2 * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
4 * Copyright (c) 1991 Regents of the University of California.
6 * Copyright (c) 1998 Matthew Dillon. All Rights Reserved.
8 * This code is derived from software contributed to Berkeley by
9 * The Mach Operating System project at Carnegie-Mellon University.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
14 * 1. Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
16 * 2. Redistributions in binary form must reproduce the above copyright
17 * notice, this list of conditions and the following disclaimer in the
18 * documentation and/or other materials provided with the distribution.
19 * 3. Neither the name of the University nor the names of its contributors
20 * may be used to endorse or promote products derived from this software
21 * without specific prior written permission.
23 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
24 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
27 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91
39 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40 * All rights reserved.
42 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
44 * Permission to use, copy, modify and distribute this software and
45 * its documentation is hereby granted, provided that both the copyright
46 * notice and this permission notice appear in all copies of the
47 * software, derivative works or modified versions, and any portions
48 * thereof, and that both notices appear in supporting documentation.
50 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
54 * Carnegie Mellon requests users of this software to return to
56 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
57 * School of Computer Science
58 * Carnegie Mellon University
59 * Pittsburgh PA 15213-3890
61 * any improvements or extensions that they make and grant Carnegie the
62 * rights to redistribute these changes.
66 * Resident memory management module.
69 #include <sys/cdefs.h>
70 __FBSDID("$FreeBSD$");
74 #include <sys/param.h>
75 #include <sys/systm.h>
76 #include <sys/counter.h>
77 #include <sys/domainset.h>
78 #include <sys/kernel.h>
79 #include <sys/limits.h>
80 #include <sys/linker.h>
82 #include <sys/malloc.h>
84 #include <sys/msgbuf.h>
85 #include <sys/mutex.h>
87 #include <sys/rwlock.h>
88 #include <sys/sleepqueue.h>
90 #include <sys/sched.h>
92 #include <sys/sysctl.h>
93 #include <sys/vmmeter.h>
94 #include <sys/vnode.h>
98 #include <vm/vm_param.h>
99 #include <vm/vm_domainset.h>
100 #include <vm/vm_kern.h>
101 #include <vm/vm_map.h>
102 #include <vm/vm_object.h>
103 #include <vm/vm_page.h>
104 #include <vm/vm_pageout.h>
105 #include <vm/vm_phys.h>
106 #include <vm/vm_pagequeue.h>
107 #include <vm/vm_pager.h>
108 #include <vm/vm_radix.h>
109 #include <vm/vm_reserv.h>
110 #include <vm/vm_extern.h>
112 #include <vm/uma_int.h>
114 #include <machine/md_var.h>
116 struct vm_domain vm_dom[MAXMEMDOM];
118 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]);
120 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
122 struct mtx_padalign __exclusive_cache_line vm_domainset_lock;
123 /* The following fields are protected by the domainset lock. */
124 domainset_t __exclusive_cache_line vm_min_domains;
125 domainset_t __exclusive_cache_line vm_severe_domains;
126 static int vm_min_waiters;
127 static int vm_severe_waiters;
128 static int vm_pageproc_waiters;
130 static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD, 0,
131 "VM page statistics");
133 static counter_u64_t pqstate_commit_retries = EARLY_COUNTER;
134 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries,
135 CTLFLAG_RD, &pqstate_commit_retries,
136 "Number of failed per-page atomic queue state updates");
138 static counter_u64_t queue_ops = EARLY_COUNTER;
139 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops,
140 CTLFLAG_RD, &queue_ops,
141 "Number of batched queue operations");
143 static counter_u64_t queue_nops = EARLY_COUNTER;
144 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops,
145 CTLFLAG_RD, &queue_nops,
146 "Number of batched queue operations with no effects");
149 counter_startup(void)
152 pqstate_commit_retries = counter_u64_alloc(M_WAITOK);
153 queue_ops = counter_u64_alloc(M_WAITOK);
154 queue_nops = counter_u64_alloc(M_WAITOK);
156 SYSINIT(page_counters, SI_SUB_CPU, SI_ORDER_ANY, counter_startup, NULL);
159 * bogus page -- for I/O to/from partially complete buffers,
160 * or for paging into sparsely invalid regions.
162 vm_page_t bogus_page;
164 vm_page_t vm_page_array;
165 long vm_page_array_size;
168 static TAILQ_HEAD(, vm_page) blacklist_head;
169 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
170 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
171 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
173 static uma_zone_t fakepg_zone;
175 static void vm_page_alloc_check(vm_page_t m);
176 static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m,
177 vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked);
178 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
179 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
180 static bool vm_page_free_prep(vm_page_t m);
181 static void vm_page_free_toq(vm_page_t m);
182 static void vm_page_init(void *dummy);
183 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
184 vm_pindex_t pindex, vm_page_t mpred);
185 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
187 static void vm_page_mvqueue(vm_page_t m, const uint8_t queue,
188 const uint16_t nflag);
189 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
190 vm_page_t m_run, vm_paddr_t high);
191 static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse);
192 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
194 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain,
196 static void vm_page_zone_release(void *arg, void **store, int cnt);
198 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
201 vm_page_init(void *dummy)
204 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
205 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
206 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
207 VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
211 * The cache page zone is initialized later since we need to be able to allocate
212 * pages before UMA is fully initialized.
215 vm_page_init_cache_zones(void *dummy __unused)
217 struct vm_domain *vmd;
218 struct vm_pgcache *pgcache;
219 int cache, domain, maxcache, pool;
222 TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &maxcache);
223 maxcache *= mp_ncpus;
224 for (domain = 0; domain < vm_ndomains; domain++) {
225 vmd = VM_DOMAIN(domain);
226 for (pool = 0; pool < VM_NFREEPOOL; pool++) {
227 pgcache = &vmd->vmd_pgcache[pool];
228 pgcache->domain = domain;
229 pgcache->pool = pool;
230 pgcache->zone = uma_zcache_create("vm pgcache",
231 PAGE_SIZE, NULL, NULL, NULL, NULL,
232 vm_page_zone_import, vm_page_zone_release, pgcache,
236 * Limit each pool's zone to 0.1% of the pages in the
239 cache = maxcache != 0 ? maxcache :
240 vmd->vmd_page_count / 1000;
241 uma_zone_set_maxcache(pgcache->zone, cache);
245 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
247 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
248 #if PAGE_SIZE == 32768
250 CTASSERT(sizeof(u_long) >= 8);
257 * Sets the page size, perhaps based upon the memory
258 * size. Must be called before any use of page-size
259 * dependent functions.
262 vm_set_page_size(void)
264 if (vm_cnt.v_page_size == 0)
265 vm_cnt.v_page_size = PAGE_SIZE;
266 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
267 panic("vm_set_page_size: page size not a power of two");
271 * vm_page_blacklist_next:
273 * Find the next entry in the provided string of blacklist
274 * addresses. Entries are separated by space, comma, or newline.
275 * If an invalid integer is encountered then the rest of the
276 * string is skipped. Updates the list pointer to the next
277 * character, or NULL if the string is exhausted or invalid.
280 vm_page_blacklist_next(char **list, char *end)
285 if (list == NULL || *list == NULL)
293 * If there's no end pointer then the buffer is coming from
294 * the kenv and we know it's null-terminated.
297 end = *list + strlen(*list);
299 /* Ensure that strtoq() won't walk off the end */
301 if (*end == '\n' || *end == ' ' || *end == ',')
304 printf("Blacklist not terminated, skipping\n");
310 for (pos = *list; *pos != '\0'; pos = cp) {
311 bad = strtoq(pos, &cp, 0);
312 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
321 if (*cp == '\0' || ++cp >= end)
325 return (trunc_page(bad));
327 printf("Garbage in RAM blacklist, skipping\n");
333 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
335 struct vm_domain *vmd;
339 m = vm_phys_paddr_to_vm_page(pa);
341 return (true); /* page does not exist, no failure */
343 vmd = vm_pagequeue_domain(m);
344 vm_domain_free_lock(vmd);
345 ret = vm_phys_unfree_page(m);
346 vm_domain_free_unlock(vmd);
348 vm_domain_freecnt_inc(vmd, -1);
349 TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
351 printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
357 * vm_page_blacklist_check:
359 * Iterate through the provided string of blacklist addresses, pulling
360 * each entry out of the physical allocator free list and putting it
361 * onto a list for reporting via the vm.page_blacklist sysctl.
364 vm_page_blacklist_check(char *list, char *end)
370 while (next != NULL) {
371 if ((pa = vm_page_blacklist_next(&next, end)) == 0)
373 vm_page_blacklist_add(pa, bootverbose);
378 * vm_page_blacklist_load:
380 * Search for a special module named "ram_blacklist". It'll be a
381 * plain text file provided by the user via the loader directive
385 vm_page_blacklist_load(char **list, char **end)
394 mod = preload_search_by_type("ram_blacklist");
396 ptr = preload_fetch_addr(mod);
397 len = preload_fetch_size(mod);
408 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
415 error = sysctl_wire_old_buffer(req, 0);
418 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
419 TAILQ_FOREACH(m, &blacklist_head, listq) {
420 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
421 (uintmax_t)m->phys_addr);
424 error = sbuf_finish(&sbuf);
430 * Initialize a dummy page for use in scans of the specified paging queue.
431 * In principle, this function only needs to set the flag PG_MARKER.
432 * Nonetheless, it write busies the page as a safety precaution.
435 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags)
438 bzero(marker, sizeof(*marker));
439 marker->flags = PG_MARKER;
440 marker->a.flags = aflags;
441 marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
442 marker->a.queue = queue;
446 vm_page_domain_init(int domain)
448 struct vm_domain *vmd;
449 struct vm_pagequeue *pq;
452 vmd = VM_DOMAIN(domain);
453 bzero(vmd, sizeof(*vmd));
454 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
455 "vm inactive pagequeue";
456 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
457 "vm active pagequeue";
458 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
459 "vm laundry pagequeue";
460 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
461 "vm unswappable pagequeue";
462 vmd->vmd_domain = domain;
463 vmd->vmd_page_count = 0;
464 vmd->vmd_free_count = 0;
466 vmd->vmd_oom = FALSE;
467 for (i = 0; i < PQ_COUNT; i++) {
468 pq = &vmd->vmd_pagequeues[i];
469 TAILQ_INIT(&pq->pq_pl);
470 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
471 MTX_DEF | MTX_DUPOK);
473 vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
475 mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
476 mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
477 snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
480 * inacthead is used to provide FIFO ordering for LRU-bypassing
483 vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
484 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
485 &vmd->vmd_inacthead, plinks.q);
488 * The clock pages are used to implement active queue scanning without
489 * requeues. Scans start at clock[0], which is advanced after the scan
490 * ends. When the two clock hands meet, they are reset and scanning
491 * resumes from the head of the queue.
493 vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
494 vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
495 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
496 &vmd->vmd_clock[0], plinks.q);
497 TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
498 &vmd->vmd_clock[1], plinks.q);
502 * Initialize a physical page in preparation for adding it to the free
506 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
511 m->busy_lock = VPB_FREED;
512 m->flags = m->a.flags = 0;
514 m->a.queue = PQ_NONE;
517 m->order = VM_NFREEORDER;
518 m->pool = VM_FREEPOOL_DEFAULT;
519 m->valid = m->dirty = 0;
523 #ifndef PMAP_HAS_PAGE_ARRAY
525 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range)
530 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
531 * However, because this page is allocated from KVM, out-of-bounds
532 * accesses using the direct map will not be trapped.
537 * Allocate physical memory for the page structures, and map it.
539 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
540 vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end,
541 VM_PROT_READ | VM_PROT_WRITE);
542 vm_page_array_size = page_range;
551 * Initializes the resident memory module. Allocates physical memory for
552 * bootstrapping UMA and some data structures that are used to manage
553 * physical pages. Initializes these structures, and populates the free
557 vm_page_startup(vm_offset_t vaddr)
559 struct vm_phys_seg *seg;
561 char *list, *listend;
562 vm_paddr_t end, high_avail, low_avail, new_end, size;
563 vm_paddr_t page_range __unused;
564 vm_paddr_t last_pa, pa;
566 int biggestone, i, segind;
571 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
575 vaddr = round_page(vaddr);
577 vm_phys_early_startup();
578 biggestone = vm_phys_avail_largest();
579 end = phys_avail[biggestone+1];
582 * Initialize the page and queue locks.
584 mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
585 for (i = 0; i < PA_LOCK_COUNT; i++)
586 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
587 for (i = 0; i < vm_ndomains; i++)
588 vm_page_domain_init(i);
592 witness_size = round_page(witness_startup_count());
593 new_end -= witness_size;
594 mapped = pmap_map(&vaddr, new_end, new_end + witness_size,
595 VM_PROT_READ | VM_PROT_WRITE);
596 bzero((void *)mapped, witness_size);
597 witness_startup((void *)mapped);
600 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
601 defined(__i386__) || defined(__mips__) || defined(__riscv) || \
602 defined(__powerpc64__)
604 * Allocate a bitmap to indicate that a random physical page
605 * needs to be included in a minidump.
607 * The amd64 port needs this to indicate which direct map pages
608 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
610 * However, i386 still needs this workspace internally within the
611 * minidump code. In theory, they are not needed on i386, but are
612 * included should the sf_buf code decide to use them.
615 for (i = 0; dump_avail[i + 1] != 0; i += 2)
616 if (dump_avail[i + 1] > last_pa)
617 last_pa = dump_avail[i + 1];
618 page_range = last_pa / PAGE_SIZE;
619 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
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 to the free lists only if it is covered by
775 * one of the ranges in phys_avail. Because we've added the
776 * ranges to the vm_phys_segs array, we can assume that each
777 * segment is either entirely contained in one of the ranges,
778 * or doesn't overlap any of them.
780 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
781 struct vm_domain *vmd;
783 if (seg->start < phys_avail[i] ||
784 seg->end > phys_avail[i + 1])
788 pagecount = (u_long)atop(seg->end - seg->start);
790 vmd = VM_DOMAIN(seg->domain);
791 vm_domain_free_lock(vmd);
792 vm_phys_enqueue_contig(m, pagecount);
793 vm_domain_free_unlock(vmd);
794 vm_domain_freecnt_inc(vmd, pagecount);
795 vm_cnt.v_page_count += (u_int)pagecount;
797 vmd = VM_DOMAIN(seg->domain);
798 vmd->vmd_page_count += (u_int)pagecount;
799 vmd->vmd_segs |= 1UL << m->segind;
805 * Remove blacklisted pages from the physical memory allocator.
807 TAILQ_INIT(&blacklist_head);
808 vm_page_blacklist_load(&list, &listend);
809 vm_page_blacklist_check(list, listend);
811 list = kern_getenv("vm.blacklist");
812 vm_page_blacklist_check(list, NULL);
815 #if VM_NRESERVLEVEL > 0
817 * Initialize the reservation management system.
826 vm_page_reference(vm_page_t m)
829 vm_page_aflag_set(m, PGA_REFERENCED);
833 vm_page_acquire_flags(vm_page_t m, int allocflags)
837 if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0)
838 locked = vm_page_trysbusy(m);
840 locked = vm_page_tryxbusy(m);
841 if (locked && (allocflags & VM_ALLOC_WIRED) != 0)
847 * vm_page_busy_sleep_flags
849 * Sleep for busy according to VM_ALLOC_ parameters. Returns true
850 * if the caller should retry and false otherwise.
853 vm_page_busy_sleep_flags(vm_object_t object, vm_page_t m, const char *wmesg,
857 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
861 * Reference the page before unlocking and sleeping so that
862 * the page daemon is less likely to reclaim it.
864 if ((allocflags & VM_ALLOC_NOCREAT) == 0)
865 vm_page_reference(m);
867 if (_vm_page_busy_sleep(object, m, m->pindex, wmesg, allocflags, true))
868 VM_OBJECT_WLOCK(object);
869 if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
876 * vm_page_busy_acquire:
878 * Acquire the busy lock as described by VM_ALLOC_* flags. Will loop
879 * and drop the object lock if necessary.
882 vm_page_busy_acquire(vm_page_t m, int allocflags)
888 * The page-specific object must be cached because page
889 * identity can change during the sleep, causing the
890 * re-lock of a different object.
891 * It is assumed that a reference to the object is already
892 * held by the callers.
896 if (vm_page_acquire_flags(m, allocflags))
898 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
901 locked = VM_OBJECT_WOWNED(obj);
904 MPASS(locked || vm_page_wired(m));
905 if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags,
907 VM_OBJECT_WLOCK(obj);
908 if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
910 KASSERT(m->object == obj || m->object == NULL,
911 ("vm_page_busy_acquire: page %p does not belong to %p",
917 * vm_page_busy_downgrade:
919 * Downgrade an exclusive busy page into a single shared busy page.
922 vm_page_busy_downgrade(vm_page_t m)
926 vm_page_assert_xbusied(m);
930 if (atomic_fcmpset_rel_int(&m->busy_lock,
931 &x, VPB_SHARERS_WORD(1)))
934 if ((x & VPB_BIT_WAITERS) != 0)
940 * vm_page_busy_tryupgrade:
942 * Attempt to upgrade a single shared busy into an exclusive busy.
945 vm_page_busy_tryupgrade(vm_page_t m)
949 vm_page_assert_sbusied(m);
952 ce = VPB_CURTHREAD_EXCLUSIVE;
954 if (VPB_SHARERS(x) > 1)
956 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
957 ("vm_page_busy_tryupgrade: invalid lock state"));
958 if (!atomic_fcmpset_acq_int(&m->busy_lock, &x,
959 ce | (x & VPB_BIT_WAITERS)))
968 * Return a positive value if the page is shared busied, 0 otherwise.
971 vm_page_sbusied(vm_page_t m)
976 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
982 * Shared unbusy a page.
985 vm_page_sunbusy(vm_page_t m)
989 vm_page_assert_sbusied(m);
993 KASSERT(x != VPB_FREED,
994 ("vm_page_sunbusy: Unlocking freed page."));
995 if (VPB_SHARERS(x) > 1) {
996 if (atomic_fcmpset_int(&m->busy_lock, &x,
1001 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
1002 ("vm_page_sunbusy: invalid lock state"));
1003 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1005 if ((x & VPB_BIT_WAITERS) == 0)
1013 * vm_page_busy_sleep:
1015 * Sleep if the page is busy, using the page pointer as wchan.
1016 * This is used to implement the hard-path of busying mechanism.
1018 * If nonshared is true, sleep only if the page is xbusy.
1020 * The object lock must be held on entry and will be released on exit.
1023 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
1028 VM_OBJECT_ASSERT_LOCKED(obj);
1029 vm_page_lock_assert(m, MA_NOTOWNED);
1031 if (!_vm_page_busy_sleep(obj, m, m->pindex, wmesg,
1032 nonshared ? VM_ALLOC_SBUSY : 0 , true))
1033 VM_OBJECT_DROP(obj);
1037 * vm_page_busy_sleep_unlocked:
1039 * Sleep if the page is busy, using the page pointer as wchan.
1040 * This is used to implement the hard-path of busying mechanism.
1042 * If nonshared is true, sleep only if the page is xbusy.
1044 * The object lock must not be held on entry. The operation will
1045 * return if the page changes identity.
1048 vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1049 const char *wmesg, bool nonshared)
1052 VM_OBJECT_ASSERT_UNLOCKED(obj);
1053 vm_page_lock_assert(m, MA_NOTOWNED);
1055 _vm_page_busy_sleep(obj, m, pindex, wmesg,
1056 nonshared ? VM_ALLOC_SBUSY : 0, false);
1060 * _vm_page_busy_sleep:
1062 * Internal busy sleep function. Verifies the page identity and
1063 * lockstate against parameters. Returns true if it sleeps and
1066 * If locked is true the lock will be dropped for any true returns
1067 * and held for any false returns.
1070 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1071 const char *wmesg, int allocflags, bool locked)
1077 * If the object is busy we must wait for that to drain to zero
1078 * before trying the page again.
1080 if (obj != NULL && vm_object_busied(obj)) {
1082 VM_OBJECT_DROP(obj);
1083 vm_object_busy_wait(obj, wmesg);
1087 if (!vm_page_busied(m))
1090 xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0;
1092 x = atomic_load_int(&m->busy_lock);
1095 * If the page changes objects or becomes unlocked we can
1098 if (x == VPB_UNBUSIED ||
1099 (xsleep && (x & VPB_BIT_SHARED) != 0) ||
1100 m->object != obj || m->pindex != pindex) {
1104 if ((x & VPB_BIT_WAITERS) != 0)
1106 } while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS));
1108 VM_OBJECT_DROP(obj);
1110 sleepq_add(m, NULL, wmesg, 0, 0);
1111 sleepq_wait(m, PVM);
1119 * Try to shared busy a page.
1120 * If the operation succeeds 1 is returned otherwise 0.
1121 * The operation never sleeps.
1124 vm_page_trysbusy(vm_page_t m)
1132 if ((x & VPB_BIT_SHARED) == 0)
1135 * Reduce the window for transient busies that will trigger
1136 * false negatives in vm_page_ps_test().
1138 if (obj != NULL && vm_object_busied(obj))
1140 if (atomic_fcmpset_acq_int(&m->busy_lock, &x,
1141 x + VPB_ONE_SHARER))
1145 /* Refetch the object now that we're guaranteed that it is stable. */
1147 if (obj != NULL && vm_object_busied(obj)) {
1157 * Try to exclusive busy a page.
1158 * If the operation succeeds 1 is returned otherwise 0.
1159 * The operation never sleeps.
1162 vm_page_tryxbusy(vm_page_t m)
1166 if (atomic_cmpset_acq_int(&(m)->busy_lock, VPB_UNBUSIED,
1167 VPB_CURTHREAD_EXCLUSIVE) == 0)
1171 if (obj != NULL && vm_object_busied(obj)) {
1179 vm_page_xunbusy_hard_tail(vm_page_t m)
1181 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1182 /* Wake the waiter. */
1187 * vm_page_xunbusy_hard:
1189 * Called when unbusy has failed because there is a waiter.
1192 vm_page_xunbusy_hard(vm_page_t m)
1194 vm_page_assert_xbusied(m);
1195 vm_page_xunbusy_hard_tail(m);
1199 vm_page_xunbusy_hard_unchecked(vm_page_t m)
1201 vm_page_assert_xbusied_unchecked(m);
1202 vm_page_xunbusy_hard_tail(m);
1206 vm_page_busy_free(vm_page_t m)
1210 atomic_thread_fence_rel();
1211 x = atomic_swap_int(&m->busy_lock, VPB_FREED);
1212 if ((x & VPB_BIT_WAITERS) != 0)
1217 * vm_page_unhold_pages:
1219 * Unhold each of the pages that is referenced by the given array.
1222 vm_page_unhold_pages(vm_page_t *ma, int count)
1225 for (; count != 0; count--) {
1226 vm_page_unwire(*ma, PQ_ACTIVE);
1232 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1236 #ifdef VM_PHYSSEG_SPARSE
1237 m = vm_phys_paddr_to_vm_page(pa);
1239 m = vm_phys_fictitious_to_vm_page(pa);
1241 #elif defined(VM_PHYSSEG_DENSE)
1245 if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1246 m = &vm_page_array[pi - first_page];
1249 return (vm_phys_fictitious_to_vm_page(pa));
1251 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1258 * Create a fictitious page with the specified physical address and
1259 * memory attribute. The memory attribute is the only the machine-
1260 * dependent aspect of a fictitious page that must be initialized.
1263 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1267 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1268 vm_page_initfake(m, paddr, memattr);
1273 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1276 if ((m->flags & PG_FICTITIOUS) != 0) {
1278 * The page's memattr might have changed since the
1279 * previous initialization. Update the pmap to the
1284 m->phys_addr = paddr;
1285 m->a.queue = PQ_NONE;
1286 /* Fictitious pages don't use "segind". */
1287 m->flags = PG_FICTITIOUS;
1288 /* Fictitious pages don't use "order" or "pool". */
1289 m->oflags = VPO_UNMANAGED;
1290 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
1291 /* Fictitious pages are unevictable. */
1295 pmap_page_set_memattr(m, memattr);
1301 * Release a fictitious page.
1304 vm_page_putfake(vm_page_t m)
1307 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1308 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1309 ("vm_page_putfake: bad page %p", m));
1310 vm_page_assert_xbusied(m);
1311 vm_page_busy_free(m);
1312 uma_zfree(fakepg_zone, m);
1316 * vm_page_updatefake:
1318 * Update the given fictitious page to the specified physical address and
1322 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1325 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1326 ("vm_page_updatefake: bad page %p", m));
1327 m->phys_addr = paddr;
1328 pmap_page_set_memattr(m, memattr);
1337 vm_page_free(vm_page_t m)
1340 m->flags &= ~PG_ZERO;
1341 vm_page_free_toq(m);
1345 * vm_page_free_zero:
1347 * Free a page to the zerod-pages queue
1350 vm_page_free_zero(vm_page_t m)
1353 m->flags |= PG_ZERO;
1354 vm_page_free_toq(m);
1358 * Unbusy and handle the page queueing for a page from a getpages request that
1359 * was optionally read ahead or behind.
1362 vm_page_readahead_finish(vm_page_t m)
1365 /* We shouldn't put invalid pages on queues. */
1366 KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m));
1369 * Since the page is not the actually needed one, whether it should
1370 * be activated or deactivated is not obvious. Empirical results
1371 * have shown that deactivating the page is usually the best choice,
1372 * unless the page is wanted by another thread.
1374 if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1375 vm_page_activate(m);
1377 vm_page_deactivate(m);
1378 vm_page_xunbusy_unchecked(m);
1382 * vm_page_sleep_if_busy:
1384 * Sleep and release the object lock if the page is busied.
1385 * Returns TRUE if the thread slept.
1387 * The given page must be unlocked and object containing it must
1391 vm_page_sleep_if_busy(vm_page_t m, const char *wmesg)
1395 vm_page_lock_assert(m, MA_NOTOWNED);
1396 VM_OBJECT_ASSERT_WLOCKED(m->object);
1399 * The page-specific object must be cached because page
1400 * identity can change during the sleep, causing the
1401 * re-lock of a different object.
1402 * It is assumed that a reference to the object is already
1403 * held by the callers.
1406 if (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, 0, true)) {
1407 VM_OBJECT_WLOCK(obj);
1414 * vm_page_sleep_if_xbusy:
1416 * Sleep and release the object lock if the page is xbusied.
1417 * Returns TRUE if the thread slept.
1419 * The given page must be unlocked and object containing it must
1423 vm_page_sleep_if_xbusy(vm_page_t m, const char *wmesg)
1427 vm_page_lock_assert(m, MA_NOTOWNED);
1428 VM_OBJECT_ASSERT_WLOCKED(m->object);
1431 * The page-specific object must be cached because page
1432 * identity can change during the sleep, causing the
1433 * re-lock of a different object.
1434 * It is assumed that a reference to the object is already
1435 * held by the callers.
1438 if (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, VM_ALLOC_SBUSY,
1440 VM_OBJECT_WLOCK(obj);
1447 * vm_page_dirty_KBI: [ internal use only ]
1449 * Set all bits in the page's dirty field.
1451 * The object containing the specified page must be locked if the
1452 * call is made from the machine-independent layer.
1454 * See vm_page_clear_dirty_mask().
1456 * This function should only be called by vm_page_dirty().
1459 vm_page_dirty_KBI(vm_page_t m)
1462 /* Refer to this operation by its public name. */
1463 KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!"));
1464 m->dirty = VM_PAGE_BITS_ALL;
1468 * vm_page_insert: [ internal use only ]
1470 * Inserts the given mem entry into the object and object list.
1472 * The object must be locked.
1475 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1479 VM_OBJECT_ASSERT_WLOCKED(object);
1480 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1481 return (vm_page_insert_after(m, object, pindex, mpred));
1485 * vm_page_insert_after:
1487 * Inserts the page "m" into the specified object at offset "pindex".
1489 * The page "mpred" must immediately precede the offset "pindex" within
1490 * the specified object.
1492 * The object must be locked.
1495 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1500 VM_OBJECT_ASSERT_WLOCKED(object);
1501 KASSERT(m->object == NULL,
1502 ("vm_page_insert_after: page already inserted"));
1503 if (mpred != NULL) {
1504 KASSERT(mpred->object == object,
1505 ("vm_page_insert_after: object doesn't contain mpred"));
1506 KASSERT(mpred->pindex < pindex,
1507 ("vm_page_insert_after: mpred doesn't precede pindex"));
1508 msucc = TAILQ_NEXT(mpred, listq);
1510 msucc = TAILQ_FIRST(&object->memq);
1512 KASSERT(msucc->pindex > pindex,
1513 ("vm_page_insert_after: msucc doesn't succeed pindex"));
1516 * Record the object/offset pair in this page.
1520 m->ref_count |= VPRC_OBJREF;
1523 * Now link into the object's ordered list of backed pages.
1525 if (vm_radix_insert(&object->rtree, m)) {
1528 m->ref_count &= ~VPRC_OBJREF;
1531 vm_page_insert_radixdone(m, object, mpred);
1536 * vm_page_insert_radixdone:
1538 * Complete page "m" insertion into the specified object after the
1539 * radix trie hooking.
1541 * The page "mpred" must precede the offset "m->pindex" within the
1544 * The object must be locked.
1547 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1550 VM_OBJECT_ASSERT_WLOCKED(object);
1551 KASSERT(object != NULL && m->object == object,
1552 ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1553 KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1554 ("vm_page_insert_radixdone: page %p is missing object ref", m));
1555 if (mpred != NULL) {
1556 KASSERT(mpred->object == object,
1557 ("vm_page_insert_radixdone: object doesn't contain mpred"));
1558 KASSERT(mpred->pindex < m->pindex,
1559 ("vm_page_insert_radixdone: mpred doesn't precede pindex"));
1563 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1565 TAILQ_INSERT_HEAD(&object->memq, m, listq);
1568 * Show that the object has one more resident page.
1570 object->resident_page_count++;
1573 * Hold the vnode until the last page is released.
1575 if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1576 vhold(object->handle);
1579 * Since we are inserting a new and possibly dirty page,
1580 * update the object's generation count.
1582 if (pmap_page_is_write_mapped(m))
1583 vm_object_set_writeable_dirty(object);
1587 * Do the work to remove a page from its object. The caller is responsible for
1588 * updating the page's fields to reflect this removal.
1591 vm_page_object_remove(vm_page_t m)
1596 vm_page_assert_xbusied(m);
1598 VM_OBJECT_ASSERT_WLOCKED(object);
1599 KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1600 ("page %p is missing its object ref", m));
1602 /* Deferred free of swap space. */
1603 if ((m->a.flags & PGA_SWAP_FREE) != 0)
1604 vm_pager_page_unswapped(m);
1606 mrem = vm_radix_remove(&object->rtree, m->pindex);
1607 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1610 * Now remove from the object's list of backed pages.
1612 TAILQ_REMOVE(&object->memq, m, listq);
1615 * And show that the object has one fewer resident page.
1617 object->resident_page_count--;
1620 * The vnode may now be recycled.
1622 if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1623 vdrop(object->handle);
1629 * Removes the specified page from its containing object, but does not
1630 * invalidate any backing storage. Returns true if the object's reference
1631 * was the last reference to the page, and false otherwise.
1633 * The object must be locked and the page must be exclusively busied.
1634 * The exclusive busy will be released on return. If this is not the
1635 * final ref and the caller does not hold a wire reference it may not
1636 * continue to access the page.
1639 vm_page_remove(vm_page_t m)
1643 dropped = vm_page_remove_xbusy(m);
1650 * vm_page_remove_xbusy
1652 * Removes the page but leaves the xbusy held. Returns true if this
1653 * removed the final ref and false otherwise.
1656 vm_page_remove_xbusy(vm_page_t m)
1659 vm_page_object_remove(m);
1661 return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1667 * Returns the page associated with the object/offset
1668 * pair specified; if none is found, NULL is returned.
1670 * The object must be locked.
1673 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1676 VM_OBJECT_ASSERT_LOCKED(object);
1677 return (vm_radix_lookup(&object->rtree, pindex));
1681 * vm_page_find_least:
1683 * Returns the page associated with the object with least pindex
1684 * greater than or equal to the parameter pindex, or NULL.
1686 * The object must be locked.
1689 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1693 VM_OBJECT_ASSERT_LOCKED(object);
1694 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1695 m = vm_radix_lookup_ge(&object->rtree, pindex);
1700 * Returns the given page's successor (by pindex) within the object if it is
1701 * resident; if none is found, NULL is returned.
1703 * The object must be locked.
1706 vm_page_next(vm_page_t m)
1710 VM_OBJECT_ASSERT_LOCKED(m->object);
1711 if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1712 MPASS(next->object == m->object);
1713 if (next->pindex != m->pindex + 1)
1720 * Returns the given page's predecessor (by pindex) within the object if it is
1721 * resident; if none is found, NULL is returned.
1723 * The object must be locked.
1726 vm_page_prev(vm_page_t m)
1730 VM_OBJECT_ASSERT_LOCKED(m->object);
1731 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1732 MPASS(prev->object == m->object);
1733 if (prev->pindex != m->pindex - 1)
1740 * Uses the page mnew as a replacement for an existing page at index
1741 * pindex which must be already present in the object.
1743 * Both pages must be exclusively busied on enter. The old page is
1746 * A return value of true means mold is now free. If this is not the
1747 * final ref and the caller does not hold a wire reference it may not
1748 * continue to access the page.
1751 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1757 VM_OBJECT_ASSERT_WLOCKED(object);
1758 vm_page_assert_xbusied(mold);
1759 KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
1760 ("vm_page_replace: page %p already in object", mnew));
1763 * This function mostly follows vm_page_insert() and
1764 * vm_page_remove() without the radix, object count and vnode
1765 * dance. Double check such functions for more comments.
1768 mnew->object = object;
1769 mnew->pindex = pindex;
1770 atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
1771 mret = vm_radix_replace(&object->rtree, mnew);
1772 KASSERT(mret == mold,
1773 ("invalid page replacement, mold=%p, mret=%p", mold, mret));
1774 KASSERT((mold->oflags & VPO_UNMANAGED) ==
1775 (mnew->oflags & VPO_UNMANAGED),
1776 ("vm_page_replace: mismatched VPO_UNMANAGED"));
1778 /* Keep the resident page list in sorted order. */
1779 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1780 TAILQ_REMOVE(&object->memq, mold, listq);
1781 mold->object = NULL;
1784 * The object's resident_page_count does not change because we have
1785 * swapped one page for another, but the generation count should
1786 * change if the page is dirty.
1788 if (pmap_page_is_write_mapped(mnew))
1789 vm_object_set_writeable_dirty(object);
1790 dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF;
1791 vm_page_xunbusy(mold);
1797 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1801 vm_page_assert_xbusied(mnew);
1803 if (vm_page_replace_hold(mnew, object, pindex, mold))
1810 * Move the given memory entry from its
1811 * current object to the specified target object/offset.
1813 * Note: swap associated with the page must be invalidated by the move. We
1814 * have to do this for several reasons: (1) we aren't freeing the
1815 * page, (2) we are dirtying the page, (3) the VM system is probably
1816 * moving the page from object A to B, and will then later move
1817 * the backing store from A to B and we can't have a conflict.
1819 * Note: we *always* dirty the page. It is necessary both for the
1820 * fact that we moved it, and because we may be invalidating
1823 * The objects must be locked.
1826 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1831 VM_OBJECT_ASSERT_WLOCKED(new_object);
1833 KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m));
1834 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1835 KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1836 ("vm_page_rename: pindex already renamed"));
1839 * Create a custom version of vm_page_insert() which does not depend
1840 * by m_prev and can cheat on the implementation aspects of the
1844 m->pindex = new_pindex;
1845 if (vm_radix_insert(&new_object->rtree, m)) {
1851 * The operation cannot fail anymore. The removal must happen before
1852 * the listq iterator is tainted.
1855 vm_page_object_remove(m);
1857 /* Return back to the new pindex to complete vm_page_insert(). */
1858 m->pindex = new_pindex;
1859 m->object = new_object;
1861 vm_page_insert_radixdone(m, new_object, mpred);
1869 * Allocate and return a page that is associated with the specified
1870 * object and offset pair. By default, this page is exclusive busied.
1872 * The caller must always specify an allocation class.
1874 * allocation classes:
1875 * VM_ALLOC_NORMAL normal process request
1876 * VM_ALLOC_SYSTEM system *really* needs a page
1877 * VM_ALLOC_INTERRUPT interrupt time request
1879 * optional allocation flags:
1880 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1881 * intends to allocate
1882 * VM_ALLOC_NOBUSY do not exclusive busy the page
1883 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1884 * VM_ALLOC_NOOBJ page is not associated with an object and
1885 * should not be exclusive busy
1886 * VM_ALLOC_SBUSY shared busy the allocated page
1887 * VM_ALLOC_WIRED wire the allocated page
1888 * VM_ALLOC_ZERO prefer a zeroed page
1891 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1894 return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1895 vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1899 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1903 return (vm_page_alloc_domain_after(object, pindex, domain, req,
1904 object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) :
1909 * Allocate a page in the specified object with the given page index. To
1910 * optimize insertion of the page into the object, the caller must also specifiy
1911 * the resident page in the object with largest index smaller than the given
1912 * page index, or NULL if no such page exists.
1915 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1916 int req, vm_page_t mpred)
1918 struct vm_domainset_iter di;
1922 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1924 m = vm_page_alloc_domain_after(object, pindex, domain, req,
1928 } while (vm_domainset_iter_page(&di, object, &domain) == 0);
1934 * Returns true if the number of free pages exceeds the minimum
1935 * for the request class and false otherwise.
1938 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
1940 u_int limit, old, new;
1942 if (req_class == VM_ALLOC_INTERRUPT)
1944 else if (req_class == VM_ALLOC_SYSTEM)
1945 limit = vmd->vmd_interrupt_free_min;
1947 limit = vmd->vmd_free_reserved;
1950 * Attempt to reserve the pages. Fail if we're below the limit.
1953 old = vmd->vmd_free_count;
1958 } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
1960 /* Wake the page daemon if we've crossed the threshold. */
1961 if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
1962 pagedaemon_wakeup(vmd->vmd_domain);
1964 /* Only update bitsets on transitions. */
1965 if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
1966 (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
1973 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
1978 * The page daemon is allowed to dig deeper into the free page list.
1980 req_class = req & VM_ALLOC_CLASS_MASK;
1981 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1982 req_class = VM_ALLOC_SYSTEM;
1983 return (_vm_domain_allocate(vmd, req_class, npages));
1987 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
1988 int req, vm_page_t mpred)
1990 struct vm_domain *vmd;
1994 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1995 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1996 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1997 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1998 ("inconsistent object(%p)/req(%x)", object, req));
1999 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2000 ("Can't sleep and retry object insertion."));
2001 KASSERT(mpred == NULL || mpred->pindex < pindex,
2002 ("mpred %p doesn't precede pindex 0x%jx", mpred,
2003 (uintmax_t)pindex));
2005 VM_OBJECT_ASSERT_WLOCKED(object);
2009 pool = object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT;
2011 #if VM_NRESERVLEVEL > 0
2013 * Can we allocate the page from a reservation?
2015 if (vm_object_reserv(object) &&
2016 (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
2018 domain = vm_phys_domain(m);
2019 vmd = VM_DOMAIN(domain);
2023 vmd = VM_DOMAIN(domain);
2024 if (vmd->vmd_pgcache[pool].zone != NULL) {
2025 m = uma_zalloc(vmd->vmd_pgcache[pool].zone, M_NOWAIT);
2027 flags |= PG_PCPU_CACHE;
2031 if (vm_domain_allocate(vmd, req, 1)) {
2033 * If not, allocate it from the free page queues.
2035 vm_domain_free_lock(vmd);
2036 m = vm_phys_alloc_pages(domain, pool, 0);
2037 vm_domain_free_unlock(vmd);
2039 vm_domain_freecnt_inc(vmd, 1);
2040 #if VM_NRESERVLEVEL > 0
2041 if (vm_reserv_reclaim_inactive(domain))
2048 * Not allocatable, give up.
2050 if (vm_domain_alloc_fail(vmd, object, req))
2056 * At this point we had better have found a good page.
2060 vm_page_alloc_check(m);
2063 * Initialize the page. Only the PG_ZERO flag is inherited.
2065 if ((req & VM_ALLOC_ZERO) != 0)
2066 flags |= (m->flags & PG_ZERO);
2067 if ((req & VM_ALLOC_NODUMP) != 0)
2071 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2073 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2074 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2075 else if ((req & VM_ALLOC_SBUSY) != 0)
2076 m->busy_lock = VPB_SHARERS_WORD(1);
2078 m->busy_lock = VPB_UNBUSIED;
2079 if (req & VM_ALLOC_WIRED) {
2085 if (object != NULL) {
2086 if (vm_page_insert_after(m, object, pindex, mpred)) {
2087 if (req & VM_ALLOC_WIRED) {
2091 KASSERT(m->object == NULL, ("page %p has object", m));
2092 m->oflags = VPO_UNMANAGED;
2093 m->busy_lock = VPB_UNBUSIED;
2094 /* Don't change PG_ZERO. */
2095 vm_page_free_toq(m);
2096 if (req & VM_ALLOC_WAITFAIL) {
2097 VM_OBJECT_WUNLOCK(object);
2099 VM_OBJECT_WLOCK(object);
2104 /* Ignore device objects; the pager sets "memattr" for them. */
2105 if (object->memattr != VM_MEMATTR_DEFAULT &&
2106 (object->flags & OBJ_FICTITIOUS) == 0)
2107 pmap_page_set_memattr(m, object->memattr);
2115 * vm_page_alloc_contig:
2117 * Allocate a contiguous set of physical pages of the given size "npages"
2118 * from the free lists. All of the physical pages must be at or above
2119 * the given physical address "low" and below the given physical address
2120 * "high". The given value "alignment" determines the alignment of the
2121 * first physical page in the set. If the given value "boundary" is
2122 * non-zero, then the set of physical pages cannot cross any physical
2123 * address boundary that is a multiple of that value. Both "alignment"
2124 * and "boundary" must be a power of two.
2126 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
2127 * then the memory attribute setting for the physical pages is configured
2128 * to the object's memory attribute setting. Otherwise, the memory
2129 * attribute setting for the physical pages is configured to "memattr",
2130 * overriding the object's memory attribute setting. However, if the
2131 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
2132 * memory attribute setting for the physical pages cannot be configured
2133 * to VM_MEMATTR_DEFAULT.
2135 * The specified object may not contain fictitious pages.
2137 * The caller must always specify an allocation class.
2139 * allocation classes:
2140 * VM_ALLOC_NORMAL normal process request
2141 * VM_ALLOC_SYSTEM system *really* needs a page
2142 * VM_ALLOC_INTERRUPT interrupt time request
2144 * optional allocation flags:
2145 * VM_ALLOC_NOBUSY do not exclusive busy the page
2146 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
2147 * VM_ALLOC_NOOBJ page is not associated with an object and
2148 * should not be exclusive busy
2149 * VM_ALLOC_SBUSY shared busy the allocated page
2150 * VM_ALLOC_WIRED wire the allocated page
2151 * VM_ALLOC_ZERO prefer a zeroed page
2154 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
2155 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2156 vm_paddr_t boundary, vm_memattr_t memattr)
2158 struct vm_domainset_iter di;
2162 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2164 m = vm_page_alloc_contig_domain(object, pindex, domain, req,
2165 npages, low, high, alignment, boundary, memattr);
2168 } while (vm_domainset_iter_page(&di, object, &domain) == 0);
2174 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2175 int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2176 vm_paddr_t boundary, vm_memattr_t memattr)
2178 struct vm_domain *vmd;
2179 vm_page_t m, m_ret, mpred;
2180 u_int busy_lock, flags, oflags;
2182 mpred = NULL; /* XXX: pacify gcc */
2183 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
2184 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
2185 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2186 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2187 ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
2189 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2190 ("Can't sleep and retry object insertion."));
2191 if (object != NULL) {
2192 VM_OBJECT_ASSERT_WLOCKED(object);
2193 KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2194 ("vm_page_alloc_contig: object %p has fictitious pages",
2197 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2199 if (object != NULL) {
2200 mpred = vm_radix_lookup_le(&object->rtree, pindex);
2201 KASSERT(mpred == NULL || mpred->pindex != pindex,
2202 ("vm_page_alloc_contig: pindex already allocated"));
2206 * Can we allocate the pages without the number of free pages falling
2207 * below the lower bound for the allocation class?
2211 #if VM_NRESERVLEVEL > 0
2213 * Can we allocate the pages from a reservation?
2215 if (vm_object_reserv(object) &&
2216 (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
2217 mpred, npages, low, high, alignment, boundary)) != NULL) {
2218 domain = vm_phys_domain(m_ret);
2219 vmd = VM_DOMAIN(domain);
2223 vmd = VM_DOMAIN(domain);
2224 if (vm_domain_allocate(vmd, req, npages)) {
2226 * allocate them from the free page queues.
2228 vm_domain_free_lock(vmd);
2229 m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2230 alignment, boundary);
2231 vm_domain_free_unlock(vmd);
2232 if (m_ret == NULL) {
2233 vm_domain_freecnt_inc(vmd, npages);
2234 #if VM_NRESERVLEVEL > 0
2235 if (vm_reserv_reclaim_contig(domain, npages, low,
2236 high, alignment, boundary))
2241 if (m_ret == NULL) {
2242 if (vm_domain_alloc_fail(vmd, object, req))
2246 #if VM_NRESERVLEVEL > 0
2249 for (m = m_ret; m < &m_ret[npages]; m++) {
2251 vm_page_alloc_check(m);
2255 * Initialize the pages. Only the PG_ZERO flag is inherited.
2258 if ((req & VM_ALLOC_ZERO) != 0)
2260 if ((req & VM_ALLOC_NODUMP) != 0)
2262 oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2264 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2265 busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2266 else if ((req & VM_ALLOC_SBUSY) != 0)
2267 busy_lock = VPB_SHARERS_WORD(1);
2269 busy_lock = VPB_UNBUSIED;
2270 if ((req & VM_ALLOC_WIRED) != 0)
2271 vm_wire_add(npages);
2272 if (object != NULL) {
2273 if (object->memattr != VM_MEMATTR_DEFAULT &&
2274 memattr == VM_MEMATTR_DEFAULT)
2275 memattr = object->memattr;
2277 for (m = m_ret; m < &m_ret[npages]; m++) {
2279 m->flags = (m->flags | PG_NODUMP) & flags;
2280 m->busy_lock = busy_lock;
2281 if ((req & VM_ALLOC_WIRED) != 0)
2285 if (object != NULL) {
2286 if (vm_page_insert_after(m, object, pindex, mpred)) {
2287 if ((req & VM_ALLOC_WIRED) != 0)
2288 vm_wire_sub(npages);
2289 KASSERT(m->object == NULL,
2290 ("page %p has object", m));
2292 for (m = m_ret; m < &m_ret[npages]; m++) {
2294 (req & VM_ALLOC_WIRED) != 0)
2296 m->oflags = VPO_UNMANAGED;
2297 m->busy_lock = VPB_UNBUSIED;
2298 /* Don't change PG_ZERO. */
2299 vm_page_free_toq(m);
2301 if (req & VM_ALLOC_WAITFAIL) {
2302 VM_OBJECT_WUNLOCK(object);
2304 VM_OBJECT_WLOCK(object);
2311 if (memattr != VM_MEMATTR_DEFAULT)
2312 pmap_page_set_memattr(m, memattr);
2319 * Check a page that has been freshly dequeued from a freelist.
2322 vm_page_alloc_check(vm_page_t m)
2325 KASSERT(m->object == NULL, ("page %p has object", m));
2326 KASSERT(m->a.queue == PQ_NONE &&
2327 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
2328 ("page %p has unexpected queue %d, flags %#x",
2329 m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK)));
2330 KASSERT(m->ref_count == 0, ("page %p has references", m));
2331 KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m));
2332 KASSERT(m->dirty == 0, ("page %p is dirty", m));
2333 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2334 ("page %p has unexpected memattr %d",
2335 m, pmap_page_get_memattr(m)));
2336 KASSERT(m->valid == 0, ("free page %p is valid", m));
2340 * vm_page_alloc_freelist:
2342 * Allocate a physical page from the specified free page list.
2344 * The caller must always specify an allocation class.
2346 * allocation classes:
2347 * VM_ALLOC_NORMAL normal process request
2348 * VM_ALLOC_SYSTEM system *really* needs a page
2349 * VM_ALLOC_INTERRUPT interrupt time request
2351 * optional allocation flags:
2352 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
2353 * intends to allocate
2354 * VM_ALLOC_WIRED wire the allocated page
2355 * VM_ALLOC_ZERO prefer a zeroed page
2358 vm_page_alloc_freelist(int freelist, int req)
2360 struct vm_domainset_iter di;
2364 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2366 m = vm_page_alloc_freelist_domain(domain, freelist, req);
2369 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2375 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2377 struct vm_domain *vmd;
2382 vmd = VM_DOMAIN(domain);
2384 if (vm_domain_allocate(vmd, req, 1)) {
2385 vm_domain_free_lock(vmd);
2386 m = vm_phys_alloc_freelist_pages(domain, freelist,
2387 VM_FREEPOOL_DIRECT, 0);
2388 vm_domain_free_unlock(vmd);
2390 vm_domain_freecnt_inc(vmd, 1);
2393 if (vm_domain_alloc_fail(vmd, NULL, req))
2398 vm_page_alloc_check(m);
2401 * Initialize the page. Only the PG_ZERO flag is inherited.
2405 if ((req & VM_ALLOC_ZERO) != 0)
2408 if ((req & VM_ALLOC_WIRED) != 0) {
2412 /* Unmanaged pages don't use "act_count". */
2413 m->oflags = VPO_UNMANAGED;
2418 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2420 struct vm_domain *vmd;
2421 struct vm_pgcache *pgcache;
2425 vmd = VM_DOMAIN(pgcache->domain);
2428 * The page daemon should avoid creating extra memory pressure since its
2429 * main purpose is to replenish the store of free pages.
2431 if (vmd->vmd_severeset || curproc == pageproc ||
2432 !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2434 domain = vmd->vmd_domain;
2435 vm_domain_free_lock(vmd);
2436 i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2437 (vm_page_t *)store);
2438 vm_domain_free_unlock(vmd);
2440 vm_domain_freecnt_inc(vmd, cnt - i);
2446 vm_page_zone_release(void *arg, void **store, int cnt)
2448 struct vm_domain *vmd;
2449 struct vm_pgcache *pgcache;
2454 vmd = VM_DOMAIN(pgcache->domain);
2455 vm_domain_free_lock(vmd);
2456 for (i = 0; i < cnt; i++) {
2457 m = (vm_page_t)store[i];
2458 vm_phys_free_pages(m, 0);
2460 vm_domain_free_unlock(vmd);
2461 vm_domain_freecnt_inc(vmd, cnt);
2464 #define VPSC_ANY 0 /* No restrictions. */
2465 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */
2466 #define VPSC_NOSUPER 2 /* Skip superpages. */
2469 * vm_page_scan_contig:
2471 * Scan vm_page_array[] between the specified entries "m_start" and
2472 * "m_end" for a run of contiguous physical pages that satisfy the
2473 * specified conditions, and return the lowest page in the run. The
2474 * specified "alignment" determines the alignment of the lowest physical
2475 * page in the run. If the specified "boundary" is non-zero, then the
2476 * run of physical pages cannot span a physical address that is a
2477 * multiple of "boundary".
2479 * "m_end" is never dereferenced, so it need not point to a vm_page
2480 * structure within vm_page_array[].
2482 * "npages" must be greater than zero. "m_start" and "m_end" must not
2483 * span a hole (or discontiguity) in the physical address space. Both
2484 * "alignment" and "boundary" must be a power of two.
2487 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2488 u_long alignment, vm_paddr_t boundary, int options)
2493 #if VM_NRESERVLEVEL > 0
2496 int m_inc, order, run_ext, run_len;
2498 KASSERT(npages > 0, ("npages is 0"));
2499 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2500 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2503 for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2504 KASSERT((m->flags & PG_MARKER) == 0,
2505 ("page %p is PG_MARKER", m));
2506 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1,
2507 ("fictitious page %p has invalid ref count", m));
2510 * If the current page would be the start of a run, check its
2511 * physical address against the end, alignment, and boundary
2512 * conditions. If it doesn't satisfy these conditions, either
2513 * terminate the scan or advance to the next page that
2514 * satisfies the failed condition.
2517 KASSERT(m_run == NULL, ("m_run != NULL"));
2518 if (m + npages > m_end)
2520 pa = VM_PAGE_TO_PHYS(m);
2521 if ((pa & (alignment - 1)) != 0) {
2522 m_inc = atop(roundup2(pa, alignment) - pa);
2525 if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2527 m_inc = atop(roundup2(pa, boundary) - pa);
2531 KASSERT(m_run != NULL, ("m_run == NULL"));
2535 if (vm_page_wired(m))
2537 #if VM_NRESERVLEVEL > 0
2538 else if ((level = vm_reserv_level(m)) >= 0 &&
2539 (options & VPSC_NORESERV) != 0) {
2541 /* Advance to the end of the reservation. */
2542 pa = VM_PAGE_TO_PHYS(m);
2543 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2547 else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2549 * The page is considered eligible for relocation if
2550 * and only if it could be laundered or reclaimed by
2553 VM_OBJECT_RLOCK(object);
2554 if (object != m->object) {
2555 VM_OBJECT_RUNLOCK(object);
2558 /* Don't care: PG_NODUMP, PG_ZERO. */
2559 if (object->type != OBJT_DEFAULT &&
2560 object->type != OBJT_SWAP &&
2561 object->type != OBJT_VNODE) {
2563 #if VM_NRESERVLEVEL > 0
2564 } else if ((options & VPSC_NOSUPER) != 0 &&
2565 (level = vm_reserv_level_iffullpop(m)) >= 0) {
2567 /* Advance to the end of the superpage. */
2568 pa = VM_PAGE_TO_PHYS(m);
2569 m_inc = atop(roundup2(pa + 1,
2570 vm_reserv_size(level)) - pa);
2572 } else if (object->memattr == VM_MEMATTR_DEFAULT &&
2573 vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) {
2575 * The page is allocated but eligible for
2576 * relocation. Extend the current run by one
2579 KASSERT(pmap_page_get_memattr(m) ==
2581 ("page %p has an unexpected memattr", m));
2582 KASSERT((m->oflags & (VPO_SWAPINPROG |
2583 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2584 ("page %p has unexpected oflags", m));
2585 /* Don't care: PGA_NOSYNC. */
2589 VM_OBJECT_RUNLOCK(object);
2590 #if VM_NRESERVLEVEL > 0
2591 } else if (level >= 0) {
2593 * The page is reserved but not yet allocated. In
2594 * other words, it is still free. Extend the current
2599 } else if ((order = m->order) < VM_NFREEORDER) {
2601 * The page is enqueued in the physical memory
2602 * allocator's free page queues. Moreover, it is the
2603 * first page in a power-of-two-sized run of
2604 * contiguous free pages. Add these pages to the end
2605 * of the current run, and jump ahead.
2607 run_ext = 1 << order;
2611 * Skip the page for one of the following reasons: (1)
2612 * It is enqueued in the physical memory allocator's
2613 * free page queues. However, it is not the first
2614 * page in a run of contiguous free pages. (This case
2615 * rarely occurs because the scan is performed in
2616 * ascending order.) (2) It is not reserved, and it is
2617 * transitioning from free to allocated. (Conversely,
2618 * the transition from allocated to free for managed
2619 * pages is blocked by the page lock.) (3) It is
2620 * allocated but not contained by an object and not
2621 * wired, e.g., allocated by Xen's balloon driver.
2627 * Extend or reset the current run of pages.
2640 if (run_len >= npages)
2646 * vm_page_reclaim_run:
2648 * Try to relocate each of the allocated virtual pages within the
2649 * specified run of physical pages to a new physical address. Free the
2650 * physical pages underlying the relocated virtual pages. A virtual page
2651 * is relocatable if and only if it could be laundered or reclaimed by
2652 * the page daemon. Whenever possible, a virtual page is relocated to a
2653 * physical address above "high".
2655 * Returns 0 if every physical page within the run was already free or
2656 * just freed by a successful relocation. Otherwise, returns a non-zero
2657 * value indicating why the last attempt to relocate a virtual page was
2660 * "req_class" must be an allocation class.
2663 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2666 struct vm_domain *vmd;
2667 struct spglist free;
2670 vm_page_t m, m_end, m_new;
2671 int error, order, req;
2673 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2674 ("req_class is not an allocation class"));
2678 m_end = m_run + npages;
2679 for (; error == 0 && m < m_end; m++) {
2680 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2681 ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2684 * Racily check for wirings. Races are handled once the object
2685 * lock is held and the page is unmapped.
2687 if (vm_page_wired(m))
2689 else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2691 * The page is relocated if and only if it could be
2692 * laundered or reclaimed by the page daemon.
2694 VM_OBJECT_WLOCK(object);
2695 /* Don't care: PG_NODUMP, PG_ZERO. */
2696 if (m->object != object ||
2697 (object->type != OBJT_DEFAULT &&
2698 object->type != OBJT_SWAP &&
2699 object->type != OBJT_VNODE))
2701 else if (object->memattr != VM_MEMATTR_DEFAULT)
2703 else if (vm_page_queue(m) != PQ_NONE &&
2704 vm_page_tryxbusy(m) != 0) {
2705 if (vm_page_wired(m)) {
2710 KASSERT(pmap_page_get_memattr(m) ==
2712 ("page %p has an unexpected memattr", m));
2713 KASSERT(m->oflags == 0,
2714 ("page %p has unexpected oflags", m));
2715 /* Don't care: PGA_NOSYNC. */
2716 if (!vm_page_none_valid(m)) {
2718 * First, try to allocate a new page
2719 * that is above "high". Failing
2720 * that, try to allocate a new page
2721 * that is below "m_run". Allocate
2722 * the new page between the end of
2723 * "m_run" and "high" only as a last
2726 req = req_class | VM_ALLOC_NOOBJ;
2727 if ((m->flags & PG_NODUMP) != 0)
2728 req |= VM_ALLOC_NODUMP;
2729 if (trunc_page(high) !=
2730 ~(vm_paddr_t)PAGE_MASK) {
2731 m_new = vm_page_alloc_contig(
2736 VM_MEMATTR_DEFAULT);
2739 if (m_new == NULL) {
2740 pa = VM_PAGE_TO_PHYS(m_run);
2741 m_new = vm_page_alloc_contig(
2743 0, pa - 1, PAGE_SIZE, 0,
2744 VM_MEMATTR_DEFAULT);
2746 if (m_new == NULL) {
2748 m_new = vm_page_alloc_contig(
2750 pa, high, PAGE_SIZE, 0,
2751 VM_MEMATTR_DEFAULT);
2753 if (m_new == NULL) {
2760 * Unmap the page and check for new
2761 * wirings that may have been acquired
2762 * through a pmap lookup.
2764 if (object->ref_count != 0 &&
2765 !vm_page_try_remove_all(m)) {
2767 vm_page_free(m_new);
2773 * Replace "m" with the new page. For
2774 * vm_page_replace(), "m" must be busy
2775 * and dequeued. Finally, change "m"
2776 * as if vm_page_free() was called.
2778 m_new->a.flags = m->a.flags &
2779 ~PGA_QUEUE_STATE_MASK;
2780 KASSERT(m_new->oflags == VPO_UNMANAGED,
2781 ("page %p is managed", m_new));
2783 pmap_copy_page(m, m_new);
2784 m_new->valid = m->valid;
2785 m_new->dirty = m->dirty;
2786 m->flags &= ~PG_ZERO;
2788 if (vm_page_replace_hold(m_new, object,
2790 vm_page_free_prep(m))
2791 SLIST_INSERT_HEAD(&free, m,
2795 * The new page must be deactivated
2796 * before the object is unlocked.
2798 vm_page_deactivate(m_new);
2800 m->flags &= ~PG_ZERO;
2802 if (vm_page_free_prep(m))
2803 SLIST_INSERT_HEAD(&free, m,
2805 KASSERT(m->dirty == 0,
2806 ("page %p is dirty", m));
2811 VM_OBJECT_WUNLOCK(object);
2813 MPASS(vm_phys_domain(m) == domain);
2814 vmd = VM_DOMAIN(domain);
2815 vm_domain_free_lock(vmd);
2817 if (order < VM_NFREEORDER) {
2819 * The page is enqueued in the physical memory
2820 * allocator's free page queues. Moreover, it
2821 * is the first page in a power-of-two-sized
2822 * run of contiguous free pages. Jump ahead
2823 * to the last page within that run, and
2824 * continue from there.
2826 m += (1 << order) - 1;
2828 #if VM_NRESERVLEVEL > 0
2829 else if (vm_reserv_is_page_free(m))
2832 vm_domain_free_unlock(vmd);
2833 if (order == VM_NFREEORDER)
2837 if ((m = SLIST_FIRST(&free)) != NULL) {
2840 vmd = VM_DOMAIN(domain);
2842 vm_domain_free_lock(vmd);
2844 MPASS(vm_phys_domain(m) == domain);
2845 SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2846 vm_phys_free_pages(m, 0);
2848 } while ((m = SLIST_FIRST(&free)) != NULL);
2849 vm_domain_free_unlock(vmd);
2850 vm_domain_freecnt_inc(vmd, cnt);
2857 CTASSERT(powerof2(NRUNS));
2859 #define RUN_INDEX(count) ((count) & (NRUNS - 1))
2861 #define MIN_RECLAIM 8
2864 * vm_page_reclaim_contig:
2866 * Reclaim allocated, contiguous physical memory satisfying the specified
2867 * conditions by relocating the virtual pages using that physical memory.
2868 * Returns true if reclamation is successful and false otherwise. Since
2869 * relocation requires the allocation of physical pages, reclamation may
2870 * fail due to a shortage of free pages. When reclamation fails, callers
2871 * are expected to perform vm_wait() before retrying a failed allocation
2872 * operation, e.g., vm_page_alloc_contig().
2874 * The caller must always specify an allocation class through "req".
2876 * allocation classes:
2877 * VM_ALLOC_NORMAL normal process request
2878 * VM_ALLOC_SYSTEM system *really* needs a page
2879 * VM_ALLOC_INTERRUPT interrupt time request
2881 * The optional allocation flags are ignored.
2883 * "npages" must be greater than zero. Both "alignment" and "boundary"
2884 * must be a power of two.
2887 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
2888 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2890 struct vm_domain *vmd;
2891 vm_paddr_t curr_low;
2892 vm_page_t m_run, m_runs[NRUNS];
2893 u_long count, reclaimed;
2894 int error, i, options, req_class;
2896 KASSERT(npages > 0, ("npages is 0"));
2897 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2898 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2899 req_class = req & VM_ALLOC_CLASS_MASK;
2902 * The page daemon is allowed to dig deeper into the free page list.
2904 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2905 req_class = VM_ALLOC_SYSTEM;
2908 * Return if the number of free pages cannot satisfy the requested
2911 vmd = VM_DOMAIN(domain);
2912 count = vmd->vmd_free_count;
2913 if (count < npages + vmd->vmd_free_reserved || (count < npages +
2914 vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2915 (count < npages && req_class == VM_ALLOC_INTERRUPT))
2919 * Scan up to three times, relaxing the restrictions ("options") on
2920 * the reclamation of reservations and superpages each time.
2922 for (options = VPSC_NORESERV;;) {
2924 * Find the highest runs that satisfy the given constraints
2925 * and restrictions, and record them in "m_runs".
2930 m_run = vm_phys_scan_contig(domain, npages, curr_low,
2931 high, alignment, boundary, options);
2934 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2935 m_runs[RUN_INDEX(count)] = m_run;
2940 * Reclaim the highest runs in LIFO (descending) order until
2941 * the number of reclaimed pages, "reclaimed", is at least
2942 * MIN_RECLAIM. Reset "reclaimed" each time because each
2943 * reclamation is idempotent, and runs will (likely) recur
2944 * from one scan to the next as restrictions are relaxed.
2947 for (i = 0; count > 0 && i < NRUNS; i++) {
2949 m_run = m_runs[RUN_INDEX(count)];
2950 error = vm_page_reclaim_run(req_class, domain, npages,
2953 reclaimed += npages;
2954 if (reclaimed >= MIN_RECLAIM)
2960 * Either relax the restrictions on the next scan or return if
2961 * the last scan had no restrictions.
2963 if (options == VPSC_NORESERV)
2964 options = VPSC_NOSUPER;
2965 else if (options == VPSC_NOSUPER)
2967 else if (options == VPSC_ANY)
2968 return (reclaimed != 0);
2973 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
2974 u_long alignment, vm_paddr_t boundary)
2976 struct vm_domainset_iter di;
2980 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2982 ret = vm_page_reclaim_contig_domain(domain, req, npages, low,
2983 high, alignment, boundary);
2986 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2992 * Set the domain in the appropriate page level domainset.
2995 vm_domain_set(struct vm_domain *vmd)
2998 mtx_lock(&vm_domainset_lock);
2999 if (!vmd->vmd_minset && vm_paging_min(vmd)) {
3000 vmd->vmd_minset = 1;
3001 DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
3003 if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
3004 vmd->vmd_severeset = 1;
3005 DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
3007 mtx_unlock(&vm_domainset_lock);
3011 * Clear the domain from the appropriate page level domainset.
3014 vm_domain_clear(struct vm_domain *vmd)
3017 mtx_lock(&vm_domainset_lock);
3018 if (vmd->vmd_minset && !vm_paging_min(vmd)) {
3019 vmd->vmd_minset = 0;
3020 DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
3021 if (vm_min_waiters != 0) {
3023 wakeup(&vm_min_domains);
3026 if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
3027 vmd->vmd_severeset = 0;
3028 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
3029 if (vm_severe_waiters != 0) {
3030 vm_severe_waiters = 0;
3031 wakeup(&vm_severe_domains);
3036 * If pageout daemon needs pages, then tell it that there are
3039 if (vmd->vmd_pageout_pages_needed &&
3040 vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
3041 wakeup(&vmd->vmd_pageout_pages_needed);
3042 vmd->vmd_pageout_pages_needed = 0;
3045 /* See comments in vm_wait_doms(). */
3046 if (vm_pageproc_waiters) {
3047 vm_pageproc_waiters = 0;
3048 wakeup(&vm_pageproc_waiters);
3050 mtx_unlock(&vm_domainset_lock);
3054 * Wait for free pages to exceed the min threshold globally.
3060 mtx_lock(&vm_domainset_lock);
3061 while (vm_page_count_min()) {
3063 msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
3065 mtx_unlock(&vm_domainset_lock);
3069 * Wait for free pages to exceed the severe threshold globally.
3072 vm_wait_severe(void)
3075 mtx_lock(&vm_domainset_lock);
3076 while (vm_page_count_severe()) {
3077 vm_severe_waiters++;
3078 msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
3081 mtx_unlock(&vm_domainset_lock);
3088 return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
3092 vm_wait_doms(const domainset_t *wdoms)
3096 * We use racey wakeup synchronization to avoid expensive global
3097 * locking for the pageproc when sleeping with a non-specific vm_wait.
3098 * To handle this, we only sleep for one tick in this instance. It
3099 * is expected that most allocations for the pageproc will come from
3100 * kmem or vm_page_grab* which will use the more specific and
3101 * race-free vm_wait_domain().
3103 if (curproc == pageproc) {
3104 mtx_lock(&vm_domainset_lock);
3105 vm_pageproc_waiters++;
3106 msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP,
3110 * XXX Ideally we would wait only until the allocation could
3111 * be satisfied. This condition can cause new allocators to
3112 * consume all freed pages while old allocators wait.
3114 mtx_lock(&vm_domainset_lock);
3115 if (vm_page_count_min_set(wdoms)) {
3117 msleep(&vm_min_domains, &vm_domainset_lock,
3118 PVM | PDROP, "vmwait", 0);
3120 mtx_unlock(&vm_domainset_lock);
3127 * Sleep until free pages are available for allocation.
3128 * - Called in various places after failed memory allocations.
3131 vm_wait_domain(int domain)
3133 struct vm_domain *vmd;
3136 vmd = VM_DOMAIN(domain);
3137 vm_domain_free_assert_unlocked(vmd);
3139 if (curproc == pageproc) {
3140 mtx_lock(&vm_domainset_lock);
3141 if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3142 vmd->vmd_pageout_pages_needed = 1;
3143 msleep(&vmd->vmd_pageout_pages_needed,
3144 &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3146 mtx_unlock(&vm_domainset_lock);
3148 if (pageproc == NULL)
3149 panic("vm_wait in early boot");
3150 DOMAINSET_ZERO(&wdom);
3151 DOMAINSET_SET(vmd->vmd_domain, &wdom);
3152 vm_wait_doms(&wdom);
3159 * Sleep until free pages are available for allocation in the
3160 * affinity domains of the obj. If obj is NULL, the domain set
3161 * for the calling thread is used.
3162 * Called in various places after failed memory allocations.
3165 vm_wait(vm_object_t obj)
3167 struct domainset *d;
3172 * Carefully fetch pointers only once: the struct domainset
3173 * itself is ummutable but the pointer might change.
3176 d = obj->domain.dr_policy;
3178 d = curthread->td_domain.dr_policy;
3180 vm_wait_doms(&d->ds_mask);
3184 * vm_domain_alloc_fail:
3186 * Called when a page allocation function fails. Informs the
3187 * pagedaemon and performs the requested wait. Requires the
3188 * domain_free and object lock on entry. Returns with the
3189 * object lock held and free lock released. Returns an error when
3190 * retry is necessary.
3194 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3197 vm_domain_free_assert_unlocked(vmd);
3199 atomic_add_int(&vmd->vmd_pageout_deficit,
3200 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3201 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3203 VM_OBJECT_WUNLOCK(object);
3204 vm_wait_domain(vmd->vmd_domain);
3206 VM_OBJECT_WLOCK(object);
3207 if (req & VM_ALLOC_WAITOK)
3217 * Sleep until free pages are available for allocation.
3218 * - Called only in vm_fault so that processes page faulting
3219 * can be easily tracked.
3220 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3221 * processes will be able to grab memory first. Do not change
3222 * this balance without careful testing first.
3225 vm_waitpfault(struct domainset *dset, int timo)
3229 * XXX Ideally we would wait only until the allocation could
3230 * be satisfied. This condition can cause new allocators to
3231 * consume all freed pages while old allocators wait.
3233 mtx_lock(&vm_domainset_lock);
3234 if (vm_page_count_min_set(&dset->ds_mask)) {
3236 msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3239 mtx_unlock(&vm_domainset_lock);
3242 static struct vm_pagequeue *
3243 _vm_page_pagequeue(vm_page_t m, uint8_t queue)
3246 return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3250 static struct vm_pagequeue *
3251 vm_page_pagequeue(vm_page_t m)
3254 return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue));
3258 static __always_inline bool
3259 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3261 vm_page_astate_t tmp;
3265 if (__predict_true(vm_page_astate_fcmpset(m, old, new)))
3267 counter_u64_add(pqstate_commit_retries, 1);
3268 } while (old->_bits == tmp._bits);
3274 * Do the work of committing a queue state update that moves the page out of
3275 * its current queue.
3278 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m,
3279 vm_page_astate_t *old, vm_page_astate_t new)
3283 vm_pagequeue_assert_locked(pq);
3284 KASSERT(vm_page_pagequeue(m) == pq,
3285 ("%s: queue %p does not match page %p", __func__, pq, m));
3286 KASSERT(old->queue != PQ_NONE && new.queue != old->queue,
3287 ("%s: invalid queue indices %d %d",
3288 __func__, old->queue, new.queue));
3291 * Once the queue index of the page changes there is nothing
3292 * synchronizing with further updates to the page's physical
3293 * queue state. Therefore we must speculatively remove the page
3294 * from the queue now and be prepared to roll back if the queue
3295 * state update fails. If the page is not physically enqueued then
3296 * we just update its queue index.
3298 if ((old->flags & PGA_ENQUEUED) != 0) {
3299 new.flags &= ~PGA_ENQUEUED;
3300 next = TAILQ_NEXT(m, plinks.q);
3301 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3302 vm_pagequeue_cnt_dec(pq);
3303 if (!vm_page_pqstate_fcmpset(m, old, new)) {
3305 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3307 TAILQ_INSERT_BEFORE(next, m, plinks.q);
3308 vm_pagequeue_cnt_inc(pq);
3314 return (vm_page_pqstate_fcmpset(m, old, new));
3319 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old,
3320 vm_page_astate_t new)
3322 struct vm_pagequeue *pq;
3323 vm_page_astate_t as;
3326 pq = _vm_page_pagequeue(m, old->queue);
3329 * The queue field and PGA_ENQUEUED flag are stable only so long as the
3330 * corresponding page queue lock is held.
3332 vm_pagequeue_lock(pq);
3333 as = vm_page_astate_load(m);
3334 if (__predict_false(as._bits != old->_bits)) {
3338 ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new);
3340 vm_pagequeue_unlock(pq);
3345 * Commit a queue state update that enqueues or requeues a page.
3348 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m,
3349 vm_page_astate_t *old, vm_page_astate_t new)
3351 struct vm_domain *vmd;
3353 vm_pagequeue_assert_locked(pq);
3354 KASSERT(old->queue != PQ_NONE && new.queue == old->queue,
3355 ("%s: invalid queue indices %d %d",
3356 __func__, old->queue, new.queue));
3358 new.flags |= PGA_ENQUEUED;
3359 if (!vm_page_pqstate_fcmpset(m, old, new))
3362 if ((old->flags & PGA_ENQUEUED) != 0)
3363 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3365 vm_pagequeue_cnt_inc(pq);
3368 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE. In particular, if
3369 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be
3370 * applied, even if it was set first.
3372 if ((old->flags & PGA_REQUEUE_HEAD) != 0) {
3373 vmd = vm_pagequeue_domain(m);
3374 KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE],
3375 ("%s: invalid page queue for page %p", __func__, m));
3376 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3378 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3384 * Commit a queue state update that encodes a request for a deferred queue
3388 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old,
3389 vm_page_astate_t new)
3392 KASSERT(old->queue == new.queue || new.queue != PQ_NONE,
3393 ("%s: invalid state, queue %d flags %x",
3394 __func__, new.queue, new.flags));
3396 if (old->_bits != new._bits &&
3397 !vm_page_pqstate_fcmpset(m, old, new))
3399 vm_page_pqbatch_submit(m, new.queue);
3404 * A generic queue state update function. This handles more cases than the
3405 * specialized functions above.
3408 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3411 if (old->_bits == new._bits)
3414 if (old->queue != PQ_NONE && new.queue != old->queue) {
3415 if (!vm_page_pqstate_commit_dequeue(m, old, new))
3417 if (new.queue != PQ_NONE)
3418 vm_page_pqbatch_submit(m, new.queue);
3420 if (!vm_page_pqstate_fcmpset(m, old, new))
3422 if (new.queue != PQ_NONE &&
3423 ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0)
3424 vm_page_pqbatch_submit(m, new.queue);
3430 * Apply deferred queue state updates to a page.
3433 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue)
3435 vm_page_astate_t new, old;
3437 CRITICAL_ASSERT(curthread);
3438 vm_pagequeue_assert_locked(pq);
3439 KASSERT(queue < PQ_COUNT,
3440 ("%s: invalid queue index %d", __func__, queue));
3441 KASSERT(pq == _vm_page_pagequeue(m, queue),
3442 ("%s: page %p does not belong to queue %p", __func__, m, pq));
3444 for (old = vm_page_astate_load(m);;) {
3445 if (__predict_false(old.queue != queue ||
3446 (old.flags & PGA_QUEUE_OP_MASK) == 0)) {
3447 counter_u64_add(queue_nops, 1);
3450 KASSERT(old.queue != PQ_NONE || (old.flags & PGA_QUEUE_STATE_MASK) == 0,
3451 ("%s: page %p has unexpected queue state", __func__, m));
3454 if ((old.flags & PGA_DEQUEUE) != 0) {
3455 new.flags &= ~PGA_QUEUE_OP_MASK;
3456 new.queue = PQ_NONE;
3457 if (__predict_true(_vm_page_pqstate_commit_dequeue(pq,
3459 counter_u64_add(queue_ops, 1);
3463 new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD);
3464 if (__predict_true(_vm_page_pqstate_commit_requeue(pq,
3466 counter_u64_add(queue_ops, 1);
3474 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3479 for (i = 0; i < bq->bq_cnt; i++)
3480 vm_pqbatch_process_page(pq, bq->bq_pa[i], queue);
3481 vm_batchqueue_init(bq);
3485 * vm_page_pqbatch_submit: [ internal use only ]
3487 * Enqueue a page in the specified page queue's batched work queue.
3488 * The caller must have encoded the requested operation in the page
3489 * structure's a.flags field.
3492 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
3494 struct vm_batchqueue *bq;
3495 struct vm_pagequeue *pq;
3498 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3499 ("page %p is unmanaged", m));
3500 KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3502 domain = vm_phys_domain(m);
3503 pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue];
3506 bq = DPCPU_PTR(pqbatch[domain][queue]);
3507 if (vm_batchqueue_insert(bq, m)) {
3512 vm_pagequeue_lock(pq);
3514 bq = DPCPU_PTR(pqbatch[domain][queue]);
3515 vm_pqbatch_process(pq, bq, queue);
3516 vm_pqbatch_process_page(pq, m, queue);
3517 vm_pagequeue_unlock(pq);
3522 * vm_page_pqbatch_drain: [ internal use only ]
3524 * Force all per-CPU page queue batch queues to be drained. This is
3525 * intended for use in severe memory shortages, to ensure that pages
3526 * do not remain stuck in the batch queues.
3529 vm_page_pqbatch_drain(void)
3532 struct vm_domain *vmd;
3533 struct vm_pagequeue *pq;
3534 int cpu, domain, queue;
3539 sched_bind(td, cpu);
3542 for (domain = 0; domain < vm_ndomains; domain++) {
3543 vmd = VM_DOMAIN(domain);
3544 for (queue = 0; queue < PQ_COUNT; queue++) {
3545 pq = &vmd->vmd_pagequeues[queue];
3546 vm_pagequeue_lock(pq);
3548 vm_pqbatch_process(pq,
3549 DPCPU_PTR(pqbatch[domain][queue]), queue);
3551 vm_pagequeue_unlock(pq);
3561 * vm_page_dequeue_deferred: [ internal use only ]
3563 * Request removal of the given page from its current page
3564 * queue. Physical removal from the queue may be deferred
3567 * The page must be locked.
3570 vm_page_dequeue_deferred(vm_page_t m)
3572 vm_page_astate_t new, old;
3574 old = vm_page_astate_load(m);
3576 if (old.queue == PQ_NONE) {
3577 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3578 ("%s: page %p has unexpected queue state",
3583 new.flags |= PGA_DEQUEUE;
3584 } while (!vm_page_pqstate_commit_request(m, &old, new));
3590 * Remove the page from whichever page queue it's in, if any, before
3594 vm_page_dequeue(vm_page_t m)
3596 vm_page_astate_t new, old;
3598 old = vm_page_astate_load(m);
3600 if (old.queue == PQ_NONE) {
3601 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3602 ("%s: page %p has unexpected queue state",
3607 new.flags &= ~PGA_QUEUE_OP_MASK;
3608 new.queue = PQ_NONE;
3609 } while (!vm_page_pqstate_commit_dequeue(m, &old, new));
3614 * Schedule the given page for insertion into the specified page queue.
3615 * Physical insertion of the page may be deferred indefinitely.
3618 vm_page_enqueue(vm_page_t m, uint8_t queue)
3621 KASSERT(m->a.queue == PQ_NONE &&
3622 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
3623 ("%s: page %p is already enqueued", __func__, m));
3624 KASSERT(m->ref_count > 0,
3625 ("%s: page %p does not carry any references", __func__, m));
3628 if ((m->a.flags & PGA_REQUEUE) == 0)
3629 vm_page_aflag_set(m, PGA_REQUEUE);
3630 vm_page_pqbatch_submit(m, queue);
3634 * vm_page_free_prep:
3636 * Prepares the given page to be put on the free list,
3637 * disassociating it from any VM object. The caller may return
3638 * the page to the free list only if this function returns true.
3640 * The object must be locked. The page must be locked if it is
3644 vm_page_free_prep(vm_page_t m)
3648 * Synchronize with threads that have dropped a reference to this
3651 atomic_thread_fence_acq();
3653 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3654 if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3657 p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3658 for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3659 KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3660 m, i, (uintmax_t)*p));
3663 if ((m->oflags & VPO_UNMANAGED) == 0) {
3664 KASSERT(!pmap_page_is_mapped(m),
3665 ("vm_page_free_prep: freeing mapped page %p", m));
3666 KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
3667 ("vm_page_free_prep: mapping flags set in page %p", m));
3669 KASSERT(m->a.queue == PQ_NONE,
3670 ("vm_page_free_prep: unmanaged page %p is queued", m));
3672 VM_CNT_INC(v_tfree);
3674 if (m->object != NULL) {
3675 KASSERT(((m->oflags & VPO_UNMANAGED) != 0) ==
3676 ((m->object->flags & OBJ_UNMANAGED) != 0),
3677 ("vm_page_free_prep: managed flag mismatch for page %p",
3679 vm_page_assert_xbusied(m);
3682 * The object reference can be released without an atomic
3685 KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
3686 m->ref_count == VPRC_OBJREF,
3687 ("vm_page_free_prep: page %p has unexpected ref_count %u",
3689 vm_page_object_remove(m);
3691 m->ref_count -= VPRC_OBJREF;
3693 vm_page_assert_unbusied(m);
3695 vm_page_busy_free(m);
3698 * If fictitious remove object association and
3701 if ((m->flags & PG_FICTITIOUS) != 0) {
3702 KASSERT(m->ref_count == 1,
3703 ("fictitious page %p is referenced", m));
3704 KASSERT(m->a.queue == PQ_NONE,
3705 ("fictitious page %p is queued", m));
3710 * Pages need not be dequeued before they are returned to the physical
3711 * memory allocator, but they must at least be marked for a deferred
3714 if ((m->oflags & VPO_UNMANAGED) == 0)
3715 vm_page_dequeue_deferred(m);
3720 if (m->ref_count != 0)
3721 panic("vm_page_free_prep: page %p has references", m);
3724 * Restore the default memory attribute to the page.
3726 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3727 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3729 #if VM_NRESERVLEVEL > 0
3731 * Determine whether the page belongs to a reservation. If the page was
3732 * allocated from a per-CPU cache, it cannot belong to a reservation, so
3733 * as an optimization, we avoid the check in that case.
3735 if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
3745 * Returns the given page to the free list, disassociating it
3746 * from any VM object.
3748 * The object must be locked. The page must be locked if it is
3752 vm_page_free_toq(vm_page_t m)
3754 struct vm_domain *vmd;
3757 if (!vm_page_free_prep(m))
3760 vmd = vm_pagequeue_domain(m);
3761 zone = vmd->vmd_pgcache[m->pool].zone;
3762 if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
3766 vm_domain_free_lock(vmd);
3767 vm_phys_free_pages(m, 0);
3768 vm_domain_free_unlock(vmd);
3769 vm_domain_freecnt_inc(vmd, 1);
3773 * vm_page_free_pages_toq:
3775 * Returns a list of pages to the free list, disassociating it
3776 * from any VM object. In other words, this is equivalent to
3777 * calling vm_page_free_toq() for each page of a list of VM objects.
3779 * The objects must be locked. The pages must be locked if it is
3783 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
3788 if (SLIST_EMPTY(free))
3792 while ((m = SLIST_FIRST(free)) != NULL) {
3794 SLIST_REMOVE_HEAD(free, plinks.s.ss);
3795 vm_page_free_toq(m);
3798 if (update_wire_count)
3803 * Mark this page as wired down, preventing reclamation by the page daemon
3804 * or when the containing object is destroyed.
3807 vm_page_wire(vm_page_t m)
3811 KASSERT(m->object != NULL,
3812 ("vm_page_wire: page %p does not belong to an object", m));
3813 if (!vm_page_busied(m) && !vm_object_busied(m->object))
3814 VM_OBJECT_ASSERT_LOCKED(m->object);
3815 KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
3816 VPRC_WIRE_COUNT(m->ref_count) >= 1,
3817 ("vm_page_wire: fictitious page %p has zero wirings", m));
3819 old = atomic_fetchadd_int(&m->ref_count, 1);
3820 KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
3821 ("vm_page_wire: counter overflow for page %p", m));
3822 if (VPRC_WIRE_COUNT(old) == 0) {
3823 if ((m->oflags & VPO_UNMANAGED) == 0)
3824 vm_page_aflag_set(m, PGA_DEQUEUE);
3830 * Attempt to wire a mapped page following a pmap lookup of that page.
3831 * This may fail if a thread is concurrently tearing down mappings of the page.
3832 * The transient failure is acceptable because it translates to the
3833 * failure of the caller pmap_extract_and_hold(), which should be then
3834 * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages().
3837 vm_page_wire_mapped(vm_page_t m)
3844 ("vm_page_wire_mapped: wiring unreferenced page %p", m));
3845 if ((old & VPRC_BLOCKED) != 0)
3847 } while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
3849 if (VPRC_WIRE_COUNT(old) == 0) {
3850 if ((m->oflags & VPO_UNMANAGED) == 0)
3851 vm_page_aflag_set(m, PGA_DEQUEUE);
3858 * Release a wiring reference to a managed page. If the page still belongs to
3859 * an object, update its position in the page queues to reflect the reference.
3860 * If the wiring was the last reference to the page, free the page.
3863 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse)
3867 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3868 ("%s: page %p is unmanaged", __func__, m));
3871 * Update LRU state before releasing the wiring reference.
3872 * Use a release store when updating the reference count to
3873 * synchronize with vm_page_free_prep().
3877 KASSERT(VPRC_WIRE_COUNT(old) > 0,
3878 ("vm_page_unwire: wire count underflow for page %p", m));
3880 if (old > VPRC_OBJREF + 1) {
3882 * The page has at least one other wiring reference. An
3883 * earlier iteration of this loop may have called
3884 * vm_page_release_toq() and cleared PGA_DEQUEUE, so
3885 * re-set it if necessary.
3887 if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0)
3888 vm_page_aflag_set(m, PGA_DEQUEUE);
3889 } else if (old == VPRC_OBJREF + 1) {
3891 * This is the last wiring. Clear PGA_DEQUEUE and
3892 * update the page's queue state to reflect the
3893 * reference. If the page does not belong to an object
3894 * (i.e., the VPRC_OBJREF bit is clear), we only need to
3895 * clear leftover queue state.
3897 vm_page_release_toq(m, nqueue, false);
3898 } else if (old == 1) {
3899 vm_page_aflag_clear(m, PGA_DEQUEUE);
3901 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
3903 if (VPRC_WIRE_COUNT(old) == 1) {
3911 * Release one wiring of the specified page, potentially allowing it to be
3914 * Only managed pages belonging to an object can be paged out. If the number
3915 * of wirings transitions to zero and the page is eligible for page out, then
3916 * the page is added to the specified paging queue. If the released wiring
3917 * represented the last reference to the page, the page is freed.
3919 * A managed page must be locked.
3922 vm_page_unwire(vm_page_t m, uint8_t nqueue)
3925 KASSERT(nqueue < PQ_COUNT,
3926 ("vm_page_unwire: invalid queue %u request for page %p",
3929 if ((m->oflags & VPO_UNMANAGED) != 0) {
3930 if (vm_page_unwire_noq(m) && m->ref_count == 0)
3934 vm_page_unwire_managed(m, nqueue, false);
3938 * Unwire a page without (re-)inserting it into a page queue. It is up
3939 * to the caller to enqueue, requeue, or free the page as appropriate.
3940 * In most cases involving managed pages, vm_page_unwire() should be used
3944 vm_page_unwire_noq(vm_page_t m)
3948 old = vm_page_drop(m, 1);
3949 KASSERT(VPRC_WIRE_COUNT(old) != 0,
3950 ("vm_page_unref: counter underflow for page %p", m));
3951 KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
3952 ("vm_page_unref: missing ref on fictitious page %p", m));
3954 if (VPRC_WIRE_COUNT(old) > 1)
3956 if ((m->oflags & VPO_UNMANAGED) == 0)
3957 vm_page_aflag_clear(m, PGA_DEQUEUE);
3963 * Ensure that the page ends up in the specified page queue. If the page is
3964 * active or being moved to the active queue, ensure that its act_count is
3965 * at least ACT_INIT but do not otherwise mess with it.
3967 * A managed page must be locked.
3969 static __always_inline void
3970 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag)
3972 vm_page_astate_t old, new;
3974 KASSERT(m->ref_count > 0,
3975 ("%s: page %p does not carry any references", __func__, m));
3976 KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD,
3977 ("%s: invalid flags %x", __func__, nflag));
3979 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
3982 old = vm_page_astate_load(m);
3984 if ((old.flags & PGA_DEQUEUE) != 0)
3987 new.flags &= ~PGA_QUEUE_OP_MASK;
3988 if (nqueue == PQ_ACTIVE)
3989 new.act_count = max(old.act_count, ACT_INIT);
3990 if (old.queue == nqueue) {
3991 if (nqueue != PQ_ACTIVE)
3997 } while (!vm_page_pqstate_commit(m, &old, new));
4001 * Put the specified page on the active list (if appropriate).
4004 vm_page_activate(vm_page_t m)
4007 vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE);
4011 * Move the specified page to the tail of the inactive queue, or requeue
4012 * the page if it is already in the inactive queue.
4015 vm_page_deactivate(vm_page_t m)
4018 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE);
4022 vm_page_deactivate_noreuse(vm_page_t m)
4025 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD);
4029 * Put a page in the laundry, or requeue it if it is already there.
4032 vm_page_launder(vm_page_t m)
4035 vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE);
4039 * Put a page in the PQ_UNSWAPPABLE holding queue.
4042 vm_page_unswappable(vm_page_t m)
4045 KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0,
4046 ("page %p already unswappable", m));
4049 vm_page_enqueue(m, PQ_UNSWAPPABLE);
4053 * Release a page back to the page queues in preparation for unwiring.
4056 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse)
4058 vm_page_astate_t old, new;
4062 * Use a check of the valid bits to determine whether we should
4063 * accelerate reclamation of the page. The object lock might not be
4064 * held here, in which case the check is racy. At worst we will either
4065 * accelerate reclamation of a valid page and violate LRU, or
4066 * unnecessarily defer reclamation of an invalid page.
4068 * If we were asked to not cache the page, place it near the head of the
4069 * inactive queue so that is reclaimed sooner.
4071 if (noreuse || m->valid == 0) {
4072 nqueue = PQ_INACTIVE;
4073 nflag = PGA_REQUEUE_HEAD;
4075 nflag = PGA_REQUEUE;
4078 old = vm_page_astate_load(m);
4083 * If the page is already in the active queue and we are not
4084 * trying to accelerate reclamation, simply mark it as
4085 * referenced and avoid any queue operations.
4087 new.flags &= ~PGA_QUEUE_OP_MASK;
4088 if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE)
4089 new.flags |= PGA_REFERENCED;
4094 } while (!vm_page_pqstate_commit(m, &old, new));
4098 * Unwire a page and either attempt to free it or re-add it to the page queues.
4101 vm_page_release(vm_page_t m, int flags)
4105 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4106 ("vm_page_release: page %p is unmanaged", m));
4108 if ((flags & VPR_TRYFREE) != 0) {
4110 object = atomic_load_ptr(&m->object);
4113 /* Depends on type-stability. */
4114 if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object))
4116 if (object == m->object) {
4117 vm_page_release_locked(m, flags);
4118 VM_OBJECT_WUNLOCK(object);
4121 VM_OBJECT_WUNLOCK(object);
4124 vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0);
4127 /* See vm_page_release(). */
4129 vm_page_release_locked(vm_page_t m, int flags)
4132 VM_OBJECT_ASSERT_WLOCKED(m->object);
4133 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4134 ("vm_page_release_locked: page %p is unmanaged", m));
4136 if (vm_page_unwire_noq(m)) {
4137 if ((flags & VPR_TRYFREE) != 0 &&
4138 (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
4139 m->dirty == 0 && vm_page_tryxbusy(m)) {
4142 vm_page_release_toq(m, PQ_INACTIVE, flags != 0);
4148 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
4152 KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
4153 ("vm_page_try_blocked_op: page %p has no object", m));
4154 KASSERT(vm_page_busied(m),
4155 ("vm_page_try_blocked_op: page %p is not busy", m));
4156 VM_OBJECT_ASSERT_LOCKED(m->object);
4161 ("vm_page_try_blocked_op: page %p has no references", m));
4162 if (VPRC_WIRE_COUNT(old) != 0)
4164 } while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
4169 * If the object is read-locked, new wirings may be created via an
4172 old = vm_page_drop(m, VPRC_BLOCKED);
4173 KASSERT(!VM_OBJECT_WOWNED(m->object) ||
4174 old == (VPRC_BLOCKED | VPRC_OBJREF),
4175 ("vm_page_try_blocked_op: unexpected refcount value %u for %p",
4181 * Atomically check for wirings and remove all mappings of the page.
4184 vm_page_try_remove_all(vm_page_t m)
4187 return (vm_page_try_blocked_op(m, pmap_remove_all));
4191 * Atomically check for wirings and remove all writeable mappings of the page.
4194 vm_page_try_remove_write(vm_page_t m)
4197 return (vm_page_try_blocked_op(m, pmap_remove_write));
4203 * Apply the specified advice to the given page.
4205 * The object and page must be locked.
4208 vm_page_advise(vm_page_t m, int advice)
4211 VM_OBJECT_ASSERT_WLOCKED(m->object);
4212 if (advice == MADV_FREE)
4214 * Mark the page clean. This will allow the page to be freed
4215 * without first paging it out. MADV_FREE pages are often
4216 * quickly reused by malloc(3), so we do not do anything that
4217 * would result in a page fault on a later access.
4220 else if (advice != MADV_DONTNEED) {
4221 if (advice == MADV_WILLNEED)
4222 vm_page_activate(m);
4226 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
4230 * Clear any references to the page. Otherwise, the page daemon will
4231 * immediately reactivate the page.
4233 vm_page_aflag_clear(m, PGA_REFERENCED);
4236 * Place clean pages near the head of the inactive queue rather than
4237 * the tail, thus defeating the queue's LRU operation and ensuring that
4238 * the page will be reused quickly. Dirty pages not already in the
4239 * laundry are moved there.
4242 vm_page_deactivate_noreuse(m);
4243 else if (!vm_page_in_laundry(m))
4248 vm_page_grab_pflags(int allocflags)
4252 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4253 (allocflags & VM_ALLOC_WIRED) != 0,
4254 ("vm_page_grab_pflags: the pages must be busied or wired"));
4255 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4256 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4257 ("vm_page_grab_pflags: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY "
4259 pflags = allocflags &
4260 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL |
4262 if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4263 pflags |= VM_ALLOC_WAITFAIL;
4264 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4265 pflags |= VM_ALLOC_SBUSY;
4271 * Grab a page, waiting until we are waken up due to the page
4272 * changing state. We keep on waiting, if the page continues
4273 * to be in the object. If the page doesn't exist, first allocate it
4274 * and then conditionally zero it.
4276 * This routine may sleep.
4278 * The object must be locked on entry. The lock will, however, be released
4279 * and reacquired if the routine sleeps.
4282 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
4287 VM_OBJECT_ASSERT_WLOCKED(object);
4288 pflags = vm_page_grab_pflags(allocflags);
4290 if ((m = vm_page_lookup(object, pindex)) != NULL) {
4291 if (!vm_page_acquire_flags(m, allocflags)) {
4292 if (vm_page_busy_sleep_flags(object, m, "pgrbwt",
4299 if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4301 m = vm_page_alloc(object, pindex, pflags);
4303 if ((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0)
4307 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
4311 if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4312 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4321 * Grab a page and make it valid, paging in if necessary. Pages missing from
4322 * their pager are zero filled and validated. If a VM_ALLOC_COUNT is supplied
4323 * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought
4324 * in simultaneously. Additional pages will be left on a paging queue but
4325 * will neither be wired nor busy regardless of allocflags.
4328 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags)
4331 vm_page_t ma[VM_INITIAL_PAGEIN];
4333 int after, i, pflags, rv;
4335 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4336 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4337 ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4338 KASSERT((allocflags &
4339 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4340 ("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
4341 VM_OBJECT_ASSERT_WLOCKED(object);
4342 pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY);
4343 pflags |= VM_ALLOC_WAITFAIL;
4347 if ((m = vm_page_lookup(object, pindex)) != NULL) {
4349 * If the page is fully valid it can only become invalid
4350 * with the object lock held. If it is not valid it can
4351 * become valid with the busy lock held. Therefore, we
4352 * may unnecessarily lock the exclusive busy here if we
4353 * race with I/O completion not using the object lock.
4354 * However, we will not end up with an invalid page and a
4357 if (!vm_page_all_valid(m) ||
4358 (allocflags & (VM_ALLOC_IGN_SBUSY | VM_ALLOC_SBUSY)) == 0) {
4359 sleep = !vm_page_tryxbusy(m);
4362 sleep = !vm_page_trysbusy(m);
4364 (void)vm_page_busy_sleep_flags(object, m, "pgrbwt",
4368 if ((allocflags & VM_ALLOC_NOCREAT) != 0 &&
4369 !vm_page_all_valid(m)) {
4375 return (VM_PAGER_FAIL);
4377 if ((allocflags & VM_ALLOC_WIRED) != 0)
4379 if (vm_page_all_valid(m))
4381 } else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4383 return (VM_PAGER_FAIL);
4384 } else if ((m = vm_page_alloc(object, pindex, pflags)) != NULL) {
4390 vm_page_assert_xbusied(m);
4392 if (vm_pager_has_page(object, pindex, NULL, &after)) {
4393 after = MIN(after, VM_INITIAL_PAGEIN);
4394 after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT);
4395 after = MAX(after, 1);
4397 for (i = 1; i < after; i++) {
4398 if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) {
4399 if (ma[i]->valid || !vm_page_tryxbusy(ma[i]))
4402 ma[i] = vm_page_alloc(object, m->pindex + i,
4409 vm_object_pip_add(object, after);
4410 VM_OBJECT_WUNLOCK(object);
4411 rv = vm_pager_get_pages(object, ma, after, NULL, NULL);
4412 VM_OBJECT_WLOCK(object);
4413 vm_object_pip_wakeupn(object, after);
4414 /* Pager may have replaced a page. */
4416 if (rv != VM_PAGER_OK) {
4417 if ((allocflags & VM_ALLOC_WIRED) != 0)
4418 vm_page_unwire_noq(m);
4419 for (i = 0; i < after; i++) {
4420 if (!vm_page_wired(ma[i]))
4421 vm_page_free(ma[i]);
4423 vm_page_xunbusy(ma[i]);
4428 for (i = 1; i < after; i++)
4429 vm_page_readahead_finish(ma[i]);
4430 MPASS(vm_page_all_valid(m));
4432 vm_page_zero_invalid(m, TRUE);
4435 if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4441 if ((allocflags & VM_ALLOC_SBUSY) != 0 && xbusy)
4442 vm_page_busy_downgrade(m);
4444 return (VM_PAGER_OK);
4448 * Return the specified range of pages from the given object. For each
4449 * page offset within the range, if a page already exists within the object
4450 * at that offset and it is busy, then wait for it to change state. If,
4451 * instead, the page doesn't exist, then allocate it.
4453 * The caller must always specify an allocation class.
4455 * allocation classes:
4456 * VM_ALLOC_NORMAL normal process request
4457 * VM_ALLOC_SYSTEM system *really* needs the pages
4459 * The caller must always specify that the pages are to be busied and/or
4462 * optional allocation flags:
4463 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages
4464 * VM_ALLOC_NOBUSY do not exclusive busy the page
4465 * VM_ALLOC_NOWAIT do not sleep
4466 * VM_ALLOC_SBUSY set page to sbusy state
4467 * VM_ALLOC_WIRED wire the pages
4468 * VM_ALLOC_ZERO zero and validate any invalid pages
4470 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it
4471 * may return a partial prefix of the requested range.
4474 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
4475 vm_page_t *ma, int count)
4481 VM_OBJECT_ASSERT_WLOCKED(object);
4482 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
4483 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
4485 pflags = vm_page_grab_pflags(allocflags);
4491 m = vm_radix_lookup_le(&object->rtree, pindex + i);
4492 if (m == NULL || m->pindex != pindex + i) {
4496 mpred = TAILQ_PREV(m, pglist, listq);
4497 for (; i < count; i++) {
4499 if (!vm_page_acquire_flags(m, allocflags)) {
4500 if (vm_page_busy_sleep_flags(object, m,
4501 "grbmaw", allocflags))
4506 if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4508 m = vm_page_alloc_after(object, pindex + i,
4509 pflags | VM_ALLOC_COUNT(count - i), mpred);
4511 if ((allocflags & (VM_ALLOC_NOWAIT |
4512 VM_ALLOC_WAITFAIL)) != 0)
4517 if (vm_page_none_valid(m) &&
4518 (allocflags & VM_ALLOC_ZERO) != 0) {
4519 if ((m->flags & PG_ZERO) == 0)
4523 if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4524 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4530 m = vm_page_next(m);
4536 * Mapping function for valid or dirty bits in a page.
4538 * Inputs are required to range within a page.
4541 vm_page_bits(int base, int size)
4547 base + size <= PAGE_SIZE,
4548 ("vm_page_bits: illegal base/size %d/%d", base, size)
4551 if (size == 0) /* handle degenerate case */
4554 first_bit = base >> DEV_BSHIFT;
4555 last_bit = (base + size - 1) >> DEV_BSHIFT;
4557 return (((vm_page_bits_t)2 << last_bit) -
4558 ((vm_page_bits_t)1 << first_bit));
4562 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set)
4565 #if PAGE_SIZE == 32768
4566 atomic_set_64((uint64_t *)bits, set);
4567 #elif PAGE_SIZE == 16384
4568 atomic_set_32((uint32_t *)bits, set);
4569 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16)
4570 atomic_set_16((uint16_t *)bits, set);
4571 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8)
4572 atomic_set_8((uint8_t *)bits, set);
4573 #else /* PAGE_SIZE <= 8192 */
4577 addr = (uintptr_t)bits;
4579 * Use a trick to perform a 32-bit atomic on the
4580 * containing aligned word, to not depend on the existence
4581 * of atomic_{set, clear}_{8, 16}.
4583 shift = addr & (sizeof(uint32_t) - 1);
4584 #if BYTE_ORDER == BIG_ENDIAN
4585 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4589 addr &= ~(sizeof(uint32_t) - 1);
4590 atomic_set_32((uint32_t *)addr, set << shift);
4591 #endif /* PAGE_SIZE */
4595 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear)
4598 #if PAGE_SIZE == 32768
4599 atomic_clear_64((uint64_t *)bits, clear);
4600 #elif PAGE_SIZE == 16384
4601 atomic_clear_32((uint32_t *)bits, clear);
4602 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16)
4603 atomic_clear_16((uint16_t *)bits, clear);
4604 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8)
4605 atomic_clear_8((uint8_t *)bits, clear);
4606 #else /* PAGE_SIZE <= 8192 */
4610 addr = (uintptr_t)bits;
4612 * Use a trick to perform a 32-bit atomic on the
4613 * containing aligned word, to not depend on the existence
4614 * of atomic_{set, clear}_{8, 16}.
4616 shift = addr & (sizeof(uint32_t) - 1);
4617 #if BYTE_ORDER == BIG_ENDIAN
4618 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4622 addr &= ~(sizeof(uint32_t) - 1);
4623 atomic_clear_32((uint32_t *)addr, clear << shift);
4624 #endif /* PAGE_SIZE */
4627 static inline vm_page_bits_t
4628 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits)
4630 #if PAGE_SIZE == 32768
4634 while (atomic_fcmpset_64(bits, &old, newbits) == 0);
4636 #elif PAGE_SIZE == 16384
4640 while (atomic_fcmpset_32(bits, &old, newbits) == 0);
4642 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16)
4646 while (atomic_fcmpset_16(bits, &old, newbits) == 0);
4648 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8)
4652 while (atomic_fcmpset_8(bits, &old, newbits) == 0);
4654 #else /* PAGE_SIZE <= 4096*/
4656 uint32_t old, new, mask;
4659 addr = (uintptr_t)bits;
4661 * Use a trick to perform a 32-bit atomic on the
4662 * containing aligned word, to not depend on the existence
4663 * of atomic_{set, swap, clear}_{8, 16}.
4665 shift = addr & (sizeof(uint32_t) - 1);
4666 #if BYTE_ORDER == BIG_ENDIAN
4667 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4671 addr &= ~(sizeof(uint32_t) - 1);
4672 mask = VM_PAGE_BITS_ALL << shift;
4677 new |= newbits << shift;
4678 } while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0);
4679 return (old >> shift);
4680 #endif /* PAGE_SIZE */
4684 * vm_page_set_valid_range:
4686 * Sets portions of a page valid. The arguments are expected
4687 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4688 * of any partial chunks touched by the range. The invalid portion of
4689 * such chunks will be zeroed.
4691 * (base + size) must be less then or equal to PAGE_SIZE.
4694 vm_page_set_valid_range(vm_page_t m, int base, int size)
4697 vm_page_bits_t pagebits;
4699 vm_page_assert_busied(m);
4700 if (size == 0) /* handle degenerate case */
4704 * If the base is not DEV_BSIZE aligned and the valid
4705 * bit is clear, we have to zero out a portion of the
4708 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4709 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
4710 pmap_zero_page_area(m, frag, base - frag);
4713 * If the ending offset is not DEV_BSIZE aligned and the
4714 * valid bit is clear, we have to zero out a portion of
4717 endoff = base + size;
4718 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4719 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
4720 pmap_zero_page_area(m, endoff,
4721 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4724 * Assert that no previously invalid block that is now being validated
4727 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
4728 ("vm_page_set_valid_range: page %p is dirty", m));
4731 * Set valid bits inclusive of any overlap.
4733 pagebits = vm_page_bits(base, size);
4734 if (vm_page_xbusied(m))
4735 m->valid |= pagebits;
4737 vm_page_bits_set(m, &m->valid, pagebits);
4741 * Set the page dirty bits and free the invalid swap space if
4742 * present. Returns the previous dirty bits.
4745 vm_page_set_dirty(vm_page_t m)
4749 VM_PAGE_OBJECT_BUSY_ASSERT(m);
4751 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) {
4753 m->dirty = VM_PAGE_BITS_ALL;
4755 old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL);
4756 if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0)
4757 vm_pager_page_unswapped(m);
4763 * Clear the given bits from the specified page's dirty field.
4765 static __inline void
4766 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
4769 vm_page_assert_busied(m);
4772 * If the page is xbusied and not write mapped we are the
4773 * only thread that can modify dirty bits. Otherwise, The pmap
4774 * layer can call vm_page_dirty() without holding a distinguished
4775 * lock. The combination of page busy and atomic operations
4776 * suffice to guarantee consistency of the page dirty field.
4778 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
4779 m->dirty &= ~pagebits;
4781 vm_page_bits_clear(m, &m->dirty, pagebits);
4785 * vm_page_set_validclean:
4787 * Sets portions of a page valid and clean. The arguments are expected
4788 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4789 * of any partial chunks touched by the range. The invalid portion of
4790 * such chunks will be zero'd.
4792 * (base + size) must be less then or equal to PAGE_SIZE.
4795 vm_page_set_validclean(vm_page_t m, int base, int size)
4797 vm_page_bits_t oldvalid, pagebits;
4800 vm_page_assert_busied(m);
4801 if (size == 0) /* handle degenerate case */
4805 * If the base is not DEV_BSIZE aligned and the valid
4806 * bit is clear, we have to zero out a portion of the
4809 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4810 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
4811 pmap_zero_page_area(m, frag, base - frag);
4814 * If the ending offset is not DEV_BSIZE aligned and the
4815 * valid bit is clear, we have to zero out a portion of
4818 endoff = base + size;
4819 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4820 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
4821 pmap_zero_page_area(m, endoff,
4822 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4825 * Set valid, clear dirty bits. If validating the entire
4826 * page we can safely clear the pmap modify bit. We also
4827 * use this opportunity to clear the PGA_NOSYNC flag. If a process
4828 * takes a write fault on a MAP_NOSYNC memory area the flag will
4831 * We set valid bits inclusive of any overlap, but we can only
4832 * clear dirty bits for DEV_BSIZE chunks that are fully within
4835 oldvalid = m->valid;
4836 pagebits = vm_page_bits(base, size);
4837 if (vm_page_xbusied(m))
4838 m->valid |= pagebits;
4840 vm_page_bits_set(m, &m->valid, pagebits);
4842 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
4843 frag = DEV_BSIZE - frag;
4849 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
4851 if (base == 0 && size == PAGE_SIZE) {
4853 * The page can only be modified within the pmap if it is
4854 * mapped, and it can only be mapped if it was previously
4857 if (oldvalid == VM_PAGE_BITS_ALL)
4859 * Perform the pmap_clear_modify() first. Otherwise,
4860 * a concurrent pmap operation, such as
4861 * pmap_protect(), could clear a modification in the
4862 * pmap and set the dirty field on the page before
4863 * pmap_clear_modify() had begun and after the dirty
4864 * field was cleared here.
4866 pmap_clear_modify(m);
4868 vm_page_aflag_clear(m, PGA_NOSYNC);
4869 } else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m))
4870 m->dirty &= ~pagebits;
4872 vm_page_clear_dirty_mask(m, pagebits);
4876 vm_page_clear_dirty(vm_page_t m, int base, int size)
4879 vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
4883 * vm_page_set_invalid:
4885 * Invalidates DEV_BSIZE'd chunks within a page. Both the
4886 * valid and dirty bits for the effected areas are cleared.
4889 vm_page_set_invalid(vm_page_t m, int base, int size)
4891 vm_page_bits_t bits;
4895 * The object lock is required so that pages can't be mapped
4896 * read-only while we're in the process of invalidating them.
4899 VM_OBJECT_ASSERT_WLOCKED(object);
4900 vm_page_assert_busied(m);
4902 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
4903 size >= object->un_pager.vnp.vnp_size)
4904 bits = VM_PAGE_BITS_ALL;
4906 bits = vm_page_bits(base, size);
4907 if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0)
4909 KASSERT((bits == 0 && vm_page_all_valid(m)) ||
4910 !pmap_page_is_mapped(m),
4911 ("vm_page_set_invalid: page %p is mapped", m));
4912 if (vm_page_xbusied(m)) {
4916 vm_page_bits_clear(m, &m->valid, bits);
4917 vm_page_bits_clear(m, &m->dirty, bits);
4924 * Invalidates the entire page. The page must be busy, unmapped, and
4925 * the enclosing object must be locked. The object locks protects
4926 * against concurrent read-only pmap enter which is done without
4930 vm_page_invalid(vm_page_t m)
4933 vm_page_assert_busied(m);
4934 VM_OBJECT_ASSERT_LOCKED(m->object);
4935 MPASS(!pmap_page_is_mapped(m));
4937 if (vm_page_xbusied(m))
4940 vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL);
4944 * vm_page_zero_invalid()
4946 * The kernel assumes that the invalid portions of a page contain
4947 * garbage, but such pages can be mapped into memory by user code.
4948 * When this occurs, we must zero out the non-valid portions of the
4949 * page so user code sees what it expects.
4951 * Pages are most often semi-valid when the end of a file is mapped
4952 * into memory and the file's size is not page aligned.
4955 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
4961 * Scan the valid bits looking for invalid sections that
4962 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the
4963 * valid bit may be set ) have already been zeroed by
4964 * vm_page_set_validclean().
4966 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
4967 if (i == (PAGE_SIZE / DEV_BSIZE) ||
4968 (m->valid & ((vm_page_bits_t)1 << i))) {
4970 pmap_zero_page_area(m,
4971 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
4978 * setvalid is TRUE when we can safely set the zero'd areas
4979 * as being valid. We can do this if there are no cache consistancy
4980 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
4989 * Is (partial) page valid? Note that the case where size == 0
4990 * will return FALSE in the degenerate case where the page is
4991 * entirely invalid, and TRUE otherwise.
4993 * Some callers envoke this routine without the busy lock held and
4994 * handle races via higher level locks. Typical callers should
4995 * hold a busy lock to prevent invalidation.
4998 vm_page_is_valid(vm_page_t m, int base, int size)
5000 vm_page_bits_t bits;
5002 bits = vm_page_bits(base, size);
5003 return (m->valid != 0 && (m->valid & bits) == bits);
5007 * Returns true if all of the specified predicates are true for the entire
5008 * (super)page and false otherwise.
5011 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
5017 if (skip_m != NULL && skip_m->object != object)
5019 VM_OBJECT_ASSERT_LOCKED(object);
5020 npages = atop(pagesizes[m->psind]);
5023 * The physically contiguous pages that make up a superpage, i.e., a
5024 * page with a page size index ("psind") greater than zero, will
5025 * occupy adjacent entries in vm_page_array[].
5027 for (i = 0; i < npages; i++) {
5028 /* Always test object consistency, including "skip_m". */
5029 if (m[i].object != object)
5031 if (&m[i] == skip_m)
5033 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
5035 if ((flags & PS_ALL_DIRTY) != 0) {
5037 * Calling vm_page_test_dirty() or pmap_is_modified()
5038 * might stop this case from spuriously returning
5039 * "false". However, that would require a write lock
5040 * on the object containing "m[i]".
5042 if (m[i].dirty != VM_PAGE_BITS_ALL)
5045 if ((flags & PS_ALL_VALID) != 0 &&
5046 m[i].valid != VM_PAGE_BITS_ALL)
5053 * Set the page's dirty bits if the page is modified.
5056 vm_page_test_dirty(vm_page_t m)
5059 vm_page_assert_busied(m);
5060 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
5065 vm_page_valid(vm_page_t m)
5068 vm_page_assert_busied(m);
5069 if (vm_page_xbusied(m))
5070 m->valid = VM_PAGE_BITS_ALL;
5072 vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL);
5076 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
5079 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
5083 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
5086 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
5090 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
5093 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
5096 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
5098 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
5101 vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
5105 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
5108 mtx_assert_(vm_page_lockptr(m), a, file, line);
5114 vm_page_object_busy_assert(vm_page_t m)
5118 * Certain of the page's fields may only be modified by the
5119 * holder of a page or object busy.
5121 if (m->object != NULL && !vm_page_busied(m))
5122 VM_OBJECT_ASSERT_BUSY(m->object);
5126 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits)
5129 if ((bits & PGA_WRITEABLE) == 0)
5133 * The PGA_WRITEABLE flag can only be set if the page is
5134 * managed, is exclusively busied or the object is locked.
5135 * Currently, this flag is only set by pmap_enter().
5137 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5138 ("PGA_WRITEABLE on unmanaged page"));
5139 if (!vm_page_xbusied(m))
5140 VM_OBJECT_ASSERT_BUSY(m->object);
5144 #include "opt_ddb.h"
5146 #include <sys/kernel.h>
5148 #include <ddb/ddb.h>
5150 DB_SHOW_COMMAND(page, vm_page_print_page_info)
5153 db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
5154 db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
5155 db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
5156 db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
5157 db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
5158 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
5159 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
5160 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
5161 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
5164 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
5168 db_printf("pq_free %d\n", vm_free_count());
5169 for (dom = 0; dom < vm_ndomains; dom++) {
5171 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
5173 vm_dom[dom].vmd_page_count,
5174 vm_dom[dom].vmd_free_count,
5175 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
5176 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
5177 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
5178 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
5182 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
5185 boolean_t phys, virt;
5188 db_printf("show pginfo addr\n");
5192 phys = strchr(modif, 'p') != NULL;
5193 virt = strchr(modif, 'v') != NULL;
5195 m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
5197 m = PHYS_TO_VM_PAGE(addr);
5199 m = (vm_page_t)addr;
5201 "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref %u\n"
5202 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
5203 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
5204 m->a.queue, m->ref_count, m->a.flags, m->oflags,
5205 m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty);