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
37 * $FreeBSD: src/sys/vm/vm_page.c,v 1.147.2.18 2002/03/10 05:03:19 alc Exp $
38 * $DragonFly: src/sys/vm/vm_page.c,v 1.4 2003/06/22 17:39:48 dillon Exp $
42 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
43 * All rights reserved.
45 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
47 * Permission to use, copy, modify and distribute this software and
48 * its documentation is hereby granted, provided that both the copyright
49 * notice and this permission notice appear in all copies of the
50 * software, derivative works or modified versions, and any portions
51 * thereof, and that both notices appear in supporting documentation.
53 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
54 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
55 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
57 * Carnegie Mellon requests users of this software to return to
59 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
60 * School of Computer Science
61 * Carnegie Mellon University
62 * Pittsburgh PA 15213-3890
64 * any improvements or extensions that they make and grant Carnegie the
65 * rights to redistribute these changes.
69 * Resident memory management module.
72 #include <sys/param.h>
73 #include <sys/systm.h>
74 #include <sys/malloc.h>
76 #include <sys/vmmeter.h>
77 #include <sys/vnode.h>
80 #include <vm/vm_param.h>
82 #include <vm/vm_kern.h>
84 #include <vm/vm_map.h>
85 #include <vm/vm_object.h>
86 #include <vm/vm_page.h>
87 #include <vm/vm_pageout.h>
88 #include <vm/vm_pager.h>
89 #include <vm/vm_extern.h>
91 static void vm_page_queue_init (void);
92 static vm_page_t vm_page_select_cache (vm_object_t, vm_pindex_t);
95 * Associated with page of user-allocatable memory is a
99 static struct vm_page **vm_page_buckets; /* Array of buckets */
100 static int vm_page_bucket_count; /* How big is array? */
101 static int vm_page_hash_mask; /* Mask for hash function */
102 static volatile int vm_page_bucket_generation;
104 struct vpgqueues vm_page_queues[PQ_COUNT];
107 vm_page_queue_init(void) {
110 for(i=0;i<PQ_L2_SIZE;i++) {
111 vm_page_queues[PQ_FREE+i].cnt = &cnt.v_free_count;
113 vm_page_queues[PQ_INACTIVE].cnt = &cnt.v_inactive_count;
115 vm_page_queues[PQ_ACTIVE].cnt = &cnt.v_active_count;
116 vm_page_queues[PQ_HOLD].cnt = &cnt.v_active_count;
117 for(i=0;i<PQ_L2_SIZE;i++) {
118 vm_page_queues[PQ_CACHE+i].cnt = &cnt.v_cache_count;
120 for(i=0;i<PQ_COUNT;i++) {
121 TAILQ_INIT(&vm_page_queues[i].pl);
125 vm_page_t vm_page_array = 0;
126 int vm_page_array_size = 0;
128 int vm_page_zero_count = 0;
130 static __inline int vm_page_hash (vm_object_t object, vm_pindex_t pindex);
131 static void vm_page_free_wakeup (void);
136 * Sets the page size, perhaps based upon the memory
137 * size. Must be called before any use of page-size
138 * dependent functions.
141 vm_set_page_size(void)
143 if (cnt.v_page_size == 0)
144 cnt.v_page_size = PAGE_SIZE;
145 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0)
146 panic("vm_set_page_size: page size not a power of two");
152 * Add a new page to the freelist for use by the system.
153 * Must be called at splhigh().
156 vm_add_new_page(vm_offset_t pa)
162 m = PHYS_TO_VM_PAGE(pa);
165 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK;
166 m->queue = m->pc + PQ_FREE;
167 TAILQ_INSERT_HEAD(&vm_page_queues[m->queue].pl, m, pageq);
168 vm_page_queues[m->queue].lcnt++;
175 * Initializes the resident memory module.
177 * Allocates memory for the page cells, and
178 * for the object/offset-to-page hash table headers.
179 * Each page cell is initialized and placed on the free list.
183 vm_page_startup(vm_offset_t starta, vm_offset_t enda, vm_offset_t vaddr)
186 struct vm_page **bucket;
187 vm_size_t npages, page_range;
194 /* the biggest memory array is the second group of pages */
196 vm_offset_t biggestone, biggestsize;
204 vaddr = round_page(vaddr);
206 for (i = 0; phys_avail[i + 1]; i += 2) {
207 phys_avail[i] = round_page(phys_avail[i]);
208 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
211 for (i = 0; phys_avail[i + 1]; i += 2) {
212 int size = phys_avail[i + 1] - phys_avail[i];
214 if (size > biggestsize) {
222 end = phys_avail[biggestone+1];
225 * Initialize the queue headers for the free queue, the active queue
226 * and the inactive queue.
229 vm_page_queue_init();
232 * Allocate (and initialize) the hash table buckets.
234 * The number of buckets MUST BE a power of 2, and the actual value is
235 * the next power of 2 greater than the number of physical pages in
238 * We make the hash table approximately 2x the number of pages to
239 * reduce the chain length. This is about the same size using the
240 * singly-linked list as the 1x hash table we were using before
241 * using TAILQ but the chain length will be smaller.
243 * Note: This computation can be tweaked if desired.
245 vm_page_buckets = (struct vm_page **)vaddr;
246 bucket = vm_page_buckets;
247 if (vm_page_bucket_count == 0) {
248 vm_page_bucket_count = 1;
249 while (vm_page_bucket_count < atop(total))
250 vm_page_bucket_count <<= 1;
252 vm_page_bucket_count <<= 1;
253 vm_page_hash_mask = vm_page_bucket_count - 1;
256 * Validate these addresses.
258 new_end = end - vm_page_bucket_count * sizeof(struct vm_page *);
259 new_end = trunc_page(new_end);
260 mapped = round_page(vaddr);
261 vaddr = pmap_map(mapped, new_end, end,
262 VM_PROT_READ | VM_PROT_WRITE);
263 vaddr = round_page(vaddr);
264 bzero((caddr_t) mapped, vaddr - mapped);
266 for (i = 0; i < vm_page_bucket_count; i++) {
272 * Compute the number of pages of memory that will be available for
273 * use (taking into account the overhead of a page structure per
277 first_page = phys_avail[0] / PAGE_SIZE;
279 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
280 npages = (total - (page_range * sizeof(struct vm_page)) -
281 (end - new_end)) / PAGE_SIZE;
285 * Initialize the mem entry structures now, and put them in the free
288 vm_page_array = (vm_page_t) vaddr;
292 * Validate these addresses.
295 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
296 mapped = pmap_map(mapped, new_end, end,
297 VM_PROT_READ | VM_PROT_WRITE);
300 * Clear all of the page structures
302 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
303 vm_page_array_size = page_range;
306 * Construct the free queue(s) in descending order (by physical
307 * address) so that the first 16MB of physical memory is allocated
308 * last rather than first. On large-memory machines, this avoids
309 * the exhaustion of low physical memory before isa_dmainit has run.
311 cnt.v_page_count = 0;
312 cnt.v_free_count = 0;
313 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
318 last_pa = phys_avail[i + 1];
319 while (pa < last_pa && npages-- > 0) {
330 * Distributes the object/offset key pair among hash buckets.
332 * NOTE: This macro depends on vm_page_bucket_count being a power of 2.
333 * This routine may not block.
335 * We try to randomize the hash based on the object to spread the pages
336 * out in the hash table without it costing us too much.
339 vm_page_hash(vm_object_t object, vm_pindex_t pindex)
341 int i = ((uintptr_t)object + pindex) ^ object->hash_rand;
343 return(i & vm_page_hash_mask);
347 vm_page_unhold(vm_page_t mem)
350 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
351 if (mem->hold_count == 0 && mem->queue == PQ_HOLD)
352 vm_page_free_toq(mem);
356 * vm_page_insert: [ internal use only ]
358 * Inserts the given mem entry into the object and object list.
360 * The pagetables are not updated but will presumably fault the page
361 * in if necessary, or if a kernel page the caller will at some point
362 * enter the page into the kernel's pmap. We are not allowed to block
363 * here so we *can't* do this anyway.
365 * The object and page must be locked, and must be splhigh.
366 * This routine may not block.
370 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
372 struct vm_page **bucket;
374 if (m->object != NULL)
375 panic("vm_page_insert: already inserted");
378 * Record the object/offset pair in this page
385 * Insert it into the object_object/offset hash table
388 bucket = &vm_page_buckets[vm_page_hash(object, pindex)];
391 vm_page_bucket_generation++;
394 * Now link into the object's list of backed pages.
397 TAILQ_INSERT_TAIL(&object->memq, m, listq);
398 object->generation++;
401 * show that the object has one more resident page.
404 object->resident_page_count++;
407 * Since we are inserting a new and possibly dirty page,
408 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
410 if (m->flags & PG_WRITEABLE)
411 vm_object_set_writeable_dirty(object);
416 * NOTE: used by device pager as well -wfj
418 * Removes the given mem entry from the object/offset-page
419 * table and the object page list, but do not invalidate/terminate
422 * The object and page must be locked, and at splhigh.
423 * The underlying pmap entry (if any) is NOT removed here.
424 * This routine may not block.
428 vm_page_remove(vm_page_t m)
432 if (m->object == NULL)
435 if ((m->flags & PG_BUSY) == 0) {
436 panic("vm_page_remove: page not busy");
440 * Basically destroy the page.
448 * Remove from the object_object/offset hash table. The object
449 * must be on the hash queue, we will panic if it isn't
451 * Note: we must NULL-out m->hnext to prevent loops in detached
452 * buffers with vm_page_lookup().
456 struct vm_page **bucket;
458 bucket = &vm_page_buckets[vm_page_hash(m->object, m->pindex)];
459 while (*bucket != m) {
461 panic("vm_page_remove(): page not found in hash");
462 bucket = &(*bucket)->hnext;
466 vm_page_bucket_generation++;
470 * Now remove from the object's list of backed pages.
473 TAILQ_REMOVE(&object->memq, m, listq);
476 * And show that the object has one fewer resident page.
479 object->resident_page_count--;
480 object->generation++;
488 * Returns the page associated with the object/offset
489 * pair specified; if none is found, NULL is returned.
491 * NOTE: the code below does not lock. It will operate properly if
492 * an interrupt makes a change, but the generation algorithm will not
493 * operate properly in an SMP environment where both cpu's are able to run
494 * kernel code simultaneously.
496 * The object must be locked. No side effects.
497 * This routine may not block.
498 * This is a critical path routine
502 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
505 struct vm_page **bucket;
509 * Search the hash table for this object/offset pair
513 generation = vm_page_bucket_generation;
514 bucket = &vm_page_buckets[vm_page_hash(object, pindex)];
515 for (m = *bucket; m != NULL; m = m->hnext) {
516 if ((m->object == object) && (m->pindex == pindex)) {
517 if (vm_page_bucket_generation != generation)
522 if (vm_page_bucket_generation != generation)
530 * Move the given memory entry from its
531 * current object to the specified target object/offset.
533 * The object must be locked.
534 * This routine may not block.
536 * Note: this routine will raise itself to splvm(), the caller need not.
538 * Note: swap associated with the page must be invalidated by the move. We
539 * have to do this for several reasons: (1) we aren't freeing the
540 * page, (2) we are dirtying the page, (3) the VM system is probably
541 * moving the page from object A to B, and will then later move
542 * the backing store from A to B and we can't have a conflict.
544 * Note: we *always* dirty the page. It is necessary both for the
545 * fact that we moved it, and because we may be invalidating
546 * swap. If the page is on the cache, we have to deactivate it
547 * or vm_page_dirty() will panic. Dirty pages are not allowed
552 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
558 vm_page_insert(m, new_object, new_pindex);
559 if (m->queue - m->pc == PQ_CACHE)
560 vm_page_deactivate(m);
566 * vm_page_unqueue_nowakeup:
568 * vm_page_unqueue() without any wakeup
570 * This routine must be called at splhigh().
571 * This routine may not block.
575 vm_page_unqueue_nowakeup(vm_page_t m)
577 int queue = m->queue;
578 struct vpgqueues *pq;
579 if (queue != PQ_NONE) {
580 pq = &vm_page_queues[queue];
582 TAILQ_REMOVE(&pq->pl, m, pageq);
591 * Remove a page from its queue.
593 * This routine must be called at splhigh().
594 * This routine may not block.
598 vm_page_unqueue(vm_page_t m)
600 int queue = m->queue;
601 struct vpgqueues *pq;
602 if (queue != PQ_NONE) {
604 pq = &vm_page_queues[queue];
605 TAILQ_REMOVE(&pq->pl, m, pageq);
608 if ((queue - m->pc) == PQ_CACHE) {
609 if (vm_paging_needed())
620 * Find a page on the specified queue with color optimization.
622 * The page coloring optimization attempts to locate a page
623 * that does not overload other nearby pages in the object in
624 * the cpu's L1 or L2 caches. We need this optimization because
625 * cpu caches tend to be physical caches, while object spaces tend
628 * This routine must be called at splvm().
629 * This routine may not block.
631 * This routine may only be called from the vm_page_list_find() macro
635 _vm_page_list_find(int basequeue, int index)
639 struct vpgqueues *pq;
641 pq = &vm_page_queues[basequeue];
644 * Note that for the first loop, index+i and index-i wind up at the
645 * same place. Even though this is not totally optimal, we've already
646 * blown it by missing the cache case so we do not care.
649 for(i = PQ_L2_SIZE / 2; i > 0; --i) {
650 if ((m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl)) != NULL)
653 if ((m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl)) != NULL)
662 * vm_page_select_cache:
664 * Find a page on the cache queue with color optimization. As pages
665 * might be found, but not applicable, they are deactivated. This
666 * keeps us from using potentially busy cached pages.
668 * This routine must be called at splvm().
669 * This routine may not block.
672 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex)
677 m = vm_page_list_find(
679 (pindex + object->pg_color) & PQ_L2_MASK,
682 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
683 m->hold_count || m->wire_count)) {
684 vm_page_deactivate(m);
692 * vm_page_select_free:
694 * Find a free or zero page, with specified preference. We attempt to
695 * inline the nominal case and fall back to _vm_page_select_free()
698 * This routine must be called at splvm().
699 * This routine may not block.
702 static __inline vm_page_t
703 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero)
707 m = vm_page_list_find(
709 (pindex + object->pg_color) & PQ_L2_MASK,
718 * Allocate and return a memory cell associated
719 * with this VM object/offset pair.
722 * VM_ALLOC_NORMAL normal process request
723 * VM_ALLOC_SYSTEM system *really* needs a page
724 * VM_ALLOC_INTERRUPT interrupt time request
725 * VM_ALLOC_ZERO zero page
727 * Object must be locked.
728 * This routine may not block.
730 * Additional special handling is required when called from an
731 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with
732 * the page cache in this case.
736 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
741 KASSERT(!vm_page_lookup(object, pindex),
742 ("vm_page_alloc: page already allocated"));
745 * The pager is allowed to eat deeper into the free page list.
748 if ((curthread == pagethread) && (page_req != VM_ALLOC_INTERRUPT)) {
749 page_req = VM_ALLOC_SYSTEM;
755 if (cnt.v_free_count > cnt.v_free_reserved) {
757 * Allocate from the free queue if there are plenty of pages
760 if (page_req == VM_ALLOC_ZERO)
761 m = vm_page_select_free(object, pindex, TRUE);
763 m = vm_page_select_free(object, pindex, FALSE);
765 (page_req == VM_ALLOC_SYSTEM &&
766 cnt.v_cache_count == 0 &&
767 cnt.v_free_count > cnt.v_interrupt_free_min) ||
768 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0)
771 * Interrupt or system, dig deeper into the free list.
773 m = vm_page_select_free(object, pindex, FALSE);
774 } else if (page_req != VM_ALLOC_INTERRUPT) {
776 * Allocatable from cache (non-interrupt only). On success,
777 * we must free the page and try again, thus ensuring that
778 * cnt.v_*_free_min counters are replenished.
780 m = vm_page_select_cache(object, pindex);
783 #if defined(DIAGNOSTIC)
784 if (cnt.v_cache_count > 0)
785 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count);
787 vm_pageout_deficit++;
791 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m));
793 vm_page_protect(m, VM_PROT_NONE);
798 * Not allocatable from cache from interrupt, give up.
801 vm_pageout_deficit++;
807 * At this point we had better have found a good page.
812 ("vm_page_alloc(): missing page on free queue\n")
816 * Remove from free queue
819 vm_page_unqueue_nowakeup(m);
822 * Initialize structure. Only the PG_ZERO flag is inherited.
825 if (m->flags & PG_ZERO) {
826 vm_page_zero_count--;
827 m->flags = PG_ZERO | PG_BUSY;
836 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m));
839 * vm_page_insert() is safe prior to the splx(). Note also that
840 * inserting a page here does not insert it into the pmap (which
841 * could cause us to block allocating memory). We cannot block
845 vm_page_insert(m, object, pindex);
848 * Don't wakeup too often - wakeup the pageout daemon when
849 * we would be nearly out of memory.
851 if (vm_paging_needed())
860 * vm_wait: (also see VM_WAIT macro)
862 * Block until free pages are available for allocation
863 * - Called in various places before memory allocations.
872 if (curthread == pagethread) {
873 vm_pageout_pages_needed = 1;
874 tsleep(&vm_pageout_pages_needed, PSWP, "VMWait", 0);
876 if (!vm_pages_needed) {
878 wakeup(&vm_pages_needed);
880 tsleep(&cnt.v_free_count, PVM, "vmwait", 0);
886 * vm_waitpfault: (also see VM_WAITPFAULT macro)
888 * Block until free pages are available for allocation
889 * - Called only in vm_fault so that processes page faulting
890 * can be easily tracked.
891 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
892 * processes will be able to grab memory first. Do not change
893 * this balance without careful testing first.
902 if (!vm_pages_needed) {
904 wakeup(&vm_pages_needed);
906 tsleep(&cnt.v_free_count, PUSER, "pfault", 0);
913 * Put the specified page on the active list (if appropriate).
914 * Ensure that act_count is at least ACT_INIT but do not otherwise
917 * The page queues must be locked.
918 * This routine may not block.
921 vm_page_activate(vm_page_t m)
926 if (m->queue != PQ_ACTIVE) {
927 if ((m->queue - m->pc) == PQ_CACHE)
932 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
933 m->queue = PQ_ACTIVE;
934 vm_page_queues[PQ_ACTIVE].lcnt++;
935 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
936 if (m->act_count < ACT_INIT)
937 m->act_count = ACT_INIT;
938 cnt.v_active_count++;
941 if (m->act_count < ACT_INIT)
942 m->act_count = ACT_INIT;
949 * vm_page_free_wakeup:
951 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
952 * routine is called when a page has been added to the cache or free
955 * This routine may not block.
956 * This routine must be called at splvm()
959 vm_page_free_wakeup(void)
962 * if pageout daemon needs pages, then tell it that there are
965 if (vm_pageout_pages_needed &&
966 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) {
967 wakeup(&vm_pageout_pages_needed);
968 vm_pageout_pages_needed = 0;
971 * wakeup processes that are waiting on memory if we hit a
972 * high water mark. And wakeup scheduler process if we have
973 * lots of memory. this process will swapin processes.
975 if (vm_pages_needed && !vm_page_count_min()) {
977 wakeup(&cnt.v_free_count);
984 * Returns the given page to the PQ_FREE list,
985 * disassociating it with any VM object.
987 * Object and page must be locked prior to entry.
988 * This routine may not block.
992 vm_page_free_toq(vm_page_t m)
995 struct vpgqueues *pq;
996 vm_object_t object = m->object;
1002 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
1004 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
1005 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
1007 if ((m->queue - m->pc) == PQ_FREE)
1008 panic("vm_page_free: freeing free page");
1010 panic("vm_page_free: freeing busy page");
1014 * unqueue, then remove page. Note that we cannot destroy
1015 * the page here because we do not want to call the pager's
1016 * callback routine until after we've put the page on the
1017 * appropriate free queue.
1020 vm_page_unqueue_nowakeup(m);
1024 * If fictitious remove object association and
1025 * return, otherwise delay object association removal.
1028 if ((m->flags & PG_FICTITIOUS) != 0) {
1036 if (m->wire_count != 0) {
1037 if (m->wire_count > 1) {
1038 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1039 m->wire_count, (long)m->pindex);
1041 panic("vm_page_free: freeing wired page\n");
1045 * If we've exhausted the object's resident pages we want to free
1050 (object->type == OBJT_VNODE) &&
1051 ((object->flags & OBJ_DEAD) == 0)
1053 struct vnode *vp = (struct vnode *)object->handle;
1055 if (vp && VSHOULDFREE(vp))
1060 * Clear the UNMANAGED flag when freeing an unmanaged page.
1063 if (m->flags & PG_UNMANAGED) {
1064 m->flags &= ~PG_UNMANAGED;
1067 pmap_page_is_free(m);
1071 if (m->hold_count != 0) {
1072 m->flags &= ~PG_ZERO;
1075 m->queue = PQ_FREE + m->pc;
1076 pq = &vm_page_queues[m->queue];
1081 * Put zero'd pages on the end ( where we look for zero'd pages
1082 * first ) and non-zerod pages at the head.
1085 if (m->flags & PG_ZERO) {
1086 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
1087 ++vm_page_zero_count;
1089 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1092 vm_page_free_wakeup();
1100 * Prevent PV management from being done on the page. The page is
1101 * removed from the paging queues as if it were wired, and as a
1102 * consequence of no longer being managed the pageout daemon will not
1103 * touch it (since there is no way to locate the pte mappings for the
1104 * page). madvise() calls that mess with the pmap will also no longer
1105 * operate on the page.
1107 * Beyond that the page is still reasonably 'normal'. Freeing the page
1108 * will clear the flag.
1110 * This routine is used by OBJT_PHYS objects - objects using unswappable
1111 * physical memory as backing store rather then swap-backed memory and
1112 * will eventually be extended to support 4MB unmanaged physical
1117 vm_page_unmanage(vm_page_t m)
1122 if ((m->flags & PG_UNMANAGED) == 0) {
1123 if (m->wire_count == 0)
1126 vm_page_flag_set(m, PG_UNMANAGED);
1133 * Mark this page as wired down by yet
1134 * another map, removing it from paging queues
1137 * The page queues must be locked.
1138 * This routine may not block.
1141 vm_page_wire(vm_page_t m)
1146 * Only bump the wire statistics if the page is not already wired,
1147 * and only unqueue the page if it is on some queue (if it is unmanaged
1148 * it is already off the queues).
1151 if (m->wire_count == 0) {
1152 if ((m->flags & PG_UNMANAGED) == 0)
1157 KASSERT(m->wire_count != 0,
1158 ("vm_page_wire: wire_count overflow m=%p", m));
1161 vm_page_flag_set(m, PG_MAPPED);
1167 * Release one wiring of this page, potentially
1168 * enabling it to be paged again.
1170 * Many pages placed on the inactive queue should actually go
1171 * into the cache, but it is difficult to figure out which. What
1172 * we do instead, if the inactive target is well met, is to put
1173 * clean pages at the head of the inactive queue instead of the tail.
1174 * This will cause them to be moved to the cache more quickly and
1175 * if not actively re-referenced, freed more quickly. If we just
1176 * stick these pages at the end of the inactive queue, heavy filesystem
1177 * meta-data accesses can cause an unnecessary paging load on memory bound
1178 * processes. This optimization causes one-time-use metadata to be
1179 * reused more quickly.
1181 * BUT, if we are in a low-memory situation we have no choice but to
1182 * put clean pages on the cache queue.
1184 * A number of routines use vm_page_unwire() to guarantee that the page
1185 * will go into either the inactive or active queues, and will NEVER
1186 * be placed in the cache - for example, just after dirtying a page.
1187 * dirty pages in the cache are not allowed.
1189 * The page queues must be locked.
1190 * This routine may not block.
1193 vm_page_unwire(vm_page_t m, int activate)
1199 if (m->wire_count > 0) {
1201 if (m->wire_count == 0) {
1203 if (m->flags & PG_UNMANAGED) {
1205 } else if (activate) {
1206 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1207 m->queue = PQ_ACTIVE;
1208 vm_page_queues[PQ_ACTIVE].lcnt++;
1209 cnt.v_active_count++;
1211 vm_page_flag_clear(m, PG_WINATCFLS);
1212 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1213 m->queue = PQ_INACTIVE;
1214 vm_page_queues[PQ_INACTIVE].lcnt++;
1215 cnt.v_inactive_count++;
1219 panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count);
1226 * Move the specified page to the inactive queue. If the page has
1227 * any associated swap, the swap is deallocated.
1229 * Normally athead is 0 resulting in LRU operation. athead is set
1230 * to 1 if we want this page to be 'as if it were placed in the cache',
1231 * except without unmapping it from the process address space.
1233 * This routine may not block.
1235 static __inline void
1236 _vm_page_deactivate(vm_page_t m, int athead)
1241 * Ignore if already inactive.
1243 if (m->queue == PQ_INACTIVE)
1247 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1248 if ((m->queue - m->pc) == PQ_CACHE)
1249 cnt.v_reactivated++;
1250 vm_page_flag_clear(m, PG_WINATCFLS);
1253 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1255 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1256 m->queue = PQ_INACTIVE;
1257 vm_page_queues[PQ_INACTIVE].lcnt++;
1258 cnt.v_inactive_count++;
1264 vm_page_deactivate(vm_page_t m)
1266 _vm_page_deactivate(m, 0);
1270 * vm_page_try_to_cache:
1272 * Returns 0 on failure, 1 on success
1275 vm_page_try_to_cache(vm_page_t m)
1277 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1278 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1281 vm_page_test_dirty(m);
1289 * vm_page_try_to_free()
1291 * Attempt to free the page. If we cannot free it, we do nothing.
1292 * 1 is returned on success, 0 on failure.
1296 vm_page_try_to_free(vm_page_t m)
1298 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1299 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1302 vm_page_test_dirty(m);
1306 vm_page_protect(m, VM_PROT_NONE);
1315 * Put the specified page onto the page cache queue (if appropriate).
1317 * This routine may not block.
1320 vm_page_cache(vm_page_t m)
1324 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || m->wire_count) {
1325 printf("vm_page_cache: attempting to cache busy page\n");
1328 if ((m->queue - m->pc) == PQ_CACHE)
1332 * Remove all pmaps and indicate that the page is not
1333 * writeable or mapped.
1336 vm_page_protect(m, VM_PROT_NONE);
1337 if (m->dirty != 0) {
1338 panic("vm_page_cache: caching a dirty page, pindex: %ld",
1342 vm_page_unqueue_nowakeup(m);
1343 m->queue = PQ_CACHE + m->pc;
1344 vm_page_queues[m->queue].lcnt++;
1345 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq);
1346 cnt.v_cache_count++;
1347 vm_page_free_wakeup();
1354 * Cache, deactivate, or do nothing as appropriate. This routine
1355 * is typically used by madvise() MADV_DONTNEED.
1357 * Generally speaking we want to move the page into the cache so
1358 * it gets reused quickly. However, this can result in a silly syndrome
1359 * due to the page recycling too quickly. Small objects will not be
1360 * fully cached. On the otherhand, if we move the page to the inactive
1361 * queue we wind up with a problem whereby very large objects
1362 * unnecessarily blow away our inactive and cache queues.
1364 * The solution is to move the pages based on a fixed weighting. We
1365 * either leave them alone, deactivate them, or move them to the cache,
1366 * where moving them to the cache has the highest weighting.
1367 * By forcing some pages into other queues we eventually force the
1368 * system to balance the queues, potentially recovering other unrelated
1369 * space from active. The idea is to not force this to happen too
1374 vm_page_dontneed(vm_page_t m)
1376 static int dnweight;
1383 * occassionally leave the page alone
1386 if ((dnw & 0x01F0) == 0 ||
1387 m->queue == PQ_INACTIVE ||
1388 m->queue - m->pc == PQ_CACHE
1390 if (m->act_count >= ACT_INIT)
1396 vm_page_test_dirty(m);
1398 if (m->dirty || (dnw & 0x0070) == 0) {
1400 * Deactivate the page 3 times out of 32.
1405 * Cache the page 28 times out of every 32. Note that
1406 * the page is deactivated instead of cached, but placed
1407 * at the head of the queue instead of the tail.
1411 _vm_page_deactivate(m, head);
1415 * Grab a page, waiting until we are waken up due to the page
1416 * changing state. We keep on waiting, if the page continues
1417 * to be in the object. If the page doesn't exist, allocate it.
1419 * This routine may block.
1422 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
1429 if ((m = vm_page_lookup(object, pindex)) != NULL) {
1430 if (m->busy || (m->flags & PG_BUSY)) {
1431 generation = object->generation;
1434 while ((object->generation == generation) &&
1435 (m->busy || (m->flags & PG_BUSY))) {
1436 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
1437 tsleep(m, PVM, "pgrbwt", 0);
1438 if ((allocflags & VM_ALLOC_RETRY) == 0) {
1451 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1454 if ((allocflags & VM_ALLOC_RETRY) == 0)
1463 * Mapping function for valid bits or for dirty bits in
1464 * a page. May not block.
1466 * Inputs are required to range within a page.
1470 vm_page_bits(int base, int size)
1476 base + size <= PAGE_SIZE,
1477 ("vm_page_bits: illegal base/size %d/%d", base, size)
1480 if (size == 0) /* handle degenerate case */
1483 first_bit = base >> DEV_BSHIFT;
1484 last_bit = (base + size - 1) >> DEV_BSHIFT;
1486 return ((2 << last_bit) - (1 << first_bit));
1490 * vm_page_set_validclean:
1492 * Sets portions of a page valid and clean. The arguments are expected
1493 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1494 * of any partial chunks touched by the range. The invalid portion of
1495 * such chunks will be zero'd.
1497 * This routine may not block.
1499 * (base + size) must be less then or equal to PAGE_SIZE.
1502 vm_page_set_validclean(vm_page_t m, int base, int size)
1508 if (size == 0) /* handle degenerate case */
1512 * If the base is not DEV_BSIZE aligned and the valid
1513 * bit is clear, we have to zero out a portion of the
1517 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
1518 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
1520 pmap_zero_page_area(
1528 * If the ending offset is not DEV_BSIZE aligned and the
1529 * valid bit is clear, we have to zero out a portion of
1533 endoff = base + size;
1535 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
1536 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
1538 pmap_zero_page_area(
1541 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
1546 * Set valid, clear dirty bits. If validating the entire
1547 * page we can safely clear the pmap modify bit. We also
1548 * use this opportunity to clear the PG_NOSYNC flag. If a process
1549 * takes a write fault on a MAP_NOSYNC memory area the flag will
1552 * We set valid bits inclusive of any overlap, but we can only
1553 * clear dirty bits for DEV_BSIZE chunks that are fully within
1557 pagebits = vm_page_bits(base, size);
1558 m->valid |= pagebits;
1560 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
1561 frag = DEV_BSIZE - frag;
1567 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
1569 m->dirty &= ~pagebits;
1570 if (base == 0 && size == PAGE_SIZE) {
1571 pmap_clear_modify(m);
1572 vm_page_flag_clear(m, PG_NOSYNC);
1579 vm_page_set_dirty(vm_page_t m, int base, int size)
1581 m->dirty |= vm_page_bits(base, size);
1587 vm_page_clear_dirty(vm_page_t m, int base, int size)
1589 m->dirty &= ~vm_page_bits(base, size);
1593 * vm_page_set_invalid:
1595 * Invalidates DEV_BSIZE'd chunks within a page. Both the
1596 * valid and dirty bits for the effected areas are cleared.
1601 vm_page_set_invalid(vm_page_t m, int base, int size)
1605 bits = vm_page_bits(base, size);
1608 m->object->generation++;
1612 * vm_page_zero_invalid()
1614 * The kernel assumes that the invalid portions of a page contain
1615 * garbage, but such pages can be mapped into memory by user code.
1616 * When this occurs, we must zero out the non-valid portions of the
1617 * page so user code sees what it expects.
1619 * Pages are most often semi-valid when the end of a file is mapped
1620 * into memory and the file's size is not page aligned.
1624 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
1630 * Scan the valid bits looking for invalid sections that
1631 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
1632 * valid bit may be set ) have already been zerod by
1633 * vm_page_set_validclean().
1636 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
1637 if (i == (PAGE_SIZE / DEV_BSIZE) ||
1638 (m->valid & (1 << i))
1641 pmap_zero_page_area(
1644 (i - b) << DEV_BSHIFT
1652 * setvalid is TRUE when we can safely set the zero'd areas
1653 * as being valid. We can do this if there are no cache consistency
1654 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
1658 m->valid = VM_PAGE_BITS_ALL;
1664 * Is (partial) page valid? Note that the case where size == 0
1665 * will return FALSE in the degenerate case where the page is
1666 * entirely invalid, and TRUE otherwise.
1672 vm_page_is_valid(vm_page_t m, int base, int size)
1674 int bits = vm_page_bits(base, size);
1676 if (m->valid && ((m->valid & bits) == bits))
1683 * update dirty bits from pmap/mmu. May not block.
1687 vm_page_test_dirty(vm_page_t m)
1689 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
1695 * This interface is for merging with malloc() someday.
1696 * Even if we never implement compaction so that contiguous allocation
1697 * works after initialization time, malloc()'s data structures are good
1698 * for statistics and for allocations of less than a page.
1702 unsigned long size, /* should be size_t here and for malloc() */
1703 struct malloc_type *type,
1707 unsigned long alignment,
1708 unsigned long boundary,
1712 vm_offset_t addr, phys, tmp_addr;
1714 vm_page_t pga = vm_page_array;
1716 size = round_page(size);
1718 panic("contigmalloc1: size must not be 0");
1719 if ((alignment & (alignment - 1)) != 0)
1720 panic("contigmalloc1: alignment must be a power of 2");
1721 if ((boundary & (boundary - 1)) != 0)
1722 panic("contigmalloc1: boundary must be a power of 2");
1725 for (pass = 0; pass <= 1; pass++) {
1729 * Find first page in array that is free, within range, aligned, and
1730 * such that the boundary won't be crossed.
1732 for (i = start; i < cnt.v_page_count; i++) {
1734 phys = VM_PAGE_TO_PHYS(&pga[i]);
1735 pqtype = pga[i].queue - pga[i].pc;
1736 if (((pqtype == PQ_FREE) || (pqtype == PQ_CACHE)) &&
1737 (phys >= low) && (phys < high) &&
1738 ((phys & (alignment - 1)) == 0) &&
1739 (((phys ^ (phys + size - 1)) & ~(boundary - 1)) == 0))
1744 * If the above failed or we will exceed the upper bound, fail.
1746 if ((i == cnt.v_page_count) ||
1747 ((VM_PAGE_TO_PHYS(&pga[i]) + size) > high)) {
1751 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
1755 KASSERT(m->queue == PQ_INACTIVE,
1756 ("contigmalloc1: page %p is not PQ_INACTIVE", m));
1758 next = TAILQ_NEXT(m, pageq);
1759 if (vm_page_sleep_busy(m, TRUE, "vpctw0"))
1761 vm_page_test_dirty(m);
1763 if (m->object->type == OBJT_VNODE) {
1764 vn_lock(m->object->handle, LK_EXCLUSIVE | LK_RETRY, curproc);
1765 vm_object_page_clean(m->object, 0, 0, OBJPC_SYNC);
1766 VOP_UNLOCK(m->object->handle, 0, curproc);
1768 } else if (m->object->type == OBJT_SWAP ||
1769 m->object->type == OBJT_DEFAULT) {
1770 vm_pageout_flush(&m, 1, 0);
1774 if ((m->dirty == 0) && (m->busy == 0) && (m->hold_count == 0))
1778 for (m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1782 KASSERT(m->queue == PQ_ACTIVE,
1783 ("contigmalloc1: page %p is not PQ_ACTIVE", m));
1785 next = TAILQ_NEXT(m, pageq);
1786 if (vm_page_sleep_busy(m, TRUE, "vpctw1"))
1788 vm_page_test_dirty(m);
1790 if (m->object->type == OBJT_VNODE) {
1791 vn_lock(m->object->handle, LK_EXCLUSIVE | LK_RETRY, curproc);
1792 vm_object_page_clean(m->object, 0, 0, OBJPC_SYNC);
1793 VOP_UNLOCK(m->object->handle, 0, curproc);
1795 } else if (m->object->type == OBJT_SWAP ||
1796 m->object->type == OBJT_DEFAULT) {
1797 vm_pageout_flush(&m, 1, 0);
1801 if ((m->dirty == 0) && (m->busy == 0) && (m->hold_count == 0))
1811 * Check successive pages for contiguous and free.
1813 for (i = start + 1; i < (start + size / PAGE_SIZE); i++) {
1815 pqtype = pga[i].queue - pga[i].pc;
1816 if ((VM_PAGE_TO_PHYS(&pga[i]) !=
1817 (VM_PAGE_TO_PHYS(&pga[i - 1]) + PAGE_SIZE)) ||
1818 ((pqtype != PQ_FREE) && (pqtype != PQ_CACHE))) {
1824 for (i = start; i < (start + size / PAGE_SIZE); i++) {
1826 vm_page_t m = &pga[i];
1828 pqtype = m->queue - m->pc;
1829 if (pqtype == PQ_CACHE) {
1833 vm_page_unqueue_nowakeup(m);
1834 m->valid = VM_PAGE_BITS_ALL;
1835 if (m->flags & PG_ZERO)
1836 vm_page_zero_count--;
1838 KASSERT(m->dirty == 0, ("contigmalloc1: page %p was dirty", m));
1845 * We've found a contiguous chunk that meets are requirements.
1846 * Allocate kernel VM, unfree and assign the physical pages to it and
1847 * return kernel VM pointer.
1850 if (vm_map_findspace(map, vm_map_min(map), size, &addr) !=
1853 * XXX We almost never run out of kernel virtual
1854 * space, so we don't make the allocated memory
1861 vm_object_reference(kernel_object);
1862 vm_map_insert(map, kernel_object, addr - VM_MIN_KERNEL_ADDRESS,
1863 addr, addr + size, VM_PROT_ALL, VM_PROT_ALL, 0);
1867 for (i = start; i < (start + size / PAGE_SIZE); i++) {
1868 vm_page_t m = &pga[i];
1869 vm_page_insert(m, kernel_object,
1870 OFF_TO_IDX(tmp_addr - VM_MIN_KERNEL_ADDRESS));
1871 tmp_addr += PAGE_SIZE;
1873 vm_map_pageable(map, addr, addr + size, FALSE);
1876 return ((void *)addr);
1883 unsigned long size, /* should be size_t here and for malloc() */
1884 struct malloc_type *type,
1888 unsigned long alignment,
1889 unsigned long boundary)
1891 return contigmalloc1(size, type, flags, low, high, alignment, boundary,
1896 contigfree(void *addr, unsigned long size, struct malloc_type *type)
1898 kmem_free(kernel_map, (vm_offset_t)addr, size);
1902 vm_page_alloc_contig(
1906 vm_offset_t alignment)
1908 return ((vm_offset_t)contigmalloc1(size, M_DEVBUF, M_NOWAIT, low, high,
1909 alignment, 0ul, kernel_map));
1912 #include "opt_ddb.h"
1914 #include <sys/kernel.h>
1916 #include <ddb/ddb.h>
1918 DB_SHOW_COMMAND(page, vm_page_print_page_info)
1920 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count);
1921 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count);
1922 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count);
1923 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count);
1924 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count);
1925 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved);
1926 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min);
1927 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target);
1928 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min);
1929 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target);
1932 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
1935 db_printf("PQ_FREE:");
1936 for(i=0;i<PQ_L2_SIZE;i++) {
1937 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
1941 db_printf("PQ_CACHE:");
1942 for(i=0;i<PQ_L2_SIZE;i++) {
1943 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
1947 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
1948 vm_page_queues[PQ_ACTIVE].lcnt,
1949 vm_page_queues[PQ_INACTIVE].lcnt);