1022 vm_page_and_queue_spin_lock(m);
1023 (void)_vm_page_rem_queue_spinlocked(m);
1024 vm_page_spin_unlock(m);
1028 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
1031 * This routine may not block.
1034 vm_page_unqueue(vm_page_t m)
1038 vm_page_and_queue_spin_lock(m);
1039 queue = _vm_page_rem_queue_spinlocked(m);
1040 if (queue == PQ_FREE || queue == PQ_CACHE) {
1041 vm_page_spin_unlock(m);
1042 pagedaemon_wakeup();
1044 vm_page_spin_unlock(m);
1049 * vm_page_list_find()
1051 * Find a page on the specified queue with color optimization.
1053 * The page coloring optimization attempts to locate a page that does
1054 * not overload other nearby pages in the object in the cpu's L1 or L2
1055 * caches. We need this optimization because cpu caches tend to be
1056 * physical caches, while object spaces tend to be virtual.
1058 * On MP systems each PQ_FREE and PQ_CACHE color queue has its own spinlock
1059 * and the algorithm is adjusted to localize allocations on a per-core basis.
1060 * This is done by 'twisting' the colors.
1062 * The page is returned spinlocked and removed from its queue (it will
1063 * be on PQ_NONE), or NULL. The page is not PG_BUSY'd. The caller
1064 * is responsible for dealing with the busy-page case (usually by
1065 * deactivating the page and looping).
1067 * NOTE: This routine is carefully inlined. A non-inlined version
1068 * is available for outside callers but the only critical path is
1069 * from within this source file.
1071 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE
1072 * represent stable storage, allowing us to order our locks vm_page
1073 * first, then queue.
1077 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
1083 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist);
1085 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl);
1087 m = _vm_page_list_find2(basequeue, index);
1090 vm_page_and_queue_spin_lock(m);
1091 if (m->queue == basequeue + index) {
1092 _vm_page_rem_queue_spinlocked(m);
1093 /* vm_page_t spin held, no queue spin */
1096 vm_page_and_queue_spin_unlock(m);
1102 _vm_page_list_find2(int basequeue, int index)
1106 struct vpgqueues *pq;
1108 pq = &vm_page_queues[basequeue];
1111 * Note that for the first loop, index+i and index-i wind up at the
1112 * same place. Even though this is not totally optimal, we've already
1113 * blown it by missing the cache case so we do not care.
1115 for (i = PQ_L2_SIZE / 2; i > 0; --i) {
1117 m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl);
1119 _vm_page_and_queue_spin_lock(m);
1121 basequeue + ((index + i) & PQ_L2_MASK)) {
1122 _vm_page_rem_queue_spinlocked(m);
1125 _vm_page_and_queue_spin_unlock(m);
1128 m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl);
1130 _vm_page_and_queue_spin_lock(m);
1132 basequeue + ((index - i) & PQ_L2_MASK)) {
1133 _vm_page_rem_queue_spinlocked(m);
1136 _vm_page_and_queue_spin_unlock(m);
1146 * Returns a vm_page candidate for allocation. The page is not busied so
1147 * it can move around. The caller must busy the page (and typically
1148 * deactivate it if it cannot be busied!)
1150 * Returns a spinlocked vm_page that has been removed from its queue.
1153 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
1155 return(_vm_page_list_find(basequeue, index, prefer_zero));
1159 * Find a page on the cache queue with color optimization, remove it
1160 * from the queue, and busy it. The returned page will not be spinlocked.
1162 * A candidate failure will be deactivated. Candidates can fail due to
1163 * being busied by someone else, in which case they will be deactivated.
1165 * This routine may not block.
1169 vm_page_select_cache(u_short pg_color)
1174 m = _vm_page_list_find(PQ_CACHE, pg_color & PQ_L2_MASK, FALSE);
1178 * (m) has been removed from its queue and spinlocked
1180 if (vm_page_busy_try(m, TRUE)) {
1181 _vm_page_deactivate_locked(m, 0);
1182 vm_page_spin_unlock(m);
1184 kprintf("Warning: busy page %p found in cache\n", m);
1188 * We successfully busied the page
1190 if ((m->flags & PG_UNMANAGED) == 0 &&
1191 m->hold_count == 0 &&
1192 m->wire_count == 0) {
1193 vm_page_spin_unlock(m);
1194 pagedaemon_wakeup();
1197 _vm_page_deactivate_locked(m, 0);
1198 if (_vm_page_wakeup(m)) {
1199 vm_page_spin_unlock(m);
1202 vm_page_spin_unlock(m);
1210 * Find a free or zero page, with specified preference. We attempt to
1211 * inline the nominal case and fall back to _vm_page_select_free()
1212 * otherwise. A busied page is removed from the queue and returned.
1214 * This routine may not block.
1216 static __inline vm_page_t
1217 vm_page_select_free(u_short pg_color, boolean_t prefer_zero)
1222 m = _vm_page_list_find(PQ_FREE, pg_color & PQ_L2_MASK,
1226 if (vm_page_busy_try(m, TRUE)) {
1227 _vm_page_deactivate_locked(m, 0);
1228 vm_page_spin_unlock(m);
1230 kprintf("Warning: busy page %p found in cache\n", m);
1233 KKASSERT((m->flags & PG_UNMANAGED) == 0);
1234 KKASSERT(m->hold_count == 0);
1235 KKASSERT(m->wire_count == 0);
1236 vm_page_spin_unlock(m);
1237 pagedaemon_wakeup();
1239 /* return busied and removed page */
1249 * Allocate and return a memory cell associated with this VM object/offset
1250 * pair. If object is NULL an unassociated page will be allocated.
1252 * The returned page will be busied and removed from its queues. This
1253 * routine can block and may return NULL if a race occurs and the page
1254 * is found to already exist at the specified (object, pindex).
1256 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain
1257 * VM_ALLOC_QUICK like normal but cannot use cache
1258 * VM_ALLOC_SYSTEM greater free drain
1259 * VM_ALLOC_INTERRUPT allow free list to be completely drained
1260 * VM_ALLOC_ZERO advisory request for pre-zero'd page only
1261 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only
1262 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision
1263 * (see vm_page_grab())
1264 * The object must be held if not NULL
1265 * This routine may not block
1267 * Additional special handling is required when called from an interrupt
1268 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache
1272 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
1279 * Cpu twist - cpu localization algorithm
1282 pg_color = mycpu->gd_cpuid + (pindex & ~ncpus_fit_mask) +
1283 (object->pg_color & ~ncpus_fit_mask);
1285 pg_color = mycpu->gd_cpuid + (pindex & ~ncpus_fit_mask);
1289 * Normal page coloring algorithm
1292 pg_color = object->pg_color + pindex;
1298 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK|
1299 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
1302 * Certain system threads (pageout daemon, buf_daemon's) are
1303 * allowed to eat deeper into the free page list.
1305 if (curthread->td_flags & TDF_SYSTHREAD)
1306 page_req |= VM_ALLOC_SYSTEM;
1309 if (vmstats.v_free_count > vmstats.v_free_reserved ||
1310 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) ||
1311 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 &&
1312 vmstats.v_free_count > vmstats.v_interrupt_free_min)
1315 * The free queue has sufficient free pages to take one out.
1317 if (page_req & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO))
1318 m = vm_page_select_free(pg_color, TRUE);
1320 m = vm_page_select_free(pg_color, FALSE);
1321 } else if (page_req & VM_ALLOC_NORMAL) {
1323 * Allocatable from the cache (non-interrupt only). On
1324 * success, we must free the page and try again, thus
1325 * ensuring that vmstats.v_*_free_min counters are replenished.
1328 if (curthread->td_preempted) {
1329 kprintf("vm_page_alloc(): warning, attempt to allocate"
1330 " cache page from preempting interrupt\n");
1333 m = vm_page_select_cache(pg_color);
1336 m = vm_page_select_cache(pg_color);
1339 * On success move the page into the free queue and loop.
1342 KASSERT(m->dirty == 0,
1343 ("Found dirty cache page %p", m));
1344 vm_page_protect(m, VM_PROT_NONE);
1350 * On failure return NULL
1352 #if defined(DIAGNOSTIC)
1353 if (vmstats.v_cache_count > 0)
1354 kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count);
1356 vm_pageout_deficit++;
1357 pagedaemon_wakeup();
1361 * No pages available, wakeup the pageout daemon and give up.
1363 vm_pageout_deficit++;
1364 pagedaemon_wakeup();
1369 * v_free_count can race so loop if we don't find the expected
1376 * Good page found. The page has already been busied for us and
1377 * removed from its queues.
1379 KASSERT(m->dirty == 0,
1380 ("vm_page_alloc: free/cache page %p was dirty", m));
1381 KKASSERT(m->queue == PQ_NONE);
1384 * Initialize the structure, inheriting some flags but clearing
1385 * all the rest. The page has already been busied for us.
1387 vm_page_flag_clear(m, ~(PG_ZERO | PG_BUSY | PG_SBUSY));
1388 KKASSERT(m->wire_count == 0);
1389 KKASSERT(m->busy == 0);
1394 * Caller must be holding the object lock (asserted by
1395 * vm_page_insert()).
1397 * NOTE: Inserting a page here does not insert it into any pmaps
1398 * (which could cause us to block allocating memory).
1400 * NOTE: If no object an unassociated page is allocated, m->pindex
1401 * can be used by the caller for any purpose.
1404 if (vm_page_insert(m, object, pindex) == FALSE) {
1405 kprintf("PAGE RACE (%p:%d,%"PRIu64")\n",
1406 object, object->type, pindex);
1409 if ((page_req & VM_ALLOC_NULL_OK) == 0)
1417 * Don't wakeup too often - wakeup the pageout daemon when
1418 * we would be nearly out of memory.
1420 pagedaemon_wakeup();
1423 * A PG_BUSY page is returned.
1429 * Wait for sufficient free memory for nominal heavy memory use kernel
1433 vm_wait_nominal(void)
1435 while (vm_page_count_min(0))
1440 * Test if vm_wait_nominal() would block.
1443 vm_test_nominal(void)
1445 if (vm_page_count_min(0))
1451 * Block until free pages are available for allocation, called in various
1452 * places before memory allocations.
1454 * The caller may loop if vm_page_count_min() == FALSE so we cannot be
1455 * more generous then that.
1461 * never wait forever
1465 lwkt_gettoken(&vm_token);
1467 if (curthread == pagethread) {
1469 * The pageout daemon itself needs pages, this is bad.
1471 if (vm_page_count_min(0)) {
1472 vm_pageout_pages_needed = 1;
1473 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo);
1477 * Wakeup the pageout daemon if necessary and wait.
1479 if (vm_page_count_target()) {
1480 if (vm_pages_needed == 0) {
1481 vm_pages_needed = 1;
1482 wakeup(&vm_pages_needed);
1484 ++vm_pages_waiting; /* SMP race ok */
1485 tsleep(&vmstats.v_free_count, 0, "vmwait", timo);
1488 lwkt_reltoken(&vm_token);
1492 * Block until free pages are available for allocation
1494 * Called only from vm_fault so that processes page faulting can be
1501 * Wakeup the pageout daemon if necessary and wait.
1503 if (vm_page_count_target()) {
1504 lwkt_gettoken(&vm_token);
1505 if (vm_page_count_target()) {
1506 if (vm_pages_needed == 0) {
1507 vm_pages_needed = 1;
1508 wakeup(&vm_pages_needed);
1510 ++vm_pages_waiting; /* SMP race ok */
1511 tsleep(&vmstats.v_free_count, 0, "pfault", hz);
1513 lwkt_reltoken(&vm_token);
1518 * Put the specified page on the active list (if appropriate). Ensure
1519 * that act_count is at least ACT_INIT but do not otherwise mess with it.
1521 * The caller should be holding the page busied ? XXX
1522 * This routine may not block.
1525 vm_page_activate(vm_page_t m)
1529 vm_page_spin_lock(m);
1530 if (m->queue != PQ_ACTIVE) {
1531 _vm_page_queue_spin_lock(m);
1532 oqueue = _vm_page_rem_queue_spinlocked(m);
1533 /* page is left spinlocked, queue is unlocked */
1535 if (oqueue == PQ_CACHE)
1536 mycpu->gd_cnt.v_reactivated++;
1537 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1538 if (m->act_count < ACT_INIT)
1539 m->act_count = ACT_INIT;
1540 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE, 0);
1542 _vm_page_and_queue_spin_unlock(m);
1543 if (oqueue == PQ_CACHE || oqueue == PQ_FREE)
1544 pagedaemon_wakeup();
1546 if (m->act_count < ACT_INIT)
1547 m->act_count = ACT_INIT;
1548 vm_page_spin_unlock(m);
1553 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
1554 * routine is called when a page has been added to the cache or free
1557 * This routine may not block.
1559 static __inline void
1560 vm_page_free_wakeup(void)
1563 * If the pageout daemon itself needs pages, then tell it that
1564 * there are some free.
1566 if (vm_pageout_pages_needed &&
1567 vmstats.v_cache_count + vmstats.v_free_count >=
1568 vmstats.v_pageout_free_min
1570 wakeup(&vm_pageout_pages_needed);
1571 vm_pageout_pages_needed = 0;
1575 * Wakeup processes that are waiting on memory.
1577 * NOTE: vm_paging_target() is the pageout daemon's target, while
1578 * vm_page_count_target() is somewhere inbetween. We want
1579 * to wake processes up prior to the pageout daemon reaching
1580 * its target to provide some hysteresis.
1582 if (vm_pages_waiting) {
1583 if (!vm_page_count_target()) {
1585 * Plenty of pages are free, wakeup everyone.
1587 vm_pages_waiting = 0;
1588 wakeup(&vmstats.v_free_count);
1589 ++mycpu->gd_cnt.v_ppwakeups;
1590 } else if (!vm_page_count_min(0)) {
1592 * Some pages are free, wakeup someone.
1594 int wcount = vm_pages_waiting;
1597 vm_pages_waiting = wcount;
1598 wakeup_one(&vmstats.v_free_count);
1599 ++mycpu->gd_cnt.v_ppwakeups;
1605 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates
1606 * it from its VM object.
1608 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on
1609 * return (the page will have been freed).
1612 vm_page_free_toq(vm_page_t m)
1614 mycpu->gd_cnt.v_tfree++;
1615 KKASSERT((m->flags & PG_MAPPED) == 0);
1616 KKASSERT(m->flags & PG_BUSY);
1618 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
1620 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
1621 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
1623 if ((m->queue - m->pc) == PQ_FREE)
1624 panic("vm_page_free: freeing free page");
1626 panic("vm_page_free: freeing busy page");
1630 * Remove from object, spinlock the page and its queues and
1631 * remove from any queue. No queue spinlock will be held
1632 * after this section (because the page was removed from any
1636 vm_page_and_queue_spin_lock(m);
1637 _vm_page_rem_queue_spinlocked(m);
1640 * No further management of fictitious pages occurs beyond object
1641 * and queue removal.
1643 if ((m->flags & PG_FICTITIOUS) != 0) {
1644 vm_page_spin_unlock(m);
1652 if (m->wire_count != 0) {
1653 if (m->wire_count > 1) {
1655 "vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1656 m->wire_count, (long)m->pindex);
1658 panic("vm_page_free: freeing wired page");
1662 * Clear the UNMANAGED flag when freeing an unmanaged page.
1664 if (m->flags & PG_UNMANAGED) {
1665 vm_page_flag_clear(m, PG_UNMANAGED);
1668 if (m->hold_count != 0) {
1669 vm_page_flag_clear(m, PG_ZERO);
1670 _vm_page_add_queue_spinlocked(m, PQ_HOLD, 0);
1672 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0);
1676 * This sequence allows us to clear PG_BUSY while still holding
1677 * its spin lock, which reduces contention vs allocators. We
1678 * must not leave the queue locked or _vm_page_wakeup() may
1681 _vm_page_queue_spin_unlock(m);
1682 if (_vm_page_wakeup(m)) {
1683 vm_page_spin_unlock(m);
1686 vm_page_spin_unlock(m);
1688 vm_page_free_wakeup();
1692 * vm_page_free_fromq_fast()
1694 * Remove a non-zero page from one of the free queues; the page is removed for
1695 * zeroing, so do not issue a wakeup.
1698 vm_page_free_fromq_fast(void)
1704 for (i = 0; i < PQ_L2_SIZE; ++i) {
1705 m = vm_page_list_find(PQ_FREE, qi, FALSE);
1706 /* page is returned spinlocked and removed from its queue */
1708 if (vm_page_busy_try(m, TRUE)) {
1710 * We were unable to busy the page, deactivate
1713 _vm_page_deactivate_locked(m, 0);
1714 vm_page_spin_unlock(m);
1715 } else if ((m->flags & PG_ZERO) == 0) {
1717 * The page is not PG_ZERO'd so return it.
1719 vm_page_spin_unlock(m);
1723 * The page is PG_ZERO, requeue it and loop
1725 _vm_page_add_queue_spinlocked(m,
1728 vm_page_queue_spin_unlock(m);
1729 if (_vm_page_wakeup(m)) {
1730 vm_page_spin_unlock(m);
1733 vm_page_spin_unlock(m);
1738 qi = (qi + PQ_PRIME2) & PQ_L2_MASK;
1744 * vm_page_unmanage()
1746 * Prevent PV management from being done on the page. The page is
1747 * removed from the paging queues as if it were wired, and as a
1748 * consequence of no longer being managed the pageout daemon will not
1749 * touch it (since there is no way to locate the pte mappings for the
1750 * page). madvise() calls that mess with the pmap will also no longer
1751 * operate on the page.
1753 * Beyond that the page is still reasonably 'normal'. Freeing the page
1754 * will clear the flag.
1756 * This routine is used by OBJT_PHYS objects - objects using unswappable
1757 * physical memory as backing store rather then swap-backed memory and
1758 * will eventually be extended to support 4MB unmanaged physical
1761 * Caller must be holding the page busy.
1764 vm_page_unmanage(vm_page_t m)
1766 KKASSERT(m->flags & PG_BUSY);
1767 if ((m->flags & PG_UNMANAGED) == 0) {
1768 if (m->wire_count == 0)
1771 vm_page_flag_set(m, PG_UNMANAGED);
1775 * Mark this page as wired down by yet another map, removing it from
1776 * paging queues as necessary.
1778 * Caller must be holding the page busy.
1781 vm_page_wire(vm_page_t m)
1784 * Only bump the wire statistics if the page is not already wired,
1785 * and only unqueue the page if it is on some queue (if it is unmanaged
1786 * it is already off the queues). Don't do anything with fictitious
1787 * pages because they are always wired.
1789 KKASSERT(m->flags & PG_BUSY);
1790 if ((m->flags & PG_FICTITIOUS) == 0) {
1791 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) {
1792 if ((m->flags & PG_UNMANAGED) == 0)
1794 atomic_add_int(&vmstats.v_wire_count, 1);
1796 KASSERT(m->wire_count != 0,
1797 ("vm_page_wire: wire_count overflow m=%p", m));
1802 * Release one wiring of this page, potentially enabling it to be paged again.
1804 * Many pages placed on the inactive queue should actually go
1805 * into the cache, but it is difficult to figure out which. What
1806 * we do instead, if the inactive target is well met, is to put
1807 * clean pages at the head of the inactive queue instead of the tail.
1808 * This will cause them to be moved to the cache more quickly and
1809 * if not actively re-referenced, freed more quickly. If we just
1810 * stick these pages at the end of the inactive queue, heavy filesystem
1811 * meta-data accesses can cause an unnecessary paging load on memory bound
1812 * processes. This optimization causes one-time-use metadata to be
1813 * reused more quickly.
1815 * BUT, if we are in a low-memory situation we have no choice but to
1816 * put clean pages on the cache queue.
1818 * A number of routines use vm_page_unwire() to guarantee that the page
1819 * will go into either the inactive or active queues, and will NEVER
1820 * be placed in the cache - for example, just after dirtying a page.
1821 * dirty pages in the cache are not allowed.
1823 * The page queues must be locked.
1824 * This routine may not block.
1827 vm_page_unwire(vm_page_t m, int activate)
1829 KKASSERT(m->flags & PG_BUSY);
1830 if (m->flags & PG_FICTITIOUS) {
1832 } else if (m->wire_count <= 0) {
1833 panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
1835 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) {
1836 atomic_add_int(&vmstats.v_wire_count, -1);
1837 if (m->flags & PG_UNMANAGED) {
1839 } else if (activate) {
1840 vm_page_spin_lock(m);
1841 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE, 0);
1842 _vm_page_and_queue_spin_unlock(m);
1844 vm_page_spin_lock(m);
1845 vm_page_flag_clear(m, PG_WINATCFLS);
1846 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE,
1848 ++vm_swapcache_inactive_heuristic;
1849 _vm_page_and_queue_spin_unlock(m);
1856 * Move the specified page to the inactive queue. If the page has
1857 * any associated swap, the swap is deallocated.
1859 * Normally athead is 0 resulting in LRU operation. athead is set
1860 * to 1 if we want this page to be 'as if it were placed in the cache',
1861 * except without unmapping it from the process address space.
1863 * vm_page's spinlock must be held on entry and will remain held on return.
1864 * This routine may not block.
1867 _vm_page_deactivate_locked(vm_page_t m, int athead)
1872 * Ignore if already inactive.
1874 if (m->queue == PQ_INACTIVE)
1876 _vm_page_queue_spin_lock(m);
1877 oqueue = _vm_page_rem_queue_spinlocked(m);
1879 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1880 if (oqueue == PQ_CACHE)
1881 mycpu->gd_cnt.v_reactivated++;
1882 vm_page_flag_clear(m, PG_WINATCFLS);
1883 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE, athead);
1885 ++vm_swapcache_inactive_heuristic;
1887 _vm_page_queue_spin_unlock(m);
1888 /* leaves vm_page spinlocked */
1892 * Attempt to deactivate a page.
1897 vm_page_deactivate(vm_page_t m)
1899 vm_page_spin_lock(m);
1900 _vm_page_deactivate_locked(m, 0);
1901 vm_page_spin_unlock(m);
1905 vm_page_deactivate_locked(vm_page_t m)
1907 _vm_page_deactivate_locked(m, 0);
1911 * Attempt to move a page to PQ_CACHE.
1913 * Returns 0 on failure, 1 on success
1915 * The page should NOT be busied by the caller. This function will validate
1916 * whether the page can be safely moved to the cache.
1919 vm_page_try_to_cache(vm_page_t m)
1921 vm_page_spin_lock(m);
1922 if (vm_page_busy_try(m, TRUE)) {
1923 vm_page_spin_unlock(m);
1926 if (m->dirty || m->hold_count || m->wire_count ||
1927 (m->flags & PG_UNMANAGED)) {
1928 if (_vm_page_wakeup(m)) {
1929 vm_page_spin_unlock(m);
1932 vm_page_spin_unlock(m);
1936 vm_page_spin_unlock(m);
1939 * Page busied by us and no longer spinlocked. Dirty pages cannot
1940 * be moved to the cache.
1942 vm_page_test_dirty(m);
1952 * Attempt to free the page. If we cannot free it, we do nothing.
1953 * 1 is returned on success, 0 on failure.
1958 vm_page_try_to_free(vm_page_t m)
1960 vm_page_spin_lock(m);
1961 if (vm_page_busy_try(m, TRUE)) {
1962 vm_page_spin_unlock(m);
1965 if (m->dirty || m->hold_count || m->wire_count ||
1966 (m->flags & PG_UNMANAGED)) {
1967 if (_vm_page_wakeup(m)) {
1968 vm_page_spin_unlock(m);
1971 vm_page_spin_unlock(m);
1975 vm_page_spin_unlock(m);
1978 * Page busied by us and no longer spinlocked. Dirty pages will
1979 * not be freed by this function. We have to re-test the
1980 * dirty bit after cleaning out the pmaps.
1982 vm_page_test_dirty(m);
1987 vm_page_protect(m, VM_PROT_NONE);
1999 * Put the specified page onto the page cache queue (if appropriate).
2001 * The page must be busy, and this routine will release the busy and
2002 * possibly even free the page.
2005 vm_page_cache(vm_page_t m)
2007 if ((m->flags & PG_UNMANAGED) || m->busy ||
2008 m->wire_count || m->hold_count) {
2009 kprintf("vm_page_cache: attempting to cache busy/held page\n");
2015 * Already in the cache (and thus not mapped)
2017 if ((m->queue - m->pc) == PQ_CACHE) {
2018 KKASSERT((m->flags & PG_MAPPED) == 0);
2024 * Caller is required to test m->dirty, but note that the act of
2025 * removing the page from its maps can cause it to become dirty
2026 * on an SMP system due to another cpu running in usermode.
2029 panic("vm_page_cache: caching a dirty page, pindex: %ld",
2034 * Remove all pmaps and indicate that the page is not
2035 * writeable or mapped. Our vm_page_protect() call may
2036 * have blocked (especially w/ VM_PROT_NONE), so recheck
2039 vm_page_protect(m, VM_PROT_NONE);
2040 if ((m->flags & (PG_UNMANAGED|PG_MAPPED)) || m->busy ||
2041 m->wire_count || m->hold_count) {
2043 } else if (m->dirty) {
2044 vm_page_deactivate(m);
2047 _vm_page_and_queue_spin_lock(m);
2048 _vm_page_rem_queue_spinlocked(m);
2049 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0);
2050 _vm_page_queue_spin_unlock(m);
2051 if (_vm_page_wakeup(m)) {
2052 vm_page_spin_unlock(m);
2055 vm_page_spin_unlock(m);
2057 vm_page_free_wakeup();
2062 * vm_page_dontneed()
2064 * Cache, deactivate, or do nothing as appropriate. This routine
2065 * is typically used by madvise() MADV_DONTNEED.
2067 * Generally speaking we want to move the page into the cache so
2068 * it gets reused quickly. However, this can result in a silly syndrome
2069 * due to the page recycling too quickly. Small objects will not be
2070 * fully cached. On the otherhand, if we move the page to the inactive
2071 * queue we wind up with a problem whereby very large objects
2072 * unnecessarily blow away our inactive and cache queues.
2074 * The solution is to move the pages based on a fixed weighting. We
2075 * either leave them alone, deactivate them, or move them to the cache,
2076 * where moving them to the cache has the highest weighting.
2077 * By forcing some pages into other queues we eventually force the
2078 * system to balance the queues, potentially recovering other unrelated
2079 * space from active. The idea is to not force this to happen too
2082 * The page must be busied.
2085 vm_page_dontneed(vm_page_t m)
2087 static int dnweight;
2094 * occassionally leave the page alone
2096 if ((dnw & 0x01F0) == 0 ||
2097 m->queue == PQ_INACTIVE ||
2098 m->queue - m->pc == PQ_CACHE
2100 if (m->act_count >= ACT_INIT)
2106 * If vm_page_dontneed() is inactivating a page, it must clear
2107 * the referenced flag; otherwise the pagedaemon will see references
2108 * on the page in the inactive queue and reactivate it. Until the
2109 * page can move to the cache queue, madvise's job is not done.
2111 vm_page_flag_clear(m, PG_REFERENCED);
2112 pmap_clear_reference(m);
2115 vm_page_test_dirty(m);
2117 if (m->dirty || (dnw & 0x0070) == 0) {
2119 * Deactivate the page 3 times out of 32.
2124 * Cache the page 28 times out of every 32. Note that
2125 * the page is deactivated instead of cached, but placed
2126 * at the head of the queue instead of the tail.
2130 vm_page_spin_lock(m);
2131 _vm_page_deactivate_locked(m, head);
2132 vm_page_spin_unlock(m);
2136 * These routines manipulate the 'soft busy' count for a page. A soft busy
2137 * is almost like PG_BUSY except that it allows certain compatible operations
2138 * to occur on the page while it is busy. For example, a page undergoing a
2139 * write can still be mapped read-only.
2141 * Because vm_pages can overlap buffers m->busy can be > 1. m->busy is only
2142 * adjusted while the vm_page is PG_BUSY so the flash will occur when the
2143 * busy bit is cleared.
2146 vm_page_io_start(vm_page_t m)
2148 KASSERT(m->flags & PG_BUSY, ("vm_page_io_start: page not busy!!!"));
2149 atomic_add_char(&m->busy, 1);
2150 vm_page_flag_set(m, PG_SBUSY);
2154 vm_page_io_finish(vm_page_t m)
2156 KASSERT(m->flags & PG_BUSY, ("vm_page_io_finish: page not busy!!!"));
2157 atomic_subtract_char(&m->busy, 1);
2159 vm_page_flag_clear(m, PG_SBUSY);
2163 * Grab a page, blocking if it is busy and allocating a page if necessary.
2164 * A busy page is returned or NULL. The page may or may not be valid and
2165 * might not be on a queue (the caller is responsible for the disposition of
2168 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the
2169 * page will be zero'd and marked valid.
2171 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked
2172 * valid even if it already exists.
2174 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also
2175 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified.
2177 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
2178 * always returned if we had blocked.
2180 * This routine may not be called from an interrupt.
2182 * PG_ZERO is *ALWAYS* cleared by this routine.
2184 * No other requirements.
2187 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2192 KKASSERT(allocflags &
2193 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
2194 vm_object_hold(object);
2196 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error);
2198 vm_page_sleep_busy(m, TRUE, "pgrbwt");
2199 if ((allocflags & VM_ALLOC_RETRY) == 0) {
2204 } else if (m == NULL) {
2205 m = vm_page_alloc(object, pindex,
2206 allocflags & ~VM_ALLOC_RETRY);
2210 if ((allocflags & VM_ALLOC_RETRY) == 0)
2219 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid.
2221 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set
2222 * valid even if already valid.
2224 if (m->valid == 0) {
2225 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) {
2226 if ((m->flags & PG_ZERO) == 0)
2227 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2228 m->valid = VM_PAGE_BITS_ALL;
2230 } else if (allocflags & VM_ALLOC_FORCE_ZERO) {
2231 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2232 m->valid = VM_PAGE_BITS_ALL;
2234 vm_page_flag_clear(m, PG_ZERO);
2236 vm_object_drop(object);
2241 * Mapping function for valid bits or for dirty bits in
2242 * a page. May not block.
2244 * Inputs are required to range within a page.
2250 vm_page_bits(int base, int size)
2256 base + size <= PAGE_SIZE,
2257 ("vm_page_bits: illegal base/size %d/%d", base, size)
2260 if (size == 0) /* handle degenerate case */
2263 first_bit = base >> DEV_BSHIFT;
2264 last_bit = (base + size - 1) >> DEV_BSHIFT;
2266 return ((2 << last_bit) - (1 << first_bit));
2270 * Sets portions of a page valid and clean. The arguments are expected
2271 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2272 * of any partial chunks touched by the range. The invalid portion of
2273 * such chunks will be zero'd.
2275 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically
2276 * align base to DEV_BSIZE so as not to mark clean a partially
2277 * truncated device block. Otherwise the dirty page status might be
2280 * This routine may not block.
2282 * (base + size) must be less then or equal to PAGE_SIZE.
2285 _vm_page_zero_valid(vm_page_t m, int base, int size)
2290 if (size == 0) /* handle degenerate case */
2294 * If the base is not DEV_BSIZE aligned and the valid
2295 * bit is clear, we have to zero out a portion of the
2299 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2300 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
2302 pmap_zero_page_area(
2310 * If the ending offset is not DEV_BSIZE aligned and the
2311 * valid bit is clear, we have to zero out a portion of
2315 endoff = base + size;
2317 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2318 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
2320 pmap_zero_page_area(
2323 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
2329 * Set valid, clear dirty bits. If validating the entire
2330 * page we can safely clear the pmap modify bit. We also
2331 * use this opportunity to clear the PG_NOSYNC flag. If a process
2332 * takes a write fault on a MAP_NOSYNC memory area the flag will
2335 * We set valid bits inclusive of any overlap, but we can only
2336 * clear dirty bits for DEV_BSIZE chunks that are fully within
2339 * Page must be busied?
2340 * No other requirements.
2343 vm_page_set_valid(vm_page_t m, int base, int size)
2345 _vm_page_zero_valid(m, base, size);
2346 m->valid |= vm_page_bits(base, size);
2351 * Set valid bits and clear dirty bits.
2353 * NOTE: This function does not clear the pmap modified bit.
2354 * Also note that e.g. NFS may use a byte-granular base
2357 * WARNING: Page must be busied? But vfs_clean_one_page() will call
2358 * this without necessarily busying the page (via bdwrite()).
2359 * So for now vm_token must also be held.
2361 * No other requirements.
2364 vm_page_set_validclean(vm_page_t m, int base, int size)
2368 _vm_page_zero_valid(m, base, size);
2369 pagebits = vm_page_bits(base, size);
2370 m->valid |= pagebits;
2371 m->dirty &= ~pagebits;
2372 if (base == 0 && size == PAGE_SIZE) {
2373 /*pmap_clear_modify(m);*/
2374 vm_page_flag_clear(m, PG_NOSYNC);
2379 * Set valid & dirty. Used by buwrite()
2381 * WARNING: Page must be busied? But vfs_dirty_one_page() will
2382 * call this function in buwrite() so for now vm_token must
2385 * No other requirements.
2388 vm_page_set_validdirty(vm_page_t m, int base, int size)
2392 pagebits = vm_page_bits(base, size);
2393 m->valid |= pagebits;
2394 m->dirty |= pagebits;
2396 vm_object_set_writeable_dirty(m->object);
2402 * NOTE: This function does not clear the pmap modified bit.
2403 * Also note that e.g. NFS may use a byte-granular base
2406 * Page must be busied?
2407 * No other requirements.
2410 vm_page_clear_dirty(vm_page_t m, int base, int size)
2412 m->dirty &= ~vm_page_bits(base, size);
2413 if (base == 0 && size == PAGE_SIZE) {
2414 /*pmap_clear_modify(m);*/
2415 vm_page_flag_clear(m, PG_NOSYNC);
2420 * Make the page all-dirty.
2422 * Also make sure the related object and vnode reflect the fact that the
2423 * object may now contain a dirty page.
2425 * Page must be busied?
2426 * No other requirements.
2429 vm_page_dirty(vm_page_t m)
2432 int pqtype = m->queue - m->pc;
2434 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE,
2435 ("vm_page_dirty: page in free/cache queue!"));
2436 if (m->dirty != VM_PAGE_BITS_ALL) {
2437 m->dirty = VM_PAGE_BITS_ALL;
2439 vm_object_set_writeable_dirty(m->object);
2444 * Invalidates DEV_BSIZE'd chunks within a page. Both the
2445 * valid and dirty bits for the effected areas are cleared.
2447 * Page must be busied?
2449 * No other requirements.
2452 vm_page_set_invalid(vm_page_t m, int base, int size)
2456 bits = vm_page_bits(base, size);
2459 m->object->generation++;
2463 * The kernel assumes that the invalid portions of a page contain
2464 * garbage, but such pages can be mapped into memory by user code.
2465 * When this occurs, we must zero out the non-valid portions of the
2466 * page so user code sees what it expects.
2468 * Pages are most often semi-valid when the end of a file is mapped
2469 * into memory and the file's size is not page aligned.
2471 * Page must be busied?
2472 * No other requirements.
2475 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
2481 * Scan the valid bits looking for invalid sections that
2482 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
2483 * valid bit may be set ) have already been zerod by
2484 * vm_page_set_validclean().
2486 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
2487 if (i == (PAGE_SIZE / DEV_BSIZE) ||
2488 (m->valid & (1 << i))
2491 pmap_zero_page_area(
2494 (i - b) << DEV_BSHIFT
2502 * setvalid is TRUE when we can safely set the zero'd areas
2503 * as being valid. We can do this if there are no cache consistency
2504 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
2507 m->valid = VM_PAGE_BITS_ALL;
2511 * Is a (partial) page valid? Note that the case where size == 0
2512 * will return FALSE in the degenerate case where the page is entirely
2513 * invalid, and TRUE otherwise.
2516 * No other requirements.
2519 vm_page_is_valid(vm_page_t m, int base, int size)
2521 int bits = vm_page_bits(base, size);
2523 if (m->valid && ((m->valid & bits) == bits))
2530 * update dirty bits from pmap/mmu. May not block.
2532 * Caller must hold the page busy
2535 vm_page_test_dirty(vm_page_t m)
2537 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
2543 * Register an action, associating it with its vm_page
2546 vm_page_register_action(vm_page_action_t action, vm_page_event_t event)
2548 struct vm_page_action_list *list;
2551 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
2552 list = &action_list[hv];
2554 lwkt_gettoken(&vm_token);
2555 vm_page_flag_set(action->m, PG_ACTIONLIST);
2556 action->event = event;
2557 LIST_INSERT_HEAD(list, action, entry);
2558 lwkt_reltoken(&vm_token);
2562 * Unregister an action, disassociating it from its related vm_page
2565 vm_page_unregister_action(vm_page_action_t action)
2567 struct vm_page_action_list *list;
2570 lwkt_gettoken(&vm_token);
2571 if (action->event != VMEVENT_NONE) {
2572 action->event = VMEVENT_NONE;
2573 LIST_REMOVE(action, entry);
2575 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
2576 list = &action_list[hv];
2577 if (LIST_EMPTY(list))
2578 vm_page_flag_clear(action->m, PG_ACTIONLIST);
2580 lwkt_reltoken(&vm_token);
2584 * Issue an event on a VM page. Corresponding action structures are
2585 * removed from the page's list and called.
2587 * If the vm_page has no more pending action events we clear its
2588 * PG_ACTIONLIST flag.
2591 vm_page_event_internal(vm_page_t m, vm_page_event_t event)
2593 struct vm_page_action_list *list;
2594 struct vm_page_action *scan;
2595 struct vm_page_action *next;
2599 hv = (int)((intptr_t)m >> 8) & VMACTION_HMASK;
2600 list = &action_list[hv];
2603 lwkt_gettoken(&vm_token);
2604 LIST_FOREACH_MUTABLE(scan, list, entry, next) {
2606 if (scan->event == event) {
2607 scan->event = VMEVENT_NONE;
2608 LIST_REMOVE(scan, entry);
2609 scan->func(m, scan);
2617 vm_page_flag_clear(m, PG_ACTIONLIST);
2618 lwkt_reltoken(&vm_token);
2621 #include "opt_ddb.h"
2623 #include <sys/kernel.h>
2625 #include <ddb/ddb.h>
2627 DB_SHOW_COMMAND(page, vm_page_print_page_info)
2629 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count);
2630 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count);
2631 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count);
2632 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count);
2633 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count);
2634 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved);
2635 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min);
2636 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target);
2637 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min);
2638 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target);
2641 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
2644 db_printf("PQ_FREE:");
2645 for(i=0;i<PQ_L2_SIZE;i++) {
2646 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
2650 db_printf("PQ_CACHE:");
2651 for(i=0;i<PQ_L2_SIZE;i++) {
2652 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
2656 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
2657 vm_page_queues[PQ_ACTIVE].lcnt,
2658 vm_page_queues[PQ_INACTIVE].lcnt);