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);
316 for (i = 0; i < PA_LOCK_COUNT; i++)
317 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
318 for (i = 0; i < PQ_COUNT; i++)
319 vm_pagequeue_init_lock(&vm_pagequeues[i]);
322 * Allocate memory for use when boot strapping the kernel memory
325 new_end = end - (boot_pages * UMA_SLAB_SIZE);
326 new_end = trunc_page(new_end);
327 mapped = pmap_map(&vaddr, new_end, end,
328 VM_PROT_READ | VM_PROT_WRITE);
329 bzero((void *)mapped, end - new_end);
330 uma_startup((void *)mapped, boot_pages);
332 #if defined(__amd64__) || defined(__i386__) || defined(__arm__) || \
335 * Allocate a bitmap to indicate that a random physical page
336 * needs to be included in a minidump.
338 * The amd64 port needs this to indicate which direct map pages
339 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
341 * However, i386 still needs this workspace internally within the
342 * minidump code. In theory, they are not needed on i386, but are
343 * included should the sf_buf code decide to use them.
346 for (i = 0; dump_avail[i + 1] != 0; i += 2)
347 if (dump_avail[i + 1] > last_pa)
348 last_pa = dump_avail[i + 1];
349 page_range = last_pa / PAGE_SIZE;
350 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
351 new_end -= vm_page_dump_size;
352 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
353 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
354 bzero((void *)vm_page_dump, vm_page_dump_size);
358 * Request that the physical pages underlying the message buffer be
359 * included in a crash dump. Since the message buffer is accessed
360 * through the direct map, they are not automatically included.
362 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
363 last_pa = pa + round_page(msgbufsize);
364 while (pa < last_pa) {
370 * Compute the number of pages of memory that will be available for
371 * use (taking into account the overhead of a page structure per
374 first_page = low_water / PAGE_SIZE;
375 #ifdef VM_PHYSSEG_SPARSE
377 for (i = 0; phys_avail[i + 1] != 0; i += 2)
378 page_range += atop(phys_avail[i + 1] - phys_avail[i]);
379 #elif defined(VM_PHYSSEG_DENSE)
380 page_range = high_water / PAGE_SIZE - first_page;
382 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
387 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
392 * Initialize the mem entry structures now, and put them in the free
395 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
396 mapped = pmap_map(&vaddr, new_end, end,
397 VM_PROT_READ | VM_PROT_WRITE);
398 vm_page_array = (vm_page_t) mapped;
399 #if VM_NRESERVLEVEL > 0
401 * Allocate memory for the reservation management system's data
404 new_end = vm_reserv_startup(&vaddr, new_end, high_water);
406 #if defined(__amd64__) || defined(__mips__)
408 * pmap_map on amd64 and mips can come out of the direct-map, not kvm
409 * like i386, so the pages must be tracked for a crashdump to include
410 * this data. This includes the vm_page_array and the early UMA
413 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
416 phys_avail[biggestone + 1] = new_end;
419 * Clear all of the page structures
421 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
422 for (i = 0; i < page_range; i++)
423 vm_page_array[i].order = VM_NFREEORDER;
424 vm_page_array_size = page_range;
427 * Initialize the physical memory allocator.
432 * Add every available physical page that is not blacklisted to
435 cnt.v_page_count = 0;
436 cnt.v_free_count = 0;
437 list = getenv("vm.blacklist");
438 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
440 last_pa = phys_avail[i + 1];
441 while (pa < last_pa) {
443 vm_page_blacklist_lookup(list, pa))
444 printf("Skipping page with pa 0x%jx\n",
447 vm_phys_add_page(pa);
452 #if VM_NRESERVLEVEL > 0
454 * Initialize the reservation management system.
462 vm_page_reference(vm_page_t m)
465 vm_page_aflag_set(m, PGA_REFERENCED);
469 vm_page_busy(vm_page_t m)
472 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
473 KASSERT((m->oflags & VPO_BUSY) == 0,
474 ("vm_page_busy: page already busy!!!"));
475 m->oflags |= VPO_BUSY;
481 * wakeup anyone waiting for the page.
484 vm_page_flash(vm_page_t m)
487 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
488 if (m->oflags & VPO_WANTED) {
489 m->oflags &= ~VPO_WANTED;
497 * clear the VPO_BUSY flag and wakeup anyone waiting for the
502 vm_page_wakeup(vm_page_t m)
505 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
506 KASSERT(m->oflags & VPO_BUSY, ("vm_page_wakeup: page not busy!!!"));
507 m->oflags &= ~VPO_BUSY;
512 vm_page_io_start(vm_page_t m)
515 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
520 vm_page_io_finish(vm_page_t m)
523 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
524 KASSERT(m->busy > 0, ("vm_page_io_finish: page %p is not busy", m));
531 * Keep page from being freed by the page daemon
532 * much of the same effect as wiring, except much lower
533 * overhead and should be used only for *very* temporary
534 * holding ("wiring").
537 vm_page_hold(vm_page_t mem)
540 vm_page_lock_assert(mem, MA_OWNED);
545 vm_page_unhold(vm_page_t mem)
548 vm_page_lock_assert(mem, MA_OWNED);
550 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
551 if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0)
552 vm_page_free_toq(mem);
556 * vm_page_unhold_pages:
558 * Unhold each of the pages that is referenced by the given array.
561 vm_page_unhold_pages(vm_page_t *ma, int count)
563 struct mtx *mtx, *new_mtx;
566 for (; count != 0; count--) {
568 * Avoid releasing and reacquiring the same page lock.
570 new_mtx = vm_page_lockptr(*ma);
571 if (mtx != new_mtx) {
585 PHYS_TO_VM_PAGE(vm_paddr_t pa)
589 #ifdef VM_PHYSSEG_SPARSE
590 m = vm_phys_paddr_to_vm_page(pa);
592 m = vm_phys_fictitious_to_vm_page(pa);
594 #elif defined(VM_PHYSSEG_DENSE)
598 if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
599 m = &vm_page_array[pi - first_page];
602 return (vm_phys_fictitious_to_vm_page(pa));
604 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
611 * Create a fictitious page with the specified physical address and
612 * memory attribute. The memory attribute is the only the machine-
613 * dependent aspect of a fictitious page that must be initialized.
616 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
620 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
621 vm_page_initfake(m, paddr, memattr);
626 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
629 if ((m->flags & PG_FICTITIOUS) != 0) {
631 * The page's memattr might have changed since the
632 * previous initialization. Update the pmap to the
637 m->phys_addr = paddr;
639 /* Fictitious pages don't use "segind". */
640 m->flags = PG_FICTITIOUS;
641 /* Fictitious pages don't use "order" or "pool". */
642 m->oflags = VPO_BUSY | VPO_UNMANAGED;
645 pmap_page_set_memattr(m, memattr);
651 * Release a fictitious page.
654 vm_page_putfake(vm_page_t m)
657 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
658 KASSERT((m->flags & PG_FICTITIOUS) != 0,
659 ("vm_page_putfake: bad page %p", m));
660 uma_zfree(fakepg_zone, m);
664 * vm_page_updatefake:
666 * Update the given fictitious page to the specified physical address and
670 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
673 KASSERT((m->flags & PG_FICTITIOUS) != 0,
674 ("vm_page_updatefake: bad page %p", m));
675 m->phys_addr = paddr;
676 pmap_page_set_memattr(m, memattr);
685 vm_page_free(vm_page_t m)
688 m->flags &= ~PG_ZERO;
695 * Free a page to the zerod-pages queue
698 vm_page_free_zero(vm_page_t m)
706 * Unbusy and handle the page queueing for a page from the VOP_GETPAGES()
707 * array which is not the request page.
710 vm_page_readahead_finish(vm_page_t m)
715 * Since the page is not the requested page, whether
716 * it should be activated or deactivated is not
717 * obvious. Empirical results have shown that
718 * deactivating the page is usually the best choice,
719 * unless the page is wanted by another thread.
