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>
123 * Associated with page of user-allocatable memory is a
127 static struct vm_page **vm_page_buckets; /* Array of buckets */
128 static int vm_page_bucket_count; /* How big is array? */
129 static int vm_page_hash_mask; /* Mask for hash function */
130 static volatile int vm_page_bucket_generation;
131 static struct mtx vm_buckets_mtx[BUCKET_HASH_SIZE];
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.
165 vm_page_startup(vm_offset_t starta, vm_offset_t enda, vm_offset_t vaddr)
168 struct vm_page **bucket;
169 vm_size_t npages, page_range;
176 /* the biggest memory array is the second group of pages */
178 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 int 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.
214 * Allocate (and initialize) the hash table buckets.
216 * The number of buckets MUST BE a power of 2, and the actual value is
217 * the next power of 2 greater than the number of physical pages in
220 * We make the hash table approximately 2x the number of pages to
221 * reduce the chain length. This is about the same size using the
222 * singly-linked list as the 1x hash table we were using before
223 * using TAILQ but the chain length will be smaller.
225 * Note: This computation can be tweaked if desired.
227 if (vm_page_bucket_count == 0) {
228 vm_page_bucket_count = 1;
229 while (vm_page_bucket_count < atop(total))
230 vm_page_bucket_count <<= 1;
232 vm_page_bucket_count <<= 1;
233 vm_page_hash_mask = vm_page_bucket_count - 1;
236 * Validate these addresses.
238 new_end = end - vm_page_bucket_count * sizeof(struct vm_page *);
239 new_end = trunc_page(new_end);
240 mapped = pmap_map(&vaddr, new_end, end,
241 VM_PROT_READ | VM_PROT_WRITE);
242 bzero((caddr_t) mapped, end - new_end);
244 vm_page_buckets = (struct vm_page **)mapped;
245 bucket = vm_page_buckets;
246 for (i = 0; i < vm_page_bucket_count; i++) {
250 for (i = 0; i < BUCKET_HASH_SIZE; ++i)
251 mtx_init(&vm_buckets_mtx[i], "vm buckets hash mutexes", MTX_DEF);
254 * Compute the number of pages of memory that will be available for
255 * use (taking into account the overhead of a page structure per
259 first_page = phys_avail[0] / PAGE_SIZE;
261 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
262 npages = (total - (page_range * sizeof(struct vm_page)) -
263 (end - new_end)) / PAGE_SIZE;
268 * Initialize the mem entry structures now, and put them in the free
271 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
272 mapped = pmap_map(&vaddr, new_end, end,
273 VM_PROT_READ | VM_PROT_WRITE);
274 vm_page_array = (vm_page_t) mapped;
277 * Clear all of the page structures
279 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
280 vm_page_array_size = page_range;
283 * Construct the free queue(s) in descending order (by physical
284 * address) so that the first 16MB of physical memory is allocated
285 * last rather than first. On large-memory machines, this avoids
286 * the exhaustion of low physical memory before isa_dmainit has run.
288 cnt.v_page_count = 0;
289 cnt.v_free_count = 0;
290 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
295 last_pa = phys_avail[i + 1];
296 while (pa < last_pa && npages-- > 0) {
297 vm_pageq_add_new_page(pa);
307 * Distributes the object/offset key pair among hash buckets.
309 * NOTE: This macro depends on vm_page_bucket_count being a power of 2.
310 * This routine may not block.
312 * We try to randomize the hash based on the object to spread the pages
313 * out in the hash table without it costing us too much.
316 vm_page_hash(vm_object_t object, vm_pindex_t pindex)
318 int i = ((uintptr_t)object + pindex) ^ object->hash_rand;
320 return(i & vm_page_hash_mask);
324 vm_page_flag_set(vm_page_t m, unsigned short bits)
327 atomic_set_short(&(m)->flags, bits);
328 /* m->flags |= bits; */
332 vm_page_flag_clear(vm_page_t m, unsigned short bits)
335 atomic_clear_short(&(m)->flags, bits);
336 /* m->flags &= ~bits; */
340 vm_page_busy(vm_page_t m)
342 KASSERT((m->flags & PG_BUSY) == 0,
343 ("vm_page_busy: page already busy!!!"));
344 vm_page_flag_set(m, PG_BUSY);
350 * wakeup anyone waiting for the page.
354 vm_page_flash(vm_page_t m)
356 if (m->flags & PG_WANTED) {
357 vm_page_flag_clear(m, PG_WANTED);
365 * clear the PG_BUSY flag and wakeup anyone waiting for the
371 vm_page_wakeup(vm_page_t m)
373 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!"));
374 vm_page_flag_clear(m, PG_BUSY);
384 vm_page_io_start(vm_page_t m)
387 atomic_add_char(&(m)->busy, 1);
391 vm_page_io_finish(vm_page_t m)
394 atomic_subtract_char(&(m)->busy, 1);
400 * Keep page from being freed by the page daemon
401 * much of the same effect as wiring, except much lower
402 * overhead and should be used only for *very* temporary
403 * holding ("wiring").
406 vm_page_hold(vm_page_t mem)
413 vm_page_unhold(vm_page_t mem)
417 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
423 * Reduce the protection of a page. This routine never raises the
424 * protection and therefore can be safely called if the page is already
425 * at VM_PROT_NONE (it will be a NOP effectively ).
429 vm_page_protect(vm_page_t mem, int prot)
431 if (prot == VM_PROT_NONE) {
432 if (mem->flags & (PG_WRITEABLE|PG_MAPPED)) {
433 pmap_page_protect(mem, VM_PROT_NONE);
434 vm_page_flag_clear(mem, PG_WRITEABLE|PG_MAPPED);
436 } else if ((prot == VM_PROT_READ) && (mem->flags & PG_WRITEABLE)) {
437 pmap_page_protect(mem, VM_PROT_READ);
438 vm_page_flag_clear(mem, PG_WRITEABLE);
444 * Zero-fill the specified page.
445 * Written as a standard pagein routine, to
446 * be used by the zero-fill object.
449 vm_page_zero_fill(vm_page_t m)
451 pmap_zero_page(VM_PAGE_TO_PHYS(m));
458 * Copy one page to another
461 vm_page_copy(vm_page_t src_m, vm_page_t dest_m)
463 pmap_copy_page(VM_PAGE_TO_PHYS(src_m), VM_PAGE_TO_PHYS(dest_m));
464 dest_m->valid = VM_PAGE_BITS_ALL;
472 * The clearing of PG_ZERO is a temporary safety until the code can be
473 * reviewed to determine that PG_ZERO is being properly cleared on
474 * write faults or maps. PG_ZERO was previously cleared in
478 vm_page_free(vm_page_t m)
480 vm_page_flag_clear(m, PG_ZERO);
487 * Free a page to the zerod-pages queue
490 vm_page_free_zero(vm_page_t m)
492 vm_page_flag_set(m, PG_ZERO);
497 * vm_page_sleep_busy:
499 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE)
500 * m->busy is zero. Returns TRUE if it had to sleep ( including if
501 * it almost had to sleep and made temporary spl*() mods), FALSE
504 * This routine assumes that interrupts can only remove the busy
505 * status from a page, not set the busy status or change it from
506 * PG_BUSY to m->busy or vise versa (which would create a timing
511 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg)
514 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) {
516 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) {
518 * Page is busy. Wait and retry.
520 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
521 tsleep(m, PVM, msg, 0);
532 * make page all dirty
536 vm_page_dirty(vm_page_t m)
538 KASSERT(m->queue - m->pc != PQ_CACHE,
539 ("vm_page_dirty: page in cache!"));
540 m->dirty = VM_PAGE_BITS_ALL;
546 * Set page to not be dirty. Note: does not clear pmap modify bits
550 vm_page_undirty(vm_page_t m)
556 * vm_page_insert: [ internal use only ]
558 * Inserts the given mem entry into the object and object list.
560 * The pagetables are not updated but will presumably fault the page
561 * in if necessary, or if a kernel page the caller will at some point
562 * enter the page into the kernel's pmap. We are not allowed to block
563 * here so we *can't* do this anyway.
565 * The object and page must be locked, and must be splhigh.
566 * This routine may not block.
570 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
572 struct vm_page **bucket;
576 if (m->object != NULL)
577 panic("vm_page_insert: already inserted");
580 * Record the object/offset pair in this page
587 * Insert it into the object_object/offset hash table
590 bucket = &vm_page_buckets[vm_page_hash(object, pindex)];
593 vm_page_bucket_generation++;
596 * Now link into the object's list of backed pages.
599 TAILQ_INSERT_TAIL(&object->memq, m, listq);
600 object->generation++;
603 * show that the object has one more resident page.
606 object->resident_page_count++;
609 * Since we are inserting a new and possibly dirty page,
610 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
612 if (m->flags & PG_WRITEABLE)
613 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
618 * NOTE: used by device pager as well -wfj
620 * Removes the given mem entry from the object/offset-page
621 * table and the object page list, but do not invalidate/terminate
624 * The object and page must be locked, and at splhigh.
625 * The underlying pmap entry (if any) is NOT removed here.
626 * This routine may not block.
630 vm_page_remove(vm_page_t m)
636 if (m->object == NULL)
639 if ((m->flags & PG_BUSY) == 0) {
640 panic("vm_page_remove: page not busy");
644 * Basically destroy the page.
652 * Remove from the object_object/offset hash table. The object
653 * must be on the hash queue, we will panic if it isn't
655 * Note: we must NULL-out m->hnext to prevent loops in detached
656 * buffers with vm_page_lookup().
660 struct vm_page **bucket;
662 bucket = &vm_page_buckets[vm_page_hash(m->object, m->pindex)];
663 while (*bucket != m) {
665 panic("vm_page_remove(): page not found in hash");
666 bucket = &(*bucket)->hnext;
670 vm_page_bucket_generation++;
674 * Now remove from the object's list of backed pages.
677 TAILQ_REMOVE(&object->memq, m, listq);
680 * 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 * NOTE: the code below does not lock. It will operate properly if
696 * an interrupt makes a change, but the generation algorithm will not
697 * operate properly in an SMP environment where both cpu's are able to run
698 * kernel code simultaneously.
700 * The object must be locked. No side effects.
701 * This routine may not block.
702 * This is a critical path routine
706 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
709 struct vm_page **bucket;
713 * Search the hash table for this object/offset pair
717 generation = vm_page_bucket_generation;
718 bucket = &vm_page_buckets[vm_page_hash(object, pindex)];
719 for (m = *bucket; m != NULL; m = m->hnext) {
720 if ((m->object == object) && (m->pindex == pindex)) {
721 if (vm_page_bucket_generation != generation)
726 if (vm_page_bucket_generation != generation)
734 * Move the given memory entry from its
735 * current object to the specified target object/offset.
737 * The object must be locked.
738 * This routine may not block.
740 * Note: this routine will raise itself to splvm(), the caller need not.
742 * Note: swap associated with the page must be invalidated by the move. We
743 * have to do this for several reasons: (1) we aren't freeing the
744 * page, (2) we are dirtying the page, (3) the VM system is probably
745 * moving the page from object A to B, and will then later move
746 * the backing store from A to B and we can't have a conflict.
748 * Note: we *always* dirty the page. It is necessary both for the
749 * fact that we moved it, and because we may be invalidating
750 * swap. If the page is on the cache, we have to deactivate it
751 * or vm_page_dirty() will panic. Dirty pages are not allowed
756 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
762 vm_page_insert(m, new_object, new_pindex);
763 if (m->queue - m->pc == PQ_CACHE)
764 vm_page_deactivate(m);
770 * vm_page_select_cache:
772 * Find a page on the cache queue with color optimization. As pages
773 * might be found, but not applicable, they are deactivated. This
774 * keeps us from using potentially busy cached pages.
776 * This routine must be called at splvm().
777 * This routine may not block.
780 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex)
788 (pindex + object->pg_color) & PQ_L2_MASK,
791 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
792 m->hold_count || m->wire_count)) {
793 vm_page_deactivate(m);
801 * vm_page_select_free:
803 * Find a free or zero page, with specified preference.
805 * This routine must be called at splvm().
806 * This routine may not block.
809 static __inline vm_page_t
810 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero)
816 (pindex + object->pg_color) & PQ_L2_MASK,
825 * Allocate and return a memory cell associated
826 * with this VM object/offset pair.
829 * VM_ALLOC_NORMAL normal process request
830 * VM_ALLOC_SYSTEM system *really* needs a page
831 * VM_ALLOC_INTERRUPT interrupt time request
832 * VM_ALLOC_ZERO zero page
834 * This routine may not block.
836 * Additional special handling is required when called from an
837 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with
838 * the page cache in this case.
842 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
849 KASSERT(!vm_page_lookup(object, pindex),
850 ("vm_page_alloc: page already allocated"));
853 * The pager is allowed to eat deeper into the free page list.
856 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) {
857 page_req = VM_ALLOC_SYSTEM;
863 if (cnt.v_free_count > cnt.v_free_reserved) {
865 * Allocate from the free queue if there are plenty of pages
868 if (page_req == VM_ALLOC_ZERO)
869 m = vm_page_select_free(object, pindex, TRUE);
871 m = vm_page_select_free(object, pindex, FALSE);
873 (page_req == VM_ALLOC_SYSTEM &&
874 cnt.v_cache_count == 0 &&
875 cnt.v_free_count > cnt.v_interrupt_free_min) ||
876 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0)
879 * Interrupt or system, dig deeper into the free list.
881 m = vm_page_select_free(object, pindex, FALSE);
882 } else if (page_req != VM_ALLOC_INTERRUPT) {
884 * Allocatable from cache (non-interrupt only). On success,
885 * we must free the page and try again, thus ensuring that
886 * cnt.v_*_free_min counters are replenished.
888 m = vm_page_select_cache(object, pindex);
891 #if defined(DIAGNOSTIC)
892 if (cnt.v_cache_count > 0)
893 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count);
895 vm_pageout_deficit++;
899 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m));
901 vm_page_protect(m, VM_PROT_NONE);
906 * Not allocatable from cache from interrupt, give up.
909 vm_pageout_deficit++;
915 * At this point we had better have found a good page.
920 ("vm_page_alloc(): missing page on free queue\n")
924 * Remove from free queue
927 vm_pageq_remove_nowakeup(m);
930 * Initialize structure. Only the PG_ZERO flag is inherited.
933 if (m->flags & PG_ZERO) {
934 vm_page_zero_count--;
935 m->flags = PG_ZERO | PG_BUSY;
944 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m));
947 * vm_page_insert() is safe prior to the splx(). Note also that
948 * inserting a page here does not insert it into the pmap (which
949 * could cause us to block allocating memory). We cannot block
953 vm_page_insert(m, object, pindex);
956 * Don't wakeup too often - wakeup the pageout daemon when
957 * we would be nearly out of memory.
959 if (vm_paging_needed())
968 * vm_wait: (also see VM_WAIT macro)
970 * Block until free pages are available for allocation
979 if (curproc == pageproc) {
980 vm_pageout_pages_needed = 1;
981 tsleep(&vm_pageout_pages_needed, PSWP, "VMWait", 0);
983 if (!vm_pages_needed) {
985 wakeup(&vm_pages_needed);
987 tsleep(&cnt.v_free_count, PVM, "vmwait", 0);
993 * vm_await: (also see VM_AWAIT macro)
995 * asleep on an event that will signal when free pages are available
1005 if (curproc == pageproc) {
1006 vm_pageout_pages_needed = 1;
1007 asleep(&vm_pageout_pages_needed, PSWP, "vmwait", 0);
1009 if (!vm_pages_needed) {
1011 wakeup(&vm_pages_needed);
1013 asleep(&cnt.v_free_count, PVM, "vmwait", 0);
1021 * Put the specified page on the active list (if appropriate).
1022 * Ensure that act_count is at least ACT_INIT but do not otherwise
1025 * The page queues must be locked.
1026 * This routine may not block.
1029 vm_page_activate(vm_page_t m)
1036 if (m->queue != PQ_ACTIVE) {
1037 if ((m->queue - m->pc) == PQ_CACHE)
1038 cnt.v_reactivated++;
1042 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1043 m->queue = PQ_ACTIVE;
1044 vm_page_queues[PQ_ACTIVE].lcnt++;
1045 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1046 if (m->act_count < ACT_INIT)
1047 m->act_count = ACT_INIT;
1048 cnt.v_active_count++;
1051 if (m->act_count < ACT_INIT)
1052 m->act_count = ACT_INIT;
1059 * vm_page_free_wakeup:
1061 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
1062 * routine is called when a page has been added to the cache or free
1065 * This routine may not block.
1066 * This routine must be called at splvm()
1068 static __inline void
1069 vm_page_free_wakeup(void)
1072 * if pageout daemon needs pages, then tell it that there are
1075 if (vm_pageout_pages_needed &&
1076 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) {
1077 wakeup(&vm_pageout_pages_needed);
1078 vm_pageout_pages_needed = 0;
1081 * wakeup processes that are waiting on memory if we hit a
1082 * high water mark. And wakeup scheduler process if we have
1083 * lots of memory. this process will swapin processes.
1085 if (vm_pages_needed && !vm_page_count_min()) {
1086 vm_pages_needed = 0;
1087 wakeup(&cnt.v_free_count);
1094 * Returns the given page to the PQ_FREE list,
1095 * disassociating it with any VM object.
1097 * Object and page must be locked prior to entry.
1098 * This routine may not block.
1102 vm_page_free_toq(vm_page_t m)
1105 struct vpgqueues *pq;
1106 vm_object_t object = m->object;
1112 if (m->busy || ((m->queue - m->pc) == PQ_FREE) ||
1113 (m->hold_count != 0)) {
1115 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
1116 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
1118 if ((m->queue - m->pc) == PQ_FREE)
1119 panic("vm_page_free: freeing free page");
1121 panic("vm_page_free: freeing busy page");
1125 * unqueue, then remove page. Note that we cannot destroy
1126 * the page here because we do not want to call the pager's
1127 * callback routine until after we've put the page on the
1128 * appropriate free queue.
1131 vm_pageq_remove_nowakeup(m);
1135 * If fictitious remove object association and
1136 * return, otherwise delay object association removal.
1139 if ((m->flags & PG_FICTITIOUS) != 0) {
1147 if (m->wire_count != 0) {
1148 if (m->wire_count > 1) {
1149 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1150 m->wire_count, (long)m->pindex);
1152 panic("vm_page_free: freeing wired page\n");
1156 * If we've exhausted the object's resident pages we want to free
1161 (object->type == OBJT_VNODE) &&
1162 ((object->flags & OBJ_DEAD) == 0)
1164 struct vnode *vp = (struct vnode *)object->handle;
1166 if (vp && VSHOULDFREE(vp))
1171 * Clear the UNMANAGED flag when freeing an unmanaged page.
1174 if (m->flags & PG_UNMANAGED) {
1175 m->flags &= ~PG_UNMANAGED;
1178 pmap_page_is_free(m);
1182 m->queue = PQ_FREE + m->pc;
1183 pq = &vm_page_queues[m->queue];
1188 * Put zero'd pages on the end ( where we look for zero'd pages
1189 * first ) and non-zerod pages at the head.
1192 if (m->flags & PG_ZERO) {
1193 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
1194 ++vm_page_zero_count;
1196 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1199 vm_page_free_wakeup();
1207 * Prevent PV management from being done on the page. The page is
1208 * removed from the paging queues as if it were wired, and as a
1209 * consequence of no longer being managed the pageout daemon will not
1210 * touch it (since there is no way to locate the pte mappings for the
1211 * page). madvise() calls that mess with the pmap will also no longer
1212 * operate on the page.
1214 * Beyond that the page is still reasonably 'normal'. Freeing the page
1215 * will clear the flag.
1217 * This routine is used by OBJT_PHYS objects - objects using unswappable
1218 * physical memory as backing store rather then swap-backed memory and
1219 * will eventually be extended to support 4MB unmanaged physical
1224 vm_page_unmanage(vm_page_t m)
1229 if ((m->flags & PG_UNMANAGED) == 0) {
1230 if (m->wire_count == 0)
1233 vm_page_flag_set(m, PG_UNMANAGED);
1240 * Mark this page as wired down by yet
1241 * another map, removing it from paging queues
1244 * The page queues must be locked.
1245 * This routine may not block.
1248 vm_page_wire(vm_page_t m)
1253 * Only bump the wire statistics if the page is not already wired,
1254 * and only unqueue the page if it is on some queue (if it is unmanaged
1255 * it is already off the queues).
1258 if (m->wire_count == 0) {
1259 if ((m->flags & PG_UNMANAGED) == 0)
1265 vm_page_flag_set(m, PG_MAPPED);
1271 * Release one wiring of this page, potentially
1272 * enabling it to be paged again.
1274 * Many pages placed on the inactive queue should actually go
1275 * into the cache, but it is difficult to figure out which. What
1276 * we do instead, if the inactive target is well met, is to put
1277 * clean pages at the head of the inactive queue instead of the tail.
1278 * This will cause them to be moved to the cache more quickly and
1279 * if not actively re-referenced, freed more quickly. If we just
1280 * stick these pages at the end of the inactive queue, heavy filesystem
1281 * meta-data accesses can cause an unnecessary paging load on memory bound
1282 * processes. This optimization causes one-time-use metadata to be
1283 * reused more quickly.
1285 * BUT, if we are in a low-memory situation we have no choice but to
1286 * put clean pages on the cache queue.
1288 * A number of routines use vm_page_unwire() to guarantee that the page
1289 * will go into either the inactive or active queues, and will NEVER
1290 * be placed in the cache - for example, just after dirtying a page.
1291 * dirty pages in the cache are not allowed.
1293 * The page queues must be locked.
1294 * This routine may not block.
1297 vm_page_unwire(vm_page_t m, int activate)
1303 if (m->wire_count > 0) {
1305 if (m->wire_count == 0) {
1307 if (m->flags & PG_UNMANAGED) {
1309 } else if (activate) {
1310 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1311 m->queue = PQ_ACTIVE;
1312 vm_page_queues[PQ_ACTIVE].lcnt++;
1313 cnt.v_active_count++;
1315 vm_page_flag_clear(m, PG_WINATCFLS);
1316 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1317 m->queue = PQ_INACTIVE;
1318 vm_page_queues[PQ_INACTIVE].lcnt++;
1319 cnt.v_inactive_count++;
1323 panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count);
1330 * Move the specified page to the inactive queue. If the page has
1331 * any associated swap, the swap is deallocated.
1333 * Normally athead is 0 resulting in LRU operation. athead is set
1334 * to 1 if we want this page to be 'as if it were placed in the cache',
1335 * except without unmapping it from the process address space.
1337 * This routine may not block.
1339 static __inline void
1340 _vm_page_deactivate(vm_page_t m, int athead)
1346 * Ignore if already inactive.
1348 if (m->queue == PQ_INACTIVE)
1352 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1353 if ((m->queue - m->pc) == PQ_CACHE)
1354 cnt.v_reactivated++;
1355 vm_page_flag_clear(m, PG_WINATCFLS);
1358 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1360 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1361 m->queue = PQ_INACTIVE;
1362 vm_page_queues[PQ_INACTIVE].lcnt++;
1363 cnt.v_inactive_count++;
1369 vm_page_deactivate(vm_page_t m)
1371 _vm_page_deactivate(m, 0);
1375 * vm_page_try_to_cache:
1377 * Returns 0 on failure, 1 on success
1380 vm_page_try_to_cache(vm_page_t m)
1384 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1385 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1388 vm_page_test_dirty(m);
1396 * vm_page_try_to_free()
1398 * Attempt to free the page. If we cannot free it, we do nothing.
1399 * 1 is returned on success, 0 on failure.
1402 vm_page_try_to_free(vm_page_t m)
1404 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1405 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1408 vm_page_test_dirty(m);
1412 vm_page_protect(m, VM_PROT_NONE);
1420 * Put the specified page onto the page cache queue (if appropriate).
1422 * This routine may not block.
1425 vm_page_cache(vm_page_t m)
1430 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || m->wire_count) {
1431 printf("vm_page_cache: attempting to cache busy page\n");
1434 if ((m->queue - m->pc) == PQ_CACHE)
1438 * Remove all pmaps and indicate that the page is not
1439 * writeable or mapped.
1442 vm_page_protect(m, VM_PROT_NONE);
1443 if (m->dirty != 0) {
1444 panic("vm_page_cache: caching a dirty page, pindex: %ld",
1448 vm_pageq_remove_nowakeup(m);
1449 m->queue = PQ_CACHE + m->pc;
1450 vm_page_queues[m->queue].lcnt++;
1451 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq);
1452 cnt.v_cache_count++;
1453 vm_page_free_wakeup();
1460 * Cache, deactivate, or do nothing as appropriate. This routine
1461 * is typically used by madvise() MADV_DONTNEED.
1463 * Generally speaking we want to move the page into the cache so
1464 * it gets reused quickly. However, this can result in a silly syndrome
1465 * due to the page recycling too quickly. Small objects will not be
1466 * fully cached. On the otherhand, if we move the page to the inactive
1467 * queue we wind up with a problem whereby very large objects
1468 * unnecessarily blow away our inactive and cache queues.
1470 * The solution is to move the pages based on a fixed weighting. We
1471 * either leave them alone, deactivate them, or move them to the cache,
1472 * where moving them to the cache has the highest weighting.
1473 * By forcing some pages into other queues we eventually force the
1474 * system to balance the queues, potentially recovering other unrelated
1475 * space from active. The idea is to not force this to happen too
1480 vm_page_dontneed(vm_page_t m)
1482 static int dnweight;
1490 * occassionally leave the page alone
1493 if ((dnw & 0x01F0) == 0 ||
1494 m->queue == PQ_INACTIVE ||
1495 m->queue - m->pc == PQ_CACHE
1497 if (m->act_count >= ACT_INIT)
1503 vm_page_test_dirty(m);
1505 if (m->dirty || (dnw & 0x0070) == 0) {
1507 * Deactivate the page 3 times out of 32.
1512 * Cache the page 28 times out of every 32. Note that
1513 * the page is deactivated instead of cached, but placed
1514 * at the head of the queue instead of the tail.
1518 _vm_page_deactivate(m, head);
1522 * Grab a page, waiting until we are waken up due to the page
1523 * changing state. We keep on waiting, if the page continues
1524 * to be in the object. If the page doesn't exist, allocate it.
1526 * This routine may block.
1529 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
1536 if ((m = vm_page_lookup(object, pindex)) != NULL) {
1537 if (m->busy || (m->flags & PG_BUSY)) {
1538 generation = object->generation;
1541 while ((object->generation == generation) &&
1542 (m->busy || (m->flags & PG_BUSY))) {
1543 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
1544 tsleep(m, PVM, "pgrbwt", 0);
1545 if ((allocflags & VM_ALLOC_RETRY) == 0) {
1558 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1561 if ((allocflags & VM_ALLOC_RETRY) == 0)
1570 * Mapping function for valid bits or for dirty bits in
1571 * a page. May not block.
1573 * Inputs are required to range within a page.
1577 vm_page_bits(int base, int size)
1583 base + size <= PAGE_SIZE,
1584 ("vm_page_bits: illegal base/size %d/%d", base, size)
1587 if (size == 0) /* handle degenerate case */
1590 first_bit = base >> DEV_BSHIFT;
1591 last_bit = (base + size - 1) >> DEV_BSHIFT;
1593 return ((2 << last_bit) - (1 << first_bit));
1597 * vm_page_set_validclean:
1599 * Sets portions of a page valid and clean. The arguments are expected
1600 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1601 * of any partial chunks touched by the range. The invalid portion of
1602 * such chunks will be zero'd.
1604 * This routine may not block.
1606 * (base + size) must be less then or equal to PAGE_SIZE.
1609 vm_page_set_validclean(vm_page_t m, int base, int size)
1616 if (size == 0) /* handle degenerate case */
1620 * If the base is not DEV_BSIZE aligned and the valid
1621 * bit is clear, we have to zero out a portion of the
1625 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
1626 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
1628 pmap_zero_page_area(
1636 * If the ending offset is not DEV_BSIZE aligned and the
1637 * valid bit is clear, we have to zero out a portion of
1641 endoff = base + size;
1643 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
1644 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
1646 pmap_zero_page_area(
1649 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
1654 * Set valid, clear dirty bits. If validating the entire
1655 * page we can safely clear the pmap modify bit. We also
1656 * use this opportunity to clear the PG_NOSYNC flag. If a process
1657 * takes a write fault on a MAP_NOSYNC memory area the flag will
1661 pagebits = vm_page_bits(base, size);
1662 m->valid |= pagebits;
1663 m->dirty &= ~pagebits;
1664 if (base == 0 && size == PAGE_SIZE) {
1665 pmap_clear_modify(m);
1666 vm_page_flag_clear(m, PG_NOSYNC);
1673 vm_page_set_dirty(vm_page_t m, int base, int size)
1675 m->dirty |= vm_page_bits(base, size);
1681 vm_page_clear_dirty(vm_page_t m, int base, int size)
1684 m->dirty &= ~vm_page_bits(base, size);
1688 * vm_page_set_invalid:
1690 * Invalidates DEV_BSIZE'd chunks within a page. Both the
1691 * valid and dirty bits for the effected areas are cleared.
1696 vm_page_set_invalid(vm_page_t m, int base, int size)
1701 bits = vm_page_bits(base, size);
1704 m->object->generation++;
1708 * vm_page_zero_invalid()
1710 * The kernel assumes that the invalid portions of a page contain
1711 * garbage, but such pages can be mapped into memory by user code.
1712 * When this occurs, we must zero out the non-valid portions of the
1713 * page so user code sees what it expects.
1715 * Pages are most often semi-valid when the end of a file is mapped
1716 * into memory and the file's size is not page aligned.
1720 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
1726 * Scan the valid bits looking for invalid sections that
1727 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
1728 * valid bit may be set ) have already been zerod by
1729 * vm_page_set_validclean().
1732 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
1733 if (i == (PAGE_SIZE / DEV_BSIZE) ||
1734 (m->valid & (1 << i))
1737 pmap_zero_page_area(
1740 (i - b) << DEV_BSHIFT
1748 * setvalid is TRUE when we can safely set the zero'd areas
1749 * as being valid. We can do this if there are no cache consistancy
1750 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
1754 m->valid = VM_PAGE_BITS_ALL;
1760 * Is (partial) page valid? Note that the case where size == 0
1761 * will return FALSE in the degenerate case where the page is
1762 * 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.
1783 vm_page_test_dirty(vm_page_t m)
1785 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
1790 #include "opt_ddb.h"
1792 #include <sys/kernel.h>
1794 #include <ddb/ddb.h>
1796 DB_SHOW_COMMAND(page, vm_page_print_page_info)
1798 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count);
1799 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count);
1800 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count);
1801 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count);
1802 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count);
1803 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved);
1804 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min);
1805 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target);
1806 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min);
1807 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target);
1810 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
1813 db_printf("PQ_FREE:");
1814 for (i = 0; i < PQ_L2_SIZE; i++) {
1815 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
1819 db_printf("PQ_CACHE:");
1820 for (i = 0; i < PQ_L2_SIZE; i++) {
1821 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
1825 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
1826 vm_page_queues[PQ_ACTIVE].lcnt,
1827 vm_page_queues[PQ_INACTIVE].lcnt);