2 * Copyright (c) 1991 Regents of the University of California.
5 * This code is derived from software contributed to Berkeley by
6 * The Mach Operating System project at Carnegie-Mellon University.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 3. All advertising materials mentioning features or use of this software
17 * must display the following acknowledgement:
18 * This product includes software developed by the University of
19 * California, Berkeley and its contributors.
20 * 4. Neither the name of the University nor the names of its contributors
21 * may be used to endorse or promote products derived from this software
22 * without specific prior written permission.
24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91
41 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
42 * All rights reserved.
44 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
46 * Permission to use, copy, modify and distribute this software and
47 * its documentation is hereby granted, provided that both the copyright
48 * notice and this permission notice appear in all copies of the
49 * software, derivative works or modified versions, and any portions
50 * thereof, and that both notices appear in supporting documentation.
52 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
53 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
54 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
56 * Carnegie Mellon requests users of this software to return to
58 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
59 * School of Computer Science
60 * Carnegie Mellon University
61 * Pittsburgh PA 15213-3890
63 * any improvements or extensions that they make and grant Carnegie the
64 * rights to redistribute these changes.
68 * GENERAL RULES ON VM_PAGE MANIPULATION
70 * - a pageq mutex is required when adding or removing a page from a
71 * page queue (vm_page_queue[]), regardless of other mutexes or the
72 * busy state of a page.
74 * - a hash chain mutex is required when associating or disassociating
75 * a page from the VM PAGE CACHE hash table (vm_page_buckets),
76 * regardless of other mutexes or the busy state of a page.
78 * - either a hash chain mutex OR a busied page is required in order
79 * to modify the page flags. A hash chain mutex must be obtained in
80 * order to busy a page. A page's flags cannot be modified by a
81 * hash chain mutex if the page is marked busy.
83 * - The object memq mutex is held when inserting or removing
84 * pages from an object (vm_page_insert() or vm_page_remove()). This
85 * is different from the object's main mutex.
87 * Generally speaking, you have to be aware of side effects when running
88 * vm_page ops. A vm_page_lookup() will return with the hash chain
89 * locked, whether it was able to lookup the page or not. vm_page_free(),
90 * vm_page_cache(), vm_page_activate(), and a number of other routines
91 * will release the hash chain mutex for you. Intermediate manipulation
92 * routines such as vm_page_flag_set() expect the hash chain to be held
93 * on entry and the hash chain will remain held on return.
95 * pageq scanning can only occur with the pageq in question locked.
96 * We have a known bottleneck with the active queue, but the cache
97 * and free queues are actually arrays already.
101 * Resident memory management module.
104 #include <sys/param.h>
105 #include <sys/systm.h>
106 #include <sys/lock.h>
107 #include <sys/malloc.h>
108 #include <sys/mutex.h>
109 #include <sys/proc.h>
110 #include <sys/vmmeter.h>
111 #include <sys/vnode.h>
114 #include <vm/vm_param.h>
115 #include <vm/vm_kern.h>
116 #include <vm/vm_object.h>
117 #include <vm/vm_page.h>
118 #include <vm/vm_pageout.h>
119 #include <vm/vm_pager.h>
120 #include <vm/vm_extern.h>
122 #include <vm/uma_int.h>
125 * Associated with page of user-allocatable memory is a
129 struct mtx vm_page_queue_mtx;
130 struct mtx vm_page_queue_free_mtx;
132 vm_page_t vm_page_array = 0;
133 int vm_page_array_size = 0;
135 int vm_page_zero_count = 0;
140 * Sets the page size, perhaps based upon the memory
141 * size. Must be called before any use of page-size
142 * dependent functions.
145 vm_set_page_size(void)
147 if (cnt.v_page_size == 0)
148 cnt.v_page_size = PAGE_SIZE;
149 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0)
150 panic("vm_set_page_size: page size not a power of two");
156 * Initializes the resident memory module.
158 * Allocates memory for the page cells, and
159 * for the object/offset-to-page hash table headers.
160 * Each page cell is initialized and placed on the free list.
163 vm_page_startup(vm_offset_t starta, vm_offset_t enda, vm_offset_t vaddr)
166 vm_size_t npages, page_range;
173 /* the biggest memory array is the second group of pages */
175 vm_offset_t biggestone, 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_size_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);
208 mtx_init(&vm_page_queue_free_mtx, "vm page queue free mutex", NULL,
212 * Initialize the queue headers for the free queue, the active queue
213 * and the inactive queue.
218 * Allocate memory for use when boot strapping the kernel memory
221 bootpages = UMA_BOOT_PAGES * UMA_SLAB_SIZE;
222 new_end = end - bootpages;
223 new_end = trunc_page(new_end);
224 mapped = pmap_map(&vaddr, new_end, end,
225 VM_PROT_READ | VM_PROT_WRITE);
226 bzero((caddr_t) mapped, end - new_end);
227 uma_startup((caddr_t)mapped);
230 * Compute the number of pages of memory that will be available for
231 * use (taking into account the overhead of a page structure per
234 first_page = phys_avail[0] / PAGE_SIZE;
235 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
236 npages = (total - (page_range * sizeof(struct vm_page)) -
237 (end - new_end)) / PAGE_SIZE;
241 * Initialize the mem entry structures now, and put them in the free
244 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
245 mapped = pmap_map(&vaddr, new_end, end,
246 VM_PROT_READ | VM_PROT_WRITE);
247 vm_page_array = (vm_page_t) mapped;
250 * Clear all of the page structures
252 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
253 vm_page_array_size = page_range;
256 * Construct the free queue(s) in descending order (by physical
257 * address) so that the first 16MB of physical memory is allocated
258 * last rather than first. On large-memory machines, this avoids
259 * the exhaustion of low physical memory before isa_dmainit has run.
261 cnt.v_page_count = 0;
262 cnt.v_free_count = 0;
263 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
268 last_pa = phys_avail[i + 1];
269 while (pa < last_pa && npages-- > 0) {
270 vm_pageq_add_new_page(pa);
278 vm_page_flag_set(vm_page_t m, unsigned short bits)
281 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
286 vm_page_flag_clear(vm_page_t m, unsigned short bits)
289 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
294 vm_page_busy(vm_page_t m)
296 KASSERT((m->flags & PG_BUSY) == 0,
297 ("vm_page_busy: page already busy!!!"));
298 vm_page_flag_set(m, PG_BUSY);
304 * wakeup anyone waiting for the page.
307 vm_page_flash(vm_page_t m)
309 if (m->flags & PG_WANTED) {
310 vm_page_flag_clear(m, PG_WANTED);
318 * clear the PG_BUSY flag and wakeup anyone waiting for the
323 vm_page_wakeup(vm_page_t m)
325 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!"));
326 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)
366 vm_page_unhold(vm_page_t mem)
369 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
371 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
372 if (mem->hold_count == 0 && mem->queue == PQ_HOLD)
373 vm_page_free_toq(mem);
379 * Copy one page to another
382 vm_page_copy(vm_page_t src_m, vm_page_t dest_m)
384 pmap_copy_page(src_m, dest_m);
385 dest_m->valid = VM_PAGE_BITS_ALL;
393 * The clearing of PG_ZERO is a temporary safety until the code can be
394 * reviewed to determine that PG_ZERO is being properly cleared on
395 * write faults or maps. PG_ZERO was previously cleared in
399 vm_page_free(vm_page_t m)
401 vm_page_flag_clear(m, PG_ZERO);
403 vm_page_zero_idle_wakeup();
409 * Free a page to the zerod-pages queue
412 vm_page_free_zero(vm_page_t m)
414 vm_page_flag_set(m, PG_ZERO);
419 * vm_page_sleep_if_busy:
421 * Sleep and release the page queues lock if PG_BUSY is set or,
422 * if also_m_busy is TRUE, busy is non-zero. Returns TRUE if the
423 * thread slept and the page queues lock was released.
424 * Otherwise, retains the page queues lock and returns FALSE.
427 vm_page_sleep_if_busy(vm_page_t m, int also_m_busy, const char *msg)
430 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
431 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) {
432 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
433 msleep(m, &vm_page_queue_mtx, PDROP | PVM, msg, 0);
442 * make page all dirty
445 vm_page_dirty(vm_page_t m)
447 KASSERT(m->queue - m->pc != PQ_CACHE,
448 ("vm_page_dirty: page in cache!"));
449 m->dirty = VM_PAGE_BITS_ALL;
455 * Implements Sleator and Tarjan's top-down splay algorithm. Returns
456 * the vm_page containing the given pindex. If, however, that
457 * pindex is not found in the vm_object, returns a vm_page that is
458 * adjacent to the pindex, coming before or after it.
461 vm_page_splay(vm_pindex_t pindex, vm_page_t root)
463 struct vm_page dummy;
464 vm_page_t lefttreemax, righttreemin, y;
468 lefttreemax = righttreemin = &dummy;
470 if (pindex < root->pindex) {
471 if ((y = root->left) == NULL)
473 if (pindex < y->pindex) {
475 root->left = y->right;
478 if ((y = root->left) == NULL)
481 /* Link into the new root's right tree. */
482 righttreemin->left = root;
484 } else if (pindex > root->pindex) {
485 if ((y = root->right) == NULL)
487 if (pindex > y->pindex) {
489 root->right = y->left;
492 if ((y = root->right) == NULL)
495 /* Link into the new root's left tree. */
496 lefttreemax->right = root;
501 /* Assemble the new root. */
502 lefttreemax->right = root->left;
503 righttreemin->left = root->right;
504 root->left = dummy.right;
505 root->right = dummy.left;
510 * vm_page_insert: [ internal use only ]
512 * Inserts the given mem entry into the object and object list.
514 * The pagetables are not updated but will presumably fault the page
515 * in if necessary, or if a kernel page the caller will at some point
516 * enter the page into the kernel's pmap. We are not allowed to block
517 * here so we *can't* do this anyway.
519 * The object and page must be locked, and must be splhigh.
520 * This routine may not block.
523 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
527 if (m->object != NULL)
528 panic("vm_page_insert: already inserted");
531 * Record the object/offset pair in this page
536 mtx_assert(object == kmem_object ? &object->mtx : &Giant, MA_OWNED);
538 * Now link into the object's ordered list of backed pages.
544 TAILQ_INSERT_TAIL(&object->memq, m, listq);
546 root = vm_page_splay(pindex, root);
547 if (pindex < root->pindex) {
548 m->left = root->left;
551 TAILQ_INSERT_BEFORE(root, m, listq);
553 m->right = root->right;
556 TAILQ_INSERT_AFTER(&object->memq, root, m, listq);
560 object->generation++;
563 * show that the object has one more resident page.
565 object->resident_page_count++;
568 * Since we are inserting a new and possibly dirty page,
569 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
571 if (m->flags & PG_WRITEABLE)
572 vm_object_set_writeable_dirty(object);
577 * NOTE: used by device pager as well -wfj
579 * Removes the given mem entry from the object/offset-page
580 * table and the object page list, but do not invalidate/terminate
583 * The object and page must be locked, and at splhigh.
584 * The underlying pmap entry (if any) is NOT removed here.
585 * This routine may not block.
588 vm_page_remove(vm_page_t m)
594 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
595 if (m->object == NULL)
598 if ((m->flags & PG_BUSY) == 0) {
599 panic("vm_page_remove: page not busy");
603 * Basically destroy the page.
610 * Now remove from the object's list of backed pages.
612 if (m != object->root)
613 vm_page_splay(m->pindex, object->root);
617 root = vm_page_splay(m->pindex, m->left);
618 root->right = m->right;
621 TAILQ_REMOVE(&object->memq, m, listq);
624 * And show that the object has one fewer resident page.
626 object->resident_page_count--;
627 object->generation++;
635 * Returns the page associated with the object/offset
636 * pair specified; if none is found, NULL is returned.
638 * The object must be locked.
639 * This routine may not block.
640 * This is a critical path routine
643 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
647 mtx_assert(object == kmem_object ? &object->mtx : &Giant, MA_OWNED);
648 m = vm_page_splay(pindex, object->root);
649 if ((object->root = m) != NULL && m->pindex != pindex)
657 * Move the given memory entry from its
658 * current object to the specified target object/offset.
660 * The object must be locked.
661 * This routine may not block.
663 * Note: this routine will raise itself to splvm(), the caller need not.
665 * Note: swap associated with the page must be invalidated by the move. We
666 * have to do this for several reasons: (1) we aren't freeing the
667 * page, (2) we are dirtying the page, (3) the VM system is probably
668 * moving the page from object A to B, and will then later move
669 * the backing store from A to B and we can't have a conflict.
671 * Note: we *always* dirty the page. It is necessary both for the
672 * fact that we moved it, and because we may be invalidating
673 * swap. If the page is on the cache, we have to deactivate it
674 * or vm_page_dirty() will panic. Dirty pages are not allowed
678 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
684 vm_page_insert(m, new_object, new_pindex);
685 if (m->queue - m->pc == PQ_CACHE)
686 vm_page_deactivate(m);
692 * vm_page_select_cache:
694 * Find a page on the cache queue with color optimization. As pages
695 * might be found, but not applicable, they are deactivated. This
696 * keeps us from using potentially busy cached pages.
698 * This routine must be called at splvm().
699 * This routine may not block.
702 vm_page_select_cache(vm_pindex_t color)
706 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
708 m = vm_pageq_find(PQ_CACHE, color & PQ_L2_MASK, FALSE);
709 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
710 m->hold_count || m->wire_count)) {
711 vm_page_deactivate(m);
719 * vm_page_select_free:
721 * Find a free or zero page, with specified preference.
723 * This routine must be called at splvm().
724 * This routine may not block.
726 static __inline vm_page_t
727 vm_page_select_free(vm_pindex_t color, boolean_t prefer_zero)
731 m = vm_pageq_find(PQ_FREE, color & PQ_L2_MASK, prefer_zero);
738 * Allocate and return a memory cell associated
739 * with this VM object/offset pair.
742 * VM_ALLOC_NORMAL normal process request
743 * VM_ALLOC_SYSTEM system *really* needs a page
744 * VM_ALLOC_INTERRUPT interrupt time request
745 * VM_ALLOC_ZERO zero page
747 * This routine may not block.
749 * Additional special handling is required when called from an
750 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with
751 * the page cache in this case.
754 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
758 int flags, page_req, s;
760 page_req = req & VM_ALLOC_CLASS_MASK;
762 if ((req & VM_ALLOC_NOOBJ) == 0) {
763 KASSERT(object != NULL,
764 ("vm_page_alloc: NULL object."));
765 mtx_assert(object == kmem_object ? &object->mtx : &Giant,
767 KASSERT(!vm_page_lookup(object, pindex),
768 ("vm_page_alloc: page already allocated"));
769 color = pindex + object->pg_color;
774 * The pager is allowed to eat deeper into the free page list.
776 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) {
777 page_req = VM_ALLOC_SYSTEM;
782 mtx_lock_spin(&vm_page_queue_free_mtx);
783 if (cnt.v_free_count > cnt.v_free_reserved ||
784 (page_req == VM_ALLOC_SYSTEM &&
785 cnt.v_cache_count == 0 &&
786 cnt.v_free_count > cnt.v_interrupt_free_min) ||
787 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0)) {
789 * Allocate from the free queue if the number of free pages
790 * exceeds the minimum for the request class.
792 m = vm_page_select_free(color, (req & VM_ALLOC_ZERO) != 0);
793 } else if (page_req != VM_ALLOC_INTERRUPT) {
794 mtx_unlock_spin(&vm_page_queue_free_mtx);
796 * Allocatable from cache (non-interrupt only). On success,
797 * we must free the page and try again, thus ensuring that
798 * cnt.v_*_free_min counters are replenished.
800 vm_page_lock_queues();
801 if ((m = vm_page_select_cache(color)) == NULL) {
802 vm_page_unlock_queues();
804 #if defined(DIAGNOSTIC)
805 if (cnt.v_cache_count > 0)
806 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count);
808 atomic_add_int(&vm_pageout_deficit, 1);
812 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m));
816 vm_page_unlock_queues();
820 * Not allocatable from cache from interrupt, give up.
822 mtx_unlock_spin(&vm_page_queue_free_mtx);
824 atomic_add_int(&vm_pageout_deficit, 1);
830 * At this point we had better have found a good page.
835 ("vm_page_alloc(): missing page on free queue\n")
839 * Remove from free queue
842 vm_pageq_remove_nowakeup(m);
845 * Initialize structure. Only the PG_ZERO flag is inherited.
848 if (m->flags & PG_ZERO) {
849 vm_page_zero_count--;
850 if (req & VM_ALLOC_ZERO)
851 flags = PG_ZERO | PG_BUSY;
854 if (req & VM_ALLOC_WIRED) {
855 atomic_add_int(&cnt.v_wire_count, 1);
863 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m));
864 mtx_unlock_spin(&vm_page_queue_free_mtx);
867 * vm_page_insert() is safe prior to the splx(). Note also that
868 * inserting a page here does not insert it into the pmap (which
869 * could cause us to block allocating memory). We cannot block
872 if ((req & VM_ALLOC_NOOBJ) == 0)
873 vm_page_insert(m, object, pindex);
876 * Don't wakeup too often - wakeup the pageout daemon when
877 * we would be nearly out of memory.
879 if (vm_paging_needed())
887 * vm_wait: (also see VM_WAIT macro)
889 * Block until free pages are available for allocation
890 * - Called in various places before memory allocations.
898 if (curproc == pageproc) {
899 vm_pageout_pages_needed = 1;
900 tsleep(&vm_pageout_pages_needed, PSWP, "VMWait", 0);
902 if (!vm_pages_needed) {
904 wakeup(&vm_pages_needed);
906 tsleep(&cnt.v_free_count, PVM, "vmwait", 0);
912 * vm_waitpfault: (also see VM_WAITPFAULT macro)
914 * Block until free pages are available for allocation
915 * - Called only in vm_fault so that processes page faulting
916 * can be easily tracked.
917 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
918 * processes will be able to grab memory first. Do not change
919 * this balance without careful testing first.
927 if (!vm_pages_needed) {
929 wakeup(&vm_pages_needed);
931 tsleep(&cnt.v_free_count, PUSER, "pfault", 0);
938 * Put the specified page on the active list (if appropriate).
939 * Ensure that act_count is at least ACT_INIT but do not otherwise
942 * The page queues must be locked.
943 * This routine may not block.
946 vm_page_activate(vm_page_t m)
950 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
952 if (m->queue != PQ_ACTIVE) {
953 if ((m->queue - m->pc) == PQ_CACHE)
956 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
957 if (m->act_count < ACT_INIT)
958 m->act_count = ACT_INIT;
959 vm_pageq_enqueue(PQ_ACTIVE, m);
962 if (m->act_count < ACT_INIT)
963 m->act_count = ACT_INIT;
969 * vm_page_free_wakeup:
971 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
972 * routine is called when a page has been added to the cache or free
975 * This routine may not block.
976 * This routine must be called at splvm()
979 vm_page_free_wakeup(void)
982 * if pageout daemon needs pages, then tell it that there are
985 if (vm_pageout_pages_needed &&
986 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) {
987 wakeup(&vm_pageout_pages_needed);
988 vm_pageout_pages_needed = 0;
991 * wakeup processes that are waiting on memory if we hit a
992 * high water mark. And wakeup scheduler process if we have
993 * lots of memory. this process will swapin processes.
995 if (vm_pages_needed && !vm_page_count_min()) {
997 wakeup(&cnt.v_free_count);
1004 * Returns the given page to the PQ_FREE list,
1005 * disassociating it with any VM object.
1007 * Object and page must be locked prior to entry.
1008 * This routine may not block.
1012 vm_page_free_toq(vm_page_t m)
1015 struct vpgqueues *pq;
1016 vm_object_t object = m->object;
1019 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1023 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
1025 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
1026 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
1028 if ((m->queue - m->pc) == PQ_FREE)
1029 panic("vm_page_free: freeing free page");
1031 panic("vm_page_free: freeing busy page");
1035 * unqueue, then remove page. Note that we cannot destroy
1036 * the page here because we do not want to call the pager's
1037 * callback routine until after we've put the page on the
1038 * appropriate free queue.
1040 vm_pageq_remove_nowakeup(m);
1044 * If fictitious remove object association and
1045 * return, otherwise delay object association removal.
1047 if ((m->flags & PG_FICTITIOUS) != 0) {
1055 if (m->wire_count != 0) {
1056 if (m->wire_count > 1) {
1057 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1058 m->wire_count, (long)m->pindex);
1060 panic("vm_page_free: freeing wired page\n");
1064 * If we've exhausted the object's resident pages we want to free
1068 (object->type == OBJT_VNODE) &&
1069 ((object->flags & OBJ_DEAD) == 0)
1071 struct vnode *vp = (struct vnode *)object->handle;
1075 if (VSHOULDFREE(vp))
1082 * Clear the UNMANAGED flag when freeing an unmanaged page.
1084 if (m->flags & PG_UNMANAGED) {
1085 m->flags &= ~PG_UNMANAGED;
1088 pmap_page_is_free(m);
1092 if (m->hold_count != 0) {
1093 m->flags &= ~PG_ZERO;
1096 m->queue = PQ_FREE + m->pc;
1097 pq = &vm_page_queues[m->queue];
1098 mtx_lock_spin(&vm_page_queue_free_mtx);
1103 * Put zero'd pages on the end ( where we look for zero'd pages
1104 * first ) and non-zerod pages at the head.
1106 if (m->flags & PG_ZERO) {
1107 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
1108 ++vm_page_zero_count;
1110 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1112 mtx_unlock_spin(&vm_page_queue_free_mtx);
1113 vm_page_free_wakeup();
1120 * Prevent PV management from being done on the page. The page is
1121 * removed from the paging queues as if it were wired, and as a
1122 * consequence of no longer being managed the pageout daemon will not
1123 * touch it (since there is no way to locate the pte mappings for the
1124 * page). madvise() calls that mess with the pmap will also no longer
1125 * operate on the page.
1127 * Beyond that the page is still reasonably 'normal'. Freeing the page
1128 * will clear the flag.
1130 * This routine is used by OBJT_PHYS objects - objects using unswappable
1131 * physical memory as backing store rather then swap-backed memory and
1132 * will eventually be extended to support 4MB unmanaged physical
1136 vm_page_unmanage(vm_page_t m)
1141 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1142 if ((m->flags & PG_UNMANAGED) == 0) {
1143 if (m->wire_count == 0)
1146 vm_page_flag_set(m, PG_UNMANAGED);
1153 * Mark this page as wired down by yet
1154 * another map, removing it from paging queues
1157 * The page queues must be locked.
1158 * This routine may not block.
1161 vm_page_wire(vm_page_t m)
1166 * Only bump the wire statistics if the page is not already wired,
1167 * and only unqueue the page if it is on some queue (if it is unmanaged
1168 * it is already off the queues).
1171 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1172 if (m->wire_count == 0) {
1173 if ((m->flags & PG_UNMANAGED) == 0)
1175 atomic_add_int(&cnt.v_wire_count, 1);
1178 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
1185 * Release one wiring of this page, potentially
1186 * enabling it to be paged again.
1188 * Many pages placed on the inactive queue should actually go
1189 * into the cache, but it is difficult to figure out which. What
1190 * we do instead, if the inactive target is well met, is to put
1191 * clean pages at the head of the inactive queue instead of the tail.
1192 * This will cause them to be moved to the cache more quickly and
1193 * if not actively re-referenced, freed more quickly. If we just
1194 * stick these pages at the end of the inactive queue, heavy filesystem
1195 * meta-data accesses can cause an unnecessary paging load on memory bound
1196 * processes. This optimization causes one-time-use metadata to be
1197 * reused more quickly.
1199 * BUT, if we are in a low-memory situation we have no choice but to
1200 * put clean pages on the cache queue.
1202 * A number of routines use vm_page_unwire() to guarantee that the page
1203 * will go into either the inactive or active queues, and will NEVER
1204 * be placed in the cache - for example, just after dirtying a page.
1205 * dirty pages in the cache are not allowed.
1207 * The page queues must be locked.
1208 * This routine may not block.
1211 vm_page_unwire(vm_page_t m, int activate)
1216 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1217 if (m->wire_count > 0) {
1219 if (m->wire_count == 0) {
1220 atomic_subtract_int(&cnt.v_wire_count, 1);
1221 if (m->flags & PG_UNMANAGED) {
1223 } else if (activate)
1224 vm_pageq_enqueue(PQ_ACTIVE, m);
1226 vm_page_flag_clear(m, PG_WINATCFLS);
1227 vm_pageq_enqueue(PQ_INACTIVE, m);
1231 panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count);
1238 * Move the specified page to the inactive queue. If the page has
1239 * any associated swap, the swap is deallocated.
1241 * Normally athead is 0 resulting in LRU operation. athead is set
1242 * to 1 if we want this page to be 'as if it were placed in the cache',
1243 * except without unmapping it from the process address space.
1245 * This routine may not block.
1247 static __inline void
1248 _vm_page_deactivate(vm_page_t m, int athead)
1252 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1254 * Ignore if already inactive.
1256 if (m->queue == PQ_INACTIVE)
1260 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1261 if ((m->queue - m->pc) == PQ_CACHE)
1262 cnt.v_reactivated++;
1263 vm_page_flag_clear(m, PG_WINATCFLS);
1266 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1268 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1269 m->queue = PQ_INACTIVE;
1270 vm_page_queues[PQ_INACTIVE].lcnt++;
1271 cnt.v_inactive_count++;
1277 vm_page_deactivate(vm_page_t m)
1279 _vm_page_deactivate(m, 0);
1283 * vm_page_try_to_cache:
1285 * Returns 0 on failure, 1 on success
1288 vm_page_try_to_cache(vm_page_t m)
1291 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1292 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1293 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1296 vm_page_test_dirty(m);
1304 * vm_page_try_to_free()
1306 * Attempt to free the page. If we cannot free it, we do nothing.
1307 * 1 is returned on success, 0 on failure.
1310 vm_page_try_to_free(vm_page_t m)
1313 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1314 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1315 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1318 vm_page_test_dirty(m);
1330 * Put the specified page onto the page cache queue (if appropriate).
1332 * This routine may not block.
1335 vm_page_cache(vm_page_t m)
1339 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1340 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || m->wire_count) {
1341 printf("vm_page_cache: attempting to cache busy page\n");
1344 if ((m->queue - m->pc) == PQ_CACHE)
1348 * Remove all pmaps and indicate that the page is not
1349 * writeable or mapped.
1352 if (m->dirty != 0) {
1353 panic("vm_page_cache: caching a dirty page, pindex: %ld",
1357 vm_pageq_remove_nowakeup(m);
1358 vm_pageq_enqueue(PQ_CACHE + m->pc, m);
1359 vm_page_free_wakeup();
1366 * Cache, deactivate, or do nothing as appropriate. This routine
1367 * is typically used by madvise() MADV_DONTNEED.
1369 * Generally speaking we want to move the page into the cache so
1370 * it gets reused quickly. However, this can result in a silly syndrome
1371 * due to the page recycling too quickly. Small objects will not be
1372 * fully cached. On the otherhand, if we move the page to the inactive
1373 * queue we wind up with a problem whereby very large objects
1374 * unnecessarily blow away our inactive and cache queues.
1376 * The solution is to move the pages based on a fixed weighting. We
1377 * either leave them alone, deactivate them, or move them to the cache,
1378 * where moving them to the cache has the highest weighting.
1379 * By forcing some pages into other queues we eventually force the
1380 * system to balance the queues, potentially recovering other unrelated
1381 * space from active. The idea is to not force this to happen too
1385 vm_page_dontneed(vm_page_t m)
1387 static int dnweight;
1391 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1395 * occassionally leave the page alone
1397 if ((dnw & 0x01F0) == 0 ||
1398 m->queue == PQ_INACTIVE ||
1399 m->queue - m->pc == PQ_CACHE
1401 if (m->act_count >= ACT_INIT)
1407 vm_page_test_dirty(m);
1409 if (m->dirty || (dnw & 0x0070) == 0) {
1411 * Deactivate the page 3 times out of 32.
1416 * Cache the page 28 times out of every 32. Note that
1417 * the page is deactivated instead of cached, but placed
1418 * at the head of the queue instead of the tail.
1422 _vm_page_deactivate(m, head);
1426 * Grab a page, waiting until we are waken up due to the page
1427 * changing state. We keep on waiting, if the page continues
1428 * to be in the object. If the page doesn't exist, allocate it.
1430 * This routine may block.
1433 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
1440 if ((m = vm_page_lookup(object, pindex)) != NULL) {
1441 vm_page_lock_queues();
1442 if (m->busy || (m->flags & PG_BUSY)) {
1443 generation = object->generation;
1446 while ((object->generation == generation) &&
1447 (m->busy || (m->flags & PG_BUSY))) {
1448 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
1449 msleep(m, &vm_page_queue_mtx, PVM, "pgrbwt", 0);
1450 if ((allocflags & VM_ALLOC_RETRY) == 0) {
1451 vm_page_unlock_queues();
1456 vm_page_unlock_queues();
1460 if (allocflags & VM_ALLOC_WIRED)
1463 vm_page_unlock_queues();
1468 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1471 if ((allocflags & VM_ALLOC_RETRY) == 0)
1480 * Mapping function for valid bits or for dirty bits in
1481 * a page. May not block.
1483 * Inputs are required to range within a page.
1486 vm_page_bits(int base, int size)
1492 base + size <= PAGE_SIZE,
1493 ("vm_page_bits: illegal base/size %d/%d", base, size)
1496 if (size == 0) /* handle degenerate case */
1499 first_bit = base >> DEV_BSHIFT;
1500 last_bit = (base + size - 1) >> DEV_BSHIFT;
1502 return ((2 << last_bit) - (1 << first_bit));
1506 * vm_page_set_validclean:
1508 * Sets portions of a page valid and clean. The arguments are expected
1509 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1510 * of any partial chunks touched by the range. The invalid portion of
1511 * such chunks will be zero'd.
1513 * This routine may not block.
1515 * (base + size) must be less then or equal to PAGE_SIZE.
1518 vm_page_set_validclean(vm_page_t m, int base, int size)
1524 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1525 if (size == 0) /* handle degenerate case */
1529 * If the base is not DEV_BSIZE aligned and the valid
1530 * bit is clear, we have to zero out a portion of the
1533 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
1534 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
1535 pmap_zero_page_area(m, frag, base - frag);
1538 * If the ending offset is not DEV_BSIZE aligned and the
1539 * valid bit is clear, we have to zero out a portion of
1542 endoff = base + size;
1543 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
1544 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
1545 pmap_zero_page_area(m, endoff,
1546 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
1549 * Set valid, clear dirty bits. If validating the entire
1550 * page we can safely clear the pmap modify bit. We also
1551 * use this opportunity to clear the PG_NOSYNC flag. If a process
1552 * takes a write fault on a MAP_NOSYNC memory area the flag will
1555 * We set valid bits inclusive of any overlap, but we can only
1556 * clear dirty bits for DEV_BSIZE chunks that are fully within
1559 pagebits = vm_page_bits(base, size);
1560 m->valid |= pagebits;
1562 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
1563 frag = DEV_BSIZE - frag;
1569 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
1571 m->dirty &= ~pagebits;
1572 if (base == 0 && size == PAGE_SIZE) {
1573 pmap_clear_modify(m);
1574 vm_page_flag_clear(m, PG_NOSYNC);
1581 vm_page_set_dirty(vm_page_t m, int base, int size)
1583 m->dirty |= vm_page_bits(base, size);
1589 vm_page_clear_dirty(vm_page_t m, int base, int size)
1592 m->dirty &= ~vm_page_bits(base, size);
1596 * vm_page_set_invalid:
1598 * Invalidates DEV_BSIZE'd chunks within a page. Both the
1599 * valid and dirty bits for the effected areas are cleared.
1604 vm_page_set_invalid(vm_page_t m, int base, int size)
1609 bits = vm_page_bits(base, size);
1612 m->object->generation++;
1616 * vm_page_zero_invalid()
1618 * The kernel assumes that the invalid portions of a page contain
1619 * garbage, but such pages can be mapped into memory by user code.
1620 * When this occurs, we must zero out the non-valid portions of the
1621 * page so user code sees what it expects.
1623 * Pages are most often semi-valid when the end of a file is mapped
1624 * into memory and the file's size is not page aligned.
1627 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
1633 * Scan the valid bits looking for invalid sections that
1634 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
1635 * valid bit may be set ) have already been zerod by
1636 * vm_page_set_validclean().
1638 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
1639 if (i == (PAGE_SIZE / DEV_BSIZE) ||
1640 (m->valid & (1 << i))
1643 pmap_zero_page_area(m,
1644 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
1651 * setvalid is TRUE when we can safely set the zero'd areas
1652 * as being valid. We can do this if there are no cache consistancy
1653 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
1656 m->valid = VM_PAGE_BITS_ALL;
1662 * Is (partial) page valid? Note that the case where size == 0
1663 * will return FALSE in the degenerate case where the page is
1664 * entirely invalid, and TRUE otherwise.
1669 vm_page_is_valid(vm_page_t m, int base, int size)
1671 int bits = vm_page_bits(base, size);
1673 if (m->valid && ((m->valid & bits) == bits))
1680 * update dirty bits from pmap/mmu. May not block.
1683 vm_page_test_dirty(vm_page_t m)
1685 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
1690 int so_zerocp_fullpage = 0;
1693 vm_page_cowfault(vm_page_t m)
1706 * An interrupt allocation is requested because the page
1707 * queues lock is held.
1709 mnew = vm_page_alloc(object, pindex, VM_ALLOC_INTERRUPT);
1711 vm_page_insert(m, object, pindex);
1712 vm_page_unlock_queues();
1714 vm_page_lock_queues();
1720 * check to see if we raced with an xmit complete when
1721 * waiting to allocate a page. If so, put things back
1726 vm_page_insert(m, object, pindex);
1727 } else { /* clear COW & copy page */
1728 if (so_zerocp_fullpage) {
1729 mnew->valid = VM_PAGE_BITS_ALL;
1731 vm_page_copy(m, mnew);
1733 vm_page_dirty(mnew);
1734 vm_page_flag_clear(mnew, PG_BUSY);
1739 vm_page_cowclear(vm_page_t m)
1742 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1746 * let vm_fault add back write permission lazily
1750 * sf_buf_free() will free the page, so we needn't do it here
1755 vm_page_cowsetup(vm_page_t m)
1758 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1760 pmap_page_protect(m, VM_PROT_READ);
1763 #include "opt_ddb.h"
1765 #include <sys/kernel.h>
1767 #include <ddb/ddb.h>
1769 DB_SHOW_COMMAND(page, vm_page_print_page_info)
1771 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count);
1772 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count);
1773 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count);
1774 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count);
1775 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count);
1776 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved);
1777 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min);
1778 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target);
1779 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min);
1780 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target);
1783 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
1786 db_printf("PQ_FREE:");
1787 for (i = 0; i < PQ_L2_SIZE; i++) {
1788 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
1792 db_printf("PQ_CACHE:");
1793 for (i = 0; i < PQ_L2_SIZE; i++) {
1794 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
1798 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
1799 vm_page_queues[PQ_ACTIVE].lcnt,
1800 vm_page_queues[PQ_INACTIVE].lcnt);