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) || \
662 defined(__powerpc64__)
664 * Allocate a bitmap to indicate that a random physical page
665 * needs to be included in a minidump.
667 * The amd64 port needs this to indicate which direct map pages
668 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
670 * However, i386 still needs this workspace internally within the
671 * minidump code. In theory, they are not needed on i386, but are
672 * included should the sf_buf code decide to use them.
675 for (i = 0; dump_avail[i + 1] != 0; i += 2)
676 if (dump_avail[i + 1] > last_pa)
677 last_pa = dump_avail[i + 1];
678 page_range = last_pa / PAGE_SIZE;
679 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
680 new_end -= vm_page_dump_size;
681 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
682 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
683 bzero((void *)vm_page_dump, vm_page_dump_size);
687 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
688 defined(__riscv) || defined(__powerpc64__)
690 * Include the UMA bootstrap pages, witness pages and vm_page_dump
691 * in a crash dump. When pmap_map() uses the direct map, they are
692 * not automatically included.
694 for (pa = new_end; pa < end; pa += PAGE_SIZE)
697 phys_avail[biggestone + 1] = new_end;
700 * Request that the physical pages underlying the message buffer be
701 * included in a crash dump. Since the message buffer is accessed
702 * through the direct map, they are not automatically included.
704 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
705 last_pa = pa + round_page(msgbufsize);
706 while (pa < last_pa) {
712 * Compute the number of pages of memory that will be available for
713 * use, taking into account the overhead of a page structure per page.
714 * In other words, solve
715 * "available physical memory" - round_page(page_range *
716 * sizeof(struct vm_page)) = page_range * PAGE_SIZE
719 low_avail = phys_avail[0];
720 high_avail = phys_avail[1];
721 for (i = 0; i < vm_phys_nsegs; i++) {
722 if (vm_phys_segs[i].start < low_avail)
723 low_avail = vm_phys_segs[i].start;
724 if (vm_phys_segs[i].end > high_avail)
725 high_avail = vm_phys_segs[i].end;
727 /* Skip the first chunk. It is already accounted for. */
728 for (i = 2; phys_avail[i + 1] != 0; i += 2) {
729 if (phys_avail[i] < low_avail)
730 low_avail = phys_avail[i];
731 if (phys_avail[i + 1] > high_avail)
732 high_avail = phys_avail[i + 1];
734 first_page = low_avail / PAGE_SIZE;
735 #ifdef VM_PHYSSEG_SPARSE
737 for (i = 0; i < vm_phys_nsegs; i++)
738 size += vm_phys_segs[i].end - vm_phys_segs[i].start;
739 for (i = 0; phys_avail[i + 1] != 0; i += 2)
740 size += phys_avail[i + 1] - phys_avail[i];
741 #elif defined(VM_PHYSSEG_DENSE)
742 size = high_avail - low_avail;
744 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
747 #ifdef PMAP_HAS_PAGE_ARRAY
748 pmap_page_array_startup(size / PAGE_SIZE);
749 biggestone = vm_phys_avail_largest();
750 end = new_end = phys_avail[biggestone + 1];
752 #ifdef VM_PHYSSEG_DENSE
754 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
755 * the overhead of a page structure per page only if vm_page_array is
756 * allocated from the last physical memory chunk. Otherwise, we must
757 * allocate page structures representing the physical memory
758 * underlying vm_page_array, even though they will not be used.
760 if (new_end != high_avail)
761 page_range = size / PAGE_SIZE;
765 page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
768 * If the partial bytes remaining are large enough for
769 * a page (PAGE_SIZE) without a corresponding
770 * 'struct vm_page', then new_end will contain an
771 * extra page after subtracting the length of the VM
772 * page array. Compensate by subtracting an extra
775 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
776 if (new_end == high_avail)
777 high_avail -= PAGE_SIZE;
778 new_end -= PAGE_SIZE;
782 new_end = vm_page_array_alloc(&vaddr, end, page_range);
785 #if VM_NRESERVLEVEL > 0
787 * Allocate physical memory for the reservation management system's
788 * data structures, and map it.
790 new_end = vm_reserv_startup(&vaddr, new_end);
792 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
793 defined(__riscv) || defined(__powerpc64__)
795 * Include vm_page_array and vm_reserv_array in a crash dump.
797 for (pa = new_end; pa < end; pa += PAGE_SIZE)
800 phys_avail[biggestone + 1] = new_end;
803 * Add physical memory segments corresponding to the available
806 for (i = 0; phys_avail[i + 1] != 0; i += 2)
807 if (vm_phys_avail_size(i) != 0)
808 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
811 * Initialize the physical memory allocator.
816 * Initialize the page structures and add every available page to the
817 * physical memory allocator's free lists.
819 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
820 for (ii = 0; ii < vm_page_array_size; ii++) {
821 m = &vm_page_array[ii];
822 vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0);
823 m->flags = PG_FICTITIOUS;
826 vm_cnt.v_page_count = 0;
827 for (segind = 0; segind < vm_phys_nsegs; segind++) {
828 seg = &vm_phys_segs[segind];
829 for (m = seg->first_page, pa = seg->start; pa < seg->end;
830 m++, pa += PAGE_SIZE)
831 vm_page_init_page(m, pa, segind);
834 * Add the segment to the free lists only if it is covered by
835 * one of the ranges in phys_avail. Because we've added the
836 * ranges to the vm_phys_segs array, we can assume that each
837 * segment is either entirely contained in one of the ranges,
838 * or doesn't overlap any of them.
840 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
841 struct vm_domain *vmd;
843 if (seg->start < phys_avail[i] ||
844 seg->end > phys_avail[i + 1])
848 pagecount = (u_long)atop(seg->end - seg->start);
850 vmd = VM_DOMAIN(seg->domain);
851 vm_domain_free_lock(vmd);
852 vm_phys_enqueue_contig(m, pagecount);
853 vm_domain_free_unlock(vmd);
854 vm_domain_freecnt_inc(vmd, pagecount);
855 vm_cnt.v_page_count += (u_int)pagecount;
857 vmd = VM_DOMAIN(seg->domain);
858 vmd->vmd_page_count += (u_int)pagecount;
859 vmd->vmd_segs |= 1UL << m->segind;
865 * Remove blacklisted pages from the physical memory allocator.
867 TAILQ_INIT(&blacklist_head);
868 vm_page_blacklist_load(&list, &listend);
869 vm_page_blacklist_check(list, listend);
871 list = kern_getenv("vm.blacklist");
872 vm_page_blacklist_check(list, NULL);
875 #if VM_NRESERVLEVEL > 0
877 * Initialize the reservation management system.
886 vm_page_reference(vm_page_t m)
889 vm_page_aflag_set(m, PGA_REFERENCED);
893 * vm_page_busy_acquire:
895 * Acquire the busy lock as described by VM_ALLOC_* flags. Will loop
896 * and drop the object lock if necessary.
899 vm_page_busy_acquire(vm_page_t m, int allocflags)
906 * The page-specific object must be cached because page
907 * identity can change during the sleep, causing the
908 * re-lock of a different object.
909 * It is assumed that a reference to the object is already
910 * held by the callers.
914 if ((allocflags & VM_ALLOC_SBUSY) == 0) {
915 if (vm_page_tryxbusy(m))
918 if (vm_page_trysbusy(m))
921 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
924 locked = VM_OBJECT_WOWNED(obj);
926 MPASS(vm_page_wired(m));
931 if (x == VPB_UNBUSIED ||
932 ((allocflags & VM_ALLOC_SBUSY) != 0 &&
933 (x & VPB_BIT_SHARED) != 0) ||
934 ((x & VPB_BIT_WAITERS) == 0 &&
935 !atomic_cmpset_int(&m->busy_lock, x,
936 x | VPB_BIT_WAITERS))) {
941 VM_OBJECT_WUNLOCK(obj);
942 sleepq_add(m, NULL, "vmpba", 0, 0);
945 VM_OBJECT_WLOCK(obj);
946 MPASS(m->object == obj || m->object == NULL);
947 if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
953 * vm_page_busy_downgrade:
955 * Downgrade an exclusive busy page into a single shared busy page.
958 vm_page_busy_downgrade(vm_page_t m)
962 vm_page_assert_xbusied(m);
966 if (atomic_fcmpset_rel_int(&m->busy_lock,
967 &x, VPB_SHARERS_WORD(1)))
970 if ((x & VPB_BIT_WAITERS) != 0)
977 * Return a positive value if the page is shared busied, 0 otherwise.
980 vm_page_sbusied(vm_page_t m)
985 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
991 * Shared unbusy a page.
994 vm_page_sunbusy(vm_page_t m)
998 vm_page_assert_sbusied(m);
1002 if (VPB_SHARERS(x) > 1) {
1003 if (atomic_fcmpset_int(&m->busy_lock, &x,
1004 x - VPB_ONE_SHARER))
1008 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
1009 ("vm_page_sunbusy: invalid lock state"));
1010 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1012 if ((x & VPB_BIT_WAITERS) == 0)
1020 * vm_page_busy_sleep:
1022 * Sleep if the page is busy, using the page pointer as wchan.
1023 * This is used to implement the hard-path of busying mechanism.
1025 * If nonshared is true, sleep only if the page is xbusy.
1027 * The object lock must be held on entry and will be released on exit.
1030 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
1036 vm_page_lock_assert(m, MA_NOTOWNED);
1037 VM_OBJECT_ASSERT_LOCKED(obj);
1041 if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) ||
1042 ((x & VPB_BIT_WAITERS) == 0 &&
1043 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) {
1044 VM_OBJECT_DROP(obj);
1048 VM_OBJECT_DROP(obj);
1049 sleepq_add(m, NULL, wmesg, 0, 0);
1050 sleepq_wait(m, PVM);
1056 * Try to shared busy a page.
1057 * If the operation succeeds 1 is returned otherwise 0.
1058 * The operation never sleeps.
1061 vm_page_trysbusy(vm_page_t m)
1067 if ((x & VPB_BIT_SHARED) == 0)
1069 if (atomic_fcmpset_acq_int(&m->busy_lock, &x,
1070 x + VPB_ONE_SHARER))
1076 * vm_page_xunbusy_hard:
1078 * Called when unbusy has failed because there is a waiter.
1081 vm_page_xunbusy_hard(vm_page_t m)
1084 vm_page_assert_xbusied(m);
1089 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1094 * Avoid releasing and reacquiring the same page lock.
1097 vm_page_change_lock(vm_page_t m, struct mtx **mtx)
1101 mtx1 = vm_page_lockptr(m);
1111 * vm_page_unhold_pages:
1113 * Unhold each of the pages that is referenced by the given array.
1116 vm_page_unhold_pages(vm_page_t *ma, int count)
1119 for (; count != 0; count--) {
1120 vm_page_unwire(*ma, PQ_ACTIVE);
1126 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1130 #ifdef VM_PHYSSEG_SPARSE
1131 m = vm_phys_paddr_to_vm_page(pa);
1133 m = vm_phys_fictitious_to_vm_page(pa);
1135 #elif defined(VM_PHYSSEG_DENSE)
1139 if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1140 m = &vm_page_array[pi - first_page];
1143 return (vm_phys_fictitious_to_vm_page(pa));
1145 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1152 * Create a fictitious page with the specified physical address and
1153 * memory attribute. The memory attribute is the only the machine-
1154 * dependent aspect of a fictitious page that must be initialized.
1157 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1161 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1162 vm_page_initfake(m, paddr, memattr);
1167 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1170 if ((m->flags & PG_FICTITIOUS) != 0) {
1172 * The page's memattr might have changed since the
1173 * previous initialization. Update the pmap to the
1178 m->phys_addr = paddr;
1180 /* Fictitious pages don't use "segind". */
1181 m->flags = PG_FICTITIOUS;
1182 /* Fictitious pages don't use "order" or "pool". */
1183 m->oflags = VPO_UNMANAGED;
1184 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1185 /* Fictitious pages are unevictable. */
1189 pmap_page_set_memattr(m, memattr);
1195 * Release a fictitious page.
1198 vm_page_putfake(vm_page_t m)
1201 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1202 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1203 ("vm_page_putfake: bad page %p", m));
1204 uma_zfree(fakepg_zone, m);
1208 * vm_page_updatefake:
1210 * Update the given fictitious page to the specified physical address and
1214 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1217 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1218 ("vm_page_updatefake: bad page %p", m));
1219 m->phys_addr = paddr;
1220 pmap_page_set_memattr(m, memattr);
1229 vm_page_free(vm_page_t m)
1232 m->flags &= ~PG_ZERO;
1233 vm_page_free_toq(m);
1237 * vm_page_free_zero:
1239 * Free a page to the zerod-pages queue
1242 vm_page_free_zero(vm_page_t m)
1245 m->flags |= PG_ZERO;
1246 vm_page_free_toq(m);
1250 * Unbusy and handle the page queueing for a page from a getpages request that
1251 * was optionally read ahead or behind.
1254 vm_page_readahead_finish(vm_page_t m)
1257 /* We shouldn't put invalid pages on queues. */
1258 KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m));
1261 * Since the page is not the actually needed one, whether it should
1262 * be activated or deactivated is not obvious. Empirical results
1263 * have shown that deactivating the page is usually the best choice,
1264 * unless the page is wanted by another thread.
1267 if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1268 vm_page_activate(m);
1270 vm_page_deactivate(m);
1276 * vm_page_sleep_if_busy:
1278 * Sleep and release the object lock if the page is busied.
1279 * Returns TRUE if the thread slept.
1281 * The given page must be unlocked and object containing it must
1285 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1289 vm_page_lock_assert(m, MA_NOTOWNED);
1290 VM_OBJECT_ASSERT_WLOCKED(m->object);
1292 if (vm_page_busied(m)) {
1294 * The page-specific object must be cached because page
1295 * identity can change during the sleep, causing the
1296 * re-lock of a different object.
1297 * It is assumed that a reference to the object is already
1298 * held by the callers.
1301 vm_page_busy_sleep(m, msg, false);
1302 VM_OBJECT_WLOCK(obj);
1309 * vm_page_sleep_if_xbusy:
1311 * Sleep and release the object lock if the page is xbusied.
1312 * Returns TRUE if the thread slept.
1314 * The given page must be unlocked and object containing it must
1318 vm_page_sleep_if_xbusy(vm_page_t m, const char *msg)
1322 vm_page_lock_assert(m, MA_NOTOWNED);
1323 VM_OBJECT_ASSERT_WLOCKED(m->object);
1325 if (vm_page_xbusied(m)) {
1327 * The page-specific object must be cached because page
1328 * identity can change during the sleep, causing the
1329 * re-lock of a different object.
1330 * It is assumed that a reference to the object is already
1331 * held by the callers.
1334 vm_page_busy_sleep(m, msg, true);
1335 VM_OBJECT_WLOCK(obj);
1342 * vm_page_dirty_KBI: [ internal use only ]
1344 * Set all bits in the page's dirty field.
1346 * The object containing the specified page must be locked if the
1347 * call is made from the machine-independent layer.
1349 * See vm_page_clear_dirty_mask().
1351 * This function should only be called by vm_page_dirty().
1354 vm_page_dirty_KBI(vm_page_t m)
1357 /* Refer to this operation by its public name. */
1358 KASSERT(m->valid == VM_PAGE_BITS_ALL,
1359 ("vm_page_dirty: page is invalid!"));
1360 m->dirty = VM_PAGE_BITS_ALL;
1364 * vm_page_insert: [ internal use only ]
1366 * Inserts the given mem entry into the object and object list.
1368 * The object must be locked.
1371 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1375 VM_OBJECT_ASSERT_WLOCKED(object);
1376 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1377 return (vm_page_insert_after(m, object, pindex, mpred));
1381 * vm_page_insert_after:
1383 * Inserts the page "m" into the specified object at offset "pindex".
1385 * The page "mpred" must immediately precede the offset "pindex" within
1386 * the specified object.
1388 * The object must be locked.
1391 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1396 VM_OBJECT_ASSERT_WLOCKED(object);
1397 KASSERT(m->object == NULL,
1398 ("vm_page_insert_after: page already inserted"));
1399 if (mpred != NULL) {
1400 KASSERT(mpred->object == object,
1401 ("vm_page_insert_after: object doesn't contain mpred"));
1402 KASSERT(mpred->pindex < pindex,
1403 ("vm_page_insert_after: mpred doesn't precede pindex"));
1404 msucc = TAILQ_NEXT(mpred, listq);
1406 msucc = TAILQ_FIRST(&object->memq);
1408 KASSERT(msucc->pindex > pindex,
1409 ("vm_page_insert_after: msucc doesn't succeed pindex"));
1412 * Record the object/offset pair in this page.
1416 m->ref_count |= VPRC_OBJREF;
1419 * Now link into the object's ordered list of backed pages.
1421 if (vm_radix_insert(&object->rtree, m)) {
1424 m->ref_count &= ~VPRC_OBJREF;
1427 vm_page_insert_radixdone(m, object, mpred);
1432 * vm_page_insert_radixdone:
1434 * Complete page "m" insertion into the specified object after the
1435 * radix trie hooking.
1437 * The page "mpred" must precede the offset "m->pindex" within the
1440 * The object must be locked.
1443 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1446 VM_OBJECT_ASSERT_WLOCKED(object);
1447 KASSERT(object != NULL && m->object == object,
1448 ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1449 KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1450 ("vm_page_insert_radixdone: page %p is missing object ref", m));
1451 if (mpred != NULL) {
1452 KASSERT(mpred->object == object,
1453 ("vm_page_insert_radixdone: object doesn't contain mpred"));
1454 KASSERT(mpred->pindex < m->pindex,
1455 ("vm_page_insert_radixdone: mpred doesn't precede pindex"));
1459 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1461 TAILQ_INSERT_HEAD(&object->memq, m, listq);
1464 * Show that the object has one more resident page.
1466 object->resident_page_count++;
1469 * Hold the vnode until the last page is released.
1471 if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1472 vhold(object->handle);
1475 * Since we are inserting a new and possibly dirty page,
1476 * update the object's OBJ_MIGHTBEDIRTY flag.
1478 if (pmap_page_is_write_mapped(m))
1479 vm_object_set_writeable_dirty(object);
1483 * Do the work to remove a page from its object. The caller is responsible for
1484 * updating the page's fields to reflect this removal.
1487 vm_page_object_remove(vm_page_t m)
1493 VM_OBJECT_ASSERT_WLOCKED(object);
1494 KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1495 ("page %p is missing its object ref", m));
1496 if (vm_page_xbusied(m))
1498 mrem = vm_radix_remove(&object->rtree, m->pindex);
1499 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1502 * Now remove from the object's list of backed pages.
1504 TAILQ_REMOVE(&object->memq, m, listq);
1507 * And show that the object has one fewer resident page.
1509 object->resident_page_count--;
1512 * The vnode may now be recycled.
1514 if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1515 vdrop(object->handle);
1521 * Removes the specified page from its containing object, but does not
1522 * invalidate any backing storage. Returns true if the object's reference
1523 * was the last reference to the page, and false otherwise.
1525 * The object must be locked.
1528 vm_page_remove(vm_page_t m)
1531 vm_page_object_remove(m);
1533 return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1539 * Returns the page associated with the object/offset
1540 * pair specified; if none is found, NULL is returned.
1542 * The object must be locked.
1545 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1548 VM_OBJECT_ASSERT_LOCKED(object);
1549 return (vm_radix_lookup(&object->rtree, pindex));
1553 * vm_page_find_least:
1555 * Returns the page associated with the object with least pindex
1556 * greater than or equal to the parameter pindex, or NULL.
1558 * The object must be locked.
1561 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1565 VM_OBJECT_ASSERT_LOCKED(object);
1566 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1567 m = vm_radix_lookup_ge(&object->rtree, pindex);
1572 * Returns the given page's successor (by pindex) within the object if it is
1573 * resident; if none is found, NULL is returned.
1575 * The object must be locked.
1578 vm_page_next(vm_page_t m)
1582 VM_OBJECT_ASSERT_LOCKED(m->object);
1583 if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1584 MPASS(next->object == m->object);
1585 if (next->pindex != m->pindex + 1)
1592 * Returns the given page's predecessor (by pindex) within the object if it is
1593 * resident; if none is found, NULL is returned.
1595 * The object must be locked.
1598 vm_page_prev(vm_page_t m)
1602 VM_OBJECT_ASSERT_LOCKED(m->object);
1603 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1604 MPASS(prev->object == m->object);
1605 if (prev->pindex != m->pindex - 1)
1612 * Uses the page mnew as a replacement for an existing page at index
1613 * pindex which must be already present in the object.
1616 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1620 VM_OBJECT_ASSERT_WLOCKED(object);
1621 KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
1622 ("vm_page_replace: page %p already in object", mnew));
1625 * This function mostly follows vm_page_insert() and
1626 * vm_page_remove() without the radix, object count and vnode
1627 * dance. Double check such functions for more comments.
1630 mnew->object = object;
1631 mnew->pindex = pindex;
1632 atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
1633 mold = vm_radix_replace(&object->rtree, mnew);
1634 KASSERT(mold->queue == PQ_NONE,
1635 ("vm_page_replace: old page %p is on a paging queue", mold));
1637 /* Keep the resident page list in sorted order. */
1638 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1639 TAILQ_REMOVE(&object->memq, mold, listq);
1641 mold->object = NULL;
1642 atomic_clear_int(&mold->ref_count, VPRC_OBJREF);
1643 vm_page_xunbusy(mold);
1646 * The object's resident_page_count does not change because we have
1647 * swapped one page for another, but OBJ_MIGHTBEDIRTY.
1649 if (pmap_page_is_write_mapped(mnew))
1650 vm_object_set_writeable_dirty(object);
1657 * Move the given memory entry from its
1658 * current object to the specified target object/offset.
1660 * Note: swap associated with the page must be invalidated by the move. We
1661 * have to do this for several reasons: (1) we aren't freeing the
1662 * page, (2) we are dirtying the page, (3) the VM system is probably
1663 * moving the page from object A to B, and will then later move
1664 * the backing store from A to B and we can't have a conflict.
1666 * Note: we *always* dirty the page. It is necessary both for the
1667 * fact that we moved it, and because we may be invalidating
1670 * The objects must be locked.
1673 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1678 VM_OBJECT_ASSERT_WLOCKED(new_object);
1680 KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m));
1681 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1682 KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1683 ("vm_page_rename: pindex already renamed"));
1686 * Create a custom version of vm_page_insert() which does not depend
1687 * by m_prev and can cheat on the implementation aspects of the
1691 m->pindex = new_pindex;
1692 if (vm_radix_insert(&new_object->rtree, m)) {
1698 * The operation cannot fail anymore. The removal must happen before
1699 * the listq iterator is tainted.
1702 vm_page_object_remove(m);
1704 /* Return back to the new pindex to complete vm_page_insert(). */
1705 m->pindex = new_pindex;
1706 m->object = new_object;
1708 vm_page_insert_radixdone(m, new_object, mpred);
1716 * Allocate and return a page that is associated with the specified
1717 * object and offset pair. By default, this page is exclusive busied.
1719 * The caller must always specify an allocation class.
1721 * allocation classes:
1722 * VM_ALLOC_NORMAL normal process request
1723 * VM_ALLOC_SYSTEM system *really* needs a page
1724 * VM_ALLOC_INTERRUPT interrupt time request
1726 * optional allocation flags:
1727 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1728 * intends to allocate
1729 * VM_ALLOC_NOBUSY do not exclusive busy the page
1730 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1731 * VM_ALLOC_NOOBJ page is not associated with an object and
1732 * should not be exclusive busy
1733 * VM_ALLOC_SBUSY shared busy the allocated page
1734 * VM_ALLOC_WIRED wire the allocated page
1735 * VM_ALLOC_ZERO prefer a zeroed page
1738 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1741 return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1742 vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1746 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1750 return (vm_page_alloc_domain_after(object, pindex, domain, req,
1751 object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) :
1756 * Allocate a page in the specified object with the given page index. To
1757 * optimize insertion of the page into the object, the caller must also specifiy
1758 * the resident page in the object with largest index smaller than the given
1759 * page index, or NULL if no such page exists.
1762 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1763 int req, vm_page_t mpred)
1765 struct vm_domainset_iter di;
1769 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1771 m = vm_page_alloc_domain_after(object, pindex, domain, req,
1775 } while (vm_domainset_iter_page(&di, object, &domain) == 0);
1781 * Returns true if the number of free pages exceeds the minimum
1782 * for the request class and false otherwise.
1785 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
1787 u_int limit, old, new;
1789 req = req & VM_ALLOC_CLASS_MASK;
1792 * The page daemon is allowed to dig deeper into the free page list.
1794 if (curproc == pageproc && req != VM_ALLOC_INTERRUPT)
1795 req = VM_ALLOC_SYSTEM;
1796 if (req == VM_ALLOC_INTERRUPT)
1798 else if (req == VM_ALLOC_SYSTEM)
1799 limit = vmd->vmd_interrupt_free_min;
1801 limit = vmd->vmd_free_reserved;
1804 * Attempt to reserve the pages. Fail if we're below the limit.
1807 old = vmd->vmd_free_count;
1812 } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
1814 /* Wake the page daemon if we've crossed the threshold. */
1815 if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
1816 pagedaemon_wakeup(vmd->vmd_domain);
1818 /* Only update bitsets on transitions. */
1819 if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
1820 (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
1827 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
1828 int req, vm_page_t mpred)
1830 struct vm_domain *vmd;
1834 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1835 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1836 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1837 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1838 ("inconsistent object(%p)/req(%x)", object, req));
1839 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
1840 ("Can't sleep and retry object insertion."));
1841 KASSERT(mpred == NULL || mpred->pindex < pindex,
1842 ("mpred %p doesn't precede pindex 0x%jx", mpred,
1843 (uintmax_t)pindex));
1845 VM_OBJECT_ASSERT_WLOCKED(object);
1849 pool = object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT;
1851 #if VM_NRESERVLEVEL > 0
1853 * Can we allocate the page from a reservation?
1855 if (vm_object_reserv(object) &&
1856 (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
1858 domain = vm_phys_domain(m);
1859 vmd = VM_DOMAIN(domain);
1863 vmd = VM_DOMAIN(domain);
1864 if (vmd->vmd_pgcache[pool].zone != NULL) {
1865 m = uma_zalloc(vmd->vmd_pgcache[pool].zone, M_NOWAIT);
1867 flags |= PG_PCPU_CACHE;
1871 if (vm_domain_allocate(vmd, req, 1)) {
1873 * If not, allocate it from the free page queues.
1875 vm_domain_free_lock(vmd);
1876 m = vm_phys_alloc_pages(domain, pool, 0);
1877 vm_domain_free_unlock(vmd);
1879 vm_domain_freecnt_inc(vmd, 1);
1880 #if VM_NRESERVLEVEL > 0
1881 if (vm_reserv_reclaim_inactive(domain))
1888 * Not allocatable, give up.
1890 if (vm_domain_alloc_fail(vmd, object, req))
1896 * At this point we had better have found a good page.
1900 vm_page_alloc_check(m);
1903 * Initialize the page. Only the PG_ZERO flag is inherited.
1905 if ((req & VM_ALLOC_ZERO) != 0)
1906 flags |= (m->flags & PG_ZERO);
1907 if ((req & VM_ALLOC_NODUMP) != 0)
1911 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1913 m->busy_lock = VPB_UNBUSIED;
1914 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1915 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1916 if ((req & VM_ALLOC_SBUSY) != 0)
1917 m->busy_lock = VPB_SHARERS_WORD(1);
1918 if (req & VM_ALLOC_WIRED) {
1920 * The page lock is not required for wiring a page until that
1921 * page is inserted into the object.
1928 if (object != NULL) {
1929 if (vm_page_insert_after(m, object, pindex, mpred)) {
1930 if (req & VM_ALLOC_WIRED) {
1934 KASSERT(m->object == NULL, ("page %p has object", m));
1935 m->oflags = VPO_UNMANAGED;
1936 m->busy_lock = VPB_UNBUSIED;
1937 /* Don't change PG_ZERO. */
1938 vm_page_free_toq(m);
1939 if (req & VM_ALLOC_WAITFAIL) {
1940 VM_OBJECT_WUNLOCK(object);
1942 VM_OBJECT_WLOCK(object);
1947 /* Ignore device objects; the pager sets "memattr" for them. */
1948 if (object->memattr != VM_MEMATTR_DEFAULT &&
1949 (object->flags & OBJ_FICTITIOUS) == 0)
1950 pmap_page_set_memattr(m, object->memattr);
1958 * vm_page_alloc_contig:
1960 * Allocate a contiguous set of physical pages of the given size "npages"
1961 * from the free lists. All of the physical pages must be at or above
1962 * the given physical address "low" and below the given physical address
1963 * "high". The given value "alignment" determines the alignment of the
1964 * first physical page in the set. If the given value "boundary" is
1965 * non-zero, then the set of physical pages cannot cross any physical
1966 * address boundary that is a multiple of that value. Both "alignment"
1967 * and "boundary" must be a power of two.
1969 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1970 * then the memory attribute setting for the physical pages is configured
1971 * to the object's memory attribute setting. Otherwise, the memory
1972 * attribute setting for the physical pages is configured to "memattr",
1973 * overriding the object's memory attribute setting. However, if the
1974 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1975 * memory attribute setting for the physical pages cannot be configured
1976 * to VM_MEMATTR_DEFAULT.
1978 * The specified object may not contain fictitious pages.
1980 * The caller must always specify an allocation class.
1982 * allocation classes:
1983 * VM_ALLOC_NORMAL normal process request
1984 * VM_ALLOC_SYSTEM system *really* needs a page
1985 * VM_ALLOC_INTERRUPT interrupt time request
1987 * optional allocation flags:
1988 * VM_ALLOC_NOBUSY do not exclusive busy the page
1989 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1990 * VM_ALLOC_NOOBJ page is not associated with an object and
1991 * should not be exclusive busy
1992 * VM_ALLOC_SBUSY shared busy the allocated page
1993 * VM_ALLOC_WIRED wire the allocated page
1994 * VM_ALLOC_ZERO prefer a zeroed page
1997 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1998 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1999 vm_paddr_t boundary, vm_memattr_t memattr)
2001 struct vm_domainset_iter di;
2005 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2007 m = vm_page_alloc_contig_domain(object, pindex, domain, req,
2008 npages, low, high, alignment, boundary, memattr);
2011 } while (vm_domainset_iter_page(&di, object, &domain) == 0);
2017 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2018 int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2019 vm_paddr_t boundary, vm_memattr_t memattr)
2021 struct vm_domain *vmd;
2022 vm_page_t m, m_ret, mpred;
2023 u_int busy_lock, flags, oflags;
2025 mpred = NULL; /* XXX: pacify gcc */
2026 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
2027 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
2028 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2029 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2030 ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
2032 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2033 ("Can't sleep and retry object insertion."));
2034 if (object != NULL) {
2035 VM_OBJECT_ASSERT_WLOCKED(object);
2036 KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2037 ("vm_page_alloc_contig: object %p has fictitious pages",
2040 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2042 if (object != NULL) {
2043 mpred = vm_radix_lookup_le(&object->rtree, pindex);
2044 KASSERT(mpred == NULL || mpred->pindex != pindex,
2045 ("vm_page_alloc_contig: pindex already allocated"));
2049 * Can we allocate the pages without the number of free pages falling
2050 * below the lower bound for the allocation class?
2054 #if VM_NRESERVLEVEL > 0
2056 * Can we allocate the pages from a reservation?
2058 if (vm_object_reserv(object) &&
2059 (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
2060 mpred, npages, low, high, alignment, boundary)) != NULL) {
2061 domain = vm_phys_domain(m_ret);
2062 vmd = VM_DOMAIN(domain);
2066 vmd = VM_DOMAIN(domain);
2067 if (vm_domain_allocate(vmd, req, npages)) {
2069 * allocate them from the free page queues.
2071 vm_domain_free_lock(vmd);
2072 m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2073 alignment, boundary);
2074 vm_domain_free_unlock(vmd);
2075 if (m_ret == NULL) {
2076 vm_domain_freecnt_inc(vmd, npages);
2077 #if VM_NRESERVLEVEL > 0
2078 if (vm_reserv_reclaim_contig(domain, npages, low,
2079 high, alignment, boundary))
2084 if (m_ret == NULL) {
2085 if (vm_domain_alloc_fail(vmd, object, req))
2089 #if VM_NRESERVLEVEL > 0
2092 for (m = m_ret; m < &m_ret[npages]; m++) {
2094 vm_page_alloc_check(m);
2098 * Initialize the pages. Only the PG_ZERO flag is inherited.
2101 if ((req & VM_ALLOC_ZERO) != 0)
2103 if ((req & VM_ALLOC_NODUMP) != 0)
2105 oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2107 busy_lock = VPB_UNBUSIED;
2108 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2109 busy_lock = VPB_SINGLE_EXCLUSIVER;
2110 if ((req & VM_ALLOC_SBUSY) != 0)
2111 busy_lock = VPB_SHARERS_WORD(1);
2112 if ((req & VM_ALLOC_WIRED) != 0)
2113 vm_wire_add(npages);
2114 if (object != NULL) {
2115 if (object->memattr != VM_MEMATTR_DEFAULT &&
2116 memattr == VM_MEMATTR_DEFAULT)
2117 memattr = object->memattr;
2119 for (m = m_ret; m < &m_ret[npages]; m++) {
2121 m->flags = (m->flags | PG_NODUMP) & flags;
2122 m->busy_lock = busy_lock;
2123 if ((req & VM_ALLOC_WIRED) != 0)
2127 if (object != NULL) {
2128 if (vm_page_insert_after(m, object, pindex, mpred)) {
2129 if ((req & VM_ALLOC_WIRED) != 0)
2130 vm_wire_sub(npages);
2131 KASSERT(m->object == NULL,
2132 ("page %p has object", m));
2134 for (m = m_ret; m < &m_ret[npages]; m++) {
2136 (req & VM_ALLOC_WIRED) != 0)
2138 m->oflags = VPO_UNMANAGED;
2139 m->busy_lock = VPB_UNBUSIED;
2140 /* Don't change PG_ZERO. */
2141 vm_page_free_toq(m);
2143 if (req & VM_ALLOC_WAITFAIL) {
2144 VM_OBJECT_WUNLOCK(object);
2146 VM_OBJECT_WLOCK(object);
2153 if (memattr != VM_MEMATTR_DEFAULT)
2154 pmap_page_set_memattr(m, memattr);
2161 * Check a page that has been freshly dequeued from a freelist.
2164 vm_page_alloc_check(vm_page_t m)
2167 KASSERT(m->object == NULL, ("page %p has object", m));
2168 KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0,
2169 ("page %p has unexpected queue %d, flags %#x",
2170 m, m->queue, (m->aflags & PGA_QUEUE_STATE_MASK)));
2171 KASSERT(m->ref_count == 0, ("page %p has references", m));
2172 KASSERT(!vm_page_busied(m), ("page %p is busy", m));
2173 KASSERT(m->dirty == 0, ("page %p is dirty", m));
2174 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2175 ("page %p has unexpected memattr %d",
2176 m, pmap_page_get_memattr(m)));
2177 KASSERT(m->valid == 0, ("free page %p is valid", m));
2181 * vm_page_alloc_freelist:
2183 * Allocate a physical page from the specified free page list.
2185 * The caller must always specify an allocation class.
2187 * allocation classes:
2188 * VM_ALLOC_NORMAL normal process request
2189 * VM_ALLOC_SYSTEM system *really* needs a page
2190 * VM_ALLOC_INTERRUPT interrupt time request
2192 * optional allocation flags:
2193 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
2194 * intends to allocate
2195 * VM_ALLOC_WIRED wire the allocated page
2196 * VM_ALLOC_ZERO prefer a zeroed page
2199 vm_page_alloc_freelist(int freelist, int req)
2201 struct vm_domainset_iter di;
2205 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2207 m = vm_page_alloc_freelist_domain(domain, freelist, req);
2210 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2216 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2218 struct vm_domain *vmd;
2223 vmd = VM_DOMAIN(domain);
2225 if (vm_domain_allocate(vmd, req, 1)) {
2226 vm_domain_free_lock(vmd);
2227 m = vm_phys_alloc_freelist_pages(domain, freelist,
2228 VM_FREEPOOL_DIRECT, 0);
2229 vm_domain_free_unlock(vmd);
2231 vm_domain_freecnt_inc(vmd, 1);
2234 if (vm_domain_alloc_fail(vmd, NULL, req))
2239 vm_page_alloc_check(m);
2242 * Initialize the page. Only the PG_ZERO flag is inherited.
2246 if ((req & VM_ALLOC_ZERO) != 0)
2249 if ((req & VM_ALLOC_WIRED) != 0) {
2251 * The page lock is not required for wiring a page that does
2252 * not belong to an object.
2257 /* Unmanaged pages don't use "act_count". */
2258 m->oflags = VPO_UNMANAGED;
2263 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2265 struct vm_domain *vmd;
2266 struct vm_pgcache *pgcache;
2270 vmd = VM_DOMAIN(pgcache->domain);
2271 /* Only import if we can bring in a full bucket. */
2272 if (cnt == 1 || !vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2274 domain = vmd->vmd_domain;
2275 vm_domain_free_lock(vmd);
2276 i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2277 (vm_page_t *)store);
2278 vm_domain_free_unlock(vmd);
2280 vm_domain_freecnt_inc(vmd, cnt - i);
2286 vm_page_zone_release(void *arg, void **store, int cnt)
2288 struct vm_domain *vmd;
2289 struct vm_pgcache *pgcache;
2294 vmd = VM_DOMAIN(pgcache->domain);
2295 vm_domain_free_lock(vmd);
2296 for (i = 0; i < cnt; i++) {
2297 m = (vm_page_t)store[i];
2298 vm_phys_free_pages(m, 0);
2300 vm_domain_free_unlock(vmd);
2301 vm_domain_freecnt_inc(vmd, cnt);
2304 #define VPSC_ANY 0 /* No restrictions. */
2305 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */
2306 #define VPSC_NOSUPER 2 /* Skip superpages. */
2309 * vm_page_scan_contig:
2311 * Scan vm_page_array[] between the specified entries "m_start" and
2312 * "m_end" for a run of contiguous physical pages that satisfy the
2313 * specified conditions, and return the lowest page in the run. The
2314 * specified "alignment" determines the alignment of the lowest physical
2315 * page in the run. If the specified "boundary" is non-zero, then the
2316 * run of physical pages cannot span a physical address that is a
2317 * multiple of "boundary".
2319 * "m_end" is never dereferenced, so it need not point to a vm_page
2320 * structure within vm_page_array[].
2322 * "npages" must be greater than zero. "m_start" and "m_end" must not
2323 * span a hole (or discontiguity) in the physical address space. Both
2324 * "alignment" and "boundary" must be a power of two.
2327 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2328 u_long alignment, vm_paddr_t boundary, int options)
2334 #if VM_NRESERVLEVEL > 0
2337 int m_inc, order, run_ext, run_len;
2339 KASSERT(npages > 0, ("npages is 0"));
2340 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2341 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2345 for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2346 KASSERT((m->flags & PG_MARKER) == 0,
2347 ("page %p is PG_MARKER", m));
2348 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1,
2349 ("fictitious page %p has invalid ref count", m));
2352 * If the current page would be the start of a run, check its
2353 * physical address against the end, alignment, and boundary
2354 * conditions. If it doesn't satisfy these conditions, either
2355 * terminate the scan or advance to the next page that
2356 * satisfies the failed condition.
2359 KASSERT(m_run == NULL, ("m_run != NULL"));
2360 if (m + npages > m_end)
2362 pa = VM_PAGE_TO_PHYS(m);
2363 if ((pa & (alignment - 1)) != 0) {
2364 m_inc = atop(roundup2(pa, alignment) - pa);
2367 if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2369 m_inc = atop(roundup2(pa, boundary) - pa);
2373 KASSERT(m_run != NULL, ("m_run == NULL"));
2375 vm_page_change_lock(m, &m_mtx);
2378 if (vm_page_wired(m))
2380 #if VM_NRESERVLEVEL > 0
2381 else if ((level = vm_reserv_level(m)) >= 0 &&
2382 (options & VPSC_NORESERV) != 0) {
2384 /* Advance to the end of the reservation. */
2385 pa = VM_PAGE_TO_PHYS(m);
2386 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2390 else if ((object = m->object) != NULL) {
2392 * The page is considered eligible for relocation if
2393 * and only if it could be laundered or reclaimed by
2396 if (!VM_OBJECT_TRYRLOCK(object)) {
2398 VM_OBJECT_RLOCK(object);
2400 if (m->object != object) {
2402 * The page may have been freed.
2404 VM_OBJECT_RUNLOCK(object);
2408 /* Don't care: PG_NODUMP, PG_ZERO. */
2409 if (object->type != OBJT_DEFAULT &&
2410 object->type != OBJT_SWAP &&
2411 object->type != OBJT_VNODE) {
2413 #if VM_NRESERVLEVEL > 0
2414 } else if ((options & VPSC_NOSUPER) != 0 &&
2415 (level = vm_reserv_level_iffullpop(m)) >= 0) {
2417 /* Advance to the end of the superpage. */
2418 pa = VM_PAGE_TO_PHYS(m);
2419 m_inc = atop(roundup2(pa + 1,
2420 vm_reserv_size(level)) - pa);
2422 } else if (object->memattr == VM_MEMATTR_DEFAULT &&
2423 vm_page_queue(m) != PQ_NONE && !vm_page_busied(m) &&
2424 !vm_page_wired(m)) {
2426 * The page is allocated but eligible for
2427 * relocation. Extend the current run by one
2430 KASSERT(pmap_page_get_memattr(m) ==
2432 ("page %p has an unexpected memattr", m));
2433 KASSERT((m->oflags & (VPO_SWAPINPROG |
2434 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2435 ("page %p has unexpected oflags", m));
2436 /* Don't care: VPO_NOSYNC. */
2440 VM_OBJECT_RUNLOCK(object);
2441 #if VM_NRESERVLEVEL > 0
2442 } else if (level >= 0) {
2444 * The page is reserved but not yet allocated. In
2445 * other words, it is still free. Extend the current
2450 } else if ((order = m->order) < VM_NFREEORDER) {
2452 * The page is enqueued in the physical memory
2453 * allocator's free page queues. Moreover, it is the
2454 * first page in a power-of-two-sized run of
2455 * contiguous free pages. Add these pages to the end
2456 * of the current run, and jump ahead.
2458 run_ext = 1 << order;
2462 * Skip the page for one of the following reasons: (1)
2463 * It is enqueued in the physical memory allocator's
2464 * free page queues. However, it is not the first
2465 * page in a run of contiguous free pages. (This case
2466 * rarely occurs because the scan is performed in
2467 * ascending order.) (2) It is not reserved, and it is
2468 * transitioning from free to allocated. (Conversely,
2469 * the transition from allocated to free for managed
2470 * pages is blocked by the page lock.) (3) It is
2471 * allocated but not contained by an object and not
2472 * wired, e.g., allocated by Xen's balloon driver.
2478 * Extend or reset the current run of pages.
2493 if (run_len >= npages)
2499 * vm_page_reclaim_run:
2501 * Try to relocate each of the allocated virtual pages within the
2502 * specified run of physical pages to a new physical address. Free the
2503 * physical pages underlying the relocated virtual pages. A virtual page
2504 * is relocatable if and only if it could be laundered or reclaimed by
2505 * the page daemon. Whenever possible, a virtual page is relocated to a
2506 * physical address above "high".
2508 * Returns 0 if every physical page within the run was already free or
2509 * just freed by a successful relocation. Otherwise, returns a non-zero
2510 * value indicating why the last attempt to relocate a virtual page was
2513 * "req_class" must be an allocation class.
2516 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2519 struct vm_domain *vmd;
2521 struct spglist free;
2524 vm_page_t m, m_end, m_new;
2525 int error, order, req;
2527 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2528 ("req_class is not an allocation class"));
2532 m_end = m_run + npages;
2534 for (; error == 0 && m < m_end; m++) {
2535 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2536 ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2539 * Avoid releasing and reacquiring the same page lock.
2541 vm_page_change_lock(m, &m_mtx);
2544 * Racily check for wirings. Races are handled below.
2546 if (vm_page_wired(m))
2548 else if ((object = m->object) != NULL) {
2550 * The page is relocated if and only if it could be
2551 * laundered or reclaimed by the page daemon.
2553 if (!VM_OBJECT_TRYWLOCK(object)) {
2555 VM_OBJECT_WLOCK(object);
2557 if (m->object != object) {
2559 * The page may have been freed.
2561 VM_OBJECT_WUNLOCK(object);
2565 /* Don't care: PG_NODUMP, PG_ZERO. */
2566 if (object->type != OBJT_DEFAULT &&
2567 object->type != OBJT_SWAP &&
2568 object->type != OBJT_VNODE)
2570 else if (object->memattr != VM_MEMATTR_DEFAULT)
2572 else if (vm_page_queue(m) != PQ_NONE &&
2573 !vm_page_busied(m) && !vm_page_wired(m)) {
2574 KASSERT(pmap_page_get_memattr(m) ==
2576 ("page %p has an unexpected memattr", m));
2577 KASSERT((m->oflags & (VPO_SWAPINPROG |
2578 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2579 ("page %p has unexpected oflags", m));
2580 /* Don't care: VPO_NOSYNC. */
2581 if (m->valid != 0) {
2583 * First, try to allocate a new page
2584 * that is above "high". Failing
2585 * that, try to allocate a new page
2586 * that is below "m_run". Allocate
2587 * the new page between the end of
2588 * "m_run" and "high" only as a last
2591 req = req_class | VM_ALLOC_NOOBJ;
2592 if ((m->flags & PG_NODUMP) != 0)
2593 req |= VM_ALLOC_NODUMP;
2594 if (trunc_page(high) !=
2595 ~(vm_paddr_t)PAGE_MASK) {
2596 m_new = vm_page_alloc_contig(
2601 VM_MEMATTR_DEFAULT);
2604 if (m_new == NULL) {
2605 pa = VM_PAGE_TO_PHYS(m_run);
2606 m_new = vm_page_alloc_contig(
2608 0, pa - 1, PAGE_SIZE, 0,
2609 VM_MEMATTR_DEFAULT);
2611 if (m_new == NULL) {
2613 m_new = vm_page_alloc_contig(
2615 pa, high, PAGE_SIZE, 0,
2616 VM_MEMATTR_DEFAULT);
2618 if (m_new == NULL) {
2624 * Unmap the page and check for new
2625 * wirings that may have been acquired
2626 * through a pmap lookup.
2628 if (object->ref_count != 0 &&
2629 !vm_page_try_remove_all(m)) {
2630 vm_page_free(m_new);
2636 * Replace "m" with the new page. For
2637 * vm_page_replace(), "m" must be busy
2638 * and dequeued. Finally, change "m"
2639 * as if vm_page_free() was called.
2641 m_new->aflags = m->aflags &
2642 ~PGA_QUEUE_STATE_MASK;
2643 KASSERT(m_new->oflags == VPO_UNMANAGED,
2644 ("page %p is managed", m_new));
2645 m_new->oflags = m->oflags & VPO_NOSYNC;
2646 pmap_copy_page(m, m_new);
2647 m_new->valid = m->valid;
2648 m_new->dirty = m->dirty;
2649 m->flags &= ~PG_ZERO;
2652 vm_page_replace_checked(m_new, object,
2654 if (vm_page_free_prep(m))
2655 SLIST_INSERT_HEAD(&free, m,
2659 * The new page must be deactivated
2660 * before the object is unlocked.
2662 vm_page_change_lock(m_new, &m_mtx);
2663 vm_page_deactivate(m_new);
2665 m->flags &= ~PG_ZERO;
2667 if (vm_page_free_prep(m))
2668 SLIST_INSERT_HEAD(&free, m,
2670 KASSERT(m->dirty == 0,
2671 ("page %p is dirty", m));
2676 VM_OBJECT_WUNLOCK(object);
2678 MPASS(vm_phys_domain(m) == domain);
2679 vmd = VM_DOMAIN(domain);
2680 vm_domain_free_lock(vmd);
2682 if (order < VM_NFREEORDER) {
2684 * The page is enqueued in the physical memory
2685 * allocator's free page queues. Moreover, it
2686 * is the first page in a power-of-two-sized
2687 * run of contiguous free pages. Jump ahead
2688 * to the last page within that run, and
2689 * continue from there.
2691 m += (1 << order) - 1;
2693 #if VM_NRESERVLEVEL > 0
2694 else if (vm_reserv_is_page_free(m))
2697 vm_domain_free_unlock(vmd);
2698 if (order == VM_NFREEORDER)
2704 if ((m = SLIST_FIRST(&free)) != NULL) {
2707 vmd = VM_DOMAIN(domain);
2709 vm_domain_free_lock(vmd);
2711 MPASS(vm_phys_domain(m) == domain);
2712 SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2713 vm_phys_free_pages(m, 0);
2715 } while ((m = SLIST_FIRST(&free)) != NULL);
2716 vm_domain_free_unlock(vmd);
2717 vm_domain_freecnt_inc(vmd, cnt);
2724 CTASSERT(powerof2(NRUNS));
2726 #define RUN_INDEX(count) ((count) & (NRUNS - 1))
2728 #define MIN_RECLAIM 8
2731 * vm_page_reclaim_contig:
2733 * Reclaim allocated, contiguous physical memory satisfying the specified
2734 * conditions by relocating the virtual pages using that physical memory.
2735 * Returns true if reclamation is successful and false otherwise. Since
2736 * relocation requires the allocation of physical pages, reclamation may
2737 * fail due to a shortage of free pages. When reclamation fails, callers
2738 * are expected to perform vm_wait() before retrying a failed allocation
2739 * operation, e.g., vm_page_alloc_contig().
2741 * The caller must always specify an allocation class through "req".
2743 * allocation classes:
2744 * VM_ALLOC_NORMAL normal process request
2745 * VM_ALLOC_SYSTEM system *really* needs a page
2746 * VM_ALLOC_INTERRUPT interrupt time request
2748 * The optional allocation flags are ignored.
2750 * "npages" must be greater than zero. Both "alignment" and "boundary"
2751 * must be a power of two.
2754 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
2755 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2757 struct vm_domain *vmd;
2758 vm_paddr_t curr_low;
2759 vm_page_t m_run, m_runs[NRUNS];
2760 u_long count, reclaimed;
2761 int error, i, options, req_class;
2763 KASSERT(npages > 0, ("npages is 0"));
2764 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2765 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2766 req_class = req & VM_ALLOC_CLASS_MASK;
2769 * The page daemon is allowed to dig deeper into the free page list.
2771 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2772 req_class = VM_ALLOC_SYSTEM;
2775 * Return if the number of free pages cannot satisfy the requested
2778 vmd = VM_DOMAIN(domain);
2779 count = vmd->vmd_free_count;
2780 if (count < npages + vmd->vmd_free_reserved || (count < npages +
2781 vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2782 (count < npages && req_class == VM_ALLOC_INTERRUPT))
2786 * Scan up to three times, relaxing the restrictions ("options") on
2787 * the reclamation of reservations and superpages each time.
2789 for (options = VPSC_NORESERV;;) {
2791 * Find the highest runs that satisfy the given constraints
2792 * and restrictions, and record them in "m_runs".
2797 m_run = vm_phys_scan_contig(domain, npages, curr_low,
2798 high, alignment, boundary, options);
2801 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2802 m_runs[RUN_INDEX(count)] = m_run;
2807 * Reclaim the highest runs in LIFO (descending) order until
2808 * the number of reclaimed pages, "reclaimed", is at least
2809 * MIN_RECLAIM. Reset "reclaimed" each time because each
2810 * reclamation is idempotent, and runs will (likely) recur
2811 * from one scan to the next as restrictions are relaxed.
2814 for (i = 0; count > 0 && i < NRUNS; i++) {
2816 m_run = m_runs[RUN_INDEX(count)];
2817 error = vm_page_reclaim_run(req_class, domain, npages,
2820 reclaimed += npages;
2821 if (reclaimed >= MIN_RECLAIM)
2827 * Either relax the restrictions on the next scan or return if
2828 * the last scan had no restrictions.
2830 if (options == VPSC_NORESERV)
2831 options = VPSC_NOSUPER;
2832 else if (options == VPSC_NOSUPER)
2834 else if (options == VPSC_ANY)
2835 return (reclaimed != 0);
2840 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
2841 u_long alignment, vm_paddr_t boundary)
2843 struct vm_domainset_iter di;
2847 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2849 ret = vm_page_reclaim_contig_domain(domain, req, npages, low,
2850 high, alignment, boundary);
2853 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2859 * Set the domain in the appropriate page level domainset.
2862 vm_domain_set(struct vm_domain *vmd)
2865 mtx_lock(&vm_domainset_lock);
2866 if (!vmd->vmd_minset && vm_paging_min(vmd)) {
2867 vmd->vmd_minset = 1;
2868 DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
2870 if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
2871 vmd->vmd_severeset = 1;
2872 DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
2874 mtx_unlock(&vm_domainset_lock);
2878 * Clear the domain from the appropriate page level domainset.
2881 vm_domain_clear(struct vm_domain *vmd)
2884 mtx_lock(&vm_domainset_lock);
2885 if (vmd->vmd_minset && !vm_paging_min(vmd)) {
2886 vmd->vmd_minset = 0;
2887 DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
2888 if (vm_min_waiters != 0) {
2890 wakeup(&vm_min_domains);
2893 if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
2894 vmd->vmd_severeset = 0;
2895 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
2896 if (vm_severe_waiters != 0) {
2897 vm_severe_waiters = 0;
2898 wakeup(&vm_severe_domains);
2903 * If pageout daemon needs pages, then tell it that there are
2906 if (vmd->vmd_pageout_pages_needed &&
2907 vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
2908 wakeup(&vmd->vmd_pageout_pages_needed);
2909 vmd->vmd_pageout_pages_needed = 0;
2912 /* See comments in vm_wait_doms(). */
2913 if (vm_pageproc_waiters) {
2914 vm_pageproc_waiters = 0;
2915 wakeup(&vm_pageproc_waiters);
2917 mtx_unlock(&vm_domainset_lock);
2921 * Wait for free pages to exceed the min threshold globally.
2927 mtx_lock(&vm_domainset_lock);
2928 while (vm_page_count_min()) {
2930 msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
2932 mtx_unlock(&vm_domainset_lock);
2936 * Wait for free pages to exceed the severe threshold globally.
2939 vm_wait_severe(void)
2942 mtx_lock(&vm_domainset_lock);
2943 while (vm_page_count_severe()) {
2944 vm_severe_waiters++;
2945 msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
2948 mtx_unlock(&vm_domainset_lock);
2955 return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
2959 vm_wait_doms(const domainset_t *wdoms)
2963 * We use racey wakeup synchronization to avoid expensive global
2964 * locking for the pageproc when sleeping with a non-specific vm_wait.
2965 * To handle this, we only sleep for one tick in this instance. It
2966 * is expected that most allocations for the pageproc will come from
2967 * kmem or vm_page_grab* which will use the more specific and
2968 * race-free vm_wait_domain().
2970 if (curproc == pageproc) {
2971 mtx_lock(&vm_domainset_lock);
2972 vm_pageproc_waiters++;
2973 msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP,
2977 * XXX Ideally we would wait only until the allocation could
2978 * be satisfied. This condition can cause new allocators to
2979 * consume all freed pages while old allocators wait.
2981 mtx_lock(&vm_domainset_lock);
2982 if (vm_page_count_min_set(wdoms)) {
2984 msleep(&vm_min_domains, &vm_domainset_lock,
2985 PVM | PDROP, "vmwait", 0);
2987 mtx_unlock(&vm_domainset_lock);
2994 * Sleep until free pages are available for allocation.
2995 * - Called in various places after failed memory allocations.
2998 vm_wait_domain(int domain)
3000 struct vm_domain *vmd;
3003 vmd = VM_DOMAIN(domain);
3004 vm_domain_free_assert_unlocked(vmd);
3006 if (curproc == pageproc) {
3007 mtx_lock(&vm_domainset_lock);
3008 if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3009 vmd->vmd_pageout_pages_needed = 1;
3010 msleep(&vmd->vmd_pageout_pages_needed,
3011 &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3013 mtx_unlock(&vm_domainset_lock);
3015 if (pageproc == NULL)
3016 panic("vm_wait in early boot");
3017 DOMAINSET_ZERO(&wdom);
3018 DOMAINSET_SET(vmd->vmd_domain, &wdom);
3019 vm_wait_doms(&wdom);
3026 * Sleep until free pages are available for allocation in the
3027 * affinity domains of the obj. If obj is NULL, the domain set
3028 * for the calling thread is used.
3029 * Called in various places after failed memory allocations.
3032 vm_wait(vm_object_t obj)
3034 struct domainset *d;
3039 * Carefully fetch pointers only once: the struct domainset
3040 * itself is ummutable but the pointer might change.
3043 d = obj->domain.dr_policy;
3045 d = curthread->td_domain.dr_policy;
3047 vm_wait_doms(&d->ds_mask);
3051 * vm_domain_alloc_fail:
3053 * Called when a page allocation function fails. Informs the
3054 * pagedaemon and performs the requested wait. Requires the
3055 * domain_free and object lock on entry. Returns with the
3056 * object lock held and free lock released. Returns an error when
3057 * retry is necessary.
3061 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3064 vm_domain_free_assert_unlocked(vmd);
3066 atomic_add_int(&vmd->vmd_pageout_deficit,
3067 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3068 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3070 VM_OBJECT_WUNLOCK(object);
3071 vm_wait_domain(vmd->vmd_domain);
3073 VM_OBJECT_WLOCK(object);
3074 if (req & VM_ALLOC_WAITOK)
3084 * Sleep until free pages are available for allocation.
3085 * - Called only in vm_fault so that processes page faulting
3086 * can be easily tracked.
3087 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3088 * processes will be able to grab memory first. Do not change
3089 * this balance without careful testing first.
3092 vm_waitpfault(struct domainset *dset, int timo)
3096 * XXX Ideally we would wait only until the allocation could
3097 * be satisfied. This condition can cause new allocators to
3098 * consume all freed pages while old allocators wait.
3100 mtx_lock(&vm_domainset_lock);
3101 if (vm_page_count_min_set(&dset->ds_mask)) {
3103 msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3106 mtx_unlock(&vm_domainset_lock);
3109 static struct vm_pagequeue *
3110 vm_page_pagequeue(vm_page_t m)
3115 if ((queue = atomic_load_8(&m->queue)) == PQ_NONE)
3117 return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3121 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m)
3123 struct vm_domain *vmd;
3126 CRITICAL_ASSERT(curthread);
3127 vm_pagequeue_assert_locked(pq);
3130 * The page daemon is allowed to set m->queue = PQ_NONE without
3131 * the page queue lock held. In this case it is about to free the page,
3132 * which must not have any queue state.
3134 qflags = atomic_load_8(&m->aflags);
3135 KASSERT(pq == vm_page_pagequeue(m) ||
3136 (qflags & PGA_QUEUE_STATE_MASK) == 0,
3137 ("page %p doesn't belong to queue %p but has aflags %#x",
3140 if ((qflags & PGA_DEQUEUE) != 0) {
3141 if (__predict_true((qflags & PGA_ENQUEUED) != 0))
3142 vm_pagequeue_remove(pq, m);
3143 vm_page_dequeue_complete(m);
3144 counter_u64_add(queue_ops, 1);
3145 } else if ((qflags & (PGA_REQUEUE | PGA_REQUEUE_HEAD)) != 0) {
3146 if ((qflags & PGA_ENQUEUED) != 0)
3147 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3149 vm_pagequeue_cnt_inc(pq);
3150 vm_page_aflag_set(m, PGA_ENQUEUED);
3154 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE.
3155 * In particular, if both flags are set in close succession,
3156 * only PGA_REQUEUE_HEAD will be applied, even if it was set
3159 if ((qflags & PGA_REQUEUE_HEAD) != 0) {
3160 KASSERT(m->queue == PQ_INACTIVE,
3161 ("head enqueue not supported for page %p", m));
3162 vmd = vm_pagequeue_domain(m);
3163 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3165 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3167 vm_page_aflag_clear(m, qflags & (PGA_REQUEUE |
3169 counter_u64_add(queue_ops, 1);
3171 counter_u64_add(queue_nops, 1);
3176 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3182 for (i = 0; i < bq->bq_cnt; i++) {
3184 if (__predict_false(m->queue != queue))
3186 vm_pqbatch_process_page(pq, m);
3188 vm_batchqueue_init(bq);
3192 * vm_page_pqbatch_submit: [ internal use only ]
3194 * Enqueue a page in the specified page queue's batched work queue.
3195 * The caller must have encoded the requested operation in the page
3196 * structure's aflags field.
3199 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
3201 struct vm_batchqueue *bq;
3202 struct vm_pagequeue *pq;
3205 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3206 ("page %p is unmanaged", m));
3207 KASSERT(mtx_owned(vm_page_lockptr(m)) || m->object == NULL,
3208 ("missing synchronization for page %p", m));
3209 KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3211 domain = vm_phys_domain(m);
3212 pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue];
3215 bq = DPCPU_PTR(pqbatch[domain][queue]);
3216 if (vm_batchqueue_insert(bq, m)) {
3221 vm_pagequeue_lock(pq);
3223 bq = DPCPU_PTR(pqbatch[domain][queue]);
3224 vm_pqbatch_process(pq, bq, queue);
3227 * The page may have been logically dequeued before we acquired the
3228 * page queue lock. In this case, since we either hold the page lock
3229 * or the page is being freed, a different thread cannot be concurrently
3230 * enqueuing the page.
3232 if (__predict_true(m->queue == queue))
3233 vm_pqbatch_process_page(pq, m);
3235 KASSERT(m->queue == PQ_NONE,
3236 ("invalid queue transition for page %p", m));
3237 KASSERT((m->aflags & PGA_ENQUEUED) == 0,
3238 ("page %p is enqueued with invalid queue index", m));
3240 vm_pagequeue_unlock(pq);
3245 * vm_page_pqbatch_drain: [ internal use only ]
3247 * Force all per-CPU page queue batch queues to be drained. This is
3248 * intended for use in severe memory shortages, to ensure that pages
3249 * do not remain stuck in the batch queues.
3252 vm_page_pqbatch_drain(void)
3255 struct vm_domain *vmd;
3256 struct vm_pagequeue *pq;
3257 int cpu, domain, queue;
3262 sched_bind(td, cpu);
3265 for (domain = 0; domain < vm_ndomains; domain++) {
3266 vmd = VM_DOMAIN(domain);
3267 for (queue = 0; queue < PQ_COUNT; queue++) {
3268 pq = &vmd->vmd_pagequeues[queue];
3269 vm_pagequeue_lock(pq);
3271 vm_pqbatch_process(pq,
3272 DPCPU_PTR(pqbatch[domain][queue]), queue);
3274 vm_pagequeue_unlock(pq);
3284 * Complete the logical removal of a page from a page queue. We must be
3285 * careful to synchronize with the page daemon, which may be concurrently
3286 * examining the page with only the page lock held. The page must not be
3287 * in a state where it appears to be logically enqueued.
3290 vm_page_dequeue_complete(vm_page_t m)
3294 atomic_thread_fence_rel();
3295 vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK);
3299 * vm_page_dequeue_deferred: [ internal use only ]
3301 * Request removal of the given page from its current page
3302 * queue. Physical removal from the queue may be deferred
3305 * The page must be locked.
3308 vm_page_dequeue_deferred(vm_page_t m)
3312 vm_page_assert_locked(m);
3314 if ((queue = vm_page_queue(m)) == PQ_NONE)
3318 * Set PGA_DEQUEUE if it is not already set to handle a concurrent call
3319 * to vm_page_dequeue_deferred_free(). In particular, avoid modifying
3320 * the page's queue state once vm_page_dequeue_deferred_free() has been
3321 * called. In the event of a race, two batch queue entries for the page
3322 * will be created, but the second will have no effect.
3324 if (vm_page_pqstate_cmpset(m, queue, queue, PGA_DEQUEUE, PGA_DEQUEUE))
3325 vm_page_pqbatch_submit(m, queue);
3329 * A variant of vm_page_dequeue_deferred() that does not assert the page
3330 * lock and is only to be called from vm_page_free_prep(). Because the
3331 * page is being freed, we can assume that nothing other than the page
3332 * daemon is scheduling queue operations on this page, so we get for
3333 * free the mutual exclusion that is otherwise provided by the page lock.
3334 * To handle races, the page daemon must take care to atomically check
3335 * for PGA_DEQUEUE when updating queue state.
3338 vm_page_dequeue_deferred_free(vm_page_t m)
3342 KASSERT(m->ref_count == 0, ("page %p has references", m));
3345 if ((m->aflags & PGA_DEQUEUE) != 0)
3347 atomic_thread_fence_acq();
3348 if ((queue = atomic_load_8(&m->queue)) == PQ_NONE)
3350 if (vm_page_pqstate_cmpset(m, queue, queue, PGA_DEQUEUE,
3352 vm_page_pqbatch_submit(m, queue);
3361 * Remove the page from whichever page queue it's in, if any.
3362 * The page must either be locked or unallocated. This constraint
3363 * ensures that the queue state of the page will remain consistent
3364 * after this function returns.
3367 vm_page_dequeue(vm_page_t m)
3369 struct vm_pagequeue *pq, *pq1;
3372 KASSERT(mtx_owned(vm_page_lockptr(m)) || m->object == NULL,
3373 ("page %p is allocated and unlocked", m));
3375 for (pq = vm_page_pagequeue(m);; pq = pq1) {
3378 * A thread may be concurrently executing
3379 * vm_page_dequeue_complete(). Ensure that all queue
3380 * state is cleared before we return.
3382 aflags = atomic_load_8(&m->aflags);
3383 if ((aflags & PGA_QUEUE_STATE_MASK) == 0)
3385 KASSERT((aflags & PGA_DEQUEUE) != 0,
3386 ("page %p has unexpected queue state flags %#x",
3390 * Busy wait until the thread updating queue state is
3391 * finished. Such a thread must be executing in a
3395 pq1 = vm_page_pagequeue(m);
3398 vm_pagequeue_lock(pq);
3399 if ((pq1 = vm_page_pagequeue(m)) == pq)
3401 vm_pagequeue_unlock(pq);
3403 KASSERT(pq == vm_page_pagequeue(m),
3404 ("%s: page %p migrated directly between queues", __func__, m));
3405 KASSERT((m->aflags & PGA_DEQUEUE) != 0 ||
3406 mtx_owned(vm_page_lockptr(m)),
3407 ("%s: queued unlocked page %p", __func__, m));
3409 if ((m->aflags & PGA_ENQUEUED) != 0)
3410 vm_pagequeue_remove(pq, m);
3411 vm_page_dequeue_complete(m);
3412 vm_pagequeue_unlock(pq);
3416 * Schedule the given page for insertion into the specified page queue.
3417 * Physical insertion of the page may be deferred indefinitely.
3420 vm_page_enqueue(vm_page_t m, uint8_t queue)
3423 vm_page_assert_locked(m);
3424 KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0,
3425 ("%s: page %p is already enqueued", __func__, m));
3428 if ((m->aflags & PGA_REQUEUE) == 0)
3429 vm_page_aflag_set(m, PGA_REQUEUE);
3430 vm_page_pqbatch_submit(m, queue);
3434 * vm_page_requeue: [ internal use only ]
3436 * Schedule a requeue of the given page.
3438 * The page must be locked.
3441 vm_page_requeue(vm_page_t m)
3444 vm_page_assert_locked(m);
3445 KASSERT(vm_page_queue(m) != PQ_NONE,
3446 ("%s: page %p is not logically enqueued", __func__, m));
3448 if ((m->aflags & PGA_REQUEUE) == 0)
3449 vm_page_aflag_set(m, PGA_REQUEUE);
3450 vm_page_pqbatch_submit(m, atomic_load_8(&m->queue));
3454 * vm_page_swapqueue: [ internal use only ]
3456 * Move the page from one queue to another, or to the tail of its
3457 * current queue, in the face of a possible concurrent call to
3458 * vm_page_dequeue_deferred_free().
3461 vm_page_swapqueue(vm_page_t m, uint8_t oldq, uint8_t newq)
3463 struct vm_pagequeue *pq;
3467 KASSERT(oldq < PQ_COUNT && newq < PQ_COUNT && oldq != newq,
3468 ("vm_page_swapqueue: invalid queues (%d, %d)", oldq, newq));
3469 vm_page_assert_locked(m);
3471 pq = &vm_pagequeue_domain(m)->vmd_pagequeues[oldq];
3472 vm_pagequeue_lock(pq);
3475 * The physical queue state might change at any point before the page
3476 * queue lock is acquired, so we must verify that we hold the correct
3477 * lock before proceeding.
3479 if (__predict_false(m->queue != oldq)) {
3480 vm_pagequeue_unlock(pq);
3485 * Once the queue index of the page changes, there is nothing
3486 * synchronizing with further updates to the physical queue state.
3487 * Therefore we must remove the page from the queue now in anticipation
3488 * of a successful commit, and be prepared to roll back.
3490 if (__predict_true((m->aflags & PGA_ENQUEUED) != 0)) {
3491 next = TAILQ_NEXT(m, plinks.q);
3492 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3493 vm_page_aflag_clear(m, PGA_ENQUEUED);
3500 * Atomically update the queue field and set PGA_REQUEUE while
3501 * ensuring that PGA_DEQUEUE has not been set.
3503 if (__predict_false(!vm_page_pqstate_cmpset(m, oldq, newq, PGA_DEQUEUE,
3506 vm_page_aflag_set(m, PGA_ENQUEUED);
3508 TAILQ_INSERT_BEFORE(next, m, plinks.q);
3510 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3512 vm_pagequeue_unlock(pq);
3515 vm_pagequeue_cnt_dec(pq);
3516 vm_pagequeue_unlock(pq);
3517 vm_page_pqbatch_submit(m, newq);
3521 * vm_page_free_prep:
3523 * Prepares the given page to be put on the free list,
3524 * disassociating it from any VM object. The caller may return
3525 * the page to the free list only if this function returns true.
3527 * The object must be locked. The page must be locked if it is
3531 vm_page_free_prep(vm_page_t m)
3535 * Synchronize with threads that have dropped a reference to this
3538 atomic_thread_fence_acq();
3540 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3541 if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3544 p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3545 for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3546 KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3547 m, i, (uintmax_t)*p));
3550 if ((m->oflags & VPO_UNMANAGED) == 0) {
3551 KASSERT(!pmap_page_is_mapped(m),
3552 ("vm_page_free_prep: freeing mapped page %p", m));
3553 KASSERT((m->aflags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
3554 ("vm_page_free_prep: mapping flags set in page %p", m));
3556 KASSERT(m->queue == PQ_NONE,
3557 ("vm_page_free_prep: unmanaged page %p is queued", m));
3559 VM_CNT_INC(v_tfree);
3561 if (vm_page_sbusied(m))
3562 panic("vm_page_free_prep: freeing busy page %p", m);
3564 if (m->object != NULL) {
3565 vm_page_object_remove(m);
3568 * The object reference can be released without an atomic
3571 KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
3572 m->ref_count == VPRC_OBJREF,
3573 ("vm_page_free_prep: page %p has unexpected ref_count %u",
3576 m->ref_count -= VPRC_OBJREF;
3580 * If fictitious remove object association and
3583 if ((m->flags & PG_FICTITIOUS) != 0) {
3584 KASSERT(m->ref_count == 1,
3585 ("fictitious page %p is referenced", m));
3586 KASSERT(m->queue == PQ_NONE,
3587 ("fictitious page %p is queued", m));
3592 * Pages need not be dequeued before they are returned to the physical
3593 * memory allocator, but they must at least be marked for a deferred
3596 if ((m->oflags & VPO_UNMANAGED) == 0)
3597 vm_page_dequeue_deferred_free(m);
3602 if (m->ref_count != 0)
3603 panic("vm_page_free_prep: page %p has references", m);
3606 * Restore the default memory attribute to the page.
3608 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3609 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3611 #if VM_NRESERVLEVEL > 0
3613 * Determine whether the page belongs to a reservation. If the page was
3614 * allocated from a per-CPU cache, it cannot belong to a reservation, so
3615 * as an optimization, we avoid the check in that case.
3617 if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
3627 * Returns the given page to the free list, disassociating it
3628 * from any VM object.
3630 * The object must be locked. The page must be locked if it is
3634 vm_page_free_toq(vm_page_t m)
3636 struct vm_domain *vmd;
3639 if (!vm_page_free_prep(m))
3642 vmd = vm_pagequeue_domain(m);
3643 zone = vmd->vmd_pgcache[m->pool].zone;
3644 if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
3648 vm_domain_free_lock(vmd);
3649 vm_phys_free_pages(m, 0);
3650 vm_domain_free_unlock(vmd);
3651 vm_domain_freecnt_inc(vmd, 1);
3655 * vm_page_free_pages_toq:
3657 * Returns a list of pages to the free list, disassociating it
3658 * from any VM object. In other words, this is equivalent to
3659 * calling vm_page_free_toq() for each page of a list of VM objects.
3661 * The objects must be locked. The pages must be locked if it is
3665 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
3670 if (SLIST_EMPTY(free))
3674 while ((m = SLIST_FIRST(free)) != NULL) {
3676 SLIST_REMOVE_HEAD(free, plinks.s.ss);
3677 vm_page_free_toq(m);
3680 if (update_wire_count)
3685 * Mark this page as wired down, preventing reclamation by the page daemon
3686 * or when the containing object is destroyed.
3689 vm_page_wire(vm_page_t m)
3693 KASSERT(m->object != NULL,
3694 ("vm_page_wire: page %p does not belong to an object", m));
3695 if (!vm_page_busied(m))
3696 VM_OBJECT_ASSERT_LOCKED(m->object);
3697 KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
3698 VPRC_WIRE_COUNT(m->ref_count) >= 1,
3699 ("vm_page_wire: fictitious page %p has zero wirings", m));
3701 old = atomic_fetchadd_int(&m->ref_count, 1);
3702 KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
3703 ("vm_page_wire: counter overflow for page %p", m));
3704 if (VPRC_WIRE_COUNT(old) == 0)
3709 * Attempt to wire a mapped page following a pmap lookup of that page.
3710 * This may fail if a thread is concurrently tearing down mappings of the page.
3713 vm_page_wire_mapped(vm_page_t m)
3720 ("vm_page_wire_mapped: wiring unreferenced page %p", m));
3721 if ((old & VPRC_BLOCKED) != 0)
3723 } while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
3725 if (VPRC_WIRE_COUNT(old) == 0)
3731 * Release one wiring of the specified page, potentially allowing it to be
3734 * Only managed pages belonging to an object can be paged out. If the number
3735 * of wirings transitions to zero and the page is eligible for page out, then
3736 * the page is added to the specified paging queue. If the released wiring
3737 * represented the last reference to the page, the page is freed.
3739 * A managed page must be locked.
3742 vm_page_unwire(vm_page_t m, uint8_t queue)
3747 KASSERT(queue < PQ_COUNT,
3748 ("vm_page_unwire: invalid queue %u request for page %p", queue, m));
3750 if ((m->oflags & VPO_UNMANAGED) != 0) {
3751 if (vm_page_unwire_noq(m) && m->ref_count == 0)
3757 * Update LRU state before releasing the wiring reference.
3758 * We only need to do this once since we hold the page lock.
3759 * Use a release store when updating the reference count to
3760 * synchronize with vm_page_free_prep().
3765 KASSERT(VPRC_WIRE_COUNT(old) > 0,
3766 ("vm_page_unwire: wire count underflow for page %p", m));
3767 if (!locked && VPRC_WIRE_COUNT(old) == 1) {
3770 if (queue == PQ_ACTIVE && vm_page_queue(m) == PQ_ACTIVE)
3771 vm_page_reference(m);
3773 vm_page_mvqueue(m, queue);
3775 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
3778 * Release the lock only after the wiring is released, to ensure that
3779 * the page daemon does not encounter and dequeue the page while it is
3785 if (VPRC_WIRE_COUNT(old) == 1) {
3793 * Unwire a page without (re-)inserting it into a page queue. It is up
3794 * to the caller to enqueue, requeue, or free the page as appropriate.
3795 * In most cases involving managed pages, vm_page_unwire() should be used
3799 vm_page_unwire_noq(vm_page_t m)
3803 old = vm_page_drop(m, 1);
3804 KASSERT(VPRC_WIRE_COUNT(old) != 0,
3805 ("vm_page_unref: counter underflow for page %p", m));
3806 KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
3807 ("vm_page_unref: missing ref on fictitious page %p", m));
3809 if (VPRC_WIRE_COUNT(old) > 1)
3816 * Ensure that the page is in the specified page queue. If the page is
3817 * active or being moved to the active queue, ensure that its act_count is
3818 * at least ACT_INIT but do not otherwise mess with it. Otherwise, ensure that
3819 * the page is at the tail of its page queue.
3821 * The page may be wired. The caller should release its wiring reference
3822 * before releasing the page lock, otherwise the page daemon may immediately
3825 * A managed page must be locked.
3827 static __always_inline void
3828 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue)
3831 vm_page_assert_locked(m);
3832 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3833 ("vm_page_mvqueue: page %p is unmanaged", m));
3835 if (vm_page_queue(m) != nqueue) {
3837 vm_page_enqueue(m, nqueue);
3838 } else if (nqueue != PQ_ACTIVE) {
3842 if (nqueue == PQ_ACTIVE && m->act_count < ACT_INIT)
3843 m->act_count = ACT_INIT;
3847 * Put the specified page on the active list (if appropriate).
3850 vm_page_activate(vm_page_t m)
3853 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
3855 vm_page_mvqueue(m, PQ_ACTIVE);
3859 * Move the specified page to the tail of the inactive queue, or requeue
3860 * the page if it is already in the inactive queue.
3863 vm_page_deactivate(vm_page_t m)
3866 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
3868 vm_page_mvqueue(m, PQ_INACTIVE);
3872 * Move the specified page close to the head of the inactive queue,
3873 * bypassing LRU. A marker page is used to maintain FIFO ordering.
3874 * As with regular enqueues, we use a per-CPU batch queue to reduce
3875 * contention on the page queue lock.
3878 _vm_page_deactivate_noreuse(vm_page_t m)
3881 vm_page_assert_locked(m);
3883 if (!vm_page_inactive(m)) {
3885 m->queue = PQ_INACTIVE;
3887 if ((m->aflags & PGA_REQUEUE_HEAD) == 0)
3888 vm_page_aflag_set(m, PGA_REQUEUE_HEAD);
3889 vm_page_pqbatch_submit(m, PQ_INACTIVE);
3893 vm_page_deactivate_noreuse(vm_page_t m)
3896 KASSERT(m->object != NULL,
3897 ("vm_page_deactivate_noreuse: page %p has no object", m));
3899 if ((m->oflags & VPO_UNMANAGED) == 0 && !vm_page_wired(m))
3900 _vm_page_deactivate_noreuse(m);
3904 * Put a page in the laundry, or requeue it if it is already there.
3907 vm_page_launder(vm_page_t m)
3910 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
3912 vm_page_mvqueue(m, PQ_LAUNDRY);
3916 * Put a page in the PQ_UNSWAPPABLE holding queue.
3919 vm_page_unswappable(vm_page_t m)
3922 vm_page_assert_locked(m);
3923 KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0,
3924 ("page %p already unswappable", m));
3927 vm_page_enqueue(m, PQ_UNSWAPPABLE);
3931 vm_page_release_toq(vm_page_t m, int flags)
3934 vm_page_assert_locked(m);
3937 * Use a check of the valid bits to determine whether we should
3938 * accelerate reclamation of the page. The object lock might not be
3939 * held here, in which case the check is racy. At worst we will either
3940 * accelerate reclamation of a valid page and violate LRU, or
3941 * unnecessarily defer reclamation of an invalid page.
3943 * If we were asked to not cache the page, place it near the head of the
3944 * inactive queue so that is reclaimed sooner.
3946 if ((flags & (VPR_TRYFREE | VPR_NOREUSE)) != 0 || m->valid == 0)
3947 _vm_page_deactivate_noreuse(m);
3948 else if (vm_page_active(m))
3949 vm_page_reference(m);
3951 vm_page_mvqueue(m, PQ_INACTIVE);
3955 * Unwire a page and either attempt to free it or re-add it to the page queues.
3958 vm_page_release(vm_page_t m, int flags)
3964 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3965 ("vm_page_release: page %p is unmanaged", m));
3967 if ((flags & VPR_TRYFREE) != 0) {
3969 object = (vm_object_t)atomic_load_ptr(&m->object);
3972 /* Depends on type-stability. */
3973 if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object)) {
3977 if (object == m->object)
3979 VM_OBJECT_WUNLOCK(object);
3981 if (__predict_true(object != NULL)) {
3982 vm_page_release_locked(m, flags);
3983 VM_OBJECT_WUNLOCK(object);
3989 * Update LRU state before releasing the wiring reference.
3990 * Use a release store when updating the reference count to
3991 * synchronize with vm_page_free_prep().
3996 KASSERT(VPRC_WIRE_COUNT(old) > 0,
3997 ("vm_page_unwire: wire count underflow for page %p", m));
3998 if (!locked && VPRC_WIRE_COUNT(old) == 1) {
4001 vm_page_release_toq(m, flags);
4003 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
4006 * Release the lock only after the wiring is released, to ensure that
4007 * the page daemon does not encounter and dequeue the page while it is
4013 if (VPRC_WIRE_COUNT(old) == 1) {
4020 /* See vm_page_release(). */
4022 vm_page_release_locked(vm_page_t m, int flags)
4025 VM_OBJECT_ASSERT_WLOCKED(m->object);
4026 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4027 ("vm_page_release_locked: page %p is unmanaged", m));
4029 if (vm_page_unwire_noq(m)) {
4030 if ((flags & VPR_TRYFREE) != 0 &&
4031 (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
4032 m->dirty == 0 && !vm_page_busied(m)) {
4036 vm_page_release_toq(m, flags);
4043 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
4047 KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
4048 ("vm_page_try_blocked_op: page %p has no object", m));
4049 KASSERT(!vm_page_busied(m),
4050 ("vm_page_try_blocked_op: page %p is busy", m));
4051 VM_OBJECT_ASSERT_LOCKED(m->object);
4056 ("vm_page_try_blocked_op: page %p has no references", m));
4057 if (VPRC_WIRE_COUNT(old) != 0)
4059 } while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
4064 * If the object is read-locked, new wirings may be created via an
4067 old = vm_page_drop(m, VPRC_BLOCKED);
4068 KASSERT(!VM_OBJECT_WOWNED(m->object) ||
4069 old == (VPRC_BLOCKED | VPRC_OBJREF),
4070 ("vm_page_try_blocked_op: unexpected refcount value %u for %p",
4076 * Atomically check for wirings and remove all mappings of the page.
4079 vm_page_try_remove_all(vm_page_t m)
4082 return (vm_page_try_blocked_op(m, pmap_remove_all));
4086 * Atomically check for wirings and remove all writeable mappings of the page.
4089 vm_page_try_remove_write(vm_page_t m)
4092 return (vm_page_try_blocked_op(m, pmap_remove_write));
4098 * Apply the specified advice to the given page.
4100 * The object and page must be locked.
4103 vm_page_advise(vm_page_t m, int advice)
4106 vm_page_assert_locked(m);
4107 VM_OBJECT_ASSERT_WLOCKED(m->object);
4108 if (advice == MADV_FREE)
4110 * Mark the page clean. This will allow the page to be freed
4111 * without first paging it out. MADV_FREE pages are often
4112 * quickly reused by malloc(3), so we do not do anything that
4113 * would result in a page fault on a later access.
4116 else if (advice != MADV_DONTNEED) {
4117 if (advice == MADV_WILLNEED)
4118 vm_page_activate(m);
4123 * Clear any references to the page. Otherwise, the page daemon will
4124 * immediately reactivate the page.
4126 vm_page_aflag_clear(m, PGA_REFERENCED);
4128 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
4132 * Place clean pages near the head of the inactive queue rather than
4133 * the tail, thus defeating the queue's LRU operation and ensuring that
4134 * the page will be reused quickly. Dirty pages not already in the
4135 * laundry are moved there.
4138 vm_page_deactivate_noreuse(m);
4139 else if (!vm_page_in_laundry(m))
4144 * Grab a page, waiting until we are waken up due to the page
4145 * changing state. We keep on waiting, if the page continues
4146 * to be in the object. If the page doesn't exist, first allocate it
4147 * and then conditionally zero it.
4149 * This routine may sleep.
4151 * The object must be locked on entry. The lock will, however, be released
4152 * and reacquired if the routine sleeps.
4155 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
4161 VM_OBJECT_ASSERT_WLOCKED(object);
4162 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4163 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4164 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4165 pflags = allocflags &
4166 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
4167 if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4168 pflags |= VM_ALLOC_WAITFAIL;
4170 if ((m = vm_page_lookup(object, pindex)) != NULL) {
4171 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
4172 vm_page_xbusied(m) : vm_page_busied(m);
4174 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4177 * Reference the page before unlocking and
4178 * sleeping so that the page daemon is less
4179 * likely to reclaim it.
4181 if ((allocflags & VM_ALLOC_NOCREAT) == 0)
4182 vm_page_aflag_set(m, PGA_REFERENCED);
4183 vm_page_busy_sleep(m, "pgrbwt", (allocflags &
4184 VM_ALLOC_IGN_SBUSY) != 0);
4185 VM_OBJECT_WLOCK(object);
4186 if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
4190 if ((allocflags & VM_ALLOC_WIRED) != 0)
4193 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
4195 else if ((allocflags & VM_ALLOC_SBUSY) != 0)
4200 if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4202 m = vm_page_alloc(object, pindex, pflags);
4204 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4208 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
4214 * Grab a page and make it valid, paging in if necessary. Pages missing from
4215 * their pager are zero filled and validated.
4218 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags)
4225 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4226 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4227 ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4228 KASSERT((allocflags &
4229 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4230 ("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
4231 VM_OBJECT_ASSERT_WLOCKED(object);
4232 pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY);
4233 pflags |= VM_ALLOC_WAITFAIL;
4237 if ((m = vm_page_lookup(object, pindex)) != NULL) {
4239 * If the page is fully valid it can only become invalid
4240 * with the object lock held. If it is not valid it can
4241 * become valid with the busy lock held. Therefore, we
4242 * may unnecessarily lock the exclusive busy here if we
4243 * race with I/O completion not using the object lock.
4244 * However, we will not end up with an invalid page and a
4247 if (m->valid != VM_PAGE_BITS_ALL ||
4248 (allocflags & (VM_ALLOC_IGN_SBUSY | VM_ALLOC_SBUSY)) == 0) {
4249 sleep = !vm_page_tryxbusy(m);
4252 sleep = !vm_page_trysbusy(m);
4255 * Reference the page before unlocking and
4256 * sleeping so that the page daemon is less
4257 * likely to reclaim it.
4259 if ((allocflags & VM_ALLOC_NOCREAT) == 0)
4260 vm_page_aflag_set(m, PGA_REFERENCED);
4261 vm_page_busy_sleep(m, "pgrbwt", (allocflags &
4262 VM_ALLOC_IGN_SBUSY) != 0);
4263 VM_OBJECT_WLOCK(object);
4266 if ((allocflags & VM_ALLOC_NOCREAT) != 0 &&
4267 m->valid != VM_PAGE_BITS_ALL) {
4273 return (VM_PAGER_FAIL);
4275 if ((allocflags & VM_ALLOC_WIRED) != 0)
4277 if (m->valid == VM_PAGE_BITS_ALL)
4279 } else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4281 return (VM_PAGER_FAIL);
4282 } else if ((m = vm_page_alloc(object, pindex, pflags)) != NULL) {
4288 vm_page_assert_xbusied(m);
4290 if (vm_pager_has_page(object, pindex, NULL, NULL)) {
4291 rv = vm_pager_get_pages(object, &m, 1, NULL, NULL);
4292 if (rv != VM_PAGER_OK) {
4293 if (allocflags & VM_ALLOC_WIRED)
4294 vm_page_unwire_noq(m);
4299 MPASS(m->valid == VM_PAGE_BITS_ALL);
4301 vm_page_zero_invalid(m, TRUE);
4304 if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4310 if ((allocflags & VM_ALLOC_SBUSY) != 0 && xbusy)
4311 vm_page_busy_downgrade(m);
4313 return (VM_PAGER_OK);
4317 * Return the specified range of pages from the given object. For each
4318 * page offset within the range, if a page already exists within the object
4319 * at that offset and it is busy, then wait for it to change state. If,
4320 * instead, the page doesn't exist, then allocate it.
4322 * The caller must always specify an allocation class.
4324 * allocation classes:
4325 * VM_ALLOC_NORMAL normal process request
4326 * VM_ALLOC_SYSTEM system *really* needs the pages
4328 * The caller must always specify that the pages are to be busied and/or
4331 * optional allocation flags:
4332 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages
4333 * VM_ALLOC_NOBUSY do not exclusive busy the page
4334 * VM_ALLOC_NOWAIT do not sleep
4335 * VM_ALLOC_SBUSY set page to sbusy state
4336 * VM_ALLOC_WIRED wire the pages
4337 * VM_ALLOC_ZERO zero and validate any invalid pages
4339 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it
4340 * may return a partial prefix of the requested range.
4343 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
4344 vm_page_t *ma, int count)
4351 VM_OBJECT_ASSERT_WLOCKED(object);
4352 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
4353 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
4354 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4355 (allocflags & VM_ALLOC_WIRED) != 0,
4356 ("vm_page_grab_pages: the pages must be busied or wired"));
4357 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4358 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4359 ("vm_page_grab_pages: VM_ALLOC_SBUSY/IGN_SBUSY mismatch"));
4362 pflags = allocflags & ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK |
4363 VM_ALLOC_WAITFAIL | VM_ALLOC_IGN_SBUSY);
4364 if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4365 pflags |= VM_ALLOC_WAITFAIL;
4368 m = vm_radix_lookup_le(&object->rtree, pindex + i);
4369 if (m == NULL || m->pindex != pindex + i) {
4373 mpred = TAILQ_PREV(m, pglist, listq);
4374 for (; i < count; i++) {
4376 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
4377 vm_page_xbusied(m) : vm_page_busied(m);
4379 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4382 * Reference the page before unlocking and
4383 * sleeping so that the page daemon is less
4384 * likely to reclaim it.
4386 if ((allocflags & VM_ALLOC_NOCREAT) == 0)
4387 vm_page_aflag_set(m, PGA_REFERENCED);
4388 vm_page_busy_sleep(m, "grbmaw", (allocflags &
4389 VM_ALLOC_IGN_SBUSY) != 0);
4390 VM_OBJECT_WLOCK(object);
4393 if ((allocflags & VM_ALLOC_WIRED) != 0)
4395 if ((allocflags & (VM_ALLOC_NOBUSY |
4396 VM_ALLOC_SBUSY)) == 0)
4398 if ((allocflags & VM_ALLOC_SBUSY) != 0)
4401 if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4403 m = vm_page_alloc_after(object, pindex + i,
4404 pflags | VM_ALLOC_COUNT(count - i), mpred);
4406 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4411 if (m->valid == 0 && (allocflags & VM_ALLOC_ZERO) != 0) {
4412 if ((m->flags & PG_ZERO) == 0)
4414 m->valid = VM_PAGE_BITS_ALL;
4417 m = vm_page_next(m);
4423 * Mapping function for valid or dirty bits in a page.
4425 * Inputs are required to range within a page.
4428 vm_page_bits(int base, int size)
4434 base + size <= PAGE_SIZE,
4435 ("vm_page_bits: illegal base/size %d/%d", base, size)
4438 if (size == 0) /* handle degenerate case */
4441 first_bit = base >> DEV_BSHIFT;
4442 last_bit = (base + size - 1) >> DEV_BSHIFT;
4444 return (((vm_page_bits_t)2 << last_bit) -
4445 ((vm_page_bits_t)1 << first_bit));
4449 * vm_page_set_valid_range:
4451 * Sets portions of a page valid. The arguments are expected
4452 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4453 * of any partial chunks touched by the range. The invalid portion of
4454 * such chunks will be zeroed.
4456 * (base + size) must be less then or equal to PAGE_SIZE.
4459 vm_page_set_valid_range(vm_page_t m, int base, int size)
4463 VM_OBJECT_ASSERT_WLOCKED(m->object);
4464 if (size == 0) /* handle degenerate case */
4468 * If the base is not DEV_BSIZE aligned and the valid
4469 * bit is clear, we have to zero out a portion of the
4472 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4473 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
4474 pmap_zero_page_area(m, frag, base - frag);
4477 * If the ending offset is not DEV_BSIZE aligned and the
4478 * valid bit is clear, we have to zero out a portion of
4481 endoff = base + size;
4482 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4483 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
4484 pmap_zero_page_area(m, endoff,
4485 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4488 * Assert that no previously invalid block that is now being validated
4491 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
4492 ("vm_page_set_valid_range: page %p is dirty", m));
4495 * Set valid bits inclusive of any overlap.
4497 m->valid |= vm_page_bits(base, size);
4501 * Clear the given bits from the specified page's dirty field.
4503 static __inline void
4504 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
4507 #if PAGE_SIZE < 16384
4512 * If the object is locked and the page is neither exclusive busy nor
4513 * write mapped, then the page's dirty field cannot possibly be
4514 * set by a concurrent pmap operation.
4516 VM_OBJECT_ASSERT_WLOCKED(m->object);
4517 if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
4518 m->dirty &= ~pagebits;
4521 * The pmap layer can call vm_page_dirty() without
4522 * holding a distinguished lock. The combination of
4523 * the object's lock and an atomic operation suffice
4524 * to guarantee consistency of the page dirty field.
4526 * For PAGE_SIZE == 32768 case, compiler already
4527 * properly aligns the dirty field, so no forcible
4528 * alignment is needed. Only require existence of
4529 * atomic_clear_64 when page size is 32768.
4531 addr = (uintptr_t)&m->dirty;
4532 #if PAGE_SIZE == 32768
4533 atomic_clear_64((uint64_t *)addr, pagebits);
4534 #elif PAGE_SIZE == 16384
4535 atomic_clear_32((uint32_t *)addr, pagebits);
4536 #else /* PAGE_SIZE <= 8192 */
4538 * Use a trick to perform a 32-bit atomic on the
4539 * containing aligned word, to not depend on the existence
4540 * of atomic_clear_{8, 16}.
4542 shift = addr & (sizeof(uint32_t) - 1);
4543 #if BYTE_ORDER == BIG_ENDIAN
4544 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
4548 addr &= ~(sizeof(uint32_t) - 1);
4549 atomic_clear_32((uint32_t *)addr, pagebits << shift);
4550 #endif /* PAGE_SIZE */
4555 * vm_page_set_validclean:
4557 * Sets portions of a page valid and clean. The arguments are expected
4558 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4559 * of any partial chunks touched by the range. The invalid portion of
4560 * such chunks will be zero'd.
4562 * (base + size) must be less then or equal to PAGE_SIZE.
4565 vm_page_set_validclean(vm_page_t m, int base, int size)
4567 vm_page_bits_t oldvalid, pagebits;
4570 VM_OBJECT_ASSERT_WLOCKED(m->object);
4571 if (size == 0) /* handle degenerate case */
4575 * If the base is not DEV_BSIZE aligned and the valid
4576 * bit is clear, we have to zero out a portion of the
4579 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4580 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
4581 pmap_zero_page_area(m, frag, base - frag);
4584 * If the ending offset is not DEV_BSIZE aligned and the
4585 * valid bit is clear, we have to zero out a portion of
4588 endoff = base + size;
4589 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4590 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
4591 pmap_zero_page_area(m, endoff,
4592 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4595 * Set valid, clear dirty bits. If validating the entire
4596 * page we can safely clear the pmap modify bit. We also
4597 * use this opportunity to clear the VPO_NOSYNC flag. If a process
4598 * takes a write fault on a MAP_NOSYNC memory area the flag will
4601 * We set valid bits inclusive of any overlap, but we can only
4602 * clear dirty bits for DEV_BSIZE chunks that are fully within
4605 oldvalid = m->valid;
4606 pagebits = vm_page_bits(base, size);
4607 m->valid |= pagebits;
4609 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
4610 frag = DEV_BSIZE - frag;
4616 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
4618 if (base == 0 && size == PAGE_SIZE) {
4620 * The page can only be modified within the pmap if it is
4621 * mapped, and it can only be mapped if it was previously
4624 if (oldvalid == VM_PAGE_BITS_ALL)
4626 * Perform the pmap_clear_modify() first. Otherwise,
4627 * a concurrent pmap operation, such as
4628 * pmap_protect(), could clear a modification in the
4629 * pmap and set the dirty field on the page before
4630 * pmap_clear_modify() had begun and after the dirty
4631 * field was cleared here.
4633 pmap_clear_modify(m);
4635 m->oflags &= ~VPO_NOSYNC;
4636 } else if (oldvalid != VM_PAGE_BITS_ALL)
4637 m->dirty &= ~pagebits;
4639 vm_page_clear_dirty_mask(m, pagebits);
4643 vm_page_clear_dirty(vm_page_t m, int base, int size)
4646 vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
4650 * vm_page_set_invalid:
4652 * Invalidates DEV_BSIZE'd chunks within a page. Both the
4653 * valid and dirty bits for the effected areas are cleared.
4656 vm_page_set_invalid(vm_page_t m, int base, int size)
4658 vm_page_bits_t bits;
4662 VM_OBJECT_ASSERT_WLOCKED(object);
4663 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
4664 size >= object->un_pager.vnp.vnp_size)
4665 bits = VM_PAGE_BITS_ALL;
4667 bits = vm_page_bits(base, size);
4668 if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL &&
4671 KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) ||
4672 !pmap_page_is_mapped(m),
4673 ("vm_page_set_invalid: page %p is mapped", m));
4679 * vm_page_zero_invalid()
4681 * The kernel assumes that the invalid portions of a page contain
4682 * garbage, but such pages can be mapped into memory by user code.
4683 * When this occurs, we must zero out the non-valid portions of the
4684 * page so user code sees what it expects.
4686 * Pages are most often semi-valid when the end of a file is mapped
4687 * into memory and the file's size is not page aligned.
4690 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
4695 VM_OBJECT_ASSERT_WLOCKED(m->object);
4697 * Scan the valid bits looking for invalid sections that
4698 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the
4699 * valid bit may be set ) have already been zeroed by
4700 * vm_page_set_validclean().
4702 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
4703 if (i == (PAGE_SIZE / DEV_BSIZE) ||
4704 (m->valid & ((vm_page_bits_t)1 << i))) {
4706 pmap_zero_page_area(m,
4707 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
4714 * setvalid is TRUE when we can safely set the zero'd areas
4715 * as being valid. We can do this if there are no cache consistancy
4716 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
4719 m->valid = VM_PAGE_BITS_ALL;
4725 * Is (partial) page valid? Note that the case where size == 0
4726 * will return FALSE in the degenerate case where the page is
4727 * entirely invalid, and TRUE otherwise.
4730 vm_page_is_valid(vm_page_t m, int base, int size)
4732 vm_page_bits_t bits;
4734 VM_OBJECT_ASSERT_LOCKED(m->object);
4735 bits = vm_page_bits(base, size);
4736 return (m->valid != 0 && (m->valid & bits) == bits);
4740 * Returns true if all of the specified predicates are true for the entire
4741 * (super)page and false otherwise.
4744 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
4750 if (skip_m != NULL && skip_m->object != object)
4752 VM_OBJECT_ASSERT_LOCKED(object);
4753 npages = atop(pagesizes[m->psind]);
4756 * The physically contiguous pages that make up a superpage, i.e., a
4757 * page with a page size index ("psind") greater than zero, will
4758 * occupy adjacent entries in vm_page_array[].
4760 for (i = 0; i < npages; i++) {
4761 /* Always test object consistency, including "skip_m". */
4762 if (m[i].object != object)
4764 if (&m[i] == skip_m)
4766 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
4768 if ((flags & PS_ALL_DIRTY) != 0) {
4770 * Calling vm_page_test_dirty() or pmap_is_modified()
4771 * might stop this case from spuriously returning
4772 * "false". However, that would require a write lock
4773 * on the object containing "m[i]".
4775 if (m[i].dirty != VM_PAGE_BITS_ALL)
4778 if ((flags & PS_ALL_VALID) != 0 &&
4779 m[i].valid != VM_PAGE_BITS_ALL)
4786 * Set the page's dirty bits if the page is modified.
4789 vm_page_test_dirty(vm_page_t m)
4792 VM_OBJECT_ASSERT_WLOCKED(m->object);
4793 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
4798 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
4801 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
4805 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
4808 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
4812 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
4815 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
4818 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
4820 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
4823 vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
4827 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
4830 mtx_assert_(vm_page_lockptr(m), a, file, line);
4836 vm_page_object_lock_assert(vm_page_t m)
4840 * Certain of the page's fields may only be modified by the
4841 * holder of the containing object's lock or the exclusive busy.
4842 * holder. Unfortunately, the holder of the write busy is
4843 * not recorded, and thus cannot be checked here.
4845 if (m->object != NULL && !vm_page_xbusied(m))
4846 VM_OBJECT_ASSERT_WLOCKED(m->object);
4850 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits)
4853 if ((bits & PGA_WRITEABLE) == 0)
4857 * The PGA_WRITEABLE flag can only be set if the page is
4858 * managed, is exclusively busied or the object is locked.
4859 * Currently, this flag is only set by pmap_enter().
4861 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4862 ("PGA_WRITEABLE on unmanaged page"));
4863 if (!vm_page_xbusied(m))
4864 VM_OBJECT_ASSERT_LOCKED(m->object);
4868 #include "opt_ddb.h"
4870 #include <sys/kernel.h>
4872 #include <ddb/ddb.h>
4874 DB_SHOW_COMMAND(page, vm_page_print_page_info)
4877 db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
4878 db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
4879 db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
4880 db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
4881 db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
4882 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
4883 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
4884 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
4885 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
4888 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
4892 db_printf("pq_free %d\n", vm_free_count());
4893 for (dom = 0; dom < vm_ndomains; dom++) {
4895 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
4897 vm_dom[dom].vmd_page_count,
4898 vm_dom[dom].vmd_free_count,
4899 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
4900 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
4901 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
4902 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
4906 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
4909 boolean_t phys, virt;
4912 db_printf("show pginfo addr\n");
4916 phys = strchr(modif, 'p') != NULL;
4917 virt = strchr(modif, 'v') != NULL;
4919 m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
4921 m = PHYS_TO_VM_PAGE(addr);
4923 m = (vm_page_t)addr;
4925 "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref %u\n"
4926 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
4927 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
4928 m->queue, m->ref_count, m->aflags, m->oflags,
4929 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty);