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/linker.h>
95 #include <sys/malloc.h>
97 #include <sys/msgbuf.h>
98 #include <sys/mutex.h>
100 #include <sys/rwlock.h>
101 #include <sys/sbuf.h>
102 #include <sys/sysctl.h>
103 #include <sys/vmmeter.h>
104 #include <sys/vnode.h>
108 #include <vm/vm_param.h>
109 #include <vm/vm_kern.h>
110 #include <vm/vm_object.h>
111 #include <vm/vm_page.h>
112 #include <vm/vm_pageout.h>
113 #include <vm/vm_pager.h>
114 #include <vm/vm_phys.h>
115 #include <vm/vm_radix.h>
116 #include <vm/vm_reserv.h>
117 #include <vm/vm_extern.h>
119 #include <vm/uma_int.h>
121 #include <machine/md_var.h>
124 * Associated with page of user-allocatable memory is a
128 struct vm_domain vm_dom[MAXMEMDOM];
129 struct mtx_padalign vm_page_queue_free_mtx;
131 struct mtx_padalign pa_lock[PA_LOCK_COUNT];
133 vm_page_t vm_page_array;
134 long vm_page_array_size;
136 int vm_page_zero_count;
138 static int boot_pages = UMA_BOOT_PAGES;
139 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
141 "number of pages allocated for bootstrapping the VM system");
143 static int pa_tryrelock_restart;
144 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
145 &pa_tryrelock_restart, 0, "Number of tryrelock restarts");
147 static TAILQ_HEAD(, vm_page) blacklist_head;
148 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
149 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
150 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
152 /* Is the page daemon waiting for free pages? */
153 static int vm_pageout_pages_needed;
155 static uma_zone_t fakepg_zone;
157 static struct vnode *vm_page_alloc_init(vm_page_t m);
158 static void vm_page_cache_turn_free(vm_page_t m);
159 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
160 static void vm_page_enqueue(uint8_t queue, vm_page_t m);
161 static void vm_page_free_wakeup(void);
162 static void vm_page_init_fakepg(void *dummy);
163 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
164 vm_pindex_t pindex, vm_page_t mpred);
165 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
167 static int vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run,
170 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init_fakepg, NULL);
173 vm_page_init_fakepg(void *dummy)
176 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
177 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
180 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
181 #if PAGE_SIZE == 32768
183 CTASSERT(sizeof(u_long) >= 8);
188 * Try to acquire a physical address lock while a pmap is locked. If we
189 * fail to trylock we unlock and lock the pmap directly and cache the
190 * locked pa in *locked. The caller should then restart their loop in case
191 * the virtual to physical mapping has changed.
194 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
201 PA_LOCK_ASSERT(lockpa, MA_OWNED);
202 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
209 atomic_add_int(&pa_tryrelock_restart, 1);
218 * Sets the page size, perhaps based upon the memory
219 * size. Must be called before any use of page-size
220 * dependent functions.
223 vm_set_page_size(void)
225 if (vm_cnt.v_page_size == 0)
226 vm_cnt.v_page_size = PAGE_SIZE;
227 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
228 panic("vm_set_page_size: page size not a power of two");
232 * vm_page_blacklist_next:
234 * Find the next entry in the provided string of blacklist
235 * addresses. Entries are separated by space, comma, or newline.
236 * If an invalid integer is encountered then the rest of the
237 * string is skipped. Updates the list pointer to the next
238 * character, or NULL if the string is exhausted or invalid.
241 vm_page_blacklist_next(char **list, char *end)
246 if (list == NULL || *list == NULL)
254 * If there's no end pointer then the buffer is coming from
255 * the kenv and we know it's null-terminated.
258 end = *list + strlen(*list);
260 /* Ensure that strtoq() won't walk off the end */
262 if (*end == '\n' || *end == ' ' || *end == ',')
265 printf("Blacklist not terminated, skipping\n");
271 for (pos = *list; *pos != '\0'; pos = cp) {
272 bad = strtoq(pos, &cp, 0);
273 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
282 if (*cp == '\0' || ++cp >= end)
286 return (trunc_page(bad));
288 printf("Garbage in RAM blacklist, skipping\n");
294 * vm_page_blacklist_check:
296 * Iterate through the provided string of blacklist addresses, pulling
297 * each entry out of the physical allocator free list and putting it
298 * onto a list for reporting via the vm.page_blacklist sysctl.
301 vm_page_blacklist_check(char *list, char *end)
309 while (next != NULL) {
310 if ((pa = vm_page_blacklist_next(&next, end)) == 0)
312 m = vm_phys_paddr_to_vm_page(pa);
315 mtx_lock(&vm_page_queue_free_mtx);
316 ret = vm_phys_unfree_page(m);
317 mtx_unlock(&vm_page_queue_free_mtx);
319 TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
321 printf("Skipping page with pa 0x%jx\n",
328 * vm_page_blacklist_load:
330 * Search for a special module named "ram_blacklist". It'll be a
331 * plain text file provided by the user via the loader directive
335 vm_page_blacklist_load(char **list, char **end)
344 mod = preload_search_by_type("ram_blacklist");
346 ptr = preload_fetch_addr(mod);
347 len = preload_fetch_size(mod);
358 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
365 error = sysctl_wire_old_buffer(req, 0);
368 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
369 TAILQ_FOREACH(m, &blacklist_head, listq) {
370 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
371 (uintmax_t)m->phys_addr);
374 error = sbuf_finish(&sbuf);
380 vm_page_domain_init(struct vm_domain *vmd)
382 struct vm_pagequeue *pq;
385 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
386 "vm inactive pagequeue";
387 *__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_vcnt) =
388 &vm_cnt.v_inactive_count;
389 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
390 "vm active pagequeue";
391 *__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_vcnt) =
392 &vm_cnt.v_active_count;
393 vmd->vmd_page_count = 0;
394 vmd->vmd_free_count = 0;
396 vmd->vmd_oom = FALSE;
398 for (i = 0; i < PQ_COUNT; i++) {
399 pq = &vmd->vmd_pagequeues[i];
400 TAILQ_INIT(&pq->pq_pl);
401 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
402 MTX_DEF | MTX_DUPOK);
409 * Initializes the resident memory module.
411 * Allocates memory for the page cells, and
412 * for the object/offset-to-page hash table headers.
413 * Each page cell is initialized and placed on the free list.
416 vm_page_startup(vm_offset_t vaddr)
419 vm_paddr_t page_range;
424 char *list, *listend;
426 vm_paddr_t biggestsize;
427 vm_paddr_t low_water, high_water;
432 vaddr = round_page(vaddr);
434 for (i = 0; phys_avail[i + 1]; i += 2) {
435 phys_avail[i] = round_page(phys_avail[i]);
436 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
439 low_water = phys_avail[0];
440 high_water = phys_avail[1];
442 for (i = 0; i < vm_phys_nsegs; i++) {
443 if (vm_phys_segs[i].start < low_water)
444 low_water = vm_phys_segs[i].start;
445 if (vm_phys_segs[i].end > high_water)
446 high_water = vm_phys_segs[i].end;
448 for (i = 0; phys_avail[i + 1]; i += 2) {
449 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
451 if (size > biggestsize) {
455 if (phys_avail[i] < low_water)
456 low_water = phys_avail[i];
457 if (phys_avail[i + 1] > high_water)
458 high_water = phys_avail[i + 1];
461 end = phys_avail[biggestone+1];
464 * Initialize the page and queue locks.
466 mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF);
467 for (i = 0; i < PA_LOCK_COUNT; i++)
468 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
469 for (i = 0; i < vm_ndomains; i++)
470 vm_page_domain_init(&vm_dom[i]);
473 * Allocate memory for use when boot strapping the kernel memory
476 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
477 * manually fetch the value.
479 TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
480 new_end = end - (boot_pages * UMA_SLAB_SIZE);
481 new_end = trunc_page(new_end);
482 mapped = pmap_map(&vaddr, new_end, end,
483 VM_PROT_READ | VM_PROT_WRITE);
484 bzero((void *)mapped, end - new_end);
485 uma_startup((void *)mapped, boot_pages);
487 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
488 defined(__i386__) || defined(__mips__)
490 * Allocate a bitmap to indicate that a random physical page
491 * needs to be included in a minidump.
493 * The amd64 port needs this to indicate which direct map pages
494 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
496 * However, i386 still needs this workspace internally within the
497 * minidump code. In theory, they are not needed on i386, but are
498 * included should the sf_buf code decide to use them.
501 for (i = 0; dump_avail[i + 1] != 0; i += 2)
502 if (dump_avail[i + 1] > last_pa)
503 last_pa = dump_avail[i + 1];
504 page_range = last_pa / PAGE_SIZE;
505 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
506 new_end -= vm_page_dump_size;
507 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
508 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
509 bzero((void *)vm_page_dump, vm_page_dump_size);
513 * Request that the physical pages underlying the message buffer be
514 * included in a crash dump. Since the message buffer is accessed
515 * through the direct map, they are not automatically included.
517 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
518 last_pa = pa + round_page(msgbufsize);
519 while (pa < last_pa) {
525 * Compute the number of pages of memory that will be available for
526 * use (taking into account the overhead of a page structure per
529 first_page = low_water / PAGE_SIZE;
530 #ifdef VM_PHYSSEG_SPARSE
532 for (i = 0; i < vm_phys_nsegs; i++) {
533 page_range += atop(vm_phys_segs[i].end -
534 vm_phys_segs[i].start);
536 for (i = 0; phys_avail[i + 1] != 0; i += 2)
537 page_range += atop(phys_avail[i + 1] - phys_avail[i]);
538 #elif defined(VM_PHYSSEG_DENSE)
539 page_range = high_water / PAGE_SIZE - first_page;
541 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
546 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
551 * Initialize the mem entry structures now, and put them in the free
554 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
555 mapped = pmap_map(&vaddr, new_end, end,
556 VM_PROT_READ | VM_PROT_WRITE);
557 vm_page_array = (vm_page_t) mapped;
558 #if VM_NRESERVLEVEL > 0
560 * Allocate memory for the reservation management system's data
563 new_end = vm_reserv_startup(&vaddr, new_end, high_water);
565 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
567 * pmap_map on arm64, amd64, and mips can come out of the direct-map,
568 * not kvm like i386, so the pages must be tracked for a crashdump to
569 * include this data. This includes the vm_page_array and the early
570 * UMA bootstrap pages.
572 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
575 phys_avail[biggestone + 1] = new_end;
578 * Add physical memory segments corresponding to the available
581 for (i = 0; phys_avail[i + 1] != 0; i += 2)
582 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
585 * Clear all of the page structures
587 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
588 for (i = 0; i < page_range; i++)
589 vm_page_array[i].order = VM_NFREEORDER;
590 vm_page_array_size = page_range;
593 * Initialize the physical memory allocator.
598 * Add every available physical page that is not blacklisted to
601 vm_cnt.v_page_count = 0;
602 vm_cnt.v_free_count = 0;
603 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
605 last_pa = phys_avail[i + 1];
606 while (pa < last_pa) {
607 vm_phys_add_page(pa);
612 TAILQ_INIT(&blacklist_head);
613 vm_page_blacklist_load(&list, &listend);
614 vm_page_blacklist_check(list, listend);
616 list = kern_getenv("vm.blacklist");
617 vm_page_blacklist_check(list, NULL);
620 #if VM_NRESERVLEVEL > 0
622 * Initialize the reservation management system.
630 vm_page_reference(vm_page_t m)
633 vm_page_aflag_set(m, PGA_REFERENCED);
637 * vm_page_busy_downgrade:
639 * Downgrade an exclusive busy page into a single shared busy page.
642 vm_page_busy_downgrade(vm_page_t m)
646 vm_page_assert_xbusied(m);
650 x &= VPB_BIT_WAITERS;
651 if (atomic_cmpset_rel_int(&m->busy_lock,
652 VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1) | x))
660 * Return a positive value if the page is shared busied, 0 otherwise.
663 vm_page_sbusied(vm_page_t m)
668 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
674 * Shared unbusy a page.
677 vm_page_sunbusy(vm_page_t m)
681 vm_page_assert_sbusied(m);
685 if (VPB_SHARERS(x) > 1) {
686 if (atomic_cmpset_int(&m->busy_lock, x,
691 if ((x & VPB_BIT_WAITERS) == 0) {
692 KASSERT(x == VPB_SHARERS_WORD(1),
693 ("vm_page_sunbusy: invalid lock state"));
694 if (atomic_cmpset_int(&m->busy_lock,
695 VPB_SHARERS_WORD(1), VPB_UNBUSIED))
699 KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS),
700 ("vm_page_sunbusy: invalid lock state for waiters"));
703 if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) {
714 * vm_page_busy_sleep:
716 * Sleep and release the page lock, using the page pointer as wchan.
717 * This is used to implement the hard-path of busying mechanism.
719 * The given page must be locked.
722 vm_page_busy_sleep(vm_page_t m, const char *wmesg)
726 vm_page_lock_assert(m, MA_OWNED);
729 if (x == VPB_UNBUSIED) {
733 if ((x & VPB_BIT_WAITERS) == 0 &&
734 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS)) {
738 msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0);
744 * Try to shared busy a page.
745 * If the operation succeeds 1 is returned otherwise 0.
746 * The operation never sleeps.
749 vm_page_trysbusy(vm_page_t m)
755 if ((x & VPB_BIT_SHARED) == 0)
757 if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER))
763 vm_page_xunbusy_maybelocked(vm_page_t m)
767 vm_page_assert_xbusied(m);
769 lockacq = !mtx_owned(vm_page_lockptr(m));
773 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
779 * vm_page_xunbusy_hard:
781 * Called after the first try the exclusive unbusy of a page failed.
782 * It is assumed that the waiters bit is on.
785 vm_page_xunbusy_hard(vm_page_t m)
788 vm_page_assert_xbusied(m);
791 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
799 * Wakeup anyone waiting for the page.
800 * The ownership bits do not change.
802 * The given page must be locked.
805 vm_page_flash(vm_page_t m)
809 vm_page_lock_assert(m, MA_OWNED);
813 if ((x & VPB_BIT_WAITERS) == 0)
815 if (atomic_cmpset_int(&m->busy_lock, x,
816 x & (~VPB_BIT_WAITERS)))
823 * Keep page from being freed by the page daemon
824 * much of the same effect as wiring, except much lower
825 * overhead and should be used only for *very* temporary
826 * holding ("wiring").
829 vm_page_hold(vm_page_t mem)
832 vm_page_lock_assert(mem, MA_OWNED);
837 vm_page_unhold(vm_page_t mem)
840 vm_page_lock_assert(mem, MA_OWNED);
841 KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!"));
843 if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0)
844 vm_page_free_toq(mem);
848 * vm_page_unhold_pages:
850 * Unhold each of the pages that is referenced by the given array.
853 vm_page_unhold_pages(vm_page_t *ma, int count)
855 struct mtx *mtx, *new_mtx;
858 for (; count != 0; count--) {
860 * Avoid releasing and reacquiring the same page lock.
862 new_mtx = vm_page_lockptr(*ma);
863 if (mtx != new_mtx) {
877 PHYS_TO_VM_PAGE(vm_paddr_t pa)
881 #ifdef VM_PHYSSEG_SPARSE
882 m = vm_phys_paddr_to_vm_page(pa);
884 m = vm_phys_fictitious_to_vm_page(pa);
886 #elif defined(VM_PHYSSEG_DENSE)
890 if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
891 m = &vm_page_array[pi - first_page];
894 return (vm_phys_fictitious_to_vm_page(pa));
896 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
903 * Create a fictitious page with the specified physical address and
904 * memory attribute. The memory attribute is the only the machine-
905 * dependent aspect of a fictitious page that must be initialized.
908 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
912 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
913 vm_page_initfake(m, paddr, memattr);
918 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
921 if ((m->flags & PG_FICTITIOUS) != 0) {
923 * The page's memattr might have changed since the
924 * previous initialization. Update the pmap to the
929 m->phys_addr = paddr;
931 /* Fictitious pages don't use "segind". */
932 m->flags = PG_FICTITIOUS;
933 /* Fictitious pages don't use "order" or "pool". */
934 m->oflags = VPO_UNMANAGED;
935 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
939 pmap_page_set_memattr(m, memattr);
945 * Release a fictitious page.
948 vm_page_putfake(vm_page_t m)
951 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
952 KASSERT((m->flags & PG_FICTITIOUS) != 0,
953 ("vm_page_putfake: bad page %p", m));
954 uma_zfree(fakepg_zone, m);
958 * vm_page_updatefake:
960 * Update the given fictitious page to the specified physical address and
964 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
967 KASSERT((m->flags & PG_FICTITIOUS) != 0,
968 ("vm_page_updatefake: bad page %p", m));
969 m->phys_addr = paddr;
970 pmap_page_set_memattr(m, memattr);
979 vm_page_free(vm_page_t m)
982 m->flags &= ~PG_ZERO;
989 * Free a page to the zerod-pages queue
992 vm_page_free_zero(vm_page_t m)
1000 * Unbusy and handle the page queueing for a page from the VOP_GETPAGES()
1001 * array which was optionally read ahead or behind.
1004 vm_page_readahead_finish(vm_page_t m)
1007 /* We shouldn't put invalid pages on queues. */
1008 KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m));
1011 * Since the page is not the actually needed one, whether it should
1012 * be activated or deactivated is not obvious. Empirical results
1013 * have shown that deactivating the page is usually the best choice,
1014 * unless the page is wanted by another thread.
1017 if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1018 vm_page_activate(m);
1020 vm_page_deactivate(m);
1026 * vm_page_sleep_if_busy:
1028 * Sleep and release the page queues lock if the page is busied.
1029 * Returns TRUE if the thread slept.
1031 * The given page must be unlocked and object containing it must
1035 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1039 vm_page_lock_assert(m, MA_NOTOWNED);
1040 VM_OBJECT_ASSERT_WLOCKED(m->object);
1042 if (vm_page_busied(m)) {
1044 * The page-specific object must be cached because page
1045 * identity can change during the sleep, causing the
1046 * re-lock of a different object.
1047 * It is assumed that a reference to the object is already
1048 * held by the callers.
1052 VM_OBJECT_WUNLOCK(obj);
1053 vm_page_busy_sleep(m, msg);
1054 VM_OBJECT_WLOCK(obj);
1061 * vm_page_dirty_KBI: [ internal use only ]
1063 * Set all bits in the page's dirty field.
1065 * The object containing the specified page must be locked if the
1066 * call is made from the machine-independent layer.
1068 * See vm_page_clear_dirty_mask().
1070 * This function should only be called by vm_page_dirty().
1073 vm_page_dirty_KBI(vm_page_t m)
1076 /* These assertions refer to this operation by its public name. */
1077 KASSERT((m->flags & PG_CACHED) == 0,
1078 ("vm_page_dirty: page in cache!"));
1079 KASSERT(m->valid == VM_PAGE_BITS_ALL,
1080 ("vm_page_dirty: page is invalid!"));
1081 m->dirty = VM_PAGE_BITS_ALL;
1085 * vm_page_insert: [ internal use only ]
1087 * Inserts the given mem entry into the object and object list.
1089 * The object must be locked.
1092 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1096 VM_OBJECT_ASSERT_WLOCKED(object);
1097 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1098 return (vm_page_insert_after(m, object, pindex, mpred));
1102 * vm_page_insert_after:
1104 * Inserts the page "m" into the specified object at offset "pindex".
1106 * The page "mpred" must immediately precede the offset "pindex" within
1107 * the specified object.
1109 * The object must be locked.
1112 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1119 VM_OBJECT_ASSERT_WLOCKED(object);
1120 KASSERT(m->object == NULL,
1121 ("vm_page_insert_after: page already inserted"));
1122 if (mpred != NULL) {
1123 KASSERT(mpred->object == object,
1124 ("vm_page_insert_after: object doesn't contain mpred"));
1125 KASSERT(mpred->pindex < pindex,
1126 ("vm_page_insert_after: mpred doesn't precede pindex"));
1127 msucc = TAILQ_NEXT(mpred, listq);
1129 msucc = TAILQ_FIRST(&object->memq);
1131 KASSERT(msucc->pindex > pindex,
1132 ("vm_page_insert_after: msucc doesn't succeed pindex"));
1135 * Record the object/offset pair in this page
1143 * Now link into the object's ordered list of backed pages.
1145 if (vm_radix_insert(&object->rtree, m)) {
1150 vm_page_insert_radixdone(m, object, mpred);
1155 * vm_page_insert_radixdone:
1157 * Complete page "m" insertion into the specified object after the
1158 * radix trie hooking.
1160 * The page "mpred" must precede the offset "m->pindex" within the
1163 * The object must be locked.
1166 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1169 VM_OBJECT_ASSERT_WLOCKED(object);
1170 KASSERT(object != NULL && m->object == object,
1171 ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1172 if (mpred != NULL) {
1173 KASSERT(mpred->object == object,
1174 ("vm_page_insert_after: object doesn't contain mpred"));
1175 KASSERT(mpred->pindex < m->pindex,
1176 ("vm_page_insert_after: mpred doesn't precede pindex"));
1180 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1182 TAILQ_INSERT_HEAD(&object->memq, m, listq);
1185 * Show that the object has one more resident page.
1187 object->resident_page_count++;
1190 * Hold the vnode until the last page is released.
1192 if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1193 vhold(object->handle);
1196 * Since we are inserting a new and possibly dirty page,
1197 * update the object's OBJ_MIGHTBEDIRTY flag.
1199 if (pmap_page_is_write_mapped(m))
1200 vm_object_set_writeable_dirty(object);
1206 * Removes the given mem entry from the object/offset-page
1207 * table and the object page list, but do not invalidate/terminate
1208 * the backing store.
1210 * The object must be locked. The page must be locked if it is managed.
1213 vm_page_remove(vm_page_t m)
1217 if ((m->oflags & VPO_UNMANAGED) == 0)
1218 vm_page_assert_locked(m);
1219 if ((object = m->object) == NULL)
1221 VM_OBJECT_ASSERT_WLOCKED(object);
1222 if (vm_page_xbusied(m))
1223 vm_page_xunbusy_maybelocked(m);
1226 * Now remove from the object's list of backed pages.
1228 vm_radix_remove(&object->rtree, m->pindex);
1229 TAILQ_REMOVE(&object->memq, m, listq);
1232 * And show that the object has one fewer resident page.
1234 object->resident_page_count--;
1237 * The vnode may now be recycled.
1239 if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1240 vdrop(object->handle);
1248 * Returns the page associated with the object/offset
1249 * pair specified; if none is found, NULL is returned.
1251 * The object must be locked.
1254 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1257 VM_OBJECT_ASSERT_LOCKED(object);
1258 return (vm_radix_lookup(&object->rtree, pindex));
1262 * vm_page_find_least:
1264 * Returns the page associated with the object with least pindex
1265 * greater than or equal to the parameter pindex, or NULL.
1267 * The object must be locked.
1270 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1274 VM_OBJECT_ASSERT_LOCKED(object);
1275 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1276 m = vm_radix_lookup_ge(&object->rtree, pindex);
1281 * Returns the given page's successor (by pindex) within the object if it is
1282 * resident; if none is found, NULL is returned.
1284 * The object must be locked.
1287 vm_page_next(vm_page_t m)
1291 VM_OBJECT_ASSERT_LOCKED(m->object);
1292 if ((next = TAILQ_NEXT(m, listq)) != NULL &&
1293 next->pindex != m->pindex + 1)
1299 * Returns the given page's predecessor (by pindex) within the object if it is
1300 * resident; if none is found, NULL is returned.
1302 * The object must be locked.
1305 vm_page_prev(vm_page_t m)
1309 VM_OBJECT_ASSERT_LOCKED(m->object);
1310 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
1311 prev->pindex != m->pindex - 1)
1317 * Uses the page mnew as a replacement for an existing page at index
1318 * pindex which must be already present in the object.
1320 * The existing page must not be on a paging queue.
1323 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1327 VM_OBJECT_ASSERT_WLOCKED(object);
1328 KASSERT(mnew->object == NULL,
1329 ("vm_page_replace: page already in object"));
1332 * This function mostly follows vm_page_insert() and
1333 * vm_page_remove() without the radix, object count and vnode
1334 * dance. Double check such functions for more comments.
1337 mnew->object = object;
1338 mnew->pindex = pindex;
1339 mold = vm_radix_replace(&object->rtree, mnew);
1340 KASSERT(mold->queue == PQ_NONE,
1341 ("vm_page_replace: mold is on a paging queue"));
1343 /* Keep the resident page list in sorted order. */
1344 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1345 TAILQ_REMOVE(&object->memq, mold, listq);
1347 mold->object = NULL;
1348 vm_page_xunbusy_maybelocked(mold);
1351 * The object's resident_page_count does not change because we have
1352 * swapped one page for another, but OBJ_MIGHTBEDIRTY.
1354 if (pmap_page_is_write_mapped(mnew))
1355 vm_object_set_writeable_dirty(object);
1362 * Move the given memory entry from its
1363 * current object to the specified target object/offset.
1365 * Note: swap associated with the page must be invalidated by the move. We
1366 * have to do this for several reasons: (1) we aren't freeing the
1367 * page, (2) we are dirtying the page, (3) the VM system is probably
1368 * moving the page from object A to B, and will then later move
1369 * the backing store from A to B and we can't have a conflict.
1371 * Note: we *always* dirty the page. It is necessary both for the
1372 * fact that we moved it, and because we may be invalidating
1373 * swap. If the page is on the cache, we have to deactivate it
1374 * or vm_page_dirty() will panic. Dirty pages are not allowed
1377 * The objects must be locked.
1380 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1385 VM_OBJECT_ASSERT_WLOCKED(new_object);
1387 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1388 KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1389 ("vm_page_rename: pindex already renamed"));
1392 * Create a custom version of vm_page_insert() which does not depend
1393 * by m_prev and can cheat on the implementation aspects of the
1397 m->pindex = new_pindex;
1398 if (vm_radix_insert(&new_object->rtree, m)) {
1404 * The operation cannot fail anymore. The removal must happen before
1405 * the listq iterator is tainted.
1411 /* Return back to the new pindex to complete vm_page_insert(). */
1412 m->pindex = new_pindex;
1413 m->object = new_object;
1415 vm_page_insert_radixdone(m, new_object, mpred);
1421 * Convert all of the given object's cached pages that have a
1422 * pindex within the given range into free pages. If the value
1423 * zero is given for "end", then the range's upper bound is
1424 * infinity. If the given object is backed by a vnode and it
1425 * transitions from having one or more cached pages to none, the
1426 * vnode's hold count is reduced.
1429 vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1434 mtx_lock(&vm_page_queue_free_mtx);
1435 if (__predict_false(vm_radix_is_empty(&object->cache))) {
1436 mtx_unlock(&vm_page_queue_free_mtx);
1439 while ((m = vm_radix_lookup_ge(&object->cache, start)) != NULL) {
1440 if (end != 0 && m->pindex >= end)
1442 vm_radix_remove(&object->cache, m->pindex);
1443 vm_page_cache_turn_free(m);
1445 empty = vm_radix_is_empty(&object->cache);
1446 mtx_unlock(&vm_page_queue_free_mtx);
1447 if (object->type == OBJT_VNODE && empty)
1448 vdrop(object->handle);
1452 * Returns the cached page that is associated with the given
1453 * object and offset. If, however, none exists, returns NULL.
1455 * The free page queue must be locked.
1457 static inline vm_page_t
1458 vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex)
1461 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1462 return (vm_radix_lookup(&object->cache, pindex));
1466 * Remove the given cached page from its containing object's
1467 * collection of cached pages.
1469 * The free page queue must be locked.
1472 vm_page_cache_remove(vm_page_t m)
1475 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1476 KASSERT((m->flags & PG_CACHED) != 0,
1477 ("vm_page_cache_remove: page %p is not cached", m));
1478 vm_radix_remove(&m->object->cache, m->pindex);
1480 vm_cnt.v_cache_count--;
1484 * Transfer all of the cached pages with offset greater than or
1485 * equal to 'offidxstart' from the original object's cache to the
1486 * new object's cache. However, any cached pages with offset
1487 * greater than or equal to the new object's size are kept in the
1488 * original object. Initially, the new object's cache must be
1489 * empty. Offset 'offidxstart' in the original object must
1490 * correspond to offset zero in the new object.
1492 * The new object must be locked.
1495 vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart,
1496 vm_object_t new_object)
1501 * Insertion into an object's collection of cached pages
1502 * requires the object to be locked. In contrast, removal does
1505 VM_OBJECT_ASSERT_WLOCKED(new_object);
1506 KASSERT(vm_radix_is_empty(&new_object->cache),
1507 ("vm_page_cache_transfer: object %p has cached pages",
1509 mtx_lock(&vm_page_queue_free_mtx);
1510 while ((m = vm_radix_lookup_ge(&orig_object->cache,
1511 offidxstart)) != NULL) {
1513 * Transfer all of the pages with offset greater than or
1514 * equal to 'offidxstart' from the original object's
1515 * cache to the new object's cache.
1517 if ((m->pindex - offidxstart) >= new_object->size)
1519 vm_radix_remove(&orig_object->cache, m->pindex);
1520 /* Update the page's object and offset. */
1521 m->object = new_object;
1522 m->pindex -= offidxstart;
1523 if (vm_radix_insert(&new_object->cache, m))
1524 vm_page_cache_turn_free(m);
1526 mtx_unlock(&vm_page_queue_free_mtx);
1530 * Returns TRUE if a cached page is associated with the given object and
1531 * offset, and FALSE otherwise.
1533 * The object must be locked.
1536 vm_page_is_cached(vm_object_t object, vm_pindex_t pindex)
1541 * Insertion into an object's collection of cached pages requires the
1542 * object to be locked. Therefore, if the object is locked and the
1543 * object's collection is empty, there is no need to acquire the free
1544 * page queues lock in order to prove that the specified page doesn't
1547 VM_OBJECT_ASSERT_WLOCKED(object);
1548 if (__predict_true(vm_object_cache_is_empty(object)))
1550 mtx_lock(&vm_page_queue_free_mtx);
1551 m = vm_page_cache_lookup(object, pindex);
1552 mtx_unlock(&vm_page_queue_free_mtx);
1559 * Allocate and return a page that is associated with the specified
1560 * object and offset pair. By default, this page is exclusive busied.
1562 * The caller must always specify an allocation class.
1564 * allocation classes:
1565 * VM_ALLOC_NORMAL normal process request
1566 * VM_ALLOC_SYSTEM system *really* needs a page
1567 * VM_ALLOC_INTERRUPT interrupt time request
1569 * optional allocation flags:
1570 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1571 * intends to allocate
1572 * VM_ALLOC_IFCACHED return page only if it is cached
1573 * VM_ALLOC_IFNOTCACHED return NULL, do not reactivate if the page
1575 * VM_ALLOC_NOBUSY do not exclusive busy the page
1576 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1577 * VM_ALLOC_NOOBJ page is not associated with an object and
1578 * should not be exclusive busy
1579 * VM_ALLOC_SBUSY shared busy the allocated page
1580 * VM_ALLOC_WIRED wire the allocated page
1581 * VM_ALLOC_ZERO prefer a zeroed page
1583 * This routine may not sleep.
1586 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1588 struct vnode *vp = NULL;
1589 vm_object_t m_object;
1591 int flags, req_class;
1593 mpred = 0; /* XXX: pacify gcc */
1594 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1595 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1596 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1597 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1598 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object,
1601 VM_OBJECT_ASSERT_WLOCKED(object);
1603 req_class = req & VM_ALLOC_CLASS_MASK;
1606 * The page daemon is allowed to dig deeper into the free page list.
1608 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1609 req_class = VM_ALLOC_SYSTEM;
1611 if (object != NULL) {
1612 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1613 KASSERT(mpred == NULL || mpred->pindex != pindex,
1614 ("vm_page_alloc: pindex already allocated"));
1618 * The page allocation request can came from consumers which already
1619 * hold the free page queue mutex, like vm_page_insert() in
1622 mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE);
1623 if (vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_free_reserved ||
1624 (req_class == VM_ALLOC_SYSTEM &&
1625 vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_interrupt_free_min) ||
1626 (req_class == VM_ALLOC_INTERRUPT &&
1627 vm_cnt.v_free_count + vm_cnt.v_cache_count > 0)) {
1629 * Allocate from the free queue if the number of free pages
1630 * exceeds the minimum for the request class.
1632 if (object != NULL &&
1633 (m = vm_page_cache_lookup(object, pindex)) != NULL) {
1634 if ((req & VM_ALLOC_IFNOTCACHED) != 0) {
1635 mtx_unlock(&vm_page_queue_free_mtx);
1638 if (vm_phys_unfree_page(m))
1639 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0);
1640 #if VM_NRESERVLEVEL > 0
1641 else if (!vm_reserv_reactivate_page(m))
1645 panic("vm_page_alloc: cache page %p is missing"
1646 " from the free queue", m);
1647 } else if ((req & VM_ALLOC_IFCACHED) != 0) {
1648 mtx_unlock(&vm_page_queue_free_mtx);
1650 #if VM_NRESERVLEVEL > 0
1651 } else if (object == NULL || (object->flags & (OBJ_COLORED |
1652 OBJ_FICTITIOUS)) != OBJ_COLORED || (m =
1653 vm_reserv_alloc_page(object, pindex, mpred)) == NULL) {
1657 m = vm_phys_alloc_pages(object != NULL ?
1658 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1659 #if VM_NRESERVLEVEL > 0
1660 if (m == NULL && vm_reserv_reclaim_inactive()) {
1661 m = vm_phys_alloc_pages(object != NULL ?
1662 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1669 * Not allocatable, give up.
1671 mtx_unlock(&vm_page_queue_free_mtx);
1672 atomic_add_int(&vm_pageout_deficit,
1673 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1674 pagedaemon_wakeup();
1679 * At this point we had better have found a good page.
1681 KASSERT(m != NULL, ("vm_page_alloc: missing page"));
1682 KASSERT(m->queue == PQ_NONE,
1683 ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue));
1684 KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m));
1685 KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m));
1686 KASSERT(!vm_page_sbusied(m),
1687 ("vm_page_alloc: page %p is busy", m));
1688 KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m));
1689 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1690 ("vm_page_alloc: page %p has unexpected memattr %d", m,
1691 pmap_page_get_memattr(m)));
1692 if ((m->flags & PG_CACHED) != 0) {
1693 KASSERT((m->flags & PG_ZERO) == 0,
1694 ("vm_page_alloc: cached page %p is PG_ZERO", m));
1695 KASSERT(m->valid != 0,
1696 ("vm_page_alloc: cached page %p is invalid", m));
1697 if (m->object == object && m->pindex == pindex)
1698 vm_cnt.v_reactivated++;
1701 m_object = m->object;
1702 vm_page_cache_remove(m);
1703 if (m_object->type == OBJT_VNODE &&
1704 vm_object_cache_is_empty(m_object))
1705 vp = m_object->handle;
1707 KASSERT(m->valid == 0,
1708 ("vm_page_alloc: free page %p is valid", m));
1709 vm_phys_freecnt_adj(m, -1);
1710 if ((m->flags & PG_ZERO) != 0)
1711 vm_page_zero_count--;
1713 mtx_unlock(&vm_page_queue_free_mtx);
1716 * Initialize the page. Only the PG_ZERO flag is inherited.
1719 if ((req & VM_ALLOC_ZERO) != 0)
1722 if ((req & VM_ALLOC_NODUMP) != 0)
1726 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1728 m->busy_lock = VPB_UNBUSIED;
1729 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1730 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1731 if ((req & VM_ALLOC_SBUSY) != 0)
1732 m->busy_lock = VPB_SHARERS_WORD(1);
1733 if (req & VM_ALLOC_WIRED) {
1735 * The page lock is not required for wiring a page until that
1736 * page is inserted into the object.
1738 atomic_add_int(&vm_cnt.v_wire_count, 1);
1743 if (object != NULL) {
1744 if (vm_page_insert_after(m, object, pindex, mpred)) {
1745 /* See the comment below about hold count. */
1748 pagedaemon_wakeup();
1749 if (req & VM_ALLOC_WIRED) {
1750 atomic_subtract_int(&vm_cnt.v_wire_count, 1);
1754 m->oflags = VPO_UNMANAGED;
1759 /* Ignore device objects; the pager sets "memattr" for them. */
1760 if (object->memattr != VM_MEMATTR_DEFAULT &&
1761 (object->flags & OBJ_FICTITIOUS) == 0)
1762 pmap_page_set_memattr(m, object->memattr);
1767 * The following call to vdrop() must come after the above call
1768 * to vm_page_insert() in case both affect the same object and
1769 * vnode. Otherwise, the affected vnode's hold count could
1770 * temporarily become zero.
1776 * Don't wakeup too often - wakeup the pageout daemon when
1777 * we would be nearly out of memory.
1779 if (vm_paging_needed())
1780 pagedaemon_wakeup();
1786 vm_page_alloc_contig_vdrop(struct spglist *lst)
1789 while (!SLIST_EMPTY(lst)) {
1790 vdrop((struct vnode *)SLIST_FIRST(lst)-> plinks.s.pv);
1791 SLIST_REMOVE_HEAD(lst, plinks.s.ss);
1796 * vm_page_alloc_contig:
1798 * Allocate a contiguous set of physical pages of the given size "npages"
1799 * from the free lists. All of the physical pages must be at or above
1800 * the given physical address "low" and below the given physical address
1801 * "high". The given value "alignment" determines the alignment of the
1802 * first physical page in the set. If the given value "boundary" is
1803 * non-zero, then the set of physical pages cannot cross any physical
1804 * address boundary that is a multiple of that value. Both "alignment"
1805 * and "boundary" must be a power of two.
1807 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1808 * then the memory attribute setting for the physical pages is configured
1809 * to the object's memory attribute setting. Otherwise, the memory
1810 * attribute setting for the physical pages is configured to "memattr",
1811 * overriding the object's memory attribute setting. However, if the
1812 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1813 * memory attribute setting for the physical pages cannot be configured
1814 * to VM_MEMATTR_DEFAULT.
1816 * The caller must always specify an allocation class.
1818 * allocation classes:
1819 * VM_ALLOC_NORMAL normal process request
1820 * VM_ALLOC_SYSTEM system *really* needs a page
1821 * VM_ALLOC_INTERRUPT interrupt time request
1823 * optional allocation flags:
1824 * VM_ALLOC_NOBUSY do not exclusive busy the page
1825 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1826 * VM_ALLOC_NOOBJ page is not associated with an object and
1827 * should not be exclusive busy
1828 * VM_ALLOC_SBUSY shared busy the allocated page
1829 * VM_ALLOC_WIRED wire the allocated page
1830 * VM_ALLOC_ZERO prefer a zeroed page
1832 * This routine may not sleep.
1835 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1836 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1837 vm_paddr_t boundary, vm_memattr_t memattr)
1840 struct spglist deferred_vdrop_list;
1841 vm_page_t m, m_tmp, m_ret;
1845 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1846 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1847 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1848 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1849 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object,
1851 if (object != NULL) {
1852 VM_OBJECT_ASSERT_WLOCKED(object);
1853 KASSERT(object->type == OBJT_PHYS,
1854 ("vm_page_alloc_contig: object %p isn't OBJT_PHYS",
1857 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
1858 req_class = req & VM_ALLOC_CLASS_MASK;
1861 * The page daemon is allowed to dig deeper into the free page list.
1863 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1864 req_class = VM_ALLOC_SYSTEM;
1866 SLIST_INIT(&deferred_vdrop_list);
1867 mtx_lock(&vm_page_queue_free_mtx);
1868 if (vm_cnt.v_free_count + vm_cnt.v_cache_count >= npages +
1869 vm_cnt.v_free_reserved || (req_class == VM_ALLOC_SYSTEM &&
1870 vm_cnt.v_free_count + vm_cnt.v_cache_count >= npages +
1871 vm_cnt.v_interrupt_free_min) || (req_class == VM_ALLOC_INTERRUPT &&
1872 vm_cnt.v_free_count + vm_cnt.v_cache_count >= npages)) {
1873 #if VM_NRESERVLEVEL > 0
1875 if (object == NULL || (object->flags & OBJ_COLORED) == 0 ||
1876 (m_ret = vm_reserv_alloc_contig(object, pindex, npages,
1877 low, high, alignment, boundary)) == NULL)
1879 m_ret = vm_phys_alloc_contig(npages, low, high,
1880 alignment, boundary);
1882 mtx_unlock(&vm_page_queue_free_mtx);
1883 atomic_add_int(&vm_pageout_deficit, npages);
1884 pagedaemon_wakeup();
1888 for (m = m_ret; m < &m_ret[npages]; m++) {
1889 drop = vm_page_alloc_init(m);
1892 * Enqueue the vnode for deferred vdrop().
1894 m->plinks.s.pv = drop;
1895 SLIST_INSERT_HEAD(&deferred_vdrop_list, m,
1900 #if VM_NRESERVLEVEL > 0
1901 if (vm_reserv_reclaim_contig(npages, low, high, alignment,
1906 mtx_unlock(&vm_page_queue_free_mtx);
1911 * Initialize the pages. Only the PG_ZERO flag is inherited.
1914 if ((req & VM_ALLOC_ZERO) != 0)
1916 if ((req & VM_ALLOC_NODUMP) != 0)
1918 if ((req & VM_ALLOC_WIRED) != 0)
1919 atomic_add_int(&vm_cnt.v_wire_count, npages);
1920 if (object != NULL) {
1921 if (object->memattr != VM_MEMATTR_DEFAULT &&
1922 memattr == VM_MEMATTR_DEFAULT)
1923 memattr = object->memattr;
1925 for (m = m_ret; m < &m_ret[npages]; m++) {
1927 m->flags = (m->flags | PG_NODUMP) & flags;
1928 m->busy_lock = VPB_UNBUSIED;
1929 if (object != NULL) {
1930 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
1931 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1932 if ((req & VM_ALLOC_SBUSY) != 0)
1933 m->busy_lock = VPB_SHARERS_WORD(1);
1935 if ((req & VM_ALLOC_WIRED) != 0)
1937 /* Unmanaged pages don't use "act_count". */
1938 m->oflags = VPO_UNMANAGED;
1939 if (object != NULL) {
1940 if (vm_page_insert(m, object, pindex)) {
1941 vm_page_alloc_contig_vdrop(
1942 &deferred_vdrop_list);
1943 if (vm_paging_needed())
1944 pagedaemon_wakeup();
1945 if ((req & VM_ALLOC_WIRED) != 0)
1946 atomic_subtract_int(&vm_cnt.v_wire_count,
1948 for (m_tmp = m, m = m_ret;
1949 m < &m_ret[npages]; m++) {
1950 if ((req & VM_ALLOC_WIRED) != 0)
1954 m->oflags |= VPO_UNMANAGED;
1962 if (memattr != VM_MEMATTR_DEFAULT)
1963 pmap_page_set_memattr(m, memattr);
1966 vm_page_alloc_contig_vdrop(&deferred_vdrop_list);
1967 if (vm_paging_needed())
1968 pagedaemon_wakeup();
1973 * Initialize a page that has been freshly dequeued from a freelist.
1974 * The caller has to drop the vnode returned, if it is not NULL.
1976 * This function may only be used to initialize unmanaged pages.
1978 * To be called with vm_page_queue_free_mtx held.
1980 static struct vnode *
1981 vm_page_alloc_init(vm_page_t m)
1984 vm_object_t m_object;
1986 KASSERT(m->queue == PQ_NONE,
1987 ("vm_page_alloc_init: page %p has unexpected queue %d",
1989 KASSERT(m->wire_count == 0,
1990 ("vm_page_alloc_init: page %p is wired", m));
1991 KASSERT(m->hold_count == 0,
1992 ("vm_page_alloc_init: page %p is held", m));
1993 KASSERT(!vm_page_sbusied(m),
1994 ("vm_page_alloc_init: page %p is busy", m));
1995 KASSERT(m->dirty == 0,
1996 ("vm_page_alloc_init: page %p is dirty", m));
1997 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1998 ("vm_page_alloc_init: page %p has unexpected memattr %d",
1999 m, pmap_page_get_memattr(m)));
2000 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2002 if ((m->flags & PG_CACHED) != 0) {
2003 KASSERT((m->flags & PG_ZERO) == 0,
2004 ("vm_page_alloc_init: cached page %p is PG_ZERO", m));
2006 m_object = m->object;
2007 vm_page_cache_remove(m);
2008 if (m_object->type == OBJT_VNODE &&
2009 vm_object_cache_is_empty(m_object))
2010 drop = m_object->handle;
2012 KASSERT(m->valid == 0,
2013 ("vm_page_alloc_init: free page %p is valid", m));
2014 vm_phys_freecnt_adj(m, -1);
2015 if ((m->flags & PG_ZERO) != 0)
2016 vm_page_zero_count--;
2022 * vm_page_alloc_freelist:
2024 * Allocate a physical page from the specified free page list.
2026 * The caller must always specify an allocation class.
2028 * allocation classes:
2029 * VM_ALLOC_NORMAL normal process request
2030 * VM_ALLOC_SYSTEM system *really* needs a page
2031 * VM_ALLOC_INTERRUPT interrupt time request
2033 * optional allocation flags:
2034 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
2035 * intends to allocate
2036 * VM_ALLOC_WIRED wire the allocated page
2037 * VM_ALLOC_ZERO prefer a zeroed page
2039 * This routine may not sleep.
2042 vm_page_alloc_freelist(int flind, int req)
2049 req_class = req & VM_ALLOC_CLASS_MASK;
2052 * The page daemon is allowed to dig deeper into the free page list.
2054 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2055 req_class = VM_ALLOC_SYSTEM;
2058 * Do not allocate reserved pages unless the req has asked for it.
2060 mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE);
2061 if (vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_free_reserved ||
2062 (req_class == VM_ALLOC_SYSTEM &&
2063 vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_interrupt_free_min) ||
2064 (req_class == VM_ALLOC_INTERRUPT &&
2065 vm_cnt.v_free_count + vm_cnt.v_cache_count > 0))
2066 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0);
2068 mtx_unlock(&vm_page_queue_free_mtx);
2069 atomic_add_int(&vm_pageout_deficit,
2070 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
2071 pagedaemon_wakeup();
2075 mtx_unlock(&vm_page_queue_free_mtx);
2078 drop = vm_page_alloc_init(m);
2079 mtx_unlock(&vm_page_queue_free_mtx);
2082 * Initialize the page. Only the PG_ZERO flag is inherited.
2086 if ((req & VM_ALLOC_ZERO) != 0)
2089 if ((req & VM_ALLOC_WIRED) != 0) {
2091 * The page lock is not required for wiring a page that does
2092 * not belong to an object.
2094 atomic_add_int(&vm_cnt.v_wire_count, 1);
2097 /* Unmanaged pages don't use "act_count". */
2098 m->oflags = VPO_UNMANAGED;
2101 if (vm_paging_needed())
2102 pagedaemon_wakeup();
2106 #define VPSC_ANY 0 /* No restrictions. */
2107 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */
2108 #define VPSC_NOSUPER 2 /* Skip superpages. */
2111 * vm_page_scan_contig:
2113 * Scan vm_page_array[] between the specified entries "m_start" and
2114 * "m_end" for a run of contiguous physical pages that satisfy the
2115 * specified conditions, and return the lowest page in the run. The
2116 * specified "alignment" determines the alignment of the lowest physical
2117 * page in the run. If the specified "boundary" is non-zero, then the
2118 * run of physical pages cannot span a physical address that is a
2119 * multiple of "boundary".
2121 * "m_end" is never dereferenced, so it need not point to a vm_page
2122 * structure within vm_page_array[].
2124 * "npages" must be greater than zero. "m_start" and "m_end" must not
2125 * span a hole (or discontiguity) in the physical address space. Both
2126 * "alignment" and "boundary" must be a power of two.
2129 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2130 u_long alignment, vm_paddr_t boundary, int options)
2132 struct mtx *m_mtx, *new_mtx;
2136 #if VM_NRESERVLEVEL > 0
2139 int m_inc, order, run_ext, run_len;
2141 KASSERT(npages > 0, ("npages is 0"));
2142 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2143 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2147 for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2148 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2149 ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2152 * If the current page would be the start of a run, check its
2153 * physical address against the end, alignment, and boundary
2154 * conditions. If it doesn't satisfy these conditions, either
2155 * terminate the scan or advance to the next page that
2156 * satisfies the failed condition.
2159 KASSERT(m_run == NULL, ("m_run != NULL"));
2160 if (m + npages > m_end)
2162 pa = VM_PAGE_TO_PHYS(m);
2163 if ((pa & (alignment - 1)) != 0) {
2164 m_inc = atop(roundup2(pa, alignment) - pa);
2167 if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2169 m_inc = atop(roundup2(pa, boundary) - pa);
2173 KASSERT(m_run != NULL, ("m_run == NULL"));
2176 * Avoid releasing and reacquiring the same page lock.
2178 new_mtx = vm_page_lockptr(m);
2179 if (m_mtx != new_mtx) {
2187 if (m->wire_count != 0 || m->hold_count != 0)
2189 #if VM_NRESERVLEVEL > 0
2190 else if ((level = vm_reserv_level(m)) >= 0 &&
2191 (options & VPSC_NORESERV) != 0) {
2193 /* Advance to the end of the reservation. */
2194 pa = VM_PAGE_TO_PHYS(m);
2195 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2199 else if ((object = m->object) != NULL) {
2201 * The page is considered eligible for relocation if
2202 * and only if it could be laundered or reclaimed by
2205 if (!VM_OBJECT_TRYRLOCK(object)) {
2207 VM_OBJECT_RLOCK(object);
2209 if (m->object != object) {
2211 * The page may have been freed.
2213 VM_OBJECT_RUNLOCK(object);
2215 } else if (m->wire_count != 0 ||
2216 m->hold_count != 0) {
2221 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2222 ("page %p is PG_UNHOLDFREE", m));
2223 /* Don't care: PG_NODUMP, PG_WINATCFLS, PG_ZERO. */
2224 if (object->type != OBJT_DEFAULT &&
2225 object->type != OBJT_SWAP &&
2226 object->type != OBJT_VNODE)
2228 else if ((m->flags & PG_CACHED) != 0 ||
2229 m != vm_page_lookup(object, m->pindex)) {
2231 * The page is cached or recently converted
2232 * from cached to free.
2234 #if VM_NRESERVLEVEL > 0
2237 * The page is reserved. Extend the
2238 * current run by one page.
2243 if ((order = m->order) < VM_NFREEORDER) {
2245 * The page is enqueued in the
2246 * physical memory allocator's cache/
2247 * free page queues. Moreover, it is
2248 * the first page in a power-of-two-
2249 * sized run of contiguous cache/free
2250 * pages. Add these pages to the end
2251 * of the current run, and jump
2254 run_ext = 1 << order;
2258 #if VM_NRESERVLEVEL > 0
2259 } else if ((options & VPSC_NOSUPER) != 0 &&
2260 (level = vm_reserv_level_iffullpop(m)) >= 0) {
2262 /* Advance to the end of the superpage. */
2263 pa = VM_PAGE_TO_PHYS(m);
2264 m_inc = atop(roundup2(pa + 1,
2265 vm_reserv_size(level)) - pa);
2267 } else if (object->memattr == VM_MEMATTR_DEFAULT &&
2268 m->queue != PQ_NONE && !vm_page_busied(m)) {
2270 * The page is allocated but eligible for
2271 * relocation. Extend the current run by one
2274 KASSERT(pmap_page_get_memattr(m) ==
2276 ("page %p has an unexpected memattr", m));
2277 KASSERT((m->oflags & (VPO_SWAPINPROG |
2278 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2279 ("page %p has unexpected oflags", m));
2280 /* Don't care: VPO_NOSYNC. */
2285 VM_OBJECT_RUNLOCK(object);
2286 #if VM_NRESERVLEVEL > 0
2287 } else if (level >= 0) {
2289 * The page is reserved but not yet allocated. In
2290 * other words, it is still cached or free. Extend
2291 * the current run by one page.
2295 } else if ((order = m->order) < VM_NFREEORDER) {
2297 * The page is enqueued in the physical memory
2298 * allocator's cache/free page queues. Moreover, it
2299 * is the first page in a power-of-two-sized run of
2300 * contiguous cache/free pages. Add these pages to
2301 * the end of the current run, and jump ahead.
2303 run_ext = 1 << order;
2307 * Skip the page for one of the following reasons: (1)
2308 * It is enqueued in the physical memory allocator's
2309 * cache/free page queues. However, it is not the
2310 * first page in a run of contiguous cache/free pages.
2311 * (This case rarely occurs because the scan is
2312 * performed in ascending order.) (2) It is not
2313 * reserved, and it is transitioning from free to
2314 * allocated. (Conversely, the transition from
2315 * allocated to free for managed pages is blocked by
2316 * the page lock.) (3) It is allocated but not
2317 * contained by an object and not wired, e.g.,
2318 * allocated by Xen's balloon driver.
2324 * Extend or reset the current run of pages.
2339 if (run_len >= npages)
2345 * vm_page_reclaim_run:
2347 * Try to relocate each of the allocated virtual pages within the
2348 * specified run of physical pages to a new physical address. Free the
2349 * physical pages underlying the relocated virtual pages. A virtual page
2350 * is relocatable if and only if it could be laundered or reclaimed by
2351 * the page daemon. Whenever possible, a virtual page is relocated to a
2352 * physical address above "high".
2354 * Returns 0 if every physical page within the run was already free or
2355 * just freed by a successful relocation. Otherwise, returns a non-zero
2356 * value indicating why the last attempt to relocate a virtual page was
2359 * "req_class" must be an allocation class.
2362 vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run,
2365 struct mtx *m_mtx, *new_mtx;
2366 struct spglist free;
2369 vm_page_t m, m_end, m_new;
2370 int error, order, req;
2372 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2373 ("req_class is not an allocation class"));
2377 m_end = m_run + npages;
2379 for (; error == 0 && m < m_end; m++) {
2380 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2381 ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2384 * Avoid releasing and reacquiring the same page lock.
2386 new_mtx = vm_page_lockptr(m);
2387 if (m_mtx != new_mtx) {
2394 if (m->wire_count != 0 || m->hold_count != 0)
2396 else if ((object = m->object) != NULL) {
2398 * The page is relocated if and only if it could be
2399 * laundered or reclaimed by the page daemon.
2401 if (!VM_OBJECT_TRYWLOCK(object)) {
2403 VM_OBJECT_WLOCK(object);
2405 if (m->object != object) {
2407 * The page may have been freed.
2409 VM_OBJECT_WUNLOCK(object);
2411 } else if (m->wire_count != 0 ||
2412 m->hold_count != 0) {
2417 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2418 ("page %p is PG_UNHOLDFREE", m));
2419 /* Don't care: PG_NODUMP, PG_WINATCFLS, PG_ZERO. */
2420 if (object->type != OBJT_DEFAULT &&
2421 object->type != OBJT_SWAP &&
2422 object->type != OBJT_VNODE)
2424 else if ((m->flags & PG_CACHED) != 0 ||
2425 m != vm_page_lookup(object, m->pindex)) {
2427 * The page is cached or recently converted
2428 * from cached to free.
2430 VM_OBJECT_WUNLOCK(object);
2432 } else if (object->memattr != VM_MEMATTR_DEFAULT)
2434 else if (m->queue != PQ_NONE && !vm_page_busied(m)) {
2435 KASSERT(pmap_page_get_memattr(m) ==
2437 ("page %p has an unexpected memattr", m));
2438 KASSERT((m->oflags & (VPO_SWAPINPROG |
2439 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2440 ("page %p has unexpected oflags", m));
2441 /* Don't care: VPO_NOSYNC. */
2442 if (m->valid != 0) {
2444 * First, try to allocate a new page
2445 * that is above "high". Failing
2446 * that, try to allocate a new page
2447 * that is below "m_run". Allocate
2448 * the new page between the end of
2449 * "m_run" and "high" only as a last
2452 req = req_class | VM_ALLOC_NOOBJ;
2453 if ((m->flags & PG_NODUMP) != 0)
2454 req |= VM_ALLOC_NODUMP;
2455 if (trunc_page(high) !=
2456 ~(vm_paddr_t)PAGE_MASK) {
2457 m_new = vm_page_alloc_contig(
2462 VM_MEMATTR_DEFAULT);
2465 if (m_new == NULL) {
2466 pa = VM_PAGE_TO_PHYS(m_run);
2467 m_new = vm_page_alloc_contig(
2469 0, pa - 1, PAGE_SIZE, 0,
2470 VM_MEMATTR_DEFAULT);
2472 if (m_new == NULL) {
2474 m_new = vm_page_alloc_contig(
2476 pa, high, PAGE_SIZE, 0,
2477 VM_MEMATTR_DEFAULT);
2479 if (m_new == NULL) {
2483 KASSERT(m_new->wire_count == 0,
2484 ("page %p is wired", m));
2487 * Replace "m" with the new page. For
2488 * vm_page_replace(), "m" must be busy
2489 * and dequeued. Finally, change "m"
2490 * as if vm_page_free() was called.
2492 if (object->ref_count != 0)
2494 m_new->aflags = m->aflags;
2495 KASSERT(m_new->oflags == VPO_UNMANAGED,
2496 ("page %p is managed", m));
2497 m_new->oflags = m->oflags & VPO_NOSYNC;
2498 pmap_copy_page(m, m_new);
2499 m_new->valid = m->valid;
2500 m_new->dirty = m->dirty;
2501 m->flags &= ~PG_ZERO;
2504 vm_page_replace_checked(m_new, object,
2510 * The new page must be deactivated
2511 * before the object is unlocked.
2513 new_mtx = vm_page_lockptr(m_new);
2514 if (m_mtx != new_mtx) {
2519 vm_page_deactivate(m_new);
2521 m->flags &= ~PG_ZERO;
2524 KASSERT(m->dirty == 0,
2525 ("page %p is dirty", m));
2527 SLIST_INSERT_HEAD(&free, m, plinks.s.ss);
2531 VM_OBJECT_WUNLOCK(object);
2534 mtx_lock(&vm_page_queue_free_mtx);
2536 if (order < VM_NFREEORDER) {
2538 * The page is enqueued in the physical memory
2539 * allocator's cache/free page queues.
2540 * Moreover, it is the first page in a power-
2541 * of-two-sized run of contiguous cache/free
2542 * pages. Jump ahead to the last page within
2543 * that run, and continue from there.
2545 m += (1 << order) - 1;
2547 #if VM_NRESERVLEVEL > 0
2548 else if (vm_reserv_is_page_free(m))
2551 mtx_unlock(&vm_page_queue_free_mtx);
2552 if (order == VM_NFREEORDER)
2558 if ((m = SLIST_FIRST(&free)) != NULL) {
2559 mtx_lock(&vm_page_queue_free_mtx);
2561 SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2562 vm_phys_freecnt_adj(m, 1);
2563 #if VM_NRESERVLEVEL > 0
2564 if (!vm_reserv_free_page(m))
2568 vm_phys_free_pages(m, 0);
2569 } while ((m = SLIST_FIRST(&free)) != NULL);
2570 vm_page_zero_idle_wakeup();
2571 vm_page_free_wakeup();
2572 mtx_unlock(&vm_page_queue_free_mtx);
2579 CTASSERT(powerof2(NRUNS));
2581 #define RUN_INDEX(count) ((count) & (NRUNS - 1))
2583 #define MIN_RECLAIM 8
2586 * vm_page_reclaim_contig:
2588 * Reclaim allocated, contiguous physical memory satisfying the specified
2589 * conditions by relocating the virtual pages using that physical memory.
2590 * Returns true if reclamation is successful and false otherwise. Since
2591 * relocation requires the allocation of physical pages, reclamation may
2592 * fail due to a shortage of cache/free pages. When reclamation fails,
2593 * callers are expected to perform VM_WAIT before retrying a failed
2594 * allocation operation, e.g., vm_page_alloc_contig().
2596 * The caller must always specify an allocation class through "req".
2598 * allocation classes:
2599 * VM_ALLOC_NORMAL normal process request
2600 * VM_ALLOC_SYSTEM system *really* needs a page
2601 * VM_ALLOC_INTERRUPT interrupt time request
2603 * The optional allocation flags are ignored.
2605 * "npages" must be greater than zero. Both "alignment" and "boundary"
2606 * must be a power of two.
2609 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
2610 u_long alignment, vm_paddr_t boundary)
2612 vm_paddr_t curr_low;
2613 vm_page_t m_run, m_runs[NRUNS];
2614 u_long count, reclaimed;
2615 int error, i, options, req_class;
2617 KASSERT(npages > 0, ("npages is 0"));
2618 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2619 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2620 req_class = req & VM_ALLOC_CLASS_MASK;
2623 * The page daemon is allowed to dig deeper into the free page list.
2625 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2626 req_class = VM_ALLOC_SYSTEM;
2629 * Return if the number of cached and free pages cannot satisfy the
2630 * requested allocation.
2632 count = vm_cnt.v_free_count + vm_cnt.v_cache_count;
2633 if (count < npages + vm_cnt.v_free_reserved || (count < npages +
2634 vm_cnt.v_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2635 (count < npages && req_class == VM_ALLOC_INTERRUPT))
2639 * Scan up to three times, relaxing the restrictions ("options") on
2640 * the reclamation of reservations and superpages each time.
2642 for (options = VPSC_NORESERV;;) {
2644 * Find the highest runs that satisfy the given constraints
2645 * and restrictions, and record them in "m_runs".
2650 m_run = vm_phys_scan_contig(npages, curr_low, high,
2651 alignment, boundary, options);
2654 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2655 m_runs[RUN_INDEX(count)] = m_run;
2660 * Reclaim the highest runs in LIFO (descending) order until
2661 * the number of reclaimed pages, "reclaimed", is at least
2662 * MIN_RECLAIM. Reset "reclaimed" each time because each
2663 * reclamation is idempotent, and runs will (likely) recur
2664 * from one scan to the next as restrictions are relaxed.
2667 for (i = 0; count > 0 && i < NRUNS; i++) {
2669 m_run = m_runs[RUN_INDEX(count)];
2670 error = vm_page_reclaim_run(req_class, npages, m_run,
2673 reclaimed += npages;
2674 if (reclaimed >= MIN_RECLAIM)
2680 * Either relax the restrictions on the next scan or return if
2681 * the last scan had no restrictions.
2683 if (options == VPSC_NORESERV)
2684 options = VPSC_NOSUPER;
2685 else if (options == VPSC_NOSUPER)
2687 else if (options == VPSC_ANY)
2688 return (reclaimed != 0);
2693 * vm_wait: (also see VM_WAIT macro)
2695 * Sleep until free pages are available for allocation.
2696 * - Called in various places before memory allocations.
2702 mtx_lock(&vm_page_queue_free_mtx);
2703 if (curproc == pageproc) {
2704 vm_pageout_pages_needed = 1;
2705 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx,
2706 PDROP | PSWP, "VMWait", 0);
2708 if (!vm_pageout_wanted) {
2709 vm_pageout_wanted = true;
2710 wakeup(&vm_pageout_wanted);
2712 vm_pages_needed = true;
2713 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM,
2719 * vm_waitpfault: (also see VM_WAITPFAULT macro)
2721 * Sleep until free pages are available for allocation.
2722 * - Called only in vm_fault so that processes page faulting
2723 * can be easily tracked.
2724 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
2725 * processes will be able to grab memory first. Do not change
2726 * this balance without careful testing first.
2732 mtx_lock(&vm_page_queue_free_mtx);
2733 if (!vm_pageout_wanted) {
2734 vm_pageout_wanted = true;
2735 wakeup(&vm_pageout_wanted);
2737 vm_pages_needed = true;
2738 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER,
2742 struct vm_pagequeue *
2743 vm_page_pagequeue(vm_page_t m)
2746 return (&vm_phys_domain(m)->vmd_pagequeues[m->queue]);
2752 * Remove the given page from its current page queue.
2754 * The page must be locked.
2757 vm_page_dequeue(vm_page_t m)
2759 struct vm_pagequeue *pq;
2761 vm_page_assert_locked(m);
2762 KASSERT(m->queue < PQ_COUNT, ("vm_page_dequeue: page %p is not queued",
2764 pq = vm_page_pagequeue(m);
2765 vm_pagequeue_lock(pq);
2767 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2768 vm_pagequeue_cnt_dec(pq);
2769 vm_pagequeue_unlock(pq);
2773 * vm_page_dequeue_locked:
2775 * Remove the given page from its current page queue.
2777 * The page and page queue must be locked.
2780 vm_page_dequeue_locked(vm_page_t m)
2782 struct vm_pagequeue *pq;
2784 vm_page_lock_assert(m, MA_OWNED);
2785 pq = vm_page_pagequeue(m);
2786 vm_pagequeue_assert_locked(pq);
2788 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2789 vm_pagequeue_cnt_dec(pq);
2795 * Add the given page to the specified page queue.
2797 * The page must be locked.
2800 vm_page_enqueue(uint8_t queue, vm_page_t m)
2802 struct vm_pagequeue *pq;
2804 vm_page_lock_assert(m, MA_OWNED);
2805 KASSERT(queue < PQ_COUNT,
2806 ("vm_page_enqueue: invalid queue %u request for page %p",
2808 pq = &vm_phys_domain(m)->vmd_pagequeues[queue];
2809 vm_pagequeue_lock(pq);
2811 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2812 vm_pagequeue_cnt_inc(pq);
2813 vm_pagequeue_unlock(pq);
2819 * Move the given page to the tail of its current page queue.
2821 * The page must be locked.
2824 vm_page_requeue(vm_page_t m)
2826 struct vm_pagequeue *pq;
2828 vm_page_lock_assert(m, MA_OWNED);
2829 KASSERT(m->queue != PQ_NONE,
2830 ("vm_page_requeue: page %p is not queued", m));
2831 pq = vm_page_pagequeue(m);
2832 vm_pagequeue_lock(pq);
2833 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2834 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2835 vm_pagequeue_unlock(pq);
2839 * vm_page_requeue_locked:
2841 * Move the given page to the tail of its current page queue.
2843 * The page queue must be locked.
2846 vm_page_requeue_locked(vm_page_t m)
2848 struct vm_pagequeue *pq;
2850 KASSERT(m->queue != PQ_NONE,
2851 ("vm_page_requeue_locked: page %p is not queued", m));
2852 pq = vm_page_pagequeue(m);
2853 vm_pagequeue_assert_locked(pq);
2854 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2855 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2861 * Put the specified page on the active list (if appropriate).
2862 * Ensure that act_count is at least ACT_INIT but do not otherwise
2865 * The page must be locked.
2868 vm_page_activate(vm_page_t m)
2872 vm_page_lock_assert(m, MA_OWNED);
2873 if ((queue = m->queue) != PQ_ACTIVE) {
2874 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2875 if (m->act_count < ACT_INIT)
2876 m->act_count = ACT_INIT;
2877 if (queue != PQ_NONE)
2879 vm_page_enqueue(PQ_ACTIVE, m);
2881 KASSERT(queue == PQ_NONE,
2882 ("vm_page_activate: wired page %p is queued", m));
2884 if (m->act_count < ACT_INIT)
2885 m->act_count = ACT_INIT;
2890 * vm_page_free_wakeup:
2892 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
2893 * routine is called when a page has been added to the cache or free
2896 * The page queues must be locked.
2899 vm_page_free_wakeup(void)
2902 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2904 * if pageout daemon needs pages, then tell it that there are
2907 if (vm_pageout_pages_needed &&
2908 vm_cnt.v_cache_count + vm_cnt.v_free_count >= vm_cnt.v_pageout_free_min) {
2909 wakeup(&vm_pageout_pages_needed);
2910 vm_pageout_pages_needed = 0;
2913 * wakeup processes that are waiting on memory if we hit a
2914 * high water mark. And wakeup scheduler process if we have
2915 * lots of memory. this process will swapin processes.
2917 if (vm_pages_needed && !vm_page_count_min()) {
2918 vm_pages_needed = false;
2919 wakeup(&vm_cnt.v_free_count);
2924 * Turn a cached page into a free page, by changing its attributes.
2925 * Keep the statistics up-to-date.
2927 * The free page queue must be locked.
2930 vm_page_cache_turn_free(vm_page_t m)
2933 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2937 KASSERT((m->flags & PG_CACHED) != 0,
2938 ("vm_page_cache_turn_free: page %p is not cached", m));
2939 m->flags &= ~PG_CACHED;
2940 vm_cnt.v_cache_count--;
2941 vm_phys_freecnt_adj(m, 1);
2947 * Returns the given page to the free list,
2948 * disassociating it with any VM object.
2950 * The object must be locked. The page must be locked if it is managed.
2953 vm_page_free_toq(vm_page_t m)
2956 if ((m->oflags & VPO_UNMANAGED) == 0) {
2957 vm_page_lock_assert(m, MA_OWNED);
2958 KASSERT(!pmap_page_is_mapped(m),
2959 ("vm_page_free_toq: freeing mapped page %p", m));
2961 KASSERT(m->queue == PQ_NONE,
2962 ("vm_page_free_toq: unmanaged page %p is queued", m));
2963 PCPU_INC(cnt.v_tfree);
2965 if (vm_page_sbusied(m))
2966 panic("vm_page_free: freeing busy page %p", m);
2969 * Unqueue, then remove page. Note that we cannot destroy
2970 * the page here because we do not want to call the pager's
2971 * callback routine until after we've put the page on the
2972 * appropriate free queue.
2978 * If fictitious remove object association and
2979 * return, otherwise delay object association removal.
2981 if ((m->flags & PG_FICTITIOUS) != 0) {
2988 if (m->wire_count != 0)
2989 panic("vm_page_free: freeing wired page %p", m);
2990 if (m->hold_count != 0) {
2991 m->flags &= ~PG_ZERO;
2992 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2993 ("vm_page_free: freeing PG_UNHOLDFREE page %p", m));
2994 m->flags |= PG_UNHOLDFREE;
2997 * Restore the default memory attribute to the page.
2999 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3000 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3003 * Insert the page into the physical memory allocator's
3004 * cache/free page queues.
3006 mtx_lock(&vm_page_queue_free_mtx);
3007 vm_phys_freecnt_adj(m, 1);
3008 #if VM_NRESERVLEVEL > 0
3009 if (!vm_reserv_free_page(m))
3013 vm_phys_free_pages(m, 0);
3014 if ((m->flags & PG_ZERO) != 0)
3015 ++vm_page_zero_count;
3017 vm_page_zero_idle_wakeup();
3018 vm_page_free_wakeup();
3019 mtx_unlock(&vm_page_queue_free_mtx);
3026 * Mark this page as wired down by yet
3027 * another map, removing it from paging queues
3030 * If the page is fictitious, then its wire count must remain one.
3032 * The page must be locked.
3035 vm_page_wire(vm_page_t m)
3039 * Only bump the wire statistics if the page is not already wired,
3040 * and only unqueue the page if it is on some queue (if it is unmanaged
3041 * it is already off the queues).
3043 vm_page_lock_assert(m, MA_OWNED);
3044 if ((m->flags & PG_FICTITIOUS) != 0) {
3045 KASSERT(m->wire_count == 1,
3046 ("vm_page_wire: fictitious page %p's wire count isn't one",
3050 if (m->wire_count == 0) {
3051 KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
3052 m->queue == PQ_NONE,
3053 ("vm_page_wire: unmanaged page %p is queued", m));
3055 atomic_add_int(&vm_cnt.v_wire_count, 1);
3058 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
3064 * Release one wiring of the specified page, potentially allowing it to be
3065 * paged out. Returns TRUE if the number of wirings transitions to zero and
3068 * Only managed pages belonging to an object can be paged out. If the number
3069 * of wirings transitions to zero and the page is eligible for page out, then
3070 * the page is added to the specified paging queue (unless PQ_NONE is
3073 * If a page is fictitious, then its wire count must always be one.
3075 * A managed page must be locked.
3078 vm_page_unwire(vm_page_t m, uint8_t queue)
3081 KASSERT(queue < PQ_COUNT || queue == PQ_NONE,
3082 ("vm_page_unwire: invalid queue %u request for page %p",
3084 if ((m->oflags & VPO_UNMANAGED) == 0)
3085 vm_page_assert_locked(m);
3086 if ((m->flags & PG_FICTITIOUS) != 0) {
3087 KASSERT(m->wire_count == 1,
3088 ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
3091 if (m->wire_count > 0) {
3093 if (m->wire_count == 0) {
3094 atomic_subtract_int(&vm_cnt.v_wire_count, 1);
3095 if ((m->oflags & VPO_UNMANAGED) == 0 &&
3096 m->object != NULL && queue != PQ_NONE) {
3097 if (queue == PQ_INACTIVE)
3098 m->flags &= ~PG_WINATCFLS;
3099 vm_page_enqueue(queue, m);
3105 panic("vm_page_unwire: page %p's wire count is zero", m);
3109 * Move the specified page to the inactive queue.
3111 * Many pages placed on the inactive queue should actually go
3112 * into the cache, but it is difficult to figure out which. What
3113 * we do instead, if the inactive target is well met, is to put
3114 * clean pages at the head of the inactive queue instead of the tail.
3115 * This will cause them to be moved to the cache more quickly and
3116 * if not actively re-referenced, reclaimed more quickly. If we just
3117 * stick these pages at the end of the inactive queue, heavy filesystem
3118 * meta-data accesses can cause an unnecessary paging load on memory bound
3119 * processes. This optimization causes one-time-use metadata to be
3120 * reused more quickly.
3122 * Normally noreuse is FALSE, resulting in LRU operation. noreuse is set
3123 * to TRUE if we want this page to be 'as if it were placed in the cache',
3124 * except without unmapping it from the process address space. In
3125 * practice this is implemented by inserting the page at the head of the
3126 * queue, using a marker page to guide FIFO insertion ordering.
3128 * The page must be locked.
3131 _vm_page_deactivate(vm_page_t m, boolean_t noreuse)
3133 struct vm_pagequeue *pq;
3136 vm_page_assert_locked(m);
3139 * Ignore if the page is already inactive, unless it is unlikely to be
3142 if ((queue = m->queue) == PQ_INACTIVE && !noreuse)
3144 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
3145 pq = &vm_phys_domain(m)->vmd_pagequeues[PQ_INACTIVE];
3146 /* Avoid multiple acquisitions of the inactive queue lock. */
3147 if (queue == PQ_INACTIVE) {
3148 vm_pagequeue_lock(pq);
3149 vm_page_dequeue_locked(m);
3151 if (queue != PQ_NONE)
3153 m->flags &= ~PG_WINATCFLS;
3154 vm_pagequeue_lock(pq);
3156 m->queue = PQ_INACTIVE;
3158 TAILQ_INSERT_BEFORE(&vm_phys_domain(m)->vmd_inacthead,
3161 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3162 vm_pagequeue_cnt_inc(pq);
3163 vm_pagequeue_unlock(pq);
3168 * Move the specified page to the inactive queue.
3170 * The page must be locked.
3173 vm_page_deactivate(vm_page_t m)
3176 _vm_page_deactivate(m, FALSE);
3180 * Move the specified page to the inactive queue with the expectation
3181 * that it is unlikely to be reused.
3183 * The page must be locked.
3186 vm_page_deactivate_noreuse(vm_page_t m)
3189 _vm_page_deactivate(m, TRUE);
3193 * vm_page_try_to_cache:
3195 * Returns 0 on failure, 1 on success
3198 vm_page_try_to_cache(vm_page_t m)
3201 vm_page_lock_assert(m, MA_OWNED);
3202 VM_OBJECT_ASSERT_WLOCKED(m->object);
3203 if (m->dirty || m->hold_count || m->wire_count ||
3204 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m))
3214 * vm_page_try_to_free()
3216 * Attempt to free the page. If we cannot free it, we do nothing.
3217 * 1 is returned on success, 0 on failure.
3220 vm_page_try_to_free(vm_page_t m)
3223 vm_page_lock_assert(m, MA_OWNED);
3224 if (m->object != NULL)
3225 VM_OBJECT_ASSERT_WLOCKED(m->object);
3226 if (m->dirty || m->hold_count || m->wire_count ||
3227 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m))
3239 * Put the specified page onto the page cache queue (if appropriate).
3241 * The object and page must be locked.
3244 vm_page_cache(vm_page_t m)
3247 boolean_t cache_was_empty;
3249 vm_page_lock_assert(m, MA_OWNED);
3251 VM_OBJECT_ASSERT_WLOCKED(object);
3252 if (vm_page_busied(m) || (m->oflags & VPO_UNMANAGED) ||
3253 m->hold_count || m->wire_count)
3254 panic("vm_page_cache: attempting to cache busy page");
3255 KASSERT(!pmap_page_is_mapped(m),
3256 ("vm_page_cache: page %p is mapped", m));
3257 KASSERT(m->dirty == 0, ("vm_page_cache: page %p is dirty", m));
3258 if (m->valid == 0 || object->type == OBJT_DEFAULT ||
3259 (object->type == OBJT_SWAP &&
3260 !vm_pager_has_page(object, m->pindex, NULL, NULL))) {
3262 * Hypothesis: A cache-eligible page belonging to a
3263 * default object or swap object but without a backing
3264 * store must be zero filled.
3269 KASSERT((m->flags & PG_CACHED) == 0,
3270 ("vm_page_cache: page %p is already cached", m));
3273 * Remove the page from the paging queues.
3278 * Remove the page from the object's collection of resident
3281 vm_radix_remove(&object->rtree, m->pindex);
3282 TAILQ_REMOVE(&object->memq, m, listq);
3283 object->resident_page_count--;
3286 * Restore the default memory attribute to the page.
3288 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3289 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3292 * Insert the page into the object's collection of cached pages
3293 * and the physical memory allocator's cache/free page queues.
3295 m->flags &= ~PG_ZERO;
3296 mtx_lock(&vm_page_queue_free_mtx);
3297 cache_was_empty = vm_radix_is_empty(&object->cache);
3298 if (vm_radix_insert(&object->cache, m)) {
3299 mtx_unlock(&vm_page_queue_free_mtx);
3300 if (object->type == OBJT_VNODE &&
3301 object->resident_page_count == 0)
3302 vdrop(object->handle);
3309 * The above call to vm_radix_insert() could reclaim the one pre-
3310 * existing cached page from this object, resulting in a call to
3313 if (!cache_was_empty)
3314 cache_was_empty = vm_radix_is_singleton(&object->cache);
3316 m->flags |= PG_CACHED;
3317 vm_cnt.v_cache_count++;
3318 PCPU_INC(cnt.v_tcached);
3319 #if VM_NRESERVLEVEL > 0
3320 if (!vm_reserv_free_page(m)) {
3324 vm_phys_free_pages(m, 0);
3326 vm_page_free_wakeup();
3327 mtx_unlock(&vm_page_queue_free_mtx);
3330 * Increment the vnode's hold count if this is the object's only
3331 * cached page. Decrement the vnode's hold count if this was
3332 * the object's only resident page.
3334 if (object->type == OBJT_VNODE) {
3335 if (cache_was_empty && object->resident_page_count != 0)
3336 vhold(object->handle);
3337 else if (!cache_was_empty && object->resident_page_count == 0)
3338 vdrop(object->handle);
3345 * Deactivate or do nothing, as appropriate.
3347 * The object and page must be locked.
3350 vm_page_advise(vm_page_t m, int advice)
3353 vm_page_assert_locked(m);
3354 VM_OBJECT_ASSERT_WLOCKED(m->object);
3355 if (advice == MADV_FREE)
3357 * Mark the page clean. This will allow the page to be freed
3358 * up by the system. However, such pages are often reused
3359 * quickly by malloc() so we do not do anything that would
3360 * cause a page fault if we can help it.
3362 * Specifically, we do not try to actually free the page now
3363 * nor do we try to put it in the cache (which would cause a
3364 * page fault on reuse).
3366 * But we do make the page as freeable as we can without
3367 * actually taking the step of unmapping it.
3370 else if (advice != MADV_DONTNEED)
3374 * Clear any references to the page. Otherwise, the page daemon will
3375 * immediately reactivate the page.
3377 vm_page_aflag_clear(m, PGA_REFERENCED);
3379 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
3383 * Place clean pages at the head of the inactive queue rather than the
3384 * tail, thus defeating the queue's LRU operation and ensuring that the
3385 * page will be reused quickly.
3387 _vm_page_deactivate(m, m->dirty == 0);
3391 * Grab a page, waiting until we are waken up due to the page
3392 * changing state. We keep on waiting, if the page continues
3393 * to be in the object. If the page doesn't exist, first allocate it
3394 * and then conditionally zero it.
3396 * This routine may sleep.
3398 * The object must be locked on entry. The lock will, however, be released
3399 * and reacquired if the routine sleeps.
3402 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
3407 VM_OBJECT_ASSERT_WLOCKED(object);
3408 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3409 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3410 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
3412 if ((m = vm_page_lookup(object, pindex)) != NULL) {
3413 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3414 vm_page_xbusied(m) : vm_page_busied(m);
3416 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3419 * Reference the page before unlocking and
3420 * sleeping so that the page daemon is less
3421 * likely to reclaim it.
3423 vm_page_aflag_set(m, PGA_REFERENCED);
3425 VM_OBJECT_WUNLOCK(object);
3426 vm_page_busy_sleep(m, "pgrbwt");
3427 VM_OBJECT_WLOCK(object);
3430 if ((allocflags & VM_ALLOC_WIRED) != 0) {
3436 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
3438 if ((allocflags & VM_ALLOC_SBUSY) != 0)
3443 m = vm_page_alloc(object, pindex, allocflags);
3445 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3447 VM_OBJECT_WUNLOCK(object);
3449 VM_OBJECT_WLOCK(object);
3451 } else if (m->valid != 0)
3453 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
3459 * Mapping function for valid or dirty bits in a page.
3461 * Inputs are required to range within a page.
3464 vm_page_bits(int base, int size)
3470 base + size <= PAGE_SIZE,
3471 ("vm_page_bits: illegal base/size %d/%d", base, size)
3474 if (size == 0) /* handle degenerate case */
3477 first_bit = base >> DEV_BSHIFT;
3478 last_bit = (base + size - 1) >> DEV_BSHIFT;
3480 return (((vm_page_bits_t)2 << last_bit) -
3481 ((vm_page_bits_t)1 << first_bit));
3485 * vm_page_set_valid_range:
3487 * Sets portions of a page valid. The arguments are expected
3488 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3489 * of any partial chunks touched by the range. The invalid portion of
3490 * such chunks will be zeroed.
3492 * (base + size) must be less then or equal to PAGE_SIZE.
3495 vm_page_set_valid_range(vm_page_t m, int base, int size)
3499 VM_OBJECT_ASSERT_WLOCKED(m->object);
3500 if (size == 0) /* handle degenerate case */
3504 * If the base is not DEV_BSIZE aligned and the valid
3505 * bit is clear, we have to zero out a portion of the
3508 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
3509 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
3510 pmap_zero_page_area(m, frag, base - frag);
3513 * If the ending offset is not DEV_BSIZE aligned and the
3514 * valid bit is clear, we have to zero out a portion of
3517 endoff = base + size;
3518 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
3519 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
3520 pmap_zero_page_area(m, endoff,
3521 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
3524 * Assert that no previously invalid block that is now being validated
3527 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
3528 ("vm_page_set_valid_range: page %p is dirty", m));
3531 * Set valid bits inclusive of any overlap.
3533 m->valid |= vm_page_bits(base, size);
3537 * Clear the given bits from the specified page's dirty field.
3539 static __inline void
3540 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
3543 #if PAGE_SIZE < 16384
3548 * If the object is locked and the page is neither exclusive busy nor
3549 * write mapped, then the page's dirty field cannot possibly be
3550 * set by a concurrent pmap operation.
3552 VM_OBJECT_ASSERT_WLOCKED(m->object);
3553 if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
3554 m->dirty &= ~pagebits;
3557 * The pmap layer can call vm_page_dirty() without
3558 * holding a distinguished lock. The combination of
3559 * the object's lock and an atomic operation suffice
3560 * to guarantee consistency of the page dirty field.
3562 * For PAGE_SIZE == 32768 case, compiler already
3563 * properly aligns the dirty field, so no forcible
3564 * alignment is needed. Only require existence of
3565 * atomic_clear_64 when page size is 32768.
3567 addr = (uintptr_t)&m->dirty;
3568 #if PAGE_SIZE == 32768
3569 atomic_clear_64((uint64_t *)addr, pagebits);
3570 #elif PAGE_SIZE == 16384
3571 atomic_clear_32((uint32_t *)addr, pagebits);
3572 #else /* PAGE_SIZE <= 8192 */
3574 * Use a trick to perform a 32-bit atomic on the
3575 * containing aligned word, to not depend on the existence
3576 * of atomic_clear_{8, 16}.
3578 shift = addr & (sizeof(uint32_t) - 1);
3579 #if BYTE_ORDER == BIG_ENDIAN
3580 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
3584 addr &= ~(sizeof(uint32_t) - 1);
3585 atomic_clear_32((uint32_t *)addr, pagebits << shift);
3586 #endif /* PAGE_SIZE */
3591 * vm_page_set_validclean:
3593 * Sets portions of a page valid and clean. The arguments are expected
3594 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3595 * of any partial chunks touched by the range. The invalid portion of
3596 * such chunks will be zero'd.
3598 * (base + size) must be less then or equal to PAGE_SIZE.
3601 vm_page_set_validclean(vm_page_t m, int base, int size)
3603 vm_page_bits_t oldvalid, pagebits;
3606 VM_OBJECT_ASSERT_WLOCKED(m->object);
3607 if (size == 0) /* handle degenerate case */
3611 * If the base is not DEV_BSIZE aligned and the valid
3612 * bit is clear, we have to zero out a portion of the
3615 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
3616 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
3617 pmap_zero_page_area(m, frag, base - frag);
3620 * If the ending offset is not DEV_BSIZE aligned and the
3621 * valid bit is clear, we have to zero out a portion of
3624 endoff = base + size;
3625 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
3626 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
3627 pmap_zero_page_area(m, endoff,
3628 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
3631 * Set valid, clear dirty bits. If validating the entire
3632 * page we can safely clear the pmap modify bit. We also
3633 * use this opportunity to clear the VPO_NOSYNC flag. If a process
3634 * takes a write fault on a MAP_NOSYNC memory area the flag will
3637 * We set valid bits inclusive of any overlap, but we can only
3638 * clear dirty bits for DEV_BSIZE chunks that are fully within
3641 oldvalid = m->valid;
3642 pagebits = vm_page_bits(base, size);
3643 m->valid |= pagebits;
3645 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
3646 frag = DEV_BSIZE - frag;
3652 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
3654 if (base == 0 && size == PAGE_SIZE) {
3656 * The page can only be modified within the pmap if it is
3657 * mapped, and it can only be mapped if it was previously
3660 if (oldvalid == VM_PAGE_BITS_ALL)
3662 * Perform the pmap_clear_modify() first. Otherwise,
3663 * a concurrent pmap operation, such as
3664 * pmap_protect(), could clear a modification in the
3665 * pmap and set the dirty field on the page before
3666 * pmap_clear_modify() had begun and after the dirty
3667 * field was cleared here.
3669 pmap_clear_modify(m);
3671 m->oflags &= ~VPO_NOSYNC;
3672 } else if (oldvalid != VM_PAGE_BITS_ALL)
3673 m->dirty &= ~pagebits;
3675 vm_page_clear_dirty_mask(m, pagebits);
3679 vm_page_clear_dirty(vm_page_t m, int base, int size)
3682 vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
3686 * vm_page_set_invalid:
3688 * Invalidates DEV_BSIZE'd chunks within a page. Both the
3689 * valid and dirty bits for the effected areas are cleared.
3692 vm_page_set_invalid(vm_page_t m, int base, int size)
3694 vm_page_bits_t bits;
3698 VM_OBJECT_ASSERT_WLOCKED(object);
3699 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
3700 size >= object->un_pager.vnp.vnp_size)
3701 bits = VM_PAGE_BITS_ALL;
3703 bits = vm_page_bits(base, size);
3704 if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL &&
3707 KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) ||
3708 !pmap_page_is_mapped(m),
3709 ("vm_page_set_invalid: page %p is mapped", m));
3715 * vm_page_zero_invalid()
3717 * The kernel assumes that the invalid portions of a page contain
3718 * garbage, but such pages can be mapped into memory by user code.
3719 * When this occurs, we must zero out the non-valid portions of the
3720 * page so user code sees what it expects.
3722 * Pages are most often semi-valid when the end of a file is mapped
3723 * into memory and the file's size is not page aligned.
3726 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
3731 VM_OBJECT_ASSERT_WLOCKED(m->object);
3733 * Scan the valid bits looking for invalid sections that
3734 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the
3735 * valid bit may be set ) have already been zeroed by
3736 * vm_page_set_validclean().
3738 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
3739 if (i == (PAGE_SIZE / DEV_BSIZE) ||
3740 (m->valid & ((vm_page_bits_t)1 << i))) {
3742 pmap_zero_page_area(m,
3743 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
3750 * setvalid is TRUE when we can safely set the zero'd areas
3751 * as being valid. We can do this if there are no cache consistancy
3752 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
3755 m->valid = VM_PAGE_BITS_ALL;
3761 * Is (partial) page valid? Note that the case where size == 0
3762 * will return FALSE in the degenerate case where the page is
3763 * entirely invalid, and TRUE otherwise.
3766 vm_page_is_valid(vm_page_t m, int base, int size)
3768 vm_page_bits_t bits;
3770 VM_OBJECT_ASSERT_LOCKED(m->object);
3771 bits = vm_page_bits(base, size);
3772 return (m->valid != 0 && (m->valid & bits) == bits);
3776 * vm_page_ps_is_valid:
3778 * Returns TRUE if the entire (super)page is valid and FALSE otherwise.
3781 vm_page_ps_is_valid(vm_page_t m)
3785 VM_OBJECT_ASSERT_LOCKED(m->object);
3786 npages = atop(pagesizes[m->psind]);
3789 * The physically contiguous pages that make up a superpage, i.e., a
3790 * page with a page size index ("psind") greater than zero, will
3791 * occupy adjacent entries in vm_page_array[].
3793 for (i = 0; i < npages; i++) {
3794 if (m[i].valid != VM_PAGE_BITS_ALL)
3801 * Set the page's dirty bits if the page is modified.
3804 vm_page_test_dirty(vm_page_t m)
3807 VM_OBJECT_ASSERT_WLOCKED(m->object);
3808 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
3813 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
3816 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
3820 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
3823 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
3827 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
3830 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
3833 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
3835 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
3838 vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
3842 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
3845 mtx_assert_(vm_page_lockptr(m), a, file, line);
3851 vm_page_object_lock_assert(vm_page_t m)
3855 * Certain of the page's fields may only be modified by the
3856 * holder of the containing object's lock or the exclusive busy.
3857 * holder. Unfortunately, the holder of the write busy is
3858 * not recorded, and thus cannot be checked here.
3860 if (m->object != NULL && !vm_page_xbusied(m))
3861 VM_OBJECT_ASSERT_WLOCKED(m->object);
3865 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits)
3868 if ((bits & PGA_WRITEABLE) == 0)
3872 * The PGA_WRITEABLE flag can only be set if the page is
3873 * managed, is exclusively busied or the object is locked.
3874 * Currently, this flag is only set by pmap_enter().
3876 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3877 ("PGA_WRITEABLE on unmanaged page"));
3878 if (!vm_page_xbusied(m))
3879 VM_OBJECT_ASSERT_LOCKED(m->object);
3883 #include "opt_ddb.h"
3885 #include <sys/kernel.h>
3887 #include <ddb/ddb.h>
3889 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3891 db_printf("vm_cnt.v_free_count: %d\n", vm_cnt.v_free_count);
3892 db_printf("vm_cnt.v_cache_count: %d\n", vm_cnt.v_cache_count);
3893 db_printf("vm_cnt.v_inactive_count: %d\n", vm_cnt.v_inactive_count);
3894 db_printf("vm_cnt.v_active_count: %d\n", vm_cnt.v_active_count);
3895 db_printf("vm_cnt.v_wire_count: %d\n", vm_cnt.v_wire_count);
3896 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
3897 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
3898 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
3899 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
3902 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3906 db_printf("pq_free %d pq_cache %d\n",
3907 vm_cnt.v_free_count, vm_cnt.v_cache_count);
3908 for (dom = 0; dom < vm_ndomains; dom++) {
3910 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pass %d\n",
3912 vm_dom[dom].vmd_page_count,
3913 vm_dom[dom].vmd_free_count,
3914 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
3915 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
3916 vm_dom[dom].vmd_pass);
3920 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
3926 db_printf("show pginfo addr\n");
3930 phys = strchr(modif, 'p') != NULL;
3932 m = PHYS_TO_VM_PAGE(addr);
3934 m = (vm_page_t)addr;
3936 "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n"
3937 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
3938 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
3939 m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags,
3940 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty);