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,
1117 VM_OBJECT_ASSERT_WLOCKED(object);
1118 KASSERT(m->object == NULL,
1119 ("vm_page_insert_after: page already inserted"));
1120 if (mpred != NULL) {
1121 KASSERT(mpred->object == object,
1122 ("vm_page_insert_after: object doesn't contain mpred"));
1123 KASSERT(mpred->pindex < pindex,
1124 ("vm_page_insert_after: mpred doesn't precede pindex"));
1125 msucc = TAILQ_NEXT(mpred, listq);
1127 msucc = TAILQ_FIRST(&object->memq);
1129 KASSERT(msucc->pindex > pindex,
1130 ("vm_page_insert_after: msucc doesn't succeed pindex"));
1133 * Record the object/offset pair in this page
1139 * Now link into the object's ordered list of backed pages.
1141 if (vm_radix_insert(&object->rtree, m)) {
1146 vm_page_insert_radixdone(m, object, mpred);
1151 * vm_page_insert_radixdone:
1153 * Complete page "m" insertion into the specified object after the
1154 * radix trie hooking.
1156 * The page "mpred" must precede the offset "m->pindex" within the
1159 * The object must be locked.
1162 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1165 VM_OBJECT_ASSERT_WLOCKED(object);
1166 KASSERT(object != NULL && m->object == object,
1167 ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1168 if (mpred != NULL) {
1169 KASSERT(mpred->object == object,
1170 ("vm_page_insert_after: object doesn't contain mpred"));
1171 KASSERT(mpred->pindex < m->pindex,
1172 ("vm_page_insert_after: mpred doesn't precede pindex"));
1176 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1178 TAILQ_INSERT_HEAD(&object->memq, m, listq);
1181 * Show that the object has one more resident page.
1183 object->resident_page_count++;
1186 * Hold the vnode until the last page is released.
1188 if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1189 vhold(object->handle);
1192 * Since we are inserting a new and possibly dirty page,
1193 * update the object's OBJ_MIGHTBEDIRTY flag.
1195 if (pmap_page_is_write_mapped(m))
1196 vm_object_set_writeable_dirty(object);
1202 * Removes the given mem entry from the object/offset-page
1203 * table and the object page list, but do not invalidate/terminate
1204 * the backing store.
1206 * The object must be locked. The page must be locked if it is managed.
1209 vm_page_remove(vm_page_t m)
1213 if ((m->oflags & VPO_UNMANAGED) == 0)
1214 vm_page_assert_locked(m);
1215 if ((object = m->object) == NULL)
1217 VM_OBJECT_ASSERT_WLOCKED(object);
1218 if (vm_page_xbusied(m))
1219 vm_page_xunbusy_maybelocked(m);
1222 * Now remove from the object's list of backed pages.
1224 vm_radix_remove(&object->rtree, m->pindex);
1225 TAILQ_REMOVE(&object->memq, m, listq);
1228 * And show that the object has one fewer resident page.
1230 object->resident_page_count--;
1233 * The vnode may now be recycled.
1235 if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1236 vdrop(object->handle);
1244 * Returns the page associated with the object/offset
1245 * pair specified; if none is found, NULL is returned.
1247 * The object must be locked.
1250 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1253 VM_OBJECT_ASSERT_LOCKED(object);
1254 return (vm_radix_lookup(&object->rtree, pindex));
1258 * vm_page_find_least:
1260 * Returns the page associated with the object with least pindex
1261 * greater than or equal to the parameter pindex, or NULL.
1263 * The object must be locked.
1266 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1270 VM_OBJECT_ASSERT_LOCKED(object);
1271 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1272 m = vm_radix_lookup_ge(&object->rtree, pindex);
1277 * Returns the given page's successor (by pindex) within the object if it is
1278 * resident; if none is found, NULL is returned.
1280 * The object must be locked.
1283 vm_page_next(vm_page_t m)
1287 VM_OBJECT_ASSERT_LOCKED(m->object);
1288 if ((next = TAILQ_NEXT(m, listq)) != NULL &&
1289 next->pindex != m->pindex + 1)
1295 * Returns the given page's predecessor (by pindex) within the object if it is
1296 * resident; if none is found, NULL is returned.
1298 * The object must be locked.
1301 vm_page_prev(vm_page_t m)
1305 VM_OBJECT_ASSERT_LOCKED(m->object);
1306 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
1307 prev->pindex != m->pindex - 1)
1313 * Uses the page mnew as a replacement for an existing page at index
1314 * pindex which must be already present in the object.
1316 * The existing page must not be on a paging queue.
1319 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1323 VM_OBJECT_ASSERT_WLOCKED(object);
1324 KASSERT(mnew->object == NULL,
1325 ("vm_page_replace: page already in object"));
1328 * This function mostly follows vm_page_insert() and
1329 * vm_page_remove() without the radix, object count and vnode
1330 * dance. Double check such functions for more comments.
1333 mnew->object = object;
1334 mnew->pindex = pindex;
1335 mold = vm_radix_replace(&object->rtree, mnew);
1336 KASSERT(mold->queue == PQ_NONE,
1337 ("vm_page_replace: mold is on a paging queue"));
1339 /* Keep the resident page list in sorted order. */
1340 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1341 TAILQ_REMOVE(&object->memq, mold, listq);
1343 mold->object = NULL;
1344 vm_page_xunbusy_maybelocked(mold);
1347 * The object's resident_page_count does not change because we have
1348 * swapped one page for another, but OBJ_MIGHTBEDIRTY.
1350 if (pmap_page_is_write_mapped(mnew))
1351 vm_object_set_writeable_dirty(object);
1358 * Move the given memory entry from its
1359 * current object to the specified target object/offset.
1361 * Note: swap associated with the page must be invalidated by the move. We
1362 * have to do this for several reasons: (1) we aren't freeing the
1363 * page, (2) we are dirtying the page, (3) the VM system is probably
1364 * moving the page from object A to B, and will then later move
1365 * the backing store from A to B and we can't have a conflict.
1367 * Note: we *always* dirty the page. It is necessary both for the
1368 * fact that we moved it, and because we may be invalidating
1369 * swap. If the page is on the cache, we have to deactivate it
1370 * or vm_page_dirty() will panic. Dirty pages are not allowed
1373 * The objects must be locked.
1376 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1381 VM_OBJECT_ASSERT_WLOCKED(new_object);
1383 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1384 KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1385 ("vm_page_rename: pindex already renamed"));
1388 * Create a custom version of vm_page_insert() which does not depend
1389 * by m_prev and can cheat on the implementation aspects of the
1393 m->pindex = new_pindex;
1394 if (vm_radix_insert(&new_object->rtree, m)) {
1400 * The operation cannot fail anymore. The removal must happen before
1401 * the listq iterator is tainted.
1407 /* Return back to the new pindex to complete vm_page_insert(). */
1408 m->pindex = new_pindex;
1409 m->object = new_object;
1411 vm_page_insert_radixdone(m, new_object, mpred);
1417 * Convert all of the given object's cached pages that have a
1418 * pindex within the given range into free pages. If the value
1419 * zero is given for "end", then the range's upper bound is
1420 * infinity. If the given object is backed by a vnode and it
1421 * transitions from having one or more cached pages to none, the
1422 * vnode's hold count is reduced.
1425 vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1430 mtx_lock(&vm_page_queue_free_mtx);
1431 if (__predict_false(vm_radix_is_empty(&object->cache))) {
1432 mtx_unlock(&vm_page_queue_free_mtx);
1435 while ((m = vm_radix_lookup_ge(&object->cache, start)) != NULL) {
1436 if (end != 0 && m->pindex >= end)
1438 vm_radix_remove(&object->cache, m->pindex);
1439 vm_page_cache_turn_free(m);
1441 empty = vm_radix_is_empty(&object->cache);
1442 mtx_unlock(&vm_page_queue_free_mtx);
1443 if (object->type == OBJT_VNODE && empty)
1444 vdrop(object->handle);
1448 * Returns the cached page that is associated with the given
1449 * object and offset. If, however, none exists, returns NULL.
1451 * The free page queue must be locked.
1453 static inline vm_page_t
1454 vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex)
1457 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1458 return (vm_radix_lookup(&object->cache, pindex));
1462 * Remove the given cached page from its containing object's
1463 * collection of cached pages.
1465 * The free page queue must be locked.
1468 vm_page_cache_remove(vm_page_t m)
1471 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1472 KASSERT((m->flags & PG_CACHED) != 0,
1473 ("vm_page_cache_remove: page %p is not cached", m));
1474 vm_radix_remove(&m->object->cache, m->pindex);
1476 vm_cnt.v_cache_count--;
1480 * Transfer all of the cached pages with offset greater than or
1481 * equal to 'offidxstart' from the original object's cache to the
1482 * new object's cache. However, any cached pages with offset
1483 * greater than or equal to the new object's size are kept in the
1484 * original object. Initially, the new object's cache must be
1485 * empty. Offset 'offidxstart' in the original object must
1486 * correspond to offset zero in the new object.
1488 * The new object must be locked.
1491 vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart,
1492 vm_object_t new_object)
1497 * Insertion into an object's collection of cached pages
1498 * requires the object to be locked. In contrast, removal does
1501 VM_OBJECT_ASSERT_WLOCKED(new_object);
1502 KASSERT(vm_radix_is_empty(&new_object->cache),
1503 ("vm_page_cache_transfer: object %p has cached pages",
1505 mtx_lock(&vm_page_queue_free_mtx);
1506 while ((m = vm_radix_lookup_ge(&orig_object->cache,
1507 offidxstart)) != NULL) {
1509 * Transfer all of the pages with offset greater than or
1510 * equal to 'offidxstart' from the original object's
1511 * cache to the new object's cache.
1513 if ((m->pindex - offidxstart) >= new_object->size)
1515 vm_radix_remove(&orig_object->cache, m->pindex);
1516 /* Update the page's object and offset. */
1517 m->object = new_object;
1518 m->pindex -= offidxstart;
1519 if (vm_radix_insert(&new_object->cache, m))
1520 vm_page_cache_turn_free(m);
1522 mtx_unlock(&vm_page_queue_free_mtx);
1526 * Returns TRUE if a cached page is associated with the given object and
1527 * offset, and FALSE otherwise.
1529 * The object must be locked.
1532 vm_page_is_cached(vm_object_t object, vm_pindex_t pindex)
1537 * Insertion into an object's collection of cached pages requires the
1538 * object to be locked. Therefore, if the object is locked and the
1539 * object's collection is empty, there is no need to acquire the free
1540 * page queues lock in order to prove that the specified page doesn't
1543 VM_OBJECT_ASSERT_WLOCKED(object);
1544 if (__predict_true(vm_object_cache_is_empty(object)))
1546 mtx_lock(&vm_page_queue_free_mtx);
1547 m = vm_page_cache_lookup(object, pindex);
1548 mtx_unlock(&vm_page_queue_free_mtx);
1555 * Allocate and return a page that is associated with the specified
1556 * object and offset pair. By default, this page is exclusive busied.
1558 * The caller must always specify an allocation class.
1560 * allocation classes:
1561 * VM_ALLOC_NORMAL normal process request
1562 * VM_ALLOC_SYSTEM system *really* needs a page
1563 * VM_ALLOC_INTERRUPT interrupt time request
1565 * optional allocation flags:
1566 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1567 * intends to allocate
1568 * VM_ALLOC_IFCACHED return page only if it is cached
1569 * VM_ALLOC_IFNOTCACHED return NULL, do not reactivate if the page
1571 * VM_ALLOC_NOBUSY do not exclusive busy the page
1572 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1573 * VM_ALLOC_NOOBJ page is not associated with an object and
1574 * should not be exclusive busy
1575 * VM_ALLOC_SBUSY shared busy the allocated page
1576 * VM_ALLOC_WIRED wire the allocated page
1577 * VM_ALLOC_ZERO prefer a zeroed page
1579 * This routine may not sleep.
1582 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1584 struct vnode *vp = NULL;
1585 vm_object_t m_object;
1587 int flags, req_class;
1589 mpred = 0; /* XXX: pacify gcc */
1590 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1591 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1592 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1593 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1594 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object,
1597 VM_OBJECT_ASSERT_WLOCKED(object);
1599 req_class = req & VM_ALLOC_CLASS_MASK;
1602 * The page daemon is allowed to dig deeper into the free page list.
1604 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1605 req_class = VM_ALLOC_SYSTEM;
1607 if (object != NULL) {
1608 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1609 KASSERT(mpred == NULL || mpred->pindex != pindex,
1610 ("vm_page_alloc: pindex already allocated"));
1614 * The page allocation request can came from consumers which already
1615 * hold the free page queue mutex, like vm_page_insert() in
1618 mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE);
1619 if (vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_free_reserved ||
1620 (req_class == VM_ALLOC_SYSTEM &&
1621 vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_interrupt_free_min) ||
1622 (req_class == VM_ALLOC_INTERRUPT &&
1623 vm_cnt.v_free_count + vm_cnt.v_cache_count > 0)) {
1625 * Allocate from the free queue if the number of free pages
1626 * exceeds the minimum for the request class.
1628 if (object != NULL &&
1629 (m = vm_page_cache_lookup(object, pindex)) != NULL) {
1630 if ((req & VM_ALLOC_IFNOTCACHED) != 0) {
1631 mtx_unlock(&vm_page_queue_free_mtx);
1634 if (vm_phys_unfree_page(m))
1635 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0);
1636 #if VM_NRESERVLEVEL > 0
1637 else if (!vm_reserv_reactivate_page(m))
1641 panic("vm_page_alloc: cache page %p is missing"
1642 " from the free queue", m);
1643 } else if ((req & VM_ALLOC_IFCACHED) != 0) {
1644 mtx_unlock(&vm_page_queue_free_mtx);
1646 #if VM_NRESERVLEVEL > 0
1647 } else if (object == NULL || (object->flags & (OBJ_COLORED |
1648 OBJ_FICTITIOUS)) != OBJ_COLORED || (m =
1649 vm_reserv_alloc_page(object, pindex, mpred)) == NULL) {
1653 m = vm_phys_alloc_pages(object != NULL ?
1654 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1655 #if VM_NRESERVLEVEL > 0
1656 if (m == NULL && vm_reserv_reclaim_inactive()) {
1657 m = vm_phys_alloc_pages(object != NULL ?
1658 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1665 * Not allocatable, give up.
1667 mtx_unlock(&vm_page_queue_free_mtx);
1668 atomic_add_int(&vm_pageout_deficit,
1669 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1670 pagedaemon_wakeup();
1675 * At this point we had better have found a good page.
1677 KASSERT(m != NULL, ("vm_page_alloc: missing page"));
1678 KASSERT(m->queue == PQ_NONE,
1679 ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue));
1680 KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m));
1681 KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m));
1682 KASSERT(!vm_page_sbusied(m),
1683 ("vm_page_alloc: page %p is busy", m));
1684 KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m));
1685 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1686 ("vm_page_alloc: page %p has unexpected memattr %d", m,
1687 pmap_page_get_memattr(m)));
1688 if ((m->flags & PG_CACHED) != 0) {
1689 KASSERT((m->flags & PG_ZERO) == 0,
1690 ("vm_page_alloc: cached page %p is PG_ZERO", m));
1691 KASSERT(m->valid != 0,
1692 ("vm_page_alloc: cached page %p is invalid", m));
1693 if (m->object == object && m->pindex == pindex)
1694 vm_cnt.v_reactivated++;
1697 m_object = m->object;
1698 vm_page_cache_remove(m);
1699 if (m_object->type == OBJT_VNODE &&
1700 vm_object_cache_is_empty(m_object))
1701 vp = m_object->handle;
1703 KASSERT(m->valid == 0,
1704 ("vm_page_alloc: free page %p is valid", m));
1705 vm_phys_freecnt_adj(m, -1);
1706 if ((m->flags & PG_ZERO) != 0)
1707 vm_page_zero_count--;
1709 mtx_unlock(&vm_page_queue_free_mtx);
1712 * Initialize the page. Only the PG_ZERO flag is inherited.
1715 if ((req & VM_ALLOC_ZERO) != 0)
1718 if ((req & VM_ALLOC_NODUMP) != 0)
1722 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1724 m->busy_lock = VPB_UNBUSIED;
1725 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1726 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1727 if ((req & VM_ALLOC_SBUSY) != 0)
1728 m->busy_lock = VPB_SHARERS_WORD(1);
1729 if (req & VM_ALLOC_WIRED) {
1731 * The page lock is not required for wiring a page until that
1732 * page is inserted into the object.
1734 atomic_add_int(&vm_cnt.v_wire_count, 1);
1739 if (object != NULL) {
1740 if (vm_page_insert_after(m, object, pindex, mpred)) {
1741 /* See the comment below about hold count. */
1744 pagedaemon_wakeup();
1745 if (req & VM_ALLOC_WIRED) {
1746 atomic_subtract_int(&vm_cnt.v_wire_count, 1);
1750 m->oflags = VPO_UNMANAGED;
1751 m->busy_lock = VPB_UNBUSIED;
1756 /* Ignore device objects; the pager sets "memattr" for them. */
1757 if (object->memattr != VM_MEMATTR_DEFAULT &&
1758 (object->flags & OBJ_FICTITIOUS) == 0)
1759 pmap_page_set_memattr(m, object->memattr);
1764 * The following call to vdrop() must come after the above call
1765 * to vm_page_insert() in case both affect the same object and
1766 * vnode. Otherwise, the affected vnode's hold count could
1767 * temporarily become zero.
1773 * Don't wakeup too often - wakeup the pageout daemon when
1774 * we would be nearly out of memory.
1776 if (vm_paging_needed())
1777 pagedaemon_wakeup();
1783 vm_page_alloc_contig_vdrop(struct spglist *lst)
1786 while (!SLIST_EMPTY(lst)) {
1787 vdrop((struct vnode *)SLIST_FIRST(lst)-> plinks.s.pv);
1788 SLIST_REMOVE_HEAD(lst, plinks.s.ss);
1793 * vm_page_alloc_contig:
1795 * Allocate a contiguous set of physical pages of the given size "npages"
1796 * from the free lists. All of the physical pages must be at or above
1797 * the given physical address "low" and below the given physical address
1798 * "high". The given value "alignment" determines the alignment of the
1799 * first physical page in the set. If the given value "boundary" is
1800 * non-zero, then the set of physical pages cannot cross any physical
1801 * address boundary that is a multiple of that value. Both "alignment"
1802 * and "boundary" must be a power of two.
1804 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1805 * then the memory attribute setting for the physical pages is configured
1806 * to the object's memory attribute setting. Otherwise, the memory
1807 * attribute setting for the physical pages is configured to "memattr",
1808 * overriding the object's memory attribute setting. However, if the
1809 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1810 * memory attribute setting for the physical pages cannot be configured
1811 * to VM_MEMATTR_DEFAULT.
1813 * The caller must always specify an allocation class.
1815 * allocation classes:
1816 * VM_ALLOC_NORMAL normal process request
1817 * VM_ALLOC_SYSTEM system *really* needs a page
1818 * VM_ALLOC_INTERRUPT interrupt time request
1820 * optional allocation flags:
1821 * VM_ALLOC_NOBUSY do not exclusive busy the page
1822 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1823 * VM_ALLOC_NOOBJ page is not associated with an object and
1824 * should not be exclusive busy
1825 * VM_ALLOC_SBUSY shared busy the allocated page
1826 * VM_ALLOC_WIRED wire the allocated page
1827 * VM_ALLOC_ZERO prefer a zeroed page
1829 * This routine may not sleep.
1832 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1833 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1834 vm_paddr_t boundary, vm_memattr_t memattr)
1837 struct spglist deferred_vdrop_list;
1838 vm_page_t m, m_tmp, m_ret;
1842 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1843 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1844 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1845 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1846 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object,
1848 if (object != NULL) {
1849 VM_OBJECT_ASSERT_WLOCKED(object);
1850 KASSERT(object->type == OBJT_PHYS,
1851 ("vm_page_alloc_contig: object %p isn't OBJT_PHYS",
1854 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
1855 req_class = req & VM_ALLOC_CLASS_MASK;
1858 * The page daemon is allowed to dig deeper into the free page list.
1860 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1861 req_class = VM_ALLOC_SYSTEM;
1863 SLIST_INIT(&deferred_vdrop_list);
1864 mtx_lock(&vm_page_queue_free_mtx);
1865 if (vm_cnt.v_free_count + vm_cnt.v_cache_count >= npages +
1866 vm_cnt.v_free_reserved || (req_class == VM_ALLOC_SYSTEM &&
1867 vm_cnt.v_free_count + vm_cnt.v_cache_count >= npages +
1868 vm_cnt.v_interrupt_free_min) || (req_class == VM_ALLOC_INTERRUPT &&
1869 vm_cnt.v_free_count + vm_cnt.v_cache_count >= npages)) {
1870 #if VM_NRESERVLEVEL > 0
1872 if (object == NULL || (object->flags & OBJ_COLORED) == 0 ||
1873 (m_ret = vm_reserv_alloc_contig(object, pindex, npages,
1874 low, high, alignment, boundary)) == NULL)
1876 m_ret = vm_phys_alloc_contig(npages, low, high,
1877 alignment, boundary);
1879 mtx_unlock(&vm_page_queue_free_mtx);
1880 atomic_add_int(&vm_pageout_deficit, npages);
1881 pagedaemon_wakeup();
1885 for (m = m_ret; m < &m_ret[npages]; m++) {
1886 drop = vm_page_alloc_init(m);
1889 * Enqueue the vnode for deferred vdrop().
1891 m->plinks.s.pv = drop;
1892 SLIST_INSERT_HEAD(&deferred_vdrop_list, m,
1897 #if VM_NRESERVLEVEL > 0
1898 if (vm_reserv_reclaim_contig(npages, low, high, alignment,
1903 mtx_unlock(&vm_page_queue_free_mtx);
1908 * Initialize the pages. Only the PG_ZERO flag is inherited.
1911 if ((req & VM_ALLOC_ZERO) != 0)
1913 if ((req & VM_ALLOC_NODUMP) != 0)
1915 if ((req & VM_ALLOC_WIRED) != 0)
1916 atomic_add_int(&vm_cnt.v_wire_count, npages);
1917 if (object != NULL) {
1918 if (object->memattr != VM_MEMATTR_DEFAULT &&
1919 memattr == VM_MEMATTR_DEFAULT)
1920 memattr = object->memattr;
1922 for (m = m_ret; m < &m_ret[npages]; m++) {
1924 m->flags = (m->flags | PG_NODUMP) & flags;
1925 m->busy_lock = VPB_UNBUSIED;
1926 if (object != NULL) {
1927 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
1928 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1929 if ((req & VM_ALLOC_SBUSY) != 0)
1930 m->busy_lock = VPB_SHARERS_WORD(1);
1932 if ((req & VM_ALLOC_WIRED) != 0)
1934 /* Unmanaged pages don't use "act_count". */
1935 m->oflags = VPO_UNMANAGED;
1936 if (object != NULL) {
1937 if (vm_page_insert(m, object, pindex)) {
1938 vm_page_alloc_contig_vdrop(
1939 &deferred_vdrop_list);
1940 if (vm_paging_needed())
1941 pagedaemon_wakeup();
1942 if ((req & VM_ALLOC_WIRED) != 0)
1943 atomic_subtract_int(&vm_cnt.v_wire_count,
1945 for (m_tmp = m, m = m_ret;
1946 m < &m_ret[npages]; m++) {
1947 if ((req & VM_ALLOC_WIRED) != 0)
1951 m->oflags |= VPO_UNMANAGED;
1953 m->busy_lock = VPB_UNBUSIED;
1960 if (memattr != VM_MEMATTR_DEFAULT)
1961 pmap_page_set_memattr(m, memattr);
1964 vm_page_alloc_contig_vdrop(&deferred_vdrop_list);
1965 if (vm_paging_needed())
1966 pagedaemon_wakeup();
1971 * Initialize a page that has been freshly dequeued from a freelist.
1972 * The caller has to drop the vnode returned, if it is not NULL.
1974 * This function may only be used to initialize unmanaged pages.
1976 * To be called with vm_page_queue_free_mtx held.
1978 static struct vnode *
1979 vm_page_alloc_init(vm_page_t m)
1982 vm_object_t m_object;
1984 KASSERT(m->queue == PQ_NONE,
1985 ("vm_page_alloc_init: page %p has unexpected queue %d",
1987 KASSERT(m->wire_count == 0,
1988 ("vm_page_alloc_init: page %p is wired", m));
1989 KASSERT(m->hold_count == 0,
1990 ("vm_page_alloc_init: page %p is held", m));
1991 KASSERT(!vm_page_sbusied(m),
1992 ("vm_page_alloc_init: page %p is busy", m));
1993 KASSERT(m->dirty == 0,
1994 ("vm_page_alloc_init: page %p is dirty", m));
1995 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1996 ("vm_page_alloc_init: page %p has unexpected memattr %d",
1997 m, pmap_page_get_memattr(m)));
1998 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2000 if ((m->flags & PG_CACHED) != 0) {
2001 KASSERT((m->flags & PG_ZERO) == 0,
2002 ("vm_page_alloc_init: cached page %p is PG_ZERO", m));
2004 m_object = m->object;
2005 vm_page_cache_remove(m);
2006 if (m_object->type == OBJT_VNODE &&
2007 vm_object_cache_is_empty(m_object))
2008 drop = m_object->handle;
2010 KASSERT(m->valid == 0,
2011 ("vm_page_alloc_init: free page %p is valid", m));
2012 vm_phys_freecnt_adj(m, -1);
2013 if ((m->flags & PG_ZERO) != 0)
2014 vm_page_zero_count--;
2020 * vm_page_alloc_freelist:
2022 * Allocate a physical page from the specified free page list.
2024 * The caller must always specify an allocation class.
2026 * allocation classes:
2027 * VM_ALLOC_NORMAL normal process request
2028 * VM_ALLOC_SYSTEM system *really* needs a page
2029 * VM_ALLOC_INTERRUPT interrupt time request
2031 * optional allocation flags:
2032 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
2033 * intends to allocate
2034 * VM_ALLOC_WIRED wire the allocated page
2035 * VM_ALLOC_ZERO prefer a zeroed page
2037 * This routine may not sleep.
2040 vm_page_alloc_freelist(int flind, int req)
2047 req_class = req & VM_ALLOC_CLASS_MASK;
2050 * The page daemon is allowed to dig deeper into the free page list.
2052 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2053 req_class = VM_ALLOC_SYSTEM;
2056 * Do not allocate reserved pages unless the req has asked for it.
2058 mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE);
2059 if (vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_free_reserved ||
2060 (req_class == VM_ALLOC_SYSTEM &&
2061 vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_interrupt_free_min) ||
2062 (req_class == VM_ALLOC_INTERRUPT &&
2063 vm_cnt.v_free_count + vm_cnt.v_cache_count > 0))
2064 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0);
2066 mtx_unlock(&vm_page_queue_free_mtx);
2067 atomic_add_int(&vm_pageout_deficit,
2068 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
2069 pagedaemon_wakeup();
2073 mtx_unlock(&vm_page_queue_free_mtx);
2076 drop = vm_page_alloc_init(m);
2077 mtx_unlock(&vm_page_queue_free_mtx);
2080 * Initialize the page. Only the PG_ZERO flag is inherited.
2084 if ((req & VM_ALLOC_ZERO) != 0)
2087 if ((req & VM_ALLOC_WIRED) != 0) {
2089 * The page lock is not required for wiring a page that does
2090 * not belong to an object.
2092 atomic_add_int(&vm_cnt.v_wire_count, 1);
2095 /* Unmanaged pages don't use "act_count". */
2096 m->oflags = VPO_UNMANAGED;
2099 if (vm_paging_needed())
2100 pagedaemon_wakeup();
2104 #define VPSC_ANY 0 /* No restrictions. */
2105 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */
2106 #define VPSC_NOSUPER 2 /* Skip superpages. */
2109 * vm_page_scan_contig:
2111 * Scan vm_page_array[] between the specified entries "m_start" and
2112 * "m_end" for a run of contiguous physical pages that satisfy the
2113 * specified conditions, and return the lowest page in the run. The
2114 * specified "alignment" determines the alignment of the lowest physical
2115 * page in the run. If the specified "boundary" is non-zero, then the
2116 * run of physical pages cannot span a physical address that is a
2117 * multiple of "boundary".
2119 * "m_end" is never dereferenced, so it need not point to a vm_page
2120 * structure within vm_page_array[].
2122 * "npages" must be greater than zero. "m_start" and "m_end" must not
2123 * span a hole (or discontiguity) in the physical address space. Both
2124 * "alignment" and "boundary" must be a power of two.
2127 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2128 u_long alignment, vm_paddr_t boundary, int options)
2130 struct mtx *m_mtx, *new_mtx;
2134 #if VM_NRESERVLEVEL > 0
2137 int m_inc, order, run_ext, run_len;
2139 KASSERT(npages > 0, ("npages is 0"));
2140 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2141 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2145 for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2146 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2147 ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2150 * If the current page would be the start of a run, check its
2151 * physical address against the end, alignment, and boundary
2152 * conditions. If it doesn't satisfy these conditions, either
2153 * terminate the scan or advance to the next page that
2154 * satisfies the failed condition.
2157 KASSERT(m_run == NULL, ("m_run != NULL"));
2158 if (m + npages > m_end)
2160 pa = VM_PAGE_TO_PHYS(m);
2161 if ((pa & (alignment - 1)) != 0) {
2162 m_inc = atop(roundup2(pa, alignment) - pa);
2165 if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2167 m_inc = atop(roundup2(pa, boundary) - pa);
2171 KASSERT(m_run != NULL, ("m_run == NULL"));
2174 * Avoid releasing and reacquiring the same page lock.
2176 new_mtx = vm_page_lockptr(m);
2177 if (m_mtx != new_mtx) {
2185 if (m->wire_count != 0 || m->hold_count != 0)
2187 #if VM_NRESERVLEVEL > 0
2188 else if ((level = vm_reserv_level(m)) >= 0 &&
2189 (options & VPSC_NORESERV) != 0) {
2191 /* Advance to the end of the reservation. */
2192 pa = VM_PAGE_TO_PHYS(m);
2193 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2197 else if ((object = m->object) != NULL) {
2199 * The page is considered eligible for relocation if
2200 * and only if it could be laundered or reclaimed by
2203 if (!VM_OBJECT_TRYRLOCK(object)) {
2205 VM_OBJECT_RLOCK(object);
2207 if (m->object != object) {
2209 * The page may have been freed.
2211 VM_OBJECT_RUNLOCK(object);
2213 } else if (m->wire_count != 0 ||
2214 m->hold_count != 0) {
2219 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2220 ("page %p is PG_UNHOLDFREE", m));
2221 /* Don't care: PG_NODUMP, PG_WINATCFLS, PG_ZERO. */
2222 if (object->type != OBJT_DEFAULT &&
2223 object->type != OBJT_SWAP &&
2224 object->type != OBJT_VNODE)
2226 else if ((m->flags & PG_CACHED) != 0 ||
2227 m != vm_page_lookup(object, m->pindex)) {
2229 * The page is cached or recently converted
2230 * from cached to free.
2232 #if VM_NRESERVLEVEL > 0
2235 * The page is reserved. Extend the
2236 * current run by one page.
2241 if ((order = m->order) < VM_NFREEORDER) {
2243 * The page is enqueued in the
2244 * physical memory allocator's cache/
2245 * free page queues. Moreover, it is
2246 * the first page in a power-of-two-
2247 * sized run of contiguous cache/free
2248 * pages. Add these pages to the end
2249 * of the current run, and jump
2252 run_ext = 1 << order;
2256 #if VM_NRESERVLEVEL > 0
2257 } else if ((options & VPSC_NOSUPER) != 0 &&
2258 (level = vm_reserv_level_iffullpop(m)) >= 0) {
2260 /* Advance to the end of the superpage. */
2261 pa = VM_PAGE_TO_PHYS(m);
2262 m_inc = atop(roundup2(pa + 1,
2263 vm_reserv_size(level)) - pa);
2265 } else if (object->memattr == VM_MEMATTR_DEFAULT &&
2266 m->queue != PQ_NONE && !vm_page_busied(m)) {
2268 * The page is allocated but eligible for
2269 * relocation. Extend the current run by one
2272 KASSERT(pmap_page_get_memattr(m) ==
2274 ("page %p has an unexpected memattr", m));
2275 KASSERT((m->oflags & (VPO_SWAPINPROG |
2276 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2277 ("page %p has unexpected oflags", m));
2278 /* Don't care: VPO_NOSYNC. */
2283 VM_OBJECT_RUNLOCK(object);
2284 #if VM_NRESERVLEVEL > 0
2285 } else if (level >= 0) {
2287 * The page is reserved but not yet allocated. In
2288 * other words, it is still cached or free. Extend
2289 * the current run by one page.
2293 } else if ((order = m->order) < VM_NFREEORDER) {
2295 * The page is enqueued in the physical memory
2296 * allocator's cache/free page queues. Moreover, it
2297 * is the first page in a power-of-two-sized run of
2298 * contiguous cache/free pages. Add these pages to
2299 * the end of the current run, and jump ahead.
2301 run_ext = 1 << order;
2305 * Skip the page for one of the following reasons: (1)
2306 * It is enqueued in the physical memory allocator's
2307 * cache/free page queues. However, it is not the
2308 * first page in a run of contiguous cache/free pages.
2309 * (This case rarely occurs because the scan is
2310 * performed in ascending order.) (2) It is not
2311 * reserved, and it is transitioning from free to
2312 * allocated. (Conversely, the transition from
2313 * allocated to free for managed pages is blocked by
2314 * the page lock.) (3) It is allocated but not
2315 * contained by an object and not wired, e.g.,
2316 * allocated by Xen's balloon driver.
2322 * Extend or reset the current run of pages.
2337 if (run_len >= npages)
2343 * vm_page_reclaim_run:
2345 * Try to relocate each of the allocated virtual pages within the
2346 * specified run of physical pages to a new physical address. Free the
2347 * physical pages underlying the relocated virtual pages. A virtual page
2348 * is relocatable if and only if it could be laundered or reclaimed by
2349 * the page daemon. Whenever possible, a virtual page is relocated to a
2350 * physical address above "high".
2352 * Returns 0 if every physical page within the run was already free or
2353 * just freed by a successful relocation. Otherwise, returns a non-zero
2354 * value indicating why the last attempt to relocate a virtual page was
2357 * "req_class" must be an allocation class.
2360 vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run,
2363 struct mtx *m_mtx, *new_mtx;
2364 struct spglist free;
2367 vm_page_t m, m_end, m_new;
2368 int error, order, req;
2370 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2371 ("req_class is not an allocation class"));
2375 m_end = m_run + npages;
2377 for (; error == 0 && m < m_end; m++) {
2378 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2379 ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2382 * Avoid releasing and reacquiring the same page lock.
2384 new_mtx = vm_page_lockptr(m);
2385 if (m_mtx != new_mtx) {
2392 if (m->wire_count != 0 || m->hold_count != 0)
2394 else if ((object = m->object) != NULL) {
2396 * The page is relocated if and only if it could be
2397 * laundered or reclaimed by the page daemon.
2399 if (!VM_OBJECT_TRYWLOCK(object)) {
2401 VM_OBJECT_WLOCK(object);
2403 if (m->object != object) {
2405 * The page may have been freed.
2407 VM_OBJECT_WUNLOCK(object);
2409 } else if (m->wire_count != 0 ||
2410 m->hold_count != 0) {
2415 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2416 ("page %p is PG_UNHOLDFREE", m));
2417 /* Don't care: PG_NODUMP, PG_WINATCFLS, PG_ZERO. */
2418 if (object->type != OBJT_DEFAULT &&
2419 object->type != OBJT_SWAP &&
2420 object->type != OBJT_VNODE)
2422 else if ((m->flags & PG_CACHED) != 0 ||
2423 m != vm_page_lookup(object, m->pindex)) {
2425 * The page is cached or recently converted
2426 * from cached to free.
2428 VM_OBJECT_WUNLOCK(object);
2430 } else if (object->memattr != VM_MEMATTR_DEFAULT)
2432 else if (m->queue != PQ_NONE && !vm_page_busied(m)) {
2433 KASSERT(pmap_page_get_memattr(m) ==
2435 ("page %p has an unexpected memattr", m));
2436 KASSERT((m->oflags & (VPO_SWAPINPROG |
2437 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2438 ("page %p has unexpected oflags", m));
2439 /* Don't care: VPO_NOSYNC. */
2440 if (m->valid != 0) {
2442 * First, try to allocate a new page
2443 * that is above "high". Failing
2444 * that, try to allocate a new page
2445 * that is below "m_run". Allocate
2446 * the new page between the end of
2447 * "m_run" and "high" only as a last
2450 req = req_class | VM_ALLOC_NOOBJ;
2451 if ((m->flags & PG_NODUMP) != 0)
2452 req |= VM_ALLOC_NODUMP;
2453 if (trunc_page(high) !=
2454 ~(vm_paddr_t)PAGE_MASK) {
2455 m_new = vm_page_alloc_contig(
2460 VM_MEMATTR_DEFAULT);
2463 if (m_new == NULL) {
2464 pa = VM_PAGE_TO_PHYS(m_run);
2465 m_new = vm_page_alloc_contig(
2467 0, pa - 1, PAGE_SIZE, 0,
2468 VM_MEMATTR_DEFAULT);
2470 if (m_new == NULL) {
2472 m_new = vm_page_alloc_contig(
2474 pa, high, PAGE_SIZE, 0,
2475 VM_MEMATTR_DEFAULT);
2477 if (m_new == NULL) {
2481 KASSERT(m_new->wire_count == 0,
2482 ("page %p is wired", m));
2485 * Replace "m" with the new page. For
2486 * vm_page_replace(), "m" must be busy
2487 * and dequeued. Finally, change "m"
2488 * as if vm_page_free() was called.
2490 if (object->ref_count != 0)
2492 m_new->aflags = m->aflags;
2493 KASSERT(m_new->oflags == VPO_UNMANAGED,
2494 ("page %p is managed", m));
2495 m_new->oflags = m->oflags & VPO_NOSYNC;
2496 pmap_copy_page(m, m_new);
2497 m_new->valid = m->valid;
2498 m_new->dirty = m->dirty;
2499 m->flags &= ~PG_ZERO;
2502 vm_page_replace_checked(m_new, object,
2508 * The new page must be deactivated
2509 * before the object is unlocked.
2511 new_mtx = vm_page_lockptr(m_new);
2512 if (m_mtx != new_mtx) {
2517 vm_page_deactivate(m_new);
2519 m->flags &= ~PG_ZERO;
2522 KASSERT(m->dirty == 0,
2523 ("page %p is dirty", m));
2525 SLIST_INSERT_HEAD(&free, m, plinks.s.ss);
2529 VM_OBJECT_WUNLOCK(object);
2532 mtx_lock(&vm_page_queue_free_mtx);
2534 if (order < VM_NFREEORDER) {
2536 * The page is enqueued in the physical memory
2537 * allocator's cache/free page queues.
2538 * Moreover, it is the first page in a power-
2539 * of-two-sized run of contiguous cache/free
2540 * pages. Jump ahead to the last page within
2541 * that run, and continue from there.
2543 m += (1 << order) - 1;
2545 #if VM_NRESERVLEVEL > 0
2546 else if (vm_reserv_is_page_free(m))
2549 mtx_unlock(&vm_page_queue_free_mtx);
2550 if (order == VM_NFREEORDER)
2556 if ((m = SLIST_FIRST(&free)) != NULL) {
2557 mtx_lock(&vm_page_queue_free_mtx);
2559 SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2560 vm_phys_freecnt_adj(m, 1);
2561 #if VM_NRESERVLEVEL > 0
2562 if (!vm_reserv_free_page(m))
2566 vm_phys_free_pages(m, 0);
2567 } while ((m = SLIST_FIRST(&free)) != NULL);
2568 vm_page_zero_idle_wakeup();
2569 vm_page_free_wakeup();
2570 mtx_unlock(&vm_page_queue_free_mtx);
2577 CTASSERT(powerof2(NRUNS));
2579 #define RUN_INDEX(count) ((count) & (NRUNS - 1))
2581 #define MIN_RECLAIM 8
2584 * vm_page_reclaim_contig:
2586 * Reclaim allocated, contiguous physical memory satisfying the specified
2587 * conditions by relocating the virtual pages using that physical memory.
2588 * Returns true if reclamation is successful and false otherwise. Since
2589 * relocation requires the allocation of physical pages, reclamation may
2590 * fail due to a shortage of cache/free pages. When reclamation fails,
2591 * callers are expected to perform VM_WAIT before retrying a failed
2592 * allocation operation, e.g., vm_page_alloc_contig().
2594 * The caller must always specify an allocation class through "req".
2596 * allocation classes:
2597 * VM_ALLOC_NORMAL normal process request
2598 * VM_ALLOC_SYSTEM system *really* needs a page
2599 * VM_ALLOC_INTERRUPT interrupt time request
2601 * The optional allocation flags are ignored.
2603 * "npages" must be greater than zero. Both "alignment" and "boundary"
2604 * must be a power of two.
2607 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
2608 u_long alignment, vm_paddr_t boundary)
2610 vm_paddr_t curr_low;
2611 vm_page_t m_run, m_runs[NRUNS];
2612 u_long count, reclaimed;
2613 int error, i, options, req_class;
2615 KASSERT(npages > 0, ("npages is 0"));
2616 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2617 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2618 req_class = req & VM_ALLOC_CLASS_MASK;
2621 * The page daemon is allowed to dig deeper into the free page list.
2623 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2624 req_class = VM_ALLOC_SYSTEM;
2627 * Return if the number of cached and free pages cannot satisfy the
2628 * requested allocation.
2630 count = vm_cnt.v_free_count + vm_cnt.v_cache_count;
2631 if (count < npages + vm_cnt.v_free_reserved || (count < npages +
2632 vm_cnt.v_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2633 (count < npages && req_class == VM_ALLOC_INTERRUPT))
2637 * Scan up to three times, relaxing the restrictions ("options") on
2638 * the reclamation of reservations and superpages each time.
2640 for (options = VPSC_NORESERV;;) {
2642 * Find the highest runs that satisfy the given constraints
2643 * and restrictions, and record them in "m_runs".
2648 m_run = vm_phys_scan_contig(npages, curr_low, high,
2649 alignment, boundary, options);
2652 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2653 m_runs[RUN_INDEX(count)] = m_run;
2658 * Reclaim the highest runs in LIFO (descending) order until
2659 * the number of reclaimed pages, "reclaimed", is at least
2660 * MIN_RECLAIM. Reset "reclaimed" each time because each
2661 * reclamation is idempotent, and runs will (likely) recur
2662 * from one scan to the next as restrictions are relaxed.
2665 for (i = 0; count > 0 && i < NRUNS; i++) {
2667 m_run = m_runs[RUN_INDEX(count)];
2668 error = vm_page_reclaim_run(req_class, npages, m_run,
2671 reclaimed += npages;
2672 if (reclaimed >= MIN_RECLAIM)
2678 * Either relax the restrictions on the next scan or return if
2679 * the last scan had no restrictions.
2681 if (options == VPSC_NORESERV)
2682 options = VPSC_NOSUPER;
2683 else if (options == VPSC_NOSUPER)
2685 else if (options == VPSC_ANY)
2686 return (reclaimed != 0);
2691 * vm_wait: (also see VM_WAIT macro)
2693 * Sleep until free pages are available for allocation.
2694 * - Called in various places before memory allocations.
2700 mtx_lock(&vm_page_queue_free_mtx);
2701 if (curproc == pageproc) {
2702 vm_pageout_pages_needed = 1;
2703 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx,
2704 PDROP | PSWP, "VMWait", 0);
2706 if (!vm_pageout_wanted) {
2707 vm_pageout_wanted = true;
2708 wakeup(&vm_pageout_wanted);
2710 vm_pages_needed = true;
2711 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM,
2717 * vm_waitpfault: (also see VM_WAITPFAULT macro)
2719 * Sleep until free pages are available for allocation.
2720 * - Called only in vm_fault so that processes page faulting
2721 * can be easily tracked.
2722 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
2723 * processes will be able to grab memory first. Do not change
2724 * this balance without careful testing first.
2730 mtx_lock(&vm_page_queue_free_mtx);
2731 if (!vm_pageout_wanted) {
2732 vm_pageout_wanted = true;
2733 wakeup(&vm_pageout_wanted);
2735 vm_pages_needed = true;
2736 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER,
2740 struct vm_pagequeue *
2741 vm_page_pagequeue(vm_page_t m)
2744 return (&vm_phys_domain(m)->vmd_pagequeues[m->queue]);
2750 * Remove the given page from its current page queue.
2752 * The page must be locked.
2755 vm_page_dequeue(vm_page_t m)
2757 struct vm_pagequeue *pq;
2759 vm_page_assert_locked(m);
2760 KASSERT(m->queue < PQ_COUNT, ("vm_page_dequeue: page %p is not queued",
2762 pq = vm_page_pagequeue(m);
2763 vm_pagequeue_lock(pq);
2765 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2766 vm_pagequeue_cnt_dec(pq);
2767 vm_pagequeue_unlock(pq);
2771 * vm_page_dequeue_locked:
2773 * Remove the given page from its current page queue.
2775 * The page and page queue must be locked.
2778 vm_page_dequeue_locked(vm_page_t m)
2780 struct vm_pagequeue *pq;
2782 vm_page_lock_assert(m, MA_OWNED);
2783 pq = vm_page_pagequeue(m);
2784 vm_pagequeue_assert_locked(pq);
2786 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2787 vm_pagequeue_cnt_dec(pq);
2793 * Add the given page to the specified page queue.
2795 * The page must be locked.
2798 vm_page_enqueue(uint8_t queue, vm_page_t m)
2800 struct vm_pagequeue *pq;
2802 vm_page_lock_assert(m, MA_OWNED);
2803 KASSERT(queue < PQ_COUNT,
2804 ("vm_page_enqueue: invalid queue %u request for page %p",
2806 pq = &vm_phys_domain(m)->vmd_pagequeues[queue];
2807 vm_pagequeue_lock(pq);
2809 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2810 vm_pagequeue_cnt_inc(pq);
2811 vm_pagequeue_unlock(pq);
2817 * Move the given page to the tail of its current page queue.
2819 * The page must be locked.
2822 vm_page_requeue(vm_page_t m)
2824 struct vm_pagequeue *pq;
2826 vm_page_lock_assert(m, MA_OWNED);
2827 KASSERT(m->queue != PQ_NONE,
2828 ("vm_page_requeue: page %p is not queued", m));
2829 pq = vm_page_pagequeue(m);
2830 vm_pagequeue_lock(pq);
2831 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2832 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2833 vm_pagequeue_unlock(pq);
2837 * vm_page_requeue_locked:
2839 * Move the given page to the tail of its current page queue.
2841 * The page queue must be locked.
2844 vm_page_requeue_locked(vm_page_t m)
2846 struct vm_pagequeue *pq;
2848 KASSERT(m->queue != PQ_NONE,
2849 ("vm_page_requeue_locked: page %p is not queued", m));
2850 pq = vm_page_pagequeue(m);
2851 vm_pagequeue_assert_locked(pq);
2852 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2853 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2859 * Put the specified page on the active list (if appropriate).
2860 * Ensure that act_count is at least ACT_INIT but do not otherwise
2863 * The page must be locked.
2866 vm_page_activate(vm_page_t m)
2870 vm_page_lock_assert(m, MA_OWNED);
2871 if ((queue = m->queue) != PQ_ACTIVE) {
2872 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2873 if (m->act_count < ACT_INIT)
2874 m->act_count = ACT_INIT;
2875 if (queue != PQ_NONE)
2877 vm_page_enqueue(PQ_ACTIVE, m);
2879 KASSERT(queue == PQ_NONE,
2880 ("vm_page_activate: wired page %p is queued", m));
2882 if (m->act_count < ACT_INIT)
2883 m->act_count = ACT_INIT;
2888 * vm_page_free_wakeup:
2890 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
2891 * routine is called when a page has been added to the cache or free
2894 * The page queues must be locked.
2897 vm_page_free_wakeup(void)
2900 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2902 * if pageout daemon needs pages, then tell it that there are
2905 if (vm_pageout_pages_needed &&
2906 vm_cnt.v_cache_count + vm_cnt.v_free_count >= vm_cnt.v_pageout_free_min) {
2907 wakeup(&vm_pageout_pages_needed);
2908 vm_pageout_pages_needed = 0;
2911 * wakeup processes that are waiting on memory if we hit a
2912 * high water mark. And wakeup scheduler process if we have
2913 * lots of memory. this process will swapin processes.
2915 if (vm_pages_needed && !vm_page_count_min()) {
2916 vm_pages_needed = false;
2917 wakeup(&vm_cnt.v_free_count);
2922 * Turn a cached page into a free page, by changing its attributes.
2923 * Keep the statistics up-to-date.
2925 * The free page queue must be locked.
2928 vm_page_cache_turn_free(vm_page_t m)
2931 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2935 KASSERT((m->flags & PG_CACHED) != 0,
2936 ("vm_page_cache_turn_free: page %p is not cached", m));
2937 m->flags &= ~PG_CACHED;
2938 vm_cnt.v_cache_count--;
2939 vm_phys_freecnt_adj(m, 1);
2945 * Returns the given page to the free list,
2946 * disassociating it with any VM object.
2948 * The object must be locked. The page must be locked if it is managed.
2951 vm_page_free_toq(vm_page_t m)
2954 if ((m->oflags & VPO_UNMANAGED) == 0) {
2955 vm_page_lock_assert(m, MA_OWNED);
2956 KASSERT(!pmap_page_is_mapped(m),
2957 ("vm_page_free_toq: freeing mapped page %p", m));
2959 KASSERT(m->queue == PQ_NONE,
2960 ("vm_page_free_toq: unmanaged page %p is queued", m));
2961 PCPU_INC(cnt.v_tfree);
2963 if (vm_page_sbusied(m))
2964 panic("vm_page_free: freeing busy page %p", m);
2967 * Unqueue, then remove page. Note that we cannot destroy
2968 * the page here because we do not want to call the pager's
2969 * callback routine until after we've put the page on the
2970 * appropriate free queue.
2976 * If fictitious remove object association and
2977 * return, otherwise delay object association removal.
2979 if ((m->flags & PG_FICTITIOUS) != 0) {
2986 if (m->wire_count != 0)
2987 panic("vm_page_free: freeing wired page %p", m);
2988 if (m->hold_count != 0) {
2989 m->flags &= ~PG_ZERO;
2990 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2991 ("vm_page_free: freeing PG_UNHOLDFREE page %p", m));
2992 m->flags |= PG_UNHOLDFREE;
2995 * Restore the default memory attribute to the page.
2997 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2998 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3001 * Insert the page into the physical memory allocator's
3002 * cache/free page queues.
3004 mtx_lock(&vm_page_queue_free_mtx);
3005 vm_phys_freecnt_adj(m, 1);
3006 #if VM_NRESERVLEVEL > 0
3007 if (!vm_reserv_free_page(m))
3011 vm_phys_free_pages(m, 0);
3012 if ((m->flags & PG_ZERO) != 0)
3013 ++vm_page_zero_count;
3015 vm_page_zero_idle_wakeup();
3016 vm_page_free_wakeup();
3017 mtx_unlock(&vm_page_queue_free_mtx);
3024 * Mark this page as wired down by yet
3025 * another map, removing it from paging queues
3028 * If the page is fictitious, then its wire count must remain one.
3030 * The page must be locked.
3033 vm_page_wire(vm_page_t m)
3037 * Only bump the wire statistics if the page is not already wired,
3038 * and only unqueue the page if it is on some queue (if it is unmanaged
3039 * it is already off the queues).
3041 vm_page_lock_assert(m, MA_OWNED);
3042 if ((m->flags & PG_FICTITIOUS) != 0) {
3043 KASSERT(m->wire_count == 1,
3044 ("vm_page_wire: fictitious page %p's wire count isn't one",
3048 if (m->wire_count == 0) {
3049 KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
3050 m->queue == PQ_NONE,
3051 ("vm_page_wire: unmanaged page %p is queued", m));
3053 atomic_add_int(&vm_cnt.v_wire_count, 1);
3056 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
3062 * Release one wiring of the specified page, potentially allowing it to be
3063 * paged out. Returns TRUE if the number of wirings transitions to zero and
3066 * Only managed pages belonging to an object can be paged out. If the number
3067 * of wirings transitions to zero and the page is eligible for page out, then
3068 * the page is added to the specified paging queue (unless PQ_NONE is
3071 * If a page is fictitious, then its wire count must always be one.
3073 * A managed page must be locked.
3076 vm_page_unwire(vm_page_t m, uint8_t queue)
3079 KASSERT(queue < PQ_COUNT || queue == PQ_NONE,
3080 ("vm_page_unwire: invalid queue %u request for page %p",
3082 if ((m->oflags & VPO_UNMANAGED) == 0)
3083 vm_page_assert_locked(m);
3084 if ((m->flags & PG_FICTITIOUS) != 0) {
3085 KASSERT(m->wire_count == 1,
3086 ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
3089 if (m->wire_count > 0) {
3091 if (m->wire_count == 0) {
3092 atomic_subtract_int(&vm_cnt.v_wire_count, 1);
3093 if ((m->oflags & VPO_UNMANAGED) == 0 &&
3094 m->object != NULL && queue != PQ_NONE) {
3095 if (queue == PQ_INACTIVE)
3096 m->flags &= ~PG_WINATCFLS;
3097 vm_page_enqueue(queue, m);
3103 panic("vm_page_unwire: page %p's wire count is zero", m);
3107 * Move the specified page to the inactive queue.
3109 * Many pages placed on the inactive queue should actually go
3110 * into the cache, but it is difficult to figure out which. What
3111 * we do instead, if the inactive target is well met, is to put
3112 * clean pages at the head of the inactive queue instead of the tail.
3113 * This will cause them to be moved to the cache more quickly and
3114 * if not actively re-referenced, reclaimed more quickly. If we just
3115 * stick these pages at the end of the inactive queue, heavy filesystem
3116 * meta-data accesses can cause an unnecessary paging load on memory bound
3117 * processes. This optimization causes one-time-use metadata to be
3118 * reused more quickly.
3120 * Normally noreuse is FALSE, resulting in LRU operation. noreuse is set
3121 * to TRUE if we want this page to be 'as if it were placed in the cache',
3122 * except without unmapping it from the process address space. In
3123 * practice this is implemented by inserting the page at the head of the
3124 * queue, using a marker page to guide FIFO insertion ordering.
3126 * The page must be locked.
3129 _vm_page_deactivate(vm_page_t m, boolean_t noreuse)
3131 struct vm_pagequeue *pq;
3134 vm_page_assert_locked(m);
3137 * Ignore if the page is already inactive, unless it is unlikely to be
3140 if ((queue = m->queue) == PQ_INACTIVE && !noreuse)
3142 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
3143 pq = &vm_phys_domain(m)->vmd_pagequeues[PQ_INACTIVE];
3144 /* Avoid multiple acquisitions of the inactive queue lock. */
3145 if (queue == PQ_INACTIVE) {
3146 vm_pagequeue_lock(pq);
3147 vm_page_dequeue_locked(m);
3149 if (queue != PQ_NONE)
3151 m->flags &= ~PG_WINATCFLS;
3152 vm_pagequeue_lock(pq);
3154 m->queue = PQ_INACTIVE;
3156 TAILQ_INSERT_BEFORE(&vm_phys_domain(m)->vmd_inacthead,
3159 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3160 vm_pagequeue_cnt_inc(pq);
3161 vm_pagequeue_unlock(pq);
3166 * Move the specified page to the inactive queue.
3168 * The page must be locked.
3171 vm_page_deactivate(vm_page_t m)
3174 _vm_page_deactivate(m, FALSE);
3178 * Move the specified page to the inactive queue with the expectation
3179 * that it is unlikely to be reused.
3181 * The page must be locked.
3184 vm_page_deactivate_noreuse(vm_page_t m)
3187 _vm_page_deactivate(m, TRUE);
3191 * vm_page_try_to_cache:
3193 * Returns 0 on failure, 1 on success
3196 vm_page_try_to_cache(vm_page_t m)
3199 vm_page_lock_assert(m, MA_OWNED);
3200 VM_OBJECT_ASSERT_WLOCKED(m->object);
3201 if (m->dirty || m->hold_count || m->wire_count ||
3202 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m))
3212 * vm_page_try_to_free()
3214 * Attempt to free the page. If we cannot free it, we do nothing.
3215 * 1 is returned on success, 0 on failure.
3218 vm_page_try_to_free(vm_page_t m)
3221 vm_page_lock_assert(m, MA_OWNED);
3222 if (m->object != NULL)
3223 VM_OBJECT_ASSERT_WLOCKED(m->object);
3224 if (m->dirty || m->hold_count || m->wire_count ||
3225 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m))
3237 * Put the specified page onto the page cache queue (if appropriate).
3239 * The object and page must be locked.
3242 vm_page_cache(vm_page_t m)
3245 boolean_t cache_was_empty;
3247 vm_page_lock_assert(m, MA_OWNED);
3249 VM_OBJECT_ASSERT_WLOCKED(object);
3250 if (vm_page_busied(m) || (m->oflags & VPO_UNMANAGED) ||
3251 m->hold_count || m->wire_count)
3252 panic("vm_page_cache: attempting to cache busy page");
3253 KASSERT(!pmap_page_is_mapped(m),
3254 ("vm_page_cache: page %p is mapped", m));
3255 KASSERT(m->dirty == 0, ("vm_page_cache: page %p is dirty", m));
3256 if (m->valid == 0 || object->type == OBJT_DEFAULT ||
3257 (object->type == OBJT_SWAP &&
3258 !vm_pager_has_page(object, m->pindex, NULL, NULL))) {
3260 * Hypothesis: A cache-eligible page belonging to a
3261 * default object or swap object but without a backing
3262 * store must be zero filled.
3267 KASSERT((m->flags & PG_CACHED) == 0,
3268 ("vm_page_cache: page %p is already cached", m));
3271 * Remove the page from the paging queues.
3276 * Remove the page from the object's collection of resident
3279 vm_radix_remove(&object->rtree, m->pindex);
3280 TAILQ_REMOVE(&object->memq, m, listq);
3281 object->resident_page_count--;
3284 * Restore the default memory attribute to the page.
3286 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3287 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3290 * Insert the page into the object's collection of cached pages
3291 * and the physical memory allocator's cache/free page queues.
3293 m->flags &= ~PG_ZERO;
3294 mtx_lock(&vm_page_queue_free_mtx);
3295 cache_was_empty = vm_radix_is_empty(&object->cache);
3296 if (vm_radix_insert(&object->cache, m)) {
3297 mtx_unlock(&vm_page_queue_free_mtx);
3298 if (object->type == OBJT_VNODE &&
3299 object->resident_page_count == 0)
3300 vdrop(object->handle);
3307 * The above call to vm_radix_insert() could reclaim the one pre-
3308 * existing cached page from this object, resulting in a call to
3311 if (!cache_was_empty)
3312 cache_was_empty = vm_radix_is_singleton(&object->cache);
3314 m->flags |= PG_CACHED;
3315 vm_cnt.v_cache_count++;
3316 PCPU_INC(cnt.v_tcached);
3317 #if VM_NRESERVLEVEL > 0
3318 if (!vm_reserv_free_page(m)) {
3322 vm_phys_free_pages(m, 0);
3324 vm_page_free_wakeup();
3325 mtx_unlock(&vm_page_queue_free_mtx);
3328 * Increment the vnode's hold count if this is the object's only
3329 * cached page. Decrement the vnode's hold count if this was
3330 * the object's only resident page.
3332 if (object->type == OBJT_VNODE) {
3333 if (cache_was_empty && object->resident_page_count != 0)
3334 vhold(object->handle);
3335 else if (!cache_was_empty && object->resident_page_count == 0)
3336 vdrop(object->handle);
3343 * Deactivate or do nothing, as appropriate.
3345 * The object and page must be locked.
3348 vm_page_advise(vm_page_t m, int advice)
3351 vm_page_assert_locked(m);
3352 VM_OBJECT_ASSERT_WLOCKED(m->object);
3353 if (advice == MADV_FREE)
3355 * Mark the page clean. This will allow the page to be freed
3356 * up by the system. However, such pages are often reused
3357 * quickly by malloc() so we do not do anything that would
3358 * cause a page fault if we can help it.
3360 * Specifically, we do not try to actually free the page now
3361 * nor do we try to put it in the cache (which would cause a
3362 * page fault on reuse).
3364 * But we do make the page as freeable as we can without
3365 * actually taking the step of unmapping it.
3368 else if (advice != MADV_DONTNEED)
3372 * Clear any references to the page. Otherwise, the page daemon will
3373 * immediately reactivate the page.
3375 vm_page_aflag_clear(m, PGA_REFERENCED);
3377 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
3381 * Place clean pages at the head of the inactive queue rather than the
3382 * tail, thus defeating the queue's LRU operation and ensuring that the
3383 * page will be reused quickly.
3385 _vm_page_deactivate(m, m->dirty == 0);
3389 * Grab a page, waiting until we are waken up due to the page
3390 * changing state. We keep on waiting, if the page continues
3391 * to be in the object. If the page doesn't exist, first allocate it
3392 * and then conditionally zero it.
3394 * This routine may sleep.
3396 * The object must be locked on entry. The lock will, however, be released
3397 * and reacquired if the routine sleeps.
3400 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
3405 VM_OBJECT_ASSERT_WLOCKED(object);
3406 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3407 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3408 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
3410 if ((m = vm_page_lookup(object, pindex)) != NULL) {
3411 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3412 vm_page_xbusied(m) : vm_page_busied(m);
3414 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3417 * Reference the page before unlocking and
3418 * sleeping so that the page daemon is less
3419 * likely to reclaim it.
3421 vm_page_aflag_set(m, PGA_REFERENCED);
3423 VM_OBJECT_WUNLOCK(object);
3424 vm_page_busy_sleep(m, "pgrbwt");
3425 VM_OBJECT_WLOCK(object);
3428 if ((allocflags & VM_ALLOC_WIRED) != 0) {
3434 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
3436 if ((allocflags & VM_ALLOC_SBUSY) != 0)
3441 m = vm_page_alloc(object, pindex, allocflags);
3443 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3445 VM_OBJECT_WUNLOCK(object);
3447 VM_OBJECT_WLOCK(object);
3449 } else if (m->valid != 0)
3451 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
3457 * Mapping function for valid or dirty bits in a page.
3459 * Inputs are required to range within a page.
3462 vm_page_bits(int base, int size)
3468 base + size <= PAGE_SIZE,
3469 ("vm_page_bits: illegal base/size %d/%d", base, size)
3472 if (size == 0) /* handle degenerate case */
3475 first_bit = base >> DEV_BSHIFT;
3476 last_bit = (base + size - 1) >> DEV_BSHIFT;
3478 return (((vm_page_bits_t)2 << last_bit) -
3479 ((vm_page_bits_t)1 << first_bit));
3483 * vm_page_set_valid_range:
3485 * Sets portions of a page valid. The arguments are expected
3486 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3487 * of any partial chunks touched by the range. The invalid portion of
3488 * such chunks will be zeroed.
3490 * (base + size) must be less then or equal to PAGE_SIZE.
3493 vm_page_set_valid_range(vm_page_t m, int base, int size)
3497 VM_OBJECT_ASSERT_WLOCKED(m->object);
3498 if (size == 0) /* handle degenerate case */
3502 * If the base is not DEV_BSIZE aligned and the valid
3503 * bit is clear, we have to zero out a portion of the
3506 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
3507 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
3508 pmap_zero_page_area(m, frag, base - frag);
3511 * If the ending offset is not DEV_BSIZE aligned and the
3512 * valid bit is clear, we have to zero out a portion of
3515 endoff = base + size;
3516 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
3517 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
3518 pmap_zero_page_area(m, endoff,
3519 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
3522 * Assert that no previously invalid block that is now being validated
3525 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
3526 ("vm_page_set_valid_range: page %p is dirty", m));
3529 * Set valid bits inclusive of any overlap.
3531 m->valid |= vm_page_bits(base, size);
3535 * Clear the given bits from the specified page's dirty field.
3537 static __inline void
3538 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
3541 #if PAGE_SIZE < 16384
3546 * If the object is locked and the page is neither exclusive busy nor
3547 * write mapped, then the page's dirty field cannot possibly be
3548 * set by a concurrent pmap operation.
3550 VM_OBJECT_ASSERT_WLOCKED(m->object);
3551 if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
3552 m->dirty &= ~pagebits;
3555 * The pmap layer can call vm_page_dirty() without
3556 * holding a distinguished lock. The combination of
3557 * the object's lock and an atomic operation suffice
3558 * to guarantee consistency of the page dirty field.
3560 * For PAGE_SIZE == 32768 case, compiler already
3561 * properly aligns the dirty field, so no forcible
3562 * alignment is needed. Only require existence of
3563 * atomic_clear_64 when page size is 32768.
3565 addr = (uintptr_t)&m->dirty;
3566 #if PAGE_SIZE == 32768
3567 atomic_clear_64((uint64_t *)addr, pagebits);
3568 #elif PAGE_SIZE == 16384
3569 atomic_clear_32((uint32_t *)addr, pagebits);
3570 #else /* PAGE_SIZE <= 8192 */
3572 * Use a trick to perform a 32-bit atomic on the
3573 * containing aligned word, to not depend on the existence
3574 * of atomic_clear_{8, 16}.
3576 shift = addr & (sizeof(uint32_t) - 1);
3577 #if BYTE_ORDER == BIG_ENDIAN
3578 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
3582 addr &= ~(sizeof(uint32_t) - 1);
3583 atomic_clear_32((uint32_t *)addr, pagebits << shift);
3584 #endif /* PAGE_SIZE */
3589 * vm_page_set_validclean:
3591 * Sets portions of a page valid and clean. The arguments are expected
3592 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3593 * of any partial chunks touched by the range. The invalid portion of
3594 * such chunks will be zero'd.
3596 * (base + size) must be less then or equal to PAGE_SIZE.
3599 vm_page_set_validclean(vm_page_t m, int base, int size)
3601 vm_page_bits_t oldvalid, pagebits;
3604 VM_OBJECT_ASSERT_WLOCKED(m->object);
3605 if (size == 0) /* handle degenerate case */
3609 * If the base is not DEV_BSIZE aligned and the valid
3610 * bit is clear, we have to zero out a portion of the
3613 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
3614 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
3615 pmap_zero_page_area(m, frag, base - frag);
3618 * If the ending offset is not DEV_BSIZE aligned and the
3619 * valid bit is clear, we have to zero out a portion of
3622 endoff = base + size;
3623 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
3624 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
3625 pmap_zero_page_area(m, endoff,
3626 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
3629 * Set valid, clear dirty bits. If validating the entire
3630 * page we can safely clear the pmap modify bit. We also
3631 * use this opportunity to clear the VPO_NOSYNC flag. If a process
3632 * takes a write fault on a MAP_NOSYNC memory area the flag will
3635 * We set valid bits inclusive of any overlap, but we can only
3636 * clear dirty bits for DEV_BSIZE chunks that are fully within
3639 oldvalid = m->valid;
3640 pagebits = vm_page_bits(base, size);
3641 m->valid |= pagebits;
3643 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
3644 frag = DEV_BSIZE - frag;
3650 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
3652 if (base == 0 && size == PAGE_SIZE) {
3654 * The page can only be modified within the pmap if it is
3655 * mapped, and it can only be mapped if it was previously
3658 if (oldvalid == VM_PAGE_BITS_ALL)
3660 * Perform the pmap_clear_modify() first. Otherwise,
3661 * a concurrent pmap operation, such as
3662 * pmap_protect(), could clear a modification in the
3663 * pmap and set the dirty field on the page before
3664 * pmap_clear_modify() had begun and after the dirty
3665 * field was cleared here.
3667 pmap_clear_modify(m);
3669 m->oflags &= ~VPO_NOSYNC;
3670 } else if (oldvalid != VM_PAGE_BITS_ALL)
3671 m->dirty &= ~pagebits;
3673 vm_page_clear_dirty_mask(m, pagebits);
3677 vm_page_clear_dirty(vm_page_t m, int base, int size)
3680 vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
3684 * vm_page_set_invalid:
3686 * Invalidates DEV_BSIZE'd chunks within a page. Both the
3687 * valid and dirty bits for the effected areas are cleared.
3690 vm_page_set_invalid(vm_page_t m, int base, int size)
3692 vm_page_bits_t bits;
3696 VM_OBJECT_ASSERT_WLOCKED(object);
3697 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
3698 size >= object->un_pager.vnp.vnp_size)
3699 bits = VM_PAGE_BITS_ALL;
3701 bits = vm_page_bits(base, size);
3702 if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL &&
3705 KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) ||
3706 !pmap_page_is_mapped(m),
3707 ("vm_page_set_invalid: page %p is mapped", m));
3713 * vm_page_zero_invalid()
3715 * The kernel assumes that the invalid portions of a page contain
3716 * garbage, but such pages can be mapped into memory by user code.
3717 * When this occurs, we must zero out the non-valid portions of the
3718 * page so user code sees what it expects.
3720 * Pages are most often semi-valid when the end of a file is mapped
3721 * into memory and the file's size is not page aligned.
3724 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
3729 VM_OBJECT_ASSERT_WLOCKED(m->object);
3731 * Scan the valid bits looking for invalid sections that
3732 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the
3733 * valid bit may be set ) have already been zeroed by
3734 * vm_page_set_validclean().
3736 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
3737 if (i == (PAGE_SIZE / DEV_BSIZE) ||
3738 (m->valid & ((vm_page_bits_t)1 << i))) {
3740 pmap_zero_page_area(m,
3741 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
3748 * setvalid is TRUE when we can safely set the zero'd areas
3749 * as being valid. We can do this if there are no cache consistancy
3750 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
3753 m->valid = VM_PAGE_BITS_ALL;
3759 * Is (partial) page valid? Note that the case where size == 0
3760 * will return FALSE in the degenerate case where the page is
3761 * entirely invalid, and TRUE otherwise.
3764 vm_page_is_valid(vm_page_t m, int base, int size)
3766 vm_page_bits_t bits;
3768 VM_OBJECT_ASSERT_LOCKED(m->object);
3769 bits = vm_page_bits(base, size);
3770 return (m->valid != 0 && (m->valid & bits) == bits);
3774 * vm_page_ps_is_valid:
3776 * Returns TRUE if the entire (super)page is valid and FALSE otherwise.
3779 vm_page_ps_is_valid(vm_page_t m)
3783 VM_OBJECT_ASSERT_LOCKED(m->object);
3784 npages = atop(pagesizes[m->psind]);
3787 * The physically contiguous pages that make up a superpage, i.e., a
3788 * page with a page size index ("psind") greater than zero, will
3789 * occupy adjacent entries in vm_page_array[].
3791 for (i = 0; i < npages; i++) {
3792 if (m[i].valid != VM_PAGE_BITS_ALL)
3799 * Set the page's dirty bits if the page is modified.
3802 vm_page_test_dirty(vm_page_t m)
3805 VM_OBJECT_ASSERT_WLOCKED(m->object);
3806 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
3811 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
3814 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
3818 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
3821 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
3825 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
3828 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
3831 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
3833 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
3836 vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
3840 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
3843 mtx_assert_(vm_page_lockptr(m), a, file, line);
3849 vm_page_object_lock_assert(vm_page_t m)
3853 * Certain of the page's fields may only be modified by the
3854 * holder of the containing object's lock or the exclusive busy.
3855 * holder. Unfortunately, the holder of the write busy is
3856 * not recorded, and thus cannot be checked here.
3858 if (m->object != NULL && !vm_page_xbusied(m))
3859 VM_OBJECT_ASSERT_WLOCKED(m->object);
3863 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits)
3866 if ((bits & PGA_WRITEABLE) == 0)
3870 * The PGA_WRITEABLE flag can only be set if the page is
3871 * managed, is exclusively busied or the object is locked.
3872 * Currently, this flag is only set by pmap_enter().
3874 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3875 ("PGA_WRITEABLE on unmanaged page"));
3876 if (!vm_page_xbusied(m))
3877 VM_OBJECT_ASSERT_LOCKED(m->object);
3881 #include "opt_ddb.h"
3883 #include <sys/kernel.h>
3885 #include <ddb/ddb.h>
3887 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3889 db_printf("vm_cnt.v_free_count: %d\n", vm_cnt.v_free_count);
3890 db_printf("vm_cnt.v_cache_count: %d\n", vm_cnt.v_cache_count);
3891 db_printf("vm_cnt.v_inactive_count: %d\n", vm_cnt.v_inactive_count);
3892 db_printf("vm_cnt.v_active_count: %d\n", vm_cnt.v_active_count);
3893 db_printf("vm_cnt.v_wire_count: %d\n", vm_cnt.v_wire_count);
3894 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
3895 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
3896 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
3897 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
3900 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3904 db_printf("pq_free %d pq_cache %d\n",
3905 vm_cnt.v_free_count, vm_cnt.v_cache_count);
3906 for (dom = 0; dom < vm_ndomains; dom++) {
3908 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pass %d\n",
3910 vm_dom[dom].vmd_page_count,
3911 vm_dom[dom].vmd_free_count,
3912 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
3913 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
3914 vm_dom[dom].vmd_pass);
3918 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
3924 db_printf("show pginfo addr\n");
3928 phys = strchr(modif, 'p') != NULL;
3930 m = PHYS_TO_VM_PAGE(addr);
3932 m = (vm_page_t)addr;
3934 "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n"
3935 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
3936 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
3937 m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags,
3938 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty);