2 * Copyright (c) 1991 Regents of the University of California.
5 * This code is derived from software contributed to Berkeley by
6 * The Mach Operating System project at Carnegie-Mellon University.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 3. All advertising materials mentioning features or use of this software
17 * must display the following acknowledgement:
18 * This product includes software developed by the University of
19 * California, Berkeley and its contributors.
20 * 4. Neither the name of the University nor the names of its contributors
21 * may be used to endorse or promote products derived from this software
22 * without specific prior written permission.
24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91
41 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
42 * All rights reserved.
44 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
46 * Permission to use, copy, modify and distribute this software and
47 * its documentation is hereby granted, provided that both the copyright
48 * notice and this permission notice appear in all copies of the
49 * software, derivative works or modified versions, and any portions
50 * thereof, and that both notices appear in supporting documentation.
52 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
53 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
54 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
56 * Carnegie Mellon requests users of this software to return to
58 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
59 * School of Computer Science
60 * Carnegie Mellon University
61 * Pittsburgh PA 15213-3890
63 * any improvements or extensions that they make and grant Carnegie the
64 * rights to redistribute these changes.
68 * GENERAL RULES ON VM_PAGE MANIPULATION
70 * - a pageq mutex is required when adding or removing a page from a
71 * page queue (vm_page_queue[]), regardless of other mutexes or the
72 * busy state of a page.
74 * - a hash chain mutex is required when associating or disassociating
75 * a page from the VM PAGE CACHE hash table (vm_page_buckets),
76 * regardless of other mutexes or the busy state of a page.
78 * - either a hash chain mutex OR a busied page is required in order
79 * to modify the page flags. A hash chain mutex must be obtained in
80 * order to busy a page. A page's flags cannot be modified by a
81 * hash chain mutex if the page is marked busy.
83 * - The object memq mutex is held when inserting or removing
84 * pages from an object (vm_page_insert() or vm_page_remove()). This
85 * is different from the object's main mutex.
87 * Generally speaking, you have to be aware of side effects when running
88 * vm_page ops. A vm_page_lookup() will return with the hash chain
89 * locked, whether it was able to lookup the page or not. vm_page_free(),
90 * vm_page_cache(), vm_page_activate(), and a number of other routines
91 * will release the hash chain mutex for you. Intermediate manipulation
92 * routines such as vm_page_flag_set() expect the hash chain to be held
93 * on entry and the hash chain will remain held on return.
95 * pageq scanning can only occur with the pageq in question locked.
96 * We have a known bottleneck with the active queue, but the cache
97 * and free queues are actually arrays already.
101 * Resident memory management module.
104 #include <sys/param.h>
105 #include <sys/systm.h>
106 #include <sys/lock.h>
107 #include <sys/malloc.h>
108 #include <sys/mutex.h>
109 #include <sys/proc.h>
110 #include <sys/vmmeter.h>
111 #include <sys/vnode.h>
114 #include <vm/vm_param.h>
115 #include <vm/vm_kern.h>
116 #include <vm/vm_object.h>
117 #include <vm/vm_page.h>
118 #include <vm/vm_pageout.h>
119 #include <vm/vm_pager.h>
120 #include <vm/vm_extern.h>
122 #include <vm/uma_int.h>
125 * Associated with page of user-allocatable memory is a
128 static struct mtx vm_page_buckets_mtx;
129 static struct vm_page **vm_page_buckets; /* Array of buckets */
130 static int vm_page_bucket_count; /* How big is array? */
131 static int vm_page_hash_mask; /* Mask for hash function */
133 struct mtx vm_page_queue_mtx;
134 struct mtx vm_page_queue_free_mtx;
136 vm_page_t vm_page_array = 0;
137 int vm_page_array_size = 0;
139 int vm_page_zero_count = 0;
144 * Sets the page size, perhaps based upon the memory
145 * size. Must be called before any use of page-size
146 * dependent functions.
149 vm_set_page_size(void)
151 if (cnt.v_page_size == 0)
152 cnt.v_page_size = PAGE_SIZE;
153 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0)
154 panic("vm_set_page_size: page size not a power of two");
160 * Initializes the resident memory module.
162 * Allocates memory for the page cells, and
163 * for the object/offset-to-page hash table headers.
164 * Each page cell is initialized and placed on the free list.
167 vm_page_startup(vm_offset_t starta, vm_offset_t enda, vm_offset_t vaddr)
170 struct vm_page **bucket;
171 vm_size_t npages, page_range;
178 /* the biggest memory array is the second group of pages */
180 vm_offset_t biggestone, biggestsize;
189 vaddr = round_page(vaddr);
191 for (i = 0; phys_avail[i + 1]; i += 2) {
192 phys_avail[i] = round_page(phys_avail[i]);
193 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
196 for (i = 0; phys_avail[i + 1]; i += 2) {
197 vm_size_t size = phys_avail[i + 1] - phys_avail[i];
199 if (size > biggestsize) {
207 end = phys_avail[biggestone+1];
210 * Initialize the locks.
212 mtx_init(&vm_page_queue_mtx, "vm page queue mutex", NULL, MTX_DEF);
213 mtx_init(&vm_page_queue_free_mtx, "vm page queue free mutex", NULL,
217 * Initialize the queue headers for the free queue, the active queue
218 * and the inactive queue.
223 * Allocate memory for use when boot strapping the kernel memory allocator
225 bootpages = UMA_BOOT_PAGES * UMA_SLAB_SIZE;
226 new_end = end - bootpages;
227 new_end = trunc_page(new_end);
228 mapped = pmap_map(&vaddr, new_end, end,
229 VM_PROT_READ | VM_PROT_WRITE);
230 bzero((caddr_t) mapped, end - new_end);
231 uma_startup((caddr_t)mapped);
236 * Allocate (and initialize) the hash table buckets.
238 * The number of buckets MUST BE a power of 2, and the actual value is
239 * the next power of 2 greater than the number of physical pages in
242 * We make the hash table approximately 2x the number of pages to
243 * reduce the chain length. This is about the same size using the
244 * singly-linked list as the 1x hash table we were using before
245 * using TAILQ but the chain length will be smaller.
247 * Note: This computation can be tweaked if desired.
249 if (vm_page_bucket_count == 0) {
250 vm_page_bucket_count = 1;
251 while (vm_page_bucket_count < atop(total))
252 vm_page_bucket_count <<= 1;
254 vm_page_bucket_count <<= 1;
255 vm_page_hash_mask = vm_page_bucket_count - 1;
258 * Validate these addresses.
260 new_end = end - vm_page_bucket_count * sizeof(struct vm_page *);
261 new_end = trunc_page(new_end);
262 mapped = pmap_map(&vaddr, new_end, end,
263 VM_PROT_READ | VM_PROT_WRITE);
264 bzero((caddr_t) mapped, end - new_end);
266 mtx_init(&vm_page_buckets_mtx, "vm page buckets mutex", NULL, MTX_SPIN);
267 vm_page_buckets = (struct vm_page **)mapped;
268 bucket = vm_page_buckets;
269 for (i = 0; i < vm_page_bucket_count; i++) {
275 * Compute the number of pages of memory that will be available for
276 * use (taking into account the overhead of a page structure per
279 first_page = phys_avail[0] / PAGE_SIZE;
280 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
281 npages = (total - (page_range * sizeof(struct vm_page)) -
282 (end - new_end)) / PAGE_SIZE;
286 * Initialize the mem entry structures now, and put them in the free
289 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
290 mapped = pmap_map(&vaddr, new_end, end,
291 VM_PROT_READ | VM_PROT_WRITE);
292 vm_page_array = (vm_page_t) mapped;
295 * Clear all of the page structures
297 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
298 vm_page_array_size = page_range;
301 * Construct the free queue(s) in descending order (by physical
302 * address) so that the first 16MB of physical memory is allocated
303 * last rather than first. On large-memory machines, this avoids
304 * the exhaustion of low physical memory before isa_dmainit has run.
306 cnt.v_page_count = 0;
307 cnt.v_free_count = 0;
308 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
313 last_pa = phys_avail[i + 1];
314 while (pa < last_pa && npages-- > 0) {
315 vm_pageq_add_new_page(pa);
325 * Distributes the object/offset key pair among hash buckets.
327 * NOTE: This macro depends on vm_page_bucket_count being a power of 2.
328 * This routine may not block.
330 * We try to randomize the hash based on the object to spread the pages
331 * out in the hash table without it costing us too much.
334 vm_page_hash(vm_object_t object, vm_pindex_t pindex)
336 int i = ((uintptr_t)object + pindex) ^ object->hash_rand;
338 return (i & vm_page_hash_mask);
342 vm_page_flag_set(vm_page_t m, unsigned short bits)
349 vm_page_flag_clear(vm_page_t m, unsigned short bits)
356 vm_page_busy(vm_page_t m)
358 KASSERT((m->flags & PG_BUSY) == 0,
359 ("vm_page_busy: page already busy!!!"));
360 vm_page_flag_set(m, PG_BUSY);
366 * wakeup anyone waiting for the page.
369 vm_page_flash(vm_page_t m)
371 if (m->flags & PG_WANTED) {
372 vm_page_flag_clear(m, PG_WANTED);
380 * clear the PG_BUSY flag and wakeup anyone waiting for the
385 vm_page_wakeup(vm_page_t m)
387 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!"));
388 vm_page_flag_clear(m, PG_BUSY);
397 vm_page_io_start(vm_page_t m)
400 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
405 vm_page_io_finish(vm_page_t m)
414 * Keep page from being freed by the page daemon
415 * much of the same effect as wiring, except much lower
416 * overhead and should be used only for *very* temporary
417 * holding ("wiring").
420 vm_page_hold(vm_page_t mem)
427 vm_page_unhold(vm_page_t mem)
431 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
432 if (mem->hold_count == 0 && mem->queue == PQ_HOLD)
433 vm_page_free_toq(mem);
439 * Reduce the protection of a page. This routine never raises the
440 * protection and therefore can be safely called if the page is already
441 * at VM_PROT_NONE (it will be a NOP effectively ).
444 vm_page_protect(vm_page_t mem, int prot)
446 if (prot == VM_PROT_NONE) {
447 if (mem->flags & (PG_WRITEABLE|PG_MAPPED)) {
448 pmap_page_protect(mem, VM_PROT_NONE);
449 vm_page_flag_clear(mem, PG_WRITEABLE|PG_MAPPED);
451 } else if ((prot == VM_PROT_READ) && (mem->flags & PG_WRITEABLE)) {
452 pmap_page_protect(mem, VM_PROT_READ);
453 vm_page_flag_clear(mem, PG_WRITEABLE);
459 * Zero-fill the specified page.
460 * Written as a standard pagein routine, to
461 * be used by the zero-fill object.
464 vm_page_zero_fill(vm_page_t m)
471 * vm_page_zero_fill_area:
473 * Like vm_page_zero_fill but only fill the specified area.
476 vm_page_zero_fill_area(vm_page_t m, int off, int size)
478 pmap_zero_page_area(m, off, size);
485 * Copy one page to another
488 vm_page_copy(vm_page_t src_m, vm_page_t dest_m)
490 pmap_copy_page(src_m, dest_m);
491 dest_m->valid = VM_PAGE_BITS_ALL;
499 * The clearing of PG_ZERO is a temporary safety until the code can be
500 * reviewed to determine that PG_ZERO is being properly cleared on
501 * write faults or maps. PG_ZERO was previously cleared in
505 vm_page_free(vm_page_t m)
507 vm_page_flag_clear(m, PG_ZERO);
509 vm_page_zero_idle_wakeup();
515 * Free a page to the zerod-pages queue
518 vm_page_free_zero(vm_page_t m)
520 vm_page_flag_set(m, PG_ZERO);
525 * vm_page_sleep_busy:
527 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE)
528 * m->busy is zero. Returns TRUE if it had to sleep ( including if
529 * it almost had to sleep and made temporary spl*() mods), FALSE
532 * This routine assumes that interrupts can only remove the busy
533 * status from a page, not set the busy status or change it from
534 * PG_BUSY to m->busy or vise versa (which would create a timing
538 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg)
541 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) {
543 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) {
545 * Page is busy. Wait and retry.
547 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
548 tsleep(m, PVM, msg, 0);
558 * vm_page_sleep_if_busy:
560 * Sleep and release the page queues lock if PG_BUSY is set or,
561 * if also_m_busy is TRUE, busy is non-zero. Returns TRUE if the
562 * thread slept and the page queues lock was released.
563 * Otherwise, retains the page queues lock and returns FALSE.
566 vm_page_sleep_if_busy(vm_page_t m, int also_m_busy, const char *msg)
569 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
570 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) {
571 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
572 msleep(m, &vm_page_queue_mtx, PDROP | PVM, msg, 0);
581 * make page all dirty
584 vm_page_dirty(vm_page_t m)
586 KASSERT(m->queue - m->pc != PQ_CACHE,
587 ("vm_page_dirty: page in cache!"));
588 m->dirty = VM_PAGE_BITS_ALL;
594 * Set page to not be dirty. Note: does not clear pmap modify bits
597 vm_page_undirty(vm_page_t m)
603 * vm_page_insert: [ internal use only ]
605 * Inserts the given mem entry into the object and object list.
607 * The pagetables are not updated but will presumably fault the page
608 * in if necessary, or if a kernel page the caller will at some point
609 * enter the page into the kernel's pmap. We are not allowed to block
610 * here so we *can't* do this anyway.
612 * The object and page must be locked, and must be splhigh.
613 * This routine may not block.
616 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
618 struct vm_page **bucket;
622 if (m->object != NULL)
623 panic("vm_page_insert: already inserted");
626 * Record the object/offset pair in this page
632 * Insert it into the object_object/offset hash table
634 bucket = &vm_page_buckets[vm_page_hash(object, pindex)];
635 mtx_lock_spin(&vm_page_buckets_mtx);
638 mtx_unlock_spin(&vm_page_buckets_mtx);
641 * Now link into the object's list of backed pages.
643 TAILQ_INSERT_TAIL(&object->memq, m, listq);
644 object->generation++;
647 * show that the object has one more resident page.
649 object->resident_page_count++;
652 * Since we are inserting a new and possibly dirty page,
653 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
655 if (m->flags & PG_WRITEABLE)
656 vm_object_set_writeable_dirty(object);
661 * NOTE: used by device pager as well -wfj
663 * Removes the given mem entry from the object/offset-page
664 * table and the object page list, but do not invalidate/terminate
667 * The object and page must be locked, and at splhigh.
668 * The underlying pmap entry (if any) is NOT removed here.
669 * This routine may not block.
672 vm_page_remove(vm_page_t m)
679 if (m->object == NULL)
682 if ((m->flags & PG_BUSY) == 0) {
683 panic("vm_page_remove: page not busy");
687 * Basically destroy the page.
694 * Remove from the object_object/offset hash table. The object
695 * must be on the hash queue, we will panic if it isn't
697 bucket = &vm_page_buckets[vm_page_hash(m->object, m->pindex)];
698 mtx_lock_spin(&vm_page_buckets_mtx);
699 while (*bucket != m) {
701 panic("vm_page_remove(): page not found in hash");
702 bucket = &(*bucket)->hnext;
706 mtx_unlock_spin(&vm_page_buckets_mtx);
709 * Now remove from the object's list of backed pages.
711 TAILQ_REMOVE(&object->memq, m, listq);
714 * And show that the object has one fewer resident page.
716 object->resident_page_count--;
717 object->generation++;
725 * Returns the page associated with the object/offset
726 * pair specified; if none is found, NULL is returned.
728 * The object must be locked. No side effects.
729 * This routine may not block.
730 * This is a critical path routine
733 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
736 struct vm_page **bucket;
739 * Search the hash table for this object/offset pair
741 bucket = &vm_page_buckets[vm_page_hash(object, pindex)];
742 mtx_lock_spin(&vm_page_buckets_mtx);
743 for (m = *bucket; m != NULL; m = m->hnext)
744 if (m->object == object && m->pindex == pindex)
746 mtx_unlock_spin(&vm_page_buckets_mtx);
753 * Move the given memory entry from its
754 * current object to the specified target object/offset.
756 * The object must be locked.
757 * This routine may not block.
759 * Note: this routine will raise itself to splvm(), the caller need not.
761 * Note: swap associated with the page must be invalidated by the move. We
762 * have to do this for several reasons: (1) we aren't freeing the
763 * page, (2) we are dirtying the page, (3) the VM system is probably
764 * moving the page from object A to B, and will then later move
765 * the backing store from A to B and we can't have a conflict.
767 * Note: we *always* dirty the page. It is necessary both for the
768 * fact that we moved it, and because we may be invalidating
769 * swap. If the page is on the cache, we have to deactivate it
770 * or vm_page_dirty() will panic. Dirty pages are not allowed
774 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
779 vm_page_lock_queues();
781 vm_page_insert(m, new_object, new_pindex);
782 if (m->queue - m->pc == PQ_CACHE)
783 vm_page_deactivate(m);
785 vm_page_unlock_queues();
790 * vm_page_select_cache:
792 * Find a page on the cache queue with color optimization. As pages
793 * might be found, but not applicable, they are deactivated. This
794 * keeps us from using potentially busy cached pages.
796 * This routine must be called at splvm().
797 * This routine may not block.
800 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex)
804 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
808 (pindex + object->pg_color) & PQ_L2_MASK,
811 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
812 m->hold_count || m->wire_count)) {
813 vm_page_deactivate(m);
821 * vm_page_select_free:
823 * Find a free or zero page, with specified preference.
825 * This routine must be called at splvm().
826 * This routine may not block.
828 static __inline vm_page_t
829 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero)
835 (pindex + object->pg_color) & PQ_L2_MASK,
844 * Allocate and return a memory cell associated
845 * with this VM object/offset pair.
848 * VM_ALLOC_NORMAL normal process request
849 * VM_ALLOC_SYSTEM system *really* needs a page
850 * VM_ALLOC_INTERRUPT interrupt time request
851 * VM_ALLOC_ZERO zero page
853 * This routine may not block.
855 * Additional special handling is required when called from an
856 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with
857 * the page cache in this case.
860 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
867 KASSERT(!vm_page_lookup(object, pindex),
868 ("vm_page_alloc: page already allocated"));
870 page_req = req & VM_ALLOC_CLASS_MASK;
873 * The pager is allowed to eat deeper into the free page list.
875 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) {
876 page_req = VM_ALLOC_SYSTEM;
881 mtx_lock_spin(&vm_page_queue_free_mtx);
882 if (cnt.v_free_count > cnt.v_free_reserved) {
884 * Allocate from the free queue if there are plenty of pages
887 m = vm_page_select_free(object, pindex,
888 (req & VM_ALLOC_ZERO) != 0);
890 (page_req == VM_ALLOC_SYSTEM &&
891 cnt.v_cache_count == 0 &&
892 cnt.v_free_count > cnt.v_interrupt_free_min) ||
893 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0)
896 * Interrupt or system, dig deeper into the free list.
898 m = vm_page_select_free(object, pindex, FALSE);
899 } else if (page_req != VM_ALLOC_INTERRUPT) {
900 mtx_unlock_spin(&vm_page_queue_free_mtx);
902 * Allocatable from cache (non-interrupt only). On success,
903 * we must free the page and try again, thus ensuring that
904 * cnt.v_*_free_min counters are replenished.
906 vm_page_lock_queues();
907 if ((m = vm_page_select_cache(object, pindex)) == NULL) {
908 vm_page_unlock_queues();
910 #if defined(DIAGNOSTIC)
911 if (cnt.v_cache_count > 0)
912 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count);
914 vm_pageout_deficit++;
918 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m));
920 vm_page_protect(m, VM_PROT_NONE);
922 vm_page_unlock_queues();
926 * Not allocatable from cache from interrupt, give up.
928 mtx_unlock_spin(&vm_page_queue_free_mtx);
930 vm_pageout_deficit++;
936 * At this point we had better have found a good page.
941 ("vm_page_alloc(): missing page on free queue\n")
945 * Remove from free queue
948 vm_pageq_remove_nowakeup(m);
951 * Initialize structure. Only the PG_ZERO flag is inherited.
953 if (m->flags & PG_ZERO) {
954 vm_page_zero_count--;
955 m->flags = PG_ZERO | PG_BUSY;
959 if (req & VM_ALLOC_WIRED) {
961 m->flags |= PG_MAPPED; /* XXX this does not belong here */
969 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m));
970 mtx_unlock_spin(&vm_page_queue_free_mtx);
973 * vm_page_insert() is safe prior to the splx(). Note also that
974 * inserting a page here does not insert it into the pmap (which
975 * could cause us to block allocating memory). We cannot block
978 vm_page_insert(m, object, pindex);
981 * Don't wakeup too often - wakeup the pageout daemon when
982 * we would be nearly out of memory.
984 if (vm_paging_needed())
992 * vm_wait: (also see VM_WAIT macro)
994 * Block until free pages are available for allocation
995 * - Called in various places before memory allocations.
1003 if (curproc == pageproc) {
1004 vm_pageout_pages_needed = 1;
1005 tsleep(&vm_pageout_pages_needed, PSWP, "VMWait", 0);
1007 if (!vm_pages_needed) {
1008 vm_pages_needed = 1;
1009 wakeup(&vm_pages_needed);
1011 tsleep(&cnt.v_free_count, PVM, "vmwait", 0);
1017 * vm_waitpfault: (also see VM_WAITPFAULT macro)
1019 * Block until free pages are available for allocation
1020 * - Called only in vm_fault so that processes page faulting
1021 * can be easily tracked.
1022 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
1023 * processes will be able to grab memory first. Do not change
1024 * this balance without careful testing first.
1032 if (!vm_pages_needed) {
1033 vm_pages_needed = 1;
1034 wakeup(&vm_pages_needed);
1036 tsleep(&cnt.v_free_count, PUSER, "pfault", 0);
1043 * Put the specified page on the active list (if appropriate).
1044 * Ensure that act_count is at least ACT_INIT but do not otherwise
1047 * The page queues must be locked.
1048 * This routine may not block.
1051 vm_page_activate(vm_page_t m)
1057 if (m->queue != PQ_ACTIVE) {
1058 if ((m->queue - m->pc) == PQ_CACHE)
1059 cnt.v_reactivated++;
1061 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1062 if (m->act_count < ACT_INIT)
1063 m->act_count = ACT_INIT;
1064 vm_pageq_enqueue(PQ_ACTIVE, m);
1067 if (m->act_count < ACT_INIT)
1068 m->act_count = ACT_INIT;
1074 * vm_page_free_wakeup:
1076 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
1077 * routine is called when a page has been added to the cache or free
1080 * This routine may not block.
1081 * This routine must be called at splvm()
1083 static __inline void
1084 vm_page_free_wakeup(void)
1087 * if pageout daemon needs pages, then tell it that there are
1090 if (vm_pageout_pages_needed &&
1091 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) {
1092 wakeup(&vm_pageout_pages_needed);
1093 vm_pageout_pages_needed = 0;
1096 * wakeup processes that are waiting on memory if we hit a
1097 * high water mark. And wakeup scheduler process if we have
1098 * lots of memory. this process will swapin processes.
1100 if (vm_pages_needed && !vm_page_count_min()) {
1101 vm_pages_needed = 0;
1102 wakeup(&cnt.v_free_count);
1109 * Returns the given page to the PQ_FREE list,
1110 * disassociating it with any VM object.
1112 * Object and page must be locked prior to entry.
1113 * This routine may not block.
1117 vm_page_free_toq(vm_page_t m)
1120 struct vpgqueues *pq;
1121 vm_object_t object = m->object;
1127 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
1129 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
1130 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
1132 if ((m->queue - m->pc) == PQ_FREE)
1133 panic("vm_page_free: freeing free page");
1135 panic("vm_page_free: freeing busy page");
1139 * unqueue, then remove page. Note that we cannot destroy
1140 * the page here because we do not want to call the pager's
1141 * callback routine until after we've put the page on the
1142 * appropriate free queue.
1144 vm_pageq_remove_nowakeup(m);
1148 * If fictitious remove object association and
1149 * return, otherwise delay object association removal.
1151 if ((m->flags & PG_FICTITIOUS) != 0) {
1159 if (m->wire_count != 0) {
1160 if (m->wire_count > 1) {
1161 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1162 m->wire_count, (long)m->pindex);
1164 panic("vm_page_free: freeing wired page\n");
1168 * If we've exhausted the object's resident pages we want to free
1172 (object->type == OBJT_VNODE) &&
1173 ((object->flags & OBJ_DEAD) == 0)
1175 struct vnode *vp = (struct vnode *)object->handle;
1177 if (vp && VSHOULDFREE(vp))
1182 * Clear the UNMANAGED flag when freeing an unmanaged page.
1184 if (m->flags & PG_UNMANAGED) {
1185 m->flags &= ~PG_UNMANAGED;
1188 pmap_page_is_free(m);
1192 if (m->hold_count != 0) {
1193 m->flags &= ~PG_ZERO;
1196 m->queue = PQ_FREE + m->pc;
1197 pq = &vm_page_queues[m->queue];
1198 mtx_lock_spin(&vm_page_queue_free_mtx);
1203 * Put zero'd pages on the end ( where we look for zero'd pages
1204 * first ) and non-zerod pages at the head.
1206 if (m->flags & PG_ZERO) {
1207 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
1208 ++vm_page_zero_count;
1210 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1212 mtx_unlock_spin(&vm_page_queue_free_mtx);
1213 vm_page_free_wakeup();
1220 * Prevent PV management from being done on the page. The page is
1221 * removed from the paging queues as if it were wired, and as a
1222 * consequence of no longer being managed the pageout daemon will not
1223 * touch it (since there is no way to locate the pte mappings for the
1224 * page). madvise() calls that mess with the pmap will also no longer
1225 * operate on the page.
1227 * Beyond that the page is still reasonably 'normal'. Freeing the page
1228 * will clear the flag.
1230 * This routine is used by OBJT_PHYS objects - objects using unswappable
1231 * physical memory as backing store rather then swap-backed memory and
1232 * will eventually be extended to support 4MB unmanaged physical
1236 vm_page_unmanage(vm_page_t m)
1241 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1242 if ((m->flags & PG_UNMANAGED) == 0) {
1243 if (m->wire_count == 0)
1246 vm_page_flag_set(m, PG_UNMANAGED);
1253 * Mark this page as wired down by yet
1254 * another map, removing it from paging queues
1257 * The page queues must be locked.
1258 * This routine may not block.
1261 vm_page_wire(vm_page_t m)
1266 * Only bump the wire statistics if the page is not already wired,
1267 * and only unqueue the page if it is on some queue (if it is unmanaged
1268 * it is already off the queues).
1271 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1272 if (m->wire_count == 0) {
1273 if ((m->flags & PG_UNMANAGED) == 0)
1278 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
1280 vm_page_flag_set(m, PG_MAPPED); /* XXX this does not belong here */
1286 * Release one wiring of this page, potentially
1287 * enabling it to be paged again.
1289 * Many pages placed on the inactive queue should actually go
1290 * into the cache, but it is difficult to figure out which. What
1291 * we do instead, if the inactive target is well met, is to put
1292 * clean pages at the head of the inactive queue instead of the tail.
1293 * This will cause them to be moved to the cache more quickly and
1294 * if not actively re-referenced, freed more quickly. If we just
1295 * stick these pages at the end of the inactive queue, heavy filesystem
1296 * meta-data accesses can cause an unnecessary paging load on memory bound
1297 * processes. This optimization causes one-time-use metadata to be
1298 * reused more quickly.
1300 * BUT, if we are in a low-memory situation we have no choice but to
1301 * put clean pages on the cache queue.
1303 * A number of routines use vm_page_unwire() to guarantee that the page
1304 * will go into either the inactive or active queues, and will NEVER
1305 * be placed in the cache - for example, just after dirtying a page.
1306 * dirty pages in the cache are not allowed.
1308 * The page queues must be locked.
1309 * This routine may not block.
1312 vm_page_unwire(vm_page_t m, int activate)
1317 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1318 if (m->wire_count > 0) {
1320 if (m->wire_count == 0) {
1322 if (m->flags & PG_UNMANAGED) {
1324 } else if (activate)
1325 vm_pageq_enqueue(PQ_ACTIVE, m);
1327 vm_page_flag_clear(m, PG_WINATCFLS);
1328 vm_pageq_enqueue(PQ_INACTIVE, m);
1332 panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count);
1339 * Move the specified page to the inactive queue. If the page has
1340 * any associated swap, the swap is deallocated.
1342 * Normally athead is 0 resulting in LRU operation. athead is set
1343 * to 1 if we want this page to be 'as if it were placed in the cache',
1344 * except without unmapping it from the process address space.
1346 * This routine may not block.
1348 static __inline void
1349 _vm_page_deactivate(vm_page_t m, int athead)
1355 * Ignore if already inactive.
1357 if (m->queue == PQ_INACTIVE)
1361 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1362 if ((m->queue - m->pc) == PQ_CACHE)
1363 cnt.v_reactivated++;
1364 vm_page_flag_clear(m, PG_WINATCFLS);
1367 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1369 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1370 m->queue = PQ_INACTIVE;
1371 vm_page_queues[PQ_INACTIVE].lcnt++;
1372 cnt.v_inactive_count++;
1378 vm_page_deactivate(vm_page_t m)
1380 _vm_page_deactivate(m, 0);
1384 * vm_page_try_to_cache:
1386 * Returns 0 on failure, 1 on success
1389 vm_page_try_to_cache(vm_page_t m)
1392 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1393 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1394 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1397 vm_page_test_dirty(m);
1405 * vm_page_try_to_free()
1407 * Attempt to free the page. If we cannot free it, we do nothing.
1408 * 1 is returned on success, 0 on failure.
1411 vm_page_try_to_free(vm_page_t m)
1414 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1415 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1416 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1419 vm_page_test_dirty(m);
1423 vm_page_protect(m, VM_PROT_NONE);
1431 * Put the specified page onto the page cache queue (if appropriate).
1433 * This routine may not block.
1436 vm_page_cache(vm_page_t m)
1440 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1441 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || m->wire_count) {
1442 printf("vm_page_cache: attempting to cache busy page\n");
1445 if ((m->queue - m->pc) == PQ_CACHE)
1449 * Remove all pmaps and indicate that the page is not
1450 * writeable or mapped.
1452 vm_page_protect(m, VM_PROT_NONE);
1453 if (m->dirty != 0) {
1454 panic("vm_page_cache: caching a dirty page, pindex: %ld",
1458 vm_pageq_remove_nowakeup(m);
1459 vm_pageq_enqueue(PQ_CACHE + m->pc, m);
1460 vm_page_free_wakeup();
1467 * Cache, deactivate, or do nothing as appropriate. This routine
1468 * is typically used by madvise() MADV_DONTNEED.
1470 * Generally speaking we want to move the page into the cache so
1471 * it gets reused quickly. However, this can result in a silly syndrome
1472 * due to the page recycling too quickly. Small objects will not be
1473 * fully cached. On the otherhand, if we move the page to the inactive
1474 * queue we wind up with a problem whereby very large objects
1475 * unnecessarily blow away our inactive and cache queues.
1477 * The solution is to move the pages based on a fixed weighting. We
1478 * either leave them alone, deactivate them, or move them to the cache,
1479 * where moving them to the cache has the highest weighting.
1480 * By forcing some pages into other queues we eventually force the
1481 * system to balance the queues, potentially recovering other unrelated
1482 * space from active. The idea is to not force this to happen too
1486 vm_page_dontneed(vm_page_t m)
1488 static int dnweight;
1492 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1496 * occassionally leave the page alone
1498 if ((dnw & 0x01F0) == 0 ||
1499 m->queue == PQ_INACTIVE ||
1500 m->queue - m->pc == PQ_CACHE
1502 if (m->act_count >= ACT_INIT)
1508 vm_page_test_dirty(m);
1510 if (m->dirty || (dnw & 0x0070) == 0) {
1512 * Deactivate the page 3 times out of 32.
1517 * Cache the page 28 times out of every 32. Note that
1518 * the page is deactivated instead of cached, but placed
1519 * at the head of the queue instead of the tail.
1523 _vm_page_deactivate(m, head);
1527 * Grab a page, waiting until we are waken up due to the page
1528 * changing state. We keep on waiting, if the page continues
1529 * to be in the object. If the page doesn't exist, allocate it.
1531 * This routine may block.
1534 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
1541 if ((m = vm_page_lookup(object, pindex)) != NULL) {
1542 if (m->busy || (m->flags & PG_BUSY)) {
1543 generation = object->generation;
1546 while ((object->generation == generation) &&
1547 (m->busy || (m->flags & PG_BUSY))) {
1548 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
1549 tsleep(m, PVM, "pgrbwt", 0);
1550 if ((allocflags & VM_ALLOC_RETRY) == 0) {
1558 vm_page_lock_queues();
1559 if (allocflags & VM_ALLOC_WIRED)
1562 vm_page_unlock_queues();
1567 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1570 if ((allocflags & VM_ALLOC_RETRY) == 0)
1579 * Mapping function for valid bits or for dirty bits in
1580 * a page. May not block.
1582 * Inputs are required to range within a page.
1585 vm_page_bits(int base, int size)
1591 base + size <= PAGE_SIZE,
1592 ("vm_page_bits: illegal base/size %d/%d", base, size)
1595 if (size == 0) /* handle degenerate case */
1598 first_bit = base >> DEV_BSHIFT;
1599 last_bit = (base + size - 1) >> DEV_BSHIFT;
1601 return ((2 << last_bit) - (1 << first_bit));
1605 * vm_page_set_validclean:
1607 * Sets portions of a page valid and clean. The arguments are expected
1608 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1609 * of any partial chunks touched by the range. The invalid portion of
1610 * such chunks will be zero'd.
1612 * This routine may not block.
1614 * (base + size) must be less then or equal to PAGE_SIZE.
1617 vm_page_set_validclean(vm_page_t m, int base, int size)
1624 if (size == 0) /* handle degenerate case */
1628 * If the base is not DEV_BSIZE aligned and the valid
1629 * bit is clear, we have to zero out a portion of the
1632 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
1633 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
1634 pmap_zero_page_area(m, frag, base - frag);
1637 * If the ending offset is not DEV_BSIZE aligned and the
1638 * valid bit is clear, we have to zero out a portion of
1641 endoff = base + size;
1642 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
1643 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
1644 pmap_zero_page_area(m, endoff,
1645 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
1648 * Set valid, clear dirty bits. If validating the entire
1649 * page we can safely clear the pmap modify bit. We also
1650 * use this opportunity to clear the PG_NOSYNC flag. If a process
1651 * takes a write fault on a MAP_NOSYNC memory area the flag will
1654 * We set valid bits inclusive of any overlap, but we can only
1655 * clear dirty bits for DEV_BSIZE chunks that are fully within
1658 pagebits = vm_page_bits(base, size);
1659 m->valid |= pagebits;
1661 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
1662 frag = DEV_BSIZE - frag;
1668 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
1670 m->dirty &= ~pagebits;
1671 if (base == 0 && size == PAGE_SIZE) {
1672 pmap_clear_modify(m);
1673 vm_page_flag_clear(m, PG_NOSYNC);
1680 vm_page_set_dirty(vm_page_t m, int base, int size)
1682 m->dirty |= vm_page_bits(base, size);
1688 vm_page_clear_dirty(vm_page_t m, int base, int size)
1691 m->dirty &= ~vm_page_bits(base, size);
1695 * vm_page_set_invalid:
1697 * Invalidates DEV_BSIZE'd chunks within a page. Both the
1698 * valid and dirty bits for the effected areas are cleared.
1703 vm_page_set_invalid(vm_page_t m, int base, int size)
1708 bits = vm_page_bits(base, size);
1711 m->object->generation++;
1715 * vm_page_zero_invalid()
1717 * The kernel assumes that the invalid portions of a page contain
1718 * garbage, but such pages can be mapped into memory by user code.
1719 * When this occurs, we must zero out the non-valid portions of the
1720 * page so user code sees what it expects.
1722 * Pages are most often semi-valid when the end of a file is mapped
1723 * into memory and the file's size is not page aligned.
1726 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
1732 * Scan the valid bits looking for invalid sections that
1733 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
1734 * valid bit may be set ) have already been zerod by
1735 * vm_page_set_validclean().
1737 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
1738 if (i == (PAGE_SIZE / DEV_BSIZE) ||
1739 (m->valid & (1 << i))
1742 pmap_zero_page_area(m,
1743 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
1750 * setvalid is TRUE when we can safely set the zero'd areas
1751 * as being valid. We can do this if there are no cache consistancy
1752 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
1755 m->valid = VM_PAGE_BITS_ALL;
1761 * Is (partial) page valid? Note that the case where size == 0
1762 * will return FALSE in the degenerate case where the page is
1763 * entirely invalid, and TRUE otherwise.
1768 vm_page_is_valid(vm_page_t m, int base, int size)
1770 int bits = vm_page_bits(base, size);
1772 if (m->valid && ((m->valid & bits) == bits))
1779 * update dirty bits from pmap/mmu. May not block.
1782 vm_page_test_dirty(vm_page_t m)
1784 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
1789 int so_zerocp_fullpage = 0;
1792 vm_page_cowfault(vm_page_t m)
1804 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL);
1806 vm_page_insert(m, object, pindex);
1813 * check to see if we raced with an xmit complete when
1814 * waiting to allocate a page. If so, put things back
1819 vm_page_insert(m, object, pindex);
1820 } else { /* clear COW & copy page */
1821 if (so_zerocp_fullpage) {
1822 mnew->valid = VM_PAGE_BITS_ALL;
1824 vm_page_copy(m, mnew);
1826 vm_page_dirty(mnew);
1827 vm_page_flag_clear(mnew, PG_BUSY);
1832 vm_page_cowclear(vm_page_t m)
1835 /* XXX KDM find out if giant is required here. */
1838 atomic_subtract_int(&m->cow, 1);
1840 * let vm_fault add back write permission lazily
1844 * sf_buf_free() will free the page, so we needn't do it here
1849 vm_page_cowsetup(vm_page_t m)
1851 /* XXX KDM find out if giant is required here */
1853 atomic_add_int(&m->cow, 1);
1854 vm_page_protect(m, VM_PROT_READ);
1857 #include "opt_ddb.h"
1859 #include <sys/kernel.h>
1861 #include <ddb/ddb.h>
1863 DB_SHOW_COMMAND(page, vm_page_print_page_info)
1865 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count);
1866 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count);
1867 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count);
1868 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count);
1869 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count);
1870 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved);
1871 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min);
1872 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target);
1873 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min);
1874 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target);
1877 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
1880 db_printf("PQ_FREE:");
1881 for (i = 0; i < PQ_L2_SIZE; i++) {
1882 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
1886 db_printf("PQ_CACHE:");
1887 for (i = 0; i < PQ_L2_SIZE; i++) {
1888 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
1892 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
1893 vm_page_queues[PQ_ACTIVE].lcnt,
1894 vm_page_queues[PQ_INACTIVE].lcnt);