721 if (m->oflags & VPO_WANTED) {
727 vm_page_deactivate(m);
733 * Free the completely invalid page. Such page state
734 * occurs due to the short read operation which did
735 * not covered our page at all, or in case when a read
747 * Sleep and release the page lock.
749 * The object containing the given page must be locked.
752 vm_page_sleep(vm_page_t m, const char *msg)
755 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
756 if (mtx_owned(vm_page_lockptr(m)))
760 * It's possible that while we sleep, the page will get
761 * unbusied and freed. If we are holding the object
762 * lock, we will assume we hold a reference to the object
763 * such that even if m->object changes, we can re-lock
766 m->oflags |= VPO_WANTED;
767 msleep(m, VM_OBJECT_MTX(m->object), PVM, msg, 0);
771 * vm_page_dirty_KBI: [ internal use only ]
773 * Set all bits in the page's dirty field.
775 * The object containing the specified page must be locked if the
776 * call is made from the machine-independent layer.
778 * See vm_page_clear_dirty_mask().
780 * This function should only be called by vm_page_dirty().
783 vm_page_dirty_KBI(vm_page_t m)
786 /* These assertions refer to this operation by its public name. */
787 KASSERT((m->flags & PG_CACHED) == 0,
788 ("vm_page_dirty: page in cache!"));
789 KASSERT(!VM_PAGE_IS_FREE(m),
790 ("vm_page_dirty: page is free!"));
791 KASSERT(m->valid == VM_PAGE_BITS_ALL,
792 ("vm_page_dirty: page is invalid!"));
793 m->dirty = VM_PAGE_BITS_ALL;
797 * vm_page_insert: [ internal use only ]
799 * Inserts the given mem entry into the object and object list.
801 * The pagetables are not updated but will presumably fault the page
802 * in if necessary, or if a kernel page the caller will at some point
803 * enter the page into the kernel's pmap. We are not allowed to sleep
804 * here so we *can't* do this anyway.
806 * The object must be locked.
809 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
813 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
814 if (m->object != NULL)
815 panic("vm_page_insert: page already inserted");
818 * Record the object/offset pair in this page
823 if (object->resident_page_count == 0) {
824 TAILQ_INSERT_TAIL(&object->memq, m, listq);
826 neighbor = vm_radix_lookup_ge(&object->rtree, pindex);
827 if (neighbor != NULL) {
828 KASSERT(pindex < neighbor->pindex,
829 ("vm_page_insert: offset %ju not minor than %ju",
830 (uintmax_t)pindex, (uintmax_t)neighbor->pindex));
831 TAILQ_INSERT_BEFORE(neighbor, m, listq);
833 TAILQ_INSERT_TAIL(&object->memq, m, listq);
835 vm_radix_insert(&object->rtree, pindex, m);
838 * Show that the object has one more resident page.
840 object->resident_page_count++;
843 * Hold the vnode until the last page is released.
845 if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
846 vhold(object->handle);
849 * Since we are inserting a new and possibly dirty page,
850 * update the object's OBJ_MIGHTBEDIRTY flag.
852 if (pmap_page_is_write_mapped(m))
853 vm_object_set_writeable_dirty(object);
859 * Removes the given mem entry from the object/offset-page
860 * table and the object page list, but do not invalidate/terminate
863 * The underlying pmap entry (if any) is NOT removed here.
865 * The object must be locked. The page must be locked if it is managed.
868 vm_page_remove(vm_page_t m)
872 if ((m->oflags & VPO_UNMANAGED) == 0)
873 vm_page_lock_assert(m, MA_OWNED);
874 if ((object = m->object) == NULL)
876 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
877 if (m->oflags & VPO_BUSY) {
878 m->oflags &= ~VPO_BUSY;
882 vm_radix_remove(&object->rtree, m->pindex);
883 TAILQ_REMOVE(&object->memq, m, listq);
886 * And show that the object has one fewer resident page.
888 object->resident_page_count--;
891 * The vnode may now be recycled.
893 if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
894 vdrop(object->handle);
902 * Returns the page associated with the object/offset
903 * pair specified; if none is found, NULL is returned.
905 * The object must be locked.
908 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
911 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
913 return (vm_radix_lookup(&object->rtree, pindex));
917 * vm_page_find_least:
919 * Returns the page associated with the object with least pindex
920 * greater than or equal to the parameter pindex, or NULL.
922 * The object must be locked.
925 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
928 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
929 return (vm_radix_lookup_ge(&object->rtree, pindex));
933 * Returns the given page's successor (by pindex) within the object if it is
934 * resident; if none is found, NULL is returned.
936 * The object must be locked.
939 vm_page_next(vm_page_t m)
943 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
944 if ((next = TAILQ_NEXT(m, listq)) != NULL &&
945 next->pindex != m->pindex + 1)
951 * Returns the given page's predecessor (by pindex) within the object if it is
952 * resident; if none is found, NULL is returned.
954 * The object must be locked.
957 vm_page_prev(vm_page_t m)
961 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
962 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
963 prev->pindex != m->pindex - 1)
971 * Move the given memory entry from its
972 * current object to the specified target object/offset.
974 * Note: swap associated with the page must be invalidated by the move. We
975 * have to do this for several reasons: (1) we aren't freeing the
976 * page, (2) we are dirtying the page, (3) the VM system is probably
977 * moving the page from object A to B, and will then later move
978 * the backing store from A to B and we can't have a conflict.
980 * Note: we *always* dirty the page. It is necessary both for the
981 * fact that we moved it, and because we may be invalidating
982 * swap. If the page is on the cache, we have to deactivate it
983 * or vm_page_dirty() will panic. Dirty pages are not allowed
986 * The objects must be locked. The page must be locked if it is managed.
989 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
993 vm_page_insert(m, new_object, new_pindex);
998 * Convert all of the given object's cached pages that have a
999 * pindex within the given range into free pages. If the value
1000 * zero is given for "end", then the range's upper bound is
1001 * infinity. If the given object is backed by a vnode and it
1002 * transitions from having one or more cached pages to none, the
1003 * vnode's hold count is reduced.
1005 * The object must be locked.
1008 vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1013 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1015 mtx_lock(&vm_page_queue_free_mtx);
1016 if (vm_object_cache_is_empty(object)) {
1017 mtx_unlock(&vm_page_queue_free_mtx);
1020 while ((m = vm_radix_lookup_ge(&object->cache, start)) != NULL) {
1021 if (end != 0 && m->pindex >= end)
1023 vm_radix_remove(&object->cache, m->pindex);
1026 /* Clear PG_CACHED and set PG_FREE. */
1027 m->flags ^= PG_CACHED | PG_FREE;
1028 KASSERT((m->flags & (PG_CACHED | PG_FREE)) == PG_FREE,
1029 ("vm_page_cache_free: page %p has inconsistent flags", m));
1030 cnt.v_cache_count--;
1033 empty = vm_object_cache_is_empty(object);
1034 mtx_unlock(&vm_page_queue_free_mtx);
1035 if (object->type == OBJT_VNODE && empty)
1036 vdrop(object->handle);
1040 * Returns the cached page that is associated with the given
1041 * object and offset. If, however, none exists, returns NULL.
1043 * The free page queue and object must be locked.
1045 static inline vm_page_t
1046 vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex)
1049 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1050 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1051 if (!vm_object_cache_is_empty(object))
1052 return (vm_radix_lookup(&object->cache, pindex));
1057 * Remove the given cached page from its containing object's
1058 * collection of cached pages.
1060 * The free page queue must be locked.
1063 vm_page_cache_remove(vm_page_t m)
1066 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1067 KASSERT((m->flags & PG_CACHED) != 0,
1068 ("vm_page_cache_remove: page %p is not cached", m));
1069 vm_radix_remove(&m->object->cache, m->pindex);
1071 cnt.v_cache_count--;
1075 * Transfer all of the cached pages with offset greater than or
1076 * equal to 'offidxstart' from the original object's cache to the
1077 * new object's cache. However, any cached pages with offset
1078 * greater than or equal to the new object's size are kept in the
1079 * original object. Initially, the new object's cache must be
1080 * empty. Offset 'offidxstart' in the original object must
1081 * correspond to offset zero in the new object.
1083 * The new object and original object must be locked.
1086 vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart,
1087 vm_object_t new_object)
1092 * Insertion into an object's collection of cached pages
1093 * requires the object to be locked. In contrast, removal does
1096 VM_OBJECT_LOCK_ASSERT(new_object, MA_OWNED);
1097 VM_OBJECT_LOCK_ASSERT(orig_object, MA_OWNED);
1098 KASSERT(vm_object_cache_is_empty(new_object),
1099 ("vm_page_cache_transfer: object %p has cached pages",
1101 mtx_lock(&vm_page_queue_free_mtx);
1102 while ((m = vm_radix_lookup_ge(&orig_object->cache,
1103 offidxstart)) != NULL) {
1104 if ((m->pindex - offidxstart) >= new_object->size)
1106 vm_radix_remove(&orig_object->cache, m->pindex);
1107 vm_radix_insert(&new_object->cache, m->pindex - offidxstart, m);
1108 m->object = new_object;
1109 m->pindex -= offidxstart;
1111 mtx_unlock(&vm_page_queue_free_mtx);
1115 * Returns TRUE if a cached page is associated with the given object and
1116 * offset, and FALSE otherwise.
1118 * The object must be locked.
1121 vm_page_is_cached(vm_object_t object, vm_pindex_t pindex)
1126 * Insertion into an object's collection of cached pages requires the
1127 * object to be locked. Therefore, if the object is locked and the
1128 * object's collection is empty, there is no need to acquire the free
1129 * page queues lock in order to prove that the specified page doesn't
1132 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1133 if (vm_object_cache_is_empty(object))
1135 mtx_lock(&vm_page_queue_free_mtx);
1136 m = vm_page_cache_lookup(object, pindex);
1137 mtx_unlock(&vm_page_queue_free_mtx);
1144 * Allocate and return a page that is associated with the specified
1145 * object and offset pair. By default, this page has the flag VPO_BUSY
1148 * The caller must always specify an allocation class.
1150 * allocation classes:
1151 * VM_ALLOC_NORMAL normal process request
1152 * VM_ALLOC_SYSTEM system *really* needs a page
1153 * VM_ALLOC_INTERRUPT interrupt time request
1155 * optional allocation flags:
1156 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1157 * intends to allocate
1158 * VM_ALLOC_IFCACHED return page only if it is cached
1159 * VM_ALLOC_IFNOTCACHED return NULL, do not reactivate if the page
1161 * VM_ALLOC_NOBUSY do not set the flag VPO_BUSY on the page
1162 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1163 * VM_ALLOC_NOOBJ page is not associated with an object and
1164 * should not have the flag VPO_BUSY set
1165 * VM_ALLOC_WIRED wire the allocated page
1166 * VM_ALLOC_ZERO prefer a zeroed page
1168 * This routine may not sleep.
1171 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1173 struct vnode *vp = NULL;
1174 vm_object_t m_object;
1176 int flags, req_class;
1178 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0),
1179 ("vm_page_alloc: inconsistent object/req"));
1181 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1183 req_class = req & VM_ALLOC_CLASS_MASK;
1186 * The page daemon is allowed to dig deeper into the free page list.
1188 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1189 req_class = VM_ALLOC_SYSTEM;
1191 mtx_lock(&vm_page_queue_free_mtx);
1192 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved ||
1193 (req_class == VM_ALLOC_SYSTEM &&
1194 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) ||
1195 (req_class == VM_ALLOC_INTERRUPT &&
1196 cnt.v_free_count + cnt.v_cache_count > 0)) {
1198 * Allocate from the free queue if the number of free pages
1199 * exceeds the minimum for the request class.
1201 if (object != NULL &&
1202 (m = vm_page_cache_lookup(object, pindex)) != NULL) {
1203 if ((req & VM_ALLOC_IFNOTCACHED) != 0) {
1204 mtx_unlock(&vm_page_queue_free_mtx);
1207 if (vm_phys_unfree_page(m))
1208 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0);
1209 #if VM_NRESERVLEVEL > 0
1210 else if (!vm_reserv_reactivate_page(m))
1214 panic("vm_page_alloc: cache page %p is missing"
1215 " from the free queue", m);
1216 } else if ((req & VM_ALLOC_IFCACHED) != 0) {
1217 mtx_unlock(&vm_page_queue_free_mtx);
1219 #if VM_NRESERVLEVEL > 0
1220 } else if (object == NULL || (object->flags & (OBJ_COLORED |
1221 OBJ_FICTITIOUS)) != OBJ_COLORED ||
1222 (m = vm_reserv_alloc_page(object, pindex)) == NULL) {
1226 m = vm_phys_alloc_pages(object != NULL ?
1227 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1228 #if VM_NRESERVLEVEL > 0
1229 if (m == NULL && vm_reserv_reclaim_inactive()) {
1230 m = vm_phys_alloc_pages(object != NULL ?
1231 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1238 * Not allocatable, give up.
1240 mtx_unlock(&vm_page_queue_free_mtx);
1241 atomic_add_int(&vm_pageout_deficit,
1242 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1243 pagedaemon_wakeup();
1248 * At this point we had better have found a good page.
1250 KASSERT(m != NULL, ("vm_page_alloc: missing page"));
1251 KASSERT(m->queue == PQ_NONE,
1252 ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue));
1253 KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m));
1254 KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m));
1255 KASSERT(m->busy == 0, ("vm_page_alloc: page %p is busy", m));
1256 KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m));
1257 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1258 ("vm_page_alloc: page %p has unexpected memattr %d", m,
1259 pmap_page_get_memattr(m)));
1260 if ((m->flags & PG_CACHED) != 0) {
1261 KASSERT((m->flags & PG_ZERO) == 0,
1262 ("vm_page_alloc: cached page %p is PG_ZERO", m));
1263 KASSERT(m->valid != 0,
1264 ("vm_page_alloc: cached page %p is invalid", m));
1265 if (m->object == object && m->pindex == pindex)
1266 cnt.v_reactivated++;
1269 m_object = m->object;
1270 vm_page_cache_remove(m);
1271 if (m_object->type == OBJT_VNODE &&
1272 vm_object_cache_is_empty(m_object))
1273 vp = m_object->handle;
1275 KASSERT(VM_PAGE_IS_FREE(m),
1276 ("vm_page_alloc: page %p is not free", m));
1277 KASSERT(m->valid == 0,
1278 ("vm_page_alloc: free page %p is valid", m));
1283 * Only the PG_ZERO flag is inherited. The PG_CACHED or PG_FREE flag
1284 * must be cleared before the free page queues lock is released.
1287 if (m->flags & PG_ZERO) {
1288 vm_page_zero_count--;
1289 if (req & VM_ALLOC_ZERO)
1292 if (req & VM_ALLOC_NODUMP)
1295 mtx_unlock(&vm_page_queue_free_mtx);
1297 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1299 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ)) == 0)
1300 m->oflags |= VPO_BUSY;
1301 if (req & VM_ALLOC_WIRED) {
1303 * The page lock is not required for wiring a page until that
1304 * page is inserted into the object.
1306 atomic_add_int(&cnt.v_wire_count, 1);
1311 if (object != NULL) {
1312 /* Ignore device objects; the pager sets "memattr" for them. */
1313 if (object->memattr != VM_MEMATTR_DEFAULT &&
1314 (object->flags & OBJ_FICTITIOUS) == 0)
1315 pmap_page_set_memattr(m, object->memattr);
1316 vm_page_insert(m, object, pindex);
1321 * The following call to vdrop() must come after the above call
1322 * to vm_page_insert() in case both affect the same object and
1323 * vnode. Otherwise, the affected vnode's hold count could
1324 * temporarily become zero.
1330 * Don't wakeup too often - wakeup the pageout daemon when
1331 * we would be nearly out of memory.
1333 if (vm_paging_needed())
1334 pagedaemon_wakeup();
1340 * vm_page_alloc_contig:
1342 * Allocate a contiguous set of physical pages of the given size "npages"
1343 * from the free lists. All of the physical pages must be at or above
1344 * the given physical address "low" and below the given physical address
1345 * "high". The given value "alignment" determines the alignment of the
1346 * first physical page in the set. If the given value "boundary" is
1347 * non-zero, then the set of physical pages cannot cross any physical
1348 * address boundary that is a multiple of that value. Both "alignment"
1349 * and "boundary" must be a power of two.
1351 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1352 * then the memory attribute setting for the physical pages is configured
1353 * to the object's memory attribute setting. Otherwise, the memory
1354 * attribute setting for the physical pages is configured to "memattr",
1355 * overriding the object's memory attribute setting. However, if the
1356 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1357 * memory attribute setting for the physical pages cannot be configured
1358 * to VM_MEMATTR_DEFAULT.
1360 * The caller must always specify an allocation class.
1362 * allocation classes:
1363 * VM_ALLOC_NORMAL normal process request
1364 * VM_ALLOC_SYSTEM system *really* needs a page
1365 * VM_ALLOC_INTERRUPT interrupt time request
1367 * optional allocation flags:
1368 * VM_ALLOC_NOBUSY do not set the flag VPO_BUSY on the page
1369 * VM_ALLOC_NOOBJ page is not associated with an object and
1370 * should not have the flag VPO_BUSY set
1371 * VM_ALLOC_WIRED wire the allocated page
1372 * VM_ALLOC_ZERO prefer a zeroed page
1374 * This routine may not sleep.
1377 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1378 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1379 vm_paddr_t boundary, vm_memattr_t memattr)
1382 vm_page_t deferred_vdrop_list, m, m_ret;
1383 u_int flags, oflags;
1386 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0),
1387 ("vm_page_alloc_contig: inconsistent object/req"));
1388 if (object != NULL) {
1389 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1390 KASSERT(object->type == OBJT_PHYS,
1391 ("vm_page_alloc_contig: object %p isn't OBJT_PHYS",
1394 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
1395 req_class = req & VM_ALLOC_CLASS_MASK;
1398 * The page daemon is allowed to dig deeper into the free page list.
1400 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1401 req_class = VM_ALLOC_SYSTEM;
1403 deferred_vdrop_list = NULL;
1404 mtx_lock(&vm_page_queue_free_mtx);
1405 if (cnt.v_free_count + cnt.v_cache_count >= npages +
1406 cnt.v_free_reserved || (req_class == VM_ALLOC_SYSTEM &&
1407 cnt.v_free_count + cnt.v_cache_count >= npages +
1408 cnt.v_interrupt_free_min) || (req_class == VM_ALLOC_INTERRUPT &&
1409 cnt.v_free_count + cnt.v_cache_count >= npages)) {
1410 #if VM_NRESERVLEVEL > 0
1412 if (object == NULL || (object->flags & OBJ_COLORED) == 0 ||
1413 (m_ret = vm_reserv_alloc_contig(object, pindex, npages,
1414 low, high, alignment, boundary)) == NULL)
1416 m_ret = vm_phys_alloc_contig(npages, low, high,
1417 alignment, boundary);
1419 mtx_unlock(&vm_page_queue_free_mtx);
1420 atomic_add_int(&vm_pageout_deficit, npages);
1421 pagedaemon_wakeup();
1425 for (m = m_ret; m < &m_ret[npages]; m++) {
1426 drop = vm_page_alloc_init(m);
1429 * Enqueue the vnode for deferred vdrop().
1431 * Once the pages are removed from the free
1432 * page list, "pageq" can be safely abused to
1433 * construct a short-lived list of vnodes.
1435 m->pageq.tqe_prev = (void *)drop;
1436 m->pageq.tqe_next = deferred_vdrop_list;
1437 deferred_vdrop_list = m;
1441 #if VM_NRESERVLEVEL > 0
1442 if (vm_reserv_reclaim_contig(npages, low, high, alignment,
1447 mtx_unlock(&vm_page_queue_free_mtx);
1452 * Initialize the pages. Only the PG_ZERO flag is inherited.
1455 if ((req & VM_ALLOC_ZERO) != 0)
1457 if ((req & VM_ALLOC_NODUMP) != 0)
1459 if ((req & VM_ALLOC_WIRED) != 0)
1460 atomic_add_int(&cnt.v_wire_count, npages);
1461 oflags = VPO_UNMANAGED;
1462 if (object != NULL) {
1463 if ((req & VM_ALLOC_NOBUSY) == 0)
1465 if (object->memattr != VM_MEMATTR_DEFAULT &&
1466 memattr == VM_MEMATTR_DEFAULT)
1467 memattr = object->memattr;
1469 for (m = m_ret; m < &m_ret[npages]; m++) {
1471 m->flags = (m->flags | PG_NODUMP) & flags;
1472 if ((req & VM_ALLOC_WIRED) != 0)
1474 /* Unmanaged pages don't use "act_count". */
1476 if (memattr != VM_MEMATTR_DEFAULT)
1477 pmap_page_set_memattr(m, memattr);
1479 vm_page_insert(m, object, pindex);
1484 while (deferred_vdrop_list != NULL) {
1485 vdrop((struct vnode *)deferred_vdrop_list->pageq.tqe_prev);
1486 deferred_vdrop_list = deferred_vdrop_list->pageq.tqe_next;
1488 if (vm_paging_needed())
1489 pagedaemon_wakeup();
1494 * Initialize a page that has been freshly dequeued from a freelist.
1495 * The caller has to drop the vnode returned, if it is not NULL.
1497 * This function may only be used to initialize unmanaged pages.
1499 * To be called with vm_page_queue_free_mtx held.
1501 static struct vnode *
1502 vm_page_alloc_init(vm_page_t m)
1505 vm_object_t m_object;
1507 KASSERT(m->queue == PQ_NONE,
1508 ("vm_page_alloc_init: page %p has unexpected queue %d",
1510 KASSERT(m->wire_count == 0,
1511 ("vm_page_alloc_init: page %p is wired", m));
1512 KASSERT(m->hold_count == 0,
1513 ("vm_page_alloc_init: page %p is held", m));
1514 KASSERT(m->busy == 0,
1515 ("vm_page_alloc_init: page %p is busy", m));
1516 KASSERT(m->dirty == 0,
1517 ("vm_page_alloc_init: page %p is dirty", m));
1518 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1519 ("vm_page_alloc_init: page %p has unexpected memattr %d",
1520 m, pmap_page_get_memattr(m)));
1521 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1523 if ((m->flags & PG_CACHED) != 0) {
1524 KASSERT((m->flags & PG_ZERO) == 0,
1525 ("vm_page_alloc_init: cached page %p is PG_ZERO", m));
1527 m_object = m->object;
1528 vm_page_cache_remove(m);
1529 if (m_object->type == OBJT_VNODE &&
1530 vm_object_cache_is_empty(m_object))
1531 drop = m_object->handle;
1533 KASSERT(VM_PAGE_IS_FREE(m),
1534 ("vm_page_alloc_init: page %p is not free", m));
1535 KASSERT(m->valid == 0,
1536 ("vm_page_alloc_init: free page %p is valid", m));
1538 if ((m->flags & PG_ZERO) != 0)
1539 vm_page_zero_count--;
1541 /* Don't clear the PG_ZERO flag; we'll need it later. */
1542 m->flags &= PG_ZERO;
1547 * vm_page_alloc_freelist:
1549 * Allocate a physical page from the specified free page list.
1551 * The caller must always specify an allocation class.
1553 * allocation classes:
1554 * VM_ALLOC_NORMAL normal process request
1555 * VM_ALLOC_SYSTEM system *really* needs a page
1556 * VM_ALLOC_INTERRUPT interrupt time request
1558 * optional allocation flags:
1559 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1560 * intends to allocate
1561 * VM_ALLOC_WIRED wire the allocated page
1562 * VM_ALLOC_ZERO prefer a zeroed page
1564 * This routine may not sleep.
1567 vm_page_alloc_freelist(int flind, int req)
1574 req_class = req & VM_ALLOC_CLASS_MASK;
1577 * The page daemon is allowed to dig deeper into the free page list.
1579 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1580 req_class = VM_ALLOC_SYSTEM;
1583 * Do not allocate reserved pages unless the req has asked for it.
1585 mtx_lock(&vm_page_queue_free_mtx);
1586 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved ||
1587 (req_class == VM_ALLOC_SYSTEM &&
1588 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) ||
1589 (req_class == VM_ALLOC_INTERRUPT &&
1590 cnt.v_free_count + cnt.v_cache_count > 0))
1591 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0);
1593 mtx_unlock(&vm_page_queue_free_mtx);
1594 atomic_add_int(&vm_pageout_deficit,
1595 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1596 pagedaemon_wakeup();
1600 mtx_unlock(&vm_page_queue_free_mtx);
1603 drop = vm_page_alloc_init(m);
1604 mtx_unlock(&vm_page_queue_free_mtx);
1607 * Initialize the page. Only the PG_ZERO flag is inherited.
1611 if ((req & VM_ALLOC_ZERO) != 0)
1614 if ((req & VM_ALLOC_WIRED) != 0) {
1616 * The page lock is not required for wiring a page that does
1617 * not belong to an object.
1619 atomic_add_int(&cnt.v_wire_count, 1);
1622 /* Unmanaged pages don't use "act_count". */
1623 m->oflags = VPO_UNMANAGED;
1626 if (vm_paging_needed())
1627 pagedaemon_wakeup();
1632 * vm_wait: (also see VM_WAIT macro)
1634 * Sleep until free pages are available for allocation.
1635 * - Called in various places before memory allocations.
1641 mtx_lock(&vm_page_queue_free_mtx);
1642 if (curproc == pageproc) {
1643 vm_pageout_pages_needed = 1;
1644 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx,
1645 PDROP | PSWP, "VMWait", 0);
1647 if (!vm_pages_needed) {
1648 vm_pages_needed = 1;
1649 wakeup(&vm_pages_needed);
1651 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM,
1657 * vm_waitpfault: (also see VM_WAITPFAULT macro)
1659 * Sleep until free pages are available for allocation.
1660 * - Called only in vm_fault so that processes page faulting
1661 * can be easily tracked.
1662 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
1663 * processes will be able to grab memory first. Do not change
1664 * this balance without careful testing first.
1670 mtx_lock(&vm_page_queue_free_mtx);
1671 if (!vm_pages_needed) {
1672 vm_pages_needed = 1;
1673 wakeup(&vm_pages_needed);
1675 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER,
1682 * Remove the given page from its current page queue.
1684 * The page must be locked.
1687 vm_page_dequeue(vm_page_t m)
1689 struct vm_pagequeue *pq;
1691 vm_page_lock_assert(m, MA_OWNED);
1692 KASSERT(m->queue != PQ_NONE,
1693 ("vm_page_dequeue: page %p is not queued", m));
1694 pq = &vm_pagequeues[m->queue];
1695 vm_pagequeue_lock(pq);
1697 TAILQ_REMOVE(&pq->pq_pl, m, pageq);
1699 vm_pagequeue_unlock(pq);
1703 * vm_page_dequeue_locked:
1705 * Remove the given page from its current page queue.
1707 * The page and page queue must be locked.
1710 vm_page_dequeue_locked(vm_page_t m)
1712 struct vm_pagequeue *pq;
1714 vm_page_lock_assert(m, MA_OWNED);
1715 pq = &vm_pagequeues[m->queue];
1716 vm_pagequeue_assert_locked(pq);
1718 TAILQ_REMOVE(&pq->pq_pl, m, pageq);
1725 * Add the given page to the specified page queue.
1727 * The page must be locked.
1730 vm_page_enqueue(int queue, vm_page_t m)
1732 struct vm_pagequeue *pq;
1734 vm_page_lock_assert(m, MA_OWNED);
1735 pq = &vm_pagequeues[queue];
1736 vm_pagequeue_lock(pq);
1738 TAILQ_INSERT_TAIL(&pq->pq_pl, m, pageq);
1740 vm_pagequeue_unlock(pq);
1746 * Move the given page to the tail of its current page queue.
1748 * The page must be locked.
1751 vm_page_requeue(vm_page_t m)
1753 struct vm_pagequeue *pq;
1755 vm_page_lock_assert(m, MA_OWNED);
1756 KASSERT(m->queue != PQ_NONE,
1757 ("vm_page_requeue: page %p is not queued", m));
1758 pq = &vm_pagequeues[m->queue];
1759 vm_pagequeue_lock(pq);
1760 TAILQ_REMOVE(&pq->pq_pl, m, pageq);
1761 TAILQ_INSERT_TAIL(&pq->pq_pl, m, pageq);
1762 vm_pagequeue_unlock(pq);
1766 * vm_page_requeue_locked:
1768 * Move the given page to the tail of its current page queue.
1770 * The page queue must be locked.
1773 vm_page_requeue_locked(vm_page_t m)
1775 struct vm_pagequeue *pq;
1777 KASSERT(m->queue != PQ_NONE,
1778 ("vm_page_requeue_locked: page %p is not queued", m));
1779 pq = &vm_pagequeues[m->queue];
1780 vm_pagequeue_assert_locked(pq);
1781 TAILQ_REMOVE(&pq->pq_pl, m, pageq);
1782 TAILQ_INSERT_TAIL(&pq->pq_pl, m, pageq);
1788 * Put the specified page on the active list (if appropriate).
1789 * Ensure that act_count is at least ACT_INIT but do not otherwise
1792 * The page must be locked.
1795 vm_page_activate(vm_page_t m)
1799 vm_page_lock_assert(m, MA_OWNED);
1800 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1801 if ((queue = m->queue) != PQ_ACTIVE) {
1802 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
1803 if (m->act_count < ACT_INIT)
1804 m->act_count = ACT_INIT;
1805 if (queue != PQ_NONE)
1807 vm_page_enqueue(PQ_ACTIVE, m);
1809 KASSERT(queue == PQ_NONE,
1810 ("vm_page_activate: wired page %p is queued", m));
1812 if (m->act_count < ACT_INIT)
1813 m->act_count = ACT_INIT;
1818 * vm_page_free_wakeup:
1820 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
1821 * routine is called when a page has been added to the cache or free
1824 * The page queues must be locked.
1827 vm_page_free_wakeup(void)
1830 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1832 * if pageout daemon needs pages, then tell it that there are
1835 if (vm_pageout_pages_needed &&
1836 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) {
1837 wakeup(&vm_pageout_pages_needed);
1838 vm_pageout_pages_needed = 0;
1841 * wakeup processes that are waiting on memory if we hit a
1842 * high water mark. And wakeup scheduler process if we have
1843 * lots of memory. this process will swapin processes.
1845 if (vm_pages_needed && !vm_page_count_min()) {
1846 vm_pages_needed = 0;
1847 wakeup(&cnt.v_free_count);
1854 * Returns the given page to the free list,
1855 * disassociating it with any VM object.
1857 * The object must be locked. The page must be locked if it is managed.
1860 vm_page_free_toq(vm_page_t m)
1863 if ((m->oflags & VPO_UNMANAGED) == 0) {
1864 vm_page_lock_assert(m, MA_OWNED);
1865 KASSERT(!pmap_page_is_mapped(m),
1866 ("vm_page_free_toq: freeing mapped page %p", m));
1868 KASSERT(m->queue == PQ_NONE,
1869 ("vm_page_free_toq: unmanaged page %p is queued", m));
1870 PCPU_INC(cnt.v_tfree);
1872 if (VM_PAGE_IS_FREE(m))
1873 panic("vm_page_free: freeing free page %p", m);
1874 else if (m->busy != 0)
1875 panic("vm_page_free: freeing busy page %p", m);
1878 * Unqueue, then remove page. Note that we cannot destroy
1879 * the page here because we do not want to call the pager's
1880 * callback routine until after we've put the page on the
1881 * appropriate free queue.
1887 * If fictitious remove object association and
1888 * return, otherwise delay object association removal.
1890 if ((m->flags & PG_FICTITIOUS) != 0) {
1897 if (m->wire_count != 0)
1898 panic("vm_page_free: freeing wired page %p", m);
1899 if (m->hold_count != 0) {
1900 m->flags &= ~PG_ZERO;
1901 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
1902 ("vm_page_free: freeing PG_UNHOLDFREE page %p", m));
1903 m->flags |= PG_UNHOLDFREE;
1906 * Restore the default memory attribute to the page.
1908 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
1909 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
1912 * Insert the page into the physical memory allocator's
1913 * cache/free page queues.
1915 mtx_lock(&vm_page_queue_free_mtx);
1916 m->flags |= PG_FREE;
1918 #if VM_NRESERVLEVEL > 0
1919 if (!vm_reserv_free_page(m))
1923 vm_phys_free_pages(m, 0);
1924 if ((m->flags & PG_ZERO) != 0)
1925 ++vm_page_zero_count;
1927 vm_page_zero_idle_wakeup();
1928 vm_page_free_wakeup();
1929 mtx_unlock(&vm_page_queue_free_mtx);
1936 * Mark this page as wired down by yet
1937 * another map, removing it from paging queues
1940 * If the page is fictitious, then its wire count must remain one.
1942 * The page must be locked.
1945 vm_page_wire(vm_page_t m)
1949 * Only bump the wire statistics if the page is not already wired,
1950 * and only unqueue the page if it is on some queue (if it is unmanaged
1951 * it is already off the queues).
1953 vm_page_lock_assert(m, MA_OWNED);
1954 if ((m->flags & PG_FICTITIOUS) != 0) {
1955 KASSERT(m->wire_count == 1,
1956 ("vm_page_wire: fictitious page %p's wire count isn't one",
1960 if (m->wire_count == 0) {
1961 KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
1962 m->queue == PQ_NONE,
1963 ("vm_page_wire: unmanaged page %p is queued", m));
1965 atomic_add_int(&cnt.v_wire_count, 1);
1968 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
1974 * Release one wiring of the specified page, potentially enabling it to be
1975 * paged again. If paging is enabled, then the value of the parameter
1976 * "activate" determines to which queue the page is added. If "activate" is
1977 * non-zero, then the page is added to the active queue. Otherwise, it is
1978 * added to the inactive queue.
1980 * However, unless the page belongs to an object, it is not enqueued because
1981 * it cannot be paged out.
1983 * If a page is fictitious, then its wire count must alway be one.
1985 * A managed page must be locked.
1988 vm_page_unwire(vm_page_t m, int activate)
1991 if ((m->oflags & VPO_UNMANAGED) == 0)
1992 vm_page_lock_assert(m, MA_OWNED);
1993 if ((m->flags & PG_FICTITIOUS) != 0) {
1994 KASSERT(m->wire_count == 1,
1995 ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
1998 if (m->wire_count > 0) {
2000 if (m->wire_count == 0) {
2001 atomic_subtract_int(&cnt.v_wire_count, 1);
2002 if ((m->oflags & VPO_UNMANAGED) != 0 ||
2006 m->flags &= ~PG_WINATCFLS;
2007 vm_page_enqueue(activate ? PQ_ACTIVE : PQ_INACTIVE, m);
2010 panic("vm_page_unwire: page %p's wire count is zero", m);
2014 * Move the specified page to the inactive queue.
2016 * Many pages placed on the inactive queue should actually go
2017 * into the cache, but it is difficult to figure out which. What
2018 * we do instead, if the inactive target is well met, is to put
2019 * clean pages at the head of the inactive queue instead of the tail.
2020 * This will cause them to be moved to the cache more quickly and
2021 * if not actively re-referenced, reclaimed more quickly. If we just
2022 * stick these pages at the end of the inactive queue, heavy filesystem
2023 * meta-data accesses can cause an unnecessary paging load on memory bound
2024 * processes. This optimization causes one-time-use metadata to be
2025 * reused more quickly.
2027 * Normally athead is 0 resulting in LRU operation. athead is set
2028 * to 1 if we want this page to be 'as if it were placed in the cache',
2029 * except without unmapping it from the process address space.
2031 * The page must be locked.
2034 _vm_page_deactivate(vm_page_t m, int athead)
2036 struct vm_pagequeue *pq;
2039 vm_page_lock_assert(m, MA_OWNED);
2042 * Ignore if already inactive.
2044 if ((queue = m->queue) == PQ_INACTIVE)
2046 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2047 if (queue != PQ_NONE)
2049 m->flags &= ~PG_WINATCFLS;
2050 pq = &vm_pagequeues[PQ_INACTIVE];
2051 vm_pagequeue_lock(pq);
2052 m->queue = PQ_INACTIVE;
2054 TAILQ_INSERT_HEAD(&pq->pq_pl, m, pageq);
2056 TAILQ_INSERT_TAIL(&pq->pq_pl, m, pageq);
2057 cnt.v_inactive_count++;
2058 vm_pagequeue_unlock(pq);
2063 * Move the specified page to the inactive queue.
2065 * The page must be locked.
2068 vm_page_deactivate(vm_page_t m)
2071 _vm_page_deactivate(m, 0);
2075 * vm_page_try_to_cache:
2077 * Returns 0 on failure, 1 on success
2080 vm_page_try_to_cache(vm_page_t m)
2083 vm_page_lock_assert(m, MA_OWNED);
2084 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2085 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
2086 (m->oflags & (VPO_BUSY | VPO_UNMANAGED)) != 0)
2096 * vm_page_try_to_free()
2098 * Attempt to free the page. If we cannot free it, we do nothing.
2099 * 1 is returned on success, 0 on failure.
2102 vm_page_try_to_free(vm_page_t m)
2105 vm_page_lock_assert(m, MA_OWNED);
2106 if (m->object != NULL)
2107 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2108 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
2109 (m->oflags & (VPO_BUSY | VPO_UNMANAGED)) != 0)
2121 * Put the specified page onto the page cache queue (if appropriate).
2123 * The object and page must be locked.
2126 vm_page_cache(vm_page_t m)
2129 int old_empty_cache;
2131 vm_page_lock_assert(m, MA_OWNED);
2133 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2134 if ((m->oflags & (VPO_UNMANAGED | VPO_BUSY)) || m->busy ||
2135 m->hold_count || m->wire_count)
2136 panic("vm_page_cache: attempting to cache busy page");
2137 KASSERT(!pmap_page_is_mapped(m),
2138 ("vm_page_cache: page %p is mapped", m));
2139 KASSERT(m->dirty == 0, ("vm_page_cache: page %p is dirty", m));
2140 if (m->valid == 0 || object->type == OBJT_DEFAULT ||
2141 (object->type == OBJT_SWAP &&
2142 !vm_pager_has_page(object, m->pindex, NULL, NULL))) {
2144 * Hypothesis: A cache-elgible page belonging to a
2145 * default object or swap object but without a backing
2146 * store must be zero filled.
2151 KASSERT((m->flags & PG_CACHED) == 0,
2152 ("vm_page_cache: page %p is already cached", m));
2153 PCPU_INC(cnt.v_tcached);
2156 * Remove the page from the paging queues.
2160 vm_radix_remove(&object->rtree, m->pindex);
2161 TAILQ_REMOVE(&object->memq, m, listq);
2162 object->resident_page_count--;
2165 * Restore the default memory attribute to the page.
2167 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2168 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2171 * Insert the page into the object's collection of cached pages
2172 * and the physical memory allocator's cache/free page queues.
2174 m->flags &= ~PG_ZERO;
2175 mtx_lock(&vm_page_queue_free_mtx);
2176 m->flags |= PG_CACHED;
2177 old_empty_cache = vm_object_cache_is_empty(object);
2178 cnt.v_cache_count++;
2179 vm_radix_insert(&object->cache, m->pindex, m);
2180 #if VM_NRESERVLEVEL > 0
2181 if (!vm_reserv_free_page(m)) {
2185 vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0);
2186 vm_phys_free_pages(m, 0);
2188 vm_page_free_wakeup();
2189 mtx_unlock(&vm_page_queue_free_mtx);
2192 * Increment the vnode's hold count if this is the object's only
2193 * cached page. Decrement the vnode's hold count if this was
2194 * the object's only resident page.
2196 if (object->type == OBJT_VNODE) {
2197 if (old_empty_cache != 0 && object->resident_page_count != 0)
2198 vhold(object->handle);
2199 else if (old_empty_cache == 0 &&
2200 object->resident_page_count == 0)
2201 vdrop(object->handle);
2208 * Cache, deactivate, or do nothing as appropriate. This routine
2209 * is typically used by madvise() MADV_DONTNEED.
2211 * Generally speaking we want to move the page into the cache so
2212 * it gets reused quickly. However, this can result in a silly syndrome
2213 * due to the page recycling too quickly. Small objects will not be
2214 * fully cached. On the otherhand, if we move the page to the inactive
2215 * queue we wind up with a problem whereby very large objects
2216 * unnecessarily blow away our inactive and cache queues.
2218 * The solution is to move the pages based on a fixed weighting. We
2219 * either leave them alone, deactivate them, or move them to the cache,
2220 * where moving them to the cache has the highest weighting.
2221 * By forcing some pages into other queues we eventually force the
2222 * system to balance the queues, potentially recovering other unrelated
2223 * space from active. The idea is to not force this to happen too
2226 * The object and page must be locked.
2229 vm_page_dontneed(vm_page_t m)
2234 vm_page_lock_assert(m, MA_OWNED);
2235 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2236 dnw = PCPU_GET(dnweight);
2240 * Occasionally leave the page alone.
2242 if ((dnw & 0x01F0) == 0 || m->queue == PQ_INACTIVE) {
2243 if (m->act_count >= ACT_INIT)
2249 * Clear any references to the page. Otherwise, the page daemon will
2250 * immediately reactivate the page.
2252 * Perform the pmap_clear_reference() first. Otherwise, a concurrent
2253 * pmap operation, such as pmap_remove(), could clear a reference in
2254 * the pmap and set PGA_REFERENCED on the page before the
2255 * pmap_clear_reference() had completed. Consequently, the page would
2256 * appear referenced based upon an old reference that occurred before
2257 * this function ran.
2259 pmap_clear_reference(m);
2260 vm_page_aflag_clear(m, PGA_REFERENCED);
2262 if (m->dirty == 0 && pmap_is_modified(m))
2265 if (m->dirty || (dnw & 0x0070) == 0) {
2267 * Deactivate the page 3 times out of 32.
2272 * Cache the page 28 times out of every 32. Note that
2273 * the page is deactivated instead of cached, but placed
2274 * at the head of the queue instead of the tail.
2278 _vm_page_deactivate(m, head);
2282 * Grab a page, waiting until we are waken up due to the page
2283 * changing state. We keep on waiting, if the page continues
2284 * to be in the object. If the page doesn't exist, first allocate it
2285 * and then conditionally zero it.
2287 * The caller must always specify the VM_ALLOC_RETRY flag. This is intended
2288 * to facilitate its eventual removal.
2290 * This routine may sleep.
2292 * The object must be locked on entry. The lock will, however, be released
2293 * and reacquired if the routine sleeps.
2296 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2300 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2301 KASSERT((allocflags & VM_ALLOC_RETRY) != 0,
2302 ("vm_page_grab: VM_ALLOC_RETRY is required"));
2304 if ((m = vm_page_lookup(object, pindex)) != NULL) {
2305 if ((m->oflags & VPO_BUSY) != 0 ||
2306 ((allocflags & VM_ALLOC_IGN_SBUSY) == 0 && m->busy != 0)) {
2308 * Reference the page before unlocking and
2309 * sleeping so that the page daemon is less
2310 * likely to reclaim it.
2312 vm_page_aflag_set(m, PGA_REFERENCED);
2313 vm_page_sleep(m, "pgrbwt");
2316 if ((allocflags & VM_ALLOC_WIRED) != 0) {
2321 if ((allocflags & VM_ALLOC_NOBUSY) == 0)
2326 m = vm_page_alloc(object, pindex, allocflags & ~(VM_ALLOC_RETRY |
2327 VM_ALLOC_IGN_SBUSY));
2329 VM_OBJECT_UNLOCK(object);
2331 VM_OBJECT_LOCK(object);
2333 } else if (m->valid != 0)
2335 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
2341 * Mapping function for valid or dirty bits in a page.
2343 * Inputs are required to range within a page.
2346 vm_page_bits(int base, int size)
2352 base + size <= PAGE_SIZE,
2353 ("vm_page_bits: illegal base/size %d/%d", base, size)
2356 if (size == 0) /* handle degenerate case */
2359 first_bit = base >> DEV_BSHIFT;
2360 last_bit = (base + size - 1) >> DEV_BSHIFT;
2362 return (((vm_page_bits_t)2 << last_bit) -
2363 ((vm_page_bits_t)1 << first_bit));
2367 * vm_page_set_valid_range:
2369 * Sets portions of a page valid. The arguments are expected
2370 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2371 * of any partial chunks touched by the range. The invalid portion of
2372 * such chunks will be zeroed.
2374 * (base + size) must be less then or equal to PAGE_SIZE.
2377 vm_page_set_valid_range(vm_page_t m, int base, int size)
2381 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2382 if (size == 0) /* handle degenerate case */
2386 * If the base is not DEV_BSIZE aligned and the valid
2387 * bit is clear, we have to zero out a portion of the
2390 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2391 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
2392 pmap_zero_page_area(m, frag, base - frag);
2395 * If the ending offset is not DEV_BSIZE aligned and the
2396 * valid bit is clear, we have to zero out a portion of
2399 endoff = base + size;
2400 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2401 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
2402 pmap_zero_page_area(m, endoff,
2403 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
2406 * Assert that no previously invalid block that is now being validated
2409 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
2410 ("vm_page_set_valid_range: page %p is dirty", m));
2413 * Set valid bits inclusive of any overlap.
2415 m->valid |= vm_page_bits(base, size);
2419 * Clear the given bits from the specified page's dirty field.
2421 static __inline void
2422 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
2425 #if PAGE_SIZE < 16384
2430 * If the object is locked and the page is neither VPO_BUSY nor
2431 * write mapped, then the page's dirty field cannot possibly be
2432 * set by a concurrent pmap operation.
2434 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2435 if ((m->oflags & VPO_BUSY) == 0 && !pmap_page_is_write_mapped(m))
2436 m->dirty &= ~pagebits;
2439 * The pmap layer can call vm_page_dirty() without
2440 * holding a distinguished lock. The combination of
2441 * the object's lock and an atomic operation suffice
2442 * to guarantee consistency of the page dirty field.
2444 * For PAGE_SIZE == 32768 case, compiler already
2445 * properly aligns the dirty field, so no forcible
2446 * alignment is needed. Only require existence of
2447 * atomic_clear_64 when page size is 32768.
2449 addr = (uintptr_t)&m->dirty;
2450 #if PAGE_SIZE == 32768
2451 atomic_clear_64((uint64_t *)addr, pagebits);
2452 #elif PAGE_SIZE == 16384
2453 atomic_clear_32((uint32_t *)addr, pagebits);
2454 #else /* PAGE_SIZE <= 8192 */
2456 * Use a trick to perform a 32-bit atomic on the
2457 * containing aligned word, to not depend on the existence
2458 * of atomic_clear_{8, 16}.
2460 shift = addr & (sizeof(uint32_t) - 1);
2461 #if BYTE_ORDER == BIG_ENDIAN
2462 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
2466 addr &= ~(sizeof(uint32_t) - 1);
2467 atomic_clear_32((uint32_t *)addr, pagebits << shift);
2468 #endif /* PAGE_SIZE */
2473 * vm_page_set_validclean:
2475 * Sets portions of a page valid and clean. The arguments are expected
2476 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2477 * of any partial chunks touched by the range. The invalid portion of
2478 * such chunks will be zero'd.
2480 * (base + size) must be less then or equal to PAGE_SIZE.
2483 vm_page_set_validclean(vm_page_t m, int base, int size)
2485 vm_page_bits_t oldvalid, pagebits;
2488 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2489 if (size == 0) /* handle degenerate case */
2493 * If the base is not DEV_BSIZE aligned and the valid
2494 * bit is clear, we have to zero out a portion of the
2497 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2498 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
2499 pmap_zero_page_area(m, frag, base - frag);
2502 * If the ending offset is not DEV_BSIZE aligned and the
2503 * valid bit is clear, we have to zero out a portion of
2506 endoff = base + size;
2507 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2508 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
2509 pmap_zero_page_area(m, endoff,
2510 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
2513 * Set valid, clear dirty bits. If validating the entire
2514 * page we can safely clear the pmap modify bit. We also
2515 * use this opportunity to clear the VPO_NOSYNC flag. If a process
2516 * takes a write fault on a MAP_NOSYNC memory area the flag will
2519 * We set valid bits inclusive of any overlap, but we can only
2520 * clear dirty bits for DEV_BSIZE chunks that are fully within
2523 oldvalid = m->valid;
2524 pagebits = vm_page_bits(base, size);
2525 m->valid |= pagebits;
2527 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
2528 frag = DEV_BSIZE - frag;
2534 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
2536 if (base == 0 && size == PAGE_SIZE) {
2538 * The page can only be modified within the pmap if it is
2539 * mapped, and it can only be mapped if it was previously
2542 if (oldvalid == VM_PAGE_BITS_ALL)
2544 * Perform the pmap_clear_modify() first. Otherwise,
2545 * a concurrent pmap operation, such as
2546 * pmap_protect(), could clear a modification in the
2547 * pmap and set the dirty field on the page before
2548 * pmap_clear_modify() had begun and after the dirty
2549 * field was cleared here.
2551 pmap_clear_modify(m);
2553 m->oflags &= ~VPO_NOSYNC;
2554 } else if (oldvalid != VM_PAGE_BITS_ALL)
2555 m->dirty &= ~pagebits;
2557 vm_page_clear_dirty_mask(m, pagebits);
2561 vm_page_clear_dirty(vm_page_t m, int base, int size)
2564 vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
2568 * vm_page_set_invalid:
2570 * Invalidates DEV_BSIZE'd chunks within a page. Both the
2571 * valid and dirty bits for the effected areas are cleared.
2574 vm_page_set_invalid(vm_page_t m, int base, int size)
2576 vm_page_bits_t bits;
2578 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2579 KASSERT((m->oflags & VPO_BUSY) == 0,
2580 ("vm_page_set_invalid: page %p is busy", m));
2581 bits = vm_page_bits(base, size);
2582 if (m->valid == VM_PAGE_BITS_ALL && bits != 0)
2584 KASSERT(!pmap_page_is_mapped(m),
2585 ("vm_page_set_invalid: page %p is mapped", m));
2591 * vm_page_zero_invalid()
2593 * The kernel assumes that the invalid portions of a page contain
2594 * garbage, but such pages can be mapped into memory by user code.
2595 * When this occurs, we must zero out the non-valid portions of the
2596 * page so user code sees what it expects.
2598 * Pages are most often semi-valid when the end of a file is mapped
2599 * into memory and the file's size is not page aligned.
2602 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
2607 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2609 * Scan the valid bits looking for invalid sections that
2610 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
2611 * valid bit may be set ) have already been zerod by
2612 * vm_page_set_validclean().
2614 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
2615 if (i == (PAGE_SIZE / DEV_BSIZE) ||
2616 (m->valid & ((vm_page_bits_t)1 << i))) {
2618 pmap_zero_page_area(m,
2619 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
2626 * setvalid is TRUE when we can safely set the zero'd areas
2627 * as being valid. We can do this if there are no cache consistancy
2628 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
2631 m->valid = VM_PAGE_BITS_ALL;
2637 * Is (partial) page valid? Note that the case where size == 0
2638 * will return FALSE in the degenerate case where the page is
2639 * entirely invalid, and TRUE otherwise.
2642 vm_page_is_valid(vm_page_t m, int base, int size)
2644 vm_page_bits_t bits;
2646 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2647 bits = vm_page_bits(base, size);
2648 if (m->valid && ((m->valid & bits) == bits))
2655 * Set the page's dirty bits if the page is modified.
2658 vm_page_test_dirty(vm_page_t m)
2661 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2662 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
2667 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
2670 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
2674 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
2677 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
2681 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
2684 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
2687 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
2689 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
2692 mtx_assert_(vm_page_lockptr(m), a, file, line);
2696 int so_zerocp_fullpage = 0;
2699 * Replace the given page with a copy. The copied page assumes
2700 * the portion of the given page's "wire_count" that is not the
2701 * responsibility of this copy-on-write mechanism.
2703 * The object containing the given page must have a non-zero
2704 * paging-in-progress count and be locked.
2707 vm_page_cowfault(vm_page_t m)
2713 vm_page_lock_assert(m, MA_OWNED);
2715 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2716 KASSERT(object->paging_in_progress != 0,
2717 ("vm_page_cowfault: object %p's paging-in-progress count is zero.",
2724 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY);
2726 vm_page_insert(m, object, pindex);
2728 VM_OBJECT_UNLOCK(object);
2730 VM_OBJECT_LOCK(object);
2731 if (m == vm_page_lookup(object, pindex)) {
2736 * Page disappeared during the wait.
2744 * check to see if we raced with an xmit complete when
2745 * waiting to allocate a page. If so, put things back
2751 vm_page_unlock(mnew);
2752 vm_page_insert(m, object, pindex);
2753 } else { /* clear COW & copy page */
2754 if (!so_zerocp_fullpage)
2755 pmap_copy_page(m, mnew);
2756 mnew->valid = VM_PAGE_BITS_ALL;
2757 vm_page_dirty(mnew);
2758 mnew->wire_count = m->wire_count - m->cow;
2759 m->wire_count = m->cow;
2765 vm_page_cowclear(vm_page_t m)
2768 vm_page_lock_assert(m, MA_OWNED);
2772 * let vm_fault add back write permission lazily
2776 * sf_buf_free() will free the page, so we needn't do it here
2781 vm_page_cowsetup(vm_page_t m)
2784 vm_page_lock_assert(m, MA_OWNED);
2785 if ((m->flags & PG_FICTITIOUS) != 0 ||
2786 (m->oflags & VPO_UNMANAGED) != 0 ||
2787 m->cow == USHRT_MAX - 1 || !VM_OBJECT_TRYLOCK(m->object))
2790 pmap_remove_write(m);
2791 VM_OBJECT_UNLOCK(m->object);
2797 vm_page_object_lock_assert(vm_page_t m)
2801 * Certain of the page's fields may only be modified by the
2802 * holder of the containing object's lock or the setter of the
2803 * page's VPO_BUSY flag. Unfortunately, the setter of the
2804 * VPO_BUSY flag is not recorded, and thus cannot be checked
2807 if (m->object != NULL && (m->oflags & VPO_BUSY) == 0)
2808 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2812 #include "opt_ddb.h"
2814 #include <sys/kernel.h>
2816 #include <ddb/ddb.h>
2818 DB_SHOW_COMMAND(page, vm_page_print_page_info)
2820 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count);
2821 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count);
2822 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count);
2823 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count);
2824 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count);
2825 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved);
2826 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min);
2827 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target);
2828 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min);
2829 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target);
2832 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
2835 db_printf("PQ_FREE:");
2836 db_printf(" %d", cnt.v_free_count);
2839 db_printf("PQ_CACHE:");
2840 db_printf(" %d", cnt.v_cache_count);
2843 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
2844 *vm_pagequeues[PQ_ACTIVE].pq_cnt,
2845 *vm_pagequeues[PQ_INACTIVE].pq_cnt);