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>
111 #include <vm/vm_dumpset.h>
113 #include <vm/uma_int.h>
115 #include <machine/md_var.h>
117 struct vm_domain vm_dom[MAXMEMDOM];
119 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]);
121 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
123 struct mtx_padalign __exclusive_cache_line vm_domainset_lock;
124 /* The following fields are protected by the domainset lock. */
125 domainset_t __exclusive_cache_line vm_min_domains;
126 domainset_t __exclusive_cache_line vm_severe_domains;
127 static int vm_min_waiters;
128 static int vm_severe_waiters;
129 static int vm_pageproc_waiters;
131 static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
132 "VM page statistics");
134 static COUNTER_U64_DEFINE_EARLY(pqstate_commit_retries);
135 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries,
136 CTLFLAG_RD, &pqstate_commit_retries,
137 "Number of failed per-page atomic queue state updates");
139 static COUNTER_U64_DEFINE_EARLY(queue_ops);
140 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops,
141 CTLFLAG_RD, &queue_ops,
142 "Number of batched queue operations");
144 static COUNTER_U64_DEFINE_EARLY(queue_nops);
145 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops,
146 CTLFLAG_RD, &queue_nops,
147 "Number of batched queue operations with no effects");
150 * bogus page -- for I/O to/from partially complete buffers,
151 * or for paging into sparsely invalid regions.
153 vm_page_t bogus_page;
155 vm_page_t vm_page_array;
156 long vm_page_array_size;
159 struct bitset *vm_page_dump;
160 long vm_page_dump_pages;
162 static TAILQ_HEAD(, vm_page) blacklist_head;
163 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
164 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
165 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
167 static uma_zone_t fakepg_zone;
169 static void vm_page_alloc_check(vm_page_t m);
170 static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m,
171 vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked);
172 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
173 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
174 static bool vm_page_free_prep(vm_page_t m);
175 static void vm_page_free_toq(vm_page_t m);
176 static void vm_page_init(void *dummy);
177 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
178 vm_pindex_t pindex, vm_page_t mpred);
179 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
181 static void vm_page_mvqueue(vm_page_t m, const uint8_t queue,
182 const uint16_t nflag);
183 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
184 vm_page_t m_run, vm_paddr_t high);
185 static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse);
186 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
188 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain,
190 static void vm_page_zone_release(void *arg, void **store, int cnt);
192 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
195 vm_page_init(void *dummy)
198 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
199 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
200 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
201 VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
205 * The cache page zone is initialized later since we need to be able to allocate
206 * pages before UMA is fully initialized.
209 vm_page_init_cache_zones(void *dummy __unused)
211 struct vm_domain *vmd;
212 struct vm_pgcache *pgcache;
213 int cache, domain, maxcache, pool;
216 TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &maxcache);
217 maxcache *= mp_ncpus;
218 for (domain = 0; domain < vm_ndomains; domain++) {
219 vmd = VM_DOMAIN(domain);
220 for (pool = 0; pool < VM_NFREEPOOL; pool++) {
221 pgcache = &vmd->vmd_pgcache[pool];
222 pgcache->domain = domain;
223 pgcache->pool = pool;
224 pgcache->zone = uma_zcache_create("vm pgcache",
225 PAGE_SIZE, NULL, NULL, NULL, NULL,
226 vm_page_zone_import, vm_page_zone_release, pgcache,
230 * Limit each pool's zone to 0.1% of the pages in the
233 cache = maxcache != 0 ? maxcache :
234 vmd->vmd_page_count / 1000;
235 uma_zone_set_maxcache(pgcache->zone, cache);
239 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
241 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
242 #if PAGE_SIZE == 32768
244 CTASSERT(sizeof(u_long) >= 8);
251 * Sets the page size, perhaps based upon the memory
252 * size. Must be called before any use of page-size
253 * dependent functions.
256 vm_set_page_size(void)
258 if (vm_cnt.v_page_size == 0)
259 vm_cnt.v_page_size = PAGE_SIZE;
260 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
261 panic("vm_set_page_size: page size not a power of two");
265 * vm_page_blacklist_next:
267 * Find the next entry in the provided string of blacklist
268 * addresses. Entries are separated by space, comma, or newline.
269 * If an invalid integer is encountered then the rest of the
270 * string is skipped. Updates the list pointer to the next
271 * character, or NULL if the string is exhausted or invalid.
274 vm_page_blacklist_next(char **list, char *end)
279 if (list == NULL || *list == NULL)
287 * If there's no end pointer then the buffer is coming from
288 * the kenv and we know it's null-terminated.
291 end = *list + strlen(*list);
293 /* Ensure that strtoq() won't walk off the end */
295 if (*end == '\n' || *end == ' ' || *end == ',')
298 printf("Blacklist not terminated, skipping\n");
304 for (pos = *list; *pos != '\0'; pos = cp) {
305 bad = strtoq(pos, &cp, 0);
306 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
315 if (*cp == '\0' || ++cp >= end)
319 return (trunc_page(bad));
321 printf("Garbage in RAM blacklist, skipping\n");
327 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
329 struct vm_domain *vmd;
333 m = vm_phys_paddr_to_vm_page(pa);
335 return (true); /* page does not exist, no failure */
337 vmd = vm_pagequeue_domain(m);
338 vm_domain_free_lock(vmd);
339 ret = vm_phys_unfree_page(m);
340 vm_domain_free_unlock(vmd);
342 vm_domain_freecnt_inc(vmd, -1);
343 TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
345 printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
351 * vm_page_blacklist_check:
353 * Iterate through the provided string of blacklist addresses, pulling
354 * each entry out of the physical allocator free list and putting it
355 * onto a list for reporting via the vm.page_blacklist sysctl.
358 vm_page_blacklist_check(char *list, char *end)
364 while (next != NULL) {
365 if ((pa = vm_page_blacklist_next(&next, end)) == 0)
367 vm_page_blacklist_add(pa, bootverbose);
372 * vm_page_blacklist_load:
374 * Search for a special module named "ram_blacklist". It'll be a
375 * plain text file provided by the user via the loader directive
379 vm_page_blacklist_load(char **list, char **end)
388 mod = preload_search_by_type("ram_blacklist");
390 ptr = preload_fetch_addr(mod);
391 len = preload_fetch_size(mod);
402 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
409 error = sysctl_wire_old_buffer(req, 0);
412 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
413 TAILQ_FOREACH(m, &blacklist_head, listq) {
414 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
415 (uintmax_t)m->phys_addr);
418 error = sbuf_finish(&sbuf);
424 * Initialize a dummy page for use in scans of the specified paging queue.
425 * In principle, this function only needs to set the flag PG_MARKER.
426 * Nonetheless, it write busies the page as a safety precaution.
429 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags)
432 bzero(marker, sizeof(*marker));
433 marker->flags = PG_MARKER;
434 marker->a.flags = aflags;
435 marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
436 marker->a.queue = queue;
440 vm_page_domain_init(int domain)
442 struct vm_domain *vmd;
443 struct vm_pagequeue *pq;
446 vmd = VM_DOMAIN(domain);
447 bzero(vmd, sizeof(*vmd));
448 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
449 "vm inactive pagequeue";
450 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
451 "vm active pagequeue";
452 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
453 "vm laundry pagequeue";
454 *__DECONST(const char **,
455 &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
456 "vm unswappable pagequeue";
457 vmd->vmd_domain = domain;
458 vmd->vmd_page_count = 0;
459 vmd->vmd_free_count = 0;
461 vmd->vmd_oom = FALSE;
462 for (i = 0; i < PQ_COUNT; i++) {
463 pq = &vmd->vmd_pagequeues[i];
464 TAILQ_INIT(&pq->pq_pl);
465 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
466 MTX_DEF | MTX_DUPOK);
468 vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
470 mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
471 mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
472 snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
475 * inacthead is used to provide FIFO ordering for LRU-bypassing
478 vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
479 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
480 &vmd->vmd_inacthead, plinks.q);
483 * The clock pages are used to implement active queue scanning without
484 * requeues. Scans start at clock[0], which is advanced after the scan
485 * ends. When the two clock hands meet, they are reset and scanning
486 * resumes from the head of the queue.
488 vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
489 vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
490 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
491 &vmd->vmd_clock[0], plinks.q);
492 TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
493 &vmd->vmd_clock[1], plinks.q);
497 * Initialize a physical page in preparation for adding it to the free
501 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
506 m->busy_lock = VPB_FREED;
507 m->flags = m->a.flags = 0;
509 m->a.queue = PQ_NONE;
512 m->order = VM_NFREEORDER;
513 m->pool = VM_FREEPOOL_DEFAULT;
514 m->valid = m->dirty = 0;
518 #ifndef PMAP_HAS_PAGE_ARRAY
520 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range)
525 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
526 * However, because this page is allocated from KVM, out-of-bounds
527 * accesses using the direct map will not be trapped.
532 * Allocate physical memory for the page structures, and map it.
534 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
535 vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end,
536 VM_PROT_READ | VM_PROT_WRITE);
537 vm_page_array_size = page_range;
546 * Initializes the resident memory module. Allocates physical memory for
547 * bootstrapping UMA and some data structures that are used to manage
548 * physical pages. Initializes these structures, and populates the free
552 vm_page_startup(vm_offset_t vaddr)
554 struct vm_phys_seg *seg;
555 struct vm_domain *vmd;
557 char *list, *listend;
558 vm_paddr_t end, high_avail, low_avail, new_end, size;
559 vm_paddr_t page_range __unused;
560 vm_paddr_t last_pa, pa, startp, endp;
562 #if MINIDUMP_PAGE_TRACKING
563 u_long vm_page_dump_size;
565 int biggestone, i, segind;
570 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
574 vaddr = round_page(vaddr);
576 vm_phys_early_startup();
577 biggestone = vm_phys_avail_largest();
578 end = phys_avail[biggestone+1];
581 * Initialize the page and queue locks.
583 mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
584 for (i = 0; i < PA_LOCK_COUNT; i++)
585 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
586 for (i = 0; i < vm_ndomains; i++)
587 vm_page_domain_init(i);
591 witness_size = round_page(witness_startup_count());
592 new_end -= witness_size;
593 mapped = pmap_map(&vaddr, new_end, new_end + witness_size,
594 VM_PROT_READ | VM_PROT_WRITE);
595 bzero((void *)mapped, witness_size);
596 witness_startup((void *)mapped);
599 #if MINIDUMP_PAGE_TRACKING
601 * Allocate a bitmap to indicate that a random physical page
602 * needs to be included in a minidump.
604 * The amd64 port needs this to indicate which direct map pages
605 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
607 * However, i386 still needs this workspace internally within the
608 * minidump code. In theory, they are not needed on i386, but are
609 * included should the sf_buf code decide to use them.
612 vm_page_dump_pages = 0;
613 for (i = 0; dump_avail[i + 1] != 0; i += 2) {
614 vm_page_dump_pages += howmany(dump_avail[i + 1], PAGE_SIZE) -
615 dump_avail[i] / PAGE_SIZE;
616 if (dump_avail[i + 1] > last_pa)
617 last_pa = dump_avail[i + 1];
619 vm_page_dump_size = round_page(BITSET_SIZE(vm_page_dump_pages));
620 new_end -= vm_page_dump_size;
621 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
622 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
623 bzero((void *)vm_page_dump, vm_page_dump_size);
627 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
628 defined(__riscv) || defined(__powerpc64__)
630 * Include the UMA bootstrap pages, witness pages and vm_page_dump
631 * in a crash dump. When pmap_map() uses the direct map, they are
632 * not automatically included.
634 for (pa = new_end; pa < end; pa += PAGE_SIZE)
637 phys_avail[biggestone + 1] = new_end;
640 * Request that the physical pages underlying the message buffer be
641 * included in a crash dump. Since the message buffer is accessed
642 * through the direct map, they are not automatically included.
644 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
645 last_pa = pa + round_page(msgbufsize);
646 while (pa < last_pa) {
652 * Compute the number of pages of memory that will be available for
653 * use, taking into account the overhead of a page structure per page.
654 * In other words, solve
655 * "available physical memory" - round_page(page_range *
656 * sizeof(struct vm_page)) = page_range * PAGE_SIZE
659 low_avail = phys_avail[0];
660 high_avail = phys_avail[1];
661 for (i = 0; i < vm_phys_nsegs; i++) {
662 if (vm_phys_segs[i].start < low_avail)
663 low_avail = vm_phys_segs[i].start;
664 if (vm_phys_segs[i].end > high_avail)
665 high_avail = vm_phys_segs[i].end;
667 /* Skip the first chunk. It is already accounted for. */
668 for (i = 2; phys_avail[i + 1] != 0; i += 2) {
669 if (phys_avail[i] < low_avail)
670 low_avail = phys_avail[i];
671 if (phys_avail[i + 1] > high_avail)
672 high_avail = phys_avail[i + 1];
674 first_page = low_avail / PAGE_SIZE;
675 #ifdef VM_PHYSSEG_SPARSE
677 for (i = 0; i < vm_phys_nsegs; i++)
678 size += vm_phys_segs[i].end - vm_phys_segs[i].start;
679 for (i = 0; phys_avail[i + 1] != 0; i += 2)
680 size += phys_avail[i + 1] - phys_avail[i];
681 #elif defined(VM_PHYSSEG_DENSE)
682 size = high_avail - low_avail;
684 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
687 #ifdef PMAP_HAS_PAGE_ARRAY
688 pmap_page_array_startup(size / PAGE_SIZE);
689 biggestone = vm_phys_avail_largest();
690 end = new_end = phys_avail[biggestone + 1];
692 #ifdef VM_PHYSSEG_DENSE
694 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
695 * the overhead of a page structure per page only if vm_page_array is
696 * allocated from the last physical memory chunk. Otherwise, we must
697 * allocate page structures representing the physical memory
698 * underlying vm_page_array, even though they will not be used.
700 if (new_end != high_avail)
701 page_range = size / PAGE_SIZE;
705 page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
708 * If the partial bytes remaining are large enough for
709 * a page (PAGE_SIZE) without a corresponding
710 * 'struct vm_page', then new_end will contain an
711 * extra page after subtracting the length of the VM
712 * page array. Compensate by subtracting an extra
715 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
716 if (new_end == high_avail)
717 high_avail -= PAGE_SIZE;
718 new_end -= PAGE_SIZE;
722 new_end = vm_page_array_alloc(&vaddr, end, page_range);
725 #if VM_NRESERVLEVEL > 0
727 * Allocate physical memory for the reservation management system's
728 * data structures, and map it.
730 new_end = vm_reserv_startup(&vaddr, new_end);
732 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
733 defined(__riscv) || defined(__powerpc64__)
735 * Include vm_page_array and vm_reserv_array in a crash dump.
737 for (pa = new_end; pa < end; pa += PAGE_SIZE)
740 phys_avail[biggestone + 1] = new_end;
743 * Add physical memory segments corresponding to the available
746 for (i = 0; phys_avail[i + 1] != 0; i += 2)
747 if (vm_phys_avail_size(i) != 0)
748 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
751 * Initialize the physical memory allocator.
756 * Initialize the page structures and add every available page to the
757 * physical memory allocator's free lists.
759 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
760 for (ii = 0; ii < vm_page_array_size; ii++) {
761 m = &vm_page_array[ii];
762 vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0);
763 m->flags = PG_FICTITIOUS;
766 vm_cnt.v_page_count = 0;
767 for (segind = 0; segind < vm_phys_nsegs; segind++) {
768 seg = &vm_phys_segs[segind];
769 for (m = seg->first_page, pa = seg->start; pa < seg->end;
770 m++, pa += PAGE_SIZE)
771 vm_page_init_page(m, pa, segind);
774 * Add the segment's pages that are covered by one of
775 * phys_avail's ranges to the free lists.
777 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
778 if (seg->end <= phys_avail[i] ||
779 seg->start >= phys_avail[i + 1])
782 startp = MAX(seg->start, phys_avail[i]);
783 endp = MIN(seg->end, phys_avail[i + 1]);
784 pagecount = (u_long)atop(endp - startp);
788 m = seg->first_page + atop(startp - seg->start);
789 vmd = VM_DOMAIN(seg->domain);
790 vm_domain_free_lock(vmd);
791 vm_phys_enqueue_contig(m, pagecount);
792 vm_domain_free_unlock(vmd);
793 vm_domain_freecnt_inc(vmd, pagecount);
794 vm_cnt.v_page_count += (u_int)pagecount;
795 vmd->vmd_page_count += (u_int)pagecount;
796 vmd->vmd_segs |= 1UL << segind;
801 * Remove blacklisted pages from the physical memory allocator.
803 TAILQ_INIT(&blacklist_head);
804 vm_page_blacklist_load(&list, &listend);
805 vm_page_blacklist_check(list, listend);
807 list = kern_getenv("vm.blacklist");
808 vm_page_blacklist_check(list, NULL);
811 #if VM_NRESERVLEVEL > 0
813 * Initialize the reservation management system.
822 vm_page_reference(vm_page_t m)
825 vm_page_aflag_set(m, PGA_REFERENCED);
831 * Helper routine for grab functions to trylock busy.
833 * Returns true on success and false on failure.
836 vm_page_trybusy(vm_page_t m, int allocflags)
839 if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0)
840 return (vm_page_trysbusy(m));
842 return (vm_page_tryxbusy(m));
848 * Helper routine for grab functions to trylock busy and wire.
850 * Returns true on success and false on failure.
853 vm_page_tryacquire(vm_page_t m, int allocflags)
857 locked = vm_page_trybusy(m, allocflags);
858 if (locked && (allocflags & VM_ALLOC_WIRED) != 0)
864 * vm_page_busy_acquire:
866 * Acquire the busy lock as described by VM_ALLOC_* flags. Will loop
867 * and drop the object lock if necessary.
870 vm_page_busy_acquire(vm_page_t m, int allocflags)
876 * The page-specific object must be cached because page
877 * identity can change during the sleep, causing the
878 * re-lock of a different object.
879 * It is assumed that a reference to the object is already
880 * held by the callers.
882 obj = atomic_load_ptr(&m->object);
884 if (vm_page_tryacquire(m, allocflags))
886 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
889 locked = VM_OBJECT_WOWNED(obj);
892 MPASS(locked || vm_page_wired(m));
893 if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags,
895 VM_OBJECT_WLOCK(obj);
896 if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
898 KASSERT(m->object == obj || m->object == NULL,
899 ("vm_page_busy_acquire: page %p does not belong to %p",
905 * vm_page_busy_downgrade:
907 * Downgrade an exclusive busy page into a single shared busy page.
910 vm_page_busy_downgrade(vm_page_t m)
914 vm_page_assert_xbusied(m);
916 x = vm_page_busy_fetch(m);
918 if (atomic_fcmpset_rel_int(&m->busy_lock,
919 &x, VPB_SHARERS_WORD(1)))
922 if ((x & VPB_BIT_WAITERS) != 0)
928 * vm_page_busy_tryupgrade:
930 * Attempt to upgrade a single shared busy into an exclusive busy.
933 vm_page_busy_tryupgrade(vm_page_t m)
937 vm_page_assert_sbusied(m);
939 x = vm_page_busy_fetch(m);
940 ce = VPB_CURTHREAD_EXCLUSIVE;
942 if (VPB_SHARERS(x) > 1)
944 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
945 ("vm_page_busy_tryupgrade: invalid lock state"));
946 if (!atomic_fcmpset_acq_int(&m->busy_lock, &x,
947 ce | (x & VPB_BIT_WAITERS)))
956 * Return a positive value if the page is shared busied, 0 otherwise.
959 vm_page_sbusied(vm_page_t m)
963 x = vm_page_busy_fetch(m);
964 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
970 * Shared unbusy a page.
973 vm_page_sunbusy(vm_page_t m)
977 vm_page_assert_sbusied(m);
979 x = vm_page_busy_fetch(m);
981 KASSERT(x != VPB_FREED,
982 ("vm_page_sunbusy: Unlocking freed page."));
983 if (VPB_SHARERS(x) > 1) {
984 if (atomic_fcmpset_int(&m->busy_lock, &x,
989 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
990 ("vm_page_sunbusy: invalid lock state"));
991 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
993 if ((x & VPB_BIT_WAITERS) == 0)
1001 * vm_page_busy_sleep:
1003 * Sleep if the page is busy, using the page pointer as wchan.
1004 * This is used to implement the hard-path of busying mechanism.
1006 * If nonshared is true, sleep only if the page is xbusy.
1008 * The object lock must be held on entry and will be released on exit.
1011 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
1016 VM_OBJECT_ASSERT_LOCKED(obj);
1017 vm_page_lock_assert(m, MA_NOTOWNED);
1019 if (!_vm_page_busy_sleep(obj, m, m->pindex, wmesg,
1020 nonshared ? VM_ALLOC_SBUSY : 0 , true))
1021 VM_OBJECT_DROP(obj);
1025 * vm_page_busy_sleep_unlocked:
1027 * Sleep if the page is busy, using the page pointer as wchan.
1028 * This is used to implement the hard-path of busying mechanism.
1030 * If nonshared is true, sleep only if the page is xbusy.
1032 * The object lock must not be held on entry. The operation will
1033 * return if the page changes identity.
1036 vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1037 const char *wmesg, bool nonshared)
1040 VM_OBJECT_ASSERT_UNLOCKED(obj);
1041 vm_page_lock_assert(m, MA_NOTOWNED);
1043 _vm_page_busy_sleep(obj, m, pindex, wmesg,
1044 nonshared ? VM_ALLOC_SBUSY : 0, false);
1048 * _vm_page_busy_sleep:
1050 * Internal busy sleep function. Verifies the page identity and
1051 * lockstate against parameters. Returns true if it sleeps and
1054 * If locked is true the lock will be dropped for any true returns
1055 * and held for any false returns.
1058 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1059 const char *wmesg, int allocflags, bool locked)
1065 * If the object is busy we must wait for that to drain to zero
1066 * before trying the page again.
1068 if (obj != NULL && vm_object_busied(obj)) {
1070 VM_OBJECT_DROP(obj);
1071 vm_object_busy_wait(obj, wmesg);
1075 if (!vm_page_busied(m))
1078 xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0;
1080 x = vm_page_busy_fetch(m);
1083 * If the page changes objects or becomes unlocked we can
1086 if (x == VPB_UNBUSIED ||
1087 (xsleep && (x & VPB_BIT_SHARED) != 0) ||
1088 m->object != obj || m->pindex != pindex) {
1092 if ((x & VPB_BIT_WAITERS) != 0)
1094 } while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS));
1096 VM_OBJECT_DROP(obj);
1098 sleepq_add(m, NULL, wmesg, 0, 0);
1099 sleepq_wait(m, PVM);
1107 * Try to shared busy a page.
1108 * If the operation succeeds 1 is returned otherwise 0.
1109 * The operation never sleeps.
1112 vm_page_trysbusy(vm_page_t m)
1118 x = vm_page_busy_fetch(m);
1120 if ((x & VPB_BIT_SHARED) == 0)
1123 * Reduce the window for transient busies that will trigger
1124 * false negatives in vm_page_ps_test().
1126 if (obj != NULL && vm_object_busied(obj))
1128 if (atomic_fcmpset_acq_int(&m->busy_lock, &x,
1129 x + VPB_ONE_SHARER))
1133 /* Refetch the object now that we're guaranteed that it is stable. */
1135 if (obj != NULL && vm_object_busied(obj)) {
1145 * Try to exclusive busy a page.
1146 * If the operation succeeds 1 is returned otherwise 0.
1147 * The operation never sleeps.
1150 vm_page_tryxbusy(vm_page_t m)
1154 if (atomic_cmpset_acq_int(&m->busy_lock, VPB_UNBUSIED,
1155 VPB_CURTHREAD_EXCLUSIVE) == 0)
1159 if (obj != NULL && vm_object_busied(obj)) {
1167 vm_page_xunbusy_hard_tail(vm_page_t m)
1169 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1170 /* Wake the waiter. */
1175 * vm_page_xunbusy_hard:
1177 * Called when unbusy has failed because there is a waiter.
1180 vm_page_xunbusy_hard(vm_page_t m)
1182 vm_page_assert_xbusied(m);
1183 vm_page_xunbusy_hard_tail(m);
1187 vm_page_xunbusy_hard_unchecked(vm_page_t m)
1189 vm_page_assert_xbusied_unchecked(m);
1190 vm_page_xunbusy_hard_tail(m);
1194 vm_page_busy_free(vm_page_t m)
1198 atomic_thread_fence_rel();
1199 x = atomic_swap_int(&m->busy_lock, VPB_FREED);
1200 if ((x & VPB_BIT_WAITERS) != 0)
1205 * vm_page_unhold_pages:
1207 * Unhold each of the pages that is referenced by the given array.
1210 vm_page_unhold_pages(vm_page_t *ma, int count)
1213 for (; count != 0; count--) {
1214 vm_page_unwire(*ma, PQ_ACTIVE);
1220 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1224 #ifdef VM_PHYSSEG_SPARSE
1225 m = vm_phys_paddr_to_vm_page(pa);
1227 m = vm_phys_fictitious_to_vm_page(pa);
1229 #elif defined(VM_PHYSSEG_DENSE)
1233 if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1234 m = &vm_page_array[pi - first_page];
1237 return (vm_phys_fictitious_to_vm_page(pa));
1239 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1246 * Create a fictitious page with the specified physical address and
1247 * memory attribute. The memory attribute is the only the machine-
1248 * dependent aspect of a fictitious page that must be initialized.
1251 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1255 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1256 vm_page_initfake(m, paddr, memattr);
1261 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1264 if ((m->flags & PG_FICTITIOUS) != 0) {
1266 * The page's memattr might have changed since the
1267 * previous initialization. Update the pmap to the
1272 m->phys_addr = paddr;
1273 m->a.queue = PQ_NONE;
1274 /* Fictitious pages don't use "segind". */
1275 m->flags = PG_FICTITIOUS;
1276 /* Fictitious pages don't use "order" or "pool". */
1277 m->oflags = VPO_UNMANAGED;
1278 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
1279 /* Fictitious pages are unevictable. */
1283 pmap_page_set_memattr(m, memattr);
1289 * Release a fictitious page.
1292 vm_page_putfake(vm_page_t m)
1295 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1296 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1297 ("vm_page_putfake: bad page %p", m));
1298 vm_page_assert_xbusied(m);
1299 vm_page_busy_free(m);
1300 uma_zfree(fakepg_zone, m);
1304 * vm_page_updatefake:
1306 * Update the given fictitious page to the specified physical address and
1310 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1313 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1314 ("vm_page_updatefake: bad page %p", m));
1315 m->phys_addr = paddr;
1316 pmap_page_set_memattr(m, memattr);
1325 vm_page_free(vm_page_t m)
1328 m->flags &= ~PG_ZERO;
1329 vm_page_free_toq(m);
1333 * vm_page_free_zero:
1335 * Free a page to the zerod-pages queue
1338 vm_page_free_zero(vm_page_t m)
1341 m->flags |= PG_ZERO;
1342 vm_page_free_toq(m);
1346 * Unbusy and handle the page queueing for a page from a getpages request that
1347 * was optionally read ahead or behind.
1350 vm_page_readahead_finish(vm_page_t m)
1353 /* We shouldn't put invalid pages on queues. */
1354 KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m));
1357 * Since the page is not the actually needed one, whether it should
1358 * be activated or deactivated is not obvious. Empirical results
1359 * have shown that deactivating the page is usually the best choice,
1360 * unless the page is wanted by another thread.
1362 if ((vm_page_busy_fetch(m) & VPB_BIT_WAITERS) != 0)
1363 vm_page_activate(m);
1365 vm_page_deactivate(m);
1366 vm_page_xunbusy_unchecked(m);
1370 * Destroy the identity of an invalid page and free it if possible.
1371 * This is intended to be used when reading a page from backing store fails.
1374 vm_page_free_invalid(vm_page_t m)
1377 KASSERT(vm_page_none_valid(m), ("page %p is valid", m));
1378 KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m));
1379 KASSERT(m->object != NULL, ("page %p has no object", m));
1380 VM_OBJECT_ASSERT_WLOCKED(m->object);
1383 * We may be attempting to free the page as part of the handling for an
1384 * I/O error, in which case the page was xbusied by a different thread.
1386 vm_page_xbusy_claim(m);
1389 * If someone has wired this page while the object lock
1390 * was not held, then the thread that unwires is responsible
1391 * for freeing the page. Otherwise just free the page now.
1392 * The wire count of this unmapped page cannot change while
1393 * we have the page xbusy and the page's object wlocked.
1395 if (vm_page_remove(m))
1400 * vm_page_sleep_if_busy:
1402 * Sleep and release the object lock if the page is busied.
1403 * Returns TRUE if the thread slept.
1405 * The given page must be unlocked and object containing it must
1409 vm_page_sleep_if_busy(vm_page_t m, const char *wmesg)
1413 vm_page_lock_assert(m, MA_NOTOWNED);
1414 VM_OBJECT_ASSERT_WLOCKED(m->object);
1417 * The page-specific object must be cached because page
1418 * identity can change during the sleep, causing the
1419 * re-lock of a different object.
1420 * It is assumed that a reference to the object is already
1421 * held by the callers.
1424 if (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, 0, true)) {
1425 VM_OBJECT_WLOCK(obj);
1432 * vm_page_sleep_if_xbusy:
1434 * Sleep and release the object lock if the page is xbusied.
1435 * Returns TRUE if the thread slept.
1437 * The given page must be unlocked and object containing it must
1441 vm_page_sleep_if_xbusy(vm_page_t m, const char *wmesg)
1445 vm_page_lock_assert(m, MA_NOTOWNED);
1446 VM_OBJECT_ASSERT_WLOCKED(m->object);
1449 * The page-specific object must be cached because page
1450 * identity can change during the sleep, causing the
1451 * re-lock of a different object.
1452 * It is assumed that a reference to the object is already
1453 * held by the callers.
1456 if (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, VM_ALLOC_SBUSY,
1458 VM_OBJECT_WLOCK(obj);
1465 * vm_page_dirty_KBI: [ internal use only ]
1467 * Set all bits in the page's dirty field.
1469 * The object containing the specified page must be locked if the
1470 * call is made from the machine-independent layer.
1472 * See vm_page_clear_dirty_mask().
1474 * This function should only be called by vm_page_dirty().
1477 vm_page_dirty_KBI(vm_page_t m)
1480 /* Refer to this operation by its public name. */
1481 KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!"));
1482 m->dirty = VM_PAGE_BITS_ALL;
1486 * vm_page_insert: [ internal use only ]
1488 * Inserts the given mem entry into the object and object list.
1490 * The object must be locked.
1493 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1497 VM_OBJECT_ASSERT_WLOCKED(object);
1498 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1499 return (vm_page_insert_after(m, object, pindex, mpred));
1503 * vm_page_insert_after:
1505 * Inserts the page "m" into the specified object at offset "pindex".
1507 * The page "mpred" must immediately precede the offset "pindex" within
1508 * the specified object.
1510 * The object must be locked.
1513 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1518 VM_OBJECT_ASSERT_WLOCKED(object);
1519 KASSERT(m->object == NULL,
1520 ("vm_page_insert_after: page already inserted"));
1521 if (mpred != NULL) {
1522 KASSERT(mpred->object == object,
1523 ("vm_page_insert_after: object doesn't contain mpred"));
1524 KASSERT(mpred->pindex < pindex,
1525 ("vm_page_insert_after: mpred doesn't precede pindex"));
1526 msucc = TAILQ_NEXT(mpred, listq);
1528 msucc = TAILQ_FIRST(&object->memq);
1530 KASSERT(msucc->pindex > pindex,
1531 ("vm_page_insert_after: msucc doesn't succeed pindex"));
1534 * Record the object/offset pair in this page.
1538 m->ref_count |= VPRC_OBJREF;
1541 * Now link into the object's ordered list of backed pages.
1543 if (vm_radix_insert(&object->rtree, m)) {
1546 m->ref_count &= ~VPRC_OBJREF;
1549 vm_page_insert_radixdone(m, object, mpred);
1554 * vm_page_insert_radixdone:
1556 * Complete page "m" insertion into the specified object after the
1557 * radix trie hooking.
1559 * The page "mpred" must precede the offset "m->pindex" within the
1562 * The object must be locked.
1565 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1568 VM_OBJECT_ASSERT_WLOCKED(object);
1569 KASSERT(object != NULL && m->object == object,
1570 ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1571 KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1572 ("vm_page_insert_radixdone: page %p is missing object ref", m));
1573 if (mpred != NULL) {
1574 KASSERT(mpred->object == object,
1575 ("vm_page_insert_radixdone: object doesn't contain mpred"));
1576 KASSERT(mpred->pindex < m->pindex,
1577 ("vm_page_insert_radixdone: mpred doesn't precede pindex"));
1581 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1583 TAILQ_INSERT_HEAD(&object->memq, m, listq);
1586 * Show that the object has one more resident page.
1588 object->resident_page_count++;
1591 * Hold the vnode until the last page is released.
1593 if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1594 vhold(object->handle);
1597 * Since we are inserting a new and possibly dirty page,
1598 * update the object's generation count.
1600 if (pmap_page_is_write_mapped(m))
1601 vm_object_set_writeable_dirty(object);
1605 * Do the work to remove a page from its object. The caller is responsible for
1606 * updating the page's fields to reflect this removal.
1609 vm_page_object_remove(vm_page_t m)
1614 vm_page_assert_xbusied(m);
1616 VM_OBJECT_ASSERT_WLOCKED(object);
1617 KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1618 ("page %p is missing its object ref", m));
1620 /* Deferred free of swap space. */
1621 if ((m->a.flags & PGA_SWAP_FREE) != 0)
1622 vm_pager_page_unswapped(m);
1625 mrem = vm_radix_remove(&object->rtree, m->pindex);
1626 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1629 * Now remove from the object's list of backed pages.
1631 TAILQ_REMOVE(&object->memq, m, listq);
1634 * And show that the object has one fewer resident page.
1636 object->resident_page_count--;
1639 * The vnode may now be recycled.
1641 if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1642 vdrop(object->handle);
1648 * Removes the specified page from its containing object, but does not
1649 * invalidate any backing storage. Returns true if the object's reference
1650 * was the last reference to the page, and false otherwise.
1652 * The object must be locked and the page must be exclusively busied.
1653 * The exclusive busy will be released on return. If this is not the
1654 * final ref and the caller does not hold a wire reference it may not
1655 * continue to access the page.
1658 vm_page_remove(vm_page_t m)
1662 dropped = vm_page_remove_xbusy(m);
1669 * vm_page_remove_xbusy
1671 * Removes the page but leaves the xbusy held. Returns true if this
1672 * removed the final ref and false otherwise.
1675 vm_page_remove_xbusy(vm_page_t m)
1678 vm_page_object_remove(m);
1679 return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1685 * Returns the page associated with the object/offset
1686 * pair specified; if none is found, NULL is returned.
1688 * The object must be locked.
1691 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1694 VM_OBJECT_ASSERT_LOCKED(object);
1695 return (vm_radix_lookup(&object->rtree, pindex));
1699 * vm_page_lookup_unlocked:
1701 * Returns the page associated with the object/offset pair specified;
1702 * if none is found, NULL is returned. The page may be no longer be
1703 * present in the object at the time that this function returns. Only
1704 * useful for opportunistic checks such as inmem().
1707 vm_page_lookup_unlocked(vm_object_t object, vm_pindex_t pindex)
1710 return (vm_radix_lookup_unlocked(&object->rtree, pindex));
1716 * Returns a page that must already have been busied by
1717 * the caller. Used for bogus page replacement.
1720 vm_page_relookup(vm_object_t object, vm_pindex_t pindex)
1724 m = vm_radix_lookup_unlocked(&object->rtree, pindex);
1725 KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) &&
1726 m->object == object && m->pindex == pindex,
1727 ("vm_page_relookup: Invalid page %p", m));
1732 * This should only be used by lockless functions for releasing transient
1733 * incorrect acquires. The page may have been freed after we acquired a
1734 * busy lock. In this case busy_lock == VPB_FREED and we have nothing
1738 vm_page_busy_release(vm_page_t m)
1742 x = vm_page_busy_fetch(m);
1746 if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) {
1747 if (atomic_fcmpset_int(&m->busy_lock, &x,
1748 x - VPB_ONE_SHARER))
1752 KASSERT((x & VPB_BIT_SHARED) != 0 ||
1753 (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE,
1754 ("vm_page_busy_release: %p xbusy not owned.", m));
1755 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1757 if ((x & VPB_BIT_WAITERS) != 0)
1764 * vm_page_find_least:
1766 * Returns the page associated with the object with least pindex
1767 * greater than or equal to the parameter pindex, or NULL.
1769 * The object must be locked.
1772 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1776 VM_OBJECT_ASSERT_LOCKED(object);
1777 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1778 m = vm_radix_lookup_ge(&object->rtree, pindex);
1783 * Returns the given page's successor (by pindex) within the object if it is
1784 * resident; if none is found, NULL is returned.
1786 * The object must be locked.
1789 vm_page_next(vm_page_t m)
1793 VM_OBJECT_ASSERT_LOCKED(m->object);
1794 if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1795 MPASS(next->object == m->object);
1796 if (next->pindex != m->pindex + 1)
1803 * Returns the given page's predecessor (by pindex) within the object if it is
1804 * resident; if none is found, NULL is returned.
1806 * The object must be locked.
1809 vm_page_prev(vm_page_t m)
1813 VM_OBJECT_ASSERT_LOCKED(m->object);
1814 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1815 MPASS(prev->object == m->object);
1816 if (prev->pindex != m->pindex - 1)
1823 * Uses the page mnew as a replacement for an existing page at index
1824 * pindex which must be already present in the object.
1826 * Both pages must be exclusively busied on enter. The old page is
1829 * A return value of true means mold is now free. If this is not the
1830 * final ref and the caller does not hold a wire reference it may not
1831 * continue to access the page.
1834 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1840 VM_OBJECT_ASSERT_WLOCKED(object);
1841 vm_page_assert_xbusied(mold);
1842 KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
1843 ("vm_page_replace: page %p already in object", mnew));
1846 * This function mostly follows vm_page_insert() and
1847 * vm_page_remove() without the radix, object count and vnode
1848 * dance. Double check such functions for more comments.
1851 mnew->object = object;
1852 mnew->pindex = pindex;
1853 atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
1854 mret = vm_radix_replace(&object->rtree, mnew);
1855 KASSERT(mret == mold,
1856 ("invalid page replacement, mold=%p, mret=%p", mold, mret));
1857 KASSERT((mold->oflags & VPO_UNMANAGED) ==
1858 (mnew->oflags & VPO_UNMANAGED),
1859 ("vm_page_replace: mismatched VPO_UNMANAGED"));
1861 /* Keep the resident page list in sorted order. */
1862 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1863 TAILQ_REMOVE(&object->memq, mold, listq);
1864 mold->object = NULL;
1867 * The object's resident_page_count does not change because we have
1868 * swapped one page for another, but the generation count should
1869 * change if the page is dirty.
1871 if (pmap_page_is_write_mapped(mnew))
1872 vm_object_set_writeable_dirty(object);
1873 dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF;
1874 vm_page_xunbusy(mold);
1880 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1884 vm_page_assert_xbusied(mnew);
1886 if (vm_page_replace_hold(mnew, object, pindex, mold))
1893 * Move the given memory entry from its
1894 * current object to the specified target object/offset.
1896 * Note: swap associated with the page must be invalidated by the move. We
1897 * have to do this for several reasons: (1) we aren't freeing the
1898 * page, (2) we are dirtying the page, (3) the VM system is probably
1899 * moving the page from object A to B, and will then later move
1900 * the backing store from A to B and we can't have a conflict.
1902 * Note: we *always* dirty the page. It is necessary both for the
1903 * fact that we moved it, and because we may be invalidating
1906 * The objects must be locked.
1909 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1914 VM_OBJECT_ASSERT_WLOCKED(new_object);
1916 KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m));
1917 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1918 KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1919 ("vm_page_rename: pindex already renamed"));
1922 * Create a custom version of vm_page_insert() which does not depend
1923 * by m_prev and can cheat on the implementation aspects of the
1927 m->pindex = new_pindex;
1928 if (vm_radix_insert(&new_object->rtree, m)) {
1934 * The operation cannot fail anymore. The removal must happen before
1935 * the listq iterator is tainted.
1938 vm_page_object_remove(m);
1940 /* Return back to the new pindex to complete vm_page_insert(). */
1941 m->pindex = new_pindex;
1942 m->object = new_object;
1944 vm_page_insert_radixdone(m, new_object, mpred);
1952 * Allocate and return a page that is associated with the specified
1953 * object and offset pair. By default, this page is exclusive busied.
1955 * The caller must always specify an allocation class.
1957 * allocation classes:
1958 * VM_ALLOC_NORMAL normal process request
1959 * VM_ALLOC_SYSTEM system *really* needs a page
1960 * VM_ALLOC_INTERRUPT interrupt time request
1962 * optional allocation flags:
1963 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1964 * intends to allocate
1965 * VM_ALLOC_NOBUSY do not exclusive busy the page
1966 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1967 * VM_ALLOC_NOOBJ page is not associated with an object and
1968 * should not be exclusive busy
1969 * VM_ALLOC_SBUSY shared busy the allocated page
1970 * VM_ALLOC_WIRED wire the allocated page
1971 * VM_ALLOC_ZERO prefer a zeroed page
1974 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1977 return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1978 vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1982 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1986 return (vm_page_alloc_domain_after(object, pindex, domain, req,
1987 object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) :
1992 * Allocate a page in the specified object with the given page index. To
1993 * optimize insertion of the page into the object, the caller must also specifiy
1994 * the resident page in the object with largest index smaller than the given
1995 * page index, or NULL if no such page exists.
1998 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1999 int req, vm_page_t mpred)
2001 struct vm_domainset_iter di;
2005 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2007 m = vm_page_alloc_domain_after(object, pindex, domain, req,
2011 } while (vm_domainset_iter_page(&di, object, &domain) == 0);
2017 * Returns true if the number of free pages exceeds the minimum
2018 * for the request class and false otherwise.
2021 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
2023 u_int limit, old, new;
2025 if (req_class == VM_ALLOC_INTERRUPT)
2027 else if (req_class == VM_ALLOC_SYSTEM)
2028 limit = vmd->vmd_interrupt_free_min;
2030 limit = vmd->vmd_free_reserved;
2033 * Attempt to reserve the pages. Fail if we're below the limit.
2036 old = vmd->vmd_free_count;
2041 } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
2043 /* Wake the page daemon if we've crossed the threshold. */
2044 if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
2045 pagedaemon_wakeup(vmd->vmd_domain);
2047 /* Only update bitsets on transitions. */
2048 if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
2049 (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
2056 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
2061 * The page daemon is allowed to dig deeper into the free page list.
2063 req_class = req & VM_ALLOC_CLASS_MASK;
2064 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2065 req_class = VM_ALLOC_SYSTEM;
2066 return (_vm_domain_allocate(vmd, req_class, npages));
2070 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
2071 int req, vm_page_t mpred)
2073 struct vm_domain *vmd;
2077 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
2078 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
2079 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2080 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2081 ("inconsistent object(%p)/req(%x)", object, req));
2082 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2083 ("Can't sleep and retry object insertion."));
2084 KASSERT(mpred == NULL || mpred->pindex < pindex,
2085 ("mpred %p doesn't precede pindex 0x%jx", mpred,
2086 (uintmax_t)pindex));
2088 VM_OBJECT_ASSERT_WLOCKED(object);
2092 pool = object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT;
2094 #if VM_NRESERVLEVEL > 0
2096 * Can we allocate the page from a reservation?
2098 if (vm_object_reserv(object) &&
2099 (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
2104 vmd = VM_DOMAIN(domain);
2105 if (vmd->vmd_pgcache[pool].zone != NULL) {
2106 m = uma_zalloc(vmd->vmd_pgcache[pool].zone, M_NOWAIT | M_NOVM);
2108 flags |= PG_PCPU_CACHE;
2112 if (vm_domain_allocate(vmd, req, 1)) {
2114 * If not, allocate it from the free page queues.
2116 vm_domain_free_lock(vmd);
2117 m = vm_phys_alloc_pages(domain, pool, 0);
2118 vm_domain_free_unlock(vmd);
2120 vm_domain_freecnt_inc(vmd, 1);
2121 #if VM_NRESERVLEVEL > 0
2122 if (vm_reserv_reclaim_inactive(domain))
2129 * Not allocatable, give up.
2131 if (vm_domain_alloc_fail(vmd, object, req))
2137 * At this point we had better have found a good page.
2141 vm_page_alloc_check(m);
2144 * Initialize the page. Only the PG_ZERO flag is inherited.
2146 if ((req & VM_ALLOC_ZERO) != 0)
2147 flags |= (m->flags & PG_ZERO);
2148 if ((req & VM_ALLOC_NODUMP) != 0)
2152 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2154 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2155 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2156 else if ((req & VM_ALLOC_SBUSY) != 0)
2157 m->busy_lock = VPB_SHARERS_WORD(1);
2159 m->busy_lock = VPB_UNBUSIED;
2160 if (req & VM_ALLOC_WIRED) {
2166 if (object != NULL) {
2167 if (vm_page_insert_after(m, object, pindex, mpred)) {
2168 if (req & VM_ALLOC_WIRED) {
2172 KASSERT(m->object == NULL, ("page %p has object", m));
2173 m->oflags = VPO_UNMANAGED;
2174 m->busy_lock = VPB_UNBUSIED;
2175 /* Don't change PG_ZERO. */
2176 vm_page_free_toq(m);
2177 if (req & VM_ALLOC_WAITFAIL) {
2178 VM_OBJECT_WUNLOCK(object);
2180 VM_OBJECT_WLOCK(object);
2185 /* Ignore device objects; the pager sets "memattr" for them. */
2186 if (object->memattr != VM_MEMATTR_DEFAULT &&
2187 (object->flags & OBJ_FICTITIOUS) == 0)
2188 pmap_page_set_memattr(m, object->memattr);
2196 * vm_page_alloc_contig:
2198 * Allocate a contiguous set of physical pages of the given size "npages"
2199 * from the free lists. All of the physical pages must be at or above
2200 * the given physical address "low" and below the given physical address
2201 * "high". The given value "alignment" determines the alignment of the
2202 * first physical page in the set. If the given value "boundary" is
2203 * non-zero, then the set of physical pages cannot cross any physical
2204 * address boundary that is a multiple of that value. Both "alignment"
2205 * and "boundary" must be a power of two.
2207 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
2208 * then the memory attribute setting for the physical pages is configured
2209 * to the object's memory attribute setting. Otherwise, the memory
2210 * attribute setting for the physical pages is configured to "memattr",
2211 * overriding the object's memory attribute setting. However, if the
2212 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
2213 * memory attribute setting for the physical pages cannot be configured
2214 * to VM_MEMATTR_DEFAULT.
2216 * The specified object may not contain fictitious pages.
2218 * The caller must always specify an allocation class.
2220 * allocation classes:
2221 * VM_ALLOC_NORMAL normal process request
2222 * VM_ALLOC_SYSTEM system *really* needs a page
2223 * VM_ALLOC_INTERRUPT interrupt time request
2225 * optional allocation flags:
2226 * VM_ALLOC_NOBUSY do not exclusive busy the page
2227 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
2228 * VM_ALLOC_NOOBJ page is not associated with an object and
2229 * should not be exclusive busy
2230 * VM_ALLOC_SBUSY shared busy the allocated page
2231 * VM_ALLOC_WIRED wire the allocated page
2232 * VM_ALLOC_ZERO prefer a zeroed page
2235 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
2236 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2237 vm_paddr_t boundary, vm_memattr_t memattr)
2239 struct vm_domainset_iter di;
2243 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2245 m = vm_page_alloc_contig_domain(object, pindex, domain, req,
2246 npages, low, high, alignment, boundary, memattr);
2249 } while (vm_domainset_iter_page(&di, object, &domain) == 0);
2255 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2256 int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2257 vm_paddr_t boundary, vm_memattr_t memattr)
2259 struct vm_domain *vmd;
2260 vm_page_t m, m_ret, mpred;
2261 u_int busy_lock, flags, oflags;
2263 mpred = NULL; /* XXX: pacify gcc */
2264 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
2265 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
2266 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2267 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2268 ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
2270 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2271 ("Can't sleep and retry object insertion."));
2272 if (object != NULL) {
2273 VM_OBJECT_ASSERT_WLOCKED(object);
2274 KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2275 ("vm_page_alloc_contig: object %p has fictitious pages",
2278 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2280 if (object != NULL) {
2281 mpred = vm_radix_lookup_le(&object->rtree, pindex);
2282 KASSERT(mpred == NULL || mpred->pindex != pindex,
2283 ("vm_page_alloc_contig: pindex already allocated"));
2287 * Can we allocate the pages without the number of free pages falling
2288 * below the lower bound for the allocation class?
2292 #if VM_NRESERVLEVEL > 0
2294 * Can we allocate the pages from a reservation?
2296 if (vm_object_reserv(object) &&
2297 (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
2298 mpred, npages, low, high, alignment, boundary)) != NULL) {
2302 vmd = VM_DOMAIN(domain);
2303 if (vm_domain_allocate(vmd, req, npages)) {
2305 * allocate them from the free page queues.
2307 vm_domain_free_lock(vmd);
2308 m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2309 alignment, boundary);
2310 vm_domain_free_unlock(vmd);
2311 if (m_ret == NULL) {
2312 vm_domain_freecnt_inc(vmd, npages);
2313 #if VM_NRESERVLEVEL > 0
2314 if (vm_reserv_reclaim_contig(domain, npages, low,
2315 high, alignment, boundary))
2320 if (m_ret == NULL) {
2321 if (vm_domain_alloc_fail(vmd, object, req))
2325 #if VM_NRESERVLEVEL > 0
2328 for (m = m_ret; m < &m_ret[npages]; m++) {
2330 vm_page_alloc_check(m);
2334 * Initialize the pages. Only the PG_ZERO flag is inherited.
2337 if ((req & VM_ALLOC_ZERO) != 0)
2339 if ((req & VM_ALLOC_NODUMP) != 0)
2341 oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2343 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2344 busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2345 else if ((req & VM_ALLOC_SBUSY) != 0)
2346 busy_lock = VPB_SHARERS_WORD(1);
2348 busy_lock = VPB_UNBUSIED;
2349 if ((req & VM_ALLOC_WIRED) != 0)
2350 vm_wire_add(npages);
2351 if (object != NULL) {
2352 if (object->memattr != VM_MEMATTR_DEFAULT &&
2353 memattr == VM_MEMATTR_DEFAULT)
2354 memattr = object->memattr;
2356 for (m = m_ret; m < &m_ret[npages]; m++) {
2358 m->flags = (m->flags | PG_NODUMP) & flags;
2359 m->busy_lock = busy_lock;
2360 if ((req & VM_ALLOC_WIRED) != 0)
2364 if (object != NULL) {
2365 if (vm_page_insert_after(m, object, pindex, mpred)) {
2366 if ((req & VM_ALLOC_WIRED) != 0)
2367 vm_wire_sub(npages);
2368 KASSERT(m->object == NULL,
2369 ("page %p has object", m));
2371 for (m = m_ret; m < &m_ret[npages]; m++) {
2373 (req & VM_ALLOC_WIRED) != 0)
2375 m->oflags = VPO_UNMANAGED;
2376 m->busy_lock = VPB_UNBUSIED;
2377 /* Don't change PG_ZERO. */
2378 vm_page_free_toq(m);
2380 if (req & VM_ALLOC_WAITFAIL) {
2381 VM_OBJECT_WUNLOCK(object);
2383 VM_OBJECT_WLOCK(object);
2390 if (memattr != VM_MEMATTR_DEFAULT)
2391 pmap_page_set_memattr(m, memattr);
2398 * Check a page that has been freshly dequeued from a freelist.
2401 vm_page_alloc_check(vm_page_t m)
2404 KASSERT(m->object == NULL, ("page %p has object", m));
2405 KASSERT(m->a.queue == PQ_NONE &&
2406 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
2407 ("page %p has unexpected queue %d, flags %#x",
2408 m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK)));
2409 KASSERT(m->ref_count == 0, ("page %p has references", m));
2410 KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m));
2411 KASSERT(m->dirty == 0, ("page %p is dirty", m));
2412 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2413 ("page %p has unexpected memattr %d",
2414 m, pmap_page_get_memattr(m)));
2415 KASSERT(m->valid == 0, ("free page %p is valid", m));
2416 pmap_vm_page_alloc_check(m);
2420 * vm_page_alloc_freelist:
2422 * Allocate a physical page from the specified free page list.
2424 * The caller must always specify an allocation class.
2426 * allocation classes:
2427 * VM_ALLOC_NORMAL normal process request
2428 * VM_ALLOC_SYSTEM system *really* needs a page
2429 * VM_ALLOC_INTERRUPT interrupt time request
2431 * optional allocation flags:
2432 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
2433 * intends to allocate
2434 * VM_ALLOC_WIRED wire the allocated page
2435 * VM_ALLOC_ZERO prefer a zeroed page
2438 vm_page_alloc_freelist(int freelist, int req)
2440 struct vm_domainset_iter di;
2444 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2446 m = vm_page_alloc_freelist_domain(domain, freelist, req);
2449 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2455 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2457 struct vm_domain *vmd;
2462 vmd = VM_DOMAIN(domain);
2464 if (vm_domain_allocate(vmd, req, 1)) {
2465 vm_domain_free_lock(vmd);
2466 m = vm_phys_alloc_freelist_pages(domain, freelist,
2467 VM_FREEPOOL_DIRECT, 0);
2468 vm_domain_free_unlock(vmd);
2470 vm_domain_freecnt_inc(vmd, 1);
2473 if (vm_domain_alloc_fail(vmd, NULL, req))
2478 vm_page_alloc_check(m);
2481 * Initialize the page. Only the PG_ZERO flag is inherited.
2485 if ((req & VM_ALLOC_ZERO) != 0)
2488 if ((req & VM_ALLOC_WIRED) != 0) {
2492 /* Unmanaged pages don't use "act_count". */
2493 m->oflags = VPO_UNMANAGED;
2498 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2500 struct vm_domain *vmd;
2501 struct vm_pgcache *pgcache;
2505 vmd = VM_DOMAIN(pgcache->domain);
2508 * The page daemon should avoid creating extra memory pressure since its
2509 * main purpose is to replenish the store of free pages.
2511 if (vmd->vmd_severeset || curproc == pageproc ||
2512 !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2514 domain = vmd->vmd_domain;
2515 vm_domain_free_lock(vmd);
2516 i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2517 (vm_page_t *)store);
2518 vm_domain_free_unlock(vmd);
2520 vm_domain_freecnt_inc(vmd, cnt - i);
2526 vm_page_zone_release(void *arg, void **store, int cnt)
2528 struct vm_domain *vmd;
2529 struct vm_pgcache *pgcache;
2534 vmd = VM_DOMAIN(pgcache->domain);
2535 vm_domain_free_lock(vmd);
2536 for (i = 0; i < cnt; i++) {
2537 m = (vm_page_t)store[i];
2538 vm_phys_free_pages(m, 0);
2540 vm_domain_free_unlock(vmd);
2541 vm_domain_freecnt_inc(vmd, cnt);
2544 #define VPSC_ANY 0 /* No restrictions. */
2545 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */
2546 #define VPSC_NOSUPER 2 /* Skip superpages. */
2549 * vm_page_scan_contig:
2551 * Scan vm_page_array[] between the specified entries "m_start" and
2552 * "m_end" for a run of contiguous physical pages that satisfy the
2553 * specified conditions, and return the lowest page in the run. The
2554 * specified "alignment" determines the alignment of the lowest physical
2555 * page in the run. If the specified "boundary" is non-zero, then the
2556 * run of physical pages cannot span a physical address that is a
2557 * multiple of "boundary".
2559 * "m_end" is never dereferenced, so it need not point to a vm_page
2560 * structure within vm_page_array[].
2562 * "npages" must be greater than zero. "m_start" and "m_end" must not
2563 * span a hole (or discontiguity) in the physical address space. Both
2564 * "alignment" and "boundary" must be a power of two.
2567 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2568 u_long alignment, vm_paddr_t boundary, int options)
2573 #if VM_NRESERVLEVEL > 0
2576 int m_inc, order, run_ext, run_len;
2578 KASSERT(npages > 0, ("npages is 0"));
2579 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2580 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2583 for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2584 KASSERT((m->flags & PG_MARKER) == 0,
2585 ("page %p is PG_MARKER", m));
2586 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1,
2587 ("fictitious page %p has invalid ref count", m));
2590 * If the current page would be the start of a run, check its
2591 * physical address against the end, alignment, and boundary
2592 * conditions. If it doesn't satisfy these conditions, either
2593 * terminate the scan or advance to the next page that
2594 * satisfies the failed condition.
2597 KASSERT(m_run == NULL, ("m_run != NULL"));
2598 if (m + npages > m_end)
2600 pa = VM_PAGE_TO_PHYS(m);
2601 if ((pa & (alignment - 1)) != 0) {
2602 m_inc = atop(roundup2(pa, alignment) - pa);
2605 if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2607 m_inc = atop(roundup2(pa, boundary) - pa);
2611 KASSERT(m_run != NULL, ("m_run == NULL"));
2615 if (vm_page_wired(m))
2617 #if VM_NRESERVLEVEL > 0
2618 else if ((level = vm_reserv_level(m)) >= 0 &&
2619 (options & VPSC_NORESERV) != 0) {
2621 /* Advance to the end of the reservation. */
2622 pa = VM_PAGE_TO_PHYS(m);
2623 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2627 else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2629 * The page is considered eligible for relocation if
2630 * and only if it could be laundered or reclaimed by
2633 VM_OBJECT_RLOCK(object);
2634 if (object != m->object) {
2635 VM_OBJECT_RUNLOCK(object);
2638 /* Don't care: PG_NODUMP, PG_ZERO. */
2639 if (object->type != OBJT_DEFAULT &&
2640 (object->flags & OBJ_SWAP) == 0 &&
2641 object->type != OBJT_VNODE) {
2643 #if VM_NRESERVLEVEL > 0
2644 } else if ((options & VPSC_NOSUPER) != 0 &&
2645 (level = vm_reserv_level_iffullpop(m)) >= 0) {
2647 /* Advance to the end of the superpage. */
2648 pa = VM_PAGE_TO_PHYS(m);
2649 m_inc = atop(roundup2(pa + 1,
2650 vm_reserv_size(level)) - pa);
2652 } else if (object->memattr == VM_MEMATTR_DEFAULT &&
2653 vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) {
2655 * The page is allocated but eligible for
2656 * relocation. Extend the current run by one
2659 KASSERT(pmap_page_get_memattr(m) ==
2661 ("page %p has an unexpected memattr", m));
2662 KASSERT((m->oflags & (VPO_SWAPINPROG |
2663 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2664 ("page %p has unexpected oflags", m));
2665 /* Don't care: PGA_NOSYNC. */
2669 VM_OBJECT_RUNLOCK(object);
2670 #if VM_NRESERVLEVEL > 0
2671 } else if (level >= 0) {
2673 * The page is reserved but not yet allocated. In
2674 * other words, it is still free. Extend the current
2679 } else if ((order = m->order) < VM_NFREEORDER) {
2681 * The page is enqueued in the physical memory
2682 * allocator's free page queues. Moreover, it is the
2683 * first page in a power-of-two-sized run of
2684 * contiguous free pages. Add these pages to the end
2685 * of the current run, and jump ahead.
2687 run_ext = 1 << order;
2691 * Skip the page for one of the following reasons: (1)
2692 * It is enqueued in the physical memory allocator's
2693 * free page queues. However, it is not the first
2694 * page in a run of contiguous free pages. (This case
2695 * rarely occurs because the scan is performed in
2696 * ascending order.) (2) It is not reserved, and it is
2697 * transitioning from free to allocated. (Conversely,
2698 * the transition from allocated to free for managed
2699 * pages is blocked by the page busy lock.) (3) It is
2700 * allocated but not contained by an object and not
2701 * wired, e.g., allocated by Xen's balloon driver.
2707 * Extend or reset the current run of pages.
2720 if (run_len >= npages)
2726 * vm_page_reclaim_run:
2728 * Try to relocate each of the allocated virtual pages within the
2729 * specified run of physical pages to a new physical address. Free the
2730 * physical pages underlying the relocated virtual pages. A virtual page
2731 * is relocatable if and only if it could be laundered or reclaimed by
2732 * the page daemon. Whenever possible, a virtual page is relocated to a
2733 * physical address above "high".
2735 * Returns 0 if every physical page within the run was already free or
2736 * just freed by a successful relocation. Otherwise, returns a non-zero
2737 * value indicating why the last attempt to relocate a virtual page was
2740 * "req_class" must be an allocation class.
2743 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2746 struct vm_domain *vmd;
2747 struct spglist free;
2750 vm_page_t m, m_end, m_new;
2751 int error, order, req;
2753 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2754 ("req_class is not an allocation class"));
2758 m_end = m_run + npages;
2759 for (; error == 0 && m < m_end; m++) {
2760 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2761 ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2764 * Racily check for wirings. Races are handled once the object
2765 * lock is held and the page is unmapped.
2767 if (vm_page_wired(m))
2769 else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2771 * The page is relocated if and only if it could be
2772 * laundered or reclaimed by the page daemon.
2774 VM_OBJECT_WLOCK(object);
2775 /* Don't care: PG_NODUMP, PG_ZERO. */
2776 if (m->object != object ||
2777 (object->type != OBJT_DEFAULT &&
2778 (object->flags & OBJ_SWAP) == 0 &&
2779 object->type != OBJT_VNODE))
2781 else if (object->memattr != VM_MEMATTR_DEFAULT)
2783 else if (vm_page_queue(m) != PQ_NONE &&
2784 vm_page_tryxbusy(m) != 0) {
2785 if (vm_page_wired(m)) {
2790 KASSERT(pmap_page_get_memattr(m) ==
2792 ("page %p has an unexpected memattr", m));
2793 KASSERT(m->oflags == 0,
2794 ("page %p has unexpected oflags", m));
2795 /* Don't care: PGA_NOSYNC. */
2796 if (!vm_page_none_valid(m)) {
2798 * First, try to allocate a new page
2799 * that is above "high". Failing
2800 * that, try to allocate a new page
2801 * that is below "m_run". Allocate
2802 * the new page between the end of
2803 * "m_run" and "high" only as a last
2806 req = req_class | VM_ALLOC_NOOBJ;
2807 if ((m->flags & PG_NODUMP) != 0)
2808 req |= VM_ALLOC_NODUMP;
2809 if (trunc_page(high) !=
2810 ~(vm_paddr_t)PAGE_MASK) {
2811 m_new = vm_page_alloc_contig(
2816 VM_MEMATTR_DEFAULT);
2819 if (m_new == NULL) {
2820 pa = VM_PAGE_TO_PHYS(m_run);
2821 m_new = vm_page_alloc_contig(
2823 0, pa - 1, PAGE_SIZE, 0,
2824 VM_MEMATTR_DEFAULT);
2826 if (m_new == NULL) {
2828 m_new = vm_page_alloc_contig(
2830 pa, high, PAGE_SIZE, 0,
2831 VM_MEMATTR_DEFAULT);
2833 if (m_new == NULL) {
2840 * Unmap the page and check for new
2841 * wirings that may have been acquired
2842 * through a pmap lookup.
2844 if (object->ref_count != 0 &&
2845 !vm_page_try_remove_all(m)) {
2847 vm_page_free(m_new);
2853 * Replace "m" with the new page. For
2854 * vm_page_replace(), "m" must be busy
2855 * and dequeued. Finally, change "m"
2856 * as if vm_page_free() was called.
2858 m_new->a.flags = m->a.flags &
2859 ~PGA_QUEUE_STATE_MASK;
2860 KASSERT(m_new->oflags == VPO_UNMANAGED,
2861 ("page %p is managed", m_new));
2863 pmap_copy_page(m, m_new);
2864 m_new->valid = m->valid;
2865 m_new->dirty = m->dirty;
2866 m->flags &= ~PG_ZERO;
2868 if (vm_page_replace_hold(m_new, object,
2870 vm_page_free_prep(m))
2871 SLIST_INSERT_HEAD(&free, m,
2875 * The new page must be deactivated
2876 * before the object is unlocked.
2878 vm_page_deactivate(m_new);
2880 m->flags &= ~PG_ZERO;
2882 if (vm_page_free_prep(m))
2883 SLIST_INSERT_HEAD(&free, m,
2885 KASSERT(m->dirty == 0,
2886 ("page %p is dirty", m));
2891 VM_OBJECT_WUNLOCK(object);
2893 MPASS(vm_page_domain(m) == domain);
2894 vmd = VM_DOMAIN(domain);
2895 vm_domain_free_lock(vmd);
2897 if (order < VM_NFREEORDER) {
2899 * The page is enqueued in the physical memory
2900 * allocator's free page queues. Moreover, it
2901 * is the first page in a power-of-two-sized
2902 * run of contiguous free pages. Jump ahead
2903 * to the last page within that run, and
2904 * continue from there.
2906 m += (1 << order) - 1;
2908 #if VM_NRESERVLEVEL > 0
2909 else if (vm_reserv_is_page_free(m))
2912 vm_domain_free_unlock(vmd);
2913 if (order == VM_NFREEORDER)
2917 if ((m = SLIST_FIRST(&free)) != NULL) {
2920 vmd = VM_DOMAIN(domain);
2922 vm_domain_free_lock(vmd);
2924 MPASS(vm_page_domain(m) == domain);
2925 SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2926 vm_phys_free_pages(m, 0);
2928 } while ((m = SLIST_FIRST(&free)) != NULL);
2929 vm_domain_free_unlock(vmd);
2930 vm_domain_freecnt_inc(vmd, cnt);
2937 CTASSERT(powerof2(NRUNS));
2939 #define RUN_INDEX(count) ((count) & (NRUNS - 1))
2941 #define MIN_RECLAIM 8
2944 * vm_page_reclaim_contig:
2946 * Reclaim allocated, contiguous physical memory satisfying the specified
2947 * conditions by relocating the virtual pages using that physical memory.
2948 * Returns true if reclamation is successful and false otherwise. Since
2949 * relocation requires the allocation of physical pages, reclamation may
2950 * fail due to a shortage of free pages. When reclamation fails, callers
2951 * are expected to perform vm_wait() before retrying a failed allocation
2952 * operation, e.g., vm_page_alloc_contig().
2954 * The caller must always specify an allocation class through "req".
2956 * allocation classes:
2957 * VM_ALLOC_NORMAL normal process request
2958 * VM_ALLOC_SYSTEM system *really* needs a page
2959 * VM_ALLOC_INTERRUPT interrupt time request
2961 * The optional allocation flags are ignored.
2963 * "npages" must be greater than zero. Both "alignment" and "boundary"
2964 * must be a power of two.
2967 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
2968 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2970 struct vm_domain *vmd;
2971 vm_paddr_t curr_low;
2972 vm_page_t m_run, m_runs[NRUNS];
2973 u_long count, minalign, reclaimed;
2974 int error, i, options, req_class;
2976 KASSERT(npages > 0, ("npages is 0"));
2977 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2978 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2981 * The caller will attempt an allocation after some runs have been
2982 * reclaimed and added to the vm_phys buddy lists. Due to limitations
2983 * of vm_phys_alloc_contig(), round up the requested length to the next
2984 * power of two or maximum chunk size, and ensure that each run is
2987 minalign = 1ul << imin(flsl(npages - 1), VM_NFREEORDER - 1);
2988 npages = roundup2(npages, minalign);
2989 if (alignment < ptoa(minalign))
2990 alignment = ptoa(minalign);
2993 * The page daemon is allowed to dig deeper into the free page list.
2995 req_class = req & VM_ALLOC_CLASS_MASK;
2996 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2997 req_class = VM_ALLOC_SYSTEM;
3000 * Return if the number of free pages cannot satisfy the requested
3003 vmd = VM_DOMAIN(domain);
3004 count = vmd->vmd_free_count;
3005 if (count < npages + vmd->vmd_free_reserved || (count < npages +
3006 vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
3007 (count < npages && req_class == VM_ALLOC_INTERRUPT))
3011 * Scan up to three times, relaxing the restrictions ("options") on
3012 * the reclamation of reservations and superpages each time.
3014 for (options = VPSC_NORESERV;;) {
3016 * Find the highest runs that satisfy the given constraints
3017 * and restrictions, and record them in "m_runs".
3022 m_run = vm_phys_scan_contig(domain, npages, curr_low,
3023 high, alignment, boundary, options);
3026 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
3027 m_runs[RUN_INDEX(count)] = m_run;
3032 * Reclaim the highest runs in LIFO (descending) order until
3033 * the number of reclaimed pages, "reclaimed", is at least
3034 * MIN_RECLAIM. Reset "reclaimed" each time because each
3035 * reclamation is idempotent, and runs will (likely) recur
3036 * from one scan to the next as restrictions are relaxed.
3039 for (i = 0; count > 0 && i < NRUNS; i++) {
3041 m_run = m_runs[RUN_INDEX(count)];
3042 error = vm_page_reclaim_run(req_class, domain, npages,
3045 reclaimed += npages;
3046 if (reclaimed >= MIN_RECLAIM)
3052 * Either relax the restrictions on the next scan or return if
3053 * the last scan had no restrictions.
3055 if (options == VPSC_NORESERV)
3056 options = VPSC_NOSUPER;
3057 else if (options == VPSC_NOSUPER)
3059 else if (options == VPSC_ANY)
3060 return (reclaimed != 0);
3065 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
3066 u_long alignment, vm_paddr_t boundary)
3068 struct vm_domainset_iter di;
3072 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
3074 ret = vm_page_reclaim_contig_domain(domain, req, npages, low,
3075 high, alignment, boundary);
3078 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
3084 * Set the domain in the appropriate page level domainset.
3087 vm_domain_set(struct vm_domain *vmd)
3090 mtx_lock(&vm_domainset_lock);
3091 if (!vmd->vmd_minset && vm_paging_min(vmd)) {
3092 vmd->vmd_minset = 1;
3093 DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
3095 if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
3096 vmd->vmd_severeset = 1;
3097 DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
3099 mtx_unlock(&vm_domainset_lock);
3103 * Clear the domain from the appropriate page level domainset.
3106 vm_domain_clear(struct vm_domain *vmd)
3109 mtx_lock(&vm_domainset_lock);
3110 if (vmd->vmd_minset && !vm_paging_min(vmd)) {
3111 vmd->vmd_minset = 0;
3112 DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
3113 if (vm_min_waiters != 0) {
3115 wakeup(&vm_min_domains);
3118 if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
3119 vmd->vmd_severeset = 0;
3120 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
3121 if (vm_severe_waiters != 0) {
3122 vm_severe_waiters = 0;
3123 wakeup(&vm_severe_domains);
3128 * If pageout daemon needs pages, then tell it that there are
3131 if (vmd->vmd_pageout_pages_needed &&
3132 vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
3133 wakeup(&vmd->vmd_pageout_pages_needed);
3134 vmd->vmd_pageout_pages_needed = 0;
3137 /* See comments in vm_wait_doms(). */
3138 if (vm_pageproc_waiters) {
3139 vm_pageproc_waiters = 0;
3140 wakeup(&vm_pageproc_waiters);
3142 mtx_unlock(&vm_domainset_lock);
3146 * Wait for free pages to exceed the min threshold globally.
3152 mtx_lock(&vm_domainset_lock);
3153 while (vm_page_count_min()) {
3155 msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
3157 mtx_unlock(&vm_domainset_lock);
3161 * Wait for free pages to exceed the severe threshold globally.
3164 vm_wait_severe(void)
3167 mtx_lock(&vm_domainset_lock);
3168 while (vm_page_count_severe()) {
3169 vm_severe_waiters++;
3170 msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
3173 mtx_unlock(&vm_domainset_lock);
3180 return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
3184 vm_wait_doms(const domainset_t *wdoms, int mflags)
3191 * We use racey wakeup synchronization to avoid expensive global
3192 * locking for the pageproc when sleeping with a non-specific vm_wait.
3193 * To handle this, we only sleep for one tick in this instance. It
3194 * is expected that most allocations for the pageproc will come from
3195 * kmem or vm_page_grab* which will use the more specific and
3196 * race-free vm_wait_domain().
3198 if (curproc == pageproc) {
3199 mtx_lock(&vm_domainset_lock);
3200 vm_pageproc_waiters++;
3201 error = msleep(&vm_pageproc_waiters, &vm_domainset_lock,
3202 PVM | PDROP | mflags, "pageprocwait", 1);
3205 * XXX Ideally we would wait only until the allocation could
3206 * be satisfied. This condition can cause new allocators to
3207 * consume all freed pages while old allocators wait.
3209 mtx_lock(&vm_domainset_lock);
3210 if (vm_page_count_min_set(wdoms)) {
3212 error = msleep(&vm_min_domains, &vm_domainset_lock,
3213 PVM | PDROP | mflags, "vmwait", 0);
3215 mtx_unlock(&vm_domainset_lock);
3223 * Sleep until free pages are available for allocation.
3224 * - Called in various places after failed memory allocations.
3227 vm_wait_domain(int domain)
3229 struct vm_domain *vmd;
3232 vmd = VM_DOMAIN(domain);
3233 vm_domain_free_assert_unlocked(vmd);
3235 if (curproc == pageproc) {
3236 mtx_lock(&vm_domainset_lock);
3237 if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3238 vmd->vmd_pageout_pages_needed = 1;
3239 msleep(&vmd->vmd_pageout_pages_needed,
3240 &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3242 mtx_unlock(&vm_domainset_lock);
3244 if (pageproc == NULL)
3245 panic("vm_wait in early boot");
3246 DOMAINSET_ZERO(&wdom);
3247 DOMAINSET_SET(vmd->vmd_domain, &wdom);
3248 vm_wait_doms(&wdom, 0);
3253 vm_wait_flags(vm_object_t obj, int mflags)
3255 struct domainset *d;
3260 * Carefully fetch pointers only once: the struct domainset
3261 * itself is ummutable but the pointer might change.
3264 d = obj->domain.dr_policy;
3266 d = curthread->td_domain.dr_policy;
3268 return (vm_wait_doms(&d->ds_mask, mflags));
3274 * Sleep until free pages are available for allocation in the
3275 * affinity domains of the obj. If obj is NULL, the domain set
3276 * for the calling thread is used.
3277 * Called in various places after failed memory allocations.
3280 vm_wait(vm_object_t obj)
3282 (void)vm_wait_flags(obj, 0);
3286 vm_wait_intr(vm_object_t obj)
3288 return (vm_wait_flags(obj, PCATCH));
3292 * vm_domain_alloc_fail:
3294 * Called when a page allocation function fails. Informs the
3295 * pagedaemon and performs the requested wait. Requires the
3296 * domain_free and object lock on entry. Returns with the
3297 * object lock held and free lock released. Returns an error when
3298 * retry is necessary.
3302 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3305 vm_domain_free_assert_unlocked(vmd);
3307 atomic_add_int(&vmd->vmd_pageout_deficit,
3308 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3309 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3311 VM_OBJECT_WUNLOCK(object);
3312 vm_wait_domain(vmd->vmd_domain);
3314 VM_OBJECT_WLOCK(object);
3315 if (req & VM_ALLOC_WAITOK)
3325 * Sleep until free pages are available for allocation.
3326 * - Called only in vm_fault so that processes page faulting
3327 * can be easily tracked.
3328 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3329 * processes will be able to grab memory first. Do not change
3330 * this balance without careful testing first.
3333 vm_waitpfault(struct domainset *dset, int timo)
3337 * XXX Ideally we would wait only until the allocation could
3338 * be satisfied. This condition can cause new allocators to
3339 * consume all freed pages while old allocators wait.
3341 mtx_lock(&vm_domainset_lock);
3342 if (vm_page_count_min_set(&dset->ds_mask)) {
3344 msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3347 mtx_unlock(&vm_domainset_lock);
3350 static struct vm_pagequeue *
3351 _vm_page_pagequeue(vm_page_t m, uint8_t queue)
3354 return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3358 static struct vm_pagequeue *
3359 vm_page_pagequeue(vm_page_t m)
3362 return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue));
3366 static __always_inline bool
3367 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3369 vm_page_astate_t tmp;
3373 if (__predict_true(vm_page_astate_fcmpset(m, old, new)))
3375 counter_u64_add(pqstate_commit_retries, 1);
3376 } while (old->_bits == tmp._bits);
3382 * Do the work of committing a queue state update that moves the page out of
3383 * its current queue.
3386 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m,
3387 vm_page_astate_t *old, vm_page_astate_t new)
3391 vm_pagequeue_assert_locked(pq);
3392 KASSERT(vm_page_pagequeue(m) == pq,
3393 ("%s: queue %p does not match page %p", __func__, pq, m));
3394 KASSERT(old->queue != PQ_NONE && new.queue != old->queue,
3395 ("%s: invalid queue indices %d %d",
3396 __func__, old->queue, new.queue));
3399 * Once the queue index of the page changes there is nothing
3400 * synchronizing with further updates to the page's physical
3401 * queue state. Therefore we must speculatively remove the page
3402 * from the queue now and be prepared to roll back if the queue
3403 * state update fails. If the page is not physically enqueued then
3404 * we just update its queue index.
3406 if ((old->flags & PGA_ENQUEUED) != 0) {
3407 new.flags &= ~PGA_ENQUEUED;
3408 next = TAILQ_NEXT(m, plinks.q);
3409 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3410 vm_pagequeue_cnt_dec(pq);
3411 if (!vm_page_pqstate_fcmpset(m, old, new)) {
3413 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3415 TAILQ_INSERT_BEFORE(next, m, plinks.q);
3416 vm_pagequeue_cnt_inc(pq);
3422 return (vm_page_pqstate_fcmpset(m, old, new));
3427 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old,
3428 vm_page_astate_t new)
3430 struct vm_pagequeue *pq;
3431 vm_page_astate_t as;
3434 pq = _vm_page_pagequeue(m, old->queue);
3437 * The queue field and PGA_ENQUEUED flag are stable only so long as the
3438 * corresponding page queue lock is held.
3440 vm_pagequeue_lock(pq);
3441 as = vm_page_astate_load(m);
3442 if (__predict_false(as._bits != old->_bits)) {
3446 ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new);
3448 vm_pagequeue_unlock(pq);
3453 * Commit a queue state update that enqueues or requeues a page.
3456 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m,
3457 vm_page_astate_t *old, vm_page_astate_t new)
3459 struct vm_domain *vmd;
3461 vm_pagequeue_assert_locked(pq);
3462 KASSERT(old->queue != PQ_NONE && new.queue == old->queue,
3463 ("%s: invalid queue indices %d %d",
3464 __func__, old->queue, new.queue));
3466 new.flags |= PGA_ENQUEUED;
3467 if (!vm_page_pqstate_fcmpset(m, old, new))
3470 if ((old->flags & PGA_ENQUEUED) != 0)
3471 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3473 vm_pagequeue_cnt_inc(pq);
3476 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE. In particular, if
3477 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be
3478 * applied, even if it was set first.
3480 if ((old->flags & PGA_REQUEUE_HEAD) != 0) {
3481 vmd = vm_pagequeue_domain(m);
3482 KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE],
3483 ("%s: invalid page queue for page %p", __func__, m));
3484 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3486 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3492 * Commit a queue state update that encodes a request for a deferred queue
3496 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old,
3497 vm_page_astate_t new)
3500 KASSERT(old->queue == new.queue || new.queue != PQ_NONE,
3501 ("%s: invalid state, queue %d flags %x",
3502 __func__, new.queue, new.flags));
3504 if (old->_bits != new._bits &&
3505 !vm_page_pqstate_fcmpset(m, old, new))
3507 vm_page_pqbatch_submit(m, new.queue);
3512 * A generic queue state update function. This handles more cases than the
3513 * specialized functions above.
3516 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3519 if (old->_bits == new._bits)
3522 if (old->queue != PQ_NONE && new.queue != old->queue) {
3523 if (!vm_page_pqstate_commit_dequeue(m, old, new))
3525 if (new.queue != PQ_NONE)
3526 vm_page_pqbatch_submit(m, new.queue);
3528 if (!vm_page_pqstate_fcmpset(m, old, new))
3530 if (new.queue != PQ_NONE &&
3531 ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0)
3532 vm_page_pqbatch_submit(m, new.queue);
3538 * Apply deferred queue state updates to a page.
3541 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue)
3543 vm_page_astate_t new, old;
3545 CRITICAL_ASSERT(curthread);
3546 vm_pagequeue_assert_locked(pq);
3547 KASSERT(queue < PQ_COUNT,
3548 ("%s: invalid queue index %d", __func__, queue));
3549 KASSERT(pq == _vm_page_pagequeue(m, queue),
3550 ("%s: page %p does not belong to queue %p", __func__, m, pq));
3552 for (old = vm_page_astate_load(m);;) {
3553 if (__predict_false(old.queue != queue ||
3554 (old.flags & PGA_QUEUE_OP_MASK) == 0)) {
3555 counter_u64_add(queue_nops, 1);
3558 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3559 ("%s: page %p is unmanaged", __func__, m));
3562 if ((old.flags & PGA_DEQUEUE) != 0) {
3563 new.flags &= ~PGA_QUEUE_OP_MASK;
3564 new.queue = PQ_NONE;
3565 if (__predict_true(_vm_page_pqstate_commit_dequeue(pq,
3567 counter_u64_add(queue_ops, 1);
3571 new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD);
3572 if (__predict_true(_vm_page_pqstate_commit_requeue(pq,
3574 counter_u64_add(queue_ops, 1);
3582 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3587 for (i = 0; i < bq->bq_cnt; i++)
3588 vm_pqbatch_process_page(pq, bq->bq_pa[i], queue);
3589 vm_batchqueue_init(bq);
3593 * vm_page_pqbatch_submit: [ internal use only ]
3595 * Enqueue a page in the specified page queue's batched work queue.
3596 * The caller must have encoded the requested operation in the page
3597 * structure's a.flags field.
3600 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
3602 struct vm_batchqueue *bq;
3603 struct vm_pagequeue *pq;
3606 KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3608 domain = vm_page_domain(m);
3610 bq = DPCPU_PTR(pqbatch[domain][queue]);
3611 if (vm_batchqueue_insert(bq, m)) {
3617 pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue];
3618 vm_pagequeue_lock(pq);
3620 bq = DPCPU_PTR(pqbatch[domain][queue]);
3621 vm_pqbatch_process(pq, bq, queue);
3622 vm_pqbatch_process_page(pq, m, queue);
3623 vm_pagequeue_unlock(pq);
3628 * vm_page_pqbatch_drain: [ internal use only ]
3630 * Force all per-CPU page queue batch queues to be drained. This is
3631 * intended for use in severe memory shortages, to ensure that pages
3632 * do not remain stuck in the batch queues.
3635 vm_page_pqbatch_drain(void)
3638 struct vm_domain *vmd;
3639 struct vm_pagequeue *pq;
3640 int cpu, domain, queue;
3645 sched_bind(td, cpu);
3648 for (domain = 0; domain < vm_ndomains; domain++) {
3649 vmd = VM_DOMAIN(domain);
3650 for (queue = 0; queue < PQ_COUNT; queue++) {
3651 pq = &vmd->vmd_pagequeues[queue];
3652 vm_pagequeue_lock(pq);
3654 vm_pqbatch_process(pq,
3655 DPCPU_PTR(pqbatch[domain][queue]), queue);
3657 vm_pagequeue_unlock(pq);
3667 * vm_page_dequeue_deferred: [ internal use only ]
3669 * Request removal of the given page from its current page
3670 * queue. Physical removal from the queue may be deferred
3674 vm_page_dequeue_deferred(vm_page_t m)
3676 vm_page_astate_t new, old;
3678 old = vm_page_astate_load(m);
3680 if (old.queue == PQ_NONE) {
3681 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3682 ("%s: page %p has unexpected queue state",
3687 new.flags |= PGA_DEQUEUE;
3688 } while (!vm_page_pqstate_commit_request(m, &old, new));
3694 * Remove the page from whichever page queue it's in, if any, before
3698 vm_page_dequeue(vm_page_t m)
3700 vm_page_astate_t new, old;
3702 old = vm_page_astate_load(m);
3704 if (old.queue == PQ_NONE) {
3705 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3706 ("%s: page %p has unexpected queue state",
3711 new.flags &= ~PGA_QUEUE_OP_MASK;
3712 new.queue = PQ_NONE;
3713 } while (!vm_page_pqstate_commit_dequeue(m, &old, new));
3718 * Schedule the given page for insertion into the specified page queue.
3719 * Physical insertion of the page may be deferred indefinitely.
3722 vm_page_enqueue(vm_page_t m, uint8_t queue)
3725 KASSERT(m->a.queue == PQ_NONE &&
3726 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
3727 ("%s: page %p is already enqueued", __func__, m));
3728 KASSERT(m->ref_count > 0,
3729 ("%s: page %p does not carry any references", __func__, m));
3732 if ((m->a.flags & PGA_REQUEUE) == 0)
3733 vm_page_aflag_set(m, PGA_REQUEUE);
3734 vm_page_pqbatch_submit(m, queue);
3738 * vm_page_free_prep:
3740 * Prepares the given page to be put on the free list,
3741 * disassociating it from any VM object. The caller may return
3742 * the page to the free list only if this function returns true.
3744 * The object, if it exists, must be locked, and then the page must
3745 * be xbusy. Otherwise the page must be not busied. A managed
3746 * page must be unmapped.
3749 vm_page_free_prep(vm_page_t m)
3753 * Synchronize with threads that have dropped a reference to this
3756 atomic_thread_fence_acq();
3758 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3759 if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3762 p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3763 for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3764 KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3765 m, i, (uintmax_t)*p));
3768 if ((m->oflags & VPO_UNMANAGED) == 0) {
3769 KASSERT(!pmap_page_is_mapped(m),
3770 ("vm_page_free_prep: freeing mapped page %p", m));
3771 KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
3772 ("vm_page_free_prep: mapping flags set in page %p", m));
3774 KASSERT(m->a.queue == PQ_NONE,
3775 ("vm_page_free_prep: unmanaged page %p is queued", m));
3777 VM_CNT_INC(v_tfree);
3779 if (m->object != NULL) {
3780 KASSERT(((m->oflags & VPO_UNMANAGED) != 0) ==
3781 ((m->object->flags & OBJ_UNMANAGED) != 0),
3782 ("vm_page_free_prep: managed flag mismatch for page %p",
3784 vm_page_assert_xbusied(m);
3787 * The object reference can be released without an atomic
3790 KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
3791 m->ref_count == VPRC_OBJREF,
3792 ("vm_page_free_prep: page %p has unexpected ref_count %u",
3794 vm_page_object_remove(m);
3795 m->ref_count -= VPRC_OBJREF;
3797 vm_page_assert_unbusied(m);
3799 vm_page_busy_free(m);
3802 * If fictitious remove object association and
3805 if ((m->flags & PG_FICTITIOUS) != 0) {
3806 KASSERT(m->ref_count == 1,
3807 ("fictitious page %p is referenced", m));
3808 KASSERT(m->a.queue == PQ_NONE,
3809 ("fictitious page %p is queued", m));
3814 * Pages need not be dequeued before they are returned to the physical
3815 * memory allocator, but they must at least be marked for a deferred
3818 if ((m->oflags & VPO_UNMANAGED) == 0)
3819 vm_page_dequeue_deferred(m);
3824 if (m->ref_count != 0)
3825 panic("vm_page_free_prep: page %p has references", m);
3828 * Restore the default memory attribute to the page.
3830 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3831 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3833 #if VM_NRESERVLEVEL > 0
3835 * Determine whether the page belongs to a reservation. If the page was
3836 * allocated from a per-CPU cache, it cannot belong to a reservation, so
3837 * as an optimization, we avoid the check in that case.
3839 if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
3849 * Returns the given page to the free list, disassociating it
3850 * from any VM object.
3852 * The object must be locked. The page must be exclusively busied if it
3853 * belongs to an object.
3856 vm_page_free_toq(vm_page_t m)
3858 struct vm_domain *vmd;
3861 if (!vm_page_free_prep(m))
3864 vmd = vm_pagequeue_domain(m);
3865 zone = vmd->vmd_pgcache[m->pool].zone;
3866 if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
3870 vm_domain_free_lock(vmd);
3871 vm_phys_free_pages(m, 0);
3872 vm_domain_free_unlock(vmd);
3873 vm_domain_freecnt_inc(vmd, 1);
3877 * vm_page_free_pages_toq:
3879 * Returns a list of pages to the free list, disassociating it
3880 * from any VM object. In other words, this is equivalent to
3881 * calling vm_page_free_toq() for each page of a list of VM objects.
3884 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
3889 if (SLIST_EMPTY(free))
3893 while ((m = SLIST_FIRST(free)) != NULL) {
3895 SLIST_REMOVE_HEAD(free, plinks.s.ss);
3896 vm_page_free_toq(m);
3899 if (update_wire_count)
3904 * Mark this page as wired down. For managed pages, this prevents reclamation
3905 * by the page daemon, or when the containing object, if any, is destroyed.
3908 vm_page_wire(vm_page_t m)
3913 if (m->object != NULL && !vm_page_busied(m) &&
3914 !vm_object_busied(m->object))
3915 VM_OBJECT_ASSERT_LOCKED(m->object);
3917 KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
3918 VPRC_WIRE_COUNT(m->ref_count) >= 1,
3919 ("vm_page_wire: fictitious page %p has zero wirings", m));
3921 old = atomic_fetchadd_int(&m->ref_count, 1);
3922 KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
3923 ("vm_page_wire: counter overflow for page %p", m));
3924 if (VPRC_WIRE_COUNT(old) == 0) {
3925 if ((m->oflags & VPO_UNMANAGED) == 0)
3926 vm_page_aflag_set(m, PGA_DEQUEUE);
3932 * Attempt to wire a mapped page following a pmap lookup of that page.
3933 * This may fail if a thread is concurrently tearing down mappings of the page.
3934 * The transient failure is acceptable because it translates to the
3935 * failure of the caller pmap_extract_and_hold(), which should be then
3936 * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages().
3939 vm_page_wire_mapped(vm_page_t m)
3946 ("vm_page_wire_mapped: wiring unreferenced page %p", m));
3947 if ((old & VPRC_BLOCKED) != 0)
3949 } while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
3951 if (VPRC_WIRE_COUNT(old) == 0) {
3952 if ((m->oflags & VPO_UNMANAGED) == 0)
3953 vm_page_aflag_set(m, PGA_DEQUEUE);
3960 * Release a wiring reference to a managed page. If the page still belongs to
3961 * an object, update its position in the page queues to reflect the reference.
3962 * If the wiring was the last reference to the page, free the page.
3965 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse)
3969 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3970 ("%s: page %p is unmanaged", __func__, m));
3973 * Update LRU state before releasing the wiring reference.
3974 * Use a release store when updating the reference count to
3975 * synchronize with vm_page_free_prep().
3979 KASSERT(VPRC_WIRE_COUNT(old) > 0,
3980 ("vm_page_unwire: wire count underflow for page %p", m));
3982 if (old > VPRC_OBJREF + 1) {
3984 * The page has at least one other wiring reference. An
3985 * earlier iteration of this loop may have called
3986 * vm_page_release_toq() and cleared PGA_DEQUEUE, so
3987 * re-set it if necessary.
3989 if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0)
3990 vm_page_aflag_set(m, PGA_DEQUEUE);
3991 } else if (old == VPRC_OBJREF + 1) {
3993 * This is the last wiring. Clear PGA_DEQUEUE and
3994 * update the page's queue state to reflect the
3995 * reference. If the page does not belong to an object
3996 * (i.e., the VPRC_OBJREF bit is clear), we only need to
3997 * clear leftover queue state.
3999 vm_page_release_toq(m, nqueue, noreuse);
4000 } else if (old == 1) {
4001 vm_page_aflag_clear(m, PGA_DEQUEUE);
4003 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
4005 if (VPRC_WIRE_COUNT(old) == 1) {
4013 * Release one wiring of the specified page, potentially allowing it to be
4016 * Only managed pages belonging to an object can be paged out. If the number
4017 * of wirings transitions to zero and the page is eligible for page out, then
4018 * the page is added to the specified paging queue. If the released wiring
4019 * represented the last reference to the page, the page is freed.
4022 vm_page_unwire(vm_page_t m, uint8_t nqueue)
4025 KASSERT(nqueue < PQ_COUNT,
4026 ("vm_page_unwire: invalid queue %u request for page %p",
4029 if ((m->oflags & VPO_UNMANAGED) != 0) {
4030 if (vm_page_unwire_noq(m) && m->ref_count == 0)
4034 vm_page_unwire_managed(m, nqueue, false);
4038 * Unwire a page without (re-)inserting it into a page queue. It is up
4039 * to the caller to enqueue, requeue, or free the page as appropriate.
4040 * In most cases involving managed pages, vm_page_unwire() should be used
4044 vm_page_unwire_noq(vm_page_t m)
4048 old = vm_page_drop(m, 1);
4049 KASSERT(VPRC_WIRE_COUNT(old) != 0,
4050 ("%s: counter underflow for page %p", __func__, m));
4051 KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
4052 ("%s: missing ref on fictitious page %p", __func__, m));
4054 if (VPRC_WIRE_COUNT(old) > 1)
4056 if ((m->oflags & VPO_UNMANAGED) == 0)
4057 vm_page_aflag_clear(m, PGA_DEQUEUE);
4063 * Ensure that the page ends up in the specified page queue. If the page is
4064 * active or being moved to the active queue, ensure that its act_count is
4065 * at least ACT_INIT but do not otherwise mess with it.
4067 static __always_inline void
4068 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag)
4070 vm_page_astate_t old, new;
4072 KASSERT(m->ref_count > 0,
4073 ("%s: page %p does not carry any references", __func__, m));
4074 KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD,
4075 ("%s: invalid flags %x", __func__, nflag));
4077 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
4080 old = vm_page_astate_load(m);
4082 if ((old.flags & PGA_DEQUEUE) != 0)
4085 new.flags &= ~PGA_QUEUE_OP_MASK;
4086 if (nqueue == PQ_ACTIVE)
4087 new.act_count = max(old.act_count, ACT_INIT);
4088 if (old.queue == nqueue) {
4089 if (nqueue != PQ_ACTIVE)
4095 } while (!vm_page_pqstate_commit(m, &old, new));
4099 * Put the specified page on the active list (if appropriate).
4102 vm_page_activate(vm_page_t m)
4105 vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE);
4109 * Move the specified page to the tail of the inactive queue, or requeue
4110 * the page if it is already in the inactive queue.
4113 vm_page_deactivate(vm_page_t m)
4116 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE);
4120 vm_page_deactivate_noreuse(vm_page_t m)
4123 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD);
4127 * Put a page in the laundry, or requeue it if it is already there.
4130 vm_page_launder(vm_page_t m)
4133 vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE);
4137 * Put a page in the PQ_UNSWAPPABLE holding queue.
4140 vm_page_unswappable(vm_page_t m)
4143 KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0,
4144 ("page %p already unswappable", m));
4147 vm_page_enqueue(m, PQ_UNSWAPPABLE);
4151 * Release a page back to the page queues in preparation for unwiring.
4154 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse)
4156 vm_page_astate_t old, new;
4160 * Use a check of the valid bits to determine whether we should
4161 * accelerate reclamation of the page. The object lock might not be
4162 * held here, in which case the check is racy. At worst we will either
4163 * accelerate reclamation of a valid page and violate LRU, or
4164 * unnecessarily defer reclamation of an invalid page.
4166 * If we were asked to not cache the page, place it near the head of the
4167 * inactive queue so that is reclaimed sooner.
4169 if (noreuse || m->valid == 0) {
4170 nqueue = PQ_INACTIVE;
4171 nflag = PGA_REQUEUE_HEAD;
4173 nflag = PGA_REQUEUE;
4176 old = vm_page_astate_load(m);
4181 * If the page is already in the active queue and we are not
4182 * trying to accelerate reclamation, simply mark it as
4183 * referenced and avoid any queue operations.
4185 new.flags &= ~PGA_QUEUE_OP_MASK;
4186 if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE)
4187 new.flags |= PGA_REFERENCED;
4192 } while (!vm_page_pqstate_commit(m, &old, new));
4196 * Unwire a page and either attempt to free it or re-add it to the page queues.
4199 vm_page_release(vm_page_t m, int flags)
4203 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4204 ("vm_page_release: page %p is unmanaged", m));
4206 if ((flags & VPR_TRYFREE) != 0) {
4208 object = atomic_load_ptr(&m->object);
4211 /* Depends on type-stability. */
4212 if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object))
4214 if (object == m->object) {
4215 vm_page_release_locked(m, flags);
4216 VM_OBJECT_WUNLOCK(object);
4219 VM_OBJECT_WUNLOCK(object);
4222 vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0);
4225 /* See vm_page_release(). */
4227 vm_page_release_locked(vm_page_t m, int flags)
4230 VM_OBJECT_ASSERT_WLOCKED(m->object);
4231 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4232 ("vm_page_release_locked: page %p is unmanaged", m));
4234 if (vm_page_unwire_noq(m)) {
4235 if ((flags & VPR_TRYFREE) != 0 &&
4236 (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
4237 m->dirty == 0 && vm_page_tryxbusy(m)) {
4239 * An unlocked lookup may have wired the page before the
4240 * busy lock was acquired, in which case the page must
4243 if (__predict_true(!vm_page_wired(m))) {
4249 vm_page_release_toq(m, PQ_INACTIVE, flags != 0);
4255 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
4259 KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
4260 ("vm_page_try_blocked_op: page %p has no object", m));
4261 KASSERT(vm_page_busied(m),
4262 ("vm_page_try_blocked_op: page %p is not busy", m));
4263 VM_OBJECT_ASSERT_LOCKED(m->object);
4268 ("vm_page_try_blocked_op: page %p has no references", m));
4269 if (VPRC_WIRE_COUNT(old) != 0)
4271 } while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
4276 * If the object is read-locked, new wirings may be created via an
4279 old = vm_page_drop(m, VPRC_BLOCKED);
4280 KASSERT(!VM_OBJECT_WOWNED(m->object) ||
4281 old == (VPRC_BLOCKED | VPRC_OBJREF),
4282 ("vm_page_try_blocked_op: unexpected refcount value %u for %p",
4288 * Atomically check for wirings and remove all mappings of the page.
4291 vm_page_try_remove_all(vm_page_t m)
4294 return (vm_page_try_blocked_op(m, pmap_remove_all));
4298 * Atomically check for wirings and remove all writeable mappings of the page.
4301 vm_page_try_remove_write(vm_page_t m)
4304 return (vm_page_try_blocked_op(m, pmap_remove_write));
4310 * Apply the specified advice to the given page.
4313 vm_page_advise(vm_page_t m, int advice)
4316 VM_OBJECT_ASSERT_WLOCKED(m->object);
4317 vm_page_assert_xbusied(m);
4319 if (advice == MADV_FREE)
4321 * Mark the page clean. This will allow the page to be freed
4322 * without first paging it out. MADV_FREE pages are often
4323 * quickly reused by malloc(3), so we do not do anything that
4324 * would result in a page fault on a later access.
4327 else if (advice != MADV_DONTNEED) {
4328 if (advice == MADV_WILLNEED)
4329 vm_page_activate(m);
4333 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
4337 * Clear any references to the page. Otherwise, the page daemon will
4338 * immediately reactivate the page.
4340 vm_page_aflag_clear(m, PGA_REFERENCED);
4343 * Place clean pages near the head of the inactive queue rather than
4344 * the tail, thus defeating the queue's LRU operation and ensuring that
4345 * the page will be reused quickly. Dirty pages not already in the
4346 * laundry are moved there.
4349 vm_page_deactivate_noreuse(m);
4350 else if (!vm_page_in_laundry(m))
4355 * vm_page_grab_release
4357 * Helper routine for grab functions to release busy on return.
4360 vm_page_grab_release(vm_page_t m, int allocflags)
4363 if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4364 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4372 * vm_page_grab_sleep
4374 * Sleep for busy according to VM_ALLOC_ parameters. Returns true
4375 * if the caller should retry and false otherwise.
4377 * If the object is locked on entry the object will be unlocked with
4378 * false returns and still locked but possibly having been dropped
4379 * with true returns.
4382 vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex,
4383 const char *wmesg, int allocflags, bool locked)
4386 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4390 * Reference the page before unlocking and sleeping so that
4391 * the page daemon is less likely to reclaim it.
4393 if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0)
4394 vm_page_reference(m);
4396 if (_vm_page_busy_sleep(object, m, pindex, wmesg, allocflags, locked) &&
4398 VM_OBJECT_WLOCK(object);
4399 if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
4406 * Assert that the grab flags are valid.
4409 vm_page_grab_check(int allocflags)
4412 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4413 (allocflags & VM_ALLOC_WIRED) != 0,
4414 ("vm_page_grab*: the pages must be busied or wired"));
4416 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4417 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4418 ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4422 * Calculate the page allocation flags for grab.
4425 vm_page_grab_pflags(int allocflags)
4429 pflags = allocflags &
4430 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL |
4432 if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4433 pflags |= VM_ALLOC_WAITFAIL;
4434 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4435 pflags |= VM_ALLOC_SBUSY;
4441 * Grab a page, waiting until we are waken up due to the page
4442 * changing state. We keep on waiting, if the page continues
4443 * to be in the object. If the page doesn't exist, first allocate it
4444 * and then conditionally zero it.
4446 * This routine may sleep.
4448 * The object must be locked on entry. The lock will, however, be released
4449 * and reacquired if the routine sleeps.
4452 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
4456 VM_OBJECT_ASSERT_WLOCKED(object);
4457 vm_page_grab_check(allocflags);
4460 if ((m = vm_page_lookup(object, pindex)) != NULL) {
4461 if (!vm_page_tryacquire(m, allocflags)) {
4462 if (vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4469 if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4471 m = vm_page_alloc(object, pindex, vm_page_grab_pflags(allocflags));
4473 if ((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0)
4477 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
4481 vm_page_grab_release(m, allocflags);
4487 * Locklessly attempt to acquire a page given a (object, pindex) tuple
4488 * and an optional previous page to avoid the radix lookup. The resulting
4489 * page will be validated against the identity tuple and busied or wired
4490 * as requested. A NULL *mp return guarantees that the page was not in
4491 * radix at the time of the call but callers must perform higher level
4492 * synchronization or retry the operation under a lock if they require
4493 * an atomic answer. This is the only lock free validation routine,
4494 * other routines can depend on the resulting page state.
4496 * The return value indicates whether the operation failed due to caller
4497 * flags. The return is tri-state with mp:
4499 * (true, *mp != NULL) - The operation was successful.
4500 * (true, *mp == NULL) - The page was not found in tree.
4501 * (false, *mp == NULL) - WAITFAIL or NOWAIT prevented acquisition.
4504 vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex,
4505 vm_page_t prev, vm_page_t *mp, int allocflags)
4509 vm_page_grab_check(allocflags);
4510 MPASS(prev == NULL || vm_page_busied(prev) || vm_page_wired(prev));
4515 * We may see a false NULL here because the previous page
4516 * has been removed or just inserted and the list is loaded
4517 * without barriers. Switch to radix to verify.
4519 if (prev == NULL || (m = TAILQ_NEXT(prev, listq)) == NULL ||
4520 QMD_IS_TRASHED(m) || m->pindex != pindex ||
4521 atomic_load_ptr(&m->object) != object) {
4524 * This guarantees the result is instantaneously
4527 m = vm_radix_lookup_unlocked(&object->rtree, pindex);
4531 if (vm_page_trybusy(m, allocflags)) {
4532 if (m->object == object && m->pindex == pindex)
4535 vm_page_busy_release(m);
4539 if (!vm_page_grab_sleep(object, m, pindex, "pgnslp",
4543 if ((allocflags & VM_ALLOC_WIRED) != 0)
4545 vm_page_grab_release(m, allocflags);
4551 * Try to locklessly grab a page and fall back to the object lock if NOCREAT
4555 vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags)
4559 vm_page_grab_check(allocflags);
4561 if (!vm_page_acquire_unlocked(object, pindex, NULL, &m, allocflags))
4567 * The radix lockless lookup should never return a false negative
4568 * errors. If the user specifies NOCREAT they are guaranteed there
4569 * was no page present at the instant of the call. A NOCREAT caller
4570 * must handle create races gracefully.
4572 if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4575 VM_OBJECT_WLOCK(object);
4576 m = vm_page_grab(object, pindex, allocflags);
4577 VM_OBJECT_WUNLOCK(object);
4583 * Grab a page and make it valid, paging in if necessary. Pages missing from
4584 * their pager are zero filled and validated. If a VM_ALLOC_COUNT is supplied
4585 * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought
4586 * in simultaneously. Additional pages will be left on a paging queue but
4587 * will neither be wired nor busy regardless of allocflags.
4590 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags)
4593 vm_page_t ma[VM_INITIAL_PAGEIN];
4594 int after, i, pflags, rv;
4596 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4597 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4598 ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4599 KASSERT((allocflags &
4600 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4601 ("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
4602 VM_OBJECT_ASSERT_WLOCKED(object);
4603 pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY |
4605 pflags |= VM_ALLOC_WAITFAIL;
4608 if ((m = vm_page_lookup(object, pindex)) != NULL) {
4610 * If the page is fully valid it can only become invalid
4611 * with the object lock held. If it is not valid it can
4612 * become valid with the busy lock held. Therefore, we
4613 * may unnecessarily lock the exclusive busy here if we
4614 * race with I/O completion not using the object lock.
4615 * However, we will not end up with an invalid page and a
4618 if (!vm_page_trybusy(m,
4619 vm_page_all_valid(m) ? allocflags : 0)) {
4620 (void)vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4624 if (vm_page_all_valid(m))
4626 if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4627 vm_page_busy_release(m);
4629 return (VM_PAGER_FAIL);
4631 } else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4633 return (VM_PAGER_FAIL);
4634 } else if ((m = vm_page_alloc(object, pindex, pflags)) == NULL) {
4638 vm_page_assert_xbusied(m);
4639 if (vm_pager_has_page(object, pindex, NULL, &after)) {
4640 after = MIN(after, VM_INITIAL_PAGEIN);
4641 after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT);
4642 after = MAX(after, 1);
4644 for (i = 1; i < after; i++) {
4645 if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) {
4646 if (ma[i]->valid || !vm_page_tryxbusy(ma[i]))
4649 ma[i] = vm_page_alloc(object, m->pindex + i,
4656 vm_object_pip_add(object, after);
4657 VM_OBJECT_WUNLOCK(object);
4658 rv = vm_pager_get_pages(object, ma, after, NULL, NULL);
4659 VM_OBJECT_WLOCK(object);
4660 vm_object_pip_wakeupn(object, after);
4661 /* Pager may have replaced a page. */
4663 if (rv != VM_PAGER_OK) {
4664 for (i = 0; i < after; i++) {
4665 if (!vm_page_wired(ma[i]))
4666 vm_page_free(ma[i]);
4668 vm_page_xunbusy(ma[i]);
4673 for (i = 1; i < after; i++)
4674 vm_page_readahead_finish(ma[i]);
4675 MPASS(vm_page_all_valid(m));
4677 vm_page_zero_invalid(m, TRUE);
4680 if ((allocflags & VM_ALLOC_WIRED) != 0)
4682 if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m))
4683 vm_page_busy_downgrade(m);
4684 else if ((allocflags & VM_ALLOC_NOBUSY) != 0)
4685 vm_page_busy_release(m);
4687 return (VM_PAGER_OK);
4691 * Locklessly grab a valid page. If the page is not valid or not yet
4692 * allocated this will fall back to the object lock method.
4695 vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object,
4696 vm_pindex_t pindex, int allocflags)
4702 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4703 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4704 ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY "
4706 KASSERT((allocflags &
4707 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4708 ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags));
4711 * Attempt a lockless lookup and busy. We need at least an sbusy
4712 * before we can inspect the valid field and return a wired page.
4714 flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED);
4715 if (!vm_page_acquire_unlocked(object, pindex, NULL, mp, flags))
4716 return (VM_PAGER_FAIL);
4717 if ((m = *mp) != NULL) {
4718 if (vm_page_all_valid(m)) {
4719 if ((allocflags & VM_ALLOC_WIRED) != 0)
4721 vm_page_grab_release(m, allocflags);
4722 return (VM_PAGER_OK);
4724 vm_page_busy_release(m);
4726 if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4728 return (VM_PAGER_FAIL);
4730 VM_OBJECT_WLOCK(object);
4731 error = vm_page_grab_valid(mp, object, pindex, allocflags);
4732 VM_OBJECT_WUNLOCK(object);
4738 * Return the specified range of pages from the given object. For each
4739 * page offset within the range, if a page already exists within the object
4740 * at that offset and it is busy, then wait for it to change state. If,
4741 * instead, the page doesn't exist, then allocate it.
4743 * The caller must always specify an allocation class.
4745 * allocation classes:
4746 * VM_ALLOC_NORMAL normal process request
4747 * VM_ALLOC_SYSTEM system *really* needs the pages
4749 * The caller must always specify that the pages are to be busied and/or
4752 * optional allocation flags:
4753 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages
4754 * VM_ALLOC_NOBUSY do not exclusive busy the page
4755 * VM_ALLOC_NOWAIT do not sleep
4756 * VM_ALLOC_SBUSY set page to sbusy state
4757 * VM_ALLOC_WIRED wire the pages
4758 * VM_ALLOC_ZERO zero and validate any invalid pages
4760 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it
4761 * may return a partial prefix of the requested range.
4764 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
4765 vm_page_t *ma, int count)
4771 VM_OBJECT_ASSERT_WLOCKED(object);
4772 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
4773 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
4775 ("vm_page_grab_pages: invalid page count %d", count));
4776 vm_page_grab_check(allocflags);
4778 pflags = vm_page_grab_pflags(allocflags);
4781 m = vm_radix_lookup_le(&object->rtree, pindex + i);
4782 if (m == NULL || m->pindex != pindex + i) {
4786 mpred = TAILQ_PREV(m, pglist, listq);
4787 for (; i < count; i++) {
4789 if (!vm_page_tryacquire(m, allocflags)) {
4790 if (vm_page_grab_sleep(object, m, pindex + i,
4791 "grbmaw", allocflags, true))
4796 if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4798 m = vm_page_alloc_after(object, pindex + i,
4799 pflags | VM_ALLOC_COUNT(count - i), mpred);
4801 if ((allocflags & (VM_ALLOC_NOWAIT |
4802 VM_ALLOC_WAITFAIL)) != 0)
4807 if (vm_page_none_valid(m) &&
4808 (allocflags & VM_ALLOC_ZERO) != 0) {
4809 if ((m->flags & PG_ZERO) == 0)
4813 vm_page_grab_release(m, allocflags);
4815 m = vm_page_next(m);
4821 * Unlocked variant of vm_page_grab_pages(). This accepts the same flags
4822 * and will fall back to the locked variant to handle allocation.
4825 vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex,
4826 int allocflags, vm_page_t *ma, int count)
4833 ("vm_page_grab_pages_unlocked: invalid page count %d", count));
4834 vm_page_grab_check(allocflags);
4837 * Modify flags for lockless acquire to hold the page until we
4838 * set it valid if necessary.
4840 flags = allocflags & ~VM_ALLOC_NOBUSY;
4842 for (i = 0; i < count; i++, pindex++) {
4843 if (!vm_page_acquire_unlocked(object, pindex, pred, &m, flags))
4847 if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) {
4848 if ((m->flags & PG_ZERO) == 0)
4852 /* m will still be wired or busy according to flags. */
4853 vm_page_grab_release(m, allocflags);
4856 if (i == count || (allocflags & VM_ALLOC_NOCREAT) != 0)
4859 VM_OBJECT_WLOCK(object);
4860 i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count);
4861 VM_OBJECT_WUNLOCK(object);
4867 * Mapping function for valid or dirty bits in a page.
4869 * Inputs are required to range within a page.
4872 vm_page_bits(int base, int size)
4878 base + size <= PAGE_SIZE,
4879 ("vm_page_bits: illegal base/size %d/%d", base, size)
4882 if (size == 0) /* handle degenerate case */
4885 first_bit = base >> DEV_BSHIFT;
4886 last_bit = (base + size - 1) >> DEV_BSHIFT;
4888 return (((vm_page_bits_t)2 << last_bit) -
4889 ((vm_page_bits_t)1 << first_bit));
4893 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set)
4896 #if PAGE_SIZE == 32768
4897 atomic_set_64((uint64_t *)bits, set);
4898 #elif PAGE_SIZE == 16384
4899 atomic_set_32((uint32_t *)bits, set);
4900 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16)
4901 atomic_set_16((uint16_t *)bits, set);
4902 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8)
4903 atomic_set_8((uint8_t *)bits, set);
4904 #else /* PAGE_SIZE <= 8192 */
4908 addr = (uintptr_t)bits;
4910 * Use a trick to perform a 32-bit atomic on the
4911 * containing aligned word, to not depend on the existence
4912 * of atomic_{set, clear}_{8, 16}.
4914 shift = addr & (sizeof(uint32_t) - 1);
4915 #if BYTE_ORDER == BIG_ENDIAN
4916 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4920 addr &= ~(sizeof(uint32_t) - 1);
4921 atomic_set_32((uint32_t *)addr, set << shift);
4922 #endif /* PAGE_SIZE */
4926 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear)
4929 #if PAGE_SIZE == 32768
4930 atomic_clear_64((uint64_t *)bits, clear);
4931 #elif PAGE_SIZE == 16384
4932 atomic_clear_32((uint32_t *)bits, clear);
4933 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16)
4934 atomic_clear_16((uint16_t *)bits, clear);
4935 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8)
4936 atomic_clear_8((uint8_t *)bits, clear);
4937 #else /* PAGE_SIZE <= 8192 */
4941 addr = (uintptr_t)bits;
4943 * Use a trick to perform a 32-bit atomic on the
4944 * containing aligned word, to not depend on the existence
4945 * of atomic_{set, clear}_{8, 16}.
4947 shift = addr & (sizeof(uint32_t) - 1);
4948 #if BYTE_ORDER == BIG_ENDIAN
4949 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4953 addr &= ~(sizeof(uint32_t) - 1);
4954 atomic_clear_32((uint32_t *)addr, clear << shift);
4955 #endif /* PAGE_SIZE */
4958 static inline vm_page_bits_t
4959 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits)
4961 #if PAGE_SIZE == 32768
4965 while (atomic_fcmpset_64(bits, &old, newbits) == 0);
4967 #elif PAGE_SIZE == 16384
4971 while (atomic_fcmpset_32(bits, &old, newbits) == 0);
4973 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16)
4977 while (atomic_fcmpset_16(bits, &old, newbits) == 0);
4979 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8)
4983 while (atomic_fcmpset_8(bits, &old, newbits) == 0);
4985 #else /* PAGE_SIZE <= 4096*/
4987 uint32_t old, new, mask;
4990 addr = (uintptr_t)bits;
4992 * Use a trick to perform a 32-bit atomic on the
4993 * containing aligned word, to not depend on the existence
4994 * of atomic_{set, swap, clear}_{8, 16}.
4996 shift = addr & (sizeof(uint32_t) - 1);
4997 #if BYTE_ORDER == BIG_ENDIAN
4998 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5002 addr &= ~(sizeof(uint32_t) - 1);
5003 mask = VM_PAGE_BITS_ALL << shift;
5008 new |= newbits << shift;
5009 } while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0);
5010 return (old >> shift);
5011 #endif /* PAGE_SIZE */
5015 * vm_page_set_valid_range:
5017 * Sets portions of a page valid. The arguments are expected
5018 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5019 * of any partial chunks touched by the range. The invalid portion of
5020 * such chunks will be zeroed.
5022 * (base + size) must be less then or equal to PAGE_SIZE.
5025 vm_page_set_valid_range(vm_page_t m, int base, int size)
5028 vm_page_bits_t pagebits;
5030 vm_page_assert_busied(m);
5031 if (size == 0) /* handle degenerate case */
5035 * If the base is not DEV_BSIZE aligned and the valid
5036 * bit is clear, we have to zero out a portion of the
5039 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5040 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
5041 pmap_zero_page_area(m, frag, base - frag);
5044 * If the ending offset is not DEV_BSIZE aligned and the
5045 * valid bit is clear, we have to zero out a portion of
5048 endoff = base + size;
5049 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5050 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
5051 pmap_zero_page_area(m, endoff,
5052 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5055 * Assert that no previously invalid block that is now being validated
5058 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
5059 ("vm_page_set_valid_range: page %p is dirty", m));
5062 * Set valid bits inclusive of any overlap.
5064 pagebits = vm_page_bits(base, size);
5065 if (vm_page_xbusied(m))
5066 m->valid |= pagebits;
5068 vm_page_bits_set(m, &m->valid, pagebits);
5072 * Set the page dirty bits and free the invalid swap space if
5073 * present. Returns the previous dirty bits.
5076 vm_page_set_dirty(vm_page_t m)
5080 VM_PAGE_OBJECT_BUSY_ASSERT(m);
5082 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) {
5084 m->dirty = VM_PAGE_BITS_ALL;
5086 old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL);
5087 if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0)
5088 vm_pager_page_unswapped(m);
5094 * Clear the given bits from the specified page's dirty field.
5096 static __inline void
5097 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
5100 vm_page_assert_busied(m);
5103 * If the page is xbusied and not write mapped we are the
5104 * only thread that can modify dirty bits. Otherwise, The pmap
5105 * layer can call vm_page_dirty() without holding a distinguished
5106 * lock. The combination of page busy and atomic operations
5107 * suffice to guarantee consistency of the page dirty field.
5109 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
5110 m->dirty &= ~pagebits;
5112 vm_page_bits_clear(m, &m->dirty, pagebits);
5116 * vm_page_set_validclean:
5118 * Sets portions of a page valid and clean. The arguments are expected
5119 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5120 * of any partial chunks touched by the range. The invalid portion of
5121 * such chunks will be zero'd.
5123 * (base + size) must be less then or equal to PAGE_SIZE.
5126 vm_page_set_validclean(vm_page_t m, int base, int size)
5128 vm_page_bits_t oldvalid, pagebits;
5131 vm_page_assert_busied(m);
5132 if (size == 0) /* handle degenerate case */
5136 * If the base is not DEV_BSIZE aligned and the valid
5137 * bit is clear, we have to zero out a portion of the
5140 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5141 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
5142 pmap_zero_page_area(m, frag, base - frag);
5145 * If the ending offset is not DEV_BSIZE aligned and the
5146 * valid bit is clear, we have to zero out a portion of
5149 endoff = base + size;
5150 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5151 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
5152 pmap_zero_page_area(m, endoff,
5153 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5156 * Set valid, clear dirty bits. If validating the entire
5157 * page we can safely clear the pmap modify bit. We also
5158 * use this opportunity to clear the PGA_NOSYNC flag. If a process
5159 * takes a write fault on a MAP_NOSYNC memory area the flag will
5162 * We set valid bits inclusive of any overlap, but we can only
5163 * clear dirty bits for DEV_BSIZE chunks that are fully within
5166 oldvalid = m->valid;
5167 pagebits = vm_page_bits(base, size);
5168 if (vm_page_xbusied(m))
5169 m->valid |= pagebits;
5171 vm_page_bits_set(m, &m->valid, pagebits);
5173 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
5174 frag = DEV_BSIZE - frag;
5180 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
5182 if (base == 0 && size == PAGE_SIZE) {
5184 * The page can only be modified within the pmap if it is
5185 * mapped, and it can only be mapped if it was previously
5188 if (oldvalid == VM_PAGE_BITS_ALL)
5190 * Perform the pmap_clear_modify() first. Otherwise,
5191 * a concurrent pmap operation, such as
5192 * pmap_protect(), could clear a modification in the
5193 * pmap and set the dirty field on the page before
5194 * pmap_clear_modify() had begun and after the dirty
5195 * field was cleared here.
5197 pmap_clear_modify(m);
5199 vm_page_aflag_clear(m, PGA_NOSYNC);
5200 } else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m))
5201 m->dirty &= ~pagebits;
5203 vm_page_clear_dirty_mask(m, pagebits);
5207 vm_page_clear_dirty(vm_page_t m, int base, int size)
5210 vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
5214 * vm_page_set_invalid:
5216 * Invalidates DEV_BSIZE'd chunks within a page. Both the
5217 * valid and dirty bits for the effected areas are cleared.
5220 vm_page_set_invalid(vm_page_t m, int base, int size)
5222 vm_page_bits_t bits;
5226 * The object lock is required so that pages can't be mapped
5227 * read-only while we're in the process of invalidating them.
5230 VM_OBJECT_ASSERT_WLOCKED(object);
5231 vm_page_assert_busied(m);
5233 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
5234 size >= object->un_pager.vnp.vnp_size)
5235 bits = VM_PAGE_BITS_ALL;
5237 bits = vm_page_bits(base, size);
5238 if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0)
5240 KASSERT((bits == 0 && vm_page_all_valid(m)) ||
5241 !pmap_page_is_mapped(m),
5242 ("vm_page_set_invalid: page %p is mapped", m));
5243 if (vm_page_xbusied(m)) {
5247 vm_page_bits_clear(m, &m->valid, bits);
5248 vm_page_bits_clear(m, &m->dirty, bits);
5255 * Invalidates the entire page. The page must be busy, unmapped, and
5256 * the enclosing object must be locked. The object locks protects
5257 * against concurrent read-only pmap enter which is done without
5261 vm_page_invalid(vm_page_t m)
5264 vm_page_assert_busied(m);
5265 VM_OBJECT_ASSERT_LOCKED(m->object);
5266 MPASS(!pmap_page_is_mapped(m));
5268 if (vm_page_xbusied(m))
5271 vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL);
5275 * vm_page_zero_invalid()
5277 * The kernel assumes that the invalid portions of a page contain
5278 * garbage, but such pages can be mapped into memory by user code.
5279 * When this occurs, we must zero out the non-valid portions of the
5280 * page so user code sees what it expects.
5282 * Pages are most often semi-valid when the end of a file is mapped
5283 * into memory and the file's size is not page aligned.
5286 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
5292 * Scan the valid bits looking for invalid sections that
5293 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the
5294 * valid bit may be set ) have already been zeroed by
5295 * vm_page_set_validclean().
5297 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
5298 if (i == (PAGE_SIZE / DEV_BSIZE) ||
5299 (m->valid & ((vm_page_bits_t)1 << i))) {
5301 pmap_zero_page_area(m,
5302 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
5309 * setvalid is TRUE when we can safely set the zero'd areas
5310 * as being valid. We can do this if there are no cache consistancy
5311 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
5320 * Is (partial) page valid? Note that the case where size == 0
5321 * will return FALSE in the degenerate case where the page is
5322 * entirely invalid, and TRUE otherwise.
5324 * Some callers envoke this routine without the busy lock held and
5325 * handle races via higher level locks. Typical callers should
5326 * hold a busy lock to prevent invalidation.
5329 vm_page_is_valid(vm_page_t m, int base, int size)
5331 vm_page_bits_t bits;
5333 bits = vm_page_bits(base, size);
5334 return (m->valid != 0 && (m->valid & bits) == bits);
5338 * Returns true if all of the specified predicates are true for the entire
5339 * (super)page and false otherwise.
5342 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
5348 if (skip_m != NULL && skip_m->object != object)
5350 VM_OBJECT_ASSERT_LOCKED(object);
5351 npages = atop(pagesizes[m->psind]);
5354 * The physically contiguous pages that make up a superpage, i.e., a
5355 * page with a page size index ("psind") greater than zero, will
5356 * occupy adjacent entries in vm_page_array[].
5358 for (i = 0; i < npages; i++) {
5359 /* Always test object consistency, including "skip_m". */
5360 if (m[i].object != object)
5362 if (&m[i] == skip_m)
5364 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
5366 if ((flags & PS_ALL_DIRTY) != 0) {
5368 * Calling vm_page_test_dirty() or pmap_is_modified()
5369 * might stop this case from spuriously returning
5370 * "false". However, that would require a write lock
5371 * on the object containing "m[i]".
5373 if (m[i].dirty != VM_PAGE_BITS_ALL)
5376 if ((flags & PS_ALL_VALID) != 0 &&
5377 m[i].valid != VM_PAGE_BITS_ALL)
5384 * Set the page's dirty bits if the page is modified.
5387 vm_page_test_dirty(vm_page_t m)
5390 vm_page_assert_busied(m);
5391 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
5396 vm_page_valid(vm_page_t m)
5399 vm_page_assert_busied(m);
5400 if (vm_page_xbusied(m))
5401 m->valid = VM_PAGE_BITS_ALL;
5403 vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL);
5407 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
5410 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
5414 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
5417 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
5421 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
5424 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
5427 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
5429 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
5432 vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
5436 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
5439 mtx_assert_(vm_page_lockptr(m), a, file, line);
5445 vm_page_object_busy_assert(vm_page_t m)
5449 * Certain of the page's fields may only be modified by the
5450 * holder of a page or object busy.
5452 if (m->object != NULL && !vm_page_busied(m))
5453 VM_OBJECT_ASSERT_BUSY(m->object);
5457 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits)
5460 if ((bits & PGA_WRITEABLE) == 0)
5464 * The PGA_WRITEABLE flag can only be set if the page is
5465 * managed, is exclusively busied or the object is locked.
5466 * Currently, this flag is only set by pmap_enter().
5468 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5469 ("PGA_WRITEABLE on unmanaged page"));
5470 if (!vm_page_xbusied(m))
5471 VM_OBJECT_ASSERT_BUSY(m->object);
5475 #include "opt_ddb.h"
5477 #include <sys/kernel.h>
5479 #include <ddb/ddb.h>
5481 DB_SHOW_COMMAND(page, vm_page_print_page_info)
5484 db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
5485 db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
5486 db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
5487 db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
5488 db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
5489 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
5490 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
5491 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
5492 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
5495 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
5499 db_printf("pq_free %d\n", vm_free_count());
5500 for (dom = 0; dom < vm_ndomains; dom++) {
5502 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
5504 vm_dom[dom].vmd_page_count,
5505 vm_dom[dom].vmd_free_count,
5506 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
5507 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
5508 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
5509 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
5513 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
5516 boolean_t phys, virt;
5519 db_printf("show pginfo addr\n");
5523 phys = strchr(modif, 'p') != NULL;
5524 virt = strchr(modif, 'v') != NULL;
5526 m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
5528 m = PHYS_TO_VM_PAGE(addr);
5530 m = (vm_page_t)addr;
5532 "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n"
5533 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
5534 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
5535 m->a.queue, m->ref_count, m->a.flags, m->oflags,
5536 m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty);