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 * 4. Neither the name of the University nor the names of its contributors
17 * may be used to endorse or promote products derived from this software
18 * without specific prior written permission.
20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91
36 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
37 * All rights reserved.
39 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
41 * Permission to use, copy, modify and distribute this software and
42 * its documentation is hereby granted, provided that both the copyright
43 * notice and this permission notice appear in all copies of the
44 * software, derivative works or modified versions, and any portions
45 * thereof, and that both notices appear in supporting documentation.
47 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
48 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
49 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
51 * Carnegie Mellon requests users of this software to return to
53 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
54 * School of Computer Science
55 * Carnegie Mellon University
56 * Pittsburgh PA 15213-3890
58 * any improvements or extensions that they make and grant Carnegie the
59 * rights to redistribute these changes.
63 * GENERAL RULES ON VM_PAGE MANIPULATION
65 * - a pageq mutex is required when adding or removing a page from a
66 * page queue (vm_page_queue[]), regardless of other mutexes or the
67 * busy state of a page.
69 * - a hash chain mutex is required when associating or disassociating
70 * a page from the VM PAGE CACHE hash table (vm_page_buckets),
71 * regardless of other mutexes or the busy state of a page.
73 * - either a hash chain mutex OR a busied page is required in order
74 * to modify the page flags. A hash chain mutex must be obtained in
75 * order to busy a page. A page's flags cannot be modified by a
76 * hash chain mutex if the page is marked busy.
78 * - The object memq mutex is held when inserting or removing
79 * pages from an object (vm_page_insert() or vm_page_remove()). This
80 * is different from the object's main mutex.
82 * Generally speaking, you have to be aware of side effects when running
83 * vm_page ops. A vm_page_lookup() will return with the hash chain
84 * locked, whether it was able to lookup the page or not. vm_page_free(),
85 * vm_page_cache(), vm_page_activate(), and a number of other routines
86 * will release the hash chain mutex for you. Intermediate manipulation
87 * routines such as vm_page_flag_set() expect the hash chain to be held
88 * on entry and the hash chain will remain held on return.
90 * pageq scanning can only occur with the pageq in question locked.
91 * We have a known bottleneck with the active queue, but the cache
92 * and free queues are actually arrays already.
96 * Resident memory management module.
99 #include <sys/cdefs.h>
100 __FBSDID("$FreeBSD$");
102 #include <sys/param.h>
103 #include <sys/systm.h>
104 #include <sys/lock.h>
105 #include <sys/malloc.h>
106 #include <sys/mutex.h>
107 #include <sys/proc.h>
108 #include <sys/vmmeter.h>
109 #include <sys/vnode.h>
112 #include <vm/vm_param.h>
113 #include <vm/vm_kern.h>
114 #include <vm/vm_object.h>
115 #include <vm/vm_page.h>
116 #include <vm/vm_pageout.h>
117 #include <vm/vm_pager.h>
118 #include <vm/vm_extern.h>
120 #include <vm/uma_int.h>
123 * Associated with page of user-allocatable memory is a
127 struct mtx vm_page_queue_mtx;
128 struct mtx vm_page_queue_free_mtx;
130 vm_page_t vm_page_array = 0;
131 int vm_page_array_size = 0;
133 int vm_page_zero_count = 0;
138 * Sets the page size, perhaps based upon the memory
139 * size. Must be called before any use of page-size
140 * dependent functions.
143 vm_set_page_size(void)
145 if (cnt.v_page_size == 0)
146 cnt.v_page_size = PAGE_SIZE;
147 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0)
148 panic("vm_set_page_size: page size not a power of two");
154 * Initializes the resident memory module.
156 * Allocates memory for the page cells, and
157 * for the object/offset-to-page hash table headers.
158 * Each page cell is initialized and placed on the free list.
161 vm_page_startup(vm_offset_t vaddr)
165 vm_paddr_t page_range;
172 /* the biggest memory array is the second group of pages */
174 vm_paddr_t biggestsize;
184 vaddr = round_page(vaddr);
186 for (i = 0; phys_avail[i + 1]; i += 2) {
187 phys_avail[i] = round_page(phys_avail[i]);
188 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
191 for (i = 0; phys_avail[i + 1]; i += 2) {
192 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
194 if (size > biggestsize) {
202 end = phys_avail[biggestone+1];
205 * Initialize the locks.
207 mtx_init(&vm_page_queue_mtx, "vm page queue mutex", NULL, MTX_DEF |
209 mtx_init(&vm_page_queue_free_mtx, "vm page queue free mutex", NULL,
213 * Initialize the queue headers for the free queue, the active queue
214 * and the inactive queue.
219 * Allocate memory for use when boot strapping the kernel memory
222 bootpages = UMA_BOOT_PAGES * UMA_SLAB_SIZE;
223 new_end = end - bootpages;
224 new_end = trunc_page(new_end);
225 mapped = pmap_map(&vaddr, new_end, end,
226 VM_PROT_READ | VM_PROT_WRITE);
227 bzero((caddr_t) mapped, end - new_end);
228 uma_startup((caddr_t)mapped);
231 * Compute the number of pages of memory that will be available for
232 * use (taking into account the overhead of a page structure per
235 first_page = phys_avail[0] / PAGE_SIZE;
236 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
237 npages = (total - (page_range * sizeof(struct vm_page)) -
238 (end - new_end)) / PAGE_SIZE;
242 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
247 * Initialize the mem entry structures now, and put them in the free
250 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
251 mapped = pmap_map(&vaddr, new_end, end,
252 VM_PROT_READ | VM_PROT_WRITE);
253 vm_page_array = (vm_page_t) mapped;
254 phys_avail[biggestone + 1] = new_end;
257 * Clear all of the page structures
259 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
260 vm_page_array_size = page_range;
263 * Construct the free queue(s) in descending order (by physical
264 * address) so that the first 16MB of physical memory is allocated
265 * last rather than first. On large-memory machines, this avoids
266 * the exhaustion of low physical memory before isa_dma_init has run.
268 cnt.v_page_count = 0;
269 cnt.v_free_count = 0;
270 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
272 last_pa = phys_avail[i + 1];
273 while (pa < last_pa && npages-- > 0) {
274 vm_pageq_add_new_page(pa);
282 vm_page_flag_set(vm_page_t m, unsigned short bits)
285 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
290 vm_page_flag_clear(vm_page_t m, unsigned short bits)
293 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
298 vm_page_busy(vm_page_t m)
300 KASSERT((m->flags & PG_BUSY) == 0,
301 ("vm_page_busy: page already busy!!!"));
302 vm_page_flag_set(m, PG_BUSY);
308 * wakeup anyone waiting for the page.
311 vm_page_flash(vm_page_t m)
313 if (m->flags & PG_WANTED) {
314 vm_page_flag_clear(m, PG_WANTED);
322 * clear the PG_BUSY flag and wakeup anyone waiting for the
327 vm_page_wakeup(vm_page_t m)
329 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!"));
330 vm_page_flag_clear(m, PG_BUSY);
335 vm_page_io_start(vm_page_t m)
338 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
343 vm_page_io_finish(vm_page_t m)
346 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
353 * Keep page from being freed by the page daemon
354 * much of the same effect as wiring, except much lower
355 * overhead and should be used only for *very* temporary
356 * holding ("wiring").
359 vm_page_hold(vm_page_t mem)
362 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
367 vm_page_unhold(vm_page_t mem)
370 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
372 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
373 if (mem->hold_count == 0 && mem->queue == PQ_HOLD)
374 vm_page_free_toq(mem);
382 * The clearing of PG_ZERO is a temporary safety until the code can be
383 * reviewed to determine that PG_ZERO is being properly cleared on
384 * write faults or maps. PG_ZERO was previously cleared in
388 vm_page_free(vm_page_t m)
390 vm_page_flag_clear(m, PG_ZERO);
392 vm_page_zero_idle_wakeup();
398 * Free a page to the zerod-pages queue
401 vm_page_free_zero(vm_page_t m)
403 vm_page_flag_set(m, PG_ZERO);
408 * vm_page_sleep_if_busy:
410 * Sleep and release the page queues lock if PG_BUSY is set or,
411 * if also_m_busy is TRUE, busy is non-zero. Returns TRUE if the
412 * thread slept and the page queues lock was released.
413 * Otherwise, retains the page queues lock and returns FALSE.
416 vm_page_sleep_if_busy(vm_page_t m, int also_m_busy, const char *msg)
419 int is_object_locked;
421 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
422 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) {
423 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
425 * It's possible that while we sleep, the page will get
426 * unbusied and freed. If we are holding the object
427 * lock, we will assume we hold a reference to the object
428 * such that even if m->object changes, we can re-lock
431 * Remove mtx_owned() after vm_object locking is finished.
434 if ((is_object_locked = object != NULL &&
435 mtx_owned(&object->mtx)))
436 mtx_unlock(&object->mtx);
437 msleep(m, &vm_page_queue_mtx, PDROP | PVM, msg, 0);
438 if (is_object_locked)
439 mtx_lock(&object->mtx);
448 * make page all dirty
451 vm_page_dirty(vm_page_t m)
453 KASSERT(m->queue - m->pc != PQ_CACHE,
454 ("vm_page_dirty: page in cache!"));
455 KASSERT(m->queue - m->pc != PQ_FREE,
456 ("vm_page_dirty: page is free!"));
457 m->dirty = VM_PAGE_BITS_ALL;
463 * Implements Sleator and Tarjan's top-down splay algorithm. Returns
464 * the vm_page containing the given pindex. If, however, that
465 * pindex is not found in the vm_object, returns a vm_page that is
466 * adjacent to the pindex, coming before or after it.
469 vm_page_splay(vm_pindex_t pindex, vm_page_t root)
471 struct vm_page dummy;
472 vm_page_t lefttreemax, righttreemin, y;
476 lefttreemax = righttreemin = &dummy;
478 if (pindex < root->pindex) {
479 if ((y = root->left) == NULL)
481 if (pindex < y->pindex) {
483 root->left = y->right;
486 if ((y = root->left) == NULL)
489 /* Link into the new root's right tree. */
490 righttreemin->left = root;
492 } else if (pindex > root->pindex) {
493 if ((y = root->right) == NULL)
495 if (pindex > y->pindex) {
497 root->right = y->left;
500 if ((y = root->right) == NULL)
503 /* Link into the new root's left tree. */
504 lefttreemax->right = root;
509 /* Assemble the new root. */
510 lefttreemax->right = root->left;
511 righttreemin->left = root->right;
512 root->left = dummy.right;
513 root->right = dummy.left;
518 * vm_page_insert: [ internal use only ]
520 * Inserts the given mem entry into the object and object list.
522 * The pagetables are not updated but will presumably fault the page
523 * in if necessary, or if a kernel page the caller will at some point
524 * enter the page into the kernel's pmap. We are not allowed to block
525 * here so we *can't* do this anyway.
527 * The object and page must be locked.
528 * This routine may not block.
531 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
535 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
536 if (m->object != NULL)
537 panic("vm_page_insert: page already inserted");
540 * Record the object/offset pair in this page
546 * Now link into the object's ordered list of backed pages.
552 TAILQ_INSERT_TAIL(&object->memq, m, listq);
554 root = vm_page_splay(pindex, root);
555 if (pindex < root->pindex) {
556 m->left = root->left;
559 TAILQ_INSERT_BEFORE(root, m, listq);
560 } else if (pindex == root->pindex)
561 panic("vm_page_insert: offset already allocated");
563 m->right = root->right;
566 TAILQ_INSERT_AFTER(&object->memq, root, m, listq);
570 object->generation++;
573 * show that the object has one more resident page.
575 object->resident_page_count++;
578 * Since we are inserting a new and possibly dirty page,
579 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
581 if (m->flags & PG_WRITEABLE)
582 vm_object_set_writeable_dirty(object);
587 * NOTE: used by device pager as well -wfj
589 * Removes the given mem entry from the object/offset-page
590 * table and the object page list, but do not invalidate/terminate
593 * The object and page must be locked.
594 * The underlying pmap entry (if any) is NOT removed here.
595 * This routine may not block.
598 vm_page_remove(vm_page_t m)
603 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
604 if (m->object == NULL)
606 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
607 if ((m->flags & PG_BUSY) == 0) {
608 panic("vm_page_remove: page not busy");
612 * Basically destroy the page.
619 * Now remove from the object's list of backed pages.
621 if (m != object->root)
622 vm_page_splay(m->pindex, object->root);
626 root = vm_page_splay(m->pindex, m->left);
627 root->right = m->right;
630 TAILQ_REMOVE(&object->memq, m, listq);
633 * And show that the object has one fewer resident page.
635 object->resident_page_count--;
636 object->generation++;
644 * Returns the page associated with the object/offset
645 * pair specified; if none is found, NULL is returned.
647 * The object must be locked.
648 * This routine may not block.
649 * This is a critical path routine
652 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
656 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
657 if ((m = object->root) != NULL && m->pindex != pindex) {
658 m = vm_page_splay(pindex, m);
659 if ((object->root = m)->pindex != pindex)
668 * Move the given memory entry from its
669 * current object to the specified target object/offset.
671 * The object must be locked.
672 * This routine may not block.
674 * Note: swap associated with the page must be invalidated by the move. We
675 * have to do this for several reasons: (1) we aren't freeing the
676 * page, (2) we are dirtying the page, (3) the VM system is probably
677 * moving the page from object A to B, and will then later move
678 * the backing store from A to B and we can't have a conflict.
680 * Note: we *always* dirty the page. It is necessary both for the
681 * fact that we moved it, and because we may be invalidating
682 * swap. If the page is on the cache, we have to deactivate it
683 * or vm_page_dirty() will panic. Dirty pages are not allowed
687 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
691 vm_page_insert(m, new_object, new_pindex);
692 if (m->queue - m->pc == PQ_CACHE)
693 vm_page_deactivate(m);
698 * vm_page_select_cache:
700 * Find a page on the cache queue with color optimization. As pages
701 * might be found, but not applicable, they are deactivated. This
702 * keeps us from using potentially busy cached pages.
704 * This routine may not block.
707 vm_page_select_cache(int color)
711 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
712 while ((m = vm_pageq_find(PQ_CACHE, color, FALSE)) != NULL) {
713 if ((m->flags & PG_BUSY) == 0 && m->busy == 0 &&
714 m->hold_count == 0 && (VM_OBJECT_TRYLOCK(m->object) ||
715 VM_OBJECT_LOCKED(m->object))) {
716 KASSERT(m->dirty == 0,
717 ("Found dirty cache page %p", m));
718 KASSERT(!pmap_page_is_mapped(m),
719 ("Found mapped cache page %p", m));
720 KASSERT((m->flags & PG_UNMANAGED) == 0,
721 ("Found unmanaged cache page %p", m));
722 KASSERT(m->wire_count == 0,
723 ("Found wired cache page %p", m));
726 vm_page_deactivate(m);
734 * Allocate and return a memory cell associated
735 * with this VM object/offset pair.
738 * VM_ALLOC_NORMAL normal process request
739 * VM_ALLOC_SYSTEM system *really* needs a page
740 * VM_ALLOC_INTERRUPT interrupt time request
741 * VM_ALLOC_ZERO zero page
743 * This routine may not block.
745 * Additional special handling is required when called from an
746 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with
747 * the page cache in this case.
750 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
752 vm_object_t m_object;
754 int color, flags, page_req;
756 page_req = req & VM_ALLOC_CLASS_MASK;
758 if ((req & VM_ALLOC_NOOBJ) == 0) {
759 KASSERT(object != NULL,
760 ("vm_page_alloc: NULL object."));
761 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
762 color = (pindex + object->pg_color) & PQ_L2_MASK;
764 color = pindex & PQ_L2_MASK;
767 * The pager is allowed to eat deeper into the free page list.
769 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) {
770 page_req = VM_ALLOC_SYSTEM;
774 mtx_lock_spin(&vm_page_queue_free_mtx);
775 if (cnt.v_free_count > cnt.v_free_reserved ||
776 (page_req == VM_ALLOC_SYSTEM &&
777 cnt.v_cache_count == 0 &&
778 cnt.v_free_count > cnt.v_interrupt_free_min) ||
779 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0)) {
781 * Allocate from the free queue if the number of free pages
782 * exceeds the minimum for the request class.
784 m = vm_pageq_find(PQ_FREE, color, (req & VM_ALLOC_ZERO) != 0);
785 } else if (page_req != VM_ALLOC_INTERRUPT) {
786 mtx_unlock_spin(&vm_page_queue_free_mtx);
788 * Allocatable from cache (non-interrupt only). On success,
789 * we must free the page and try again, thus ensuring that
790 * cnt.v_*_free_min counters are replenished.
792 vm_page_lock_queues();
793 if ((m = vm_page_select_cache(color)) == NULL) {
794 #if defined(DIAGNOSTIC)
795 if (cnt.v_cache_count > 0)
796 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count);
798 vm_page_unlock_queues();
799 atomic_add_int(&vm_pageout_deficit, 1);
803 m_object = m->object;
804 VM_OBJECT_LOCK_ASSERT(m_object, MA_OWNED);
807 vm_page_unlock_queues();
808 if (m_object != object)
809 VM_OBJECT_UNLOCK(m_object);
813 * Not allocatable from cache from interrupt, give up.
815 mtx_unlock_spin(&vm_page_queue_free_mtx);
816 atomic_add_int(&vm_pageout_deficit, 1);
822 * At this point we had better have found a good page.
827 ("vm_page_alloc(): missing page on free queue")
831 * Remove from free queue
833 vm_pageq_remove_nowakeup(m);
836 * Initialize structure. Only the PG_ZERO flag is inherited.
839 if (m->flags & PG_ZERO) {
840 vm_page_zero_count--;
841 if (req & VM_ALLOC_ZERO)
842 flags = PG_ZERO | PG_BUSY;
844 if (req & VM_ALLOC_NOOBJ)
847 if (req & VM_ALLOC_WIRED) {
848 atomic_add_int(&cnt.v_wire_count, 1);
856 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m));
857 mtx_unlock_spin(&vm_page_queue_free_mtx);
859 if ((req & VM_ALLOC_NOOBJ) == 0)
860 vm_page_insert(m, object, pindex);
865 * Don't wakeup too often - wakeup the pageout daemon when
866 * we would be nearly out of memory.
868 if (vm_paging_needed())
875 * vm_wait: (also see VM_WAIT macro)
877 * Block until free pages are available for allocation
878 * - Called in various places before memory allocations.
884 vm_page_lock_queues();
885 if (curproc == pageproc) {
886 vm_pageout_pages_needed = 1;
887 msleep(&vm_pageout_pages_needed, &vm_page_queue_mtx,
888 PDROP | PSWP, "VMWait", 0);
890 if (!vm_pages_needed) {
892 wakeup(&vm_pages_needed);
894 msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PVM,
900 * vm_waitpfault: (also see VM_WAITPFAULT macro)
902 * Block until free pages are available for allocation
903 * - Called only in vm_fault so that processes page faulting
904 * can be easily tracked.
905 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
906 * processes will be able to grab memory first. Do not change
907 * this balance without careful testing first.
913 vm_page_lock_queues();
914 if (!vm_pages_needed) {
916 wakeup(&vm_pages_needed);
918 msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PUSER,
925 * Put the specified page on the active list (if appropriate).
926 * Ensure that act_count is at least ACT_INIT but do not otherwise
929 * The page queues must be locked.
930 * This routine may not block.
933 vm_page_activate(vm_page_t m)
936 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
937 if (m->queue != PQ_ACTIVE) {
938 if ((m->queue - m->pc) == PQ_CACHE)
941 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
942 if (m->act_count < ACT_INIT)
943 m->act_count = ACT_INIT;
944 vm_pageq_enqueue(PQ_ACTIVE, m);
947 if (m->act_count < ACT_INIT)
948 m->act_count = ACT_INIT;
953 * vm_page_free_wakeup:
955 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
956 * routine is called when a page has been added to the cache or free
959 * The page queues must be locked.
960 * This routine may not block.
963 vm_page_free_wakeup(void)
966 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
968 * if pageout daemon needs pages, then tell it that there are
971 if (vm_pageout_pages_needed &&
972 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) {
973 wakeup(&vm_pageout_pages_needed);
974 vm_pageout_pages_needed = 0;
977 * wakeup processes that are waiting on memory if we hit a
978 * high water mark. And wakeup scheduler process if we have
979 * lots of memory. this process will swapin processes.
981 if (vm_pages_needed && !vm_page_count_min()) {
983 wakeup(&cnt.v_free_count);
990 * Returns the given page to the PQ_FREE list,
991 * disassociating it with any VM object.
993 * Object and page must be locked prior to entry.
994 * This routine may not block.
998 vm_page_free_toq(vm_page_t m)
1000 struct vpgqueues *pq;
1001 vm_object_t object = m->object;
1003 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1006 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
1008 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
1009 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
1011 if ((m->queue - m->pc) == PQ_FREE)
1012 panic("vm_page_free: freeing free page");
1014 panic("vm_page_free: freeing busy page");
1018 * unqueue, then remove page. Note that we cannot destroy
1019 * the page here because we do not want to call the pager's
1020 * callback routine until after we've put the page on the
1021 * appropriate free queue.
1023 vm_pageq_remove_nowakeup(m);
1027 * If fictitious remove object association and
1028 * return, otherwise delay object association removal.
1030 if ((m->flags & PG_FICTITIOUS) != 0) {
1037 if (m->wire_count != 0) {
1038 if (m->wire_count > 1) {
1039 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1040 m->wire_count, (long)m->pindex);
1042 panic("vm_page_free: freeing wired page");
1046 * 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;
1057 if (VSHOULDFREE(vp))
1064 * Clear the UNMANAGED flag when freeing an unmanaged page.
1066 if (m->flags & PG_UNMANAGED) {
1067 m->flags &= ~PG_UNMANAGED;
1070 if (m->hold_count != 0) {
1071 m->flags &= ~PG_ZERO;
1074 m->queue = PQ_FREE + m->pc;
1075 pq = &vm_page_queues[m->queue];
1076 mtx_lock_spin(&vm_page_queue_free_mtx);
1081 * Put zero'd pages on the end ( where we look for zero'd pages
1082 * first ) and non-zerod pages at the head.
1084 if (m->flags & PG_ZERO) {
1085 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
1086 ++vm_page_zero_count;
1088 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1090 mtx_unlock_spin(&vm_page_queue_free_mtx);
1091 vm_page_free_wakeup();
1097 * Prevent PV management from being done on the page. The page is
1098 * removed from the paging queues as if it were wired, and as a
1099 * consequence of no longer being managed the pageout daemon will not
1100 * touch it (since there is no way to locate the pte mappings for the
1101 * page). madvise() calls that mess with the pmap will also no longer
1102 * operate on the page.
1104 * Beyond that the page is still reasonably 'normal'. Freeing the page
1105 * will clear the flag.
1107 * This routine is used by OBJT_PHYS objects - objects using unswappable
1108 * physical memory as backing store rather then swap-backed memory and
1109 * will eventually be extended to support 4MB unmanaged physical
1113 vm_page_unmanage(vm_page_t m)
1116 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1117 if ((m->flags & PG_UNMANAGED) == 0) {
1118 if (m->wire_count == 0)
1121 vm_page_flag_set(m, PG_UNMANAGED);
1127 * Mark this page as wired down by yet
1128 * another map, removing it from paging queues
1131 * The page queues must be locked.
1132 * This routine may not block.
1135 vm_page_wire(vm_page_t m)
1139 * Only bump the wire statistics if the page is not already wired,
1140 * and only unqueue the page if it is on some queue (if it is unmanaged
1141 * it is already off the queues).
1143 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1144 if (m->flags & PG_FICTITIOUS)
1146 if (m->wire_count == 0) {
1147 if ((m->flags & PG_UNMANAGED) == 0)
1149 atomic_add_int(&cnt.v_wire_count, 1);
1152 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
1158 * Release one wiring of this page, potentially
1159 * enabling it to be paged again.
1161 * Many pages placed on the inactive queue should actually go
1162 * into the cache, but it is difficult to figure out which. What
1163 * we do instead, if the inactive target is well met, is to put
1164 * clean pages at the head of the inactive queue instead of the tail.
1165 * This will cause them to be moved to the cache more quickly and
1166 * if not actively re-referenced, freed more quickly. If we just
1167 * stick these pages at the end of the inactive queue, heavy filesystem
1168 * meta-data accesses can cause an unnecessary paging load on memory bound
1169 * processes. This optimization causes one-time-use metadata to be
1170 * reused more quickly.
1172 * BUT, if we are in a low-memory situation we have no choice but to
1173 * put clean pages on the cache queue.
1175 * A number of routines use vm_page_unwire() to guarantee that the page
1176 * will go into either the inactive or active queues, and will NEVER
1177 * be placed in the cache - for example, just after dirtying a page.
1178 * dirty pages in the cache are not allowed.
1180 * The page queues must be locked.
1181 * This routine may not block.
1184 vm_page_unwire(vm_page_t m, int activate)
1187 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1188 if (m->flags & PG_FICTITIOUS)
1190 if (m->wire_count > 0) {
1192 if (m->wire_count == 0) {
1193 atomic_subtract_int(&cnt.v_wire_count, 1);
1194 if (m->flags & PG_UNMANAGED) {
1196 } else if (activate)
1197 vm_pageq_enqueue(PQ_ACTIVE, m);
1199 vm_page_flag_clear(m, PG_WINATCFLS);
1200 vm_pageq_enqueue(PQ_INACTIVE, m);
1204 panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
1210 * Move the specified page to the inactive queue. If the page has
1211 * any associated swap, the swap is deallocated.
1213 * Normally athead is 0 resulting in LRU operation. athead is set
1214 * to 1 if we want this page to be 'as if it were placed in the cache',
1215 * except without unmapping it from the process address space.
1217 * This routine may not block.
1219 static __inline void
1220 _vm_page_deactivate(vm_page_t m, int athead)
1223 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1226 * Ignore if already inactive.
1228 if (m->queue == PQ_INACTIVE)
1230 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1231 if ((m->queue - m->pc) == PQ_CACHE)
1232 cnt.v_reactivated++;
1233 vm_page_flag_clear(m, PG_WINATCFLS);
1236 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1238 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1239 m->queue = PQ_INACTIVE;
1240 vm_page_queues[PQ_INACTIVE].lcnt++;
1241 cnt.v_inactive_count++;
1246 vm_page_deactivate(vm_page_t m)
1248 _vm_page_deactivate(m, 0);
1252 * vm_page_try_to_cache:
1254 * Returns 0 on failure, 1 on success
1257 vm_page_try_to_cache(vm_page_t m)
1260 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1261 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1262 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1273 * vm_page_try_to_free()
1275 * Attempt to free the page. If we cannot free it, we do nothing.
1276 * 1 is returned on success, 0 on failure.
1279 vm_page_try_to_free(vm_page_t m)
1282 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1283 if (m->object != NULL)
1284 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1285 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1286 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1300 * Put the specified page onto the page cache queue (if appropriate).
1302 * This routine may not block.
1305 vm_page_cache(vm_page_t m)
1308 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1309 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
1310 m->hold_count || m->wire_count) {
1311 printf("vm_page_cache: attempting to cache busy page\n");
1314 if ((m->queue - m->pc) == PQ_CACHE)
1318 * Remove all pmaps and indicate that the page is not
1319 * writeable or mapped.
1322 if (m->dirty != 0) {
1323 panic("vm_page_cache: caching a dirty page, pindex: %ld",
1326 vm_pageq_remove_nowakeup(m);
1327 vm_pageq_enqueue(PQ_CACHE + m->pc, m);
1328 vm_page_free_wakeup();
1334 * Cache, deactivate, or do nothing as appropriate. This routine
1335 * is typically used by madvise() MADV_DONTNEED.
1337 * Generally speaking we want to move the page into the cache so
1338 * it gets reused quickly. However, this can result in a silly syndrome
1339 * due to the page recycling too quickly. Small objects will not be
1340 * fully cached. On the otherhand, if we move the page to the inactive
1341 * queue we wind up with a problem whereby very large objects
1342 * unnecessarily blow away our inactive and cache queues.
1344 * The solution is to move the pages based on a fixed weighting. We
1345 * either leave them alone, deactivate them, or move them to the cache,
1346 * where moving them to the cache has the highest weighting.
1347 * By forcing some pages into other queues we eventually force the
1348 * system to balance the queues, potentially recovering other unrelated
1349 * space from active. The idea is to not force this to happen too
1353 vm_page_dontneed(vm_page_t m)
1355 static int dnweight;
1359 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1363 * occassionally leave the page alone
1365 if ((dnw & 0x01F0) == 0 ||
1366 m->queue == PQ_INACTIVE ||
1367 m->queue - m->pc == PQ_CACHE
1369 if (m->act_count >= ACT_INIT)
1374 if (m->dirty == 0 && pmap_is_modified(m))
1377 if (m->dirty || (dnw & 0x0070) == 0) {
1379 * Deactivate the page 3 times out of 32.
1384 * Cache the page 28 times out of every 32. Note that
1385 * the page is deactivated instead of cached, but placed
1386 * at the head of the queue instead of the tail.
1390 _vm_page_deactivate(m, head);
1394 * Grab a page, waiting until we are waken up due to the page
1395 * changing state. We keep on waiting, if the page continues
1396 * to be in the object. If the page doesn't exist, first allocate it
1397 * and then conditionally zero it.
1399 * This routine may block.
1402 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
1406 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1408 if ((m = vm_page_lookup(object, pindex)) != NULL) {
1409 vm_page_lock_queues();
1410 if (m->busy || (m->flags & PG_BUSY)) {
1411 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
1412 VM_OBJECT_UNLOCK(object);
1413 msleep(m, &vm_page_queue_mtx, PDROP | PVM, "pgrbwt", 0);
1414 VM_OBJECT_LOCK(object);
1415 if ((allocflags & VM_ALLOC_RETRY) == 0)
1419 if (allocflags & VM_ALLOC_WIRED)
1422 vm_page_unlock_queues();
1426 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1428 VM_OBJECT_UNLOCK(object);
1430 VM_OBJECT_LOCK(object);
1431 if ((allocflags & VM_ALLOC_RETRY) == 0)
1435 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
1441 * Mapping function for valid bits or for dirty bits in
1442 * a page. May not block.
1444 * Inputs are required to range within a page.
1447 vm_page_bits(int base, int size)
1453 base + size <= PAGE_SIZE,
1454 ("vm_page_bits: illegal base/size %d/%d", base, size)
1457 if (size == 0) /* handle degenerate case */
1460 first_bit = base >> DEV_BSHIFT;
1461 last_bit = (base + size - 1) >> DEV_BSHIFT;
1463 return ((2 << last_bit) - (1 << first_bit));
1467 * vm_page_set_validclean:
1469 * Sets portions of a page valid and clean. The arguments are expected
1470 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1471 * of any partial chunks touched by the range. The invalid portion of
1472 * such chunks will be zero'd.
1474 * This routine may not block.
1476 * (base + size) must be less then or equal to PAGE_SIZE.
1479 vm_page_set_validclean(vm_page_t m, int base, int size)
1485 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1486 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1487 if (size == 0) /* handle degenerate case */
1491 * If the base is not DEV_BSIZE aligned and the valid
1492 * bit is clear, we have to zero out a portion of the
1495 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
1496 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
1497 pmap_zero_page_area(m, frag, base - frag);
1500 * If the ending offset is not DEV_BSIZE aligned and the
1501 * valid bit is clear, we have to zero out a portion of
1504 endoff = base + size;
1505 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
1506 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
1507 pmap_zero_page_area(m, endoff,
1508 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
1511 * Set valid, clear dirty bits. If validating the entire
1512 * page we can safely clear the pmap modify bit. We also
1513 * use this opportunity to clear the PG_NOSYNC flag. If a process
1514 * takes a write fault on a MAP_NOSYNC memory area the flag will
1517 * We set valid bits inclusive of any overlap, but we can only
1518 * clear dirty bits for DEV_BSIZE chunks that are fully within
1521 pagebits = vm_page_bits(base, size);
1522 m->valid |= pagebits;
1524 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
1525 frag = DEV_BSIZE - frag;
1531 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
1533 m->dirty &= ~pagebits;
1534 if (base == 0 && size == PAGE_SIZE) {
1535 pmap_clear_modify(m);
1536 vm_page_flag_clear(m, PG_NOSYNC);
1541 vm_page_clear_dirty(vm_page_t m, int base, int size)
1544 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1545 m->dirty &= ~vm_page_bits(base, size);
1549 * vm_page_set_invalid:
1551 * Invalidates DEV_BSIZE'd chunks within a page. Both the
1552 * valid and dirty bits for the effected areas are cleared.
1557 vm_page_set_invalid(vm_page_t m, int base, int size)
1561 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1562 bits = vm_page_bits(base, size);
1563 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1566 m->object->generation++;
1570 * vm_page_zero_invalid()
1572 * The kernel assumes that the invalid portions of a page contain
1573 * garbage, but such pages can be mapped into memory by user code.
1574 * When this occurs, we must zero out the non-valid portions of the
1575 * page so user code sees what it expects.
1577 * Pages are most often semi-valid when the end of a file is mapped
1578 * into memory and the file's size is not page aligned.
1581 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
1586 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1588 * Scan the valid bits looking for invalid sections that
1589 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
1590 * valid bit may be set ) have already been zerod by
1591 * vm_page_set_validclean().
1593 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
1594 if (i == (PAGE_SIZE / DEV_BSIZE) ||
1595 (m->valid & (1 << i))
1598 pmap_zero_page_area(m,
1599 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
1606 * setvalid is TRUE when we can safely set the zero'd areas
1607 * as being valid. We can do this if there are no cache consistancy
1608 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
1611 m->valid = VM_PAGE_BITS_ALL;
1617 * Is (partial) page valid? Note that the case where size == 0
1618 * will return FALSE in the degenerate case where the page is
1619 * entirely invalid, and TRUE otherwise.
1624 vm_page_is_valid(vm_page_t m, int base, int size)
1626 int bits = vm_page_bits(base, size);
1628 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1629 if (m->valid && ((m->valid & bits) == bits))
1636 * update dirty bits from pmap/mmu. May not block.
1639 vm_page_test_dirty(vm_page_t m)
1641 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
1646 int so_zerocp_fullpage = 0;
1649 vm_page_cowfault(vm_page_t m)
1661 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL);
1663 vm_page_insert(m, object, pindex);
1664 vm_page_unlock_queues();
1665 VM_OBJECT_UNLOCK(object);
1667 VM_OBJECT_LOCK(object);
1668 vm_page_lock_queues();
1674 * check to see if we raced with an xmit complete when
1675 * waiting to allocate a page. If so, put things back
1680 vm_page_insert(m, object, pindex);
1681 } else { /* clear COW & copy page */
1682 if (!so_zerocp_fullpage)
1683 pmap_copy_page(m, mnew);
1684 mnew->valid = VM_PAGE_BITS_ALL;
1685 vm_page_dirty(mnew);
1686 vm_page_flag_clear(mnew, PG_BUSY);
1691 vm_page_cowclear(vm_page_t m)
1694 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1698 * let vm_fault add back write permission lazily
1702 * sf_buf_free() will free the page, so we needn't do it here
1707 vm_page_cowsetup(vm_page_t m)
1710 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1712 pmap_page_protect(m, VM_PROT_READ);
1715 #include "opt_ddb.h"
1717 #include <sys/kernel.h>
1719 #include <ddb/ddb.h>
1721 DB_SHOW_COMMAND(page, vm_page_print_page_info)
1723 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count);
1724 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count);
1725 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count);
1726 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count);
1727 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count);
1728 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved);
1729 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min);
1730 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target);
1731 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min);
1732 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target);
1735 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
1738 db_printf("PQ_FREE:");
1739 for (i = 0; i < PQ_L2_SIZE; i++) {
1740 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
1744 db_printf("PQ_CACHE:");
1745 for (i = 0; i < PQ_L2_SIZE; i++) {
1746 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
1750 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
1751 vm_page_queues[PQ_ACTIVE].lcnt,
1752 vm_page_queues[PQ_INACTIVE].lcnt);