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 regardless of other locks or the busy state of a page.
69 * * In general, no thread besides the page daemon can acquire or
70 * hold more than one page queue lock at a time.
72 * * The page daemon can acquire and hold any pair of page queue
75 * - The object lock is required when inserting or removing
76 * pages from an object (vm_page_insert() or vm_page_remove()).
81 * Resident memory management module.
84 #include <sys/cdefs.h>
85 __FBSDID("$FreeBSD$");
89 #include <sys/param.h>
90 #include <sys/systm.h>
92 #include <sys/kernel.h>
93 #include <sys/limits.h>
94 #include <sys/malloc.h>
96 #include <sys/msgbuf.h>
97 #include <sys/mutex.h>
99 #include <sys/rwlock.h>
100 #include <sys/sysctl.h>
101 #include <sys/vmmeter.h>
102 #include <sys/vnode.h>
106 #include <vm/vm_param.h>
107 #include <vm/vm_kern.h>
108 #include <vm/vm_object.h>
109 #include <vm/vm_page.h>
110 #include <vm/vm_pageout.h>
111 #include <vm/vm_pager.h>
112 #include <vm/vm_phys.h>
113 #include <vm/vm_radix.h>
114 #include <vm/vm_reserv.h>
115 #include <vm/vm_extern.h>
117 #include <vm/uma_int.h>
119 #include <machine/md_var.h>
122 * Associated with page of user-allocatable memory is a
126 struct vm_domain vm_dom[MAXMEMDOM];
127 struct mtx_padalign vm_page_queue_free_mtx;
129 struct mtx_padalign pa_lock[PA_LOCK_COUNT];
131 vm_page_t vm_page_array;
132 long vm_page_array_size;
134 int vm_page_zero_count;
136 static int boot_pages = UMA_BOOT_PAGES;
137 TUNABLE_INT("vm.boot_pages", &boot_pages);
138 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RD, &boot_pages, 0,
139 "number of pages allocated for bootstrapping the VM system");
141 static int pa_tryrelock_restart;
142 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
143 &pa_tryrelock_restart, 0, "Number of tryrelock restarts");
145 static uma_zone_t fakepg_zone;
147 static struct vnode *vm_page_alloc_init(vm_page_t m);
148 static void vm_page_cache_turn_free(vm_page_t m);
149 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
150 static void vm_page_enqueue(int queue, vm_page_t m);
151 static void vm_page_init_fakepg(void *dummy);
152 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
153 vm_pindex_t pindex, vm_page_t mpred);
154 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
157 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init_fakepg, NULL);
160 vm_page_init_fakepg(void *dummy)
163 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
164 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
167 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
168 #if PAGE_SIZE == 32768
170 CTASSERT(sizeof(u_long) >= 8);
175 * Try to acquire a physical address lock while a pmap is locked. If we
176 * fail to trylock we unlock and lock the pmap directly and cache the
177 * locked pa in *locked. The caller should then restart their loop in case
178 * the virtual to physical mapping has changed.
181 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
188 PA_LOCK_ASSERT(lockpa, MA_OWNED);
189 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
196 atomic_add_int(&pa_tryrelock_restart, 1);
205 * Sets the page size, perhaps based upon the memory
206 * size. Must be called before any use of page-size
207 * dependent functions.
210 vm_set_page_size(void)
212 if (vm_cnt.v_page_size == 0)
213 vm_cnt.v_page_size = PAGE_SIZE;
214 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
215 panic("vm_set_page_size: page size not a power of two");
219 * vm_page_blacklist_lookup:
221 * See if a physical address in this page has been listed
222 * in the blacklist tunable. Entries in the tunable are
223 * separated by spaces or commas. If an invalid integer is
224 * encountered then the rest of the string is skipped.
227 vm_page_blacklist_lookup(char *list, vm_paddr_t pa)
232 for (pos = list; *pos != '\0'; pos = cp) {
233 bad = strtoq(pos, &cp, 0);
235 if (*cp == ' ' || *cp == ',') {
242 if (pa == trunc_page(bad))
249 vm_page_domain_init(struct vm_domain *vmd)
251 struct vm_pagequeue *pq;
254 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
255 "vm inactive pagequeue";
256 *__DECONST(int **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_vcnt) =
257 &vm_cnt.v_inactive_count;
258 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
259 "vm active pagequeue";
260 *__DECONST(int **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_vcnt) =
261 &vm_cnt.v_active_count;
262 vmd->vmd_page_count = 0;
263 vmd->vmd_free_count = 0;
265 vmd->vmd_oom = FALSE;
267 for (i = 0; i < PQ_COUNT; i++) {
268 pq = &vmd->vmd_pagequeues[i];
269 TAILQ_INIT(&pq->pq_pl);
270 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
271 MTX_DEF | MTX_DUPOK);
278 * Initializes the resident memory module.
280 * Allocates memory for the page cells, and
281 * for the object/offset-to-page hash table headers.
282 * Each page cell is initialized and placed on the free list.
285 vm_page_startup(vm_offset_t vaddr)
288 vm_paddr_t page_range;
295 /* the biggest memory array is the second group of pages */
297 vm_paddr_t biggestsize;
298 vm_paddr_t low_water, high_water;
303 vaddr = round_page(vaddr);
305 for (i = 0; phys_avail[i + 1]; i += 2) {
306 phys_avail[i] = round_page(phys_avail[i]);
307 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
310 low_water = phys_avail[0];
311 high_water = phys_avail[1];
313 for (i = 0; phys_avail[i + 1]; i += 2) {
314 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
316 if (size > biggestsize) {
320 if (phys_avail[i] < low_water)
321 low_water = phys_avail[i];
322 if (phys_avail[i + 1] > high_water)
323 high_water = phys_avail[i + 1];
330 end = phys_avail[biggestone+1];
333 * Initialize the page and queue locks.
335 mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF);
336 for (i = 0; i < PA_LOCK_COUNT; i++)
337 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
338 for (i = 0; i < vm_ndomains; i++)
339 vm_page_domain_init(&vm_dom[i]);
342 * Allocate memory for use when boot strapping the kernel memory
345 new_end = end - (boot_pages * UMA_SLAB_SIZE);
346 new_end = trunc_page(new_end);
347 mapped = pmap_map(&vaddr, new_end, end,
348 VM_PROT_READ | VM_PROT_WRITE);
349 bzero((void *)mapped, end - new_end);
350 uma_startup((void *)mapped, boot_pages);
352 #if defined(__amd64__) || defined(__i386__) || defined(__arm__) || \
355 * Allocate a bitmap to indicate that a random physical page
356 * needs to be included in a minidump.
358 * The amd64 port needs this to indicate which direct map pages
359 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
361 * However, i386 still needs this workspace internally within the
362 * minidump code. In theory, they are not needed on i386, but are
363 * included should the sf_buf code decide to use them.
366 for (i = 0; dump_avail[i + 1] != 0; i += 2)
367 if (dump_avail[i + 1] > last_pa)
368 last_pa = dump_avail[i + 1];
369 page_range = last_pa / PAGE_SIZE;
370 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
371 new_end -= vm_page_dump_size;
372 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
373 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
374 bzero((void *)vm_page_dump, vm_page_dump_size);
378 * Request that the physical pages underlying the message buffer be
379 * included in a crash dump. Since the message buffer is accessed
380 * through the direct map, they are not automatically included.
382 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
383 last_pa = pa + round_page(msgbufsize);
384 while (pa < last_pa) {
390 * Compute the number of pages of memory that will be available for
391 * use (taking into account the overhead of a page structure per
394 first_page = low_water / PAGE_SIZE;
395 #ifdef VM_PHYSSEG_SPARSE
397 for (i = 0; phys_avail[i + 1] != 0; i += 2)
398 page_range += atop(phys_avail[i + 1] - phys_avail[i]);
399 #elif defined(VM_PHYSSEG_DENSE)
400 page_range = high_water / PAGE_SIZE - first_page;
402 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
407 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
412 * Initialize the mem entry structures now, and put them in the free
415 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
416 mapped = pmap_map(&vaddr, new_end, end,
417 VM_PROT_READ | VM_PROT_WRITE);
418 vm_page_array = (vm_page_t) mapped;
419 #if VM_NRESERVLEVEL > 0
421 * Allocate memory for the reservation management system's data
424 new_end = vm_reserv_startup(&vaddr, new_end, high_water);
426 #if defined(__amd64__) || defined(__mips__)
428 * pmap_map on amd64 and mips can come out of the direct-map, not kvm
429 * like i386, so the pages must be tracked for a crashdump to include
430 * this data. This includes the vm_page_array and the early UMA
433 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
436 phys_avail[biggestone + 1] = new_end;
439 * Clear all of the page structures
441 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
442 for (i = 0; i < page_range; i++)
443 vm_page_array[i].order = VM_NFREEORDER;
444 vm_page_array_size = page_range;
447 * Initialize the physical memory allocator.
452 * Add every available physical page that is not blacklisted to
455 vm_cnt.v_page_count = 0;
456 vm_cnt.v_free_count = 0;
457 list = getenv("vm.blacklist");
458 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
460 last_pa = phys_avail[i + 1];
461 while (pa < last_pa) {
463 vm_page_blacklist_lookup(list, pa))
464 printf("Skipping page with pa 0x%jx\n",
467 vm_phys_add_page(pa);
472 #if VM_NRESERVLEVEL > 0
474 * Initialize the reservation management system.
482 vm_page_reference(vm_page_t m)
485 vm_page_aflag_set(m, PGA_REFERENCED);
489 * vm_page_busy_downgrade:
491 * Downgrade an exclusive busy page into a single shared busy page.
494 vm_page_busy_downgrade(vm_page_t m)
498 vm_page_assert_xbusied(m);
502 x &= VPB_BIT_WAITERS;
503 if (atomic_cmpset_rel_int(&m->busy_lock,
504 VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1) | x))
512 * Return a positive value if the page is shared busied, 0 otherwise.
515 vm_page_sbusied(vm_page_t m)
520 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
526 * Shared unbusy a page.
529 vm_page_sunbusy(vm_page_t m)
533 vm_page_assert_sbusied(m);
537 if (VPB_SHARERS(x) > 1) {
538 if (atomic_cmpset_int(&m->busy_lock, x,
543 if ((x & VPB_BIT_WAITERS) == 0) {
544 KASSERT(x == VPB_SHARERS_WORD(1),
545 ("vm_page_sunbusy: invalid lock state"));
546 if (atomic_cmpset_int(&m->busy_lock,
547 VPB_SHARERS_WORD(1), VPB_UNBUSIED))
551 KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS),
552 ("vm_page_sunbusy: invalid lock state for waiters"));
555 if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) {
566 * vm_page_busy_sleep:
568 * Sleep and release the page lock, using the page pointer as wchan.
569 * This is used to implement the hard-path of busying mechanism.
571 * The given page must be locked.
574 vm_page_busy_sleep(vm_page_t m, const char *wmesg)
578 vm_page_lock_assert(m, MA_OWNED);
581 if (x == VPB_UNBUSIED) {
585 if ((x & VPB_BIT_WAITERS) == 0 &&
586 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS)) {
590 msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0);
596 * Try to shared busy a page.
597 * If the operation succeeds 1 is returned otherwise 0.
598 * The operation never sleeps.
601 vm_page_trysbusy(vm_page_t m)
607 if ((x & VPB_BIT_SHARED) == 0)
609 if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER))
615 * vm_page_xunbusy_hard:
617 * Called after the first try the exclusive unbusy of a page failed.
618 * It is assumed that the waiters bit is on.
621 vm_page_xunbusy_hard(vm_page_t m)
624 vm_page_assert_xbusied(m);
627 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
635 * Wakeup anyone waiting for the page.
636 * The ownership bits do not change.
638 * The given page must be locked.
641 vm_page_flash(vm_page_t m)
645 vm_page_lock_assert(m, MA_OWNED);
649 if ((x & VPB_BIT_WAITERS) == 0)
651 if (atomic_cmpset_int(&m->busy_lock, x,
652 x & (~VPB_BIT_WAITERS)))
659 * Keep page from being freed by the page daemon
660 * much of the same effect as wiring, except much lower
661 * overhead and should be used only for *very* temporary
662 * holding ("wiring").
665 vm_page_hold(vm_page_t mem)
668 vm_page_lock_assert(mem, MA_OWNED);
673 vm_page_unhold(vm_page_t mem)
676 vm_page_lock_assert(mem, MA_OWNED);
677 KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!"));
679 if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0)
680 vm_page_free_toq(mem);
684 * vm_page_unhold_pages:
686 * Unhold each of the pages that is referenced by the given array.
689 vm_page_unhold_pages(vm_page_t *ma, int count)
691 struct mtx *mtx, *new_mtx;
694 for (; count != 0; count--) {
696 * Avoid releasing and reacquiring the same page lock.
698 new_mtx = vm_page_lockptr(*ma);
699 if (mtx != new_mtx) {
713 PHYS_TO_VM_PAGE(vm_paddr_t pa)
717 #ifdef VM_PHYSSEG_SPARSE
718 m = vm_phys_paddr_to_vm_page(pa);
720 m = vm_phys_fictitious_to_vm_page(pa);
722 #elif defined(VM_PHYSSEG_DENSE)
726 if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
727 m = &vm_page_array[pi - first_page];
730 return (vm_phys_fictitious_to_vm_page(pa));
732 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
739 * Create a fictitious page with the specified physical address and
740 * memory attribute. The memory attribute is the only the machine-
741 * dependent aspect of a fictitious page that must be initialized.
744 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
748 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
749 vm_page_initfake(m, paddr, memattr);
754 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
757 if ((m->flags & PG_FICTITIOUS) != 0) {
759 * The page's memattr might have changed since the
760 * previous initialization. Update the pmap to the
765 m->phys_addr = paddr;
767 /* Fictitious pages don't use "segind". */
768 m->flags = PG_FICTITIOUS;
769 /* Fictitious pages don't use "order" or "pool". */
770 m->oflags = VPO_UNMANAGED;
771 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
775 pmap_page_set_memattr(m, memattr);
781 * Release a fictitious page.
784 vm_page_putfake(vm_page_t m)
787 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
788 KASSERT((m->flags & PG_FICTITIOUS) != 0,
789 ("vm_page_putfake: bad page %p", m));
790 uma_zfree(fakepg_zone, m);
794 * vm_page_updatefake:
796 * Update the given fictitious page to the specified physical address and
800 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
803 KASSERT((m->flags & PG_FICTITIOUS) != 0,
804 ("vm_page_updatefake: bad page %p", m));
805 m->phys_addr = paddr;
806 pmap_page_set_memattr(m, memattr);
815 vm_page_free(vm_page_t m)
818 m->flags &= ~PG_ZERO;
825 * Free a page to the zerod-pages queue
828 vm_page_free_zero(vm_page_t m)
836 * Unbusy and handle the page queueing for a page from the VOP_GETPAGES()
837 * array which is not the request page.
840 vm_page_readahead_finish(vm_page_t m)
845 * Since the page is not the requested page, whether
846 * it should be activated or deactivated is not
847 * obvious. Empirical results have shown that
848 * deactivating the page is usually the best choice,
849 * unless the page is wanted by another thread.
852 if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
855 vm_page_deactivate(m);
860 * Free the completely invalid page. Such page state
861 * occurs due to the short read operation which did
862 * not covered our page at all, or in case when a read
872 * vm_page_sleep_if_busy:
874 * Sleep and release the page queues lock if the page is busied.
875 * Returns TRUE if the thread slept.
877 * The given page must be unlocked and object containing it must
881 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
885 vm_page_lock_assert(m, MA_NOTOWNED);
886 VM_OBJECT_ASSERT_WLOCKED(m->object);
888 if (vm_page_busied(m)) {
890 * The page-specific object must be cached because page
891 * identity can change during the sleep, causing the
892 * re-lock of a different object.
893 * It is assumed that a reference to the object is already
894 * held by the callers.
898 VM_OBJECT_WUNLOCK(obj);
899 vm_page_busy_sleep(m, msg);
900 VM_OBJECT_WLOCK(obj);
907 * vm_page_dirty_KBI: [ internal use only ]
909 * Set all bits in the page's dirty field.
911 * The object containing the specified page must be locked if the
912 * call is made from the machine-independent layer.
914 * See vm_page_clear_dirty_mask().
916 * This function should only be called by vm_page_dirty().
919 vm_page_dirty_KBI(vm_page_t m)
922 /* These assertions refer to this operation by its public name. */
923 KASSERT((m->flags & PG_CACHED) == 0,
924 ("vm_page_dirty: page in cache!"));
925 KASSERT(m->valid == VM_PAGE_BITS_ALL,
926 ("vm_page_dirty: page is invalid!"));
927 m->dirty = VM_PAGE_BITS_ALL;
931 * vm_page_insert: [ internal use only ]
933 * Inserts the given mem entry into the object and object list.
935 * The object must be locked.
938 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
942 VM_OBJECT_ASSERT_WLOCKED(object);
943 mpred = vm_radix_lookup_le(&object->rtree, pindex);
944 return (vm_page_insert_after(m, object, pindex, mpred));
948 * vm_page_insert_after:
950 * Inserts the page "m" into the specified object at offset "pindex".
952 * The page "mpred" must immediately precede the offset "pindex" within
953 * the specified object.
955 * The object must be locked.
958 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
965 VM_OBJECT_ASSERT_WLOCKED(object);
966 KASSERT(m->object == NULL,
967 ("vm_page_insert_after: page already inserted"));
969 KASSERT(mpred->object == object,
970 ("vm_page_insert_after: object doesn't contain mpred"));
971 KASSERT(mpred->pindex < pindex,
972 ("vm_page_insert_after: mpred doesn't precede pindex"));
973 msucc = TAILQ_NEXT(mpred, listq);
975 msucc = TAILQ_FIRST(&object->memq);
977 KASSERT(msucc->pindex > pindex,
978 ("vm_page_insert_after: msucc doesn't succeed pindex"));
981 * Record the object/offset pair in this page
989 * Now link into the object's ordered list of backed pages.
991 if (vm_radix_insert(&object->rtree, m)) {
996 vm_page_insert_radixdone(m, object, mpred);
1001 * vm_page_insert_radixdone:
1003 * Complete page "m" insertion into the specified object after the
1004 * radix trie hooking.
1006 * The page "mpred" must precede the offset "m->pindex" within the
1009 * The object must be locked.
1012 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1015 VM_OBJECT_ASSERT_WLOCKED(object);
1016 KASSERT(object != NULL && m->object == object,
1017 ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1018 if (mpred != NULL) {
1019 KASSERT(mpred->object == object,
1020 ("vm_page_insert_after: object doesn't contain mpred"));
1021 KASSERT(mpred->pindex < m->pindex,
1022 ("vm_page_insert_after: mpred doesn't precede pindex"));
1026 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1028 TAILQ_INSERT_HEAD(&object->memq, m, listq);
1031 * Show that the object has one more resident page.
1033 object->resident_page_count++;
1036 * Hold the vnode until the last page is released.
1038 if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1039 vhold(object->handle);
1042 * Since we are inserting a new and possibly dirty page,
1043 * update the object's OBJ_MIGHTBEDIRTY flag.
1045 if (pmap_page_is_write_mapped(m))
1046 vm_object_set_writeable_dirty(object);
1052 * Removes the given mem entry from the object/offset-page
1053 * table and the object page list, but do not invalidate/terminate
1054 * the backing store.
1056 * The object must be locked. The page must be locked if it is managed.
1059 vm_page_remove(vm_page_t m)
1064 if ((m->oflags & VPO_UNMANAGED) == 0)
1065 vm_page_lock_assert(m, MA_OWNED);
1066 if ((object = m->object) == NULL)
1068 VM_OBJECT_ASSERT_WLOCKED(object);
1069 if (vm_page_xbusied(m)) {
1071 if ((m->oflags & VPO_UNMANAGED) != 0 &&
1072 !mtx_owned(vm_page_lockptr(m))) {
1077 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1083 * Now remove from the object's list of backed pages.
1085 vm_radix_remove(&object->rtree, m->pindex);
1086 TAILQ_REMOVE(&object->memq, m, listq);
1089 * And show that the object has one fewer resident page.
1091 object->resident_page_count--;
1094 * The vnode may now be recycled.
1096 if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1097 vdrop(object->handle);
1105 * Returns the page associated with the object/offset
1106 * pair specified; if none is found, NULL is returned.
1108 * The object must be locked.
1111 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1114 VM_OBJECT_ASSERT_LOCKED(object);
1115 return (vm_radix_lookup(&object->rtree, pindex));
1119 * vm_page_find_least:
1121 * Returns the page associated with the object with least pindex
1122 * greater than or equal to the parameter pindex, or NULL.
1124 * The object must be locked.
1127 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1131 VM_OBJECT_ASSERT_LOCKED(object);
1132 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1133 m = vm_radix_lookup_ge(&object->rtree, pindex);
1138 * Returns the given page's successor (by pindex) within the object if it is
1139 * resident; if none is found, NULL is returned.
1141 * The object must be locked.
1144 vm_page_next(vm_page_t m)
1148 VM_OBJECT_ASSERT_WLOCKED(m->object);
1149 if ((next = TAILQ_NEXT(m, listq)) != NULL &&
1150 next->pindex != m->pindex + 1)
1156 * Returns the given page's predecessor (by pindex) within the object if it is
1157 * resident; if none is found, NULL is returned.
1159 * The object must be locked.
1162 vm_page_prev(vm_page_t m)
1166 VM_OBJECT_ASSERT_WLOCKED(m->object);
1167 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
1168 prev->pindex != m->pindex - 1)
1174 * Uses the page mnew as a replacement for an existing page at index
1175 * pindex which must be already present in the object.
1177 * The existing page must not be on a paging queue.
1180 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1182 vm_page_t mold, mpred;
1184 VM_OBJECT_ASSERT_WLOCKED(object);
1187 * This function mostly follows vm_page_insert() and
1188 * vm_page_remove() without the radix, object count and vnode
1189 * dance. Double check such functions for more comments.
1191 mpred = vm_radix_lookup(&object->rtree, pindex);
1192 KASSERT(mpred != NULL,
1193 ("vm_page_replace: replacing page not present with pindex"));
1194 mpred = TAILQ_PREV(mpred, respgs, listq);
1196 KASSERT(mpred->pindex < pindex,
1197 ("vm_page_insert_after: mpred doesn't precede pindex"));
1199 mnew->object = object;
1200 mnew->pindex = pindex;
1201 mold = vm_radix_replace(&object->rtree, mnew);
1202 KASSERT(mold->queue == PQ_NONE,
1203 ("vm_page_replace: mold is on a paging queue"));
1205 /* Detach the old page from the resident tailq. */
1206 TAILQ_REMOVE(&object->memq, mold, listq);
1208 mold->object = NULL;
1209 vm_page_xunbusy(mold);
1211 /* Insert the new page in the resident tailq. */
1213 TAILQ_INSERT_AFTER(&object->memq, mpred, mnew, listq);
1215 TAILQ_INSERT_HEAD(&object->memq, mnew, listq);
1216 if (pmap_page_is_write_mapped(mnew))
1217 vm_object_set_writeable_dirty(object);
1224 * Move the given memory entry from its
1225 * current object to the specified target object/offset.
1227 * Note: swap associated with the page must be invalidated by the move. We
1228 * have to do this for several reasons: (1) we aren't freeing the
1229 * page, (2) we are dirtying the page, (3) the VM system is probably
1230 * moving the page from object A to B, and will then later move
1231 * the backing store from A to B and we can't have a conflict.
1233 * Note: we *always* dirty the page. It is necessary both for the
1234 * fact that we moved it, and because we may be invalidating
1235 * swap. If the page is on the cache, we have to deactivate it
1236 * or vm_page_dirty() will panic. Dirty pages are not allowed
1239 * The objects must be locked.
1242 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1247 VM_OBJECT_ASSERT_WLOCKED(new_object);
1249 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1250 KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1251 ("vm_page_rename: pindex already renamed"));
1254 * Create a custom version of vm_page_insert() which does not depend
1255 * by m_prev and can cheat on the implementation aspects of the
1259 m->pindex = new_pindex;
1260 if (vm_radix_insert(&new_object->rtree, m)) {
1266 * The operation cannot fail anymore. The removal must happen before
1267 * the listq iterator is tainted.
1273 /* Return back to the new pindex to complete vm_page_insert(). */
1274 m->pindex = new_pindex;
1275 m->object = new_object;
1277 vm_page_insert_radixdone(m, new_object, mpred);
1283 * Convert all of the given object's cached pages that have a
1284 * pindex within the given range into free pages. If the value
1285 * zero is given for "end", then the range's upper bound is
1286 * infinity. If the given object is backed by a vnode and it
1287 * transitions from having one or more cached pages to none, the
1288 * vnode's hold count is reduced.
1291 vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1296 mtx_lock(&vm_page_queue_free_mtx);
1297 if (__predict_false(vm_radix_is_empty(&object->cache))) {
1298 mtx_unlock(&vm_page_queue_free_mtx);
1301 while ((m = vm_radix_lookup_ge(&object->cache, start)) != NULL) {
1302 if (end != 0 && m->pindex >= end)
1304 vm_radix_remove(&object->cache, m->pindex);
1305 vm_page_cache_turn_free(m);
1307 empty = vm_radix_is_empty(&object->cache);
1308 mtx_unlock(&vm_page_queue_free_mtx);
1309 if (object->type == OBJT_VNODE && empty)
1310 vdrop(object->handle);
1314 * Returns the cached page that is associated with the given
1315 * object and offset. If, however, none exists, returns NULL.
1317 * The free page queue must be locked.
1319 static inline vm_page_t
1320 vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex)
1323 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1324 return (vm_radix_lookup(&object->cache, pindex));
1328 * Remove the given cached page from its containing object's
1329 * collection of cached pages.
1331 * The free page queue must be locked.
1334 vm_page_cache_remove(vm_page_t m)
1337 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1338 KASSERT((m->flags & PG_CACHED) != 0,
1339 ("vm_page_cache_remove: page %p is not cached", m));
1340 vm_radix_remove(&m->object->cache, m->pindex);
1342 vm_cnt.v_cache_count--;
1346 * Transfer all of the cached pages with offset greater than or
1347 * equal to 'offidxstart' from the original object's cache to the
1348 * new object's cache. However, any cached pages with offset
1349 * greater than or equal to the new object's size are kept in the
1350 * original object. Initially, the new object's cache must be
1351 * empty. Offset 'offidxstart' in the original object must
1352 * correspond to offset zero in the new object.
1354 * The new object must be locked.
1357 vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart,
1358 vm_object_t new_object)
1363 * Insertion into an object's collection of cached pages
1364 * requires the object to be locked. In contrast, removal does
1367 VM_OBJECT_ASSERT_WLOCKED(new_object);
1368 KASSERT(vm_radix_is_empty(&new_object->cache),
1369 ("vm_page_cache_transfer: object %p has cached pages",
1371 mtx_lock(&vm_page_queue_free_mtx);
1372 while ((m = vm_radix_lookup_ge(&orig_object->cache,
1373 offidxstart)) != NULL) {
1375 * Transfer all of the pages with offset greater than or
1376 * equal to 'offidxstart' from the original object's
1377 * cache to the new object's cache.
1379 if ((m->pindex - offidxstart) >= new_object->size)
1381 vm_radix_remove(&orig_object->cache, m->pindex);
1382 /* Update the page's object and offset. */
1383 m->object = new_object;
1384 m->pindex -= offidxstart;
1385 if (vm_radix_insert(&new_object->cache, m))
1386 vm_page_cache_turn_free(m);
1388 mtx_unlock(&vm_page_queue_free_mtx);
1392 * Returns TRUE if a cached page is associated with the given object and
1393 * offset, and FALSE otherwise.
1395 * The object must be locked.
1398 vm_page_is_cached(vm_object_t object, vm_pindex_t pindex)
1403 * Insertion into an object's collection of cached pages requires the
1404 * object to be locked. Therefore, if the object is locked and the
1405 * object's collection is empty, there is no need to acquire the free
1406 * page queues lock in order to prove that the specified page doesn't
1409 VM_OBJECT_ASSERT_WLOCKED(object);
1410 if (__predict_true(vm_object_cache_is_empty(object)))
1412 mtx_lock(&vm_page_queue_free_mtx);
1413 m = vm_page_cache_lookup(object, pindex);
1414 mtx_unlock(&vm_page_queue_free_mtx);
1421 * Allocate and return a page that is associated with the specified
1422 * object and offset pair. By default, this page is exclusive busied.
1424 * The caller must always specify an allocation class.
1426 * allocation classes:
1427 * VM_ALLOC_NORMAL normal process request
1428 * VM_ALLOC_SYSTEM system *really* needs a page
1429 * VM_ALLOC_INTERRUPT interrupt time request
1431 * optional allocation flags:
1432 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1433 * intends to allocate
1434 * VM_ALLOC_IFCACHED return page only if it is cached
1435 * VM_ALLOC_IFNOTCACHED return NULL, do not reactivate if the page
1437 * VM_ALLOC_NOBUSY do not exclusive busy the page
1438 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1439 * VM_ALLOC_NOOBJ page is not associated with an object and
1440 * should not be exclusive busy
1441 * VM_ALLOC_SBUSY shared busy the allocated page
1442 * VM_ALLOC_WIRED wire the allocated page
1443 * VM_ALLOC_ZERO prefer a zeroed page
1445 * This routine may not sleep.
1448 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1450 struct vnode *vp = NULL;
1451 vm_object_t m_object;
1453 int flags, req_class;
1455 mpred = 0; /* XXX: pacify gcc */
1456 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1457 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1458 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1459 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1460 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object,
1463 VM_OBJECT_ASSERT_WLOCKED(object);
1465 req_class = req & VM_ALLOC_CLASS_MASK;
1468 * The page daemon is allowed to dig deeper into the free page list.
1470 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1471 req_class = VM_ALLOC_SYSTEM;
1473 if (object != NULL) {
1474 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1475 KASSERT(mpred == NULL || mpred->pindex != pindex,
1476 ("vm_page_alloc: pindex already allocated"));
1480 * The page allocation request can came from consumers which already
1481 * hold the free page queue mutex, like vm_page_insert() in
1484 mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE);
1485 if (vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_free_reserved ||
1486 (req_class == VM_ALLOC_SYSTEM &&
1487 vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_interrupt_free_min) ||
1488 (req_class == VM_ALLOC_INTERRUPT &&
1489 vm_cnt.v_free_count + vm_cnt.v_cache_count > 0)) {
1491 * Allocate from the free queue if the number of free pages
1492 * exceeds the minimum for the request class.
1494 if (object != NULL &&
1495 (m = vm_page_cache_lookup(object, pindex)) != NULL) {
1496 if ((req & VM_ALLOC_IFNOTCACHED) != 0) {
1497 mtx_unlock(&vm_page_queue_free_mtx);
1500 if (vm_phys_unfree_page(m))
1501 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0);
1502 #if VM_NRESERVLEVEL > 0
1503 else if (!vm_reserv_reactivate_page(m))
1507 panic("vm_page_alloc: cache page %p is missing"
1508 " from the free queue", m);
1509 } else if ((req & VM_ALLOC_IFCACHED) != 0) {
1510 mtx_unlock(&vm_page_queue_free_mtx);
1512 #if VM_NRESERVLEVEL > 0
1513 } else if (object == NULL || (object->flags & (OBJ_COLORED |
1514 OBJ_FICTITIOUS)) != OBJ_COLORED || (m =
1515 vm_reserv_alloc_page(object, pindex, mpred)) == NULL) {
1519 m = vm_phys_alloc_pages(object != NULL ?
1520 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1521 #if VM_NRESERVLEVEL > 0
1522 if (m == NULL && vm_reserv_reclaim_inactive()) {
1523 m = vm_phys_alloc_pages(object != NULL ?
1524 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1531 * Not allocatable, give up.
1533 mtx_unlock(&vm_page_queue_free_mtx);
1534 atomic_add_int(&vm_pageout_deficit,
1535 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1536 pagedaemon_wakeup();
1541 * At this point we had better have found a good page.
1543 KASSERT(m != NULL, ("vm_page_alloc: missing page"));
1544 KASSERT(m->queue == PQ_NONE,
1545 ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue));
1546 KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m));
1547 KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m));
1548 KASSERT(!vm_page_sbusied(m),
1549 ("vm_page_alloc: page %p is busy", m));
1550 KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m));
1551 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1552 ("vm_page_alloc: page %p has unexpected memattr %d", m,
1553 pmap_page_get_memattr(m)));
1554 if ((m->flags & PG_CACHED) != 0) {
1555 KASSERT((m->flags & PG_ZERO) == 0,
1556 ("vm_page_alloc: cached page %p is PG_ZERO", m));
1557 KASSERT(m->valid != 0,
1558 ("vm_page_alloc: cached page %p is invalid", m));
1559 if (m->object == object && m->pindex == pindex)
1560 vm_cnt.v_reactivated++;
1563 m_object = m->object;
1564 vm_page_cache_remove(m);
1565 if (m_object->type == OBJT_VNODE &&
1566 vm_object_cache_is_empty(m_object))
1567 vp = m_object->handle;
1569 KASSERT(m->valid == 0,
1570 ("vm_page_alloc: free page %p is valid", m));
1571 vm_phys_freecnt_adj(m, -1);
1572 if ((m->flags & PG_ZERO) != 0)
1573 vm_page_zero_count--;
1575 mtx_unlock(&vm_page_queue_free_mtx);
1578 * Initialize the page. Only the PG_ZERO flag is inherited.
1581 if ((req & VM_ALLOC_ZERO) != 0)
1584 if ((req & VM_ALLOC_NODUMP) != 0)
1588 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1590 m->busy_lock = VPB_UNBUSIED;
1591 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1592 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1593 if ((req & VM_ALLOC_SBUSY) != 0)
1594 m->busy_lock = VPB_SHARERS_WORD(1);
1595 if (req & VM_ALLOC_WIRED) {
1597 * The page lock is not required for wiring a page until that
1598 * page is inserted into the object.
1600 atomic_add_int(&vm_cnt.v_wire_count, 1);
1605 if (object != NULL) {
1606 if (vm_page_insert_after(m, object, pindex, mpred)) {
1607 /* See the comment below about hold count. */
1610 pagedaemon_wakeup();
1611 if (req & VM_ALLOC_WIRED) {
1612 atomic_subtract_int(&vm_cnt.v_wire_count, 1);
1620 /* Ignore device objects; the pager sets "memattr" for them. */
1621 if (object->memattr != VM_MEMATTR_DEFAULT &&
1622 (object->flags & OBJ_FICTITIOUS) == 0)
1623 pmap_page_set_memattr(m, object->memattr);
1628 * The following call to vdrop() must come after the above call
1629 * to vm_page_insert() in case both affect the same object and
1630 * vnode. Otherwise, the affected vnode's hold count could
1631 * temporarily become zero.
1637 * Don't wakeup too often - wakeup the pageout daemon when
1638 * we would be nearly out of memory.
1640 if (vm_paging_needed())
1641 pagedaemon_wakeup();
1647 vm_page_alloc_contig_vdrop(struct spglist *lst)
1650 while (!SLIST_EMPTY(lst)) {
1651 vdrop((struct vnode *)SLIST_FIRST(lst)-> plinks.s.pv);
1652 SLIST_REMOVE_HEAD(lst, plinks.s.ss);
1657 * vm_page_alloc_contig:
1659 * Allocate a contiguous set of physical pages of the given size "npages"
1660 * from the free lists. All of the physical pages must be at or above
1661 * the given physical address "low" and below the given physical address
1662 * "high". The given value "alignment" determines the alignment of the
1663 * first physical page in the set. If the given value "boundary" is
1664 * non-zero, then the set of physical pages cannot cross any physical
1665 * address boundary that is a multiple of that value. Both "alignment"
1666 * and "boundary" must be a power of two.
1668 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1669 * then the memory attribute setting for the physical pages is configured
1670 * to the object's memory attribute setting. Otherwise, the memory
1671 * attribute setting for the physical pages is configured to "memattr",
1672 * overriding the object's memory attribute setting. However, if the
1673 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1674 * memory attribute setting for the physical pages cannot be configured
1675 * to VM_MEMATTR_DEFAULT.
1677 * The caller must always specify an allocation class.
1679 * allocation classes:
1680 * VM_ALLOC_NORMAL normal process request
1681 * VM_ALLOC_SYSTEM system *really* needs a page
1682 * VM_ALLOC_INTERRUPT interrupt time request
1684 * optional allocation flags:
1685 * VM_ALLOC_NOBUSY do not exclusive busy the page
1686 * VM_ALLOC_NOOBJ page is not associated with an object and
1687 * should not be exclusive busy
1688 * VM_ALLOC_SBUSY shared busy the allocated page
1689 * VM_ALLOC_WIRED wire the allocated page
1690 * VM_ALLOC_ZERO prefer a zeroed page
1692 * This routine may not sleep.
1695 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1696 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1697 vm_paddr_t boundary, vm_memattr_t memattr)
1700 struct spglist deferred_vdrop_list;
1701 vm_page_t m, m_tmp, m_ret;
1705 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1706 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1707 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1708 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1709 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object,
1711 if (object != NULL) {
1712 VM_OBJECT_ASSERT_WLOCKED(object);
1713 KASSERT(object->type == OBJT_PHYS,
1714 ("vm_page_alloc_contig: object %p isn't OBJT_PHYS",
1717 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
1718 req_class = req & VM_ALLOC_CLASS_MASK;
1721 * The page daemon is allowed to dig deeper into the free page list.
1723 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1724 req_class = VM_ALLOC_SYSTEM;
1726 SLIST_INIT(&deferred_vdrop_list);
1727 mtx_lock(&vm_page_queue_free_mtx);
1728 if (vm_cnt.v_free_count + vm_cnt.v_cache_count >= npages +
1729 vm_cnt.v_free_reserved || (req_class == VM_ALLOC_SYSTEM &&
1730 vm_cnt.v_free_count + vm_cnt.v_cache_count >= npages +
1731 vm_cnt.v_interrupt_free_min) || (req_class == VM_ALLOC_INTERRUPT &&
1732 vm_cnt.v_free_count + vm_cnt.v_cache_count >= npages)) {
1733 #if VM_NRESERVLEVEL > 0
1735 if (object == NULL || (object->flags & OBJ_COLORED) == 0 ||
1736 (m_ret = vm_reserv_alloc_contig(object, pindex, npages,
1737 low, high, alignment, boundary)) == NULL)
1739 m_ret = vm_phys_alloc_contig(npages, low, high,
1740 alignment, boundary);
1742 mtx_unlock(&vm_page_queue_free_mtx);
1743 atomic_add_int(&vm_pageout_deficit, npages);
1744 pagedaemon_wakeup();
1748 for (m = m_ret; m < &m_ret[npages]; m++) {
1749 drop = vm_page_alloc_init(m);
1752 * Enqueue the vnode for deferred vdrop().
1754 m->plinks.s.pv = drop;
1755 SLIST_INSERT_HEAD(&deferred_vdrop_list, m,
1760 #if VM_NRESERVLEVEL > 0
1761 if (vm_reserv_reclaim_contig(npages, low, high, alignment,
1766 mtx_unlock(&vm_page_queue_free_mtx);
1771 * Initialize the pages. Only the PG_ZERO flag is inherited.
1774 if ((req & VM_ALLOC_ZERO) != 0)
1776 if ((req & VM_ALLOC_NODUMP) != 0)
1778 if ((req & VM_ALLOC_WIRED) != 0)
1779 atomic_add_int(&vm_cnt.v_wire_count, npages);
1780 if (object != NULL) {
1781 if (object->memattr != VM_MEMATTR_DEFAULT &&
1782 memattr == VM_MEMATTR_DEFAULT)
1783 memattr = object->memattr;
1785 for (m = m_ret; m < &m_ret[npages]; m++) {
1787 m->flags = (m->flags | PG_NODUMP) & flags;
1788 m->busy_lock = VPB_UNBUSIED;
1789 if (object != NULL) {
1790 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
1791 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1792 if ((req & VM_ALLOC_SBUSY) != 0)
1793 m->busy_lock = VPB_SHARERS_WORD(1);
1795 if ((req & VM_ALLOC_WIRED) != 0)
1797 /* Unmanaged pages don't use "act_count". */
1798 m->oflags = VPO_UNMANAGED;
1799 if (object != NULL) {
1800 if (vm_page_insert(m, object, pindex)) {
1801 vm_page_alloc_contig_vdrop(
1802 &deferred_vdrop_list);
1803 if (vm_paging_needed())
1804 pagedaemon_wakeup();
1805 if ((req & VM_ALLOC_WIRED) != 0)
1806 atomic_subtract_int(&vm_cnt.v_wire_count,
1808 for (m_tmp = m, m = m_ret;
1809 m < &m_ret[npages]; m++) {
1810 if ((req & VM_ALLOC_WIRED) != 0)
1820 if (memattr != VM_MEMATTR_DEFAULT)
1821 pmap_page_set_memattr(m, memattr);
1824 vm_page_alloc_contig_vdrop(&deferred_vdrop_list);
1825 if (vm_paging_needed())
1826 pagedaemon_wakeup();
1831 * Initialize a page that has been freshly dequeued from a freelist.
1832 * The caller has to drop the vnode returned, if it is not NULL.
1834 * This function may only be used to initialize unmanaged pages.
1836 * To be called with vm_page_queue_free_mtx held.
1838 static struct vnode *
1839 vm_page_alloc_init(vm_page_t m)
1842 vm_object_t m_object;
1844 KASSERT(m->queue == PQ_NONE,
1845 ("vm_page_alloc_init: page %p has unexpected queue %d",
1847 KASSERT(m->wire_count == 0,
1848 ("vm_page_alloc_init: page %p is wired", m));
1849 KASSERT(m->hold_count == 0,
1850 ("vm_page_alloc_init: page %p is held", m));
1851 KASSERT(!vm_page_sbusied(m),
1852 ("vm_page_alloc_init: page %p is busy", m));
1853 KASSERT(m->dirty == 0,
1854 ("vm_page_alloc_init: page %p is dirty", m));
1855 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1856 ("vm_page_alloc_init: page %p has unexpected memattr %d",
1857 m, pmap_page_get_memattr(m)));
1858 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1860 if ((m->flags & PG_CACHED) != 0) {
1861 KASSERT((m->flags & PG_ZERO) == 0,
1862 ("vm_page_alloc_init: cached page %p is PG_ZERO", m));
1864 m_object = m->object;
1865 vm_page_cache_remove(m);
1866 if (m_object->type == OBJT_VNODE &&
1867 vm_object_cache_is_empty(m_object))
1868 drop = m_object->handle;
1870 KASSERT(m->valid == 0,
1871 ("vm_page_alloc_init: free page %p is valid", m));
1872 vm_phys_freecnt_adj(m, -1);
1873 if ((m->flags & PG_ZERO) != 0)
1874 vm_page_zero_count--;
1880 * vm_page_alloc_freelist:
1882 * Allocate a physical page from the specified free page list.
1884 * The caller must always specify an allocation class.
1886 * allocation classes:
1887 * VM_ALLOC_NORMAL normal process request
1888 * VM_ALLOC_SYSTEM system *really* needs a page
1889 * VM_ALLOC_INTERRUPT interrupt time request
1891 * optional allocation flags:
1892 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1893 * intends to allocate
1894 * VM_ALLOC_WIRED wire the allocated page
1895 * VM_ALLOC_ZERO prefer a zeroed page
1897 * This routine may not sleep.
1900 vm_page_alloc_freelist(int flind, int req)
1907 req_class = req & VM_ALLOC_CLASS_MASK;
1910 * The page daemon is allowed to dig deeper into the free page list.
1912 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1913 req_class = VM_ALLOC_SYSTEM;
1916 * Do not allocate reserved pages unless the req has asked for it.
1918 mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE);
1919 if (vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_free_reserved ||
1920 (req_class == VM_ALLOC_SYSTEM &&
1921 vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_interrupt_free_min) ||
1922 (req_class == VM_ALLOC_INTERRUPT &&
1923 vm_cnt.v_free_count + vm_cnt.v_cache_count > 0))
1924 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0);
1926 mtx_unlock(&vm_page_queue_free_mtx);
1927 atomic_add_int(&vm_pageout_deficit,
1928 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1929 pagedaemon_wakeup();
1933 mtx_unlock(&vm_page_queue_free_mtx);
1936 drop = vm_page_alloc_init(m);
1937 mtx_unlock(&vm_page_queue_free_mtx);
1940 * Initialize the page. Only the PG_ZERO flag is inherited.
1944 if ((req & VM_ALLOC_ZERO) != 0)
1947 if ((req & VM_ALLOC_WIRED) != 0) {
1949 * The page lock is not required for wiring a page that does
1950 * not belong to an object.
1952 atomic_add_int(&vm_cnt.v_wire_count, 1);
1955 /* Unmanaged pages don't use "act_count". */
1956 m->oflags = VPO_UNMANAGED;
1959 if (vm_paging_needed())
1960 pagedaemon_wakeup();
1965 * vm_wait: (also see VM_WAIT macro)
1967 * Sleep until free pages are available for allocation.
1968 * - Called in various places before memory allocations.
1974 mtx_lock(&vm_page_queue_free_mtx);
1975 if (curproc == pageproc) {
1976 vm_pageout_pages_needed = 1;
1977 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx,
1978 PDROP | PSWP, "VMWait", 0);
1980 if (!vm_pages_needed) {
1981 vm_pages_needed = 1;
1982 wakeup(&vm_pages_needed);
1984 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM,
1990 * vm_waitpfault: (also see VM_WAITPFAULT macro)
1992 * Sleep until free pages are available for allocation.
1993 * - Called only in vm_fault so that processes page faulting
1994 * can be easily tracked.
1995 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
1996 * processes will be able to grab memory first. Do not change
1997 * this balance without careful testing first.
2003 mtx_lock(&vm_page_queue_free_mtx);
2004 if (!vm_pages_needed) {
2005 vm_pages_needed = 1;
2006 wakeup(&vm_pages_needed);
2008 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER,
2012 struct vm_pagequeue *
2013 vm_page_pagequeue(vm_page_t m)
2016 return (&vm_phys_domain(m)->vmd_pagequeues[m->queue]);
2022 * Remove the given page from its current page queue.
2024 * The page must be locked.
2027 vm_page_dequeue(vm_page_t m)
2029 struct vm_pagequeue *pq;
2031 vm_page_assert_locked(m);
2032 KASSERT(m->queue == PQ_ACTIVE || m->queue == PQ_INACTIVE,
2033 ("vm_page_dequeue: page %p is not queued", m));
2034 pq = vm_page_pagequeue(m);
2035 vm_pagequeue_lock(pq);
2037 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2038 vm_pagequeue_cnt_dec(pq);
2039 vm_pagequeue_unlock(pq);
2043 * vm_page_dequeue_locked:
2045 * Remove the given page from its current page queue.
2047 * The page and page queue must be locked.
2050 vm_page_dequeue_locked(vm_page_t m)
2052 struct vm_pagequeue *pq;
2054 vm_page_lock_assert(m, MA_OWNED);
2055 pq = vm_page_pagequeue(m);
2056 vm_pagequeue_assert_locked(pq);
2058 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2059 vm_pagequeue_cnt_dec(pq);
2065 * Add the given page to the specified page queue.
2067 * The page must be locked.
2070 vm_page_enqueue(int queue, vm_page_t m)
2072 struct vm_pagequeue *pq;
2074 vm_page_lock_assert(m, MA_OWNED);
2075 pq = &vm_phys_domain(m)->vmd_pagequeues[queue];
2076 vm_pagequeue_lock(pq);
2078 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2079 vm_pagequeue_cnt_inc(pq);
2080 vm_pagequeue_unlock(pq);
2086 * Move the given page to the tail of its current page queue.
2088 * The page must be locked.
2091 vm_page_requeue(vm_page_t m)
2093 struct vm_pagequeue *pq;
2095 vm_page_lock_assert(m, MA_OWNED);
2096 KASSERT(m->queue != PQ_NONE,
2097 ("vm_page_requeue: page %p is not queued", m));
2098 pq = vm_page_pagequeue(m);
2099 vm_pagequeue_lock(pq);
2100 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2101 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2102 vm_pagequeue_unlock(pq);
2106 * vm_page_requeue_locked:
2108 * Move the given page to the tail of its current page queue.
2110 * The page queue must be locked.
2113 vm_page_requeue_locked(vm_page_t m)
2115 struct vm_pagequeue *pq;
2117 KASSERT(m->queue != PQ_NONE,
2118 ("vm_page_requeue_locked: page %p is not queued", m));
2119 pq = vm_page_pagequeue(m);
2120 vm_pagequeue_assert_locked(pq);
2121 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2122 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2128 * Put the specified page on the active list (if appropriate).
2129 * Ensure that act_count is at least ACT_INIT but do not otherwise
2132 * The page must be locked.
2135 vm_page_activate(vm_page_t m)
2139 vm_page_lock_assert(m, MA_OWNED);
2140 if ((queue = m->queue) != PQ_ACTIVE) {
2141 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2142 if (m->act_count < ACT_INIT)
2143 m->act_count = ACT_INIT;
2144 if (queue != PQ_NONE)
2146 vm_page_enqueue(PQ_ACTIVE, m);
2148 KASSERT(queue == PQ_NONE,
2149 ("vm_page_activate: wired page %p is queued", m));
2151 if (m->act_count < ACT_INIT)
2152 m->act_count = ACT_INIT;
2157 * vm_page_free_wakeup:
2159 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
2160 * routine is called when a page has been added to the cache or free
2163 * The page queues must be locked.
2166 vm_page_free_wakeup(void)
2169 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2171 * if pageout daemon needs pages, then tell it that there are
2174 if (vm_pageout_pages_needed &&
2175 vm_cnt.v_cache_count + vm_cnt.v_free_count >= vm_cnt.v_pageout_free_min) {
2176 wakeup(&vm_pageout_pages_needed);
2177 vm_pageout_pages_needed = 0;
2180 * wakeup processes that are waiting on memory if we hit a
2181 * high water mark. And wakeup scheduler process if we have
2182 * lots of memory. this process will swapin processes.
2184 if (vm_pages_needed && !vm_page_count_min()) {
2185 vm_pages_needed = 0;
2186 wakeup(&vm_cnt.v_free_count);
2191 * Turn a cached page into a free page, by changing its attributes.
2192 * Keep the statistics up-to-date.
2194 * The free page queue must be locked.
2197 vm_page_cache_turn_free(vm_page_t m)
2200 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2204 KASSERT((m->flags & PG_CACHED) != 0,
2205 ("vm_page_cache_turn_free: page %p is not cached", m));
2206 m->flags &= ~PG_CACHED;
2207 vm_cnt.v_cache_count--;
2208 vm_phys_freecnt_adj(m, 1);
2214 * Returns the given page to the free list,
2215 * disassociating it with any VM object.
2217 * The object must be locked. The page must be locked if it is managed.
2220 vm_page_free_toq(vm_page_t m)
2223 if ((m->oflags & VPO_UNMANAGED) == 0) {
2224 vm_page_lock_assert(m, MA_OWNED);
2225 KASSERT(!pmap_page_is_mapped(m),
2226 ("vm_page_free_toq: freeing mapped page %p", m));
2228 KASSERT(m->queue == PQ_NONE,
2229 ("vm_page_free_toq: unmanaged page %p is queued", m));
2230 PCPU_INC(cnt.v_tfree);
2232 if (vm_page_sbusied(m))
2233 panic("vm_page_free: freeing busy page %p", m);
2236 * Unqueue, then remove page. Note that we cannot destroy
2237 * the page here because we do not want to call the pager's
2238 * callback routine until after we've put the page on the
2239 * appropriate free queue.
2245 * If fictitious remove object association and
2246 * return, otherwise delay object association removal.
2248 if ((m->flags & PG_FICTITIOUS) != 0) {
2255 if (m->wire_count != 0)
2256 panic("vm_page_free: freeing wired page %p", m);
2257 if (m->hold_count != 0) {
2258 m->flags &= ~PG_ZERO;
2259 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2260 ("vm_page_free: freeing PG_UNHOLDFREE page %p", m));
2261 m->flags |= PG_UNHOLDFREE;
2264 * Restore the default memory attribute to the page.
2266 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2267 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2270 * Insert the page into the physical memory allocator's
2271 * cache/free page queues.
2273 mtx_lock(&vm_page_queue_free_mtx);
2274 vm_phys_freecnt_adj(m, 1);
2275 #if VM_NRESERVLEVEL > 0
2276 if (!vm_reserv_free_page(m))
2280 vm_phys_free_pages(m, 0);
2281 if ((m->flags & PG_ZERO) != 0)
2282 ++vm_page_zero_count;
2284 vm_page_zero_idle_wakeup();
2285 vm_page_free_wakeup();
2286 mtx_unlock(&vm_page_queue_free_mtx);
2293 * Mark this page as wired down by yet
2294 * another map, removing it from paging queues
2297 * If the page is fictitious, then its wire count must remain one.
2299 * The page must be locked.
2302 vm_page_wire(vm_page_t m)
2306 * Only bump the wire statistics if the page is not already wired,
2307 * and only unqueue the page if it is on some queue (if it is unmanaged
2308 * it is already off the queues).
2310 vm_page_lock_assert(m, MA_OWNED);
2311 if ((m->flags & PG_FICTITIOUS) != 0) {
2312 KASSERT(m->wire_count == 1,
2313 ("vm_page_wire: fictitious page %p's wire count isn't one",
2317 if (m->wire_count == 0) {
2318 KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
2319 m->queue == PQ_NONE,
2320 ("vm_page_wire: unmanaged page %p is queued", m));
2322 atomic_add_int(&vm_cnt.v_wire_count, 1);
2325 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
2331 * Release one wiring of the specified page, potentially enabling it to be
2332 * paged again. If paging is enabled, then the value of the parameter
2333 * "activate" determines to which queue the page is added. If "activate" is
2334 * non-zero, then the page is added to the active queue. Otherwise, it is
2335 * added to the inactive queue.
2337 * However, unless the page belongs to an object, it is not enqueued because
2338 * it cannot be paged out.
2340 * If a page is fictitious, then its wire count must always be one.
2342 * A managed page must be locked.
2345 vm_page_unwire(vm_page_t m, int activate)
2348 if ((m->oflags & VPO_UNMANAGED) == 0)
2349 vm_page_lock_assert(m, MA_OWNED);
2350 if ((m->flags & PG_FICTITIOUS) != 0) {
2351 KASSERT(m->wire_count == 1,
2352 ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
2355 if (m->wire_count > 0) {
2357 if (m->wire_count == 0) {
2358 atomic_subtract_int(&vm_cnt.v_wire_count, 1);
2359 if ((m->oflags & VPO_UNMANAGED) != 0 ||
2363 m->flags &= ~PG_WINATCFLS;
2364 vm_page_enqueue(activate ? PQ_ACTIVE : PQ_INACTIVE, m);
2367 panic("vm_page_unwire: page %p's wire count is zero", m);
2371 * Move the specified page to the inactive queue.
2373 * Many pages placed on the inactive queue should actually go
2374 * into the cache, but it is difficult to figure out which. What
2375 * we do instead, if the inactive target is well met, is to put
2376 * clean pages at the head of the inactive queue instead of the tail.
2377 * This will cause them to be moved to the cache more quickly and
2378 * if not actively re-referenced, reclaimed more quickly. If we just
2379 * stick these pages at the end of the inactive queue, heavy filesystem
2380 * meta-data accesses can cause an unnecessary paging load on memory bound
2381 * processes. This optimization causes one-time-use metadata to be
2382 * reused more quickly.
2384 * Normally athead is 0 resulting in LRU operation. athead is set
2385 * to 1 if we want this page to be 'as if it were placed in the cache',
2386 * except without unmapping it from the process address space.
2388 * The page must be locked.
2391 _vm_page_deactivate(vm_page_t m, int athead)
2393 struct vm_pagequeue *pq;
2396 vm_page_lock_assert(m, MA_OWNED);
2399 * Ignore if already inactive.
2401 if ((queue = m->queue) == PQ_INACTIVE)
2403 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2404 if (queue != PQ_NONE)
2406 m->flags &= ~PG_WINATCFLS;
2407 pq = &vm_phys_domain(m)->vmd_pagequeues[PQ_INACTIVE];
2408 vm_pagequeue_lock(pq);
2409 m->queue = PQ_INACTIVE;
2411 TAILQ_INSERT_HEAD(&pq->pq_pl, m, plinks.q);
2413 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2414 vm_pagequeue_cnt_inc(pq);
2415 vm_pagequeue_unlock(pq);
2420 * Move the specified page to the inactive queue.
2422 * The page must be locked.
2425 vm_page_deactivate(vm_page_t m)
2428 _vm_page_deactivate(m, 0);
2432 * vm_page_try_to_cache:
2434 * Returns 0 on failure, 1 on success
2437 vm_page_try_to_cache(vm_page_t m)
2440 vm_page_lock_assert(m, MA_OWNED);
2441 VM_OBJECT_ASSERT_WLOCKED(m->object);
2442 if (m->dirty || m->hold_count || m->wire_count ||
2443 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m))
2453 * vm_page_try_to_free()
2455 * Attempt to free the page. If we cannot free it, we do nothing.
2456 * 1 is returned on success, 0 on failure.
2459 vm_page_try_to_free(vm_page_t m)
2462 vm_page_lock_assert(m, MA_OWNED);
2463 if (m->object != NULL)
2464 VM_OBJECT_ASSERT_WLOCKED(m->object);
2465 if (m->dirty || m->hold_count || m->wire_count ||
2466 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m))
2478 * Put the specified page onto the page cache queue (if appropriate).
2480 * The object and page must be locked.
2483 vm_page_cache(vm_page_t m)
2486 boolean_t cache_was_empty;
2488 vm_page_lock_assert(m, MA_OWNED);
2490 VM_OBJECT_ASSERT_WLOCKED(object);
2491 if (vm_page_busied(m) || (m->oflags & VPO_UNMANAGED) ||
2492 m->hold_count || m->wire_count)
2493 panic("vm_page_cache: attempting to cache busy page");
2494 KASSERT(!pmap_page_is_mapped(m),
2495 ("vm_page_cache: page %p is mapped", m));
2496 KASSERT(m->dirty == 0, ("vm_page_cache: page %p is dirty", m));
2497 if (m->valid == 0 || object->type == OBJT_DEFAULT ||
2498 (object->type == OBJT_SWAP &&
2499 !vm_pager_has_page(object, m->pindex, NULL, NULL))) {
2501 * Hypothesis: A cache-elgible page belonging to a
2502 * default object or swap object but without a backing
2503 * store must be zero filled.
2508 KASSERT((m->flags & PG_CACHED) == 0,
2509 ("vm_page_cache: page %p is already cached", m));
2512 * Remove the page from the paging queues.
2517 * Remove the page from the object's collection of resident
2520 vm_radix_remove(&object->rtree, m->pindex);
2521 TAILQ_REMOVE(&object->memq, m, listq);
2522 object->resident_page_count--;
2525 * Restore the default memory attribute to the page.
2527 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2528 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2531 * Insert the page into the object's collection of cached pages
2532 * and the physical memory allocator's cache/free page queues.
2534 m->flags &= ~PG_ZERO;
2535 mtx_lock(&vm_page_queue_free_mtx);
2536 cache_was_empty = vm_radix_is_empty(&object->cache);
2537 if (vm_radix_insert(&object->cache, m)) {
2538 mtx_unlock(&vm_page_queue_free_mtx);
2539 if (object->resident_page_count == 0)
2540 vdrop(object->handle);
2547 * The above call to vm_radix_insert() could reclaim the one pre-
2548 * existing cached page from this object, resulting in a call to
2551 if (!cache_was_empty)
2552 cache_was_empty = vm_radix_is_singleton(&object->cache);
2554 m->flags |= PG_CACHED;
2555 vm_cnt.v_cache_count++;
2556 PCPU_INC(cnt.v_tcached);
2557 #if VM_NRESERVLEVEL > 0
2558 if (!vm_reserv_free_page(m)) {
2562 vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0);
2563 vm_phys_free_pages(m, 0);
2565 vm_page_free_wakeup();
2566 mtx_unlock(&vm_page_queue_free_mtx);
2569 * Increment the vnode's hold count if this is the object's only
2570 * cached page. Decrement the vnode's hold count if this was
2571 * the object's only resident page.
2573 if (object->type == OBJT_VNODE) {
2574 if (cache_was_empty && object->resident_page_count != 0)
2575 vhold(object->handle);
2576 else if (!cache_was_empty && object->resident_page_count == 0)
2577 vdrop(object->handle);
2584 * Cache, deactivate, or do nothing as appropriate. This routine
2585 * is used by madvise().
2587 * Generally speaking we want to move the page into the cache so
2588 * it gets reused quickly. However, this can result in a silly syndrome
2589 * due to the page recycling too quickly. Small objects will not be
2590 * fully cached. On the other hand, if we move the page to the inactive
2591 * queue we wind up with a problem whereby very large objects
2592 * unnecessarily blow away our inactive and cache queues.
2594 * The solution is to move the pages based on a fixed weighting. We
2595 * either leave them alone, deactivate them, or move them to the cache,
2596 * where moving them to the cache has the highest weighting.
2597 * By forcing some pages into other queues we eventually force the
2598 * system to balance the queues, potentially recovering other unrelated
2599 * space from active. The idea is to not force this to happen too
2602 * The object and page must be locked.
2605 vm_page_advise(vm_page_t m, int advice)
2609 vm_page_assert_locked(m);
2610 VM_OBJECT_ASSERT_WLOCKED(m->object);
2611 if (advice == MADV_FREE) {
2613 * Mark the page clean. This will allow the page to be freed
2614 * up by the system. However, such pages are often reused
2615 * quickly by malloc() so we do not do anything that would
2616 * cause a page fault if we can help it.
2618 * Specifically, we do not try to actually free the page now
2619 * nor do we try to put it in the cache (which would cause a
2620 * page fault on reuse).
2622 * But we do make the page is freeable as we can without
2623 * actually taking the step of unmapping it.
2627 } else if (advice != MADV_DONTNEED)
2629 dnw = PCPU_GET(dnweight);
2633 * Occasionally leave the page alone.
2635 if ((dnw & 0x01F0) == 0 || m->queue == PQ_INACTIVE) {
2636 if (m->act_count >= ACT_INIT)
2642 * Clear any references to the page. Otherwise, the page daemon will
2643 * immediately reactivate the page.
2645 vm_page_aflag_clear(m, PGA_REFERENCED);
2647 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
2650 if (m->dirty || (dnw & 0x0070) == 0) {
2652 * Deactivate the page 3 times out of 32.
2657 * Cache the page 28 times out of every 32. Note that
2658 * the page is deactivated instead of cached, but placed
2659 * at the head of the queue instead of the tail.
2663 _vm_page_deactivate(m, head);
2667 * Grab a page, waiting until we are waken up due to the page
2668 * changing state. We keep on waiting, if the page continues
2669 * to be in the object. If the page doesn't exist, first allocate it
2670 * and then conditionally zero it.
2672 * This routine may sleep.
2674 * The object must be locked on entry. The lock will, however, be released
2675 * and reacquired if the routine sleeps.
2678 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2683 VM_OBJECT_ASSERT_WLOCKED(object);
2684 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
2685 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
2686 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
2688 if ((m = vm_page_lookup(object, pindex)) != NULL) {
2689 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
2690 vm_page_xbusied(m) : vm_page_busied(m);
2693 * Reference the page before unlocking and
2694 * sleeping so that the page daemon is less
2695 * likely to reclaim it.
2697 vm_page_aflag_set(m, PGA_REFERENCED);
2699 VM_OBJECT_WUNLOCK(object);
2700 vm_page_busy_sleep(m, "pgrbwt");
2701 VM_OBJECT_WLOCK(object);
2704 if ((allocflags & VM_ALLOC_WIRED) != 0) {
2710 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
2712 if ((allocflags & VM_ALLOC_SBUSY) != 0)
2717 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_IGN_SBUSY);
2719 VM_OBJECT_WUNLOCK(object);
2721 VM_OBJECT_WLOCK(object);
2723 } else if (m->valid != 0)
2725 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
2731 * Mapping function for valid or dirty bits in a page.
2733 * Inputs are required to range within a page.
2736 vm_page_bits(int base, int size)
2742 base + size <= PAGE_SIZE,
2743 ("vm_page_bits: illegal base/size %d/%d", base, size)
2746 if (size == 0) /* handle degenerate case */
2749 first_bit = base >> DEV_BSHIFT;
2750 last_bit = (base + size - 1) >> DEV_BSHIFT;
2752 return (((vm_page_bits_t)2 << last_bit) -
2753 ((vm_page_bits_t)1 << first_bit));
2757 * vm_page_set_valid_range:
2759 * Sets portions of a page valid. The arguments are expected
2760 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2761 * of any partial chunks touched by the range. The invalid portion of
2762 * such chunks will be zeroed.
2764 * (base + size) must be less then or equal to PAGE_SIZE.
2767 vm_page_set_valid_range(vm_page_t m, int base, int size)
2771 VM_OBJECT_ASSERT_WLOCKED(m->object);
2772 if (size == 0) /* handle degenerate case */
2776 * If the base is not DEV_BSIZE aligned and the valid
2777 * bit is clear, we have to zero out a portion of the
2780 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2781 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
2782 pmap_zero_page_area(m, frag, base - frag);
2785 * If the ending offset is not DEV_BSIZE aligned and the
2786 * valid bit is clear, we have to zero out a portion of
2789 endoff = base + size;
2790 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2791 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
2792 pmap_zero_page_area(m, endoff,
2793 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
2796 * Assert that no previously invalid block that is now being validated
2799 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
2800 ("vm_page_set_valid_range: page %p is dirty", m));
2803 * Set valid bits inclusive of any overlap.
2805 m->valid |= vm_page_bits(base, size);
2809 * Clear the given bits from the specified page's dirty field.
2811 static __inline void
2812 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
2815 #if PAGE_SIZE < 16384
2820 * If the object is locked and the page is neither exclusive busy nor
2821 * write mapped, then the page's dirty field cannot possibly be
2822 * set by a concurrent pmap operation.
2824 VM_OBJECT_ASSERT_WLOCKED(m->object);
2825 if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
2826 m->dirty &= ~pagebits;
2829 * The pmap layer can call vm_page_dirty() without
2830 * holding a distinguished lock. The combination of
2831 * the object's lock and an atomic operation suffice
2832 * to guarantee consistency of the page dirty field.
2834 * For PAGE_SIZE == 32768 case, compiler already
2835 * properly aligns the dirty field, so no forcible
2836 * alignment is needed. Only require existence of
2837 * atomic_clear_64 when page size is 32768.
2839 addr = (uintptr_t)&m->dirty;
2840 #if PAGE_SIZE == 32768
2841 atomic_clear_64((uint64_t *)addr, pagebits);
2842 #elif PAGE_SIZE == 16384
2843 atomic_clear_32((uint32_t *)addr, pagebits);
2844 #else /* PAGE_SIZE <= 8192 */
2846 * Use a trick to perform a 32-bit atomic on the
2847 * containing aligned word, to not depend on the existence
2848 * of atomic_clear_{8, 16}.
2850 shift = addr & (sizeof(uint32_t) - 1);
2851 #if BYTE_ORDER == BIG_ENDIAN
2852 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
2856 addr &= ~(sizeof(uint32_t) - 1);
2857 atomic_clear_32((uint32_t *)addr, pagebits << shift);
2858 #endif /* PAGE_SIZE */
2863 * vm_page_set_validclean:
2865 * Sets portions of a page valid and clean. The arguments are expected
2866 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2867 * of any partial chunks touched by the range. The invalid portion of
2868 * such chunks will be zero'd.
2870 * (base + size) must be less then or equal to PAGE_SIZE.
2873 vm_page_set_validclean(vm_page_t m, int base, int size)
2875 vm_page_bits_t oldvalid, pagebits;
2878 VM_OBJECT_ASSERT_WLOCKED(m->object);
2879 if (size == 0) /* handle degenerate case */
2883 * If the base is not DEV_BSIZE aligned and the valid
2884 * bit is clear, we have to zero out a portion of the
2887 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2888 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
2889 pmap_zero_page_area(m, frag, base - frag);
2892 * If the ending offset is not DEV_BSIZE aligned and the
2893 * valid bit is clear, we have to zero out a portion of
2896 endoff = base + size;
2897 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2898 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
2899 pmap_zero_page_area(m, endoff,
2900 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
2903 * Set valid, clear dirty bits. If validating the entire
2904 * page we can safely clear the pmap modify bit. We also
2905 * use this opportunity to clear the VPO_NOSYNC flag. If a process
2906 * takes a write fault on a MAP_NOSYNC memory area the flag will
2909 * We set valid bits inclusive of any overlap, but we can only
2910 * clear dirty bits for DEV_BSIZE chunks that are fully within
2913 oldvalid = m->valid;
2914 pagebits = vm_page_bits(base, size);
2915 m->valid |= pagebits;
2917 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
2918 frag = DEV_BSIZE - frag;
2924 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
2926 if (base == 0 && size == PAGE_SIZE) {
2928 * The page can only be modified within the pmap if it is
2929 * mapped, and it can only be mapped if it was previously
2932 if (oldvalid == VM_PAGE_BITS_ALL)
2934 * Perform the pmap_clear_modify() first. Otherwise,
2935 * a concurrent pmap operation, such as
2936 * pmap_protect(), could clear a modification in the
2937 * pmap and set the dirty field on the page before
2938 * pmap_clear_modify() had begun and after the dirty
2939 * field was cleared here.
2941 pmap_clear_modify(m);
2943 m->oflags &= ~VPO_NOSYNC;
2944 } else if (oldvalid != VM_PAGE_BITS_ALL)
2945 m->dirty &= ~pagebits;
2947 vm_page_clear_dirty_mask(m, pagebits);
2951 vm_page_clear_dirty(vm_page_t m, int base, int size)
2954 vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
2958 * vm_page_set_invalid:
2960 * Invalidates DEV_BSIZE'd chunks within a page. Both the
2961 * valid and dirty bits for the effected areas are cleared.
2964 vm_page_set_invalid(vm_page_t m, int base, int size)
2966 vm_page_bits_t bits;
2970 VM_OBJECT_ASSERT_WLOCKED(object);
2971 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
2972 size >= object->un_pager.vnp.vnp_size)
2973 bits = VM_PAGE_BITS_ALL;
2975 bits = vm_page_bits(base, size);
2976 if (m->valid == VM_PAGE_BITS_ALL && bits != 0)
2978 KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) ||
2979 !pmap_page_is_mapped(m),
2980 ("vm_page_set_invalid: page %p is mapped", m));
2986 * vm_page_zero_invalid()
2988 * The kernel assumes that the invalid portions of a page contain
2989 * garbage, but such pages can be mapped into memory by user code.
2990 * When this occurs, we must zero out the non-valid portions of the
2991 * page so user code sees what it expects.
2993 * Pages are most often semi-valid when the end of a file is mapped
2994 * into memory and the file's size is not page aligned.
2997 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
3002 VM_OBJECT_ASSERT_WLOCKED(m->object);
3004 * Scan the valid bits looking for invalid sections that
3005 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
3006 * valid bit may be set ) have already been zerod by
3007 * vm_page_set_validclean().
3009 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
3010 if (i == (PAGE_SIZE / DEV_BSIZE) ||
3011 (m->valid & ((vm_page_bits_t)1 << i))) {
3013 pmap_zero_page_area(m,
3014 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
3021 * setvalid is TRUE when we can safely set the zero'd areas
3022 * as being valid. We can do this if there are no cache consistancy
3023 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
3026 m->valid = VM_PAGE_BITS_ALL;
3032 * Is (partial) page valid? Note that the case where size == 0
3033 * will return FALSE in the degenerate case where the page is
3034 * entirely invalid, and TRUE otherwise.
3037 vm_page_is_valid(vm_page_t m, int base, int size)
3039 vm_page_bits_t bits;
3041 VM_OBJECT_ASSERT_LOCKED(m->object);
3042 bits = vm_page_bits(base, size);
3043 return (m->valid != 0 && (m->valid & bits) == bits);
3047 * Set the page's dirty bits if the page is modified.
3050 vm_page_test_dirty(vm_page_t m)
3053 VM_OBJECT_ASSERT_WLOCKED(m->object);
3054 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
3059 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
3062 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
3066 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
3069 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
3073 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
3076 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
3079 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
3081 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
3084 vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
3088 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
3091 mtx_assert_(vm_page_lockptr(m), a, file, line);
3097 vm_page_object_lock_assert(vm_page_t m)
3101 * Certain of the page's fields may only be modified by the
3102 * holder of the containing object's lock or the exclusive busy.
3103 * holder. Unfortunately, the holder of the write busy is
3104 * not recorded, and thus cannot be checked here.
3106 if (m->object != NULL && !vm_page_xbusied(m))
3107 VM_OBJECT_ASSERT_WLOCKED(m->object);
3111 #include "opt_ddb.h"
3113 #include <sys/kernel.h>
3115 #include <ddb/ddb.h>
3117 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3119 db_printf("vm_cnt.v_free_count: %d\n", vm_cnt.v_free_count);
3120 db_printf("vm_cnt.v_cache_count: %d\n", vm_cnt.v_cache_count);
3121 db_printf("vm_cnt.v_inactive_count: %d\n", vm_cnt.v_inactive_count);
3122 db_printf("vm_cnt.v_active_count: %d\n", vm_cnt.v_active_count);
3123 db_printf("vm_cnt.v_wire_count: %d\n", vm_cnt.v_wire_count);
3124 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
3125 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
3126 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
3127 db_printf("vm_cnt.v_cache_min: %d\n", vm_cnt.v_cache_min);
3128 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
3131 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3135 db_printf("pq_free %d pq_cache %d\n",
3136 vm_cnt.v_free_count, vm_cnt.v_cache_count);
3137 for (dom = 0; dom < vm_ndomains; dom++) {
3139 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pass %d\n",
3141 vm_dom[dom].vmd_page_count,
3142 vm_dom[dom].vmd_free_count,
3143 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
3144 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
3145 vm_dom[dom].vmd_pass);
3149 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
3155 db_printf("show pginfo addr\n");
3159 phys = strchr(modif, 'p') != NULL;
3161 m = PHYS_TO_VM_PAGE(addr);
3163 m = (vm_page_t)addr;
3165 "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n"
3166 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
3167 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
3168 m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags,
3169 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty);