2 * Copyright (c) 1991 Regents of the University of California.
4 * Copyright (c) 1998 Matthew Dillon. All Rights Reserved.
6 * This code is derived from software contributed to Berkeley by
7 * The Mach Operating System project at Carnegie-Mellon University.
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
17 * 3. Neither the name of the University nor the names of its contributors
18 * may be used to endorse or promote products derived from this software
19 * without specific prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91
37 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38 * All rights reserved.
40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
42 * Permission to use, copy, modify and distribute this software and
43 * its documentation is hereby granted, provided that both the copyright
44 * notice and this permission notice appear in all copies of the
45 * software, derivative works or modified versions, and any portions
46 * thereof, and that both notices appear in supporting documentation.
48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
52 * Carnegie Mellon requests users of this software to return to
54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
55 * School of Computer Science
56 * Carnegie Mellon University
57 * Pittsburgh PA 15213-3890
59 * any improvements or extensions that they make and grant Carnegie the
60 * rights to redistribute these changes.
64 * GENERAL RULES ON VM_PAGE MANIPULATION
66 * - A page queue lock is required when adding or removing a page from a
67 * page queue regardless of other locks or the busy state of a page.
69 * * In general, no thread besides the page daemon can acquire or
70 * hold more than one page queue lock at a time.
72 * * The page daemon can acquire and hold any pair of page queue
75 * - The object lock is required when inserting or removing
76 * pages from an object (vm_page_insert() or vm_page_remove()).
81 * Resident memory management module.
84 #include <sys/cdefs.h>
85 __FBSDID("$FreeBSD$");
89 #include <sys/param.h>
90 #include <sys/systm.h>
92 #include <sys/kernel.h>
93 #include <sys/limits.h>
94 #include <sys/linker.h>
95 #include <sys/malloc.h>
97 #include <sys/msgbuf.h>
98 #include <sys/mutex.h>
100 #include <sys/rwlock.h>
101 #include <sys/sbuf.h>
103 #include <sys/sysctl.h>
104 #include <sys/vmmeter.h>
105 #include <sys/vnode.h>
109 #include <vm/vm_param.h>
110 #include <vm/vm_kern.h>
111 #include <vm/vm_object.h>
112 #include <vm/vm_page.h>
113 #include <vm/vm_pageout.h>
114 #include <vm/vm_pager.h>
115 #include <vm/vm_phys.h>
116 #include <vm/vm_radix.h>
117 #include <vm/vm_reserv.h>
118 #include <vm/vm_extern.h>
120 #include <vm/uma_int.h>
122 #include <machine/md_var.h>
125 * Associated with page of user-allocatable memory is a
129 struct vm_domain vm_dom[MAXMEMDOM];
130 struct mtx_padalign __exclusive_cache_line vm_page_queue_free_mtx;
132 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
135 * bogus page -- for I/O to/from partially complete buffers,
136 * or for paging into sparsely invalid regions.
138 vm_page_t bogus_page;
140 vm_page_t vm_page_array;
141 long vm_page_array_size;
144 static int boot_pages = UMA_BOOT_PAGES;
145 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
147 "number of pages allocated for bootstrapping the VM system");
149 static int pa_tryrelock_restart;
150 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
151 &pa_tryrelock_restart, 0, "Number of tryrelock restarts");
153 static TAILQ_HEAD(, vm_page) blacklist_head;
154 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
155 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
156 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
158 /* Is the page daemon waiting for free pages? */
159 static int vm_pageout_pages_needed;
161 static uma_zone_t fakepg_zone;
163 static void vm_page_alloc_check(vm_page_t m);
164 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
165 static void vm_page_enqueue(uint8_t queue, vm_page_t m);
166 static void vm_page_free_wakeup(void);
167 static void vm_page_init(void *dummy);
168 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
169 vm_pindex_t pindex, vm_page_t mpred);
170 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
172 static int vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run,
175 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
178 vm_page_init(void *dummy)
181 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
182 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
183 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
184 VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
187 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
188 #if PAGE_SIZE == 32768
190 CTASSERT(sizeof(u_long) >= 8);
195 * Try to acquire a physical address lock while a pmap is locked. If we
196 * fail to trylock we unlock and lock the pmap directly and cache the
197 * locked pa in *locked. The caller should then restart their loop in case
198 * the virtual to physical mapping has changed.
201 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
208 PA_LOCK_ASSERT(lockpa, MA_OWNED);
209 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
216 atomic_add_int(&pa_tryrelock_restart, 1);
225 * Sets the page size, perhaps based upon the memory
226 * size. Must be called before any use of page-size
227 * dependent functions.
230 vm_set_page_size(void)
232 if (vm_cnt.v_page_size == 0)
233 vm_cnt.v_page_size = PAGE_SIZE;
234 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
235 panic("vm_set_page_size: page size not a power of two");
239 * vm_page_blacklist_next:
241 * Find the next entry in the provided string of blacklist
242 * addresses. Entries are separated by space, comma, or newline.
243 * If an invalid integer is encountered then the rest of the
244 * string is skipped. Updates the list pointer to the next
245 * character, or NULL if the string is exhausted or invalid.
248 vm_page_blacklist_next(char **list, char *end)
253 if (list == NULL || *list == NULL)
261 * If there's no end pointer then the buffer is coming from
262 * the kenv and we know it's null-terminated.
265 end = *list + strlen(*list);
267 /* Ensure that strtoq() won't walk off the end */
269 if (*end == '\n' || *end == ' ' || *end == ',')
272 printf("Blacklist not terminated, skipping\n");
278 for (pos = *list; *pos != '\0'; pos = cp) {
279 bad = strtoq(pos, &cp, 0);
280 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
289 if (*cp == '\0' || ++cp >= end)
293 return (trunc_page(bad));
295 printf("Garbage in RAM blacklist, skipping\n");
301 * vm_page_blacklist_check:
303 * Iterate through the provided string of blacklist addresses, pulling
304 * each entry out of the physical allocator free list and putting it
305 * onto a list for reporting via the vm.page_blacklist sysctl.
308 vm_page_blacklist_check(char *list, char *end)
316 while (next != NULL) {
317 if ((pa = vm_page_blacklist_next(&next, end)) == 0)
319 m = vm_phys_paddr_to_vm_page(pa);
322 mtx_lock(&vm_page_queue_free_mtx);
323 ret = vm_phys_unfree_page(m);
324 mtx_unlock(&vm_page_queue_free_mtx);
326 TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
328 printf("Skipping page with pa 0x%jx\n",
335 * vm_page_blacklist_load:
337 * Search for a special module named "ram_blacklist". It'll be a
338 * plain text file provided by the user via the loader directive
342 vm_page_blacklist_load(char **list, char **end)
351 mod = preload_search_by_type("ram_blacklist");
353 ptr = preload_fetch_addr(mod);
354 len = preload_fetch_size(mod);
365 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
372 error = sysctl_wire_old_buffer(req, 0);
375 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
376 TAILQ_FOREACH(m, &blacklist_head, listq) {
377 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
378 (uintmax_t)m->phys_addr);
381 error = sbuf_finish(&sbuf);
387 vm_page_domain_init(struct vm_domain *vmd)
389 struct vm_pagequeue *pq;
392 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
393 "vm inactive pagequeue";
394 *__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_vcnt) =
395 &vm_cnt.v_inactive_count;
396 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
397 "vm active pagequeue";
398 *__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_vcnt) =
399 &vm_cnt.v_active_count;
400 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
401 "vm laundry pagequeue";
402 *__DECONST(int **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_vcnt) =
403 &vm_cnt.v_laundry_count;
404 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
405 "vm unswappable pagequeue";
406 /* Unswappable dirty pages are counted as being in the laundry. */
407 *__DECONST(int **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_vcnt) =
408 &vm_cnt.v_laundry_count;
409 vmd->vmd_page_count = 0;
410 vmd->vmd_free_count = 0;
412 vmd->vmd_oom = FALSE;
413 for (i = 0; i < PQ_COUNT; i++) {
414 pq = &vmd->vmd_pagequeues[i];
415 TAILQ_INIT(&pq->pq_pl);
416 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
417 MTX_DEF | MTX_DUPOK);
424 * Initializes the resident memory module. Allocates physical memory for
425 * bootstrapping UMA and some data structures that are used to manage
426 * physical pages. Initializes these structures, and populates the free
430 vm_page_startup(vm_offset_t vaddr)
432 struct vm_domain *vmd;
433 struct vm_phys_seg *seg;
435 char *list, *listend;
437 vm_paddr_t end, high_avail, low_avail, new_end, page_range, size;
438 vm_paddr_t biggestsize, last_pa, pa;
440 int biggestone, i, pages_per_zone, segind;
444 vaddr = round_page(vaddr);
446 for (i = 0; phys_avail[i + 1]; i += 2) {
447 phys_avail[i] = round_page(phys_avail[i]);
448 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
450 for (i = 0; phys_avail[i + 1]; i += 2) {
451 size = phys_avail[i + 1] - phys_avail[i];
452 if (size > biggestsize) {
458 end = phys_avail[biggestone+1];
461 * Initialize the page and queue locks.
463 mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF);
464 for (i = 0; i < PA_LOCK_COUNT; i++)
465 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
466 for (i = 0; i < vm_ndomains; i++)
467 vm_page_domain_init(&vm_dom[i]);
470 * Almost all of the pages needed for bootstrapping UMA are used
471 * for zone structures, so if the number of CPUs results in those
472 * structures taking more than one page each, we set aside more pages
473 * in proportion to the zone structure size.
475 pages_per_zone = howmany(sizeof(struct uma_zone) +
476 sizeof(struct uma_cache) * (mp_maxid + 1) +
477 roundup2(sizeof(struct uma_slab), sizeof(void *)), UMA_SLAB_SIZE);
478 if (pages_per_zone > 1) {
479 /* Reserve more pages so that we don't run out. */
480 boot_pages = UMA_BOOT_PAGES_ZONES * pages_per_zone;
484 * Allocate memory for use when boot strapping the kernel memory
487 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
488 * manually fetch the value.
490 TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
491 new_end = end - (boot_pages * UMA_SLAB_SIZE);
492 new_end = trunc_page(new_end);
493 mapped = pmap_map(&vaddr, new_end, end,
494 VM_PROT_READ | VM_PROT_WRITE);
495 bzero((void *)mapped, end - new_end);
496 uma_startup((void *)mapped, boot_pages);
498 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
499 defined(__i386__) || defined(__mips__)
501 * Allocate a bitmap to indicate that a random physical page
502 * needs to be included in a minidump.
504 * The amd64 port needs this to indicate which direct map pages
505 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
507 * However, i386 still needs this workspace internally within the
508 * minidump code. In theory, they are not needed on i386, but are
509 * included should the sf_buf code decide to use them.
512 for (i = 0; dump_avail[i + 1] != 0; i += 2)
513 if (dump_avail[i + 1] > last_pa)
514 last_pa = dump_avail[i + 1];
515 page_range = last_pa / PAGE_SIZE;
516 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
517 new_end -= vm_page_dump_size;
518 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
519 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
520 bzero((void *)vm_page_dump, vm_page_dump_size);
524 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
526 * Include the UMA bootstrap pages and vm_page_dump in a crash dump.
527 * When pmap_map() uses the direct map, they are not automatically
530 for (pa = new_end; pa < end; pa += PAGE_SIZE)
533 phys_avail[biggestone + 1] = new_end;
536 * Request that the physical pages underlying the message buffer be
537 * included in a crash dump. Since the message buffer is accessed
538 * through the direct map, they are not automatically included.
540 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
541 last_pa = pa + round_page(msgbufsize);
542 while (pa < last_pa) {
548 * Compute the number of pages of memory that will be available for
549 * use, taking into account the overhead of a page structure per page.
550 * In other words, solve
551 * "available physical memory" - round_page(page_range *
552 * sizeof(struct vm_page)) = page_range * PAGE_SIZE
555 low_avail = phys_avail[0];
556 high_avail = phys_avail[1];
557 for (i = 0; i < vm_phys_nsegs; i++) {
558 if (vm_phys_segs[i].start < low_avail)
559 low_avail = vm_phys_segs[i].start;
560 if (vm_phys_segs[i].end > high_avail)
561 high_avail = vm_phys_segs[i].end;
563 /* Skip the first chunk. It is already accounted for. */
564 for (i = 2; phys_avail[i + 1] != 0; i += 2) {
565 if (phys_avail[i] < low_avail)
566 low_avail = phys_avail[i];
567 if (phys_avail[i + 1] > high_avail)
568 high_avail = phys_avail[i + 1];
570 first_page = low_avail / PAGE_SIZE;
571 #ifdef VM_PHYSSEG_SPARSE
573 for (i = 0; i < vm_phys_nsegs; i++)
574 size += vm_phys_segs[i].end - vm_phys_segs[i].start;
575 for (i = 0; phys_avail[i + 1] != 0; i += 2)
576 size += phys_avail[i + 1] - phys_avail[i];
577 #elif defined(VM_PHYSSEG_DENSE)
578 size = high_avail - low_avail;
580 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
583 #ifdef VM_PHYSSEG_DENSE
585 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
586 * the overhead of a page structure per page only if vm_page_array is
587 * allocated from the last physical memory chunk. Otherwise, we must
588 * allocate page structures representing the physical memory
589 * underlying vm_page_array, even though they will not be used.
591 if (new_end != high_avail)
592 page_range = size / PAGE_SIZE;
596 page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
599 * If the partial bytes remaining are large enough for
600 * a page (PAGE_SIZE) without a corresponding
601 * 'struct vm_page', then new_end will contain an
602 * extra page after subtracting the length of the VM
603 * page array. Compensate by subtracting an extra
606 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
607 if (new_end == high_avail)
608 high_avail -= PAGE_SIZE;
609 new_end -= PAGE_SIZE;
615 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
616 * However, because this page is allocated from KVM, out-of-bounds
617 * accesses using the direct map will not be trapped.
622 * Allocate physical memory for the page structures, and map it.
624 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
625 mapped = pmap_map(&vaddr, new_end, end,
626 VM_PROT_READ | VM_PROT_WRITE);
627 vm_page_array = (vm_page_t)mapped;
628 vm_page_array_size = page_range;
630 #if VM_NRESERVLEVEL > 0
632 * Allocate physical memory for the reservation management system's
633 * data structures, and map it.
635 if (high_avail == end)
636 high_avail = new_end;
637 new_end = vm_reserv_startup(&vaddr, new_end, high_avail);
639 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
641 * Include vm_page_array and vm_reserv_array in a crash dump.
643 for (pa = new_end; pa < end; pa += PAGE_SIZE)
646 phys_avail[biggestone + 1] = new_end;
649 * Add physical memory segments corresponding to the available
652 for (i = 0; phys_avail[i + 1] != 0; i += 2)
653 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
656 * Initialize the physical memory allocator.
661 * Initialize the page structures and add every available page to the
662 * physical memory allocator's free lists.
664 vm_cnt.v_page_count = 0;
665 vm_cnt.v_free_count = 0;
666 for (segind = 0; segind < vm_phys_nsegs; segind++) {
667 seg = &vm_phys_segs[segind];
668 for (pa = seg->start; pa < seg->end; pa += PAGE_SIZE)
669 vm_phys_init_page(pa);
672 * Add the segment to the free lists only if it is covered by
673 * one of the ranges in phys_avail. Because we've added the
674 * ranges to the vm_phys_segs array, we can assume that each
675 * segment is either entirely contained in one of the ranges,
676 * or doesn't overlap any of them.
678 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
679 if (seg->start < phys_avail[i] ||
680 seg->end > phys_avail[i + 1])
684 pagecount = (u_long)atop(seg->end - seg->start);
686 mtx_lock(&vm_page_queue_free_mtx);
687 vm_phys_free_contig(m, pagecount);
688 vm_phys_freecnt_adj(m, (int)pagecount);
689 mtx_unlock(&vm_page_queue_free_mtx);
690 vm_cnt.v_page_count += (u_int)pagecount;
692 vmd = &vm_dom[seg->domain];
693 vmd->vmd_page_count += (u_int)pagecount;
694 vmd->vmd_segs |= 1UL << m->segind;
700 * Remove blacklisted pages from the physical memory allocator.
702 TAILQ_INIT(&blacklist_head);
703 vm_page_blacklist_load(&list, &listend);
704 vm_page_blacklist_check(list, listend);
706 list = kern_getenv("vm.blacklist");
707 vm_page_blacklist_check(list, NULL);
710 #if VM_NRESERVLEVEL > 0
712 * Initialize the reservation management system.
720 vm_page_reference(vm_page_t m)
723 vm_page_aflag_set(m, PGA_REFERENCED);
727 * vm_page_busy_downgrade:
729 * Downgrade an exclusive busy page into a single shared busy page.
732 vm_page_busy_downgrade(vm_page_t m)
737 vm_page_assert_xbusied(m);
738 locked = mtx_owned(vm_page_lockptr(m));
742 x &= VPB_BIT_WAITERS;
743 if (x != 0 && !locked)
745 if (atomic_cmpset_rel_int(&m->busy_lock,
746 VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1)))
748 if (x != 0 && !locked)
761 * Return a positive value if the page is shared busied, 0 otherwise.
764 vm_page_sbusied(vm_page_t m)
769 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
775 * Shared unbusy a page.
778 vm_page_sunbusy(vm_page_t m)
782 vm_page_lock_assert(m, MA_NOTOWNED);
783 vm_page_assert_sbusied(m);
787 if (VPB_SHARERS(x) > 1) {
788 if (atomic_cmpset_int(&m->busy_lock, x,
793 if ((x & VPB_BIT_WAITERS) == 0) {
794 KASSERT(x == VPB_SHARERS_WORD(1),
795 ("vm_page_sunbusy: invalid lock state"));
796 if (atomic_cmpset_int(&m->busy_lock,
797 VPB_SHARERS_WORD(1), VPB_UNBUSIED))
801 KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS),
802 ("vm_page_sunbusy: invalid lock state for waiters"));
805 if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) {
816 * vm_page_busy_sleep:
818 * Sleep and release the page lock, using the page pointer as wchan.
819 * This is used to implement the hard-path of busying mechanism.
821 * The given page must be locked.
823 * If nonshared is true, sleep only if the page is xbusy.
826 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
830 vm_page_assert_locked(m);
833 if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) ||
834 ((x & VPB_BIT_WAITERS) == 0 &&
835 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) {
839 msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0);
845 * Try to shared busy a page.
846 * If the operation succeeds 1 is returned otherwise 0.
847 * The operation never sleeps.
850 vm_page_trysbusy(vm_page_t m)
856 if ((x & VPB_BIT_SHARED) == 0)
858 if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER))
864 vm_page_xunbusy_locked(vm_page_t m)
867 vm_page_assert_xbusied(m);
868 vm_page_assert_locked(m);
870 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
871 /* There is a waiter, do wakeup() instead of vm_page_flash(). */
876 vm_page_xunbusy_maybelocked(vm_page_t m)
880 vm_page_assert_xbusied(m);
883 * Fast path for unbusy. If it succeeds, we know that there
884 * are no waiters, so we do not need a wakeup.
886 if (atomic_cmpset_rel_int(&m->busy_lock, VPB_SINGLE_EXCLUSIVER,
890 lockacq = !mtx_owned(vm_page_lockptr(m));
893 vm_page_xunbusy_locked(m);
899 * vm_page_xunbusy_hard:
901 * Called after the first try the exclusive unbusy of a page failed.
902 * It is assumed that the waiters bit is on.
905 vm_page_xunbusy_hard(vm_page_t m)
908 vm_page_assert_xbusied(m);
911 vm_page_xunbusy_locked(m);
918 * Wakeup anyone waiting for the page.
919 * The ownership bits do not change.
921 * The given page must be locked.
924 vm_page_flash(vm_page_t m)
928 vm_page_lock_assert(m, MA_OWNED);
932 if ((x & VPB_BIT_WAITERS) == 0)
934 if (atomic_cmpset_int(&m->busy_lock, x,
935 x & (~VPB_BIT_WAITERS)))
942 * Avoid releasing and reacquiring the same page lock.
945 vm_page_change_lock(vm_page_t m, struct mtx **mtx)
949 mtx1 = vm_page_lockptr(m);
959 * Keep page from being freed by the page daemon
960 * much of the same effect as wiring, except much lower
961 * overhead and should be used only for *very* temporary
962 * holding ("wiring").
965 vm_page_hold(vm_page_t mem)
968 vm_page_lock_assert(mem, MA_OWNED);
973 vm_page_unhold(vm_page_t mem)
976 vm_page_lock_assert(mem, MA_OWNED);
977 KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!"));
979 if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0)
980 vm_page_free_toq(mem);
984 * vm_page_unhold_pages:
986 * Unhold each of the pages that is referenced by the given array.
989 vm_page_unhold_pages(vm_page_t *ma, int count)
994 for (; count != 0; count--) {
995 vm_page_change_lock(*ma, &mtx);
1004 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1008 #ifdef VM_PHYSSEG_SPARSE
1009 m = vm_phys_paddr_to_vm_page(pa);
1011 m = vm_phys_fictitious_to_vm_page(pa);
1013 #elif defined(VM_PHYSSEG_DENSE)
1017 if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1018 m = &vm_page_array[pi - first_page];
1021 return (vm_phys_fictitious_to_vm_page(pa));
1023 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1030 * Create a fictitious page with the specified physical address and
1031 * memory attribute. The memory attribute is the only the machine-
1032 * dependent aspect of a fictitious page that must be initialized.
1035 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1039 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1040 vm_page_initfake(m, paddr, memattr);
1045 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1048 if ((m->flags & PG_FICTITIOUS) != 0) {
1050 * The page's memattr might have changed since the
1051 * previous initialization. Update the pmap to the
1056 m->phys_addr = paddr;
1058 /* Fictitious pages don't use "segind". */
1059 m->flags = PG_FICTITIOUS;
1060 /* Fictitious pages don't use "order" or "pool". */
1061 m->oflags = VPO_UNMANAGED;
1062 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1066 pmap_page_set_memattr(m, memattr);
1072 * Release a fictitious page.
1075 vm_page_putfake(vm_page_t m)
1078 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1079 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1080 ("vm_page_putfake: bad page %p", m));
1081 uma_zfree(fakepg_zone, m);
1085 * vm_page_updatefake:
1087 * Update the given fictitious page to the specified physical address and
1091 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1094 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1095 ("vm_page_updatefake: bad page %p", m));
1096 m->phys_addr = paddr;
1097 pmap_page_set_memattr(m, memattr);
1106 vm_page_free(vm_page_t m)
1109 m->flags &= ~PG_ZERO;
1110 vm_page_free_toq(m);
1114 * vm_page_free_zero:
1116 * Free a page to the zerod-pages queue
1119 vm_page_free_zero(vm_page_t m)
1122 m->flags |= PG_ZERO;
1123 vm_page_free_toq(m);
1127 * Unbusy and handle the page queueing for a page from a getpages request that
1128 * was optionally read ahead or behind.
1131 vm_page_readahead_finish(vm_page_t m)
1134 /* We shouldn't put invalid pages on queues. */
1135 KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m));
1138 * Since the page is not the actually needed one, whether it should
1139 * be activated or deactivated is not obvious. Empirical results
1140 * have shown that deactivating the page is usually the best choice,
1141 * unless the page is wanted by another thread.
1144 if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1145 vm_page_activate(m);
1147 vm_page_deactivate(m);
1153 * vm_page_sleep_if_busy:
1155 * Sleep and release the page queues lock if the page is busied.
1156 * Returns TRUE if the thread slept.
1158 * The given page must be unlocked and object containing it must
1162 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1166 vm_page_lock_assert(m, MA_NOTOWNED);
1167 VM_OBJECT_ASSERT_WLOCKED(m->object);
1169 if (vm_page_busied(m)) {
1171 * The page-specific object must be cached because page
1172 * identity can change during the sleep, causing the
1173 * re-lock of a different object.
1174 * It is assumed that a reference to the object is already
1175 * held by the callers.
1179 VM_OBJECT_WUNLOCK(obj);
1180 vm_page_busy_sleep(m, msg, false);
1181 VM_OBJECT_WLOCK(obj);
1188 * vm_page_dirty_KBI: [ internal use only ]
1190 * Set all bits in the page's dirty field.
1192 * The object containing the specified page must be locked if the
1193 * call is made from the machine-independent layer.
1195 * See vm_page_clear_dirty_mask().
1197 * This function should only be called by vm_page_dirty().
1200 vm_page_dirty_KBI(vm_page_t m)
1203 /* Refer to this operation by its public name. */
1204 KASSERT(m->valid == VM_PAGE_BITS_ALL,
1205 ("vm_page_dirty: page is invalid!"));
1206 m->dirty = VM_PAGE_BITS_ALL;
1210 * vm_page_insert: [ internal use only ]
1212 * Inserts the given mem entry into the object and object list.
1214 * The object must be locked.
1217 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1221 VM_OBJECT_ASSERT_WLOCKED(object);
1222 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1223 return (vm_page_insert_after(m, object, pindex, mpred));
1227 * vm_page_insert_after:
1229 * Inserts the page "m" into the specified object at offset "pindex".
1231 * The page "mpred" must immediately precede the offset "pindex" within
1232 * the specified object.
1234 * The object must be locked.
1237 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1242 VM_OBJECT_ASSERT_WLOCKED(object);
1243 KASSERT(m->object == NULL,
1244 ("vm_page_insert_after: page already inserted"));
1245 if (mpred != NULL) {
1246 KASSERT(mpred->object == object,
1247 ("vm_page_insert_after: object doesn't contain mpred"));
1248 KASSERT(mpred->pindex < pindex,
1249 ("vm_page_insert_after: mpred doesn't precede pindex"));
1250 msucc = TAILQ_NEXT(mpred, listq);
1252 msucc = TAILQ_FIRST(&object->memq);
1254 KASSERT(msucc->pindex > pindex,
1255 ("vm_page_insert_after: msucc doesn't succeed pindex"));
1258 * Record the object/offset pair in this page
1264 * Now link into the object's ordered list of backed pages.
1266 if (vm_radix_insert(&object->rtree, m)) {
1271 vm_page_insert_radixdone(m, object, mpred);
1276 * vm_page_insert_radixdone:
1278 * Complete page "m" insertion into the specified object after the
1279 * radix trie hooking.
1281 * The page "mpred" must precede the offset "m->pindex" within the
1284 * The object must be locked.
1287 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1290 VM_OBJECT_ASSERT_WLOCKED(object);
1291 KASSERT(object != NULL && m->object == object,
1292 ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1293 if (mpred != NULL) {
1294 KASSERT(mpred->object == object,
1295 ("vm_page_insert_after: object doesn't contain mpred"));
1296 KASSERT(mpred->pindex < m->pindex,
1297 ("vm_page_insert_after: mpred doesn't precede pindex"));
1301 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1303 TAILQ_INSERT_HEAD(&object->memq, m, listq);
1306 * Show that the object has one more resident page.
1308 object->resident_page_count++;
1311 * Hold the vnode until the last page is released.
1313 if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1314 vhold(object->handle);
1317 * Since we are inserting a new and possibly dirty page,
1318 * update the object's OBJ_MIGHTBEDIRTY flag.
1320 if (pmap_page_is_write_mapped(m))
1321 vm_object_set_writeable_dirty(object);
1327 * Removes the specified page from its containing object, but does not
1328 * invalidate any backing storage.
1330 * The object must be locked. The page must be locked if it is managed.
1333 vm_page_remove(vm_page_t m)
1338 if ((m->oflags & VPO_UNMANAGED) == 0)
1339 vm_page_assert_locked(m);
1340 if ((object = m->object) == NULL)
1342 VM_OBJECT_ASSERT_WLOCKED(object);
1343 if (vm_page_xbusied(m))
1344 vm_page_xunbusy_maybelocked(m);
1345 mrem = vm_radix_remove(&object->rtree, m->pindex);
1346 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1349 * Now remove from the object's list of backed pages.
1351 TAILQ_REMOVE(&object->memq, m, listq);
1354 * And show that the object has one fewer resident page.
1356 object->resident_page_count--;
1359 * The vnode may now be recycled.
1361 if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1362 vdrop(object->handle);
1370 * Returns the page associated with the object/offset
1371 * pair specified; if none is found, NULL is returned.
1373 * The object must be locked.
1376 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1379 VM_OBJECT_ASSERT_LOCKED(object);
1380 return (vm_radix_lookup(&object->rtree, pindex));
1384 * vm_page_find_least:
1386 * Returns the page associated with the object with least pindex
1387 * greater than or equal to the parameter pindex, or NULL.
1389 * The object must be locked.
1392 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1396 VM_OBJECT_ASSERT_LOCKED(object);
1397 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1398 m = vm_radix_lookup_ge(&object->rtree, pindex);
1403 * Returns the given page's successor (by pindex) within the object if it is
1404 * resident; if none is found, NULL is returned.
1406 * The object must be locked.
1409 vm_page_next(vm_page_t m)
1413 VM_OBJECT_ASSERT_LOCKED(m->object);
1414 if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1415 MPASS(next->object == m->object);
1416 if (next->pindex != m->pindex + 1)
1423 * Returns the given page's predecessor (by pindex) within the object if it is
1424 * resident; if none is found, NULL is returned.
1426 * The object must be locked.
1429 vm_page_prev(vm_page_t m)
1433 VM_OBJECT_ASSERT_LOCKED(m->object);
1434 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1435 MPASS(prev->object == m->object);
1436 if (prev->pindex != m->pindex - 1)
1443 * Uses the page mnew as a replacement for an existing page at index
1444 * pindex which must be already present in the object.
1446 * The existing page must not be on a paging queue.
1449 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1453 VM_OBJECT_ASSERT_WLOCKED(object);
1454 KASSERT(mnew->object == NULL,
1455 ("vm_page_replace: page already in object"));
1458 * This function mostly follows vm_page_insert() and
1459 * vm_page_remove() without the radix, object count and vnode
1460 * dance. Double check such functions for more comments.
1463 mnew->object = object;
1464 mnew->pindex = pindex;
1465 mold = vm_radix_replace(&object->rtree, mnew);
1466 KASSERT(mold->queue == PQ_NONE,
1467 ("vm_page_replace: mold is on a paging queue"));
1469 /* Keep the resident page list in sorted order. */
1470 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1471 TAILQ_REMOVE(&object->memq, mold, listq);
1473 mold->object = NULL;
1474 vm_page_xunbusy_maybelocked(mold);
1477 * The object's resident_page_count does not change because we have
1478 * swapped one page for another, but OBJ_MIGHTBEDIRTY.
1480 if (pmap_page_is_write_mapped(mnew))
1481 vm_object_set_writeable_dirty(object);
1488 * Move the given memory entry from its
1489 * current object to the specified target object/offset.
1491 * Note: swap associated with the page must be invalidated by the move. We
1492 * have to do this for several reasons: (1) we aren't freeing the
1493 * page, (2) we are dirtying the page, (3) the VM system is probably
1494 * moving the page from object A to B, and will then later move
1495 * the backing store from A to B and we can't have a conflict.
1497 * Note: we *always* dirty the page. It is necessary both for the
1498 * fact that we moved it, and because we may be invalidating
1501 * The objects must be locked.
1504 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1509 VM_OBJECT_ASSERT_WLOCKED(new_object);
1511 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1512 KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1513 ("vm_page_rename: pindex already renamed"));
1516 * Create a custom version of vm_page_insert() which does not depend
1517 * by m_prev and can cheat on the implementation aspects of the
1521 m->pindex = new_pindex;
1522 if (vm_radix_insert(&new_object->rtree, m)) {
1528 * The operation cannot fail anymore. The removal must happen before
1529 * the listq iterator is tainted.
1535 /* Return back to the new pindex to complete vm_page_insert(). */
1536 m->pindex = new_pindex;
1537 m->object = new_object;
1539 vm_page_insert_radixdone(m, new_object, mpred);
1547 * Allocate and return a page that is associated with the specified
1548 * object and offset pair. By default, this page is exclusive busied.
1550 * The caller must always specify an allocation class.
1552 * allocation classes:
1553 * VM_ALLOC_NORMAL normal process request
1554 * VM_ALLOC_SYSTEM system *really* needs a page
1555 * VM_ALLOC_INTERRUPT interrupt time request
1557 * optional allocation flags:
1558 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1559 * intends to allocate
1560 * VM_ALLOC_NOBUSY do not exclusive busy the page
1561 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1562 * VM_ALLOC_NOOBJ page is not associated with an object and
1563 * should not be exclusive busy
1564 * VM_ALLOC_SBUSY shared busy the allocated page
1565 * VM_ALLOC_WIRED wire the allocated page
1566 * VM_ALLOC_ZERO prefer a zeroed page
1568 * This routine may not sleep.
1571 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1574 return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1575 vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1579 * Allocate a page in the specified object with the given page index. To
1580 * optimize insertion of the page into the object, the caller must also specifiy
1581 * the resident page in the object with largest index smaller than the given
1582 * page index, or NULL if no such page exists.
1585 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex, int req,
1589 int flags, req_class;
1591 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1592 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1593 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1594 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1595 ("inconsistent object(%p)/req(%x)", object, req));
1596 KASSERT(mpred == NULL || mpred->pindex < pindex,
1597 ("mpred %p doesn't precede pindex 0x%jx", mpred,
1598 (uintmax_t)pindex));
1600 VM_OBJECT_ASSERT_WLOCKED(object);
1602 req_class = req & VM_ALLOC_CLASS_MASK;
1605 * The page daemon is allowed to dig deeper into the free page list.
1607 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1608 req_class = VM_ALLOC_SYSTEM;
1611 * Allocate a page if the number of free pages exceeds the minimum
1612 * for the request class.
1614 mtx_lock(&vm_page_queue_free_mtx);
1615 if (vm_cnt.v_free_count > vm_cnt.v_free_reserved ||
1616 (req_class == VM_ALLOC_SYSTEM &&
1617 vm_cnt.v_free_count > vm_cnt.v_interrupt_free_min) ||
1618 (req_class == VM_ALLOC_INTERRUPT &&
1619 vm_cnt.v_free_count > 0)) {
1621 * Can we allocate the page from a reservation?
1623 #if VM_NRESERVLEVEL > 0
1624 if (object == NULL || (object->flags & (OBJ_COLORED |
1625 OBJ_FICTITIOUS)) != OBJ_COLORED || (m =
1626 vm_reserv_alloc_page(object, pindex, mpred)) == NULL)
1630 * If not, allocate it from the free page queues.
1632 m = vm_phys_alloc_pages(object != NULL ?
1633 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1634 #if VM_NRESERVLEVEL > 0
1635 if (m == NULL && vm_reserv_reclaim_inactive()) {
1636 m = vm_phys_alloc_pages(object != NULL ?
1637 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1644 * Not allocatable, give up.
1646 mtx_unlock(&vm_page_queue_free_mtx);
1647 atomic_add_int(&vm_pageout_deficit,
1648 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1649 pagedaemon_wakeup();
1654 * At this point we had better have found a good page.
1656 KASSERT(m != NULL, ("missing page"));
1657 vm_phys_freecnt_adj(m, -1);
1658 mtx_unlock(&vm_page_queue_free_mtx);
1659 vm_page_alloc_check(m);
1662 * Initialize the page. Only the PG_ZERO flag is inherited.
1665 if ((req & VM_ALLOC_ZERO) != 0)
1668 if ((req & VM_ALLOC_NODUMP) != 0)
1672 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1674 m->busy_lock = VPB_UNBUSIED;
1675 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1676 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1677 if ((req & VM_ALLOC_SBUSY) != 0)
1678 m->busy_lock = VPB_SHARERS_WORD(1);
1679 if (req & VM_ALLOC_WIRED) {
1681 * The page lock is not required for wiring a page until that
1682 * page is inserted into the object.
1684 atomic_add_int(&vm_cnt.v_wire_count, 1);
1689 if (object != NULL) {
1690 if (vm_page_insert_after(m, object, pindex, mpred)) {
1691 pagedaemon_wakeup();
1692 if (req & VM_ALLOC_WIRED) {
1693 atomic_subtract_int(&vm_cnt.v_wire_count, 1);
1696 KASSERT(m->object == NULL, ("page %p has object", m));
1697 m->oflags = VPO_UNMANAGED;
1698 m->busy_lock = VPB_UNBUSIED;
1699 /* Don't change PG_ZERO. */
1700 vm_page_free_toq(m);
1704 /* Ignore device objects; the pager sets "memattr" for them. */
1705 if (object->memattr != VM_MEMATTR_DEFAULT &&
1706 (object->flags & OBJ_FICTITIOUS) == 0)
1707 pmap_page_set_memattr(m, object->memattr);
1712 * Don't wakeup too often - wakeup the pageout daemon when
1713 * we would be nearly out of memory.
1715 if (vm_paging_needed())
1716 pagedaemon_wakeup();
1722 * vm_page_alloc_contig:
1724 * Allocate a contiguous set of physical pages of the given size "npages"
1725 * from the free lists. All of the physical pages must be at or above
1726 * the given physical address "low" and below the given physical address
1727 * "high". The given value "alignment" determines the alignment of the
1728 * first physical page in the set. If the given value "boundary" is
1729 * non-zero, then the set of physical pages cannot cross any physical
1730 * address boundary that is a multiple of that value. Both "alignment"
1731 * and "boundary" must be a power of two.
1733 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1734 * then the memory attribute setting for the physical pages is configured
1735 * to the object's memory attribute setting. Otherwise, the memory
1736 * attribute setting for the physical pages is configured to "memattr",
1737 * overriding the object's memory attribute setting. However, if the
1738 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1739 * memory attribute setting for the physical pages cannot be configured
1740 * to VM_MEMATTR_DEFAULT.
1742 * The specified object may not contain fictitious pages.
1744 * The caller must always specify an allocation class.
1746 * allocation classes:
1747 * VM_ALLOC_NORMAL normal process request
1748 * VM_ALLOC_SYSTEM system *really* needs a page
1749 * VM_ALLOC_INTERRUPT interrupt time request
1751 * optional allocation flags:
1752 * VM_ALLOC_NOBUSY do not exclusive busy the page
1753 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1754 * VM_ALLOC_NOOBJ page is not associated with an object and
1755 * should not be exclusive busy
1756 * VM_ALLOC_SBUSY shared busy the allocated page
1757 * VM_ALLOC_WIRED wire the allocated page
1758 * VM_ALLOC_ZERO prefer a zeroed page
1760 * This routine may not sleep.
1763 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1764 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1765 vm_paddr_t boundary, vm_memattr_t memattr)
1767 vm_page_t m, m_ret, mpred;
1768 u_int busy_lock, flags, oflags;
1771 mpred = NULL; /* XXX: pacify gcc */
1772 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1773 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1774 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1775 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1776 ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
1778 if (object != NULL) {
1779 VM_OBJECT_ASSERT_WLOCKED(object);
1780 KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
1781 ("vm_page_alloc_contig: object %p has fictitious pages",
1784 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
1785 req_class = req & VM_ALLOC_CLASS_MASK;
1788 * The page daemon is allowed to dig deeper into the free page list.
1790 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1791 req_class = VM_ALLOC_SYSTEM;
1793 if (object != NULL) {
1794 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1795 KASSERT(mpred == NULL || mpred->pindex != pindex,
1796 ("vm_page_alloc_contig: pindex already allocated"));
1800 * Can we allocate the pages without the number of free pages falling
1801 * below the lower bound for the allocation class?
1803 mtx_lock(&vm_page_queue_free_mtx);
1804 if (vm_cnt.v_free_count >= npages + vm_cnt.v_free_reserved ||
1805 (req_class == VM_ALLOC_SYSTEM &&
1806 vm_cnt.v_free_count >= npages + vm_cnt.v_interrupt_free_min) ||
1807 (req_class == VM_ALLOC_INTERRUPT &&
1808 vm_cnt.v_free_count >= npages)) {
1810 * Can we allocate the pages from a reservation?
1812 #if VM_NRESERVLEVEL > 0
1814 if (object == NULL || (object->flags & OBJ_COLORED) == 0 ||
1815 (m_ret = vm_reserv_alloc_contig(object, pindex, npages,
1816 low, high, alignment, boundary, mpred)) == NULL)
1819 * If not, allocate them from the free page queues.
1821 m_ret = vm_phys_alloc_contig(npages, low, high,
1822 alignment, boundary);
1824 mtx_unlock(&vm_page_queue_free_mtx);
1825 atomic_add_int(&vm_pageout_deficit, npages);
1826 pagedaemon_wakeup();
1830 vm_phys_freecnt_adj(m_ret, -npages);
1832 #if VM_NRESERVLEVEL > 0
1833 if (vm_reserv_reclaim_contig(npages, low, high, alignment,
1838 mtx_unlock(&vm_page_queue_free_mtx);
1841 for (m = m_ret; m < &m_ret[npages]; m++)
1842 vm_page_alloc_check(m);
1845 * Initialize the pages. Only the PG_ZERO flag is inherited.
1848 if ((req & VM_ALLOC_ZERO) != 0)
1850 if ((req & VM_ALLOC_NODUMP) != 0)
1852 oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1854 busy_lock = VPB_UNBUSIED;
1855 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1856 busy_lock = VPB_SINGLE_EXCLUSIVER;
1857 if ((req & VM_ALLOC_SBUSY) != 0)
1858 busy_lock = VPB_SHARERS_WORD(1);
1859 if ((req & VM_ALLOC_WIRED) != 0)
1860 atomic_add_int(&vm_cnt.v_wire_count, npages);
1861 if (object != NULL) {
1862 if (object->memattr != VM_MEMATTR_DEFAULT &&
1863 memattr == VM_MEMATTR_DEFAULT)
1864 memattr = object->memattr;
1866 for (m = m_ret; m < &m_ret[npages]; m++) {
1868 m->flags = (m->flags | PG_NODUMP) & flags;
1869 m->busy_lock = busy_lock;
1870 if ((req & VM_ALLOC_WIRED) != 0)
1874 if (object != NULL) {
1875 if (vm_page_insert_after(m, object, pindex, mpred)) {
1876 pagedaemon_wakeup();
1877 if ((req & VM_ALLOC_WIRED) != 0)
1878 atomic_subtract_int(
1879 &vm_cnt.v_wire_count, npages);
1880 KASSERT(m->object == NULL,
1881 ("page %p has object", m));
1883 for (m = m_ret; m < &m_ret[npages]; m++) {
1885 (req & VM_ALLOC_WIRED) != 0)
1887 m->oflags = VPO_UNMANAGED;
1888 m->busy_lock = VPB_UNBUSIED;
1889 /* Don't change PG_ZERO. */
1890 vm_page_free_toq(m);
1897 if (memattr != VM_MEMATTR_DEFAULT)
1898 pmap_page_set_memattr(m, memattr);
1901 if (vm_paging_needed())
1902 pagedaemon_wakeup();
1907 * Check a page that has been freshly dequeued from a freelist.
1910 vm_page_alloc_check(vm_page_t m)
1913 KASSERT(m->object == NULL, ("page %p has object", m));
1914 KASSERT(m->queue == PQ_NONE,
1915 ("page %p has unexpected queue %d", m, m->queue));
1916 KASSERT(m->wire_count == 0, ("page %p is wired", m));
1917 KASSERT(m->hold_count == 0, ("page %p is held", m));
1918 KASSERT(!vm_page_busied(m), ("page %p is busy", m));
1919 KASSERT(m->dirty == 0, ("page %p is dirty", m));
1920 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1921 ("page %p has unexpected memattr %d",
1922 m, pmap_page_get_memattr(m)));
1923 KASSERT(m->valid == 0, ("free page %p is valid", m));
1927 * vm_page_alloc_freelist:
1929 * Allocate a physical page from the specified free page list.
1931 * The caller must always specify an allocation class.
1933 * allocation classes:
1934 * VM_ALLOC_NORMAL normal process request
1935 * VM_ALLOC_SYSTEM system *really* needs a page
1936 * VM_ALLOC_INTERRUPT interrupt time request
1938 * optional allocation flags:
1939 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1940 * intends to allocate
1941 * VM_ALLOC_WIRED wire the allocated page
1942 * VM_ALLOC_ZERO prefer a zeroed page
1944 * This routine may not sleep.
1947 vm_page_alloc_freelist(int flind, int req)
1953 req_class = req & VM_ALLOC_CLASS_MASK;
1956 * The page daemon is allowed to dig deeper into the free page list.
1958 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1959 req_class = VM_ALLOC_SYSTEM;
1962 * Do not allocate reserved pages unless the req has asked for it.
1964 mtx_lock(&vm_page_queue_free_mtx);
1965 if (vm_cnt.v_free_count > vm_cnt.v_free_reserved ||
1966 (req_class == VM_ALLOC_SYSTEM &&
1967 vm_cnt.v_free_count > vm_cnt.v_interrupt_free_min) ||
1968 (req_class == VM_ALLOC_INTERRUPT &&
1969 vm_cnt.v_free_count > 0))
1970 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0);
1972 mtx_unlock(&vm_page_queue_free_mtx);
1973 atomic_add_int(&vm_pageout_deficit,
1974 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1975 pagedaemon_wakeup();
1979 mtx_unlock(&vm_page_queue_free_mtx);
1982 vm_phys_freecnt_adj(m, -1);
1983 mtx_unlock(&vm_page_queue_free_mtx);
1984 vm_page_alloc_check(m);
1987 * Initialize the page. Only the PG_ZERO flag is inherited.
1991 if ((req & VM_ALLOC_ZERO) != 0)
1994 if ((req & VM_ALLOC_WIRED) != 0) {
1996 * The page lock is not required for wiring a page that does
1997 * not belong to an object.
1999 atomic_add_int(&vm_cnt.v_wire_count, 1);
2002 /* Unmanaged pages don't use "act_count". */
2003 m->oflags = VPO_UNMANAGED;
2004 if (vm_paging_needed())
2005 pagedaemon_wakeup();
2009 #define VPSC_ANY 0 /* No restrictions. */
2010 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */
2011 #define VPSC_NOSUPER 2 /* Skip superpages. */
2014 * vm_page_scan_contig:
2016 * Scan vm_page_array[] between the specified entries "m_start" and
2017 * "m_end" for a run of contiguous physical pages that satisfy the
2018 * specified conditions, and return the lowest page in the run. The
2019 * specified "alignment" determines the alignment of the lowest physical
2020 * page in the run. If the specified "boundary" is non-zero, then the
2021 * run of physical pages cannot span a physical address that is a
2022 * multiple of "boundary".
2024 * "m_end" is never dereferenced, so it need not point to a vm_page
2025 * structure within vm_page_array[].
2027 * "npages" must be greater than zero. "m_start" and "m_end" must not
2028 * span a hole (or discontiguity) in the physical address space. Both
2029 * "alignment" and "boundary" must be a power of two.
2032 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2033 u_long alignment, vm_paddr_t boundary, int options)
2039 #if VM_NRESERVLEVEL > 0
2042 int m_inc, order, run_ext, run_len;
2044 KASSERT(npages > 0, ("npages is 0"));
2045 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2046 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2050 for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2051 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2052 ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2055 * If the current page would be the start of a run, check its
2056 * physical address against the end, alignment, and boundary
2057 * conditions. If it doesn't satisfy these conditions, either
2058 * terminate the scan or advance to the next page that
2059 * satisfies the failed condition.
2062 KASSERT(m_run == NULL, ("m_run != NULL"));
2063 if (m + npages > m_end)
2065 pa = VM_PAGE_TO_PHYS(m);
2066 if ((pa & (alignment - 1)) != 0) {
2067 m_inc = atop(roundup2(pa, alignment) - pa);
2070 if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2072 m_inc = atop(roundup2(pa, boundary) - pa);
2076 KASSERT(m_run != NULL, ("m_run == NULL"));
2078 vm_page_change_lock(m, &m_mtx);
2081 if (m->wire_count != 0 || m->hold_count != 0)
2083 #if VM_NRESERVLEVEL > 0
2084 else if ((level = vm_reserv_level(m)) >= 0 &&
2085 (options & VPSC_NORESERV) != 0) {
2087 /* Advance to the end of the reservation. */
2088 pa = VM_PAGE_TO_PHYS(m);
2089 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2093 else if ((object = m->object) != NULL) {
2095 * The page is considered eligible for relocation if
2096 * and only if it could be laundered or reclaimed by
2099 if (!VM_OBJECT_TRYRLOCK(object)) {
2101 VM_OBJECT_RLOCK(object);
2103 if (m->object != object) {
2105 * The page may have been freed.
2107 VM_OBJECT_RUNLOCK(object);
2109 } else if (m->wire_count != 0 ||
2110 m->hold_count != 0) {
2115 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2116 ("page %p is PG_UNHOLDFREE", m));
2117 /* Don't care: PG_NODUMP, PG_ZERO. */
2118 if (object->type != OBJT_DEFAULT &&
2119 object->type != OBJT_SWAP &&
2120 object->type != OBJT_VNODE) {
2122 #if VM_NRESERVLEVEL > 0
2123 } else if ((options & VPSC_NOSUPER) != 0 &&
2124 (level = vm_reserv_level_iffullpop(m)) >= 0) {
2126 /* Advance to the end of the superpage. */
2127 pa = VM_PAGE_TO_PHYS(m);
2128 m_inc = atop(roundup2(pa + 1,
2129 vm_reserv_size(level)) - pa);
2131 } else if (object->memattr == VM_MEMATTR_DEFAULT &&
2132 m->queue != PQ_NONE && !vm_page_busied(m)) {
2134 * The page is allocated but eligible for
2135 * relocation. Extend the current run by one
2138 KASSERT(pmap_page_get_memattr(m) ==
2140 ("page %p has an unexpected memattr", m));
2141 KASSERT((m->oflags & (VPO_SWAPINPROG |
2142 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2143 ("page %p has unexpected oflags", m));
2144 /* Don't care: VPO_NOSYNC. */
2149 VM_OBJECT_RUNLOCK(object);
2150 #if VM_NRESERVLEVEL > 0
2151 } else if (level >= 0) {
2153 * The page is reserved but not yet allocated. In
2154 * other words, it is still free. Extend the current
2159 } else if ((order = m->order) < VM_NFREEORDER) {
2161 * The page is enqueued in the physical memory
2162 * allocator's free page queues. Moreover, it is the
2163 * first page in a power-of-two-sized run of
2164 * contiguous free pages. Add these pages to the end
2165 * of the current run, and jump ahead.
2167 run_ext = 1 << order;
2171 * Skip the page for one of the following reasons: (1)
2172 * It is enqueued in the physical memory allocator's
2173 * free page queues. However, it is not the first
2174 * page in a run of contiguous free pages. (This case
2175 * rarely occurs because the scan is performed in
2176 * ascending order.) (2) It is not reserved, and it is
2177 * transitioning from free to allocated. (Conversely,
2178 * the transition from allocated to free for managed
2179 * pages is blocked by the page lock.) (3) It is
2180 * allocated but not contained by an object and not
2181 * wired, e.g., allocated by Xen's balloon driver.
2187 * Extend or reset the current run of pages.
2202 if (run_len >= npages)
2208 * vm_page_reclaim_run:
2210 * Try to relocate each of the allocated virtual pages within the
2211 * specified run of physical pages to a new physical address. Free the
2212 * physical pages underlying the relocated virtual pages. A virtual page
2213 * is relocatable if and only if it could be laundered or reclaimed by
2214 * the page daemon. Whenever possible, a virtual page is relocated to a
2215 * physical address above "high".
2217 * Returns 0 if every physical page within the run was already free or
2218 * just freed by a successful relocation. Otherwise, returns a non-zero
2219 * value indicating why the last attempt to relocate a virtual page was
2222 * "req_class" must be an allocation class.
2225 vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run,
2229 struct spglist free;
2232 vm_page_t m, m_end, m_new;
2233 int error, order, req;
2235 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2236 ("req_class is not an allocation class"));
2240 m_end = m_run + npages;
2242 for (; error == 0 && m < m_end; m++) {
2243 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2244 ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2247 * Avoid releasing and reacquiring the same page lock.
2249 vm_page_change_lock(m, &m_mtx);
2251 if (m->wire_count != 0 || m->hold_count != 0)
2253 else if ((object = m->object) != NULL) {
2255 * The page is relocated if and only if it could be
2256 * laundered or reclaimed by the page daemon.
2258 if (!VM_OBJECT_TRYWLOCK(object)) {
2260 VM_OBJECT_WLOCK(object);
2262 if (m->object != object) {
2264 * The page may have been freed.
2266 VM_OBJECT_WUNLOCK(object);
2268 } else if (m->wire_count != 0 ||
2269 m->hold_count != 0) {
2274 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2275 ("page %p is PG_UNHOLDFREE", m));
2276 /* Don't care: PG_NODUMP, PG_ZERO. */
2277 if (object->type != OBJT_DEFAULT &&
2278 object->type != OBJT_SWAP &&
2279 object->type != OBJT_VNODE)
2281 else if (object->memattr != VM_MEMATTR_DEFAULT)
2283 else if (m->queue != PQ_NONE && !vm_page_busied(m)) {
2284 KASSERT(pmap_page_get_memattr(m) ==
2286 ("page %p has an unexpected memattr", m));
2287 KASSERT((m->oflags & (VPO_SWAPINPROG |
2288 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2289 ("page %p has unexpected oflags", m));
2290 /* Don't care: VPO_NOSYNC. */
2291 if (m->valid != 0) {
2293 * First, try to allocate a new page
2294 * that is above "high". Failing
2295 * that, try to allocate a new page
2296 * that is below "m_run". Allocate
2297 * the new page between the end of
2298 * "m_run" and "high" only as a last
2301 req = req_class | VM_ALLOC_NOOBJ;
2302 if ((m->flags & PG_NODUMP) != 0)
2303 req |= VM_ALLOC_NODUMP;
2304 if (trunc_page(high) !=
2305 ~(vm_paddr_t)PAGE_MASK) {
2306 m_new = vm_page_alloc_contig(
2311 VM_MEMATTR_DEFAULT);
2314 if (m_new == NULL) {
2315 pa = VM_PAGE_TO_PHYS(m_run);
2316 m_new = vm_page_alloc_contig(
2318 0, pa - 1, PAGE_SIZE, 0,
2319 VM_MEMATTR_DEFAULT);
2321 if (m_new == NULL) {
2323 m_new = vm_page_alloc_contig(
2325 pa, high, PAGE_SIZE, 0,
2326 VM_MEMATTR_DEFAULT);
2328 if (m_new == NULL) {
2332 KASSERT(m_new->wire_count == 0,
2333 ("page %p is wired", m));
2336 * Replace "m" with the new page. For
2337 * vm_page_replace(), "m" must be busy
2338 * and dequeued. Finally, change "m"
2339 * as if vm_page_free() was called.
2341 if (object->ref_count != 0)
2343 m_new->aflags = m->aflags;
2344 KASSERT(m_new->oflags == VPO_UNMANAGED,
2345 ("page %p is managed", m));
2346 m_new->oflags = m->oflags & VPO_NOSYNC;
2347 pmap_copy_page(m, m_new);
2348 m_new->valid = m->valid;
2349 m_new->dirty = m->dirty;
2350 m->flags &= ~PG_ZERO;
2353 vm_page_replace_checked(m_new, object,
2359 * The new page must be deactivated
2360 * before the object is unlocked.
2362 vm_page_change_lock(m_new, &m_mtx);
2363 vm_page_deactivate(m_new);
2365 m->flags &= ~PG_ZERO;
2368 KASSERT(m->dirty == 0,
2369 ("page %p is dirty", m));
2371 SLIST_INSERT_HEAD(&free, m, plinks.s.ss);
2375 VM_OBJECT_WUNLOCK(object);
2377 mtx_lock(&vm_page_queue_free_mtx);
2379 if (order < VM_NFREEORDER) {
2381 * The page is enqueued in the physical memory
2382 * allocator's free page queues. Moreover, it
2383 * is the first page in a power-of-two-sized
2384 * run of contiguous free pages. Jump ahead
2385 * to the last page within that run, and
2386 * continue from there.
2388 m += (1 << order) - 1;
2390 #if VM_NRESERVLEVEL > 0
2391 else if (vm_reserv_is_page_free(m))
2394 mtx_unlock(&vm_page_queue_free_mtx);
2395 if (order == VM_NFREEORDER)
2401 if ((m = SLIST_FIRST(&free)) != NULL) {
2402 mtx_lock(&vm_page_queue_free_mtx);
2404 SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2405 vm_phys_freecnt_adj(m, 1);
2406 #if VM_NRESERVLEVEL > 0
2407 if (!vm_reserv_free_page(m))
2411 vm_phys_free_pages(m, 0);
2412 } while ((m = SLIST_FIRST(&free)) != NULL);
2413 vm_page_free_wakeup();
2414 mtx_unlock(&vm_page_queue_free_mtx);
2421 CTASSERT(powerof2(NRUNS));
2423 #define RUN_INDEX(count) ((count) & (NRUNS - 1))
2425 #define MIN_RECLAIM 8
2428 * vm_page_reclaim_contig:
2430 * Reclaim allocated, contiguous physical memory satisfying the specified
2431 * conditions by relocating the virtual pages using that physical memory.
2432 * Returns true if reclamation is successful and false otherwise. Since
2433 * relocation requires the allocation of physical pages, reclamation may
2434 * fail due to a shortage of free pages. When reclamation fails, callers
2435 * are expected to perform VM_WAIT before retrying a failed allocation
2436 * operation, e.g., vm_page_alloc_contig().
2438 * The caller must always specify an allocation class through "req".
2440 * allocation classes:
2441 * VM_ALLOC_NORMAL normal process request
2442 * VM_ALLOC_SYSTEM system *really* needs a page
2443 * VM_ALLOC_INTERRUPT interrupt time request
2445 * The optional allocation flags are ignored.
2447 * "npages" must be greater than zero. Both "alignment" and "boundary"
2448 * must be a power of two.
2451 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
2452 u_long alignment, vm_paddr_t boundary)
2454 vm_paddr_t curr_low;
2455 vm_page_t m_run, m_runs[NRUNS];
2456 u_long count, reclaimed;
2457 int error, i, options, req_class;
2459 KASSERT(npages > 0, ("npages is 0"));
2460 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2461 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2462 req_class = req & VM_ALLOC_CLASS_MASK;
2465 * The page daemon is allowed to dig deeper into the free page list.
2467 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2468 req_class = VM_ALLOC_SYSTEM;
2471 * Return if the number of free pages cannot satisfy the requested
2474 count = vm_cnt.v_free_count;
2475 if (count < npages + vm_cnt.v_free_reserved || (count < npages +
2476 vm_cnt.v_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2477 (count < npages && req_class == VM_ALLOC_INTERRUPT))
2481 * Scan up to three times, relaxing the restrictions ("options") on
2482 * the reclamation of reservations and superpages each time.
2484 for (options = VPSC_NORESERV;;) {
2486 * Find the highest runs that satisfy the given constraints
2487 * and restrictions, and record them in "m_runs".
2492 m_run = vm_phys_scan_contig(npages, curr_low, high,
2493 alignment, boundary, options);
2496 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2497 m_runs[RUN_INDEX(count)] = m_run;
2502 * Reclaim the highest runs in LIFO (descending) order until
2503 * the number of reclaimed pages, "reclaimed", is at least
2504 * MIN_RECLAIM. Reset "reclaimed" each time because each
2505 * reclamation is idempotent, and runs will (likely) recur
2506 * from one scan to the next as restrictions are relaxed.
2509 for (i = 0; count > 0 && i < NRUNS; i++) {
2511 m_run = m_runs[RUN_INDEX(count)];
2512 error = vm_page_reclaim_run(req_class, npages, m_run,
2515 reclaimed += npages;
2516 if (reclaimed >= MIN_RECLAIM)
2522 * Either relax the restrictions on the next scan or return if
2523 * the last scan had no restrictions.
2525 if (options == VPSC_NORESERV)
2526 options = VPSC_NOSUPER;
2527 else if (options == VPSC_NOSUPER)
2529 else if (options == VPSC_ANY)
2530 return (reclaimed != 0);
2535 * vm_wait: (also see VM_WAIT macro)
2537 * Sleep until free pages are available for allocation.
2538 * - Called in various places before memory allocations.
2544 mtx_lock(&vm_page_queue_free_mtx);
2545 if (curproc == pageproc) {
2546 vm_pageout_pages_needed = 1;
2547 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx,
2548 PDROP | PSWP, "VMWait", 0);
2550 if (__predict_false(pageproc == NULL))
2551 panic("vm_wait in early boot");
2552 if (!vm_pageout_wanted) {
2553 vm_pageout_wanted = true;
2554 wakeup(&vm_pageout_wanted);
2556 vm_pages_needed = true;
2557 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM,
2563 * vm_waitpfault: (also see VM_WAITPFAULT macro)
2565 * Sleep until free pages are available for allocation.
2566 * - Called only in vm_fault so that processes page faulting
2567 * can be easily tracked.
2568 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
2569 * processes will be able to grab memory first. Do not change
2570 * this balance without careful testing first.
2576 mtx_lock(&vm_page_queue_free_mtx);
2577 if (!vm_pageout_wanted) {
2578 vm_pageout_wanted = true;
2579 wakeup(&vm_pageout_wanted);
2581 vm_pages_needed = true;
2582 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER,
2586 struct vm_pagequeue *
2587 vm_page_pagequeue(vm_page_t m)
2590 if (vm_page_in_laundry(m))
2591 return (&vm_dom[0].vmd_pagequeues[m->queue]);
2593 return (&vm_phys_domain(m)->vmd_pagequeues[m->queue]);
2599 * Remove the given page from its current page queue.
2601 * The page must be locked.
2604 vm_page_dequeue(vm_page_t m)
2606 struct vm_pagequeue *pq;
2608 vm_page_assert_locked(m);
2609 KASSERT(m->queue < PQ_COUNT, ("vm_page_dequeue: page %p is not queued",
2611 pq = vm_page_pagequeue(m);
2612 vm_pagequeue_lock(pq);
2614 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2615 vm_pagequeue_cnt_dec(pq);
2616 vm_pagequeue_unlock(pq);
2620 * vm_page_dequeue_locked:
2622 * Remove the given page from its current page queue.
2624 * The page and page queue must be locked.
2627 vm_page_dequeue_locked(vm_page_t m)
2629 struct vm_pagequeue *pq;
2631 vm_page_lock_assert(m, MA_OWNED);
2632 pq = vm_page_pagequeue(m);
2633 vm_pagequeue_assert_locked(pq);
2635 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2636 vm_pagequeue_cnt_dec(pq);
2642 * Add the given page to the specified page queue.
2644 * The page must be locked.
2647 vm_page_enqueue(uint8_t queue, vm_page_t m)
2649 struct vm_pagequeue *pq;
2651 vm_page_lock_assert(m, MA_OWNED);
2652 KASSERT(queue < PQ_COUNT,
2653 ("vm_page_enqueue: invalid queue %u request for page %p",
2655 if (queue == PQ_LAUNDRY || queue == PQ_UNSWAPPABLE)
2656 pq = &vm_dom[0].vmd_pagequeues[queue];
2658 pq = &vm_phys_domain(m)->vmd_pagequeues[queue];
2659 vm_pagequeue_lock(pq);
2661 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2662 vm_pagequeue_cnt_inc(pq);
2663 vm_pagequeue_unlock(pq);
2669 * Move the given page to the tail of its current page queue.
2671 * The page must be locked.
2674 vm_page_requeue(vm_page_t m)
2676 struct vm_pagequeue *pq;
2678 vm_page_lock_assert(m, MA_OWNED);
2679 KASSERT(m->queue != PQ_NONE,
2680 ("vm_page_requeue: page %p is not queued", m));
2681 pq = vm_page_pagequeue(m);
2682 vm_pagequeue_lock(pq);
2683 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2684 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2685 vm_pagequeue_unlock(pq);
2689 * vm_page_requeue_locked:
2691 * Move the given page to the tail of its current page queue.
2693 * The page queue must be locked.
2696 vm_page_requeue_locked(vm_page_t m)
2698 struct vm_pagequeue *pq;
2700 KASSERT(m->queue != PQ_NONE,
2701 ("vm_page_requeue_locked: page %p is not queued", m));
2702 pq = vm_page_pagequeue(m);
2703 vm_pagequeue_assert_locked(pq);
2704 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2705 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2711 * Put the specified page on the active list (if appropriate).
2712 * Ensure that act_count is at least ACT_INIT but do not otherwise
2715 * The page must be locked.
2718 vm_page_activate(vm_page_t m)
2722 vm_page_lock_assert(m, MA_OWNED);
2723 if ((queue = m->queue) != PQ_ACTIVE) {
2724 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2725 if (m->act_count < ACT_INIT)
2726 m->act_count = ACT_INIT;
2727 if (queue != PQ_NONE)
2729 vm_page_enqueue(PQ_ACTIVE, m);
2731 KASSERT(queue == PQ_NONE,
2732 ("vm_page_activate: wired page %p is queued", m));
2734 if (m->act_count < ACT_INIT)
2735 m->act_count = ACT_INIT;
2740 * vm_page_free_wakeup:
2742 * Helper routine for vm_page_free_toq(). This routine is called
2743 * when a page is added to the free queues.
2745 * The page queues must be locked.
2748 vm_page_free_wakeup(void)
2751 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2753 * if pageout daemon needs pages, then tell it that there are
2756 if (vm_pageout_pages_needed &&
2757 vm_cnt.v_free_count >= vm_cnt.v_pageout_free_min) {
2758 wakeup(&vm_pageout_pages_needed);
2759 vm_pageout_pages_needed = 0;
2762 * wakeup processes that are waiting on memory if we hit a
2763 * high water mark. And wakeup scheduler process if we have
2764 * lots of memory. this process will swapin processes.
2766 if (vm_pages_needed && !vm_page_count_min()) {
2767 vm_pages_needed = false;
2768 wakeup(&vm_cnt.v_free_count);
2775 * Returns the given page to the free list,
2776 * disassociating it with any VM object.
2778 * The object must be locked. The page must be locked if it is managed.
2781 vm_page_free_toq(vm_page_t m)
2784 if ((m->oflags & VPO_UNMANAGED) == 0) {
2785 vm_page_lock_assert(m, MA_OWNED);
2786 KASSERT(!pmap_page_is_mapped(m),
2787 ("vm_page_free_toq: freeing mapped page %p", m));
2789 KASSERT(m->queue == PQ_NONE,
2790 ("vm_page_free_toq: unmanaged page %p is queued", m));
2791 VM_CNT_INC(v_tfree);
2793 if (vm_page_sbusied(m))
2794 panic("vm_page_free: freeing busy page %p", m);
2797 * Unqueue, then remove page. Note that we cannot destroy
2798 * the page here because we do not want to call the pager's
2799 * callback routine until after we've put the page on the
2800 * appropriate free queue.
2806 * If fictitious remove object association and
2807 * return, otherwise delay object association removal.
2809 if ((m->flags & PG_FICTITIOUS) != 0) {
2816 if (m->wire_count != 0)
2817 panic("vm_page_free: freeing wired page %p", m);
2818 if (m->hold_count != 0) {
2819 m->flags &= ~PG_ZERO;
2820 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2821 ("vm_page_free: freeing PG_UNHOLDFREE page %p", m));
2822 m->flags |= PG_UNHOLDFREE;
2825 * Restore the default memory attribute to the page.
2827 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2828 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2831 * Insert the page into the physical memory allocator's free
2834 mtx_lock(&vm_page_queue_free_mtx);
2835 vm_phys_freecnt_adj(m, 1);
2836 #if VM_NRESERVLEVEL > 0
2837 if (!vm_reserv_free_page(m))
2841 vm_phys_free_pages(m, 0);
2842 vm_page_free_wakeup();
2843 mtx_unlock(&vm_page_queue_free_mtx);
2850 * Mark this page as wired down by yet
2851 * another map, removing it from paging queues
2854 * If the page is fictitious, then its wire count must remain one.
2856 * The page must be locked.
2859 vm_page_wire(vm_page_t m)
2863 * Only bump the wire statistics if the page is not already wired,
2864 * and only unqueue the page if it is on some queue (if it is unmanaged
2865 * it is already off the queues).
2867 vm_page_lock_assert(m, MA_OWNED);
2868 if ((m->flags & PG_FICTITIOUS) != 0) {
2869 KASSERT(m->wire_count == 1,
2870 ("vm_page_wire: fictitious page %p's wire count isn't one",
2874 if (m->wire_count == 0) {
2875 KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
2876 m->queue == PQ_NONE,
2877 ("vm_page_wire: unmanaged page %p is queued", m));
2879 atomic_add_int(&vm_cnt.v_wire_count, 1);
2882 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
2888 * Release one wiring of the specified page, potentially allowing it to be
2889 * paged out. Returns TRUE if the number of wirings transitions to zero and
2892 * Only managed pages belonging to an object can be paged out. If the number
2893 * of wirings transitions to zero and the page is eligible for page out, then
2894 * the page is added to the specified paging queue (unless PQ_NONE is
2897 * If a page is fictitious, then its wire count must always be one.
2899 * A managed page must be locked.
2902 vm_page_unwire(vm_page_t m, uint8_t queue)
2905 KASSERT(queue < PQ_COUNT || queue == PQ_NONE,
2906 ("vm_page_unwire: invalid queue %u request for page %p",
2908 if ((m->oflags & VPO_UNMANAGED) == 0)
2909 vm_page_assert_locked(m);
2910 if ((m->flags & PG_FICTITIOUS) != 0) {
2911 KASSERT(m->wire_count == 1,
2912 ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
2915 if (m->wire_count > 0) {
2917 if (m->wire_count == 0) {
2918 atomic_subtract_int(&vm_cnt.v_wire_count, 1);
2919 if ((m->oflags & VPO_UNMANAGED) == 0 &&
2920 m->object != NULL && queue != PQ_NONE)
2921 vm_page_enqueue(queue, m);
2926 panic("vm_page_unwire: page %p's wire count is zero", m);
2930 * Move the specified page to the inactive queue.
2932 * Normally, "noreuse" is FALSE, resulting in LRU ordering of the inactive
2933 * queue. However, setting "noreuse" to TRUE will accelerate the specified
2934 * page's reclamation, but it will not unmap the page from any address space.
2935 * This is implemented by inserting the page near the head of the inactive
2936 * queue, using a marker page to guide FIFO insertion ordering.
2938 * The page must be locked.
2941 _vm_page_deactivate(vm_page_t m, boolean_t noreuse)
2943 struct vm_pagequeue *pq;
2946 vm_page_assert_locked(m);
2949 * Ignore if the page is already inactive, unless it is unlikely to be
2952 if ((queue = m->queue) == PQ_INACTIVE && !noreuse)
2954 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2955 pq = &vm_phys_domain(m)->vmd_pagequeues[PQ_INACTIVE];
2956 /* Avoid multiple acquisitions of the inactive queue lock. */
2957 if (queue == PQ_INACTIVE) {
2958 vm_pagequeue_lock(pq);
2959 vm_page_dequeue_locked(m);
2961 if (queue != PQ_NONE)
2963 vm_pagequeue_lock(pq);
2965 m->queue = PQ_INACTIVE;
2967 TAILQ_INSERT_BEFORE(&vm_phys_domain(m)->vmd_inacthead,
2970 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2971 vm_pagequeue_cnt_inc(pq);
2972 vm_pagequeue_unlock(pq);
2977 * Move the specified page to the inactive queue.
2979 * The page must be locked.
2982 vm_page_deactivate(vm_page_t m)
2985 _vm_page_deactivate(m, FALSE);
2989 * Move the specified page to the inactive queue with the expectation
2990 * that it is unlikely to be reused.
2992 * The page must be locked.
2995 vm_page_deactivate_noreuse(vm_page_t m)
2998 _vm_page_deactivate(m, TRUE);
3004 * Put a page in the laundry.
3007 vm_page_launder(vm_page_t m)
3011 vm_page_assert_locked(m);
3012 if ((queue = m->queue) != PQ_LAUNDRY) {
3013 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
3014 if (queue != PQ_NONE)
3016 vm_page_enqueue(PQ_LAUNDRY, m);
3018 KASSERT(queue == PQ_NONE,
3019 ("wired page %p is queued", m));
3024 * vm_page_unswappable
3026 * Put a page in the PQ_UNSWAPPABLE holding queue.
3029 vm_page_unswappable(vm_page_t m)
3032 vm_page_assert_locked(m);
3033 KASSERT(m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0,
3034 ("page %p already unswappable", m));
3035 if (m->queue != PQ_NONE)
3037 vm_page_enqueue(PQ_UNSWAPPABLE, m);
3041 * vm_page_try_to_free()
3043 * Attempt to free the page. If we cannot free it, we do nothing.
3044 * 1 is returned on success, 0 on failure.
3047 vm_page_try_to_free(vm_page_t m)
3050 vm_page_lock_assert(m, MA_OWNED);
3051 if (m->object != NULL)
3052 VM_OBJECT_ASSERT_WLOCKED(m->object);
3053 if (m->dirty || m->hold_count || m->wire_count ||
3054 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m))
3066 * Apply the specified advice to the given page.
3068 * The object and page must be locked.
3071 vm_page_advise(vm_page_t m, int advice)
3074 vm_page_assert_locked(m);
3075 VM_OBJECT_ASSERT_WLOCKED(m->object);
3076 if (advice == MADV_FREE)
3078 * Mark the page clean. This will allow the page to be freed
3079 * without first paging it out. MADV_FREE pages are often
3080 * quickly reused by malloc(3), so we do not do anything that
3081 * would result in a page fault on a later access.
3084 else if (advice != MADV_DONTNEED) {
3085 if (advice == MADV_WILLNEED)
3086 vm_page_activate(m);
3091 * Clear any references to the page. Otherwise, the page daemon will
3092 * immediately reactivate the page.
3094 vm_page_aflag_clear(m, PGA_REFERENCED);
3096 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
3100 * Place clean pages near the head of the inactive queue rather than
3101 * the tail, thus defeating the queue's LRU operation and ensuring that
3102 * the page will be reused quickly. Dirty pages not already in the
3103 * laundry are moved there.
3106 vm_page_deactivate_noreuse(m);
3112 * Grab a page, waiting until we are waken up due to the page
3113 * changing state. We keep on waiting, if the page continues
3114 * to be in the object. If the page doesn't exist, first allocate it
3115 * and then conditionally zero it.
3117 * This routine may sleep.
3119 * The object must be locked on entry. The lock will, however, be released
3120 * and reacquired if the routine sleeps.
3123 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
3128 VM_OBJECT_ASSERT_WLOCKED(object);
3129 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3130 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3131 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
3133 if ((m = vm_page_lookup(object, pindex)) != NULL) {
3134 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3135 vm_page_xbusied(m) : vm_page_busied(m);
3137 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3140 * Reference the page before unlocking and
3141 * sleeping so that the page daemon is less
3142 * likely to reclaim it.
3144 vm_page_aflag_set(m, PGA_REFERENCED);
3146 VM_OBJECT_WUNLOCK(object);
3147 vm_page_busy_sleep(m, "pgrbwt", (allocflags &
3148 VM_ALLOC_IGN_SBUSY) != 0);
3149 VM_OBJECT_WLOCK(object);
3152 if ((allocflags & VM_ALLOC_WIRED) != 0) {
3158 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
3160 if ((allocflags & VM_ALLOC_SBUSY) != 0)
3165 m = vm_page_alloc(object, pindex, allocflags);
3167 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3169 VM_OBJECT_WUNLOCK(object);
3171 VM_OBJECT_WLOCK(object);
3174 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
3180 * Return the specified range of pages from the given object. For each
3181 * page offset within the range, if a page already exists within the object
3182 * at that offset and it is busy, then wait for it to change state. If,
3183 * instead, the page doesn't exist, then allocate it.
3185 * The caller must always specify an allocation class.
3187 * allocation classes:
3188 * VM_ALLOC_NORMAL normal process request
3189 * VM_ALLOC_SYSTEM system *really* needs the pages
3191 * The caller must always specify that the pages are to be busied and/or
3194 * optional allocation flags:
3195 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages
3196 * VM_ALLOC_NOBUSY do not exclusive busy the page
3197 * VM_ALLOC_NOWAIT do not sleep
3198 * VM_ALLOC_SBUSY set page to sbusy state
3199 * VM_ALLOC_WIRED wire the pages
3200 * VM_ALLOC_ZERO zero and validate any invalid pages
3202 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it
3203 * may return a partial prefix of the requested range.
3206 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
3207 vm_page_t *ma, int count)
3213 VM_OBJECT_ASSERT_WLOCKED(object);
3214 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
3215 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
3216 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
3217 (allocflags & VM_ALLOC_WIRED) != 0,
3218 ("vm_page_grab_pages: the pages must be busied or wired"));
3219 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3220 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3221 ("vm_page_grab_pages: VM_ALLOC_SBUSY/IGN_SBUSY mismatch"));
3226 m = vm_radix_lookup_le(&object->rtree, pindex + i);
3227 if (m == NULL || m->pindex != pindex + i) {
3231 mpred = TAILQ_PREV(m, pglist, listq);
3232 for (; i < count; i++) {
3234 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3235 vm_page_xbusied(m) : vm_page_busied(m);
3237 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3240 * Reference the page before unlocking and
3241 * sleeping so that the page daemon is less
3242 * likely to reclaim it.
3244 vm_page_aflag_set(m, PGA_REFERENCED);
3246 VM_OBJECT_WUNLOCK(object);
3247 vm_page_busy_sleep(m, "grbmaw", (allocflags &
3248 VM_ALLOC_IGN_SBUSY) != 0);
3249 VM_OBJECT_WLOCK(object);
3252 if ((allocflags & VM_ALLOC_WIRED) != 0) {
3257 if ((allocflags & (VM_ALLOC_NOBUSY |
3258 VM_ALLOC_SBUSY)) == 0)
3260 if ((allocflags & VM_ALLOC_SBUSY) != 0)
3263 m = vm_page_alloc_after(object, pindex + i,
3264 (allocflags & ~VM_ALLOC_IGN_SBUSY) |
3265 VM_ALLOC_COUNT(count - i), mpred);
3267 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3269 VM_OBJECT_WUNLOCK(object);
3271 VM_OBJECT_WLOCK(object);
3275 if (m->valid == 0 && (allocflags & VM_ALLOC_ZERO) != 0) {
3276 if ((m->flags & PG_ZERO) == 0)
3278 m->valid = VM_PAGE_BITS_ALL;
3281 m = vm_page_next(m);
3287 * Mapping function for valid or dirty bits in a page.
3289 * Inputs are required to range within a page.
3292 vm_page_bits(int base, int size)
3298 base + size <= PAGE_SIZE,
3299 ("vm_page_bits: illegal base/size %d/%d", base, size)
3302 if (size == 0) /* handle degenerate case */
3305 first_bit = base >> DEV_BSHIFT;
3306 last_bit = (base + size - 1) >> DEV_BSHIFT;
3308 return (((vm_page_bits_t)2 << last_bit) -
3309 ((vm_page_bits_t)1 << first_bit));
3313 * vm_page_set_valid_range:
3315 * Sets portions of a page valid. The arguments are expected
3316 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3317 * of any partial chunks touched by the range. The invalid portion of
3318 * such chunks will be zeroed.
3320 * (base + size) must be less then or equal to PAGE_SIZE.
3323 vm_page_set_valid_range(vm_page_t m, int base, int size)
3327 VM_OBJECT_ASSERT_WLOCKED(m->object);
3328 if (size == 0) /* handle degenerate case */
3332 * If the base is not DEV_BSIZE aligned and the valid
3333 * bit is clear, we have to zero out a portion of the
3336 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
3337 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
3338 pmap_zero_page_area(m, frag, base - frag);
3341 * If the ending offset is not DEV_BSIZE aligned and the
3342 * valid bit is clear, we have to zero out a portion of
3345 endoff = base + size;
3346 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
3347 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
3348 pmap_zero_page_area(m, endoff,
3349 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
3352 * Assert that no previously invalid block that is now being validated
3355 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
3356 ("vm_page_set_valid_range: page %p is dirty", m));
3359 * Set valid bits inclusive of any overlap.
3361 m->valid |= vm_page_bits(base, size);
3365 * Clear the given bits from the specified page's dirty field.
3367 static __inline void
3368 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
3371 #if PAGE_SIZE < 16384
3376 * If the object is locked and the page is neither exclusive busy nor
3377 * write mapped, then the page's dirty field cannot possibly be
3378 * set by a concurrent pmap operation.
3380 VM_OBJECT_ASSERT_WLOCKED(m->object);
3381 if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
3382 m->dirty &= ~pagebits;
3385 * The pmap layer can call vm_page_dirty() without
3386 * holding a distinguished lock. The combination of
3387 * the object's lock and an atomic operation suffice
3388 * to guarantee consistency of the page dirty field.
3390 * For PAGE_SIZE == 32768 case, compiler already
3391 * properly aligns the dirty field, so no forcible
3392 * alignment is needed. Only require existence of
3393 * atomic_clear_64 when page size is 32768.
3395 addr = (uintptr_t)&m->dirty;
3396 #if PAGE_SIZE == 32768
3397 atomic_clear_64((uint64_t *)addr, pagebits);
3398 #elif PAGE_SIZE == 16384
3399 atomic_clear_32((uint32_t *)addr, pagebits);
3400 #else /* PAGE_SIZE <= 8192 */
3402 * Use a trick to perform a 32-bit atomic on the
3403 * containing aligned word, to not depend on the existence
3404 * of atomic_clear_{8, 16}.
3406 shift = addr & (sizeof(uint32_t) - 1);
3407 #if BYTE_ORDER == BIG_ENDIAN
3408 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
3412 addr &= ~(sizeof(uint32_t) - 1);
3413 atomic_clear_32((uint32_t *)addr, pagebits << shift);
3414 #endif /* PAGE_SIZE */
3419 * vm_page_set_validclean:
3421 * Sets portions of a page valid and clean. The arguments are expected
3422 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3423 * of any partial chunks touched by the range. The invalid portion of
3424 * such chunks will be zero'd.
3426 * (base + size) must be less then or equal to PAGE_SIZE.
3429 vm_page_set_validclean(vm_page_t m, int base, int size)
3431 vm_page_bits_t oldvalid, pagebits;
3434 VM_OBJECT_ASSERT_WLOCKED(m->object);
3435 if (size == 0) /* handle degenerate case */
3439 * If the base is not DEV_BSIZE aligned and the valid
3440 * bit is clear, we have to zero out a portion of the
3443 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
3444 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
3445 pmap_zero_page_area(m, frag, base - frag);
3448 * If the ending offset is not DEV_BSIZE aligned and the
3449 * valid bit is clear, we have to zero out a portion of
3452 endoff = base + size;
3453 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
3454 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
3455 pmap_zero_page_area(m, endoff,
3456 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
3459 * Set valid, clear dirty bits. If validating the entire
3460 * page we can safely clear the pmap modify bit. We also
3461 * use this opportunity to clear the VPO_NOSYNC flag. If a process
3462 * takes a write fault on a MAP_NOSYNC memory area the flag will
3465 * We set valid bits inclusive of any overlap, but we can only
3466 * clear dirty bits for DEV_BSIZE chunks that are fully within
3469 oldvalid = m->valid;
3470 pagebits = vm_page_bits(base, size);
3471 m->valid |= pagebits;
3473 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
3474 frag = DEV_BSIZE - frag;
3480 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
3482 if (base == 0 && size == PAGE_SIZE) {
3484 * The page can only be modified within the pmap if it is
3485 * mapped, and it can only be mapped if it was previously
3488 if (oldvalid == VM_PAGE_BITS_ALL)
3490 * Perform the pmap_clear_modify() first. Otherwise,
3491 * a concurrent pmap operation, such as
3492 * pmap_protect(), could clear a modification in the
3493 * pmap and set the dirty field on the page before
3494 * pmap_clear_modify() had begun and after the dirty
3495 * field was cleared here.
3497 pmap_clear_modify(m);
3499 m->oflags &= ~VPO_NOSYNC;
3500 } else if (oldvalid != VM_PAGE_BITS_ALL)
3501 m->dirty &= ~pagebits;
3503 vm_page_clear_dirty_mask(m, pagebits);
3507 vm_page_clear_dirty(vm_page_t m, int base, int size)
3510 vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
3514 * vm_page_set_invalid:
3516 * Invalidates DEV_BSIZE'd chunks within a page. Both the
3517 * valid and dirty bits for the effected areas are cleared.
3520 vm_page_set_invalid(vm_page_t m, int base, int size)
3522 vm_page_bits_t bits;
3526 VM_OBJECT_ASSERT_WLOCKED(object);
3527 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
3528 size >= object->un_pager.vnp.vnp_size)
3529 bits = VM_PAGE_BITS_ALL;
3531 bits = vm_page_bits(base, size);
3532 if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL &&
3535 KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) ||
3536 !pmap_page_is_mapped(m),
3537 ("vm_page_set_invalid: page %p is mapped", m));
3543 * vm_page_zero_invalid()
3545 * The kernel assumes that the invalid portions of a page contain
3546 * garbage, but such pages can be mapped into memory by user code.
3547 * When this occurs, we must zero out the non-valid portions of the
3548 * page so user code sees what it expects.
3550 * Pages are most often semi-valid when the end of a file is mapped
3551 * into memory and the file's size is not page aligned.
3554 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
3559 VM_OBJECT_ASSERT_WLOCKED(m->object);
3561 * Scan the valid bits looking for invalid sections that
3562 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the
3563 * valid bit may be set ) have already been zeroed by
3564 * vm_page_set_validclean().
3566 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
3567 if (i == (PAGE_SIZE / DEV_BSIZE) ||
3568 (m->valid & ((vm_page_bits_t)1 << i))) {
3570 pmap_zero_page_area(m,
3571 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
3578 * setvalid is TRUE when we can safely set the zero'd areas
3579 * as being valid. We can do this if there are no cache consistancy
3580 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
3583 m->valid = VM_PAGE_BITS_ALL;
3589 * Is (partial) page valid? Note that the case where size == 0
3590 * will return FALSE in the degenerate case where the page is
3591 * entirely invalid, and TRUE otherwise.
3594 vm_page_is_valid(vm_page_t m, int base, int size)
3596 vm_page_bits_t bits;
3598 VM_OBJECT_ASSERT_LOCKED(m->object);
3599 bits = vm_page_bits(base, size);
3600 return (m->valid != 0 && (m->valid & bits) == bits);
3604 * Returns true if all of the specified predicates are true for the entire
3605 * (super)page and false otherwise.
3608 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
3614 VM_OBJECT_ASSERT_LOCKED(object);
3615 npages = atop(pagesizes[m->psind]);
3618 * The physically contiguous pages that make up a superpage, i.e., a
3619 * page with a page size index ("psind") greater than zero, will
3620 * occupy adjacent entries in vm_page_array[].
3622 for (i = 0; i < npages; i++) {
3623 /* Always test object consistency, including "skip_m". */
3624 if (m[i].object != object)
3626 if (&m[i] == skip_m)
3628 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
3630 if ((flags & PS_ALL_DIRTY) != 0) {
3632 * Calling vm_page_test_dirty() or pmap_is_modified()
3633 * might stop this case from spuriously returning
3634 * "false". However, that would require a write lock
3635 * on the object containing "m[i]".
3637 if (m[i].dirty != VM_PAGE_BITS_ALL)
3640 if ((flags & PS_ALL_VALID) != 0 &&
3641 m[i].valid != VM_PAGE_BITS_ALL)
3648 * Set the page's dirty bits if the page is modified.
3651 vm_page_test_dirty(vm_page_t m)
3654 VM_OBJECT_ASSERT_WLOCKED(m->object);
3655 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
3660 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
3663 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
3667 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
3670 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
3674 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
3677 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
3680 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
3682 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
3685 vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
3689 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
3692 mtx_assert_(vm_page_lockptr(m), a, file, line);
3698 vm_page_object_lock_assert(vm_page_t m)
3702 * Certain of the page's fields may only be modified by the
3703 * holder of the containing object's lock or the exclusive busy.
3704 * holder. Unfortunately, the holder of the write busy is
3705 * not recorded, and thus cannot be checked here.
3707 if (m->object != NULL && !vm_page_xbusied(m))
3708 VM_OBJECT_ASSERT_WLOCKED(m->object);
3712 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits)
3715 if ((bits & PGA_WRITEABLE) == 0)
3719 * The PGA_WRITEABLE flag can only be set if the page is
3720 * managed, is exclusively busied or the object is locked.
3721 * Currently, this flag is only set by pmap_enter().
3723 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3724 ("PGA_WRITEABLE on unmanaged page"));
3725 if (!vm_page_xbusied(m))
3726 VM_OBJECT_ASSERT_LOCKED(m->object);
3730 #include "opt_ddb.h"
3732 #include <sys/kernel.h>
3734 #include <ddb/ddb.h>
3736 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3739 db_printf("vm_cnt.v_free_count: %d\n", vm_cnt.v_free_count);
3740 db_printf("vm_cnt.v_inactive_count: %d\n", vm_cnt.v_inactive_count);
3741 db_printf("vm_cnt.v_active_count: %d\n", vm_cnt.v_active_count);
3742 db_printf("vm_cnt.v_laundry_count: %d\n", vm_cnt.v_laundry_count);
3743 db_printf("vm_cnt.v_wire_count: %d\n", vm_cnt.v_wire_count);
3744 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
3745 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
3746 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
3747 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
3750 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3754 db_printf("pq_free %d\n", vm_cnt.v_free_count);
3755 for (dom = 0; dom < vm_ndomains; dom++) {
3757 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
3759 vm_dom[dom].vmd_page_count,
3760 vm_dom[dom].vmd_free_count,
3761 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
3762 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
3763 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
3764 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
3768 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
3774 db_printf("show pginfo addr\n");
3778 phys = strchr(modif, 'p') != NULL;
3780 m = PHYS_TO_VM_PAGE(addr);
3782 m = (vm_page_t)addr;
3784 "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n"
3785 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
3786 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
3787 m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags,
3788 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty);