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 | CTLFLAG_MPSAFE, 0,
131 "VM page statistics");
133 static COUNTER_U64_DEFINE_EARLY(pqstate_commit_retries);
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_DEFINE_EARLY(queue_ops);
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_DEFINE_EARLY(queue_nops);
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 * bogus page -- for I/O to/from partially complete buffers,
150 * or for paging into sparsely invalid regions.
152 vm_page_t bogus_page;
154 vm_page_t vm_page_array;
155 long vm_page_array_size;
158 static TAILQ_HEAD(, vm_page) blacklist_head;
159 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
160 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
161 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
163 static uma_zone_t fakepg_zone;
165 static void vm_page_alloc_check(vm_page_t m);
166 static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m,
167 vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked);
168 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
169 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
170 static bool vm_page_free_prep(vm_page_t m);
171 static void vm_page_free_toq(vm_page_t m);
172 static void vm_page_init(void *dummy);
173 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
174 vm_pindex_t pindex, vm_page_t mpred);
175 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
177 static void vm_page_mvqueue(vm_page_t m, const uint8_t queue,
178 const uint16_t nflag);
179 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
180 vm_page_t m_run, vm_paddr_t high);
181 static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse);
182 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
184 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain,
186 static void vm_page_zone_release(void *arg, void **store, int cnt);
188 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
191 vm_page_init(void *dummy)
194 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
195 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
196 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
197 VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
201 * The cache page zone is initialized later since we need to be able to allocate
202 * pages before UMA is fully initialized.
205 vm_page_init_cache_zones(void *dummy __unused)
207 struct vm_domain *vmd;
208 struct vm_pgcache *pgcache;
209 int cache, domain, maxcache, pool;
212 TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &maxcache);
213 maxcache *= mp_ncpus;
214 for (domain = 0; domain < vm_ndomains; domain++) {
215 vmd = VM_DOMAIN(domain);
216 for (pool = 0; pool < VM_NFREEPOOL; pool++) {
217 pgcache = &vmd->vmd_pgcache[pool];
218 pgcache->domain = domain;
219 pgcache->pool = pool;
220 pgcache->zone = uma_zcache_create("vm pgcache",
221 PAGE_SIZE, NULL, NULL, NULL, NULL,
222 vm_page_zone_import, vm_page_zone_release, pgcache,
226 * Limit each pool's zone to 0.1% of the pages in the
229 cache = maxcache != 0 ? maxcache :
230 vmd->vmd_page_count / 1000;
231 uma_zone_set_maxcache(pgcache->zone, cache);
235 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
237 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
238 #if PAGE_SIZE == 32768
240 CTASSERT(sizeof(u_long) >= 8);
247 * Sets the page size, perhaps based upon the memory
248 * size. Must be called before any use of page-size
249 * dependent functions.
252 vm_set_page_size(void)
254 if (vm_cnt.v_page_size == 0)
255 vm_cnt.v_page_size = PAGE_SIZE;
256 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
257 panic("vm_set_page_size: page size not a power of two");
261 * vm_page_blacklist_next:
263 * Find the next entry in the provided string of blacklist
264 * addresses. Entries are separated by space, comma, or newline.
265 * If an invalid integer is encountered then the rest of the
266 * string is skipped. Updates the list pointer to the next
267 * character, or NULL if the string is exhausted or invalid.
270 vm_page_blacklist_next(char **list, char *end)
275 if (list == NULL || *list == NULL)
283 * If there's no end pointer then the buffer is coming from
284 * the kenv and we know it's null-terminated.
287 end = *list + strlen(*list);
289 /* Ensure that strtoq() won't walk off the end */
291 if (*end == '\n' || *end == ' ' || *end == ',')
294 printf("Blacklist not terminated, skipping\n");
300 for (pos = *list; *pos != '\0'; pos = cp) {
301 bad = strtoq(pos, &cp, 0);
302 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
311 if (*cp == '\0' || ++cp >= end)
315 return (trunc_page(bad));
317 printf("Garbage in RAM blacklist, skipping\n");
323 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
325 struct vm_domain *vmd;
329 m = vm_phys_paddr_to_vm_page(pa);
331 return (true); /* page does not exist, no failure */
333 vmd = vm_pagequeue_domain(m);
334 vm_domain_free_lock(vmd);
335 ret = vm_phys_unfree_page(m);
336 vm_domain_free_unlock(vmd);
338 vm_domain_freecnt_inc(vmd, -1);
339 TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
341 printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
347 * vm_page_blacklist_check:
349 * Iterate through the provided string of blacklist addresses, pulling
350 * each entry out of the physical allocator free list and putting it
351 * onto a list for reporting via the vm.page_blacklist sysctl.
354 vm_page_blacklist_check(char *list, char *end)
360 while (next != NULL) {
361 if ((pa = vm_page_blacklist_next(&next, end)) == 0)
363 vm_page_blacklist_add(pa, bootverbose);
368 * vm_page_blacklist_load:
370 * Search for a special module named "ram_blacklist". It'll be a
371 * plain text file provided by the user via the loader directive
375 vm_page_blacklist_load(char **list, char **end)
384 mod = preload_search_by_type("ram_blacklist");
386 ptr = preload_fetch_addr(mod);
387 len = preload_fetch_size(mod);
398 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
405 error = sysctl_wire_old_buffer(req, 0);
408 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
409 TAILQ_FOREACH(m, &blacklist_head, listq) {
410 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
411 (uintmax_t)m->phys_addr);
414 error = sbuf_finish(&sbuf);
420 * Initialize a dummy page for use in scans of the specified paging queue.
421 * In principle, this function only needs to set the flag PG_MARKER.
422 * Nonetheless, it write busies the page as a safety precaution.
425 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags)
428 bzero(marker, sizeof(*marker));
429 marker->flags = PG_MARKER;
430 marker->a.flags = aflags;
431 marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
432 marker->a.queue = queue;
436 vm_page_domain_init(int domain)
438 struct vm_domain *vmd;
439 struct vm_pagequeue *pq;
442 vmd = VM_DOMAIN(domain);
443 bzero(vmd, sizeof(*vmd));
444 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
445 "vm inactive pagequeue";
446 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
447 "vm active pagequeue";
448 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
449 "vm laundry pagequeue";
450 *__DECONST(const char **,
451 &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
452 "vm unswappable pagequeue";
453 vmd->vmd_domain = domain;
454 vmd->vmd_page_count = 0;
455 vmd->vmd_free_count = 0;
457 vmd->vmd_oom = FALSE;
458 for (i = 0; i < PQ_COUNT; i++) {
459 pq = &vmd->vmd_pagequeues[i];
460 TAILQ_INIT(&pq->pq_pl);
461 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
462 MTX_DEF | MTX_DUPOK);
464 vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
466 mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
467 mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
468 snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
471 * inacthead is used to provide FIFO ordering for LRU-bypassing
474 vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
475 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
476 &vmd->vmd_inacthead, plinks.q);
479 * The clock pages are used to implement active queue scanning without
480 * requeues. Scans start at clock[0], which is advanced after the scan
481 * ends. When the two clock hands meet, they are reset and scanning
482 * resumes from the head of the queue.
484 vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
485 vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
486 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
487 &vmd->vmd_clock[0], plinks.q);
488 TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
489 &vmd->vmd_clock[1], plinks.q);
493 * Initialize a physical page in preparation for adding it to the free
497 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
502 m->busy_lock = VPB_FREED;
503 m->flags = m->a.flags = 0;
505 m->a.queue = PQ_NONE;
508 m->order = VM_NFREEORDER;
509 m->pool = VM_FREEPOOL_DEFAULT;
510 m->valid = m->dirty = 0;
514 #ifndef PMAP_HAS_PAGE_ARRAY
516 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range)
521 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
522 * However, because this page is allocated from KVM, out-of-bounds
523 * accesses using the direct map will not be trapped.
528 * Allocate physical memory for the page structures, and map it.
530 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
531 vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end,
532 VM_PROT_READ | VM_PROT_WRITE);
533 vm_page_array_size = page_range;
542 * Initializes the resident memory module. Allocates physical memory for
543 * bootstrapping UMA and some data structures that are used to manage
544 * physical pages. Initializes these structures, and populates the free
548 vm_page_startup(vm_offset_t vaddr)
550 struct vm_phys_seg *seg;
552 char *list, *listend;
553 vm_paddr_t end, high_avail, low_avail, new_end, size;
554 vm_paddr_t page_range __unused;
555 vm_paddr_t last_pa, pa;
557 int biggestone, i, segind;
562 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
566 vaddr = round_page(vaddr);
568 vm_phys_early_startup();
569 biggestone = vm_phys_avail_largest();
570 end = phys_avail[biggestone+1];
573 * Initialize the page and queue locks.
575 mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
576 for (i = 0; i < PA_LOCK_COUNT; i++)
577 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
578 for (i = 0; i < vm_ndomains; i++)
579 vm_page_domain_init(i);
583 witness_size = round_page(witness_startup_count());
584 new_end -= witness_size;
585 mapped = pmap_map(&vaddr, new_end, new_end + witness_size,
586 VM_PROT_READ | VM_PROT_WRITE);
587 bzero((void *)mapped, witness_size);
588 witness_startup((void *)mapped);
591 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
592 defined(__i386__) || defined(__mips__) || defined(__riscv) || \
593 defined(__powerpc64__)
595 * Allocate a bitmap to indicate that a random physical page
596 * needs to be included in a minidump.
598 * The amd64 port needs this to indicate which direct map pages
599 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
601 * However, i386 still needs this workspace internally within the
602 * minidump code. In theory, they are not needed on i386, but are
603 * included should the sf_buf code decide to use them.
606 for (i = 0; dump_avail[i + 1] != 0; i += 2)
607 if (dump_avail[i + 1] > last_pa)
608 last_pa = dump_avail[i + 1];
609 page_range = last_pa / PAGE_SIZE;
610 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
611 new_end -= vm_page_dump_size;
612 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
613 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
614 bzero((void *)vm_page_dump, vm_page_dump_size);
618 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
619 defined(__riscv) || defined(__powerpc64__)
621 * Include the UMA bootstrap pages, witness pages and vm_page_dump
622 * in a crash dump. When pmap_map() uses the direct map, they are
623 * not automatically included.
625 for (pa = new_end; pa < end; pa += PAGE_SIZE)
628 phys_avail[biggestone + 1] = new_end;
631 * Request that the physical pages underlying the message buffer be
632 * included in a crash dump. Since the message buffer is accessed
633 * through the direct map, they are not automatically included.
635 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
636 last_pa = pa + round_page(msgbufsize);
637 while (pa < last_pa) {
643 * Compute the number of pages of memory that will be available for
644 * use, taking into account the overhead of a page structure per page.
645 * In other words, solve
646 * "available physical memory" - round_page(page_range *
647 * sizeof(struct vm_page)) = page_range * PAGE_SIZE
650 low_avail = phys_avail[0];
651 high_avail = phys_avail[1];
652 for (i = 0; i < vm_phys_nsegs; i++) {
653 if (vm_phys_segs[i].start < low_avail)
654 low_avail = vm_phys_segs[i].start;
655 if (vm_phys_segs[i].end > high_avail)
656 high_avail = vm_phys_segs[i].end;
658 /* Skip the first chunk. It is already accounted for. */
659 for (i = 2; phys_avail[i + 1] != 0; i += 2) {
660 if (phys_avail[i] < low_avail)
661 low_avail = phys_avail[i];
662 if (phys_avail[i + 1] > high_avail)
663 high_avail = phys_avail[i + 1];
665 first_page = low_avail / PAGE_SIZE;
666 #ifdef VM_PHYSSEG_SPARSE
668 for (i = 0; i < vm_phys_nsegs; i++)
669 size += vm_phys_segs[i].end - vm_phys_segs[i].start;
670 for (i = 0; phys_avail[i + 1] != 0; i += 2)
671 size += phys_avail[i + 1] - phys_avail[i];
672 #elif defined(VM_PHYSSEG_DENSE)
673 size = high_avail - low_avail;
675 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
678 #ifdef PMAP_HAS_PAGE_ARRAY
679 pmap_page_array_startup(size / PAGE_SIZE);
680 biggestone = vm_phys_avail_largest();
681 end = new_end = phys_avail[biggestone + 1];
683 #ifdef VM_PHYSSEG_DENSE
685 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
686 * the overhead of a page structure per page only if vm_page_array is
687 * allocated from the last physical memory chunk. Otherwise, we must
688 * allocate page structures representing the physical memory
689 * underlying vm_page_array, even though they will not be used.
691 if (new_end != high_avail)
692 page_range = size / PAGE_SIZE;
696 page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
699 * If the partial bytes remaining are large enough for
700 * a page (PAGE_SIZE) without a corresponding
701 * 'struct vm_page', then new_end will contain an
702 * extra page after subtracting the length of the VM
703 * page array. Compensate by subtracting an extra
706 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
707 if (new_end == high_avail)
708 high_avail -= PAGE_SIZE;
709 new_end -= PAGE_SIZE;
713 new_end = vm_page_array_alloc(&vaddr, end, page_range);
716 #if VM_NRESERVLEVEL > 0
718 * Allocate physical memory for the reservation management system's
719 * data structures, and map it.
721 new_end = vm_reserv_startup(&vaddr, new_end);
723 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
724 defined(__riscv) || defined(__powerpc64__)
726 * Include vm_page_array and vm_reserv_array in a crash dump.
728 for (pa = new_end; pa < end; pa += PAGE_SIZE)
731 phys_avail[biggestone + 1] = new_end;
734 * Add physical memory segments corresponding to the available
737 for (i = 0; phys_avail[i + 1] != 0; i += 2)
738 if (vm_phys_avail_size(i) != 0)
739 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
742 * Initialize the physical memory allocator.
747 * Initialize the page structures and add every available page to the
748 * physical memory allocator's free lists.
750 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
751 for (ii = 0; ii < vm_page_array_size; ii++) {
752 m = &vm_page_array[ii];
753 vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0);
754 m->flags = PG_FICTITIOUS;
757 vm_cnt.v_page_count = 0;
758 for (segind = 0; segind < vm_phys_nsegs; segind++) {
759 seg = &vm_phys_segs[segind];
760 for (m = seg->first_page, pa = seg->start; pa < seg->end;
761 m++, pa += PAGE_SIZE)
762 vm_page_init_page(m, pa, segind);
765 * Add the segment to the free lists only if it is covered by
766 * one of the ranges in phys_avail. Because we've added the
767 * ranges to the vm_phys_segs array, we can assume that each
768 * segment is either entirely contained in one of the ranges,
769 * or doesn't overlap any of them.
771 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
772 struct vm_domain *vmd;
774 if (seg->start < phys_avail[i] ||
775 seg->end > phys_avail[i + 1])
779 pagecount = (u_long)atop(seg->end - seg->start);
781 vmd = VM_DOMAIN(seg->domain);
782 vm_domain_free_lock(vmd);
783 vm_phys_enqueue_contig(m, pagecount);
784 vm_domain_free_unlock(vmd);
785 vm_domain_freecnt_inc(vmd, pagecount);
786 vm_cnt.v_page_count += (u_int)pagecount;
788 vmd = VM_DOMAIN(seg->domain);
789 vmd->vmd_page_count += (u_int)pagecount;
790 vmd->vmd_segs |= 1UL << m->segind;
796 * Remove blacklisted pages from the physical memory allocator.
798 TAILQ_INIT(&blacklist_head);
799 vm_page_blacklist_load(&list, &listend);
800 vm_page_blacklist_check(list, listend);
802 list = kern_getenv("vm.blacklist");
803 vm_page_blacklist_check(list, NULL);
806 #if VM_NRESERVLEVEL > 0
808 * Initialize the reservation management system.
817 vm_page_reference(vm_page_t m)
820 vm_page_aflag_set(m, PGA_REFERENCED);
826 * Helper routine for grab functions to trylock busy.
828 * Returns true on success and false on failure.
831 vm_page_trybusy(vm_page_t m, int allocflags)
834 if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0)
835 return (vm_page_trysbusy(m));
837 return (vm_page_tryxbusy(m));
843 * Helper routine for grab functions to trylock busy and wire.
845 * Returns true on success and false on failure.
848 vm_page_tryacquire(vm_page_t m, int allocflags)
852 locked = vm_page_trybusy(m, allocflags);
853 if (locked && (allocflags & VM_ALLOC_WIRED) != 0)
859 * vm_page_busy_acquire:
861 * Acquire the busy lock as described by VM_ALLOC_* flags. Will loop
862 * and drop the object lock if necessary.
865 vm_page_busy_acquire(vm_page_t m, int allocflags)
871 * The page-specific object must be cached because page
872 * identity can change during the sleep, causing the
873 * re-lock of a different object.
874 * It is assumed that a reference to the object is already
875 * held by the callers.
879 if (vm_page_tryacquire(m, allocflags))
881 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
884 locked = VM_OBJECT_WOWNED(obj);
887 MPASS(locked || vm_page_wired(m));
888 if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags,
890 VM_OBJECT_WLOCK(obj);
891 if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
893 KASSERT(m->object == obj || m->object == NULL,
894 ("vm_page_busy_acquire: page %p does not belong to %p",
900 * vm_page_busy_downgrade:
902 * Downgrade an exclusive busy page into a single shared busy page.
905 vm_page_busy_downgrade(vm_page_t m)
909 vm_page_assert_xbusied(m);
913 if (atomic_fcmpset_rel_int(&m->busy_lock,
914 &x, VPB_SHARERS_WORD(1)))
917 if ((x & VPB_BIT_WAITERS) != 0)
923 * vm_page_busy_tryupgrade:
925 * Attempt to upgrade a single shared busy into an exclusive busy.
928 vm_page_busy_tryupgrade(vm_page_t m)
932 vm_page_assert_sbusied(m);
935 ce = VPB_CURTHREAD_EXCLUSIVE;
937 if (VPB_SHARERS(x) > 1)
939 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
940 ("vm_page_busy_tryupgrade: invalid lock state"));
941 if (!atomic_fcmpset_acq_int(&m->busy_lock, &x,
942 ce | (x & VPB_BIT_WAITERS)))
951 * Return a positive value if the page is shared busied, 0 otherwise.
954 vm_page_sbusied(vm_page_t m)
959 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
965 * Shared unbusy a page.
968 vm_page_sunbusy(vm_page_t m)
972 vm_page_assert_sbusied(m);
976 KASSERT(x != VPB_FREED,
977 ("vm_page_sunbusy: Unlocking freed page."));
978 if (VPB_SHARERS(x) > 1) {
979 if (atomic_fcmpset_int(&m->busy_lock, &x,
984 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
985 ("vm_page_sunbusy: invalid lock state"));
986 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
988 if ((x & VPB_BIT_WAITERS) == 0)
996 * vm_page_busy_sleep:
998 * Sleep if the page is busy, using the page pointer as wchan.
999 * This is used to implement the hard-path of busying mechanism.
1001 * If nonshared is true, sleep only if the page is xbusy.
1003 * The object lock must be held on entry and will be released on exit.
1006 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
1011 VM_OBJECT_ASSERT_LOCKED(obj);
1012 vm_page_lock_assert(m, MA_NOTOWNED);
1014 if (!_vm_page_busy_sleep(obj, m, m->pindex, wmesg,
1015 nonshared ? VM_ALLOC_SBUSY : 0 , true))
1016 VM_OBJECT_DROP(obj);
1020 * vm_page_busy_sleep_unlocked:
1022 * Sleep if the page is busy, using the page pointer as wchan.
1023 * This is used to implement the hard-path of busying mechanism.
1025 * If nonshared is true, sleep only if the page is xbusy.
1027 * The object lock must not be held on entry. The operation will
1028 * return if the page changes identity.
1031 vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1032 const char *wmesg, bool nonshared)
1035 VM_OBJECT_ASSERT_UNLOCKED(obj);
1036 vm_page_lock_assert(m, MA_NOTOWNED);
1038 _vm_page_busy_sleep(obj, m, pindex, wmesg,
1039 nonshared ? VM_ALLOC_SBUSY : 0, false);
1043 * _vm_page_busy_sleep:
1045 * Internal busy sleep function. Verifies the page identity and
1046 * lockstate against parameters. Returns true if it sleeps and
1049 * If locked is true the lock will be dropped for any true returns
1050 * and held for any false returns.
1053 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1054 const char *wmesg, int allocflags, bool locked)
1060 * If the object is busy we must wait for that to drain to zero
1061 * before trying the page again.
1063 if (obj != NULL && vm_object_busied(obj)) {
1065 VM_OBJECT_DROP(obj);
1066 vm_object_busy_wait(obj, wmesg);
1070 if (!vm_page_busied(m))
1073 xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0;
1075 x = atomic_load_int(&m->busy_lock);
1078 * If the page changes objects or becomes unlocked we can
1081 if (x == VPB_UNBUSIED ||
1082 (xsleep && (x & VPB_BIT_SHARED) != 0) ||
1083 m->object != obj || m->pindex != pindex) {
1087 if ((x & VPB_BIT_WAITERS) != 0)
1089 } while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS));
1091 VM_OBJECT_DROP(obj);
1093 sleepq_add(m, NULL, wmesg, 0, 0);
1094 sleepq_wait(m, PVM);
1102 * Try to shared busy a page.
1103 * If the operation succeeds 1 is returned otherwise 0.
1104 * The operation never sleeps.
1107 vm_page_trysbusy(vm_page_t m)
1115 if ((x & VPB_BIT_SHARED) == 0)
1118 * Reduce the window for transient busies that will trigger
1119 * false negatives in vm_page_ps_test().
1121 if (obj != NULL && vm_object_busied(obj))
1123 if (atomic_fcmpset_acq_int(&m->busy_lock, &x,
1124 x + VPB_ONE_SHARER))
1128 /* Refetch the object now that we're guaranteed that it is stable. */
1130 if (obj != NULL && vm_object_busied(obj)) {
1140 * Try to exclusive busy a page.
1141 * If the operation succeeds 1 is returned otherwise 0.
1142 * The operation never sleeps.
1145 vm_page_tryxbusy(vm_page_t m)
1149 if (atomic_cmpset_acq_int(&(m)->busy_lock, VPB_UNBUSIED,
1150 VPB_CURTHREAD_EXCLUSIVE) == 0)
1154 if (obj != NULL && vm_object_busied(obj)) {
1162 vm_page_xunbusy_hard_tail(vm_page_t m)
1164 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1165 /* Wake the waiter. */
1170 * vm_page_xunbusy_hard:
1172 * Called when unbusy has failed because there is a waiter.
1175 vm_page_xunbusy_hard(vm_page_t m)
1177 vm_page_assert_xbusied(m);
1178 vm_page_xunbusy_hard_tail(m);
1182 vm_page_xunbusy_hard_unchecked(vm_page_t m)
1184 vm_page_assert_xbusied_unchecked(m);
1185 vm_page_xunbusy_hard_tail(m);
1189 vm_page_busy_free(vm_page_t m)
1193 atomic_thread_fence_rel();
1194 x = atomic_swap_int(&m->busy_lock, VPB_FREED);
1195 if ((x & VPB_BIT_WAITERS) != 0)
1200 * vm_page_unhold_pages:
1202 * Unhold each of the pages that is referenced by the given array.
1205 vm_page_unhold_pages(vm_page_t *ma, int count)
1208 for (; count != 0; count--) {
1209 vm_page_unwire(*ma, PQ_ACTIVE);
1215 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1219 #ifdef VM_PHYSSEG_SPARSE
1220 m = vm_phys_paddr_to_vm_page(pa);
1222 m = vm_phys_fictitious_to_vm_page(pa);
1224 #elif defined(VM_PHYSSEG_DENSE)
1228 if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1229 m = &vm_page_array[pi - first_page];
1232 return (vm_phys_fictitious_to_vm_page(pa));
1234 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1241 * Create a fictitious page with the specified physical address and
1242 * memory attribute. The memory attribute is the only the machine-
1243 * dependent aspect of a fictitious page that must be initialized.
1246 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1250 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1251 vm_page_initfake(m, paddr, memattr);
1256 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1259 if ((m->flags & PG_FICTITIOUS) != 0) {
1261 * The page's memattr might have changed since the
1262 * previous initialization. Update the pmap to the
1267 m->phys_addr = paddr;
1268 m->a.queue = PQ_NONE;
1269 /* Fictitious pages don't use "segind". */
1270 m->flags = PG_FICTITIOUS;
1271 /* Fictitious pages don't use "order" or "pool". */
1272 m->oflags = VPO_UNMANAGED;
1273 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
1274 /* Fictitious pages are unevictable. */
1278 pmap_page_set_memattr(m, memattr);
1284 * Release a fictitious page.
1287 vm_page_putfake(vm_page_t m)
1290 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1291 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1292 ("vm_page_putfake: bad page %p", m));
1293 vm_page_assert_xbusied(m);
1294 vm_page_busy_free(m);
1295 uma_zfree(fakepg_zone, m);
1299 * vm_page_updatefake:
1301 * Update the given fictitious page to the specified physical address and
1305 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1308 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1309 ("vm_page_updatefake: bad page %p", m));
1310 m->phys_addr = paddr;
1311 pmap_page_set_memattr(m, memattr);
1320 vm_page_free(vm_page_t m)
1323 m->flags &= ~PG_ZERO;
1324 vm_page_free_toq(m);
1328 * vm_page_free_zero:
1330 * Free a page to the zerod-pages queue
1333 vm_page_free_zero(vm_page_t m)
1336 m->flags |= PG_ZERO;
1337 vm_page_free_toq(m);
1341 * Unbusy and handle the page queueing for a page from a getpages request that
1342 * was optionally read ahead or behind.
1345 vm_page_readahead_finish(vm_page_t m)
1348 /* We shouldn't put invalid pages on queues. */
1349 KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m));
1352 * Since the page is not the actually needed one, whether it should
1353 * be activated or deactivated is not obvious. Empirical results
1354 * have shown that deactivating the page is usually the best choice,
1355 * unless the page is wanted by another thread.
1357 if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1358 vm_page_activate(m);
1360 vm_page_deactivate(m);
1361 vm_page_xunbusy_unchecked(m);
1365 * vm_page_sleep_if_busy:
1367 * Sleep and release the object lock if the page is busied.
1368 * Returns TRUE if the thread slept.
1370 * The given page must be unlocked and object containing it must
1374 vm_page_sleep_if_busy(vm_page_t m, const char *wmesg)
1378 vm_page_lock_assert(m, MA_NOTOWNED);
1379 VM_OBJECT_ASSERT_WLOCKED(m->object);
1382 * The page-specific object must be cached because page
1383 * identity can change during the sleep, causing the
1384 * re-lock of a different object.
1385 * It is assumed that a reference to the object is already
1386 * held by the callers.
1389 if (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, 0, true)) {
1390 VM_OBJECT_WLOCK(obj);
1397 * vm_page_sleep_if_xbusy:
1399 * Sleep and release the object lock if the page is xbusied.
1400 * Returns TRUE if the thread slept.
1402 * The given page must be unlocked and object containing it must
1406 vm_page_sleep_if_xbusy(vm_page_t m, const char *wmesg)
1410 vm_page_lock_assert(m, MA_NOTOWNED);
1411 VM_OBJECT_ASSERT_WLOCKED(m->object);
1414 * The page-specific object must be cached because page
1415 * identity can change during the sleep, causing the
1416 * re-lock of a different object.
1417 * It is assumed that a reference to the object is already
1418 * held by the callers.
1421 if (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, VM_ALLOC_SBUSY,
1423 VM_OBJECT_WLOCK(obj);
1430 * vm_page_dirty_KBI: [ internal use only ]
1432 * Set all bits in the page's dirty field.
1434 * The object containing the specified page must be locked if the
1435 * call is made from the machine-independent layer.
1437 * See vm_page_clear_dirty_mask().
1439 * This function should only be called by vm_page_dirty().
1442 vm_page_dirty_KBI(vm_page_t m)
1445 /* Refer to this operation by its public name. */
1446 KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!"));
1447 m->dirty = VM_PAGE_BITS_ALL;
1451 * vm_page_insert: [ internal use only ]
1453 * Inserts the given mem entry into the object and object list.
1455 * The object must be locked.
1458 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1462 VM_OBJECT_ASSERT_WLOCKED(object);
1463 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1464 return (vm_page_insert_after(m, object, pindex, mpred));
1468 * vm_page_insert_after:
1470 * Inserts the page "m" into the specified object at offset "pindex".
1472 * The page "mpred" must immediately precede the offset "pindex" within
1473 * the specified object.
1475 * The object must be locked.
1478 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1483 VM_OBJECT_ASSERT_WLOCKED(object);
1484 KASSERT(m->object == NULL,
1485 ("vm_page_insert_after: page already inserted"));
1486 if (mpred != NULL) {
1487 KASSERT(mpred->object == object,
1488 ("vm_page_insert_after: object doesn't contain mpred"));
1489 KASSERT(mpred->pindex < pindex,
1490 ("vm_page_insert_after: mpred doesn't precede pindex"));
1491 msucc = TAILQ_NEXT(mpred, listq);
1493 msucc = TAILQ_FIRST(&object->memq);
1495 KASSERT(msucc->pindex > pindex,
1496 ("vm_page_insert_after: msucc doesn't succeed pindex"));
1499 * Record the object/offset pair in this page.
1503 m->ref_count |= VPRC_OBJREF;
1506 * Now link into the object's ordered list of backed pages.
1508 if (vm_radix_insert(&object->rtree, m)) {
1511 m->ref_count &= ~VPRC_OBJREF;
1514 vm_page_insert_radixdone(m, object, mpred);
1519 * vm_page_insert_radixdone:
1521 * Complete page "m" insertion into the specified object after the
1522 * radix trie hooking.
1524 * The page "mpred" must precede the offset "m->pindex" within the
1527 * The object must be locked.
1530 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1533 VM_OBJECT_ASSERT_WLOCKED(object);
1534 KASSERT(object != NULL && m->object == object,
1535 ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1536 KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1537 ("vm_page_insert_radixdone: page %p is missing object ref", m));
1538 if (mpred != NULL) {
1539 KASSERT(mpred->object == object,
1540 ("vm_page_insert_radixdone: object doesn't contain mpred"));
1541 KASSERT(mpred->pindex < m->pindex,
1542 ("vm_page_insert_radixdone: mpred doesn't precede pindex"));
1546 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1548 TAILQ_INSERT_HEAD(&object->memq, m, listq);
1551 * Show that the object has one more resident page.
1553 object->resident_page_count++;
1556 * Hold the vnode until the last page is released.
1558 if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1559 vhold(object->handle);
1562 * Since we are inserting a new and possibly dirty page,
1563 * update the object's generation count.
1565 if (pmap_page_is_write_mapped(m))
1566 vm_object_set_writeable_dirty(object);
1570 * Do the work to remove a page from its object. The caller is responsible for
1571 * updating the page's fields to reflect this removal.
1574 vm_page_object_remove(vm_page_t m)
1579 vm_page_assert_xbusied(m);
1581 VM_OBJECT_ASSERT_WLOCKED(object);
1582 KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1583 ("page %p is missing its object ref", m));
1585 /* Deferred free of swap space. */
1586 if ((m->a.flags & PGA_SWAP_FREE) != 0)
1587 vm_pager_page_unswapped(m);
1590 mrem = vm_radix_remove(&object->rtree, m->pindex);
1591 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1594 * Now remove from the object's list of backed pages.
1596 TAILQ_REMOVE(&object->memq, m, listq);
1599 * And show that the object has one fewer resident page.
1601 object->resident_page_count--;
1604 * The vnode may now be recycled.
1606 if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1607 vdrop(object->handle);
1613 * Removes the specified page from its containing object, but does not
1614 * invalidate any backing storage. Returns true if the object's reference
1615 * was the last reference to the page, and false otherwise.
1617 * The object must be locked and the page must be exclusively busied.
1618 * The exclusive busy will be released on return. If this is not the
1619 * final ref and the caller does not hold a wire reference it may not
1620 * continue to access the page.
1623 vm_page_remove(vm_page_t m)
1627 dropped = vm_page_remove_xbusy(m);
1634 * vm_page_remove_xbusy
1636 * Removes the page but leaves the xbusy held. Returns true if this
1637 * removed the final ref and false otherwise.
1640 vm_page_remove_xbusy(vm_page_t m)
1643 vm_page_object_remove(m);
1644 return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1650 * Returns the page associated with the object/offset
1651 * pair specified; if none is found, NULL is returned.
1653 * The object must be locked.
1656 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1659 VM_OBJECT_ASSERT_LOCKED(object);
1660 return (vm_radix_lookup(&object->rtree, pindex));
1666 * Returns a page that must already have been busied by
1667 * the caller. Used for bogus page replacement.
1670 vm_page_relookup(vm_object_t object, vm_pindex_t pindex)
1674 m = vm_radix_lookup_unlocked(&object->rtree, pindex);
1675 KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) &&
1676 m->object == object && m->pindex == pindex,
1677 ("vm_page_relookup: Invalid page %p", m));
1682 * This should only be used by lockless functions for releasing transient
1683 * incorrect acquires. The page may have been freed after we acquired a
1684 * busy lock. In this case busy_lock == VPB_FREED and we have nothing
1688 vm_page_busy_release(vm_page_t m)
1692 x = atomic_load_int(&m->busy_lock);
1696 if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) {
1697 if (atomic_fcmpset_int(&m->busy_lock, &x,
1698 x - VPB_ONE_SHARER))
1702 KASSERT((x & VPB_BIT_SHARED) != 0 ||
1703 (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE,
1704 ("vm_page_busy_release: %p xbusy not owned.", m));
1705 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1707 if ((x & VPB_BIT_WAITERS) != 0)
1714 * vm_page_find_least:
1716 * Returns the page associated with the object with least pindex
1717 * greater than or equal to the parameter pindex, or NULL.
1719 * The object must be locked.
1722 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1726 VM_OBJECT_ASSERT_LOCKED(object);
1727 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1728 m = vm_radix_lookup_ge(&object->rtree, pindex);
1733 * Returns the given page's successor (by pindex) within the object if it is
1734 * resident; if none is found, NULL is returned.
1736 * The object must be locked.
1739 vm_page_next(vm_page_t m)
1743 VM_OBJECT_ASSERT_LOCKED(m->object);
1744 if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1745 MPASS(next->object == m->object);
1746 if (next->pindex != m->pindex + 1)
1753 * Returns the given page's predecessor (by pindex) within the object if it is
1754 * resident; if none is found, NULL is returned.
1756 * The object must be locked.
1759 vm_page_prev(vm_page_t m)
1763 VM_OBJECT_ASSERT_LOCKED(m->object);
1764 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1765 MPASS(prev->object == m->object);
1766 if (prev->pindex != m->pindex - 1)
1773 * Uses the page mnew as a replacement for an existing page at index
1774 * pindex which must be already present in the object.
1776 * Both pages must be exclusively busied on enter. The old page is
1779 * A return value of true means mold is now free. If this is not the
1780 * final ref and the caller does not hold a wire reference it may not
1781 * continue to access the page.
1784 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1790 VM_OBJECT_ASSERT_WLOCKED(object);
1791 vm_page_assert_xbusied(mold);
1792 KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
1793 ("vm_page_replace: page %p already in object", mnew));
1796 * This function mostly follows vm_page_insert() and
1797 * vm_page_remove() without the radix, object count and vnode
1798 * dance. Double check such functions for more comments.
1801 mnew->object = object;
1802 mnew->pindex = pindex;
1803 atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
1804 mret = vm_radix_replace(&object->rtree, mnew);
1805 KASSERT(mret == mold,
1806 ("invalid page replacement, mold=%p, mret=%p", mold, mret));
1807 KASSERT((mold->oflags & VPO_UNMANAGED) ==
1808 (mnew->oflags & VPO_UNMANAGED),
1809 ("vm_page_replace: mismatched VPO_UNMANAGED"));
1811 /* Keep the resident page list in sorted order. */
1812 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1813 TAILQ_REMOVE(&object->memq, mold, listq);
1814 mold->object = NULL;
1817 * The object's resident_page_count does not change because we have
1818 * swapped one page for another, but the generation count should
1819 * change if the page is dirty.
1821 if (pmap_page_is_write_mapped(mnew))
1822 vm_object_set_writeable_dirty(object);
1823 dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF;
1824 vm_page_xunbusy(mold);
1830 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1834 vm_page_assert_xbusied(mnew);
1836 if (vm_page_replace_hold(mnew, object, pindex, mold))
1843 * Move the given memory entry from its
1844 * current object to the specified target object/offset.
1846 * Note: swap associated with the page must be invalidated by the move. We
1847 * have to do this for several reasons: (1) we aren't freeing the
1848 * page, (2) we are dirtying the page, (3) the VM system is probably
1849 * moving the page from object A to B, and will then later move
1850 * the backing store from A to B and we can't have a conflict.
1852 * Note: we *always* dirty the page. It is necessary both for the
1853 * fact that we moved it, and because we may be invalidating
1856 * The objects must be locked.
1859 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1864 VM_OBJECT_ASSERT_WLOCKED(new_object);
1866 KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m));
1867 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1868 KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1869 ("vm_page_rename: pindex already renamed"));
1872 * Create a custom version of vm_page_insert() which does not depend
1873 * by m_prev and can cheat on the implementation aspects of the
1877 m->pindex = new_pindex;
1878 if (vm_radix_insert(&new_object->rtree, m)) {
1884 * The operation cannot fail anymore. The removal must happen before
1885 * the listq iterator is tainted.
1888 vm_page_object_remove(m);
1890 /* Return back to the new pindex to complete vm_page_insert(). */
1891 m->pindex = new_pindex;
1892 m->object = new_object;
1894 vm_page_insert_radixdone(m, new_object, mpred);
1902 * Allocate and return a page that is associated with the specified
1903 * object and offset pair. By default, this page is exclusive busied.
1905 * The caller must always specify an allocation class.
1907 * allocation classes:
1908 * VM_ALLOC_NORMAL normal process request
1909 * VM_ALLOC_SYSTEM system *really* needs a page
1910 * VM_ALLOC_INTERRUPT interrupt time request
1912 * optional allocation flags:
1913 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1914 * intends to allocate
1915 * VM_ALLOC_NOBUSY do not exclusive busy the page
1916 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1917 * VM_ALLOC_NOOBJ page is not associated with an object and
1918 * should not be exclusive busy
1919 * VM_ALLOC_SBUSY shared busy the allocated page
1920 * VM_ALLOC_WIRED wire the allocated page
1921 * VM_ALLOC_ZERO prefer a zeroed page
1924 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1927 return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1928 vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1932 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1936 return (vm_page_alloc_domain_after(object, pindex, domain, req,
1937 object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) :
1942 * Allocate a page in the specified object with the given page index. To
1943 * optimize insertion of the page into the object, the caller must also specifiy
1944 * the resident page in the object with largest index smaller than the given
1945 * page index, or NULL if no such page exists.
1948 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1949 int req, vm_page_t mpred)
1951 struct vm_domainset_iter di;
1955 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1957 m = vm_page_alloc_domain_after(object, pindex, domain, req,
1961 } while (vm_domainset_iter_page(&di, object, &domain) == 0);
1967 * Returns true if the number of free pages exceeds the minimum
1968 * for the request class and false otherwise.
1971 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
1973 u_int limit, old, new;
1975 if (req_class == VM_ALLOC_INTERRUPT)
1977 else if (req_class == VM_ALLOC_SYSTEM)
1978 limit = vmd->vmd_interrupt_free_min;
1980 limit = vmd->vmd_free_reserved;
1983 * Attempt to reserve the pages. Fail if we're below the limit.
1986 old = vmd->vmd_free_count;
1991 } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
1993 /* Wake the page daemon if we've crossed the threshold. */
1994 if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
1995 pagedaemon_wakeup(vmd->vmd_domain);
1997 /* Only update bitsets on transitions. */
1998 if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
1999 (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
2006 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
2011 * The page daemon is allowed to dig deeper into the free page list.
2013 req_class = req & VM_ALLOC_CLASS_MASK;
2014 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2015 req_class = VM_ALLOC_SYSTEM;
2016 return (_vm_domain_allocate(vmd, req_class, npages));
2020 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
2021 int req, vm_page_t mpred)
2023 struct vm_domain *vmd;
2027 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
2028 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
2029 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2030 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2031 ("inconsistent object(%p)/req(%x)", object, req));
2032 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2033 ("Can't sleep and retry object insertion."));
2034 KASSERT(mpred == NULL || mpred->pindex < pindex,
2035 ("mpred %p doesn't precede pindex 0x%jx", mpred,
2036 (uintmax_t)pindex));
2038 VM_OBJECT_ASSERT_WLOCKED(object);
2042 pool = object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT;
2044 #if VM_NRESERVLEVEL > 0
2046 * Can we allocate the page from a reservation?
2048 if (vm_object_reserv(object) &&
2049 (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
2054 vmd = VM_DOMAIN(domain);
2055 if (vmd->vmd_pgcache[pool].zone != NULL) {
2056 m = uma_zalloc(vmd->vmd_pgcache[pool].zone, M_NOWAIT | M_NOVM);
2058 flags |= PG_PCPU_CACHE;
2062 if (vm_domain_allocate(vmd, req, 1)) {
2064 * If not, allocate it from the free page queues.
2066 vm_domain_free_lock(vmd);
2067 m = vm_phys_alloc_pages(domain, pool, 0);
2068 vm_domain_free_unlock(vmd);
2070 vm_domain_freecnt_inc(vmd, 1);
2071 #if VM_NRESERVLEVEL > 0
2072 if (vm_reserv_reclaim_inactive(domain))
2079 * Not allocatable, give up.
2081 if (vm_domain_alloc_fail(vmd, object, req))
2087 * At this point we had better have found a good page.
2091 vm_page_alloc_check(m);
2094 * Initialize the page. Only the PG_ZERO flag is inherited.
2096 if ((req & VM_ALLOC_ZERO) != 0)
2097 flags |= (m->flags & PG_ZERO);
2098 if ((req & VM_ALLOC_NODUMP) != 0)
2102 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2104 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2105 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2106 else if ((req & VM_ALLOC_SBUSY) != 0)
2107 m->busy_lock = VPB_SHARERS_WORD(1);
2109 m->busy_lock = VPB_UNBUSIED;
2110 if (req & VM_ALLOC_WIRED) {
2116 if (object != NULL) {
2117 if (vm_page_insert_after(m, object, pindex, mpred)) {
2118 if (req & VM_ALLOC_WIRED) {
2122 KASSERT(m->object == NULL, ("page %p has object", m));
2123 m->oflags = VPO_UNMANAGED;
2124 m->busy_lock = VPB_UNBUSIED;
2125 /* Don't change PG_ZERO. */
2126 vm_page_free_toq(m);
2127 if (req & VM_ALLOC_WAITFAIL) {
2128 VM_OBJECT_WUNLOCK(object);
2130 VM_OBJECT_WLOCK(object);
2135 /* Ignore device objects; the pager sets "memattr" for them. */
2136 if (object->memattr != VM_MEMATTR_DEFAULT &&
2137 (object->flags & OBJ_FICTITIOUS) == 0)
2138 pmap_page_set_memattr(m, object->memattr);
2146 * vm_page_alloc_contig:
2148 * Allocate a contiguous set of physical pages of the given size "npages"
2149 * from the free lists. All of the physical pages must be at or above
2150 * the given physical address "low" and below the given physical address
2151 * "high". The given value "alignment" determines the alignment of the
2152 * first physical page in the set. If the given value "boundary" is
2153 * non-zero, then the set of physical pages cannot cross any physical
2154 * address boundary that is a multiple of that value. Both "alignment"
2155 * and "boundary" must be a power of two.
2157 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
2158 * then the memory attribute setting for the physical pages is configured
2159 * to the object's memory attribute setting. Otherwise, the memory
2160 * attribute setting for the physical pages is configured to "memattr",
2161 * overriding the object's memory attribute setting. However, if the
2162 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
2163 * memory attribute setting for the physical pages cannot be configured
2164 * to VM_MEMATTR_DEFAULT.
2166 * The specified object may not contain fictitious pages.
2168 * The caller must always specify an allocation class.
2170 * allocation classes:
2171 * VM_ALLOC_NORMAL normal process request
2172 * VM_ALLOC_SYSTEM system *really* needs a page
2173 * VM_ALLOC_INTERRUPT interrupt time request
2175 * optional allocation flags:
2176 * VM_ALLOC_NOBUSY do not exclusive busy the page
2177 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
2178 * VM_ALLOC_NOOBJ page is not associated with an object and
2179 * should not be exclusive busy
2180 * VM_ALLOC_SBUSY shared busy the allocated page
2181 * VM_ALLOC_WIRED wire the allocated page
2182 * VM_ALLOC_ZERO prefer a zeroed page
2185 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
2186 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2187 vm_paddr_t boundary, vm_memattr_t memattr)
2189 struct vm_domainset_iter di;
2193 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2195 m = vm_page_alloc_contig_domain(object, pindex, domain, req,
2196 npages, low, high, alignment, boundary, memattr);
2199 } while (vm_domainset_iter_page(&di, object, &domain) == 0);
2205 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2206 int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2207 vm_paddr_t boundary, vm_memattr_t memattr)
2209 struct vm_domain *vmd;
2210 vm_page_t m, m_ret, mpred;
2211 u_int busy_lock, flags, oflags;
2213 mpred = NULL; /* XXX: pacify gcc */
2214 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
2215 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
2216 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2217 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2218 ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
2220 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2221 ("Can't sleep and retry object insertion."));
2222 if (object != NULL) {
2223 VM_OBJECT_ASSERT_WLOCKED(object);
2224 KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2225 ("vm_page_alloc_contig: object %p has fictitious pages",
2228 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2230 if (object != NULL) {
2231 mpred = vm_radix_lookup_le(&object->rtree, pindex);
2232 KASSERT(mpred == NULL || mpred->pindex != pindex,
2233 ("vm_page_alloc_contig: pindex already allocated"));
2237 * Can we allocate the pages without the number of free pages falling
2238 * below the lower bound for the allocation class?
2242 #if VM_NRESERVLEVEL > 0
2244 * Can we allocate the pages from a reservation?
2246 if (vm_object_reserv(object) &&
2247 (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
2248 mpred, npages, low, high, alignment, boundary)) != NULL) {
2252 vmd = VM_DOMAIN(domain);
2253 if (vm_domain_allocate(vmd, req, npages)) {
2255 * allocate them from the free page queues.
2257 vm_domain_free_lock(vmd);
2258 m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2259 alignment, boundary);
2260 vm_domain_free_unlock(vmd);
2261 if (m_ret == NULL) {
2262 vm_domain_freecnt_inc(vmd, npages);
2263 #if VM_NRESERVLEVEL > 0
2264 if (vm_reserv_reclaim_contig(domain, npages, low,
2265 high, alignment, boundary))
2270 if (m_ret == NULL) {
2271 if (vm_domain_alloc_fail(vmd, object, req))
2275 #if VM_NRESERVLEVEL > 0
2278 for (m = m_ret; m < &m_ret[npages]; m++) {
2280 vm_page_alloc_check(m);
2284 * Initialize the pages. Only the PG_ZERO flag is inherited.
2287 if ((req & VM_ALLOC_ZERO) != 0)
2289 if ((req & VM_ALLOC_NODUMP) != 0)
2291 oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2293 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2294 busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2295 else if ((req & VM_ALLOC_SBUSY) != 0)
2296 busy_lock = VPB_SHARERS_WORD(1);
2298 busy_lock = VPB_UNBUSIED;
2299 if ((req & VM_ALLOC_WIRED) != 0)
2300 vm_wire_add(npages);
2301 if (object != NULL) {
2302 if (object->memattr != VM_MEMATTR_DEFAULT &&
2303 memattr == VM_MEMATTR_DEFAULT)
2304 memattr = object->memattr;
2306 for (m = m_ret; m < &m_ret[npages]; m++) {
2308 m->flags = (m->flags | PG_NODUMP) & flags;
2309 m->busy_lock = busy_lock;
2310 if ((req & VM_ALLOC_WIRED) != 0)
2314 if (object != NULL) {
2315 if (vm_page_insert_after(m, object, pindex, mpred)) {
2316 if ((req & VM_ALLOC_WIRED) != 0)
2317 vm_wire_sub(npages);
2318 KASSERT(m->object == NULL,
2319 ("page %p has object", m));
2321 for (m = m_ret; m < &m_ret[npages]; m++) {
2323 (req & VM_ALLOC_WIRED) != 0)
2325 m->oflags = VPO_UNMANAGED;
2326 m->busy_lock = VPB_UNBUSIED;
2327 /* Don't change PG_ZERO. */
2328 vm_page_free_toq(m);
2330 if (req & VM_ALLOC_WAITFAIL) {
2331 VM_OBJECT_WUNLOCK(object);
2333 VM_OBJECT_WLOCK(object);
2340 if (memattr != VM_MEMATTR_DEFAULT)
2341 pmap_page_set_memattr(m, memattr);
2348 * Check a page that has been freshly dequeued from a freelist.
2351 vm_page_alloc_check(vm_page_t m)
2354 KASSERT(m->object == NULL, ("page %p has object", m));
2355 KASSERT(m->a.queue == PQ_NONE &&
2356 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
2357 ("page %p has unexpected queue %d, flags %#x",
2358 m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK)));
2359 KASSERT(m->ref_count == 0, ("page %p has references", m));
2360 KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m));
2361 KASSERT(m->dirty == 0, ("page %p is dirty", m));
2362 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2363 ("page %p has unexpected memattr %d",
2364 m, pmap_page_get_memattr(m)));
2365 KASSERT(m->valid == 0, ("free page %p is valid", m));
2369 * vm_page_alloc_freelist:
2371 * Allocate a physical page from the specified free page list.
2373 * The caller must always specify an allocation class.
2375 * allocation classes:
2376 * VM_ALLOC_NORMAL normal process request
2377 * VM_ALLOC_SYSTEM system *really* needs a page
2378 * VM_ALLOC_INTERRUPT interrupt time request
2380 * optional allocation flags:
2381 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
2382 * intends to allocate
2383 * VM_ALLOC_WIRED wire the allocated page
2384 * VM_ALLOC_ZERO prefer a zeroed page
2387 vm_page_alloc_freelist(int freelist, int req)
2389 struct vm_domainset_iter di;
2393 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2395 m = vm_page_alloc_freelist_domain(domain, freelist, req);
2398 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2404 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2406 struct vm_domain *vmd;
2411 vmd = VM_DOMAIN(domain);
2413 if (vm_domain_allocate(vmd, req, 1)) {
2414 vm_domain_free_lock(vmd);
2415 m = vm_phys_alloc_freelist_pages(domain, freelist,
2416 VM_FREEPOOL_DIRECT, 0);
2417 vm_domain_free_unlock(vmd);
2419 vm_domain_freecnt_inc(vmd, 1);
2422 if (vm_domain_alloc_fail(vmd, NULL, req))
2427 vm_page_alloc_check(m);
2430 * Initialize the page. Only the PG_ZERO flag is inherited.
2434 if ((req & VM_ALLOC_ZERO) != 0)
2437 if ((req & VM_ALLOC_WIRED) != 0) {
2441 /* Unmanaged pages don't use "act_count". */
2442 m->oflags = VPO_UNMANAGED;
2447 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2449 struct vm_domain *vmd;
2450 struct vm_pgcache *pgcache;
2454 vmd = VM_DOMAIN(pgcache->domain);
2457 * The page daemon should avoid creating extra memory pressure since its
2458 * main purpose is to replenish the store of free pages.
2460 if (vmd->vmd_severeset || curproc == pageproc ||
2461 !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2463 domain = vmd->vmd_domain;
2464 vm_domain_free_lock(vmd);
2465 i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2466 (vm_page_t *)store);
2467 vm_domain_free_unlock(vmd);
2469 vm_domain_freecnt_inc(vmd, cnt - i);
2475 vm_page_zone_release(void *arg, void **store, int cnt)
2477 struct vm_domain *vmd;
2478 struct vm_pgcache *pgcache;
2483 vmd = VM_DOMAIN(pgcache->domain);
2484 vm_domain_free_lock(vmd);
2485 for (i = 0; i < cnt; i++) {
2486 m = (vm_page_t)store[i];
2487 vm_phys_free_pages(m, 0);
2489 vm_domain_free_unlock(vmd);
2490 vm_domain_freecnt_inc(vmd, cnt);
2493 #define VPSC_ANY 0 /* No restrictions. */
2494 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */
2495 #define VPSC_NOSUPER 2 /* Skip superpages. */
2498 * vm_page_scan_contig:
2500 * Scan vm_page_array[] between the specified entries "m_start" and
2501 * "m_end" for a run of contiguous physical pages that satisfy the
2502 * specified conditions, and return the lowest page in the run. The
2503 * specified "alignment" determines the alignment of the lowest physical
2504 * page in the run. If the specified "boundary" is non-zero, then the
2505 * run of physical pages cannot span a physical address that is a
2506 * multiple of "boundary".
2508 * "m_end" is never dereferenced, so it need not point to a vm_page
2509 * structure within vm_page_array[].
2511 * "npages" must be greater than zero. "m_start" and "m_end" must not
2512 * span a hole (or discontiguity) in the physical address space. Both
2513 * "alignment" and "boundary" must be a power of two.
2516 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2517 u_long alignment, vm_paddr_t boundary, int options)
2522 #if VM_NRESERVLEVEL > 0
2525 int m_inc, order, run_ext, run_len;
2527 KASSERT(npages > 0, ("npages is 0"));
2528 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2529 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2532 for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2533 KASSERT((m->flags & PG_MARKER) == 0,
2534 ("page %p is PG_MARKER", m));
2535 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1,
2536 ("fictitious page %p has invalid ref count", m));
2539 * If the current page would be the start of a run, check its
2540 * physical address against the end, alignment, and boundary
2541 * conditions. If it doesn't satisfy these conditions, either
2542 * terminate the scan or advance to the next page that
2543 * satisfies the failed condition.
2546 KASSERT(m_run == NULL, ("m_run != NULL"));
2547 if (m + npages > m_end)
2549 pa = VM_PAGE_TO_PHYS(m);
2550 if ((pa & (alignment - 1)) != 0) {
2551 m_inc = atop(roundup2(pa, alignment) - pa);
2554 if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2556 m_inc = atop(roundup2(pa, boundary) - pa);
2560 KASSERT(m_run != NULL, ("m_run == NULL"));
2564 if (vm_page_wired(m))
2566 #if VM_NRESERVLEVEL > 0
2567 else if ((level = vm_reserv_level(m)) >= 0 &&
2568 (options & VPSC_NORESERV) != 0) {
2570 /* Advance to the end of the reservation. */
2571 pa = VM_PAGE_TO_PHYS(m);
2572 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2576 else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2578 * The page is considered eligible for relocation if
2579 * and only if it could be laundered or reclaimed by
2582 VM_OBJECT_RLOCK(object);
2583 if (object != m->object) {
2584 VM_OBJECT_RUNLOCK(object);
2587 /* Don't care: PG_NODUMP, PG_ZERO. */
2588 if (object->type != OBJT_DEFAULT &&
2589 object->type != OBJT_SWAP &&
2590 object->type != OBJT_VNODE) {
2592 #if VM_NRESERVLEVEL > 0
2593 } else if ((options & VPSC_NOSUPER) != 0 &&
2594 (level = vm_reserv_level_iffullpop(m)) >= 0) {
2596 /* Advance to the end of the superpage. */
2597 pa = VM_PAGE_TO_PHYS(m);
2598 m_inc = atop(roundup2(pa + 1,
2599 vm_reserv_size(level)) - pa);
2601 } else if (object->memattr == VM_MEMATTR_DEFAULT &&
2602 vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) {
2604 * The page is allocated but eligible for
2605 * relocation. Extend the current run by one
2608 KASSERT(pmap_page_get_memattr(m) ==
2610 ("page %p has an unexpected memattr", m));
2611 KASSERT((m->oflags & (VPO_SWAPINPROG |
2612 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2613 ("page %p has unexpected oflags", m));
2614 /* Don't care: PGA_NOSYNC. */
2618 VM_OBJECT_RUNLOCK(object);
2619 #if VM_NRESERVLEVEL > 0
2620 } else if (level >= 0) {
2622 * The page is reserved but not yet allocated. In
2623 * other words, it is still free. Extend the current
2628 } else if ((order = m->order) < VM_NFREEORDER) {
2630 * The page is enqueued in the physical memory
2631 * allocator's free page queues. Moreover, it is the
2632 * first page in a power-of-two-sized run of
2633 * contiguous free pages. Add these pages to the end
2634 * of the current run, and jump ahead.
2636 run_ext = 1 << order;
2640 * Skip the page for one of the following reasons: (1)
2641 * It is enqueued in the physical memory allocator's
2642 * free page queues. However, it is not the first
2643 * page in a run of contiguous free pages. (This case
2644 * rarely occurs because the scan is performed in
2645 * ascending order.) (2) It is not reserved, and it is
2646 * transitioning from free to allocated. (Conversely,
2647 * the transition from allocated to free for managed
2648 * pages is blocked by the page lock.) (3) It is
2649 * allocated but not contained by an object and not
2650 * wired, e.g., allocated by Xen's balloon driver.
2656 * Extend or reset the current run of pages.
2669 if (run_len >= npages)
2675 * vm_page_reclaim_run:
2677 * Try to relocate each of the allocated virtual pages within the
2678 * specified run of physical pages to a new physical address. Free the
2679 * physical pages underlying the relocated virtual pages. A virtual page
2680 * is relocatable if and only if it could be laundered or reclaimed by
2681 * the page daemon. Whenever possible, a virtual page is relocated to a
2682 * physical address above "high".
2684 * Returns 0 if every physical page within the run was already free or
2685 * just freed by a successful relocation. Otherwise, returns a non-zero
2686 * value indicating why the last attempt to relocate a virtual page was
2689 * "req_class" must be an allocation class.
2692 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2695 struct vm_domain *vmd;
2696 struct spglist free;
2699 vm_page_t m, m_end, m_new;
2700 int error, order, req;
2702 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2703 ("req_class is not an allocation class"));
2707 m_end = m_run + npages;
2708 for (; error == 0 && m < m_end; m++) {
2709 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2710 ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2713 * Racily check for wirings. Races are handled once the object
2714 * lock is held and the page is unmapped.
2716 if (vm_page_wired(m))
2718 else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2720 * The page is relocated if and only if it could be
2721 * laundered or reclaimed by the page daemon.
2723 VM_OBJECT_WLOCK(object);
2724 /* Don't care: PG_NODUMP, PG_ZERO. */
2725 if (m->object != object ||
2726 (object->type != OBJT_DEFAULT &&
2727 object->type != OBJT_SWAP &&
2728 object->type != OBJT_VNODE))
2730 else if (object->memattr != VM_MEMATTR_DEFAULT)
2732 else if (vm_page_queue(m) != PQ_NONE &&
2733 vm_page_tryxbusy(m) != 0) {
2734 if (vm_page_wired(m)) {
2739 KASSERT(pmap_page_get_memattr(m) ==
2741 ("page %p has an unexpected memattr", m));
2742 KASSERT(m->oflags == 0,
2743 ("page %p has unexpected oflags", m));
2744 /* Don't care: PGA_NOSYNC. */
2745 if (!vm_page_none_valid(m)) {
2747 * First, try to allocate a new page
2748 * that is above "high". Failing
2749 * that, try to allocate a new page
2750 * that is below "m_run". Allocate
2751 * the new page between the end of
2752 * "m_run" and "high" only as a last
2755 req = req_class | VM_ALLOC_NOOBJ;
2756 if ((m->flags & PG_NODUMP) != 0)
2757 req |= VM_ALLOC_NODUMP;
2758 if (trunc_page(high) !=
2759 ~(vm_paddr_t)PAGE_MASK) {
2760 m_new = vm_page_alloc_contig(
2765 VM_MEMATTR_DEFAULT);
2768 if (m_new == NULL) {
2769 pa = VM_PAGE_TO_PHYS(m_run);
2770 m_new = vm_page_alloc_contig(
2772 0, pa - 1, PAGE_SIZE, 0,
2773 VM_MEMATTR_DEFAULT);
2775 if (m_new == NULL) {
2777 m_new = vm_page_alloc_contig(
2779 pa, high, PAGE_SIZE, 0,
2780 VM_MEMATTR_DEFAULT);
2782 if (m_new == NULL) {
2789 * Unmap the page and check for new
2790 * wirings that may have been acquired
2791 * through a pmap lookup.
2793 if (object->ref_count != 0 &&
2794 !vm_page_try_remove_all(m)) {
2796 vm_page_free(m_new);
2802 * Replace "m" with the new page. For
2803 * vm_page_replace(), "m" must be busy
2804 * and dequeued. Finally, change "m"
2805 * as if vm_page_free() was called.
2807 m_new->a.flags = m->a.flags &
2808 ~PGA_QUEUE_STATE_MASK;
2809 KASSERT(m_new->oflags == VPO_UNMANAGED,
2810 ("page %p is managed", m_new));
2812 pmap_copy_page(m, m_new);
2813 m_new->valid = m->valid;
2814 m_new->dirty = m->dirty;
2815 m->flags &= ~PG_ZERO;
2817 if (vm_page_replace_hold(m_new, object,
2819 vm_page_free_prep(m))
2820 SLIST_INSERT_HEAD(&free, m,
2824 * The new page must be deactivated
2825 * before the object is unlocked.
2827 vm_page_deactivate(m_new);
2829 m->flags &= ~PG_ZERO;
2831 if (vm_page_free_prep(m))
2832 SLIST_INSERT_HEAD(&free, m,
2834 KASSERT(m->dirty == 0,
2835 ("page %p is dirty", m));
2840 VM_OBJECT_WUNLOCK(object);
2842 MPASS(vm_phys_domain(m) == domain);
2843 vmd = VM_DOMAIN(domain);
2844 vm_domain_free_lock(vmd);
2846 if (order < VM_NFREEORDER) {
2848 * The page is enqueued in the physical memory
2849 * allocator's free page queues. Moreover, it
2850 * is the first page in a power-of-two-sized
2851 * run of contiguous free pages. Jump ahead
2852 * to the last page within that run, and
2853 * continue from there.
2855 m += (1 << order) - 1;
2857 #if VM_NRESERVLEVEL > 0
2858 else if (vm_reserv_is_page_free(m))
2861 vm_domain_free_unlock(vmd);
2862 if (order == VM_NFREEORDER)
2866 if ((m = SLIST_FIRST(&free)) != NULL) {
2869 vmd = VM_DOMAIN(domain);
2871 vm_domain_free_lock(vmd);
2873 MPASS(vm_phys_domain(m) == domain);
2874 SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2875 vm_phys_free_pages(m, 0);
2877 } while ((m = SLIST_FIRST(&free)) != NULL);
2878 vm_domain_free_unlock(vmd);
2879 vm_domain_freecnt_inc(vmd, cnt);
2886 CTASSERT(powerof2(NRUNS));
2888 #define RUN_INDEX(count) ((count) & (NRUNS - 1))
2890 #define MIN_RECLAIM 8
2893 * vm_page_reclaim_contig:
2895 * Reclaim allocated, contiguous physical memory satisfying the specified
2896 * conditions by relocating the virtual pages using that physical memory.
2897 * Returns true if reclamation is successful and false otherwise. Since
2898 * relocation requires the allocation of physical pages, reclamation may
2899 * fail due to a shortage of free pages. When reclamation fails, callers
2900 * are expected to perform vm_wait() before retrying a failed allocation
2901 * operation, e.g., vm_page_alloc_contig().
2903 * The caller must always specify an allocation class through "req".
2905 * allocation classes:
2906 * VM_ALLOC_NORMAL normal process request
2907 * VM_ALLOC_SYSTEM system *really* needs a page
2908 * VM_ALLOC_INTERRUPT interrupt time request
2910 * The optional allocation flags are ignored.
2912 * "npages" must be greater than zero. Both "alignment" and "boundary"
2913 * must be a power of two.
2916 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
2917 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2919 struct vm_domain *vmd;
2920 vm_paddr_t curr_low;
2921 vm_page_t m_run, m_runs[NRUNS];
2922 u_long count, reclaimed;
2923 int error, i, options, req_class;
2925 KASSERT(npages > 0, ("npages is 0"));
2926 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2927 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2928 req_class = req & VM_ALLOC_CLASS_MASK;
2931 * The page daemon is allowed to dig deeper into the free page list.
2933 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2934 req_class = VM_ALLOC_SYSTEM;
2937 * Return if the number of free pages cannot satisfy the requested
2940 vmd = VM_DOMAIN(domain);
2941 count = vmd->vmd_free_count;
2942 if (count < npages + vmd->vmd_free_reserved || (count < npages +
2943 vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2944 (count < npages && req_class == VM_ALLOC_INTERRUPT))
2948 * Scan up to three times, relaxing the restrictions ("options") on
2949 * the reclamation of reservations and superpages each time.
2951 for (options = VPSC_NORESERV;;) {
2953 * Find the highest runs that satisfy the given constraints
2954 * and restrictions, and record them in "m_runs".
2959 m_run = vm_phys_scan_contig(domain, npages, curr_low,
2960 high, alignment, boundary, options);
2963 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2964 m_runs[RUN_INDEX(count)] = m_run;
2969 * Reclaim the highest runs in LIFO (descending) order until
2970 * the number of reclaimed pages, "reclaimed", is at least
2971 * MIN_RECLAIM. Reset "reclaimed" each time because each
2972 * reclamation is idempotent, and runs will (likely) recur
2973 * from one scan to the next as restrictions are relaxed.
2976 for (i = 0; count > 0 && i < NRUNS; i++) {
2978 m_run = m_runs[RUN_INDEX(count)];
2979 error = vm_page_reclaim_run(req_class, domain, npages,
2982 reclaimed += npages;
2983 if (reclaimed >= MIN_RECLAIM)
2989 * Either relax the restrictions on the next scan or return if
2990 * the last scan had no restrictions.
2992 if (options == VPSC_NORESERV)
2993 options = VPSC_NOSUPER;
2994 else if (options == VPSC_NOSUPER)
2996 else if (options == VPSC_ANY)
2997 return (reclaimed != 0);
3002 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
3003 u_long alignment, vm_paddr_t boundary)
3005 struct vm_domainset_iter di;
3009 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
3011 ret = vm_page_reclaim_contig_domain(domain, req, npages, low,
3012 high, alignment, boundary);
3015 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
3021 * Set the domain in the appropriate page level domainset.
3024 vm_domain_set(struct vm_domain *vmd)
3027 mtx_lock(&vm_domainset_lock);
3028 if (!vmd->vmd_minset && vm_paging_min(vmd)) {
3029 vmd->vmd_minset = 1;
3030 DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
3032 if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
3033 vmd->vmd_severeset = 1;
3034 DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
3036 mtx_unlock(&vm_domainset_lock);
3040 * Clear the domain from the appropriate page level domainset.
3043 vm_domain_clear(struct vm_domain *vmd)
3046 mtx_lock(&vm_domainset_lock);
3047 if (vmd->vmd_minset && !vm_paging_min(vmd)) {
3048 vmd->vmd_minset = 0;
3049 DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
3050 if (vm_min_waiters != 0) {
3052 wakeup(&vm_min_domains);
3055 if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
3056 vmd->vmd_severeset = 0;
3057 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
3058 if (vm_severe_waiters != 0) {
3059 vm_severe_waiters = 0;
3060 wakeup(&vm_severe_domains);
3065 * If pageout daemon needs pages, then tell it that there are
3068 if (vmd->vmd_pageout_pages_needed &&
3069 vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
3070 wakeup(&vmd->vmd_pageout_pages_needed);
3071 vmd->vmd_pageout_pages_needed = 0;
3074 /* See comments in vm_wait_doms(). */
3075 if (vm_pageproc_waiters) {
3076 vm_pageproc_waiters = 0;
3077 wakeup(&vm_pageproc_waiters);
3079 mtx_unlock(&vm_domainset_lock);
3083 * Wait for free pages to exceed the min threshold globally.
3089 mtx_lock(&vm_domainset_lock);
3090 while (vm_page_count_min()) {
3092 msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
3094 mtx_unlock(&vm_domainset_lock);
3098 * Wait for free pages to exceed the severe threshold globally.
3101 vm_wait_severe(void)
3104 mtx_lock(&vm_domainset_lock);
3105 while (vm_page_count_severe()) {
3106 vm_severe_waiters++;
3107 msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
3110 mtx_unlock(&vm_domainset_lock);
3117 return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
3121 vm_wait_doms(const domainset_t *wdoms)
3125 * We use racey wakeup synchronization to avoid expensive global
3126 * locking for the pageproc when sleeping with a non-specific vm_wait.
3127 * To handle this, we only sleep for one tick in this instance. It
3128 * is expected that most allocations for the pageproc will come from
3129 * kmem or vm_page_grab* which will use the more specific and
3130 * race-free vm_wait_domain().
3132 if (curproc == pageproc) {
3133 mtx_lock(&vm_domainset_lock);
3134 vm_pageproc_waiters++;
3135 msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP,
3139 * XXX Ideally we would wait only until the allocation could
3140 * be satisfied. This condition can cause new allocators to
3141 * consume all freed pages while old allocators wait.
3143 mtx_lock(&vm_domainset_lock);
3144 if (vm_page_count_min_set(wdoms)) {
3146 msleep(&vm_min_domains, &vm_domainset_lock,
3147 PVM | PDROP, "vmwait", 0);
3149 mtx_unlock(&vm_domainset_lock);
3156 * Sleep until free pages are available for allocation.
3157 * - Called in various places after failed memory allocations.
3160 vm_wait_domain(int domain)
3162 struct vm_domain *vmd;
3165 vmd = VM_DOMAIN(domain);
3166 vm_domain_free_assert_unlocked(vmd);
3168 if (curproc == pageproc) {
3169 mtx_lock(&vm_domainset_lock);
3170 if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3171 vmd->vmd_pageout_pages_needed = 1;
3172 msleep(&vmd->vmd_pageout_pages_needed,
3173 &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3175 mtx_unlock(&vm_domainset_lock);
3177 if (pageproc == NULL)
3178 panic("vm_wait in early boot");
3179 DOMAINSET_ZERO(&wdom);
3180 DOMAINSET_SET(vmd->vmd_domain, &wdom);
3181 vm_wait_doms(&wdom);
3188 * Sleep until free pages are available for allocation in the
3189 * affinity domains of the obj. If obj is NULL, the domain set
3190 * for the calling thread is used.
3191 * Called in various places after failed memory allocations.
3194 vm_wait(vm_object_t obj)
3196 struct domainset *d;
3201 * Carefully fetch pointers only once: the struct domainset
3202 * itself is ummutable but the pointer might change.
3205 d = obj->domain.dr_policy;
3207 d = curthread->td_domain.dr_policy;
3209 vm_wait_doms(&d->ds_mask);
3213 * vm_domain_alloc_fail:
3215 * Called when a page allocation function fails. Informs the
3216 * pagedaemon and performs the requested wait. Requires the
3217 * domain_free and object lock on entry. Returns with the
3218 * object lock held and free lock released. Returns an error when
3219 * retry is necessary.
3223 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3226 vm_domain_free_assert_unlocked(vmd);
3228 atomic_add_int(&vmd->vmd_pageout_deficit,
3229 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3230 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3232 VM_OBJECT_WUNLOCK(object);
3233 vm_wait_domain(vmd->vmd_domain);
3235 VM_OBJECT_WLOCK(object);
3236 if (req & VM_ALLOC_WAITOK)
3246 * Sleep until free pages are available for allocation.
3247 * - Called only in vm_fault so that processes page faulting
3248 * can be easily tracked.
3249 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3250 * processes will be able to grab memory first. Do not change
3251 * this balance without careful testing first.
3254 vm_waitpfault(struct domainset *dset, int timo)
3258 * XXX Ideally we would wait only until the allocation could
3259 * be satisfied. This condition can cause new allocators to
3260 * consume all freed pages while old allocators wait.
3262 mtx_lock(&vm_domainset_lock);
3263 if (vm_page_count_min_set(&dset->ds_mask)) {
3265 msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3268 mtx_unlock(&vm_domainset_lock);
3271 static struct vm_pagequeue *
3272 _vm_page_pagequeue(vm_page_t m, uint8_t queue)
3275 return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3279 static struct vm_pagequeue *
3280 vm_page_pagequeue(vm_page_t m)
3283 return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue));
3287 static __always_inline bool
3288 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3290 vm_page_astate_t tmp;
3294 if (__predict_true(vm_page_astate_fcmpset(m, old, new)))
3296 counter_u64_add(pqstate_commit_retries, 1);
3297 } while (old->_bits == tmp._bits);
3303 * Do the work of committing a queue state update that moves the page out of
3304 * its current queue.
3307 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m,
3308 vm_page_astate_t *old, vm_page_astate_t new)
3312 vm_pagequeue_assert_locked(pq);
3313 KASSERT(vm_page_pagequeue(m) == pq,
3314 ("%s: queue %p does not match page %p", __func__, pq, m));
3315 KASSERT(old->queue != PQ_NONE && new.queue != old->queue,
3316 ("%s: invalid queue indices %d %d",
3317 __func__, old->queue, new.queue));
3320 * Once the queue index of the page changes there is nothing
3321 * synchronizing with further updates to the page's physical
3322 * queue state. Therefore we must speculatively remove the page
3323 * from the queue now and be prepared to roll back if the queue
3324 * state update fails. If the page is not physically enqueued then
3325 * we just update its queue index.
3327 if ((old->flags & PGA_ENQUEUED) != 0) {
3328 new.flags &= ~PGA_ENQUEUED;
3329 next = TAILQ_NEXT(m, plinks.q);
3330 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3331 vm_pagequeue_cnt_dec(pq);
3332 if (!vm_page_pqstate_fcmpset(m, old, new)) {
3334 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3336 TAILQ_INSERT_BEFORE(next, m, plinks.q);
3337 vm_pagequeue_cnt_inc(pq);
3343 return (vm_page_pqstate_fcmpset(m, old, new));
3348 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old,
3349 vm_page_astate_t new)
3351 struct vm_pagequeue *pq;
3352 vm_page_astate_t as;
3355 pq = _vm_page_pagequeue(m, old->queue);
3358 * The queue field and PGA_ENQUEUED flag are stable only so long as the
3359 * corresponding page queue lock is held.
3361 vm_pagequeue_lock(pq);
3362 as = vm_page_astate_load(m);
3363 if (__predict_false(as._bits != old->_bits)) {
3367 ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new);
3369 vm_pagequeue_unlock(pq);
3374 * Commit a queue state update that enqueues or requeues a page.
3377 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m,
3378 vm_page_astate_t *old, vm_page_astate_t new)
3380 struct vm_domain *vmd;
3382 vm_pagequeue_assert_locked(pq);
3383 KASSERT(old->queue != PQ_NONE && new.queue == old->queue,
3384 ("%s: invalid queue indices %d %d",
3385 __func__, old->queue, new.queue));
3387 new.flags |= PGA_ENQUEUED;
3388 if (!vm_page_pqstate_fcmpset(m, old, new))
3391 if ((old->flags & PGA_ENQUEUED) != 0)
3392 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3394 vm_pagequeue_cnt_inc(pq);
3397 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE. In particular, if
3398 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be
3399 * applied, even if it was set first.
3401 if ((old->flags & PGA_REQUEUE_HEAD) != 0) {
3402 vmd = vm_pagequeue_domain(m);
3403 KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE],
3404 ("%s: invalid page queue for page %p", __func__, m));
3405 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3407 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3413 * Commit a queue state update that encodes a request for a deferred queue
3417 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old,
3418 vm_page_astate_t new)
3421 KASSERT(old->queue == new.queue || new.queue != PQ_NONE,
3422 ("%s: invalid state, queue %d flags %x",
3423 __func__, new.queue, new.flags));
3425 if (old->_bits != new._bits &&
3426 !vm_page_pqstate_fcmpset(m, old, new))
3428 vm_page_pqbatch_submit(m, new.queue);
3433 * A generic queue state update function. This handles more cases than the
3434 * specialized functions above.
3437 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3440 if (old->_bits == new._bits)
3443 if (old->queue != PQ_NONE && new.queue != old->queue) {
3444 if (!vm_page_pqstate_commit_dequeue(m, old, new))
3446 if (new.queue != PQ_NONE)
3447 vm_page_pqbatch_submit(m, new.queue);
3449 if (!vm_page_pqstate_fcmpset(m, old, new))
3451 if (new.queue != PQ_NONE &&
3452 ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0)
3453 vm_page_pqbatch_submit(m, new.queue);
3459 * Apply deferred queue state updates to a page.
3462 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue)
3464 vm_page_astate_t new, old;
3466 CRITICAL_ASSERT(curthread);
3467 vm_pagequeue_assert_locked(pq);
3468 KASSERT(queue < PQ_COUNT,
3469 ("%s: invalid queue index %d", __func__, queue));
3470 KASSERT(pq == _vm_page_pagequeue(m, queue),
3471 ("%s: page %p does not belong to queue %p", __func__, m, pq));
3473 for (old = vm_page_astate_load(m);;) {
3474 if (__predict_false(old.queue != queue ||
3475 (old.flags & PGA_QUEUE_OP_MASK) == 0)) {
3476 counter_u64_add(queue_nops, 1);
3479 KASSERT(old.queue != PQ_NONE || (old.flags & PGA_QUEUE_STATE_MASK) == 0,
3480 ("%s: page %p has unexpected queue state", __func__, m));
3483 if ((old.flags & PGA_DEQUEUE) != 0) {
3484 new.flags &= ~PGA_QUEUE_OP_MASK;
3485 new.queue = PQ_NONE;
3486 if (__predict_true(_vm_page_pqstate_commit_dequeue(pq,
3488 counter_u64_add(queue_ops, 1);
3492 new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD);
3493 if (__predict_true(_vm_page_pqstate_commit_requeue(pq,
3495 counter_u64_add(queue_ops, 1);
3503 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3508 for (i = 0; i < bq->bq_cnt; i++)
3509 vm_pqbatch_process_page(pq, bq->bq_pa[i], queue);
3510 vm_batchqueue_init(bq);
3514 * vm_page_pqbatch_submit: [ internal use only ]
3516 * Enqueue a page in the specified page queue's batched work queue.
3517 * The caller must have encoded the requested operation in the page
3518 * structure's a.flags field.
3521 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
3523 struct vm_batchqueue *bq;
3524 struct vm_pagequeue *pq;
3527 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3528 ("page %p is unmanaged", m));
3529 KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3531 domain = vm_phys_domain(m);
3532 pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue];
3535 bq = DPCPU_PTR(pqbatch[domain][queue]);
3536 if (vm_batchqueue_insert(bq, m)) {
3541 vm_pagequeue_lock(pq);
3543 bq = DPCPU_PTR(pqbatch[domain][queue]);
3544 vm_pqbatch_process(pq, bq, queue);
3545 vm_pqbatch_process_page(pq, m, queue);
3546 vm_pagequeue_unlock(pq);
3551 * vm_page_pqbatch_drain: [ internal use only ]
3553 * Force all per-CPU page queue batch queues to be drained. This is
3554 * intended for use in severe memory shortages, to ensure that pages
3555 * do not remain stuck in the batch queues.
3558 vm_page_pqbatch_drain(void)
3561 struct vm_domain *vmd;
3562 struct vm_pagequeue *pq;
3563 int cpu, domain, queue;
3568 sched_bind(td, cpu);
3571 for (domain = 0; domain < vm_ndomains; domain++) {
3572 vmd = VM_DOMAIN(domain);
3573 for (queue = 0; queue < PQ_COUNT; queue++) {
3574 pq = &vmd->vmd_pagequeues[queue];
3575 vm_pagequeue_lock(pq);
3577 vm_pqbatch_process(pq,
3578 DPCPU_PTR(pqbatch[domain][queue]), queue);
3580 vm_pagequeue_unlock(pq);
3590 * vm_page_dequeue_deferred: [ internal use only ]
3592 * Request removal of the given page from its current page
3593 * queue. Physical removal from the queue may be deferred
3596 * The page must be locked.
3599 vm_page_dequeue_deferred(vm_page_t m)
3601 vm_page_astate_t new, old;
3603 old = vm_page_astate_load(m);
3605 if (old.queue == PQ_NONE) {
3606 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3607 ("%s: page %p has unexpected queue state",
3612 new.flags |= PGA_DEQUEUE;
3613 } while (!vm_page_pqstate_commit_request(m, &old, new));
3619 * Remove the page from whichever page queue it's in, if any, before
3623 vm_page_dequeue(vm_page_t m)
3625 vm_page_astate_t new, old;
3627 old = vm_page_astate_load(m);
3629 if (old.queue == PQ_NONE) {
3630 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3631 ("%s: page %p has unexpected queue state",
3636 new.flags &= ~PGA_QUEUE_OP_MASK;
3637 new.queue = PQ_NONE;
3638 } while (!vm_page_pqstate_commit_dequeue(m, &old, new));
3643 * Schedule the given page for insertion into the specified page queue.
3644 * Physical insertion of the page may be deferred indefinitely.
3647 vm_page_enqueue(vm_page_t m, uint8_t queue)
3650 KASSERT(m->a.queue == PQ_NONE &&
3651 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
3652 ("%s: page %p is already enqueued", __func__, m));
3653 KASSERT(m->ref_count > 0,
3654 ("%s: page %p does not carry any references", __func__, m));
3657 if ((m->a.flags & PGA_REQUEUE) == 0)
3658 vm_page_aflag_set(m, PGA_REQUEUE);
3659 vm_page_pqbatch_submit(m, queue);
3663 * vm_page_free_prep:
3665 * Prepares the given page to be put on the free list,
3666 * disassociating it from any VM object. The caller may return
3667 * the page to the free list only if this function returns true.
3669 * The object, if it exists, must be locked, and then the page must
3670 * be xbusy. Otherwise the page must be not busied. A managed
3671 * page must be unmapped.
3674 vm_page_free_prep(vm_page_t m)
3678 * Synchronize with threads that have dropped a reference to this
3681 atomic_thread_fence_acq();
3683 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3684 if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3687 p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3688 for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3689 KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3690 m, i, (uintmax_t)*p));
3693 if ((m->oflags & VPO_UNMANAGED) == 0) {
3694 KASSERT(!pmap_page_is_mapped(m),
3695 ("vm_page_free_prep: freeing mapped page %p", m));
3696 KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
3697 ("vm_page_free_prep: mapping flags set in page %p", m));
3699 KASSERT(m->a.queue == PQ_NONE,
3700 ("vm_page_free_prep: unmanaged page %p is queued", m));
3702 VM_CNT_INC(v_tfree);
3704 if (m->object != NULL) {
3705 KASSERT(((m->oflags & VPO_UNMANAGED) != 0) ==
3706 ((m->object->flags & OBJ_UNMANAGED) != 0),
3707 ("vm_page_free_prep: managed flag mismatch for page %p",
3709 vm_page_assert_xbusied(m);
3712 * The object reference can be released without an atomic
3715 KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
3716 m->ref_count == VPRC_OBJREF,
3717 ("vm_page_free_prep: page %p has unexpected ref_count %u",
3719 vm_page_object_remove(m);
3720 m->ref_count -= VPRC_OBJREF;
3722 vm_page_assert_unbusied(m);
3724 vm_page_busy_free(m);
3727 * If fictitious remove object association and
3730 if ((m->flags & PG_FICTITIOUS) != 0) {
3731 KASSERT(m->ref_count == 1,
3732 ("fictitious page %p is referenced", m));
3733 KASSERT(m->a.queue == PQ_NONE,
3734 ("fictitious page %p is queued", m));
3739 * Pages need not be dequeued before they are returned to the physical
3740 * memory allocator, but they must at least be marked for a deferred
3743 if ((m->oflags & VPO_UNMANAGED) == 0)
3744 vm_page_dequeue_deferred(m);
3749 if (m->ref_count != 0)
3750 panic("vm_page_free_prep: page %p has references", m);
3753 * Restore the default memory attribute to the page.
3755 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3756 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3758 #if VM_NRESERVLEVEL > 0
3760 * Determine whether the page belongs to a reservation. If the page was
3761 * allocated from a per-CPU cache, it cannot belong to a reservation, so
3762 * as an optimization, we avoid the check in that case.
3764 if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
3774 * Returns the given page to the free list, disassociating it
3775 * from any VM object.
3777 * The object must be locked. The page must be locked if it is
3781 vm_page_free_toq(vm_page_t m)
3783 struct vm_domain *vmd;
3786 if (!vm_page_free_prep(m))
3789 vmd = vm_pagequeue_domain(m);
3790 zone = vmd->vmd_pgcache[m->pool].zone;
3791 if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
3795 vm_domain_free_lock(vmd);
3796 vm_phys_free_pages(m, 0);
3797 vm_domain_free_unlock(vmd);
3798 vm_domain_freecnt_inc(vmd, 1);
3802 * vm_page_free_pages_toq:
3804 * Returns a list of pages to the free list, disassociating it
3805 * from any VM object. In other words, this is equivalent to
3806 * calling vm_page_free_toq() for each page of a list of VM objects.
3808 * The objects must be locked. The pages must be locked if it is
3812 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
3817 if (SLIST_EMPTY(free))
3821 while ((m = SLIST_FIRST(free)) != NULL) {
3823 SLIST_REMOVE_HEAD(free, plinks.s.ss);
3824 vm_page_free_toq(m);
3827 if (update_wire_count)
3832 * Mark this page as wired down, preventing reclamation by the page daemon
3833 * or when the containing object is destroyed.
3836 vm_page_wire(vm_page_t m)
3840 KASSERT(m->object != NULL,
3841 ("vm_page_wire: page %p does not belong to an object", m));
3842 if (!vm_page_busied(m) && !vm_object_busied(m->object))
3843 VM_OBJECT_ASSERT_LOCKED(m->object);
3844 KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
3845 VPRC_WIRE_COUNT(m->ref_count) >= 1,
3846 ("vm_page_wire: fictitious page %p has zero wirings", m));
3848 old = atomic_fetchadd_int(&m->ref_count, 1);
3849 KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
3850 ("vm_page_wire: counter overflow for page %p", m));
3851 if (VPRC_WIRE_COUNT(old) == 0) {
3852 if ((m->oflags & VPO_UNMANAGED) == 0)
3853 vm_page_aflag_set(m, PGA_DEQUEUE);
3859 * Attempt to wire a mapped page following a pmap lookup of that page.
3860 * This may fail if a thread is concurrently tearing down mappings of the page.
3861 * The transient failure is acceptable because it translates to the
3862 * failure of the caller pmap_extract_and_hold(), which should be then
3863 * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages().
3866 vm_page_wire_mapped(vm_page_t m)
3873 ("vm_page_wire_mapped: wiring unreferenced page %p", m));
3874 if ((old & VPRC_BLOCKED) != 0)
3876 } while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
3878 if (VPRC_WIRE_COUNT(old) == 0) {
3879 if ((m->oflags & VPO_UNMANAGED) == 0)
3880 vm_page_aflag_set(m, PGA_DEQUEUE);
3887 * Release a wiring reference to a managed page. If the page still belongs to
3888 * an object, update its position in the page queues to reflect the reference.
3889 * If the wiring was the last reference to the page, free the page.
3892 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse)
3896 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3897 ("%s: page %p is unmanaged", __func__, m));
3900 * Update LRU state before releasing the wiring reference.
3901 * Use a release store when updating the reference count to
3902 * synchronize with vm_page_free_prep().
3906 KASSERT(VPRC_WIRE_COUNT(old) > 0,
3907 ("vm_page_unwire: wire count underflow for page %p", m));
3909 if (old > VPRC_OBJREF + 1) {
3911 * The page has at least one other wiring reference. An
3912 * earlier iteration of this loop may have called
3913 * vm_page_release_toq() and cleared PGA_DEQUEUE, so
3914 * re-set it if necessary.
3916 if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0)
3917 vm_page_aflag_set(m, PGA_DEQUEUE);
3918 } else if (old == VPRC_OBJREF + 1) {
3920 * This is the last wiring. Clear PGA_DEQUEUE and
3921 * update the page's queue state to reflect the
3922 * reference. If the page does not belong to an object
3923 * (i.e., the VPRC_OBJREF bit is clear), we only need to
3924 * clear leftover queue state.
3926 vm_page_release_toq(m, nqueue, false);
3927 } else if (old == 1) {
3928 vm_page_aflag_clear(m, PGA_DEQUEUE);
3930 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
3932 if (VPRC_WIRE_COUNT(old) == 1) {
3940 * Release one wiring of the specified page, potentially allowing it to be
3943 * Only managed pages belonging to an object can be paged out. If the number
3944 * of wirings transitions to zero and the page is eligible for page out, then
3945 * the page is added to the specified paging queue. If the released wiring
3946 * represented the last reference to the page, the page is freed.
3948 * A managed page must be locked.
3951 vm_page_unwire(vm_page_t m, uint8_t nqueue)
3954 KASSERT(nqueue < PQ_COUNT,
3955 ("vm_page_unwire: invalid queue %u request for page %p",
3958 if ((m->oflags & VPO_UNMANAGED) != 0) {
3959 if (vm_page_unwire_noq(m) && m->ref_count == 0)
3963 vm_page_unwire_managed(m, nqueue, false);
3967 * Unwire a page without (re-)inserting it into a page queue. It is up
3968 * to the caller to enqueue, requeue, or free the page as appropriate.
3969 * In most cases involving managed pages, vm_page_unwire() should be used
3973 vm_page_unwire_noq(vm_page_t m)
3977 old = vm_page_drop(m, 1);
3978 KASSERT(VPRC_WIRE_COUNT(old) != 0,
3979 ("vm_page_unref: counter underflow for page %p", m));
3980 KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
3981 ("vm_page_unref: missing ref on fictitious page %p", m));
3983 if (VPRC_WIRE_COUNT(old) > 1)
3985 if ((m->oflags & VPO_UNMANAGED) == 0)
3986 vm_page_aflag_clear(m, PGA_DEQUEUE);
3992 * Ensure that the page ends up in the specified page queue. If the page is
3993 * active or being moved to the active queue, ensure that its act_count is
3994 * at least ACT_INIT but do not otherwise mess with it.
3996 * A managed page must be locked.
3998 static __always_inline void
3999 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag)
4001 vm_page_astate_t old, new;
4003 KASSERT(m->ref_count > 0,
4004 ("%s: page %p does not carry any references", __func__, m));
4005 KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD,
4006 ("%s: invalid flags %x", __func__, nflag));
4008 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
4011 old = vm_page_astate_load(m);
4013 if ((old.flags & PGA_DEQUEUE) != 0)
4016 new.flags &= ~PGA_QUEUE_OP_MASK;
4017 if (nqueue == PQ_ACTIVE)
4018 new.act_count = max(old.act_count, ACT_INIT);
4019 if (old.queue == nqueue) {
4020 if (nqueue != PQ_ACTIVE)
4026 } while (!vm_page_pqstate_commit(m, &old, new));
4030 * Put the specified page on the active list (if appropriate).
4033 vm_page_activate(vm_page_t m)
4036 vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE);
4040 * Move the specified page to the tail of the inactive queue, or requeue
4041 * the page if it is already in the inactive queue.
4044 vm_page_deactivate(vm_page_t m)
4047 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE);
4051 vm_page_deactivate_noreuse(vm_page_t m)
4054 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD);
4058 * Put a page in the laundry, or requeue it if it is already there.
4061 vm_page_launder(vm_page_t m)
4064 vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE);
4068 * Put a page in the PQ_UNSWAPPABLE holding queue.
4071 vm_page_unswappable(vm_page_t m)
4074 KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0,
4075 ("page %p already unswappable", m));
4078 vm_page_enqueue(m, PQ_UNSWAPPABLE);
4082 * Release a page back to the page queues in preparation for unwiring.
4085 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse)
4087 vm_page_astate_t old, new;
4091 * Use a check of the valid bits to determine whether we should
4092 * accelerate reclamation of the page. The object lock might not be
4093 * held here, in which case the check is racy. At worst we will either
4094 * accelerate reclamation of a valid page and violate LRU, or
4095 * unnecessarily defer reclamation of an invalid page.
4097 * If we were asked to not cache the page, place it near the head of the
4098 * inactive queue so that is reclaimed sooner.
4100 if (noreuse || m->valid == 0) {
4101 nqueue = PQ_INACTIVE;
4102 nflag = PGA_REQUEUE_HEAD;
4104 nflag = PGA_REQUEUE;
4107 old = vm_page_astate_load(m);
4112 * If the page is already in the active queue and we are not
4113 * trying to accelerate reclamation, simply mark it as
4114 * referenced and avoid any queue operations.
4116 new.flags &= ~PGA_QUEUE_OP_MASK;
4117 if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE)
4118 new.flags |= PGA_REFERENCED;
4123 } while (!vm_page_pqstate_commit(m, &old, new));
4127 * Unwire a page and either attempt to free it or re-add it to the page queues.
4130 vm_page_release(vm_page_t m, int flags)
4134 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4135 ("vm_page_release: page %p is unmanaged", m));
4137 if ((flags & VPR_TRYFREE) != 0) {
4139 object = atomic_load_ptr(&m->object);
4142 /* Depends on type-stability. */
4143 if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object))
4145 if (object == m->object) {
4146 vm_page_release_locked(m, flags);
4147 VM_OBJECT_WUNLOCK(object);
4150 VM_OBJECT_WUNLOCK(object);
4153 vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0);
4156 /* See vm_page_release(). */
4158 vm_page_release_locked(vm_page_t m, int flags)
4161 VM_OBJECT_ASSERT_WLOCKED(m->object);
4162 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4163 ("vm_page_release_locked: page %p is unmanaged", m));
4165 if (vm_page_unwire_noq(m)) {
4166 if ((flags & VPR_TRYFREE) != 0 &&
4167 (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
4168 m->dirty == 0 && vm_page_tryxbusy(m)) {
4170 * An unlocked lookup may have wired the page before the
4171 * busy lock was acquired, in which case the page must
4174 if (__predict_true(!vm_page_wired(m))) {
4180 vm_page_release_toq(m, PQ_INACTIVE, flags != 0);
4186 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
4190 KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
4191 ("vm_page_try_blocked_op: page %p has no object", m));
4192 KASSERT(vm_page_busied(m),
4193 ("vm_page_try_blocked_op: page %p is not busy", m));
4194 VM_OBJECT_ASSERT_LOCKED(m->object);
4199 ("vm_page_try_blocked_op: page %p has no references", m));
4200 if (VPRC_WIRE_COUNT(old) != 0)
4202 } while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
4207 * If the object is read-locked, new wirings may be created via an
4210 old = vm_page_drop(m, VPRC_BLOCKED);
4211 KASSERT(!VM_OBJECT_WOWNED(m->object) ||
4212 old == (VPRC_BLOCKED | VPRC_OBJREF),
4213 ("vm_page_try_blocked_op: unexpected refcount value %u for %p",
4219 * Atomically check for wirings and remove all mappings of the page.
4222 vm_page_try_remove_all(vm_page_t m)
4225 return (vm_page_try_blocked_op(m, pmap_remove_all));
4229 * Atomically check for wirings and remove all writeable mappings of the page.
4232 vm_page_try_remove_write(vm_page_t m)
4235 return (vm_page_try_blocked_op(m, pmap_remove_write));
4241 * Apply the specified advice to the given page.
4243 * The object and page must be locked.
4246 vm_page_advise(vm_page_t m, int advice)
4249 VM_OBJECT_ASSERT_WLOCKED(m->object);
4250 if (advice == MADV_FREE)
4252 * Mark the page clean. This will allow the page to be freed
4253 * without first paging it out. MADV_FREE pages are often
4254 * quickly reused by malloc(3), so we do not do anything that
4255 * would result in a page fault on a later access.
4258 else if (advice != MADV_DONTNEED) {
4259 if (advice == MADV_WILLNEED)
4260 vm_page_activate(m);
4264 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
4268 * Clear any references to the page. Otherwise, the page daemon will
4269 * immediately reactivate the page.
4271 vm_page_aflag_clear(m, PGA_REFERENCED);
4274 * Place clean pages near the head of the inactive queue rather than
4275 * the tail, thus defeating the queue's LRU operation and ensuring that
4276 * the page will be reused quickly. Dirty pages not already in the
4277 * laundry are moved there.
4280 vm_page_deactivate_noreuse(m);
4281 else if (!vm_page_in_laundry(m))
4286 * vm_page_grab_release
4288 * Helper routine for grab functions to release busy on return.
4291 vm_page_grab_release(vm_page_t m, int allocflags)
4294 if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4295 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4303 * vm_page_grab_sleep
4305 * Sleep for busy according to VM_ALLOC_ parameters. Returns true
4306 * if the caller should retry and false otherwise.
4308 * If the object is locked on entry the object will be unlocked with
4309 * false returns and still locked but possibly having been dropped
4310 * with true returns.
4313 vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex,
4314 const char *wmesg, int allocflags, bool locked)
4317 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4321 * Reference the page before unlocking and sleeping so that
4322 * the page daemon is less likely to reclaim it.
4324 if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0)
4325 vm_page_reference(m);
4327 if (_vm_page_busy_sleep(object, m, m->pindex, wmesg, allocflags,
4329 VM_OBJECT_WLOCK(object);
4330 if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
4337 * Assert that the grab flags are valid.
4340 vm_page_grab_check(int allocflags)
4343 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4344 (allocflags & VM_ALLOC_WIRED) != 0,
4345 ("vm_page_grab*: the pages must be busied or wired"));
4347 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4348 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4349 ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4353 * Calculate the page allocation flags for grab.
4356 vm_page_grab_pflags(int allocflags)
4360 pflags = allocflags &
4361 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL |
4363 if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4364 pflags |= VM_ALLOC_WAITFAIL;
4365 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4366 pflags |= VM_ALLOC_SBUSY;
4372 * Grab a page, waiting until we are waken up due to the page
4373 * changing state. We keep on waiting, if the page continues
4374 * to be in the object. If the page doesn't exist, first allocate it
4375 * and then conditionally zero it.
4377 * This routine may sleep.
4379 * The object must be locked on entry. The lock will, however, be released
4380 * and reacquired if the routine sleeps.
4383 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
4387 VM_OBJECT_ASSERT_WLOCKED(object);
4388 vm_page_grab_check(allocflags);
4391 if ((m = vm_page_lookup(object, pindex)) != NULL) {
4392 if (!vm_page_tryacquire(m, allocflags)) {
4393 if (vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4400 if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4402 m = vm_page_alloc(object, pindex, vm_page_grab_pflags(allocflags));
4404 if ((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0)
4408 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
4412 vm_page_grab_release(m, allocflags);
4418 * Locklessly attempt to acquire a page given a (object, pindex) tuple
4419 * and an optional previous page to avoid the radix lookup. The resulting
4420 * page will be validated against the identity tuple and busied or wired
4421 * as requested. A NULL *mp return guarantees that the page was not in
4422 * radix at the time of the call but callers must perform higher level
4423 * synchronization or retry the operation under a lock if they require
4424 * an atomic answer. This is the only lock free validation routine,
4425 * other routines can depend on the resulting page state.
4427 * The return value indicates whether the operation failed due to caller
4428 * flags. The return is tri-state with mp:
4430 * (true, *mp != NULL) - The operation was successful.
4431 * (true, *mp == NULL) - The page was not found in tree.
4432 * (false, *mp == NULL) - WAITFAIL or NOWAIT prevented acquisition.
4435 vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex,
4436 vm_page_t prev, vm_page_t *mp, int allocflags)
4440 vm_page_grab_check(allocflags);
4441 MPASS(prev == NULL || vm_page_busied(prev) || vm_page_wired(prev));
4446 * We may see a false NULL here because the previous page
4447 * has been removed or just inserted and the list is loaded
4448 * without barriers. Switch to radix to verify.
4450 if (prev == NULL || (m = TAILQ_NEXT(prev, listq)) == NULL ||
4451 QMD_IS_TRASHED(m) || m->pindex != pindex ||
4452 atomic_load_ptr(&m->object) != object) {
4455 * This guarantees the result is instantaneously
4458 m = vm_radix_lookup_unlocked(&object->rtree, pindex);
4462 if (vm_page_trybusy(m, allocflags)) {
4463 if (m->object == object && m->pindex == pindex)
4466 vm_page_busy_release(m);
4470 if (!vm_page_grab_sleep(object, m, pindex, "pgnslp",
4474 if ((allocflags & VM_ALLOC_WIRED) != 0)
4476 vm_page_grab_release(m, allocflags);
4482 * Try to locklessly grab a page and fall back to the object lock if NOCREAT
4486 vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags)
4490 vm_page_grab_check(allocflags);
4492 if (!vm_page_acquire_unlocked(object, pindex, NULL, &m, allocflags))
4498 * The radix lockless lookup should never return a false negative
4499 * errors. If the user specifies NOCREAT they are guaranteed there
4500 * was no page present at the instant of the call. A NOCREAT caller
4501 * must handle create races gracefully.
4503 if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4506 VM_OBJECT_WLOCK(object);
4507 m = vm_page_grab(object, pindex, allocflags);
4508 VM_OBJECT_WUNLOCK(object);
4514 * Grab a page and make it valid, paging in if necessary. Pages missing from
4515 * their pager are zero filled and validated. If a VM_ALLOC_COUNT is supplied
4516 * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought
4517 * in simultaneously. Additional pages will be left on a paging queue but
4518 * will neither be wired nor busy regardless of allocflags.
4521 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags)
4524 vm_page_t ma[VM_INITIAL_PAGEIN];
4525 int after, i, pflags, rv;
4527 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4528 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4529 ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4530 KASSERT((allocflags &
4531 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4532 ("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
4533 VM_OBJECT_ASSERT_WLOCKED(object);
4534 pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY |
4536 pflags |= VM_ALLOC_WAITFAIL;
4539 if ((m = vm_page_lookup(object, pindex)) != NULL) {
4541 * If the page is fully valid it can only become invalid
4542 * with the object lock held. If it is not valid it can
4543 * become valid with the busy lock held. Therefore, we
4544 * may unnecessarily lock the exclusive busy here if we
4545 * race with I/O completion not using the object lock.
4546 * However, we will not end up with an invalid page and a
4549 if (!vm_page_trybusy(m,
4550 vm_page_all_valid(m) ? allocflags : 0)) {
4551 (void)vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4555 if (vm_page_all_valid(m))
4557 if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4558 vm_page_busy_release(m);
4560 return (VM_PAGER_FAIL);
4562 } else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4564 return (VM_PAGER_FAIL);
4565 } else if ((m = vm_page_alloc(object, pindex, pflags)) == NULL) {
4569 vm_page_assert_xbusied(m);
4570 if (vm_pager_has_page(object, pindex, NULL, &after)) {
4571 after = MIN(after, VM_INITIAL_PAGEIN);
4572 after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT);
4573 after = MAX(after, 1);
4575 for (i = 1; i < after; i++) {
4576 if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) {
4577 if (ma[i]->valid || !vm_page_tryxbusy(ma[i]))
4580 ma[i] = vm_page_alloc(object, m->pindex + i,
4587 vm_object_pip_add(object, after);
4588 VM_OBJECT_WUNLOCK(object);
4589 rv = vm_pager_get_pages(object, ma, after, NULL, NULL);
4590 VM_OBJECT_WLOCK(object);
4591 vm_object_pip_wakeupn(object, after);
4592 /* Pager may have replaced a page. */
4594 if (rv != VM_PAGER_OK) {
4595 for (i = 0; i < after; i++) {
4596 if (!vm_page_wired(ma[i]))
4597 vm_page_free(ma[i]);
4599 vm_page_xunbusy(ma[i]);
4604 for (i = 1; i < after; i++)
4605 vm_page_readahead_finish(ma[i]);
4606 MPASS(vm_page_all_valid(m));
4608 vm_page_zero_invalid(m, TRUE);
4611 if ((allocflags & VM_ALLOC_WIRED) != 0)
4613 if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m))
4614 vm_page_busy_downgrade(m);
4615 else if ((allocflags & VM_ALLOC_NOBUSY) != 0)
4616 vm_page_busy_release(m);
4618 return (VM_PAGER_OK);
4622 * Locklessly grab a valid page. If the page is not valid or not yet
4623 * allocated this will fall back to the object lock method.
4626 vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object,
4627 vm_pindex_t pindex, int allocflags)
4633 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4634 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4635 ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY "
4637 KASSERT((allocflags &
4638 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4639 ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags));
4642 * Attempt a lockless lookup and busy. We need at least an sbusy
4643 * before we can inspect the valid field and return a wired page.
4645 flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED);
4646 if (!vm_page_acquire_unlocked(object, pindex, NULL, mp, flags))
4647 return (VM_PAGER_FAIL);
4648 if ((m = *mp) != NULL) {
4649 if (vm_page_all_valid(m)) {
4650 if ((allocflags & VM_ALLOC_WIRED) != 0)
4652 vm_page_grab_release(m, allocflags);
4653 return (VM_PAGER_OK);
4655 vm_page_busy_release(m);
4657 if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4659 return (VM_PAGER_FAIL);
4661 VM_OBJECT_WLOCK(object);
4662 error = vm_page_grab_valid(mp, object, pindex, allocflags);
4663 VM_OBJECT_WUNLOCK(object);
4669 * Return the specified range of pages from the given object. For each
4670 * page offset within the range, if a page already exists within the object
4671 * at that offset and it is busy, then wait for it to change state. If,
4672 * instead, the page doesn't exist, then allocate it.
4674 * The caller must always specify an allocation class.
4676 * allocation classes:
4677 * VM_ALLOC_NORMAL normal process request
4678 * VM_ALLOC_SYSTEM system *really* needs the pages
4680 * The caller must always specify that the pages are to be busied and/or
4683 * optional allocation flags:
4684 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages
4685 * VM_ALLOC_NOBUSY do not exclusive busy the page
4686 * VM_ALLOC_NOWAIT do not sleep
4687 * VM_ALLOC_SBUSY set page to sbusy state
4688 * VM_ALLOC_WIRED wire the pages
4689 * VM_ALLOC_ZERO zero and validate any invalid pages
4691 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it
4692 * may return a partial prefix of the requested range.
4695 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
4696 vm_page_t *ma, int count)
4702 VM_OBJECT_ASSERT_WLOCKED(object);
4703 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
4704 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
4705 vm_page_grab_check(allocflags);
4707 pflags = vm_page_grab_pflags(allocflags);
4713 m = vm_radix_lookup_le(&object->rtree, pindex + i);
4714 if (m == NULL || m->pindex != pindex + i) {
4718 mpred = TAILQ_PREV(m, pglist, listq);
4719 for (; i < count; i++) {
4721 if (!vm_page_tryacquire(m, allocflags)) {
4722 if (vm_page_grab_sleep(object, m, pindex,
4723 "grbmaw", allocflags, true))
4728 if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4730 m = vm_page_alloc_after(object, pindex + i,
4731 pflags | VM_ALLOC_COUNT(count - i), mpred);
4733 if ((allocflags & (VM_ALLOC_NOWAIT |
4734 VM_ALLOC_WAITFAIL)) != 0)
4739 if (vm_page_none_valid(m) &&
4740 (allocflags & VM_ALLOC_ZERO) != 0) {
4741 if ((m->flags & PG_ZERO) == 0)
4745 vm_page_grab_release(m, allocflags);
4747 m = vm_page_next(m);
4753 * Unlocked variant of vm_page_grab_pages(). This accepts the same flags
4754 * and will fall back to the locked variant to handle allocation.
4757 vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex,
4758 int allocflags, vm_page_t *ma, int count)
4764 vm_page_grab_check(allocflags);
4767 * Modify flags for lockless acquire to hold the page until we
4768 * set it valid if necessary.
4770 flags = allocflags & ~VM_ALLOC_NOBUSY;
4772 for (i = 0; i < count; i++, pindex++) {
4773 if (!vm_page_acquire_unlocked(object, pindex, pred, &m, flags))
4777 if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) {
4778 if ((m->flags & PG_ZERO) == 0)
4782 /* m will still be wired or busy according to flags. */
4783 vm_page_grab_release(m, allocflags);
4786 if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4789 VM_OBJECT_WLOCK(object);
4790 i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count);
4791 VM_OBJECT_WUNLOCK(object);
4797 * Mapping function for valid or dirty bits in a page.
4799 * Inputs are required to range within a page.
4802 vm_page_bits(int base, int size)
4808 base + size <= PAGE_SIZE,
4809 ("vm_page_bits: illegal base/size %d/%d", base, size)
4812 if (size == 0) /* handle degenerate case */
4815 first_bit = base >> DEV_BSHIFT;
4816 last_bit = (base + size - 1) >> DEV_BSHIFT;
4818 return (((vm_page_bits_t)2 << last_bit) -
4819 ((vm_page_bits_t)1 << first_bit));
4823 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set)
4826 #if PAGE_SIZE == 32768
4827 atomic_set_64((uint64_t *)bits, set);
4828 #elif PAGE_SIZE == 16384
4829 atomic_set_32((uint32_t *)bits, set);
4830 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16)
4831 atomic_set_16((uint16_t *)bits, set);
4832 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8)
4833 atomic_set_8((uint8_t *)bits, set);
4834 #else /* PAGE_SIZE <= 8192 */
4838 addr = (uintptr_t)bits;
4840 * Use a trick to perform a 32-bit atomic on the
4841 * containing aligned word, to not depend on the existence
4842 * of atomic_{set, clear}_{8, 16}.
4844 shift = addr & (sizeof(uint32_t) - 1);
4845 #if BYTE_ORDER == BIG_ENDIAN
4846 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4850 addr &= ~(sizeof(uint32_t) - 1);
4851 atomic_set_32((uint32_t *)addr, set << shift);
4852 #endif /* PAGE_SIZE */
4856 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear)
4859 #if PAGE_SIZE == 32768
4860 atomic_clear_64((uint64_t *)bits, clear);
4861 #elif PAGE_SIZE == 16384
4862 atomic_clear_32((uint32_t *)bits, clear);
4863 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16)
4864 atomic_clear_16((uint16_t *)bits, clear);
4865 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8)
4866 atomic_clear_8((uint8_t *)bits, clear);
4867 #else /* PAGE_SIZE <= 8192 */
4871 addr = (uintptr_t)bits;
4873 * Use a trick to perform a 32-bit atomic on the
4874 * containing aligned word, to not depend on the existence
4875 * of atomic_{set, clear}_{8, 16}.
4877 shift = addr & (sizeof(uint32_t) - 1);
4878 #if BYTE_ORDER == BIG_ENDIAN
4879 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4883 addr &= ~(sizeof(uint32_t) - 1);
4884 atomic_clear_32((uint32_t *)addr, clear << shift);
4885 #endif /* PAGE_SIZE */
4888 static inline vm_page_bits_t
4889 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits)
4891 #if PAGE_SIZE == 32768
4895 while (atomic_fcmpset_64(bits, &old, newbits) == 0);
4897 #elif PAGE_SIZE == 16384
4901 while (atomic_fcmpset_32(bits, &old, newbits) == 0);
4903 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16)
4907 while (atomic_fcmpset_16(bits, &old, newbits) == 0);
4909 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8)
4913 while (atomic_fcmpset_8(bits, &old, newbits) == 0);
4915 #else /* PAGE_SIZE <= 4096*/
4917 uint32_t old, new, mask;
4920 addr = (uintptr_t)bits;
4922 * Use a trick to perform a 32-bit atomic on the
4923 * containing aligned word, to not depend on the existence
4924 * of atomic_{set, swap, clear}_{8, 16}.
4926 shift = addr & (sizeof(uint32_t) - 1);
4927 #if BYTE_ORDER == BIG_ENDIAN
4928 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4932 addr &= ~(sizeof(uint32_t) - 1);
4933 mask = VM_PAGE_BITS_ALL << shift;
4938 new |= newbits << shift;
4939 } while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0);
4940 return (old >> shift);
4941 #endif /* PAGE_SIZE */
4945 * vm_page_set_valid_range:
4947 * Sets portions of a page valid. The arguments are expected
4948 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4949 * of any partial chunks touched by the range. The invalid portion of
4950 * such chunks will be zeroed.
4952 * (base + size) must be less then or equal to PAGE_SIZE.
4955 vm_page_set_valid_range(vm_page_t m, int base, int size)
4958 vm_page_bits_t pagebits;
4960 vm_page_assert_busied(m);
4961 if (size == 0) /* handle degenerate case */
4965 * If the base is not DEV_BSIZE aligned and the valid
4966 * bit is clear, we have to zero out a portion of the
4969 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4970 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
4971 pmap_zero_page_area(m, frag, base - frag);
4974 * If the ending offset is not DEV_BSIZE aligned and the
4975 * valid bit is clear, we have to zero out a portion of
4978 endoff = base + size;
4979 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4980 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
4981 pmap_zero_page_area(m, endoff,
4982 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4985 * Assert that no previously invalid block that is now being validated
4988 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
4989 ("vm_page_set_valid_range: page %p is dirty", m));
4992 * Set valid bits inclusive of any overlap.
4994 pagebits = vm_page_bits(base, size);
4995 if (vm_page_xbusied(m))
4996 m->valid |= pagebits;
4998 vm_page_bits_set(m, &m->valid, pagebits);
5002 * Set the page dirty bits and free the invalid swap space if
5003 * present. Returns the previous dirty bits.
5006 vm_page_set_dirty(vm_page_t m)
5010 VM_PAGE_OBJECT_BUSY_ASSERT(m);
5012 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) {
5014 m->dirty = VM_PAGE_BITS_ALL;
5016 old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL);
5017 if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0)
5018 vm_pager_page_unswapped(m);
5024 * Clear the given bits from the specified page's dirty field.
5026 static __inline void
5027 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
5030 vm_page_assert_busied(m);
5033 * If the page is xbusied and not write mapped we are the
5034 * only thread that can modify dirty bits. Otherwise, The pmap
5035 * layer can call vm_page_dirty() without holding a distinguished
5036 * lock. The combination of page busy and atomic operations
5037 * suffice to guarantee consistency of the page dirty field.
5039 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
5040 m->dirty &= ~pagebits;
5042 vm_page_bits_clear(m, &m->dirty, pagebits);
5046 * vm_page_set_validclean:
5048 * Sets portions of a page valid and clean. The arguments are expected
5049 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5050 * of any partial chunks touched by the range. The invalid portion of
5051 * such chunks will be zero'd.
5053 * (base + size) must be less then or equal to PAGE_SIZE.
5056 vm_page_set_validclean(vm_page_t m, int base, int size)
5058 vm_page_bits_t oldvalid, pagebits;
5061 vm_page_assert_busied(m);
5062 if (size == 0) /* handle degenerate case */
5066 * If the base is not DEV_BSIZE aligned and the valid
5067 * bit is clear, we have to zero out a portion of the
5070 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5071 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
5072 pmap_zero_page_area(m, frag, base - frag);
5075 * If the ending offset is not DEV_BSIZE aligned and the
5076 * valid bit is clear, we have to zero out a portion of
5079 endoff = base + size;
5080 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5081 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
5082 pmap_zero_page_area(m, endoff,
5083 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5086 * Set valid, clear dirty bits. If validating the entire
5087 * page we can safely clear the pmap modify bit. We also
5088 * use this opportunity to clear the PGA_NOSYNC flag. If a process
5089 * takes a write fault on a MAP_NOSYNC memory area the flag will
5092 * We set valid bits inclusive of any overlap, but we can only
5093 * clear dirty bits for DEV_BSIZE chunks that are fully within
5096 oldvalid = m->valid;
5097 pagebits = vm_page_bits(base, size);
5098 if (vm_page_xbusied(m))
5099 m->valid |= pagebits;
5101 vm_page_bits_set(m, &m->valid, pagebits);
5103 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
5104 frag = DEV_BSIZE - frag;
5110 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
5112 if (base == 0 && size == PAGE_SIZE) {
5114 * The page can only be modified within the pmap if it is
5115 * mapped, and it can only be mapped if it was previously
5118 if (oldvalid == VM_PAGE_BITS_ALL)
5120 * Perform the pmap_clear_modify() first. Otherwise,
5121 * a concurrent pmap operation, such as
5122 * pmap_protect(), could clear a modification in the
5123 * pmap and set the dirty field on the page before
5124 * pmap_clear_modify() had begun and after the dirty
5125 * field was cleared here.
5127 pmap_clear_modify(m);
5129 vm_page_aflag_clear(m, PGA_NOSYNC);
5130 } else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m))
5131 m->dirty &= ~pagebits;
5133 vm_page_clear_dirty_mask(m, pagebits);
5137 vm_page_clear_dirty(vm_page_t m, int base, int size)
5140 vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
5144 * vm_page_set_invalid:
5146 * Invalidates DEV_BSIZE'd chunks within a page. Both the
5147 * valid and dirty bits for the effected areas are cleared.
5150 vm_page_set_invalid(vm_page_t m, int base, int size)
5152 vm_page_bits_t bits;
5156 * The object lock is required so that pages can't be mapped
5157 * read-only while we're in the process of invalidating them.
5160 VM_OBJECT_ASSERT_WLOCKED(object);
5161 vm_page_assert_busied(m);
5163 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
5164 size >= object->un_pager.vnp.vnp_size)
5165 bits = VM_PAGE_BITS_ALL;
5167 bits = vm_page_bits(base, size);
5168 if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0)
5170 KASSERT((bits == 0 && vm_page_all_valid(m)) ||
5171 !pmap_page_is_mapped(m),
5172 ("vm_page_set_invalid: page %p is mapped", m));
5173 if (vm_page_xbusied(m)) {
5177 vm_page_bits_clear(m, &m->valid, bits);
5178 vm_page_bits_clear(m, &m->dirty, bits);
5185 * Invalidates the entire page. The page must be busy, unmapped, and
5186 * the enclosing object must be locked. The object locks protects
5187 * against concurrent read-only pmap enter which is done without
5191 vm_page_invalid(vm_page_t m)
5194 vm_page_assert_busied(m);
5195 VM_OBJECT_ASSERT_LOCKED(m->object);
5196 MPASS(!pmap_page_is_mapped(m));
5198 if (vm_page_xbusied(m))
5201 vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL);
5205 * vm_page_zero_invalid()
5207 * The kernel assumes that the invalid portions of a page contain
5208 * garbage, but such pages can be mapped into memory by user code.
5209 * When this occurs, we must zero out the non-valid portions of the
5210 * page so user code sees what it expects.
5212 * Pages are most often semi-valid when the end of a file is mapped
5213 * into memory and the file's size is not page aligned.
5216 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
5222 * Scan the valid bits looking for invalid sections that
5223 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the
5224 * valid bit may be set ) have already been zeroed by
5225 * vm_page_set_validclean().
5227 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
5228 if (i == (PAGE_SIZE / DEV_BSIZE) ||
5229 (m->valid & ((vm_page_bits_t)1 << i))) {
5231 pmap_zero_page_area(m,
5232 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
5239 * setvalid is TRUE when we can safely set the zero'd areas
5240 * as being valid. We can do this if there are no cache consistancy
5241 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
5250 * Is (partial) page valid? Note that the case where size == 0
5251 * will return FALSE in the degenerate case where the page is
5252 * entirely invalid, and TRUE otherwise.
5254 * Some callers envoke this routine without the busy lock held and
5255 * handle races via higher level locks. Typical callers should
5256 * hold a busy lock to prevent invalidation.
5259 vm_page_is_valid(vm_page_t m, int base, int size)
5261 vm_page_bits_t bits;
5263 bits = vm_page_bits(base, size);
5264 return (m->valid != 0 && (m->valid & bits) == bits);
5268 * Returns true if all of the specified predicates are true for the entire
5269 * (super)page and false otherwise.
5272 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
5278 if (skip_m != NULL && skip_m->object != object)
5280 VM_OBJECT_ASSERT_LOCKED(object);
5281 npages = atop(pagesizes[m->psind]);
5284 * The physically contiguous pages that make up a superpage, i.e., a
5285 * page with a page size index ("psind") greater than zero, will
5286 * occupy adjacent entries in vm_page_array[].
5288 for (i = 0; i < npages; i++) {
5289 /* Always test object consistency, including "skip_m". */
5290 if (m[i].object != object)
5292 if (&m[i] == skip_m)
5294 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
5296 if ((flags & PS_ALL_DIRTY) != 0) {
5298 * Calling vm_page_test_dirty() or pmap_is_modified()
5299 * might stop this case from spuriously returning
5300 * "false". However, that would require a write lock
5301 * on the object containing "m[i]".
5303 if (m[i].dirty != VM_PAGE_BITS_ALL)
5306 if ((flags & PS_ALL_VALID) != 0 &&
5307 m[i].valid != VM_PAGE_BITS_ALL)
5314 * Set the page's dirty bits if the page is modified.
5317 vm_page_test_dirty(vm_page_t m)
5320 vm_page_assert_busied(m);
5321 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
5326 vm_page_valid(vm_page_t m)
5329 vm_page_assert_busied(m);
5330 if (vm_page_xbusied(m))
5331 m->valid = VM_PAGE_BITS_ALL;
5333 vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL);
5337 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
5340 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
5344 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
5347 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
5351 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
5354 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
5357 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
5359 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
5362 vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
5366 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
5369 mtx_assert_(vm_page_lockptr(m), a, file, line);
5375 vm_page_object_busy_assert(vm_page_t m)
5379 * Certain of the page's fields may only be modified by the
5380 * holder of a page or object busy.
5382 if (m->object != NULL && !vm_page_busied(m))
5383 VM_OBJECT_ASSERT_BUSY(m->object);
5387 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits)
5390 if ((bits & PGA_WRITEABLE) == 0)
5394 * The PGA_WRITEABLE flag can only be set if the page is
5395 * managed, is exclusively busied or the object is locked.
5396 * Currently, this flag is only set by pmap_enter().
5398 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5399 ("PGA_WRITEABLE on unmanaged page"));
5400 if (!vm_page_xbusied(m))
5401 VM_OBJECT_ASSERT_BUSY(m->object);
5405 #include "opt_ddb.h"
5407 #include <sys/kernel.h>
5409 #include <ddb/ddb.h>
5411 DB_SHOW_COMMAND(page, vm_page_print_page_info)
5414 db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
5415 db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
5416 db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
5417 db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
5418 db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
5419 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
5420 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
5421 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
5422 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
5425 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
5429 db_printf("pq_free %d\n", vm_free_count());
5430 for (dom = 0; dom < vm_ndomains; dom++) {
5432 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
5434 vm_dom[dom].vmd_page_count,
5435 vm_dom[dom].vmd_free_count,
5436 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
5437 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
5438 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
5439 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
5443 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
5446 boolean_t phys, virt;
5449 db_printf("show pginfo addr\n");
5453 phys = strchr(modif, 'p') != NULL;
5454 virt = strchr(modif, 'v') != NULL;
5456 m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
5458 m = PHYS_TO_VM_PAGE(addr);
5460 m = (vm_page_t)addr;
5462 "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n"
5463 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
5464 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
5465 m->a.queue, m->ref_count, m->a.flags, m->oflags,
5466 m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty);