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 queue_ops = EARLY_COUNTER;
138 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops,
139 CTLFLAG_RD, &queue_ops,
140 "Number of batched queue operations");
142 static counter_u64_t queue_nops = EARLY_COUNTER;
143 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops,
144 CTLFLAG_RD, &queue_nops,
145 "Number of batched queue operations with no effects");
148 counter_startup(void)
151 queue_ops = counter_u64_alloc(M_WAITOK);
152 queue_nops = counter_u64_alloc(M_WAITOK);
154 SYSINIT(page_counters, SI_SUB_CPU, SI_ORDER_ANY, counter_startup, NULL);
157 * bogus page -- for I/O to/from partially complete buffers,
158 * or for paging into sparsely invalid regions.
160 vm_page_t bogus_page;
162 vm_page_t vm_page_array;
163 long vm_page_array_size;
166 static int boot_pages;
167 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
169 "number of pages allocated for bootstrapping the VM system");
171 static int pa_tryrelock_restart;
172 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
173 &pa_tryrelock_restart, 0, "Number of tryrelock restarts");
175 static TAILQ_HEAD(, vm_page) blacklist_head;
176 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
177 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
178 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
180 static uma_zone_t fakepg_zone;
182 static void vm_page_alloc_check(vm_page_t m);
183 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
184 static void vm_page_dequeue_complete(vm_page_t m);
185 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
186 static void vm_page_init(void *dummy);
187 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
188 vm_pindex_t pindex, vm_page_t mpred);
189 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
191 static void vm_page_mvqueue(vm_page_t m, uint8_t queue);
192 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
193 vm_page_t m_run, vm_paddr_t high);
194 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
196 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain,
198 static void vm_page_zone_release(void *arg, void **store, int cnt);
200 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
203 vm_page_init(void *dummy)
206 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
207 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
208 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
209 VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
213 * The cache page zone is initialized later since we need to be able to allocate
214 * pages before UMA is fully initialized.
217 vm_page_init_cache_zones(void *dummy __unused)
219 struct vm_domain *vmd;
220 struct vm_pgcache *pgcache;
223 for (domain = 0; domain < vm_ndomains; domain++) {
224 vmd = VM_DOMAIN(domain);
227 * Don't allow the page caches to take up more than .25% of
230 if (vmd->vmd_page_count / 400 < 256 * mp_ncpus * VM_NFREEPOOL)
232 for (pool = 0; pool < VM_NFREEPOOL; pool++) {
233 pgcache = &vmd->vmd_pgcache[pool];
234 pgcache->domain = domain;
235 pgcache->pool = pool;
236 pgcache->zone = uma_zcache_create("vm pgcache",
237 sizeof(struct vm_page), NULL, NULL, NULL, NULL,
238 vm_page_zone_import, vm_page_zone_release, pgcache,
239 UMA_ZONE_MAXBUCKET | UMA_ZONE_VM);
240 (void)uma_zone_set_maxcache(pgcache->zone, 0);
244 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
246 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
247 #if PAGE_SIZE == 32768
249 CTASSERT(sizeof(u_long) >= 8);
254 * Try to acquire a physical address lock while a pmap is locked. If we
255 * fail to trylock we unlock and lock the pmap directly and cache the
256 * locked pa in *locked. The caller should then restart their loop in case
257 * the virtual to physical mapping has changed.
260 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
267 PA_LOCK_ASSERT(lockpa, MA_OWNED);
268 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
275 atomic_add_int(&pa_tryrelock_restart, 1);
284 * Sets the page size, perhaps based upon the memory
285 * size. Must be called before any use of page-size
286 * dependent functions.
289 vm_set_page_size(void)
291 if (vm_cnt.v_page_size == 0)
292 vm_cnt.v_page_size = PAGE_SIZE;
293 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
294 panic("vm_set_page_size: page size not a power of two");
298 * vm_page_blacklist_next:
300 * Find the next entry in the provided string of blacklist
301 * addresses. Entries are separated by space, comma, or newline.
302 * If an invalid integer is encountered then the rest of the
303 * string is skipped. Updates the list pointer to the next
304 * character, or NULL if the string is exhausted or invalid.
307 vm_page_blacklist_next(char **list, char *end)
312 if (list == NULL || *list == NULL)
320 * If there's no end pointer then the buffer is coming from
321 * the kenv and we know it's null-terminated.
324 end = *list + strlen(*list);
326 /* Ensure that strtoq() won't walk off the end */
328 if (*end == '\n' || *end == ' ' || *end == ',')
331 printf("Blacklist not terminated, skipping\n");
337 for (pos = *list; *pos != '\0'; pos = cp) {
338 bad = strtoq(pos, &cp, 0);
339 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
348 if (*cp == '\0' || ++cp >= end)
352 return (trunc_page(bad));
354 printf("Garbage in RAM blacklist, skipping\n");
360 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
362 struct vm_domain *vmd;
366 m = vm_phys_paddr_to_vm_page(pa);
368 return (true); /* page does not exist, no failure */
370 vmd = vm_pagequeue_domain(m);
371 vm_domain_free_lock(vmd);
372 ret = vm_phys_unfree_page(m);
373 vm_domain_free_unlock(vmd);
375 vm_domain_freecnt_inc(vmd, -1);
376 TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
378 printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
384 * vm_page_blacklist_check:
386 * Iterate through the provided string of blacklist addresses, pulling
387 * each entry out of the physical allocator free list and putting it
388 * onto a list for reporting via the vm.page_blacklist sysctl.
391 vm_page_blacklist_check(char *list, char *end)
397 while (next != NULL) {
398 if ((pa = vm_page_blacklist_next(&next, end)) == 0)
400 vm_page_blacklist_add(pa, bootverbose);
405 * vm_page_blacklist_load:
407 * Search for a special module named "ram_blacklist". It'll be a
408 * plain text file provided by the user via the loader directive
412 vm_page_blacklist_load(char **list, char **end)
421 mod = preload_search_by_type("ram_blacklist");
423 ptr = preload_fetch_addr(mod);
424 len = preload_fetch_size(mod);
435 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
442 error = sysctl_wire_old_buffer(req, 0);
445 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
446 TAILQ_FOREACH(m, &blacklist_head, listq) {
447 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
448 (uintmax_t)m->phys_addr);
451 error = sbuf_finish(&sbuf);
457 * Initialize a dummy page for use in scans of the specified paging queue.
458 * In principle, this function only needs to set the flag PG_MARKER.
459 * Nonetheless, it write busies the page as a safety precaution.
462 vm_page_init_marker(vm_page_t marker, int queue, uint8_t aflags)
465 bzero(marker, sizeof(*marker));
466 marker->flags = PG_MARKER;
467 marker->aflags = aflags;
468 marker->busy_lock = VPB_SINGLE_EXCLUSIVER;
469 marker->queue = queue;
473 vm_page_domain_init(int domain)
475 struct vm_domain *vmd;
476 struct vm_pagequeue *pq;
479 vmd = VM_DOMAIN(domain);
480 bzero(vmd, sizeof(*vmd));
481 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
482 "vm inactive pagequeue";
483 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
484 "vm active pagequeue";
485 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
486 "vm laundry pagequeue";
487 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
488 "vm unswappable pagequeue";
489 vmd->vmd_domain = domain;
490 vmd->vmd_page_count = 0;
491 vmd->vmd_free_count = 0;
493 vmd->vmd_oom = FALSE;
494 for (i = 0; i < PQ_COUNT; i++) {
495 pq = &vmd->vmd_pagequeues[i];
496 TAILQ_INIT(&pq->pq_pl);
497 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
498 MTX_DEF | MTX_DUPOK);
500 vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
502 mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
503 mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
504 snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
507 * inacthead is used to provide FIFO ordering for LRU-bypassing
510 vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
511 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
512 &vmd->vmd_inacthead, plinks.q);
515 * The clock pages are used to implement active queue scanning without
516 * requeues. Scans start at clock[0], which is advanced after the scan
517 * ends. When the two clock hands meet, they are reset and scanning
518 * resumes from the head of the queue.
520 vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
521 vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
522 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
523 &vmd->vmd_clock[0], plinks.q);
524 TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
525 &vmd->vmd_clock[1], plinks.q);
529 * Initialize a physical page in preparation for adding it to the free
533 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
538 m->busy_lock = VPB_UNBUSIED;
539 m->flags = m->aflags = 0;
544 m->order = VM_NFREEORDER;
545 m->pool = VM_FREEPOOL_DEFAULT;
546 m->valid = m->dirty = 0;
550 #ifndef PMAP_HAS_PAGE_ARRAY
552 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range)
557 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
558 * However, because this page is allocated from KVM, out-of-bounds
559 * accesses using the direct map will not be trapped.
564 * Allocate physical memory for the page structures, and map it.
566 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
567 vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end,
568 VM_PROT_READ | VM_PROT_WRITE);
569 vm_page_array_size = page_range;
578 * Initializes the resident memory module. Allocates physical memory for
579 * bootstrapping UMA and some data structures that are used to manage
580 * physical pages. Initializes these structures, and populates the free
584 vm_page_startup(vm_offset_t vaddr)
586 struct vm_phys_seg *seg;
588 char *list, *listend;
590 vm_paddr_t end, high_avail, low_avail, new_end, page_range, size;
591 vm_paddr_t last_pa, pa;
593 int biggestone, i, segind;
597 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
601 vaddr = round_page(vaddr);
603 vm_phys_early_startup();
604 biggestone = vm_phys_avail_largest();
605 end = phys_avail[biggestone+1];
608 * Initialize the page and queue locks.
610 mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
611 for (i = 0; i < PA_LOCK_COUNT; i++)
612 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
613 for (i = 0; i < vm_ndomains; i++)
614 vm_page_domain_init(i);
617 * Allocate memory for use when boot strapping the kernel memory
618 * allocator. Tell UMA how many zones we are going to create
619 * before going fully functional. UMA will add its zones.
621 * VM startup zones: vmem, vmem_btag, VM OBJECT, RADIX NODE, MAP,
622 * KMAP ENTRY, MAP ENTRY, VMSPACE.
624 boot_pages = uma_startup_count(8);
626 #ifndef UMA_MD_SMALL_ALLOC
627 /* vmem_startup() calls uma_prealloc(). */
628 boot_pages += vmem_startup_count();
629 /* vm_map_startup() calls uma_prealloc(). */
630 boot_pages += howmany(MAX_KMAP,
631 UMA_SLAB_SPACE / sizeof(struct vm_map));
634 * Before going fully functional kmem_init() does allocation
635 * from "KMAP ENTRY" and vmem_create() does allocation from "vmem".
640 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
641 * manually fetch the value.
643 TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
644 new_end = end - (boot_pages * UMA_SLAB_SIZE);
645 new_end = trunc_page(new_end);
646 mapped = pmap_map(&vaddr, new_end, end,
647 VM_PROT_READ | VM_PROT_WRITE);
648 bzero((void *)mapped, end - new_end);
649 uma_startup((void *)mapped, boot_pages);
652 witness_size = round_page(witness_startup_count());
653 new_end -= witness_size;
654 mapped = pmap_map(&vaddr, new_end, new_end + witness_size,
655 VM_PROT_READ | VM_PROT_WRITE);
656 bzero((void *)mapped, witness_size);
657 witness_startup((void *)mapped);
660 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
661 defined(__i386__) || defined(__mips__) || defined(__riscv)
663 * Allocate a bitmap to indicate that a random physical page
664 * needs to be included in a minidump.
666 * The amd64 port needs this to indicate which direct map pages
667 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
669 * However, i386 still needs this workspace internally within the
670 * minidump code. In theory, they are not needed on i386, but are
671 * included should the sf_buf code decide to use them.
674 for (i = 0; dump_avail[i + 1] != 0; i += 2)
675 if (dump_avail[i + 1] > last_pa)
676 last_pa = dump_avail[i + 1];
677 page_range = last_pa / PAGE_SIZE;
678 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
679 new_end -= vm_page_dump_size;
680 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
681 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
682 bzero((void *)vm_page_dump, vm_page_dump_size);
686 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
689 * Include the UMA bootstrap pages, witness pages and vm_page_dump
690 * in a crash dump. When pmap_map() uses the direct map, they are
691 * not automatically included.
693 for (pa = new_end; pa < end; pa += PAGE_SIZE)
696 phys_avail[biggestone + 1] = new_end;
699 * Request that the physical pages underlying the message buffer be
700 * included in a crash dump. Since the message buffer is accessed
701 * through the direct map, they are not automatically included.
703 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
704 last_pa = pa + round_page(msgbufsize);
705 while (pa < last_pa) {
711 * Compute the number of pages of memory that will be available for
712 * use, taking into account the overhead of a page structure per page.
713 * In other words, solve
714 * "available physical memory" - round_page(page_range *
715 * sizeof(struct vm_page)) = page_range * PAGE_SIZE
718 low_avail = phys_avail[0];
719 high_avail = phys_avail[1];
720 for (i = 0; i < vm_phys_nsegs; i++) {
721 if (vm_phys_segs[i].start < low_avail)
722 low_avail = vm_phys_segs[i].start;
723 if (vm_phys_segs[i].end > high_avail)
724 high_avail = vm_phys_segs[i].end;
726 /* Skip the first chunk. It is already accounted for. */
727 for (i = 2; phys_avail[i + 1] != 0; i += 2) {
728 if (phys_avail[i] < low_avail)
729 low_avail = phys_avail[i];
730 if (phys_avail[i + 1] > high_avail)
731 high_avail = phys_avail[i + 1];
733 first_page = low_avail / PAGE_SIZE;
734 #ifdef VM_PHYSSEG_SPARSE
736 for (i = 0; i < vm_phys_nsegs; i++)
737 size += vm_phys_segs[i].end - vm_phys_segs[i].start;
738 for (i = 0; phys_avail[i + 1] != 0; i += 2)
739 size += phys_avail[i + 1] - phys_avail[i];
740 #elif defined(VM_PHYSSEG_DENSE)
741 size = high_avail - low_avail;
743 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
746 #ifdef PMAP_HAS_PAGE_ARRAY
747 pmap_page_array_startup(size / PAGE_SIZE);
748 biggestone = vm_phys_avail_largest();
749 end = new_end = phys_avail[biggestone + 1];
751 #ifdef VM_PHYSSEG_DENSE
753 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
754 * the overhead of a page structure per page only if vm_page_array is
755 * allocated from the last physical memory chunk. Otherwise, we must
756 * allocate page structures representing the physical memory
757 * underlying vm_page_array, even though they will not be used.
759 if (new_end != high_avail)
760 page_range = size / PAGE_SIZE;
764 page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
767 * If the partial bytes remaining are large enough for
768 * a page (PAGE_SIZE) without a corresponding
769 * 'struct vm_page', then new_end will contain an
770 * extra page after subtracting the length of the VM
771 * page array. Compensate by subtracting an extra
774 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
775 if (new_end == high_avail)
776 high_avail -= PAGE_SIZE;
777 new_end -= PAGE_SIZE;
781 new_end = vm_page_array_alloc(&vaddr, end, page_range);
784 #if VM_NRESERVLEVEL > 0
786 * Allocate physical memory for the reservation management system's
787 * data structures, and map it.
789 new_end = vm_reserv_startup(&vaddr, new_end);
791 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
794 * Include vm_page_array and vm_reserv_array in a crash dump.
796 for (pa = new_end; pa < end; pa += PAGE_SIZE)
799 phys_avail[biggestone + 1] = new_end;
802 * Add physical memory segments corresponding to the available
805 for (i = 0; phys_avail[i + 1] != 0; i += 2)
806 if (vm_phys_avail_size(i) != 0)
807 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
810 * Initialize the physical memory allocator.
815 * Initialize the page structures and add every available page to the
816 * physical memory allocator's free lists.
818 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
819 for (ii = 0; ii < vm_page_array_size; ii++) {
820 m = &vm_page_array[ii];
821 vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0);
822 m->flags = PG_FICTITIOUS;
825 vm_cnt.v_page_count = 0;
826 for (segind = 0; segind < vm_phys_nsegs; segind++) {
827 seg = &vm_phys_segs[segind];
828 for (m = seg->first_page, pa = seg->start; pa < seg->end;
829 m++, pa += PAGE_SIZE)
830 vm_page_init_page(m, pa, segind);
833 * Add the segment to the free lists only if it is covered by
834 * one of the ranges in phys_avail. Because we've added the
835 * ranges to the vm_phys_segs array, we can assume that each
836 * segment is either entirely contained in one of the ranges,
837 * or doesn't overlap any of them.
839 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
840 struct vm_domain *vmd;
842 if (seg->start < phys_avail[i] ||
843 seg->end > phys_avail[i + 1])
847 pagecount = (u_long)atop(seg->end - seg->start);
849 vmd = VM_DOMAIN(seg->domain);
850 vm_domain_free_lock(vmd);
851 vm_phys_enqueue_contig(m, pagecount);
852 vm_domain_free_unlock(vmd);
853 vm_domain_freecnt_inc(vmd, pagecount);
854 vm_cnt.v_page_count += (u_int)pagecount;
856 vmd = VM_DOMAIN(seg->domain);
857 vmd->vmd_page_count += (u_int)pagecount;
858 vmd->vmd_segs |= 1UL << m->segind;
864 * Remove blacklisted pages from the physical memory allocator.
866 TAILQ_INIT(&blacklist_head);
867 vm_page_blacklist_load(&list, &listend);
868 vm_page_blacklist_check(list, listend);
870 list = kern_getenv("vm.blacklist");
871 vm_page_blacklist_check(list, NULL);
874 #if VM_NRESERVLEVEL > 0
876 * Initialize the reservation management system.
885 vm_page_reference(vm_page_t m)
888 vm_page_aflag_set(m, PGA_REFERENCED);
892 * vm_page_busy_acquire:
894 * Acquire the busy lock as described by VM_ALLOC_* flags. Will loop
895 * and drop the object lock if necessary.
898 vm_page_busy_acquire(vm_page_t m, int allocflags)
905 * The page-specific object must be cached because page
906 * identity can change during the sleep, causing the
907 * re-lock of a different object.
908 * It is assumed that a reference to the object is already
909 * held by the callers.
913 if ((allocflags & VM_ALLOC_SBUSY) == 0) {
914 if (vm_page_tryxbusy(m))
917 if (vm_page_trysbusy(m))
920 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
923 locked = VM_OBJECT_WOWNED(obj);
925 MPASS(vm_page_wired(m));
930 if (x == VPB_UNBUSIED ||
931 ((allocflags & VM_ALLOC_SBUSY) != 0 &&
932 (x & VPB_BIT_SHARED) != 0) ||
933 ((x & VPB_BIT_WAITERS) == 0 &&
934 !atomic_cmpset_int(&m->busy_lock, x,
935 x | VPB_BIT_WAITERS))) {
940 VM_OBJECT_WUNLOCK(obj);
941 sleepq_add(m, NULL, "vmpba", 0, 0);
944 VM_OBJECT_WLOCK(obj);
945 MPASS(m->object == obj || m->object == NULL);
946 if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
952 * vm_page_busy_downgrade:
954 * Downgrade an exclusive busy page into a single shared busy page.
957 vm_page_busy_downgrade(vm_page_t m)
961 vm_page_assert_xbusied(m);
965 if (atomic_fcmpset_rel_int(&m->busy_lock,
966 &x, VPB_SHARERS_WORD(1)))
969 if ((x & VPB_BIT_WAITERS) != 0)
976 * Return a positive value if the page is shared busied, 0 otherwise.
979 vm_page_sbusied(vm_page_t m)
984 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
990 * Shared unbusy a page.
993 vm_page_sunbusy(vm_page_t m)
997 vm_page_assert_sbusied(m);
1001 if (VPB_SHARERS(x) > 1) {
1002 if (atomic_fcmpset_int(&m->busy_lock, &x,
1003 x - VPB_ONE_SHARER))
1007 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
1008 ("vm_page_sunbusy: invalid lock state"));
1009 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1011 if ((x & VPB_BIT_WAITERS) == 0)
1019 * vm_page_busy_sleep:
1021 * Sleep if the page is busy, using the page pointer as wchan.
1022 * This is used to implement the hard-path of busying mechanism.
1024 * If nonshared is true, sleep only if the page is xbusy.
1026 * The object lock must be held on entry and will be released on exit.
1029 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
1035 vm_page_lock_assert(m, MA_NOTOWNED);
1036 VM_OBJECT_ASSERT_LOCKED(obj);
1040 if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) ||
1041 ((x & VPB_BIT_WAITERS) == 0 &&
1042 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) {
1043 VM_OBJECT_DROP(obj);
1047 VM_OBJECT_DROP(obj);
1048 sleepq_add(m, NULL, wmesg, 0, 0);
1049 sleepq_wait(m, PVM);
1055 * Try to shared busy a page.
1056 * If the operation succeeds 1 is returned otherwise 0.
1057 * The operation never sleeps.
1060 vm_page_trysbusy(vm_page_t m)
1066 if ((x & VPB_BIT_SHARED) == 0)
1068 if (atomic_fcmpset_acq_int(&m->busy_lock, &x,
1069 x + VPB_ONE_SHARER))
1075 * vm_page_xunbusy_hard:
1077 * Called when unbusy has failed because there is a waiter.
1080 vm_page_xunbusy_hard(vm_page_t m)
1083 vm_page_assert_xbusied(m);
1088 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1093 * Avoid releasing and reacquiring the same page lock.
1096 vm_page_change_lock(vm_page_t m, struct mtx **mtx)
1100 mtx1 = vm_page_lockptr(m);
1110 * vm_page_unhold_pages:
1112 * Unhold each of the pages that is referenced by the given array.
1115 vm_page_unhold_pages(vm_page_t *ma, int count)
1118 for (; count != 0; count--) {
1119 vm_page_unwire(*ma, PQ_ACTIVE);
1125 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1129 #ifdef VM_PHYSSEG_SPARSE
1130 m = vm_phys_paddr_to_vm_page(pa);
1132 m = vm_phys_fictitious_to_vm_page(pa);
1134 #elif defined(VM_PHYSSEG_DENSE)
1138 if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1139 m = &vm_page_array[pi - first_page];
1142 return (vm_phys_fictitious_to_vm_page(pa));
1144 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1151 * Create a fictitious page with the specified physical address and
1152 * memory attribute. The memory attribute is the only the machine-
1153 * dependent aspect of a fictitious page that must be initialized.
1156 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1160 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1161 vm_page_initfake(m, paddr, memattr);
1166 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1169 if ((m->flags & PG_FICTITIOUS) != 0) {
1171 * The page's memattr might have changed since the
1172 * previous initialization. Update the pmap to the
1177 m->phys_addr = paddr;
1179 /* Fictitious pages don't use "segind". */
1180 m->flags = PG_FICTITIOUS;
1181 /* Fictitious pages don't use "order" or "pool". */
1182 m->oflags = VPO_UNMANAGED;
1183 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1184 /* Fictitious pages are unevictable. */
1188 pmap_page_set_memattr(m, memattr);
1194 * Release a fictitious page.
1197 vm_page_putfake(vm_page_t m)
1200 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1201 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1202 ("vm_page_putfake: bad page %p", m));
1203 uma_zfree(fakepg_zone, m);
1207 * vm_page_updatefake:
1209 * Update the given fictitious page to the specified physical address and
1213 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1216 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1217 ("vm_page_updatefake: bad page %p", m));
1218 m->phys_addr = paddr;
1219 pmap_page_set_memattr(m, memattr);
1228 vm_page_free(vm_page_t m)
1231 m->flags &= ~PG_ZERO;
1232 vm_page_free_toq(m);
1236 * vm_page_free_zero:
1238 * Free a page to the zerod-pages queue
1241 vm_page_free_zero(vm_page_t m)
1244 m->flags |= PG_ZERO;
1245 vm_page_free_toq(m);
1249 * Unbusy and handle the page queueing for a page from a getpages request that
1250 * was optionally read ahead or behind.
1253 vm_page_readahead_finish(vm_page_t m)
1256 /* We shouldn't put invalid pages on queues. */
1257 KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m));
1260 * Since the page is not the actually needed one, whether it should
1261 * be activated or deactivated is not obvious. Empirical results
1262 * have shown that deactivating the page is usually the best choice,
1263 * unless the page is wanted by another thread.
1266 if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1267 vm_page_activate(m);
1269 vm_page_deactivate(m);
1275 * vm_page_sleep_if_busy:
1277 * Sleep and release the object lock if the page is busied.
1278 * Returns TRUE if the thread slept.
1280 * The given page must be unlocked and object containing it must
1284 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1288 vm_page_lock_assert(m, MA_NOTOWNED);
1289 VM_OBJECT_ASSERT_WLOCKED(m->object);
1291 if (vm_page_busied(m)) {
1293 * The page-specific object must be cached because page
1294 * identity can change during the sleep, causing the
1295 * re-lock of a different object.
1296 * It is assumed that a reference to the object is already
1297 * held by the callers.
1300 vm_page_busy_sleep(m, msg, false);
1301 VM_OBJECT_WLOCK(obj);
1308 * vm_page_sleep_if_xbusy:
1310 * Sleep and release the object lock if the page is xbusied.
1311 * Returns TRUE if the thread slept.
1313 * The given page must be unlocked and object containing it must
1317 vm_page_sleep_if_xbusy(vm_page_t m, const char *msg)
1321 vm_page_lock_assert(m, MA_NOTOWNED);
1322 VM_OBJECT_ASSERT_WLOCKED(m->object);
1324 if (vm_page_xbusied(m)) {
1326 * The page-specific object must be cached because page
1327 * identity can change during the sleep, causing the
1328 * re-lock of a different object.
1329 * It is assumed that a reference to the object is already
1330 * held by the callers.
1333 vm_page_busy_sleep(m, msg, true);
1334 VM_OBJECT_WLOCK(obj);
1341 * vm_page_dirty_KBI: [ internal use only ]
1343 * Set all bits in the page's dirty field.
1345 * The object containing the specified page must be locked if the
1346 * call is made from the machine-independent layer.
1348 * See vm_page_clear_dirty_mask().
1350 * This function should only be called by vm_page_dirty().
1353 vm_page_dirty_KBI(vm_page_t m)
1356 /* Refer to this operation by its public name. */
1357 KASSERT(m->valid == VM_PAGE_BITS_ALL,
1358 ("vm_page_dirty: page is invalid!"));
1359 m->dirty = VM_PAGE_BITS_ALL;
1363 * vm_page_insert: [ internal use only ]
1365 * Inserts the given mem entry into the object and object list.
1367 * The object must be locked.
1370 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1374 VM_OBJECT_ASSERT_WLOCKED(object);
1375 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1376 return (vm_page_insert_after(m, object, pindex, mpred));
1380 * vm_page_insert_after:
1382 * Inserts the page "m" into the specified object at offset "pindex".
1384 * The page "mpred" must immediately precede the offset "pindex" within
1385 * the specified object.
1387 * The object must be locked.
1390 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1395 VM_OBJECT_ASSERT_WLOCKED(object);
1396 KASSERT(m->object == NULL,
1397 ("vm_page_insert_after: page already inserted"));
1398 if (mpred != NULL) {
1399 KASSERT(mpred->object == object,
1400 ("vm_page_insert_after: object doesn't contain mpred"));
1401 KASSERT(mpred->pindex < pindex,
1402 ("vm_page_insert_after: mpred doesn't precede pindex"));
1403 msucc = TAILQ_NEXT(mpred, listq);
1405 msucc = TAILQ_FIRST(&object->memq);
1407 KASSERT(msucc->pindex > pindex,
1408 ("vm_page_insert_after: msucc doesn't succeed pindex"));
1411 * Record the object/offset pair in this page.
1415 m->ref_count |= VPRC_OBJREF;
1418 * Now link into the object's ordered list of backed pages.
1420 if (vm_radix_insert(&object->rtree, m)) {
1423 m->ref_count &= ~VPRC_OBJREF;
1426 vm_page_insert_radixdone(m, object, mpred);
1431 * vm_page_insert_radixdone:
1433 * Complete page "m" insertion into the specified object after the
1434 * radix trie hooking.
1436 * The page "mpred" must precede the offset "m->pindex" within the
1439 * The object must be locked.
1442 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1445 VM_OBJECT_ASSERT_WLOCKED(object);
1446 KASSERT(object != NULL && m->object == object,
1447 ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1448 KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1449 ("vm_page_insert_radixdone: page %p is missing object ref", m));
1450 if (mpred != NULL) {
1451 KASSERT(mpred->object == object,
1452 ("vm_page_insert_radixdone: object doesn't contain mpred"));
1453 KASSERT(mpred->pindex < m->pindex,
1454 ("vm_page_insert_radixdone: mpred doesn't precede pindex"));
1458 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1460 TAILQ_INSERT_HEAD(&object->memq, m, listq);
1463 * Show that the object has one more resident page.
1465 object->resident_page_count++;
1468 * Hold the vnode until the last page is released.
1470 if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1471 vhold(object->handle);
1474 * Since we are inserting a new and possibly dirty page,
1475 * update the object's OBJ_MIGHTBEDIRTY flag.
1477 if (pmap_page_is_write_mapped(m))
1478 vm_object_set_writeable_dirty(object);
1482 * Do the work to remove a page from its object. The caller is responsible for
1483 * updating the page's fields to reflect this removal.
1486 vm_page_object_remove(vm_page_t m)
1492 VM_OBJECT_ASSERT_WLOCKED(object);
1493 KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1494 ("page %p is missing its object ref", m));
1495 if (vm_page_xbusied(m))
1497 mrem = vm_radix_remove(&object->rtree, m->pindex);
1498 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1501 * Now remove from the object's list of backed pages.
1503 TAILQ_REMOVE(&object->memq, m, listq);
1506 * And show that the object has one fewer resident page.
1508 object->resident_page_count--;
1511 * The vnode may now be recycled.
1513 if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1514 vdrop(object->handle);
1520 * Removes the specified page from its containing object, but does not
1521 * invalidate any backing storage. Returns true if the object's reference
1522 * was the last reference to the page, and false otherwise.
1524 * The object must be locked.
1527 vm_page_remove(vm_page_t m)
1530 vm_page_object_remove(m);
1532 return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1538 * Returns the page associated with the object/offset
1539 * pair specified; if none is found, NULL is returned.
1541 * The object must be locked.
1544 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1547 VM_OBJECT_ASSERT_LOCKED(object);
1548 return (vm_radix_lookup(&object->rtree, pindex));
1552 * vm_page_find_least:
1554 * Returns the page associated with the object with least pindex
1555 * greater than or equal to the parameter pindex, or NULL.
1557 * The object must be locked.
1560 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1564 VM_OBJECT_ASSERT_LOCKED(object);
1565 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1566 m = vm_radix_lookup_ge(&object->rtree, pindex);
1571 * Returns the given page's successor (by pindex) within the object if it is
1572 * resident; if none is found, NULL is returned.
1574 * The object must be locked.
1577 vm_page_next(vm_page_t m)
1581 VM_OBJECT_ASSERT_LOCKED(m->object);
1582 if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1583 MPASS(next->object == m->object);
1584 if (next->pindex != m->pindex + 1)
1591 * Returns the given page's predecessor (by pindex) within the object if it is
1592 * resident; if none is found, NULL is returned.
1594 * The object must be locked.
1597 vm_page_prev(vm_page_t m)
1601 VM_OBJECT_ASSERT_LOCKED(m->object);
1602 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1603 MPASS(prev->object == m->object);
1604 if (prev->pindex != m->pindex - 1)
1611 * Uses the page mnew as a replacement for an existing page at index
1612 * pindex which must be already present in the object.
1615 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1619 VM_OBJECT_ASSERT_WLOCKED(object);
1620 KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
1621 ("vm_page_replace: page %p already in object", mnew));
1624 * This function mostly follows vm_page_insert() and
1625 * vm_page_remove() without the radix, object count and vnode
1626 * dance. Double check such functions for more comments.
1629 mnew->object = object;
1630 mnew->pindex = pindex;
1631 atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
1632 mold = vm_radix_replace(&object->rtree, mnew);
1633 KASSERT(mold->queue == PQ_NONE,
1634 ("vm_page_replace: old page %p is on a paging queue", mold));
1636 /* Keep the resident page list in sorted order. */
1637 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1638 TAILQ_REMOVE(&object->memq, mold, listq);
1640 mold->object = NULL;
1641 atomic_clear_int(&mold->ref_count, VPRC_OBJREF);
1642 vm_page_xunbusy(mold);
1645 * The object's resident_page_count does not change because we have
1646 * swapped one page for another, but OBJ_MIGHTBEDIRTY.
1648 if (pmap_page_is_write_mapped(mnew))
1649 vm_object_set_writeable_dirty(object);
1656 * Move the given memory entry from its
1657 * current object to the specified target object/offset.
1659 * Note: swap associated with the page must be invalidated by the move. We
1660 * have to do this for several reasons: (1) we aren't freeing the
1661 * page, (2) we are dirtying the page, (3) the VM system is probably
1662 * moving the page from object A to B, and will then later move
1663 * the backing store from A to B and we can't have a conflict.
1665 * Note: we *always* dirty the page. It is necessary both for the
1666 * fact that we moved it, and because we may be invalidating
1669 * The objects must be locked.
1672 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1677 VM_OBJECT_ASSERT_WLOCKED(new_object);
1679 KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m));
1680 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1681 KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1682 ("vm_page_rename: pindex already renamed"));
1685 * Create a custom version of vm_page_insert() which does not depend
1686 * by m_prev and can cheat on the implementation aspects of the
1690 m->pindex = new_pindex;
1691 if (vm_radix_insert(&new_object->rtree, m)) {
1697 * The operation cannot fail anymore. The removal must happen before
1698 * the listq iterator is tainted.
1701 vm_page_object_remove(m);
1703 /* Return back to the new pindex to complete vm_page_insert(). */
1704 m->pindex = new_pindex;
1705 m->object = new_object;
1707 vm_page_insert_radixdone(m, new_object, mpred);
1715 * Allocate and return a page that is associated with the specified
1716 * object and offset pair. By default, this page is exclusive busied.
1718 * The caller must always specify an allocation class.
1720 * allocation classes:
1721 * VM_ALLOC_NORMAL normal process request
1722 * VM_ALLOC_SYSTEM system *really* needs a page
1723 * VM_ALLOC_INTERRUPT interrupt time request
1725 * optional allocation flags:
1726 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1727 * intends to allocate
1728 * VM_ALLOC_NOBUSY do not exclusive busy the page
1729 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1730 * VM_ALLOC_NOOBJ page is not associated with an object and
1731 * should not be exclusive busy
1732 * VM_ALLOC_SBUSY shared busy the allocated page
1733 * VM_ALLOC_WIRED wire the allocated page
1734 * VM_ALLOC_ZERO prefer a zeroed page
1737 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1740 return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1741 vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1745 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1749 return (vm_page_alloc_domain_after(object, pindex, domain, req,
1750 object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) :
1755 * Allocate a page in the specified object with the given page index. To
1756 * optimize insertion of the page into the object, the caller must also specifiy
1757 * the resident page in the object with largest index smaller than the given
1758 * page index, or NULL if no such page exists.
1761 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1762 int req, vm_page_t mpred)
1764 struct vm_domainset_iter di;
1768 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1770 m = vm_page_alloc_domain_after(object, pindex, domain, req,
1774 } while (vm_domainset_iter_page(&di, object, &domain) == 0);
1780 * Returns true if the number of free pages exceeds the minimum
1781 * for the request class and false otherwise.
1784 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
1786 u_int limit, old, new;
1788 req = req & VM_ALLOC_CLASS_MASK;
1791 * The page daemon is allowed to dig deeper into the free page list.
1793 if (curproc == pageproc && req != VM_ALLOC_INTERRUPT)
1794 req = VM_ALLOC_SYSTEM;
1795 if (req == VM_ALLOC_INTERRUPT)
1797 else if (req == VM_ALLOC_SYSTEM)
1798 limit = vmd->vmd_interrupt_free_min;
1800 limit = vmd->vmd_free_reserved;
1803 * Attempt to reserve the pages. Fail if we're below the limit.
1806 old = vmd->vmd_free_count;
1811 } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
1813 /* Wake the page daemon if we've crossed the threshold. */
1814 if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
1815 pagedaemon_wakeup(vmd->vmd_domain);
1817 /* Only update bitsets on transitions. */
1818 if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
1819 (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
1826 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
1827 int req, vm_page_t mpred)
1829 struct vm_domain *vmd;
1833 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1834 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1835 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1836 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1837 ("inconsistent object(%p)/req(%x)", object, req));
1838 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
1839 ("Can't sleep and retry object insertion."));
1840 KASSERT(mpred == NULL || mpred->pindex < pindex,
1841 ("mpred %p doesn't precede pindex 0x%jx", mpred,
1842 (uintmax_t)pindex));
1844 VM_OBJECT_ASSERT_WLOCKED(object);
1848 pool = object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT;
1850 #if VM_NRESERVLEVEL > 0
1852 * Can we allocate the page from a reservation?
1854 if (vm_object_reserv(object) &&
1855 (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
1857 domain = vm_phys_domain(m);
1858 vmd = VM_DOMAIN(domain);
1862 vmd = VM_DOMAIN(domain);
1863 if (vmd->vmd_pgcache[pool].zone != NULL) {
1864 m = uma_zalloc(vmd->vmd_pgcache[pool].zone, M_NOWAIT);
1866 flags |= PG_PCPU_CACHE;
1870 if (vm_domain_allocate(vmd, req, 1)) {
1872 * If not, allocate it from the free page queues.
1874 vm_domain_free_lock(vmd);
1875 m = vm_phys_alloc_pages(domain, pool, 0);
1876 vm_domain_free_unlock(vmd);
1878 vm_domain_freecnt_inc(vmd, 1);
1879 #if VM_NRESERVLEVEL > 0
1880 if (vm_reserv_reclaim_inactive(domain))
1887 * Not allocatable, give up.
1889 if (vm_domain_alloc_fail(vmd, object, req))
1895 * At this point we had better have found a good page.
1899 vm_page_alloc_check(m);
1902 * Initialize the page. Only the PG_ZERO flag is inherited.
1904 if ((req & VM_ALLOC_ZERO) != 0)
1905 flags |= (m->flags & PG_ZERO);
1906 if ((req & VM_ALLOC_NODUMP) != 0)
1910 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1912 m->busy_lock = VPB_UNBUSIED;
1913 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1914 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1915 if ((req & VM_ALLOC_SBUSY) != 0)
1916 m->busy_lock = VPB_SHARERS_WORD(1);
1917 if (req & VM_ALLOC_WIRED) {
1919 * The page lock is not required for wiring a page until that
1920 * page is inserted into the object.
1927 if (object != NULL) {
1928 if (vm_page_insert_after(m, object, pindex, mpred)) {
1929 if (req & VM_ALLOC_WIRED) {
1933 KASSERT(m->object == NULL, ("page %p has object", m));
1934 m->oflags = VPO_UNMANAGED;
1935 m->busy_lock = VPB_UNBUSIED;
1936 /* Don't change PG_ZERO. */
1937 vm_page_free_toq(m);
1938 if (req & VM_ALLOC_WAITFAIL) {
1939 VM_OBJECT_WUNLOCK(object);
1941 VM_OBJECT_WLOCK(object);
1946 /* Ignore device objects; the pager sets "memattr" for them. */
1947 if (object->memattr != VM_MEMATTR_DEFAULT &&
1948 (object->flags & OBJ_FICTITIOUS) == 0)
1949 pmap_page_set_memattr(m, object->memattr);
1957 * vm_page_alloc_contig:
1959 * Allocate a contiguous set of physical pages of the given size "npages"
1960 * from the free lists. All of the physical pages must be at or above
1961 * the given physical address "low" and below the given physical address
1962 * "high". The given value "alignment" determines the alignment of the
1963 * first physical page in the set. If the given value "boundary" is
1964 * non-zero, then the set of physical pages cannot cross any physical
1965 * address boundary that is a multiple of that value. Both "alignment"
1966 * and "boundary" must be a power of two.
1968 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1969 * then the memory attribute setting for the physical pages is configured
1970 * to the object's memory attribute setting. Otherwise, the memory
1971 * attribute setting for the physical pages is configured to "memattr",
1972 * overriding the object's memory attribute setting. However, if the
1973 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1974 * memory attribute setting for the physical pages cannot be configured
1975 * to VM_MEMATTR_DEFAULT.
1977 * The specified object may not contain fictitious pages.
1979 * The caller must always specify an allocation class.
1981 * allocation classes:
1982 * VM_ALLOC_NORMAL normal process request
1983 * VM_ALLOC_SYSTEM system *really* needs a page
1984 * VM_ALLOC_INTERRUPT interrupt time request
1986 * optional allocation flags:
1987 * VM_ALLOC_NOBUSY do not exclusive busy the page
1988 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1989 * VM_ALLOC_NOOBJ page is not associated with an object and
1990 * should not be exclusive busy
1991 * VM_ALLOC_SBUSY shared busy the allocated page
1992 * VM_ALLOC_WIRED wire the allocated page
1993 * VM_ALLOC_ZERO prefer a zeroed page
1996 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1997 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1998 vm_paddr_t boundary, vm_memattr_t memattr)
2000 struct vm_domainset_iter di;
2004 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2006 m = vm_page_alloc_contig_domain(object, pindex, domain, req,
2007 npages, low, high, alignment, boundary, memattr);
2010 } while (vm_domainset_iter_page(&di, object, &domain) == 0);
2016 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2017 int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2018 vm_paddr_t boundary, vm_memattr_t memattr)
2020 struct vm_domain *vmd;
2021 vm_page_t m, m_ret, mpred;
2022 u_int busy_lock, flags, oflags;
2024 mpred = NULL; /* XXX: pacify gcc */
2025 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
2026 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
2027 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2028 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2029 ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
2031 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2032 ("Can't sleep and retry object insertion."));
2033 if (object != NULL) {
2034 VM_OBJECT_ASSERT_WLOCKED(object);
2035 KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2036 ("vm_page_alloc_contig: object %p has fictitious pages",
2039 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2041 if (object != NULL) {
2042 mpred = vm_radix_lookup_le(&object->rtree, pindex);
2043 KASSERT(mpred == NULL || mpred->pindex != pindex,
2044 ("vm_page_alloc_contig: pindex already allocated"));
2048 * Can we allocate the pages without the number of free pages falling
2049 * below the lower bound for the allocation class?
2053 #if VM_NRESERVLEVEL > 0
2055 * Can we allocate the pages from a reservation?
2057 if (vm_object_reserv(object) &&
2058 (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
2059 mpred, npages, low, high, alignment, boundary)) != NULL) {
2060 domain = vm_phys_domain(m_ret);
2061 vmd = VM_DOMAIN(domain);
2065 vmd = VM_DOMAIN(domain);
2066 if (vm_domain_allocate(vmd, req, npages)) {
2068 * allocate them from the free page queues.
2070 vm_domain_free_lock(vmd);
2071 m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2072 alignment, boundary);
2073 vm_domain_free_unlock(vmd);
2074 if (m_ret == NULL) {
2075 vm_domain_freecnt_inc(vmd, npages);
2076 #if VM_NRESERVLEVEL > 0
2077 if (vm_reserv_reclaim_contig(domain, npages, low,
2078 high, alignment, boundary))
2083 if (m_ret == NULL) {
2084 if (vm_domain_alloc_fail(vmd, object, req))
2088 #if VM_NRESERVLEVEL > 0
2091 for (m = m_ret; m < &m_ret[npages]; m++) {
2093 vm_page_alloc_check(m);
2097 * Initialize the pages. Only the PG_ZERO flag is inherited.
2100 if ((req & VM_ALLOC_ZERO) != 0)
2102 if ((req & VM_ALLOC_NODUMP) != 0)
2104 oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2106 busy_lock = VPB_UNBUSIED;
2107 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2108 busy_lock = VPB_SINGLE_EXCLUSIVER;
2109 if ((req & VM_ALLOC_SBUSY) != 0)
2110 busy_lock = VPB_SHARERS_WORD(1);
2111 if ((req & VM_ALLOC_WIRED) != 0)
2112 vm_wire_add(npages);
2113 if (object != NULL) {
2114 if (object->memattr != VM_MEMATTR_DEFAULT &&
2115 memattr == VM_MEMATTR_DEFAULT)
2116 memattr = object->memattr;
2118 for (m = m_ret; m < &m_ret[npages]; m++) {
2120 m->flags = (m->flags | PG_NODUMP) & flags;
2121 m->busy_lock = busy_lock;
2122 if ((req & VM_ALLOC_WIRED) != 0)
2126 if (object != NULL) {
2127 if (vm_page_insert_after(m, object, pindex, mpred)) {
2128 if ((req & VM_ALLOC_WIRED) != 0)
2129 vm_wire_sub(npages);
2130 KASSERT(m->object == NULL,
2131 ("page %p has object", m));
2133 for (m = m_ret; m < &m_ret[npages]; m++) {
2135 (req & VM_ALLOC_WIRED) != 0)
2137 m->oflags = VPO_UNMANAGED;
2138 m->busy_lock = VPB_UNBUSIED;
2139 /* Don't change PG_ZERO. */
2140 vm_page_free_toq(m);
2142 if (req & VM_ALLOC_WAITFAIL) {
2143 VM_OBJECT_WUNLOCK(object);
2145 VM_OBJECT_WLOCK(object);
2152 if (memattr != VM_MEMATTR_DEFAULT)
2153 pmap_page_set_memattr(m, memattr);
2160 * Check a page that has been freshly dequeued from a freelist.
2163 vm_page_alloc_check(vm_page_t m)
2166 KASSERT(m->object == NULL, ("page %p has object", m));
2167 KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0,
2168 ("page %p has unexpected queue %d, flags %#x",
2169 m, m->queue, (m->aflags & PGA_QUEUE_STATE_MASK)));
2170 KASSERT(m->ref_count == 0, ("page %p has references", m));
2171 KASSERT(!vm_page_busied(m), ("page %p is busy", m));
2172 KASSERT(m->dirty == 0, ("page %p is dirty", m));
2173 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2174 ("page %p has unexpected memattr %d",
2175 m, pmap_page_get_memattr(m)));
2176 KASSERT(m->valid == 0, ("free page %p is valid", m));
2180 * vm_page_alloc_freelist:
2182 * Allocate a physical page from the specified free page list.
2184 * The caller must always specify an allocation class.
2186 * allocation classes:
2187 * VM_ALLOC_NORMAL normal process request
2188 * VM_ALLOC_SYSTEM system *really* needs a page
2189 * VM_ALLOC_INTERRUPT interrupt time request
2191 * optional allocation flags:
2192 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
2193 * intends to allocate
2194 * VM_ALLOC_WIRED wire the allocated page
2195 * VM_ALLOC_ZERO prefer a zeroed page
2198 vm_page_alloc_freelist(int freelist, int req)
2200 struct vm_domainset_iter di;
2204 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2206 m = vm_page_alloc_freelist_domain(domain, freelist, req);
2209 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2215 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2217 struct vm_domain *vmd;
2222 vmd = VM_DOMAIN(domain);
2224 if (vm_domain_allocate(vmd, req, 1)) {
2225 vm_domain_free_lock(vmd);
2226 m = vm_phys_alloc_freelist_pages(domain, freelist,
2227 VM_FREEPOOL_DIRECT, 0);
2228 vm_domain_free_unlock(vmd);
2230 vm_domain_freecnt_inc(vmd, 1);
2233 if (vm_domain_alloc_fail(vmd, NULL, req))
2238 vm_page_alloc_check(m);
2241 * Initialize the page. Only the PG_ZERO flag is inherited.
2245 if ((req & VM_ALLOC_ZERO) != 0)
2248 if ((req & VM_ALLOC_WIRED) != 0) {
2250 * The page lock is not required for wiring a page that does
2251 * not belong to an object.
2256 /* Unmanaged pages don't use "act_count". */
2257 m->oflags = VPO_UNMANAGED;
2262 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2264 struct vm_domain *vmd;
2265 struct vm_pgcache *pgcache;
2269 vmd = VM_DOMAIN(pgcache->domain);
2270 /* Only import if we can bring in a full bucket. */
2271 if (cnt == 1 || !vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2273 domain = vmd->vmd_domain;
2274 vm_domain_free_lock(vmd);
2275 i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2276 (vm_page_t *)store);
2277 vm_domain_free_unlock(vmd);
2279 vm_domain_freecnt_inc(vmd, cnt - i);
2285 vm_page_zone_release(void *arg, void **store, int cnt)
2287 struct vm_domain *vmd;
2288 struct vm_pgcache *pgcache;
2293 vmd = VM_DOMAIN(pgcache->domain);
2294 vm_domain_free_lock(vmd);
2295 for (i = 0; i < cnt; i++) {
2296 m = (vm_page_t)store[i];
2297 vm_phys_free_pages(m, 0);
2299 vm_domain_free_unlock(vmd);
2300 vm_domain_freecnt_inc(vmd, cnt);
2303 #define VPSC_ANY 0 /* No restrictions. */
2304 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */
2305 #define VPSC_NOSUPER 2 /* Skip superpages. */
2308 * vm_page_scan_contig:
2310 * Scan vm_page_array[] between the specified entries "m_start" and
2311 * "m_end" for a run of contiguous physical pages that satisfy the
2312 * specified conditions, and return the lowest page in the run. The
2313 * specified "alignment" determines the alignment of the lowest physical
2314 * page in the run. If the specified "boundary" is non-zero, then the
2315 * run of physical pages cannot span a physical address that is a
2316 * multiple of "boundary".
2318 * "m_end" is never dereferenced, so it need not point to a vm_page
2319 * structure within vm_page_array[].
2321 * "npages" must be greater than zero. "m_start" and "m_end" must not
2322 * span a hole (or discontiguity) in the physical address space. Both
2323 * "alignment" and "boundary" must be a power of two.
2326 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2327 u_long alignment, vm_paddr_t boundary, int options)
2333 #if VM_NRESERVLEVEL > 0
2336 int m_inc, order, run_ext, run_len;
2338 KASSERT(npages > 0, ("npages is 0"));
2339 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2340 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2344 for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2345 KASSERT((m->flags & PG_MARKER) == 0,
2346 ("page %p is PG_MARKER", m));
2347 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1,
2348 ("fictitious page %p has invalid ref count", m));
2351 * If the current page would be the start of a run, check its
2352 * physical address against the end, alignment, and boundary
2353 * conditions. If it doesn't satisfy these conditions, either
2354 * terminate the scan or advance to the next page that
2355 * satisfies the failed condition.
2358 KASSERT(m_run == NULL, ("m_run != NULL"));
2359 if (m + npages > m_end)
2361 pa = VM_PAGE_TO_PHYS(m);
2362 if ((pa & (alignment - 1)) != 0) {
2363 m_inc = atop(roundup2(pa, alignment) - pa);
2366 if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2368 m_inc = atop(roundup2(pa, boundary) - pa);
2372 KASSERT(m_run != NULL, ("m_run == NULL"));
2374 vm_page_change_lock(m, &m_mtx);
2377 if (vm_page_wired(m))
2379 #if VM_NRESERVLEVEL > 0
2380 else if ((level = vm_reserv_level(m)) >= 0 &&
2381 (options & VPSC_NORESERV) != 0) {
2383 /* Advance to the end of the reservation. */
2384 pa = VM_PAGE_TO_PHYS(m);
2385 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2389 else if ((object = m->object) != NULL) {
2391 * The page is considered eligible for relocation if
2392 * and only if it could be laundered or reclaimed by
2395 if (!VM_OBJECT_TRYRLOCK(object)) {
2397 VM_OBJECT_RLOCK(object);
2399 if (m->object != object) {
2401 * The page may have been freed.
2403 VM_OBJECT_RUNLOCK(object);
2407 /* Don't care: PG_NODUMP, PG_ZERO. */
2408 if (object->type != OBJT_DEFAULT &&
2409 object->type != OBJT_SWAP &&
2410 object->type != OBJT_VNODE) {
2412 #if VM_NRESERVLEVEL > 0
2413 } else if ((options & VPSC_NOSUPER) != 0 &&
2414 (level = vm_reserv_level_iffullpop(m)) >= 0) {
2416 /* Advance to the end of the superpage. */
2417 pa = VM_PAGE_TO_PHYS(m);
2418 m_inc = atop(roundup2(pa + 1,
2419 vm_reserv_size(level)) - pa);
2421 } else if (object->memattr == VM_MEMATTR_DEFAULT &&
2422 vm_page_queue(m) != PQ_NONE && !vm_page_busied(m) &&
2423 !vm_page_wired(m)) {
2425 * The page is allocated but eligible for
2426 * relocation. Extend the current run by one
2429 KASSERT(pmap_page_get_memattr(m) ==
2431 ("page %p has an unexpected memattr", m));
2432 KASSERT((m->oflags & (VPO_SWAPINPROG |
2433 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2434 ("page %p has unexpected oflags", m));
2435 /* Don't care: VPO_NOSYNC. */
2439 VM_OBJECT_RUNLOCK(object);
2440 #if VM_NRESERVLEVEL > 0
2441 } else if (level >= 0) {
2443 * The page is reserved but not yet allocated. In
2444 * other words, it is still free. Extend the current
2449 } else if ((order = m->order) < VM_NFREEORDER) {
2451 * The page is enqueued in the physical memory
2452 * allocator's free page queues. Moreover, it is the
2453 * first page in a power-of-two-sized run of
2454 * contiguous free pages. Add these pages to the end
2455 * of the current run, and jump ahead.
2457 run_ext = 1 << order;
2461 * Skip the page for one of the following reasons: (1)
2462 * It is enqueued in the physical memory allocator's
2463 * free page queues. However, it is not the first
2464 * page in a run of contiguous free pages. (This case
2465 * rarely occurs because the scan is performed in
2466 * ascending order.) (2) It is not reserved, and it is
2467 * transitioning from free to allocated. (Conversely,
2468 * the transition from allocated to free for managed
2469 * pages is blocked by the page lock.) (3) It is
2470 * allocated but not contained by an object and not
2471 * wired, e.g., allocated by Xen's balloon driver.
2477 * Extend or reset the current run of pages.
2492 if (run_len >= npages)
2498 * vm_page_reclaim_run:
2500 * Try to relocate each of the allocated virtual pages within the
2501 * specified run of physical pages to a new physical address. Free the
2502 * physical pages underlying the relocated virtual pages. A virtual page
2503 * is relocatable if and only if it could be laundered or reclaimed by
2504 * the page daemon. Whenever possible, a virtual page is relocated to a
2505 * physical address above "high".
2507 * Returns 0 if every physical page within the run was already free or
2508 * just freed by a successful relocation. Otherwise, returns a non-zero
2509 * value indicating why the last attempt to relocate a virtual page was
2512 * "req_class" must be an allocation class.
2515 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2518 struct vm_domain *vmd;
2520 struct spglist free;
2523 vm_page_t m, m_end, m_new;
2524 int error, order, req;
2526 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2527 ("req_class is not an allocation class"));
2531 m_end = m_run + npages;
2533 for (; error == 0 && m < m_end; m++) {
2534 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2535 ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2538 * Avoid releasing and reacquiring the same page lock.
2540 vm_page_change_lock(m, &m_mtx);
2543 * Racily check for wirings. Races are handled below.
2545 if (vm_page_wired(m))
2547 else if ((object = m->object) != NULL) {
2549 * The page is relocated if and only if it could be
2550 * laundered or reclaimed by the page daemon.
2552 if (!VM_OBJECT_TRYWLOCK(object)) {
2554 VM_OBJECT_WLOCK(object);
2556 if (m->object != object) {
2558 * The page may have been freed.
2560 VM_OBJECT_WUNLOCK(object);
2564 /* Don't care: PG_NODUMP, PG_ZERO. */
2565 if (object->type != OBJT_DEFAULT &&
2566 object->type != OBJT_SWAP &&
2567 object->type != OBJT_VNODE)
2569 else if (object->memattr != VM_MEMATTR_DEFAULT)
2571 else if (vm_page_queue(m) != PQ_NONE &&
2572 !vm_page_busied(m) && !vm_page_wired(m)) {
2573 KASSERT(pmap_page_get_memattr(m) ==
2575 ("page %p has an unexpected memattr", m));
2576 KASSERT((m->oflags & (VPO_SWAPINPROG |
2577 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2578 ("page %p has unexpected oflags", m));
2579 /* Don't care: VPO_NOSYNC. */
2580 if (m->valid != 0) {
2582 * First, try to allocate a new page
2583 * that is above "high". Failing
2584 * that, try to allocate a new page
2585 * that is below "m_run". Allocate
2586 * the new page between the end of
2587 * "m_run" and "high" only as a last
2590 req = req_class | VM_ALLOC_NOOBJ;
2591 if ((m->flags & PG_NODUMP) != 0)
2592 req |= VM_ALLOC_NODUMP;
2593 if (trunc_page(high) !=
2594 ~(vm_paddr_t)PAGE_MASK) {
2595 m_new = vm_page_alloc_contig(
2600 VM_MEMATTR_DEFAULT);
2603 if (m_new == NULL) {
2604 pa = VM_PAGE_TO_PHYS(m_run);
2605 m_new = vm_page_alloc_contig(
2607 0, pa - 1, PAGE_SIZE, 0,
2608 VM_MEMATTR_DEFAULT);
2610 if (m_new == NULL) {
2612 m_new = vm_page_alloc_contig(
2614 pa, high, PAGE_SIZE, 0,
2615 VM_MEMATTR_DEFAULT);
2617 if (m_new == NULL) {
2623 * Unmap the page and check for new
2624 * wirings that may have been acquired
2625 * through a pmap lookup.
2627 if (object->ref_count != 0 &&
2628 !vm_page_try_remove_all(m)) {
2629 vm_page_free(m_new);
2635 * Replace "m" with the new page. For
2636 * vm_page_replace(), "m" must be busy
2637 * and dequeued. Finally, change "m"
2638 * as if vm_page_free() was called.
2640 m_new->aflags = m->aflags &
2641 ~PGA_QUEUE_STATE_MASK;
2642 KASSERT(m_new->oflags == VPO_UNMANAGED,
2643 ("page %p is managed", m_new));
2644 m_new->oflags = m->oflags & VPO_NOSYNC;
2645 pmap_copy_page(m, m_new);
2646 m_new->valid = m->valid;
2647 m_new->dirty = m->dirty;
2648 m->flags &= ~PG_ZERO;
2651 vm_page_replace_checked(m_new, object,
2653 if (vm_page_free_prep(m))
2654 SLIST_INSERT_HEAD(&free, m,
2658 * The new page must be deactivated
2659 * before the object is unlocked.
2661 vm_page_change_lock(m_new, &m_mtx);
2662 vm_page_deactivate(m_new);
2664 m->flags &= ~PG_ZERO;
2666 if (vm_page_free_prep(m))
2667 SLIST_INSERT_HEAD(&free, m,
2669 KASSERT(m->dirty == 0,
2670 ("page %p is dirty", m));
2675 VM_OBJECT_WUNLOCK(object);
2677 MPASS(vm_phys_domain(m) == domain);
2678 vmd = VM_DOMAIN(domain);
2679 vm_domain_free_lock(vmd);
2681 if (order < VM_NFREEORDER) {
2683 * The page is enqueued in the physical memory
2684 * allocator's free page queues. Moreover, it
2685 * is the first page in a power-of-two-sized
2686 * run of contiguous free pages. Jump ahead
2687 * to the last page within that run, and
2688 * continue from there.
2690 m += (1 << order) - 1;
2692 #if VM_NRESERVLEVEL > 0
2693 else if (vm_reserv_is_page_free(m))
2696 vm_domain_free_unlock(vmd);
2697 if (order == VM_NFREEORDER)
2703 if ((m = SLIST_FIRST(&free)) != NULL) {
2706 vmd = VM_DOMAIN(domain);
2708 vm_domain_free_lock(vmd);
2710 MPASS(vm_phys_domain(m) == domain);
2711 SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2712 vm_phys_free_pages(m, 0);
2714 } while ((m = SLIST_FIRST(&free)) != NULL);
2715 vm_domain_free_unlock(vmd);
2716 vm_domain_freecnt_inc(vmd, cnt);
2723 CTASSERT(powerof2(NRUNS));
2725 #define RUN_INDEX(count) ((count) & (NRUNS - 1))
2727 #define MIN_RECLAIM 8
2730 * vm_page_reclaim_contig:
2732 * Reclaim allocated, contiguous physical memory satisfying the specified
2733 * conditions by relocating the virtual pages using that physical memory.
2734 * Returns true if reclamation is successful and false otherwise. Since
2735 * relocation requires the allocation of physical pages, reclamation may
2736 * fail due to a shortage of free pages. When reclamation fails, callers
2737 * are expected to perform vm_wait() before retrying a failed allocation
2738 * operation, e.g., vm_page_alloc_contig().
2740 * The caller must always specify an allocation class through "req".
2742 * allocation classes:
2743 * VM_ALLOC_NORMAL normal process request
2744 * VM_ALLOC_SYSTEM system *really* needs a page
2745 * VM_ALLOC_INTERRUPT interrupt time request
2747 * The optional allocation flags are ignored.
2749 * "npages" must be greater than zero. Both "alignment" and "boundary"
2750 * must be a power of two.
2753 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
2754 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2756 struct vm_domain *vmd;
2757 vm_paddr_t curr_low;
2758 vm_page_t m_run, m_runs[NRUNS];
2759 u_long count, reclaimed;
2760 int error, i, options, req_class;
2762 KASSERT(npages > 0, ("npages is 0"));
2763 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2764 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2765 req_class = req & VM_ALLOC_CLASS_MASK;
2768 * The page daemon is allowed to dig deeper into the free page list.
2770 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2771 req_class = VM_ALLOC_SYSTEM;
2774 * Return if the number of free pages cannot satisfy the requested
2777 vmd = VM_DOMAIN(domain);
2778 count = vmd->vmd_free_count;
2779 if (count < npages + vmd->vmd_free_reserved || (count < npages +
2780 vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2781 (count < npages && req_class == VM_ALLOC_INTERRUPT))
2785 * Scan up to three times, relaxing the restrictions ("options") on
2786 * the reclamation of reservations and superpages each time.
2788 for (options = VPSC_NORESERV;;) {
2790 * Find the highest runs that satisfy the given constraints
2791 * and restrictions, and record them in "m_runs".
2796 m_run = vm_phys_scan_contig(domain, npages, curr_low,
2797 high, alignment, boundary, options);
2800 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2801 m_runs[RUN_INDEX(count)] = m_run;
2806 * Reclaim the highest runs in LIFO (descending) order until
2807 * the number of reclaimed pages, "reclaimed", is at least
2808 * MIN_RECLAIM. Reset "reclaimed" each time because each
2809 * reclamation is idempotent, and runs will (likely) recur
2810 * from one scan to the next as restrictions are relaxed.
2813 for (i = 0; count > 0 && i < NRUNS; i++) {
2815 m_run = m_runs[RUN_INDEX(count)];
2816 error = vm_page_reclaim_run(req_class, domain, npages,
2819 reclaimed += npages;
2820 if (reclaimed >= MIN_RECLAIM)
2826 * Either relax the restrictions on the next scan or return if
2827 * the last scan had no restrictions.
2829 if (options == VPSC_NORESERV)
2830 options = VPSC_NOSUPER;
2831 else if (options == VPSC_NOSUPER)
2833 else if (options == VPSC_ANY)
2834 return (reclaimed != 0);
2839 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
2840 u_long alignment, vm_paddr_t boundary)
2842 struct vm_domainset_iter di;
2846 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2848 ret = vm_page_reclaim_contig_domain(domain, req, npages, low,
2849 high, alignment, boundary);
2852 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2858 * Set the domain in the appropriate page level domainset.
2861 vm_domain_set(struct vm_domain *vmd)
2864 mtx_lock(&vm_domainset_lock);
2865 if (!vmd->vmd_minset && vm_paging_min(vmd)) {
2866 vmd->vmd_minset = 1;
2867 DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
2869 if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
2870 vmd->vmd_severeset = 1;
2871 DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
2873 mtx_unlock(&vm_domainset_lock);
2877 * Clear the domain from the appropriate page level domainset.
2880 vm_domain_clear(struct vm_domain *vmd)
2883 mtx_lock(&vm_domainset_lock);
2884 if (vmd->vmd_minset && !vm_paging_min(vmd)) {
2885 vmd->vmd_minset = 0;
2886 DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
2887 if (vm_min_waiters != 0) {
2889 wakeup(&vm_min_domains);
2892 if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
2893 vmd->vmd_severeset = 0;
2894 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
2895 if (vm_severe_waiters != 0) {
2896 vm_severe_waiters = 0;
2897 wakeup(&vm_severe_domains);
2902 * If pageout daemon needs pages, then tell it that there are
2905 if (vmd->vmd_pageout_pages_needed &&
2906 vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
2907 wakeup(&vmd->vmd_pageout_pages_needed);
2908 vmd->vmd_pageout_pages_needed = 0;
2911 /* See comments in vm_wait_doms(). */
2912 if (vm_pageproc_waiters) {
2913 vm_pageproc_waiters = 0;
2914 wakeup(&vm_pageproc_waiters);
2916 mtx_unlock(&vm_domainset_lock);
2920 * Wait for free pages to exceed the min threshold globally.
2926 mtx_lock(&vm_domainset_lock);
2927 while (vm_page_count_min()) {
2929 msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
2931 mtx_unlock(&vm_domainset_lock);
2935 * Wait for free pages to exceed the severe threshold globally.
2938 vm_wait_severe(void)
2941 mtx_lock(&vm_domainset_lock);
2942 while (vm_page_count_severe()) {
2943 vm_severe_waiters++;
2944 msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
2947 mtx_unlock(&vm_domainset_lock);
2954 return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
2958 vm_wait_doms(const domainset_t *wdoms)
2962 * We use racey wakeup synchronization to avoid expensive global
2963 * locking for the pageproc when sleeping with a non-specific vm_wait.
2964 * To handle this, we only sleep for one tick in this instance. It
2965 * is expected that most allocations for the pageproc will come from
2966 * kmem or vm_page_grab* which will use the more specific and
2967 * race-free vm_wait_domain().
2969 if (curproc == pageproc) {
2970 mtx_lock(&vm_domainset_lock);
2971 vm_pageproc_waiters++;
2972 msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP,
2976 * XXX Ideally we would wait only until the allocation could
2977 * be satisfied. This condition can cause new allocators to
2978 * consume all freed pages while old allocators wait.
2980 mtx_lock(&vm_domainset_lock);
2981 if (vm_page_count_min_set(wdoms)) {
2983 msleep(&vm_min_domains, &vm_domainset_lock,
2984 PVM | PDROP, "vmwait", 0);
2986 mtx_unlock(&vm_domainset_lock);
2993 * Sleep until free pages are available for allocation.
2994 * - Called in various places after failed memory allocations.
2997 vm_wait_domain(int domain)
2999 struct vm_domain *vmd;
3002 vmd = VM_DOMAIN(domain);
3003 vm_domain_free_assert_unlocked(vmd);
3005 if (curproc == pageproc) {
3006 mtx_lock(&vm_domainset_lock);
3007 if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3008 vmd->vmd_pageout_pages_needed = 1;
3009 msleep(&vmd->vmd_pageout_pages_needed,
3010 &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3012 mtx_unlock(&vm_domainset_lock);
3014 if (pageproc == NULL)
3015 panic("vm_wait in early boot");
3016 DOMAINSET_ZERO(&wdom);
3017 DOMAINSET_SET(vmd->vmd_domain, &wdom);
3018 vm_wait_doms(&wdom);
3025 * Sleep until free pages are available for allocation in the
3026 * affinity domains of the obj. If obj is NULL, the domain set
3027 * for the calling thread is used.
3028 * Called in various places after failed memory allocations.
3031 vm_wait(vm_object_t obj)
3033 struct domainset *d;
3038 * Carefully fetch pointers only once: the struct domainset
3039 * itself is ummutable but the pointer might change.
3042 d = obj->domain.dr_policy;
3044 d = curthread->td_domain.dr_policy;
3046 vm_wait_doms(&d->ds_mask);
3050 * vm_domain_alloc_fail:
3052 * Called when a page allocation function fails. Informs the
3053 * pagedaemon and performs the requested wait. Requires the
3054 * domain_free and object lock on entry. Returns with the
3055 * object lock held and free lock released. Returns an error when
3056 * retry is necessary.
3060 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3063 vm_domain_free_assert_unlocked(vmd);
3065 atomic_add_int(&vmd->vmd_pageout_deficit,
3066 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3067 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3069 VM_OBJECT_WUNLOCK(object);
3070 vm_wait_domain(vmd->vmd_domain);
3072 VM_OBJECT_WLOCK(object);
3073 if (req & VM_ALLOC_WAITOK)
3083 * Sleep until free pages are available for allocation.
3084 * - Called only in vm_fault so that processes page faulting
3085 * can be easily tracked.
3086 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3087 * processes will be able to grab memory first. Do not change
3088 * this balance without careful testing first.
3091 vm_waitpfault(struct domainset *dset, int timo)
3095 * XXX Ideally we would wait only until the allocation could
3096 * be satisfied. This condition can cause new allocators to
3097 * consume all freed pages while old allocators wait.
3099 mtx_lock(&vm_domainset_lock);
3100 if (vm_page_count_min_set(&dset->ds_mask)) {
3102 msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3105 mtx_unlock(&vm_domainset_lock);
3108 static struct vm_pagequeue *
3109 vm_page_pagequeue(vm_page_t m)
3114 if ((queue = atomic_load_8(&m->queue)) == PQ_NONE)
3116 return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3120 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m)
3122 struct vm_domain *vmd;
3125 CRITICAL_ASSERT(curthread);
3126 vm_pagequeue_assert_locked(pq);
3129 * The page daemon is allowed to set m->queue = PQ_NONE without
3130 * the page queue lock held. In this case it is about to free the page,
3131 * which must not have any queue state.
3133 qflags = atomic_load_8(&m->aflags);
3134 KASSERT(pq == vm_page_pagequeue(m) ||
3135 (qflags & PGA_QUEUE_STATE_MASK) == 0,
3136 ("page %p doesn't belong to queue %p but has aflags %#x",
3139 if ((qflags & PGA_DEQUEUE) != 0) {
3140 if (__predict_true((qflags & PGA_ENQUEUED) != 0))
3141 vm_pagequeue_remove(pq, m);
3142 vm_page_dequeue_complete(m);
3143 counter_u64_add(queue_ops, 1);
3144 } else if ((qflags & (PGA_REQUEUE | PGA_REQUEUE_HEAD)) != 0) {
3145 if ((qflags & PGA_ENQUEUED) != 0)
3146 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3148 vm_pagequeue_cnt_inc(pq);
3149 vm_page_aflag_set(m, PGA_ENQUEUED);
3153 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE.
3154 * In particular, if both flags are set in close succession,
3155 * only PGA_REQUEUE_HEAD will be applied, even if it was set
3158 if ((qflags & PGA_REQUEUE_HEAD) != 0) {
3159 KASSERT(m->queue == PQ_INACTIVE,
3160 ("head enqueue not supported for page %p", m));
3161 vmd = vm_pagequeue_domain(m);
3162 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3164 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3166 vm_page_aflag_clear(m, qflags & (PGA_REQUEUE |
3168 counter_u64_add(queue_ops, 1);
3170 counter_u64_add(queue_nops, 1);
3175 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3181 for (i = 0; i < bq->bq_cnt; i++) {
3183 if (__predict_false(m->queue != queue))
3185 vm_pqbatch_process_page(pq, m);
3187 vm_batchqueue_init(bq);
3191 * vm_page_pqbatch_submit: [ internal use only ]
3193 * Enqueue a page in the specified page queue's batched work queue.
3194 * The caller must have encoded the requested operation in the page
3195 * structure's aflags field.
3198 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
3200 struct vm_batchqueue *bq;
3201 struct vm_pagequeue *pq;
3204 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3205 ("page %p is unmanaged", m));
3206 KASSERT(mtx_owned(vm_page_lockptr(m)) || m->object == NULL,
3207 ("missing synchronization for page %p", m));
3208 KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3210 domain = vm_phys_domain(m);
3211 pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue];
3214 bq = DPCPU_PTR(pqbatch[domain][queue]);
3215 if (vm_batchqueue_insert(bq, m)) {
3219 if (!vm_pagequeue_trylock(pq)) {
3221 vm_pagequeue_lock(pq);
3223 bq = DPCPU_PTR(pqbatch[domain][queue]);
3225 vm_pqbatch_process(pq, bq, queue);
3228 * The page may have been logically dequeued before we acquired the
3229 * page queue lock. In this case, since we either hold the page lock
3230 * or the page is being freed, a different thread cannot be concurrently
3231 * enqueuing the page.
3233 if (__predict_true(m->queue == queue))
3234 vm_pqbatch_process_page(pq, m);
3236 KASSERT(m->queue == PQ_NONE,
3237 ("invalid queue transition for page %p", m));
3238 KASSERT((m->aflags & PGA_ENQUEUED) == 0,
3239 ("page %p is enqueued with invalid queue index", m));
3241 vm_pagequeue_unlock(pq);
3246 * vm_page_pqbatch_drain: [ internal use only ]
3248 * Force all per-CPU page queue batch queues to be drained. This is
3249 * intended for use in severe memory shortages, to ensure that pages
3250 * do not remain stuck in the batch queues.
3253 vm_page_pqbatch_drain(void)
3256 struct vm_domain *vmd;
3257 struct vm_pagequeue *pq;
3258 int cpu, domain, queue;
3263 sched_bind(td, cpu);
3266 for (domain = 0; domain < vm_ndomains; domain++) {
3267 vmd = VM_DOMAIN(domain);
3268 for (queue = 0; queue < PQ_COUNT; queue++) {
3269 pq = &vmd->vmd_pagequeues[queue];
3270 vm_pagequeue_lock(pq);
3272 vm_pqbatch_process(pq,
3273 DPCPU_PTR(pqbatch[domain][queue]), queue);
3275 vm_pagequeue_unlock(pq);
3285 * Complete the logical removal of a page from a page queue. We must be
3286 * careful to synchronize with the page daemon, which may be concurrently
3287 * examining the page with only the page lock held. The page must not be
3288 * in a state where it appears to be logically enqueued.
3291 vm_page_dequeue_complete(vm_page_t m)
3295 atomic_thread_fence_rel();
3296 vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK);
3300 * vm_page_dequeue_deferred: [ internal use only ]
3302 * Request removal of the given page from its current page
3303 * queue. Physical removal from the queue may be deferred
3306 * The page must be locked.
3309 vm_page_dequeue_deferred(vm_page_t m)
3313 vm_page_assert_locked(m);
3315 if ((queue = vm_page_queue(m)) == PQ_NONE)
3319 * Set PGA_DEQUEUE if it is not already set to handle a concurrent call
3320 * to vm_page_dequeue_deferred_free(). In particular, avoid modifying
3321 * the page's queue state once vm_page_dequeue_deferred_free() has been
3322 * called. In the event of a race, two batch queue entries for the page
3323 * will be created, but the second will have no effect.
3325 if (vm_page_pqstate_cmpset(m, queue, queue, PGA_DEQUEUE, PGA_DEQUEUE))
3326 vm_page_pqbatch_submit(m, queue);
3330 * A variant of vm_page_dequeue_deferred() that does not assert the page
3331 * lock and is only to be called from vm_page_free_prep(). Because the
3332 * page is being freed, we can assume that nothing other than the page
3333 * daemon is scheduling queue operations on this page, so we get for
3334 * free the mutual exclusion that is otherwise provided by the page lock.
3335 * To handle races, the page daemon must take care to atomically check
3336 * for PGA_DEQUEUE when updating queue state.
3339 vm_page_dequeue_deferred_free(vm_page_t m)
3343 KASSERT(m->ref_count == 0, ("page %p has references", m));
3346 if ((m->aflags & PGA_DEQUEUE) != 0)
3348 atomic_thread_fence_acq();
3349 if ((queue = atomic_load_8(&m->queue)) == PQ_NONE)
3351 if (vm_page_pqstate_cmpset(m, queue, queue, PGA_DEQUEUE,
3353 vm_page_pqbatch_submit(m, queue);
3362 * Remove the page from whichever page queue it's in, if any.
3363 * The page must either be locked or unallocated. This constraint
3364 * ensures that the queue state of the page will remain consistent
3365 * after this function returns.
3368 vm_page_dequeue(vm_page_t m)
3370 struct vm_pagequeue *pq, *pq1;
3373 KASSERT(mtx_owned(vm_page_lockptr(m)) || m->object == NULL,
3374 ("page %p is allocated and unlocked", m));
3376 for (pq = vm_page_pagequeue(m);; pq = pq1) {
3379 * A thread may be concurrently executing
3380 * vm_page_dequeue_complete(). Ensure that all queue
3381 * state is cleared before we return.
3383 aflags = atomic_load_8(&m->aflags);
3384 if ((aflags & PGA_QUEUE_STATE_MASK) == 0)
3386 KASSERT((aflags & PGA_DEQUEUE) != 0,
3387 ("page %p has unexpected queue state flags %#x",
3391 * Busy wait until the thread updating queue state is
3392 * finished. Such a thread must be executing in a
3396 pq1 = vm_page_pagequeue(m);
3399 vm_pagequeue_lock(pq);
3400 if ((pq1 = vm_page_pagequeue(m)) == pq)
3402 vm_pagequeue_unlock(pq);
3404 KASSERT(pq == vm_page_pagequeue(m),
3405 ("%s: page %p migrated directly between queues", __func__, m));
3406 KASSERT((m->aflags & PGA_DEQUEUE) != 0 ||
3407 mtx_owned(vm_page_lockptr(m)),
3408 ("%s: queued unlocked page %p", __func__, m));
3410 if ((m->aflags & PGA_ENQUEUED) != 0)
3411 vm_pagequeue_remove(pq, m);
3412 vm_page_dequeue_complete(m);
3413 vm_pagequeue_unlock(pq);
3417 * Schedule the given page for insertion into the specified page queue.
3418 * Physical insertion of the page may be deferred indefinitely.
3421 vm_page_enqueue(vm_page_t m, uint8_t queue)
3424 vm_page_assert_locked(m);
3425 KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0,
3426 ("%s: page %p is already enqueued", __func__, m));
3429 if ((m->aflags & PGA_REQUEUE) == 0)
3430 vm_page_aflag_set(m, PGA_REQUEUE);
3431 vm_page_pqbatch_submit(m, queue);
3435 * vm_page_requeue: [ internal use only ]
3437 * Schedule a requeue of the given page.
3439 * The page must be locked.
3442 vm_page_requeue(vm_page_t m)
3445 vm_page_assert_locked(m);
3446 KASSERT(vm_page_queue(m) != PQ_NONE,
3447 ("%s: page %p is not logically enqueued", __func__, m));
3449 if ((m->aflags & PGA_REQUEUE) == 0)
3450 vm_page_aflag_set(m, PGA_REQUEUE);
3451 vm_page_pqbatch_submit(m, atomic_load_8(&m->queue));
3455 * vm_page_swapqueue: [ internal use only ]
3457 * Move the page from one queue to another, or to the tail of its
3458 * current queue, in the face of a possible concurrent call to
3459 * vm_page_dequeue_deferred_free().
3462 vm_page_swapqueue(vm_page_t m, uint8_t oldq, uint8_t newq)
3464 struct vm_pagequeue *pq;
3468 KASSERT(oldq < PQ_COUNT && newq < PQ_COUNT && oldq != newq,
3469 ("vm_page_swapqueue: invalid queues (%d, %d)", oldq, newq));
3470 vm_page_assert_locked(m);
3472 pq = &vm_pagequeue_domain(m)->vmd_pagequeues[oldq];
3473 vm_pagequeue_lock(pq);
3476 * The physical queue state might change at any point before the page
3477 * queue lock is acquired, so we must verify that we hold the correct
3478 * lock before proceeding.
3480 if (__predict_false(m->queue != oldq)) {
3481 vm_pagequeue_unlock(pq);
3486 * Once the queue index of the page changes, there is nothing
3487 * synchronizing with further updates to the physical queue state.
3488 * Therefore we must remove the page from the queue now in anticipation
3489 * of a successful commit, and be prepared to roll back.
3491 if (__predict_true((m->aflags & PGA_ENQUEUED) != 0)) {
3492 next = TAILQ_NEXT(m, plinks.q);
3493 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3494 vm_page_aflag_clear(m, PGA_ENQUEUED);
3501 * Atomically update the queue field and set PGA_REQUEUE while
3502 * ensuring that PGA_DEQUEUE has not been set.
3504 if (__predict_false(!vm_page_pqstate_cmpset(m, oldq, newq, PGA_DEQUEUE,
3507 vm_page_aflag_set(m, PGA_ENQUEUED);
3509 TAILQ_INSERT_BEFORE(next, m, plinks.q);
3511 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3513 vm_pagequeue_unlock(pq);
3516 vm_pagequeue_cnt_dec(pq);
3517 vm_pagequeue_unlock(pq);
3518 vm_page_pqbatch_submit(m, newq);
3522 * vm_page_free_prep:
3524 * Prepares the given page to be put on the free list,
3525 * disassociating it from any VM object. The caller may return
3526 * the page to the free list only if this function returns true.
3528 * The object must be locked. The page must be locked if it is
3532 vm_page_free_prep(vm_page_t m)
3536 * Synchronize with threads that have dropped a reference to this
3539 atomic_thread_fence_acq();
3541 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3542 if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3545 p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3546 for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3547 KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3548 m, i, (uintmax_t)*p));
3551 if ((m->oflags & VPO_UNMANAGED) == 0)
3552 KASSERT(!pmap_page_is_mapped(m),
3553 ("vm_page_free_prep: freeing mapped page %p", m));
3555 KASSERT(m->queue == PQ_NONE,
3556 ("vm_page_free_prep: unmanaged page %p is queued", m));
3557 VM_CNT_INC(v_tfree);
3559 if (vm_page_sbusied(m))
3560 panic("vm_page_free_prep: freeing busy page %p", m);
3562 if (m->object != NULL) {
3563 vm_page_object_remove(m);
3566 * The object reference can be released without an atomic
3569 KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
3570 m->ref_count == VPRC_OBJREF,
3571 ("vm_page_free_prep: page %p has unexpected ref_count %u",
3574 m->ref_count -= VPRC_OBJREF;
3578 * If fictitious remove object association and
3581 if ((m->flags & PG_FICTITIOUS) != 0) {
3582 KASSERT(m->ref_count == 1,
3583 ("fictitious page %p is referenced", m));
3584 KASSERT(m->queue == PQ_NONE,
3585 ("fictitious page %p is queued", m));
3590 * Pages need not be dequeued before they are returned to the physical
3591 * memory allocator, but they must at least be marked for a deferred
3594 if ((m->oflags & VPO_UNMANAGED) == 0)
3595 vm_page_dequeue_deferred_free(m);
3600 if (m->ref_count != 0)
3601 panic("vm_page_free_prep: page %p has references", m);
3604 * Restore the default memory attribute to the page.
3606 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3607 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3609 #if VM_NRESERVLEVEL > 0
3611 * Determine whether the page belongs to a reservation. If the page was
3612 * allocated from a per-CPU cache, it cannot belong to a reservation, so
3613 * as an optimization, we avoid the check in that case.
3615 if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
3625 * Returns the given page to the free list, disassociating it
3626 * from any VM object.
3628 * The object must be locked. The page must be locked if it is
3632 vm_page_free_toq(vm_page_t m)
3634 struct vm_domain *vmd;
3637 if (!vm_page_free_prep(m))
3640 vmd = vm_pagequeue_domain(m);
3641 zone = vmd->vmd_pgcache[m->pool].zone;
3642 if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
3646 vm_domain_free_lock(vmd);
3647 vm_phys_free_pages(m, 0);
3648 vm_domain_free_unlock(vmd);
3649 vm_domain_freecnt_inc(vmd, 1);
3653 * vm_page_free_pages_toq:
3655 * Returns a list of pages to the free list, disassociating it
3656 * from any VM object. In other words, this is equivalent to
3657 * calling vm_page_free_toq() for each page of a list of VM objects.
3659 * The objects must be locked. The pages must be locked if it is
3663 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
3668 if (SLIST_EMPTY(free))
3672 while ((m = SLIST_FIRST(free)) != NULL) {
3674 SLIST_REMOVE_HEAD(free, plinks.s.ss);
3675 vm_page_free_toq(m);
3678 if (update_wire_count)
3683 * Mark this page as wired down, preventing reclamation by the page daemon
3684 * or when the containing object is destroyed.
3687 vm_page_wire(vm_page_t m)
3691 KASSERT(m->object != NULL,
3692 ("vm_page_wire: page %p does not belong to an object", m));
3693 if (!vm_page_busied(m))
3694 VM_OBJECT_ASSERT_LOCKED(m->object);
3695 KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
3696 VPRC_WIRE_COUNT(m->ref_count) >= 1,
3697 ("vm_page_wire: fictitious page %p has zero wirings", m));
3699 old = atomic_fetchadd_int(&m->ref_count, 1);
3700 KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
3701 ("vm_page_wire: counter overflow for page %p", m));
3702 if (VPRC_WIRE_COUNT(old) == 0)
3707 * Attempt to wire a mapped page following a pmap lookup of that page.
3708 * This may fail if a thread is concurrently tearing down mappings of the page.
3711 vm_page_wire_mapped(vm_page_t m)
3718 ("vm_page_wire_mapped: wiring unreferenced page %p", m));
3719 if ((old & VPRC_BLOCKED) != 0)
3721 } while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
3723 if (VPRC_WIRE_COUNT(old) == 0)
3729 * Release one wiring of the specified page, potentially allowing it to be
3732 * Only managed pages belonging to an object can be paged out. If the number
3733 * of wirings transitions to zero and the page is eligible for page out, then
3734 * the page is added to the specified paging queue. If the released wiring
3735 * represented the last reference to the page, the page is freed.
3737 * A managed page must be locked.
3740 vm_page_unwire(vm_page_t m, uint8_t queue)
3745 KASSERT(queue < PQ_COUNT,
3746 ("vm_page_unwire: invalid queue %u request for page %p", queue, m));
3748 if ((m->oflags & VPO_UNMANAGED) != 0) {
3749 if (vm_page_unwire_noq(m) && m->ref_count == 0)
3755 * Update LRU state before releasing the wiring reference.
3756 * We only need to do this once since we hold the page lock.
3757 * Use a release store when updating the reference count to
3758 * synchronize with vm_page_free_prep().
3763 KASSERT(VPRC_WIRE_COUNT(old) > 0,
3764 ("vm_page_unwire: wire count underflow for page %p", m));
3765 if (!locked && VPRC_WIRE_COUNT(old) == 1) {
3768 if (queue == PQ_ACTIVE && vm_page_queue(m) == PQ_ACTIVE)
3769 vm_page_reference(m);
3771 vm_page_mvqueue(m, queue);
3773 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
3776 * Release the lock only after the wiring is released, to ensure that
3777 * the page daemon does not encounter and dequeue the page while it is
3783 if (VPRC_WIRE_COUNT(old) == 1) {
3791 * Unwire a page without (re-)inserting it into a page queue. It is up
3792 * to the caller to enqueue, requeue, or free the page as appropriate.
3793 * In most cases involving managed pages, vm_page_unwire() should be used
3797 vm_page_unwire_noq(vm_page_t m)
3801 old = vm_page_drop(m, 1);
3802 KASSERT(VPRC_WIRE_COUNT(old) != 0,
3803 ("vm_page_unref: counter underflow for page %p", m));
3804 KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
3805 ("vm_page_unref: missing ref on fictitious page %p", m));
3807 if (VPRC_WIRE_COUNT(old) > 1)
3814 * Ensure that the page is in the specified page queue. If the page is
3815 * active or being moved to the active queue, ensure that its act_count is
3816 * at least ACT_INIT but do not otherwise mess with it. Otherwise, ensure that
3817 * the page is at the tail of its page queue.
3819 * The page may be wired. The caller should release its wiring reference
3820 * before releasing the page lock, otherwise the page daemon may immediately
3823 * A managed page must be locked.
3825 static __always_inline void
3826 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue)
3829 vm_page_assert_locked(m);
3830 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3831 ("vm_page_mvqueue: page %p is unmanaged", m));
3833 if (vm_page_queue(m) != nqueue) {
3835 vm_page_enqueue(m, nqueue);
3836 } else if (nqueue != PQ_ACTIVE) {
3840 if (nqueue == PQ_ACTIVE && m->act_count < ACT_INIT)
3841 m->act_count = ACT_INIT;
3845 * Put the specified page on the active list (if appropriate).
3848 vm_page_activate(vm_page_t m)
3851 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
3853 vm_page_mvqueue(m, PQ_ACTIVE);
3857 * Move the specified page to the tail of the inactive queue, or requeue
3858 * the page if it is already in the inactive queue.
3861 vm_page_deactivate(vm_page_t m)
3864 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
3866 vm_page_mvqueue(m, PQ_INACTIVE);
3870 * Move the specified page close to the head of the inactive queue,
3871 * bypassing LRU. A marker page is used to maintain FIFO ordering.
3872 * As with regular enqueues, we use a per-CPU batch queue to reduce
3873 * contention on the page queue lock.
3876 _vm_page_deactivate_noreuse(vm_page_t m)
3879 vm_page_assert_locked(m);
3881 if (!vm_page_inactive(m)) {
3883 m->queue = PQ_INACTIVE;
3885 if ((m->aflags & PGA_REQUEUE_HEAD) == 0)
3886 vm_page_aflag_set(m, PGA_REQUEUE_HEAD);
3887 vm_page_pqbatch_submit(m, PQ_INACTIVE);
3891 vm_page_deactivate_noreuse(vm_page_t m)
3894 KASSERT(m->object != NULL,
3895 ("vm_page_deactivate_noreuse: page %p has no object", m));
3897 if ((m->oflags & VPO_UNMANAGED) == 0 && !vm_page_wired(m))
3898 _vm_page_deactivate_noreuse(m);
3902 * Put a page in the laundry, or requeue it if it is already there.
3905 vm_page_launder(vm_page_t m)
3908 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
3910 vm_page_mvqueue(m, PQ_LAUNDRY);
3914 * Put a page in the PQ_UNSWAPPABLE holding queue.
3917 vm_page_unswappable(vm_page_t m)
3920 vm_page_assert_locked(m);
3921 KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0,
3922 ("page %p already unswappable", m));
3925 vm_page_enqueue(m, PQ_UNSWAPPABLE);
3929 vm_page_release_toq(vm_page_t m, int flags)
3932 vm_page_assert_locked(m);
3935 * Use a check of the valid bits to determine whether we should
3936 * accelerate reclamation of the page. The object lock might not be
3937 * held here, in which case the check is racy. At worst we will either
3938 * accelerate reclamation of a valid page and violate LRU, or
3939 * unnecessarily defer reclamation of an invalid page.
3941 * If we were asked to not cache the page, place it near the head of the
3942 * inactive queue so that is reclaimed sooner.
3944 if ((flags & (VPR_TRYFREE | VPR_NOREUSE)) != 0 || m->valid == 0)
3945 _vm_page_deactivate_noreuse(m);
3946 else if (vm_page_active(m))
3947 vm_page_reference(m);
3949 vm_page_mvqueue(m, PQ_INACTIVE);
3953 * Unwire a page and either attempt to free it or re-add it to the page queues.
3956 vm_page_release(vm_page_t m, int flags)
3962 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3963 ("vm_page_release: page %p is unmanaged", m));
3965 if ((flags & VPR_TRYFREE) != 0) {
3967 object = (vm_object_t)atomic_load_ptr(&m->object);
3970 /* Depends on type-stability. */
3971 if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object)) {
3975 if (object == m->object)
3977 VM_OBJECT_WUNLOCK(object);
3979 if (__predict_true(object != NULL)) {
3980 vm_page_release_locked(m, flags);
3981 VM_OBJECT_WUNLOCK(object);
3987 * Update LRU state before releasing the wiring reference.
3988 * Use a release store when updating the reference count to
3989 * synchronize with vm_page_free_prep().
3994 KASSERT(VPRC_WIRE_COUNT(old) > 0,
3995 ("vm_page_unwire: wire count underflow for page %p", m));
3996 if (!locked && VPRC_WIRE_COUNT(old) == 1) {
3999 vm_page_release_toq(m, flags);
4001 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
4004 * Release the lock only after the wiring is released, to ensure that
4005 * the page daemon does not encounter and dequeue the page while it is
4011 if (VPRC_WIRE_COUNT(old) == 1) {
4018 /* See vm_page_release(). */
4020 vm_page_release_locked(vm_page_t m, int flags)
4023 VM_OBJECT_ASSERT_WLOCKED(m->object);
4024 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4025 ("vm_page_release_locked: page %p is unmanaged", m));
4027 if (vm_page_unwire_noq(m)) {
4028 if ((flags & VPR_TRYFREE) != 0 &&
4029 (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
4030 m->dirty == 0 && !vm_page_busied(m)) {
4034 vm_page_release_toq(m, flags);
4041 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
4045 KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
4046 ("vm_page_try_blocked_op: page %p has no object", m));
4047 KASSERT(!vm_page_busied(m),
4048 ("vm_page_try_blocked_op: page %p is busy", m));
4049 VM_OBJECT_ASSERT_LOCKED(m->object);
4054 ("vm_page_try_blocked_op: page %p has no references", m));
4055 if (VPRC_WIRE_COUNT(old) != 0)
4057 } while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
4062 * If the object is read-locked, new wirings may be created via an
4065 old = vm_page_drop(m, VPRC_BLOCKED);
4066 KASSERT(!VM_OBJECT_WOWNED(m->object) ||
4067 old == (VPRC_BLOCKED | VPRC_OBJREF),
4068 ("vm_page_try_blocked_op: unexpected refcount value %u for %p",
4074 * Atomically check for wirings and remove all mappings of the page.
4077 vm_page_try_remove_all(vm_page_t m)
4080 return (vm_page_try_blocked_op(m, pmap_remove_all));
4084 * Atomically check for wirings and remove all writeable mappings of the page.
4087 vm_page_try_remove_write(vm_page_t m)
4090 return (vm_page_try_blocked_op(m, pmap_remove_write));
4096 * Apply the specified advice to the given page.
4098 * The object and page must be locked.
4101 vm_page_advise(vm_page_t m, int advice)
4104 vm_page_assert_locked(m);
4105 VM_OBJECT_ASSERT_WLOCKED(m->object);
4106 if (advice == MADV_FREE)
4108 * Mark the page clean. This will allow the page to be freed
4109 * without first paging it out. MADV_FREE pages are often
4110 * quickly reused by malloc(3), so we do not do anything that
4111 * would result in a page fault on a later access.
4114 else if (advice != MADV_DONTNEED) {
4115 if (advice == MADV_WILLNEED)
4116 vm_page_activate(m);
4121 * Clear any references to the page. Otherwise, the page daemon will
4122 * immediately reactivate the page.
4124 vm_page_aflag_clear(m, PGA_REFERENCED);
4126 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
4130 * Place clean pages near the head of the inactive queue rather than
4131 * the tail, thus defeating the queue's LRU operation and ensuring that
4132 * the page will be reused quickly. Dirty pages not already in the
4133 * laundry are moved there.
4136 vm_page_deactivate_noreuse(m);
4137 else if (!vm_page_in_laundry(m))
4142 * Grab a page, waiting until we are waken up due to the page
4143 * changing state. We keep on waiting, if the page continues
4144 * to be in the object. If the page doesn't exist, first allocate it
4145 * and then conditionally zero it.
4147 * This routine may sleep.
4149 * The object must be locked on entry. The lock will, however, be released
4150 * and reacquired if the routine sleeps.
4153 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
4159 VM_OBJECT_ASSERT_WLOCKED(object);
4160 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4161 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4162 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4163 pflags = allocflags &
4164 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
4165 if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4166 pflags |= VM_ALLOC_WAITFAIL;
4168 if ((m = vm_page_lookup(object, pindex)) != NULL) {
4169 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
4170 vm_page_xbusied(m) : vm_page_busied(m);
4172 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4175 * Reference the page before unlocking and
4176 * sleeping so that the page daemon is less
4177 * likely to reclaim it.
4179 if ((allocflags & VM_ALLOC_NOCREAT) == 0)
4180 vm_page_aflag_set(m, PGA_REFERENCED);
4181 vm_page_busy_sleep(m, "pgrbwt", (allocflags &
4182 VM_ALLOC_IGN_SBUSY) != 0);
4183 VM_OBJECT_WLOCK(object);
4184 if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
4188 if ((allocflags & VM_ALLOC_WIRED) != 0)
4191 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
4193 else if ((allocflags & VM_ALLOC_SBUSY) != 0)
4198 if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4200 m = vm_page_alloc(object, pindex, pflags);
4202 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4206 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
4212 * Grab a page and make it valid, paging in if necessary. Pages missing from
4213 * their pager are zero filled and validated.
4216 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags)
4223 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4224 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4225 ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4226 KASSERT((allocflags &
4227 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4228 ("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
4229 VM_OBJECT_ASSERT_WLOCKED(object);
4230 pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY);
4231 pflags |= VM_ALLOC_WAITFAIL;
4235 if ((m = vm_page_lookup(object, pindex)) != NULL) {
4237 * If the page is fully valid it can only become invalid
4238 * with the object lock held. If it is not valid it can
4239 * become valid with the busy lock held. Therefore, we
4240 * may unnecessarily lock the exclusive busy here if we
4241 * race with I/O completion not using the object lock.
4242 * However, we will not end up with an invalid page and a
4245 if (m->valid != VM_PAGE_BITS_ALL ||
4246 (allocflags & (VM_ALLOC_IGN_SBUSY | VM_ALLOC_SBUSY)) == 0) {
4247 sleep = !vm_page_tryxbusy(m);
4250 sleep = !vm_page_trysbusy(m);
4253 * Reference the page before unlocking and
4254 * sleeping so that the page daemon is less
4255 * likely to reclaim it.
4257 if ((allocflags & VM_ALLOC_NOCREAT) == 0)
4258 vm_page_aflag_set(m, PGA_REFERENCED);
4259 vm_page_busy_sleep(m, "pgrbwt", (allocflags &
4260 VM_ALLOC_IGN_SBUSY) != 0);
4261 VM_OBJECT_WLOCK(object);
4264 if ((allocflags & VM_ALLOC_NOCREAT) != 0 &&
4265 m->valid != VM_PAGE_BITS_ALL) {
4271 return (VM_PAGER_FAIL);
4273 if ((allocflags & VM_ALLOC_WIRED) != 0)
4275 if (m->valid == VM_PAGE_BITS_ALL)
4277 } else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4279 return (VM_PAGER_FAIL);
4280 } else if ((m = vm_page_alloc(object, pindex, pflags)) != NULL) {
4286 vm_page_assert_xbusied(m);
4288 if (vm_pager_has_page(object, pindex, NULL, NULL)) {
4289 rv = vm_pager_get_pages(object, &m, 1, NULL, NULL);
4290 if (rv != VM_PAGER_OK) {
4291 if (allocflags & VM_ALLOC_WIRED)
4292 vm_page_unwire_noq(m);
4297 MPASS(m->valid == VM_PAGE_BITS_ALL);
4299 vm_page_zero_invalid(m, TRUE);
4302 if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4308 if ((allocflags & VM_ALLOC_SBUSY) != 0 && xbusy)
4309 vm_page_busy_downgrade(m);
4311 return (VM_PAGER_OK);
4315 * Return the specified range of pages from the given object. For each
4316 * page offset within the range, if a page already exists within the object
4317 * at that offset and it is busy, then wait for it to change state. If,
4318 * instead, the page doesn't exist, then allocate it.
4320 * The caller must always specify an allocation class.
4322 * allocation classes:
4323 * VM_ALLOC_NORMAL normal process request
4324 * VM_ALLOC_SYSTEM system *really* needs the pages
4326 * The caller must always specify that the pages are to be busied and/or
4329 * optional allocation flags:
4330 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages
4331 * VM_ALLOC_NOBUSY do not exclusive busy the page
4332 * VM_ALLOC_NOWAIT do not sleep
4333 * VM_ALLOC_SBUSY set page to sbusy state
4334 * VM_ALLOC_WIRED wire the pages
4335 * VM_ALLOC_ZERO zero and validate any invalid pages
4337 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it
4338 * may return a partial prefix of the requested range.
4341 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
4342 vm_page_t *ma, int count)
4349 VM_OBJECT_ASSERT_WLOCKED(object);
4350 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
4351 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
4352 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4353 (allocflags & VM_ALLOC_WIRED) != 0,
4354 ("vm_page_grab_pages: the pages must be busied or wired"));
4355 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4356 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4357 ("vm_page_grab_pages: VM_ALLOC_SBUSY/IGN_SBUSY mismatch"));
4360 pflags = allocflags & ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK |
4361 VM_ALLOC_WAITFAIL | VM_ALLOC_IGN_SBUSY);
4362 if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4363 pflags |= VM_ALLOC_WAITFAIL;
4366 m = vm_radix_lookup_le(&object->rtree, pindex + i);
4367 if (m == NULL || m->pindex != pindex + i) {
4371 mpred = TAILQ_PREV(m, pglist, listq);
4372 for (; i < count; i++) {
4374 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
4375 vm_page_xbusied(m) : vm_page_busied(m);
4377 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4380 * Reference the page before unlocking and
4381 * sleeping so that the page daemon is less
4382 * likely to reclaim it.
4384 if ((allocflags & VM_ALLOC_NOCREAT) == 0)
4385 vm_page_aflag_set(m, PGA_REFERENCED);
4386 vm_page_busy_sleep(m, "grbmaw", (allocflags &
4387 VM_ALLOC_IGN_SBUSY) != 0);
4388 VM_OBJECT_WLOCK(object);
4391 if ((allocflags & VM_ALLOC_WIRED) != 0)
4393 if ((allocflags & (VM_ALLOC_NOBUSY |
4394 VM_ALLOC_SBUSY)) == 0)
4396 if ((allocflags & VM_ALLOC_SBUSY) != 0)
4399 if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4401 m = vm_page_alloc_after(object, pindex + i,
4402 pflags | VM_ALLOC_COUNT(count - i), mpred);
4404 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4409 if (m->valid == 0 && (allocflags & VM_ALLOC_ZERO) != 0) {
4410 if ((m->flags & PG_ZERO) == 0)
4412 m->valid = VM_PAGE_BITS_ALL;
4415 m = vm_page_next(m);
4421 * Mapping function for valid or dirty bits in a page.
4423 * Inputs are required to range within a page.
4426 vm_page_bits(int base, int size)
4432 base + size <= PAGE_SIZE,
4433 ("vm_page_bits: illegal base/size %d/%d", base, size)
4436 if (size == 0) /* handle degenerate case */
4439 first_bit = base >> DEV_BSHIFT;
4440 last_bit = (base + size - 1) >> DEV_BSHIFT;
4442 return (((vm_page_bits_t)2 << last_bit) -
4443 ((vm_page_bits_t)1 << first_bit));
4447 * vm_page_set_valid_range:
4449 * Sets portions of a page valid. The arguments are expected
4450 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4451 * of any partial chunks touched by the range. The invalid portion of
4452 * such chunks will be zeroed.
4454 * (base + size) must be less then or equal to PAGE_SIZE.
4457 vm_page_set_valid_range(vm_page_t m, int base, int size)
4461 VM_OBJECT_ASSERT_WLOCKED(m->object);
4462 if (size == 0) /* handle degenerate case */
4466 * If the base is not DEV_BSIZE aligned and the valid
4467 * bit is clear, we have to zero out a portion of the
4470 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4471 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
4472 pmap_zero_page_area(m, frag, base - frag);
4475 * If the ending offset is not DEV_BSIZE aligned and the
4476 * valid bit is clear, we have to zero out a portion of
4479 endoff = base + size;
4480 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4481 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
4482 pmap_zero_page_area(m, endoff,
4483 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4486 * Assert that no previously invalid block that is now being validated
4489 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
4490 ("vm_page_set_valid_range: page %p is dirty", m));
4493 * Set valid bits inclusive of any overlap.
4495 m->valid |= vm_page_bits(base, size);
4499 * Clear the given bits from the specified page's dirty field.
4501 static __inline void
4502 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
4505 #if PAGE_SIZE < 16384
4510 * If the object is locked and the page is neither exclusive busy nor
4511 * write mapped, then the page's dirty field cannot possibly be
4512 * set by a concurrent pmap operation.
4514 VM_OBJECT_ASSERT_WLOCKED(m->object);
4515 if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
4516 m->dirty &= ~pagebits;
4519 * The pmap layer can call vm_page_dirty() without
4520 * holding a distinguished lock. The combination of
4521 * the object's lock and an atomic operation suffice
4522 * to guarantee consistency of the page dirty field.
4524 * For PAGE_SIZE == 32768 case, compiler already
4525 * properly aligns the dirty field, so no forcible
4526 * alignment is needed. Only require existence of
4527 * atomic_clear_64 when page size is 32768.
4529 addr = (uintptr_t)&m->dirty;
4530 #if PAGE_SIZE == 32768
4531 atomic_clear_64((uint64_t *)addr, pagebits);
4532 #elif PAGE_SIZE == 16384
4533 atomic_clear_32((uint32_t *)addr, pagebits);
4534 #else /* PAGE_SIZE <= 8192 */
4536 * Use a trick to perform a 32-bit atomic on the
4537 * containing aligned word, to not depend on the existence
4538 * of atomic_clear_{8, 16}.
4540 shift = addr & (sizeof(uint32_t) - 1);
4541 #if BYTE_ORDER == BIG_ENDIAN
4542 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
4546 addr &= ~(sizeof(uint32_t) - 1);
4547 atomic_clear_32((uint32_t *)addr, pagebits << shift);
4548 #endif /* PAGE_SIZE */
4553 * vm_page_set_validclean:
4555 * Sets portions of a page valid and clean. The arguments are expected
4556 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4557 * of any partial chunks touched by the range. The invalid portion of
4558 * such chunks will be zero'd.
4560 * (base + size) must be less then or equal to PAGE_SIZE.
4563 vm_page_set_validclean(vm_page_t m, int base, int size)
4565 vm_page_bits_t oldvalid, pagebits;
4568 VM_OBJECT_ASSERT_WLOCKED(m->object);
4569 if (size == 0) /* handle degenerate case */
4573 * If the base is not DEV_BSIZE aligned and the valid
4574 * bit is clear, we have to zero out a portion of the
4577 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4578 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
4579 pmap_zero_page_area(m, frag, base - frag);
4582 * If the ending offset is not DEV_BSIZE aligned and the
4583 * valid bit is clear, we have to zero out a portion of
4586 endoff = base + size;
4587 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4588 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
4589 pmap_zero_page_area(m, endoff,
4590 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4593 * Set valid, clear dirty bits. If validating the entire
4594 * page we can safely clear the pmap modify bit. We also
4595 * use this opportunity to clear the VPO_NOSYNC flag. If a process
4596 * takes a write fault on a MAP_NOSYNC memory area the flag will
4599 * We set valid bits inclusive of any overlap, but we can only
4600 * clear dirty bits for DEV_BSIZE chunks that are fully within
4603 oldvalid = m->valid;
4604 pagebits = vm_page_bits(base, size);
4605 m->valid |= pagebits;
4607 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
4608 frag = DEV_BSIZE - frag;
4614 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
4616 if (base == 0 && size == PAGE_SIZE) {
4618 * The page can only be modified within the pmap if it is
4619 * mapped, and it can only be mapped if it was previously
4622 if (oldvalid == VM_PAGE_BITS_ALL)
4624 * Perform the pmap_clear_modify() first. Otherwise,
4625 * a concurrent pmap operation, such as
4626 * pmap_protect(), could clear a modification in the
4627 * pmap and set the dirty field on the page before
4628 * pmap_clear_modify() had begun and after the dirty
4629 * field was cleared here.
4631 pmap_clear_modify(m);
4633 m->oflags &= ~VPO_NOSYNC;
4634 } else if (oldvalid != VM_PAGE_BITS_ALL)
4635 m->dirty &= ~pagebits;
4637 vm_page_clear_dirty_mask(m, pagebits);
4641 vm_page_clear_dirty(vm_page_t m, int base, int size)
4644 vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
4648 * vm_page_set_invalid:
4650 * Invalidates DEV_BSIZE'd chunks within a page. Both the
4651 * valid and dirty bits for the effected areas are cleared.
4654 vm_page_set_invalid(vm_page_t m, int base, int size)
4656 vm_page_bits_t bits;
4660 VM_OBJECT_ASSERT_WLOCKED(object);
4661 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
4662 size >= object->un_pager.vnp.vnp_size)
4663 bits = VM_PAGE_BITS_ALL;
4665 bits = vm_page_bits(base, size);
4666 if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL &&
4669 KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) ||
4670 !pmap_page_is_mapped(m),
4671 ("vm_page_set_invalid: page %p is mapped", m));
4677 * vm_page_zero_invalid()
4679 * The kernel assumes that the invalid portions of a page contain
4680 * garbage, but such pages can be mapped into memory by user code.
4681 * When this occurs, we must zero out the non-valid portions of the
4682 * page so user code sees what it expects.
4684 * Pages are most often semi-valid when the end of a file is mapped
4685 * into memory and the file's size is not page aligned.
4688 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
4693 VM_OBJECT_ASSERT_WLOCKED(m->object);
4695 * Scan the valid bits looking for invalid sections that
4696 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the
4697 * valid bit may be set ) have already been zeroed by
4698 * vm_page_set_validclean().
4700 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
4701 if (i == (PAGE_SIZE / DEV_BSIZE) ||
4702 (m->valid & ((vm_page_bits_t)1 << i))) {
4704 pmap_zero_page_area(m,
4705 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
4712 * setvalid is TRUE when we can safely set the zero'd areas
4713 * as being valid. We can do this if there are no cache consistancy
4714 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
4717 m->valid = VM_PAGE_BITS_ALL;
4723 * Is (partial) page valid? Note that the case where size == 0
4724 * will return FALSE in the degenerate case where the page is
4725 * entirely invalid, and TRUE otherwise.
4728 vm_page_is_valid(vm_page_t m, int base, int size)
4730 vm_page_bits_t bits;
4732 VM_OBJECT_ASSERT_LOCKED(m->object);
4733 bits = vm_page_bits(base, size);
4734 return (m->valid != 0 && (m->valid & bits) == bits);
4738 * Returns true if all of the specified predicates are true for the entire
4739 * (super)page and false otherwise.
4742 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
4748 if (skip_m != NULL && skip_m->object != object)
4750 VM_OBJECT_ASSERT_LOCKED(object);
4751 npages = atop(pagesizes[m->psind]);
4754 * The physically contiguous pages that make up a superpage, i.e., a
4755 * page with a page size index ("psind") greater than zero, will
4756 * occupy adjacent entries in vm_page_array[].
4758 for (i = 0; i < npages; i++) {
4759 /* Always test object consistency, including "skip_m". */
4760 if (m[i].object != object)
4762 if (&m[i] == skip_m)
4764 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
4766 if ((flags & PS_ALL_DIRTY) != 0) {
4768 * Calling vm_page_test_dirty() or pmap_is_modified()
4769 * might stop this case from spuriously returning
4770 * "false". However, that would require a write lock
4771 * on the object containing "m[i]".
4773 if (m[i].dirty != VM_PAGE_BITS_ALL)
4776 if ((flags & PS_ALL_VALID) != 0 &&
4777 m[i].valid != VM_PAGE_BITS_ALL)
4784 * Set the page's dirty bits if the page is modified.
4787 vm_page_test_dirty(vm_page_t m)
4790 VM_OBJECT_ASSERT_WLOCKED(m->object);
4791 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
4796 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
4799 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
4803 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
4806 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
4810 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
4813 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
4816 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
4818 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
4821 vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
4825 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
4828 mtx_assert_(vm_page_lockptr(m), a, file, line);
4834 vm_page_object_lock_assert(vm_page_t m)
4838 * Certain of the page's fields may only be modified by the
4839 * holder of the containing object's lock or the exclusive busy.
4840 * holder. Unfortunately, the holder of the write busy is
4841 * not recorded, and thus cannot be checked here.
4843 if (m->object != NULL && !vm_page_xbusied(m))
4844 VM_OBJECT_ASSERT_WLOCKED(m->object);
4848 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits)
4851 if ((bits & PGA_WRITEABLE) == 0)
4855 * The PGA_WRITEABLE flag can only be set if the page is
4856 * managed, is exclusively busied or the object is locked.
4857 * Currently, this flag is only set by pmap_enter().
4859 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4860 ("PGA_WRITEABLE on unmanaged page"));
4861 if (!vm_page_xbusied(m))
4862 VM_OBJECT_ASSERT_LOCKED(m->object);
4866 #include "opt_ddb.h"
4868 #include <sys/kernel.h>
4870 #include <ddb/ddb.h>
4872 DB_SHOW_COMMAND(page, vm_page_print_page_info)
4875 db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
4876 db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
4877 db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
4878 db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
4879 db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
4880 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
4881 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
4882 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
4883 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
4886 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
4890 db_printf("pq_free %d\n", vm_free_count());
4891 for (dom = 0; dom < vm_ndomains; dom++) {
4893 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
4895 vm_dom[dom].vmd_page_count,
4896 vm_dom[dom].vmd_free_count,
4897 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
4898 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
4899 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
4900 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
4904 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
4907 boolean_t phys, virt;
4910 db_printf("show pginfo addr\n");
4914 phys = strchr(modif, 'p') != NULL;
4915 virt = strchr(modif, 'v') != NULL;
4917 m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
4919 m = PHYS_TO_VM_PAGE(addr);
4921 m = (vm_page_t)addr;
4923 "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref %u\n"
4924 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
4925 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
4926 m->queue, m->ref_count, m->aflags, m->oflags,
4927 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty);