2 * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
4 * Copyright (c) 1991 Regents of the University of California.
6 * Copyright (c) 1998 Matthew Dillon. All Rights Reserved.
8 * This code is derived from software contributed to Berkeley by
9 * The Mach Operating System project at Carnegie-Mellon University.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
14 * 1. Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
16 * 2. Redistributions in binary form must reproduce the above copyright
17 * notice, this list of conditions and the following disclaimer in the
18 * documentation and/or other materials provided with the distribution.
19 * 3. Neither the name of the University nor the names of its contributors
20 * may be used to endorse or promote products derived from this software
21 * without specific prior written permission.
23 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
24 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
27 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91
39 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40 * All rights reserved.
42 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
44 * Permission to use, copy, modify and distribute this software and
45 * its documentation is hereby granted, provided that both the copyright
46 * notice and this permission notice appear in all copies of the
47 * software, derivative works or modified versions, and any portions
48 * thereof, and that both notices appear in supporting documentation.
50 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
54 * Carnegie Mellon requests users of this software to return to
56 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
57 * School of Computer Science
58 * Carnegie Mellon University
59 * Pittsburgh PA 15213-3890
61 * any improvements or extensions that they make and grant Carnegie the
62 * rights to redistribute these changes.
66 * Resident memory management module.
69 #include <sys/cdefs.h>
70 __FBSDID("$FreeBSD$");
74 #include <sys/param.h>
75 #include <sys/systm.h>
76 #include <sys/counter.h>
77 #include <sys/domainset.h>
78 #include <sys/kernel.h>
79 #include <sys/limits.h>
80 #include <sys/linker.h>
82 #include <sys/malloc.h>
84 #include <sys/msgbuf.h>
85 #include <sys/mutex.h>
87 #include <sys/rwlock.h>
88 #include <sys/sleepqueue.h>
90 #include <sys/sched.h>
92 #include <sys/sysctl.h>
93 #include <sys/vmmeter.h>
94 #include <sys/vnode.h>
98 #include <vm/vm_param.h>
99 #include <vm/vm_domainset.h>
100 #include <vm/vm_kern.h>
101 #include <vm/vm_map.h>
102 #include <vm/vm_object.h>
103 #include <vm/vm_page.h>
104 #include <vm/vm_pageout.h>
105 #include <vm/vm_phys.h>
106 #include <vm/vm_pagequeue.h>
107 #include <vm/vm_pager.h>
108 #include <vm/vm_radix.h>
109 #include <vm/vm_reserv.h>
110 #include <vm/vm_extern.h>
112 #include <vm/uma_int.h>
114 #include <machine/md_var.h>
116 extern int uma_startup_count(int);
117 extern void uma_startup(void *, int);
118 extern int vmem_startup_count(void);
120 struct vm_domain vm_dom[MAXMEMDOM];
122 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]);
124 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
126 struct mtx_padalign __exclusive_cache_line vm_domainset_lock;
127 /* The following fields are protected by the domainset lock. */
128 domainset_t __exclusive_cache_line vm_min_domains;
129 domainset_t __exclusive_cache_line vm_severe_domains;
130 static int vm_min_waiters;
131 static int vm_severe_waiters;
132 static int vm_pageproc_waiters;
134 static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD, 0,
135 "VM page statistics");
137 static counter_u64_t queue_ops = EARLY_COUNTER;
138 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops,
139 CTLFLAG_RD, &queue_ops,
140 "Number of batched queue operations");
142 static counter_u64_t queue_nops = EARLY_COUNTER;
143 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops,
144 CTLFLAG_RD, &queue_nops,
145 "Number of batched queue operations with no effects");
148 counter_startup(void)
151 queue_ops = counter_u64_alloc(M_WAITOK);
152 queue_nops = counter_u64_alloc(M_WAITOK);
154 SYSINIT(page_counters, SI_SUB_CPU, SI_ORDER_ANY, counter_startup, NULL);
157 * bogus page -- for I/O to/from partially complete buffers,
158 * or for paging into sparsely invalid regions.
160 vm_page_t bogus_page;
162 vm_page_t vm_page_array;
163 long vm_page_array_size;
166 static int boot_pages;
167 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
169 "number of pages allocated for bootstrapping the VM system");
171 static TAILQ_HEAD(, vm_page) blacklist_head;
172 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
173 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
174 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
176 static uma_zone_t fakepg_zone;
178 static void vm_page_alloc_check(vm_page_t m);
179 static void _vm_page_busy_sleep(vm_object_t obj, vm_page_t m,
180 const char *wmesg, bool nonshared, bool locked);
181 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
182 static void vm_page_dequeue_complete(vm_page_t m);
183 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
184 static void vm_page_init(void *dummy);
185 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
186 vm_pindex_t pindex, vm_page_t mpred);
187 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
189 static void vm_page_mvqueue(vm_page_t m, uint8_t queue);
190 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
191 vm_page_t m_run, vm_paddr_t high);
192 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
194 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain,
196 static void vm_page_zone_release(void *arg, void **store, int cnt);
198 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
201 vm_page_init(void *dummy)
204 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
205 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
206 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
207 VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
211 * The cache page zone is initialized later since we need to be able to allocate
212 * pages before UMA is fully initialized.
215 vm_page_init_cache_zones(void *dummy __unused)
217 struct vm_domain *vmd;
218 struct vm_pgcache *pgcache;
219 int cache, domain, maxcache, pool;
222 TUNABLE_INT_FETCH("vm.pgcache_zone_max", &maxcache);
223 for (domain = 0; domain < vm_ndomains; domain++) {
224 vmd = VM_DOMAIN(domain);
225 for (pool = 0; pool < VM_NFREEPOOL; pool++) {
226 pgcache = &vmd->vmd_pgcache[pool];
227 pgcache->domain = domain;
228 pgcache->pool = pool;
229 pgcache->zone = uma_zcache_create("vm pgcache",
230 PAGE_SIZE, NULL, NULL, NULL, NULL,
231 vm_page_zone_import, vm_page_zone_release, pgcache,
235 * Limit each pool's zone to 0.1% of the pages in the
238 cache = maxcache != 0 ? maxcache :
239 vmd->vmd_page_count / 1000;
240 uma_zone_set_maxcache(pgcache->zone, cache);
244 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
246 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
247 #if PAGE_SIZE == 32768
249 CTASSERT(sizeof(u_long) >= 8);
256 * Sets the page size, perhaps based upon the memory
257 * size. Must be called before any use of page-size
258 * dependent functions.
261 vm_set_page_size(void)
263 if (vm_cnt.v_page_size == 0)
264 vm_cnt.v_page_size = PAGE_SIZE;
265 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
266 panic("vm_set_page_size: page size not a power of two");
270 * vm_page_blacklist_next:
272 * Find the next entry in the provided string of blacklist
273 * addresses. Entries are separated by space, comma, or newline.
274 * If an invalid integer is encountered then the rest of the
275 * string is skipped. Updates the list pointer to the next
276 * character, or NULL if the string is exhausted or invalid.
279 vm_page_blacklist_next(char **list, char *end)
284 if (list == NULL || *list == NULL)
292 * If there's no end pointer then the buffer is coming from
293 * the kenv and we know it's null-terminated.
296 end = *list + strlen(*list);
298 /* Ensure that strtoq() won't walk off the end */
300 if (*end == '\n' || *end == ' ' || *end == ',')
303 printf("Blacklist not terminated, skipping\n");
309 for (pos = *list; *pos != '\0'; pos = cp) {
310 bad = strtoq(pos, &cp, 0);
311 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
320 if (*cp == '\0' || ++cp >= end)
324 return (trunc_page(bad));
326 printf("Garbage in RAM blacklist, skipping\n");
332 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
334 struct vm_domain *vmd;
338 m = vm_phys_paddr_to_vm_page(pa);
340 return (true); /* page does not exist, no failure */
342 vmd = vm_pagequeue_domain(m);
343 vm_domain_free_lock(vmd);
344 ret = vm_phys_unfree_page(m);
345 vm_domain_free_unlock(vmd);
347 vm_domain_freecnt_inc(vmd, -1);
348 TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
350 printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
356 * vm_page_blacklist_check:
358 * Iterate through the provided string of blacklist addresses, pulling
359 * each entry out of the physical allocator free list and putting it
360 * onto a list for reporting via the vm.page_blacklist sysctl.
363 vm_page_blacklist_check(char *list, char *end)
369 while (next != NULL) {
370 if ((pa = vm_page_blacklist_next(&next, end)) == 0)
372 vm_page_blacklist_add(pa, bootverbose);
377 * vm_page_blacklist_load:
379 * Search for a special module named "ram_blacklist". It'll be a
380 * plain text file provided by the user via the loader directive
384 vm_page_blacklist_load(char **list, char **end)
393 mod = preload_search_by_type("ram_blacklist");
395 ptr = preload_fetch_addr(mod);
396 len = preload_fetch_size(mod);
407 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
414 error = sysctl_wire_old_buffer(req, 0);
417 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
418 TAILQ_FOREACH(m, &blacklist_head, listq) {
419 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
420 (uintmax_t)m->phys_addr);
423 error = sbuf_finish(&sbuf);
429 * Initialize a dummy page for use in scans of the specified paging queue.
430 * In principle, this function only needs to set the flag PG_MARKER.
431 * Nonetheless, it write busies the page as a safety precaution.
434 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags)
437 bzero(marker, sizeof(*marker));
438 marker->flags = PG_MARKER;
439 marker->a.flags = aflags;
440 marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
441 marker->a.queue = queue;
445 vm_page_domain_init(int domain)
447 struct vm_domain *vmd;
448 struct vm_pagequeue *pq;
451 vmd = VM_DOMAIN(domain);
452 bzero(vmd, sizeof(*vmd));
453 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
454 "vm inactive pagequeue";
455 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
456 "vm active pagequeue";
457 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
458 "vm laundry pagequeue";
459 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
460 "vm unswappable pagequeue";
461 vmd->vmd_domain = domain;
462 vmd->vmd_page_count = 0;
463 vmd->vmd_free_count = 0;
465 vmd->vmd_oom = FALSE;
466 for (i = 0; i < PQ_COUNT; i++) {
467 pq = &vmd->vmd_pagequeues[i];
468 TAILQ_INIT(&pq->pq_pl);
469 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
470 MTX_DEF | MTX_DUPOK);
472 vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
474 mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
475 mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
476 snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
479 * inacthead is used to provide FIFO ordering for LRU-bypassing
482 vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
483 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
484 &vmd->vmd_inacthead, plinks.q);
487 * The clock pages are used to implement active queue scanning without
488 * requeues. Scans start at clock[0], which is advanced after the scan
489 * ends. When the two clock hands meet, they are reset and scanning
490 * resumes from the head of the queue.
492 vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
493 vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
494 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
495 &vmd->vmd_clock[0], plinks.q);
496 TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
497 &vmd->vmd_clock[1], plinks.q);
501 * Initialize a physical page in preparation for adding it to the free
505 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
510 m->busy_lock = VPB_UNBUSIED;
511 m->flags = m->a.flags = 0;
513 m->a.queue = PQ_NONE;
516 m->order = VM_NFREEORDER;
517 m->pool = VM_FREEPOOL_DEFAULT;
518 m->valid = m->dirty = 0;
522 #ifndef PMAP_HAS_PAGE_ARRAY
524 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range)
529 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
530 * However, because this page is allocated from KVM, out-of-bounds
531 * accesses using the direct map will not be trapped.
536 * Allocate physical memory for the page structures, and map it.
538 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
539 vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end,
540 VM_PROT_READ | VM_PROT_WRITE);
541 vm_page_array_size = page_range;
550 * Initializes the resident memory module. Allocates physical memory for
551 * bootstrapping UMA and some data structures that are used to manage
552 * physical pages. Initializes these structures, and populates the free
556 vm_page_startup(vm_offset_t vaddr)
558 struct vm_phys_seg *seg;
560 char *list, *listend;
562 vm_paddr_t end, high_avail, low_avail, new_end, size;
563 vm_paddr_t page_range __unused;
564 vm_paddr_t last_pa, pa;
566 int biggestone, i, segind;
570 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
574 vaddr = round_page(vaddr);
576 vm_phys_early_startup();
577 biggestone = vm_phys_avail_largest();
578 end = phys_avail[biggestone+1];
581 * Initialize the page and queue locks.
583 mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
584 for (i = 0; i < PA_LOCK_COUNT; i++)
585 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
586 for (i = 0; i < vm_ndomains; i++)
587 vm_page_domain_init(i);
590 * Allocate memory for use when boot strapping the kernel memory
591 * allocator. Tell UMA how many zones we are going to create
592 * before going fully functional. UMA will add its zones.
594 * VM startup zones: vmem, vmem_btag, VM OBJECT, RADIX NODE, MAP,
595 * KMAP ENTRY, MAP ENTRY, VMSPACE.
597 boot_pages = uma_startup_count(8);
599 #ifndef UMA_MD_SMALL_ALLOC
600 /* vmem_startup() calls uma_prealloc(). */
601 boot_pages += vmem_startup_count();
602 /* vm_map_startup() calls uma_prealloc(). */
603 boot_pages += howmany(MAX_KMAP,
604 slab_ipers(sizeof(struct vm_map), UMA_ALIGN_PTR));
607 * Before going fully functional kmem_init() does allocation
608 * from "KMAP ENTRY" and vmem_create() does allocation from "vmem".
613 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
614 * manually fetch the value.
616 TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
617 new_end = end - (boot_pages * UMA_SLAB_SIZE);
618 new_end = trunc_page(new_end);
619 mapped = pmap_map(&vaddr, new_end, end,
620 VM_PROT_READ | VM_PROT_WRITE);
621 bzero((void *)mapped, end - new_end);
622 uma_startup((void *)mapped, boot_pages);
625 witness_size = round_page(witness_startup_count());
626 new_end -= witness_size;
627 mapped = pmap_map(&vaddr, new_end, new_end + witness_size,
628 VM_PROT_READ | VM_PROT_WRITE);
629 bzero((void *)mapped, witness_size);
630 witness_startup((void *)mapped);
633 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
634 defined(__i386__) || defined(__mips__) || defined(__riscv) || \
635 defined(__powerpc64__)
637 * Allocate a bitmap to indicate that a random physical page
638 * needs to be included in a minidump.
640 * The amd64 port needs this to indicate which direct map pages
641 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
643 * However, i386 still needs this workspace internally within the
644 * minidump code. In theory, they are not needed on i386, but are
645 * included should the sf_buf code decide to use them.
648 for (i = 0; dump_avail[i + 1] != 0; i += 2)
649 if (dump_avail[i + 1] > last_pa)
650 last_pa = dump_avail[i + 1];
651 page_range = last_pa / PAGE_SIZE;
652 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
653 new_end -= vm_page_dump_size;
654 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
655 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
656 bzero((void *)vm_page_dump, vm_page_dump_size);
660 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
661 defined(__riscv) || defined(__powerpc64__)
663 * Include the UMA bootstrap pages, witness pages and vm_page_dump
664 * in a crash dump. When pmap_map() uses the direct map, they are
665 * not automatically included.
