2 * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
4 * Copyright (c) 1991 Regents of the University of California.
6 * Copyright (c) 1998 Matthew Dillon. All Rights Reserved.
8 * This code is derived from software contributed to Berkeley by
9 * The Mach Operating System project at Carnegie-Mellon University.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
14 * 1. Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
16 * 2. Redistributions in binary form must reproduce the above copyright
17 * notice, this list of conditions and the following disclaimer in the
18 * documentation and/or other materials provided with the distribution.
19 * 3. Neither the name of the University nor the names of its contributors
20 * may be used to endorse or promote products derived from this software
21 * without specific prior written permission.
23 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
24 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
27 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91
39 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40 * All rights reserved.
42 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
44 * Permission to use, copy, modify and distribute this software and
45 * its documentation is hereby granted, provided that both the copyright
46 * notice and this permission notice appear in all copies of the
47 * software, derivative works or modified versions, and any portions
48 * thereof, and that both notices appear in supporting documentation.
50 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
54 * Carnegie Mellon requests users of this software to return to
56 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
57 * School of Computer Science
58 * Carnegie Mellon University
59 * Pittsburgh PA 15213-3890
61 * any improvements or extensions that they make and grant Carnegie the
62 * rights to redistribute these changes.
66 * Resident memory management module.
69 #include <sys/cdefs.h>
70 __FBSDID("$FreeBSD$");
74 #include <sys/param.h>
75 #include <sys/systm.h>
76 #include <sys/counter.h>
77 #include <sys/domainset.h>
78 #include <sys/kernel.h>
79 #include <sys/limits.h>
80 #include <sys/linker.h>
82 #include <sys/malloc.h>
84 #include <sys/msgbuf.h>
85 #include <sys/mutex.h>
87 #include <sys/rwlock.h>
88 #include <sys/sleepqueue.h>
90 #include <sys/sched.h>
92 #include <sys/sysctl.h>
93 #include <sys/vmmeter.h>
94 #include <sys/vnode.h>
98 #include <vm/vm_param.h>
99 #include <vm/vm_domainset.h>
100 #include <vm/vm_kern.h>
101 #include <vm/vm_map.h>
102 #include <vm/vm_object.h>
103 #include <vm/vm_page.h>
104 #include <vm/vm_pageout.h>
105 #include <vm/vm_phys.h>
106 #include <vm/vm_pagequeue.h>
107 #include <vm/vm_pager.h>
108 #include <vm/vm_radix.h>
109 #include <vm/vm_reserv.h>
110 #include <vm/vm_extern.h>
112 #include <vm/uma_int.h>
114 #include <machine/md_var.h>
116 struct vm_domain vm_dom[MAXMEMDOM];
118 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]);
120 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
122 struct mtx_padalign __exclusive_cache_line vm_domainset_lock;
123 /* The following fields are protected by the domainset lock. */
124 domainset_t __exclusive_cache_line vm_min_domains;
125 domainset_t __exclusive_cache_line vm_severe_domains;
126 static int vm_min_waiters;
127 static int vm_severe_waiters;
128 static int vm_pageproc_waiters;
130 static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
131 "VM page statistics");
133 static COUNTER_U64_DEFINE_EARLY(pqstate_commit_retries);
134 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries,
135 CTLFLAG_RD, &pqstate_commit_retries,
136 "Number of failed per-page atomic queue state updates");
138 static COUNTER_U64_DEFINE_EARLY(queue_ops);
139 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops,
140 CTLFLAG_RD, &queue_ops,
141 "Number of batched queue operations");
143 static COUNTER_U64_DEFINE_EARLY(queue_nops);
144 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops,
145 CTLFLAG_RD, &queue_nops,
146 "Number of batched queue operations with no effects");
149 * bogus page -- for I/O to/from partially complete buffers,
150 * or for paging into sparsely invalid regions.
152 vm_page_t bogus_page;
154 vm_page_t vm_page_array;
155 long vm_page_array_size;
158 struct bitset *vm_page_dump;
159 long vm_page_dump_pages;
161 static TAILQ_HEAD(, vm_page) blacklist_head;
162 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
163 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
164 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
166 static uma_zone_t fakepg_zone;
168 static void vm_page_alloc_check(vm_page_t m);
169 static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m,
170 vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked);
171 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
172 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
173 static bool vm_page_free_prep(vm_page_t m);
174 static void vm_page_free_toq(vm_page_t m);
175 static void vm_page_init(void *dummy);
176 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
177 vm_pindex_t pindex, vm_page_t mpred);
178 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
180 static void vm_page_mvqueue(vm_page_t m, const uint8_t queue,
181 const uint16_t nflag);
182 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
183 vm_page_t m_run, vm_paddr_t high);
184 static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse);
185 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
187 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain,
189 static void vm_page_zone_release(void *arg, void **store, int cnt);
191 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
194 vm_page_init(void *dummy)
197 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
198 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
199 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
200 VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
204 * The cache page zone is initialized later since we need to be able to allocate
205 * pages before UMA is fully initialized.
208 vm_page_init_cache_zones(void *dummy __unused)
210 struct vm_domain *vmd;
211 struct vm_pgcache *pgcache;
212 int cache, domain, maxcache, pool;
215 TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &maxcache);
216 maxcache *= mp_ncpus;
217 for (domain = 0; domain < vm_ndomains; domain++) {
218 vmd = VM_DOMAIN(domain);
219 for (pool = 0; pool < VM_NFREEPOOL; pool++) {
220 pgcache = &vmd->vmd_pgcache[pool];
221 pgcache->domain = domain;
222 pgcache->pool = pool;
223 pgcache->zone = uma_zcache_create("vm pgcache",
224 PAGE_SIZE, NULL, NULL, NULL, NULL,
225 vm_page_zone_import, vm_page_zone_release, pgcache,
229 * Limit each pool's zone to 0.1% of the pages in the
232 cache = maxcache != 0 ? maxcache :
233 vmd->vmd_page_count / 1000;
234 uma_zone_set_maxcache(pgcache->zone, cache);
238 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
240 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
241 #if PAGE_SIZE == 32768
243 CTASSERT(sizeof(u_long) >= 8);
250 * Sets the page size, perhaps based upon the memory
251 * size. Must be called before any use of page-size
252 * dependent functions.
255 vm_set_page_size(void)
257 if (vm_cnt.v_page_size == 0)
258 vm_cnt.v_page_size = PAGE_SIZE;
259 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
260 panic("vm_set_page_size: page size not a power of two");
264 * vm_page_blacklist_next:
266 * Find the next entry in the provided string of blacklist
267 * addresses. Entries are separated by space, comma, or newline.
268 * If an invalid integer is encountered then the rest of the
269 * string is skipped. Updates the list pointer to the next
270 * character, or NULL if the string is exhausted or invalid.
273 vm_page_blacklist_next(char **list, char *end)
278 if (list == NULL || *list == NULL)
286 * If there's no end pointer then the buffer is coming from
287 * the kenv and we know it's null-terminated.
290 end = *list + strlen(*list);
292 /* Ensure that strtoq() won't walk off the end */
294 if (*end == '\n' || *end == ' ' || *end == ',')
297 printf("Blacklist not terminated, skipping\n");
303 for (pos = *list; *pos != '\0'; pos = cp) {
304 bad = strtoq(pos, &cp, 0);
305 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
314 if (*cp == '\0' || ++cp >= end)
318 return (trunc_page(bad));
320 printf("Garbage in RAM blacklist, skipping\n");
326 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
328 struct vm_domain *vmd;
332 m = vm_phys_paddr_to_vm_page(pa);
334 return (true); /* page does not exist, no failure */
336 vmd = vm_pagequeue_domain(m);
337 vm_domain_free_lock(vmd);
338 ret = vm_phys_unfree_page(m);
339 vm_domain_free_unlock(vmd);
341 vm_domain_freecnt_inc(vmd, -1);
342 TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
344 printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
350 * vm_page_blacklist_check:
352 * Iterate through the provided string of blacklist addresses, pulling
353 * each entry out of the physical allocator free list and putting it
354 * onto a list for reporting via the vm.page_blacklist sysctl.
357 vm_page_blacklist_check(char *list, char *end)
363 while (next != NULL) {
364 if ((pa = vm_page_blacklist_next(&next, end)) == 0)
366 vm_page_blacklist_add(pa, bootverbose);
371 * vm_page_blacklist_load:
373 * Search for a special module named "ram_blacklist". It'll be a
374 * plain text file provided by the user via the loader directive
378 vm_page_blacklist_load(char **list, char **end)
387 mod = preload_search_by_type("ram_blacklist");
389 ptr = preload_fetch_addr(mod);
390 len = preload_fetch_size(mod);
401 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
408 error = sysctl_wire_old_buffer(req, 0);
411 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
412 TAILQ_FOREACH(m, &blacklist_head, listq) {
413 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
414 (uintmax_t)m->phys_addr);
417 error = sbuf_finish(&sbuf);
423 * Initialize a dummy page for use in scans of the specified paging queue.
424 * In principle, this function only needs to set the flag PG_MARKER.
425 * Nonetheless, it write busies the page as a safety precaution.
428 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags)
431 bzero(marker, sizeof(*marker));
432 marker->flags = PG_MARKER;
433 marker->a.flags = aflags;
434 marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
435 marker->a.queue = queue;
439 vm_page_domain_init(int domain)
441 struct vm_domain *vmd;
442 struct vm_pagequeue *pq;
445 vmd = VM_DOMAIN(domain);
446 bzero(vmd, sizeof(*vmd));
447 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
448 "vm inactive pagequeue";
449 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
450 "vm active pagequeue";
451 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
452 "vm laundry pagequeue";
453 *__DECONST(const char **,
454 &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
455 "vm unswappable pagequeue";
456 vmd->vmd_domain = domain;
457 vmd->vmd_page_count = 0;
458 vmd->vmd_free_count = 0;
460 vmd->vmd_oom = FALSE;
461 for (i = 0; i < PQ_COUNT; i++) {
462 pq = &vmd->vmd_pagequeues[i];
463 TAILQ_INIT(&pq->pq_pl);
464 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
465 MTX_DEF | MTX_DUPOK);
467 vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
469 mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
470 mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
471 snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
474 * inacthead is used to provide FIFO ordering for LRU-bypassing
477 vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
478 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
479 &vmd->vmd_inacthead, plinks.q);
482 * The clock pages are used to implement active queue scanning without
483 * requeues. Scans start at clock[0], which is advanced after the scan
484 * ends. When the two clock hands meet, they are reset and scanning
485 * resumes from the head of the queue.
487 vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
488 vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
489 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
490 &vmd->vmd_clock[0], plinks.q);
491 TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
492 &vmd->vmd_clock[1], plinks.q);
496 * Initialize a physical page in preparation for adding it to the free
500 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
505 m->busy_lock = VPB_FREED;
506 m->flags = m->a.flags = 0;
508 m->a.queue = PQ_NONE;
511 m->order = VM_NFREEORDER;
512 m->pool = VM_FREEPOOL_DEFAULT;
513 m->valid = m->dirty = 0;
517 #ifndef PMAP_HAS_PAGE_ARRAY
519 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range)
524 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
525 * However, because this page is allocated from KVM, out-of-bounds
526 * accesses using the direct map will not be trapped.
531 * Allocate physical memory for the page structures, and map it.
533 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
534 vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end,
535 VM_PROT_READ | VM_PROT_WRITE);
536 vm_page_array_size = page_range;
545 * Initializes the resident memory module. Allocates physical memory for
546 * bootstrapping UMA and some data structures that are used to manage
547 * physical pages. Initializes these structures, and populates the free
551 vm_page_startup(vm_offset_t vaddr)
553 struct vm_phys_seg *seg;
555 char *list, *listend;
556 vm_paddr_t end, high_avail, low_avail, new_end, size;
557 vm_paddr_t page_range __unused;
558 vm_paddr_t last_pa, pa;
560 #if MINIDUMP_PAGE_TRACKING
561 u_long vm_page_dump_size;
563 int biggestone, i, segind;
568 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
572 vaddr = round_page(vaddr);
574 vm_phys_early_startup();
575 biggestone = vm_phys_avail_largest();
576 end = phys_avail[biggestone+1];
579 * Initialize the page and queue locks.
581 mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
582 for (i = 0; i < PA_LOCK_COUNT; i++)
583 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
584 for (i = 0; i < vm_ndomains; i++)
585 vm_page_domain_init(i);
589 witness_size = round_page(witness_startup_count());
590 new_end -= witness_size;
591 mapped = pmap_map(&vaddr, new_end, new_end + witness_size,
592 VM_PROT_READ | VM_PROT_WRITE);
593 bzero((void *)mapped, witness_size);
594 witness_startup((void *)mapped);
597 #if MINIDUMP_PAGE_TRACKING
599 * Allocate a bitmap to indicate that a random physical page
600 * needs to be included in a minidump.
602 * The amd64 port needs this to indicate which direct map pages
603 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
605 * However, i386 still needs this workspace internally within the
606 * minidump code. In theory, they are not needed on i386, but are
607 * included should the sf_buf code decide to use them.
610 for (i = 0; dump_avail[i + 1] != 0; i += 2)
611 if (dump_avail[i + 1] > last_pa)
612 last_pa = dump_avail[i + 1];
613 vm_page_dump_pages = last_pa / PAGE_SIZE;
614 vm_page_dump_size = round_page(BITSET_SIZE(vm_page_dump_pages));
615 new_end -= vm_page_dump_size;
616 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
617 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
618 bzero((void *)vm_page_dump, vm_page_dump_size);
622 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
623 defined(__riscv) || defined(__powerpc64__)
625 * Include the UMA bootstrap pages, witness pages and vm_page_dump
626 * in a crash dump. When pmap_map() uses the direct map, they are
627 * not automatically included.
629 for (pa = new_end; pa < end; pa += PAGE_SIZE)
632 phys_avail[biggestone + 1] = new_end;
635 * Request that the physical pages underlying the message buffer be
636 * included in a crash dump. Since the message buffer is accessed
637 * through the direct map, they are not automatically included.
639 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
640 last_pa = pa + round_page(msgbufsize);
641 while (pa < last_pa) {
647 * Compute the number of pages of memory that will be available for
648 * use, taking into account the overhead of a page structure per page.
649 * In other words, solve
650 * "available physical memory" - round_page(page_range *
651 * sizeof(struct vm_page)) = page_range * PAGE_SIZE
654 low_avail = phys_avail[0];
655 high_avail = phys_avail[1];
656 for (i = 0; i < vm_phys_nsegs; i++) {
657 if (vm_phys_segs[i].start < low_avail)
658 low_avail = vm_phys_segs[i].start;
659 if (vm_phys_segs[i].end > high_avail)
660 high_avail = vm_phys_segs[i].end;
662 /* Skip the first chunk. It is already accounted for. */
663 for (i = 2; phys_avail[i + 1] != 0; i += 2) {
664 if (phys_avail[i] < low_avail)
665 low_avail = phys_avail[i];
666 if (phys_avail[i + 1] > high_avail)
667 high_avail = phys_avail[i + 1];
669 first_page = low_avail / PAGE_SIZE;
670 #ifdef VM_PHYSSEG_SPARSE
672 for (i = 0; i < vm_phys_nsegs; i++)
673 size += vm_phys_segs[i].end - vm_phys_segs[i].start;
674 for (i = 0; phys_avail[i + 1] != 0; i += 2)
675 size += phys_avail[i + 1] - phys_avail[i];
676 #elif defined(VM_PHYSSEG_DENSE)
677 size = high_avail - low_avail;
679 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
682 #ifdef PMAP_HAS_PAGE_ARRAY
683 pmap_page_array_startup(size / PAGE_SIZE);
684 biggestone = vm_phys_avail_largest();
685 end = new_end = phys_avail[biggestone + 1];
687 #ifdef VM_PHYSSEG_DENSE
689 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
690 * the overhead of a page structure per page only if vm_page_array is
691 * allocated from the last physical memory chunk. Otherwise, we must
692 * allocate page structures representing the physical memory
693 * underlying vm_page_array, even though they will not be used.
695 if (new_end != high_avail)
696 page_range = size / PAGE_SIZE;
700 page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
703 * If the partial bytes remaining are large enough for
704 * a page (PAGE_SIZE) without a corresponding
705 * 'struct vm_page', then new_end will contain an
706 * extra page after subtracting the length of the VM
707 * page array. Compensate by subtracting an extra
710 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
711 if (new_end == high_avail)
712 high_avail -= PAGE_SIZE;
713 new_end -= PAGE_SIZE;
717 new_end = vm_page_array_alloc(&vaddr, end, page_range);
720 #if VM_NRESERVLEVEL > 0
722 * Allocate physical memory for the reservation management system's
723 * data structures, and map it.
725 new_end = vm_reserv_startup(&vaddr, new_end);
727 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
728 defined(__riscv) || defined(__powerpc64__)
730 * Include vm_page_array and vm_reserv_array in a crash dump.
732 for (pa = new_end; pa < end; pa += PAGE_SIZE)
735 phys_avail[biggestone + 1] = new_end;
738 * Add physical memory segments corresponding to the available
741 for (i = 0; phys_avail[i + 1] != 0; i += 2)
742 if (vm_phys_avail_size(i) != 0)
743 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
746 * Initialize the physical memory allocator.
751 * Initialize the page structures and add every available page to the
752 * physical memory allocator's free lists.
754 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
755 for (ii = 0; ii < vm_page_array_size; ii++) {
756 m = &vm_page_array[ii];
757 vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0);
758 m->flags = PG_FICTITIOUS;
761 vm_cnt.v_page_count = 0;
762 for (segind = 0; segind < vm_phys_nsegs; segind++) {
763 seg = &vm_phys_segs[segind];
764 for (m = seg->first_page, pa = seg->start; pa < seg->end;
765 m++, pa += PAGE_SIZE)
766 vm_page_init_page(m, pa, segind);
769 * Add the segment to the free lists only if it is covered by
770 * one of the ranges in phys_avail. Because we've added the
771 * ranges to the vm_phys_segs array, we can assume that each
772 * segment is either entirely contained in one of the ranges,
773 * or doesn't overlap any of them.
775 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
776 struct vm_domain *vmd;
778 if (seg->start < phys_avail[i] ||
779 seg->end > phys_avail[i + 1])
783 pagecount = (u_long)atop(seg->end - seg->start);
785 vmd = VM_DOMAIN(seg->domain);
786 vm_domain_free_lock(vmd);
787 vm_phys_enqueue_contig(m, pagecount);
788 vm_domain_free_unlock(vmd);
789 vm_domain_freecnt_inc(vmd, pagecount);
790 vm_cnt.v_page_count += (u_int)pagecount;
792 vmd = VM_DOMAIN(seg->domain);
793 vmd->vmd_page_count += (u_int)pagecount;
794 vmd->vmd_segs |= 1UL << m->segind;
800 * Remove blacklisted pages from the physical memory allocator.
802 TAILQ_INIT(&blacklist_head);
803 vm_page_blacklist_load(&list, &listend);
804 vm_page_blacklist_check(list, listend);
806 list = kern_getenv("vm.blacklist");
807 vm_page_blacklist_check(list, NULL);
810 #if VM_NRESERVLEVEL > 0
812 * Initialize the reservation management system.
821 vm_page_reference(vm_page_t m)
824 vm_page_aflag_set(m, PGA_REFERENCED);
830 * Helper routine for grab functions to trylock busy.
832 * Returns true on success and false on failure.
835 vm_page_trybusy(vm_page_t m, int allocflags)
838 if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0)
839 return (vm_page_trysbusy(m));
841 return (vm_page_tryxbusy(m));
847 * Helper routine for grab functions to trylock busy and wire.
849 * Returns true on success and false on failure.
852 vm_page_tryacquire(vm_page_t m, int allocflags)
856 locked = vm_page_trybusy(m, allocflags);
857 if (locked && (allocflags & VM_ALLOC_WIRED) != 0)
863 * vm_page_busy_acquire:
865 * Acquire the busy lock as described by VM_ALLOC_* flags. Will loop
866 * and drop the object lock if necessary.
869 vm_page_busy_acquire(vm_page_t m, int allocflags)
875 * The page-specific object must be cached because page
876 * identity can change during the sleep, causing the
877 * re-lock of a different object.
878 * It is assumed that a reference to the object is already
879 * held by the callers.
883 if (vm_page_tryacquire(m, allocflags))
885 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
888 locked = VM_OBJECT_WOWNED(obj);
891 MPASS(locked || vm_page_wired(m));
892 if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags,
894 VM_OBJECT_WLOCK(obj);
895 if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
897 KASSERT(m->object == obj || m->object == NULL,
898 ("vm_page_busy_acquire: page %p does not belong to %p",
904 * vm_page_busy_downgrade:
906 * Downgrade an exclusive busy page into a single shared busy page.
909 vm_page_busy_downgrade(vm_page_t m)
913 vm_page_assert_xbusied(m);
915 x = vm_page_busy_fetch(m);
917 if (atomic_fcmpset_rel_int(&m->busy_lock,
918 &x, VPB_SHARERS_WORD(1)))
921 if ((x & VPB_BIT_WAITERS) != 0)
927 * vm_page_busy_tryupgrade:
929 * Attempt to upgrade a single shared busy into an exclusive busy.
932 vm_page_busy_tryupgrade(vm_page_t m)
936 vm_page_assert_sbusied(m);
938 x = vm_page_busy_fetch(m);
939 ce = VPB_CURTHREAD_EXCLUSIVE;
941 if (VPB_SHARERS(x) > 1)
943 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
944 ("vm_page_busy_tryupgrade: invalid lock state"));
945 if (!atomic_fcmpset_acq_int(&m->busy_lock, &x,
946 ce | (x & VPB_BIT_WAITERS)))
955 * Return a positive value if the page is shared busied, 0 otherwise.
958 vm_page_sbusied(vm_page_t m)
962 x = vm_page_busy_fetch(m);
963 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
969 * Shared unbusy a page.
972 vm_page_sunbusy(vm_page_t m)
976 vm_page_assert_sbusied(m);
978 x = vm_page_busy_fetch(m);
980 KASSERT(x != VPB_FREED,
981 ("vm_page_sunbusy: Unlocking freed page."));
982 if (VPB_SHARERS(x) > 1) {
983 if (atomic_fcmpset_int(&m->busy_lock, &x,
988 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
989 ("vm_page_sunbusy: invalid lock state"));
990 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
992 if ((x & VPB_BIT_WAITERS) == 0)
1000 * vm_page_busy_sleep:
1002 * Sleep if the page is busy, using the page pointer as wchan.
1003 * This is used to implement the hard-path of busying mechanism.
1005 * If nonshared is true, sleep only if the page is xbusy.
1007 * The object lock must be held on entry and will be released on exit.
1010 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
1015 VM_OBJECT_ASSERT_LOCKED(obj);
1016 vm_page_lock_assert(m, MA_NOTOWNED);
1018 if (!_vm_page_busy_sleep(obj, m, m->pindex, wmesg,
1019 nonshared ? VM_ALLOC_SBUSY : 0 , true))
1020 VM_OBJECT_DROP(obj);
1024 * vm_page_busy_sleep_unlocked:
1026 * Sleep if the page is busy, using the page pointer as wchan.
1027 * This is used to implement the hard-path of busying mechanism.
1029 * If nonshared is true, sleep only if the page is xbusy.
1031 * The object lock must not be held on entry. The operation will
1032 * return if the page changes identity.
1035 vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1036 const char *wmesg, bool nonshared)
1039 VM_OBJECT_ASSERT_UNLOCKED(obj);
1040 vm_page_lock_assert(m, MA_NOTOWNED);
1042 _vm_page_busy_sleep(obj, m, pindex, wmesg,
1043 nonshared ? VM_ALLOC_SBUSY : 0, false);
1047 * _vm_page_busy_sleep:
1049 * Internal busy sleep function. Verifies the page identity and
1050 * lockstate against parameters. Returns true if it sleeps and
1053 * If locked is true the lock will be dropped for any true returns
1054 * and held for any false returns.
1057 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1058 const char *wmesg, int allocflags, bool locked)
1064 * If the object is busy we must wait for that to drain to zero
1065 * before trying the page again.
1067 if (obj != NULL && vm_object_busied(obj)) {
1069 VM_OBJECT_DROP(obj);
1070 vm_object_busy_wait(obj, wmesg);
1074 if (!vm_page_busied(m))
1077 xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0;
1079 x = vm_page_busy_fetch(m);
1082 * If the page changes objects or becomes unlocked we can
1085 if (x == VPB_UNBUSIED ||
1086 (xsleep && (x & VPB_BIT_SHARED) != 0) ||
1087 m->object != obj || m->pindex != pindex) {
1091 if ((x & VPB_BIT_WAITERS) != 0)
1093 } while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS));
1095 VM_OBJECT_DROP(obj);
1097 sleepq_add(m, NULL, wmesg, 0, 0);
1098 sleepq_wait(m, PVM);
1106 * Try to shared busy a page.
1107 * If the operation succeeds 1 is returned otherwise 0.
1108 * The operation never sleeps.
1111 vm_page_trysbusy(vm_page_t m)
1117 x = vm_page_busy_fetch(m);
1119 if ((x & VPB_BIT_SHARED) == 0)
1122 * Reduce the window for transient busies that will trigger
1123 * false negatives in vm_page_ps_test().
1125 if (obj != NULL && vm_object_busied(obj))
1127 if (atomic_fcmpset_acq_int(&m->busy_lock, &x,
1128 x + VPB_ONE_SHARER))
1132 /* Refetch the object now that we're guaranteed that it is stable. */
1134 if (obj != NULL && vm_object_busied(obj)) {
1144 * Try to exclusive busy a page.
1145 * If the operation succeeds 1 is returned otherwise 0.
1146 * The operation never sleeps.
1149 vm_page_tryxbusy(vm_page_t m)
1153 if (atomic_cmpset_acq_int(&m->busy_lock, VPB_UNBUSIED,
1154 VPB_CURTHREAD_EXCLUSIVE) == 0)
1158 if (obj != NULL && vm_object_busied(obj)) {
1166 vm_page_xunbusy_hard_tail(vm_page_t m)
1168 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1169 /* Wake the waiter. */
1174 * vm_page_xunbusy_hard:
1176 * Called when unbusy has failed because there is a waiter.
1179 vm_page_xunbusy_hard(vm_page_t m)
1181 vm_page_assert_xbusied(m);
1182 vm_page_xunbusy_hard_tail(m);
1186 vm_page_xunbusy_hard_unchecked(vm_page_t m)
1188 vm_page_assert_xbusied_unchecked(m);
1189 vm_page_xunbusy_hard_tail(m);
1193 vm_page_busy_free(vm_page_t m)
1197 atomic_thread_fence_rel();
1198 x = atomic_swap_int(&m->busy_lock, VPB_FREED);
1199 if ((x & VPB_BIT_WAITERS) != 0)
1204 * vm_page_unhold_pages:
1206 * Unhold each of the pages that is referenced by the given array.
1209 vm_page_unhold_pages(vm_page_t *ma, int count)
1212 for (; count != 0; count--) {
1213 vm_page_unwire(*ma, PQ_ACTIVE);
1219 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1223 #ifdef VM_PHYSSEG_SPARSE
1224 m = vm_phys_paddr_to_vm_page(pa);
1226 m = vm_phys_fictitious_to_vm_page(pa);
1228 #elif defined(VM_PHYSSEG_DENSE)
1232 if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1233 m = &vm_page_array[pi - first_page];
1236 return (vm_phys_fictitious_to_vm_page(pa));
1238 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1245 * Create a fictitious page with the specified physical address and
1246 * memory attribute. The memory attribute is the only the machine-
1247 * dependent aspect of a fictitious page that must be initialized.
1250 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1254 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1255 vm_page_initfake(m, paddr, memattr);
1260 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1263 if ((m->flags & PG_FICTITIOUS) != 0) {
1265 * The page's memattr might have changed since the
1266 * previous initialization. Update the pmap to the
1271 m->phys_addr = paddr;
1272 m->a.queue = PQ_NONE;
1273 /* Fictitious pages don't use "segind". */
1274 m->flags = PG_FICTITIOUS;
1275 /* Fictitious pages don't use "order" or "pool". */
1276 m->oflags = VPO_UNMANAGED;
1277 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
1278 /* Fictitious pages are unevictable. */
1282 pmap_page_set_memattr(m, memattr);
1288 * Release a fictitious page.
1291 vm_page_putfake(vm_page_t m)
1294 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1295 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1296 ("vm_page_putfake: bad page %p", m));
1297 vm_page_assert_xbusied(m);
1298 vm_page_busy_free(m);
1299 uma_zfree(fakepg_zone, m);
1303 * vm_page_updatefake:
1305 * Update the given fictitious page to the specified physical address and
1309 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1312 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1313 ("vm_page_updatefake: bad page %p", m));
1314 m->phys_addr = paddr;
1315 pmap_page_set_memattr(m, memattr);
1324 vm_page_free(vm_page_t m)
1327 m->flags &= ~PG_ZERO;
1328 vm_page_free_toq(m);
1332 * vm_page_free_zero:
1334 * Free a page to the zerod-pages queue
1337 vm_page_free_zero(vm_page_t m)
1340 m->flags |= PG_ZERO;
1341 vm_page_free_toq(m);
1345 * Unbusy and handle the page queueing for a page from a getpages request that
1346 * was optionally read ahead or behind.
1349 vm_page_readahead_finish(vm_page_t m)
1352 /* We shouldn't put invalid pages on queues. */
1353 KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m));
1356 * Since the page is not the actually needed one, whether it should
1357 * be activated or deactivated is not obvious. Empirical results
1358 * have shown that deactivating the page is usually the best choice,
1359 * unless the page is wanted by another thread.
1361 if ((vm_page_busy_fetch(m) & VPB_BIT_WAITERS) != 0)
1362 vm_page_activate(m);
1364 vm_page_deactivate(m);
1365 vm_page_xunbusy_unchecked(m);
1369 * Destroy the identity of an invalid page and free it if possible.
1370 * This is intended to be used when reading a page from backing store fails.
1373 vm_page_free_invalid(vm_page_t m)
1376 KASSERT(vm_page_none_valid(m), ("page %p is valid", m));
1377 KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m));
1378 KASSERT(m->object != NULL, ("page %p has no object", m));
1379 VM_OBJECT_ASSERT_WLOCKED(m->object);
1382 * We may be attempting to free the page as part of the handling for an
1383 * I/O error, in which case the page was xbusied by a different thread.
1385 vm_page_xbusy_claim(m);
1388 * If someone has wired this page while the object lock
1389 * was not held, then the thread that unwires is responsible
1390 * for freeing the page. Otherwise just free the page now.
1391 * The wire count of this unmapped page cannot change while
1392 * we have the page xbusy and the page's object wlocked.
1394 if (vm_page_remove(m))
1399 * vm_page_sleep_if_busy:
1401 * Sleep and release the object lock if the page is busied.
1402 * Returns TRUE if the thread slept.
1404 * The given page must be unlocked and object containing it must
1408 vm_page_sleep_if_busy(vm_page_t m, const char *wmesg)
1412 vm_page_lock_assert(m, MA_NOTOWNED);
1413 VM_OBJECT_ASSERT_WLOCKED(m->object);
1416 * The page-specific object must be cached because page
1417 * identity can change during the sleep, causing the
1418 * re-lock of a different object.
1419 * It is assumed that a reference to the object is already
1420 * held by the callers.
1423 if (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, 0, true)) {
1424 VM_OBJECT_WLOCK(obj);
1431 * vm_page_sleep_if_xbusy:
1433 * Sleep and release the object lock if the page is xbusied.
1434 * Returns TRUE if the thread slept.
1436 * The given page must be unlocked and object containing it must
1440 vm_page_sleep_if_xbusy(vm_page_t m, const char *wmesg)
1444 vm_page_lock_assert(m, MA_NOTOWNED);
1445 VM_OBJECT_ASSERT_WLOCKED(m->object);
1448 * The page-specific object must be cached because page
1449 * identity can change during the sleep, causing the
1450 * re-lock of a different object.
1451 * It is assumed that a reference to the object is already
1452 * held by the callers.
1455 if (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, VM_ALLOC_SBUSY,
1457 VM_OBJECT_WLOCK(obj);
1464 * vm_page_dirty_KBI: [ internal use only ]
1466 * Set all bits in the page's dirty field.
1468 * The object containing the specified page must be locked if the
1469 * call is made from the machine-independent layer.
1471 * See vm_page_clear_dirty_mask().
1473 * This function should only be called by vm_page_dirty().
1476 vm_page_dirty_KBI(vm_page_t m)
1479 /* Refer to this operation by its public name. */
1480 KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!"));
1481 m->dirty = VM_PAGE_BITS_ALL;
1485 * vm_page_insert: [ internal use only ]
1487 * Inserts the given mem entry into the object and object list.
1489 * The object must be locked.
1492 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1496 VM_OBJECT_ASSERT_WLOCKED(object);
1497 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1498 return (vm_page_insert_after(m, object, pindex, mpred));
1502 * vm_page_insert_after:
1504 * Inserts the page "m" into the specified object at offset "pindex".
1506 * The page "mpred" must immediately precede the offset "pindex" within
1507 * the specified object.
1509 * The object must be locked.
1512 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1517 VM_OBJECT_ASSERT_WLOCKED(object);
1518 KASSERT(m->object == NULL,
1519 ("vm_page_insert_after: page already inserted"));
1520 if (mpred != NULL) {
1521 KASSERT(mpred->object == object,
1522 ("vm_page_insert_after: object doesn't contain mpred"));
1523 KASSERT(mpred->pindex < pindex,
1524 ("vm_page_insert_after: mpred doesn't precede pindex"));
1525 msucc = TAILQ_NEXT(mpred, listq);
1527 msucc = TAILQ_FIRST(&object->memq);
1529 KASSERT(msucc->pindex > pindex,
1530 ("vm_page_insert_after: msucc doesn't succeed pindex"));
1533 * Record the object/offset pair in this page.
1537 m->ref_count |= VPRC_OBJREF;
1540 * Now link into the object's ordered list of backed pages.
1542 if (vm_radix_insert(&object->rtree, m)) {
1545 m->ref_count &= ~VPRC_OBJREF;
1548 vm_page_insert_radixdone(m, object, mpred);
1553 * vm_page_insert_radixdone:
1555 * Complete page "m" insertion into the specified object after the
1556 * radix trie hooking.
1558 * The page "mpred" must precede the offset "m->pindex" within the
1561 * The object must be locked.
1564 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1567 VM_OBJECT_ASSERT_WLOCKED(object);
1568 KASSERT(object != NULL && m->object == object,
1569 ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1570 KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1571 ("vm_page_insert_radixdone: page %p is missing object ref", m));
1572 if (mpred != NULL) {
1573 KASSERT(mpred->object == object,
1574 ("vm_page_insert_radixdone: object doesn't contain mpred"));
1575 KASSERT(mpred->pindex < m->pindex,
1576 ("vm_page_insert_radixdone: mpred doesn't precede pindex"));
1580 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1582 TAILQ_INSERT_HEAD(&object->memq, m, listq);
1585 * Show that the object has one more resident page.
1587 object->resident_page_count++;
1590 * Hold the vnode until the last page is released.
1592 if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1593 vhold(object->handle);
1596 * Since we are inserting a new and possibly dirty page,
1597 * update the object's generation count.
1599 if (pmap_page_is_write_mapped(m))
1600 vm_object_set_writeable_dirty(object);
1604 * Do the work to remove a page from its object. The caller is responsible for
1605 * updating the page's fields to reflect this removal.
1608 vm_page_object_remove(vm_page_t m)
1613 vm_page_assert_xbusied(m);
1615 VM_OBJECT_ASSERT_WLOCKED(object);
1616 KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1617 ("page %p is missing its object ref", m));
1619 /* Deferred free of swap space. */
1620 if ((m->a.flags & PGA_SWAP_FREE) != 0)
1621 vm_pager_page_unswapped(m);
1624 mrem = vm_radix_remove(&object->rtree, m->pindex);
1625 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1628 * Now remove from the object's list of backed pages.
1630 TAILQ_REMOVE(&object->memq, m, listq);
1633 * And show that the object has one fewer resident page.
1635 object->resident_page_count--;
1638 * The vnode may now be recycled.
1640 if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1641 vdrop(object->handle);
1647 * Removes the specified page from its containing object, but does not
1648 * invalidate any backing storage. Returns true if the object's reference
1649 * was the last reference to the page, and false otherwise.
1651 * The object must be locked and the page must be exclusively busied.
1652 * The exclusive busy will be released on return. If this is not the
1653 * final ref and the caller does not hold a wire reference it may not
1654 * continue to access the page.
1657 vm_page_remove(vm_page_t m)
1661 dropped = vm_page_remove_xbusy(m);
1668 * vm_page_remove_xbusy
1670 * Removes the page but leaves the xbusy held. Returns true if this
1671 * removed the final ref and false otherwise.
1674 vm_page_remove_xbusy(vm_page_t m)
1677 vm_page_object_remove(m);
1678 return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1684 * Returns the page associated with the object/offset
1685 * pair specified; if none is found, NULL is returned.
1687 * The object must be locked.
1690 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1693 VM_OBJECT_ASSERT_LOCKED(object);
1694 return (vm_radix_lookup(&object->rtree, pindex));
1700 * Returns a page that must already have been busied by
1701 * the caller. Used for bogus page replacement.
1704 vm_page_relookup(vm_object_t object, vm_pindex_t pindex)
1708 m = vm_radix_lookup_unlocked(&object->rtree, pindex);
1709 KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) &&
1710 m->object == object && m->pindex == pindex,
1711 ("vm_page_relookup: Invalid page %p", m));
1716 * This should only be used by lockless functions for releasing transient
1717 * incorrect acquires. The page may have been freed after we acquired a
1718 * busy lock. In this case busy_lock == VPB_FREED and we have nothing
1722 vm_page_busy_release(vm_page_t m)
1726 x = vm_page_busy_fetch(m);
1730 if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) {
1731 if (atomic_fcmpset_int(&m->busy_lock, &x,
1732 x - VPB_ONE_SHARER))
1736 KASSERT((x & VPB_BIT_SHARED) != 0 ||
1737 (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE,
1738 ("vm_page_busy_release: %p xbusy not owned.", m));
1739 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1741 if ((x & VPB_BIT_WAITERS) != 0)
1748 * vm_page_find_least:
1750 * Returns the page associated with the object with least pindex
1751 * greater than or equal to the parameter pindex, or NULL.
1753 * The object must be locked.
1756 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1760 VM_OBJECT_ASSERT_LOCKED(object);
1761 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1762 m = vm_radix_lookup_ge(&object->rtree, pindex);
1767 * Returns the given page's successor (by pindex) within the object if it is
1768 * resident; if none is found, NULL is returned.
1770 * The object must be locked.
1773 vm_page_next(vm_page_t m)
1777 VM_OBJECT_ASSERT_LOCKED(m->object);
1778 if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1779 MPASS(next->object == m->object);
1780 if (next->pindex != m->pindex + 1)
1787 * Returns the given page's predecessor (by pindex) within the object if it is
1788 * resident; if none is found, NULL is returned.
1790 * The object must be locked.
1793 vm_page_prev(vm_page_t m)
1797 VM_OBJECT_ASSERT_LOCKED(m->object);
1798 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1799 MPASS(prev->object == m->object);
1800 if (prev->pindex != m->pindex - 1)
1807 * Uses the page mnew as a replacement for an existing page at index
1808 * pindex which must be already present in the object.
1810 * Both pages must be exclusively busied on enter. The old page is
1813 * A return value of true means mold is now free. If this is not the
1814 * final ref and the caller does not hold a wire reference it may not
1815 * continue to access the page.
1818 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1824 VM_OBJECT_ASSERT_WLOCKED(object);
1825 vm_page_assert_xbusied(mold);
1826 KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
1827 ("vm_page_replace: page %p already in object", mnew));
1830 * This function mostly follows vm_page_insert() and
1831 * vm_page_remove() without the radix, object count and vnode
1832 * dance. Double check such functions for more comments.
1835 mnew->object = object;
1836 mnew->pindex = pindex;
1837 atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
1838 mret = vm_radix_replace(&object->rtree, mnew);
1839 KASSERT(mret == mold,
1840 ("invalid page replacement, mold=%p, mret=%p", mold, mret));
1841 KASSERT((mold->oflags & VPO_UNMANAGED) ==
1842 (mnew->oflags & VPO_UNMANAGED),
1843 ("vm_page_replace: mismatched VPO_UNMANAGED"));
1845 /* Keep the resident page list in sorted order. */
1846 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1847 TAILQ_REMOVE(&object->memq, mold, listq);
1848 mold->object = NULL;
1851 * The object's resident_page_count does not change because we have
1852 * swapped one page for another, but the generation count should
1853 * change if the page is dirty.
1855 if (pmap_page_is_write_mapped(mnew))
1856 vm_object_set_writeable_dirty(object);
1857 dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF;
1858 vm_page_xunbusy(mold);
1864 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1868 vm_page_assert_xbusied(mnew);
1870 if (vm_page_replace_hold(mnew, object, pindex, mold))
1877 * Move the given memory entry from its
1878 * current object to the specified target object/offset.
1880 * Note: swap associated with the page must be invalidated by the move. We
1881 * have to do this for several reasons: (1) we aren't freeing the
1882 * page, (2) we are dirtying the page, (3) the VM system is probably
1883 * moving the page from object A to B, and will then later move
1884 * the backing store from A to B and we can't have a conflict.
1886 * Note: we *always* dirty the page. It is necessary both for the
1887 * fact that we moved it, and because we may be invalidating
1890 * The objects must be locked.
1893 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1898 VM_OBJECT_ASSERT_WLOCKED(new_object);
1900 KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m));
1901 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1902 KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1903 ("vm_page_rename: pindex already renamed"));
1906 * Create a custom version of vm_page_insert() which does not depend
1907 * by m_prev and can cheat on the implementation aspects of the
1911 m->pindex = new_pindex;
1912 if (vm_radix_insert(&new_object->rtree, m)) {
1918 * The operation cannot fail anymore. The removal must happen before
1919 * the listq iterator is tainted.
1922 vm_page_object_remove(m);
1924 /* Return back to the new pindex to complete vm_page_insert(). */
1925 m->pindex = new_pindex;
1926 m->object = new_object;
1928 vm_page_insert_radixdone(m, new_object, mpred);
1936 * Allocate and return a page that is associated with the specified
1937 * object and offset pair. By default, this page is exclusive busied.
1939 * The caller must always specify an allocation class.
1941 * allocation classes:
1942 * VM_ALLOC_NORMAL normal process request
1943 * VM_ALLOC_SYSTEM system *really* needs a page
1944 * VM_ALLOC_INTERRUPT interrupt time request
1946 * optional allocation flags:
1947 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1948 * intends to allocate
1949 * VM_ALLOC_NOBUSY do not exclusive busy the page
1950 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1951 * VM_ALLOC_NOOBJ page is not associated with an object and
1952 * should not be exclusive busy
1953 * VM_ALLOC_SBUSY shared busy the allocated page
1954 * VM_ALLOC_WIRED wire the allocated page
1955 * VM_ALLOC_ZERO prefer a zeroed page
1958 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1961 return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1962 vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1966 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1970 return (vm_page_alloc_domain_after(object, pindex, domain, req,
1971 object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) :
1976 * Allocate a page in the specified object with the given page index. To
1977 * optimize insertion of the page into the object, the caller must also specifiy
1978 * the resident page in the object with largest index smaller than the given
1979 * page index, or NULL if no such page exists.
1982 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1983 int req, vm_page_t mpred)
1985 struct vm_domainset_iter di;
1989 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1991 m = vm_page_alloc_domain_after(object, pindex, domain, req,
1995 } while (vm_domainset_iter_page(&di, object, &domain) == 0);
2001 * Returns true if the number of free pages exceeds the minimum
2002 * for the request class and false otherwise.
2005 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
2007 u_int limit, old, new;
2009 if (req_class == VM_ALLOC_INTERRUPT)
2011 else if (req_class == VM_ALLOC_SYSTEM)
2012 limit = vmd->vmd_interrupt_free_min;
2014 limit = vmd->vmd_free_reserved;
2017 * Attempt to reserve the pages. Fail if we're below the limit.
2020 old = vmd->vmd_free_count;
2025 } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
2027 /* Wake the page daemon if we've crossed the threshold. */
2028 if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
2029 pagedaemon_wakeup(vmd->vmd_domain);
2031 /* Only update bitsets on transitions. */
2032 if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
2033 (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
2040 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
2045 * The page daemon is allowed to dig deeper into the free page list.
2047 req_class = req & VM_ALLOC_CLASS_MASK;
2048 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2049 req_class = VM_ALLOC_SYSTEM;
2050 return (_vm_domain_allocate(vmd, req_class, npages));
2054 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
2055 int req, vm_page_t mpred)
2057 struct vm_domain *vmd;
2061 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
2062 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
2063 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2064 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2065 ("inconsistent object(%p)/req(%x)", object, req));
2066 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2067 ("Can't sleep and retry object insertion."));
2068 KASSERT(mpred == NULL || mpred->pindex < pindex,
2069 ("mpred %p doesn't precede pindex 0x%jx", mpred,
2070 (uintmax_t)pindex));
2072 VM_OBJECT_ASSERT_WLOCKED(object);
2076 pool = object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT;
2078 #if VM_NRESERVLEVEL > 0
2080 * Can we allocate the page from a reservation?
2082 if (vm_object_reserv(object) &&
2083 (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
2088 vmd = VM_DOMAIN(domain);
2089 if (vmd->vmd_pgcache[pool].zone != NULL) {
2090 m = uma_zalloc(vmd->vmd_pgcache[pool].zone, M_NOWAIT | M_NOVM);
2092 flags |= PG_PCPU_CACHE;
2096 if (vm_domain_allocate(vmd, req, 1)) {
2098 * If not, allocate it from the free page queues.
2100 vm_domain_free_lock(vmd);
2101 m = vm_phys_alloc_pages(domain, pool, 0);
2102 vm_domain_free_unlock(vmd);
2104 vm_domain_freecnt_inc(vmd, 1);
2105 #if VM_NRESERVLEVEL > 0
2106 if (vm_reserv_reclaim_inactive(domain))
2113 * Not allocatable, give up.
2115 if (vm_domain_alloc_fail(vmd, object, req))
2121 * At this point we had better have found a good page.
2125 vm_page_alloc_check(m);
2128 * Initialize the page. Only the PG_ZERO flag is inherited.
2130 if ((req & VM_ALLOC_ZERO) != 0)
2131 flags |= (m->flags & PG_ZERO);
2132 if ((req & VM_ALLOC_NODUMP) != 0)
2136 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2138 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2139 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2140 else if ((req & VM_ALLOC_SBUSY) != 0)
2141 m->busy_lock = VPB_SHARERS_WORD(1);
2143 m->busy_lock = VPB_UNBUSIED;
2144 if (req & VM_ALLOC_WIRED) {
2150 if (object != NULL) {
2151 if (vm_page_insert_after(m, object, pindex, mpred)) {
2152 if (req & VM_ALLOC_WIRED) {
2156 KASSERT(m->object == NULL, ("page %p has object", m));
2157 m->oflags = VPO_UNMANAGED;
2158 m->busy_lock = VPB_UNBUSIED;
2159 /* Don't change PG_ZERO. */
2160 vm_page_free_toq(m);
2161 if (req & VM_ALLOC_WAITFAIL) {
2162 VM_OBJECT_WUNLOCK(object);
2164 VM_OBJECT_WLOCK(object);
2169 /* Ignore device objects; the pager sets "memattr" for them. */
2170 if (object->memattr != VM_MEMATTR_DEFAULT &&
2171 (object->flags & OBJ_FICTITIOUS) == 0)
2172 pmap_page_set_memattr(m, object->memattr);
2180 * vm_page_alloc_contig:
2182 * Allocate a contiguous set of physical pages of the given size "npages"
2183 * from the free lists. All of the physical pages must be at or above
2184 * the given physical address "low" and below the given physical address
2185 * "high". The given value "alignment" determines the alignment of the
2186 * first physical page in the set. If the given value "boundary" is
2187 * non-zero, then the set of physical pages cannot cross any physical
2188 * address boundary that is a multiple of that value. Both "alignment"
2189 * and "boundary" must be a power of two.
2191 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
2192 * then the memory attribute setting for the physical pages is configured
2193 * to the object's memory attribute setting. Otherwise, the memory
2194 * attribute setting for the physical pages is configured to "memattr",
2195 * overriding the object's memory attribute setting. However, if the
2196 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
2197 * memory attribute setting for the physical pages cannot be configured
2198 * to VM_MEMATTR_DEFAULT.
2200 * The specified object may not contain fictitious pages.
2202 * The caller must always specify an allocation class.
2204 * allocation classes:
2205 * VM_ALLOC_NORMAL normal process request
2206 * VM_ALLOC_SYSTEM system *really* needs a page
2207 * VM_ALLOC_INTERRUPT interrupt time request
2209 * optional allocation flags:
2210 * VM_ALLOC_NOBUSY do not exclusive busy the page
2211 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
2212 * VM_ALLOC_NOOBJ page is not associated with an object and
2213 * should not be exclusive busy
2214 * VM_ALLOC_SBUSY shared busy the allocated page
2215 * VM_ALLOC_WIRED wire the allocated page
2216 * VM_ALLOC_ZERO prefer a zeroed page
2219 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
2220 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2221 vm_paddr_t boundary, vm_memattr_t memattr)
2223 struct vm_domainset_iter di;
2227 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2229 m = vm_page_alloc_contig_domain(object, pindex, domain, req,
2230 npages, low, high, alignment, boundary, memattr);
2233 } while (vm_domainset_iter_page(&di, object, &domain) == 0);
2239 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2240 int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2241 vm_paddr_t boundary, vm_memattr_t memattr)
2243 struct vm_domain *vmd;
2244 vm_page_t m, m_ret, mpred;
2245 u_int busy_lock, flags, oflags;
2247 mpred = NULL; /* XXX: pacify gcc */
2248 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
2249 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
2250 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2251 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2252 ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
2254 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2255 ("Can't sleep and retry object insertion."));
2256 if (object != NULL) {
2257 VM_OBJECT_ASSERT_WLOCKED(object);
2258 KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2259 ("vm_page_alloc_contig: object %p has fictitious pages",
2262 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2264 if (object != NULL) {
2265 mpred = vm_radix_lookup_le(&object->rtree, pindex);
2266 KASSERT(mpred == NULL || mpred->pindex != pindex,
2267 ("vm_page_alloc_contig: pindex already allocated"));
2271 * Can we allocate the pages without the number of free pages falling
2272 * below the lower bound for the allocation class?
2276 #if VM_NRESERVLEVEL > 0
2278 * Can we allocate the pages from a reservation?
2280 if (vm_object_reserv(object) &&
2281 (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
2282 mpred, npages, low, high, alignment, boundary)) != NULL) {
2286 vmd = VM_DOMAIN(domain);
2287 if (vm_domain_allocate(vmd, req, npages)) {
2289 * allocate them from the free page queues.
2291 vm_domain_free_lock(vmd);
2292 m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2293 alignment, boundary);
2294 vm_domain_free_unlock(vmd);
2295 if (m_ret == NULL) {
2296 vm_domain_freecnt_inc(vmd, npages);
2297 #if VM_NRESERVLEVEL > 0
2298 if (vm_reserv_reclaim_contig(domain, npages, low,
2299 high, alignment, boundary))
2304 if (m_ret == NULL) {
2305 if (vm_domain_alloc_fail(vmd, object, req))
2309 #if VM_NRESERVLEVEL > 0
2312 for (m = m_ret; m < &m_ret[npages]; m++) {
2314 vm_page_alloc_check(m);
2318 * Initialize the pages. Only the PG_ZERO flag is inherited.
2321 if ((req & VM_ALLOC_ZERO) != 0)
2323 if ((req & VM_ALLOC_NODUMP) != 0)
2325 oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2327 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2328 busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2329 else if ((req & VM_ALLOC_SBUSY) != 0)
2330 busy_lock = VPB_SHARERS_WORD(1);
2332 busy_lock = VPB_UNBUSIED;
2333 if ((req & VM_ALLOC_WIRED) != 0)
2334 vm_wire_add(npages);
2335 if (object != NULL) {
2336 if (object->memattr != VM_MEMATTR_DEFAULT &&
2337 memattr == VM_MEMATTR_DEFAULT)
2338 memattr = object->memattr;
2340 for (m = m_ret; m < &m_ret[npages]; m++) {
2342 m->flags = (m->flags | PG_NODUMP) & flags;
2343 m->busy_lock = busy_lock;
2344 if ((req & VM_ALLOC_WIRED) != 0)
2348 if (object != NULL) {
2349 if (vm_page_insert_after(m, object, pindex, mpred)) {
2350 if ((req & VM_ALLOC_WIRED) != 0)
2351 vm_wire_sub(npages);
2352 KASSERT(m->object == NULL,
2353 ("page %p has object", m));
2355 for (m = m_ret; m < &m_ret[npages]; m++) {
2357 (req & VM_ALLOC_WIRED) != 0)
2359 m->oflags = VPO_UNMANAGED;
2360 m->busy_lock = VPB_UNBUSIED;
2361 /* Don't change PG_ZERO. */
2362 vm_page_free_toq(m);
2364 if (req & VM_ALLOC_WAITFAIL) {
2365 VM_OBJECT_WUNLOCK(object);
2367 VM_OBJECT_WLOCK(object);
2374 if (memattr != VM_MEMATTR_DEFAULT)
2375 pmap_page_set_memattr(m, memattr);
2382 * Check a page that has been freshly dequeued from a freelist.
2385 vm_page_alloc_check(vm_page_t m)
2388 KASSERT(m->object == NULL, ("page %p has object", m));
2389 KASSERT(m->a.queue == PQ_NONE &&
2390 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
2391 ("page %p has unexpected queue %d, flags %#x",
2392 m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK)));
2393 KASSERT(m->ref_count == 0, ("page %p has references", m));
2394 KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m));
2395 KASSERT(m->dirty == 0, ("page %p is dirty", m));
2396 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2397 ("page %p has unexpected memattr %d",
2398 m, pmap_page_get_memattr(m)));
2399 KASSERT(m->valid == 0, ("free page %p is valid", m));
2403 * vm_page_alloc_freelist:
2405 * Allocate a physical page from the specified free page list.
2407 * The caller must always specify an allocation class.
2409 * allocation classes:
2410 * VM_ALLOC_NORMAL normal process request
2411 * VM_ALLOC_SYSTEM system *really* needs a page
2412 * VM_ALLOC_INTERRUPT interrupt time request
2414 * optional allocation flags:
2415 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
2416 * intends to allocate
2417 * VM_ALLOC_WIRED wire the allocated page
2418 * VM_ALLOC_ZERO prefer a zeroed page
2421 vm_page_alloc_freelist(int freelist, int req)
2423 struct vm_domainset_iter di;
2427 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2429 m = vm_page_alloc_freelist_domain(domain, freelist, req);
2432 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2438 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2440 struct vm_domain *vmd;
2445 vmd = VM_DOMAIN(domain);
2447 if (vm_domain_allocate(vmd, req, 1)) {
2448 vm_domain_free_lock(vmd);
2449 m = vm_phys_alloc_freelist_pages(domain, freelist,
2450 VM_FREEPOOL_DIRECT, 0);
2451 vm_domain_free_unlock(vmd);
2453 vm_domain_freecnt_inc(vmd, 1);
2456 if (vm_domain_alloc_fail(vmd, NULL, req))
2461 vm_page_alloc_check(m);
2464 * Initialize the page. Only the PG_ZERO flag is inherited.
2468 if ((req & VM_ALLOC_ZERO) != 0)
2471 if ((req & VM_ALLOC_WIRED) != 0) {
2475 /* Unmanaged pages don't use "act_count". */
2476 m->oflags = VPO_UNMANAGED;
2481 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2483 struct vm_domain *vmd;
2484 struct vm_pgcache *pgcache;
2488 vmd = VM_DOMAIN(pgcache->domain);
2491 * The page daemon should avoid creating extra memory pressure since its
2492 * main purpose is to replenish the store of free pages.
2494 if (vmd->vmd_severeset || curproc == pageproc ||
2495 !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2497 domain = vmd->vmd_domain;
2498 vm_domain_free_lock(vmd);
2499 i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2500 (vm_page_t *)store);
2501 vm_domain_free_unlock(vmd);
2503 vm_domain_freecnt_inc(vmd, cnt - i);
2509 vm_page_zone_release(void *arg, void **store, int cnt)
2511 struct vm_domain *vmd;
2512 struct vm_pgcache *pgcache;
2517 vmd = VM_DOMAIN(pgcache->domain);
2518 vm_domain_free_lock(vmd);
2519 for (i = 0; i < cnt; i++) {
2520 m = (vm_page_t)store[i];
2521 vm_phys_free_pages(m, 0);
2523 vm_domain_free_unlock(vmd);
2524 vm_domain_freecnt_inc(vmd, cnt);
2527 #define VPSC_ANY 0 /* No restrictions. */
2528 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */
2529 #define VPSC_NOSUPER 2 /* Skip superpages. */
2532 * vm_page_scan_contig:
2534 * Scan vm_page_array[] between the specified entries "m_start" and
2535 * "m_end" for a run of contiguous physical pages that satisfy the
2536 * specified conditions, and return the lowest page in the run. The
2537 * specified "alignment" determines the alignment of the lowest physical
2538 * page in the run. If the specified "boundary" is non-zero, then the
2539 * run of physical pages cannot span a physical address that is a
2540 * multiple of "boundary".
2542 * "m_end" is never dereferenced, so it need not point to a vm_page
2543 * structure within vm_page_array[].
2545 * "npages" must be greater than zero. "m_start" and "m_end" must not
2546 * span a hole (or discontiguity) in the physical address space. Both
2547 * "alignment" and "boundary" must be a power of two.
2550 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2551 u_long alignment, vm_paddr_t boundary, int options)
2556 #if VM_NRESERVLEVEL > 0
2559 int m_inc, order, run_ext, run_len;
2561 KASSERT(npages > 0, ("npages is 0"));
2562 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2563 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2566 for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2567 KASSERT((m->flags & PG_MARKER) == 0,
2568 ("page %p is PG_MARKER", m));
2569 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1,
2570 ("fictitious page %p has invalid ref count", m));
2573 * If the current page would be the start of a run, check its
2574 * physical address against the end, alignment, and boundary
2575 * conditions. If it doesn't satisfy these conditions, either
2576 * terminate the scan or advance to the next page that
2577 * satisfies the failed condition.
2580 KASSERT(m_run == NULL, ("m_run != NULL"));
2581 if (m + npages > m_end)
2583 pa = VM_PAGE_TO_PHYS(m);
2584 if ((pa & (alignment - 1)) != 0) {
2585 m_inc = atop(roundup2(pa, alignment) - pa);
2588 if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2590 m_inc = atop(roundup2(pa, boundary) - pa);
2594 KASSERT(m_run != NULL, ("m_run == NULL"));
2598 if (vm_page_wired(m))
2600 #if VM_NRESERVLEVEL > 0
2601 else if ((level = vm_reserv_level(m)) >= 0 &&
2602 (options & VPSC_NORESERV) != 0) {
2604 /* Advance to the end of the reservation. */
2605 pa = VM_PAGE_TO_PHYS(m);
2606 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2610 else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2612 * The page is considered eligible for relocation if
2613 * and only if it could be laundered or reclaimed by
2616 VM_OBJECT_RLOCK(object);
2617 if (object != m->object) {
2618 VM_OBJECT_RUNLOCK(object);
2621 /* Don't care: PG_NODUMP, PG_ZERO. */
2622 if (object->type != OBJT_DEFAULT &&
2623 object->type != OBJT_SWAP &&
2624 object->type != OBJT_VNODE) {
2626 #if VM_NRESERVLEVEL > 0
2627 } else if ((options & VPSC_NOSUPER) != 0 &&
2628 (level = vm_reserv_level_iffullpop(m)) >= 0) {
2630 /* Advance to the end of the superpage. */
2631 pa = VM_PAGE_TO_PHYS(m);
2632 m_inc = atop(roundup2(pa + 1,
2633 vm_reserv_size(level)) - pa);
2635 } else if (object->memattr == VM_MEMATTR_DEFAULT &&
2636 vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) {
2638 * The page is allocated but eligible for
2639 * relocation. Extend the current run by one
2642 KASSERT(pmap_page_get_memattr(m) ==
2644 ("page %p has an unexpected memattr", m));
2645 KASSERT((m->oflags & (VPO_SWAPINPROG |
2646 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2647 ("page %p has unexpected oflags", m));
2648 /* Don't care: PGA_NOSYNC. */
2652 VM_OBJECT_RUNLOCK(object);
2653 #if VM_NRESERVLEVEL > 0
2654 } else if (level >= 0) {
2656 * The page is reserved but not yet allocated. In
2657 * other words, it is still free. Extend the current
2662 } else if ((order = m->order) < VM_NFREEORDER) {
2664 * The page is enqueued in the physical memory
2665 * allocator's free page queues. Moreover, it is the
2666 * first page in a power-of-two-sized run of
2667 * contiguous free pages. Add these pages to the end
2668 * of the current run, and jump ahead.
2670 run_ext = 1 << order;
2674 * Skip the page for one of the following reasons: (1)
2675 * It is enqueued in the physical memory allocator's
2676 * free page queues. However, it is not the first
2677 * page in a run of contiguous free pages. (This case
2678 * rarely occurs because the scan is performed in
2679 * ascending order.) (2) It is not reserved, and it is
2680 * transitioning from free to allocated. (Conversely,
2681 * the transition from allocated to free for managed
2682 * pages is blocked by the page busy lock.) (3) It is
2683 * allocated but not contained by an object and not
2684 * wired, e.g., allocated by Xen's balloon driver.
2690 * Extend or reset the current run of pages.
2703 if (run_len >= npages)
2709 * vm_page_reclaim_run:
2711 * Try to relocate each of the allocated virtual pages within the
2712 * specified run of physical pages to a new physical address. Free the
2713 * physical pages underlying the relocated virtual pages. A virtual page
2714 * is relocatable if and only if it could be laundered or reclaimed by
2715 * the page daemon. Whenever possible, a virtual page is relocated to a
2716 * physical address above "high".
2718 * Returns 0 if every physical page within the run was already free or
2719 * just freed by a successful relocation. Otherwise, returns a non-zero
2720 * value indicating why the last attempt to relocate a virtual page was
2723 * "req_class" must be an allocation class.
2726 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2729 struct vm_domain *vmd;
2730 struct spglist free;
2733 vm_page_t m, m_end, m_new;
2734 int error, order, req;
2736 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2737 ("req_class is not an allocation class"));
2741 m_end = m_run + npages;
2742 for (; error == 0 && m < m_end; m++) {
2743 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2744 ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2747 * Racily check for wirings. Races are handled once the object
2748 * lock is held and the page is unmapped.
2750 if (vm_page_wired(m))
2752 else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2754 * The page is relocated if and only if it could be
2755 * laundered or reclaimed by the page daemon.
2757 VM_OBJECT_WLOCK(object);
2758 /* Don't care: PG_NODUMP, PG_ZERO. */
2759 if (m->object != object ||
2760 (object->type != OBJT_DEFAULT &&
2761 object->type != OBJT_SWAP &&
2762 object->type != OBJT_VNODE))
2764 else if (object->memattr != VM_MEMATTR_DEFAULT)
2766 else if (vm_page_queue(m) != PQ_NONE &&
2767 vm_page_tryxbusy(m) != 0) {
2768 if (vm_page_wired(m)) {
2773 KASSERT(pmap_page_get_memattr(m) ==
2775 ("page %p has an unexpected memattr", m));
2776 KASSERT(m->oflags == 0,
2777 ("page %p has unexpected oflags", m));
2778 /* Don't care: PGA_NOSYNC. */
2779 if (!vm_page_none_valid(m)) {
2781 * First, try to allocate a new page
2782 * that is above "high". Failing
2783 * that, try to allocate a new page
2784 * that is below "m_run". Allocate
2785 * the new page between the end of
2786 * "m_run" and "high" only as a last
2789 req = req_class | VM_ALLOC_NOOBJ;
2790 if ((m->flags & PG_NODUMP) != 0)
2791 req |= VM_ALLOC_NODUMP;
2792 if (trunc_page(high) !=
2793 ~(vm_paddr_t)PAGE_MASK) {
2794 m_new = vm_page_alloc_contig(
2799 VM_MEMATTR_DEFAULT);
2802 if (m_new == NULL) {
2803 pa = VM_PAGE_TO_PHYS(m_run);
2804 m_new = vm_page_alloc_contig(
2806 0, pa - 1, PAGE_SIZE, 0,
2807 VM_MEMATTR_DEFAULT);
2809 if (m_new == NULL) {
2811 m_new = vm_page_alloc_contig(
2813 pa, high, PAGE_SIZE, 0,
2814 VM_MEMATTR_DEFAULT);
2816 if (m_new == NULL) {
2823 * Unmap the page and check for new
2824 * wirings that may have been acquired
2825 * through a pmap lookup.
2827 if (object->ref_count != 0 &&
2828 !vm_page_try_remove_all(m)) {
2830 vm_page_free(m_new);
2836 * Replace "m" with the new page. For
2837 * vm_page_replace(), "m" must be busy
2838 * and dequeued. Finally, change "m"
2839 * as if vm_page_free() was called.
2841 m_new->a.flags = m->a.flags &
2842 ~PGA_QUEUE_STATE_MASK;
2843 KASSERT(m_new->oflags == VPO_UNMANAGED,
2844 ("page %p is managed", m_new));
2846 pmap_copy_page(m, m_new);
2847 m_new->valid = m->valid;
2848 m_new->dirty = m->dirty;
2849 m->flags &= ~PG_ZERO;
2851 if (vm_page_replace_hold(m_new, object,
2853 vm_page_free_prep(m))
2854 SLIST_INSERT_HEAD(&free, m,
2858 * The new page must be deactivated
2859 * before the object is unlocked.
2861 vm_page_deactivate(m_new);
2863 m->flags &= ~PG_ZERO;
2865 if (vm_page_free_prep(m))
2866 SLIST_INSERT_HEAD(&free, m,
2868 KASSERT(m->dirty == 0,
2869 ("page %p is dirty", m));
2874 VM_OBJECT_WUNLOCK(object);
2876 MPASS(vm_phys_domain(m) == domain);
2877 vmd = VM_DOMAIN(domain);
2878 vm_domain_free_lock(vmd);
2880 if (order < VM_NFREEORDER) {
2882 * The page is enqueued in the physical memory
2883 * allocator's free page queues. Moreover, it
2884 * is the first page in a power-of-two-sized
2885 * run of contiguous free pages. Jump ahead
2886 * to the last page within that run, and
2887 * continue from there.
2889 m += (1 << order) - 1;
2891 #if VM_NRESERVLEVEL > 0
2892 else if (vm_reserv_is_page_free(m))
2895 vm_domain_free_unlock(vmd);
2896 if (order == VM_NFREEORDER)
2900 if ((m = SLIST_FIRST(&free)) != NULL) {
2903 vmd = VM_DOMAIN(domain);
2905 vm_domain_free_lock(vmd);
2907 MPASS(vm_phys_domain(m) == domain);
2908 SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2909 vm_phys_free_pages(m, 0);
2911 } while ((m = SLIST_FIRST(&free)) != NULL);
2912 vm_domain_free_unlock(vmd);
2913 vm_domain_freecnt_inc(vmd, cnt);
2920 CTASSERT(powerof2(NRUNS));
2922 #define RUN_INDEX(count) ((count) & (NRUNS - 1))
2924 #define MIN_RECLAIM 8
2927 * vm_page_reclaim_contig:
2929 * Reclaim allocated, contiguous physical memory satisfying the specified
2930 * conditions by relocating the virtual pages using that physical memory.
2931 * Returns true if reclamation is successful and false otherwise. Since
2932 * relocation requires the allocation of physical pages, reclamation may
2933 * fail due to a shortage of free pages. When reclamation fails, callers
2934 * are expected to perform vm_wait() before retrying a failed allocation
2935 * operation, e.g., vm_page_alloc_contig().
2937 * The caller must always specify an allocation class through "req".
2939 * allocation classes:
2940 * VM_ALLOC_NORMAL normal process request
2941 * VM_ALLOC_SYSTEM system *really* needs a page
2942 * VM_ALLOC_INTERRUPT interrupt time request
2944 * The optional allocation flags are ignored.
2946 * "npages" must be greater than zero. Both "alignment" and "boundary"
2947 * must be a power of two.
2950 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
2951 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2953 struct vm_domain *vmd;
2954 vm_paddr_t curr_low;
2955 vm_page_t m_run, m_runs[NRUNS];
2956 u_long count, reclaimed;
2957 int error, i, options, req_class;
2959 KASSERT(npages > 0, ("npages is 0"));
2960 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2961 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2962 req_class = req & VM_ALLOC_CLASS_MASK;
2965 * The page daemon is allowed to dig deeper into the free page list.
2967 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2968 req_class = VM_ALLOC_SYSTEM;
2971 * Return if the number of free pages cannot satisfy the requested
2974 vmd = VM_DOMAIN(domain);
2975 count = vmd->vmd_free_count;
2976 if (count < npages + vmd->vmd_free_reserved || (count < npages +
2977 vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2978 (count < npages && req_class == VM_ALLOC_INTERRUPT))
2982 * Scan up to three times, relaxing the restrictions ("options") on
2983 * the reclamation of reservations and superpages each time.
2985 for (options = VPSC_NORESERV;;) {
2987 * Find the highest runs that satisfy the given constraints
2988 * and restrictions, and record them in "m_runs".
2993 m_run = vm_phys_scan_contig(domain, npages, curr_low,
2994 high, alignment, boundary, options);
2997 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2998 m_runs[RUN_INDEX(count)] = m_run;
3003 * Reclaim the highest runs in LIFO (descending) order until
3004 * the number of reclaimed pages, "reclaimed", is at least
3005 * MIN_RECLAIM. Reset "reclaimed" each time because each
3006 * reclamation is idempotent, and runs will (likely) recur
3007 * from one scan to the next as restrictions are relaxed.
3010 for (i = 0; count > 0 && i < NRUNS; i++) {
3012 m_run = m_runs[RUN_INDEX(count)];
3013 error = vm_page_reclaim_run(req_class, domain, npages,
3016 reclaimed += npages;
3017 if (reclaimed >= MIN_RECLAIM)
3023 * Either relax the restrictions on the next scan or return if
3024 * the last scan had no restrictions.
3026 if (options == VPSC_NORESERV)
3027 options = VPSC_NOSUPER;
3028 else if (options == VPSC_NOSUPER)
3030 else if (options == VPSC_ANY)
3031 return (reclaimed != 0);
3036 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
3037 u_long alignment, vm_paddr_t boundary)
3039 struct vm_domainset_iter di;
3043 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
3045 ret = vm_page_reclaim_contig_domain(domain, req, npages, low,
3046 high, alignment, boundary);
3049 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
3055 * Set the domain in the appropriate page level domainset.
3058 vm_domain_set(struct vm_domain *vmd)
3061 mtx_lock(&vm_domainset_lock);
3062 if (!vmd->vmd_minset && vm_paging_min(vmd)) {
3063 vmd->vmd_minset = 1;
3064 DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
3066 if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
3067 vmd->vmd_severeset = 1;
3068 DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
3070 mtx_unlock(&vm_domainset_lock);
3074 * Clear the domain from the appropriate page level domainset.
3077 vm_domain_clear(struct vm_domain *vmd)
3080 mtx_lock(&vm_domainset_lock);
3081 if (vmd->vmd_minset && !vm_paging_min(vmd)) {
3082 vmd->vmd_minset = 0;
3083 DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
3084 if (vm_min_waiters != 0) {
3086 wakeup(&vm_min_domains);
3089 if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
3090 vmd->vmd_severeset = 0;
3091 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
3092 if (vm_severe_waiters != 0) {
3093 vm_severe_waiters = 0;
3094 wakeup(&vm_severe_domains);
3099 * If pageout daemon needs pages, then tell it that there are
3102 if (vmd->vmd_pageout_pages_needed &&
3103 vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
3104 wakeup(&vmd->vmd_pageout_pages_needed);
3105 vmd->vmd_pageout_pages_needed = 0;
3108 /* See comments in vm_wait_doms(). */
3109 if (vm_pageproc_waiters) {
3110 vm_pageproc_waiters = 0;
3111 wakeup(&vm_pageproc_waiters);
3113 mtx_unlock(&vm_domainset_lock);
3117 * Wait for free pages to exceed the min threshold globally.
3123 mtx_lock(&vm_domainset_lock);
3124 while (vm_page_count_min()) {
3126 msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
3128 mtx_unlock(&vm_domainset_lock);
3132 * Wait for free pages to exceed the severe threshold globally.
3135 vm_wait_severe(void)
3138 mtx_lock(&vm_domainset_lock);
3139 while (vm_page_count_severe()) {
3140 vm_severe_waiters++;
3141 msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
3144 mtx_unlock(&vm_domainset_lock);
3151 return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
3155 vm_wait_doms(const domainset_t *wdoms, int mflags)
3162 * We use racey wakeup synchronization to avoid expensive global
3163 * locking for the pageproc when sleeping with a non-specific vm_wait.
3164 * To handle this, we only sleep for one tick in this instance. It
3165 * is expected that most allocations for the pageproc will come from
3166 * kmem or vm_page_grab* which will use the more specific and
3167 * race-free vm_wait_domain().
3169 if (curproc == pageproc) {
3170 mtx_lock(&vm_domainset_lock);
3171 vm_pageproc_waiters++;
3172 error = msleep(&vm_pageproc_waiters, &vm_domainset_lock,
3173 PVM | PDROP | mflags, "pageprocwait", 1);
3176 * XXX Ideally we would wait only until the allocation could
3177 * be satisfied. This condition can cause new allocators to
3178 * consume all freed pages while old allocators wait.
3180 mtx_lock(&vm_domainset_lock);
3181 if (vm_page_count_min_set(wdoms)) {
3183 error = msleep(&vm_min_domains, &vm_domainset_lock,
3184 PVM | PDROP | mflags, "vmwait", 0);
3186 mtx_unlock(&vm_domainset_lock);
3194 * Sleep until free pages are available for allocation.
3195 * - Called in various places after failed memory allocations.
3198 vm_wait_domain(int domain)
3200 struct vm_domain *vmd;
3203 vmd = VM_DOMAIN(domain);
3204 vm_domain_free_assert_unlocked(vmd);
3206 if (curproc == pageproc) {
3207 mtx_lock(&vm_domainset_lock);
3208 if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3209 vmd->vmd_pageout_pages_needed = 1;
3210 msleep(&vmd->vmd_pageout_pages_needed,
3211 &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3213 mtx_unlock(&vm_domainset_lock);
3215 if (pageproc == NULL)
3216 panic("vm_wait in early boot");
3217 DOMAINSET_ZERO(&wdom);
3218 DOMAINSET_SET(vmd->vmd_domain, &wdom);
3219 vm_wait_doms(&wdom, 0);
3224 vm_wait_flags(vm_object_t obj, int mflags)
3226 struct domainset *d;
3231 * Carefully fetch pointers only once: the struct domainset
3232 * itself is ummutable but the pointer might change.
3235 d = obj->domain.dr_policy;
3237 d = curthread->td_domain.dr_policy;
3239 return (vm_wait_doms(&d->ds_mask, mflags));
3245 * Sleep until free pages are available for allocation in the
3246 * affinity domains of the obj. If obj is NULL, the domain set
3247 * for the calling thread is used.
3248 * Called in various places after failed memory allocations.
3251 vm_wait(vm_object_t obj)
3253 (void)vm_wait_flags(obj, 0);
3257 vm_wait_intr(vm_object_t obj)
3259 return (vm_wait_flags(obj, PCATCH));
3263 * vm_domain_alloc_fail:
3265 * Called when a page allocation function fails. Informs the
3266 * pagedaemon and performs the requested wait. Requires the
3267 * domain_free and object lock on entry. Returns with the
3268 * object lock held and free lock released. Returns an error when
3269 * retry is necessary.
3273 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3276 vm_domain_free_assert_unlocked(vmd);
3278 atomic_add_int(&vmd->vmd_pageout_deficit,
3279 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3280 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3282 VM_OBJECT_WUNLOCK(object);
3283 vm_wait_domain(vmd->vmd_domain);
3285 VM_OBJECT_WLOCK(object);
3286 if (req & VM_ALLOC_WAITOK)
3296 * Sleep until free pages are available for allocation.
3297 * - Called only in vm_fault so that processes page faulting
3298 * can be easily tracked.
3299 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3300 * processes will be able to grab memory first. Do not change
3301 * this balance without careful testing first.
3304 vm_waitpfault(struct domainset *dset, int timo)
3308 * XXX Ideally we would wait only until the allocation could
3309 * be satisfied. This condition can cause new allocators to
3310 * consume all freed pages while old allocators wait.
3312 mtx_lock(&vm_domainset_lock);
3313 if (vm_page_count_min_set(&dset->ds_mask)) {
3315 msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3318 mtx_unlock(&vm_domainset_lock);
3321 static struct vm_pagequeue *
3322 _vm_page_pagequeue(vm_page_t m, uint8_t queue)
3325 return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3329 static struct vm_pagequeue *
3330 vm_page_pagequeue(vm_page_t m)
3333 return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue));
3337 static __always_inline bool
3338 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3340 vm_page_astate_t tmp;
3344 if (__predict_true(vm_page_astate_fcmpset(m, old, new)))
3346 counter_u64_add(pqstate_commit_retries, 1);
3347 } while (old->_bits == tmp._bits);
3353 * Do the work of committing a queue state update that moves the page out of
3354 * its current queue.
3357 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m,
3358 vm_page_astate_t *old, vm_page_astate_t new)
3362 vm_pagequeue_assert_locked(pq);
3363 KASSERT(vm_page_pagequeue(m) == pq,
3364 ("%s: queue %p does not match page %p", __func__, pq, m));
3365 KASSERT(old->queue != PQ_NONE && new.queue != old->queue,
3366 ("%s: invalid queue indices %d %d",
3367 __func__, old->queue, new.queue));
3370 * Once the queue index of the page changes there is nothing
3371 * synchronizing with further updates to the page's physical
3372 * queue state. Therefore we must speculatively remove the page
3373 * from the queue now and be prepared to roll back if the queue
3374 * state update fails. If the page is not physically enqueued then
3375 * we just update its queue index.
3377 if ((old->flags & PGA_ENQUEUED) != 0) {
3378 new.flags &= ~PGA_ENQUEUED;
3379 next = TAILQ_NEXT(m, plinks.q);
3380 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3381 vm_pagequeue_cnt_dec(pq);
3382 if (!vm_page_pqstate_fcmpset(m, old, new)) {
3384 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3386 TAILQ_INSERT_BEFORE(next, m, plinks.q);
3387 vm_pagequeue_cnt_inc(pq);
3393 return (vm_page_pqstate_fcmpset(m, old, new));
3398 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old,
3399 vm_page_astate_t new)
3401 struct vm_pagequeue *pq;
3402 vm_page_astate_t as;
3405 pq = _vm_page_pagequeue(m, old->queue);
3408 * The queue field and PGA_ENQUEUED flag are stable only so long as the
3409 * corresponding page queue lock is held.
3411 vm_pagequeue_lock(pq);
3412 as = vm_page_astate_load(m);
3413 if (__predict_false(as._bits != old->_bits)) {
3417 ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new);
3419 vm_pagequeue_unlock(pq);
3424 * Commit a queue state update that enqueues or requeues a page.
3427 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m,
3428 vm_page_astate_t *old, vm_page_astate_t new)
3430 struct vm_domain *vmd;
3432 vm_pagequeue_assert_locked(pq);
3433 KASSERT(old->queue != PQ_NONE && new.queue == old->queue,
3434 ("%s: invalid queue indices %d %d",
3435 __func__, old->queue, new.queue));
3437 new.flags |= PGA_ENQUEUED;
3438 if (!vm_page_pqstate_fcmpset(m, old, new))
3441 if ((old->flags & PGA_ENQUEUED) != 0)
3442 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3444 vm_pagequeue_cnt_inc(pq);
3447 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE. In particular, if
3448 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be
3449 * applied, even if it was set first.
3451 if ((old->flags & PGA_REQUEUE_HEAD) != 0) {
3452 vmd = vm_pagequeue_domain(m);
3453 KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE],
3454 ("%s: invalid page queue for page %p", __func__, m));
3455 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3457 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3463 * Commit a queue state update that encodes a request for a deferred queue
3467 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old,
3468 vm_page_astate_t new)
3471 KASSERT(old->queue == new.queue || new.queue != PQ_NONE,
3472 ("%s: invalid state, queue %d flags %x",
3473 __func__, new.queue, new.flags));
3475 if (old->_bits != new._bits &&
3476 !vm_page_pqstate_fcmpset(m, old, new))
3478 vm_page_pqbatch_submit(m, new.queue);
3483 * A generic queue state update function. This handles more cases than the
3484 * specialized functions above.
3487 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3490 if (old->_bits == new._bits)
3493 if (old->queue != PQ_NONE && new.queue != old->queue) {
3494 if (!vm_page_pqstate_commit_dequeue(m, old, new))
3496 if (new.queue != PQ_NONE)
3497 vm_page_pqbatch_submit(m, new.queue);
3499 if (!vm_page_pqstate_fcmpset(m, old, new))
3501 if (new.queue != PQ_NONE &&
3502 ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0)
3503 vm_page_pqbatch_submit(m, new.queue);
3509 * Apply deferred queue state updates to a page.
3512 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue)
3514 vm_page_astate_t new, old;
3516 CRITICAL_ASSERT(curthread);
3517 vm_pagequeue_assert_locked(pq);
3518 KASSERT(queue < PQ_COUNT,
3519 ("%s: invalid queue index %d", __func__, queue));
3520 KASSERT(pq == _vm_page_pagequeue(m, queue),
3521 ("%s: page %p does not belong to queue %p", __func__, m, pq));
3523 for (old = vm_page_astate_load(m);;) {
3524 if (__predict_false(old.queue != queue ||
3525 (old.flags & PGA_QUEUE_OP_MASK) == 0)) {
3526 counter_u64_add(queue_nops, 1);
3529 KASSERT(old.queue != PQ_NONE || (old.flags & PGA_QUEUE_STATE_MASK) == 0,
3530 ("%s: page %p has unexpected queue state", __func__, m));
3533 if ((old.flags & PGA_DEQUEUE) != 0) {
3534 new.flags &= ~PGA_QUEUE_OP_MASK;
3535 new.queue = PQ_NONE;
3536 if (__predict_true(_vm_page_pqstate_commit_dequeue(pq,
3538 counter_u64_add(queue_ops, 1);
3542 new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD);
3543 if (__predict_true(_vm_page_pqstate_commit_requeue(pq,
3545 counter_u64_add(queue_ops, 1);
3553 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3558 for (i = 0; i < bq->bq_cnt; i++)
3559 vm_pqbatch_process_page(pq, bq->bq_pa[i], queue);
3560 vm_batchqueue_init(bq);
3564 * vm_page_pqbatch_submit: [ internal use only ]
3566 * Enqueue a page in the specified page queue's batched work queue.
3567 * The caller must have encoded the requested operation in the page
3568 * structure's a.flags field.
3571 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
3573 struct vm_batchqueue *bq;
3574 struct vm_pagequeue *pq;
3577 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3578 ("page %p is unmanaged", m));
3579 KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3581 domain = vm_phys_domain(m);
3582 pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue];
3585 bq = DPCPU_PTR(pqbatch[domain][queue]);
3586 if (vm_batchqueue_insert(bq, m)) {
3591 vm_pagequeue_lock(pq);
3593 bq = DPCPU_PTR(pqbatch[domain][queue]);
3594 vm_pqbatch_process(pq, bq, queue);
3595 vm_pqbatch_process_page(pq, m, queue);
3596 vm_pagequeue_unlock(pq);
3601 * vm_page_pqbatch_drain: [ internal use only ]
3603 * Force all per-CPU page queue batch queues to be drained. This is
3604 * intended for use in severe memory shortages, to ensure that pages
3605 * do not remain stuck in the batch queues.
3608 vm_page_pqbatch_drain(void)
3611 struct vm_domain *vmd;
3612 struct vm_pagequeue *pq;
3613 int cpu, domain, queue;
3618 sched_bind(td, cpu);
3621 for (domain = 0; domain < vm_ndomains; domain++) {
3622 vmd = VM_DOMAIN(domain);
3623 for (queue = 0; queue < PQ_COUNT; queue++) {
3624 pq = &vmd->vmd_pagequeues[queue];
3625 vm_pagequeue_lock(pq);
3627 vm_pqbatch_process(pq,
3628 DPCPU_PTR(pqbatch[domain][queue]), queue);
3630 vm_pagequeue_unlock(pq);
3640 * vm_page_dequeue_deferred: [ internal use only ]
3642 * Request removal of the given page from its current page
3643 * queue. Physical removal from the queue may be deferred
3647 vm_page_dequeue_deferred(vm_page_t m)
3649 vm_page_astate_t new, old;
3651 old = vm_page_astate_load(m);
3653 if (old.queue == PQ_NONE) {
3654 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3655 ("%s: page %p has unexpected queue state",
3660 new.flags |= PGA_DEQUEUE;
3661 } while (!vm_page_pqstate_commit_request(m, &old, new));
3667 * Remove the page from whichever page queue it's in, if any, before
3671 vm_page_dequeue(vm_page_t m)
3673 vm_page_astate_t new, old;
3675 old = vm_page_astate_load(m);
3677 if (old.queue == PQ_NONE) {
3678 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3679 ("%s: page %p has unexpected queue state",
3684 new.flags &= ~PGA_QUEUE_OP_MASK;
3685 new.queue = PQ_NONE;
3686 } while (!vm_page_pqstate_commit_dequeue(m, &old, new));
3691 * Schedule the given page for insertion into the specified page queue.
3692 * Physical insertion of the page may be deferred indefinitely.
3695 vm_page_enqueue(vm_page_t m, uint8_t queue)
3698 KASSERT(m->a.queue == PQ_NONE &&
3699 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
3700 ("%s: page %p is already enqueued", __func__, m));
3701 KASSERT(m->ref_count > 0,
3702 ("%s: page %p does not carry any references", __func__, m));
3705 if ((m->a.flags & PGA_REQUEUE) == 0)
3706 vm_page_aflag_set(m, PGA_REQUEUE);
3707 vm_page_pqbatch_submit(m, queue);
3711 * vm_page_free_prep:
3713 * Prepares the given page to be put on the free list,
3714 * disassociating it from any VM object. The caller may return
3715 * the page to the free list only if this function returns true.
3717 * The object, if it exists, must be locked, and then the page must
3718 * be xbusy. Otherwise the page must be not busied. A managed
3719 * page must be unmapped.
3722 vm_page_free_prep(vm_page_t m)
3726 * Synchronize with threads that have dropped a reference to this
3729 atomic_thread_fence_acq();
3731 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3732 if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3735 p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3736 for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3737 KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3738 m, i, (uintmax_t)*p));
3741 if ((m->oflags & VPO_UNMANAGED) == 0) {
3742 KASSERT(!pmap_page_is_mapped(m),
3743 ("vm_page_free_prep: freeing mapped page %p", m));
3744 KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
3745 ("vm_page_free_prep: mapping flags set in page %p", m));
3747 KASSERT(m->a.queue == PQ_NONE,
3748 ("vm_page_free_prep: unmanaged page %p is queued", m));
3750 VM_CNT_INC(v_tfree);
3752 if (m->object != NULL) {
3753 KASSERT(((m->oflags & VPO_UNMANAGED) != 0) ==
3754 ((m->object->flags & OBJ_UNMANAGED) != 0),
3755 ("vm_page_free_prep: managed flag mismatch for page %p",
3757 vm_page_assert_xbusied(m);
3760 * The object reference can be released without an atomic
3763 KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
3764 m->ref_count == VPRC_OBJREF,
3765 ("vm_page_free_prep: page %p has unexpected ref_count %u",
3767 vm_page_object_remove(m);
3768 m->ref_count -= VPRC_OBJREF;
3770 vm_page_assert_unbusied(m);
3772 vm_page_busy_free(m);
3775 * If fictitious remove object association and
3778 if ((m->flags & PG_FICTITIOUS) != 0) {
3779 KASSERT(m->ref_count == 1,
3780 ("fictitious page %p is referenced", m));
3781 KASSERT(m->a.queue == PQ_NONE,
3782 ("fictitious page %p is queued", m));
3787 * Pages need not be dequeued before they are returned to the physical
3788 * memory allocator, but they must at least be marked for a deferred
3791 if ((m->oflags & VPO_UNMANAGED) == 0)
3792 vm_page_dequeue_deferred(m);
3797 if (m->ref_count != 0)
3798 panic("vm_page_free_prep: page %p has references", m);
3801 * Restore the default memory attribute to the page.
3803 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3804 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3806 #if VM_NRESERVLEVEL > 0
3808 * Determine whether the page belongs to a reservation. If the page was
3809 * allocated from a per-CPU cache, it cannot belong to a reservation, so
3810 * as an optimization, we avoid the check in that case.
3812 if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
3822 * Returns the given page to the free list, disassociating it
3823 * from any VM object.
3825 * The object must be locked. The page must be exclusively busied if it
3826 * belongs to an object.
3829 vm_page_free_toq(vm_page_t m)
3831 struct vm_domain *vmd;
3834 if (!vm_page_free_prep(m))
3837 vmd = vm_pagequeue_domain(m);
3838 zone = vmd->vmd_pgcache[m->pool].zone;
3839 if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
3843 vm_domain_free_lock(vmd);
3844 vm_phys_free_pages(m, 0);
3845 vm_domain_free_unlock(vmd);
3846 vm_domain_freecnt_inc(vmd, 1);
3850 * vm_page_free_pages_toq:
3852 * Returns a list of pages to the free list, disassociating it
3853 * from any VM object. In other words, this is equivalent to
3854 * calling vm_page_free_toq() for each page of a list of VM objects.
3857 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
3862 if (SLIST_EMPTY(free))
3866 while ((m = SLIST_FIRST(free)) != NULL) {
3868 SLIST_REMOVE_HEAD(free, plinks.s.ss);
3869 vm_page_free_toq(m);
3872 if (update_wire_count)
3877 * Mark this page as wired down. For managed pages, this prevents reclamation
3878 * by the page daemon, or when the containing object, if any, is destroyed.
3881 vm_page_wire(vm_page_t m)
3886 if (m->object != NULL && !vm_page_busied(m) &&
3887 !vm_object_busied(m->object))
3888 VM_OBJECT_ASSERT_LOCKED(m->object);
3890 KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
3891 VPRC_WIRE_COUNT(m->ref_count) >= 1,
3892 ("vm_page_wire: fictitious page %p has zero wirings", m));
3894 old = atomic_fetchadd_int(&m->ref_count, 1);
3895 KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
3896 ("vm_page_wire: counter overflow for page %p", m));
3897 if (VPRC_WIRE_COUNT(old) == 0) {
3898 if ((m->oflags & VPO_UNMANAGED) == 0)
3899 vm_page_aflag_set(m, PGA_DEQUEUE);
3905 * Attempt to wire a mapped page following a pmap lookup of that page.
3906 * This may fail if a thread is concurrently tearing down mappings of the page.
3907 * The transient failure is acceptable because it translates to the
3908 * failure of the caller pmap_extract_and_hold(), which should be then
3909 * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages().
3912 vm_page_wire_mapped(vm_page_t m)
3919 ("vm_page_wire_mapped: wiring unreferenced page %p", m));
3920 if ((old & VPRC_BLOCKED) != 0)
3922 } while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
3924 if (VPRC_WIRE_COUNT(old) == 0) {
3925 if ((m->oflags & VPO_UNMANAGED) == 0)
3926 vm_page_aflag_set(m, PGA_DEQUEUE);
3933 * Release a wiring reference to a managed page. If the page still belongs to
3934 * an object, update its position in the page queues to reflect the reference.
3935 * If the wiring was the last reference to the page, free the page.
3938 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse)
3942 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3943 ("%s: page %p is unmanaged", __func__, m));
3946 * Update LRU state before releasing the wiring reference.
3947 * Use a release store when updating the reference count to
3948 * synchronize with vm_page_free_prep().
3952 KASSERT(VPRC_WIRE_COUNT(old) > 0,
3953 ("vm_page_unwire: wire count underflow for page %p", m));
3955 if (old > VPRC_OBJREF + 1) {
3957 * The page has at least one other wiring reference. An
3958 * earlier iteration of this loop may have called
3959 * vm_page_release_toq() and cleared PGA_DEQUEUE, so
3960 * re-set it if necessary.
3962 if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0)
3963 vm_page_aflag_set(m, PGA_DEQUEUE);
3964 } else if (old == VPRC_OBJREF + 1) {
3966 * This is the last wiring. Clear PGA_DEQUEUE and
3967 * update the page's queue state to reflect the
3968 * reference. If the page does not belong to an object
3969 * (i.e., the VPRC_OBJREF bit is clear), we only need to
3970 * clear leftover queue state.
3972 vm_page_release_toq(m, nqueue, false);
3973 } else if (old == 1) {
3974 vm_page_aflag_clear(m, PGA_DEQUEUE);
3976 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
3978 if (VPRC_WIRE_COUNT(old) == 1) {
3986 * Release one wiring of the specified page, potentially allowing it to be
3989 * Only managed pages belonging to an object can be paged out. If the number
3990 * of wirings transitions to zero and the page is eligible for page out, then
3991 * the page is added to the specified paging queue. If the released wiring
3992 * represented the last reference to the page, the page is freed.
3995 vm_page_unwire(vm_page_t m, uint8_t nqueue)
3998 KASSERT(nqueue < PQ_COUNT,
3999 ("vm_page_unwire: invalid queue %u request for page %p",
4002 if ((m->oflags & VPO_UNMANAGED) != 0) {
4003 if (vm_page_unwire_noq(m) && m->ref_count == 0)
4007 vm_page_unwire_managed(m, nqueue, false);
4011 * Unwire a page without (re-)inserting it into a page queue. It is up
4012 * to the caller to enqueue, requeue, or free the page as appropriate.
4013 * In most cases involving managed pages, vm_page_unwire() should be used
4017 vm_page_unwire_noq(vm_page_t m)
4021 old = vm_page_drop(m, 1);
4022 KASSERT(VPRC_WIRE_COUNT(old) != 0,
4023 ("vm_page_unref: counter underflow for page %p", m));
4024 KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
4025 ("vm_page_unref: missing ref on fictitious page %p", m));
4027 if (VPRC_WIRE_COUNT(old) > 1)
4029 if ((m->oflags & VPO_UNMANAGED) == 0)
4030 vm_page_aflag_clear(m, PGA_DEQUEUE);
4036 * Ensure that the page ends up in the specified page queue. If the page is
4037 * active or being moved to the active queue, ensure that its act_count is
4038 * at least ACT_INIT but do not otherwise mess with it.
4040 static __always_inline void
4041 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag)
4043 vm_page_astate_t old, new;
4045 KASSERT(m->ref_count > 0,
4046 ("%s: page %p does not carry any references", __func__, m));
4047 KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD,
4048 ("%s: invalid flags %x", __func__, nflag));
4050 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
4053 old = vm_page_astate_load(m);
4055 if ((old.flags & PGA_DEQUEUE) != 0)
4058 new.flags &= ~PGA_QUEUE_OP_MASK;
4059 if (nqueue == PQ_ACTIVE)
4060 new.act_count = max(old.act_count, ACT_INIT);
4061 if (old.queue == nqueue) {
4062 if (nqueue != PQ_ACTIVE)
4068 } while (!vm_page_pqstate_commit(m, &old, new));
4072 * Put the specified page on the active list (if appropriate).
4075 vm_page_activate(vm_page_t m)
4078 vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE);
4082 * Move the specified page to the tail of the inactive queue, or requeue
4083 * the page if it is already in the inactive queue.
4086 vm_page_deactivate(vm_page_t m)
4089 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE);
4093 vm_page_deactivate_noreuse(vm_page_t m)
4096 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD);
4100 * Put a page in the laundry, or requeue it if it is already there.
4103 vm_page_launder(vm_page_t m)
4106 vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE);
4110 * Put a page in the PQ_UNSWAPPABLE holding queue.
4113 vm_page_unswappable(vm_page_t m)
4116 KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0,
4117 ("page %p already unswappable", m));
4120 vm_page_enqueue(m, PQ_UNSWAPPABLE);
4124 * Release a page back to the page queues in preparation for unwiring.
4127 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse)
4129 vm_page_astate_t old, new;
4133 * Use a check of the valid bits to determine whether we should
4134 * accelerate reclamation of the page. The object lock might not be
4135 * held here, in which case the check is racy. At worst we will either
4136 * accelerate reclamation of a valid page and violate LRU, or
4137 * unnecessarily defer reclamation of an invalid page.
4139 * If we were asked to not cache the page, place it near the head of the
4140 * inactive queue so that is reclaimed sooner.
4142 if (noreuse || m->valid == 0) {
4143 nqueue = PQ_INACTIVE;
4144 nflag = PGA_REQUEUE_HEAD;
4146 nflag = PGA_REQUEUE;
4149 old = vm_page_astate_load(m);
4154 * If the page is already in the active queue and we are not
4155 * trying to accelerate reclamation, simply mark it as
4156 * referenced and avoid any queue operations.
4158 new.flags &= ~PGA_QUEUE_OP_MASK;
4159 if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE)
4160 new.flags |= PGA_REFERENCED;
4165 } while (!vm_page_pqstate_commit(m, &old, new));
4169 * Unwire a page and either attempt to free it or re-add it to the page queues.
4172 vm_page_release(vm_page_t m, int flags)
4176 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4177 ("vm_page_release: page %p is unmanaged", m));
4179 if ((flags & VPR_TRYFREE) != 0) {
4181 object = atomic_load_ptr(&m->object);
4184 /* Depends on type-stability. */
4185 if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object))
4187 if (object == m->object) {
4188 vm_page_release_locked(m, flags);
4189 VM_OBJECT_WUNLOCK(object);
4192 VM_OBJECT_WUNLOCK(object);
4195 vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0);
4198 /* See vm_page_release(). */
4200 vm_page_release_locked(vm_page_t m, int flags)
4203 VM_OBJECT_ASSERT_WLOCKED(m->object);
4204 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4205 ("vm_page_release_locked: page %p is unmanaged", m));
4207 if (vm_page_unwire_noq(m)) {
4208 if ((flags & VPR_TRYFREE) != 0 &&
4209 (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
4210 m->dirty == 0 && vm_page_tryxbusy(m)) {
4212 * An unlocked lookup may have wired the page before the
4213 * busy lock was acquired, in which case the page must
4216 if (__predict_true(!vm_page_wired(m))) {
4222 vm_page_release_toq(m, PQ_INACTIVE, flags != 0);
4228 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
4232 KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
4233 ("vm_page_try_blocked_op: page %p has no object", m));
4234 KASSERT(vm_page_busied(m),
4235 ("vm_page_try_blocked_op: page %p is not busy", m));
4236 VM_OBJECT_ASSERT_LOCKED(m->object);
4241 ("vm_page_try_blocked_op: page %p has no references", m));
4242 if (VPRC_WIRE_COUNT(old) != 0)
4244 } while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
4249 * If the object is read-locked, new wirings may be created via an
4252 old = vm_page_drop(m, VPRC_BLOCKED);
4253 KASSERT(!VM_OBJECT_WOWNED(m->object) ||
4254 old == (VPRC_BLOCKED | VPRC_OBJREF),
4255 ("vm_page_try_blocked_op: unexpected refcount value %u for %p",
4261 * Atomically check for wirings and remove all mappings of the page.
4264 vm_page_try_remove_all(vm_page_t m)
4267 return (vm_page_try_blocked_op(m, pmap_remove_all));
4271 * Atomically check for wirings and remove all writeable mappings of the page.
4274 vm_page_try_remove_write(vm_page_t m)
4277 return (vm_page_try_blocked_op(m, pmap_remove_write));
4283 * Apply the specified advice to the given page.
4286 vm_page_advise(vm_page_t m, int advice)
4289 VM_OBJECT_ASSERT_WLOCKED(m->object);
4290 vm_page_assert_xbusied(m);
4292 if (advice == MADV_FREE)
4294 * Mark the page clean. This will allow the page to be freed
4295 * without first paging it out. MADV_FREE pages are often
4296 * quickly reused by malloc(3), so we do not do anything that
4297 * would result in a page fault on a later access.
4300 else if (advice != MADV_DONTNEED) {
4301 if (advice == MADV_WILLNEED)
4302 vm_page_activate(m);
4306 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
4310 * Clear any references to the page. Otherwise, the page daemon will
4311 * immediately reactivate the page.
4313 vm_page_aflag_clear(m, PGA_REFERENCED);
4316 * Place clean pages near the head of the inactive queue rather than
4317 * the tail, thus defeating the queue's LRU operation and ensuring that
4318 * the page will be reused quickly. Dirty pages not already in the
4319 * laundry are moved there.
4322 vm_page_deactivate_noreuse(m);
4323 else if (!vm_page_in_laundry(m))
4328 * vm_page_grab_release
4330 * Helper routine for grab functions to release busy on return.
4333 vm_page_grab_release(vm_page_t m, int allocflags)
4336 if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4337 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4345 * vm_page_grab_sleep
4347 * Sleep for busy according to VM_ALLOC_ parameters. Returns true
4348 * if the caller should retry and false otherwise.
4350 * If the object is locked on entry the object will be unlocked with
4351 * false returns and still locked but possibly having been dropped
4352 * with true returns.
4355 vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex,
4356 const char *wmesg, int allocflags, bool locked)
4359 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4363 * Reference the page before unlocking and sleeping so that
4364 * the page daemon is less likely to reclaim it.
4366 if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0)
4367 vm_page_reference(m);
4369 if (_vm_page_busy_sleep(object, m, m->pindex, wmesg, allocflags,
4371 VM_OBJECT_WLOCK(object);
4372 if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
4379 * Assert that the grab flags are valid.
4382 vm_page_grab_check(int allocflags)
4385 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4386 (allocflags & VM_ALLOC_WIRED) != 0,
4387 ("vm_page_grab*: the pages must be busied or wired"));
4389 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4390 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4391 ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4395 * Calculate the page allocation flags for grab.
4398 vm_page_grab_pflags(int allocflags)
4402 pflags = allocflags &
4403 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL |
4405 if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4406 pflags |= VM_ALLOC_WAITFAIL;
4407 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4408 pflags |= VM_ALLOC_SBUSY;
4414 * Grab a page, waiting until we are waken up due to the page
4415 * changing state. We keep on waiting, if the page continues
4416 * to be in the object. If the page doesn't exist, first allocate it
4417 * and then conditionally zero it.
4419 * This routine may sleep.
4421 * The object must be locked on entry. The lock will, however, be released
4422 * and reacquired if the routine sleeps.
4425 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
4429 VM_OBJECT_ASSERT_WLOCKED(object);
4430 vm_page_grab_check(allocflags);
4433 if ((m = vm_page_lookup(object, pindex)) != NULL) {
4434 if (!vm_page_tryacquire(m, allocflags)) {
4435 if (vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4442 if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4444 m = vm_page_alloc(object, pindex, vm_page_grab_pflags(allocflags));
4446 if ((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0)
4450 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
4454 vm_page_grab_release(m, allocflags);
4460 * Locklessly attempt to acquire a page given a (object, pindex) tuple
4461 * and an optional previous page to avoid the radix lookup. The resulting
4462 * page will be validated against the identity tuple and busied or wired
4463 * as requested. A NULL *mp return guarantees that the page was not in
4464 * radix at the time of the call but callers must perform higher level
4465 * synchronization or retry the operation under a lock if they require
4466 * an atomic answer. This is the only lock free validation routine,
4467 * other routines can depend on the resulting page state.
4469 * The return value indicates whether the operation failed due to caller
4470 * flags. The return is tri-state with mp:
4472 * (true, *mp != NULL) - The operation was successful.
4473 * (true, *mp == NULL) - The page was not found in tree.
4474 * (false, *mp == NULL) - WAITFAIL or NOWAIT prevented acquisition.
4477 vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex,
4478 vm_page_t prev, vm_page_t *mp, int allocflags)
4482 vm_page_grab_check(allocflags);
4483 MPASS(prev == NULL || vm_page_busied(prev) || vm_page_wired(prev));
4488 * We may see a false NULL here because the previous page
4489 * has been removed or just inserted and the list is loaded
4490 * without barriers. Switch to radix to verify.
4492 if (prev == NULL || (m = TAILQ_NEXT(prev, listq)) == NULL ||
4493 QMD_IS_TRASHED(m) || m->pindex != pindex ||
4494 atomic_load_ptr(&m->object) != object) {
4497 * This guarantees the result is instantaneously
4500 m = vm_radix_lookup_unlocked(&object->rtree, pindex);
4504 if (vm_page_trybusy(m, allocflags)) {
4505 if (m->object == object && m->pindex == pindex)
4508 vm_page_busy_release(m);
4512 if (!vm_page_grab_sleep(object, m, pindex, "pgnslp",
4516 if ((allocflags & VM_ALLOC_WIRED) != 0)
4518 vm_page_grab_release(m, allocflags);
4524 * Try to locklessly grab a page and fall back to the object lock if NOCREAT
4528 vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags)
4532 vm_page_grab_check(allocflags);
4534 if (!vm_page_acquire_unlocked(object, pindex, NULL, &m, allocflags))
4540 * The radix lockless lookup should never return a false negative
4541 * errors. If the user specifies NOCREAT they are guaranteed there
4542 * was no page present at the instant of the call. A NOCREAT caller
4543 * must handle create races gracefully.
4545 if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4548 VM_OBJECT_WLOCK(object);
4549 m = vm_page_grab(object, pindex, allocflags);
4550 VM_OBJECT_WUNLOCK(object);
4556 * Grab a page and make it valid, paging in if necessary. Pages missing from
4557 * their pager are zero filled and validated. If a VM_ALLOC_COUNT is supplied
4558 * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought
4559 * in simultaneously. Additional pages will be left on a paging queue but
4560 * will neither be wired nor busy regardless of allocflags.
4563 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags)
4566 vm_page_t ma[VM_INITIAL_PAGEIN];
4567 int after, i, pflags, rv;
4569 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4570 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4571 ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4572 KASSERT((allocflags &
4573 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4574 ("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
4575 VM_OBJECT_ASSERT_WLOCKED(object);
4576 pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY |
4578 pflags |= VM_ALLOC_WAITFAIL;
4581 if ((m = vm_page_lookup(object, pindex)) != NULL) {
4583 * If the page is fully valid it can only become invalid
4584 * with the object lock held. If it is not valid it can
4585 * become valid with the busy lock held. Therefore, we
4586 * may unnecessarily lock the exclusive busy here if we
4587 * race with I/O completion not using the object lock.
4588 * However, we will not end up with an invalid page and a
4591 if (!vm_page_trybusy(m,
4592 vm_page_all_valid(m) ? allocflags : 0)) {
4593 (void)vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4597 if (vm_page_all_valid(m))
4599 if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4600 vm_page_busy_release(m);
4602 return (VM_PAGER_FAIL);
4604 } else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4606 return (VM_PAGER_FAIL);
4607 } else if ((m = vm_page_alloc(object, pindex, pflags)) == NULL) {
4611 vm_page_assert_xbusied(m);
4612 if (vm_pager_has_page(object, pindex, NULL, &after)) {
4613 after = MIN(after, VM_INITIAL_PAGEIN);
4614 after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT);
4615 after = MAX(after, 1);
4617 for (i = 1; i < after; i++) {
4618 if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) {
4619 if (ma[i]->valid || !vm_page_tryxbusy(ma[i]))
4622 ma[i] = vm_page_alloc(object, m->pindex + i,
4629 vm_object_pip_add(object, after);
4630 VM_OBJECT_WUNLOCK(object);
4631 rv = vm_pager_get_pages(object, ma, after, NULL, NULL);
4632 VM_OBJECT_WLOCK(object);
4633 vm_object_pip_wakeupn(object, after);
4634 /* Pager may have replaced a page. */
4636 if (rv != VM_PAGER_OK) {
4637 for (i = 0; i < after; i++) {
4638 if (!vm_page_wired(ma[i]))
4639 vm_page_free(ma[i]);
4641 vm_page_xunbusy(ma[i]);
4646 for (i = 1; i < after; i++)
4647 vm_page_readahead_finish(ma[i]);
4648 MPASS(vm_page_all_valid(m));
4650 vm_page_zero_invalid(m, TRUE);
4653 if ((allocflags & VM_ALLOC_WIRED) != 0)
4655 if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m))
4656 vm_page_busy_downgrade(m);
4657 else if ((allocflags & VM_ALLOC_NOBUSY) != 0)
4658 vm_page_busy_release(m);
4660 return (VM_PAGER_OK);
4664 * Locklessly grab a valid page. If the page is not valid or not yet
4665 * allocated this will fall back to the object lock method.
4668 vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object,
4669 vm_pindex_t pindex, int allocflags)
4675 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4676 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4677 ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY "
4679 KASSERT((allocflags &
4680 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4681 ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags));
4684 * Attempt a lockless lookup and busy. We need at least an sbusy
4685 * before we can inspect the valid field and return a wired page.
4687 flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED);
4688 if (!vm_page_acquire_unlocked(object, pindex, NULL, mp, flags))
4689 return (VM_PAGER_FAIL);
4690 if ((m = *mp) != NULL) {
4691 if (vm_page_all_valid(m)) {
4692 if ((allocflags & VM_ALLOC_WIRED) != 0)
4694 vm_page_grab_release(m, allocflags);
4695 return (VM_PAGER_OK);
4697 vm_page_busy_release(m);
4699 if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4701 return (VM_PAGER_FAIL);
4703 VM_OBJECT_WLOCK(object);
4704 error = vm_page_grab_valid(mp, object, pindex, allocflags);
4705 VM_OBJECT_WUNLOCK(object);
4711 * Return the specified range of pages from the given object. For each
4712 * page offset within the range, if a page already exists within the object
4713 * at that offset and it is busy, then wait for it to change state. If,
4714 * instead, the page doesn't exist, then allocate it.
4716 * The caller must always specify an allocation class.
4718 * allocation classes:
4719 * VM_ALLOC_NORMAL normal process request
4720 * VM_ALLOC_SYSTEM system *really* needs the pages
4722 * The caller must always specify that the pages are to be busied and/or
4725 * optional allocation flags:
4726 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages
4727 * VM_ALLOC_NOBUSY do not exclusive busy the page
4728 * VM_ALLOC_NOWAIT do not sleep
4729 * VM_ALLOC_SBUSY set page to sbusy state
4730 * VM_ALLOC_WIRED wire the pages
4731 * VM_ALLOC_ZERO zero and validate any invalid pages
4733 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it
4734 * may return a partial prefix of the requested range.
4737 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
4738 vm_page_t *ma, int count)
4744 VM_OBJECT_ASSERT_WLOCKED(object);
4745 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
4746 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
4748 ("vm_page_grab_pages: invalid page count %d", count));
4749 vm_page_grab_check(allocflags);
4751 pflags = vm_page_grab_pflags(allocflags);
4754 m = vm_radix_lookup_le(&object->rtree, pindex + i);
4755 if (m == NULL || m->pindex != pindex + i) {
4759 mpred = TAILQ_PREV(m, pglist, listq);
4760 for (; i < count; i++) {
4762 if (!vm_page_tryacquire(m, allocflags)) {
4763 if (vm_page_grab_sleep(object, m, pindex,
4764 "grbmaw", allocflags, true))
4769 if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4771 m = vm_page_alloc_after(object, pindex + i,
4772 pflags | VM_ALLOC_COUNT(count - i), mpred);
4774 if ((allocflags & (VM_ALLOC_NOWAIT |
4775 VM_ALLOC_WAITFAIL)) != 0)
4780 if (vm_page_none_valid(m) &&
4781 (allocflags & VM_ALLOC_ZERO) != 0) {
4782 if ((m->flags & PG_ZERO) == 0)
4786 vm_page_grab_release(m, allocflags);
4788 m = vm_page_next(m);
4794 * Unlocked variant of vm_page_grab_pages(). This accepts the same flags
4795 * and will fall back to the locked variant to handle allocation.
4798 vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex,
4799 int allocflags, vm_page_t *ma, int count)
4806 ("vm_page_grab_pages_unlocked: invalid page count %d", count));
4807 vm_page_grab_check(allocflags);
4810 * Modify flags for lockless acquire to hold the page until we
4811 * set it valid if necessary.
4813 flags = allocflags & ~VM_ALLOC_NOBUSY;
4815 for (i = 0; i < count; i++, pindex++) {
4816 if (!vm_page_acquire_unlocked(object, pindex, pred, &m, flags))
4820 if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) {
4821 if ((m->flags & PG_ZERO) == 0)
4825 /* m will still be wired or busy according to flags. */
4826 vm_page_grab_release(m, allocflags);
4829 if (i == count || (allocflags & VM_ALLOC_NOCREAT) != 0)
4832 VM_OBJECT_WLOCK(object);
4833 i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count);
4834 VM_OBJECT_WUNLOCK(object);
4840 * Mapping function for valid or dirty bits in a page.
4842 * Inputs are required to range within a page.
4845 vm_page_bits(int base, int size)
4851 base + size <= PAGE_SIZE,
4852 ("vm_page_bits: illegal base/size %d/%d", base, size)
4855 if (size == 0) /* handle degenerate case */
4858 first_bit = base >> DEV_BSHIFT;
4859 last_bit = (base + size - 1) >> DEV_BSHIFT;
4861 return (((vm_page_bits_t)2 << last_bit) -
4862 ((vm_page_bits_t)1 << first_bit));
4866 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set)
4869 #if PAGE_SIZE == 32768
4870 atomic_set_64((uint64_t *)bits, set);
4871 #elif PAGE_SIZE == 16384
4872 atomic_set_32((uint32_t *)bits, set);
4873 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16)
4874 atomic_set_16((uint16_t *)bits, set);
4875 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8)
4876 atomic_set_8((uint8_t *)bits, set);
4877 #else /* PAGE_SIZE <= 8192 */
4881 addr = (uintptr_t)bits;
4883 * Use a trick to perform a 32-bit atomic on the
4884 * containing aligned word, to not depend on the existence
4885 * of atomic_{set, clear}_{8, 16}.
4887 shift = addr & (sizeof(uint32_t) - 1);
4888 #if BYTE_ORDER == BIG_ENDIAN
4889 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4893 addr &= ~(sizeof(uint32_t) - 1);
4894 atomic_set_32((uint32_t *)addr, set << shift);
4895 #endif /* PAGE_SIZE */
4899 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear)
4902 #if PAGE_SIZE == 32768
4903 atomic_clear_64((uint64_t *)bits, clear);
4904 #elif PAGE_SIZE == 16384
4905 atomic_clear_32((uint32_t *)bits, clear);
4906 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16)
4907 atomic_clear_16((uint16_t *)bits, clear);
4908 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8)
4909 atomic_clear_8((uint8_t *)bits, clear);
4910 #else /* PAGE_SIZE <= 8192 */
4914 addr = (uintptr_t)bits;
4916 * Use a trick to perform a 32-bit atomic on the
4917 * containing aligned word, to not depend on the existence
4918 * of atomic_{set, clear}_{8, 16}.
4920 shift = addr & (sizeof(uint32_t) - 1);
4921 #if BYTE_ORDER == BIG_ENDIAN
4922 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4926 addr &= ~(sizeof(uint32_t) - 1);
4927 atomic_clear_32((uint32_t *)addr, clear << shift);
4928 #endif /* PAGE_SIZE */
4931 static inline vm_page_bits_t
4932 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits)
4934 #if PAGE_SIZE == 32768
4938 while (atomic_fcmpset_64(bits, &old, newbits) == 0);
4940 #elif PAGE_SIZE == 16384
4944 while (atomic_fcmpset_32(bits, &old, newbits) == 0);
4946 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16)
4950 while (atomic_fcmpset_16(bits, &old, newbits) == 0);
4952 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8)
4956 while (atomic_fcmpset_8(bits, &old, newbits) == 0);
4958 #else /* PAGE_SIZE <= 4096*/
4960 uint32_t old, new, mask;
4963 addr = (uintptr_t)bits;
4965 * Use a trick to perform a 32-bit atomic on the
4966 * containing aligned word, to not depend on the existence
4967 * of atomic_{set, swap, clear}_{8, 16}.
4969 shift = addr & (sizeof(uint32_t) - 1);
4970 #if BYTE_ORDER == BIG_ENDIAN
4971 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4975 addr &= ~(sizeof(uint32_t) - 1);
4976 mask = VM_PAGE_BITS_ALL << shift;
4981 new |= newbits << shift;
4982 } while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0);
4983 return (old >> shift);
4984 #endif /* PAGE_SIZE */
4988 * vm_page_set_valid_range:
4990 * Sets portions of a page valid. The arguments are expected
4991 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4992 * of any partial chunks touched by the range. The invalid portion of
4993 * such chunks will be zeroed.
4995 * (base + size) must be less then or equal to PAGE_SIZE.
4998 vm_page_set_valid_range(vm_page_t m, int base, int size)
5001 vm_page_bits_t pagebits;
5003 vm_page_assert_busied(m);
5004 if (size == 0) /* handle degenerate case */
5008 * If the base is not DEV_BSIZE aligned and the valid
5009 * bit is clear, we have to zero out a portion of the
5012 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5013 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
5014 pmap_zero_page_area(m, frag, base - frag);
5017 * If the ending offset is not DEV_BSIZE aligned and the
5018 * valid bit is clear, we have to zero out a portion of
5021 endoff = base + size;
5022 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5023 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
5024 pmap_zero_page_area(m, endoff,
5025 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5028 * Assert that no previously invalid block that is now being validated
5031 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
5032 ("vm_page_set_valid_range: page %p is dirty", m));
5035 * Set valid bits inclusive of any overlap.
5037 pagebits = vm_page_bits(base, size);
5038 if (vm_page_xbusied(m))
5039 m->valid |= pagebits;
5041 vm_page_bits_set(m, &m->valid, pagebits);
5045 * Set the page dirty bits and free the invalid swap space if
5046 * present. Returns the previous dirty bits.
5049 vm_page_set_dirty(vm_page_t m)
5053 VM_PAGE_OBJECT_BUSY_ASSERT(m);
5055 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) {
5057 m->dirty = VM_PAGE_BITS_ALL;
5059 old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL);
5060 if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0)
5061 vm_pager_page_unswapped(m);
5067 * Clear the given bits from the specified page's dirty field.
5069 static __inline void
5070 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
5073 vm_page_assert_busied(m);
5076 * If the page is xbusied and not write mapped we are the
5077 * only thread that can modify dirty bits. Otherwise, The pmap
5078 * layer can call vm_page_dirty() without holding a distinguished
5079 * lock. The combination of page busy and atomic operations
5080 * suffice to guarantee consistency of the page dirty field.
5082 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
5083 m->dirty &= ~pagebits;
5085 vm_page_bits_clear(m, &m->dirty, pagebits);
5089 * vm_page_set_validclean:
5091 * Sets portions of a page valid and clean. The arguments are expected
5092 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5093 * of any partial chunks touched by the range. The invalid portion of
5094 * such chunks will be zero'd.
5096 * (base + size) must be less then or equal to PAGE_SIZE.
5099 vm_page_set_validclean(vm_page_t m, int base, int size)
5101 vm_page_bits_t oldvalid, pagebits;
5104 vm_page_assert_busied(m);
5105 if (size == 0) /* handle degenerate case */
5109 * If the base is not DEV_BSIZE aligned and the valid
5110 * bit is clear, we have to zero out a portion of the
5113 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5114 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
5115 pmap_zero_page_area(m, frag, base - frag);
5118 * If the ending offset is not DEV_BSIZE aligned and the
5119 * valid bit is clear, we have to zero out a portion of
5122 endoff = base + size;
5123 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5124 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
5125 pmap_zero_page_area(m, endoff,
5126 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5129 * Set valid, clear dirty bits. If validating the entire
5130 * page we can safely clear the pmap modify bit. We also
5131 * use this opportunity to clear the PGA_NOSYNC flag. If a process
5132 * takes a write fault on a MAP_NOSYNC memory area the flag will
5135 * We set valid bits inclusive of any overlap, but we can only
5136 * clear dirty bits for DEV_BSIZE chunks that are fully within
5139 oldvalid = m->valid;
5140 pagebits = vm_page_bits(base, size);
5141 if (vm_page_xbusied(m))
5142 m->valid |= pagebits;
5144 vm_page_bits_set(m, &m->valid, pagebits);
5146 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
5147 frag = DEV_BSIZE - frag;
5153 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
5155 if (base == 0 && size == PAGE_SIZE) {
5157 * The page can only be modified within the pmap if it is
5158 * mapped, and it can only be mapped if it was previously
5161 if (oldvalid == VM_PAGE_BITS_ALL)
5163 * Perform the pmap_clear_modify() first. Otherwise,
5164 * a concurrent pmap operation, such as
5165 * pmap_protect(), could clear a modification in the
5166 * pmap and set the dirty field on the page before
5167 * pmap_clear_modify() had begun and after the dirty
5168 * field was cleared here.
5170 pmap_clear_modify(m);
5172 vm_page_aflag_clear(m, PGA_NOSYNC);
5173 } else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m))
5174 m->dirty &= ~pagebits;
5176 vm_page_clear_dirty_mask(m, pagebits);
5180 vm_page_clear_dirty(vm_page_t m, int base, int size)
5183 vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
5187 * vm_page_set_invalid:
5189 * Invalidates DEV_BSIZE'd chunks within a page. Both the
5190 * valid and dirty bits for the effected areas are cleared.
5193 vm_page_set_invalid(vm_page_t m, int base, int size)
5195 vm_page_bits_t bits;
5199 * The object lock is required so that pages can't be mapped
5200 * read-only while we're in the process of invalidating them.
5203 VM_OBJECT_ASSERT_WLOCKED(object);
5204 vm_page_assert_busied(m);
5206 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
5207 size >= object->un_pager.vnp.vnp_size)
5208 bits = VM_PAGE_BITS_ALL;
5210 bits = vm_page_bits(base, size);
5211 if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0)
5213 KASSERT((bits == 0 && vm_page_all_valid(m)) ||
5214 !pmap_page_is_mapped(m),
5215 ("vm_page_set_invalid: page %p is mapped", m));
5216 if (vm_page_xbusied(m)) {
5220 vm_page_bits_clear(m, &m->valid, bits);
5221 vm_page_bits_clear(m, &m->dirty, bits);
5228 * Invalidates the entire page. The page must be busy, unmapped, and
5229 * the enclosing object must be locked. The object locks protects
5230 * against concurrent read-only pmap enter which is done without
5234 vm_page_invalid(vm_page_t m)
5237 vm_page_assert_busied(m);
5238 VM_OBJECT_ASSERT_LOCKED(m->object);
5239 MPASS(!pmap_page_is_mapped(m));
5241 if (vm_page_xbusied(m))
5244 vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL);
5248 * vm_page_zero_invalid()
5250 * The kernel assumes that the invalid portions of a page contain
5251 * garbage, but such pages can be mapped into memory by user code.
5252 * When this occurs, we must zero out the non-valid portions of the
5253 * page so user code sees what it expects.
5255 * Pages are most often semi-valid when the end of a file is mapped
5256 * into memory and the file's size is not page aligned.
5259 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
5265 * Scan the valid bits looking for invalid sections that
5266 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the
5267 * valid bit may be set ) have already been zeroed by
5268 * vm_page_set_validclean().
5270 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
5271 if (i == (PAGE_SIZE / DEV_BSIZE) ||
5272 (m->valid & ((vm_page_bits_t)1 << i))) {
5274 pmap_zero_page_area(m,
5275 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
5282 * setvalid is TRUE when we can safely set the zero'd areas
5283 * as being valid. We can do this if there are no cache consistancy
5284 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
5293 * Is (partial) page valid? Note that the case where size == 0
5294 * will return FALSE in the degenerate case where the page is
5295 * entirely invalid, and TRUE otherwise.
5297 * Some callers envoke this routine without the busy lock held and
5298 * handle races via higher level locks. Typical callers should
5299 * hold a busy lock to prevent invalidation.
5302 vm_page_is_valid(vm_page_t m, int base, int size)
5304 vm_page_bits_t bits;
5306 bits = vm_page_bits(base, size);
5307 return (m->valid != 0 && (m->valid & bits) == bits);
5311 * Returns true if all of the specified predicates are true for the entire
5312 * (super)page and false otherwise.
5315 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
5321 if (skip_m != NULL && skip_m->object != object)
5323 VM_OBJECT_ASSERT_LOCKED(object);
5324 npages = atop(pagesizes[m->psind]);
5327 * The physically contiguous pages that make up a superpage, i.e., a
5328 * page with a page size index ("psind") greater than zero, will
5329 * occupy adjacent entries in vm_page_array[].
5331 for (i = 0; i < npages; i++) {
5332 /* Always test object consistency, including "skip_m". */
5333 if (m[i].object != object)
5335 if (&m[i] == skip_m)
5337 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
5339 if ((flags & PS_ALL_DIRTY) != 0) {
5341 * Calling vm_page_test_dirty() or pmap_is_modified()
5342 * might stop this case from spuriously returning
5343 * "false". However, that would require a write lock
5344 * on the object containing "m[i]".
5346 if (m[i].dirty != VM_PAGE_BITS_ALL)
5349 if ((flags & PS_ALL_VALID) != 0 &&
5350 m[i].valid != VM_PAGE_BITS_ALL)
5357 * Set the page's dirty bits if the page is modified.
5360 vm_page_test_dirty(vm_page_t m)
5363 vm_page_assert_busied(m);
5364 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
5369 vm_page_valid(vm_page_t m)
5372 vm_page_assert_busied(m);
5373 if (vm_page_xbusied(m))
5374 m->valid = VM_PAGE_BITS_ALL;
5376 vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL);
5380 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
5383 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
5387 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
5390 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
5394 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
5397 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
5400 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
5402 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
5405 vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
5409 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
5412 mtx_assert_(vm_page_lockptr(m), a, file, line);
5418 vm_page_object_busy_assert(vm_page_t m)
5422 * Certain of the page's fields may only be modified by the
5423 * holder of a page or object busy.
5425 if (m->object != NULL && !vm_page_busied(m))
5426 VM_OBJECT_ASSERT_BUSY(m->object);
5430 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits)
5433 if ((bits & PGA_WRITEABLE) == 0)
5437 * The PGA_WRITEABLE flag can only be set if the page is
5438 * managed, is exclusively busied or the object is locked.
5439 * Currently, this flag is only set by pmap_enter().
5441 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5442 ("PGA_WRITEABLE on unmanaged page"));
5443 if (!vm_page_xbusied(m))
5444 VM_OBJECT_ASSERT_BUSY(m->object);
5448 #include "opt_ddb.h"
5450 #include <sys/kernel.h>
5452 #include <ddb/ddb.h>
5454 DB_SHOW_COMMAND(page, vm_page_print_page_info)
5457 db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
5458 db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
5459 db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
5460 db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
5461 db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
5462 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
5463 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
5464 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
5465 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
5468 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
5472 db_printf("pq_free %d\n", vm_free_count());
5473 for (dom = 0; dom < vm_ndomains; dom++) {
5475 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
5477 vm_dom[dom].vmd_page_count,
5478 vm_dom[dom].vmd_free_count,
5479 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
5480 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
5481 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
5482 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
5486 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
5489 boolean_t phys, virt;
5492 db_printf("show pginfo addr\n");
5496 phys = strchr(modif, 'p') != NULL;
5497 virt = strchr(modif, 'v') != NULL;
5499 m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
5501 m = PHYS_TO_VM_PAGE(addr);
5503 m = (vm_page_t)addr;
5505 "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n"
5506 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
5507 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
5508 m->a.queue, m->ref_count, m->a.flags, m->oflags,
5509 m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty);