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 extern int uma_startup_count(int);
117 extern void uma_startup(void *, int);
118 extern int vmem_startup_count(void);
120 struct vm_domain vm_dom[MAXMEMDOM];
122 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]);
124 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
126 struct mtx_padalign __exclusive_cache_line vm_domainset_lock;
127 /* The following fields are protected by the domainset lock. */
128 domainset_t __exclusive_cache_line vm_min_domains;
129 domainset_t __exclusive_cache_line vm_severe_domains;
130 static int vm_min_waiters;
131 static int vm_severe_waiters;
132 static int vm_pageproc_waiters;
134 static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD, 0,
135 "VM page statistics");
137 static counter_u64_t pqstate_commit_retries = EARLY_COUNTER;
138 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries,
139 CTLFLAG_RD, &pqstate_commit_retries,
140 "Number of failed per-page atomic queue state updates");
142 static counter_u64_t queue_ops = EARLY_COUNTER;
143 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops,
144 CTLFLAG_RD, &queue_ops,
145 "Number of batched queue operations");
147 static counter_u64_t queue_nops = EARLY_COUNTER;
148 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops,
149 CTLFLAG_RD, &queue_nops,
150 "Number of batched queue operations with no effects");
153 counter_startup(void)
156 pqstate_commit_retries = counter_u64_alloc(M_WAITOK);
157 queue_ops = counter_u64_alloc(M_WAITOK);
158 queue_nops = counter_u64_alloc(M_WAITOK);
160 SYSINIT(page_counters, SI_SUB_CPU, SI_ORDER_ANY, counter_startup, NULL);
163 * bogus page -- for I/O to/from partially complete buffers,
164 * or for paging into sparsely invalid regions.
166 vm_page_t bogus_page;
168 vm_page_t vm_page_array;
169 long vm_page_array_size;
172 static int boot_pages;
173 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
175 "number of pages allocated for bootstrapping the VM system");
177 static TAILQ_HEAD(, vm_page) blacklist_head;
178 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
179 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
180 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
182 static uma_zone_t fakepg_zone;
184 static void vm_page_alloc_check(vm_page_t m);
185 static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m,
186 const char *wmesg, bool nonshared, bool locked);
187 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
188 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
189 static bool vm_page_free_prep(vm_page_t m);
190 static void vm_page_free_toq(vm_page_t m);
191 static void vm_page_init(void *dummy);
192 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
193 vm_pindex_t pindex, vm_page_t mpred);
194 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
196 static void vm_page_mvqueue(vm_page_t m, const uint8_t queue,
197 const uint16_t nflag);
198 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
199 vm_page_t m_run, vm_paddr_t high);
200 static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse);
201 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
203 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain,
205 static void vm_page_zone_release(void *arg, void **store, int cnt);
207 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
210 vm_page_init(void *dummy)
213 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
214 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
215 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
216 VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
220 * The cache page zone is initialized later since we need to be able to allocate
221 * pages before UMA is fully initialized.
224 vm_page_init_cache_zones(void *dummy __unused)
226 struct vm_domain *vmd;
227 struct vm_pgcache *pgcache;
228 int cache, domain, maxcache, pool;
231 TUNABLE_INT_FETCH("vm.pgcache_zone_max", &maxcache);
232 for (domain = 0; domain < vm_ndomains; domain++) {
233 vmd = VM_DOMAIN(domain);
234 for (pool = 0; pool < VM_NFREEPOOL; pool++) {
235 pgcache = &vmd->vmd_pgcache[pool];
236 pgcache->domain = domain;
237 pgcache->pool = pool;
238 pgcache->zone = uma_zcache_create("vm pgcache",
239 PAGE_SIZE, NULL, NULL, NULL, NULL,
240 vm_page_zone_import, vm_page_zone_release, pgcache,
244 * Limit each pool's zone to 0.1% of the pages in the
247 cache = maxcache != 0 ? maxcache :
248 vmd->vmd_page_count / 1000;
249 uma_zone_set_maxcache(pgcache->zone, cache);
253 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
255 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
256 #if PAGE_SIZE == 32768
258 CTASSERT(sizeof(u_long) >= 8);
265 * Sets the page size, perhaps based upon the memory
266 * size. Must be called before any use of page-size
267 * dependent functions.
270 vm_set_page_size(void)
272 if (vm_cnt.v_page_size == 0)
273 vm_cnt.v_page_size = PAGE_SIZE;
274 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
275 panic("vm_set_page_size: page size not a power of two");
279 * vm_page_blacklist_next:
281 * Find the next entry in the provided string of blacklist
282 * addresses. Entries are separated by space, comma, or newline.
283 * If an invalid integer is encountered then the rest of the
284 * string is skipped. Updates the list pointer to the next
285 * character, or NULL if the string is exhausted or invalid.
288 vm_page_blacklist_next(char **list, char *end)
293 if (list == NULL || *list == NULL)
301 * If there's no end pointer then the buffer is coming from
302 * the kenv and we know it's null-terminated.
305 end = *list + strlen(*list);
307 /* Ensure that strtoq() won't walk off the end */
309 if (*end == '\n' || *end == ' ' || *end == ',')
312 printf("Blacklist not terminated, skipping\n");
318 for (pos = *list; *pos != '\0'; pos = cp) {
319 bad = strtoq(pos, &cp, 0);
320 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
329 if (*cp == '\0' || ++cp >= end)
333 return (trunc_page(bad));
335 printf("Garbage in RAM blacklist, skipping\n");
341 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
343 struct vm_domain *vmd;
347 m = vm_phys_paddr_to_vm_page(pa);
349 return (true); /* page does not exist, no failure */
351 vmd = vm_pagequeue_domain(m);
352 vm_domain_free_lock(vmd);
353 ret = vm_phys_unfree_page(m);
354 vm_domain_free_unlock(vmd);
356 vm_domain_freecnt_inc(vmd, -1);
357 TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
359 printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
365 * vm_page_blacklist_check:
367 * Iterate through the provided string of blacklist addresses, pulling
368 * each entry out of the physical allocator free list and putting it
369 * onto a list for reporting via the vm.page_blacklist sysctl.
372 vm_page_blacklist_check(char *list, char *end)
378 while (next != NULL) {
379 if ((pa = vm_page_blacklist_next(&next, end)) == 0)
381 vm_page_blacklist_add(pa, bootverbose);
386 * vm_page_blacklist_load:
388 * Search for a special module named "ram_blacklist". It'll be a
389 * plain text file provided by the user via the loader directive
393 vm_page_blacklist_load(char **list, char **end)
402 mod = preload_search_by_type("ram_blacklist");
404 ptr = preload_fetch_addr(mod);
405 len = preload_fetch_size(mod);
416 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
423 error = sysctl_wire_old_buffer(req, 0);
426 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
427 TAILQ_FOREACH(m, &blacklist_head, listq) {
428 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
429 (uintmax_t)m->phys_addr);
432 error = sbuf_finish(&sbuf);
438 * Initialize a dummy page for use in scans of the specified paging queue.
439 * In principle, this function only needs to set the flag PG_MARKER.
440 * Nonetheless, it write busies the page as a safety precaution.
443 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags)
446 bzero(marker, sizeof(*marker));
447 marker->flags = PG_MARKER;
448 marker->a.flags = aflags;
449 marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
450 marker->a.queue = queue;
454 vm_page_domain_init(int domain)
456 struct vm_domain *vmd;
457 struct vm_pagequeue *pq;
460 vmd = VM_DOMAIN(domain);
461 bzero(vmd, sizeof(*vmd));
462 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
463 "vm inactive pagequeue";
464 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
465 "vm active pagequeue";
466 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
467 "vm laundry pagequeue";
468 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
469 "vm unswappable pagequeue";
470 vmd->vmd_domain = domain;
471 vmd->vmd_page_count = 0;
472 vmd->vmd_free_count = 0;
474 vmd->vmd_oom = FALSE;
475 for (i = 0; i < PQ_COUNT; i++) {
476 pq = &vmd->vmd_pagequeues[i];
477 TAILQ_INIT(&pq->pq_pl);
478 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
479 MTX_DEF | MTX_DUPOK);
481 vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
483 mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
484 mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
485 snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
488 * inacthead is used to provide FIFO ordering for LRU-bypassing
491 vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
492 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
493 &vmd->vmd_inacthead, plinks.q);
496 * The clock pages are used to implement active queue scanning without
497 * requeues. Scans start at clock[0], which is advanced after the scan
498 * ends. When the two clock hands meet, they are reset and scanning
499 * resumes from the head of the queue.
501 vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
502 vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
503 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
504 &vmd->vmd_clock[0], plinks.q);
505 TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
506 &vmd->vmd_clock[1], plinks.q);
510 * Initialize a physical page in preparation for adding it to the free
514 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
519 m->busy_lock = VPB_UNBUSIED;
520 m->flags = m->a.flags = 0;
522 m->a.queue = PQ_NONE;
525 m->order = VM_NFREEORDER;
526 m->pool = VM_FREEPOOL_DEFAULT;
527 m->valid = m->dirty = 0;
531 #ifndef PMAP_HAS_PAGE_ARRAY
533 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range)
538 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
539 * However, because this page is allocated from KVM, out-of-bounds
540 * accesses using the direct map will not be trapped.
545 * Allocate physical memory for the page structures, and map it.
547 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
548 vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end,
549 VM_PROT_READ | VM_PROT_WRITE);
550 vm_page_array_size = page_range;
559 * Initializes the resident memory module. Allocates physical memory for
560 * bootstrapping UMA and some data structures that are used to manage
561 * physical pages. Initializes these structures, and populates the free
565 vm_page_startup(vm_offset_t vaddr)
567 struct vm_phys_seg *seg;
569 char *list, *listend;
571 vm_paddr_t end, high_avail, low_avail, new_end, size;
572 vm_paddr_t page_range __unused;
573 vm_paddr_t last_pa, pa;
575 int biggestone, i, segind;
579 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
583 vaddr = round_page(vaddr);
585 vm_phys_early_startup();
586 biggestone = vm_phys_avail_largest();
587 end = phys_avail[biggestone+1];
590 * Initialize the page and queue locks.
592 mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
593 for (i = 0; i < PA_LOCK_COUNT; i++)
594 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
595 for (i = 0; i < vm_ndomains; i++)
596 vm_page_domain_init(i);
599 * Allocate memory for use when boot strapping the kernel memory
600 * allocator. Tell UMA how many zones we are going to create
601 * before going fully functional. UMA will add its zones.
603 * VM startup zones: vmem, vmem_btag, VM OBJECT, RADIX NODE, MAP,
604 * KMAP ENTRY, MAP ENTRY, VMSPACE.
606 boot_pages = uma_startup_count(8);
608 #ifndef UMA_MD_SMALL_ALLOC
609 /* vmem_startup() calls uma_prealloc(). */
610 boot_pages += vmem_startup_count();
611 /* vm_map_startup() calls uma_prealloc(). */
612 boot_pages += howmany(MAX_KMAP,
613 slab_ipers(sizeof(struct vm_map), UMA_ALIGN_PTR));
616 * Before going fully functional kmem_init() does allocation
617 * from "KMAP ENTRY" and vmem_create() does allocation from "vmem".
622 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
623 * manually fetch the value.
625 TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
626 new_end = end - (boot_pages * UMA_SLAB_SIZE);
627 new_end = trunc_page(new_end);
628 mapped = pmap_map(&vaddr, new_end, end,
629 VM_PROT_READ | VM_PROT_WRITE);
630 bzero((void *)mapped, end - new_end);
631 uma_startup((void *)mapped, boot_pages);
634 witness_size = round_page(witness_startup_count());
635 new_end -= witness_size;
636 mapped = pmap_map(&vaddr, new_end, new_end + witness_size,
637 VM_PROT_READ | VM_PROT_WRITE);
638 bzero((void *)mapped, witness_size);
639 witness_startup((void *)mapped);
642 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
643 defined(__i386__) || defined(__mips__) || defined(__riscv) || \
644 defined(__powerpc64__)
646 * Allocate a bitmap to indicate that a random physical page
647 * needs to be included in a minidump.
649 * The amd64 port needs this to indicate which direct map pages
650 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
652 * However, i386 still needs this workspace internally within the
653 * minidump code. In theory, they are not needed on i386, but are
654 * included should the sf_buf code decide to use them.
657 for (i = 0; dump_avail[i + 1] != 0; i += 2)
658 if (dump_avail[i + 1] > last_pa)
659 last_pa = dump_avail[i + 1];
660 page_range = last_pa / PAGE_SIZE;
661 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
662 new_end -= vm_page_dump_size;
663 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
664 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
665 bzero((void *)vm_page_dump, vm_page_dump_size);
669 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
670 defined(__riscv) || defined(__powerpc64__)
672 * Include the UMA bootstrap pages, witness pages and vm_page_dump
673 * in a crash dump. When pmap_map() uses the direct map, they are
674 * not automatically included.
676 for (pa = new_end; pa < end; pa += PAGE_SIZE)
679 phys_avail[biggestone + 1] = new_end;
682 * Request that the physical pages underlying the message buffer be
683 * included in a crash dump. Since the message buffer is accessed
684 * through the direct map, they are not automatically included.
686 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
687 last_pa = pa + round_page(msgbufsize);
688 while (pa < last_pa) {
694 * Compute the number of pages of memory that will be available for
695 * use, taking into account the overhead of a page structure per page.
696 * In other words, solve
697 * "available physical memory" - round_page(page_range *
698 * sizeof(struct vm_page)) = page_range * PAGE_SIZE
701 low_avail = phys_avail[0];
702 high_avail = phys_avail[1];
703 for (i = 0; i < vm_phys_nsegs; i++) {
704 if (vm_phys_segs[i].start < low_avail)
705 low_avail = vm_phys_segs[i].start;
706 if (vm_phys_segs[i].end > high_avail)
707 high_avail = vm_phys_segs[i].end;
709 /* Skip the first chunk. It is already accounted for. */
710 for (i = 2; phys_avail[i + 1] != 0; i += 2) {
711 if (phys_avail[i] < low_avail)
712 low_avail = phys_avail[i];
713 if (phys_avail[i + 1] > high_avail)
714 high_avail = phys_avail[i + 1];
716 first_page = low_avail / PAGE_SIZE;
717 #ifdef VM_PHYSSEG_SPARSE
719 for (i = 0; i < vm_phys_nsegs; i++)
720 size += vm_phys_segs[i].end - vm_phys_segs[i].start;
721 for (i = 0; phys_avail[i + 1] != 0; i += 2)
722 size += phys_avail[i + 1] - phys_avail[i];
723 #elif defined(VM_PHYSSEG_DENSE)
724 size = high_avail - low_avail;
726 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
729 #ifdef PMAP_HAS_PAGE_ARRAY
730 pmap_page_array_startup(size / PAGE_SIZE);
731 biggestone = vm_phys_avail_largest();
732 end = new_end = phys_avail[biggestone + 1];
734 #ifdef VM_PHYSSEG_DENSE
736 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
737 * the overhead of a page structure per page only if vm_page_array is
738 * allocated from the last physical memory chunk. Otherwise, we must
739 * allocate page structures representing the physical memory
740 * underlying vm_page_array, even though they will not be used.
742 if (new_end != high_avail)
743 page_range = size / PAGE_SIZE;
747 page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
750 * If the partial bytes remaining are large enough for
751 * a page (PAGE_SIZE) without a corresponding
752 * 'struct vm_page', then new_end will contain an
753 * extra page after subtracting the length of the VM
754 * page array. Compensate by subtracting an extra
757 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
758 if (new_end == high_avail)
759 high_avail -= PAGE_SIZE;
760 new_end -= PAGE_SIZE;
764 new_end = vm_page_array_alloc(&vaddr, end, page_range);
767 #if VM_NRESERVLEVEL > 0
769 * Allocate physical memory for the reservation management system's
770 * data structures, and map it.
772 new_end = vm_reserv_startup(&vaddr, new_end);
774 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
775 defined(__riscv) || defined(__powerpc64__)
777 * Include vm_page_array and vm_reserv_array in a crash dump.
779 for (pa = new_end; pa < end; pa += PAGE_SIZE)
782 phys_avail[biggestone + 1] = new_end;
785 * Add physical memory segments corresponding to the available
788 for (i = 0; phys_avail[i + 1] != 0; i += 2)
789 if (vm_phys_avail_size(i) != 0)
790 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
793 * Initialize the physical memory allocator.
798 * Initialize the page structures and add every available page to the
799 * physical memory allocator's free lists.
801 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
802 for (ii = 0; ii < vm_page_array_size; ii++) {
803 m = &vm_page_array[ii];
804 vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0);
805 m->flags = PG_FICTITIOUS;
808 vm_cnt.v_page_count = 0;
809 for (segind = 0; segind < vm_phys_nsegs; segind++) {
810 seg = &vm_phys_segs[segind];
811 for (m = seg->first_page, pa = seg->start; pa < seg->end;
812 m++, pa += PAGE_SIZE)
813 vm_page_init_page(m, pa, segind);
816 * Add the segment to the free lists only if it is covered by
817 * one of the ranges in phys_avail. Because we've added the
818 * ranges to the vm_phys_segs array, we can assume that each
819 * segment is either entirely contained in one of the ranges,
820 * or doesn't overlap any of them.
822 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
823 struct vm_domain *vmd;
825 if (seg->start < phys_avail[i] ||
826 seg->end > phys_avail[i + 1])
830 pagecount = (u_long)atop(seg->end - seg->start);
832 vmd = VM_DOMAIN(seg->domain);
833 vm_domain_free_lock(vmd);
834 vm_phys_enqueue_contig(m, pagecount);
835 vm_domain_free_unlock(vmd);
836 vm_domain_freecnt_inc(vmd, pagecount);
837 vm_cnt.v_page_count += (u_int)pagecount;
839 vmd = VM_DOMAIN(seg->domain);
840 vmd->vmd_page_count += (u_int)pagecount;
841 vmd->vmd_segs |= 1UL << m->segind;
847 * Remove blacklisted pages from the physical memory allocator.
849 TAILQ_INIT(&blacklist_head);
850 vm_page_blacklist_load(&list, &listend);
851 vm_page_blacklist_check(list, listend);
853 list = kern_getenv("vm.blacklist");
854 vm_page_blacklist_check(list, NULL);
857 #if VM_NRESERVLEVEL > 0
859 * Initialize the reservation management system.
868 vm_page_reference(vm_page_t m)
871 vm_page_aflag_set(m, PGA_REFERENCED);
875 vm_page_acquire_flags(vm_page_t m, int allocflags)
879 if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0)
880 locked = vm_page_trysbusy(m);
882 locked = vm_page_tryxbusy(m);
883 if (locked && (allocflags & VM_ALLOC_WIRED) != 0)
889 * vm_page_busy_sleep_flags
891 * Sleep for busy according to VM_ALLOC_ parameters.
894 vm_page_busy_sleep_flags(vm_object_t object, vm_page_t m, const char *wmesg,
898 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
901 * Reference the page before unlocking and
902 * sleeping so that the page daemon is less
903 * likely to reclaim it.
905 if ((allocflags & VM_ALLOC_NOCREAT) == 0)
906 vm_page_aflag_set(m, PGA_REFERENCED);
907 if (_vm_page_busy_sleep(object, m, wmesg, (allocflags &
908 VM_ALLOC_IGN_SBUSY) != 0, true))
909 VM_OBJECT_WLOCK(object);
910 if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
916 * vm_page_busy_acquire:
918 * Acquire the busy lock as described by VM_ALLOC_* flags. Will loop
919 * and drop the object lock if necessary.
922 vm_page_busy_acquire(vm_page_t m, int allocflags)
928 * The page-specific object must be cached because page
929 * identity can change during the sleep, causing the
930 * re-lock of a different object.
931 * It is assumed that a reference to the object is already
932 * held by the callers.
936 if (vm_page_acquire_flags(m, allocflags))
938 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
941 locked = VM_OBJECT_WOWNED(obj);
944 MPASS(locked || vm_page_wired(m));
945 if (_vm_page_busy_sleep(obj, m, "vmpba",
946 (allocflags & VM_ALLOC_SBUSY) != 0, locked))
947 VM_OBJECT_WLOCK(obj);
948 if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
950 KASSERT(m->object == obj || m->object == NULL,
951 ("vm_page_busy_acquire: page %p does not belong to %p",
957 * vm_page_busy_downgrade:
959 * Downgrade an exclusive busy page into a single shared busy page.
962 vm_page_busy_downgrade(vm_page_t m)
966 vm_page_assert_xbusied(m);
970 if (atomic_fcmpset_rel_int(&m->busy_lock,
971 &x, VPB_SHARERS_WORD(1)))
974 if ((x & VPB_BIT_WAITERS) != 0)
980 * vm_page_busy_tryupgrade:
982 * Attempt to upgrade a single shared busy into an exclusive busy.
985 vm_page_busy_tryupgrade(vm_page_t m)
989 vm_page_assert_sbusied(m);
992 ce = VPB_CURTHREAD_EXCLUSIVE;
994 if (VPB_SHARERS(x) > 1)
996 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
997 ("vm_page_busy_tryupgrade: invalid lock state"));
998 if (!atomic_fcmpset_acq_int(&m->busy_lock, &x,
999 ce | (x & VPB_BIT_WAITERS)))
1008 * Return a positive value if the page is shared busied, 0 otherwise.
1011 vm_page_sbusied(vm_page_t m)
1016 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
1022 * Shared unbusy a page.
1025 vm_page_sunbusy(vm_page_t m)
1029 vm_page_assert_sbusied(m);
1033 if (VPB_SHARERS(x) > 1) {
1034 if (atomic_fcmpset_int(&m->busy_lock, &x,
1035 x - VPB_ONE_SHARER))
1039 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
1040 ("vm_page_sunbusy: invalid lock state"));
1041 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1043 if ((x & VPB_BIT_WAITERS) == 0)
1051 * vm_page_busy_sleep:
1053 * Sleep if the page is busy, using the page pointer as wchan.
1054 * This is used to implement the hard-path of busying mechanism.
1056 * If nonshared is true, sleep only if the page is xbusy.
1058 * The object lock must be held on entry and will be released on exit.
1061 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
1066 VM_OBJECT_ASSERT_LOCKED(obj);
1067 vm_page_lock_assert(m, MA_NOTOWNED);
1069 if (!_vm_page_busy_sleep(obj, m, wmesg, nonshared, true))
1070 VM_OBJECT_DROP(obj);
1074 * _vm_page_busy_sleep:
1076 * Internal busy sleep function.
1079 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, const char *wmesg,
1080 bool nonshared, bool locked)
1085 * If the object is busy we must wait for that to drain to zero
1086 * before trying the page again.
1088 if (obj != NULL && vm_object_busied(obj)) {
1090 VM_OBJECT_DROP(obj);
1091 vm_object_busy_wait(obj, wmesg);
1096 if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) ||
1097 ((x & VPB_BIT_WAITERS) == 0 &&
1098 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) {
1103 VM_OBJECT_DROP(obj);
1105 sleepq_add(m, NULL, wmesg, 0, 0);
1106 sleepq_wait(m, PVM);
1114 * Try to shared busy a page.
1115 * If the operation succeeds 1 is returned otherwise 0.
1116 * The operation never sleeps.
1119 vm_page_trysbusy(vm_page_t m)
1127 if ((x & VPB_BIT_SHARED) == 0)
1130 * Reduce the window for transient busies that will trigger
1131 * false negatives in vm_page_ps_test().
1133 if (obj != NULL && vm_object_busied(obj))
1135 if (atomic_fcmpset_acq_int(&m->busy_lock, &x,
1136 x + VPB_ONE_SHARER))
1140 /* Refetch the object now that we're guaranteed that it is stable. */
1142 if (obj != NULL && vm_object_busied(obj)) {
1152 * Try to exclusive busy a page.
1153 * If the operation succeeds 1 is returned otherwise 0.
1154 * The operation never sleeps.
1157 vm_page_tryxbusy(vm_page_t m)
1161 if (atomic_cmpset_acq_int(&(m)->busy_lock, VPB_UNBUSIED,
1162 VPB_CURTHREAD_EXCLUSIVE) == 0)
1166 if (obj != NULL && vm_object_busied(obj)) {
1174 vm_page_xunbusy_hard_tail(vm_page_t m)
1176 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1177 /* Wake the waiter. */
1182 * vm_page_xunbusy_hard:
1184 * Called when unbusy has failed because there is a waiter.
1187 vm_page_xunbusy_hard(vm_page_t m)
1189 vm_page_assert_xbusied(m);
1190 vm_page_xunbusy_hard_tail(m);
1194 vm_page_xunbusy_hard_unchecked(vm_page_t m)
1196 vm_page_assert_xbusied_unchecked(m);
1197 vm_page_xunbusy_hard_tail(m);
1201 * Avoid releasing and reacquiring the same page lock.
1204 vm_page_change_lock(vm_page_t m, struct mtx **mtx)
1208 mtx1 = vm_page_lockptr(m);
1218 * vm_page_unhold_pages:
1220 * Unhold each of the pages that is referenced by the given array.
1223 vm_page_unhold_pages(vm_page_t *ma, int count)
1226 for (; count != 0; count--) {
1227 vm_page_unwire(*ma, PQ_ACTIVE);
1233 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1237 #ifdef VM_PHYSSEG_SPARSE
1238 m = vm_phys_paddr_to_vm_page(pa);
1240 m = vm_phys_fictitious_to_vm_page(pa);
1242 #elif defined(VM_PHYSSEG_DENSE)
1246 if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1247 m = &vm_page_array[pi - first_page];
1250 return (vm_phys_fictitious_to_vm_page(pa));
1252 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1259 * Create a fictitious page with the specified physical address and
1260 * memory attribute. The memory attribute is the only the machine-
1261 * dependent aspect of a fictitious page that must be initialized.
1264 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1268 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1269 vm_page_initfake(m, paddr, memattr);
1274 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1277 if ((m->flags & PG_FICTITIOUS) != 0) {
1279 * The page's memattr might have changed since the
1280 * previous initialization. Update the pmap to the
1285 m->phys_addr = paddr;
1286 m->a.queue = PQ_NONE;
1287 /* Fictitious pages don't use "segind". */
1288 m->flags = PG_FICTITIOUS;
1289 /* Fictitious pages don't use "order" or "pool". */
1290 m->oflags = VPO_UNMANAGED;
1291 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
1292 /* Fictitious pages are unevictable. */
1296 pmap_page_set_memattr(m, memattr);
1302 * Release a fictitious page.
1305 vm_page_putfake(vm_page_t m)
1308 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1309 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1310 ("vm_page_putfake: bad page %p", m));
1312 uma_zfree(fakepg_zone, m);
1316 * vm_page_updatefake:
1318 * Update the given fictitious page to the specified physical address and
1322 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1325 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1326 ("vm_page_updatefake: bad page %p", m));
1327 m->phys_addr = paddr;
1328 pmap_page_set_memattr(m, memattr);
1337 vm_page_free(vm_page_t m)
1340 m->flags &= ~PG_ZERO;
1341 vm_page_free_toq(m);
1345 * vm_page_free_zero:
1347 * Free a page to the zerod-pages queue
1350 vm_page_free_zero(vm_page_t m)
1353 m->flags |= PG_ZERO;
1354 vm_page_free_toq(m);
1358 * Unbusy and handle the page queueing for a page from a getpages request that
1359 * was optionally read ahead or behind.
1362 vm_page_readahead_finish(vm_page_t m)
1365 /* We shouldn't put invalid pages on queues. */
1366 KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m));
1369 * Since the page is not the actually needed one, whether it should
1370 * be activated or deactivated is not obvious. Empirical results
1371 * have shown that deactivating the page is usually the best choice,
1372 * unless the page is wanted by another thread.
1374 if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1375 vm_page_activate(m);
1377 vm_page_deactivate(m);
1378 vm_page_xunbusy_unchecked(m);
1382 * vm_page_sleep_if_busy:
1384 * Sleep and release the object lock if the page is busied.
1385 * Returns TRUE if the thread slept.
1387 * The given page must be unlocked and object containing it must
1391 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1395 vm_page_lock_assert(m, MA_NOTOWNED);
1396 VM_OBJECT_ASSERT_WLOCKED(m->object);
1399 * The page-specific object must be cached because page
1400 * identity can change during the sleep, causing the
1401 * re-lock of a different object.
1402 * It is assumed that a reference to the object is already
1403 * held by the callers.
1406 if (vm_page_busied(m) || (obj != NULL && obj->busy)) {
1407 vm_page_busy_sleep(m, msg, false);
1408 VM_OBJECT_WLOCK(obj);
1415 * vm_page_sleep_if_xbusy:
1417 * Sleep and release the object lock if the page is xbusied.
1418 * Returns TRUE if the thread slept.
1420 * The given page must be unlocked and object containing it must
1424 vm_page_sleep_if_xbusy(vm_page_t m, const char *msg)
1428 vm_page_lock_assert(m, MA_NOTOWNED);
1429 VM_OBJECT_ASSERT_WLOCKED(m->object);
1432 * The page-specific object must be cached because page
1433 * identity can change during the sleep, causing the
1434 * re-lock of a different object.
1435 * It is assumed that a reference to the object is already
1436 * held by the callers.
1439 if (vm_page_xbusied(m) || (obj != NULL && obj->busy)) {
1440 vm_page_busy_sleep(m, msg, true);
1441 VM_OBJECT_WLOCK(obj);
1448 * vm_page_dirty_KBI: [ internal use only ]
1450 * Set all bits in the page's dirty field.
1452 * The object containing the specified page must be locked if the
1453 * call is made from the machine-independent layer.
1455 * See vm_page_clear_dirty_mask().
1457 * This function should only be called by vm_page_dirty().
1460 vm_page_dirty_KBI(vm_page_t m)
1463 /* Refer to this operation by its public name. */
1464 KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!"));
1465 m->dirty = VM_PAGE_BITS_ALL;
1469 * vm_page_insert: [ internal use only ]
1471 * Inserts the given mem entry into the object and object list.
1473 * The object must be locked.
1476 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1480 VM_OBJECT_ASSERT_WLOCKED(object);
1481 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1482 return (vm_page_insert_after(m, object, pindex, mpred));
1486 * vm_page_insert_after:
1488 * Inserts the page "m" into the specified object at offset "pindex".
1490 * The page "mpred" must immediately precede the offset "pindex" within
1491 * the specified object.
1493 * The object must be locked.
1496 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1501 VM_OBJECT_ASSERT_WLOCKED(object);
1502 KASSERT(m->object == NULL,
1503 ("vm_page_insert_after: page already inserted"));
1504 if (mpred != NULL) {
1505 KASSERT(mpred->object == object,
1506 ("vm_page_insert_after: object doesn't contain mpred"));
1507 KASSERT(mpred->pindex < pindex,
1508 ("vm_page_insert_after: mpred doesn't precede pindex"));
1509 msucc = TAILQ_NEXT(mpred, listq);
1511 msucc = TAILQ_FIRST(&object->memq);
1513 KASSERT(msucc->pindex > pindex,
1514 ("vm_page_insert_after: msucc doesn't succeed pindex"));
1517 * Record the object/offset pair in this page.
1521 m->ref_count |= VPRC_OBJREF;
1524 * Now link into the object's ordered list of backed pages.
1526 if (vm_radix_insert(&object->rtree, m)) {
1529 m->ref_count &= ~VPRC_OBJREF;
1532 vm_page_insert_radixdone(m, object, mpred);
1537 * vm_page_insert_radixdone:
1539 * Complete page "m" insertion into the specified object after the
1540 * radix trie hooking.
1542 * The page "mpred" must precede the offset "m->pindex" within the
1545 * The object must be locked.
1548 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1551 VM_OBJECT_ASSERT_WLOCKED(object);
1552 KASSERT(object != NULL && m->object == object,
1553 ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1554 KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1555 ("vm_page_insert_radixdone: page %p is missing object ref", m));
1556 if (mpred != NULL) {
1557 KASSERT(mpred->object == object,
1558 ("vm_page_insert_radixdone: object doesn't contain mpred"));
1559 KASSERT(mpred->pindex < m->pindex,
1560 ("vm_page_insert_radixdone: mpred doesn't precede pindex"));
1564 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1566 TAILQ_INSERT_HEAD(&object->memq, m, listq);
1569 * Show that the object has one more resident page.
1571 object->resident_page_count++;
1574 * Hold the vnode until the last page is released.
1576 if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1577 vhold(object->handle);
1580 * Since we are inserting a new and possibly dirty page,
1581 * update the object's generation count.
1583 if (pmap_page_is_write_mapped(m))
1584 vm_object_set_writeable_dirty(object);
1588 * Do the work to remove a page from its object. The caller is responsible for
1589 * updating the page's fields to reflect this removal.
1592 vm_page_object_remove(vm_page_t m)
1597 vm_page_assert_xbusied(m);
1599 VM_OBJECT_ASSERT_WLOCKED(object);
1600 KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1601 ("page %p is missing its object ref", m));
1603 /* Deferred free of swap space. */
1604 if ((m->a.flags & PGA_SWAP_FREE) != 0)
1605 vm_pager_page_unswapped(m);
1607 mrem = vm_radix_remove(&object->rtree, m->pindex);
1608 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1611 * Now remove from the object's list of backed pages.
1613 TAILQ_REMOVE(&object->memq, m, listq);
1616 * And show that the object has one fewer resident page.
1618 object->resident_page_count--;
1621 * The vnode may now be recycled.
1623 if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1624 vdrop(object->handle);
1630 * Removes the specified page from its containing object, but does not
1631 * invalidate any backing storage. Returns true if the object's reference
1632 * was the last reference to the page, and false otherwise.
1634 * The object must be locked and the page must be exclusively busied.
1635 * The exclusive busy will be released on return. If this is not the
1636 * final ref and the caller does not hold a wire reference it may not
1637 * continue to access the page.
1640 vm_page_remove(vm_page_t m)
1644 dropped = vm_page_remove_xbusy(m);
1651 * vm_page_remove_xbusy
1653 * Removes the page but leaves the xbusy held. Returns true if this
1654 * removed the final ref and false otherwise.
1657 vm_page_remove_xbusy(vm_page_t m)
1660 vm_page_object_remove(m);
1662 return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1668 * Returns the page associated with the object/offset
1669 * pair specified; if none is found, NULL is returned.
1671 * The object must be locked.
1674 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1677 VM_OBJECT_ASSERT_LOCKED(object);
1678 return (vm_radix_lookup(&object->rtree, pindex));
1682 * vm_page_find_least:
1684 * Returns the page associated with the object with least pindex
1685 * greater than or equal to the parameter pindex, or NULL.
1687 * The object must be locked.
1690 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1694 VM_OBJECT_ASSERT_LOCKED(object);
1695 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1696 m = vm_radix_lookup_ge(&object->rtree, pindex);
1701 * Returns the given page's successor (by pindex) within the object if it is
1702 * resident; if none is found, NULL is returned.
1704 * The object must be locked.
1707 vm_page_next(vm_page_t m)
1711 VM_OBJECT_ASSERT_LOCKED(m->object);
1712 if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1713 MPASS(next->object == m->object);
1714 if (next->pindex != m->pindex + 1)
1721 * Returns the given page's predecessor (by pindex) within the object if it is
1722 * resident; if none is found, NULL is returned.
1724 * The object must be locked.
1727 vm_page_prev(vm_page_t m)
1731 VM_OBJECT_ASSERT_LOCKED(m->object);
1732 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1733 MPASS(prev->object == m->object);
1734 if (prev->pindex != m->pindex - 1)
1741 * Uses the page mnew as a replacement for an existing page at index
1742 * pindex which must be already present in the object.
1744 * Both pages must be exclusively busied on enter. The old page is
1747 * A return value of true means mold is now free. If this is not the
1748 * final ref and the caller does not hold a wire reference it may not
1749 * continue to access the page.
1752 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1758 VM_OBJECT_ASSERT_WLOCKED(object);
1759 vm_page_assert_xbusied(mold);
1760 KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
1761 ("vm_page_replace: page %p already in object", mnew));
1764 * This function mostly follows vm_page_insert() and
1765 * vm_page_remove() without the radix, object count and vnode
1766 * dance. Double check such functions for more comments.
1769 mnew->object = object;
1770 mnew->pindex = pindex;
1771 atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
1772 mret = vm_radix_replace(&object->rtree, mnew);
1773 KASSERT(mret == mold,
1774 ("invalid page replacement, mold=%p, mret=%p", mold, mret));
1775 KASSERT((mold->oflags & VPO_UNMANAGED) ==
1776 (mnew->oflags & VPO_UNMANAGED),
1777 ("vm_page_replace: mismatched VPO_UNMANAGED"));
1779 /* Keep the resident page list in sorted order. */
1780 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1781 TAILQ_REMOVE(&object->memq, mold, listq);
1782 mold->object = NULL;
1785 * The object's resident_page_count does not change because we have
1786 * swapped one page for another, but the generation count should
1787 * change if the page is dirty.
1789 if (pmap_page_is_write_mapped(mnew))
1790 vm_object_set_writeable_dirty(object);
1791 dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF;
1792 vm_page_xunbusy(mold);
1798 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1802 vm_page_assert_xbusied(mnew);
1804 if (vm_page_replace_hold(mnew, object, pindex, mold))
1811 * Move the given memory entry from its
1812 * current object to the specified target object/offset.
1814 * Note: swap associated with the page must be invalidated by the move. We
1815 * have to do this for several reasons: (1) we aren't freeing the
1816 * page, (2) we are dirtying the page, (3) the VM system is probably
1817 * moving the page from object A to B, and will then later move
1818 * the backing store from A to B and we can't have a conflict.
1820 * Note: we *always* dirty the page. It is necessary both for the
1821 * fact that we moved it, and because we may be invalidating
1824 * The objects must be locked.
1827 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1832 VM_OBJECT_ASSERT_WLOCKED(new_object);
1834 KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m));
1835 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1836 KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1837 ("vm_page_rename: pindex already renamed"));
1840 * Create a custom version of vm_page_insert() which does not depend
1841 * by m_prev and can cheat on the implementation aspects of the
1845 m->pindex = new_pindex;
1846 if (vm_radix_insert(&new_object->rtree, m)) {
1852 * The operation cannot fail anymore. The removal must happen before
1853 * the listq iterator is tainted.
1856 vm_page_object_remove(m);
1858 /* Return back to the new pindex to complete vm_page_insert(). */
1859 m->pindex = new_pindex;
1860 m->object = new_object;
1862 vm_page_insert_radixdone(m, new_object, mpred);
1870 * Allocate and return a page that is associated with the specified
1871 * object and offset pair. By default, this page is exclusive busied.
1873 * The caller must always specify an allocation class.
1875 * allocation classes:
1876 * VM_ALLOC_NORMAL normal process request
1877 * VM_ALLOC_SYSTEM system *really* needs a page
1878 * VM_ALLOC_INTERRUPT interrupt time request
1880 * optional allocation flags:
1881 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1882 * intends to allocate
1883 * VM_ALLOC_NOBUSY do not exclusive busy the page
1884 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1885 * VM_ALLOC_NOOBJ page is not associated with an object and
1886 * should not be exclusive busy
1887 * VM_ALLOC_SBUSY shared busy the allocated page
1888 * VM_ALLOC_WIRED wire the allocated page
1889 * VM_ALLOC_ZERO prefer a zeroed page
1892 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1895 return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1896 vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1900 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1904 return (vm_page_alloc_domain_after(object, pindex, domain, req,
1905 object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) :
1910 * Allocate a page in the specified object with the given page index. To
1911 * optimize insertion of the page into the object, the caller must also specifiy
1912 * the resident page in the object with largest index smaller than the given
1913 * page index, or NULL if no such page exists.
1916 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1917 int req, vm_page_t mpred)
1919 struct vm_domainset_iter di;
1923 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1925 m = vm_page_alloc_domain_after(object, pindex, domain, req,
1929 } while (vm_domainset_iter_page(&di, object, &domain) == 0);
1935 * Returns true if the number of free pages exceeds the minimum
1936 * for the request class and false otherwise.
1939 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
1941 u_int limit, old, new;
1943 if (req_class == VM_ALLOC_INTERRUPT)
1945 else if (req_class == VM_ALLOC_SYSTEM)
1946 limit = vmd->vmd_interrupt_free_min;
1948 limit = vmd->vmd_free_reserved;
1951 * Attempt to reserve the pages. Fail if we're below the limit.
1954 old = vmd->vmd_free_count;
1959 } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
1961 /* Wake the page daemon if we've crossed the threshold. */
1962 if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
1963 pagedaemon_wakeup(vmd->vmd_domain);
1965 /* Only update bitsets on transitions. */
1966 if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
1967 (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
1974 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
1979 * The page daemon is allowed to dig deeper into the free page list.
1981 req_class = req & VM_ALLOC_CLASS_MASK;
1982 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1983 req_class = VM_ALLOC_SYSTEM;
1984 return (_vm_domain_allocate(vmd, req_class, npages));
1988 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
1989 int req, vm_page_t mpred)
1991 struct vm_domain *vmd;
1995 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1996 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1997 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1998 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1999 ("inconsistent object(%p)/req(%x)", object, req));
2000 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2001 ("Can't sleep and retry object insertion."));
2002 KASSERT(mpred == NULL || mpred->pindex < pindex,
2003 ("mpred %p doesn't precede pindex 0x%jx", mpred,
2004 (uintmax_t)pindex));
2006 VM_OBJECT_ASSERT_WLOCKED(object);
2010 pool = object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT;
2012 #if VM_NRESERVLEVEL > 0
2014 * Can we allocate the page from a reservation?
2016 if (vm_object_reserv(object) &&
2017 (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
2019 domain = vm_phys_domain(m);
2020 vmd = VM_DOMAIN(domain);
2024 vmd = VM_DOMAIN(domain);
2025 if (vmd->vmd_pgcache[pool].zone != NULL) {
2026 m = uma_zalloc(vmd->vmd_pgcache[pool].zone, M_NOWAIT);
2028 flags |= PG_PCPU_CACHE;
2032 if (vm_domain_allocate(vmd, req, 1)) {
2034 * If not, allocate it from the free page queues.
2036 vm_domain_free_lock(vmd);
2037 m = vm_phys_alloc_pages(domain, pool, 0);
2038 vm_domain_free_unlock(vmd);
2040 vm_domain_freecnt_inc(vmd, 1);
2041 #if VM_NRESERVLEVEL > 0
2042 if (vm_reserv_reclaim_inactive(domain))
2049 * Not allocatable, give up.
2051 if (vm_domain_alloc_fail(vmd, object, req))
2057 * At this point we had better have found a good page.
2061 vm_page_alloc_check(m);
2064 * Initialize the page. Only the PG_ZERO flag is inherited.
2066 if ((req & VM_ALLOC_ZERO) != 0)
2067 flags |= (m->flags & PG_ZERO);
2068 if ((req & VM_ALLOC_NODUMP) != 0)
2072 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2074 m->busy_lock = VPB_UNBUSIED;
2075 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2076 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2077 if ((req & VM_ALLOC_SBUSY) != 0)
2078 m->busy_lock = VPB_SHARERS_WORD(1);
2079 if (req & VM_ALLOC_WIRED) {
2085 if (object != NULL) {
2086 if (vm_page_insert_after(m, object, pindex, mpred)) {
2087 if (req & VM_ALLOC_WIRED) {
2091 KASSERT(m->object == NULL, ("page %p has object", m));
2092 m->oflags = VPO_UNMANAGED;
2093 m->busy_lock = VPB_UNBUSIED;
2094 /* Don't change PG_ZERO. */
2095 vm_page_free_toq(m);
2096 if (req & VM_ALLOC_WAITFAIL) {
2097 VM_OBJECT_WUNLOCK(object);
2099 VM_OBJECT_WLOCK(object);
2104 /* Ignore device objects; the pager sets "memattr" for them. */
2105 if (object->memattr != VM_MEMATTR_DEFAULT &&
2106 (object->flags & OBJ_FICTITIOUS) == 0)
2107 pmap_page_set_memattr(m, object->memattr);
2115 * vm_page_alloc_contig:
2117 * Allocate a contiguous set of physical pages of the given size "npages"
2118 * from the free lists. All of the physical pages must be at or above
2119 * the given physical address "low" and below the given physical address
2120 * "high". The given value "alignment" determines the alignment of the
2121 * first physical page in the set. If the given value "boundary" is
2122 * non-zero, then the set of physical pages cannot cross any physical
2123 * address boundary that is a multiple of that value. Both "alignment"
2124 * and "boundary" must be a power of two.
2126 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
2127 * then the memory attribute setting for the physical pages is configured
2128 * to the object's memory attribute setting. Otherwise, the memory
2129 * attribute setting for the physical pages is configured to "memattr",
2130 * overriding the object's memory attribute setting. However, if the
2131 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
2132 * memory attribute setting for the physical pages cannot be configured
2133 * to VM_MEMATTR_DEFAULT.
2135 * The specified object may not contain fictitious pages.
2137 * The caller must always specify an allocation class.
2139 * allocation classes:
2140 * VM_ALLOC_NORMAL normal process request
2141 * VM_ALLOC_SYSTEM system *really* needs a page
2142 * VM_ALLOC_INTERRUPT interrupt time request
2144 * optional allocation flags:
2145 * VM_ALLOC_NOBUSY do not exclusive busy the page
2146 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
2147 * VM_ALLOC_NOOBJ page is not associated with an object and
2148 * should not be exclusive busy
2149 * VM_ALLOC_SBUSY shared busy the allocated page
2150 * VM_ALLOC_WIRED wire the allocated page
2151 * VM_ALLOC_ZERO prefer a zeroed page
2154 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
2155 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2156 vm_paddr_t boundary, vm_memattr_t memattr)
2158 struct vm_domainset_iter di;
2162 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2164 m = vm_page_alloc_contig_domain(object, pindex, domain, req,
2165 npages, low, high, alignment, boundary, memattr);
2168 } while (vm_domainset_iter_page(&di, object, &domain) == 0);
2174 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2175 int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2176 vm_paddr_t boundary, vm_memattr_t memattr)
2178 struct vm_domain *vmd;
2179 vm_page_t m, m_ret, mpred;
2180 u_int busy_lock, flags, oflags;
2182 mpred = NULL; /* XXX: pacify gcc */
2183 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
2184 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
2185 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2186 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2187 ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
2189 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2190 ("Can't sleep and retry object insertion."));
2191 if (object != NULL) {
2192 VM_OBJECT_ASSERT_WLOCKED(object);
2193 KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2194 ("vm_page_alloc_contig: object %p has fictitious pages",
2197 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2199 if (object != NULL) {
2200 mpred = vm_radix_lookup_le(&object->rtree, pindex);
2201 KASSERT(mpred == NULL || mpred->pindex != pindex,
2202 ("vm_page_alloc_contig: pindex already allocated"));
2206 * Can we allocate the pages without the number of free pages falling
2207 * below the lower bound for the allocation class?
2211 #if VM_NRESERVLEVEL > 0
2213 * Can we allocate the pages from a reservation?
2215 if (vm_object_reserv(object) &&
2216 (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
2217 mpred, npages, low, high, alignment, boundary)) != NULL) {
2218 domain = vm_phys_domain(m_ret);
2219 vmd = VM_DOMAIN(domain);
2223 vmd = VM_DOMAIN(domain);
2224 if (vm_domain_allocate(vmd, req, npages)) {
2226 * allocate them from the free page queues.
2228 vm_domain_free_lock(vmd);
2229 m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2230 alignment, boundary);
2231 vm_domain_free_unlock(vmd);
2232 if (m_ret == NULL) {
2233 vm_domain_freecnt_inc(vmd, npages);
2234 #if VM_NRESERVLEVEL > 0
2235 if (vm_reserv_reclaim_contig(domain, npages, low,
2236 high, alignment, boundary))
2241 if (m_ret == NULL) {
2242 if (vm_domain_alloc_fail(vmd, object, req))
2246 #if VM_NRESERVLEVEL > 0
2249 for (m = m_ret; m < &m_ret[npages]; m++) {
2251 vm_page_alloc_check(m);
2255 * Initialize the pages. Only the PG_ZERO flag is inherited.
2258 if ((req & VM_ALLOC_ZERO) != 0)
2260 if ((req & VM_ALLOC_NODUMP) != 0)
2262 oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2264 busy_lock = VPB_UNBUSIED;
2265 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2266 busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2267 if ((req & VM_ALLOC_SBUSY) != 0)
2268 busy_lock = VPB_SHARERS_WORD(1);
2269 if ((req & VM_ALLOC_WIRED) != 0)
2270 vm_wire_add(npages);
2271 if (object != NULL) {
2272 if (object->memattr != VM_MEMATTR_DEFAULT &&
2273 memattr == VM_MEMATTR_DEFAULT)
2274 memattr = object->memattr;
2276 for (m = m_ret; m < &m_ret[npages]; m++) {
2278 m->flags = (m->flags | PG_NODUMP) & flags;
2279 m->busy_lock = busy_lock;
2280 if ((req & VM_ALLOC_WIRED) != 0)
2284 if (object != NULL) {
2285 if (vm_page_insert_after(m, object, pindex, mpred)) {
2286 if ((req & VM_ALLOC_WIRED) != 0)
2287 vm_wire_sub(npages);
2288 KASSERT(m->object == NULL,
2289 ("page %p has object", m));
2291 for (m = m_ret; m < &m_ret[npages]; m++) {
2293 (req & VM_ALLOC_WIRED) != 0)
2295 m->oflags = VPO_UNMANAGED;
2296 m->busy_lock = VPB_UNBUSIED;
2297 /* Don't change PG_ZERO. */
2298 vm_page_free_toq(m);
2300 if (req & VM_ALLOC_WAITFAIL) {
2301 VM_OBJECT_WUNLOCK(object);
2303 VM_OBJECT_WLOCK(object);
2310 if (memattr != VM_MEMATTR_DEFAULT)
2311 pmap_page_set_memattr(m, memattr);
2318 * Check a page that has been freshly dequeued from a freelist.
2321 vm_page_alloc_check(vm_page_t m)
2324 KASSERT(m->object == NULL, ("page %p has object", m));
2325 KASSERT(m->a.queue == PQ_NONE &&
2326 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
2327 ("page %p has unexpected queue %d, flags %#x",
2328 m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK)));
2329 KASSERT(m->ref_count == 0, ("page %p has references", m));
2330 KASSERT(!vm_page_busied(m), ("page %p is busy", m));
2331 KASSERT(m->dirty == 0, ("page %p is dirty", m));
2332 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2333 ("page %p has unexpected memattr %d",
2334 m, pmap_page_get_memattr(m)));
2335 KASSERT(m->valid == 0, ("free page %p is valid", m));
2339 * vm_page_alloc_freelist:
2341 * Allocate a physical page from the specified free page list.
2343 * The caller must always specify an allocation class.
2345 * allocation classes:
2346 * VM_ALLOC_NORMAL normal process request
2347 * VM_ALLOC_SYSTEM system *really* needs a page
2348 * VM_ALLOC_INTERRUPT interrupt time request
2350 * optional allocation flags:
2351 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
2352 * intends to allocate
2353 * VM_ALLOC_WIRED wire the allocated page
2354 * VM_ALLOC_ZERO prefer a zeroed page
2357 vm_page_alloc_freelist(int freelist, int req)
2359 struct vm_domainset_iter di;
2363 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2365 m = vm_page_alloc_freelist_domain(domain, freelist, req);
2368 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2374 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2376 struct vm_domain *vmd;
2381 vmd = VM_DOMAIN(domain);
2383 if (vm_domain_allocate(vmd, req, 1)) {
2384 vm_domain_free_lock(vmd);
2385 m = vm_phys_alloc_freelist_pages(domain, freelist,
2386 VM_FREEPOOL_DIRECT, 0);
2387 vm_domain_free_unlock(vmd);
2389 vm_domain_freecnt_inc(vmd, 1);
2392 if (vm_domain_alloc_fail(vmd, NULL, req))
2397 vm_page_alloc_check(m);
2400 * Initialize the page. Only the PG_ZERO flag is inherited.
2404 if ((req & VM_ALLOC_ZERO) != 0)
2407 if ((req & VM_ALLOC_WIRED) != 0) {
2411 /* Unmanaged pages don't use "act_count". */
2412 m->oflags = VPO_UNMANAGED;
2417 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2419 struct vm_domain *vmd;
2420 struct vm_pgcache *pgcache;
2424 vmd = VM_DOMAIN(pgcache->domain);
2427 * The page daemon should avoid creating extra memory pressure since its
2428 * main purpose is to replenish the store of free pages.
2430 if (vmd->vmd_severeset || curproc == pageproc ||
2431 !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2433 domain = vmd->vmd_domain;
2434 vm_domain_free_lock(vmd);
2435 i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2436 (vm_page_t *)store);
2437 vm_domain_free_unlock(vmd);
2439 vm_domain_freecnt_inc(vmd, cnt - i);
2445 vm_page_zone_release(void *arg, void **store, int cnt)
2447 struct vm_domain *vmd;
2448 struct vm_pgcache *pgcache;
2453 vmd = VM_DOMAIN(pgcache->domain);
2454 vm_domain_free_lock(vmd);
2455 for (i = 0; i < cnt; i++) {
2456 m = (vm_page_t)store[i];
2457 vm_phys_free_pages(m, 0);
2459 vm_domain_free_unlock(vmd);
2460 vm_domain_freecnt_inc(vmd, cnt);
2463 #define VPSC_ANY 0 /* No restrictions. */
2464 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */
2465 #define VPSC_NOSUPER 2 /* Skip superpages. */
2468 * vm_page_scan_contig:
2470 * Scan vm_page_array[] between the specified entries "m_start" and
2471 * "m_end" for a run of contiguous physical pages that satisfy the
2472 * specified conditions, and return the lowest page in the run. The
2473 * specified "alignment" determines the alignment of the lowest physical
2474 * page in the run. If the specified "boundary" is non-zero, then the
2475 * run of physical pages cannot span a physical address that is a
2476 * multiple of "boundary".
2478 * "m_end" is never dereferenced, so it need not point to a vm_page
2479 * structure within vm_page_array[].
2481 * "npages" must be greater than zero. "m_start" and "m_end" must not
2482 * span a hole (or discontiguity) in the physical address space. Both
2483 * "alignment" and "boundary" must be a power of two.
2486 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2487 u_long alignment, vm_paddr_t boundary, int options)
2493 #if VM_NRESERVLEVEL > 0
2496 int m_inc, order, run_ext, run_len;
2498 KASSERT(npages > 0, ("npages is 0"));
2499 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2500 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2504 for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2505 KASSERT((m->flags & PG_MARKER) == 0,
2506 ("page %p is PG_MARKER", m));
2507 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1,
2508 ("fictitious page %p has invalid ref count", m));
2511 * If the current page would be the start of a run, check its
2512 * physical address against the end, alignment, and boundary
2513 * conditions. If it doesn't satisfy these conditions, either
2514 * terminate the scan or advance to the next page that
2515 * satisfies the failed condition.
2518 KASSERT(m_run == NULL, ("m_run != NULL"));
2519 if (m + npages > m_end)
2521 pa = VM_PAGE_TO_PHYS(m);
2522 if ((pa & (alignment - 1)) != 0) {
2523 m_inc = atop(roundup2(pa, alignment) - pa);
2526 if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2528 m_inc = atop(roundup2(pa, boundary) - pa);
2532 KASSERT(m_run != NULL, ("m_run == NULL"));
2534 vm_page_change_lock(m, &m_mtx);
2537 if (vm_page_wired(m))
2539 #if VM_NRESERVLEVEL > 0
2540 else if ((level = vm_reserv_level(m)) >= 0 &&
2541 (options & VPSC_NORESERV) != 0) {
2543 /* Advance to the end of the reservation. */
2544 pa = VM_PAGE_TO_PHYS(m);
2545 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2549 else if ((object = m->object) != NULL) {
2551 * The page is considered eligible for relocation if
2552 * and only if it could be laundered or reclaimed by
2555 if (!VM_OBJECT_TRYRLOCK(object)) {
2557 VM_OBJECT_RLOCK(object);
2559 if (m->object != object) {
2561 * The page may have been freed.
2563 VM_OBJECT_RUNLOCK(object);
2567 /* Don't care: PG_NODUMP, PG_ZERO. */
2568 if (object->type != OBJT_DEFAULT &&
2569 object->type != OBJT_SWAP &&
2570 object->type != OBJT_VNODE) {
2572 #if VM_NRESERVLEVEL > 0
2573 } else if ((options & VPSC_NOSUPER) != 0 &&
2574 (level = vm_reserv_level_iffullpop(m)) >= 0) {
2576 /* Advance to the end of the superpage. */
2577 pa = VM_PAGE_TO_PHYS(m);
2578 m_inc = atop(roundup2(pa + 1,
2579 vm_reserv_size(level)) - pa);
2581 } else if (object->memattr == VM_MEMATTR_DEFAULT &&
2582 vm_page_queue(m) != PQ_NONE && !vm_page_busied(m) &&
2583 !vm_page_wired(m)) {
2585 * The page is allocated but eligible for
2586 * relocation. Extend the current run by one
2589 KASSERT(pmap_page_get_memattr(m) ==
2591 ("page %p has an unexpected memattr", m));
2592 KASSERT((m->oflags & (VPO_SWAPINPROG |
2593 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2594 ("page %p has unexpected oflags", m));
2595 /* Don't care: PGA_NOSYNC. */
2599 VM_OBJECT_RUNLOCK(object);
2600 #if VM_NRESERVLEVEL > 0
2601 } else if (level >= 0) {
2603 * The page is reserved but not yet allocated. In
2604 * other words, it is still free. Extend the current
2609 } else if ((order = m->order) < VM_NFREEORDER) {
2611 * The page is enqueued in the physical memory
2612 * allocator's free page queues. Moreover, it is the
2613 * first page in a power-of-two-sized run of
2614 * contiguous free pages. Add these pages to the end
2615 * of the current run, and jump ahead.
2617 run_ext = 1 << order;
2621 * Skip the page for one of the following reasons: (1)
2622 * It is enqueued in the physical memory allocator's
2623 * free page queues. However, it is not the first
2624 * page in a run of contiguous free pages. (This case
2625 * rarely occurs because the scan is performed in
2626 * ascending order.) (2) It is not reserved, and it is
2627 * transitioning from free to allocated. (Conversely,
2628 * the transition from allocated to free for managed
2629 * pages is blocked by the page lock.) (3) It is
2630 * allocated but not contained by an object and not
2631 * wired, e.g., allocated by Xen's balloon driver.
2637 * Extend or reset the current run of pages.
2652 if (run_len >= npages)
2658 * vm_page_reclaim_run:
2660 * Try to relocate each of the allocated virtual pages within the
2661 * specified run of physical pages to a new physical address. Free the
2662 * physical pages underlying the relocated virtual pages. A virtual page
2663 * is relocatable if and only if it could be laundered or reclaimed by
2664 * the page daemon. Whenever possible, a virtual page is relocated to a
2665 * physical address above "high".
2667 * Returns 0 if every physical page within the run was already free or
2668 * just freed by a successful relocation. Otherwise, returns a non-zero
2669 * value indicating why the last attempt to relocate a virtual page was
2672 * "req_class" must be an allocation class.
2675 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2678 struct vm_domain *vmd;
2680 struct spglist free;
2683 vm_page_t m, m_end, m_new;
2684 int error, order, req;
2686 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2687 ("req_class is not an allocation class"));
2691 m_end = m_run + npages;
2693 for (; error == 0 && m < m_end; m++) {
2694 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2695 ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2698 * Avoid releasing and reacquiring the same page lock.
2700 vm_page_change_lock(m, &m_mtx);
2703 * Racily check for wirings. Races are handled below.
2705 if (vm_page_wired(m))
2707 else if ((object = m->object) != NULL) {
2709 * The page is relocated if and only if it could be
2710 * laundered or reclaimed by the page daemon.
2712 if (!VM_OBJECT_TRYWLOCK(object)) {
2714 VM_OBJECT_WLOCK(object);
2716 if (m->object != object) {
2718 * The page may have been freed.
2720 VM_OBJECT_WUNLOCK(object);
2724 /* Don't care: PG_NODUMP, PG_ZERO. */
2725 if (object->type != OBJT_DEFAULT &&
2726 object->type != OBJT_SWAP &&
2727 object->type != OBJT_VNODE)
2729 else if (object->memattr != VM_MEMATTR_DEFAULT)
2731 else if (vm_page_queue(m) != PQ_NONE &&
2732 vm_page_tryxbusy(m) != 0) {
2733 if (vm_page_wired(m)) {
2738 KASSERT(pmap_page_get_memattr(m) ==
2740 ("page %p has an unexpected memattr", m));
2741 KASSERT(m->oflags == 0,
2742 ("page %p has unexpected oflags", m));
2743 /* Don't care: PGA_NOSYNC. */
2744 if (!vm_page_none_valid(m)) {
2746 * First, try to allocate a new page
2747 * that is above "high". Failing
2748 * that, try to allocate a new page
2749 * that is below "m_run". Allocate
2750 * the new page between the end of
2751 * "m_run" and "high" only as a last
2754 req = req_class | VM_ALLOC_NOOBJ;
2755 if ((m->flags & PG_NODUMP) != 0)
2756 req |= VM_ALLOC_NODUMP;
2757 if (trunc_page(high) !=
2758 ~(vm_paddr_t)PAGE_MASK) {
2759 m_new = vm_page_alloc_contig(
2764 VM_MEMATTR_DEFAULT);
2767 if (m_new == NULL) {
2768 pa = VM_PAGE_TO_PHYS(m_run);
2769 m_new = vm_page_alloc_contig(
2771 0, pa - 1, PAGE_SIZE, 0,
2772 VM_MEMATTR_DEFAULT);
2774 if (m_new == NULL) {
2776 m_new = vm_page_alloc_contig(
2778 pa, high, PAGE_SIZE, 0,
2779 VM_MEMATTR_DEFAULT);
2781 if (m_new == NULL) {
2788 * Unmap the page and check for new
2789 * wirings that may have been acquired
2790 * through a pmap lookup.
2792 if (object->ref_count != 0 &&
2793 !vm_page_try_remove_all(m)) {
2795 vm_page_free(m_new);
2801 * Replace "m" with the new page. For
2802 * vm_page_replace(), "m" must be busy
2803 * and dequeued. Finally, change "m"
2804 * as if vm_page_free() was called.
2806 m_new->a.flags = m->a.flags &
2807 ~PGA_QUEUE_STATE_MASK;
2808 KASSERT(m_new->oflags == VPO_UNMANAGED,
2809 ("page %p is managed", m_new));
2811 pmap_copy_page(m, m_new);
2812 m_new->valid = m->valid;
2813 m_new->dirty = m->dirty;
2814 m->flags &= ~PG_ZERO;
2816 if (vm_page_replace_hold(m_new, object,
2818 vm_page_free_prep(m))
2819 SLIST_INSERT_HEAD(&free, m,
2823 * The new page must be deactivated
2824 * before the object is unlocked.
2826 vm_page_change_lock(m_new, &m_mtx);
2827 vm_page_deactivate(m_new);
2829 m->flags &= ~PG_ZERO;
2831 if (vm_page_free_prep(m))
2832 SLIST_INSERT_HEAD(&free, m,
2834 KASSERT(m->dirty == 0,
2835 ("page %p is dirty", m));
2840 VM_OBJECT_WUNLOCK(object);
2842 MPASS(vm_phys_domain(m) == domain);
2843 vmd = VM_DOMAIN(domain);
2844 vm_domain_free_lock(vmd);
2846 if (order < VM_NFREEORDER) {
2848 * The page is enqueued in the physical memory
2849 * allocator's free page queues. Moreover, it
2850 * is the first page in a power-of-two-sized
2851 * run of contiguous free pages. Jump ahead
2852 * to the last page within that run, and
2853 * continue from there.
2855 m += (1 << order) - 1;
2857 #if VM_NRESERVLEVEL > 0
2858 else if (vm_reserv_is_page_free(m))
2861 vm_domain_free_unlock(vmd);
2862 if (order == VM_NFREEORDER)
2868 if ((m = SLIST_FIRST(&free)) != NULL) {
2871 vmd = VM_DOMAIN(domain);
2873 vm_domain_free_lock(vmd);
2875 MPASS(vm_phys_domain(m) == domain);
2876 SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2877 vm_phys_free_pages(m, 0);
2879 } while ((m = SLIST_FIRST(&free)) != NULL);
2880 vm_domain_free_unlock(vmd);
2881 vm_domain_freecnt_inc(vmd, cnt);
2888 CTASSERT(powerof2(NRUNS));
2890 #define RUN_INDEX(count) ((count) & (NRUNS - 1))
2892 #define MIN_RECLAIM 8
2895 * vm_page_reclaim_contig:
2897 * Reclaim allocated, contiguous physical memory satisfying the specified
2898 * conditions by relocating the virtual pages using that physical memory.
2899 * Returns true if reclamation is successful and false otherwise. Since
2900 * relocation requires the allocation of physical pages, reclamation may
2901 * fail due to a shortage of free pages. When reclamation fails, callers
2902 * are expected to perform vm_wait() before retrying a failed allocation
2903 * operation, e.g., vm_page_alloc_contig().
2905 * The caller must always specify an allocation class through "req".
2907 * allocation classes:
2908 * VM_ALLOC_NORMAL normal process request
2909 * VM_ALLOC_SYSTEM system *really* needs a page
2910 * VM_ALLOC_INTERRUPT interrupt time request
2912 * The optional allocation flags are ignored.
2914 * "npages" must be greater than zero. Both "alignment" and "boundary"
2915 * must be a power of two.
2918 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
2919 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2921 struct vm_domain *vmd;
2922 vm_paddr_t curr_low;
2923 vm_page_t m_run, m_runs[NRUNS];
2924 u_long count, reclaimed;
2925 int error, i, options, req_class;
2927 KASSERT(npages > 0, ("npages is 0"));
2928 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2929 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2930 req_class = req & VM_ALLOC_CLASS_MASK;
2933 * The page daemon is allowed to dig deeper into the free page list.
2935 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2936 req_class = VM_ALLOC_SYSTEM;
2939 * Return if the number of free pages cannot satisfy the requested
2942 vmd = VM_DOMAIN(domain);
2943 count = vmd->vmd_free_count;
2944 if (count < npages + vmd->vmd_free_reserved || (count < npages +
2945 vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2946 (count < npages && req_class == VM_ALLOC_INTERRUPT))
2950 * Scan up to three times, relaxing the restrictions ("options") on
2951 * the reclamation of reservations and superpages each time.
2953 for (options = VPSC_NORESERV;;) {
2955 * Find the highest runs that satisfy the given constraints
2956 * and restrictions, and record them in "m_runs".
2961 m_run = vm_phys_scan_contig(domain, npages, curr_low,
2962 high, alignment, boundary, options);
2965 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2966 m_runs[RUN_INDEX(count)] = m_run;
2971 * Reclaim the highest runs in LIFO (descending) order until
2972 * the number of reclaimed pages, "reclaimed", is at least
2973 * MIN_RECLAIM. Reset "reclaimed" each time because each
2974 * reclamation is idempotent, and runs will (likely) recur
2975 * from one scan to the next as restrictions are relaxed.
2978 for (i = 0; count > 0 && i < NRUNS; i++) {
2980 m_run = m_runs[RUN_INDEX(count)];
2981 error = vm_page_reclaim_run(req_class, domain, npages,
2984 reclaimed += npages;
2985 if (reclaimed >= MIN_RECLAIM)
2991 * Either relax the restrictions on the next scan or return if
2992 * the last scan had no restrictions.
2994 if (options == VPSC_NORESERV)
2995 options = VPSC_NOSUPER;
2996 else if (options == VPSC_NOSUPER)
2998 else if (options == VPSC_ANY)
2999 return (reclaimed != 0);
3004 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
3005 u_long alignment, vm_paddr_t boundary)
3007 struct vm_domainset_iter di;
3011 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
3013 ret = vm_page_reclaim_contig_domain(domain, req, npages, low,
3014 high, alignment, boundary);
3017 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
3023 * Set the domain in the appropriate page level domainset.
3026 vm_domain_set(struct vm_domain *vmd)
3029 mtx_lock(&vm_domainset_lock);
3030 if (!vmd->vmd_minset && vm_paging_min(vmd)) {
3031 vmd->vmd_minset = 1;
3032 DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
3034 if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
3035 vmd->vmd_severeset = 1;
3036 DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
3038 mtx_unlock(&vm_domainset_lock);
3042 * Clear the domain from the appropriate page level domainset.
3045 vm_domain_clear(struct vm_domain *vmd)
3048 mtx_lock(&vm_domainset_lock);
3049 if (vmd->vmd_minset && !vm_paging_min(vmd)) {
3050 vmd->vmd_minset = 0;
3051 DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
3052 if (vm_min_waiters != 0) {
3054 wakeup(&vm_min_domains);
3057 if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
3058 vmd->vmd_severeset = 0;
3059 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
3060 if (vm_severe_waiters != 0) {
3061 vm_severe_waiters = 0;
3062 wakeup(&vm_severe_domains);
3067 * If pageout daemon needs pages, then tell it that there are
3070 if (vmd->vmd_pageout_pages_needed &&
3071 vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
3072 wakeup(&vmd->vmd_pageout_pages_needed);
3073 vmd->vmd_pageout_pages_needed = 0;
3076 /* See comments in vm_wait_doms(). */
3077 if (vm_pageproc_waiters) {
3078 vm_pageproc_waiters = 0;
3079 wakeup(&vm_pageproc_waiters);
3081 mtx_unlock(&vm_domainset_lock);
3085 * Wait for free pages to exceed the min threshold globally.
3091 mtx_lock(&vm_domainset_lock);
3092 while (vm_page_count_min()) {
3094 msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
3096 mtx_unlock(&vm_domainset_lock);
3100 * Wait for free pages to exceed the severe threshold globally.
3103 vm_wait_severe(void)
3106 mtx_lock(&vm_domainset_lock);
3107 while (vm_page_count_severe()) {
3108 vm_severe_waiters++;
3109 msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
3112 mtx_unlock(&vm_domainset_lock);
3119 return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
3123 vm_wait_doms(const domainset_t *wdoms)
3127 * We use racey wakeup synchronization to avoid expensive global
3128 * locking for the pageproc when sleeping with a non-specific vm_wait.
3129 * To handle this, we only sleep for one tick in this instance. It
3130 * is expected that most allocations for the pageproc will come from
3131 * kmem or vm_page_grab* which will use the more specific and
3132 * race-free vm_wait_domain().
3134 if (curproc == pageproc) {
3135 mtx_lock(&vm_domainset_lock);
3136 vm_pageproc_waiters++;
3137 msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP,
3141 * XXX Ideally we would wait only until the allocation could
3142 * be satisfied. This condition can cause new allocators to
3143 * consume all freed pages while old allocators wait.
3145 mtx_lock(&vm_domainset_lock);
3146 if (vm_page_count_min_set(wdoms)) {
3148 msleep(&vm_min_domains, &vm_domainset_lock,
3149 PVM | PDROP, "vmwait", 0);
3151 mtx_unlock(&vm_domainset_lock);
3158 * Sleep until free pages are available for allocation.
3159 * - Called in various places after failed memory allocations.
3162 vm_wait_domain(int domain)
3164 struct vm_domain *vmd;
3167 vmd = VM_DOMAIN(domain);
3168 vm_domain_free_assert_unlocked(vmd);
3170 if (curproc == pageproc) {
3171 mtx_lock(&vm_domainset_lock);
3172 if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3173 vmd->vmd_pageout_pages_needed = 1;
3174 msleep(&vmd->vmd_pageout_pages_needed,
3175 &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3177 mtx_unlock(&vm_domainset_lock);
3179 if (pageproc == NULL)
3180 panic("vm_wait in early boot");
3181 DOMAINSET_ZERO(&wdom);
3182 DOMAINSET_SET(vmd->vmd_domain, &wdom);
3183 vm_wait_doms(&wdom);
3190 * Sleep until free pages are available for allocation in the
3191 * affinity domains of the obj. If obj is NULL, the domain set
3192 * for the calling thread is used.
3193 * Called in various places after failed memory allocations.
3196 vm_wait(vm_object_t obj)
3198 struct domainset *d;
3203 * Carefully fetch pointers only once: the struct domainset
3204 * itself is ummutable but the pointer might change.
3207 d = obj->domain.dr_policy;
3209 d = curthread->td_domain.dr_policy;
3211 vm_wait_doms(&d->ds_mask);
3215 * vm_domain_alloc_fail:
3217 * Called when a page allocation function fails. Informs the
3218 * pagedaemon and performs the requested wait. Requires the
3219 * domain_free and object lock on entry. Returns with the
3220 * object lock held and free lock released. Returns an error when
3221 * retry is necessary.
3225 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3228 vm_domain_free_assert_unlocked(vmd);
3230 atomic_add_int(&vmd->vmd_pageout_deficit,
3231 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3232 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3234 VM_OBJECT_WUNLOCK(object);
3235 vm_wait_domain(vmd->vmd_domain);
3237 VM_OBJECT_WLOCK(object);
3238 if (req & VM_ALLOC_WAITOK)
3248 * Sleep until free pages are available for allocation.
3249 * - Called only in vm_fault so that processes page faulting
3250 * can be easily tracked.
3251 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3252 * processes will be able to grab memory first. Do not change
3253 * this balance without careful testing first.
3256 vm_waitpfault(struct domainset *dset, int timo)
3260 * XXX Ideally we would wait only until the allocation could
3261 * be satisfied. This condition can cause new allocators to
3262 * consume all freed pages while old allocators wait.
3264 mtx_lock(&vm_domainset_lock);
3265 if (vm_page_count_min_set(&dset->ds_mask)) {
3267 msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3270 mtx_unlock(&vm_domainset_lock);
3273 static struct vm_pagequeue *
3274 _vm_page_pagequeue(vm_page_t m, uint8_t queue)
3277 return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3281 static struct vm_pagequeue *
3282 vm_page_pagequeue(vm_page_t m)
3285 return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue));
3289 static __always_inline bool
3290 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3292 vm_page_astate_t tmp;
3296 if (__predict_true(vm_page_astate_fcmpset(m, old, new)))
3298 counter_u64_add(pqstate_commit_retries, 1);
3299 } while (old->_bits == tmp._bits);
3305 * Do the work of committing a queue state update that moves the page out of
3306 * its current queue.
3309 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m,
3310 vm_page_astate_t *old, vm_page_astate_t new)
3314 vm_pagequeue_assert_locked(pq);
3315 KASSERT(vm_page_pagequeue(m) == pq,
3316 ("%s: queue %p does not match page %p", __func__, pq, m));
3317 KASSERT(old->queue != PQ_NONE && new.queue != old->queue,
3318 ("%s: invalid queue indices %d %d",
3319 __func__, old->queue, new.queue));
3322 * Once the queue index of the page changes there is nothing
3323 * synchronizing with further updates to the page's physical
3324 * queue state. Therefore we must speculatively remove the page
3325 * from the queue now and be prepared to roll back if the queue
3326 * state update fails. If the page is not physically enqueued then
3327 * we just update its queue index.
3329 if ((old->flags & PGA_ENQUEUED) != 0) {
3330 new.flags &= ~PGA_ENQUEUED;
3331 next = TAILQ_NEXT(m, plinks.q);
3332 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3333 vm_pagequeue_cnt_dec(pq);
3334 if (!vm_page_pqstate_fcmpset(m, old, new)) {
3336 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3338 TAILQ_INSERT_BEFORE(next, m, plinks.q);
3339 vm_pagequeue_cnt_inc(pq);
3345 return (vm_page_pqstate_fcmpset(m, old, new));
3350 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old,
3351 vm_page_astate_t new)
3353 struct vm_pagequeue *pq;
3354 vm_page_astate_t as;
3357 pq = _vm_page_pagequeue(m, old->queue);
3360 * The queue field and PGA_ENQUEUED flag are stable only so long as the
3361 * corresponding page queue lock is held.
3363 vm_pagequeue_lock(pq);
3364 as = vm_page_astate_load(m);
3365 if (__predict_false(as._bits != old->_bits)) {
3369 ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new);
3371 vm_pagequeue_unlock(pq);
3376 * Commit a queue state update that enqueues or requeues a page.
3379 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m,
3380 vm_page_astate_t *old, vm_page_astate_t new)
3382 struct vm_domain *vmd;
3384 vm_pagequeue_assert_locked(pq);
3385 KASSERT(old->queue != PQ_NONE && new.queue == old->queue,
3386 ("%s: invalid queue indices %d %d",
3387 __func__, old->queue, new.queue));
3389 new.flags |= PGA_ENQUEUED;
3390 if (!vm_page_pqstate_fcmpset(m, old, new))
3393 if ((old->flags & PGA_ENQUEUED) != 0)
3394 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3396 vm_pagequeue_cnt_inc(pq);
3399 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE. In particular, if
3400 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be
3401 * applied, even if it was set first.
3403 if ((old->flags & PGA_REQUEUE_HEAD) != 0) {
3404 vmd = vm_pagequeue_domain(m);
3405 KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE],
3406 ("%s: invalid page queue for page %p", __func__, m));
3407 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3409 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3415 * Commit a queue state update that encodes a request for a deferred queue
3419 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old,
3420 vm_page_astate_t new)
3423 KASSERT(old->queue == new.queue || new.queue != PQ_NONE,
3424 ("%s: invalid state, queue %d flags %x",
3425 __func__, new.queue, new.flags));
3427 if (old->_bits != new._bits &&
3428 !vm_page_pqstate_fcmpset(m, old, new))
3430 vm_page_pqbatch_submit(m, new.queue);
3435 * A generic queue state update function. This handles more cases than the
3436 * specialized functions above.
3439 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3442 if (old->_bits == new._bits)
3445 if (old->queue != PQ_NONE && new.queue != old->queue) {
3446 if (!vm_page_pqstate_commit_dequeue(m, old, new))
3448 if (new.queue != PQ_NONE)
3449 vm_page_pqbatch_submit(m, new.queue);
3451 if (!vm_page_pqstate_fcmpset(m, old, new))
3453 if (new.queue != PQ_NONE &&
3454 ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0)
3455 vm_page_pqbatch_submit(m, new.queue);
3461 * Apply deferred queue state updates to a page.
3464 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue)
3466 vm_page_astate_t new, old;
3468 CRITICAL_ASSERT(curthread);
3469 vm_pagequeue_assert_locked(pq);
3470 KASSERT(queue < PQ_COUNT,
3471 ("%s: invalid queue index %d", __func__, queue));
3472 KASSERT(pq == _vm_page_pagequeue(m, queue),
3473 ("%s: page %p does not belong to queue %p", __func__, m, pq));
3475 for (old = vm_page_astate_load(m);;) {
3476 if (__predict_false(old.queue != queue ||
3477 (old.flags & PGA_QUEUE_OP_MASK) == 0)) {
3478 counter_u64_add(queue_nops, 1);
3481 KASSERT(old.queue != PQ_NONE || (old.flags & PGA_QUEUE_STATE_MASK) == 0,
3482 ("%s: page %p has unexpected queue state", __func__, m));
3485 if ((old.flags & PGA_DEQUEUE) != 0) {
3486 new.flags &= ~PGA_QUEUE_OP_MASK;
3487 new.queue = PQ_NONE;
3488 if (__predict_true(_vm_page_pqstate_commit_dequeue(pq,
3490 counter_u64_add(queue_ops, 1);
3494 new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD);
3495 if (__predict_true(_vm_page_pqstate_commit_requeue(pq,
3497 counter_u64_add(queue_ops, 1);
3505 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3510 for (i = 0; i < bq->bq_cnt; i++)
3511 vm_pqbatch_process_page(pq, bq->bq_pa[i], queue);
3512 vm_batchqueue_init(bq);
3516 * vm_page_pqbatch_submit: [ internal use only ]
3518 * Enqueue a page in the specified page queue's batched work queue.
3519 * The caller must have encoded the requested operation in the page
3520 * structure's a.flags field.
3523 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
3525 struct vm_batchqueue *bq;
3526 struct vm_pagequeue *pq;
3529 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3530 ("page %p is unmanaged", m));
3531 KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3533 domain = vm_phys_domain(m);
3534 pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue];
3537 bq = DPCPU_PTR(pqbatch[domain][queue]);
3538 if (vm_batchqueue_insert(bq, m)) {
3543 vm_pagequeue_lock(pq);
3545 bq = DPCPU_PTR(pqbatch[domain][queue]);
3546 vm_pqbatch_process(pq, bq, queue);
3547 vm_pqbatch_process_page(pq, m, queue);
3548 vm_pagequeue_unlock(pq);
3553 * vm_page_pqbatch_drain: [ internal use only ]
3555 * Force all per-CPU page queue batch queues to be drained. This is
3556 * intended for use in severe memory shortages, to ensure that pages
3557 * do not remain stuck in the batch queues.
3560 vm_page_pqbatch_drain(void)
3563 struct vm_domain *vmd;
3564 struct vm_pagequeue *pq;
3565 int cpu, domain, queue;
3570 sched_bind(td, cpu);
3573 for (domain = 0; domain < vm_ndomains; domain++) {
3574 vmd = VM_DOMAIN(domain);
3575 for (queue = 0; queue < PQ_COUNT; queue++) {
3576 pq = &vmd->vmd_pagequeues[queue];
3577 vm_pagequeue_lock(pq);
3579 vm_pqbatch_process(pq,
3580 DPCPU_PTR(pqbatch[domain][queue]), queue);
3582 vm_pagequeue_unlock(pq);
3592 * vm_page_dequeue_deferred: [ internal use only ]
3594 * Request removal of the given page from its current page
3595 * queue. Physical removal from the queue may be deferred
3598 * The page must be locked.
3601 vm_page_dequeue_deferred(vm_page_t m)
3603 vm_page_astate_t new, old;
3605 old = vm_page_astate_load(m);
3607 if (old.queue == PQ_NONE) {
3608 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3609 ("%s: page %p has unexpected queue state",
3614 new.flags |= PGA_DEQUEUE;
3615 } while (!vm_page_pqstate_commit_request(m, &old, new));
3621 * Remove the page from whichever page queue it's in, if any, before
3625 vm_page_dequeue(vm_page_t m)
3627 vm_page_astate_t new, old;
3629 old = vm_page_astate_load(m);
3631 if (old.queue == PQ_NONE) {
3632 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3633 ("%s: page %p has unexpected queue state",
3638 new.flags &= ~PGA_QUEUE_OP_MASK;
3639 new.queue = PQ_NONE;
3640 } while (!vm_page_pqstate_commit_dequeue(m, &old, new));
3645 * Schedule the given page for insertion into the specified page queue.
3646 * Physical insertion of the page may be deferred indefinitely.
3649 vm_page_enqueue(vm_page_t m, uint8_t queue)
3652 KASSERT(m->a.queue == PQ_NONE &&
3653 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
3654 ("%s: page %p is already enqueued", __func__, m));
3655 KASSERT(m->ref_count > 0,
3656 ("%s: page %p does not carry any references", __func__, m));
3659 if ((m->a.flags & PGA_REQUEUE) == 0)
3660 vm_page_aflag_set(m, PGA_REQUEUE);
3661 vm_page_pqbatch_submit(m, queue);
3665 * vm_page_free_prep:
3667 * Prepares the given page to be put on the free list,
3668 * disassociating it from any VM object. The caller may return
3669 * the page to the free list only if this function returns true.
3671 * The object must be locked. The page must be locked if it is
3675 vm_page_free_prep(vm_page_t m)
3679 * Synchronize with threads that have dropped a reference to this
3682 atomic_thread_fence_acq();
3684 if (vm_page_sbusied(m))
3685 panic("vm_page_free_prep: freeing shared busy page %p", m);
3687 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3688 if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3691 p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3692 for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3693 KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3694 m, i, (uintmax_t)*p));
3697 if ((m->oflags & VPO_UNMANAGED) == 0) {
3698 KASSERT(!pmap_page_is_mapped(m),
3699 ("vm_page_free_prep: freeing mapped page %p", m));
3700 KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
3701 ("vm_page_free_prep: mapping flags set in page %p", m));
3703 KASSERT(m->a.queue == PQ_NONE,
3704 ("vm_page_free_prep: unmanaged page %p is queued", m));
3706 VM_CNT_INC(v_tfree);
3708 if (m->object != NULL) {
3709 KASSERT(((m->oflags & VPO_UNMANAGED) != 0) ==
3710 ((m->object->flags & OBJ_UNMANAGED) != 0),
3711 ("vm_page_free_prep: managed flag mismatch for page %p",
3713 vm_page_object_remove(m);
3716 * The object reference can be released without an atomic
3719 KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
3720 m->ref_count == VPRC_OBJREF,
3721 ("vm_page_free_prep: page %p has unexpected ref_count %u",
3724 m->ref_count -= VPRC_OBJREF;
3728 if (vm_page_xbusied(m))
3729 panic("vm_page_free_prep: freeing exclusive busy page %p", m);
3732 * If fictitious remove object association and
3735 if ((m->flags & PG_FICTITIOUS) != 0) {
3736 KASSERT(m->ref_count == 1,
3737 ("fictitious page %p is referenced", m));
3738 KASSERT(m->a.queue == PQ_NONE,
3739 ("fictitious page %p is queued", m));
3744 * Pages need not be dequeued before they are returned to the physical
3745 * memory allocator, but they must at least be marked for a deferred
3748 if ((m->oflags & VPO_UNMANAGED) == 0)
3749 vm_page_dequeue_deferred(m);
3754 if (m->ref_count != 0)
3755 panic("vm_page_free_prep: page %p has references", m);
3758 * Restore the default memory attribute to the page.
3760 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3761 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3763 #if VM_NRESERVLEVEL > 0
3765 * Determine whether the page belongs to a reservation. If the page was
3766 * allocated from a per-CPU cache, it cannot belong to a reservation, so
3767 * as an optimization, we avoid the check in that case.
3769 if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
3779 * Returns the given page to the free list, disassociating it
3780 * from any VM object.
3782 * The object must be locked. The page must be locked if it is
3786 vm_page_free_toq(vm_page_t m)
3788 struct vm_domain *vmd;
3791 if (!vm_page_free_prep(m))
3794 vmd = vm_pagequeue_domain(m);
3795 zone = vmd->vmd_pgcache[m->pool].zone;
3796 if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
3800 vm_domain_free_lock(vmd);
3801 vm_phys_free_pages(m, 0);
3802 vm_domain_free_unlock(vmd);
3803 vm_domain_freecnt_inc(vmd, 1);
3807 * vm_page_free_pages_toq:
3809 * Returns a list of pages to the free list, disassociating it
3810 * from any VM object. In other words, this is equivalent to
3811 * calling vm_page_free_toq() for each page of a list of VM objects.
3813 * The objects must be locked. The pages must be locked if it is
3817 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
3822 if (SLIST_EMPTY(free))
3826 while ((m = SLIST_FIRST(free)) != NULL) {
3828 SLIST_REMOVE_HEAD(free, plinks.s.ss);
3829 vm_page_free_toq(m);
3832 if (update_wire_count)
3837 * Mark this page as wired down, preventing reclamation by the page daemon
3838 * or when the containing object is destroyed.
3841 vm_page_wire(vm_page_t m)
3845 KASSERT(m->object != NULL,
3846 ("vm_page_wire: page %p does not belong to an object", m));
3847 if (!vm_page_busied(m) && !vm_object_busied(m->object))
3848 VM_OBJECT_ASSERT_LOCKED(m->object);
3849 KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
3850 VPRC_WIRE_COUNT(m->ref_count) >= 1,
3851 ("vm_page_wire: fictitious page %p has zero wirings", m));
3853 old = atomic_fetchadd_int(&m->ref_count, 1);
3854 KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
3855 ("vm_page_wire: counter overflow for page %p", m));
3856 if (VPRC_WIRE_COUNT(old) == 0) {
3857 if ((m->oflags & VPO_UNMANAGED) == 0)
3858 vm_page_aflag_set(m, PGA_DEQUEUE);
3864 * Attempt to wire a mapped page following a pmap lookup of that page.
3865 * This may fail if a thread is concurrently tearing down mappings of the page.
3866 * The transient failure is acceptable because it translates to the
3867 * failure of the caller pmap_extract_and_hold(), which should be then
3868 * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages().
3871 vm_page_wire_mapped(vm_page_t m)
3878 ("vm_page_wire_mapped: wiring unreferenced page %p", m));
3879 if ((old & VPRC_BLOCKED) != 0)
3881 } while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
3883 if (VPRC_WIRE_COUNT(old) == 0) {
3884 if ((m->oflags & VPO_UNMANAGED) == 0)
3885 vm_page_aflag_set(m, PGA_DEQUEUE);
3892 * Release a wiring reference to a managed page. If the page still belongs to
3893 * an object, update its position in the page queues to reflect the reference.
3894 * If the wiring was the last reference to the page, free the page.
3897 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse)
3901 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3902 ("%s: page %p is unmanaged", __func__, m));
3905 * Update LRU state before releasing the wiring reference.
3906 * Use a release store when updating the reference count to
3907 * synchronize with vm_page_free_prep().
3911 KASSERT(VPRC_WIRE_COUNT(old) > 0,
3912 ("vm_page_unwire: wire count underflow for page %p", m));
3914 if (old > VPRC_OBJREF + 1) {
3916 * The page has at least one other wiring reference. An
3917 * earlier iteration of this loop may have called
3918 * vm_page_release_toq() and cleared PGA_DEQUEUE, so
3919 * re-set it if necessary.
3921 if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0)
3922 vm_page_aflag_set(m, PGA_DEQUEUE);
3923 } else if (old == VPRC_OBJREF + 1) {
3925 * This is the last wiring. Clear PGA_DEQUEUE and
3926 * update the page's queue state to reflect the
3927 * reference. If the page does not belong to an object
3928 * (i.e., the VPRC_OBJREF bit is clear), we only need to
3929 * clear leftover queue state.
3931 vm_page_release_toq(m, nqueue, false);
3932 } else if (old == 1) {
3933 vm_page_aflag_clear(m, PGA_DEQUEUE);
3935 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
3937 if (VPRC_WIRE_COUNT(old) == 1) {
3945 * Release one wiring of the specified page, potentially allowing it to be
3948 * Only managed pages belonging to an object can be paged out. If the number
3949 * of wirings transitions to zero and the page is eligible for page out, then
3950 * the page is added to the specified paging queue. If the released wiring
3951 * represented the last reference to the page, the page is freed.
3953 * A managed page must be locked.
3956 vm_page_unwire(vm_page_t m, uint8_t nqueue)
3959 KASSERT(nqueue < PQ_COUNT,
3960 ("vm_page_unwire: invalid queue %u request for page %p",
3963 if ((m->oflags & VPO_UNMANAGED) != 0) {
3964 if (vm_page_unwire_noq(m) && m->ref_count == 0)
3968 vm_page_unwire_managed(m, nqueue, false);
3972 * Unwire a page without (re-)inserting it into a page queue. It is up
3973 * to the caller to enqueue, requeue, or free the page as appropriate.
3974 * In most cases involving managed pages, vm_page_unwire() should be used
3978 vm_page_unwire_noq(vm_page_t m)
3982 old = vm_page_drop(m, 1);
3983 KASSERT(VPRC_WIRE_COUNT(old) != 0,
3984 ("vm_page_unref: counter underflow for page %p", m));
3985 KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
3986 ("vm_page_unref: missing ref on fictitious page %p", m));
3988 if (VPRC_WIRE_COUNT(old) > 1)
3990 if ((m->oflags & VPO_UNMANAGED) == 0)
3991 vm_page_aflag_clear(m, PGA_DEQUEUE);
3997 * Ensure that the page ends up in the specified page queue. If the page is
3998 * active or being moved to the active queue, ensure that its act_count is
3999 * at least ACT_INIT but do not otherwise mess with it.
4001 * A managed page must be locked.
4003 static __always_inline void
4004 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag)
4006 vm_page_astate_t old, new;
4008 KASSERT(m->ref_count > 0,
4009 ("%s: page %p does not carry any references", __func__, m));
4010 KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD,
4011 ("%s: invalid flags %x", __func__, nflag));
4013 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
4016 old = vm_page_astate_load(m);
4018 if ((old.flags & PGA_DEQUEUE) != 0)
4021 new.flags &= ~PGA_QUEUE_OP_MASK;
4022 if (nqueue == PQ_ACTIVE)
4023 new.act_count = max(old.act_count, ACT_INIT);
4024 if (old.queue == nqueue) {
4025 if (nqueue != PQ_ACTIVE)
4031 } while (!vm_page_pqstate_commit(m, &old, new));
4035 * Put the specified page on the active list (if appropriate).
4038 vm_page_activate(vm_page_t m)
4041 vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE);
4045 * Move the specified page to the tail of the inactive queue, or requeue
4046 * the page if it is already in the inactive queue.
4049 vm_page_deactivate(vm_page_t m)
4052 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE);
4056 vm_page_deactivate_noreuse(vm_page_t m)
4059 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD);
4063 * Put a page in the laundry, or requeue it if it is already there.
4066 vm_page_launder(vm_page_t m)
4069 vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE);
4073 * Put a page in the PQ_UNSWAPPABLE holding queue.
4076 vm_page_unswappable(vm_page_t m)
4079 KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0,
4080 ("page %p already unswappable", m));
4083 vm_page_enqueue(m, PQ_UNSWAPPABLE);
4087 * Release a page back to the page queues in preparation for unwiring.
4090 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse)
4092 vm_page_astate_t old, new;
4096 * Use a check of the valid bits to determine whether we should
4097 * accelerate reclamation of the page. The object lock might not be
4098 * held here, in which case the check is racy. At worst we will either
4099 * accelerate reclamation of a valid page and violate LRU, or
4100 * unnecessarily defer reclamation of an invalid page.
4102 * If we were asked to not cache the page, place it near the head of the
4103 * inactive queue so that is reclaimed sooner.
4105 if (noreuse || m->valid == 0) {
4106 nqueue = PQ_INACTIVE;
4107 nflag = PGA_REQUEUE_HEAD;
4109 nflag = PGA_REQUEUE;
4112 old = vm_page_astate_load(m);
4117 * If the page is already in the active queue and we are not
4118 * trying to accelerate reclamation, simply mark it as
4119 * referenced and avoid any queue operations.
4121 new.flags &= ~PGA_QUEUE_OP_MASK;
4122 if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE)
4123 new.flags |= PGA_REFERENCED;
4128 } while (!vm_page_pqstate_commit(m, &old, new));
4132 * Unwire a page and either attempt to free it or re-add it to the page queues.
4135 vm_page_release(vm_page_t m, int flags)
4139 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4140 ("vm_page_release: page %p is unmanaged", m));
4142 if ((flags & VPR_TRYFREE) != 0) {
4144 object = (vm_object_t)atomic_load_ptr(&m->object);
4147 /* Depends on type-stability. */
4148 if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object))
4150 if (object == m->object) {
4151 vm_page_release_locked(m, flags);
4152 VM_OBJECT_WUNLOCK(object);
4155 VM_OBJECT_WUNLOCK(object);
4158 vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0);
4161 /* See vm_page_release(). */
4163 vm_page_release_locked(vm_page_t m, int flags)
4166 VM_OBJECT_ASSERT_WLOCKED(m->object);
4167 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4168 ("vm_page_release_locked: page %p is unmanaged", m));
4170 if (vm_page_unwire_noq(m)) {
4171 if ((flags & VPR_TRYFREE) != 0 &&
4172 (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
4173 m->dirty == 0 && vm_page_tryxbusy(m)) {
4176 vm_page_release_toq(m, PQ_INACTIVE, flags != 0);
4182 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
4186 KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
4187 ("vm_page_try_blocked_op: page %p has no object", m));
4188 KASSERT(vm_page_busied(m),
4189 ("vm_page_try_blocked_op: page %p is not busy", m));
4190 VM_OBJECT_ASSERT_LOCKED(m->object);
4195 ("vm_page_try_blocked_op: page %p has no references", m));
4196 if (VPRC_WIRE_COUNT(old) != 0)
4198 } while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
4203 * If the object is read-locked, new wirings may be created via an
4206 old = vm_page_drop(m, VPRC_BLOCKED);
4207 KASSERT(!VM_OBJECT_WOWNED(m->object) ||
4208 old == (VPRC_BLOCKED | VPRC_OBJREF),
4209 ("vm_page_try_blocked_op: unexpected refcount value %u for %p",
4215 * Atomically check for wirings and remove all mappings of the page.
4218 vm_page_try_remove_all(vm_page_t m)
4221 return (vm_page_try_blocked_op(m, pmap_remove_all));
4225 * Atomically check for wirings and remove all writeable mappings of the page.
4228 vm_page_try_remove_write(vm_page_t m)
4231 return (vm_page_try_blocked_op(m, pmap_remove_write));
4237 * Apply the specified advice to the given page.
4239 * The object and page must be locked.
4242 vm_page_advise(vm_page_t m, int advice)
4245 VM_OBJECT_ASSERT_WLOCKED(m->object);
4246 if (advice == MADV_FREE)
4248 * Mark the page clean. This will allow the page to be freed
4249 * without first paging it out. MADV_FREE pages are often
4250 * quickly reused by malloc(3), so we do not do anything that
4251 * would result in a page fault on a later access.
4254 else if (advice != MADV_DONTNEED) {
4255 if (advice == MADV_WILLNEED)
4256 vm_page_activate(m);
4260 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
4264 * Clear any references to the page. Otherwise, the page daemon will
4265 * immediately reactivate the page.
4267 vm_page_aflag_clear(m, PGA_REFERENCED);
4270 * Place clean pages near the head of the inactive queue rather than
4271 * the tail, thus defeating the queue's LRU operation and ensuring that
4272 * the page will be reused quickly. Dirty pages not already in the
4273 * laundry are moved there.
4276 vm_page_deactivate_noreuse(m);
4277 else if (!vm_page_in_laundry(m))
4282 vm_page_grab_pflags(int allocflags)
4286 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4287 (allocflags & VM_ALLOC_WIRED) != 0,
4288 ("vm_page_grab_pflags: the pages must be busied or wired"));
4289 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4290 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4291 ("vm_page_grab_pflags: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY "
4293 pflags = allocflags &
4294 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL |
4296 if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4297 pflags |= VM_ALLOC_WAITFAIL;
4298 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4299 pflags |= VM_ALLOC_SBUSY;
4305 * Grab a page, waiting until we are waken up due to the page
4306 * changing state. We keep on waiting, if the page continues
4307 * to be in the object. If the page doesn't exist, first allocate it
4308 * and then conditionally zero it.
4310 * This routine may sleep.
4312 * The object must be locked on entry. The lock will, however, be released
4313 * and reacquired if the routine sleeps.
4316 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
4321 VM_OBJECT_ASSERT_WLOCKED(object);
4322 pflags = vm_page_grab_pflags(allocflags);
4324 if ((m = vm_page_lookup(object, pindex)) != NULL) {
4325 if (!vm_page_acquire_flags(m, allocflags)) {
4326 if (vm_page_busy_sleep_flags(object, m, "pgrbwt",
4333 if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4335 m = vm_page_alloc(object, pindex, pflags);
4337 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4341 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
4345 if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4346 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4355 * Grab a page and make it valid, paging in if necessary. Pages missing from
4356 * their pager are zero filled and validated. If a VM_ALLOC_COUNT is supplied
4357 * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought
4358 * in simultaneously. Additional pages will be left on a paging queue but
4359 * will neither be wired nor busy regardless of allocflags.
4362 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags)
4365 vm_page_t ma[VM_INITIAL_PAGEIN];
4367 int after, i, pflags, rv;
4369 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4370 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4371 ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4372 KASSERT((allocflags &
4373 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4374 ("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
4375 VM_OBJECT_ASSERT_WLOCKED(object);
4376 pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY);
4377 pflags |= VM_ALLOC_WAITFAIL;
4381 if ((m = vm_page_lookup(object, pindex)) != NULL) {
4383 * If the page is fully valid it can only become invalid
4384 * with the object lock held. If it is not valid it can
4385 * become valid with the busy lock held. Therefore, we
4386 * may unnecessarily lock the exclusive busy here if we
4387 * race with I/O completion not using the object lock.
4388 * However, we will not end up with an invalid page and a
4391 if (!vm_page_all_valid(m) ||
4392 (allocflags & (VM_ALLOC_IGN_SBUSY | VM_ALLOC_SBUSY)) == 0) {
4393 sleep = !vm_page_tryxbusy(m);
4396 sleep = !vm_page_trysbusy(m);
4398 (void)vm_page_busy_sleep_flags(object, m, "pgrbwt",
4402 if ((allocflags & VM_ALLOC_NOCREAT) != 0 &&
4403 !vm_page_all_valid(m)) {
4409 return (VM_PAGER_FAIL);
4411 if ((allocflags & VM_ALLOC_WIRED) != 0)
4413 if (vm_page_all_valid(m))
4415 } else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4417 return (VM_PAGER_FAIL);
4418 } else if ((m = vm_page_alloc(object, pindex, pflags)) != NULL) {
4424 vm_page_assert_xbusied(m);
4426 if (vm_pager_has_page(object, pindex, NULL, &after)) {
4427 after = MIN(after, VM_INITIAL_PAGEIN);
4428 after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT);
4429 after = MAX(after, 1);
4431 for (i = 1; i < after; i++) {
4432 if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) {
4433 if (ma[i]->valid || !vm_page_tryxbusy(ma[i]))
4436 ma[i] = vm_page_alloc(object, m->pindex + i,
4443 rv = vm_pager_get_pages(object, ma, after, NULL, NULL);
4444 /* Pager may have replaced a page. */
4446 if (rv != VM_PAGER_OK) {
4447 if ((allocflags & VM_ALLOC_WIRED) != 0)
4448 vm_page_unwire_noq(m);
4449 for (i = 0; i < after; i++) {
4450 if (!vm_page_wired(ma[i]))
4451 vm_page_free(ma[i]);
4453 vm_page_xunbusy(ma[i]);
4458 for (i = 1; i < after; i++)
4459 vm_page_readahead_finish(ma[i]);
4460 MPASS(vm_page_all_valid(m));
4462 vm_page_zero_invalid(m, TRUE);
4465 if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4471 if ((allocflags & VM_ALLOC_SBUSY) != 0 && xbusy)
4472 vm_page_busy_downgrade(m);
4474 return (VM_PAGER_OK);
4478 * Return the specified range of pages from the given object. For each
4479 * page offset within the range, if a page already exists within the object
4480 * at that offset and it is busy, then wait for it to change state. If,
4481 * instead, the page doesn't exist, then allocate it.
4483 * The caller must always specify an allocation class.
4485 * allocation classes:
4486 * VM_ALLOC_NORMAL normal process request
4487 * VM_ALLOC_SYSTEM system *really* needs the pages
4489 * The caller must always specify that the pages are to be busied and/or
4492 * optional allocation flags:
4493 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages
4494 * VM_ALLOC_NOBUSY do not exclusive busy the page
4495 * VM_ALLOC_NOWAIT do not sleep
4496 * VM_ALLOC_SBUSY set page to sbusy state
4497 * VM_ALLOC_WIRED wire the pages
4498 * VM_ALLOC_ZERO zero and validate any invalid pages
4500 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it
4501 * may return a partial prefix of the requested range.
4504 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
4505 vm_page_t *ma, int count)
4511 VM_OBJECT_ASSERT_WLOCKED(object);
4512 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
4513 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
4515 pflags = vm_page_grab_pflags(allocflags);
4521 m = vm_radix_lookup_le(&object->rtree, pindex + i);
4522 if (m == NULL || m->pindex != pindex + i) {
4526 mpred = TAILQ_PREV(m, pglist, listq);
4527 for (; i < count; i++) {
4529 if (!vm_page_acquire_flags(m, allocflags)) {
4530 if (vm_page_busy_sleep_flags(object, m,
4531 "grbmaw", allocflags))
4536 if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4538 m = vm_page_alloc_after(object, pindex + i,
4539 pflags | VM_ALLOC_COUNT(count - i), mpred);
4541 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4546 if (vm_page_none_valid(m) &&
4547 (allocflags & VM_ALLOC_ZERO) != 0) {
4548 if ((m->flags & PG_ZERO) == 0)
4552 if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4553 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4559 m = vm_page_next(m);
4565 * Mapping function for valid or dirty bits in a page.
4567 * Inputs are required to range within a page.
4570 vm_page_bits(int base, int size)
4576 base + size <= PAGE_SIZE,
4577 ("vm_page_bits: illegal base/size %d/%d", base, size)
4580 if (size == 0) /* handle degenerate case */
4583 first_bit = base >> DEV_BSHIFT;
4584 last_bit = (base + size - 1) >> DEV_BSHIFT;
4586 return (((vm_page_bits_t)2 << last_bit) -
4587 ((vm_page_bits_t)1 << first_bit));
4591 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set)
4594 #if PAGE_SIZE == 32768
4595 atomic_set_64((uint64_t *)bits, set);
4596 #elif PAGE_SIZE == 16384
4597 atomic_set_32((uint32_t *)bits, set);
4598 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16)
4599 atomic_set_16((uint16_t *)bits, set);
4600 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8)
4601 atomic_set_8((uint8_t *)bits, set);
4602 #else /* PAGE_SIZE <= 8192 */
4606 addr = (uintptr_t)bits;
4608 * Use a trick to perform a 32-bit atomic on the
4609 * containing aligned word, to not depend on the existence
4610 * of atomic_{set, clear}_{8, 16}.
4612 shift = addr & (sizeof(uint32_t) - 1);
4613 #if BYTE_ORDER == BIG_ENDIAN
4614 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4618 addr &= ~(sizeof(uint32_t) - 1);
4619 atomic_set_32((uint32_t *)addr, set << shift);
4620 #endif /* PAGE_SIZE */
4624 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear)
4627 #if PAGE_SIZE == 32768
4628 atomic_clear_64((uint64_t *)bits, clear);
4629 #elif PAGE_SIZE == 16384
4630 atomic_clear_32((uint32_t *)bits, clear);
4631 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16)
4632 atomic_clear_16((uint16_t *)bits, clear);
4633 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8)
4634 atomic_clear_8((uint8_t *)bits, clear);
4635 #else /* PAGE_SIZE <= 8192 */
4639 addr = (uintptr_t)bits;
4641 * Use a trick to perform a 32-bit atomic on the
4642 * containing aligned word, to not depend on the existence
4643 * of atomic_{set, clear}_{8, 16}.
4645 shift = addr & (sizeof(uint32_t) - 1);
4646 #if BYTE_ORDER == BIG_ENDIAN
4647 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4651 addr &= ~(sizeof(uint32_t) - 1);
4652 atomic_clear_32((uint32_t *)addr, clear << shift);
4653 #endif /* PAGE_SIZE */
4656 static inline vm_page_bits_t
4657 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits)
4659 #if PAGE_SIZE == 32768
4663 while (atomic_fcmpset_64(bits, &old, newbits) == 0);
4665 #elif PAGE_SIZE == 16384
4669 while (atomic_fcmpset_32(bits, &old, newbits) == 0);
4671 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16)
4675 while (atomic_fcmpset_16(bits, &old, newbits) == 0);
4677 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8)
4681 while (atomic_fcmpset_8(bits, &old, newbits) == 0);
4683 #else /* PAGE_SIZE <= 4096*/
4685 uint32_t old, new, mask;
4688 addr = (uintptr_t)bits;
4690 * Use a trick to perform a 32-bit atomic on the
4691 * containing aligned word, to not depend on the existence
4692 * of atomic_{set, swap, clear}_{8, 16}.
4694 shift = addr & (sizeof(uint32_t) - 1);
4695 #if BYTE_ORDER == BIG_ENDIAN
4696 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4700 addr &= ~(sizeof(uint32_t) - 1);
4701 mask = VM_PAGE_BITS_ALL << shift;
4706 new |= newbits << shift;
4707 } while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0);
4708 return (old >> shift);
4709 #endif /* PAGE_SIZE */
4713 * vm_page_set_valid_range:
4715 * Sets portions of a page valid. The arguments are expected
4716 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4717 * of any partial chunks touched by the range. The invalid portion of
4718 * such chunks will be zeroed.
4720 * (base + size) must be less then or equal to PAGE_SIZE.
4723 vm_page_set_valid_range(vm_page_t m, int base, int size)
4726 vm_page_bits_t pagebits;
4728 vm_page_assert_busied(m);
4729 if (size == 0) /* handle degenerate case */
4733 * If the base is not DEV_BSIZE aligned and the valid
4734 * bit is clear, we have to zero out a portion of the
4737 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4738 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
4739 pmap_zero_page_area(m, frag, base - frag);
4742 * If the ending offset is not DEV_BSIZE aligned and the
4743 * valid bit is clear, we have to zero out a portion of
4746 endoff = base + size;
4747 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4748 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
4749 pmap_zero_page_area(m, endoff,
4750 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4753 * Assert that no previously invalid block that is now being validated
4756 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
4757 ("vm_page_set_valid_range: page %p is dirty", m));
4760 * Set valid bits inclusive of any overlap.
4762 pagebits = vm_page_bits(base, size);
4763 if (vm_page_xbusied(m))
4764 m->valid |= pagebits;
4766 vm_page_bits_set(m, &m->valid, pagebits);
4770 * Set the page dirty bits and free the invalid swap space if
4771 * present. Returns the previous dirty bits.
4774 vm_page_set_dirty(vm_page_t m)
4778 VM_PAGE_OBJECT_BUSY_ASSERT(m);
4780 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) {
4782 m->dirty = VM_PAGE_BITS_ALL;
4784 old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL);
4785 if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0)
4786 vm_pager_page_unswapped(m);
4792 * Clear the given bits from the specified page's dirty field.
4794 static __inline void
4795 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
4798 vm_page_assert_busied(m);
4801 * If the page is xbusied and not write mapped we are the
4802 * only thread that can modify dirty bits. Otherwise, The pmap
4803 * layer can call vm_page_dirty() without holding a distinguished
4804 * lock. The combination of page busy and atomic operations
4805 * suffice to guarantee consistency of the page dirty field.
4807 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
4808 m->dirty &= ~pagebits;
4810 vm_page_bits_clear(m, &m->dirty, pagebits);
4814 * vm_page_set_validclean:
4816 * Sets portions of a page valid and clean. The arguments are expected
4817 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4818 * of any partial chunks touched by the range. The invalid portion of
4819 * such chunks will be zero'd.
4821 * (base + size) must be less then or equal to PAGE_SIZE.
4824 vm_page_set_validclean(vm_page_t m, int base, int size)
4826 vm_page_bits_t oldvalid, pagebits;
4829 vm_page_assert_busied(m);
4830 if (size == 0) /* handle degenerate case */
4834 * If the base is not DEV_BSIZE aligned and the valid
4835 * bit is clear, we have to zero out a portion of the
4838 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4839 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
4840 pmap_zero_page_area(m, frag, base - frag);
4843 * If the ending offset is not DEV_BSIZE aligned and the
4844 * valid bit is clear, we have to zero out a portion of
4847 endoff = base + size;
4848 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4849 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
4850 pmap_zero_page_area(m, endoff,
4851 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4854 * Set valid, clear dirty bits. If validating the entire
4855 * page we can safely clear the pmap modify bit. We also
4856 * use this opportunity to clear the PGA_NOSYNC flag. If a process
4857 * takes a write fault on a MAP_NOSYNC memory area the flag will
4860 * We set valid bits inclusive of any overlap, but we can only
4861 * clear dirty bits for DEV_BSIZE chunks that are fully within
4864 oldvalid = m->valid;
4865 pagebits = vm_page_bits(base, size);
4866 if (vm_page_xbusied(m))
4867 m->valid |= pagebits;
4869 vm_page_bits_set(m, &m->valid, pagebits);
4871 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
4872 frag = DEV_BSIZE - frag;
4878 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
4880 if (base == 0 && size == PAGE_SIZE) {
4882 * The page can only be modified within the pmap if it is
4883 * mapped, and it can only be mapped if it was previously
4886 if (oldvalid == VM_PAGE_BITS_ALL)
4888 * Perform the pmap_clear_modify() first. Otherwise,
4889 * a concurrent pmap operation, such as
4890 * pmap_protect(), could clear a modification in the
4891 * pmap and set the dirty field on the page before
4892 * pmap_clear_modify() had begun and after the dirty
4893 * field was cleared here.
4895 pmap_clear_modify(m);
4897 vm_page_aflag_clear(m, PGA_NOSYNC);
4898 } else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m))
4899 m->dirty &= ~pagebits;
4901 vm_page_clear_dirty_mask(m, pagebits);
4905 vm_page_clear_dirty(vm_page_t m, int base, int size)
4908 vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
4912 * vm_page_set_invalid:
4914 * Invalidates DEV_BSIZE'd chunks within a page. Both the
4915 * valid and dirty bits for the effected areas are cleared.
4918 vm_page_set_invalid(vm_page_t m, int base, int size)
4920 vm_page_bits_t bits;
4924 * The object lock is required so that pages can't be mapped
4925 * read-only while we're in the process of invalidating them.
4928 VM_OBJECT_ASSERT_WLOCKED(object);
4929 vm_page_assert_busied(m);
4931 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
4932 size >= object->un_pager.vnp.vnp_size)
4933 bits = VM_PAGE_BITS_ALL;
4935 bits = vm_page_bits(base, size);
4936 if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0)
4938 KASSERT((bits == 0 && vm_page_all_valid(m)) ||
4939 !pmap_page_is_mapped(m),
4940 ("vm_page_set_invalid: page %p is mapped", m));
4941 if (vm_page_xbusied(m)) {
4945 vm_page_bits_clear(m, &m->valid, bits);
4946 vm_page_bits_clear(m, &m->dirty, bits);
4953 * Invalidates the entire page. The page must be busy, unmapped, and
4954 * the enclosing object must be locked. The object locks protects
4955 * against concurrent read-only pmap enter which is done without
4959 vm_page_invalid(vm_page_t m)
4962 vm_page_assert_busied(m);
4963 VM_OBJECT_ASSERT_LOCKED(m->object);
4964 MPASS(!pmap_page_is_mapped(m));
4966 if (vm_page_xbusied(m))
4969 vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL);
4973 * vm_page_zero_invalid()
4975 * The kernel assumes that the invalid portions of a page contain
4976 * garbage, but such pages can be mapped into memory by user code.
4977 * When this occurs, we must zero out the non-valid portions of the
4978 * page so user code sees what it expects.
4980 * Pages are most often semi-valid when the end of a file is mapped
4981 * into memory and the file's size is not page aligned.
4984 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
4990 * Scan the valid bits looking for invalid sections that
4991 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the
4992 * valid bit may be set ) have already been zeroed by
4993 * vm_page_set_validclean().
4995 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
4996 if (i == (PAGE_SIZE / DEV_BSIZE) ||
4997 (m->valid & ((vm_page_bits_t)1 << i))) {
4999 pmap_zero_page_area(m,
5000 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
5007 * setvalid is TRUE when we can safely set the zero'd areas
5008 * as being valid. We can do this if there are no cache consistancy
5009 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
5018 * Is (partial) page valid? Note that the case where size == 0
5019 * will return FALSE in the degenerate case where the page is
5020 * entirely invalid, and TRUE otherwise.
5022 * Some callers envoke this routine without the busy lock held and
5023 * handle races via higher level locks. Typical callers should
5024 * hold a busy lock to prevent invalidation.
5027 vm_page_is_valid(vm_page_t m, int base, int size)
5029 vm_page_bits_t bits;
5031 bits = vm_page_bits(base, size);
5032 return (m->valid != 0 && (m->valid & bits) == bits);
5036 * Returns true if all of the specified predicates are true for the entire
5037 * (super)page and false otherwise.
5040 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
5046 if (skip_m != NULL && skip_m->object != object)
5048 VM_OBJECT_ASSERT_LOCKED(object);
5049 npages = atop(pagesizes[m->psind]);
5052 * The physically contiguous pages that make up a superpage, i.e., a
5053 * page with a page size index ("psind") greater than zero, will
5054 * occupy adjacent entries in vm_page_array[].
5056 for (i = 0; i < npages; i++) {
5057 /* Always test object consistency, including "skip_m". */
5058 if (m[i].object != object)
5060 if (&m[i] == skip_m)
5062 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
5064 if ((flags & PS_ALL_DIRTY) != 0) {
5066 * Calling vm_page_test_dirty() or pmap_is_modified()
5067 * might stop this case from spuriously returning
5068 * "false". However, that would require a write lock
5069 * on the object containing "m[i]".
5071 if (m[i].dirty != VM_PAGE_BITS_ALL)
5074 if ((flags & PS_ALL_VALID) != 0 &&
5075 m[i].valid != VM_PAGE_BITS_ALL)
5082 * Set the page's dirty bits if the page is modified.
5085 vm_page_test_dirty(vm_page_t m)
5088 vm_page_assert_busied(m);
5089 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
5094 vm_page_valid(vm_page_t m)
5097 vm_page_assert_busied(m);
5098 if (vm_page_xbusied(m))
5099 m->valid = VM_PAGE_BITS_ALL;
5101 vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL);
5105 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
5108 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
5112 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
5115 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
5119 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
5122 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
5125 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
5127 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
5130 vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
5134 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
5137 mtx_assert_(vm_page_lockptr(m), a, file, line);
5143 vm_page_object_busy_assert(vm_page_t m)
5147 * Certain of the page's fields may only be modified by the
5148 * holder of a page or object busy.
5150 if (m->object != NULL && !vm_page_busied(m))
5151 VM_OBJECT_ASSERT_BUSY(m->object);
5155 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits)
5158 if ((bits & PGA_WRITEABLE) == 0)
5162 * The PGA_WRITEABLE flag can only be set if the page is
5163 * managed, is exclusively busied or the object is locked.
5164 * Currently, this flag is only set by pmap_enter().
5166 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5167 ("PGA_WRITEABLE on unmanaged page"));
5168 if (!vm_page_xbusied(m))
5169 VM_OBJECT_ASSERT_BUSY(m->object);
5173 #include "opt_ddb.h"
5175 #include <sys/kernel.h>
5177 #include <ddb/ddb.h>
5179 DB_SHOW_COMMAND(page, vm_page_print_page_info)
5182 db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
5183 db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
5184 db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
5185 db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
5186 db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
5187 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
5188 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
5189 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
5190 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
5193 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
5197 db_printf("pq_free %d\n", vm_free_count());
5198 for (dom = 0; dom < vm_ndomains; dom++) {
5200 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
5202 vm_dom[dom].vmd_page_count,
5203 vm_dom[dom].vmd_free_count,
5204 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
5205 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
5206 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
5207 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
5211 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
5214 boolean_t phys, virt;
5217 db_printf("show pginfo addr\n");
5221 phys = strchr(modif, 'p') != NULL;
5222 virt = strchr(modif, 'v') != NULL;
5224 m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
5226 m = PHYS_TO_VM_PAGE(addr);
5228 m = (vm_page_t)addr;
5230 "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref %u\n"
5231 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
5232 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
5233 m->a.queue, m->ref_count, m->a.flags, m->oflags,
5234 m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty);