2 * Copyright (c) 1991 Regents of the University of California.
4 * Copyright (c) 1998 Matthew Dillon. All Rights Reserved.
6 * This code is derived from software contributed to Berkeley by
7 * The Mach Operating System project at Carnegie-Mellon University.
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
17 * 4. Neither the name of the University nor the names of its contributors
18 * may be used to endorse or promote products derived from this software
19 * without specific prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91
37 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38 * All rights reserved.
40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
42 * Permission to use, copy, modify and distribute this software and
43 * its documentation is hereby granted, provided that both the copyright
44 * notice and this permission notice appear in all copies of the
45 * software, derivative works or modified versions, and any portions
46 * thereof, and that both notices appear in supporting documentation.
48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
52 * Carnegie Mellon requests users of this software to return to
54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
55 * School of Computer Science
56 * Carnegie Mellon University
57 * Pittsburgh PA 15213-3890
59 * any improvements or extensions that they make and grant Carnegie the
60 * rights to redistribute these changes.
64 * GENERAL RULES ON VM_PAGE MANIPULATION
66 * - A page queue lock is required when adding or removing a page from a
67 * page queue (vm_pagequeues[]), regardless of other locks or the
68 * busy state of a page.
70 * * In general, no thread besides the page daemon can acquire or
71 * hold more than one page queue lock at a time.
73 * * The page daemon can acquire and hold any pair of page queue
76 * - The object mutex is held when inserting or removing
77 * pages from an object (vm_page_insert() or vm_page_remove()).
82 * Resident memory management module.
85 #include <sys/cdefs.h>
86 __FBSDID("$FreeBSD$");
90 #include <sys/param.h>
91 #include <sys/systm.h>
93 #include <sys/kernel.h>
94 #include <sys/limits.h>
95 #include <sys/malloc.h>
96 #include <sys/msgbuf.h>
97 #include <sys/mutex.h>
99 #include <sys/sysctl.h>
100 #include <sys/vmmeter.h>
101 #include <sys/vnode.h>
105 #include <vm/vm_param.h>
106 #include <vm/vm_kern.h>
107 #include <vm/vm_object.h>
108 #include <vm/vm_page.h>
109 #include <vm/vm_pageout.h>
110 #include <vm/vm_pager.h>
111 #include <vm/vm_phys.h>
112 #include <vm/vm_radix.h>
113 #include <vm/vm_reserv.h>
114 #include <vm/vm_extern.h>
116 #include <vm/uma_int.h>
118 #include <machine/md_var.h>
121 * Associated with page of user-allocatable memory is a
125 struct vm_pagequeue vm_pagequeues[PQ_COUNT] = {
127 .pq_pl = TAILQ_HEAD_INITIALIZER(
128 vm_pagequeues[PQ_INACTIVE].pq_pl),
129 .pq_cnt = &cnt.v_inactive_count,
130 .pq_name = "vm inactive pagequeue"
133 .pq_pl = TAILQ_HEAD_INITIALIZER(
134 vm_pagequeues[PQ_ACTIVE].pq_pl),
135 .pq_cnt = &cnt.v_active_count,
136 .pq_name = "vm active pagequeue"
139 struct mtx_padalign vm_page_queue_free_mtx;
141 struct mtx_padalign pa_lock[PA_LOCK_COUNT];
143 vm_page_t vm_page_array;
144 long vm_page_array_size;
146 int vm_page_zero_count;
148 static int boot_pages = UMA_BOOT_PAGES;
149 TUNABLE_INT("vm.boot_pages", &boot_pages);
150 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RD, &boot_pages, 0,
151 "number of pages allocated for bootstrapping the VM system");
153 static int pa_tryrelock_restart;
154 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
155 &pa_tryrelock_restart, 0, "Number of tryrelock restarts");
157 static uma_zone_t fakepg_zone;
159 static struct vnode *vm_page_alloc_init(vm_page_t m);
160 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
161 static void vm_page_enqueue(int queue, vm_page_t m);
162 static void vm_page_init_fakepg(void *dummy);
164 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init_fakepg, NULL);
167 vm_page_init_fakepg(void *dummy)
170 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
171 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
174 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
175 #if PAGE_SIZE == 32768
177 CTASSERT(sizeof(u_long) >= 8);
182 * Try to acquire a physical address lock while a pmap is locked. If we
183 * fail to trylock we unlock and lock the pmap directly and cache the
184 * locked pa in *locked. The caller should then restart their loop in case
185 * the virtual to physical mapping has changed.
188 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
195 PA_LOCK_ASSERT(lockpa, MA_OWNED);
196 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
203 atomic_add_int(&pa_tryrelock_restart, 1);
212 * Sets the page size, perhaps based upon the memory
213 * size. Must be called before any use of page-size
214 * dependent functions.
217 vm_set_page_size(void)
219 if (cnt.v_page_size == 0)
220 cnt.v_page_size = PAGE_SIZE;
221 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0)
222 panic("vm_set_page_size: page size not a power of two");
226 * vm_page_blacklist_lookup:
228 * See if a physical address in this page has been listed
229 * in the blacklist tunable. Entries in the tunable are
230 * separated by spaces or commas. If an invalid integer is
231 * encountered then the rest of the string is skipped.
234 vm_page_blacklist_lookup(char *list, vm_paddr_t pa)
239 for (pos = list; *pos != '\0'; pos = cp) {
240 bad = strtoq(pos, &cp, 0);
242 if (*cp == ' ' || *cp == ',') {
249 if (pa == trunc_page(bad))
258 * Initializes the resident memory module.
260 * Allocates memory for the page cells, and
261 * for the object/offset-to-page hash table headers.
262 * Each page cell is initialized and placed on the free list.
265 vm_page_startup(vm_offset_t vaddr)
268 vm_paddr_t page_range;
275 /* the biggest memory array is the second group of pages */
277 vm_paddr_t biggestsize;
278 vm_paddr_t low_water, high_water;
283 vaddr = round_page(vaddr);
285 for (i = 0; phys_avail[i + 1]; i += 2) {
286 phys_avail[i] = round_page(phys_avail[i]);
287 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
290 low_water = phys_avail[0];
291 high_water = phys_avail[1];
293 for (i = 0; phys_avail[i + 1]; i += 2) {
294 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
296 if (size > biggestsize) {
300 if (phys_avail[i] < low_water)
301 low_water = phys_avail[i];
302 if (phys_avail[i + 1] > high_water)
303 high_water = phys_avail[i + 1];
310 end = phys_avail[biggestone+1];
313 * Initialize the page and queue locks.
315 mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF |
317 for (i = 0; i < PA_LOCK_COUNT; i++)
318 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
319 for (i = 0; i < PQ_COUNT; i++)
320 vm_pagequeue_init_lock(&vm_pagequeues[i]);
323 * Allocate memory for use when boot strapping the kernel memory
326 new_end = end - (boot_pages * UMA_SLAB_SIZE);
327 new_end = trunc_page(new_end);
328 mapped = pmap_map(&vaddr, new_end, end,
329 VM_PROT_READ | VM_PROT_WRITE);
330 bzero((void *)mapped, end - new_end);
331 uma_startup((void *)mapped, boot_pages);
333 #if defined(__amd64__) || defined(__i386__) || defined(__arm__) || \
336 * Allocate a bitmap to indicate that a random physical page
337 * needs to be included in a minidump.
339 * The amd64 port needs this to indicate which direct map pages
340 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
342 * However, i386 still needs this workspace internally within the
343 * minidump code. In theory, they are not needed on i386, but are
344 * included should the sf_buf code decide to use them.
347 for (i = 0; dump_avail[i + 1] != 0; i += 2)
348 if (dump_avail[i + 1] > last_pa)
349 last_pa = dump_avail[i + 1];
350 page_range = last_pa / PAGE_SIZE;
351 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
352 new_end -= vm_page_dump_size;
353 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
354 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
355 bzero((void *)vm_page_dump, vm_page_dump_size);
359 * Request that the physical pages underlying the message buffer be
360 * included in a crash dump. Since the message buffer is accessed
361 * through the direct map, they are not automatically included.
363 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
364 last_pa = pa + round_page(msgbufsize);
365 while (pa < last_pa) {
371 * Compute the number of pages of memory that will be available for
372 * use (taking into account the overhead of a page structure per
375 first_page = low_water / PAGE_SIZE;
376 #ifdef VM_PHYSSEG_SPARSE
378 for (i = 0; phys_avail[i + 1] != 0; i += 2)
379 page_range += atop(phys_avail[i + 1] - phys_avail[i]);
380 #elif defined(VM_PHYSSEG_DENSE)
381 page_range = high_water / PAGE_SIZE - first_page;
383 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
388 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
393 * Initialize the mem entry structures now, and put them in the free
396 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
397 mapped = pmap_map(&vaddr, new_end, end,
398 VM_PROT_READ | VM_PROT_WRITE);
399 vm_page_array = (vm_page_t) mapped;
400 #if VM_NRESERVLEVEL > 0
402 * Allocate memory for the reservation management system's data
405 new_end = vm_reserv_startup(&vaddr, new_end, high_water);
407 #if defined(__amd64__) || defined(__mips__)
409 * pmap_map on amd64 and mips can come out of the direct-map, not kvm
410 * like i386, so the pages must be tracked for a crashdump to include
411 * this data. This includes the vm_page_array and the early UMA
414 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
417 phys_avail[biggestone + 1] = new_end;
420 * Clear all of the page structures
422 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
423 for (i = 0; i < page_range; i++)
424 vm_page_array[i].order = VM_NFREEORDER;
425 vm_page_array_size = page_range;
428 * Initialize the physical memory allocator.
433 * Add every available physical page that is not blacklisted to
436 cnt.v_page_count = 0;
437 cnt.v_free_count = 0;
438 list = getenv("vm.blacklist");
439 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
441 last_pa = phys_avail[i + 1];
442 while (pa < last_pa) {
444 vm_page_blacklist_lookup(list, pa))
445 printf("Skipping page with pa 0x%jx\n",
448 vm_phys_add_page(pa);
453 #if VM_NRESERVLEVEL > 0
455 * Initialize the reservation management system.
463 vm_page_reference(vm_page_t m)
466 vm_page_aflag_set(m, PGA_REFERENCED);
470 vm_page_busy(vm_page_t m)
473 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
474 KASSERT((m->oflags & VPO_BUSY) == 0,
475 ("vm_page_busy: page already busy!!!"));
476 m->oflags |= VPO_BUSY;
482 * wakeup anyone waiting for the page.
485 vm_page_flash(vm_page_t m)
488 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
489 if (m->oflags & VPO_WANTED) {
490 m->oflags &= ~VPO_WANTED;
498 * clear the VPO_BUSY flag and wakeup anyone waiting for the
503 vm_page_wakeup(vm_page_t m)
506 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
507 KASSERT(m->oflags & VPO_BUSY, ("vm_page_wakeup: page not busy!!!"));
508 m->oflags &= ~VPO_BUSY;
513 vm_page_io_start(vm_page_t m)
516 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
521 vm_page_io_finish(vm_page_t m)
524 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
525 KASSERT(m->busy > 0, ("vm_page_io_finish: page %p is not busy", m));
532 * Keep page from being freed by the page daemon
533 * much of the same effect as wiring, except much lower
534 * overhead and should be used only for *very* temporary
535 * holding ("wiring").
538 vm_page_hold(vm_page_t mem)
541 vm_page_lock_assert(mem, MA_OWNED);
546 vm_page_unhold(vm_page_t mem)
549 vm_page_lock_assert(mem, MA_OWNED);
551 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
552 if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0)
553 vm_page_free_toq(mem);
557 * vm_page_unhold_pages:
559 * Unhold each of the pages that is referenced by the given array.
562 vm_page_unhold_pages(vm_page_t *ma, int count)
564 struct mtx *mtx, *new_mtx;
567 for (; count != 0; count--) {
569 * Avoid releasing and reacquiring the same page lock.
571 new_mtx = vm_page_lockptr(*ma);
572 if (mtx != new_mtx) {
586 PHYS_TO_VM_PAGE(vm_paddr_t pa)
590 #ifdef VM_PHYSSEG_SPARSE
591 m = vm_phys_paddr_to_vm_page(pa);
593 m = vm_phys_fictitious_to_vm_page(pa);
595 #elif defined(VM_PHYSSEG_DENSE)
599 if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
600 m = &vm_page_array[pi - first_page];
603 return (vm_phys_fictitious_to_vm_page(pa));
605 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
612 * Create a fictitious page with the specified physical address and
613 * memory attribute. The memory attribute is the only the machine-
614 * dependent aspect of a fictitious page that must be initialized.
617 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
621 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
622 vm_page_initfake(m, paddr, memattr);
627 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
630 if ((m->flags & PG_FICTITIOUS) != 0) {
632 * The page's memattr might have changed since the
633 * previous initialization. Update the pmap to the
638 m->phys_addr = paddr;
640 /* Fictitious pages don't use "segind". */
641 m->flags = PG_FICTITIOUS;
642 /* Fictitious pages don't use "order" or "pool". */
643 m->oflags = VPO_BUSY | VPO_UNMANAGED;
646 pmap_page_set_memattr(m, memattr);
652 * Release a fictitious page.
655 vm_page_putfake(vm_page_t m)
658 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
659 KASSERT((m->flags & PG_FICTITIOUS) != 0,
660 ("vm_page_putfake: bad page %p", m));
661 uma_zfree(fakepg_zone, m);
665 * vm_page_updatefake:
667 * Update the given fictitious page to the specified physical address and
671 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
674 KASSERT((m->flags & PG_FICTITIOUS) != 0,
675 ("vm_page_updatefake: bad page %p", m));
676 m->phys_addr = paddr;
677 pmap_page_set_memattr(m, memattr);
686 vm_page_free(vm_page_t m)
689 m->flags &= ~PG_ZERO;
696 * Free a page to the zerod-pages queue
699 vm_page_free_zero(vm_page_t m)
707 * Unbusy and handle the page queueing for a page from the VOP_GETPAGES()
708 * array which is not the request page.
711 vm_page_readahead_finish(vm_page_t m)
716 * Since the page is not the requested page, whether
717 * it should be activated or deactivated is not
718 * obvious. Empirical results have shown that
719 * deactivating the page is usually the best choice,
720 * unless the page is wanted by another thread.
722 if (m->oflags & VPO_WANTED) {
728 vm_page_deactivate(m);
734 * Free the completely invalid page. Such page state
735 * occurs due to the short read operation which did
736 * not covered our page at all, or in case when a read
748 * Sleep and release the page lock.
750 * The object containing the given page must be locked.
753 vm_page_sleep(vm_page_t m, const char *msg)
756 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
757 if (mtx_owned(vm_page_lockptr(m)))
761 * It's possible that while we sleep, the page will get
762 * unbusied and freed. If we are holding the object
763 * lock, we will assume we hold a reference to the object
764 * such that even if m->object changes, we can re-lock
767 m->oflags |= VPO_WANTED;
768 msleep(m, VM_OBJECT_MTX(m->object), PVM, msg, 0);
772 * vm_page_dirty_KBI: [ internal use only ]
774 * Set all bits in the page's dirty field.
776 * The object containing the specified page must be locked if the
777 * call is made from the machine-independent layer.
779 * See vm_page_clear_dirty_mask().
781 * This function should only be called by vm_page_dirty().
784 vm_page_dirty_KBI(vm_page_t m)
787 /* These assertions refer to this operation by its public name. */
788 KASSERT((m->flags & PG_CACHED) == 0,
789 ("vm_page_dirty: page in cache!"));
790 KASSERT(!VM_PAGE_IS_FREE(m),
791 ("vm_page_dirty: page is free!"));
792 KASSERT(m->valid == VM_PAGE_BITS_ALL,
793 ("vm_page_dirty: page is invalid!"));
794 m->dirty = VM_PAGE_BITS_ALL;
798 * vm_page_insert: [ internal use only ]
800 * Inserts the given mem entry into the object and object list.
802 * The pagetables are not updated but will presumably fault the page
803 * in if necessary, or if a kernel page the caller will at some point
804 * enter the page into the kernel's pmap. We are not allowed to sleep
805 * here so we *can't* do this anyway.
807 * The object must be locked.
810 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
814 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
815 if (m->object != NULL)
816 panic("vm_page_insert: page already inserted");
819 * Record the object/offset pair in this page
824 if (object->resident_page_count == 0) {
825 TAILQ_INSERT_TAIL(&object->memq, m, listq);
827 neighbor = vm_radix_lookup_ge(&object->rtree, pindex);
828 if (neighbor != NULL) {
829 KASSERT(pindex < neighbor->pindex,
830 ("vm_page_insert: offset %ju not minor than %ju",
831 (uintmax_t)pindex, (uintmax_t)neighbor->pindex));
832 TAILQ_INSERT_BEFORE(neighbor, m, listq);
834 TAILQ_INSERT_TAIL(&object->memq, m, listq);
836 if (vm_radix_insert(&object->rtree, pindex, m) != 0)
837 panic("vm_page_insert: unable to insert the new page");
840 * Show that the object has one more resident page.
842 object->resident_page_count++;
845 * Hold the vnode until the last page is released.
847 if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
848 vhold(object->handle);
851 * Since we are inserting a new and possibly dirty page,
852 * update the object's OBJ_MIGHTBEDIRTY flag.
854 if (pmap_page_is_write_mapped(m))
855 vm_object_set_writeable_dirty(object);
861 * Removes the given mem entry from the object/offset-page
862 * table and the object page list, but do not invalidate/terminate
865 * The underlying pmap entry (if any) is NOT removed here.
867 * The object must be locked. The page must be locked if it is managed.
870 vm_page_remove(vm_page_t m)
874 if ((m->oflags & VPO_UNMANAGED) == 0)
875 vm_page_lock_assert(m, MA_OWNED);
876 if ((object = m->object) == NULL)
878 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
879 if (m->oflags & VPO_BUSY) {
880 m->oflags &= ~VPO_BUSY;
884 vm_radix_remove(&object->rtree, m->pindex);
885 TAILQ_REMOVE(&object->memq, m, listq);
888 * And show that the object has one fewer resident page.
890 object->resident_page_count--;
893 * The vnode may now be recycled.
895 if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
896 vdrop(object->handle);
904 * Returns the page associated with the object/offset
905 * pair specified; if none is found, NULL is returned.
907 * The object must be locked.
910 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
913 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
915 return vm_radix_lookup(&object->rtree, pindex);
919 * vm_page_find_least:
921 * Returns the page associated with the object with least pindex
922 * greater than or equal to the parameter pindex, or NULL.
924 * The object must be locked.
927 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
930 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
931 if (object->resident_page_count)
932 return (vm_radix_lookup_ge(&object->rtree, pindex));
937 * Returns the given page's successor (by pindex) within the object if it is
938 * resident; if none is found, NULL is returned.
940 * The object must be locked.
943 vm_page_next(vm_page_t m)
947 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
948 if ((next = TAILQ_NEXT(m, listq)) != NULL &&
949 next->pindex != m->pindex + 1)
955 * Returns the given page's predecessor (by pindex) within the object if it is
956 * resident; if none is found, NULL is returned.
958 * The object must be locked.
961 vm_page_prev(vm_page_t m)
965 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
966 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
967 prev->pindex != m->pindex - 1)
975 * Move the given memory entry from its
976 * current object to the specified target object/offset.
978 * Note: swap associated with the page must be invalidated by the move. We
979 * have to do this for several reasons: (1) we aren't freeing the
980 * page, (2) we are dirtying the page, (3) the VM system is probably
981 * moving the page from object A to B, and will then later move
982 * the backing store from A to B and we can't have a conflict.
984 * Note: we *always* dirty the page. It is necessary both for the
985 * fact that we moved it, and because we may be invalidating
986 * swap. If the page is on the cache, we have to deactivate it
987 * or vm_page_dirty() will panic. Dirty pages are not allowed
990 * The objects must be locked. The page must be locked if it is managed.
993 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
997 vm_page_insert(m, new_object, new_pindex);
1002 * Convert all of the given object's cached pages that have a
1003 * pindex within the given range into free pages. If the value
1004 * zero is given for "end", then the range's upper bound is
1005 * infinity. If the given object is backed by a vnode and it
1006 * transitions from having one or more cached pages to none, the
1007 * vnode's hold count is reduced.
1010 vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1015 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1017 mtx_lock(&vm_page_queue_free_mtx);
1018 if (vm_object_cache_is_empty(object)) {
1019 mtx_unlock(&vm_page_queue_free_mtx);
1022 while ((m = vm_radix_lookup_ge(&object->cache, start)) != NULL) {
1023 if (end != 0 && m->pindex >= end)
1025 vm_radix_remove(&object->cache, m->pindex);
1028 /* Clear PG_CACHED and set PG_FREE. */
1029 m->flags ^= PG_CACHED | PG_FREE;
1030 KASSERT((m->flags & (PG_CACHED | PG_FREE)) == PG_FREE,
1031 ("vm_page_cache_free: page %p has inconsistent flags", m));
1032 cnt.v_cache_count--;
1035 empty = vm_object_cache_is_empty(object);
1036 mtx_unlock(&vm_page_queue_free_mtx);
1037 if (object->type == OBJT_VNODE && empty)
1038 vdrop(object->handle);
1042 * Returns the cached page that is associated with the given
1043 * object and offset. If, however, none exists, returns NULL.
1045 * The free page queue must be locked.
1047 static inline vm_page_t
1048 vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex)
1051 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1052 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1053 if (!vm_object_cache_is_empty(object))
1054 return (vm_radix_lookup(&object->cache, pindex));
1059 * Remove the given cached page from its containing object's
1060 * collection of cached pages.
1062 * The free page queue must be locked.
1065 vm_page_cache_remove(vm_page_t m)
1068 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1069 KASSERT((m->flags & PG_CACHED) != 0,
1070 ("vm_page_cache_remove: page %p is not cached", m));
1071 vm_radix_remove(&m->object->cache, m->pindex);
1073 cnt.v_cache_count--;
1077 * Transfer all of the cached pages with offset greater than or
1078 * equal to 'offidxstart' from the original object's cache to the
1079 * new object's cache. However, any cached pages with offset
1080 * greater than or equal to the new object's size are kept in the
1081 * original object. Initially, the new object's cache must be
1082 * empty. Offset 'offidxstart' in the original object must
1083 * correspond to offset zero in the new object.
1085 * The new object must be locked.
1088 vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart,
1089 vm_object_t new_object)
1094 * Insertion into an object's collection of cached pages
1095 * requires the object to be locked. In contrast, removal does
1098 VM_OBJECT_LOCK_ASSERT(new_object, MA_OWNED);
1099 VM_OBJECT_LOCK_ASSERT(orig_object, MA_OWNED);
1100 KASSERT(vm_object_cache_is_empty(new_object),
1101 ("vm_page_cache_transfer: object %p has cached pages",
1103 mtx_lock(&vm_page_queue_free_mtx);
1104 while ((m = vm_radix_lookup_ge(&orig_object->cache,
1105 offidxstart)) != NULL) {
1106 if ((m->pindex - offidxstart) >= new_object->size)
1108 vm_radix_remove(&orig_object->cache, m->pindex);
1109 if (vm_radix_insert(&new_object->cache,
1110 m->pindex - offidxstart, m) != 0)
1111 panic("vm_page_cache_transfer: failed vm_radix_insert");
1112 m->object = new_object;
1113 m->pindex -= offidxstart;
1115 mtx_unlock(&vm_page_queue_free_mtx);
1119 * Returns TRUE if a cached page is associated with the given object and
1120 * offset, and FALSE otherwise.
1122 * The object must be locked.
1125 vm_page_is_cached(vm_object_t object, vm_pindex_t pindex)
1130 * Insertion into an object's collection of cached pages requires the
1131 * object to be locked. Therefore, if the object is locked and the
1132 * object's collection is empty, there is no need to acquire the free
1133 * page queues lock in order to prove that the specified page doesn't
1136 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1137 if (vm_object_cache_is_empty(object))
1139 mtx_lock(&vm_page_queue_free_mtx);
1140 m = vm_page_cache_lookup(object, pindex);
1141 mtx_unlock(&vm_page_queue_free_mtx);
1148 * Allocate and return a page that is associated with the specified
1149 * object and offset pair. By default, this page has the flag VPO_BUSY
1152 * The caller must always specify an allocation class.
1154 * allocation classes:
1155 * VM_ALLOC_NORMAL normal process request
1156 * VM_ALLOC_SYSTEM system *really* needs a page
1157 * VM_ALLOC_INTERRUPT interrupt time request
1159 * optional allocation flags:
1160 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1161 * intends to allocate
1162 * VM_ALLOC_IFCACHED return page only if it is cached
1163 * VM_ALLOC_IFNOTCACHED return NULL, do not reactivate if the page
1165 * VM_ALLOC_NOBUSY do not set the flag VPO_BUSY on the page
1166 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1167 * VM_ALLOC_NOOBJ page is not associated with an object and
1168 * should not have the flag VPO_BUSY set
1169 * VM_ALLOC_WIRED wire the allocated page
1170 * VM_ALLOC_ZERO prefer a zeroed page
1172 * This routine may not sleep.
1175 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1177 struct vnode *vp = NULL;
1178 vm_object_t m_object;
1180 int flags, req_class;
1182 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0),
1183 ("vm_page_alloc: inconsistent object/req"));
1185 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1187 req_class = req & VM_ALLOC_CLASS_MASK;
1190 * The page daemon is allowed to dig deeper into the free page list.
1192 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1193 req_class = VM_ALLOC_SYSTEM;
1195 mtx_lock(&vm_page_queue_free_mtx);
1196 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved ||
1197 (req_class == VM_ALLOC_SYSTEM &&
1198 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) ||
1199 (req_class == VM_ALLOC_INTERRUPT &&
1200 cnt.v_free_count + cnt.v_cache_count > 0)) {
1202 * Allocate from the free queue if the number of free pages
1203 * exceeds the minimum for the request class.
1205 if (object != NULL &&
1206 (m = vm_page_cache_lookup(object, pindex)) != NULL) {
1207 if ((req & VM_ALLOC_IFNOTCACHED) != 0) {
1208 mtx_unlock(&vm_page_queue_free_mtx);
1211 if (vm_phys_unfree_page(m))
1212 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0);
1213 #if VM_NRESERVLEVEL > 0
1214 else if (!vm_reserv_reactivate_page(m))
1218 panic("vm_page_alloc: cache page %p is missing"
1219 " from the free queue", m);
1220 } else if ((req & VM_ALLOC_IFCACHED) != 0) {
1221 mtx_unlock(&vm_page_queue_free_mtx);
1223 #if VM_NRESERVLEVEL > 0
1224 } else if (object == NULL || (object->flags & (OBJ_COLORED |
1225 OBJ_FICTITIOUS)) != OBJ_COLORED ||
1226 (m = vm_reserv_alloc_page(object, pindex)) == NULL) {
1230 m = vm_phys_alloc_pages(object != NULL ?
1231 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1232 #if VM_NRESERVLEVEL > 0
1233 if (m == NULL && vm_reserv_reclaim_inactive()) {
1234 m = vm_phys_alloc_pages(object != NULL ?
1235 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1242 * Not allocatable, give up.
1244 mtx_unlock(&vm_page_queue_free_mtx);
1245 atomic_add_int(&vm_pageout_deficit,
1246 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1247 pagedaemon_wakeup();
1252 * At this point we had better have found a good page.
1254 KASSERT(m != NULL, ("vm_page_alloc: missing page"));
1255 KASSERT(m->queue == PQ_NONE,
1256 ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue));
1257 KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m));
1258 KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m));
1259 KASSERT(m->busy == 0, ("vm_page_alloc: page %p is busy", m));
1260 KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m));
1261 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1262 ("vm_page_alloc: page %p has unexpected memattr %d", m,
1263 pmap_page_get_memattr(m)));
1264 if ((m->flags & PG_CACHED) != 0) {
1265 KASSERT((m->flags & PG_ZERO) == 0,
1266 ("vm_page_alloc: cached page %p is PG_ZERO", m));
1267 KASSERT(m->valid != 0,
1268 ("vm_page_alloc: cached page %p is invalid", m));
1269 if (m->object == object && m->pindex == pindex)
1270 cnt.v_reactivated++;
1273 m_object = m->object;
1274 vm_page_cache_remove(m);
1275 if (m_object->type == OBJT_VNODE &&
1276 vm_object_cache_is_empty(m_object))
1277 vp = m_object->handle;
1279 KASSERT(VM_PAGE_IS_FREE(m),
1280 ("vm_page_alloc: page %p is not free", m));
1281 KASSERT(m->valid == 0,
1282 ("vm_page_alloc: free page %p is valid", m));
1287 * Only the PG_ZERO flag is inherited. The PG_CACHED or PG_FREE flag
1288 * must be cleared before the free page queues lock is released.
1291 if (m->flags & PG_ZERO) {
1292 vm_page_zero_count--;
1293 if (req & VM_ALLOC_ZERO)
1296 if (req & VM_ALLOC_NODUMP)
1299 mtx_unlock(&vm_page_queue_free_mtx);
1301 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1303 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ)) == 0)
1304 m->oflags |= VPO_BUSY;
1305 if (req & VM_ALLOC_WIRED) {
1307 * The page lock is not required for wiring a page until that
1308 * page is inserted into the object.
1310 atomic_add_int(&cnt.v_wire_count, 1);
1315 if (object != NULL) {
1316 /* Ignore device objects; the pager sets "memattr" for them. */
1317 if (object->memattr != VM_MEMATTR_DEFAULT &&
1318 (object->flags & OBJ_FICTITIOUS) == 0)
1319 pmap_page_set_memattr(m, object->memattr);
1320 vm_page_insert(m, object, pindex);
1325 * The following call to vdrop() must come after the above call
1326 * to vm_page_insert() in case both affect the same object and
1327 * vnode. Otherwise, the affected vnode's hold count could
1328 * temporarily become zero.
1334 * Don't wakeup too often - wakeup the pageout daemon when
1335 * we would be nearly out of memory.
1337 if (vm_paging_needed())
1338 pagedaemon_wakeup();
1344 * vm_page_alloc_contig:
1346 * Allocate a contiguous set of physical pages of the given size "npages"
1347 * from the free lists. All of the physical pages must be at or above
1348 * the given physical address "low" and below the given physical address
1349 * "high". The given value "alignment" determines the alignment of the
1350 * first physical page in the set. If the given value "boundary" is
1351 * non-zero, then the set of physical pages cannot cross any physical
1352 * address boundary that is a multiple of that value. Both "alignment"
1353 * and "boundary" must be a power of two.
1355 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1356 * then the memory attribute setting for the physical pages is configured
1357 * to the object's memory attribute setting. Otherwise, the memory
1358 * attribute setting for the physical pages is configured to "memattr",
1359 * overriding the object's memory attribute setting. However, if the
1360 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1361 * memory attribute setting for the physical pages cannot be configured
1362 * to VM_MEMATTR_DEFAULT.
1364 * The caller must always specify an allocation class.
1366 * allocation classes:
1367 * VM_ALLOC_NORMAL normal process request
1368 * VM_ALLOC_SYSTEM system *really* needs a page
1369 * VM_ALLOC_INTERRUPT interrupt time request
1371 * optional allocation flags:
1372 * VM_ALLOC_NOBUSY do not set the flag VPO_BUSY on the page
1373 * VM_ALLOC_NOOBJ page is not associated with an object and
1374 * should not have the flag VPO_BUSY set
1375 * VM_ALLOC_WIRED wire the allocated page
1376 * VM_ALLOC_ZERO prefer a zeroed page
1378 * This routine may not sleep.
1381 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1382 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1383 vm_paddr_t boundary, vm_memattr_t memattr)
1386 vm_page_t deferred_vdrop_list, m, m_ret;
1387 u_int flags, oflags;
1390 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0),
1391 ("vm_page_alloc_contig: inconsistent object/req"));
1392 if (object != NULL) {
1393 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1394 KASSERT(object->type == OBJT_PHYS,
1395 ("vm_page_alloc_contig: object %p isn't OBJT_PHYS",
1398 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
1399 req_class = req & VM_ALLOC_CLASS_MASK;
1402 * The page daemon is allowed to dig deeper into the free page list.
1404 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1405 req_class = VM_ALLOC_SYSTEM;
1407 deferred_vdrop_list = NULL;
1408 mtx_lock(&vm_page_queue_free_mtx);
1409 if (cnt.v_free_count + cnt.v_cache_count >= npages +
1410 cnt.v_free_reserved || (req_class == VM_ALLOC_SYSTEM &&
1411 cnt.v_free_count + cnt.v_cache_count >= npages +
1412 cnt.v_interrupt_free_min) || (req_class == VM_ALLOC_INTERRUPT &&
1413 cnt.v_free_count + cnt.v_cache_count >= npages)) {
1414 #if VM_NRESERVLEVEL > 0
1416 if (object == NULL || (object->flags & OBJ_COLORED) == 0 ||
1417 (m_ret = vm_reserv_alloc_contig(object, pindex, npages,
1418 low, high, alignment, boundary)) == NULL)
1420 m_ret = vm_phys_alloc_contig(npages, low, high,
1421 alignment, boundary);
1423 mtx_unlock(&vm_page_queue_free_mtx);
1424 atomic_add_int(&vm_pageout_deficit, npages);
1425 pagedaemon_wakeup();
1429 for (m = m_ret; m < &m_ret[npages]; m++) {
1430 drop = vm_page_alloc_init(m);
1433 * Enqueue the vnode for deferred vdrop().
1435 * Once the pages are removed from the free
1436 * page list, "pageq" can be safely abused to
1437 * construct a short-lived list of vnodes.
1439 m->pageq.tqe_prev = (void *)drop;
1440 m->pageq.tqe_next = deferred_vdrop_list;
1441 deferred_vdrop_list = m;
1445 #if VM_NRESERVLEVEL > 0
1446 if (vm_reserv_reclaim_contig(npages, low, high, alignment,
1451 mtx_unlock(&vm_page_queue_free_mtx);
1456 * Initialize the pages. Only the PG_ZERO flag is inherited.
1459 if ((req & VM_ALLOC_ZERO) != 0)
1461 if ((req & VM_ALLOC_NODUMP) != 0)
1463 if ((req & VM_ALLOC_WIRED) != 0)
1464 atomic_add_int(&cnt.v_wire_count, npages);
1465 oflags = VPO_UNMANAGED;
1466 if (object != NULL) {
1467 if ((req & VM_ALLOC_NOBUSY) == 0)
1469 if (object->memattr != VM_MEMATTR_DEFAULT &&
1470 memattr == VM_MEMATTR_DEFAULT)
1471 memattr = object->memattr;
1473 for (m = m_ret; m < &m_ret[npages]; m++) {
1475 m->flags = (m->flags | PG_NODUMP) & flags;
1476 if ((req & VM_ALLOC_WIRED) != 0)
1478 /* Unmanaged pages don't use "act_count". */
1480 if (memattr != VM_MEMATTR_DEFAULT)
1481 pmap_page_set_memattr(m, memattr);
1483 vm_page_insert(m, object, pindex);
1488 while (deferred_vdrop_list != NULL) {
1489 vdrop((struct vnode *)deferred_vdrop_list->pageq.tqe_prev);
1490 deferred_vdrop_list = deferred_vdrop_list->pageq.tqe_next;
1492 if (vm_paging_needed())
1493 pagedaemon_wakeup();
1498 * Initialize a page that has been freshly dequeued from a freelist.
1499 * The caller has to drop the vnode returned, if it is not NULL.
1501 * This function may only be used to initialize unmanaged pages.
1503 * To be called with vm_page_queue_free_mtx held.
1505 static struct vnode *
1506 vm_page_alloc_init(vm_page_t m)
1509 vm_object_t m_object;
1511 KASSERT(m->queue == PQ_NONE,
1512 ("vm_page_alloc_init: page %p has unexpected queue %d",
1514 KASSERT(m->wire_count == 0,
1515 ("vm_page_alloc_init: page %p is wired", m));
1516 KASSERT(m->hold_count == 0,
1517 ("vm_page_alloc_init: page %p is held", m));
1518 KASSERT(m->busy == 0,
1519 ("vm_page_alloc_init: page %p is busy", m));
1520 KASSERT(m->dirty == 0,
1521 ("vm_page_alloc_init: page %p is dirty", m));
1522 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1523 ("vm_page_alloc_init: page %p has unexpected memattr %d",
1524 m, pmap_page_get_memattr(m)));
1525 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1527 if ((m->flags & PG_CACHED) != 0) {
1528 KASSERT((m->flags & PG_ZERO) == 0,
1529 ("vm_page_alloc_init: cached page %p is PG_ZERO", m));
1531 m_object = m->object;
1532 vm_page_cache_remove(m);
1533 if (m_object->type == OBJT_VNODE &&
1534 vm_object_cache_is_empty(m_object))
1535 drop = m_object->handle;
1537 KASSERT(VM_PAGE_IS_FREE(m),
1538 ("vm_page_alloc_init: page %p is not free", m));
1539 KASSERT(m->valid == 0,
1540 ("vm_page_alloc_init: free page %p is valid", m));
1542 if ((m->flags & PG_ZERO) != 0)
1543 vm_page_zero_count--;
1545 /* Don't clear the PG_ZERO flag; we'll need it later. */
1546 m->flags &= PG_ZERO;
1551 * vm_page_alloc_freelist:
1553 * Allocate a physical page from the specified free page list.
1555 * The caller must always specify an allocation class.
1557 * allocation classes:
1558 * VM_ALLOC_NORMAL normal process request
1559 * VM_ALLOC_SYSTEM system *really* needs a page
1560 * VM_ALLOC_INTERRUPT interrupt time request
1562 * optional allocation flags:
1563 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1564 * intends to allocate
1565 * VM_ALLOC_WIRED wire the allocated page
1566 * VM_ALLOC_ZERO prefer a zeroed page
1568 * This routine may not sleep.
1571 vm_page_alloc_freelist(int flind, int req)
1578 req_class = req & VM_ALLOC_CLASS_MASK;
1581 * The page daemon is allowed to dig deeper into the free page list.
1583 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1584 req_class = VM_ALLOC_SYSTEM;
1587 * Do not allocate reserved pages unless the req has asked for it.
1589 mtx_lock(&vm_page_queue_free_mtx);
1590 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved ||
1591 (req_class == VM_ALLOC_SYSTEM &&
1592 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) ||
1593 (req_class == VM_ALLOC_INTERRUPT &&
1594 cnt.v_free_count + cnt.v_cache_count > 0))
1595 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0);
1597 mtx_unlock(&vm_page_queue_free_mtx);
1598 atomic_add_int(&vm_pageout_deficit,
1599 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1600 pagedaemon_wakeup();
1604 mtx_unlock(&vm_page_queue_free_mtx);
1607 drop = vm_page_alloc_init(m);
1608 mtx_unlock(&vm_page_queue_free_mtx);
1611 * Initialize the page. Only the PG_ZERO flag is inherited.
1615 if ((req & VM_ALLOC_ZERO) != 0)
1618 if ((req & VM_ALLOC_WIRED) != 0) {
1620 * The page lock is not required for wiring a page that does
1621 * not belong to an object.
1623 atomic_add_int(&cnt.v_wire_count, 1);
1626 /* Unmanaged pages don't use "act_count". */
1627 m->oflags = VPO_UNMANAGED;
1630 if (vm_paging_needed())
1631 pagedaemon_wakeup();
1636 * vm_wait: (also see VM_WAIT macro)
1638 * Sleep until free pages are available for allocation.
1639 * - Called in various places before memory allocations.
1645 mtx_lock(&vm_page_queue_free_mtx);
1646 if (curproc == pageproc) {
1647 vm_pageout_pages_needed = 1;
1648 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx,
1649 PDROP | PSWP, "VMWait", 0);
1651 if (!vm_pages_needed) {
1652 vm_pages_needed = 1;
1653 wakeup(&vm_pages_needed);
1655 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM,
1661 * vm_waitpfault: (also see VM_WAITPFAULT macro)
1663 * Sleep until free pages are available for allocation.
1664 * - Called only in vm_fault so that processes page faulting
1665 * can be easily tracked.
1666 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
1667 * processes will be able to grab memory first. Do not change
1668 * this balance without careful testing first.
1674 mtx_lock(&vm_page_queue_free_mtx);
1675 if (!vm_pages_needed) {
1676 vm_pages_needed = 1;
1677 wakeup(&vm_pages_needed);
1679 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER,
1686 * Remove the given page from its current page queue.
1688 * The page must be locked.
1691 vm_page_dequeue(vm_page_t m)
1693 struct vm_pagequeue *pq;
1695 vm_page_lock_assert(m, MA_OWNED);
1696 KASSERT(m->queue != PQ_NONE,
1697 ("vm_page_dequeue: page %p is not queued", m));
1698 pq = &vm_pagequeues[m->queue];
1699 vm_pagequeue_lock(pq);
1701 TAILQ_REMOVE(&pq->pq_pl, m, pageq);
1703 vm_pagequeue_unlock(pq);
1707 * vm_page_dequeue_locked:
1709 * Remove the given page from its current page queue.
1711 * The page and page queue must be locked.
1714 vm_page_dequeue_locked(vm_page_t m)
1716 struct vm_pagequeue *pq;
1718 vm_page_lock_assert(m, MA_OWNED);
1719 pq = &vm_pagequeues[m->queue];
1720 vm_pagequeue_assert_locked(pq);
1722 TAILQ_REMOVE(&pq->pq_pl, m, pageq);
1729 * Add the given page to the specified page queue.
1731 * The page must be locked.
1734 vm_page_enqueue(int queue, vm_page_t m)
1736 struct vm_pagequeue *pq;
1738 vm_page_lock_assert(m, MA_OWNED);
1739 pq = &vm_pagequeues[queue];
1740 vm_pagequeue_lock(pq);
1742 TAILQ_INSERT_TAIL(&pq->pq_pl, m, pageq);
1744 vm_pagequeue_unlock(pq);
1750 * Move the given page to the tail of its current page queue.
1752 * The page must be locked.
1755 vm_page_requeue(vm_page_t m)
1757 struct vm_pagequeue *pq;
1759 vm_page_lock_assert(m, MA_OWNED);
1760 KASSERT(m->queue != PQ_NONE,
1761 ("vm_page_requeue: page %p is not queued", m));
1762 pq = &vm_pagequeues[m->queue];
1763 vm_pagequeue_lock(pq);
1764 TAILQ_REMOVE(&pq->pq_pl, m, pageq);
1765 TAILQ_INSERT_TAIL(&pq->pq_pl, m, pageq);
1766 vm_pagequeue_unlock(pq);
1770 * vm_page_requeue_locked:
1772 * Move the given page to the tail of its current page queue.
1774 * The page queue must be locked.
1777 vm_page_requeue_locked(vm_page_t m)
1779 struct vm_pagequeue *pq;
1781 KASSERT(m->queue != PQ_NONE,
1782 ("vm_page_requeue_locked: page %p is not queued", m));
1783 pq = &vm_pagequeues[m->queue];
1784 vm_pagequeue_assert_locked(pq);
1785 TAILQ_REMOVE(&pq->pq_pl, m, pageq);
1786 TAILQ_INSERT_TAIL(&pq->pq_pl, m, pageq);
1792 * Put the specified page on the active list (if appropriate).
1793 * Ensure that act_count is at least ACT_INIT but do not otherwise
1796 * The page must be locked.
1799 vm_page_activate(vm_page_t m)
1803 vm_page_lock_assert(m, MA_OWNED);
1804 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1805 if ((queue = m->queue) != PQ_ACTIVE) {
1806 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
1807 if (m->act_count < ACT_INIT)
1808 m->act_count = ACT_INIT;
1809 if (queue != PQ_NONE)
1811 vm_page_enqueue(PQ_ACTIVE, m);
1813 KASSERT(queue == PQ_NONE,
1814 ("vm_page_activate: wired page %p is queued", m));
1816 if (m->act_count < ACT_INIT)
1817 m->act_count = ACT_INIT;
1822 * vm_page_free_wakeup:
1824 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
1825 * routine is called when a page has been added to the cache or free
1828 * The page queues must be locked.
1831 vm_page_free_wakeup(void)
1834 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1836 * if pageout daemon needs pages, then tell it that there are
1839 if (vm_pageout_pages_needed &&
1840 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) {
1841 wakeup(&vm_pageout_pages_needed);
1842 vm_pageout_pages_needed = 0;
1845 * wakeup processes that are waiting on memory if we hit a
1846 * high water mark. And wakeup scheduler process if we have
1847 * lots of memory. this process will swapin processes.
1849 if (vm_pages_needed && !vm_page_count_min()) {
1850 vm_pages_needed = 0;
1851 wakeup(&cnt.v_free_count);
1858 * Returns the given page to the free list,
1859 * disassociating it with any VM object.
1861 * The object must be locked. The page must be locked if it is managed.
1864 vm_page_free_toq(vm_page_t m)
1867 if ((m->oflags & VPO_UNMANAGED) == 0) {
1868 vm_page_lock_assert(m, MA_OWNED);
1869 KASSERT(!pmap_page_is_mapped(m),
1870 ("vm_page_free_toq: freeing mapped page %p", m));
1872 KASSERT(m->queue == PQ_NONE,
1873 ("vm_page_free_toq: unmanaged page %p is queued", m));
1874 PCPU_INC(cnt.v_tfree);
1876 if (VM_PAGE_IS_FREE(m))
1877 panic("vm_page_free: freeing free page %p", m);
1878 else if (m->busy != 0)
1879 panic("vm_page_free: freeing busy page %p", m);
1882 * Unqueue, then remove page. Note that we cannot destroy
1883 * the page here because we do not want to call the pager's
1884 * callback routine until after we've put the page on the
1885 * appropriate free queue.
1891 * If fictitious remove object association and
1892 * return, otherwise delay object association removal.
1894 if ((m->flags & PG_FICTITIOUS) != 0) {
1901 if (m->wire_count != 0)
1902 panic("vm_page_free: freeing wired page %p", m);
1903 if (m->hold_count != 0) {
1904 m->flags &= ~PG_ZERO;
1905 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
1906 ("vm_page_free: freeing PG_UNHOLDFREE page %p", m));
1907 m->flags |= PG_UNHOLDFREE;
1910 * Restore the default memory attribute to the page.
1912 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
1913 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
1916 * Insert the page into the physical memory allocator's
1917 * cache/free page queues.
1919 mtx_lock(&vm_page_queue_free_mtx);
1920 m->flags |= PG_FREE;
1922 #if VM_NRESERVLEVEL > 0
1923 if (!vm_reserv_free_page(m))
1927 vm_phys_free_pages(m, 0);
1928 if ((m->flags & PG_ZERO) != 0)
1929 ++vm_page_zero_count;
1931 vm_page_zero_idle_wakeup();
1932 vm_page_free_wakeup();
1933 mtx_unlock(&vm_page_queue_free_mtx);
1940 * Mark this page as wired down by yet
1941 * another map, removing it from paging queues
1944 * If the page is fictitious, then its wire count must remain one.
1946 * The page must be locked.
1949 vm_page_wire(vm_page_t m)
1953 * Only bump the wire statistics if the page is not already wired,
1954 * and only unqueue the page if it is on some queue (if it is unmanaged
1955 * it is already off the queues).
1957 vm_page_lock_assert(m, MA_OWNED);
1958 if ((m->flags & PG_FICTITIOUS) != 0) {
1959 KASSERT(m->wire_count == 1,
1960 ("vm_page_wire: fictitious page %p's wire count isn't one",
1964 if (m->wire_count == 0) {
1965 KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
1966 m->queue == PQ_NONE,
1967 ("vm_page_wire: unmanaged page %p is queued", m));
1969 atomic_add_int(&cnt.v_wire_count, 1);
1972 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
1978 * Release one wiring of the specified page, potentially enabling it to be
1979 * paged again. If paging is enabled, then the value of the parameter
1980 * "activate" determines to which queue the page is added. If "activate" is
1981 * non-zero, then the page is added to the active queue. Otherwise, it is
1982 * added to the inactive queue.
1984 * However, unless the page belongs to an object, it is not enqueued because
1985 * it cannot be paged out.
1987 * If a page is fictitious, then its wire count must alway be one.
1989 * A managed page must be locked.
1992 vm_page_unwire(vm_page_t m, int activate)
1995 if ((m->oflags & VPO_UNMANAGED) == 0)
1996 vm_page_lock_assert(m, MA_OWNED);
1997 if ((m->flags & PG_FICTITIOUS) != 0) {
1998 KASSERT(m->wire_count == 1,
1999 ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
2002 if (m->wire_count > 0) {
2004 if (m->wire_count == 0) {
2005 atomic_subtract_int(&cnt.v_wire_count, 1);
2006 if ((m->oflags & VPO_UNMANAGED) != 0 ||
2010 m->flags &= ~PG_WINATCFLS;
2011 vm_page_enqueue(activate ? PQ_ACTIVE : PQ_INACTIVE, m);
2014 panic("vm_page_unwire: page %p's wire count is zero", m);
2018 * Move the specified page to the inactive queue.
2020 * Many pages placed on the inactive queue should actually go
2021 * into the cache, but it is difficult to figure out which. What
2022 * we do instead, if the inactive target is well met, is to put
2023 * clean pages at the head of the inactive queue instead of the tail.
2024 * This will cause them to be moved to the cache more quickly and
2025 * if not actively re-referenced, reclaimed more quickly. If we just
2026 * stick these pages at the end of the inactive queue, heavy filesystem
2027 * meta-data accesses can cause an unnecessary paging load on memory bound
2028 * processes. This optimization causes one-time-use metadata to be
2029 * reused more quickly.
2031 * Normally athead is 0 resulting in LRU operation. athead is set
2032 * to 1 if we want this page to be 'as if it were placed in the cache',
2033 * except without unmapping it from the process address space.
2035 * The page must be locked.
2038 _vm_page_deactivate(vm_page_t m, int athead)
2040 struct vm_pagequeue *pq;
2043 vm_page_lock_assert(m, MA_OWNED);
2046 * Ignore if already inactive.
2048 if ((queue = m->queue) == PQ_INACTIVE)
2050 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2051 if (queue != PQ_NONE)
2053 m->flags &= ~PG_WINATCFLS;
2054 pq = &vm_pagequeues[PQ_INACTIVE];
2055 vm_pagequeue_lock(pq);
2056 m->queue = PQ_INACTIVE;
2058 TAILQ_INSERT_HEAD(&pq->pq_pl, m, pageq);
2060 TAILQ_INSERT_TAIL(&pq->pq_pl, m, pageq);
2061 cnt.v_inactive_count++;
2062 vm_pagequeue_unlock(pq);
2067 * Move the specified page to the inactive queue.
2069 * The page must be locked.
2072 vm_page_deactivate(vm_page_t m)
2075 _vm_page_deactivate(m, 0);
2079 * vm_page_try_to_cache:
2081 * Returns 0 on failure, 1 on success
2084 vm_page_try_to_cache(vm_page_t m)
2087 vm_page_lock_assert(m, MA_OWNED);
2088 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2089 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
2090 (m->oflags & (VPO_BUSY | VPO_UNMANAGED)) != 0)
2100 * vm_page_try_to_free()
2102 * Attempt to free the page. If we cannot free it, we do nothing.
2103 * 1 is returned on success, 0 on failure.
2106 vm_page_try_to_free(vm_page_t m)
2109 vm_page_lock_assert(m, MA_OWNED);
2110 if (m->object != NULL)
2111 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2112 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
2113 (m->oflags & (VPO_BUSY | VPO_UNMANAGED)) != 0)
2125 * Put the specified page onto the page cache queue (if appropriate).
2127 * The object and page must be locked.
2130 vm_page_cache(vm_page_t m)
2133 int old_empty_cache;
2135 vm_page_lock_assert(m, MA_OWNED);
2137 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2138 if ((m->oflags & (VPO_UNMANAGED | VPO_BUSY)) || m->busy ||
2139 m->hold_count || m->wire_count)
2140 panic("vm_page_cache: attempting to cache busy page");
2141 KASSERT(!pmap_page_is_mapped(m),
2142 ("vm_page_cache: page %p is mapped", m));
2143 KASSERT(m->dirty == 0, ("vm_page_cache: page %p is dirty", m));
2144 if (m->valid == 0 || object->type == OBJT_DEFAULT ||
2145 (object->type == OBJT_SWAP &&
2146 !vm_pager_has_page(object, m->pindex, NULL, NULL))) {
2148 * Hypothesis: A cache-elgible page belonging to a
2149 * default object or swap object but without a backing
2150 * store must be zero filled.
2155 KASSERT((m->flags & PG_CACHED) == 0,
2156 ("vm_page_cache: page %p is already cached", m));
2157 PCPU_INC(cnt.v_tcached);
2160 * Remove the page from the paging queues.
2164 vm_radix_remove(&object->rtree, m->pindex);
2165 TAILQ_REMOVE(&object->memq, m, listq);
2166 object->resident_page_count--;
2169 * Restore the default memory attribute to the page.
2171 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2172 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2175 * Insert the page into the object's collection of cached pages
2176 * and the physical memory allocator's cache/free page queues.
2178 m->flags &= ~PG_ZERO;
2179 mtx_lock(&vm_page_queue_free_mtx);
2180 m->flags |= PG_CACHED;
2181 old_empty_cache = vm_object_cache_is_empty(object);
2182 cnt.v_cache_count++;
2183 if (vm_radix_insert(&object->cache, m->pindex, m) != 0)
2184 panic("vm_page_cache: vm_radix_insert failed");
2185 #if VM_NRESERVLEVEL > 0
2186 if (!vm_reserv_free_page(m)) {
2190 vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0);
2191 vm_phys_free_pages(m, 0);
2193 vm_page_free_wakeup();
2194 mtx_unlock(&vm_page_queue_free_mtx);
2197 * Increment the vnode's hold count if this is the object's only
2198 * cached page. Decrement the vnode's hold count if this was
2199 * the object's only resident page.
2201 if (object->type == OBJT_VNODE) {
2202 if (old_empty_cache != 0 && object->resident_page_count != 0)
2203 vhold(object->handle);
2204 else if (old_empty_cache == 0 &&
2205 object->resident_page_count == 0)
2206 vdrop(object->handle);
2213 * Cache, deactivate, or do nothing as appropriate. This routine
2214 * is typically used by madvise() MADV_DONTNEED.
2216 * Generally speaking we want to move the page into the cache so
2217 * it gets reused quickly. However, this can result in a silly syndrome
2218 * due to the page recycling too quickly. Small objects will not be
2219 * fully cached. On the otherhand, if we move the page to the inactive
2220 * queue we wind up with a problem whereby very large objects
2221 * unnecessarily blow away our inactive and cache queues.
2223 * The solution is to move the pages based on a fixed weighting. We
2224 * either leave them alone, deactivate them, or move them to the cache,
2225 * where moving them to the cache has the highest weighting.
2226 * By forcing some pages into other queues we eventually force the
2227 * system to balance the queues, potentially recovering other unrelated
2228 * space from active. The idea is to not force this to happen too
2231 * The object and page must be locked.
2234 vm_page_dontneed(vm_page_t m)
2239 vm_page_lock_assert(m, MA_OWNED);
2240 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2241 dnw = PCPU_GET(dnweight);
2245 * Occasionally leave the page alone.
2247 if ((dnw & 0x01F0) == 0 || m->queue == PQ_INACTIVE) {
2248 if (m->act_count >= ACT_INIT)
2254 * Clear any references to the page. Otherwise, the page daemon will
2255 * immediately reactivate the page.
2257 * Perform the pmap_clear_reference() first. Otherwise, a concurrent
2258 * pmap operation, such as pmap_remove(), could clear a reference in
2259 * the pmap and set PGA_REFERENCED on the page before the
2260 * pmap_clear_reference() had completed. Consequently, the page would
2261 * appear referenced based upon an old reference that occurred before
2262 * this function ran.
2264 pmap_clear_reference(m);
2265 vm_page_aflag_clear(m, PGA_REFERENCED);
2267 if (m->dirty == 0 && pmap_is_modified(m))
2270 if (m->dirty || (dnw & 0x0070) == 0) {
2272 * Deactivate the page 3 times out of 32.
2277 * Cache the page 28 times out of every 32. Note that
2278 * the page is deactivated instead of cached, but placed
2279 * at the head of the queue instead of the tail.
2283 _vm_page_deactivate(m, head);
2287 * Grab a page, waiting until we are waken up due to the page
2288 * changing state. We keep on waiting, if the page continues
2289 * to be in the object. If the page doesn't exist, first allocate it
2290 * and then conditionally zero it.
2292 * The caller must always specify the VM_ALLOC_RETRY flag. This is intended
2293 * to facilitate its eventual removal.
2295 * This routine may sleep.
2297 * The object must be locked on entry. The lock will, however, be released
2298 * and reacquired if the routine sleeps.
2301 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2305 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2306 KASSERT((allocflags & VM_ALLOC_RETRY) != 0,
2307 ("vm_page_grab: VM_ALLOC_RETRY is required"));
2309 if ((m = vm_page_lookup(object, pindex)) != NULL) {
2310 if ((m->oflags & VPO_BUSY) != 0 ||
2311 ((allocflags & VM_ALLOC_IGN_SBUSY) == 0 && m->busy != 0)) {
2313 * Reference the page before unlocking and
2314 * sleeping so that the page daemon is less
2315 * likely to reclaim it.
2317 vm_page_aflag_set(m, PGA_REFERENCED);
2318 vm_page_sleep(m, "pgrbwt");
2321 if ((allocflags & VM_ALLOC_WIRED) != 0) {
2326 if ((allocflags & VM_ALLOC_NOBUSY) == 0)
2331 m = vm_page_alloc(object, pindex, allocflags & ~(VM_ALLOC_RETRY |
2332 VM_ALLOC_IGN_SBUSY));
2334 VM_OBJECT_UNLOCK(object);
2336 VM_OBJECT_LOCK(object);
2338 } else if (m->valid != 0)
2340 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
2346 * Mapping function for valid or dirty bits in a page.
2348 * Inputs are required to range within a page.
2351 vm_page_bits(int base, int size)
2357 base + size <= PAGE_SIZE,
2358 ("vm_page_bits: illegal base/size %d/%d", base, size)
2361 if (size == 0) /* handle degenerate case */
2364 first_bit = base >> DEV_BSHIFT;
2365 last_bit = (base + size - 1) >> DEV_BSHIFT;
2367 return (((vm_page_bits_t)2 << last_bit) -
2368 ((vm_page_bits_t)1 << first_bit));
2372 * vm_page_set_valid_range:
2374 * Sets portions of a page valid. The arguments are expected
2375 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2376 * of any partial chunks touched by the range. The invalid portion of
2377 * such chunks will be zeroed.
2379 * (base + size) must be less then or equal to PAGE_SIZE.
2382 vm_page_set_valid_range(vm_page_t m, int base, int size)
2386 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2387 if (size == 0) /* handle degenerate case */
2391 * If the base is not DEV_BSIZE aligned and the valid
2392 * bit is clear, we have to zero out a portion of the
2395 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2396 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
2397 pmap_zero_page_area(m, frag, base - frag);
2400 * If the ending offset is not DEV_BSIZE aligned and the
2401 * valid bit is clear, we have to zero out a portion of
2404 endoff = base + size;
2405 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2406 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
2407 pmap_zero_page_area(m, endoff,
2408 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
2411 * Assert that no previously invalid block that is now being validated
2414 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
2415 ("vm_page_set_valid_range: page %p is dirty", m));
2418 * Set valid bits inclusive of any overlap.
2420 m->valid |= vm_page_bits(base, size);
2424 * Clear the given bits from the specified page's dirty field.
2426 static __inline void
2427 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
2430 #if PAGE_SIZE < 16384
2435 * If the object is locked and the page is neither VPO_BUSY nor
2436 * write mapped, then the page's dirty field cannot possibly be
2437 * set by a concurrent pmap operation.
2439 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2440 if ((m->oflags & VPO_BUSY) == 0 && !pmap_page_is_write_mapped(m))
2441 m->dirty &= ~pagebits;
2444 * The pmap layer can call vm_page_dirty() without
2445 * holding a distinguished lock. The combination of
2446 * the object's lock and an atomic operation suffice
2447 * to guarantee consistency of the page dirty field.
2449 * For PAGE_SIZE == 32768 case, compiler already
2450 * properly aligns the dirty field, so no forcible
2451 * alignment is needed. Only require existence of
2452 * atomic_clear_64 when page size is 32768.
2454 addr = (uintptr_t)&m->dirty;
2455 #if PAGE_SIZE == 32768
2456 atomic_clear_64((uint64_t *)addr, pagebits);
2457 #elif PAGE_SIZE == 16384
2458 atomic_clear_32((uint32_t *)addr, pagebits);
2459 #else /* PAGE_SIZE <= 8192 */
2461 * Use a trick to perform a 32-bit atomic on the
2462 * containing aligned word, to not depend on the existence
2463 * of atomic_clear_{8, 16}.
2465 shift = addr & (sizeof(uint32_t) - 1);
2466 #if BYTE_ORDER == BIG_ENDIAN
2467 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
2471 addr &= ~(sizeof(uint32_t) - 1);
2472 atomic_clear_32((uint32_t *)addr, pagebits << shift);
2473 #endif /* PAGE_SIZE */
2478 * vm_page_set_validclean:
2480 * Sets portions of a page valid and clean. The arguments are expected
2481 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2482 * of any partial chunks touched by the range. The invalid portion of
2483 * such chunks will be zero'd.
2485 * (base + size) must be less then or equal to PAGE_SIZE.
2488 vm_page_set_validclean(vm_page_t m, int base, int size)
2490 vm_page_bits_t oldvalid, pagebits;
2493 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2494 if (size == 0) /* handle degenerate case */
2498 * If the base is not DEV_BSIZE aligned and the valid
2499 * bit is clear, we have to zero out a portion of the
2502 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2503 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
2504 pmap_zero_page_area(m, frag, base - frag);
2507 * If the ending offset is not DEV_BSIZE aligned and the
2508 * valid bit is clear, we have to zero out a portion of
2511 endoff = base + size;
2512 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2513 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
2514 pmap_zero_page_area(m, endoff,
2515 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
2518 * Set valid, clear dirty bits. If validating the entire
2519 * page we can safely clear the pmap modify bit. We also
2520 * use this opportunity to clear the VPO_NOSYNC flag. If a process
2521 * takes a write fault on a MAP_NOSYNC memory area the flag will
2524 * We set valid bits inclusive of any overlap, but we can only
2525 * clear dirty bits for DEV_BSIZE chunks that are fully within
2528 oldvalid = m->valid;
2529 pagebits = vm_page_bits(base, size);
2530 m->valid |= pagebits;
2532 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
2533 frag = DEV_BSIZE - frag;
2539 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
2541 if (base == 0 && size == PAGE_SIZE) {
2543 * The page can only be modified within the pmap if it is
2544 * mapped, and it can only be mapped if it was previously
2547 if (oldvalid == VM_PAGE_BITS_ALL)
2549 * Perform the pmap_clear_modify() first. Otherwise,
2550 * a concurrent pmap operation, such as
2551 * pmap_protect(), could clear a modification in the
2552 * pmap and set the dirty field on the page before
2553 * pmap_clear_modify() had begun and after the dirty
2554 * field was cleared here.
2556 pmap_clear_modify(m);
2558 m->oflags &= ~VPO_NOSYNC;
2559 } else if (oldvalid != VM_PAGE_BITS_ALL)
2560 m->dirty &= ~pagebits;
2562 vm_page_clear_dirty_mask(m, pagebits);
2566 vm_page_clear_dirty(vm_page_t m, int base, int size)
2569 vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
2573 * vm_page_set_invalid:
2575 * Invalidates DEV_BSIZE'd chunks within a page. Both the
2576 * valid and dirty bits for the effected areas are cleared.
2579 vm_page_set_invalid(vm_page_t m, int base, int size)
2581 vm_page_bits_t bits;
2583 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2584 KASSERT((m->oflags & VPO_BUSY) == 0,
2585 ("vm_page_set_invalid: page %p is busy", m));
2586 bits = vm_page_bits(base, size);
2587 if (m->valid == VM_PAGE_BITS_ALL && bits != 0)
2589 KASSERT(!pmap_page_is_mapped(m),
2590 ("vm_page_set_invalid: page %p is mapped", m));
2596 * vm_page_zero_invalid()
2598 * The kernel assumes that the invalid portions of a page contain
2599 * garbage, but such pages can be mapped into memory by user code.
2600 * When this occurs, we must zero out the non-valid portions of the
2601 * page so user code sees what it expects.
2603 * Pages are most often semi-valid when the end of a file is mapped
2604 * into memory and the file's size is not page aligned.
2607 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
2612 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2614 * Scan the valid bits looking for invalid sections that
2615 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
2616 * valid bit may be set ) have already been zerod by
2617 * vm_page_set_validclean().
2619 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
2620 if (i == (PAGE_SIZE / DEV_BSIZE) ||
2621 (m->valid & ((vm_page_bits_t)1 << i))) {
2623 pmap_zero_page_area(m,
2624 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
2631 * setvalid is TRUE when we can safely set the zero'd areas
2632 * as being valid. We can do this if there are no cache consistancy
2633 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
2636 m->valid = VM_PAGE_BITS_ALL;
2642 * Is (partial) page valid? Note that the case where size == 0
2643 * will return FALSE in the degenerate case where the page is
2644 * entirely invalid, and TRUE otherwise.
2647 vm_page_is_valid(vm_page_t m, int base, int size)
2649 vm_page_bits_t bits;
2651 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2652 bits = vm_page_bits(base, size);
2653 if (m->valid && ((m->valid & bits) == bits))
2660 * Set the page's dirty bits if the page is modified.
2663 vm_page_test_dirty(vm_page_t m)
2666 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2667 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
2672 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
2675 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
2679 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
2682 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
2686 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
2689 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
2692 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
2694 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
2697 mtx_assert_(vm_page_lockptr(m), a, file, line);
2701 int so_zerocp_fullpage = 0;
2704 * Replace the given page with a copy. The copied page assumes
2705 * the portion of the given page's "wire_count" that is not the
2706 * responsibility of this copy-on-write mechanism.
2708 * The object containing the given page must have a non-zero
2709 * paging-in-progress count and be locked.
2712 vm_page_cowfault(vm_page_t m)
2718 vm_page_lock_assert(m, MA_OWNED);
2720 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2721 KASSERT(object->paging_in_progress != 0,
2722 ("vm_page_cowfault: object %p's paging-in-progress count is zero.",
2729 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY);
2731 vm_page_insert(m, object, pindex);
2733 VM_OBJECT_UNLOCK(object);
2735 VM_OBJECT_LOCK(object);
2736 if (m == vm_page_lookup(object, pindex)) {
2741 * Page disappeared during the wait.
2749 * check to see if we raced with an xmit complete when
2750 * waiting to allocate a page. If so, put things back
2756 vm_page_unlock(mnew);
2757 vm_page_insert(m, object, pindex);
2758 } else { /* clear COW & copy page */
2759 if (!so_zerocp_fullpage)
2760 pmap_copy_page(m, mnew);
2761 mnew->valid = VM_PAGE_BITS_ALL;
2762 vm_page_dirty(mnew);
2763 mnew->wire_count = m->wire_count - m->cow;
2764 m->wire_count = m->cow;
2770 vm_page_cowclear(vm_page_t m)
2773 vm_page_lock_assert(m, MA_OWNED);
2777 * let vm_fault add back write permission lazily
2781 * sf_buf_free() will free the page, so we needn't do it here
2786 vm_page_cowsetup(vm_page_t m)
2789 vm_page_lock_assert(m, MA_OWNED);
2790 if ((m->flags & PG_FICTITIOUS) != 0 ||
2791 (m->oflags & VPO_UNMANAGED) != 0 ||
2792 m->cow == USHRT_MAX - 1 || !VM_OBJECT_TRYLOCK(m->object))
2795 pmap_remove_write(m);
2796 VM_OBJECT_UNLOCK(m->object);
2802 vm_page_object_lock_assert(vm_page_t m)
2806 * Certain of the page's fields may only be modified by the
2807 * holder of the containing object's lock or the setter of the
2808 * page's VPO_BUSY flag. Unfortunately, the setter of the
2809 * VPO_BUSY flag is not recorded, and thus cannot be checked
2812 if (m->object != NULL && (m->oflags & VPO_BUSY) == 0)
2813 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2817 #include "opt_ddb.h"
2819 #include <sys/kernel.h>
2821 #include <ddb/ddb.h>
2823 DB_SHOW_COMMAND(page, vm_page_print_page_info)
2825 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count);
2826 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count);
2827 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count);
2828 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count);
2829 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count);
2830 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved);
2831 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min);
2832 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target);
2833 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min);
2834 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target);
2837 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
2840 db_printf("PQ_FREE:");
2841 db_printf(" %d", cnt.v_free_count);
2844 db_printf("PQ_CACHE:");
2845 db_printf(" %d", cnt.v_cache_count);
2848 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
2849 *vm_pagequeues[PQ_ACTIVE].pq_cnt,
2850 *vm_pagequeues[PQ_INACTIVE].pq_cnt);