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 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
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(uint8_t 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 vm_paddr_t biggestsize;
296 vm_paddr_t low_water, high_water;
301 vaddr = round_page(vaddr);
303 for (i = 0; phys_avail[i + 1]; i += 2) {
304 phys_avail[i] = round_page(phys_avail[i]);
305 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
310 * There is no obvious reason why i386 PV Xen needs vm_page structs
311 * created for these pseudo-physical addresses. XXX
313 vm_phys_add_seg(0, phys_avail[0]);
316 low_water = phys_avail[0];
317 high_water = phys_avail[1];
319 for (i = 0; i < vm_phys_nsegs; i++) {
320 if (vm_phys_segs[i].start < low_water)
321 low_water = vm_phys_segs[i].start;
322 if (vm_phys_segs[i].end > high_water)
323 high_water = vm_phys_segs[i].end;
325 for (i = 0; phys_avail[i + 1]; i += 2) {
326 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
328 if (size > biggestsize) {
332 if (phys_avail[i] < low_water)
333 low_water = phys_avail[i];
334 if (phys_avail[i + 1] > high_water)
335 high_water = phys_avail[i + 1];
338 end = phys_avail[biggestone+1];
341 * Initialize the page and queue locks.
343 mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF);
344 for (i = 0; i < PA_LOCK_COUNT; i++)
345 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
346 for (i = 0; i < vm_ndomains; i++)
347 vm_page_domain_init(&vm_dom[i]);
350 * Allocate memory for use when boot strapping the kernel memory
353 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
354 * manually fetch the value.
356 TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
357 new_end = end - (boot_pages * UMA_SLAB_SIZE);
358 new_end = trunc_page(new_end);
359 mapped = pmap_map(&vaddr, new_end, end,
360 VM_PROT_READ | VM_PROT_WRITE);
361 bzero((void *)mapped, end - new_end);
362 uma_startup((void *)mapped, boot_pages);
364 #if defined(__amd64__) || defined(__i386__) || defined(__arm__) || \
367 * Allocate a bitmap to indicate that a random physical page
368 * needs to be included in a minidump.
370 * The amd64 port needs this to indicate which direct map pages
371 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
373 * However, i386 still needs this workspace internally within the
374 * minidump code. In theory, they are not needed on i386, but are
375 * included should the sf_buf code decide to use them.
378 for (i = 0; dump_avail[i + 1] != 0; i += 2)
379 if (dump_avail[i + 1] > last_pa)
380 last_pa = dump_avail[i + 1];
381 page_range = last_pa / PAGE_SIZE;
382 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
383 new_end -= vm_page_dump_size;
384 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
385 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
386 bzero((void *)vm_page_dump, vm_page_dump_size);
390 * Request that the physical pages underlying the message buffer be
391 * included in a crash dump. Since the message buffer is accessed
392 * through the direct map, they are not automatically included.
394 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
395 last_pa = pa + round_page(msgbufsize);
396 while (pa < last_pa) {
402 * Compute the number of pages of memory that will be available for
403 * use (taking into account the overhead of a page structure per
406 first_page = low_water / PAGE_SIZE;
407 #ifdef VM_PHYSSEG_SPARSE
409 for (i = 0; i < vm_phys_nsegs; i++) {
410 page_range += atop(vm_phys_segs[i].end -
411 vm_phys_segs[i].start);
413 for (i = 0; phys_avail[i + 1] != 0; i += 2)
414 page_range += atop(phys_avail[i + 1] - phys_avail[i]);
415 #elif defined(VM_PHYSSEG_DENSE)
416 page_range = high_water / PAGE_SIZE - first_page;
418 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
423 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
428 * Initialize the mem entry structures now, and put them in the free
431 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
432 mapped = pmap_map(&vaddr, new_end, end,
433 VM_PROT_READ | VM_PROT_WRITE);
434 vm_page_array = (vm_page_t) mapped;
435 #if VM_NRESERVLEVEL > 0
437 * Allocate memory for the reservation management system's data
440 new_end = vm_reserv_startup(&vaddr, new_end, high_water);
442 #if defined(__amd64__) || defined(__mips__)
444 * pmap_map on amd64 and mips can come out of the direct-map, not kvm
445 * like i386, so the pages must be tracked for a crashdump to include
446 * this data. This includes the vm_page_array and the early UMA
449 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
452 phys_avail[biggestone + 1] = new_end;
455 * Add physical memory segments corresponding to the available
458 for (i = 0; phys_avail[i + 1] != 0; i += 2)
459 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
462 * Clear all of the page structures
464 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
465 for (i = 0; i < page_range; i++)
466 vm_page_array[i].order = VM_NFREEORDER;
467 vm_page_array_size = page_range;
470 * Initialize the physical memory allocator.
475 * Add every available physical page that is not blacklisted to
478 vm_cnt.v_page_count = 0;
479 vm_cnt.v_free_count = 0;
480 list = kern_getenv("vm.blacklist");
481 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
483 last_pa = phys_avail[i + 1];
484 while (pa < last_pa) {
486 vm_page_blacklist_lookup(list, pa))
487 printf("Skipping page with pa 0x%jx\n",
490 vm_phys_add_page(pa);
495 #if VM_NRESERVLEVEL > 0
497 * Initialize the reservation management system.
505 vm_page_reference(vm_page_t m)
508 vm_page_aflag_set(m, PGA_REFERENCED);
512 * vm_page_busy_downgrade:
514 * Downgrade an exclusive busy page into a single shared busy page.
517 vm_page_busy_downgrade(vm_page_t m)
521 vm_page_assert_xbusied(m);
525 x &= VPB_BIT_WAITERS;
526 if (atomic_cmpset_rel_int(&m->busy_lock,
527 VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1) | x))
535 * Return a positive value if the page is shared busied, 0 otherwise.
538 vm_page_sbusied(vm_page_t m)
543 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
549 * Shared unbusy a page.
552 vm_page_sunbusy(vm_page_t m)
556 vm_page_assert_sbusied(m);
560 if (VPB_SHARERS(x) > 1) {
561 if (atomic_cmpset_int(&m->busy_lock, x,
566 if ((x & VPB_BIT_WAITERS) == 0) {
567 KASSERT(x == VPB_SHARERS_WORD(1),
568 ("vm_page_sunbusy: invalid lock state"));
569 if (atomic_cmpset_int(&m->busy_lock,
570 VPB_SHARERS_WORD(1), VPB_UNBUSIED))
574 KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS),
575 ("vm_page_sunbusy: invalid lock state for waiters"));
578 if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) {
589 * vm_page_busy_sleep:
591 * Sleep and release the page lock, using the page pointer as wchan.
592 * This is used to implement the hard-path of busying mechanism.
594 * The given page must be locked.
597 vm_page_busy_sleep(vm_page_t m, const char *wmesg)
601 vm_page_lock_assert(m, MA_OWNED);
604 if (x == VPB_UNBUSIED) {
608 if ((x & VPB_BIT_WAITERS) == 0 &&
609 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS)) {
613 msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0);
619 * Try to shared busy a page.
620 * If the operation succeeds 1 is returned otherwise 0.
621 * The operation never sleeps.
624 vm_page_trysbusy(vm_page_t m)
630 if ((x & VPB_BIT_SHARED) == 0)
632 if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER))
638 * vm_page_xunbusy_hard:
640 * Called after the first try the exclusive unbusy of a page failed.
641 * It is assumed that the waiters bit is on.
644 vm_page_xunbusy_hard(vm_page_t m)
647 vm_page_assert_xbusied(m);
650 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
658 * Wakeup anyone waiting for the page.
659 * The ownership bits do not change.
661 * The given page must be locked.
664 vm_page_flash(vm_page_t m)
668 vm_page_lock_assert(m, MA_OWNED);
672 if ((x & VPB_BIT_WAITERS) == 0)
674 if (atomic_cmpset_int(&m->busy_lock, x,
675 x & (~VPB_BIT_WAITERS)))
682 * Keep page from being freed by the page daemon
683 * much of the same effect as wiring, except much lower
684 * overhead and should be used only for *very* temporary
685 * holding ("wiring").
688 vm_page_hold(vm_page_t mem)
691 vm_page_lock_assert(mem, MA_OWNED);
696 vm_page_unhold(vm_page_t mem)
699 vm_page_lock_assert(mem, MA_OWNED);
700 KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!"));
702 if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0)
703 vm_page_free_toq(mem);
707 * vm_page_unhold_pages:
709 * Unhold each of the pages that is referenced by the given array.
712 vm_page_unhold_pages(vm_page_t *ma, int count)
714 struct mtx *mtx, *new_mtx;
717 for (; count != 0; count--) {
719 * Avoid releasing and reacquiring the same page lock.
721 new_mtx = vm_page_lockptr(*ma);
722 if (mtx != new_mtx) {
736 PHYS_TO_VM_PAGE(vm_paddr_t pa)
740 #ifdef VM_PHYSSEG_SPARSE
741 m = vm_phys_paddr_to_vm_page(pa);
743 m = vm_phys_fictitious_to_vm_page(pa);
745 #elif defined(VM_PHYSSEG_DENSE)
749 if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
750 m = &vm_page_array[pi - first_page];
753 return (vm_phys_fictitious_to_vm_page(pa));
755 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
762 * Create a fictitious page with the specified physical address and
763 * memory attribute. The memory attribute is the only the machine-
764 * dependent aspect of a fictitious page that must be initialized.
767 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
771 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
772 vm_page_initfake(m, paddr, memattr);
777 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
780 if ((m->flags & PG_FICTITIOUS) != 0) {
782 * The page's memattr might have changed since the
783 * previous initialization. Update the pmap to the
788 m->phys_addr = paddr;
790 /* Fictitious pages don't use "segind". */
791 m->flags = PG_FICTITIOUS;
792 /* Fictitious pages don't use "order" or "pool". */
793 m->oflags = VPO_UNMANAGED;
794 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
798 pmap_page_set_memattr(m, memattr);
804 * Release a fictitious page.
807 vm_page_putfake(vm_page_t m)
810 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
811 KASSERT((m->flags & PG_FICTITIOUS) != 0,
812 ("vm_page_putfake: bad page %p", m));
813 uma_zfree(fakepg_zone, m);
817 * vm_page_updatefake:
819 * Update the given fictitious page to the specified physical address and
823 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
826 KASSERT((m->flags & PG_FICTITIOUS) != 0,
827 ("vm_page_updatefake: bad page %p", m));
828 m->phys_addr = paddr;
829 pmap_page_set_memattr(m, memattr);
838 vm_page_free(vm_page_t m)
841 m->flags &= ~PG_ZERO;
848 * Free a page to the zerod-pages queue
851 vm_page_free_zero(vm_page_t m)
859 * Unbusy and handle the page queueing for a page from the VOP_GETPAGES()
860 * array which is not the request page.
863 vm_page_readahead_finish(vm_page_t m)
868 * Since the page is not the requested page, whether
869 * it should be activated or deactivated is not
870 * obvious. Empirical results have shown that
871 * deactivating the page is usually the best choice,
872 * unless the page is wanted by another thread.
875 if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
878 vm_page_deactivate(m);
883 * Free the completely invalid page. Such page state
884 * occurs due to the short read operation which did
885 * not covered our page at all, or in case when a read
895 * vm_page_sleep_if_busy:
897 * Sleep and release the page queues lock if the page is busied.
898 * Returns TRUE if the thread slept.
900 * The given page must be unlocked and object containing it must
904 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
908 vm_page_lock_assert(m, MA_NOTOWNED);
909 VM_OBJECT_ASSERT_WLOCKED(m->object);
911 if (vm_page_busied(m)) {
913 * The page-specific object must be cached because page
914 * identity can change during the sleep, causing the
915 * re-lock of a different object.
916 * It is assumed that a reference to the object is already
917 * held by the callers.
921 VM_OBJECT_WUNLOCK(obj);
922 vm_page_busy_sleep(m, msg);
923 VM_OBJECT_WLOCK(obj);
930 * vm_page_dirty_KBI: [ internal use only ]
932 * Set all bits in the page's dirty field.
934 * The object containing the specified page must be locked if the
935 * call is made from the machine-independent layer.
937 * See vm_page_clear_dirty_mask().
939 * This function should only be called by vm_page_dirty().
942 vm_page_dirty_KBI(vm_page_t m)
945 /* These assertions refer to this operation by its public name. */
946 KASSERT((m->flags & PG_CACHED) == 0,
947 ("vm_page_dirty: page in cache!"));
948 KASSERT(m->valid == VM_PAGE_BITS_ALL,
949 ("vm_page_dirty: page is invalid!"));
950 m->dirty = VM_PAGE_BITS_ALL;
954 * vm_page_insert: [ internal use only ]
956 * Inserts the given mem entry into the object and object list.
958 * The object must be locked.
961 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
965 VM_OBJECT_ASSERT_WLOCKED(object);
966 mpred = vm_radix_lookup_le(&object->rtree, pindex);
967 return (vm_page_insert_after(m, object, pindex, mpred));
971 * vm_page_insert_after:
973 * Inserts the page "m" into the specified object at offset "pindex".
975 * The page "mpred" must immediately precede the offset "pindex" within
976 * the specified object.
978 * The object must be locked.
981 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
988 VM_OBJECT_ASSERT_WLOCKED(object);
989 KASSERT(m->object == NULL,
990 ("vm_page_insert_after: page already inserted"));
992 KASSERT(mpred->object == object,
993 ("vm_page_insert_after: object doesn't contain mpred"));
994 KASSERT(mpred->pindex < pindex,
995 ("vm_page_insert_after: mpred doesn't precede pindex"));
996 msucc = TAILQ_NEXT(mpred, listq);
998 msucc = TAILQ_FIRST(&object->memq);
1000 KASSERT(msucc->pindex > pindex,
1001 ("vm_page_insert_after: msucc doesn't succeed pindex"));
1004 * Record the object/offset pair in this page
1012 * Now link into the object's ordered list of backed pages.
1014 if (vm_radix_insert(&object->rtree, m)) {
1019 vm_page_insert_radixdone(m, object, mpred);
1024 * vm_page_insert_radixdone:
1026 * Complete page "m" insertion into the specified object after the
1027 * radix trie hooking.
1029 * The page "mpred" must precede the offset "m->pindex" within the
1032 * The object must be locked.
1035 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1038 VM_OBJECT_ASSERT_WLOCKED(object);
1039 KASSERT(object != NULL && m->object == object,
1040 ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1041 if (mpred != NULL) {
1042 KASSERT(mpred->object == object,
1043 ("vm_page_insert_after: object doesn't contain mpred"));
1044 KASSERT(mpred->pindex < m->pindex,
1045 ("vm_page_insert_after: mpred doesn't precede pindex"));
1049 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1051 TAILQ_INSERT_HEAD(&object->memq, m, listq);
1054 * Show that the object has one more resident page.
1056 object->resident_page_count++;
1059 * Hold the vnode until the last page is released.
1061 if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1062 vhold(object->handle);
1065 * Since we are inserting a new and possibly dirty page,
1066 * update the object's OBJ_MIGHTBEDIRTY flag.
1068 if (pmap_page_is_write_mapped(m))
1069 vm_object_set_writeable_dirty(object);
1075 * Removes the given mem entry from the object/offset-page
1076 * table and the object page list, but do not invalidate/terminate
1077 * the backing store.
1079 * The object must be locked. The page must be locked if it is managed.
1082 vm_page_remove(vm_page_t m)
1087 if ((m->oflags & VPO_UNMANAGED) == 0)
1088 vm_page_lock_assert(m, MA_OWNED);
1089 if ((object = m->object) == NULL)
1091 VM_OBJECT_ASSERT_WLOCKED(object);
1092 if (vm_page_xbusied(m)) {
1094 if ((m->oflags & VPO_UNMANAGED) != 0 &&
1095 !mtx_owned(vm_page_lockptr(m))) {
1100 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1106 * Now remove from the object's list of backed pages.
1108 vm_radix_remove(&object->rtree, m->pindex);
1109 TAILQ_REMOVE(&object->memq, m, listq);
1112 * And show that the object has one fewer resident page.
1114 object->resident_page_count--;
1117 * The vnode may now be recycled.
1119 if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1120 vdrop(object->handle);
1128 * Returns the page associated with the object/offset
1129 * pair specified; if none is found, NULL is returned.
1131 * The object must be locked.
1134 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1137 VM_OBJECT_ASSERT_LOCKED(object);
1138 return (vm_radix_lookup(&object->rtree, pindex));
1142 * vm_page_find_least:
1144 * Returns the page associated with the object with least pindex
1145 * greater than or equal to the parameter pindex, or NULL.
1147 * The object must be locked.
1150 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1154 VM_OBJECT_ASSERT_LOCKED(object);
1155 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1156 m = vm_radix_lookup_ge(&object->rtree, pindex);
1161 * Returns the given page's successor (by pindex) within the object if it is
1162 * resident; if none is found, NULL is returned.
1164 * The object must be locked.
1167 vm_page_next(vm_page_t m)
1171 VM_OBJECT_ASSERT_WLOCKED(m->object);
1172 if ((next = TAILQ_NEXT(m, listq)) != NULL &&
1173 next->pindex != m->pindex + 1)
1179 * Returns the given page's predecessor (by pindex) within the object if it is
1180 * resident; if none is found, NULL is returned.
1182 * The object must be locked.
1185 vm_page_prev(vm_page_t m)
1189 VM_OBJECT_ASSERT_WLOCKED(m->object);
1190 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
1191 prev->pindex != m->pindex - 1)
1197 * Uses the page mnew as a replacement for an existing page at index
1198 * pindex which must be already present in the object.
1200 * The existing page must not be on a paging queue.
1203 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1205 vm_page_t mold, mpred;
1207 VM_OBJECT_ASSERT_WLOCKED(object);
1210 * This function mostly follows vm_page_insert() and
1211 * vm_page_remove() without the radix, object count and vnode
1212 * dance. Double check such functions for more comments.
1214 mpred = vm_radix_lookup(&object->rtree, pindex);
1215 KASSERT(mpred != NULL,
1216 ("vm_page_replace: replacing page not present with pindex"));
1217 mpred = TAILQ_PREV(mpred, respgs, listq);
1219 KASSERT(mpred->pindex < pindex,
1220 ("vm_page_insert_after: mpred doesn't precede pindex"));
1222 mnew->object = object;
1223 mnew->pindex = pindex;
1224 mold = vm_radix_replace(&object->rtree, mnew);
1225 KASSERT(mold->queue == PQ_NONE,
1226 ("vm_page_replace: mold is on a paging queue"));
1228 /* Detach the old page from the resident tailq. */
1229 TAILQ_REMOVE(&object->memq, mold, listq);
1231 mold->object = NULL;
1232 vm_page_xunbusy(mold);
1234 /* Insert the new page in the resident tailq. */
1236 TAILQ_INSERT_AFTER(&object->memq, mpred, mnew, listq);
1238 TAILQ_INSERT_HEAD(&object->memq, mnew, listq);
1239 if (pmap_page_is_write_mapped(mnew))
1240 vm_object_set_writeable_dirty(object);
1247 * Move the given memory entry from its
1248 * current object to the specified target object/offset.
1250 * Note: swap associated with the page must be invalidated by the move. We
1251 * have to do this for several reasons: (1) we aren't freeing the
1252 * page, (2) we are dirtying the page, (3) the VM system is probably
1253 * moving the page from object A to B, and will then later move
1254 * the backing store from A to B and we can't have a conflict.
1256 * Note: we *always* dirty the page. It is necessary both for the
1257 * fact that we moved it, and because we may be invalidating
1258 * swap. If the page is on the cache, we have to deactivate it
1259 * or vm_page_dirty() will panic. Dirty pages are not allowed
1262 * The objects must be locked.
1265 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1270 VM_OBJECT_ASSERT_WLOCKED(new_object);
1272 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1273 KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1274 ("vm_page_rename: pindex already renamed"));
1277 * Create a custom version of vm_page_insert() which does not depend
1278 * by m_prev and can cheat on the implementation aspects of the
1282 m->pindex = new_pindex;
1283 if (vm_radix_insert(&new_object->rtree, m)) {
1289 * The operation cannot fail anymore. The removal must happen before
1290 * the listq iterator is tainted.
1296 /* Return back to the new pindex to complete vm_page_insert(). */
1297 m->pindex = new_pindex;
1298 m->object = new_object;
1300 vm_page_insert_radixdone(m, new_object, mpred);
1306 * Convert all of the given object's cached pages that have a
1307 * pindex within the given range into free pages. If the value
1308 * zero is given for "end", then the range's upper bound is
1309 * infinity. If the given object is backed by a vnode and it
1310 * transitions from having one or more cached pages to none, the
1311 * vnode's hold count is reduced.
1314 vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1319 mtx_lock(&vm_page_queue_free_mtx);
1320 if (__predict_false(vm_radix_is_empty(&object->cache))) {
1321 mtx_unlock(&vm_page_queue_free_mtx);
1324 while ((m = vm_radix_lookup_ge(&object->cache, start)) != NULL) {
1325 if (end != 0 && m->pindex >= end)
1327 vm_radix_remove(&object->cache, m->pindex);
1328 vm_page_cache_turn_free(m);
1330 empty = vm_radix_is_empty(&object->cache);
1331 mtx_unlock(&vm_page_queue_free_mtx);
1332 if (object->type == OBJT_VNODE && empty)
1333 vdrop(object->handle);
1337 * Returns the cached page that is associated with the given
1338 * object and offset. If, however, none exists, returns NULL.
1340 * The free page queue must be locked.
1342 static inline vm_page_t
1343 vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex)
1346 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1347 return (vm_radix_lookup(&object->cache, pindex));
1351 * Remove the given cached page from its containing object's
1352 * collection of cached pages.
1354 * The free page queue must be locked.
1357 vm_page_cache_remove(vm_page_t m)
1360 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1361 KASSERT((m->flags & PG_CACHED) != 0,
1362 ("vm_page_cache_remove: page %p is not cached", m));
1363 vm_radix_remove(&m->object->cache, m->pindex);
1365 vm_cnt.v_cache_count--;
1369 * Transfer all of the cached pages with offset greater than or
1370 * equal to 'offidxstart' from the original object's cache to the
1371 * new object's cache. However, any cached pages with offset
1372 * greater than or equal to the new object's size are kept in the
1373 * original object. Initially, the new object's cache must be
1374 * empty. Offset 'offidxstart' in the original object must
1375 * correspond to offset zero in the new object.
1377 * The new object must be locked.
1380 vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart,
1381 vm_object_t new_object)
1386 * Insertion into an object's collection of cached pages
1387 * requires the object to be locked. In contrast, removal does
1390 VM_OBJECT_ASSERT_WLOCKED(new_object);
1391 KASSERT(vm_radix_is_empty(&new_object->cache),
1392 ("vm_page_cache_transfer: object %p has cached pages",
1394 mtx_lock(&vm_page_queue_free_mtx);
1395 while ((m = vm_radix_lookup_ge(&orig_object->cache,
1396 offidxstart)) != NULL) {
1398 * Transfer all of the pages with offset greater than or
1399 * equal to 'offidxstart' from the original object's
1400 * cache to the new object's cache.
1402 if ((m->pindex - offidxstart) >= new_object->size)
1404 vm_radix_remove(&orig_object->cache, m->pindex);
1405 /* Update the page's object and offset. */
1406 m->object = new_object;
1407 m->pindex -= offidxstart;
1408 if (vm_radix_insert(&new_object->cache, m))
1409 vm_page_cache_turn_free(m);
1411 mtx_unlock(&vm_page_queue_free_mtx);
1415 * Returns TRUE if a cached page is associated with the given object and
1416 * offset, and FALSE otherwise.
1418 * The object must be locked.
1421 vm_page_is_cached(vm_object_t object, vm_pindex_t pindex)
1426 * Insertion into an object's collection of cached pages requires the
1427 * object to be locked. Therefore, if the object is locked and the
1428 * object's collection is empty, there is no need to acquire the free
1429 * page queues lock in order to prove that the specified page doesn't
1432 VM_OBJECT_ASSERT_WLOCKED(object);
1433 if (__predict_true(vm_object_cache_is_empty(object)))
1435 mtx_lock(&vm_page_queue_free_mtx);
1436 m = vm_page_cache_lookup(object, pindex);
1437 mtx_unlock(&vm_page_queue_free_mtx);
1444 * Allocate and return a page that is associated with the specified
1445 * object and offset pair. By default, this page is exclusive busied.
1447 * The caller must always specify an allocation class.
1449 * allocation classes:
1450 * VM_ALLOC_NORMAL normal process request
1451 * VM_ALLOC_SYSTEM system *really* needs a page
1452 * VM_ALLOC_INTERRUPT interrupt time request
1454 * optional allocation flags:
1455 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1456 * intends to allocate
1457 * VM_ALLOC_IFCACHED return page only if it is cached
1458 * VM_ALLOC_IFNOTCACHED return NULL, do not reactivate if the page
1460 * VM_ALLOC_NOBUSY do not exclusive busy the page
1461 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1462 * VM_ALLOC_NOOBJ page is not associated with an object and
1463 * should not be exclusive busy
1464 * VM_ALLOC_SBUSY shared busy the allocated page
1465 * VM_ALLOC_WIRED wire the allocated page
1466 * VM_ALLOC_ZERO prefer a zeroed page
1468 * This routine may not sleep.
1471 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1473 struct vnode *vp = NULL;
1474 vm_object_t m_object;
1476 int flags, req_class;
1478 mpred = 0; /* XXX: pacify gcc */
1479 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1480 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1481 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1482 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1483 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object,
1486 VM_OBJECT_ASSERT_WLOCKED(object);
1488 req_class = req & VM_ALLOC_CLASS_MASK;
1491 * The page daemon is allowed to dig deeper into the free page list.
1493 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1494 req_class = VM_ALLOC_SYSTEM;
1496 if (object != NULL) {
1497 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1498 KASSERT(mpred == NULL || mpred->pindex != pindex,
1499 ("vm_page_alloc: pindex already allocated"));
1503 * The page allocation request can came from consumers which already
1504 * hold the free page queue mutex, like vm_page_insert() in
1507 mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE);
1508 if (vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_free_reserved ||
1509 (req_class == VM_ALLOC_SYSTEM &&
1510 vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_interrupt_free_min) ||
1511 (req_class == VM_ALLOC_INTERRUPT &&
1512 vm_cnt.v_free_count + vm_cnt.v_cache_count > 0)) {
1514 * Allocate from the free queue if the number of free pages
1515 * exceeds the minimum for the request class.
1517 if (object != NULL &&
1518 (m = vm_page_cache_lookup(object, pindex)) != NULL) {
1519 if ((req & VM_ALLOC_IFNOTCACHED) != 0) {
1520 mtx_unlock(&vm_page_queue_free_mtx);
1523 if (vm_phys_unfree_page(m))
1524 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0);
1525 #if VM_NRESERVLEVEL > 0
1526 else if (!vm_reserv_reactivate_page(m))
1530 panic("vm_page_alloc: cache page %p is missing"
1531 " from the free queue", m);
1532 } else if ((req & VM_ALLOC_IFCACHED) != 0) {
1533 mtx_unlock(&vm_page_queue_free_mtx);
1535 #if VM_NRESERVLEVEL > 0
1536 } else if (object == NULL || (object->flags & (OBJ_COLORED |
1537 OBJ_FICTITIOUS)) != OBJ_COLORED || (m =
1538 vm_reserv_alloc_page(object, pindex, mpred)) == NULL) {
1542 m = vm_phys_alloc_pages(object != NULL ?
1543 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1544 #if VM_NRESERVLEVEL > 0
1545 if (m == NULL && vm_reserv_reclaim_inactive()) {
1546 m = vm_phys_alloc_pages(object != NULL ?
1547 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1554 * Not allocatable, give up.
1556 mtx_unlock(&vm_page_queue_free_mtx);
1557 atomic_add_int(&vm_pageout_deficit,
1558 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1559 pagedaemon_wakeup();
1564 * At this point we had better have found a good page.
1566 KASSERT(m != NULL, ("vm_page_alloc: missing page"));
1567 KASSERT(m->queue == PQ_NONE,
1568 ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue));
1569 KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m));
1570 KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m));
1571 KASSERT(!vm_page_sbusied(m),
1572 ("vm_page_alloc: page %p is busy", m));
1573 KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m));
1574 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1575 ("vm_page_alloc: page %p has unexpected memattr %d", m,
1576 pmap_page_get_memattr(m)));
1577 if ((m->flags & PG_CACHED) != 0) {
1578 KASSERT((m->flags & PG_ZERO) == 0,
1579 ("vm_page_alloc: cached page %p is PG_ZERO", m));
1580 KASSERT(m->valid != 0,
1581 ("vm_page_alloc: cached page %p is invalid", m));
1582 if (m->object == object && m->pindex == pindex)
1583 vm_cnt.v_reactivated++;
1586 m_object = m->object;
1587 vm_page_cache_remove(m);
1588 if (m_object->type == OBJT_VNODE &&
1589 vm_object_cache_is_empty(m_object))
1590 vp = m_object->handle;
1592 KASSERT(m->valid == 0,
1593 ("vm_page_alloc: free page %p is valid", m));
1594 vm_phys_freecnt_adj(m, -1);
1595 if ((m->flags & PG_ZERO) != 0)
1596 vm_page_zero_count--;
1598 mtx_unlock(&vm_page_queue_free_mtx);
1601 * Initialize the page. Only the PG_ZERO flag is inherited.
1604 if ((req & VM_ALLOC_ZERO) != 0)
1607 if ((req & VM_ALLOC_NODUMP) != 0)
1611 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1613 m->busy_lock = VPB_UNBUSIED;
1614 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1615 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1616 if ((req & VM_ALLOC_SBUSY) != 0)
1617 m->busy_lock = VPB_SHARERS_WORD(1);
1618 if (req & VM_ALLOC_WIRED) {
1620 * The page lock is not required for wiring a page until that
1621 * page is inserted into the object.
1623 atomic_add_int(&vm_cnt.v_wire_count, 1);
1628 if (object != NULL) {
1629 if (vm_page_insert_after(m, object, pindex, mpred)) {
1630 /* See the comment below about hold count. */
1633 pagedaemon_wakeup();
1634 if (req & VM_ALLOC_WIRED) {
1635 atomic_subtract_int(&vm_cnt.v_wire_count, 1);
1643 /* Ignore device objects; the pager sets "memattr" for them. */
1644 if (object->memattr != VM_MEMATTR_DEFAULT &&
1645 (object->flags & OBJ_FICTITIOUS) == 0)
1646 pmap_page_set_memattr(m, object->memattr);
1651 * The following call to vdrop() must come after the above call
1652 * to vm_page_insert() in case both affect the same object and
1653 * vnode. Otherwise, the affected vnode's hold count could
1654 * temporarily become zero.
1660 * Don't wakeup too often - wakeup the pageout daemon when
1661 * we would be nearly out of memory.
1663 if (vm_paging_needed())
1664 pagedaemon_wakeup();
1670 vm_page_alloc_contig_vdrop(struct spglist *lst)
1673 while (!SLIST_EMPTY(lst)) {
1674 vdrop((struct vnode *)SLIST_FIRST(lst)-> plinks.s.pv);
1675 SLIST_REMOVE_HEAD(lst, plinks.s.ss);
1680 * vm_page_alloc_contig:
1682 * Allocate a contiguous set of physical pages of the given size "npages"
1683 * from the free lists. All of the physical pages must be at or above
1684 * the given physical address "low" and below the given physical address
1685 * "high". The given value "alignment" determines the alignment of the
1686 * first physical page in the set. If the given value "boundary" is
1687 * non-zero, then the set of physical pages cannot cross any physical
1688 * address boundary that is a multiple of that value. Both "alignment"
1689 * and "boundary" must be a power of two.
1691 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1692 * then the memory attribute setting for the physical pages is configured
1693 * to the object's memory attribute setting. Otherwise, the memory
1694 * attribute setting for the physical pages is configured to "memattr",
1695 * overriding the object's memory attribute setting. However, if the
1696 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1697 * memory attribute setting for the physical pages cannot be configured
1698 * to VM_MEMATTR_DEFAULT.
1700 * The caller must always specify an allocation class.
1702 * allocation classes:
1703 * VM_ALLOC_NORMAL normal process request
1704 * VM_ALLOC_SYSTEM system *really* needs a page
1705 * VM_ALLOC_INTERRUPT interrupt time request
1707 * optional allocation flags:
1708 * VM_ALLOC_NOBUSY do not exclusive busy the page
1709 * VM_ALLOC_NOOBJ page is not associated with an object and
1710 * should not be exclusive busy
1711 * VM_ALLOC_SBUSY shared busy the allocated page
1712 * VM_ALLOC_WIRED wire the allocated page
1713 * VM_ALLOC_ZERO prefer a zeroed page
1715 * This routine may not sleep.
1718 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1719 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1720 vm_paddr_t boundary, vm_memattr_t memattr)
1723 struct spglist deferred_vdrop_list;
1724 vm_page_t m, m_tmp, m_ret;
1728 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1729 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1730 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1731 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1732 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object,
1734 if (object != NULL) {
1735 VM_OBJECT_ASSERT_WLOCKED(object);
1736 KASSERT(object->type == OBJT_PHYS,
1737 ("vm_page_alloc_contig: object %p isn't OBJT_PHYS",
1740 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
1741 req_class = req & VM_ALLOC_CLASS_MASK;
1744 * The page daemon is allowed to dig deeper into the free page list.
1746 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1747 req_class = VM_ALLOC_SYSTEM;
1749 SLIST_INIT(&deferred_vdrop_list);
1750 mtx_lock(&vm_page_queue_free_mtx);
1751 if (vm_cnt.v_free_count + vm_cnt.v_cache_count >= npages +
1752 vm_cnt.v_free_reserved || (req_class == VM_ALLOC_SYSTEM &&
1753 vm_cnt.v_free_count + vm_cnt.v_cache_count >= npages +
1754 vm_cnt.v_interrupt_free_min) || (req_class == VM_ALLOC_INTERRUPT &&
1755 vm_cnt.v_free_count + vm_cnt.v_cache_count >= npages)) {
1756 #if VM_NRESERVLEVEL > 0
1758 if (object == NULL || (object->flags & OBJ_COLORED) == 0 ||
1759 (m_ret = vm_reserv_alloc_contig(object, pindex, npages,
1760 low, high, alignment, boundary)) == NULL)
1762 m_ret = vm_phys_alloc_contig(npages, low, high,
1763 alignment, boundary);
1765 mtx_unlock(&vm_page_queue_free_mtx);
1766 atomic_add_int(&vm_pageout_deficit, npages);
1767 pagedaemon_wakeup();
1771 for (m = m_ret; m < &m_ret[npages]; m++) {
1772 drop = vm_page_alloc_init(m);
1775 * Enqueue the vnode for deferred vdrop().
1777 m->plinks.s.pv = drop;
1778 SLIST_INSERT_HEAD(&deferred_vdrop_list, m,
1783 #if VM_NRESERVLEVEL > 0
1784 if (vm_reserv_reclaim_contig(npages, low, high, alignment,
1789 mtx_unlock(&vm_page_queue_free_mtx);
1794 * Initialize the pages. Only the PG_ZERO flag is inherited.
1797 if ((req & VM_ALLOC_ZERO) != 0)
1799 if ((req & VM_ALLOC_NODUMP) != 0)
1801 if ((req & VM_ALLOC_WIRED) != 0)
1802 atomic_add_int(&vm_cnt.v_wire_count, npages);
1803 if (object != NULL) {
1804 if (object->memattr != VM_MEMATTR_DEFAULT &&
1805 memattr == VM_MEMATTR_DEFAULT)
1806 memattr = object->memattr;
1808 for (m = m_ret; m < &m_ret[npages]; m++) {
1810 m->flags = (m->flags | PG_NODUMP) & flags;
1811 m->busy_lock = VPB_UNBUSIED;
1812 if (object != NULL) {
1813 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
1814 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1815 if ((req & VM_ALLOC_SBUSY) != 0)
1816 m->busy_lock = VPB_SHARERS_WORD(1);
1818 if ((req & VM_ALLOC_WIRED) != 0)
1820 /* Unmanaged pages don't use "act_count". */
1821 m->oflags = VPO_UNMANAGED;
1822 if (object != NULL) {
1823 if (vm_page_insert(m, object, pindex)) {
1824 vm_page_alloc_contig_vdrop(
1825 &deferred_vdrop_list);
1826 if (vm_paging_needed())
1827 pagedaemon_wakeup();
1828 if ((req & VM_ALLOC_WIRED) != 0)
1829 atomic_subtract_int(&vm_cnt.v_wire_count,
1831 for (m_tmp = m, m = m_ret;
1832 m < &m_ret[npages]; m++) {
1833 if ((req & VM_ALLOC_WIRED) != 0)
1843 if (memattr != VM_MEMATTR_DEFAULT)
1844 pmap_page_set_memattr(m, memattr);
1847 vm_page_alloc_contig_vdrop(&deferred_vdrop_list);
1848 if (vm_paging_needed())
1849 pagedaemon_wakeup();
1854 * Initialize a page that has been freshly dequeued from a freelist.
1855 * The caller has to drop the vnode returned, if it is not NULL.
1857 * This function may only be used to initialize unmanaged pages.
1859 * To be called with vm_page_queue_free_mtx held.
1861 static struct vnode *
1862 vm_page_alloc_init(vm_page_t m)
1865 vm_object_t m_object;
1867 KASSERT(m->queue == PQ_NONE,
1868 ("vm_page_alloc_init: page %p has unexpected queue %d",
1870 KASSERT(m->wire_count == 0,
1871 ("vm_page_alloc_init: page %p is wired", m));
1872 KASSERT(m->hold_count == 0,
1873 ("vm_page_alloc_init: page %p is held", m));
1874 KASSERT(!vm_page_sbusied(m),
1875 ("vm_page_alloc_init: page %p is busy", m));
1876 KASSERT(m->dirty == 0,
1877 ("vm_page_alloc_init: page %p is dirty", m));
1878 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1879 ("vm_page_alloc_init: page %p has unexpected memattr %d",
1880 m, pmap_page_get_memattr(m)));
1881 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1883 if ((m->flags & PG_CACHED) != 0) {
1884 KASSERT((m->flags & PG_ZERO) == 0,
1885 ("vm_page_alloc_init: cached page %p is PG_ZERO", m));
1887 m_object = m->object;
1888 vm_page_cache_remove(m);
1889 if (m_object->type == OBJT_VNODE &&
1890 vm_object_cache_is_empty(m_object))
1891 drop = m_object->handle;
1893 KASSERT(m->valid == 0,
1894 ("vm_page_alloc_init: free page %p is valid", m));
1895 vm_phys_freecnt_adj(m, -1);
1896 if ((m->flags & PG_ZERO) != 0)
1897 vm_page_zero_count--;
1903 * vm_page_alloc_freelist:
1905 * Allocate a physical page from the specified free page list.
1907 * The caller must always specify an allocation class.
1909 * allocation classes:
1910 * VM_ALLOC_NORMAL normal process request
1911 * VM_ALLOC_SYSTEM system *really* needs a page
1912 * VM_ALLOC_INTERRUPT interrupt time request
1914 * optional allocation flags:
1915 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1916 * intends to allocate
1917 * VM_ALLOC_WIRED wire the allocated page
1918 * VM_ALLOC_ZERO prefer a zeroed page
1920 * This routine may not sleep.
1923 vm_page_alloc_freelist(int flind, int req)
1930 req_class = req & VM_ALLOC_CLASS_MASK;
1933 * The page daemon is allowed to dig deeper into the free page list.
1935 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1936 req_class = VM_ALLOC_SYSTEM;
1939 * Do not allocate reserved pages unless the req has asked for it.
1941 mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE);
1942 if (vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_free_reserved ||
1943 (req_class == VM_ALLOC_SYSTEM &&
1944 vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_interrupt_free_min) ||
1945 (req_class == VM_ALLOC_INTERRUPT &&
1946 vm_cnt.v_free_count + vm_cnt.v_cache_count > 0))
1947 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0);
1949 mtx_unlock(&vm_page_queue_free_mtx);
1950 atomic_add_int(&vm_pageout_deficit,
1951 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1952 pagedaemon_wakeup();
1956 mtx_unlock(&vm_page_queue_free_mtx);
1959 drop = vm_page_alloc_init(m);
1960 mtx_unlock(&vm_page_queue_free_mtx);
1963 * Initialize the page. Only the PG_ZERO flag is inherited.
1967 if ((req & VM_ALLOC_ZERO) != 0)
1970 if ((req & VM_ALLOC_WIRED) != 0) {
1972 * The page lock is not required for wiring a page that does
1973 * not belong to an object.
1975 atomic_add_int(&vm_cnt.v_wire_count, 1);
1978 /* Unmanaged pages don't use "act_count". */
1979 m->oflags = VPO_UNMANAGED;
1982 if (vm_paging_needed())
1983 pagedaemon_wakeup();
1988 * vm_wait: (also see VM_WAIT macro)
1990 * Sleep until free pages are available for allocation.
1991 * - Called in various places before memory allocations.
1997 mtx_lock(&vm_page_queue_free_mtx);
1998 if (curproc == pageproc) {
1999 vm_pageout_pages_needed = 1;
2000 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx,
2001 PDROP | PSWP, "VMWait", 0);
2003 if (!vm_pages_needed) {
2004 vm_pages_needed = 1;
2005 wakeup(&vm_pages_needed);
2007 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM,
2013 * vm_waitpfault: (also see VM_WAITPFAULT macro)
2015 * Sleep until free pages are available for allocation.
2016 * - Called only in vm_fault so that processes page faulting
2017 * can be easily tracked.
2018 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
2019 * processes will be able to grab memory first. Do not change
2020 * this balance without careful testing first.
2026 mtx_lock(&vm_page_queue_free_mtx);
2027 if (!vm_pages_needed) {
2028 vm_pages_needed = 1;
2029 wakeup(&vm_pages_needed);
2031 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER,
2035 struct vm_pagequeue *
2036 vm_page_pagequeue(vm_page_t m)
2039 return (&vm_phys_domain(m)->vmd_pagequeues[m->queue]);
2045 * Remove the given page from its current page queue.
2047 * The page must be locked.
2050 vm_page_dequeue(vm_page_t m)
2052 struct vm_pagequeue *pq;
2054 vm_page_assert_locked(m);
2055 KASSERT(m->queue < PQ_COUNT, ("vm_page_dequeue: page %p is not queued",
2057 pq = vm_page_pagequeue(m);
2058 vm_pagequeue_lock(pq);
2060 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2061 vm_pagequeue_cnt_dec(pq);
2062 vm_pagequeue_unlock(pq);
2066 * vm_page_dequeue_locked:
2068 * Remove the given page from its current page queue.
2070 * The page and page queue must be locked.
2073 vm_page_dequeue_locked(vm_page_t m)
2075 struct vm_pagequeue *pq;
2077 vm_page_lock_assert(m, MA_OWNED);
2078 pq = vm_page_pagequeue(m);
2079 vm_pagequeue_assert_locked(pq);
2081 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2082 vm_pagequeue_cnt_dec(pq);
2088 * Add the given page to the specified page queue.
2090 * The page must be locked.
2093 vm_page_enqueue(uint8_t queue, vm_page_t m)
2095 struct vm_pagequeue *pq;
2097 vm_page_lock_assert(m, MA_OWNED);
2098 KASSERT(queue < PQ_COUNT,
2099 ("vm_page_enqueue: invalid queue %u request for page %p",
2101 pq = &vm_phys_domain(m)->vmd_pagequeues[queue];
2102 vm_pagequeue_lock(pq);
2104 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2105 vm_pagequeue_cnt_inc(pq);
2106 vm_pagequeue_unlock(pq);
2112 * Move the given page to the tail of its current page queue.
2114 * The page must be locked.
2117 vm_page_requeue(vm_page_t m)
2119 struct vm_pagequeue *pq;
2121 vm_page_lock_assert(m, MA_OWNED);
2122 KASSERT(m->queue != PQ_NONE,
2123 ("vm_page_requeue: page %p is not queued", m));
2124 pq = vm_page_pagequeue(m);
2125 vm_pagequeue_lock(pq);
2126 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2127 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2128 vm_pagequeue_unlock(pq);
2132 * vm_page_requeue_locked:
2134 * Move the given page to the tail of its current page queue.
2136 * The page queue must be locked.
2139 vm_page_requeue_locked(vm_page_t m)
2141 struct vm_pagequeue *pq;
2143 KASSERT(m->queue != PQ_NONE,
2144 ("vm_page_requeue_locked: page %p is not queued", m));
2145 pq = vm_page_pagequeue(m);
2146 vm_pagequeue_assert_locked(pq);
2147 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2148 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2154 * Put the specified page on the active list (if appropriate).
2155 * Ensure that act_count is at least ACT_INIT but do not otherwise
2158 * The page must be locked.
2161 vm_page_activate(vm_page_t m)
2165 vm_page_lock_assert(m, MA_OWNED);
2166 if ((queue = m->queue) != PQ_ACTIVE) {
2167 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2168 if (m->act_count < ACT_INIT)
2169 m->act_count = ACT_INIT;
2170 if (queue != PQ_NONE)
2172 vm_page_enqueue(PQ_ACTIVE, m);
2174 KASSERT(queue == PQ_NONE,
2175 ("vm_page_activate: wired page %p is queued", m));
2177 if (m->act_count < ACT_INIT)
2178 m->act_count = ACT_INIT;
2183 * vm_page_free_wakeup:
2185 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
2186 * routine is called when a page has been added to the cache or free
2189 * The page queues must be locked.
2192 vm_page_free_wakeup(void)
2195 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2197 * if pageout daemon needs pages, then tell it that there are
2200 if (vm_pageout_pages_needed &&
2201 vm_cnt.v_cache_count + vm_cnt.v_free_count >= vm_cnt.v_pageout_free_min) {
2202 wakeup(&vm_pageout_pages_needed);
2203 vm_pageout_pages_needed = 0;
2206 * wakeup processes that are waiting on memory if we hit a
2207 * high water mark. And wakeup scheduler process if we have
2208 * lots of memory. this process will swapin processes.
2210 if (vm_pages_needed && !vm_page_count_min()) {
2211 vm_pages_needed = 0;
2212 wakeup(&vm_cnt.v_free_count);
2217 * Turn a cached page into a free page, by changing its attributes.
2218 * Keep the statistics up-to-date.
2220 * The free page queue must be locked.
2223 vm_page_cache_turn_free(vm_page_t m)
2226 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2230 KASSERT((m->flags & PG_CACHED) != 0,
2231 ("vm_page_cache_turn_free: page %p is not cached", m));
2232 m->flags &= ~PG_CACHED;
2233 vm_cnt.v_cache_count--;
2234 vm_phys_freecnt_adj(m, 1);
2240 * Returns the given page to the free list,
2241 * disassociating it with any VM object.
2243 * The object must be locked. The page must be locked if it is managed.
2246 vm_page_free_toq(vm_page_t m)
2249 if ((m->oflags & VPO_UNMANAGED) == 0) {
2250 vm_page_lock_assert(m, MA_OWNED);
2251 KASSERT(!pmap_page_is_mapped(m),
2252 ("vm_page_free_toq: freeing mapped page %p", m));
2254 KASSERT(m->queue == PQ_NONE,
2255 ("vm_page_free_toq: unmanaged page %p is queued", m));
2256 PCPU_INC(cnt.v_tfree);
2258 if (vm_page_sbusied(m))
2259 panic("vm_page_free: freeing busy page %p", m);
2262 * Unqueue, then remove page. Note that we cannot destroy
2263 * the page here because we do not want to call the pager's
2264 * callback routine until after we've put the page on the
2265 * appropriate free queue.
2271 * If fictitious remove object association and
2272 * return, otherwise delay object association removal.
2274 if ((m->flags & PG_FICTITIOUS) != 0) {
2281 if (m->wire_count != 0)
2282 panic("vm_page_free: freeing wired page %p", m);
2283 if (m->hold_count != 0) {
2284 m->flags &= ~PG_ZERO;
2285 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2286 ("vm_page_free: freeing PG_UNHOLDFREE page %p", m));
2287 m->flags |= PG_UNHOLDFREE;
2290 * Restore the default memory attribute to the page.
2292 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2293 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2296 * Insert the page into the physical memory allocator's
2297 * cache/free page queues.
2299 mtx_lock(&vm_page_queue_free_mtx);
2300 vm_phys_freecnt_adj(m, 1);
2301 #if VM_NRESERVLEVEL > 0
2302 if (!vm_reserv_free_page(m))
2306 vm_phys_free_pages(m, 0);
2307 if ((m->flags & PG_ZERO) != 0)
2308 ++vm_page_zero_count;
2310 vm_page_zero_idle_wakeup();
2311 vm_page_free_wakeup();
2312 mtx_unlock(&vm_page_queue_free_mtx);
2319 * Mark this page as wired down by yet
2320 * another map, removing it from paging queues
2323 * If the page is fictitious, then its wire count must remain one.
2325 * The page must be locked.
2328 vm_page_wire(vm_page_t m)
2332 * Only bump the wire statistics if the page is not already wired,
2333 * and only unqueue the page if it is on some queue (if it is unmanaged
2334 * it is already off the queues).
2336 vm_page_lock_assert(m, MA_OWNED);
2337 if ((m->flags & PG_FICTITIOUS) != 0) {
2338 KASSERT(m->wire_count == 1,
2339 ("vm_page_wire: fictitious page %p's wire count isn't one",
2343 if (m->wire_count == 0) {
2344 KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
2345 m->queue == PQ_NONE,
2346 ("vm_page_wire: unmanaged page %p is queued", m));
2348 atomic_add_int(&vm_cnt.v_wire_count, 1);
2351 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
2357 * Release one wiring of the specified page, potentially enabling it to be
2358 * paged again. If paging is enabled, then the value of the parameter
2359 * "queue" determines the queue to which the page is added.
2361 * However, unless the page belongs to an object, it is not enqueued because
2362 * it cannot be paged out.
2364 * If a page is fictitious, then its wire count must always be one.
2366 * A managed page must be locked.
2369 vm_page_unwire(vm_page_t m, uint8_t queue)
2372 KASSERT(queue < PQ_COUNT,
2373 ("vm_page_unwire: invalid queue %u request for page %p",
2375 if ((m->oflags & VPO_UNMANAGED) == 0)
2376 vm_page_lock_assert(m, MA_OWNED);
2377 if ((m->flags & PG_FICTITIOUS) != 0) {
2378 KASSERT(m->wire_count == 1,
2379 ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
2382 if (m->wire_count > 0) {
2384 if (m->wire_count == 0) {
2385 atomic_subtract_int(&vm_cnt.v_wire_count, 1);
2386 if ((m->oflags & VPO_UNMANAGED) != 0 ||
2389 if (queue == PQ_INACTIVE)
2390 m->flags &= ~PG_WINATCFLS;
2391 vm_page_enqueue(queue, m);
2394 panic("vm_page_unwire: page %p's wire count is zero", m);
2398 * Move the specified page to the inactive queue.
2400 * Many pages placed on the inactive queue should actually go
2401 * into the cache, but it is difficult to figure out which. What
2402 * we do instead, if the inactive target is well met, is to put
2403 * clean pages at the head of the inactive queue instead of the tail.
2404 * This will cause them to be moved to the cache more quickly and
2405 * if not actively re-referenced, reclaimed more quickly. If we just
2406 * stick these pages at the end of the inactive queue, heavy filesystem
2407 * meta-data accesses can cause an unnecessary paging load on memory bound
2408 * processes. This optimization causes one-time-use metadata to be
2409 * reused more quickly.
2411 * Normally athead is 0 resulting in LRU operation. athead is set
2412 * to 1 if we want this page to be 'as if it were placed in the cache',
2413 * except without unmapping it from the process address space.
2415 * The page must be locked.
2418 _vm_page_deactivate(vm_page_t m, int athead)
2420 struct vm_pagequeue *pq;
2423 vm_page_lock_assert(m, MA_OWNED);
2426 * Ignore if already inactive.
2428 if ((queue = m->queue) == PQ_INACTIVE)
2430 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2431 if (queue != PQ_NONE)
2433 m->flags &= ~PG_WINATCFLS;
2434 pq = &vm_phys_domain(m)->vmd_pagequeues[PQ_INACTIVE];
2435 vm_pagequeue_lock(pq);
2436 m->queue = PQ_INACTIVE;
2438 TAILQ_INSERT_HEAD(&pq->pq_pl, m, plinks.q);
2440 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2441 vm_pagequeue_cnt_inc(pq);
2442 vm_pagequeue_unlock(pq);
2447 * Move the specified page to the inactive queue.
2449 * The page must be locked.
2452 vm_page_deactivate(vm_page_t m)
2455 _vm_page_deactivate(m, 0);
2459 * vm_page_try_to_cache:
2461 * Returns 0 on failure, 1 on success
2464 vm_page_try_to_cache(vm_page_t m)
2467 vm_page_lock_assert(m, MA_OWNED);
2468 VM_OBJECT_ASSERT_WLOCKED(m->object);
2469 if (m->dirty || m->hold_count || m->wire_count ||
2470 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m))
2480 * vm_page_try_to_free()
2482 * Attempt to free the page. If we cannot free it, we do nothing.
2483 * 1 is returned on success, 0 on failure.
2486 vm_page_try_to_free(vm_page_t m)
2489 vm_page_lock_assert(m, MA_OWNED);
2490 if (m->object != NULL)
2491 VM_OBJECT_ASSERT_WLOCKED(m->object);
2492 if (m->dirty || m->hold_count || m->wire_count ||
2493 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m))
2505 * Put the specified page onto the page cache queue (if appropriate).
2507 * The object and page must be locked.
2510 vm_page_cache(vm_page_t m)
2513 boolean_t cache_was_empty;
2515 vm_page_lock_assert(m, MA_OWNED);
2517 VM_OBJECT_ASSERT_WLOCKED(object);
2518 if (vm_page_busied(m) || (m->oflags & VPO_UNMANAGED) ||
2519 m->hold_count || m->wire_count)
2520 panic("vm_page_cache: attempting to cache busy page");
2521 KASSERT(!pmap_page_is_mapped(m),
2522 ("vm_page_cache: page %p is mapped", m));
2523 KASSERT(m->dirty == 0, ("vm_page_cache: page %p is dirty", m));
2524 if (m->valid == 0 || object->type == OBJT_DEFAULT ||
2525 (object->type == OBJT_SWAP &&
2526 !vm_pager_has_page(object, m->pindex, NULL, NULL))) {
2528 * Hypothesis: A cache-eligible page belonging to a
2529 * default object or swap object but without a backing
2530 * store must be zero filled.
2535 KASSERT((m->flags & PG_CACHED) == 0,
2536 ("vm_page_cache: page %p is already cached", m));
2539 * Remove the page from the paging queues.
2544 * Remove the page from the object's collection of resident
2547 vm_radix_remove(&object->rtree, m->pindex);
2548 TAILQ_REMOVE(&object->memq, m, listq);
2549 object->resident_page_count--;
2552 * Restore the default memory attribute to the page.
2554 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2555 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2558 * Insert the page into the object's collection of cached pages
2559 * and the physical memory allocator's cache/free page queues.
2561 m->flags &= ~PG_ZERO;
2562 mtx_lock(&vm_page_queue_free_mtx);
2563 cache_was_empty = vm_radix_is_empty(&object->cache);
2564 if (vm_radix_insert(&object->cache, m)) {
2565 mtx_unlock(&vm_page_queue_free_mtx);
2566 if (object->resident_page_count == 0)
2567 vdrop(object->handle);
2574 * The above call to vm_radix_insert() could reclaim the one pre-
2575 * existing cached page from this object, resulting in a call to
2578 if (!cache_was_empty)
2579 cache_was_empty = vm_radix_is_singleton(&object->cache);
2581 m->flags |= PG_CACHED;
2582 vm_cnt.v_cache_count++;
2583 PCPU_INC(cnt.v_tcached);
2584 #if VM_NRESERVLEVEL > 0
2585 if (!vm_reserv_free_page(m)) {
2589 vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0);
2590 vm_phys_free_pages(m, 0);
2592 vm_page_free_wakeup();
2593 mtx_unlock(&vm_page_queue_free_mtx);
2596 * Increment the vnode's hold count if this is the object's only
2597 * cached page. Decrement the vnode's hold count if this was
2598 * the object's only resident page.
2600 if (object->type == OBJT_VNODE) {
2601 if (cache_was_empty && object->resident_page_count != 0)
2602 vhold(object->handle);
2603 else if (!cache_was_empty && object->resident_page_count == 0)
2604 vdrop(object->handle);
2611 * Cache, deactivate, or do nothing as appropriate. This routine
2612 * is used by madvise().
2614 * Generally speaking we want to move the page into the cache so
2615 * it gets reused quickly. However, this can result in a silly syndrome
2616 * due to the page recycling too quickly. Small objects will not be
2617 * fully cached. On the other hand, if we move the page to the inactive
2618 * queue we wind up with a problem whereby very large objects
2619 * unnecessarily blow away our inactive and cache queues.
2621 * The solution is to move the pages based on a fixed weighting. We
2622 * either leave them alone, deactivate them, or move them to the cache,
2623 * where moving them to the cache has the highest weighting.
2624 * By forcing some pages into other queues we eventually force the
2625 * system to balance the queues, potentially recovering other unrelated
2626 * space from active. The idea is to not force this to happen too
2629 * The object and page must be locked.
2632 vm_page_advise(vm_page_t m, int advice)
2636 vm_page_assert_locked(m);
2637 VM_OBJECT_ASSERT_WLOCKED(m->object);
2638 if (advice == MADV_FREE) {
2640 * Mark the page clean. This will allow the page to be freed
2641 * up by the system. However, such pages are often reused
2642 * quickly by malloc() so we do not do anything that would
2643 * cause a page fault if we can help it.
2645 * Specifically, we do not try to actually free the page now
2646 * nor do we try to put it in the cache (which would cause a
2647 * page fault on reuse).
2649 * But we do make the page is freeable as we can without
2650 * actually taking the step of unmapping it.
2654 } else if (advice != MADV_DONTNEED)
2656 dnw = PCPU_GET(dnweight);
2660 * Occasionally leave the page alone.
2662 if ((dnw & 0x01F0) == 0 || m->queue == PQ_INACTIVE) {
2663 if (m->act_count >= ACT_INIT)
2669 * Clear any references to the page. Otherwise, the page daemon will
2670 * immediately reactivate the page.
2672 vm_page_aflag_clear(m, PGA_REFERENCED);
2674 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
2677 if (m->dirty || (dnw & 0x0070) == 0) {
2679 * Deactivate the page 3 times out of 32.
2684 * Cache the page 28 times out of every 32. Note that
2685 * the page is deactivated instead of cached, but placed
2686 * at the head of the queue instead of the tail.
2690 _vm_page_deactivate(m, head);
2694 * Grab a page, waiting until we are waken up due to the page
2695 * changing state. We keep on waiting, if the page continues
2696 * to be in the object. If the page doesn't exist, first allocate it
2697 * and then conditionally zero it.
2699 * This routine may sleep.
2701 * The object must be locked on entry. The lock will, however, be released
2702 * and reacquired if the routine sleeps.
2705 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2710 VM_OBJECT_ASSERT_WLOCKED(object);
2711 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
2712 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
2713 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
2715 if ((m = vm_page_lookup(object, pindex)) != NULL) {
2716 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
2717 vm_page_xbusied(m) : vm_page_busied(m);
2719 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
2722 * Reference the page before unlocking and
2723 * sleeping so that the page daemon is less
2724 * likely to reclaim it.
2726 vm_page_aflag_set(m, PGA_REFERENCED);
2728 VM_OBJECT_WUNLOCK(object);
2729 vm_page_busy_sleep(m, "pgrbwt");
2730 VM_OBJECT_WLOCK(object);
2733 if ((allocflags & VM_ALLOC_WIRED) != 0) {
2739 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
2741 if ((allocflags & VM_ALLOC_SBUSY) != 0)
2746 m = vm_page_alloc(object, pindex, allocflags);
2748 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
2750 VM_OBJECT_WUNLOCK(object);
2752 VM_OBJECT_WLOCK(object);
2754 } else if (m->valid != 0)
2756 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
2762 * Mapping function for valid or dirty bits in a page.
2764 * Inputs are required to range within a page.
2767 vm_page_bits(int base, int size)
2773 base + size <= PAGE_SIZE,
2774 ("vm_page_bits: illegal base/size %d/%d", base, size)
2777 if (size == 0) /* handle degenerate case */
2780 first_bit = base >> DEV_BSHIFT;
2781 last_bit = (base + size - 1) >> DEV_BSHIFT;
2783 return (((vm_page_bits_t)2 << last_bit) -
2784 ((vm_page_bits_t)1 << first_bit));
2788 * vm_page_set_valid_range:
2790 * Sets portions of a page valid. The arguments are expected
2791 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2792 * of any partial chunks touched by the range. The invalid portion of
2793 * such chunks will be zeroed.
2795 * (base + size) must be less then or equal to PAGE_SIZE.
2798 vm_page_set_valid_range(vm_page_t m, int base, int size)
2802 VM_OBJECT_ASSERT_WLOCKED(m->object);
2803 if (size == 0) /* handle degenerate case */
2807 * If the base is not DEV_BSIZE aligned and the valid
2808 * bit is clear, we have to zero out a portion of the
2811 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2812 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
2813 pmap_zero_page_area(m, frag, base - frag);
2816 * If the ending offset is not DEV_BSIZE aligned and the
2817 * valid bit is clear, we have to zero out a portion of
2820 endoff = base + size;
2821 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2822 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
2823 pmap_zero_page_area(m, endoff,
2824 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
2827 * Assert that no previously invalid block that is now being validated
2830 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
2831 ("vm_page_set_valid_range: page %p is dirty", m));
2834 * Set valid bits inclusive of any overlap.
2836 m->valid |= vm_page_bits(base, size);
2840 * Clear the given bits from the specified page's dirty field.
2842 static __inline void
2843 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
2846 #if PAGE_SIZE < 16384
2851 * If the object is locked and the page is neither exclusive busy nor
2852 * write mapped, then the page's dirty field cannot possibly be
2853 * set by a concurrent pmap operation.
2855 VM_OBJECT_ASSERT_WLOCKED(m->object);
2856 if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
2857 m->dirty &= ~pagebits;
2860 * The pmap layer can call vm_page_dirty() without
2861 * holding a distinguished lock. The combination of
2862 * the object's lock and an atomic operation suffice
2863 * to guarantee consistency of the page dirty field.
2865 * For PAGE_SIZE == 32768 case, compiler already
2866 * properly aligns the dirty field, so no forcible
2867 * alignment is needed. Only require existence of
2868 * atomic_clear_64 when page size is 32768.
2870 addr = (uintptr_t)&m->dirty;
2871 #if PAGE_SIZE == 32768
2872 atomic_clear_64((uint64_t *)addr, pagebits);
2873 #elif PAGE_SIZE == 16384
2874 atomic_clear_32((uint32_t *)addr, pagebits);
2875 #else /* PAGE_SIZE <= 8192 */
2877 * Use a trick to perform a 32-bit atomic on the
2878 * containing aligned word, to not depend on the existence
2879 * of atomic_clear_{8, 16}.
2881 shift = addr & (sizeof(uint32_t) - 1);
2882 #if BYTE_ORDER == BIG_ENDIAN
2883 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
2887 addr &= ~(sizeof(uint32_t) - 1);
2888 atomic_clear_32((uint32_t *)addr, pagebits << shift);
2889 #endif /* PAGE_SIZE */
2894 * vm_page_set_validclean:
2896 * Sets portions of a page valid and clean. The arguments are expected
2897 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2898 * of any partial chunks touched by the range. The invalid portion of
2899 * such chunks will be zero'd.
2901 * (base + size) must be less then or equal to PAGE_SIZE.
2904 vm_page_set_validclean(vm_page_t m, int base, int size)
2906 vm_page_bits_t oldvalid, pagebits;
2909 VM_OBJECT_ASSERT_WLOCKED(m->object);
2910 if (size == 0) /* handle degenerate case */
2914 * If the base is not DEV_BSIZE aligned and the valid
2915 * bit is clear, we have to zero out a portion of the
2918 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2919 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
2920 pmap_zero_page_area(m, frag, base - frag);
2923 * If the ending offset is not DEV_BSIZE aligned and the
2924 * valid bit is clear, we have to zero out a portion of
2927 endoff = base + size;
2928 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2929 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
2930 pmap_zero_page_area(m, endoff,
2931 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
2934 * Set valid, clear dirty bits. If validating the entire
2935 * page we can safely clear the pmap modify bit. We also
2936 * use this opportunity to clear the VPO_NOSYNC flag. If a process
2937 * takes a write fault on a MAP_NOSYNC memory area the flag will
2940 * We set valid bits inclusive of any overlap, but we can only
2941 * clear dirty bits for DEV_BSIZE chunks that are fully within
2944 oldvalid = m->valid;
2945 pagebits = vm_page_bits(base, size);
2946 m->valid |= pagebits;
2948 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
2949 frag = DEV_BSIZE - frag;
2955 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
2957 if (base == 0 && size == PAGE_SIZE) {
2959 * The page can only be modified within the pmap if it is
2960 * mapped, and it can only be mapped if it was previously
2963 if (oldvalid == VM_PAGE_BITS_ALL)
2965 * Perform the pmap_clear_modify() first. Otherwise,
2966 * a concurrent pmap operation, such as
2967 * pmap_protect(), could clear a modification in the
2968 * pmap and set the dirty field on the page before
2969 * pmap_clear_modify() had begun and after the dirty
2970 * field was cleared here.
2972 pmap_clear_modify(m);
2974 m->oflags &= ~VPO_NOSYNC;
2975 } else if (oldvalid != VM_PAGE_BITS_ALL)
2976 m->dirty &= ~pagebits;
2978 vm_page_clear_dirty_mask(m, pagebits);
2982 vm_page_clear_dirty(vm_page_t m, int base, int size)
2985 vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
2989 * vm_page_set_invalid:
2991 * Invalidates DEV_BSIZE'd chunks within a page. Both the
2992 * valid and dirty bits for the effected areas are cleared.
2995 vm_page_set_invalid(vm_page_t m, int base, int size)
2997 vm_page_bits_t bits;
3001 VM_OBJECT_ASSERT_WLOCKED(object);
3002 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
3003 size >= object->un_pager.vnp.vnp_size)
3004 bits = VM_PAGE_BITS_ALL;
3006 bits = vm_page_bits(base, size);
3007 if (m->valid == VM_PAGE_BITS_ALL && bits != 0)
3009 KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) ||
3010 !pmap_page_is_mapped(m),
3011 ("vm_page_set_invalid: page %p is mapped", m));
3017 * vm_page_zero_invalid()
3019 * The kernel assumes that the invalid portions of a page contain
3020 * garbage, but such pages can be mapped into memory by user code.
3021 * When this occurs, we must zero out the non-valid portions of the
3022 * page so user code sees what it expects.
3024 * Pages are most often semi-valid when the end of a file is mapped
3025 * into memory and the file's size is not page aligned.
3028 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
3033 VM_OBJECT_ASSERT_WLOCKED(m->object);
3035 * Scan the valid bits looking for invalid sections that
3036 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
3037 * valid bit may be set ) have already been zerod by
3038 * vm_page_set_validclean().
3040 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
3041 if (i == (PAGE_SIZE / DEV_BSIZE) ||
3042 (m->valid & ((vm_page_bits_t)1 << i))) {
3044 pmap_zero_page_area(m,
3045 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
3052 * setvalid is TRUE when we can safely set the zero'd areas
3053 * as being valid. We can do this if there are no cache consistancy
3054 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
3057 m->valid = VM_PAGE_BITS_ALL;
3063 * Is (partial) page valid? Note that the case where size == 0
3064 * will return FALSE in the degenerate case where the page is
3065 * entirely invalid, and TRUE otherwise.
3068 vm_page_is_valid(vm_page_t m, int base, int size)
3070 vm_page_bits_t bits;
3072 VM_OBJECT_ASSERT_LOCKED(m->object);
3073 bits = vm_page_bits(base, size);
3074 return (m->valid != 0 && (m->valid & bits) == bits);
3078 * vm_page_ps_is_valid:
3080 * Returns TRUE if the entire (super)page is valid and FALSE otherwise.
3083 vm_page_ps_is_valid(vm_page_t m)
3087 VM_OBJECT_ASSERT_LOCKED(m->object);
3088 npages = atop(pagesizes[m->psind]);
3091 * The physically contiguous pages that make up a superpage, i.e., a
3092 * page with a page size index ("psind") greater than zero, will
3093 * occupy adjacent entries in vm_page_array[].
3095 for (i = 0; i < npages; i++) {
3096 if (m[i].valid != VM_PAGE_BITS_ALL)
3103 * Set the page's dirty bits if the page is modified.
3106 vm_page_test_dirty(vm_page_t m)
3109 VM_OBJECT_ASSERT_WLOCKED(m->object);
3110 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
3115 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
3118 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
3122 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
3125 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
3129 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
3132 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
3135 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
3137 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
3140 vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
3144 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
3147 mtx_assert_(vm_page_lockptr(m), a, file, line);
3153 vm_page_object_lock_assert(vm_page_t m)
3157 * Certain of the page's fields may only be modified by the
3158 * holder of the containing object's lock or the exclusive busy.
3159 * holder. Unfortunately, the holder of the write busy is
3160 * not recorded, and thus cannot be checked here.
3162 if (m->object != NULL && !vm_page_xbusied(m))
3163 VM_OBJECT_ASSERT_WLOCKED(m->object);
3167 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits)
3170 if ((bits & PGA_WRITEABLE) == 0)
3174 * The PGA_WRITEABLE flag can only be set if the page is
3175 * managed, is exclusively busied or the object is locked.
3176 * Currently, this flag is only set by pmap_enter().
3178 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3179 ("PGA_WRITEABLE on unmanaged page"));
3180 if (!vm_page_xbusied(m))
3181 VM_OBJECT_ASSERT_LOCKED(m->object);
3185 #include "opt_ddb.h"
3187 #include <sys/kernel.h>
3189 #include <ddb/ddb.h>
3191 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3193 db_printf("vm_cnt.v_free_count: %d\n", vm_cnt.v_free_count);
3194 db_printf("vm_cnt.v_cache_count: %d\n", vm_cnt.v_cache_count);
3195 db_printf("vm_cnt.v_inactive_count: %d\n", vm_cnt.v_inactive_count);
3196 db_printf("vm_cnt.v_active_count: %d\n", vm_cnt.v_active_count);
3197 db_printf("vm_cnt.v_wire_count: %d\n", vm_cnt.v_wire_count);
3198 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
3199 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
3200 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
3201 db_printf("vm_cnt.v_cache_min: %d\n", vm_cnt.v_cache_min);
3202 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
3205 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3209 db_printf("pq_free %d pq_cache %d\n",
3210 vm_cnt.v_free_count, vm_cnt.v_cache_count);
3211 for (dom = 0; dom < vm_ndomains; dom++) {
3213 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pass %d\n",
3215 vm_dom[dom].vmd_page_count,
3216 vm_dom[dom].vmd_free_count,
3217 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
3218 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
3219 vm_dom[dom].vmd_pass);
3223 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
3229 db_printf("show pginfo addr\n");
3233 phys = strchr(modif, 'p') != NULL;
3235 m = PHYS_TO_VM_PAGE(addr);
3237 m = (vm_page_t)addr;
3239 "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n"
3240 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
3241 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
3242 m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags,
3243 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty);