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 vm_page_t vm_page_array = 0;
134 int vm_page_array_size = 0;
136 int vm_page_zero_count = 0;
141 * Sets the page size, perhaps based upon the memory
142 * size. Must be called before any use of page-size
143 * dependent functions.
146 vm_set_page_size(void)
148 if (cnt.v_page_size == 0)
149 cnt.v_page_size = PAGE_SIZE;
150 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0)
151 panic("vm_set_page_size: page size not a power of two");
157 * Initializes the resident memory module.
159 * Allocates memory for the page cells, and
160 * for the object/offset-to-page hash table headers.
161 * Each page cell is initialized and placed on the free list.
164 vm_page_startup(vm_offset_t starta, vm_offset_t enda, vm_offset_t vaddr)
167 struct vm_page **bucket;
168 vm_size_t npages, page_range;
175 /* the biggest memory array is the second group of pages */
177 vm_offset_t biggestone, biggestsize;
186 vaddr = round_page(vaddr);
188 for (i = 0; phys_avail[i + 1]; i += 2) {
189 phys_avail[i] = round_page(phys_avail[i]);
190 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
193 for (i = 0; phys_avail[i + 1]; i += 2) {
194 vm_size_t size = phys_avail[i + 1] - phys_avail[i];
196 if (size > biggestsize) {
204 end = phys_avail[biggestone+1];
207 * Initialize the queue headers for the free queue, the active queue
208 * and the inactive queue.
213 * Allocate memory for use when boot strapping the kernel memory allocator
215 bootpages = UMA_BOOT_PAGES * UMA_SLAB_SIZE;
216 new_end = end - bootpages;
217 new_end = trunc_page(new_end);
218 mapped = pmap_map(&vaddr, new_end, end,
219 VM_PROT_READ | VM_PROT_WRITE);
220 bzero((caddr_t) mapped, end - new_end);
221 uma_startup((caddr_t)mapped);
226 * Allocate (and initialize) the hash table buckets.
228 * The number of buckets MUST BE a power of 2, and the actual value is
229 * the next power of 2 greater than the number of physical pages in
232 * We make the hash table approximately 2x the number of pages to
233 * reduce the chain length. This is about the same size using the
234 * singly-linked list as the 1x hash table we were using before
235 * using TAILQ but the chain length will be smaller.
237 * Note: This computation can be tweaked if desired.
239 if (vm_page_bucket_count == 0) {
240 vm_page_bucket_count = 1;
241 while (vm_page_bucket_count < atop(total))
242 vm_page_bucket_count <<= 1;
244 vm_page_bucket_count <<= 1;
245 vm_page_hash_mask = vm_page_bucket_count - 1;
248 * Validate these addresses.
250 new_end = end - vm_page_bucket_count * sizeof(struct vm_page *);
251 new_end = trunc_page(new_end);
252 mapped = pmap_map(&vaddr, new_end, end,
253 VM_PROT_READ | VM_PROT_WRITE);
254 bzero((caddr_t) mapped, end - new_end);
256 mtx_init(&vm_page_buckets_mtx, "vm page buckets mutex", NULL, MTX_SPIN);
257 vm_page_buckets = (struct vm_page **)mapped;
258 bucket = vm_page_buckets;
259 for (i = 0; i < vm_page_bucket_count; i++) {
265 * Compute the number of pages of memory that will be available for
266 * use (taking into account the overhead of a page structure per
269 first_page = phys_avail[0] / PAGE_SIZE;
270 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
271 npages = (total - (page_range * sizeof(struct vm_page)) -
272 (end - new_end)) / PAGE_SIZE;
276 * Initialize the mem entry structures now, and put them in the free
279 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
280 mapped = pmap_map(&vaddr, new_end, end,
281 VM_PROT_READ | VM_PROT_WRITE);
282 vm_page_array = (vm_page_t) mapped;
285 * Clear all of the page structures
287 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
288 vm_page_array_size = page_range;
291 * Construct the free queue(s) in descending order (by physical
292 * address) so that the first 16MB of physical memory is allocated
293 * last rather than first. On large-memory machines, this avoids
294 * the exhaustion of low physical memory before isa_dmainit has run.
296 cnt.v_page_count = 0;
297 cnt.v_free_count = 0;
298 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
303 last_pa = phys_avail[i + 1];
304 while (pa < last_pa && npages-- > 0) {
305 vm_pageq_add_new_page(pa);
315 * Distributes the object/offset key pair among hash buckets.
317 * NOTE: This macro depends on vm_page_bucket_count being a power of 2.
318 * This routine may not block.
320 * We try to randomize the hash based on the object to spread the pages
321 * out in the hash table without it costing us too much.
324 vm_page_hash(vm_object_t object, vm_pindex_t pindex)
326 int i = ((uintptr_t)object + pindex) ^ object->hash_rand;
328 return (i & vm_page_hash_mask);
332 vm_page_flag_set(vm_page_t m, unsigned short bits)
339 vm_page_flag_clear(vm_page_t m, unsigned short bits)
346 vm_page_busy(vm_page_t m)
348 KASSERT((m->flags & PG_BUSY) == 0,
349 ("vm_page_busy: page already busy!!!"));
350 vm_page_flag_set(m, PG_BUSY);
356 * wakeup anyone waiting for the page.
359 vm_page_flash(vm_page_t m)
361 if (m->flags & PG_WANTED) {
362 vm_page_flag_clear(m, PG_WANTED);
370 * clear the PG_BUSY flag and wakeup anyone waiting for the
375 vm_page_wakeup(vm_page_t m)
377 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!"));
378 vm_page_flag_clear(m, PG_BUSY);
387 vm_page_io_start(vm_page_t m)
394 vm_page_io_finish(vm_page_t m)
403 * Keep page from being freed by the page daemon
404 * much of the same effect as wiring, except much lower
405 * overhead and should be used only for *very* temporary
406 * holding ("wiring").
409 vm_page_hold(vm_page_t mem)
416 vm_page_unhold(vm_page_t mem)
420 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
421 if (mem->hold_count == 0 && mem->queue == PQ_HOLD)
422 vm_page_free_toq(mem);
428 * Reduce the protection of a page. This routine never raises the
429 * protection and therefore can be safely called if the page is already
430 * at VM_PROT_NONE (it will be a NOP effectively ).
433 vm_page_protect(vm_page_t mem, int prot)
435 if (prot == VM_PROT_NONE) {
436 if (mem->flags & (PG_WRITEABLE|PG_MAPPED)) {
437 pmap_page_protect(mem, VM_PROT_NONE);
438 vm_page_flag_clear(mem, PG_WRITEABLE|PG_MAPPED);
440 } else if ((prot == VM_PROT_READ) && (mem->flags & PG_WRITEABLE)) {
441 pmap_page_protect(mem, VM_PROT_READ);
442 vm_page_flag_clear(mem, PG_WRITEABLE);
448 * Zero-fill the specified page.
449 * Written as a standard pagein routine, to
450 * be used by the zero-fill object.
453 vm_page_zero_fill(vm_page_t m)
460 * vm_page_zero_fill_area:
462 * Like vm_page_zero_fill but only fill the specified area.
465 vm_page_zero_fill_area(vm_page_t m, int off, int size)
467 pmap_zero_page_area(m, off, size);
474 * Copy one page to another
477 vm_page_copy(vm_page_t src_m, vm_page_t dest_m)
479 pmap_copy_page(src_m, dest_m);
480 dest_m->valid = VM_PAGE_BITS_ALL;
488 * The clearing of PG_ZERO is a temporary safety until the code can be
489 * reviewed to determine that PG_ZERO is being properly cleared on
490 * write faults or maps. PG_ZERO was previously cleared in
494 vm_page_free(vm_page_t m)
496 vm_page_flag_clear(m, PG_ZERO);
498 vm_page_zero_idle_wakeup();
504 * Free a page to the zerod-pages queue
507 vm_page_free_zero(vm_page_t m)
509 vm_page_flag_set(m, PG_ZERO);
514 * vm_page_sleep_busy:
516 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE)
517 * m->busy is zero. Returns TRUE if it had to sleep ( including if
518 * it almost had to sleep and made temporary spl*() mods), FALSE
521 * This routine assumes that interrupts can only remove the busy
522 * status from a page, not set the busy status or change it from
523 * PG_BUSY to m->busy or vise versa (which would create a timing
527 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg)
530 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) {
532 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) {
534 * Page is busy. Wait and retry.
536 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
537 tsleep(m, PVM, msg, 0);
548 * make page all dirty
551 vm_page_dirty(vm_page_t m)
553 KASSERT(m->queue - m->pc != PQ_CACHE,
554 ("vm_page_dirty: page in cache!"));
555 m->dirty = VM_PAGE_BITS_ALL;
561 * Set page to not be dirty. Note: does not clear pmap modify bits
564 vm_page_undirty(vm_page_t m)
570 * vm_page_insert: [ internal use only ]
572 * Inserts the given mem entry into the object and object list.
574 * The pagetables are not updated but will presumably fault the page
575 * in if necessary, or if a kernel page the caller will at some point
576 * enter the page into the kernel's pmap. We are not allowed to block
577 * here so we *can't* do this anyway.
579 * The object and page must be locked, and must be splhigh.
580 * This routine may not block.
583 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
585 struct vm_page **bucket;
589 if (m->object != NULL)
590 panic("vm_page_insert: already inserted");
593 * Record the object/offset pair in this page
599 * Insert it into the object_object/offset hash table
601 bucket = &vm_page_buckets[vm_page_hash(object, pindex)];
602 mtx_lock_spin(&vm_page_buckets_mtx);
605 mtx_unlock_spin(&vm_page_buckets_mtx);
608 * Now link into the object's list of backed pages.
610 TAILQ_INSERT_TAIL(&object->memq, m, listq);
611 object->generation++;
614 * show that the object has one more resident page.
616 object->resident_page_count++;
619 * Since we are inserting a new and possibly dirty page,
620 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
622 if (m->flags & PG_WRITEABLE)
623 vm_object_set_writeable_dirty(object);
628 * NOTE: used by device pager as well -wfj
630 * Removes the given mem entry from the object/offset-page
631 * table and the object page list, but do not invalidate/terminate
634 * The object and page must be locked, and at splhigh.
635 * The underlying pmap entry (if any) is NOT removed here.
636 * This routine may not block.
639 vm_page_remove(vm_page_t m)
646 if (m->object == NULL)
649 if ((m->flags & PG_BUSY) == 0) {
650 panic("vm_page_remove: page not busy");
654 * Basically destroy the page.
661 * Remove from the object_object/offset hash table. The object
662 * must be on the hash queue, we will panic if it isn't
664 bucket = &vm_page_buckets[vm_page_hash(m->object, m->pindex)];
665 mtx_lock_spin(&vm_page_buckets_mtx);
666 while (*bucket != m) {
668 panic("vm_page_remove(): page not found in hash");
669 bucket = &(*bucket)->hnext;
673 mtx_unlock_spin(&vm_page_buckets_mtx);
676 * Now remove from the object's list of backed pages.
678 TAILQ_REMOVE(&object->memq, m, listq);
681 * And show that the object has one fewer resident page.
683 object->resident_page_count--;
684 object->generation++;
692 * Returns the page associated with the object/offset
693 * pair specified; if none is found, NULL is returned.
695 * The object must be locked. No side effects.
696 * This routine may not block.
697 * This is a critical path routine
700 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
703 struct vm_page **bucket;
706 * Search the hash table for this object/offset pair
708 bucket = &vm_page_buckets[vm_page_hash(object, pindex)];
709 mtx_lock_spin(&vm_page_buckets_mtx);
710 for (m = *bucket; m != NULL; m = m->hnext)
711 if (m->object == object && m->pindex == pindex)
713 mtx_unlock_spin(&vm_page_buckets_mtx);
720 * Move the given memory entry from its
721 * current object to the specified target object/offset.
723 * The object must be locked.
724 * This routine may not block.
726 * Note: this routine will raise itself to splvm(), the caller need not.
728 * Note: swap associated with the page must be invalidated by the move. We
729 * have to do this for several reasons: (1) we aren't freeing the
730 * page, (2) we are dirtying the page, (3) the VM system is probably
731 * moving the page from object A to B, and will then later move
732 * the backing store from A to B and we can't have a conflict.
734 * Note: we *always* dirty the page. It is necessary both for the
735 * fact that we moved it, and because we may be invalidating
736 * swap. If the page is on the cache, we have to deactivate it
737 * or vm_page_dirty() will panic. Dirty pages are not allowed
741 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
747 vm_page_insert(m, new_object, new_pindex);
748 if (m->queue - m->pc == PQ_CACHE)
749 vm_page_deactivate(m);
755 * vm_page_select_cache:
757 * Find a page on the cache queue with color optimization. As pages
758 * might be found, but not applicable, they are deactivated. This
759 * keeps us from using potentially busy cached pages.
761 * This routine must be called at splvm().
762 * This routine may not block.
765 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex)
773 (pindex + object->pg_color) & PQ_L2_MASK,
776 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
777 m->hold_count || m->wire_count)) {
778 vm_page_deactivate(m);
786 * vm_page_select_free:
788 * Find a free or zero page, with specified preference.
790 * This routine must be called at splvm().
791 * This routine may not block.
793 static __inline vm_page_t
794 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero)
800 (pindex + object->pg_color) & PQ_L2_MASK,
809 * Allocate and return a memory cell associated
810 * with this VM object/offset pair.
813 * VM_ALLOC_NORMAL normal process request
814 * VM_ALLOC_SYSTEM system *really* needs a page
815 * VM_ALLOC_INTERRUPT interrupt time request
816 * VM_ALLOC_ZERO zero page
818 * This routine may not block.
820 * Additional special handling is required when called from an
821 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with
822 * the page cache in this case.
825 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
828 boolean_t prefer_zero;
833 KASSERT(!vm_page_lookup(object, pindex),
834 ("vm_page_alloc: page already allocated"));
836 prefer_zero = (page_req & VM_ALLOC_ZERO) != 0 ? TRUE : FALSE;
837 page_req &= ~VM_ALLOC_ZERO;
840 * The pager is allowed to eat deeper into the free page list.
842 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) {
843 page_req = VM_ALLOC_SYSTEM;
849 if (cnt.v_free_count > cnt.v_free_reserved) {
851 * Allocate from the free queue if there are plenty of pages
855 m = vm_page_select_free(object, pindex, prefer_zero);
857 (page_req == VM_ALLOC_SYSTEM &&
858 cnt.v_cache_count == 0 &&
859 cnt.v_free_count > cnt.v_interrupt_free_min) ||
860 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0)
863 * Interrupt or system, dig deeper into the free list.
865 m = vm_page_select_free(object, pindex, FALSE);
866 } else if (page_req != VM_ALLOC_INTERRUPT) {
868 * Allocatable from cache (non-interrupt only). On success,
869 * we must free the page and try again, thus ensuring that
870 * cnt.v_*_free_min counters are replenished.
872 m = vm_page_select_cache(object, pindex);
875 #if defined(DIAGNOSTIC)
876 if (cnt.v_cache_count > 0)
877 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count);
879 vm_pageout_deficit++;
883 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m));
885 vm_page_protect(m, VM_PROT_NONE);
890 * Not allocatable from cache from interrupt, give up.
893 vm_pageout_deficit++;
899 * At this point we had better have found a good page.
904 ("vm_page_alloc(): missing page on free queue\n")
908 * Remove from free queue
911 vm_pageq_remove_nowakeup(m);
914 * Initialize structure. Only the PG_ZERO flag is inherited.
916 if (m->flags & PG_ZERO) {
917 vm_page_zero_count--;
918 m->flags = PG_ZERO | PG_BUSY;
927 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m));
930 * vm_page_insert() is safe prior to the splx(). Note also that
931 * inserting a page here does not insert it into the pmap (which
932 * could cause us to block allocating memory). We cannot block
935 vm_page_insert(m, object, pindex);
938 * Don't wakeup too often - wakeup the pageout daemon when
939 * we would be nearly out of memory.
941 if (vm_paging_needed())
949 * vm_wait: (also see VM_WAIT macro)
951 * Block until free pages are available for allocation
952 * - Called in various places before memory allocations.
960 if (curproc == pageproc) {
961 vm_pageout_pages_needed = 1;
962 tsleep(&vm_pageout_pages_needed, PSWP, "VMWait", 0);
964 if (!vm_pages_needed) {
966 wakeup(&vm_pages_needed);
968 tsleep(&cnt.v_free_count, PVM, "vmwait", 0);
974 * vm_waitpfault: (also see VM_WAITPFAULT macro)
976 * Block until free pages are available for allocation
977 * - Called only in vm_fault so that processes page faulting
978 * can be easily tracked.
979 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
980 * processes will be able to grab memory first. Do not change
981 * this balance without careful testing first.
989 if (!vm_pages_needed) {
991 wakeup(&vm_pages_needed);
993 tsleep(&cnt.v_free_count, PUSER, "pfault", 0);
1000 * Put the specified page on the active list (if appropriate).
1001 * Ensure that act_count is at least ACT_INIT but do not otherwise
1004 * The page queues must be locked.
1005 * This routine may not block.
1008 vm_page_activate(vm_page_t m)
1014 if (m->queue != PQ_ACTIVE) {
1015 if ((m->queue - m->pc) == PQ_CACHE)
1016 cnt.v_reactivated++;
1018 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1019 if (m->act_count < ACT_INIT)
1020 m->act_count = ACT_INIT;
1021 vm_pageq_enqueue(PQ_ACTIVE, m);
1024 if (m->act_count < ACT_INIT)
1025 m->act_count = ACT_INIT;
1031 * vm_page_free_wakeup:
1033 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
1034 * routine is called when a page has been added to the cache or free
1037 * This routine may not block.
1038 * This routine must be called at splvm()
1040 static __inline void
1041 vm_page_free_wakeup(void)
1044 * if pageout daemon needs pages, then tell it that there are
1047 if (vm_pageout_pages_needed &&
1048 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) {
1049 wakeup(&vm_pageout_pages_needed);
1050 vm_pageout_pages_needed = 0;
1053 * wakeup processes that are waiting on memory if we hit a
1054 * high water mark. And wakeup scheduler process if we have
1055 * lots of memory. this process will swapin processes.
1057 if (vm_pages_needed && !vm_page_count_min()) {
1058 vm_pages_needed = 0;
1059 wakeup(&cnt.v_free_count);
1066 * Returns the given page to the PQ_FREE list,
1067 * disassociating it with any VM object.
1069 * Object and page must be locked prior to entry.
1070 * This routine may not block.
1074 vm_page_free_toq(vm_page_t m)
1077 struct vpgqueues *pq;
1078 vm_object_t object = m->object;
1084 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
1086 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
1087 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
1089 if ((m->queue - m->pc) == PQ_FREE)
1090 panic("vm_page_free: freeing free page");
1092 panic("vm_page_free: freeing busy page");
1096 * unqueue, then remove page. Note that we cannot destroy
1097 * the page here because we do not want to call the pager's
1098 * callback routine until after we've put the page on the
1099 * appropriate free queue.
1101 vm_pageq_remove_nowakeup(m);
1105 * If fictitious remove object association and
1106 * return, otherwise delay object association removal.
1108 if ((m->flags & PG_FICTITIOUS) != 0) {
1116 if (m->wire_count != 0) {
1117 if (m->wire_count > 1) {
1118 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1119 m->wire_count, (long)m->pindex);
1121 panic("vm_page_free: freeing wired page\n");
1125 * If we've exhausted the object's resident pages we want to free
1129 (object->type == OBJT_VNODE) &&
1130 ((object->flags & OBJ_DEAD) == 0)
1132 struct vnode *vp = (struct vnode *)object->handle;
1134 if (vp && VSHOULDFREE(vp))
1139 * Clear the UNMANAGED flag when freeing an unmanaged page.
1141 if (m->flags & PG_UNMANAGED) {
1142 m->flags &= ~PG_UNMANAGED;
1145 pmap_page_is_free(m);
1149 if (m->hold_count != 0) {
1150 m->flags &= ~PG_ZERO;
1153 m->queue = PQ_FREE + m->pc;
1154 pq = &vm_page_queues[m->queue];
1159 * Put zero'd pages on the end ( where we look for zero'd pages
1160 * first ) and non-zerod pages at the head.
1162 if (m->flags & PG_ZERO) {
1163 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
1164 ++vm_page_zero_count;
1166 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1168 vm_page_free_wakeup();
1175 * Prevent PV management from being done on the page. The page is
1176 * removed from the paging queues as if it were wired, and as a
1177 * consequence of no longer being managed the pageout daemon will not
1178 * touch it (since there is no way to locate the pte mappings for the
1179 * page). madvise() calls that mess with the pmap will also no longer
1180 * operate on the page.
1182 * Beyond that the page is still reasonably 'normal'. Freeing the page
1183 * will clear the flag.
1185 * This routine is used by OBJT_PHYS objects - objects using unswappable
1186 * physical memory as backing store rather then swap-backed memory and
1187 * will eventually be extended to support 4MB unmanaged physical
1191 vm_page_unmanage(vm_page_t m)
1196 if ((m->flags & PG_UNMANAGED) == 0) {
1197 if (m->wire_count == 0)
1200 vm_page_flag_set(m, PG_UNMANAGED);
1207 * Mark this page as wired down by yet
1208 * another map, removing it from paging queues
1211 * The page queues must be locked.
1212 * This routine may not block.
1215 vm_page_wire(vm_page_t m)
1220 * Only bump the wire statistics if the page is not already wired,
1221 * and only unqueue the page if it is on some queue (if it is unmanaged
1222 * it is already off the queues).
1225 if (m->wire_count == 0) {
1226 if ((m->flags & PG_UNMANAGED) == 0)
1231 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
1233 vm_page_flag_set(m, PG_MAPPED);
1239 * Release one wiring of this page, potentially
1240 * enabling it to be paged again.
1242 * Many pages placed on the inactive queue should actually go
1243 * into the cache, but it is difficult to figure out which. What
1244 * we do instead, if the inactive target is well met, is to put
1245 * clean pages at the head of the inactive queue instead of the tail.
1246 * This will cause them to be moved to the cache more quickly and
1247 * if not actively re-referenced, freed more quickly. If we just
1248 * stick these pages at the end of the inactive queue, heavy filesystem
1249 * meta-data accesses can cause an unnecessary paging load on memory bound
1250 * processes. This optimization causes one-time-use metadata to be
1251 * reused more quickly.
1253 * BUT, if we are in a low-memory situation we have no choice but to
1254 * put clean pages on the cache queue.
1256 * A number of routines use vm_page_unwire() to guarantee that the page
1257 * will go into either the inactive or active queues, and will NEVER
1258 * be placed in the cache - for example, just after dirtying a page.
1259 * dirty pages in the cache are not allowed.
1261 * The page queues must be locked.
1262 * This routine may not block.
1265 vm_page_unwire(vm_page_t m, int activate)
1271 if (m->wire_count > 0) {
1273 if (m->wire_count == 0) {
1275 if (m->flags & PG_UNMANAGED) {
1277 } else if (activate)
1278 vm_pageq_enqueue(PQ_ACTIVE, m);
1280 vm_page_flag_clear(m, PG_WINATCFLS);
1281 vm_pageq_enqueue(PQ_INACTIVE, m);
1285 panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count);
1292 * Move the specified page to the inactive queue. If the page has
1293 * any associated swap, the swap is deallocated.
1295 * Normally athead is 0 resulting in LRU operation. athead is set
1296 * to 1 if we want this page to be 'as if it were placed in the cache',
1297 * except without unmapping it from the process address space.
1299 * This routine may not block.
1301 static __inline void
1302 _vm_page_deactivate(vm_page_t m, int athead)
1308 * Ignore if already inactive.
1310 if (m->queue == PQ_INACTIVE)
1314 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1315 if ((m->queue - m->pc) == PQ_CACHE)
1316 cnt.v_reactivated++;
1317 vm_page_flag_clear(m, PG_WINATCFLS);
1320 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1322 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1323 m->queue = PQ_INACTIVE;
1324 vm_page_queues[PQ_INACTIVE].lcnt++;
1325 cnt.v_inactive_count++;
1331 vm_page_deactivate(vm_page_t m)
1333 _vm_page_deactivate(m, 0);
1337 * vm_page_try_to_cache:
1339 * Returns 0 on failure, 1 on success
1342 vm_page_try_to_cache(vm_page_t m)
1346 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1347 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1350 vm_page_test_dirty(m);
1358 * vm_page_try_to_free()
1360 * Attempt to free the page. If we cannot free it, we do nothing.
1361 * 1 is returned on success, 0 on failure.
1364 vm_page_try_to_free(vm_page_t m)
1366 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1367 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1370 vm_page_test_dirty(m);
1374 vm_page_protect(m, VM_PROT_NONE);
1382 * Put the specified page onto the page cache queue (if appropriate).
1384 * This routine may not block.
1387 vm_page_cache(vm_page_t m)
1392 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || m->wire_count) {
1393 printf("vm_page_cache: attempting to cache busy page\n");
1396 if ((m->queue - m->pc) == PQ_CACHE)
1400 * Remove all pmaps and indicate that the page is not
1401 * writeable or mapped.
1403 vm_page_protect(m, VM_PROT_NONE);
1404 if (m->dirty != 0) {
1405 panic("vm_page_cache: caching a dirty page, pindex: %ld",
1409 vm_pageq_remove_nowakeup(m);
1410 vm_pageq_enqueue(PQ_CACHE + m->pc, m);
1411 vm_page_free_wakeup();
1418 * Cache, deactivate, or do nothing as appropriate. This routine
1419 * is typically used by madvise() MADV_DONTNEED.
1421 * Generally speaking we want to move the page into the cache so
1422 * it gets reused quickly. However, this can result in a silly syndrome
1423 * due to the page recycling too quickly. Small objects will not be
1424 * fully cached. On the otherhand, if we move the page to the inactive
1425 * queue we wind up with a problem whereby very large objects
1426 * unnecessarily blow away our inactive and cache queues.
1428 * The solution is to move the pages based on a fixed weighting. We
1429 * either leave them alone, deactivate them, or move them to the cache,
1430 * where moving them to the cache has the highest weighting.
1431 * By forcing some pages into other queues we eventually force the
1432 * system to balance the queues, potentially recovering other unrelated
1433 * space from active. The idea is to not force this to happen too
1437 vm_page_dontneed(vm_page_t m)
1439 static int dnweight;
1447 * occassionally leave the page alone
1449 if ((dnw & 0x01F0) == 0 ||
1450 m->queue == PQ_INACTIVE ||
1451 m->queue - m->pc == PQ_CACHE
1453 if (m->act_count >= ACT_INIT)
1459 vm_page_test_dirty(m);
1461 if (m->dirty || (dnw & 0x0070) == 0) {
1463 * Deactivate the page 3 times out of 32.
1468 * Cache the page 28 times out of every 32. Note that
1469 * the page is deactivated instead of cached, but placed
1470 * at the head of the queue instead of the tail.
1474 _vm_page_deactivate(m, head);
1478 * Grab a page, waiting until we are waken up due to the page
1479 * changing state. We keep on waiting, if the page continues
1480 * to be in the object. If the page doesn't exist, allocate it.
1482 * This routine may block.
1485 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
1492 if ((m = vm_page_lookup(object, pindex)) != NULL) {
1493 if (m->busy || (m->flags & PG_BUSY)) {
1494 generation = object->generation;
1497 while ((object->generation == generation) &&
1498 (m->busy || (m->flags & PG_BUSY))) {
1499 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
1500 tsleep(m, PVM, "pgrbwt", 0);
1501 if ((allocflags & VM_ALLOC_RETRY) == 0) {
1514 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1517 if ((allocflags & VM_ALLOC_RETRY) == 0)
1526 * Mapping function for valid bits or for dirty bits in
1527 * a page. May not block.
1529 * Inputs are required to range within a page.
1532 vm_page_bits(int base, int size)
1538 base + size <= PAGE_SIZE,
1539 ("vm_page_bits: illegal base/size %d/%d", base, size)
1542 if (size == 0) /* handle degenerate case */
1545 first_bit = base >> DEV_BSHIFT;
1546 last_bit = (base + size - 1) >> DEV_BSHIFT;
1548 return ((2 << last_bit) - (1 << first_bit));
1552 * vm_page_set_validclean:
1554 * Sets portions of a page valid and clean. The arguments are expected
1555 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1556 * of any partial chunks touched by the range. The invalid portion of
1557 * such chunks will be zero'd.
1559 * This routine may not block.
1561 * (base + size) must be less then or equal to PAGE_SIZE.
1564 vm_page_set_validclean(vm_page_t m, int base, int size)
1571 if (size == 0) /* handle degenerate case */
1575 * If the base is not DEV_BSIZE aligned and the valid
1576 * bit is clear, we have to zero out a portion of the
1579 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
1580 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
1581 pmap_zero_page_area(m, frag, base - frag);
1584 * If the ending offset is not DEV_BSIZE aligned and the
1585 * valid bit is clear, we have to zero out a portion of
1588 endoff = base + size;
1589 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
1590 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
1591 pmap_zero_page_area(m, endoff,
1592 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
1595 * Set valid, clear dirty bits. If validating the entire
1596 * page we can safely clear the pmap modify bit. We also
1597 * use this opportunity to clear the PG_NOSYNC flag. If a process
1598 * takes a write fault on a MAP_NOSYNC memory area the flag will
1601 * We set valid bits inclusive of any overlap, but we can only
1602 * clear dirty bits for DEV_BSIZE chunks that are fully within
1605 pagebits = vm_page_bits(base, size);
1606 m->valid |= pagebits;
1608 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
1609 frag = DEV_BSIZE - frag;
1615 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
1617 m->dirty &= ~pagebits;
1618 if (base == 0 && size == PAGE_SIZE) {
1619 pmap_clear_modify(m);
1620 vm_page_flag_clear(m, PG_NOSYNC);
1627 vm_page_set_dirty(vm_page_t m, int base, int size)
1629 m->dirty |= vm_page_bits(base, size);
1635 vm_page_clear_dirty(vm_page_t m, int base, int size)
1638 m->dirty &= ~vm_page_bits(base, size);
1642 * vm_page_set_invalid:
1644 * Invalidates DEV_BSIZE'd chunks within a page. Both the
1645 * valid and dirty bits for the effected areas are cleared.
1650 vm_page_set_invalid(vm_page_t m, int base, int size)
1655 bits = vm_page_bits(base, size);
1658 m->object->generation++;
1662 * vm_page_zero_invalid()
1664 * The kernel assumes that the invalid portions of a page contain
1665 * garbage, but such pages can be mapped into memory by user code.
1666 * When this occurs, we must zero out the non-valid portions of the
1667 * page so user code sees what it expects.
1669 * Pages are most often semi-valid when the end of a file is mapped
1670 * into memory and the file's size is not page aligned.
1673 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
1679 * Scan the valid bits looking for invalid sections that
1680 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
1681 * valid bit may be set ) have already been zerod by
1682 * vm_page_set_validclean().
1684 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
1685 if (i == (PAGE_SIZE / DEV_BSIZE) ||
1686 (m->valid & (1 << i))
1689 pmap_zero_page_area(m,
1690 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
1697 * setvalid is TRUE when we can safely set the zero'd areas
1698 * as being valid. We can do this if there are no cache consistancy
1699 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
1702 m->valid = VM_PAGE_BITS_ALL;
1708 * Is (partial) page valid? Note that the case where size == 0
1709 * will return FALSE in the degenerate case where the page is
1710 * entirely invalid, and TRUE otherwise.
1715 vm_page_is_valid(vm_page_t m, int base, int size)
1717 int bits = vm_page_bits(base, size);
1719 if (m->valid && ((m->valid & bits) == bits))
1726 * update dirty bits from pmap/mmu. May not block.
1729 vm_page_test_dirty(vm_page_t m)
1731 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
1736 int so_zerocp_fullpage = 0;
1739 vm_page_cowfault(vm_page_t m)
1751 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL);
1753 vm_page_insert(m, object, pindex);
1760 * check to see if we raced with an xmit complete when
1761 * waiting to allocate a page. If so, put things back
1766 vm_page_insert(m, object, pindex);
1767 } else { /* clear COW & copy page */
1768 if (so_zerocp_fullpage) {
1769 mnew->valid = VM_PAGE_BITS_ALL;
1771 vm_page_copy(m, mnew);
1773 vm_page_dirty(mnew);
1774 vm_page_flag_clear(mnew, PG_BUSY);
1776 vm_page_wakeup(m); /*unbusy the page */
1780 vm_page_cowclear(vm_page_t m)
1783 /* XXX KDM find out if giant is required here. */
1786 atomic_subtract_int(&m->cow, 1);
1788 * let vm_fault add back write permission lazily
1792 * sf_buf_free() will free the page, so we needn't do it here
1797 vm_page_cowsetup(vm_page_t m)
1799 /* XXX KDM find out if giant is required here */
1801 atomic_add_int(&m->cow, 1);
1802 vm_page_protect(m, VM_PROT_READ);
1805 #include "opt_ddb.h"
1807 #include <sys/kernel.h>
1809 #include <ddb/ddb.h>
1811 DB_SHOW_COMMAND(page, vm_page_print_page_info)
1813 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count);
1814 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count);
1815 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count);
1816 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count);
1817 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count);
1818 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved);
1819 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min);
1820 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target);
1821 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min);
1822 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target);
1825 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
1828 db_printf("PQ_FREE:");
1829 for (i = 0; i < PQ_L2_SIZE; i++) {
1830 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
1834 db_printf("PQ_CACHE:");
1835 for (i = 0; i < PQ_L2_SIZE; i++) {
1836 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
1840 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
1841 vm_page_queues[PQ_ACTIVE].lcnt,
1842 vm_page_queues[PQ_INACTIVE].lcnt);