667 for (pa = new_end; pa < end; pa += PAGE_SIZE)
670 phys_avail[biggestone + 1] = new_end;
673 * Request that the physical pages underlying the message buffer be
674 * included in a crash dump. Since the message buffer is accessed
675 * through the direct map, they are not automatically included.
677 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
678 last_pa = pa + round_page(msgbufsize);
679 while (pa < last_pa) {
685 * Compute the number of pages of memory that will be available for
686 * use, taking into account the overhead of a page structure per page.
687 * In other words, solve
688 * "available physical memory" - round_page(page_range *
689 * sizeof(struct vm_page)) = page_range * PAGE_SIZE
692 low_avail = phys_avail[0];
693 high_avail = phys_avail[1];
694 for (i = 0; i < vm_phys_nsegs; i++) {
695 if (vm_phys_segs[i].start < low_avail)
696 low_avail = vm_phys_segs[i].start;
697 if (vm_phys_segs[i].end > high_avail)
698 high_avail = vm_phys_segs[i].end;
700 /* Skip the first chunk. It is already accounted for. */
701 for (i = 2; phys_avail[i + 1] != 0; i += 2) {
702 if (phys_avail[i] < low_avail)
703 low_avail = phys_avail[i];
704 if (phys_avail[i + 1] > high_avail)
705 high_avail = phys_avail[i + 1];
707 first_page = low_avail / PAGE_SIZE;
708 #ifdef VM_PHYSSEG_SPARSE
710 for (i = 0; i < vm_phys_nsegs; i++)
711 size += vm_phys_segs[i].end - vm_phys_segs[i].start;
712 for (i = 0; phys_avail[i + 1] != 0; i += 2)
713 size += phys_avail[i + 1] - phys_avail[i];
714 #elif defined(VM_PHYSSEG_DENSE)
715 size = high_avail - low_avail;
717 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
720 #ifdef PMAP_HAS_PAGE_ARRAY
721 pmap_page_array_startup(size / PAGE_SIZE);
722 biggestone = vm_phys_avail_largest();
723 end = new_end = phys_avail[biggestone + 1];
725 #ifdef VM_PHYSSEG_DENSE
727 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
728 * the overhead of a page structure per page only if vm_page_array is
729 * allocated from the last physical memory chunk. Otherwise, we must
730 * allocate page structures representing the physical memory
731 * underlying vm_page_array, even though they will not be used.
733 if (new_end != high_avail)
734 page_range = size / PAGE_SIZE;
738 page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
741 * If the partial bytes remaining are large enough for
742 * a page (PAGE_SIZE) without a corresponding
743 * 'struct vm_page', then new_end will contain an
744 * extra page after subtracting the length of the VM
745 * page array. Compensate by subtracting an extra
748 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
749 if (new_end == high_avail)
750 high_avail -= PAGE_SIZE;
751 new_end -= PAGE_SIZE;
755 new_end = vm_page_array_alloc(&vaddr, end, page_range);
758 #if VM_NRESERVLEVEL > 0
760 * Allocate physical memory for the reservation management system's
761 * data structures, and map it.
763 new_end = vm_reserv_startup(&vaddr, new_end);
765 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
766 defined(__riscv) || defined(__powerpc64__)
768 * Include vm_page_array and vm_reserv_array in a crash dump.
770 for (pa = new_end; pa < end; pa += PAGE_SIZE)
773 phys_avail[biggestone + 1] = new_end;
776 * Add physical memory segments corresponding to the available
779 for (i = 0; phys_avail[i + 1] != 0; i += 2)
780 if (vm_phys_avail_size(i) != 0)
781 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
784 * Initialize the physical memory allocator.
789 * Initialize the page structures and add every available page to the
790 * physical memory allocator's free lists.
792 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
793 for (ii = 0; ii < vm_page_array_size; ii++) {
794 m = &vm_page_array[ii];
795 vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0);
796 m->flags = PG_FICTITIOUS;
799 vm_cnt.v_page_count = 0;
800 for (segind = 0; segind < vm_phys_nsegs; segind++) {
801 seg = &vm_phys_segs[segind];
802 for (m = seg->first_page, pa = seg->start; pa < seg->end;
803 m++, pa += PAGE_SIZE)
804 vm_page_init_page(m, pa, segind);
807 * Add the segment to the free lists only if it is covered by
808 * one of the ranges in phys_avail. Because we've added the
809 * ranges to the vm_phys_segs array, we can assume that each
810 * segment is either entirely contained in one of the ranges,
811 * or doesn't overlap any of them.
813 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
814 struct vm_domain *vmd;
816 if (seg->start < phys_avail[i] ||
817 seg->end > phys_avail[i + 1])
821 pagecount = (u_long)atop(seg->end - seg->start);
823 vmd = VM_DOMAIN(seg->domain);
824 vm_domain_free_lock(vmd);
825 vm_phys_enqueue_contig(m, pagecount);
826 vm_domain_free_unlock(vmd);
827 vm_domain_freecnt_inc(vmd, pagecount);
828 vm_cnt.v_page_count += (u_int)pagecount;
830 vmd = VM_DOMAIN(seg->domain);
831 vmd->vmd_page_count += (u_int)pagecount;
832 vmd->vmd_segs |= 1UL << m->segind;
838 * Remove blacklisted pages from the physical memory allocator.
840 TAILQ_INIT(&blacklist_head);
841 vm_page_blacklist_load(&list, &listend);
842 vm_page_blacklist_check(list, listend);
844 list = kern_getenv("vm.blacklist");
845 vm_page_blacklist_check(list, NULL);
848 #if VM_NRESERVLEVEL > 0
850 * Initialize the reservation management system.
859 vm_page_reference(vm_page_t m)
862 vm_page_aflag_set(m, PGA_REFERENCED);
866 vm_page_acquire_flags(vm_page_t m, int allocflags)
870 if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0)
871 locked = vm_page_trysbusy(m);
873 locked = vm_page_tryxbusy(m);
874 if (locked && (allocflags & VM_ALLOC_WIRED) != 0)
880 vm_page_busy_sleep_flags(vm_object_t object, vm_page_t m, const char *wchan,
884 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
887 * Reference the page before unlocking and
888 * sleeping so that the page daemon is less
889 * likely to reclaim it.
891 if ((allocflags & VM_ALLOC_NOCREAT) == 0)
892 vm_page_aflag_set(m, PGA_REFERENCED);
893 vm_page_busy_sleep(m, wchan, (allocflags &
894 VM_ALLOC_IGN_SBUSY) != 0);
895 VM_OBJECT_WLOCK(object);
896 if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
902 * vm_page_busy_acquire:
904 * Acquire the busy lock as described by VM_ALLOC_* flags. Will loop
905 * and drop the object lock if necessary.
908 vm_page_busy_acquire(vm_page_t m, int allocflags)
914 * The page-specific object must be cached because page
915 * identity can change during the sleep, causing the
916 * re-lock of a different object.
917 * It is assumed that a reference to the object is already
918 * held by the callers.
922 if (vm_page_acquire_flags(m, allocflags))
924 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
927 locked = VM_OBJECT_WOWNED(obj);
930 MPASS(locked || vm_page_wired(m));
931 _vm_page_busy_sleep(obj, m, "vmpba",
932 (allocflags & VM_ALLOC_SBUSY) != 0, locked);
934 VM_OBJECT_WLOCK(obj);
935 if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
937 KASSERT(m->object == obj || m->object == NULL,
938 ("vm_page_busy_acquire: page %p does not belong to %p",
944 * vm_page_busy_downgrade:
946 * Downgrade an exclusive busy page into a single shared busy page.
949 vm_page_busy_downgrade(vm_page_t m)
953 vm_page_assert_xbusied(m);
957 if (atomic_fcmpset_rel_int(&m->busy_lock,
958 &x, VPB_SHARERS_WORD(1)))
961 if ((x & VPB_BIT_WAITERS) != 0)
967 * vm_page_busy_tryupgrade:
969 * Attempt to upgrade a single shared busy into an exclusive busy.
972 vm_page_busy_tryupgrade(vm_page_t m)
976 vm_page_assert_sbusied(m);
979 ce = VPB_CURTHREAD_EXCLUSIVE;
981 if (VPB_SHARERS(x) > 1)
983 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
984 ("vm_page_busy_tryupgrade: invalid lock state"));
985 if (!atomic_fcmpset_acq_int(&m->busy_lock, &x,
986 ce | (x & VPB_BIT_WAITERS)))
995 * Return a positive value if the page is shared busied, 0 otherwise.
998 vm_page_sbusied(vm_page_t m)
1003 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
1009 * Shared unbusy a page.
1012 vm_page_sunbusy(vm_page_t m)
1016 vm_page_assert_sbusied(m);
1020 if (VPB_SHARERS(x) > 1) {
1021 if (atomic_fcmpset_int(&m->busy_lock, &x,
1022 x - VPB_ONE_SHARER))
1026 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
1027 ("vm_page_sunbusy: invalid lock state"));
1028 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1030 if ((x & VPB_BIT_WAITERS) == 0)
1038 * vm_page_busy_sleep:
1040 * Sleep if the page is busy, using the page pointer as wchan.
1041 * This is used to implement the hard-path of busying mechanism.
1043 * If nonshared is true, sleep only if the page is xbusy.
1045 * The object lock must be held on entry and will be released on exit.
1048 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
1053 VM_OBJECT_ASSERT_LOCKED(obj);
1054 vm_page_lock_assert(m, MA_NOTOWNED);
1056 _vm_page_busy_sleep(obj, m, wmesg, nonshared, true);
1060 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, const char *wmesg,
1061 bool nonshared, bool locked)
1066 * If the object is busy we must wait for that to drain to zero
1067 * before trying the page again.
1069 if (obj != NULL && vm_object_busied(obj)) {
1071 VM_OBJECT_DROP(obj);
1072 vm_object_busy_wait(obj, wmesg);
1077 if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) ||
1078 ((x & VPB_BIT_WAITERS) == 0 &&
1079 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) {
1081 VM_OBJECT_DROP(obj);
1086 VM_OBJECT_DROP(obj);
1088 sleepq_add(m, NULL, wmesg, 0, 0);
1089 sleepq_wait(m, PVM);
1096 * Try to shared busy a page.
1097 * If the operation succeeds 1 is returned otherwise 0.
1098 * The operation never sleeps.
1101 vm_page_trysbusy(vm_page_t m)
1109 if ((x & VPB_BIT_SHARED) == 0)
1112 * Reduce the window for transient busies that will trigger
1113 * false negatives in vm_page_ps_test().
1115 if (obj != NULL && vm_object_busied(obj))
1117 if (atomic_fcmpset_acq_int(&m->busy_lock, &x,
1118 x + VPB_ONE_SHARER))
1122 /* Refetch the object now that we're guaranteed that it is stable. */
1124 if (obj != NULL && vm_object_busied(obj)) {
1134 * Try to exclusive busy a page.
1135 * If the operation succeeds 1 is returned otherwise 0.
1136 * The operation never sleeps.
1139 vm_page_tryxbusy(vm_page_t m)
1143 if (atomic_cmpset_acq_int(&(m)->busy_lock, VPB_UNBUSIED,
1144 VPB_CURTHREAD_EXCLUSIVE) == 0)
1148 if (obj != NULL && vm_object_busied(obj)) {
1156 vm_page_xunbusy_hard_tail(vm_page_t m)
1158 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1159 /* Wake the waiter. */
1164 * vm_page_xunbusy_hard:
1166 * Called when unbusy has failed because there is a waiter.
1169 vm_page_xunbusy_hard(vm_page_t m)
1171 vm_page_assert_xbusied(m);
1172 vm_page_xunbusy_hard_tail(m);
1176 vm_page_xunbusy_hard_unchecked(vm_page_t m)
1178 vm_page_assert_xbusied_unchecked(m);
1179 vm_page_xunbusy_hard_tail(m);
1183 * Avoid releasing and reacquiring the same page lock.
1186 vm_page_change_lock(vm_page_t m, struct mtx **mtx)
1190 mtx1 = vm_page_lockptr(m);
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 if (vm_page_xbusied(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.
1358 if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1359 vm_page_activate(m);
1361 vm_page_deactivate(m);
1363 vm_page_xunbusy_unchecked(m);
1367 * vm_page_sleep_if_busy:
1369 * Sleep and release the object lock if the page is busied.
1370 * Returns TRUE if the thread slept.
1372 * The given page must be unlocked and object containing it must
1376 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1380 vm_page_lock_assert(m, MA_NOTOWNED);
1381 VM_OBJECT_ASSERT_WLOCKED(m->object);
1384 * The page-specific object must be cached because page
1385 * identity can change during the sleep, causing the
1386 * re-lock of a different object.
1387 * It is assumed that a reference to the object is already
1388 * held by the callers.
1391 if (vm_page_busied(m) || (obj != NULL && obj->busy)) {
1392 vm_page_busy_sleep(m, msg, false);
1393 VM_OBJECT_WLOCK(obj);
1400 * vm_page_sleep_if_xbusy:
1402 * Sleep and release the object lock if the page is xbusied.
1403 * Returns TRUE if the thread slept.
1405 * The given page must be unlocked and object containing it must
1409 vm_page_sleep_if_xbusy(vm_page_t m, const char *msg)
1413 vm_page_lock_assert(m, MA_NOTOWNED);
1414 VM_OBJECT_ASSERT_WLOCKED(m->object);
1417 * The page-specific object must be cached because page
1418 * identity can change during the sleep, causing the
1419 * re-lock of a different object.
1420 * It is assumed that a reference to the object is already
1421 * held by the callers.
1424 if (vm_page_xbusied(m) || (obj != NULL && obj->busy)) {
1425 vm_page_busy_sleep(m, msg, true);
1426 VM_OBJECT_WLOCK(obj);
1433 * vm_page_dirty_KBI: [ internal use only ]
1435 * Set all bits in the page's dirty field.
1437 * The object containing the specified page must be locked if the
1438 * call is made from the machine-independent layer.
1440 * See vm_page_clear_dirty_mask().
1442 * This function should only be called by vm_page_dirty().
1445 vm_page_dirty_KBI(vm_page_t m)
1448 /* Refer to this operation by its public name. */
1449 KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!"));
1450 m->dirty = VM_PAGE_BITS_ALL;
1454 * vm_page_insert: [ internal use only ]
1456 * Inserts the given mem entry into the object and object list.
1458 * The object must be locked.
1461 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1465 VM_OBJECT_ASSERT_WLOCKED(object);
1466 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1467 return (vm_page_insert_after(m, object, pindex, mpred));
1471 * vm_page_insert_after:
1473 * Inserts the page "m" into the specified object at offset "pindex".
1475 * The page "mpred" must immediately precede the offset "pindex" within
1476 * the specified object.
1478 * The object must be locked.
1481 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1486 VM_OBJECT_ASSERT_WLOCKED(object);
1487 KASSERT(m->object == NULL,
1488 ("vm_page_insert_after: page already inserted"));
1489 if (mpred != NULL) {
1490 KASSERT(mpred->object == object,
1491 ("vm_page_insert_after: object doesn't contain mpred"));
1492 KASSERT(mpred->pindex < pindex,
1493 ("vm_page_insert_after: mpred doesn't precede pindex"));
1494 msucc = TAILQ_NEXT(mpred, listq);
1496 msucc = TAILQ_FIRST(&object->memq);
1498 KASSERT(msucc->pindex > pindex,
1499 ("vm_page_insert_after: msucc doesn't succeed pindex"));
1502 * Record the object/offset pair in this page.
1506 m->ref_count |= VPRC_OBJREF;
1509 * Now link into the object's ordered list of backed pages.
1511 if (vm_radix_insert(&object->rtree, m)) {
1514 m->ref_count &= ~VPRC_OBJREF;
1517 vm_page_insert_radixdone(m, object, mpred);
1522 * vm_page_insert_radixdone:
1524 * Complete page "m" insertion into the specified object after the
1525 * radix trie hooking.
1527 * The page "mpred" must precede the offset "m->pindex" within the
1530 * The object must be locked.
1533 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1536 VM_OBJECT_ASSERT_WLOCKED(object);
1537 KASSERT(object != NULL && m->object == object,
1538 ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1539 KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1540 ("vm_page_insert_radixdone: page %p is missing object ref", m));
1541 if (mpred != NULL) {
1542 KASSERT(mpred->object == object,
1543 ("vm_page_insert_radixdone: object doesn't contain mpred"));
1544 KASSERT(mpred->pindex < m->pindex,
1545 ("vm_page_insert_radixdone: mpred doesn't precede pindex"));
1549 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1551 TAILQ_INSERT_HEAD(&object->memq, m, listq);
1554 * Show that the object has one more resident page.
1556 object->resident_page_count++;
1559 * Hold the vnode until the last page is released.
1561 if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1562 vhold(object->handle);
1565 * Since we are inserting a new and possibly dirty page,
1566 * update the object's generation count.
1568 if (pmap_page_is_write_mapped(m))
1569 vm_object_set_writeable_dirty(object);
1573 * Do the work to remove a page from its object. The caller is responsible for
1574 * updating the page's fields to reflect this removal.
1577 vm_page_object_remove(vm_page_t m)
1583 VM_OBJECT_ASSERT_WLOCKED(object);
1584 KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1585 ("page %p is missing its object ref", m));
1587 /* Deferred free of swap space. */
1588 if ((m->a.flags & PGA_SWAP_FREE) != 0)
1589 vm_pager_page_unswapped(m);
1591 mrem = vm_radix_remove(&object->rtree, m->pindex);
1592 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1595 * Now remove from the object's list of backed pages.
1597 TAILQ_REMOVE(&object->memq, m, listq);
1600 * And show that the object has one fewer resident page.
1602 object->resident_page_count--;
1605 * The vnode may now be recycled.
1607 if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1608 vdrop(object->handle);
1614 * Removes the specified page from its containing object, but does not
1615 * invalidate any backing storage. Returns true if the object's reference
1616 * was the last reference to the page, and false otherwise.
1618 * The object must be locked.
1621 vm_page_remove(vm_page_t m)
1624 vm_page_object_remove(m);
1626 return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1632 * Returns the page associated with the object/offset
1633 * pair specified; if none is found, NULL is returned.
1635 * The object must be locked.
1638 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1641 VM_OBJECT_ASSERT_LOCKED(object);
1642 return (vm_radix_lookup(&object->rtree, pindex));
1646 * vm_page_find_least:
1648 * Returns the page associated with the object with least pindex
1649 * greater than or equal to the parameter pindex, or NULL.
1651 * The object must be locked.
1654 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1658 VM_OBJECT_ASSERT_LOCKED(object);
1659 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1660 m = vm_radix_lookup_ge(&object->rtree, pindex);
1665 * Returns the given page's successor (by pindex) within the object if it is
1666 * resident; if none is found, NULL is returned.
1668 * The object must be locked.
1671 vm_page_next(vm_page_t m)
1675 VM_OBJECT_ASSERT_LOCKED(m->object);
1676 if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1677 MPASS(next->object == m->object);
1678 if (next->pindex != m->pindex + 1)
1685 * Returns the given page's predecessor (by pindex) within the object if it is
1686 * resident; if none is found, NULL is returned.
1688 * The object must be locked.
1691 vm_page_prev(vm_page_t m)
1695 VM_OBJECT_ASSERT_LOCKED(m->object);
1696 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1697 MPASS(prev->object == m->object);
1698 if (prev->pindex != m->pindex - 1)
1705 * Uses the page mnew as a replacement for an existing page at index
1706 * pindex which must be already present in the object.
1709 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1713 VM_OBJECT_ASSERT_WLOCKED(object);
1714 KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
1715 ("vm_page_replace: page %p already in object", mnew));
1718 * This function mostly follows vm_page_insert() and
1719 * vm_page_remove() without the radix, object count and vnode
1720 * dance. Double check such functions for more comments.
1723 mnew->object = object;
1724 mnew->pindex = pindex;
1725 atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
1726 mold = vm_radix_replace(&object->rtree, mnew);
1728 /* Keep the resident page list in sorted order. */
1729 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1730 TAILQ_REMOVE(&object->memq, mold, listq);
1732 mold->object = NULL;
1733 atomic_clear_int(&mold->ref_count, VPRC_OBJREF);
1734 vm_page_xunbusy(mold);
1737 * The object's resident_page_count does not change because we have
1738 * swapped one page for another, but the generation count should
1739 * change if the page is dirty.
1741 if (pmap_page_is_write_mapped(mnew))
1742 vm_object_set_writeable_dirty(object);
1749 * Move the given memory entry from its
1750 * current object to the specified target object/offset.
1752 * Note: swap associated with the page must be invalidated by the move. We
1753 * have to do this for several reasons: (1) we aren't freeing the
1754 * page, (2) we are dirtying the page, (3) the VM system is probably
1755 * moving the page from object A to B, and will then later move
1756 * the backing store from A to B and we can't have a conflict.
1758 * Note: we *always* dirty the page. It is necessary both for the
1759 * fact that we moved it, and because we may be invalidating
1762 * The objects must be locked.
1765 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1770 VM_OBJECT_ASSERT_WLOCKED(new_object);
1772 KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m));
1773 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1774 KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1775 ("vm_page_rename: pindex already renamed"));
1778 * Create a custom version of vm_page_insert() which does not depend
1779 * by m_prev and can cheat on the implementation aspects of the
1783 m->pindex = new_pindex;
1784 if (vm_radix_insert(&new_object->rtree, m)) {
1790 * The operation cannot fail anymore. The removal must happen before
1791 * the listq iterator is tainted.
1794 vm_page_object_remove(m);
1796 /* Return back to the new pindex to complete vm_page_insert(). */
1797 m->pindex = new_pindex;
1798 m->object = new_object;
1800 vm_page_insert_radixdone(m, new_object, mpred);
1808 * Allocate and return a page that is associated with the specified
1809 * object and offset pair. By default, this page is exclusive busied.
1811 * The caller must always specify an allocation class.
1813 * allocation classes:
1814 * VM_ALLOC_NORMAL normal process request
1815 * VM_ALLOC_SYSTEM system *really* needs a page
1816 * VM_ALLOC_INTERRUPT interrupt time request
1818 * optional allocation flags:
1819 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1820 * intends to allocate
1821 * VM_ALLOC_NOBUSY do not exclusive busy the page
1822 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1823 * VM_ALLOC_NOOBJ page is not associated with an object and
1824 * should not be exclusive busy
1825 * VM_ALLOC_SBUSY shared busy the allocated page
1826 * VM_ALLOC_WIRED wire the allocated page
1827 * VM_ALLOC_ZERO prefer a zeroed page
1830 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1833 return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1834 vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1838 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1842 return (vm_page_alloc_domain_after(object, pindex, domain, req,
1843 object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) :
1848 * Allocate a page in the specified object with the given page index. To
1849 * optimize insertion of the page into the object, the caller must also specifiy
1850 * the resident page in the object with largest index smaller than the given
1851 * page index, or NULL if no such page exists.
1854 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1855 int req, vm_page_t mpred)
1857 struct vm_domainset_iter di;
1861 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1863 m = vm_page_alloc_domain_after(object, pindex, domain, req,
1867 } while (vm_domainset_iter_page(&di, object, &domain) == 0);
1873 * Returns true if the number of free pages exceeds the minimum
1874 * for the request class and false otherwise.
1877 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
1879 u_int limit, old, new;
1881 if (req_class == VM_ALLOC_INTERRUPT)
1883 else if (req_class == VM_ALLOC_SYSTEM)
1884 limit = vmd->vmd_interrupt_free_min;
1886 limit = vmd->vmd_free_reserved;
1889 * Attempt to reserve the pages. Fail if we're below the limit.
1892 old = vmd->vmd_free_count;
1897 } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
1899 /* Wake the page daemon if we've crossed the threshold. */
1900 if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
1901 pagedaemon_wakeup(vmd->vmd_domain);
1903 /* Only update bitsets on transitions. */
1904 if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
1905 (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
1912 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
1917 * The page daemon is allowed to dig deeper into the free page list.
1919 req_class = req & VM_ALLOC_CLASS_MASK;
1920 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1921 req_class = VM_ALLOC_SYSTEM;
1922 return (_vm_domain_allocate(vmd, req_class, npages));
1926 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
1927 int req, vm_page_t mpred)
1929 struct vm_domain *vmd;
1933 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1934 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1935 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1936 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1937 ("inconsistent object(%p)/req(%x)", object, req));
1938 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
1939 ("Can't sleep and retry object insertion."));
1940 KASSERT(mpred == NULL || mpred->pindex < pindex,
1941 ("mpred %p doesn't precede pindex 0x%jx", mpred,
1942 (uintmax_t)pindex));
1944 VM_OBJECT_ASSERT_WLOCKED(object);
1948 pool = object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT;
1950 #if VM_NRESERVLEVEL > 0
1952 * Can we allocate the page from a reservation?
1954 if (vm_object_reserv(object) &&
1955 (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
1957 domain = vm_phys_domain(m);
1958 vmd = VM_DOMAIN(domain);
1962 vmd = VM_DOMAIN(domain);
1963 if (vmd->vmd_pgcache[pool].zone != NULL) {
1964 m = uma_zalloc(vmd->vmd_pgcache[pool].zone, M_NOWAIT);
1966 flags |= PG_PCPU_CACHE;
1970 if (vm_domain_allocate(vmd, req, 1)) {
1972 * If not, allocate it from the free page queues.
1974 vm_domain_free_lock(vmd);
1975 m = vm_phys_alloc_pages(domain, pool, 0);
1976 vm_domain_free_unlock(vmd);
1978 vm_domain_freecnt_inc(vmd, 1);
1979 #if VM_NRESERVLEVEL > 0
1980 if (vm_reserv_reclaim_inactive(domain))
1987 * Not allocatable, give up.
1989 if (vm_domain_alloc_fail(vmd, object, req))
1995 * At this point we had better have found a good page.
1999 vm_page_alloc_check(m);
2002 * Initialize the page. Only the PG_ZERO flag is inherited.
2004 if ((req & VM_ALLOC_ZERO) != 0)
2005 flags |= (m->flags & PG_ZERO);
2006 if ((req & VM_ALLOC_NODUMP) != 0)
2010 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2012 m->busy_lock = VPB_UNBUSIED;
2013 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2014 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2015 if ((req & VM_ALLOC_SBUSY) != 0)
2016 m->busy_lock = VPB_SHARERS_WORD(1);
2017 if (req & VM_ALLOC_WIRED) {
2019 * The page lock is not required for wiring a page until that
2020 * page is inserted into the object.
2027 if (object != NULL) {
2028 if (vm_page_insert_after(m, object, pindex, mpred)) {
2029 if (req & VM_ALLOC_WIRED) {
2033 KASSERT(m->object == NULL, ("page %p has object", m));
2034 m->oflags = VPO_UNMANAGED;
2035 m->busy_lock = VPB_UNBUSIED;
2036 /* Don't change PG_ZERO. */
2037 vm_page_free_toq(m);
2038 if (req & VM_ALLOC_WAITFAIL) {
2039 VM_OBJECT_WUNLOCK(object);
2041 VM_OBJECT_WLOCK(object);
2046 /* Ignore device objects; the pager sets "memattr" for them. */
2047 if (object->memattr != VM_MEMATTR_DEFAULT &&
2048 (object->flags & OBJ_FICTITIOUS) == 0)
2049 pmap_page_set_memattr(m, object->memattr);
2057 * vm_page_alloc_contig:
2059 * Allocate a contiguous set of physical pages of the given size "npages"
2060 * from the free lists. All of the physical pages must be at or above
2061 * the given physical address "low" and below the given physical address
2062 * "high". The given value "alignment" determines the alignment of the
2063 * first physical page in the set. If the given value "boundary" is
2064 * non-zero, then the set of physical pages cannot cross any physical
2065 * address boundary that is a multiple of that value. Both "alignment"
2066 * and "boundary" must be a power of two.
2068 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
2069 * then the memory attribute setting for the physical pages is configured
2070 * to the object's memory attribute setting. Otherwise, the memory
2071 * attribute setting for the physical pages is configured to "memattr",
2072 * overriding the object's memory attribute setting. However, if the
2073 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
2074 * memory attribute setting for the physical pages cannot be configured
2075 * to VM_MEMATTR_DEFAULT.
2077 * The specified object may not contain fictitious pages.
2079 * The caller must always specify an allocation class.
2081 * allocation classes:
2082 * VM_ALLOC_NORMAL normal process request
2083 * VM_ALLOC_SYSTEM system *really* needs a page
2084 * VM_ALLOC_INTERRUPT interrupt time request
2086 * optional allocation flags:
2087 * VM_ALLOC_NOBUSY do not exclusive busy the page
2088 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
2089 * VM_ALLOC_NOOBJ page is not associated with an object and
2090 * should not be exclusive busy
2091 * VM_ALLOC_SBUSY shared busy the allocated page
2092 * VM_ALLOC_WIRED wire the allocated page
2093 * VM_ALLOC_ZERO prefer a zeroed page
2096 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
2097 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2098 vm_paddr_t boundary, vm_memattr_t memattr)
2100 struct vm_domainset_iter di;
2104 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2106 m = vm_page_alloc_contig_domain(object, pindex, domain, req,
2107 npages, low, high, alignment, boundary, memattr);
2110 } while (vm_domainset_iter_page(&di, object, &domain) == 0);
2116 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2117 int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2118 vm_paddr_t boundary, vm_memattr_t memattr)
2120 struct vm_domain *vmd;
2121 vm_page_t m, m_ret, mpred;
2122 u_int busy_lock, flags, oflags;
2124 mpred = NULL; /* XXX: pacify gcc */
2125 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
2126 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
2127 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2128 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2129 ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
2131 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2132 ("Can't sleep and retry object insertion."));
2133 if (object != NULL) {
2134 VM_OBJECT_ASSERT_WLOCKED(object);
2135 KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2136 ("vm_page_alloc_contig: object %p has fictitious pages",
2139 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2141 if (object != NULL) {
2142 mpred = vm_radix_lookup_le(&object->rtree, pindex);
2143 KASSERT(mpred == NULL || mpred->pindex != pindex,
2144 ("vm_page_alloc_contig: pindex already allocated"));
2148 * Can we allocate the pages without the number of free pages falling
2149 * below the lower bound for the allocation class?
2153 #if VM_NRESERVLEVEL > 0
2155 * Can we allocate the pages from a reservation?
2157 if (vm_object_reserv(object) &&
2158 (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
2159 mpred, npages, low, high, alignment, boundary)) != NULL) {
2160 domain = vm_phys_domain(m_ret);
2161 vmd = VM_DOMAIN(domain);
2165 vmd = VM_DOMAIN(domain);
2166 if (vm_domain_allocate(vmd, req, npages)) {
2168 * allocate them from the free page queues.
2170 vm_domain_free_lock(vmd);
2171 m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2172 alignment, boundary);
2173 vm_domain_free_unlock(vmd);
2174 if (m_ret == NULL) {
2175 vm_domain_freecnt_inc(vmd, npages);
2176 #if VM_NRESERVLEVEL > 0
2177 if (vm_reserv_reclaim_contig(domain, npages, low,
2178 high, alignment, boundary))
2183 if (m_ret == NULL) {
2184 if (vm_domain_alloc_fail(vmd, object, req))
2188 #if VM_NRESERVLEVEL > 0
2191 for (m = m_ret; m < &m_ret[npages]; m++) {
2193 vm_page_alloc_check(m);
2197 * Initialize the pages. Only the PG_ZERO flag is inherited.
2200 if ((req & VM_ALLOC_ZERO) != 0)
2202 if ((req & VM_ALLOC_NODUMP) != 0)
2204 oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2206 busy_lock = VPB_UNBUSIED;
2207 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2208 busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2209 if ((req & VM_ALLOC_SBUSY) != 0)
2210 busy_lock = VPB_SHARERS_WORD(1);
2211 if ((req & VM_ALLOC_WIRED) != 0)
2212 vm_wire_add(npages);
2213 if (object != NULL) {
2214 if (object->memattr != VM_MEMATTR_DEFAULT &&
2215 memattr == VM_MEMATTR_DEFAULT)
2216 memattr = object->memattr;
2218 for (m = m_ret; m < &m_ret[npages]; m++) {
2220 m->flags = (m->flags | PG_NODUMP) & flags;
2221 m->busy_lock = busy_lock;
2222 if ((req & VM_ALLOC_WIRED) != 0)
2226 if (object != NULL) {
2227 if (vm_page_insert_after(m, object, pindex, mpred)) {
2228 if ((req & VM_ALLOC_WIRED) != 0)
2229 vm_wire_sub(npages);
2230 KASSERT(m->object == NULL,
2231 ("page %p has object", m));
2233 for (m = m_ret; m < &m_ret[npages]; m++) {
2235 (req & VM_ALLOC_WIRED) != 0)
2237 m->oflags = VPO_UNMANAGED;
2238 m->busy_lock = VPB_UNBUSIED;
2239 /* Don't change PG_ZERO. */
2240 vm_page_free_toq(m);
2242 if (req & VM_ALLOC_WAITFAIL) {
2243 VM_OBJECT_WUNLOCK(object);
2245 VM_OBJECT_WLOCK(object);
2252 if (memattr != VM_MEMATTR_DEFAULT)
2253 pmap_page_set_memattr(m, memattr);
2260 * Check a page that has been freshly dequeued from a freelist.
2263 vm_page_alloc_check(vm_page_t m)
2266 KASSERT(m->object == NULL, ("page %p has object", m));
2267 KASSERT(m->a.queue == PQ_NONE &&
2268 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
2269 ("page %p has unexpected queue %d, flags %#x",
2270 m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK)));
2271 KASSERT(m->ref_count == 0, ("page %p has references", m));
2272 KASSERT(!vm_page_busied(m), ("page %p is busy", m));
2273 KASSERT(m->dirty == 0, ("page %p is dirty", m));
2274 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2275 ("page %p has unexpected memattr %d",
2276 m, pmap_page_get_memattr(m)));
2277 KASSERT(m->valid == 0, ("free page %p is valid", m));
2281 * vm_page_alloc_freelist:
2283 * Allocate a physical page from the specified free page list.
2285 * The caller must always specify an allocation class.
2287 * allocation classes:
2288 * VM_ALLOC_NORMAL normal process request
2289 * VM_ALLOC_SYSTEM system *really* needs a page
2290 * VM_ALLOC_INTERRUPT interrupt time request
2292 * optional allocation flags:
2293 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
2294 * intends to allocate
2295 * VM_ALLOC_WIRED wire the allocated page
2296 * VM_ALLOC_ZERO prefer a zeroed page
2299 vm_page_alloc_freelist(int freelist, int req)
2301 struct vm_domainset_iter di;
2305 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2307 m = vm_page_alloc_freelist_domain(domain, freelist, req);
2310 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2316 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2318 struct vm_domain *vmd;
2323 vmd = VM_DOMAIN(domain);
2325 if (vm_domain_allocate(vmd, req, 1)) {
2326 vm_domain_free_lock(vmd);
2327 m = vm_phys_alloc_freelist_pages(domain, freelist,
2328 VM_FREEPOOL_DIRECT, 0);
2329 vm_domain_free_unlock(vmd);
2331 vm_domain_freecnt_inc(vmd, 1);
2334 if (vm_domain_alloc_fail(vmd, NULL, req))
2339 vm_page_alloc_check(m);
2342 * Initialize the page. Only the PG_ZERO flag is inherited.
2346 if ((req & VM_ALLOC_ZERO) != 0)
2349 if ((req & VM_ALLOC_WIRED) != 0) {
2351 * The page lock is not required for wiring a page that does
2352 * not belong to an object.
2357 /* Unmanaged pages don't use "act_count". */
2358 m->oflags = VPO_UNMANAGED;
2363 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2365 struct vm_domain *vmd;
2366 struct vm_pgcache *pgcache;
2370 vmd = VM_DOMAIN(pgcache->domain);
2373 * The page daemon should avoid creating extra memory pressure since its
2374 * main purpose is to replenish the store of free pages.
2376 if (vmd->vmd_severeset || curproc == pageproc ||
2377 !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2379 domain = vmd->vmd_domain;
2380 vm_domain_free_lock(vmd);
2381 i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2382 (vm_page_t *)store);
2383 vm_domain_free_unlock(vmd);
2385 vm_domain_freecnt_inc(vmd, cnt - i);
2391 vm_page_zone_release(void *arg, void **store, int cnt)
2393 struct vm_domain *vmd;
2394 struct vm_pgcache *pgcache;
2399 vmd = VM_DOMAIN(pgcache->domain);
2400 vm_domain_free_lock(vmd);
2401 for (i = 0; i < cnt; i++) {
2402 m = (vm_page_t)store[i];
2403 vm_phys_free_pages(m, 0);
2405 vm_domain_free_unlock(vmd);
2406 vm_domain_freecnt_inc(vmd, cnt);
2409 #define VPSC_ANY 0 /* No restrictions. */
2410 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */
2411 #define VPSC_NOSUPER 2 /* Skip superpages. */
2414 * vm_page_scan_contig:
2416 * Scan vm_page_array[] between the specified entries "m_start" and
2417 * "m_end" for a run of contiguous physical pages that satisfy the
2418 * specified conditions, and return the lowest page in the run. The
2419 * specified "alignment" determines the alignment of the lowest physical
2420 * page in the run. If the specified "boundary" is non-zero, then the
2421 * run of physical pages cannot span a physical address that is a
2422 * multiple of "boundary".
2424 * "m_end" is never dereferenced, so it need not point to a vm_page
2425 * structure within vm_page_array[].
2427 * "npages" must be greater than zero. "m_start" and "m_end" must not
2428 * span a hole (or discontiguity) in the physical address space. Both
2429 * "alignment" and "boundary" must be a power of two.
2432 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2433 u_long alignment, vm_paddr_t boundary, int options)
2439 #if VM_NRESERVLEVEL > 0
2442 int m_inc, order, run_ext, run_len;
2444 KASSERT(npages > 0, ("npages is 0"));
2445 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2446 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2450 for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2451 KASSERT((m->flags & PG_MARKER) == 0,
2452 ("page %p is PG_MARKER", m));
2453 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1,
2454 ("fictitious page %p has invalid ref count", m));
2457 * If the current page would be the start of a run, check its
2458 * physical address against the end, alignment, and boundary
2459 * conditions. If it doesn't satisfy these conditions, either
2460 * terminate the scan or advance to the next page that
2461 * satisfies the failed condition.
2464 KASSERT(m_run == NULL, ("m_run != NULL"));
2465 if (m + npages > m_end)
2467 pa = VM_PAGE_TO_PHYS(m);
2468 if ((pa & (alignment - 1)) != 0) {
2469 m_inc = atop(roundup2(pa, alignment) - pa);
2472 if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2474 m_inc = atop(roundup2(pa, boundary) - pa);
2478 KASSERT(m_run != NULL, ("m_run == NULL"));
2480 vm_page_change_lock(m, &m_mtx);
2483 if (vm_page_wired(m))
2485 #if VM_NRESERVLEVEL > 0
2486 else if ((level = vm_reserv_level(m)) >= 0 &&
2487 (options & VPSC_NORESERV) != 0) {
2489 /* Advance to the end of the reservation. */
2490 pa = VM_PAGE_TO_PHYS(m);
2491 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2495 else if ((object = m->object) != NULL) {
2497 * The page is considered eligible for relocation if
2498 * and only if it could be laundered or reclaimed by
2501 if (!VM_OBJECT_TRYRLOCK(object)) {
2503 VM_OBJECT_RLOCK(object);
2505 if (m->object != object) {
2507 * The page may have been freed.
2509 VM_OBJECT_RUNLOCK(object);
2513 /* Don't care: PG_NODUMP, PG_ZERO. */
2514 if (object->type != OBJT_DEFAULT &&
2515 object->type != OBJT_SWAP &&
2516 object->type != OBJT_VNODE) {
2518 #if VM_NRESERVLEVEL > 0
2519 } else if ((options & VPSC_NOSUPER) != 0 &&
2520 (level = vm_reserv_level_iffullpop(m)) >= 0) {
2522 /* Advance to the end of the superpage. */
2523 pa = VM_PAGE_TO_PHYS(m);
2524 m_inc = atop(roundup2(pa + 1,
2525 vm_reserv_size(level)) - pa);
2527 } else if (object->memattr == VM_MEMATTR_DEFAULT &&
2528 vm_page_queue(m) != PQ_NONE && !vm_page_busied(m) &&
2529 !vm_page_wired(m)) {
2531 * The page is allocated but eligible for
2532 * relocation. Extend the current run by one
2535 KASSERT(pmap_page_get_memattr(m) ==
2537 ("page %p has an unexpected memattr", m));
2538 KASSERT((m->oflags & (VPO_SWAPINPROG |
2539 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2540 ("page %p has unexpected oflags", m));
2541 /* Don't care: PGA_NOSYNC. */
2545 VM_OBJECT_RUNLOCK(object);
2546 #if VM_NRESERVLEVEL > 0
2547 } else if (level >= 0) {
2549 * The page is reserved but not yet allocated. In
2550 * other words, it is still free. Extend the current
2555 } else if ((order = m->order) < VM_NFREEORDER) {
2557 * The page is enqueued in the physical memory
2558 * allocator's free page queues. Moreover, it is the
2559 * first page in a power-of-two-sized run of
2560 * contiguous free pages. Add these pages to the end
2561 * of the current run, and jump ahead.
2563 run_ext = 1 << order;
2567 * Skip the page for one of the following reasons: (1)
2568 * It is enqueued in the physical memory allocator's
2569 * free page queues. However, it is not the first
2570 * page in a run of contiguous free pages. (This case
2571 * rarely occurs because the scan is performed in
2572 * ascending order.) (2) It is not reserved, and it is
2573 * transitioning from free to allocated. (Conversely,
2574 * the transition from allocated to free for managed
2575 * pages is blocked by the page lock.) (3) It is
2576 * allocated but not contained by an object and not
2577 * wired, e.g., allocated by Xen's balloon driver.
2583 * Extend or reset the current run of pages.
2598 if (run_len >= npages)
2604 * vm_page_reclaim_run:
2606 * Try to relocate each of the allocated virtual pages within the
2607 * specified run of physical pages to a new physical address. Free the
2608 * physical pages underlying the relocated virtual pages. A virtual page
2609 * is relocatable if and only if it could be laundered or reclaimed by
2610 * the page daemon. Whenever possible, a virtual page is relocated to a
2611 * physical address above "high".
2613 * Returns 0 if every physical page within the run was already free or
2614 * just freed by a successful relocation. Otherwise, returns a non-zero
2615 * value indicating why the last attempt to relocate a virtual page was
2618 * "req_class" must be an allocation class.
2621 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2624 struct vm_domain *vmd;
2626 struct spglist free;
2629 vm_page_t m, m_end, m_new;
2630 int error, order, req;
2632 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2633 ("req_class is not an allocation class"));
2637 m_end = m_run + npages;
2639 for (; error == 0 && m < m_end; m++) {
2640 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2641 ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2644 * Avoid releasing and reacquiring the same page lock.
2646 vm_page_change_lock(m, &m_mtx);
2649 * Racily check for wirings. Races are handled below.
2651 if (vm_page_wired(m))
2653 else if ((object = m->object) != NULL) {
2655 * The page is relocated if and only if it could be
2656 * laundered or reclaimed by the page daemon.
2658 if (!VM_OBJECT_TRYWLOCK(object)) {
2660 VM_OBJECT_WLOCK(object);
2662 if (m->object != object) {
2664 * The page may have been freed.
2666 VM_OBJECT_WUNLOCK(object);
2670 /* Don't care: PG_NODUMP, PG_ZERO. */
2671 if (object->type != OBJT_DEFAULT &&
2672 object->type != OBJT_SWAP &&
2673 object->type != OBJT_VNODE)
2675 else if (object->memattr != VM_MEMATTR_DEFAULT)
2677 else if (vm_page_queue(m) != PQ_NONE &&
2678 vm_page_tryxbusy(m) != 0) {
2679 if (vm_page_wired(m)) {
2684 KASSERT(pmap_page_get_memattr(m) ==
2686 ("page %p has an unexpected memattr", m));
2687 KASSERT((m->oflags & (VPO_SWAPINPROG |
2688 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2689 ("page %p has unexpected oflags", m));
2690 /* Don't care: PGA_NOSYNC. */
2691 if (!vm_page_none_valid(m)) {
2693 * First, try to allocate a new page
2694 * that is above "high". Failing
2695 * that, try to allocate a new page
2696 * that is below "m_run". Allocate
2697 * the new page between the end of
2698 * "m_run" and "high" only as a last
2701 req = req_class | VM_ALLOC_NOOBJ;
2702 if ((m->flags & PG_NODUMP) != 0)
2703 req |= VM_ALLOC_NODUMP;
2704 if (trunc_page(high) !=
2705 ~(vm_paddr_t)PAGE_MASK) {
2706 m_new = vm_page_alloc_contig(
2711 VM_MEMATTR_DEFAULT);
2714 if (m_new == NULL) {
2715 pa = VM_PAGE_TO_PHYS(m_run);
2716 m_new = vm_page_alloc_contig(
2718 0, pa - 1, PAGE_SIZE, 0,
2719 VM_MEMATTR_DEFAULT);
2721 if (m_new == NULL) {
2723 m_new = vm_page_alloc_contig(
2725 pa, high, PAGE_SIZE, 0,
2726 VM_MEMATTR_DEFAULT);
2728 if (m_new == NULL) {
2735 * Unmap the page and check for new
2736 * wirings that may have been acquired
2737 * through a pmap lookup.
2739 if (object->ref_count != 0 &&
2740 !vm_page_try_remove_all(m)) {
2741 vm_page_free(m_new);
2747 * Replace "m" with the new page. For
2748 * vm_page_replace(), "m" must be busy
2749 * and dequeued. Finally, change "m"
2750 * as if vm_page_free() was called.
2752 m_new->a.flags = m->a.flags &
2753 ~PGA_QUEUE_STATE_MASK;
2754 KASSERT(m_new->oflags == VPO_UNMANAGED,
2755 ("page %p is managed", m_new));
2756 pmap_copy_page(m, m_new);
2757 m_new->valid = m->valid;
2758 m_new->dirty = m->dirty;
2759 m->flags &= ~PG_ZERO;
2761 vm_page_replace_checked(m_new, object,
2763 if (vm_page_free_prep(m))
2764 SLIST_INSERT_HEAD(&free, m,
2768 * The new page must be deactivated
2769 * before the object is unlocked.
2771 vm_page_change_lock(m_new, &m_mtx);
2772 vm_page_deactivate(m_new);
2774 m->flags &= ~PG_ZERO;
2776 if (vm_page_free_prep(m))
2777 SLIST_INSERT_HEAD(&free, m,
2779 KASSERT(m->dirty == 0,
2780 ("page %p is dirty", m));
2785 VM_OBJECT_WUNLOCK(object);
2787 MPASS(vm_phys_domain(m) == domain);
2788 vmd = VM_DOMAIN(domain);
2789 vm_domain_free_lock(vmd);
2791 if (order < VM_NFREEORDER) {
2793 * The page is enqueued in the physical memory
2794 * allocator's free page queues. Moreover, it
2795 * is the first page in a power-of-two-sized
2796 * run of contiguous free pages. Jump ahead
2797 * to the last page within that run, and
2798 * continue from there.
2800 m += (1 << order) - 1;
2802 #if VM_NRESERVLEVEL > 0
2803 else if (vm_reserv_is_page_free(m))
2806 vm_domain_free_unlock(vmd);
2807 if (order == VM_NFREEORDER)
2813 if ((m = SLIST_FIRST(&free)) != NULL) {
2816 vmd = VM_DOMAIN(domain);
2818 vm_domain_free_lock(vmd);
2820 MPASS(vm_phys_domain(m) == domain);
2821 SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2822 vm_phys_free_pages(m, 0);
2824 } while ((m = SLIST_FIRST(&free)) != NULL);
2825 vm_domain_free_unlock(vmd);
2826 vm_domain_freecnt_inc(vmd, cnt);
2833 CTASSERT(powerof2(NRUNS));
2835 #define RUN_INDEX(count) ((count) & (NRUNS - 1))
2837 #define MIN_RECLAIM 8
2840 * vm_page_reclaim_contig:
2842 * Reclaim allocated, contiguous physical memory satisfying the specified
2843 * conditions by relocating the virtual pages using that physical memory.
2844 * Returns true if reclamation is successful and false otherwise. Since
2845 * relocation requires the allocation of physical pages, reclamation may
2846 * fail due to a shortage of free pages. When reclamation fails, callers
2847 * are expected to perform vm_wait() before retrying a failed allocation
2848 * operation, e.g., vm_page_alloc_contig().
2850 * The caller must always specify an allocation class through "req".
2852 * allocation classes:
2853 * VM_ALLOC_NORMAL normal process request
2854 * VM_ALLOC_SYSTEM system *really* needs a page
2855 * VM_ALLOC_INTERRUPT interrupt time request
2857 * The optional allocation flags are ignored.
2859 * "npages" must be greater than zero. Both "alignment" and "boundary"
2860 * must be a power of two.
2863 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
2864 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2866 struct vm_domain *vmd;
2867 vm_paddr_t curr_low;
2868 vm_page_t m_run, m_runs[NRUNS];
2869 u_long count, reclaimed;
2870 int error, i, options, req_class;
2872 KASSERT(npages > 0, ("npages is 0"));
2873 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2874 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2875 req_class = req & VM_ALLOC_CLASS_MASK;
2878 * The page daemon is allowed to dig deeper into the free page list.
2880 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2881 req_class = VM_ALLOC_SYSTEM;
2884 * Return if the number of free pages cannot satisfy the requested
2887 vmd = VM_DOMAIN(domain);
2888 count = vmd->vmd_free_count;
2889 if (count < npages + vmd->vmd_free_reserved || (count < npages +
2890 vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2891 (count < npages && req_class == VM_ALLOC_INTERRUPT))
2895 * Scan up to three times, relaxing the restrictions ("options") on
2896 * the reclamation of reservations and superpages each time.
2898 for (options = VPSC_NORESERV;;) {
2900 * Find the highest runs that satisfy the given constraints
2901 * and restrictions, and record them in "m_runs".
2906 m_run = vm_phys_scan_contig(domain, npages, curr_low,
2907 high, alignment, boundary, options);
2910 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2911 m_runs[RUN_INDEX(count)] = m_run;
2916 * Reclaim the highest runs in LIFO (descending) order until
2917 * the number of reclaimed pages, "reclaimed", is at least
2918 * MIN_RECLAIM. Reset "reclaimed" each time because each
2919 * reclamation is idempotent, and runs will (likely) recur
2920 * from one scan to the next as restrictions are relaxed.
2923 for (i = 0; count > 0 && i < NRUNS; i++) {
2925 m_run = m_runs[RUN_INDEX(count)];
2926 error = vm_page_reclaim_run(req_class, domain, npages,
2929 reclaimed += npages;
2930 if (reclaimed >= MIN_RECLAIM)
2936 * Either relax the restrictions on the next scan or return if
2937 * the last scan had no restrictions.
2939 if (options == VPSC_NORESERV)
2940 options = VPSC_NOSUPER;
2941 else if (options == VPSC_NOSUPER)
2943 else if (options == VPSC_ANY)
2944 return (reclaimed != 0);
2949 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
2950 u_long alignment, vm_paddr_t boundary)
2952 struct vm_domainset_iter di;
2956 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2958 ret = vm_page_reclaim_contig_domain(domain, req, npages, low,
2959 high, alignment, boundary);
2962 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2968 * Set the domain in the appropriate page level domainset.
2971 vm_domain_set(struct vm_domain *vmd)
2974 mtx_lock(&vm_domainset_lock);
2975 if (!vmd->vmd_minset && vm_paging_min(vmd)) {
2976 vmd->vmd_minset = 1;
2977 DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
2979 if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
2980 vmd->vmd_severeset = 1;
2981 DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
2983 mtx_unlock(&vm_domainset_lock);
2987 * Clear the domain from the appropriate page level domainset.
2990 vm_domain_clear(struct vm_domain *vmd)
2993 mtx_lock(&vm_domainset_lock);
2994 if (vmd->vmd_minset && !vm_paging_min(vmd)) {
2995 vmd->vmd_minset = 0;
2996 DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
2997 if (vm_min_waiters != 0) {
2999 wakeup(&vm_min_domains);
3002 if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
3003 vmd->vmd_severeset = 0;
3004 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
3005 if (vm_severe_waiters != 0) {
3006 vm_severe_waiters = 0;
3007 wakeup(&vm_severe_domains);
3012 * If pageout daemon needs pages, then tell it that there are
3015 if (vmd->vmd_pageout_pages_needed &&
3016 vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
3017 wakeup(&vmd->vmd_pageout_pages_needed);
3018 vmd->vmd_pageout_pages_needed = 0;
3021 /* See comments in vm_wait_doms(). */
3022 if (vm_pageproc_waiters) {
3023 vm_pageproc_waiters = 0;
3024 wakeup(&vm_pageproc_waiters);
3026 mtx_unlock(&vm_domainset_lock);
3030 * Wait for free pages to exceed the min threshold globally.
3036 mtx_lock(&vm_domainset_lock);
3037 while (vm_page_count_min()) {
3039 msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
3041 mtx_unlock(&vm_domainset_lock);
3045 * Wait for free pages to exceed the severe threshold globally.
3048 vm_wait_severe(void)
3051 mtx_lock(&vm_domainset_lock);
3052 while (vm_page_count_severe()) {
3053 vm_severe_waiters++;
3054 msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
3057 mtx_unlock(&vm_domainset_lock);
3064 return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
3068 vm_wait_doms(const domainset_t *wdoms)
3072 * We use racey wakeup synchronization to avoid expensive global
3073 * locking for the pageproc when sleeping with a non-specific vm_wait.
3074 * To handle this, we only sleep for one tick in this instance. It
3075 * is expected that most allocations for the pageproc will come from
3076 * kmem or vm_page_grab* which will use the more specific and
3077 * race-free vm_wait_domain().
3079 if (curproc == pageproc) {
3080 mtx_lock(&vm_domainset_lock);
3081 vm_pageproc_waiters++;
3082 msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP,
3086 * XXX Ideally we would wait only until the allocation could
3087 * be satisfied. This condition can cause new allocators to
3088 * consume all freed pages while old allocators wait.
3090 mtx_lock(&vm_domainset_lock);
3091 if (vm_page_count_min_set(wdoms)) {
3093 msleep(&vm_min_domains, &vm_domainset_lock,
3094 PVM | PDROP, "vmwait", 0);
3096 mtx_unlock(&vm_domainset_lock);
3103 * Sleep until free pages are available for allocation.
3104 * - Called in various places after failed memory allocations.
3107 vm_wait_domain(int domain)
3109 struct vm_domain *vmd;
3112 vmd = VM_DOMAIN(domain);
3113 vm_domain_free_assert_unlocked(vmd);
3115 if (curproc == pageproc) {
3116 mtx_lock(&vm_domainset_lock);
3117 if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3118 vmd->vmd_pageout_pages_needed = 1;
3119 msleep(&vmd->vmd_pageout_pages_needed,
3120 &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3122 mtx_unlock(&vm_domainset_lock);
3124 if (pageproc == NULL)
3125 panic("vm_wait in early boot");
3126 DOMAINSET_ZERO(&wdom);
3127 DOMAINSET_SET(vmd->vmd_domain, &wdom);
3128 vm_wait_doms(&wdom);
3135 * Sleep until free pages are available for allocation in the
3136 * affinity domains of the obj. If obj is NULL, the domain set
3137 * for the calling thread is used.
3138 * Called in various places after failed memory allocations.
3141 vm_wait(vm_object_t obj)
3143 struct domainset *d;
3148 * Carefully fetch pointers only once: the struct domainset
3149 * itself is ummutable but the pointer might change.
3152 d = obj->domain.dr_policy;
3154 d = curthread->td_domain.dr_policy;
3156 vm_wait_doms(&d->ds_mask);
3160 * vm_domain_alloc_fail:
3162 * Called when a page allocation function fails. Informs the
3163 * pagedaemon and performs the requested wait. Requires the
3164 * domain_free and object lock on entry. Returns with the
3165 * object lock held and free lock released. Returns an error when
3166 * retry is necessary.
3170 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3173 vm_domain_free_assert_unlocked(vmd);
3175 atomic_add_int(&vmd->vmd_pageout_deficit,
3176 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3177 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3179 VM_OBJECT_WUNLOCK(object);
3180 vm_wait_domain(vmd->vmd_domain);
3182 VM_OBJECT_WLOCK(object);
3183 if (req & VM_ALLOC_WAITOK)
3193 * Sleep until free pages are available for allocation.
3194 * - Called only in vm_fault so that processes page faulting
3195 * can be easily tracked.
3196 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3197 * processes will be able to grab memory first. Do not change
3198 * this balance without careful testing first.
3201 vm_waitpfault(struct domainset *dset, int timo)
3205 * XXX Ideally we would wait only until the allocation could
3206 * be satisfied. This condition can cause new allocators to
3207 * consume all freed pages while old allocators wait.
3209 mtx_lock(&vm_domainset_lock);
3210 if (vm_page_count_min_set(&dset->ds_mask)) {
3212 msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3215 mtx_unlock(&vm_domainset_lock);
3218 static struct vm_pagequeue *
3219 vm_page_pagequeue(vm_page_t m)
3224 if ((queue = atomic_load_8(&m->a.queue)) == PQ_NONE)
3226 return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3230 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m)
3232 struct vm_domain *vmd;
3235 CRITICAL_ASSERT(curthread);
3236 vm_pagequeue_assert_locked(pq);
3239 * The page daemon is allowed to set m->a.queue = PQ_NONE without
3240 * the page queue lock held. In this case it is about to free the page,
3241 * which must not have any queue state.
3243 qflags = atomic_load_16(&m->a.flags);
3244 KASSERT(pq == vm_page_pagequeue(m) ||
3245 (qflags & PGA_QUEUE_STATE_MASK) == 0,
3246 ("page %p doesn't belong to queue %p but has aflags %#x",
3249 if ((qflags & PGA_DEQUEUE) != 0) {
3250 if (__predict_true((qflags & PGA_ENQUEUED) != 0))
3251 vm_pagequeue_remove(pq, m);
3252 vm_page_dequeue_complete(m);
3253 counter_u64_add(queue_ops, 1);
3254 } else if ((qflags & (PGA_REQUEUE | PGA_REQUEUE_HEAD)) != 0) {
3255 if ((qflags & PGA_ENQUEUED) != 0)
3256 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3258 vm_pagequeue_cnt_inc(pq);
3259 vm_page_aflag_set(m, PGA_ENQUEUED);
3263 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE.
3264 * In particular, if both flags are set in close succession,
3265 * only PGA_REQUEUE_HEAD will be applied, even if it was set
3268 if ((qflags & PGA_REQUEUE_HEAD) != 0) {
3269 KASSERT(m->a.queue == PQ_INACTIVE,
3270 ("head enqueue not supported for page %p", m));
3271 vmd = vm_pagequeue_domain(m);
3272 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3274 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3276 vm_page_aflag_clear(m, qflags & (PGA_REQUEUE |
3278 counter_u64_add(queue_ops, 1);
3280 counter_u64_add(queue_nops, 1);
3285 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3291 for (i = 0; i < bq->bq_cnt; i++) {
3293 if (__predict_false(m->a.queue != queue))
3295 vm_pqbatch_process_page(pq, m);
3297 vm_batchqueue_init(bq);
3301 * vm_page_pqbatch_submit: [ internal use only ]
3303 * Enqueue a page in the specified page queue's batched work queue.
3304 * The caller must have encoded the requested operation in the page
3305 * structure's a.flags field.
3308 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
3310 struct vm_batchqueue *bq;
3311 struct vm_pagequeue *pq;
3314 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3315 ("page %p is unmanaged", m));
3316 KASSERT(mtx_owned(vm_page_lockptr(m)) || m->object == NULL,
3317 ("missing synchronization for page %p", m));
3318 KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3320 domain = vm_phys_domain(m);
3321 pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue];
3324 bq = DPCPU_PTR(pqbatch[domain][queue]);
3325 if (vm_batchqueue_insert(bq, m)) {
3330 vm_pagequeue_lock(pq);
3332 bq = DPCPU_PTR(pqbatch[domain][queue]);
3333 vm_pqbatch_process(pq, bq, queue);
3336 * The page may have been logically dequeued before we acquired the
3337 * page queue lock. In this case, since we either hold the page lock
3338 * or the page is being freed, a different thread cannot be concurrently
3339 * enqueuing the page.
3341 if (__predict_true(m->a.queue == queue))
3342 vm_pqbatch_process_page(pq, m);
3344 KASSERT(m->a.queue == PQ_NONE,
3345 ("invalid queue transition for page %p", m));
3346 KASSERT((m->a.flags & PGA_ENQUEUED) == 0,
3347 ("page %p is enqueued with invalid queue index", m));
3349 vm_pagequeue_unlock(pq);
3354 * vm_page_pqbatch_drain: [ internal use only ]
3356 * Force all per-CPU page queue batch queues to be drained. This is
3357 * intended for use in severe memory shortages, to ensure that pages
3358 * do not remain stuck in the batch queues.
3361 vm_page_pqbatch_drain(void)
3364 struct vm_domain *vmd;
3365 struct vm_pagequeue *pq;
3366 int cpu, domain, queue;
3371 sched_bind(td, cpu);
3374 for (domain = 0; domain < vm_ndomains; domain++) {
3375 vmd = VM_DOMAIN(domain);
3376 for (queue = 0; queue < PQ_COUNT; queue++) {
3377 pq = &vmd->vmd_pagequeues[queue];
3378 vm_pagequeue_lock(pq);
3380 vm_pqbatch_process(pq,
3381 DPCPU_PTR(pqbatch[domain][queue]), queue);
3383 vm_pagequeue_unlock(pq);
3393 * Complete the logical removal of a page from a page queue. We must be
3394 * careful to synchronize with the page daemon, which may be concurrently
3395 * examining the page with only the page lock held. The page must not be
3396 * in a state where it appears to be logically enqueued.
3399 vm_page_dequeue_complete(vm_page_t m)
3402 m->a.queue = PQ_NONE;
3403 atomic_thread_fence_rel();
3404 vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK);
3408 * vm_page_dequeue_deferred: [ internal use only ]
3410 * Request removal of the given page from its current page
3411 * queue. Physical removal from the queue may be deferred
3414 * The page must be locked.
3417 vm_page_dequeue_deferred(vm_page_t m)
3421 vm_page_assert_locked(m);
3423 if ((queue = vm_page_queue(m)) == PQ_NONE)
3427 * Set PGA_DEQUEUE if it is not already set to handle a concurrent call
3428 * to vm_page_dequeue_deferred_free(). In particular, avoid modifying
3429 * the page's queue state once vm_page_dequeue_deferred_free() has been
3430 * called. In the event of a race, two batch queue entries for the page
3431 * will be created, but the second will have no effect.
3433 if (vm_page_pqstate_cmpset(m, queue, queue, PGA_DEQUEUE, PGA_DEQUEUE))
3434 vm_page_pqbatch_submit(m, queue);
3438 * A variant of vm_page_dequeue_deferred() that does not assert the page
3439 * lock and is only to be called from vm_page_free_prep(). Because the
3440 * page is being freed, we can assume that nothing other than the page
3441 * daemon is scheduling queue operations on this page, so we get for
3442 * free the mutual exclusion that is otherwise provided by the page lock.
3443 * To handle races, the page daemon must take care to atomically check
3444 * for PGA_DEQUEUE when updating queue state.
3447 vm_page_dequeue_deferred_free(vm_page_t m)
3451 KASSERT(m->ref_count == 0, ("page %p has references", m));
3454 if ((m->a.flags & PGA_DEQUEUE) != 0)
3456 atomic_thread_fence_acq();
3457 if ((queue = atomic_load_8(&m->a.queue)) == PQ_NONE)
3459 if (vm_page_pqstate_cmpset(m, queue, queue, PGA_DEQUEUE,
3461 vm_page_pqbatch_submit(m, queue);
3470 * Remove the page from whichever page queue it's in, if any.
3471 * The page must either be locked or unallocated. This constraint
3472 * ensures that the queue state of the page will remain consistent
3473 * after this function returns.
3476 vm_page_dequeue(vm_page_t m)
3478 struct vm_pagequeue *pq, *pq1;
3481 KASSERT(mtx_owned(vm_page_lockptr(m)) || m->ref_count == 0,
3482 ("page %p is allocated and unlocked", m));
3484 for (pq = vm_page_pagequeue(m);; pq = pq1) {
3487 * A thread may be concurrently executing
3488 * vm_page_dequeue_complete(). Ensure that all queue
3489 * state is cleared before we return.
3491 aflags = atomic_load_16(&m->a.flags);
3492 if ((aflags & PGA_QUEUE_STATE_MASK) == 0)
3494 KASSERT((aflags & PGA_DEQUEUE) != 0,
3495 ("page %p has unexpected queue state flags %#x",
3499 * Busy wait until the thread updating queue state is
3500 * finished. Such a thread must be executing in a
3504 pq1 = vm_page_pagequeue(m);
3507 vm_pagequeue_lock(pq);
3508 if ((pq1 = vm_page_pagequeue(m)) == pq)
3510 vm_pagequeue_unlock(pq);
3512 KASSERT(pq == vm_page_pagequeue(m),
3513 ("%s: page %p migrated directly between queues", __func__, m));
3514 KASSERT((m->a.flags & PGA_DEQUEUE) != 0 ||
3515 mtx_owned(vm_page_lockptr(m)),
3516 ("%s: queued unlocked page %p", __func__, m));
3518 if ((m->a.flags & PGA_ENQUEUED) != 0)
3519 vm_pagequeue_remove(pq, m);
3520 vm_page_dequeue_complete(m);
3521 vm_pagequeue_unlock(pq);
3525 * Schedule the given page for insertion into the specified page queue.
3526 * Physical insertion of the page may be deferred indefinitely.
3529 vm_page_enqueue(vm_page_t m, uint8_t queue)
3532 vm_page_assert_locked(m);
3533 KASSERT(m->a.queue == PQ_NONE &&
3534 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
3535 ("%s: page %p is already enqueued", __func__, m));
3536 KASSERT(m->ref_count > 0,
3537 ("%s: page %p does not carry any references", __func__, m));
3540 if ((m->a.flags & PGA_REQUEUE) == 0)
3541 vm_page_aflag_set(m, PGA_REQUEUE);
3542 vm_page_pqbatch_submit(m, queue);
3546 * vm_page_requeue: [ internal use only ]
3548 * Schedule a requeue of the given page.
3550 * The page must be locked.
3553 vm_page_requeue(vm_page_t m)
3556 vm_page_assert_locked(m);
3557 KASSERT(vm_page_queue(m) != PQ_NONE,
3558 ("%s: page %p is not logically enqueued", __func__, m));
3559 KASSERT(m->ref_count > 0,
3560 ("%s: page %p does not carry any references", __func__, m));
3562 if ((m->a.flags & PGA_REQUEUE) == 0)
3563 vm_page_aflag_set(m, PGA_REQUEUE);
3564 vm_page_pqbatch_submit(m, atomic_load_8(&m->a.queue));
3568 * vm_page_swapqueue: [ internal use only ]
3570 * Move the page from one queue to another, or to the tail of its
3571 * current queue, in the face of a possible concurrent call to
3572 * vm_page_dequeue_deferred_free().
3575 vm_page_swapqueue(vm_page_t m, uint8_t oldq, uint8_t newq)
3577 struct vm_pagequeue *pq;
3581 KASSERT(oldq < PQ_COUNT && newq < PQ_COUNT && oldq != newq,
3582 ("vm_page_swapqueue: invalid queues (%d, %d)", oldq, newq));
3583 vm_page_assert_locked(m);
3585 pq = &vm_pagequeue_domain(m)->vmd_pagequeues[oldq];
3586 vm_pagequeue_lock(pq);
3589 * The physical queue state might change at any point before the page
3590 * queue lock is acquired, so we must verify that we hold the correct
3591 * lock before proceeding.
3593 if (__predict_false(m->a.queue != oldq)) {
3594 vm_pagequeue_unlock(pq);
3599 * Once the queue index of the page changes, there is nothing
3600 * synchronizing with further updates to the physical queue state.
3601 * Therefore we must remove the page from the queue now in anticipation
3602 * of a successful commit, and be prepared to roll back.
3604 if (__predict_true((m->a.flags & PGA_ENQUEUED) != 0)) {
3605 next = TAILQ_NEXT(m, plinks.q);
3606 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3607 vm_page_aflag_clear(m, PGA_ENQUEUED);
3614 * Atomically update the queue field and set PGA_REQUEUE while
3615 * ensuring that PGA_DEQUEUE has not been set.
3617 if (__predict_false(!vm_page_pqstate_cmpset(m, oldq, newq, PGA_DEQUEUE,
3620 vm_page_aflag_set(m, PGA_ENQUEUED);
3622 TAILQ_INSERT_BEFORE(next, m, plinks.q);
3624 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3626 vm_pagequeue_unlock(pq);
3629 vm_pagequeue_cnt_dec(pq);
3630 vm_pagequeue_unlock(pq);
3631 vm_page_pqbatch_submit(m, newq);
3635 * vm_page_free_prep:
3637 * Prepares the given page to be put on the free list,
3638 * disassociating it from any VM object. The caller may return
3639 * the page to the free list only if this function returns true.
3641 * The object must be locked. The page must be locked if it is
3645 vm_page_free_prep(vm_page_t m)
3649 * Synchronize with threads that have dropped a reference to this
3652 atomic_thread_fence_acq();
3654 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3655 if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3658 p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3659 for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3660 KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3661 m, i, (uintmax_t)*p));
3664 if ((m->oflags & VPO_UNMANAGED) == 0) {
3665 KASSERT(!pmap_page_is_mapped(m),
3666 ("vm_page_free_prep: freeing mapped page %p", m));
3667 KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
3668 ("vm_page_free_prep: mapping flags set in page %p", m));
3670 KASSERT(m->a.queue == PQ_NONE,
3671 ("vm_page_free_prep: unmanaged page %p is queued", m));
3673 VM_CNT_INC(v_tfree);
3675 if (vm_page_sbusied(m))
3676 panic("vm_page_free_prep: freeing shared busy page %p", m);
3678 if (m->object != NULL) {
3679 vm_page_object_remove(m);
3682 * The object reference can be released without an atomic
3685 KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
3686 m->ref_count == VPRC_OBJREF,
3687 ("vm_page_free_prep: page %p has unexpected ref_count %u",
3690 m->ref_count -= VPRC_OBJREF;
3693 if (vm_page_xbusied(m))
3697 * If fictitious remove object association and
3700 if ((m->flags & PG_FICTITIOUS) != 0) {
3701 KASSERT(m->ref_count == 1,
3702 ("fictitious page %p is referenced", m));
3703 KASSERT(m->a.queue == PQ_NONE,
3704 ("fictitious page %p is queued", m));
3709 * Pages need not be dequeued before they are returned to the physical
3710 * memory allocator, but they must at least be marked for a deferred
3713 if ((m->oflags & VPO_UNMANAGED) == 0)
3714 vm_page_dequeue_deferred_free(m);
3719 if (m->ref_count != 0)
3720 panic("vm_page_free_prep: page %p has references", m);
3723 * Restore the default memory attribute to the page.
3725 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3726 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3728 #if VM_NRESERVLEVEL > 0
3730 * Determine whether the page belongs to a reservation. If the page was
3731 * allocated from a per-CPU cache, it cannot belong to a reservation, so
3732 * as an optimization, we avoid the check in that case.
3734 if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
3744 * Returns the given page to the free list, disassociating it
3745 * from any VM object.
3747 * The object must be locked. The page must be locked if it is
3751 vm_page_free_toq(vm_page_t m)
3753 struct vm_domain *vmd;
3756 if (!vm_page_free_prep(m))
3759 vmd = vm_pagequeue_domain(m);
3760 zone = vmd->vmd_pgcache[m->pool].zone;
3761 if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
3765 vm_domain_free_lock(vmd);
3766 vm_phys_free_pages(m, 0);
3767 vm_domain_free_unlock(vmd);
3768 vm_domain_freecnt_inc(vmd, 1);
3772 * vm_page_free_pages_toq:
3774 * Returns a list of pages to the free list, disassociating it
3775 * from any VM object. In other words, this is equivalent to
3776 * calling vm_page_free_toq() for each page of a list of VM objects.
3778 * The objects must be locked. The pages must be locked if it is
3782 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
3787 if (SLIST_EMPTY(free))
3791 while ((m = SLIST_FIRST(free)) != NULL) {
3793 SLIST_REMOVE_HEAD(free, plinks.s.ss);
3794 vm_page_free_toq(m);
3797 if (update_wire_count)
3802 * Mark this page as wired down, preventing reclamation by the page daemon
3803 * or when the containing object is destroyed.
3806 vm_page_wire(vm_page_t m)
3810 KASSERT(m->object != NULL,
3811 ("vm_page_wire: page %p does not belong to an object", m));
3812 if (!vm_page_busied(m) && !vm_object_busied(m->object))
3813 VM_OBJECT_ASSERT_LOCKED(m->object);
3814 KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
3815 VPRC_WIRE_COUNT(m->ref_count) >= 1,
3816 ("vm_page_wire: fictitious page %p has zero wirings", m));
3818 old = atomic_fetchadd_int(&m->ref_count, 1);
3819 KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
3820 ("vm_page_wire: counter overflow for page %p", m));
3821 if (VPRC_WIRE_COUNT(old) == 0)
3826 * Attempt to wire a mapped page following a pmap lookup of that page.
3827 * This may fail if a thread is concurrently tearing down mappings of the page.
3828 * The transient failure is acceptable because it translates to the
3829 * failure of the caller pmap_extract_and_hold(), which should be then
3830 * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages().
3833 vm_page_wire_mapped(vm_page_t m)
3840 ("vm_page_wire_mapped: wiring unreferenced page %p", m));
3841 if ((old & VPRC_BLOCKED) != 0)
3843 } while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
3845 if (VPRC_WIRE_COUNT(old) == 0)
3851 * Release one wiring of the specified page, potentially allowing it to be
3854 * Only managed pages belonging to an object can be paged out. If the number
3855 * of wirings transitions to zero and the page is eligible for page out, then
3856 * the page is added to the specified paging queue. If the released wiring
3857 * represented the last reference to the page, the page is freed.
3859 * A managed page must be locked.
3862 vm_page_unwire(vm_page_t m, uint8_t queue)
3867 KASSERT(queue < PQ_COUNT,
3868 ("vm_page_unwire: invalid queue %u request for page %p", queue, m));
3870 if ((m->oflags & VPO_UNMANAGED) != 0) {
3871 if (vm_page_unwire_noq(m) && m->ref_count == 0)
3877 * Update LRU state before releasing the wiring reference.
3878 * We only need to do this once since we hold the page lock.
3879 * Use a release store when updating the reference count to
3880 * synchronize with vm_page_free_prep().
3885 KASSERT(VPRC_WIRE_COUNT(old) > 0,
3886 ("vm_page_unwire: wire count underflow for page %p", m));
3887 if (!locked && VPRC_WIRE_COUNT(old) == 1) {
3890 if (queue == PQ_ACTIVE && vm_page_queue(m) == PQ_ACTIVE)
3891 vm_page_reference(m);
3893 vm_page_mvqueue(m, queue);
3895 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
3898 * Release the lock only after the wiring is released, to ensure that
3899 * the page daemon does not encounter and dequeue the page while it is
3905 if (VPRC_WIRE_COUNT(old) == 1) {
3913 * Unwire a page without (re-)inserting it into a page queue. It is up
3914 * to the caller to enqueue, requeue, or free the page as appropriate.
3915 * In most cases involving managed pages, vm_page_unwire() should be used
3919 vm_page_unwire_noq(vm_page_t m)
3923 old = vm_page_drop(m, 1);
3924 KASSERT(VPRC_WIRE_COUNT(old) != 0,
3925 ("vm_page_unref: counter underflow for page %p", m));
3926 KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
3927 ("vm_page_unref: missing ref on fictitious page %p", m));
3929 if (VPRC_WIRE_COUNT(old) > 1)
3936 * Ensure that the page is in the specified page queue. If the page is
3937 * active or being moved to the active queue, ensure that its act_count is
3938 * at least ACT_INIT but do not otherwise mess with it. Otherwise, ensure that
3939 * the page is at the tail of its page queue.
3941 * The page may be wired. The caller should release its wiring reference
3942 * before releasing the page lock, otherwise the page daemon may immediately
3945 * A managed page must be locked.
3947 static __always_inline void
3948 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue)
3951 vm_page_assert_locked(m);
3952 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3953 ("vm_page_mvqueue: page %p is unmanaged", m));
3954 KASSERT(m->ref_count > 0,
3955 ("%s: page %p does not carry any references", __func__, m));
3957 if (vm_page_queue(m) != nqueue) {
3959 vm_page_enqueue(m, nqueue);
3960 } else if (nqueue != PQ_ACTIVE) {
3964 if (nqueue == PQ_ACTIVE && m->a.act_count < ACT_INIT)
3965 m->a.act_count = ACT_INIT;
3969 * Put the specified page on the active list (if appropriate).
3972 vm_page_activate(vm_page_t m)
3975 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
3977 vm_page_mvqueue(m, PQ_ACTIVE);
3981 * Move the specified page to the tail of the inactive queue, or requeue
3982 * the page if it is already in the inactive queue.
3985 vm_page_deactivate(vm_page_t m)
3988 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
3990 vm_page_mvqueue(m, PQ_INACTIVE);
3994 * Move the specified page close to the head of the inactive queue,
3995 * bypassing LRU. A marker page is used to maintain FIFO ordering.
3996 * As with regular enqueues, we use a per-CPU batch queue to reduce
3997 * contention on the page queue lock.
4000 _vm_page_deactivate_noreuse(vm_page_t m)
4003 vm_page_assert_locked(m);
4005 if (!vm_page_inactive(m)) {
4007 m->a.queue = PQ_INACTIVE;
4009 if ((m->a.flags & PGA_REQUEUE_HEAD) == 0)
4010 vm_page_aflag_set(m, PGA_REQUEUE_HEAD);
4011 vm_page_pqbatch_submit(m, PQ_INACTIVE);
4015 vm_page_deactivate_noreuse(vm_page_t m)
4018 KASSERT(m->object != NULL,
4019 ("vm_page_deactivate_noreuse: page %p has no object", m));
4021 if ((m->oflags & VPO_UNMANAGED) == 0 && !vm_page_wired(m))
4022 _vm_page_deactivate_noreuse(m);
4026 * Put a page in the laundry, or requeue it if it is already there.
4029 vm_page_launder(vm_page_t m)
4032 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
4034 vm_page_mvqueue(m, PQ_LAUNDRY);
4038 * Put a page in the PQ_UNSWAPPABLE holding queue.
4041 vm_page_unswappable(vm_page_t m)
4044 vm_page_assert_locked(m);
4045 KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0,
4046 ("page %p already unswappable", m));
4049 vm_page_enqueue(m, PQ_UNSWAPPABLE);
4053 vm_page_release_toq(vm_page_t m, int flags)
4056 vm_page_assert_locked(m);
4059 * Use a check of the valid bits to determine whether we should
4060 * accelerate reclamation of the page. The object lock might not be
4061 * held here, in which case the check is racy. At worst we will either
4062 * accelerate reclamation of a valid page and violate LRU, or
4063 * unnecessarily defer reclamation of an invalid page.
4065 * If we were asked to not cache the page, place it near the head of the
4066 * inactive queue so that is reclaimed sooner.
4068 if ((flags & (VPR_TRYFREE | VPR_NOREUSE)) != 0 || m->valid == 0)
4069 _vm_page_deactivate_noreuse(m);
4070 else if (vm_page_active(m))
4071 vm_page_reference(m);
4073 vm_page_mvqueue(m, PQ_INACTIVE);
4077 * Unwire a page and either attempt to free it or re-add it to the page queues.
4080 vm_page_release(vm_page_t m, int flags)
4086 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4087 ("vm_page_release: page %p is unmanaged", m));
4089 if ((flags & VPR_TRYFREE) != 0) {
4091 object = (vm_object_t)atomic_load_ptr(&m->object);
4094 /* Depends on type-stability. */
4095 if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object)) {
4099 if (object == m->object)
4101 VM_OBJECT_WUNLOCK(object);
4103 if (__predict_true(object != NULL)) {
4104 vm_page_release_locked(m, flags);
4105 VM_OBJECT_WUNLOCK(object);
4111 * Update LRU state before releasing the wiring reference.
4112 * Use a release store when updating the reference count to
4113 * synchronize with vm_page_free_prep().
4118 KASSERT(VPRC_WIRE_COUNT(old) > 0,
4119 ("vm_page_unwire: wire count underflow for page %p", m));
4120 if (!locked && VPRC_WIRE_COUNT(old) == 1) {
4123 vm_page_release_toq(m, flags);
4125 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
4128 * Release the lock only after the wiring is released, to ensure that
4129 * the page daemon does not encounter and dequeue the page while it is
4135 if (VPRC_WIRE_COUNT(old) == 1) {
4142 /* See vm_page_release(). */
4144 vm_page_release_locked(vm_page_t m, int flags)
4147 VM_OBJECT_ASSERT_WLOCKED(m->object);
4148 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4149 ("vm_page_release_locked: page %p is unmanaged", m));
4151 if (vm_page_unwire_noq(m)) {
4152 if ((flags & VPR_TRYFREE) != 0 &&
4153 (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
4154 m->dirty == 0 && !vm_page_busied(m)) {
4158 vm_page_release_toq(m, flags);
4165 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
4169 KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
4170 ("vm_page_try_blocked_op: page %p has no object", m));
4171 KASSERT(vm_page_busied(m),
4172 ("vm_page_try_blocked_op: page %p is not busy", m));
4173 VM_OBJECT_ASSERT_LOCKED(m->object);
4178 ("vm_page_try_blocked_op: page %p has no references", m));
4179 if (VPRC_WIRE_COUNT(old) != 0)
4181 } while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
4186 * If the object is read-locked, new wirings may be created via an
4189 old = vm_page_drop(m, VPRC_BLOCKED);
4190 KASSERT(!VM_OBJECT_WOWNED(m->object) ||
4191 old == (VPRC_BLOCKED | VPRC_OBJREF),
4192 ("vm_page_try_blocked_op: unexpected refcount value %u for %p",
4198 * Atomically check for wirings and remove all mappings of the page.
4201 vm_page_try_remove_all(vm_page_t m)
4204 return (vm_page_try_blocked_op(m, pmap_remove_all));
4208 * Atomically check for wirings and remove all writeable mappings of the page.
4211 vm_page_try_remove_write(vm_page_t m)
4214 return (vm_page_try_blocked_op(m, pmap_remove_write));
4220 * Apply the specified advice to the given page.
4222 * The object and page must be locked.
4225 vm_page_advise(vm_page_t m, int advice)
4228 vm_page_assert_locked(m);
4229 VM_OBJECT_ASSERT_WLOCKED(m->object);
4230 if (advice == MADV_FREE)
4232 * Mark the page clean. This will allow the page to be freed
4233 * without first paging it out. MADV_FREE pages are often
4234 * quickly reused by malloc(3), so we do not do anything that
4235 * would result in a page fault on a later access.
4238 else if (advice != MADV_DONTNEED) {
4239 if (advice == MADV_WILLNEED)
4240 vm_page_activate(m);
4245 * Clear any references to the page. Otherwise, the page daemon will
4246 * immediately reactivate the page.
4248 vm_page_aflag_clear(m, PGA_REFERENCED);
4250 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
4254 * Place clean pages near the head of the inactive queue rather than
4255 * the tail, thus defeating the queue's LRU operation and ensuring that
4256 * the page will be reused quickly. Dirty pages not already in the
4257 * laundry are moved there.
4260 vm_page_deactivate_noreuse(m);
4261 else if (!vm_page_in_laundry(m))
4266 vm_page_grab_pflags(int allocflags)
4270 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4271 (allocflags & VM_ALLOC_WIRED) != 0,
4272 ("vm_page_grab_pflags: the pages must be busied or wired"));
4273 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4274 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4275 ("vm_page_grab_pflags: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY "
4277 pflags = allocflags &
4278 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL |
4280 if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4281 pflags |= VM_ALLOC_WAITFAIL;
4282 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4283 pflags |= VM_ALLOC_SBUSY;
4289 * Grab a page, waiting until we are waken up due to the page
4290 * changing state. We keep on waiting, if the page continues
4291 * to be in the object. If the page doesn't exist, first allocate it
4292 * and then conditionally zero it.
4294 * This routine may sleep.
4296 * The object must be locked on entry. The lock will, however, be released
4297 * and reacquired if the routine sleeps.
4300 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
4305 VM_OBJECT_ASSERT_WLOCKED(object);
4306 pflags = vm_page_grab_pflags(allocflags);
4308 if ((m = vm_page_lookup(object, pindex)) != NULL) {
4309 if (!vm_page_acquire_flags(m, allocflags)) {
4310 if (vm_page_busy_sleep_flags(object, m, "pgrbwt",
4317 if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4319 m = vm_page_alloc(object, pindex, pflags);
4321 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4325 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
4329 if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4330 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4339 * Grab a page and make it valid, paging in if necessary. Pages missing from
4340 * their pager are zero filled and validated. If a VM_ALLOC_COUNT is supplied
4341 * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought
4342 * in simultaneously. Additional pages will be left on a paging queue but
4343 * will neither be wired nor busy regardless of allocflags.
4346 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags)
4349 vm_page_t ma[VM_INITIAL_PAGEIN];
4351 int after, i, pflags, rv;
4353 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4354 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4355 ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4356 KASSERT((allocflags &
4357 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4358 ("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
4359 VM_OBJECT_ASSERT_WLOCKED(object);
4360 pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY);
4361 pflags |= VM_ALLOC_WAITFAIL;
4365 if ((m = vm_page_lookup(object, pindex)) != NULL) {
4367 * If the page is fully valid it can only become invalid
4368 * with the object lock held. If it is not valid it can
4369 * become valid with the busy lock held. Therefore, we
4370 * may unnecessarily lock the exclusive busy here if we
4371 * race with I/O completion not using the object lock.
4372 * However, we will not end up with an invalid page and a
4375 if (!vm_page_all_valid(m) ||
4376 (allocflags & (VM_ALLOC_IGN_SBUSY | VM_ALLOC_SBUSY)) == 0) {
4377 sleep = !vm_page_tryxbusy(m);
4380 sleep = !vm_page_trysbusy(m);
4382 (void)vm_page_busy_sleep_flags(object, m, "pgrbwt",
4386 if ((allocflags & VM_ALLOC_NOCREAT) != 0 &&
4387 !vm_page_all_valid(m)) {
4393 return (VM_PAGER_FAIL);
4395 if ((allocflags & VM_ALLOC_WIRED) != 0)
4397 if (vm_page_all_valid(m))
4399 } else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4401 return (VM_PAGER_FAIL);
4402 } else if ((m = vm_page_alloc(object, pindex, pflags)) != NULL) {
4408 vm_page_assert_xbusied(m);
4410 if (vm_pager_has_page(object, pindex, NULL, &after)) {
4411 after = MIN(after, VM_INITIAL_PAGEIN);
4412 after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT);
4413 after = MAX(after, 1);
4415 for (i = 1; i < after; i++) {
4416 if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) {
4417 if (ma[i]->valid || !vm_page_tryxbusy(ma[i]))
4420 ma[i] = vm_page_alloc(object, m->pindex + i,
4427 rv = vm_pager_get_pages(object, ma, after, NULL, NULL);
4428 /* Pager may have replaced a page. */
4430 if (rv != VM_PAGER_OK) {
4431 if ((allocflags & VM_ALLOC_WIRED) != 0)
4432 vm_page_unwire_noq(m);
4433 for (i = 0; i < after; i++) {
4434 if (!vm_page_wired(ma[i]))
4435 vm_page_free(ma[i]);
4437 vm_page_xunbusy(ma[i]);
4442 for (i = 1; i < after; i++)
4443 vm_page_readahead_finish(ma[i]);
4444 MPASS(vm_page_all_valid(m));
4446 vm_page_zero_invalid(m, TRUE);
4449 if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4455 if ((allocflags & VM_ALLOC_SBUSY) != 0 && xbusy)
4456 vm_page_busy_downgrade(m);
4458 return (VM_PAGER_OK);
4462 * Return the specified range of pages from the given object. For each
4463 * page offset within the range, if a page already exists within the object
4464 * at that offset and it is busy, then wait for it to change state. If,
4465 * instead, the page doesn't exist, then allocate it.
4467 * The caller must always specify an allocation class.
4469 * allocation classes:
4470 * VM_ALLOC_NORMAL normal process request
4471 * VM_ALLOC_SYSTEM system *really* needs the pages
4473 * The caller must always specify that the pages are to be busied and/or
4476 * optional allocation flags:
4477 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages
4478 * VM_ALLOC_NOBUSY do not exclusive busy the page
4479 * VM_ALLOC_NOWAIT do not sleep
4480 * VM_ALLOC_SBUSY set page to sbusy state
4481 * VM_ALLOC_WIRED wire the pages
4482 * VM_ALLOC_ZERO zero and validate any invalid pages
4484 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it
4485 * may return a partial prefix of the requested range.
4488 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
4489 vm_page_t *ma, int count)
4495 VM_OBJECT_ASSERT_WLOCKED(object);
4496 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
4497 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
4499 pflags = vm_page_grab_pflags(allocflags);
4505 m = vm_radix_lookup_le(&object->rtree, pindex + i);
4506 if (m == NULL || m->pindex != pindex + i) {
4510 mpred = TAILQ_PREV(m, pglist, listq);
4511 for (; i < count; i++) {
4513 if (!vm_page_acquire_flags(m, allocflags)) {
4514 if (vm_page_busy_sleep_flags(object, m,
4515 "grbmaw", allocflags))
4520 if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4522 m = vm_page_alloc_after(object, pindex + i,
4523 pflags | VM_ALLOC_COUNT(count - i), mpred);
4525 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4530 if (vm_page_none_valid(m) &&
4531 (allocflags & VM_ALLOC_ZERO) != 0) {
4532 if ((m->flags & PG_ZERO) == 0)
4536 if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4537 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4543 m = vm_page_next(m);
4549 * Mapping function for valid or dirty bits in a page.
4551 * Inputs are required to range within a page.
4554 vm_page_bits(int base, int size)
4560 base + size <= PAGE_SIZE,
4561 ("vm_page_bits: illegal base/size %d/%d", base, size)
4564 if (size == 0) /* handle degenerate case */
4567 first_bit = base >> DEV_BSHIFT;
4568 last_bit = (base + size - 1) >> DEV_BSHIFT;
4570 return (((vm_page_bits_t)2 << last_bit) -
4571 ((vm_page_bits_t)1 << first_bit));
4575 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set)
4578 #if PAGE_SIZE == 32768
4579 atomic_set_64((uint64_t *)bits, set);
4580 #elif PAGE_SIZE == 16384
4581 atomic_set_32((uint32_t *)bits, set);
4582 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16)
4583 atomic_set_16((uint16_t *)bits, set);
4584 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8)
4585 atomic_set_8((uint8_t *)bits, set);
4586 #else /* PAGE_SIZE <= 8192 */
4590 addr = (uintptr_t)bits;
4592 * Use a trick to perform a 32-bit atomic on the
4593 * containing aligned word, to not depend on the existence
4594 * of atomic_{set, clear}_{8, 16}.
4596 shift = addr & (sizeof(uint32_t) - 1);
4597 #if BYTE_ORDER == BIG_ENDIAN
4598 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4602 addr &= ~(sizeof(uint32_t) - 1);
4603 atomic_set_32((uint32_t *)addr, set << shift);
4604 #endif /* PAGE_SIZE */
4608 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear)
4611 #if PAGE_SIZE == 32768
4612 atomic_clear_64((uint64_t *)bits, clear);
4613 #elif PAGE_SIZE == 16384
4614 atomic_clear_32((uint32_t *)bits, clear);
4615 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16)
4616 atomic_clear_16((uint16_t *)bits, clear);
4617 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8)
4618 atomic_clear_8((uint8_t *)bits, clear);
4619 #else /* PAGE_SIZE <= 8192 */
4623 addr = (uintptr_t)bits;
4625 * Use a trick to perform a 32-bit atomic on the
4626 * containing aligned word, to not depend on the existence
4627 * of atomic_{set, clear}_{8, 16}.
4629 shift = addr & (sizeof(uint32_t) - 1);
4630 #if BYTE_ORDER == BIG_ENDIAN
4631 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4635 addr &= ~(sizeof(uint32_t) - 1);
4636 atomic_clear_32((uint32_t *)addr, clear << shift);
4637 #endif /* PAGE_SIZE */
4640 static inline vm_page_bits_t
4641 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits)
4643 #if PAGE_SIZE == 32768
4647 while (atomic_fcmpset_64(bits, &old, newbits) == 0);
4649 #elif PAGE_SIZE == 16384
4653 while (atomic_fcmpset_32(bits, &old, newbits) == 0);
4655 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16)
4659 while (atomic_fcmpset_16(bits, &old, newbits) == 0);
4661 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8)
4665 while (atomic_fcmpset_8(bits, &old, newbits) == 0);
4667 #else /* PAGE_SIZE <= 4096*/
4669 uint32_t old, new, mask;
4672 addr = (uintptr_t)bits;
4674 * Use a trick to perform a 32-bit atomic on the
4675 * containing aligned word, to not depend on the existence
4676 * of atomic_{set, swap, clear}_{8, 16}.
4678 shift = addr & (sizeof(uint32_t) - 1);
4679 #if BYTE_ORDER == BIG_ENDIAN
4680 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4684 addr &= ~(sizeof(uint32_t) - 1);
4685 mask = VM_PAGE_BITS_ALL << shift;
4690 new |= newbits << shift;
4691 } while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0);
4692 return (old >> shift);
4693 #endif /* PAGE_SIZE */
4697 * vm_page_set_valid_range:
4699 * Sets portions of a page valid. The arguments are expected
4700 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4701 * of any partial chunks touched by the range. The invalid portion of
4702 * such chunks will be zeroed.
4704 * (base + size) must be less then or equal to PAGE_SIZE.
4707 vm_page_set_valid_range(vm_page_t m, int base, int size)
4710 vm_page_bits_t pagebits;
4712 vm_page_assert_busied(m);
4713 if (size == 0) /* handle degenerate case */
4717 * If the base is not DEV_BSIZE aligned and the valid
4718 * bit is clear, we have to zero out a portion of the
4721 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4722 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
4723 pmap_zero_page_area(m, frag, base - frag);
4726 * If the ending offset is not DEV_BSIZE aligned and the
4727 * valid bit is clear, we have to zero out a portion of
4730 endoff = base + size;
4731 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4732 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
4733 pmap_zero_page_area(m, endoff,
4734 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4737 * Assert that no previously invalid block that is now being validated
4740 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
4741 ("vm_page_set_valid_range: page %p is dirty", m));
4744 * Set valid bits inclusive of any overlap.
4746 pagebits = vm_page_bits(base, size);
4747 if (vm_page_xbusied(m))
4748 m->valid |= pagebits;
4750 vm_page_bits_set(m, &m->valid, pagebits);
4754 * Set the page dirty bits and free the invalid swap space if
4755 * present. Returns the previous dirty bits.
4758 vm_page_set_dirty(vm_page_t m)
4762 VM_PAGE_OBJECT_BUSY_ASSERT(m);
4764 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) {
4766 m->dirty = VM_PAGE_BITS_ALL;
4768 old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL);
4769 if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0)
4770 vm_pager_page_unswapped(m);
4776 * Clear the given bits from the specified page's dirty field.
4778 static __inline void
4779 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
4782 vm_page_assert_busied(m);
4785 * If the page is xbusied and not write mapped we are the
4786 * only thread that can modify dirty bits. Otherwise, The pmap
4787 * layer can call vm_page_dirty() without holding a distinguished
4788 * lock. The combination of page busy and atomic operations
4789 * suffice to guarantee consistency of the page dirty field.
4791 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
4792 m->dirty &= ~pagebits;
4794 vm_page_bits_clear(m, &m->dirty, pagebits);
4798 * vm_page_set_validclean:
4800 * Sets portions of a page valid and clean. The arguments are expected
4801 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4802 * of any partial chunks touched by the range. The invalid portion of
4803 * such chunks will be zero'd.
4805 * (base + size) must be less then or equal to PAGE_SIZE.
4808 vm_page_set_validclean(vm_page_t m, int base, int size)
4810 vm_page_bits_t oldvalid, pagebits;
4813 vm_page_assert_busied(m);
4814 if (size == 0) /* handle degenerate case */
4818 * If the base is not DEV_BSIZE aligned and the valid
4819 * bit is clear, we have to zero out a portion of the
4822 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4823 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
4824 pmap_zero_page_area(m, frag, base - frag);
4827 * If the ending offset is not DEV_BSIZE aligned and the
4828 * valid bit is clear, we have to zero out a portion of
4831 endoff = base + size;
4832 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4833 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
4834 pmap_zero_page_area(m, endoff,
4835 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4838 * Set valid, clear dirty bits. If validating the entire
4839 * page we can safely clear the pmap modify bit. We also
4840 * use this opportunity to clear the PGA_NOSYNC flag. If a process
4841 * takes a write fault on a MAP_NOSYNC memory area the flag will
4844 * We set valid bits inclusive of any overlap, but we can only
4845 * clear dirty bits for DEV_BSIZE chunks that are fully within
4848 oldvalid = m->valid;
4849 pagebits = vm_page_bits(base, size);
4850 if (vm_page_xbusied(m))
4851 m->valid |= pagebits;
4853 vm_page_bits_set(m, &m->valid, pagebits);
4855 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
4856 frag = DEV_BSIZE - frag;
4862 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
4864 if (base == 0 && size == PAGE_SIZE) {
4866 * The page can only be modified within the pmap if it is
4867 * mapped, and it can only be mapped if it was previously
4870 if (oldvalid == VM_PAGE_BITS_ALL)
4872 * Perform the pmap_clear_modify() first. Otherwise,
4873 * a concurrent pmap operation, such as
4874 * pmap_protect(), could clear a modification in the
4875 * pmap and set the dirty field on the page before
4876 * pmap_clear_modify() had begun and after the dirty
4877 * field was cleared here.
4879 pmap_clear_modify(m);
4881 vm_page_aflag_clear(m, PGA_NOSYNC);
4882 } else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m))
4883 m->dirty &= ~pagebits;
4885 vm_page_clear_dirty_mask(m, pagebits);
4889 vm_page_clear_dirty(vm_page_t m, int base, int size)
4892 vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
4896 * vm_page_set_invalid:
4898 * Invalidates DEV_BSIZE'd chunks within a page. Both the
4899 * valid and dirty bits for the effected areas are cleared.
4902 vm_page_set_invalid(vm_page_t m, int base, int size)
4904 vm_page_bits_t bits;
4908 * The object lock is required so that pages can't be mapped
4909 * read-only while we're in the process of invalidating them.
4912 VM_OBJECT_ASSERT_WLOCKED(object);
4913 vm_page_assert_busied(m);
4915 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
4916 size >= object->un_pager.vnp.vnp_size)
4917 bits = VM_PAGE_BITS_ALL;
4919 bits = vm_page_bits(base, size);
4920 if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0)
4922 KASSERT((bits == 0 && vm_page_all_valid(m)) ||
4923 !pmap_page_is_mapped(m),
4924 ("vm_page_set_invalid: page %p is mapped", m));
4925 if (vm_page_xbusied(m)) {
4929 vm_page_bits_clear(m, &m->valid, bits);
4930 vm_page_bits_clear(m, &m->dirty, bits);
4937 * Invalidates the entire page. The page must be busy, unmapped, and
4938 * the enclosing object must be locked. The object locks protects
4939 * against concurrent read-only pmap enter which is done without
4943 vm_page_invalid(vm_page_t m)
4946 vm_page_assert_busied(m);
4947 VM_OBJECT_ASSERT_LOCKED(m->object);
4948 MPASS(!pmap_page_is_mapped(m));
4950 if (vm_page_xbusied(m))
4953 vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL);
4957 * vm_page_zero_invalid()
4959 * The kernel assumes that the invalid portions of a page contain
4960 * garbage, but such pages can be mapped into memory by user code.
4961 * When this occurs, we must zero out the non-valid portions of the
4962 * page so user code sees what it expects.
4964 * Pages are most often semi-valid when the end of a file is mapped
4965 * into memory and the file's size is not page aligned.
4968 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
4974 * Scan the valid bits looking for invalid sections that
4975 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the
4976 * valid bit may be set ) have already been zeroed by
4977 * vm_page_set_validclean().
4979 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
4980 if (i == (PAGE_SIZE / DEV_BSIZE) ||
4981 (m->valid & ((vm_page_bits_t)1 << i))) {
4983 pmap_zero_page_area(m,
4984 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
4991 * setvalid is TRUE when we can safely set the zero'd areas
4992 * as being valid. We can do this if there are no cache consistancy
4993 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
5002 * Is (partial) page valid? Note that the case where size == 0
5003 * will return FALSE in the degenerate case where the page is
5004 * entirely invalid, and TRUE otherwise.
5006 * Some callers envoke this routine without the busy lock held and
5007 * handle races via higher level locks. Typical callers should
5008 * hold a busy lock to prevent invalidation.
5011 vm_page_is_valid(vm_page_t m, int base, int size)
5013 vm_page_bits_t bits;
5015 bits = vm_page_bits(base, size);
5016 return (m->valid != 0 && (m->valid & bits) == bits);
5020 * Returns true if all of the specified predicates are true for the entire
5021 * (super)page and false otherwise.
5024 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
5030 if (skip_m != NULL && skip_m->object != object)
5032 VM_OBJECT_ASSERT_LOCKED(object);
5033 npages = atop(pagesizes[m->psind]);
5036 * The physically contiguous pages that make up a superpage, i.e., a
5037 * page with a page size index ("psind") greater than zero, will
5038 * occupy adjacent entries in vm_page_array[].
5040 for (i = 0; i < npages; i++) {
5041 /* Always test object consistency, including "skip_m". */
5042 if (m[i].object != object)
5044 if (&m[i] == skip_m)
5046 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
5048 if ((flags & PS_ALL_DIRTY) != 0) {
5050 * Calling vm_page_test_dirty() or pmap_is_modified()
5051 * might stop this case from spuriously returning
5052 * "false". However, that would require a write lock
5053 * on the object containing "m[i]".
5055 if (m[i].dirty != VM_PAGE_BITS_ALL)
5058 if ((flags & PS_ALL_VALID) != 0 &&
5059 m[i].valid != VM_PAGE_BITS_ALL)
5066 * Set the page's dirty bits if the page is modified.
5069 vm_page_test_dirty(vm_page_t m)
5072 vm_page_assert_busied(m);
5073 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
5078 vm_page_valid(vm_page_t m)
5081 vm_page_assert_busied(m);
5082 if (vm_page_xbusied(m))
5083 m->valid = VM_PAGE_BITS_ALL;
5085 vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL);
5089 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
5092 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
5096 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
5099 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
5103 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
5106 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
5109 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
5111 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
5114 vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
5118 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
5121 mtx_assert_(vm_page_lockptr(m), a, file, line);
5127 vm_page_object_busy_assert(vm_page_t m)
5131 * Certain of the page's fields may only be modified by the
5132 * holder of a page or object busy.
5134 if (m->object != NULL && !vm_page_busied(m))
5135 VM_OBJECT_ASSERT_BUSY(m->object);
5139 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits)
5142 if ((bits & PGA_WRITEABLE) == 0)
5146 * The PGA_WRITEABLE flag can only be set if the page is
5147 * managed, is exclusively busied or the object is locked.
5148 * Currently, this flag is only set by pmap_enter().
5150 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5151 ("PGA_WRITEABLE on unmanaged page"));
5152 if (!vm_page_xbusied(m))
5153 VM_OBJECT_ASSERT_BUSY(m->object);
5157 #include "opt_ddb.h"
5159 #include <sys/kernel.h>
5161 #include <ddb/ddb.h>
5163 DB_SHOW_COMMAND(page, vm_page_print_page_info)
5166 db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
5167 db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
5168 db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
5169 db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
5170 db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
5171 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
5172 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
5173 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
5174 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
5177 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
5181 db_printf("pq_free %d\n", vm_free_count());
5182 for (dom = 0; dom < vm_ndomains; dom++) {
5184 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
5186 vm_dom[dom].vmd_page_count,
5187 vm_dom[dom].vmd_free_count,
5188 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
5189 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
5190 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
5191 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
5195 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
5198 boolean_t phys, virt;
5201 db_printf("show pginfo addr\n");
5205 phys = strchr(modif, 'p') != NULL;
5206 virt = strchr(modif, 'v') != NULL;
5208 m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
5210 m = PHYS_TO_VM_PAGE(addr);
5212 m = (vm_page_t)addr;
5214 "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref %u\n"
5215 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
5216 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
5217 m->a.queue, m->ref_count, m->a.flags, m->oflags,
5218 m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty);