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 * GENERAL RULES ON VM_PAGE MANIPULATION
68 * - A page queue lock is required when adding or removing a page from a
69 * page queue regardless of other locks or the busy state of a page.
71 * * In general, no thread besides the page daemon can acquire or
72 * hold more than one page queue lock at a time.
74 * * The page daemon can acquire and hold any pair of page queue
77 * - The object lock is required when inserting or removing
78 * pages from an object (vm_page_insert() or vm_page_remove()).
83 * Resident memory management module.
86 #include <sys/cdefs.h>
87 __FBSDID("$FreeBSD$");
91 #include <sys/param.h>
92 #include <sys/systm.h>
94 #include <sys/domainset.h>
95 #include <sys/kernel.h>
96 #include <sys/limits.h>
97 #include <sys/linker.h>
98 #include <sys/malloc.h>
100 #include <sys/msgbuf.h>
101 #include <sys/mutex.h>
102 #include <sys/proc.h>
103 #include <sys/rwlock.h>
104 #include <sys/sbuf.h>
106 #include <sys/sysctl.h>
107 #include <sys/vmmeter.h>
108 #include <sys/vnode.h>
112 #include <vm/vm_param.h>
113 #include <vm/vm_domainset.h>
114 #include <vm/vm_kern.h>
115 #include <vm/vm_map.h>
116 #include <vm/vm_object.h>
117 #include <vm/vm_page.h>
118 #include <vm/vm_pageout.h>
119 #include <vm/vm_phys.h>
120 #include <vm/vm_pagequeue.h>
121 #include <vm/vm_pager.h>
122 #include <vm/vm_radix.h>
123 #include <vm/vm_reserv.h>
124 #include <vm/vm_extern.h>
126 #include <vm/uma_int.h>
128 #include <machine/md_var.h>
130 extern int uma_startup_count(int);
131 extern void uma_startup(void *, int);
132 extern int vmem_startup_count(void);
135 * Associated with page of user-allocatable memory is a
139 struct vm_domain vm_dom[MAXMEMDOM];
141 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
142 struct mtx_padalign __exclusive_cache_line vm_domainset_lock;
143 domainset_t __exclusive_cache_line vm_min_domains;
144 domainset_t __exclusive_cache_line vm_severe_domains;
145 static int vm_min_waiters;
146 static int vm_severe_waiters;
147 static int vm_pageproc_waiters;
151 * bogus page -- for I/O to/from partially complete buffers,
152 * or for paging into sparsely invalid regions.
154 vm_page_t bogus_page;
156 vm_page_t vm_page_array;
157 long vm_page_array_size;
160 static int boot_pages;
161 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
163 "number of pages allocated for bootstrapping the VM system");
165 static int pa_tryrelock_restart;
166 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
167 &pa_tryrelock_restart, 0, "Number of tryrelock restarts");
169 static TAILQ_HEAD(, vm_page) blacklist_head;
170 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
171 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
172 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
174 static uma_zone_t fakepg_zone;
176 static void vm_page_alloc_check(vm_page_t m);
177 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
178 static void vm_page_enqueue(uint8_t queue, vm_page_t m);
179 static void vm_page_free_phys(struct vm_domain *vmd, vm_page_t m);
180 static void vm_page_init(void *dummy);
181 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
182 vm_pindex_t pindex, vm_page_t mpred);
183 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
185 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
186 vm_page_t m_run, vm_paddr_t high);
187 static void vm_domain_free_wakeup(struct vm_domain *);
188 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
191 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
194 vm_page_init(void *dummy)
197 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
198 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
199 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
200 VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
203 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
204 #if PAGE_SIZE == 32768
206 CTASSERT(sizeof(u_long) >= 8);
211 * Try to acquire a physical address lock while a pmap is locked. If we
212 * fail to trylock we unlock and lock the pmap directly and cache the
213 * locked pa in *locked. The caller should then restart their loop in case
214 * the virtual to physical mapping has changed.
217 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
224 PA_LOCK_ASSERT(lockpa, MA_OWNED);
225 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
232 atomic_add_int(&pa_tryrelock_restart, 1);
241 * Sets the page size, perhaps based upon the memory
242 * size. Must be called before any use of page-size
243 * dependent functions.
246 vm_set_page_size(void)
248 if (vm_cnt.v_page_size == 0)
249 vm_cnt.v_page_size = PAGE_SIZE;
250 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
251 panic("vm_set_page_size: page size not a power of two");
255 * vm_page_blacklist_next:
257 * Find the next entry in the provided string of blacklist
258 * addresses. Entries are separated by space, comma, or newline.
259 * If an invalid integer is encountered then the rest of the
260 * string is skipped. Updates the list pointer to the next
261 * character, or NULL if the string is exhausted or invalid.
264 vm_page_blacklist_next(char **list, char *end)
269 if (list == NULL || *list == NULL)
277 * If there's no end pointer then the buffer is coming from
278 * the kenv and we know it's null-terminated.
281 end = *list + strlen(*list);
283 /* Ensure that strtoq() won't walk off the end */
285 if (*end == '\n' || *end == ' ' || *end == ',')
288 printf("Blacklist not terminated, skipping\n");
294 for (pos = *list; *pos != '\0'; pos = cp) {
295 bad = strtoq(pos, &cp, 0);
296 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
305 if (*cp == '\0' || ++cp >= end)
309 return (trunc_page(bad));
311 printf("Garbage in RAM blacklist, skipping\n");
317 * vm_page_blacklist_check:
319 * Iterate through the provided string of blacklist addresses, pulling
320 * each entry out of the physical allocator free list and putting it
321 * onto a list for reporting via the vm.page_blacklist sysctl.
324 vm_page_blacklist_check(char *list, char *end)
326 struct vm_domain *vmd;
333 while (next != NULL) {
334 if ((pa = vm_page_blacklist_next(&next, end)) == 0)
336 m = vm_phys_paddr_to_vm_page(pa);
339 vmd = vm_pagequeue_domain(m);
340 vm_domain_free_lock(vmd);
341 ret = vm_phys_unfree_page(m);
342 vm_domain_free_unlock(vmd);
344 TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
346 printf("Skipping page with pa 0x%jx\n",
353 * vm_page_blacklist_load:
355 * Search for a special module named "ram_blacklist". It'll be a
356 * plain text file provided by the user via the loader directive
360 vm_page_blacklist_load(char **list, char **end)
369 mod = preload_search_by_type("ram_blacklist");
371 ptr = preload_fetch_addr(mod);
372 len = preload_fetch_size(mod);
383 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
390 error = sysctl_wire_old_buffer(req, 0);
393 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
394 TAILQ_FOREACH(m, &blacklist_head, listq) {
395 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
396 (uintmax_t)m->phys_addr);
399 error = sbuf_finish(&sbuf);
405 vm_page_domain_init(int domain)
407 struct vm_domain *vmd;
408 struct vm_pagequeue *pq;
411 vmd = VM_DOMAIN(domain);
412 bzero(vmd, sizeof(*vmd));
413 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
414 "vm inactive pagequeue";
415 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
416 "vm active pagequeue";
417 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
418 "vm laundry pagequeue";
419 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
420 "vm unswappable pagequeue";
421 vmd->vmd_domain = domain;
422 vmd->vmd_page_count = 0;
423 vmd->vmd_free_count = 0;
425 vmd->vmd_oom = FALSE;
426 for (i = 0; i < PQ_COUNT; i++) {
427 pq = &vmd->vmd_pagequeues[i];
428 TAILQ_INIT(&pq->pq_pl);
429 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
430 MTX_DEF | MTX_DUPOK);
432 mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
436 * Initialize a physical page in preparation for adding it to the free
440 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
445 m->busy_lock = VPB_UNBUSIED;
452 m->order = VM_NFREEORDER;
453 m->pool = VM_FREEPOOL_DEFAULT;
454 m->valid = m->dirty = 0;
461 * Initializes the resident memory module. Allocates physical memory for
462 * bootstrapping UMA and some data structures that are used to manage
463 * physical pages. Initializes these structures, and populates the free
467 vm_page_startup(vm_offset_t vaddr)
469 struct vm_phys_seg *seg;
471 char *list, *listend;
473 vm_paddr_t end, high_avail, low_avail, new_end, page_range, size;
474 vm_paddr_t biggestsize, last_pa, pa;
476 int biggestone, i, segind;
480 vaddr = round_page(vaddr);
482 for (i = 0; phys_avail[i + 1]; i += 2) {
483 phys_avail[i] = round_page(phys_avail[i]);
484 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
486 for (i = 0; phys_avail[i + 1]; i += 2) {
487 size = phys_avail[i + 1] - phys_avail[i];
488 if (size > biggestsize) {
494 end = phys_avail[biggestone+1];
497 * Initialize the page and queue locks.
499 mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
500 for (i = 0; i < PA_LOCK_COUNT; i++)
501 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
502 for (i = 0; i < vm_ndomains; i++)
503 vm_page_domain_init(i);
506 * Allocate memory for use when boot strapping the kernel memory
507 * allocator. Tell UMA how many zones we are going to create
508 * before going fully functional. UMA will add its zones.
510 * VM startup zones: vmem, vmem_btag, VM OBJECT, RADIX NODE, MAP,
511 * KMAP ENTRY, MAP ENTRY, VMSPACE.
513 boot_pages = uma_startup_count(8);
515 #ifndef UMA_MD_SMALL_ALLOC
516 /* vmem_startup() calls uma_prealloc(). */
517 boot_pages += vmem_startup_count();
518 /* vm_map_startup() calls uma_prealloc(). */
519 boot_pages += howmany(MAX_KMAP,
520 UMA_SLAB_SPACE / sizeof(struct vm_map));
523 * Before going fully functional kmem_init() does allocation
524 * from "KMAP ENTRY" and vmem_create() does allocation from "vmem".
529 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
530 * manually fetch the value.
532 TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
533 new_end = end - (boot_pages * UMA_SLAB_SIZE);
534 new_end = trunc_page(new_end);
535 mapped = pmap_map(&vaddr, new_end, end,
536 VM_PROT_READ | VM_PROT_WRITE);
537 bzero((void *)mapped, end - new_end);
538 uma_startup((void *)mapped, boot_pages);
540 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
541 defined(__i386__) || defined(__mips__)
543 * Allocate a bitmap to indicate that a random physical page
544 * needs to be included in a minidump.
546 * The amd64 port needs this to indicate which direct map pages
547 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
549 * However, i386 still needs this workspace internally within the
550 * minidump code. In theory, they are not needed on i386, but are
551 * included should the sf_buf code decide to use them.
554 for (i = 0; dump_avail[i + 1] != 0; i += 2)
555 if (dump_avail[i + 1] > last_pa)
556 last_pa = dump_avail[i + 1];
557 page_range = last_pa / PAGE_SIZE;
558 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
559 new_end -= vm_page_dump_size;
560 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
561 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
562 bzero((void *)vm_page_dump, vm_page_dump_size);
566 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
568 * Include the UMA bootstrap pages and vm_page_dump in a crash dump.
569 * When pmap_map() uses the direct map, they are not automatically
572 for (pa = new_end; pa < end; pa += PAGE_SIZE)
575 phys_avail[biggestone + 1] = new_end;
578 * Request that the physical pages underlying the message buffer be
579 * included in a crash dump. Since the message buffer is accessed
580 * through the direct map, they are not automatically included.
582 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
583 last_pa = pa + round_page(msgbufsize);
584 while (pa < last_pa) {
590 * Compute the number of pages of memory that will be available for
591 * use, taking into account the overhead of a page structure per page.
592 * In other words, solve
593 * "available physical memory" - round_page(page_range *
594 * sizeof(struct vm_page)) = page_range * PAGE_SIZE
597 low_avail = phys_avail[0];
598 high_avail = phys_avail[1];
599 for (i = 0; i < vm_phys_nsegs; i++) {
600 if (vm_phys_segs[i].start < low_avail)
601 low_avail = vm_phys_segs[i].start;
602 if (vm_phys_segs[i].end > high_avail)
603 high_avail = vm_phys_segs[i].end;
605 /* Skip the first chunk. It is already accounted for. */
606 for (i = 2; phys_avail[i + 1] != 0; i += 2) {
607 if (phys_avail[i] < low_avail)
608 low_avail = phys_avail[i];
609 if (phys_avail[i + 1] > high_avail)
610 high_avail = phys_avail[i + 1];
612 first_page = low_avail / PAGE_SIZE;
613 #ifdef VM_PHYSSEG_SPARSE
615 for (i = 0; i < vm_phys_nsegs; i++)
616 size += vm_phys_segs[i].end - vm_phys_segs[i].start;
617 for (i = 0; phys_avail[i + 1] != 0; i += 2)
618 size += phys_avail[i + 1] - phys_avail[i];
619 #elif defined(VM_PHYSSEG_DENSE)
620 size = high_avail - low_avail;
622 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
625 #ifdef VM_PHYSSEG_DENSE
627 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
628 * the overhead of a page structure per page only if vm_page_array is
629 * allocated from the last physical memory chunk. Otherwise, we must
630 * allocate page structures representing the physical memory
631 * underlying vm_page_array, even though they will not be used.
633 if (new_end != high_avail)
634 page_range = size / PAGE_SIZE;
638 page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
641 * If the partial bytes remaining are large enough for
642 * a page (PAGE_SIZE) without a corresponding
643 * 'struct vm_page', then new_end will contain an
644 * extra page after subtracting the length of the VM
645 * page array. Compensate by subtracting an extra
648 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
649 if (new_end == high_avail)
650 high_avail -= PAGE_SIZE;
651 new_end -= PAGE_SIZE;
657 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
658 * However, because this page is allocated from KVM, out-of-bounds
659 * accesses using the direct map will not be trapped.
664 * Allocate physical memory for the page structures, and map it.
666 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
667 mapped = pmap_map(&vaddr, new_end, end,
668 VM_PROT_READ | VM_PROT_WRITE);
669 vm_page_array = (vm_page_t)mapped;
670 vm_page_array_size = page_range;
672 #if VM_NRESERVLEVEL > 0
674 * Allocate physical memory for the reservation management system's
675 * data structures, and map it.
677 if (high_avail == end)
678 high_avail = new_end;
679 new_end = vm_reserv_startup(&vaddr, new_end, high_avail);
681 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
683 * Include vm_page_array and vm_reserv_array in a crash dump.
685 for (pa = new_end; pa < end; pa += PAGE_SIZE)
688 phys_avail[biggestone + 1] = new_end;
691 * Add physical memory segments corresponding to the available
694 for (i = 0; phys_avail[i + 1] != 0; i += 2)
695 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
698 * Initialize the physical memory allocator.
703 * Initialize the page structures and add every available page to the
704 * physical memory allocator's free lists.
706 vm_cnt.v_page_count = 0;
707 for (segind = 0; segind < vm_phys_nsegs; segind++) {
708 seg = &vm_phys_segs[segind];
709 for (m = seg->first_page, pa = seg->start; pa < seg->end;
710 m++, pa += PAGE_SIZE)
711 vm_page_init_page(m, pa, segind);
714 * Add the segment to the free lists only if it is covered by
715 * one of the ranges in phys_avail. Because we've added the
716 * ranges to the vm_phys_segs array, we can assume that each
717 * segment is either entirely contained in one of the ranges,
718 * or doesn't overlap any of them.
720 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
721 struct vm_domain *vmd;
723 if (seg->start < phys_avail[i] ||
724 seg->end > phys_avail[i + 1])
728 pagecount = (u_long)atop(seg->end - seg->start);
730 vmd = VM_DOMAIN(seg->domain);
731 vm_domain_free_lock(vmd);
732 vm_phys_free_contig(m, pagecount);
733 vm_domain_freecnt_adj(vmd, (int)pagecount);
734 vm_domain_free_unlock(vmd);
735 vm_cnt.v_page_count += (u_int)pagecount;
737 vmd = VM_DOMAIN(seg->domain);
738 vmd->vmd_page_count += (u_int)pagecount;
739 vmd->vmd_segs |= 1UL << m->segind;
745 * Remove blacklisted pages from the physical memory allocator.
747 TAILQ_INIT(&blacklist_head);
748 vm_page_blacklist_load(&list, &listend);
749 vm_page_blacklist_check(list, listend);
751 list = kern_getenv("vm.blacklist");
752 vm_page_blacklist_check(list, NULL);
755 #if VM_NRESERVLEVEL > 0
757 * Initialize the reservation management system.
762 * Set an initial domain policy for thread0 so that allocations
771 vm_page_reference(vm_page_t m)
774 vm_page_aflag_set(m, PGA_REFERENCED);
778 * vm_page_busy_downgrade:
780 * Downgrade an exclusive busy page into a single shared busy page.
783 vm_page_busy_downgrade(vm_page_t m)
788 vm_page_assert_xbusied(m);
789 locked = mtx_owned(vm_page_lockptr(m));
793 x &= VPB_BIT_WAITERS;
794 if (x != 0 && !locked)
796 if (atomic_cmpset_rel_int(&m->busy_lock,
797 VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1)))
799 if (x != 0 && !locked)
812 * Return a positive value if the page is shared busied, 0 otherwise.
815 vm_page_sbusied(vm_page_t m)
820 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
826 * Shared unbusy a page.
829 vm_page_sunbusy(vm_page_t m)
833 vm_page_lock_assert(m, MA_NOTOWNED);
834 vm_page_assert_sbusied(m);
838 if (VPB_SHARERS(x) > 1) {
839 if (atomic_cmpset_int(&m->busy_lock, x,
844 if ((x & VPB_BIT_WAITERS) == 0) {
845 KASSERT(x == VPB_SHARERS_WORD(1),
846 ("vm_page_sunbusy: invalid lock state"));
847 if (atomic_cmpset_int(&m->busy_lock,
848 VPB_SHARERS_WORD(1), VPB_UNBUSIED))
852 KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS),
853 ("vm_page_sunbusy: invalid lock state for waiters"));
856 if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) {
867 * vm_page_busy_sleep:
869 * Sleep and release the page lock, using the page pointer as wchan.
870 * This is used to implement the hard-path of busying mechanism.
872 * The given page must be locked.
874 * If nonshared is true, sleep only if the page is xbusy.
877 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
881 vm_page_assert_locked(m);
884 if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) ||
885 ((x & VPB_BIT_WAITERS) == 0 &&
886 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) {
890 msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0);
896 * Try to shared busy a page.
897 * If the operation succeeds 1 is returned otherwise 0.
898 * The operation never sleeps.
901 vm_page_trysbusy(vm_page_t m)
907 if ((x & VPB_BIT_SHARED) == 0)
909 if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER))
915 vm_page_xunbusy_locked(vm_page_t m)
918 vm_page_assert_xbusied(m);
919 vm_page_assert_locked(m);
921 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
922 /* There is a waiter, do wakeup() instead of vm_page_flash(). */
927 vm_page_xunbusy_maybelocked(vm_page_t m)
931 vm_page_assert_xbusied(m);
934 * Fast path for unbusy. If it succeeds, we know that there
935 * are no waiters, so we do not need a wakeup.
937 if (atomic_cmpset_rel_int(&m->busy_lock, VPB_SINGLE_EXCLUSIVER,
941 lockacq = !mtx_owned(vm_page_lockptr(m));
944 vm_page_xunbusy_locked(m);
950 * vm_page_xunbusy_hard:
952 * Called after the first try the exclusive unbusy of a page failed.
953 * It is assumed that the waiters bit is on.
956 vm_page_xunbusy_hard(vm_page_t m)
959 vm_page_assert_xbusied(m);
962 vm_page_xunbusy_locked(m);
969 * Wakeup anyone waiting for the page.
970 * The ownership bits do not change.
972 * The given page must be locked.
975 vm_page_flash(vm_page_t m)
979 vm_page_lock_assert(m, MA_OWNED);
983 if ((x & VPB_BIT_WAITERS) == 0)
985 if (atomic_cmpset_int(&m->busy_lock, x,
986 x & (~VPB_BIT_WAITERS)))
993 * Avoid releasing and reacquiring the same page lock.
996 vm_page_change_lock(vm_page_t m, struct mtx **mtx)
1000 mtx1 = vm_page_lockptr(m);
1010 * Keep page from being freed by the page daemon
1011 * much of the same effect as wiring, except much lower
1012 * overhead and should be used only for *very* temporary
1013 * holding ("wiring").
1016 vm_page_hold(vm_page_t mem)
1019 vm_page_lock_assert(mem, MA_OWNED);
1024 vm_page_unhold(vm_page_t mem)
1027 vm_page_lock_assert(mem, MA_OWNED);
1028 KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!"));
1030 if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0)
1031 vm_page_free_toq(mem);
1035 * vm_page_unhold_pages:
1037 * Unhold each of the pages that is referenced by the given array.
1040 vm_page_unhold_pages(vm_page_t *ma, int count)
1045 for (; count != 0; count--) {
1046 vm_page_change_lock(*ma, &mtx);
1047 vm_page_unhold(*ma);
1055 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1059 #ifdef VM_PHYSSEG_SPARSE
1060 m = vm_phys_paddr_to_vm_page(pa);
1062 m = vm_phys_fictitious_to_vm_page(pa);
1064 #elif defined(VM_PHYSSEG_DENSE)
1068 if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1069 m = &vm_page_array[pi - first_page];
1072 return (vm_phys_fictitious_to_vm_page(pa));
1074 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1081 * Create a fictitious page with the specified physical address and
1082 * memory attribute. The memory attribute is the only the machine-
1083 * dependent aspect of a fictitious page that must be initialized.
1086 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1090 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1091 vm_page_initfake(m, paddr, memattr);
1096 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1099 if ((m->flags & PG_FICTITIOUS) != 0) {
1101 * The page's memattr might have changed since the
1102 * previous initialization. Update the pmap to the
1107 m->phys_addr = paddr;
1109 /* Fictitious pages don't use "segind". */
1110 m->flags = PG_FICTITIOUS;
1111 /* Fictitious pages don't use "order" or "pool". */
1112 m->oflags = VPO_UNMANAGED;
1113 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1117 pmap_page_set_memattr(m, memattr);
1123 * Release a fictitious page.
1126 vm_page_putfake(vm_page_t m)
1129 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1130 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1131 ("vm_page_putfake: bad page %p", m));
1132 uma_zfree(fakepg_zone, m);
1136 * vm_page_updatefake:
1138 * Update the given fictitious page to the specified physical address and
1142 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1145 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1146 ("vm_page_updatefake: bad page %p", m));
1147 m->phys_addr = paddr;
1148 pmap_page_set_memattr(m, memattr);
1157 vm_page_free(vm_page_t m)
1160 m->flags &= ~PG_ZERO;
1161 vm_page_free_toq(m);
1165 * vm_page_free_zero:
1167 * Free a page to the zerod-pages queue
1170 vm_page_free_zero(vm_page_t m)
1173 m->flags |= PG_ZERO;
1174 vm_page_free_toq(m);
1178 * Unbusy and handle the page queueing for a page from a getpages request that
1179 * was optionally read ahead or behind.
1182 vm_page_readahead_finish(vm_page_t m)
1185 /* We shouldn't put invalid pages on queues. */
1186 KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m));
1189 * Since the page is not the actually needed one, whether it should
1190 * be activated or deactivated is not obvious. Empirical results
1191 * have shown that deactivating the page is usually the best choice,
1192 * unless the page is wanted by another thread.
1195 if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1196 vm_page_activate(m);
1198 vm_page_deactivate(m);
1204 * vm_page_sleep_if_busy:
1206 * Sleep and release the page queues lock if the page is busied.
1207 * Returns TRUE if the thread slept.
1209 * The given page must be unlocked and object containing it must
1213 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1217 vm_page_lock_assert(m, MA_NOTOWNED);
1218 VM_OBJECT_ASSERT_WLOCKED(m->object);
1220 if (vm_page_busied(m)) {
1222 * The page-specific object must be cached because page
1223 * identity can change during the sleep, causing the
1224 * re-lock of a different object.
1225 * It is assumed that a reference to the object is already
1226 * held by the callers.
1230 VM_OBJECT_WUNLOCK(obj);
1231 vm_page_busy_sleep(m, msg, false);
1232 VM_OBJECT_WLOCK(obj);
1239 * vm_page_dirty_KBI: [ internal use only ]
1241 * Set all bits in the page's dirty field.
1243 * The object containing the specified page must be locked if the
1244 * call is made from the machine-independent layer.
1246 * See vm_page_clear_dirty_mask().
1248 * This function should only be called by vm_page_dirty().
1251 vm_page_dirty_KBI(vm_page_t m)
1254 /* Refer to this operation by its public name. */
1255 KASSERT(m->valid == VM_PAGE_BITS_ALL,
1256 ("vm_page_dirty: page is invalid!"));
1257 m->dirty = VM_PAGE_BITS_ALL;
1261 * vm_page_insert: [ internal use only ]
1263 * Inserts the given mem entry into the object and object list.
1265 * The object must be locked.
1268 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1272 VM_OBJECT_ASSERT_WLOCKED(object);
1273 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1274 return (vm_page_insert_after(m, object, pindex, mpred));
1278 * vm_page_insert_after:
1280 * Inserts the page "m" into the specified object at offset "pindex".
1282 * The page "mpred" must immediately precede the offset "pindex" within
1283 * the specified object.
1285 * The object must be locked.
1288 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1293 VM_OBJECT_ASSERT_WLOCKED(object);
1294 KASSERT(m->object == NULL,
1295 ("vm_page_insert_after: page already inserted"));
1296 if (mpred != NULL) {
1297 KASSERT(mpred->object == object,
1298 ("vm_page_insert_after: object doesn't contain mpred"));
1299 KASSERT(mpred->pindex < pindex,
1300 ("vm_page_insert_after: mpred doesn't precede pindex"));
1301 msucc = TAILQ_NEXT(mpred, listq);
1303 msucc = TAILQ_FIRST(&object->memq);
1305 KASSERT(msucc->pindex > pindex,
1306 ("vm_page_insert_after: msucc doesn't succeed pindex"));
1309 * Record the object/offset pair in this page
1315 * Now link into the object's ordered list of backed pages.
1317 if (vm_radix_insert(&object->rtree, m)) {
1322 vm_page_insert_radixdone(m, object, mpred);
1327 * vm_page_insert_radixdone:
1329 * Complete page "m" insertion into the specified object after the
1330 * radix trie hooking.
1332 * The page "mpred" must precede the offset "m->pindex" within the
1335 * The object must be locked.
1338 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1341 VM_OBJECT_ASSERT_WLOCKED(object);
1342 KASSERT(object != NULL && m->object == object,
1343 ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1344 if (mpred != NULL) {
1345 KASSERT(mpred->object == object,
1346 ("vm_page_insert_after: object doesn't contain mpred"));
1347 KASSERT(mpred->pindex < m->pindex,
1348 ("vm_page_insert_after: mpred doesn't precede pindex"));
1352 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1354 TAILQ_INSERT_HEAD(&object->memq, m, listq);
1357 * Show that the object has one more resident page.
1359 object->resident_page_count++;
1362 * Hold the vnode until the last page is released.
1364 if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1365 vhold(object->handle);
1368 * Since we are inserting a new and possibly dirty page,
1369 * update the object's OBJ_MIGHTBEDIRTY flag.
1371 if (pmap_page_is_write_mapped(m))
1372 vm_object_set_writeable_dirty(object);
1378 * Removes the specified page from its containing object, but does not
1379 * invalidate any backing storage.
1381 * The object must be locked. The page must be locked if it is managed.
1384 vm_page_remove(vm_page_t m)
1389 if ((m->oflags & VPO_UNMANAGED) == 0)
1390 vm_page_assert_locked(m);
1391 if ((object = m->object) == NULL)
1393 VM_OBJECT_ASSERT_WLOCKED(object);
1394 if (vm_page_xbusied(m))
1395 vm_page_xunbusy_maybelocked(m);
1396 mrem = vm_radix_remove(&object->rtree, m->pindex);
1397 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1400 * Now remove from the object's list of backed pages.
1402 TAILQ_REMOVE(&object->memq, m, listq);
1405 * And show that the object has one fewer resident page.
1407 object->resident_page_count--;
1410 * The vnode may now be recycled.
1412 if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1413 vdrop(object->handle);
1421 * Returns the page associated with the object/offset
1422 * pair specified; if none is found, NULL is returned.
1424 * The object must be locked.
1427 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1430 VM_OBJECT_ASSERT_LOCKED(object);
1431 return (vm_radix_lookup(&object->rtree, pindex));
1435 * vm_page_find_least:
1437 * Returns the page associated with the object with least pindex
1438 * greater than or equal to the parameter pindex, or NULL.
1440 * The object must be locked.
1443 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1447 VM_OBJECT_ASSERT_LOCKED(object);
1448 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1449 m = vm_radix_lookup_ge(&object->rtree, pindex);
1454 * Returns the given page's successor (by pindex) within the object if it is
1455 * resident; if none is found, NULL is returned.
1457 * The object must be locked.
1460 vm_page_next(vm_page_t m)
1464 VM_OBJECT_ASSERT_LOCKED(m->object);
1465 if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1466 MPASS(next->object == m->object);
1467 if (next->pindex != m->pindex + 1)
1474 * Returns the given page's predecessor (by pindex) within the object if it is
1475 * resident; if none is found, NULL is returned.
1477 * The object must be locked.
1480 vm_page_prev(vm_page_t m)
1484 VM_OBJECT_ASSERT_LOCKED(m->object);
1485 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1486 MPASS(prev->object == m->object);
1487 if (prev->pindex != m->pindex - 1)
1494 * Uses the page mnew as a replacement for an existing page at index
1495 * pindex which must be already present in the object.
1497 * The existing page must not be on a paging queue.
1500 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1504 VM_OBJECT_ASSERT_WLOCKED(object);
1505 KASSERT(mnew->object == NULL,
1506 ("vm_page_replace: page already in object"));
1509 * This function mostly follows vm_page_insert() and
1510 * vm_page_remove() without the radix, object count and vnode
1511 * dance. Double check such functions for more comments.
1514 mnew->object = object;
1515 mnew->pindex = pindex;
1516 mold = vm_radix_replace(&object->rtree, mnew);
1517 KASSERT(mold->queue == PQ_NONE,
1518 ("vm_page_replace: mold is on a paging queue"));
1520 /* Keep the resident page list in sorted order. */
1521 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1522 TAILQ_REMOVE(&object->memq, mold, listq);
1524 mold->object = NULL;
1525 vm_page_xunbusy_maybelocked(mold);
1528 * The object's resident_page_count does not change because we have
1529 * swapped one page for another, but OBJ_MIGHTBEDIRTY.
1531 if (pmap_page_is_write_mapped(mnew))
1532 vm_object_set_writeable_dirty(object);
1539 * Move the given memory entry from its
1540 * current object to the specified target object/offset.
1542 * Note: swap associated with the page must be invalidated by the move. We
1543 * have to do this for several reasons: (1) we aren't freeing the
1544 * page, (2) we are dirtying the page, (3) the VM system is probably
1545 * moving the page from object A to B, and will then later move
1546 * the backing store from A to B and we can't have a conflict.
1548 * Note: we *always* dirty the page. It is necessary both for the
1549 * fact that we moved it, and because we may be invalidating
1552 * The objects must be locked.
1555 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1560 VM_OBJECT_ASSERT_WLOCKED(new_object);
1562 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1563 KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1564 ("vm_page_rename: pindex already renamed"));
1567 * Create a custom version of vm_page_insert() which does not depend
1568 * by m_prev and can cheat on the implementation aspects of the
1572 m->pindex = new_pindex;
1573 if (vm_radix_insert(&new_object->rtree, m)) {
1579 * The operation cannot fail anymore. The removal must happen before
1580 * the listq iterator is tainted.
1586 /* Return back to the new pindex to complete vm_page_insert(). */
1587 m->pindex = new_pindex;
1588 m->object = new_object;
1590 vm_page_insert_radixdone(m, new_object, mpred);
1598 * Allocate and return a page that is associated with the specified
1599 * object and offset pair. By default, this page is exclusive busied.
1601 * The caller must always specify an allocation class.
1603 * allocation classes:
1604 * VM_ALLOC_NORMAL normal process request
1605 * VM_ALLOC_SYSTEM system *really* needs a page
1606 * VM_ALLOC_INTERRUPT interrupt time request
1608 * optional allocation flags:
1609 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1610 * intends to allocate
1611 * VM_ALLOC_NOBUSY do not exclusive busy the page
1612 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1613 * VM_ALLOC_NOOBJ page is not associated with an object and
1614 * should not be exclusive busy
1615 * VM_ALLOC_SBUSY shared busy the allocated page
1616 * VM_ALLOC_WIRED wire the allocated page
1617 * VM_ALLOC_ZERO prefer a zeroed page
1620 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1623 return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1624 vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1628 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1632 return (vm_page_alloc_domain_after(object, pindex, domain, req,
1633 object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) :
1638 * Allocate a page in the specified object with the given page index. To
1639 * optimize insertion of the page into the object, the caller must also specifiy
1640 * the resident page in the object with largest index smaller than the given
1641 * page index, or NULL if no such page exists.
1644 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1645 int req, vm_page_t mpred)
1647 struct vm_domainset_iter di;
1651 vm_domainset_iter_page_init(&di, object, &domain, &req);
1653 m = vm_page_alloc_domain_after(object, pindex, domain, req,
1657 } while (vm_domainset_iter_page(&di, &domain, &req) == 0);
1663 * Returns true if the number of free pages exceeds the minimum
1664 * for the request class and false otherwise.
1667 vm_domain_available(struct vm_domain *vmd, int req, int npages)
1670 vm_domain_free_assert_locked(vmd);
1671 req = req & VM_ALLOC_CLASS_MASK;
1674 * The page daemon is allowed to dig deeper into the free page list.
1676 if (curproc == pageproc && req != VM_ALLOC_INTERRUPT)
1677 req = VM_ALLOC_SYSTEM;
1679 if (vmd->vmd_free_count >= npages + vmd->vmd_free_reserved ||
1680 (req == VM_ALLOC_SYSTEM &&
1681 vmd->vmd_free_count >= npages + vmd->vmd_interrupt_free_min) ||
1682 (req == VM_ALLOC_INTERRUPT &&
1683 vmd->vmd_free_count >= npages))
1690 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
1691 int req, vm_page_t mpred)
1693 struct vm_domain *vmd;
1698 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1699 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1700 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1701 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1702 ("inconsistent object(%p)/req(%x)", object, req));
1703 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
1704 ("Can't sleep and retry object insertion."));
1705 KASSERT(mpred == NULL || mpred->pindex < pindex,
1706 ("mpred %p doesn't precede pindex 0x%jx", mpred,
1707 (uintmax_t)pindex));
1709 VM_OBJECT_ASSERT_WLOCKED(object);
1713 #if VM_NRESERVLEVEL > 0
1714 if (vm_object_reserv(object) &&
1715 (m = vm_reserv_extend(req, object, pindex, domain, mpred))
1717 domain = vm_phys_domain(m);
1718 vmd = VM_DOMAIN(domain);
1722 vmd = VM_DOMAIN(domain);
1723 vm_domain_free_lock(vmd);
1724 if (vm_domain_available(vmd, req, 1)) {
1726 * Can we allocate the page from a reservation?
1728 #if VM_NRESERVLEVEL > 0
1729 if (!vm_object_reserv(object) ||
1730 (m = vm_reserv_alloc_page(object, pindex,
1731 domain, mpred)) == NULL)
1735 * If not, allocate it from the free page queues.
1737 m = vm_phys_alloc_pages(domain, object != NULL ?
1738 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1739 #if VM_NRESERVLEVEL > 0
1740 if (m == NULL && vm_reserv_reclaim_inactive(domain)) {
1741 m = vm_phys_alloc_pages(domain,
1743 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1751 * Not allocatable, give up.
1753 if (vm_domain_alloc_fail(vmd, object, req))
1759 * At this point we had better have found a good page.
1761 KASSERT(m != NULL, ("missing page"));
1762 free_count = vm_domain_freecnt_adj(vmd, -1);
1763 vm_domain_free_unlock(vmd);
1766 * Don't wakeup too often - wakeup the pageout daemon when
1767 * we would be nearly out of memory.
1769 if (vm_paging_needed(vmd, free_count))
1770 pagedaemon_wakeup(vmd->vmd_domain);
1771 #if VM_NRESERVLEVEL > 0
1774 vm_page_alloc_check(m);
1777 * Initialize the page. Only the PG_ZERO flag is inherited.
1780 if ((req & VM_ALLOC_ZERO) != 0)
1783 if ((req & VM_ALLOC_NODUMP) != 0)
1787 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1789 m->busy_lock = VPB_UNBUSIED;
1790 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1791 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1792 if ((req & VM_ALLOC_SBUSY) != 0)
1793 m->busy_lock = VPB_SHARERS_WORD(1);
1794 if (req & VM_ALLOC_WIRED) {
1796 * The page lock is not required for wiring a page until that
1797 * page is inserted into the object.
1804 if (object != NULL) {
1805 if (vm_page_insert_after(m, object, pindex, mpred)) {
1806 pagedaemon_wakeup(domain);
1807 if (req & VM_ALLOC_WIRED) {
1811 KASSERT(m->object == NULL, ("page %p has object", m));
1812 m->oflags = VPO_UNMANAGED;
1813 m->busy_lock = VPB_UNBUSIED;
1814 /* Don't change PG_ZERO. */
1815 vm_page_free_toq(m);
1816 if (req & VM_ALLOC_WAITFAIL) {
1817 VM_OBJECT_WUNLOCK(object);
1819 VM_OBJECT_WLOCK(object);
1824 /* Ignore device objects; the pager sets "memattr" for them. */
1825 if (object->memattr != VM_MEMATTR_DEFAULT &&
1826 (object->flags & OBJ_FICTITIOUS) == 0)
1827 pmap_page_set_memattr(m, object->memattr);
1835 * vm_page_alloc_contig:
1837 * Allocate a contiguous set of physical pages of the given size "npages"
1838 * from the free lists. All of the physical pages must be at or above
1839 * the given physical address "low" and below the given physical address
1840 * "high". The given value "alignment" determines the alignment of the
1841 * first physical page in the set. If the given value "boundary" is
1842 * non-zero, then the set of physical pages cannot cross any physical
1843 * address boundary that is a multiple of that value. Both "alignment"
1844 * and "boundary" must be a power of two.
1846 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1847 * then the memory attribute setting for the physical pages is configured
1848 * to the object's memory attribute setting. Otherwise, the memory
1849 * attribute setting for the physical pages is configured to "memattr",
1850 * overriding the object's memory attribute setting. However, if the
1851 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1852 * memory attribute setting for the physical pages cannot be configured
1853 * to VM_MEMATTR_DEFAULT.
1855 * The specified object may not contain fictitious pages.
1857 * The caller must always specify an allocation class.
1859 * allocation classes:
1860 * VM_ALLOC_NORMAL normal process request
1861 * VM_ALLOC_SYSTEM system *really* needs a page
1862 * VM_ALLOC_INTERRUPT interrupt time request
1864 * optional allocation flags:
1865 * VM_ALLOC_NOBUSY do not exclusive busy the page
1866 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1867 * VM_ALLOC_NOOBJ page is not associated with an object and
1868 * should not be exclusive busy
1869 * VM_ALLOC_SBUSY shared busy the allocated page
1870 * VM_ALLOC_WIRED wire the allocated page
1871 * VM_ALLOC_ZERO prefer a zeroed page
1874 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1875 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1876 vm_paddr_t boundary, vm_memattr_t memattr)
1878 struct vm_domainset_iter di;
1882 vm_domainset_iter_page_init(&di, object, &domain, &req);
1884 m = vm_page_alloc_contig_domain(object, pindex, domain, req,
1885 npages, low, high, alignment, boundary, memattr);
1888 } while (vm_domainset_iter_page(&di, &domain, &req) == 0);
1894 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1895 int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1896 vm_paddr_t boundary, vm_memattr_t memattr)
1898 struct vm_domain *vmd;
1899 vm_page_t m, m_ret, mpred;
1900 u_int busy_lock, flags, oflags;
1902 mpred = NULL; /* XXX: pacify gcc */
1903 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1904 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1905 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1906 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1907 ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
1909 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
1910 ("Can't sleep and retry object insertion."));
1911 if (object != NULL) {
1912 VM_OBJECT_ASSERT_WLOCKED(object);
1913 KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
1914 ("vm_page_alloc_contig: object %p has fictitious pages",
1917 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
1919 if (object != NULL) {
1920 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1921 KASSERT(mpred == NULL || mpred->pindex != pindex,
1922 ("vm_page_alloc_contig: pindex already allocated"));
1926 * Can we allocate the pages without the number of free pages falling
1927 * below the lower bound for the allocation class?
1930 #if VM_NRESERVLEVEL > 0
1931 if (vm_object_reserv(object) &&
1932 (m_ret = vm_reserv_extend_contig(req, object, pindex, domain,
1933 npages, low, high, alignment, boundary, mpred)) != NULL) {
1934 domain = vm_phys_domain(m_ret);
1935 vmd = VM_DOMAIN(domain);
1940 vmd = VM_DOMAIN(domain);
1941 vm_domain_free_lock(vmd);
1942 if (vm_domain_available(vmd, req, npages)) {
1944 * Can we allocate the pages from a reservation?
1946 #if VM_NRESERVLEVEL > 0
1948 if (!vm_object_reserv(object) ||
1949 (m_ret = vm_reserv_alloc_contig(object, pindex, domain,
1950 npages, low, high, alignment, boundary, mpred)) == NULL)
1953 * If not, allocate them from the free page queues.
1955 m_ret = vm_phys_alloc_contig(domain, npages, low, high,
1956 alignment, boundary);
1957 #if VM_NRESERVLEVEL > 0
1958 if (m_ret == NULL && vm_reserv_reclaim_contig(
1959 domain, npages, low, high, alignment, boundary))
1963 if (m_ret == NULL) {
1964 if (vm_domain_alloc_fail(vmd, object, req))
1968 vm_domain_freecnt_adj(vmd, -npages);
1969 vm_domain_free_unlock(vmd);
1970 #if VM_NRESERVLEVEL > 0
1973 for (m = m_ret; m < &m_ret[npages]; m++)
1974 vm_page_alloc_check(m);
1977 * Initialize the pages. Only the PG_ZERO flag is inherited.
1980 if ((req & VM_ALLOC_ZERO) != 0)
1982 if ((req & VM_ALLOC_NODUMP) != 0)
1984 oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1986 busy_lock = VPB_UNBUSIED;
1987 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1988 busy_lock = VPB_SINGLE_EXCLUSIVER;
1989 if ((req & VM_ALLOC_SBUSY) != 0)
1990 busy_lock = VPB_SHARERS_WORD(1);
1991 if ((req & VM_ALLOC_WIRED) != 0)
1992 vm_wire_add(npages);
1993 if (object != NULL) {
1994 if (object->memattr != VM_MEMATTR_DEFAULT &&
1995 memattr == VM_MEMATTR_DEFAULT)
1996 memattr = object->memattr;
1998 for (m = m_ret; m < &m_ret[npages]; m++) {
2000 m->flags = (m->flags | PG_NODUMP) & flags;
2001 m->busy_lock = busy_lock;
2002 if ((req & VM_ALLOC_WIRED) != 0)
2006 if (object != NULL) {
2007 if (vm_page_insert_after(m, object, pindex, mpred)) {
2008 pagedaemon_wakeup(domain);
2009 if ((req & VM_ALLOC_WIRED) != 0)
2010 vm_wire_sub(npages);
2011 KASSERT(m->object == NULL,
2012 ("page %p has object", m));
2014 for (m = m_ret; m < &m_ret[npages]; m++) {
2016 (req & VM_ALLOC_WIRED) != 0)
2018 m->oflags = VPO_UNMANAGED;
2019 m->busy_lock = VPB_UNBUSIED;
2020 /* Don't change PG_ZERO. */
2021 vm_page_free_toq(m);
2023 if (req & VM_ALLOC_WAITFAIL) {
2024 VM_OBJECT_WUNLOCK(object);
2026 VM_OBJECT_WLOCK(object);
2033 if (memattr != VM_MEMATTR_DEFAULT)
2034 pmap_page_set_memattr(m, memattr);
2037 vmd = VM_DOMAIN(domain);
2038 if (vm_paging_needed(vmd, vmd->vmd_free_count))
2039 pagedaemon_wakeup(domain);
2044 * Check a page that has been freshly dequeued from a freelist.
2047 vm_page_alloc_check(vm_page_t m)
2050 KASSERT(m->object == NULL, ("page %p has object", m));
2051 KASSERT(m->queue == PQ_NONE,
2052 ("page %p has unexpected queue %d", m, m->queue));
2053 KASSERT(!vm_page_held(m), ("page %p is held", m));
2054 KASSERT(!vm_page_busied(m), ("page %p is busy", m));
2055 KASSERT(m->dirty == 0, ("page %p is dirty", m));
2056 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2057 ("page %p has unexpected memattr %d",
2058 m, pmap_page_get_memattr(m)));
2059 KASSERT(m->valid == 0, ("free page %p is valid", m));
2063 * vm_page_alloc_freelist:
2065 * Allocate a physical page from the specified free page list.
2067 * The caller must always specify an allocation class.
2069 * allocation classes:
2070 * VM_ALLOC_NORMAL normal process request
2071 * VM_ALLOC_SYSTEM system *really* needs a page
2072 * VM_ALLOC_INTERRUPT interrupt time request
2074 * optional allocation flags:
2075 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
2076 * intends to allocate
2077 * VM_ALLOC_WIRED wire the allocated page
2078 * VM_ALLOC_ZERO prefer a zeroed page
2081 vm_page_alloc_freelist(int freelist, int req)
2083 struct vm_domainset_iter di;
2087 vm_domainset_iter_page_init(&di, kernel_object, &domain, &req);
2089 m = vm_page_alloc_freelist_domain(domain, freelist, req);
2092 } while (vm_domainset_iter_page(&di, &domain, &req) == 0);
2098 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2100 struct vm_domain *vmd;
2102 u_int flags, free_count;
2105 * Do not allocate reserved pages unless the req has asked for it.
2107 vmd = VM_DOMAIN(domain);
2109 vm_domain_free_lock(vmd);
2110 if (vm_domain_available(vmd, req, 1))
2111 m = vm_phys_alloc_freelist_pages(domain, freelist,
2112 VM_FREEPOOL_DIRECT, 0);
2114 if (vm_domain_alloc_fail(vmd, NULL, req))
2118 free_count = vm_domain_freecnt_adj(vmd, -1);
2119 vm_domain_free_unlock(vmd);
2120 vm_page_alloc_check(m);
2123 * Initialize the page. Only the PG_ZERO flag is inherited.
2127 if ((req & VM_ALLOC_ZERO) != 0)
2130 if ((req & VM_ALLOC_WIRED) != 0) {
2132 * The page lock is not required for wiring a page that does
2133 * not belong to an object.
2138 /* Unmanaged pages don't use "act_count". */
2139 m->oflags = VPO_UNMANAGED;
2140 if (vm_paging_needed(vmd, free_count))
2141 pagedaemon_wakeup(domain);
2145 #define VPSC_ANY 0 /* No restrictions. */
2146 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */
2147 #define VPSC_NOSUPER 2 /* Skip superpages. */
2150 * vm_page_scan_contig:
2152 * Scan vm_page_array[] between the specified entries "m_start" and
2153 * "m_end" for a run of contiguous physical pages that satisfy the
2154 * specified conditions, and return the lowest page in the run. The
2155 * specified "alignment" determines the alignment of the lowest physical
2156 * page in the run. If the specified "boundary" is non-zero, then the
2157 * run of physical pages cannot span a physical address that is a
2158 * multiple of "boundary".
2160 * "m_end" is never dereferenced, so it need not point to a vm_page
2161 * structure within vm_page_array[].
2163 * "npages" must be greater than zero. "m_start" and "m_end" must not
2164 * span a hole (or discontiguity) in the physical address space. Both
2165 * "alignment" and "boundary" must be a power of two.
2168 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2169 u_long alignment, vm_paddr_t boundary, int options)
2175 #if VM_NRESERVLEVEL > 0
2178 int m_inc, order, run_ext, run_len;
2180 KASSERT(npages > 0, ("npages is 0"));
2181 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2182 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2186 for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2187 KASSERT((m->flags & PG_MARKER) == 0,
2188 ("page %p is PG_MARKER", m));
2189 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->wire_count == 1,
2190 ("fictitious page %p has invalid wire count", m));
2193 * If the current page would be the start of a run, check its
2194 * physical address against the end, alignment, and boundary
2195 * conditions. If it doesn't satisfy these conditions, either
2196 * terminate the scan or advance to the next page that
2197 * satisfies the failed condition.
2200 KASSERT(m_run == NULL, ("m_run != NULL"));
2201 if (m + npages > m_end)
2203 pa = VM_PAGE_TO_PHYS(m);
2204 if ((pa & (alignment - 1)) != 0) {
2205 m_inc = atop(roundup2(pa, alignment) - pa);
2208 if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2210 m_inc = atop(roundup2(pa, boundary) - pa);
2214 KASSERT(m_run != NULL, ("m_run == NULL"));
2216 vm_page_change_lock(m, &m_mtx);
2219 if (vm_page_held(m))
2221 #if VM_NRESERVLEVEL > 0
2222 else if ((level = vm_reserv_level(m)) >= 0 &&
2223 (options & VPSC_NORESERV) != 0) {
2225 /* Advance to the end of the reservation. */
2226 pa = VM_PAGE_TO_PHYS(m);
2227 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2231 else if ((object = m->object) != NULL) {
2233 * The page is considered eligible for relocation if
2234 * and only if it could be laundered or reclaimed by
2237 if (!VM_OBJECT_TRYRLOCK(object)) {
2239 VM_OBJECT_RLOCK(object);
2241 if (m->object != object) {
2243 * The page may have been freed.
2245 VM_OBJECT_RUNLOCK(object);
2247 } else if (vm_page_held(m)) {
2252 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2253 ("page %p is PG_UNHOLDFREE", m));
2254 /* Don't care: PG_NODUMP, PG_ZERO. */
2255 if (object->type != OBJT_DEFAULT &&
2256 object->type != OBJT_SWAP &&
2257 object->type != OBJT_VNODE) {
2259 #if VM_NRESERVLEVEL > 0
2260 } else if ((options & VPSC_NOSUPER) != 0 &&
2261 (level = vm_reserv_level_iffullpop(m)) >= 0) {
2263 /* Advance to the end of the superpage. */
2264 pa = VM_PAGE_TO_PHYS(m);
2265 m_inc = atop(roundup2(pa + 1,
2266 vm_reserv_size(level)) - pa);
2268 } else if (object->memattr == VM_MEMATTR_DEFAULT &&
2269 m->queue != PQ_NONE && !vm_page_busied(m)) {
2271 * The page is allocated but eligible for
2272 * relocation. Extend the current run by one
2275 KASSERT(pmap_page_get_memattr(m) ==
2277 ("page %p has an unexpected memattr", m));
2278 KASSERT((m->oflags & (VPO_SWAPINPROG |
2279 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2280 ("page %p has unexpected oflags", m));
2281 /* Don't care: VPO_NOSYNC. */
2286 VM_OBJECT_RUNLOCK(object);
2287 #if VM_NRESERVLEVEL > 0
2288 } else if (level >= 0) {
2290 * The page is reserved but not yet allocated. In
2291 * other words, it is still free. Extend the current
2296 } else if ((order = m->order) < VM_NFREEORDER) {
2298 * The page is enqueued in the physical memory
2299 * allocator's free page queues. Moreover, it is the
2300 * first page in a power-of-two-sized run of
2301 * contiguous free pages. Add these pages to the end
2302 * of the current run, and jump ahead.
2304 run_ext = 1 << order;
2308 * Skip the page for one of the following reasons: (1)
2309 * It is enqueued in the physical memory allocator's
2310 * free page queues. However, it is not the first
2311 * page in a run of contiguous free pages. (This case
2312 * rarely occurs because the scan is performed in
2313 * ascending order.) (2) It is not reserved, and it is
2314 * transitioning from free to allocated. (Conversely,
2315 * the transition from allocated to free for managed
2316 * pages is blocked by the page lock.) (3) It is
2317 * allocated but not contained by an object and not
2318 * wired, e.g., allocated by Xen's balloon driver.
2324 * Extend or reset the current run of pages.
2339 if (run_len >= npages)
2345 * vm_page_reclaim_run:
2347 * Try to relocate each of the allocated virtual pages within the
2348 * specified run of physical pages to a new physical address. Free the
2349 * physical pages underlying the relocated virtual pages. A virtual page
2350 * is relocatable if and only if it could be laundered or reclaimed by
2351 * the page daemon. Whenever possible, a virtual page is relocated to a
2352 * physical address above "high".
2354 * Returns 0 if every physical page within the run was already free or
2355 * just freed by a successful relocation. Otherwise, returns a non-zero
2356 * value indicating why the last attempt to relocate a virtual page was
2359 * "req_class" must be an allocation class.
2362 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2365 struct vm_domain *vmd;
2367 struct spglist free;
2370 vm_page_t m, m_end, m_new;
2371 int error, order, req;
2373 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2374 ("req_class is not an allocation class"));
2378 m_end = m_run + npages;
2380 for (; error == 0 && m < m_end; m++) {
2381 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2382 ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2385 * Avoid releasing and reacquiring the same page lock.
2387 vm_page_change_lock(m, &m_mtx);
2389 if (vm_page_held(m))
2391 else if ((object = m->object) != NULL) {
2393 * The page is relocated if and only if it could be
2394 * laundered or reclaimed by the page daemon.
2396 if (!VM_OBJECT_TRYWLOCK(object)) {
2398 VM_OBJECT_WLOCK(object);
2400 if (m->object != object) {
2402 * The page may have been freed.
2404 VM_OBJECT_WUNLOCK(object);
2406 } else if (vm_page_held(m)) {
2411 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2412 ("page %p is PG_UNHOLDFREE", m));
2413 /* Don't care: PG_NODUMP, PG_ZERO. */
2414 if (object->type != OBJT_DEFAULT &&
2415 object->type != OBJT_SWAP &&
2416 object->type != OBJT_VNODE)
2418 else if (object->memattr != VM_MEMATTR_DEFAULT)
2420 else if (m->queue != PQ_NONE && !vm_page_busied(m)) {
2421 KASSERT(pmap_page_get_memattr(m) ==
2423 ("page %p has an unexpected memattr", m));
2424 KASSERT((m->oflags & (VPO_SWAPINPROG |
2425 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2426 ("page %p has unexpected oflags", m));
2427 /* Don't care: VPO_NOSYNC. */
2428 if (m->valid != 0) {
2430 * First, try to allocate a new page
2431 * that is above "high". Failing
2432 * that, try to allocate a new page
2433 * that is below "m_run". Allocate
2434 * the new page between the end of
2435 * "m_run" and "high" only as a last
2438 req = req_class | VM_ALLOC_NOOBJ;
2439 if ((m->flags & PG_NODUMP) != 0)
2440 req |= VM_ALLOC_NODUMP;
2441 if (trunc_page(high) !=
2442 ~(vm_paddr_t)PAGE_MASK) {
2443 m_new = vm_page_alloc_contig(
2448 VM_MEMATTR_DEFAULT);
2451 if (m_new == NULL) {
2452 pa = VM_PAGE_TO_PHYS(m_run);
2453 m_new = vm_page_alloc_contig(
2455 0, pa - 1, PAGE_SIZE, 0,
2456 VM_MEMATTR_DEFAULT);
2458 if (m_new == NULL) {
2460 m_new = vm_page_alloc_contig(
2462 pa, high, PAGE_SIZE, 0,
2463 VM_MEMATTR_DEFAULT);
2465 if (m_new == NULL) {
2469 KASSERT(m_new->wire_count == 0,
2470 ("page %p is wired", m));
2473 * Replace "m" with the new page. For
2474 * vm_page_replace(), "m" must be busy
2475 * and dequeued. Finally, change "m"
2476 * as if vm_page_free() was called.
2478 if (object->ref_count != 0)
2480 m_new->aflags = m->aflags;
2481 KASSERT(m_new->oflags == VPO_UNMANAGED,
2482 ("page %p is managed", m));
2483 m_new->oflags = m->oflags & VPO_NOSYNC;
2484 pmap_copy_page(m, m_new);
2485 m_new->valid = m->valid;
2486 m_new->dirty = m->dirty;
2487 m->flags &= ~PG_ZERO;
2490 vm_page_replace_checked(m_new, object,
2496 * The new page must be deactivated
2497 * before the object is unlocked.
2499 vm_page_change_lock(m_new, &m_mtx);
2500 vm_page_deactivate(m_new);
2502 m->flags &= ~PG_ZERO;
2505 KASSERT(m->dirty == 0,
2506 ("page %p is dirty", m));
2508 SLIST_INSERT_HEAD(&free, m, plinks.s.ss);
2512 VM_OBJECT_WUNLOCK(object);
2514 MPASS(vm_phys_domain(m) == domain);
2515 vmd = VM_DOMAIN(domain);
2516 vm_domain_free_lock(vmd);
2518 if (order < VM_NFREEORDER) {
2520 * The page is enqueued in the physical memory
2521 * allocator's free page queues. Moreover, it
2522 * is the first page in a power-of-two-sized
2523 * run of contiguous free pages. Jump ahead
2524 * to the last page within that run, and
2525 * continue from there.
2527 m += (1 << order) - 1;
2529 #if VM_NRESERVLEVEL > 0
2530 else if (vm_reserv_is_page_free(m))
2533 vm_domain_free_unlock(vmd);
2534 if (order == VM_NFREEORDER)
2540 if ((m = SLIST_FIRST(&free)) != NULL) {
2541 vmd = VM_DOMAIN(domain);
2542 vm_domain_free_lock(vmd);
2544 MPASS(vm_phys_domain(m) == domain);
2545 SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2546 vm_page_free_phys(vmd, m);
2547 } while ((m = SLIST_FIRST(&free)) != NULL);
2548 vm_domain_free_wakeup(vmd);
2549 vm_domain_free_unlock(vmd);
2556 CTASSERT(powerof2(NRUNS));
2558 #define RUN_INDEX(count) ((count) & (NRUNS - 1))
2560 #define MIN_RECLAIM 8
2563 * vm_page_reclaim_contig:
2565 * Reclaim allocated, contiguous physical memory satisfying the specified
2566 * conditions by relocating the virtual pages using that physical memory.
2567 * Returns true if reclamation is successful and false otherwise. Since
2568 * relocation requires the allocation of physical pages, reclamation may
2569 * fail due to a shortage of free pages. When reclamation fails, callers
2570 * are expected to perform VM_WAIT before retrying a failed allocation
2571 * operation, e.g., vm_page_alloc_contig().
2573 * The caller must always specify an allocation class through "req".
2575 * allocation classes:
2576 * VM_ALLOC_NORMAL normal process request
2577 * VM_ALLOC_SYSTEM system *really* needs a page
2578 * VM_ALLOC_INTERRUPT interrupt time request
2580 * The optional allocation flags are ignored.
2582 * "npages" must be greater than zero. Both "alignment" and "boundary"
2583 * must be a power of two.
2586 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
2587 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2589 struct vm_domain *vmd;
2590 vm_paddr_t curr_low;
2591 vm_page_t m_run, m_runs[NRUNS];
2592 u_long count, reclaimed;
2593 int error, i, options, req_class;
2595 KASSERT(npages > 0, ("npages is 0"));
2596 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2597 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2598 req_class = req & VM_ALLOC_CLASS_MASK;
2601 * The page daemon is allowed to dig deeper into the free page list.
2603 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2604 req_class = VM_ALLOC_SYSTEM;
2607 * Return if the number of free pages cannot satisfy the requested
2610 vmd = VM_DOMAIN(domain);
2611 count = vmd->vmd_free_count;
2612 if (count < npages + vmd->vmd_free_reserved || (count < npages +
2613 vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2614 (count < npages && req_class == VM_ALLOC_INTERRUPT))
2618 * Scan up to three times, relaxing the restrictions ("options") on
2619 * the reclamation of reservations and superpages each time.
2621 for (options = VPSC_NORESERV;;) {
2623 * Find the highest runs that satisfy the given constraints
2624 * and restrictions, and record them in "m_runs".
2629 m_run = vm_phys_scan_contig(domain, npages, curr_low,
2630 high, alignment, boundary, options);
2633 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2634 m_runs[RUN_INDEX(count)] = m_run;
2639 * Reclaim the highest runs in LIFO (descending) order until
2640 * the number of reclaimed pages, "reclaimed", is at least
2641 * MIN_RECLAIM. Reset "reclaimed" each time because each
2642 * reclamation is idempotent, and runs will (likely) recur
2643 * from one scan to the next as restrictions are relaxed.
2646 for (i = 0; count > 0 && i < NRUNS; i++) {
2648 m_run = m_runs[RUN_INDEX(count)];
2649 error = vm_page_reclaim_run(req_class, domain, npages,
2652 reclaimed += npages;
2653 if (reclaimed >= MIN_RECLAIM)
2659 * Either relax the restrictions on the next scan or return if
2660 * the last scan had no restrictions.
2662 if (options == VPSC_NORESERV)
2663 options = VPSC_NOSUPER;
2664 else if (options == VPSC_NOSUPER)
2666 else if (options == VPSC_ANY)
2667 return (reclaimed != 0);
2672 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
2673 u_long alignment, vm_paddr_t boundary)
2675 struct vm_domainset_iter di;
2679 vm_domainset_iter_page_init(&di, kernel_object, &domain, &req);
2681 ret = vm_page_reclaim_contig_domain(domain, req, npages, low,
2682 high, alignment, boundary);
2685 } while (vm_domainset_iter_page(&di, &domain, &req) == 0);
2691 * Set the domain in the appropriate page level domainset.
2694 vm_domain_set(struct vm_domain *vmd)
2697 mtx_lock(&vm_domainset_lock);
2698 if (!vmd->vmd_minset && vm_paging_min(vmd)) {
2699 vmd->vmd_minset = 1;
2700 DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
2702 if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
2703 vmd->vmd_severeset = 1;
2704 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
2706 mtx_unlock(&vm_domainset_lock);
2710 * Clear the domain from the appropriate page level domainset.
2713 vm_domain_clear(struct vm_domain *vmd)
2716 mtx_lock(&vm_domainset_lock);
2717 if (vmd->vmd_minset && !vm_paging_min(vmd)) {
2718 vmd->vmd_minset = 0;
2719 DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
2720 if (vm_min_waiters != 0) {
2722 wakeup(&vm_min_domains);
2725 if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
2726 vmd->vmd_severeset = 0;
2727 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
2728 if (vm_severe_waiters != 0) {
2729 vm_severe_waiters = 0;
2730 wakeup(&vm_severe_domains);
2733 mtx_unlock(&vm_domainset_lock);
2737 * Wait for free pages to exceed the min threshold globally.
2743 mtx_lock(&vm_domainset_lock);
2744 while (vm_page_count_min()) {
2746 msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
2748 mtx_unlock(&vm_domainset_lock);
2752 * Wait for free pages to exceed the severe threshold globally.
2755 vm_wait_severe(void)
2758 mtx_lock(&vm_domainset_lock);
2759 while (vm_page_count_severe()) {
2760 vm_severe_waiters++;
2761 msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
2764 mtx_unlock(&vm_domainset_lock);
2774 for (i = 0; i < vm_ndomains; i++)
2775 cnt += VM_DOMAIN(i)->vmd_waiters;
2776 cnt += vm_severe_waiters + vm_min_waiters;
2784 * Sleep until free pages are available for allocation.
2785 * - Called in various places after failed memory allocations.
2788 vm_wait_domain(int domain)
2790 struct vm_domain *vmd;
2792 vmd = VM_DOMAIN(domain);
2793 vm_domain_free_assert_locked(vmd);
2795 if (curproc == pageproc) {
2796 vmd->vmd_pageout_pages_needed = 1;
2797 msleep(&vmd->vmd_pageout_pages_needed,
2798 vm_domain_free_lockptr(vmd), PDROP | PSWP, "VMWait", 0);
2800 if (pageproc == NULL)
2801 panic("vm_wait in early boot");
2802 pagedaemon_wait(domain, PVM, "vmwait");
2807 * vm_wait: (also see VM_WAIT macro)
2809 * Sleep until free pages are available for allocation.
2810 * - Called in various places after failed memory allocations.
2817 * We use racey wakeup synchronization to avoid expensive global
2818 * locking for the pageproc when sleeping with a non-specific vm_wait.
2819 * To handle this, we only sleep for one tick in this instance. It
2820 * is expected that most allocations for the pageproc will come from
2821 * kmem or vm_page_grab* which will use the more specific and
2822 * race-free vm_wait_domain().
2824 if (curproc == pageproc) {
2825 mtx_lock(&vm_domainset_lock);
2826 vm_pageproc_waiters++;
2827 msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM,
2829 mtx_unlock(&vm_domainset_lock);
2832 * XXX Ideally we would wait only until the allocation could
2833 * be satisfied. This condition can cause new allocators to
2834 * consume all freed pages while old allocators wait.
2836 mtx_lock(&vm_domainset_lock);
2837 if (vm_page_count_min()) {
2839 msleep(&vm_min_domains, &vm_domainset_lock, PVM,
2842 mtx_unlock(&vm_domainset_lock);
2847 * vm_domain_alloc_fail:
2849 * Called when a page allocation function fails. Informs the
2850 * pagedaemon and performs the requested wait. Requires the
2851 * domain_free and object lock on entry. Returns with the
2852 * object lock held and free lock released. Returns an error when
2853 * retry is necessary.
2857 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
2860 vm_domain_free_assert_locked(vmd);
2862 atomic_add_int(&vmd->vmd_pageout_deficit,
2863 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
2864 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
2866 VM_OBJECT_WUNLOCK(object);
2867 vm_wait_domain(vmd->vmd_domain);
2869 VM_OBJECT_WLOCK(object);
2870 if (req & VM_ALLOC_WAITOK)
2873 vm_domain_free_unlock(vmd);
2874 pagedaemon_wakeup(vmd->vmd_domain);
2880 * vm_waitpfault: (also see VM_WAITPFAULT macro)
2882 * Sleep until free pages are available for allocation.
2883 * - Called only in vm_fault so that processes page faulting
2884 * can be easily tracked.
2885 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
2886 * processes will be able to grab memory first. Do not change
2887 * this balance without careful testing first.
2893 mtx_lock(&vm_domainset_lock);
2894 if (vm_page_count_min()) {
2896 msleep(&vm_min_domains, &vm_domainset_lock, PUSER, "pfault", 0);
2898 mtx_unlock(&vm_domainset_lock);
2901 struct vm_pagequeue *
2902 vm_page_pagequeue(vm_page_t m)
2905 return (&vm_pagequeue_domain(m)->vmd_pagequeues[m->queue]);
2911 * Remove the given page from its current page queue.
2913 * The page must be locked.
2916 vm_page_dequeue(vm_page_t m)
2918 struct vm_pagequeue *pq;
2920 vm_page_assert_locked(m);
2921 KASSERT(m->queue < PQ_COUNT, ("vm_page_dequeue: page %p is not queued",
2923 pq = vm_page_pagequeue(m);
2924 vm_pagequeue_lock(pq);
2926 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2927 vm_pagequeue_cnt_dec(pq);
2928 vm_pagequeue_unlock(pq);
2932 * vm_page_dequeue_locked:
2934 * Remove the given page from its current page queue.
2936 * The page and page queue must be locked.
2939 vm_page_dequeue_locked(vm_page_t m)
2941 struct vm_pagequeue *pq;
2943 vm_page_lock_assert(m, MA_OWNED);
2944 pq = vm_page_pagequeue(m);
2945 vm_pagequeue_assert_locked(pq);
2947 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2948 vm_pagequeue_cnt_dec(pq);
2954 * Add the given page to the specified page queue.
2956 * The page must be locked.
2959 vm_page_enqueue(uint8_t queue, vm_page_t m)
2961 struct vm_pagequeue *pq;
2963 vm_page_lock_assert(m, MA_OWNED);
2964 KASSERT(queue < PQ_COUNT,
2965 ("vm_page_enqueue: invalid queue %u request for page %p",
2967 pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue];
2968 vm_pagequeue_lock(pq);
2970 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2971 vm_pagequeue_cnt_inc(pq);
2972 vm_pagequeue_unlock(pq);
2978 * Move the given page to the tail of its current page queue.
2980 * The page must be locked.
2983 vm_page_requeue(vm_page_t m)
2985 struct vm_pagequeue *pq;
2987 vm_page_lock_assert(m, MA_OWNED);
2988 KASSERT(m->queue != PQ_NONE,
2989 ("vm_page_requeue: page %p is not queued", m));
2990 pq = vm_page_pagequeue(m);
2991 vm_pagequeue_lock(pq);
2992 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2993 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2994 vm_pagequeue_unlock(pq);
2998 * vm_page_requeue_locked:
3000 * Move the given page to the tail of its current page queue.
3002 * The page queue must be locked.
3005 vm_page_requeue_locked(vm_page_t m)
3007 struct vm_pagequeue *pq;
3009 KASSERT(m->queue != PQ_NONE,
3010 ("vm_page_requeue_locked: page %p is not queued", m));
3011 pq = vm_page_pagequeue(m);
3012 vm_pagequeue_assert_locked(pq);
3013 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3014 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3020 * Put the specified page on the active list (if appropriate).
3021 * Ensure that act_count is at least ACT_INIT but do not otherwise
3024 * The page must be locked.
3027 vm_page_activate(vm_page_t m)
3031 vm_page_lock_assert(m, MA_OWNED);
3032 if ((queue = m->queue) != PQ_ACTIVE) {
3033 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
3034 if (m->act_count < ACT_INIT)
3035 m->act_count = ACT_INIT;
3036 if (queue != PQ_NONE)
3038 vm_page_enqueue(PQ_ACTIVE, m);
3041 if (m->act_count < ACT_INIT)
3042 m->act_count = ACT_INIT;
3047 * vm_domain_free_wakeup:
3049 * Helper routine for vm_page_free_toq(). This routine is called
3050 * when a page is added to the free queues.
3052 * The page queues must be locked.
3055 vm_domain_free_wakeup(struct vm_domain *vmd)
3058 vm_domain_free_assert_locked(vmd);
3061 * if pageout daemon needs pages, then tell it that there are
3064 if (vmd->vmd_pageout_pages_needed &&
3065 vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
3066 wakeup(&vmd->vmd_pageout_pages_needed);
3067 vmd->vmd_pageout_pages_needed = 0;
3070 * wakeup processes that are waiting on memory if we hit a
3071 * high water mark. And wakeup scheduler process if we have
3072 * lots of memory. this process will swapin processes.
3074 if (vmd->vmd_pages_needed && !vm_paging_min(vmd)) {
3075 vmd->vmd_pages_needed = false;
3076 wakeup(&vmd->vmd_free_count);
3078 if ((vmd->vmd_minset && !vm_paging_min(vmd)) ||
3079 (vmd->vmd_severeset && !vm_paging_severe(vmd)))
3080 vm_domain_clear(vmd);
3082 /* See comments in vm_wait(); */
3083 if (vm_pageproc_waiters) {
3084 vm_pageproc_waiters = 0;
3085 wakeup(&vm_pageproc_waiters);
3091 * vm_page_free_prep:
3093 * Prepares the given page to be put on the free list,
3094 * disassociating it from any VM object. The caller may return
3095 * the page to the free list only if this function returns true.
3097 * The object must be locked. The page must be locked if it is
3098 * managed. For a queued managed page, the pagequeue_locked
3099 * argument specifies whether the page queue is already locked.
3102 vm_page_free_prep(vm_page_t m, bool pagequeue_locked)
3105 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3106 if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3109 p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3110 for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3111 KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3112 m, i, (uintmax_t)*p));
3115 if ((m->oflags & VPO_UNMANAGED) == 0) {
3116 vm_page_lock_assert(m, MA_OWNED);
3117 KASSERT(!pmap_page_is_mapped(m),
3118 ("vm_page_free_toq: freeing mapped page %p", m));
3120 KASSERT(m->queue == PQ_NONE,
3121 ("vm_page_free_toq: unmanaged page %p is queued", m));
3122 VM_CNT_INC(v_tfree);
3124 if (vm_page_sbusied(m))
3125 panic("vm_page_free: freeing busy page %p", m);
3130 * If fictitious remove object association and
3133 if ((m->flags & PG_FICTITIOUS) != 0) {
3134 KASSERT(m->wire_count == 1,
3135 ("fictitious page %p is not wired", m));
3136 KASSERT(m->queue == PQ_NONE,
3137 ("fictitious page %p is queued", m));
3141 if (m->queue != PQ_NONE) {
3142 if (pagequeue_locked)
3143 vm_page_dequeue_locked(m);
3150 if (m->wire_count != 0)
3151 panic("vm_page_free: freeing wired page %p", m);
3152 if (m->hold_count != 0) {
3153 m->flags &= ~PG_ZERO;
3154 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
3155 ("vm_page_free: freeing PG_UNHOLDFREE page %p", m));
3156 m->flags |= PG_UNHOLDFREE;
3161 * Restore the default memory attribute to the page.
3163 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3164 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3170 * Insert the page into the physical memory allocator's free page
3171 * queues. This is the last step to free a page.
3174 vm_page_free_phys(struct vm_domain *vmd, vm_page_t m)
3177 vm_domain_free_assert_locked(vmd);
3179 vm_domain_freecnt_adj(vmd, 1);
3180 #if VM_NRESERVLEVEL > 0
3181 if (!vm_reserv_free_page(m))
3183 vm_phys_free_pages(m, 0);
3187 vm_page_free_phys_pglist(struct pglist *tq)
3189 struct vm_domain *vmd;
3192 if (TAILQ_EMPTY(tq))
3195 TAILQ_FOREACH(m, tq, listq) {
3196 if (vmd != vm_pagequeue_domain(m)) {
3198 vm_domain_free_wakeup(vmd);
3199 vm_domain_free_unlock(vmd);
3201 vmd = vm_pagequeue_domain(m);
3202 vm_domain_free_lock(vmd);
3204 vm_page_free_phys(vmd, m);
3207 vm_domain_free_wakeup(vmd);
3208 vm_domain_free_unlock(vmd);
3215 * Returns the given page to the free list, disassociating it
3216 * from any VM object.
3218 * The object must be locked. The page must be locked if it is
3222 vm_page_free_toq(vm_page_t m)
3224 struct vm_domain *vmd;
3226 if (!vm_page_free_prep(m, false))
3228 vmd = vm_pagequeue_domain(m);
3229 vm_domain_free_lock(vmd);
3230 vm_page_free_phys(vmd, m);
3231 vm_domain_free_wakeup(vmd);
3232 vm_domain_free_unlock(vmd);
3238 * Mark this page as wired down. If the page is fictitious, then
3239 * its wire count must remain one.
3241 * The page must be locked.
3244 vm_page_wire(vm_page_t m)
3247 vm_page_assert_locked(m);
3248 if ((m->flags & PG_FICTITIOUS) != 0) {
3249 KASSERT(m->wire_count == 1,
3250 ("vm_page_wire: fictitious page %p's wire count isn't one",
3254 if (m->wire_count == 0) {
3255 KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
3256 m->queue == PQ_NONE,
3257 ("vm_page_wire: unmanaged page %p is queued", m));
3261 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
3267 * Release one wiring of the specified page, potentially allowing it to be
3268 * paged out. Returns TRUE if the number of wirings transitions to zero and
3271 * Only managed pages belonging to an object can be paged out. If the number
3272 * of wirings transitions to zero and the page is eligible for page out, then
3273 * the page is added to the specified paging queue (unless PQ_NONE is
3274 * specified, in which case the page is dequeued if it belongs to a paging
3277 * If a page is fictitious, then its wire count must always be one.
3279 * A managed page must be locked.
3282 vm_page_unwire(vm_page_t m, uint8_t queue)
3286 KASSERT(queue < PQ_COUNT || queue == PQ_NONE,
3287 ("vm_page_unwire: invalid queue %u request for page %p",
3290 unwired = vm_page_unwire_noq(m);
3291 if (unwired && (m->oflags & VPO_UNMANAGED) == 0 && m->object != NULL) {
3292 if (m->queue == queue) {
3293 if (queue == PQ_ACTIVE)
3294 vm_page_reference(m);
3295 else if (queue != PQ_NONE)
3299 if (queue != PQ_NONE) {
3300 vm_page_enqueue(queue, m);
3301 if (queue == PQ_ACTIVE)
3302 /* Initialize act_count. */
3303 vm_page_activate(m);
3312 * vm_page_unwire_noq:
3314 * Unwire a page without (re-)inserting it into a page queue. It is up
3315 * to the caller to enqueue, requeue, or free the page as appropriate.
3316 * In most cases, vm_page_unwire() should be used instead.
3319 vm_page_unwire_noq(vm_page_t m)
3322 if ((m->oflags & VPO_UNMANAGED) == 0)
3323 vm_page_assert_locked(m);
3324 if ((m->flags & PG_FICTITIOUS) != 0) {
3325 KASSERT(m->wire_count == 1,
3326 ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
3329 if (m->wire_count == 0)
3330 panic("vm_page_unwire: page %p's wire count is zero", m);
3332 if (m->wire_count == 0) {
3340 * Move the specified page to the inactive queue.
3342 * Normally, "noreuse" is FALSE, resulting in LRU ordering of the inactive
3343 * queue. However, setting "noreuse" to TRUE will accelerate the specified
3344 * page's reclamation, but it will not unmap the page from any address space.
3345 * This is implemented by inserting the page near the head of the inactive
3346 * queue, using a marker page to guide FIFO insertion ordering.
3348 * The page must be locked.
3351 _vm_page_deactivate(vm_page_t m, boolean_t noreuse)
3353 struct vm_pagequeue *pq;
3356 vm_page_assert_locked(m);
3359 * Ignore if the page is already inactive, unless it is unlikely to be
3362 if ((queue = m->queue) == PQ_INACTIVE && !noreuse)
3364 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
3365 pq = &vm_pagequeue_domain(m)->vmd_pagequeues[PQ_INACTIVE];
3366 /* Avoid multiple acquisitions of the inactive queue lock. */
3367 if (queue == PQ_INACTIVE) {
3368 vm_pagequeue_lock(pq);
3369 vm_page_dequeue_locked(m);
3371 if (queue != PQ_NONE)
3373 vm_pagequeue_lock(pq);
3375 m->queue = PQ_INACTIVE;
3377 TAILQ_INSERT_BEFORE(
3378 &vm_pagequeue_domain(m)->vmd_inacthead, m,
3381 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3382 vm_pagequeue_cnt_inc(pq);
3383 vm_pagequeue_unlock(pq);
3388 * Move the specified page to the inactive queue.
3390 * The page must be locked.
3393 vm_page_deactivate(vm_page_t m)
3396 _vm_page_deactivate(m, FALSE);
3400 * Move the specified page to the inactive queue with the expectation
3401 * that it is unlikely to be reused.
3403 * The page must be locked.
3406 vm_page_deactivate_noreuse(vm_page_t m)
3409 _vm_page_deactivate(m, TRUE);
3415 * Put a page in the laundry.
3418 vm_page_launder(vm_page_t m)
3422 vm_page_assert_locked(m);
3423 if ((queue = m->queue) != PQ_LAUNDRY) {
3424 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
3425 if (queue != PQ_NONE)
3427 vm_page_enqueue(PQ_LAUNDRY, m);
3429 KASSERT(queue == PQ_NONE,
3430 ("wired page %p is queued", m));
3435 * vm_page_unswappable
3437 * Put a page in the PQ_UNSWAPPABLE holding queue.
3440 vm_page_unswappable(vm_page_t m)
3443 vm_page_assert_locked(m);
3444 KASSERT(m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0,
3445 ("page %p already unswappable", m));
3446 if (m->queue != PQ_NONE)
3448 vm_page_enqueue(PQ_UNSWAPPABLE, m);
3452 * Attempt to free the page. If it cannot be freed, do nothing. Returns true
3453 * if the page is freed and false otherwise.
3455 * The page must be managed. The page and its containing object must be
3459 vm_page_try_to_free(vm_page_t m)
3462 vm_page_assert_locked(m);
3463 VM_OBJECT_ASSERT_WLOCKED(m->object);
3464 KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("page %p is unmanaged", m));
3465 if (m->dirty != 0 || vm_page_held(m) || vm_page_busied(m))
3467 if (m->object->ref_count != 0) {
3479 * Apply the specified advice to the given page.
3481 * The object and page must be locked.
3484 vm_page_advise(vm_page_t m, int advice)
3487 vm_page_assert_locked(m);
3488 VM_OBJECT_ASSERT_WLOCKED(m->object);
3489 if (advice == MADV_FREE)
3491 * Mark the page clean. This will allow the page to be freed
3492 * without first paging it out. MADV_FREE pages are often
3493 * quickly reused by malloc(3), so we do not do anything that
3494 * would result in a page fault on a later access.
3497 else if (advice != MADV_DONTNEED) {
3498 if (advice == MADV_WILLNEED)
3499 vm_page_activate(m);
3504 * Clear any references to the page. Otherwise, the page daemon will
3505 * immediately reactivate the page.
3507 vm_page_aflag_clear(m, PGA_REFERENCED);
3509 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
3513 * Place clean pages near the head of the inactive queue rather than
3514 * the tail, thus defeating the queue's LRU operation and ensuring that
3515 * the page will be reused quickly. Dirty pages not already in the
3516 * laundry are moved there.
3519 vm_page_deactivate_noreuse(m);
3525 * Grab a page, waiting until we are waken up due to the page
3526 * changing state. We keep on waiting, if the page continues
3527 * to be in the object. If the page doesn't exist, first allocate it
3528 * and then conditionally zero it.
3530 * This routine may sleep.
3532 * The object must be locked on entry. The lock will, however, be released
3533 * and reacquired if the routine sleeps.
3536 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
3542 VM_OBJECT_ASSERT_WLOCKED(object);
3543 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3544 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3545 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
3546 pflags = allocflags &
3547 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
3548 if ((allocflags & VM_ALLOC_NOWAIT) == 0)
3549 pflags |= VM_ALLOC_WAITFAIL;
3551 if ((m = vm_page_lookup(object, pindex)) != NULL) {
3552 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3553 vm_page_xbusied(m) : vm_page_busied(m);
3555 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3558 * Reference the page before unlocking and
3559 * sleeping so that the page daemon is less
3560 * likely to reclaim it.
3562 vm_page_aflag_set(m, PGA_REFERENCED);
3564 VM_OBJECT_WUNLOCK(object);
3565 vm_page_busy_sleep(m, "pgrbwt", (allocflags &
3566 VM_ALLOC_IGN_SBUSY) != 0);
3567 VM_OBJECT_WLOCK(object);
3570 if ((allocflags & VM_ALLOC_WIRED) != 0) {
3576 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
3578 if ((allocflags & VM_ALLOC_SBUSY) != 0)
3583 m = vm_page_alloc(object, pindex, pflags);
3585 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3589 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
3595 * Return the specified range of pages from the given object. For each
3596 * page offset within the range, if a page already exists within the object
3597 * at that offset and it is busy, then wait for it to change state. If,
3598 * instead, the page doesn't exist, then allocate it.
3600 * The caller must always specify an allocation class.
3602 * allocation classes:
3603 * VM_ALLOC_NORMAL normal process request
3604 * VM_ALLOC_SYSTEM system *really* needs the pages
3606 * The caller must always specify that the pages are to be busied and/or
3609 * optional allocation flags:
3610 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages
3611 * VM_ALLOC_NOBUSY do not exclusive busy the page
3612 * VM_ALLOC_NOWAIT do not sleep
3613 * VM_ALLOC_SBUSY set page to sbusy state
3614 * VM_ALLOC_WIRED wire the pages
3615 * VM_ALLOC_ZERO zero and validate any invalid pages
3617 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it
3618 * may return a partial prefix of the requested range.
3621 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
3622 vm_page_t *ma, int count)
3629 VM_OBJECT_ASSERT_WLOCKED(object);
3630 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
3631 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
3632 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
3633 (allocflags & VM_ALLOC_WIRED) != 0,
3634 ("vm_page_grab_pages: the pages must be busied or wired"));
3635 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3636 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3637 ("vm_page_grab_pages: VM_ALLOC_SBUSY/IGN_SBUSY mismatch"));
3640 pflags = allocflags & ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK |
3641 VM_ALLOC_WAITFAIL | VM_ALLOC_IGN_SBUSY);
3642 if ((allocflags & VM_ALLOC_NOWAIT) == 0)
3643 pflags |= VM_ALLOC_WAITFAIL;
3646 m = vm_radix_lookup_le(&object->rtree, pindex + i);
3647 if (m == NULL || m->pindex != pindex + i) {
3651 mpred = TAILQ_PREV(m, pglist, listq);
3652 for (; i < count; i++) {
3654 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3655 vm_page_xbusied(m) : vm_page_busied(m);
3657 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3660 * Reference the page before unlocking and
3661 * sleeping so that the page daemon is less
3662 * likely to reclaim it.
3664 vm_page_aflag_set(m, PGA_REFERENCED);
3666 VM_OBJECT_WUNLOCK(object);
3667 vm_page_busy_sleep(m, "grbmaw", (allocflags &
3668 VM_ALLOC_IGN_SBUSY) != 0);
3669 VM_OBJECT_WLOCK(object);
3672 if ((allocflags & VM_ALLOC_WIRED) != 0) {
3677 if ((allocflags & (VM_ALLOC_NOBUSY |
3678 VM_ALLOC_SBUSY)) == 0)
3680 if ((allocflags & VM_ALLOC_SBUSY) != 0)
3683 m = vm_page_alloc_after(object, pindex + i,
3684 pflags | VM_ALLOC_COUNT(count - i), mpred);
3686 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3691 if (m->valid == 0 && (allocflags & VM_ALLOC_ZERO) != 0) {
3692 if ((m->flags & PG_ZERO) == 0)
3694 m->valid = VM_PAGE_BITS_ALL;
3697 m = vm_page_next(m);
3703 * Mapping function for valid or dirty bits in a page.
3705 * Inputs are required to range within a page.
3708 vm_page_bits(int base, int size)
3714 base + size <= PAGE_SIZE,
3715 ("vm_page_bits: illegal base/size %d/%d", base, size)
3718 if (size == 0) /* handle degenerate case */
3721 first_bit = base >> DEV_BSHIFT;
3722 last_bit = (base + size - 1) >> DEV_BSHIFT;
3724 return (((vm_page_bits_t)2 << last_bit) -
3725 ((vm_page_bits_t)1 << first_bit));
3729 * vm_page_set_valid_range:
3731 * Sets portions of a page valid. The arguments are expected
3732 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3733 * of any partial chunks touched by the range. The invalid portion of
3734 * such chunks will be zeroed.
3736 * (base + size) must be less then or equal to PAGE_SIZE.
3739 vm_page_set_valid_range(vm_page_t m, int base, int size)
3743 VM_OBJECT_ASSERT_WLOCKED(m->object);
3744 if (size == 0) /* handle degenerate case */
3748 * If the base is not DEV_BSIZE aligned and the valid
3749 * bit is clear, we have to zero out a portion of the
3752 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
3753 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
3754 pmap_zero_page_area(m, frag, base - frag);
3757 * If the ending offset is not DEV_BSIZE aligned and the
3758 * valid bit is clear, we have to zero out a portion of
3761 endoff = base + size;
3762 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
3763 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
3764 pmap_zero_page_area(m, endoff,
3765 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
3768 * Assert that no previously invalid block that is now being validated
3771 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
3772 ("vm_page_set_valid_range: page %p is dirty", m));
3775 * Set valid bits inclusive of any overlap.
3777 m->valid |= vm_page_bits(base, size);
3781 * Clear the given bits from the specified page's dirty field.
3783 static __inline void
3784 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
3787 #if PAGE_SIZE < 16384
3792 * If the object is locked and the page is neither exclusive busy nor
3793 * write mapped, then the page's dirty field cannot possibly be
3794 * set by a concurrent pmap operation.
3796 VM_OBJECT_ASSERT_WLOCKED(m->object);
3797 if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
3798 m->dirty &= ~pagebits;
3801 * The pmap layer can call vm_page_dirty() without
3802 * holding a distinguished lock. The combination of
3803 * the object's lock and an atomic operation suffice
3804 * to guarantee consistency of the page dirty field.
3806 * For PAGE_SIZE == 32768 case, compiler already
3807 * properly aligns the dirty field, so no forcible
3808 * alignment is needed. Only require existence of
3809 * atomic_clear_64 when page size is 32768.
3811 addr = (uintptr_t)&m->dirty;
3812 #if PAGE_SIZE == 32768
3813 atomic_clear_64((uint64_t *)addr, pagebits);
3814 #elif PAGE_SIZE == 16384
3815 atomic_clear_32((uint32_t *)addr, pagebits);
3816 #else /* PAGE_SIZE <= 8192 */
3818 * Use a trick to perform a 32-bit atomic on the
3819 * containing aligned word, to not depend on the existence
3820 * of atomic_clear_{8, 16}.
3822 shift = addr & (sizeof(uint32_t) - 1);
3823 #if BYTE_ORDER == BIG_ENDIAN
3824 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
3828 addr &= ~(sizeof(uint32_t) - 1);
3829 atomic_clear_32((uint32_t *)addr, pagebits << shift);
3830 #endif /* PAGE_SIZE */
3835 * vm_page_set_validclean:
3837 * Sets portions of a page valid and clean. The arguments are expected
3838 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3839 * of any partial chunks touched by the range. The invalid portion of
3840 * such chunks will be zero'd.
3842 * (base + size) must be less then or equal to PAGE_SIZE.
3845 vm_page_set_validclean(vm_page_t m, int base, int size)
3847 vm_page_bits_t oldvalid, pagebits;
3850 VM_OBJECT_ASSERT_WLOCKED(m->object);
3851 if (size == 0) /* handle degenerate case */
3855 * If the base is not DEV_BSIZE aligned and the valid
3856 * bit is clear, we have to zero out a portion of the
3859 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
3860 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
3861 pmap_zero_page_area(m, frag, base - frag);
3864 * If the ending offset is not DEV_BSIZE aligned and the
3865 * valid bit is clear, we have to zero out a portion of
3868 endoff = base + size;
3869 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
3870 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
3871 pmap_zero_page_area(m, endoff,
3872 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
3875 * Set valid, clear dirty bits. If validating the entire
3876 * page we can safely clear the pmap modify bit. We also
3877 * use this opportunity to clear the VPO_NOSYNC flag. If a process
3878 * takes a write fault on a MAP_NOSYNC memory area the flag will
3881 * We set valid bits inclusive of any overlap, but we can only
3882 * clear dirty bits for DEV_BSIZE chunks that are fully within
3885 oldvalid = m->valid;
3886 pagebits = vm_page_bits(base, size);
3887 m->valid |= pagebits;
3889 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
3890 frag = DEV_BSIZE - frag;
3896 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
3898 if (base == 0 && size == PAGE_SIZE) {
3900 * The page can only be modified within the pmap if it is
3901 * mapped, and it can only be mapped if it was previously
3904 if (oldvalid == VM_PAGE_BITS_ALL)
3906 * Perform the pmap_clear_modify() first. Otherwise,
3907 * a concurrent pmap operation, such as
3908 * pmap_protect(), could clear a modification in the
3909 * pmap and set the dirty field on the page before
3910 * pmap_clear_modify() had begun and after the dirty
3911 * field was cleared here.
3913 pmap_clear_modify(m);
3915 m->oflags &= ~VPO_NOSYNC;
3916 } else if (oldvalid != VM_PAGE_BITS_ALL)
3917 m->dirty &= ~pagebits;
3919 vm_page_clear_dirty_mask(m, pagebits);
3923 vm_page_clear_dirty(vm_page_t m, int base, int size)
3926 vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
3930 * vm_page_set_invalid:
3932 * Invalidates DEV_BSIZE'd chunks within a page. Both the
3933 * valid and dirty bits for the effected areas are cleared.
3936 vm_page_set_invalid(vm_page_t m, int base, int size)
3938 vm_page_bits_t bits;
3942 VM_OBJECT_ASSERT_WLOCKED(object);
3943 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
3944 size >= object->un_pager.vnp.vnp_size)
3945 bits = VM_PAGE_BITS_ALL;
3947 bits = vm_page_bits(base, size);
3948 if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL &&
3951 KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) ||
3952 !pmap_page_is_mapped(m),
3953 ("vm_page_set_invalid: page %p is mapped", m));
3959 * vm_page_zero_invalid()
3961 * The kernel assumes that the invalid portions of a page contain
3962 * garbage, but such pages can be mapped into memory by user code.
3963 * When this occurs, we must zero out the non-valid portions of the
3964 * page so user code sees what it expects.
3966 * Pages are most often semi-valid when the end of a file is mapped
3967 * into memory and the file's size is not page aligned.
3970 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
3975 VM_OBJECT_ASSERT_WLOCKED(m->object);
3977 * Scan the valid bits looking for invalid sections that
3978 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the
3979 * valid bit may be set ) have already been zeroed by
3980 * vm_page_set_validclean().
3982 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
3983 if (i == (PAGE_SIZE / DEV_BSIZE) ||
3984 (m->valid & ((vm_page_bits_t)1 << i))) {
3986 pmap_zero_page_area(m,
3987 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
3994 * setvalid is TRUE when we can safely set the zero'd areas
3995 * as being valid. We can do this if there are no cache consistancy
3996 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
3999 m->valid = VM_PAGE_BITS_ALL;
4005 * Is (partial) page valid? Note that the case where size == 0
4006 * will return FALSE in the degenerate case where the page is
4007 * entirely invalid, and TRUE otherwise.
4010 vm_page_is_valid(vm_page_t m, int base, int size)
4012 vm_page_bits_t bits;
4014 VM_OBJECT_ASSERT_LOCKED(m->object);
4015 bits = vm_page_bits(base, size);
4016 return (m->valid != 0 && (m->valid & bits) == bits);
4020 * Returns true if all of the specified predicates are true for the entire
4021 * (super)page and false otherwise.
4024 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
4030 VM_OBJECT_ASSERT_LOCKED(object);
4031 npages = atop(pagesizes[m->psind]);
4034 * The physically contiguous pages that make up a superpage, i.e., a
4035 * page with a page size index ("psind") greater than zero, will
4036 * occupy adjacent entries in vm_page_array[].
4038 for (i = 0; i < npages; i++) {
4039 /* Always test object consistency, including "skip_m". */
4040 if (m[i].object != object)
4042 if (&m[i] == skip_m)
4044 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
4046 if ((flags & PS_ALL_DIRTY) != 0) {
4048 * Calling vm_page_test_dirty() or pmap_is_modified()
4049 * might stop this case from spuriously returning
4050 * "false". However, that would require a write lock
4051 * on the object containing "m[i]".
4053 if (m[i].dirty != VM_PAGE_BITS_ALL)
4056 if ((flags & PS_ALL_VALID) != 0 &&
4057 m[i].valid != VM_PAGE_BITS_ALL)
4064 * Set the page's dirty bits if the page is modified.
4067 vm_page_test_dirty(vm_page_t m)
4070 VM_OBJECT_ASSERT_WLOCKED(m->object);
4071 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
4076 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
4079 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
4083 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
4086 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
4090 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
4093 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
4096 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
4098 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
4101 vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
4105 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
4108 mtx_assert_(vm_page_lockptr(m), a, file, line);
4114 vm_page_object_lock_assert(vm_page_t m)
4118 * Certain of the page's fields may only be modified by the
4119 * holder of the containing object's lock or the exclusive busy.
4120 * holder. Unfortunately, the holder of the write busy is
4121 * not recorded, and thus cannot be checked here.
4123 if (m->object != NULL && !vm_page_xbusied(m))
4124 VM_OBJECT_ASSERT_WLOCKED(m->object);
4128 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits)
4131 if ((bits & PGA_WRITEABLE) == 0)
4135 * The PGA_WRITEABLE flag can only be set if the page is
4136 * managed, is exclusively busied or the object is locked.
4137 * Currently, this flag is only set by pmap_enter().
4139 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4140 ("PGA_WRITEABLE on unmanaged page"));
4141 if (!vm_page_xbusied(m))
4142 VM_OBJECT_ASSERT_LOCKED(m->object);
4146 #include "opt_ddb.h"
4148 #include <sys/kernel.h>
4150 #include <ddb/ddb.h>
4152 DB_SHOW_COMMAND(page, vm_page_print_page_info)
4155 db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
4156 db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
4157 db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
4158 db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
4159 db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
4160 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
4161 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
4162 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
4163 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
4166 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
4170 db_printf("pq_free %d\n", vm_free_count());
4171 for (dom = 0; dom < vm_ndomains; dom++) {
4173 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
4175 vm_dom[dom].vmd_page_count,
4176 vm_dom[dom].vmd_free_count,
4177 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
4178 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
4179 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
4180 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
4184 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
4190 db_printf("show pginfo addr\n");
4194 phys = strchr(modif, 'p') != NULL;
4196 m = PHYS_TO_VM_PAGE(addr);
4198 m = (vm_page_t)addr;
4200 "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n"
4201 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
4202 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
4203 m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags,
4204 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty);