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_phys(vm_page_t m);
167 static void vm_page_free_wakeup(void);
168 static void vm_page_init(void *dummy);
169 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
170 vm_pindex_t pindex, vm_page_t mpred);
171 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
173 static int vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run,
176 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
179 vm_page_init(void *dummy)
182 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
183 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
184 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
185 VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
188 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
189 #if PAGE_SIZE == 32768
191 CTASSERT(sizeof(u_long) >= 8);
196 * Try to acquire a physical address lock while a pmap is locked. If we
197 * fail to trylock we unlock and lock the pmap directly and cache the
198 * locked pa in *locked. The caller should then restart their loop in case
199 * the virtual to physical mapping has changed.
202 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
209 PA_LOCK_ASSERT(lockpa, MA_OWNED);
210 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
217 atomic_add_int(&pa_tryrelock_restart, 1);
226 * Sets the page size, perhaps based upon the memory
227 * size. Must be called before any use of page-size
228 * dependent functions.
231 vm_set_page_size(void)
233 if (vm_cnt.v_page_size == 0)
234 vm_cnt.v_page_size = PAGE_SIZE;
235 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
236 panic("vm_set_page_size: page size not a power of two");
240 * vm_page_blacklist_next:
242 * Find the next entry in the provided string of blacklist
243 * addresses. Entries are separated by space, comma, or newline.
244 * If an invalid integer is encountered then the rest of the
245 * string is skipped. Updates the list pointer to the next
246 * character, or NULL if the string is exhausted or invalid.
249 vm_page_blacklist_next(char **list, char *end)
254 if (list == NULL || *list == NULL)
262 * If there's no end pointer then the buffer is coming from
263 * the kenv and we know it's null-terminated.
266 end = *list + strlen(*list);
268 /* Ensure that strtoq() won't walk off the end */
270 if (*end == '\n' || *end == ' ' || *end == ',')
273 printf("Blacklist not terminated, skipping\n");
279 for (pos = *list; *pos != '\0'; pos = cp) {
280 bad = strtoq(pos, &cp, 0);
281 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
290 if (*cp == '\0' || ++cp >= end)
294 return (trunc_page(bad));
296 printf("Garbage in RAM blacklist, skipping\n");
302 * vm_page_blacklist_check:
304 * Iterate through the provided string of blacklist addresses, pulling
305 * each entry out of the physical allocator free list and putting it
306 * onto a list for reporting via the vm.page_blacklist sysctl.
309 vm_page_blacklist_check(char *list, char *end)
317 while (next != NULL) {
318 if ((pa = vm_page_blacklist_next(&next, end)) == 0)
320 m = vm_phys_paddr_to_vm_page(pa);
323 mtx_lock(&vm_page_queue_free_mtx);
324 ret = vm_phys_unfree_page(m);
325 mtx_unlock(&vm_page_queue_free_mtx);
327 TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
329 printf("Skipping page with pa 0x%jx\n",
336 * vm_page_blacklist_load:
338 * Search for a special module named "ram_blacklist". It'll be a
339 * plain text file provided by the user via the loader directive
343 vm_page_blacklist_load(char **list, char **end)
352 mod = preload_search_by_type("ram_blacklist");
354 ptr = preload_fetch_addr(mod);
355 len = preload_fetch_size(mod);
366 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
373 error = sysctl_wire_old_buffer(req, 0);
376 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
377 TAILQ_FOREACH(m, &blacklist_head, listq) {
378 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
379 (uintmax_t)m->phys_addr);
382 error = sbuf_finish(&sbuf);
388 vm_page_domain_init(struct vm_domain *vmd)
390 struct vm_pagequeue *pq;
393 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
394 "vm inactive pagequeue";
395 *__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_vcnt) =
396 &vm_cnt.v_inactive_count;
397 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
398 "vm active pagequeue";
399 *__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_vcnt) =
400 &vm_cnt.v_active_count;
401 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
402 "vm laundry pagequeue";
403 *__DECONST(int **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_vcnt) =
404 &vm_cnt.v_laundry_count;
405 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
406 "vm unswappable pagequeue";
407 /* Unswappable dirty pages are counted as being in the laundry. */
408 *__DECONST(int **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_vcnt) =
409 &vm_cnt.v_laundry_count;
410 vmd->vmd_page_count = 0;
411 vmd->vmd_free_count = 0;
413 vmd->vmd_oom = FALSE;
414 for (i = 0; i < PQ_COUNT; i++) {
415 pq = &vmd->vmd_pagequeues[i];
416 TAILQ_INIT(&pq->pq_pl);
417 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
418 MTX_DEF | MTX_DUPOK);
425 * Initializes the resident memory module. Allocates physical memory for
426 * bootstrapping UMA and some data structures that are used to manage
427 * physical pages. Initializes these structures, and populates the free
431 vm_page_startup(vm_offset_t vaddr)
433 struct vm_domain *vmd;
434 struct vm_phys_seg *seg;
436 char *list, *listend;
438 vm_paddr_t end, high_avail, low_avail, new_end, page_range, size;
439 vm_paddr_t biggestsize, last_pa, pa;
441 int biggestone, i, pages_per_zone, segind;
445 vaddr = round_page(vaddr);
447 for (i = 0; phys_avail[i + 1]; i += 2) {
448 phys_avail[i] = round_page(phys_avail[i]);
449 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
451 for (i = 0; phys_avail[i + 1]; i += 2) {
452 size = phys_avail[i + 1] - phys_avail[i];
453 if (size > biggestsize) {
459 end = phys_avail[biggestone+1];
462 * Initialize the page and queue locks.
464 mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF);
465 for (i = 0; i < PA_LOCK_COUNT; i++)
466 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
467 for (i = 0; i < vm_ndomains; i++)
468 vm_page_domain_init(&vm_dom[i]);
471 * Almost all of the pages needed for bootstrapping UMA are used
472 * for zone structures, so if the number of CPUs results in those
473 * structures taking more than one page each, we set aside more pages
474 * in proportion to the zone structure size.
476 pages_per_zone = howmany(sizeof(struct uma_zone) +
477 sizeof(struct uma_cache) * (mp_maxid + 1) +
478 roundup2(sizeof(struct uma_slab), sizeof(void *)), UMA_SLAB_SIZE);
479 if (pages_per_zone > 1) {
480 /* Reserve more pages so that we don't run out. */
481 boot_pages = UMA_BOOT_PAGES_ZONES * pages_per_zone;
485 * Allocate memory for use when boot strapping the kernel memory
488 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
489 * manually fetch the value.
491 TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
492 new_end = end - (boot_pages * UMA_SLAB_SIZE);
493 new_end = trunc_page(new_end);
494 mapped = pmap_map(&vaddr, new_end, end,
495 VM_PROT_READ | VM_PROT_WRITE);
496 bzero((void *)mapped, end - new_end);
497 uma_startup((void *)mapped, boot_pages);
499 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
500 defined(__i386__) || defined(__mips__)
502 * Allocate a bitmap to indicate that a random physical page
503 * needs to be included in a minidump.
505 * The amd64 port needs this to indicate which direct map pages
506 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
508 * However, i386 still needs this workspace internally within the
509 * minidump code. In theory, they are not needed on i386, but are
510 * included should the sf_buf code decide to use them.
513 for (i = 0; dump_avail[i + 1] != 0; i += 2)
514 if (dump_avail[i + 1] > last_pa)
515 last_pa = dump_avail[i + 1];
516 page_range = last_pa / PAGE_SIZE;
517 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
518 new_end -= vm_page_dump_size;
519 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
520 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
521 bzero((void *)vm_page_dump, vm_page_dump_size);
525 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
527 * Include the UMA bootstrap pages and vm_page_dump in a crash dump.
528 * When pmap_map() uses the direct map, they are not automatically
531 for (pa = new_end; pa < end; pa += PAGE_SIZE)
534 phys_avail[biggestone + 1] = new_end;
537 * Request that the physical pages underlying the message buffer be
538 * included in a crash dump. Since the message buffer is accessed
539 * through the direct map, they are not automatically included.
541 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
542 last_pa = pa + round_page(msgbufsize);
543 while (pa < last_pa) {
549 * Compute the number of pages of memory that will be available for
550 * use, taking into account the overhead of a page structure per page.
551 * In other words, solve
552 * "available physical memory" - round_page(page_range *
553 * sizeof(struct vm_page)) = page_range * PAGE_SIZE
556 low_avail = phys_avail[0];
557 high_avail = phys_avail[1];
558 for (i = 0; i < vm_phys_nsegs; i++) {
559 if (vm_phys_segs[i].start < low_avail)
560 low_avail = vm_phys_segs[i].start;
561 if (vm_phys_segs[i].end > high_avail)
562 high_avail = vm_phys_segs[i].end;
564 /* Skip the first chunk. It is already accounted for. */
565 for (i = 2; phys_avail[i + 1] != 0; i += 2) {
566 if (phys_avail[i] < low_avail)
567 low_avail = phys_avail[i];
568 if (phys_avail[i + 1] > high_avail)
569 high_avail = phys_avail[i + 1];
571 first_page = low_avail / PAGE_SIZE;
572 #ifdef VM_PHYSSEG_SPARSE
574 for (i = 0; i < vm_phys_nsegs; i++)
575 size += vm_phys_segs[i].end - vm_phys_segs[i].start;
576 for (i = 0; phys_avail[i + 1] != 0; i += 2)
577 size += phys_avail[i + 1] - phys_avail[i];
578 #elif defined(VM_PHYSSEG_DENSE)
579 size = high_avail - low_avail;
581 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
584 #ifdef VM_PHYSSEG_DENSE
586 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
587 * the overhead of a page structure per page only if vm_page_array is
588 * allocated from the last physical memory chunk. Otherwise, we must
589 * allocate page structures representing the physical memory
590 * underlying vm_page_array, even though they will not be used.
592 if (new_end != high_avail)
593 page_range = size / PAGE_SIZE;
597 page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
600 * If the partial bytes remaining are large enough for
601 * a page (PAGE_SIZE) without a corresponding
602 * 'struct vm_page', then new_end will contain an
603 * extra page after subtracting the length of the VM
604 * page array. Compensate by subtracting an extra
607 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
608 if (new_end == high_avail)
609 high_avail -= PAGE_SIZE;
610 new_end -= PAGE_SIZE;
616 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
617 * However, because this page is allocated from KVM, out-of-bounds
618 * accesses using the direct map will not be trapped.
623 * Allocate physical memory for the page structures, and map it.
625 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
626 mapped = pmap_map(&vaddr, new_end, end,
627 VM_PROT_READ | VM_PROT_WRITE);
628 vm_page_array = (vm_page_t)mapped;
629 vm_page_array_size = page_range;
631 #if VM_NRESERVLEVEL > 0
633 * Allocate physical memory for the reservation management system's
634 * data structures, and map it.
636 if (high_avail == end)
637 high_avail = new_end;
638 new_end = vm_reserv_startup(&vaddr, new_end, high_avail);
640 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
642 * Include vm_page_array and vm_reserv_array in a crash dump.
644 for (pa = new_end; pa < end; pa += PAGE_SIZE)
647 phys_avail[biggestone + 1] = new_end;
650 * Add physical memory segments corresponding to the available
653 for (i = 0; phys_avail[i + 1] != 0; i += 2)
654 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
657 * Initialize the physical memory allocator.
662 * Initialize the page structures and add every available page to the
663 * physical memory allocator's free lists.
665 vm_cnt.v_page_count = 0;
666 vm_cnt.v_free_count = 0;
667 for (segind = 0; segind < vm_phys_nsegs; segind++) {
668 seg = &vm_phys_segs[segind];
669 for (pa = seg->start; pa < seg->end; pa += PAGE_SIZE)
670 vm_phys_init_page(pa);
673 * Add the segment to the free lists only if it is covered by
674 * one of the ranges in phys_avail. Because we've added the
675 * ranges to the vm_phys_segs array, we can assume that each
676 * segment is either entirely contained in one of the ranges,
677 * or doesn't overlap any of them.
679 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
680 if (seg->start < phys_avail[i] ||
681 seg->end > phys_avail[i + 1])
685 pagecount = (u_long)atop(seg->end - seg->start);
687 mtx_lock(&vm_page_queue_free_mtx);
688 vm_phys_free_contig(m, pagecount);
689 vm_phys_freecnt_adj(m, (int)pagecount);
690 mtx_unlock(&vm_page_queue_free_mtx);
691 vm_cnt.v_page_count += (u_int)pagecount;
693 vmd = &vm_dom[seg->domain];
694 vmd->vmd_page_count += (u_int)pagecount;
695 vmd->vmd_segs |= 1UL << m->segind;
701 * Remove blacklisted pages from the physical memory allocator.
703 TAILQ_INIT(&blacklist_head);
704 vm_page_blacklist_load(&list, &listend);
705 vm_page_blacklist_check(list, listend);
707 list = kern_getenv("vm.blacklist");
708 vm_page_blacklist_check(list, NULL);
711 #if VM_NRESERVLEVEL > 0
713 * Initialize the reservation management system.
721 vm_page_reference(vm_page_t m)
724 vm_page_aflag_set(m, PGA_REFERENCED);
728 * vm_page_busy_downgrade:
730 * Downgrade an exclusive busy page into a single shared busy page.
733 vm_page_busy_downgrade(vm_page_t m)
738 vm_page_assert_xbusied(m);
739 locked = mtx_owned(vm_page_lockptr(m));
743 x &= VPB_BIT_WAITERS;
744 if (x != 0 && !locked)
746 if (atomic_cmpset_rel_int(&m->busy_lock,
747 VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1)))
749 if (x != 0 && !locked)
762 * Return a positive value if the page is shared busied, 0 otherwise.
765 vm_page_sbusied(vm_page_t m)
770 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
776 * Shared unbusy a page.
779 vm_page_sunbusy(vm_page_t m)
783 vm_page_lock_assert(m, MA_NOTOWNED);
784 vm_page_assert_sbusied(m);
788 if (VPB_SHARERS(x) > 1) {
789 if (atomic_cmpset_int(&m->busy_lock, x,
794 if ((x & VPB_BIT_WAITERS) == 0) {
795 KASSERT(x == VPB_SHARERS_WORD(1),
796 ("vm_page_sunbusy: invalid lock state"));
797 if (atomic_cmpset_int(&m->busy_lock,
798 VPB_SHARERS_WORD(1), VPB_UNBUSIED))
802 KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS),
803 ("vm_page_sunbusy: invalid lock state for waiters"));
806 if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) {
817 * vm_page_busy_sleep:
819 * Sleep and release the page lock, using the page pointer as wchan.
820 * This is used to implement the hard-path of busying mechanism.
822 * The given page must be locked.
824 * If nonshared is true, sleep only if the page is xbusy.
827 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
831 vm_page_assert_locked(m);
834 if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) ||
835 ((x & VPB_BIT_WAITERS) == 0 &&
836 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) {
840 msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0);
846 * Try to shared busy a page.
847 * If the operation succeeds 1 is returned otherwise 0.
848 * The operation never sleeps.
851 vm_page_trysbusy(vm_page_t m)
857 if ((x & VPB_BIT_SHARED) == 0)
859 if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER))
865 vm_page_xunbusy_locked(vm_page_t m)
868 vm_page_assert_xbusied(m);
869 vm_page_assert_locked(m);
871 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
872 /* There is a waiter, do wakeup() instead of vm_page_flash(). */
877 vm_page_xunbusy_maybelocked(vm_page_t m)
881 vm_page_assert_xbusied(m);
884 * Fast path for unbusy. If it succeeds, we know that there
885 * are no waiters, so we do not need a wakeup.
887 if (atomic_cmpset_rel_int(&m->busy_lock, VPB_SINGLE_EXCLUSIVER,
891 lockacq = !mtx_owned(vm_page_lockptr(m));
894 vm_page_xunbusy_locked(m);
900 * vm_page_xunbusy_hard:
902 * Called after the first try the exclusive unbusy of a page failed.
903 * It is assumed that the waiters bit is on.
906 vm_page_xunbusy_hard(vm_page_t m)
909 vm_page_assert_xbusied(m);
912 vm_page_xunbusy_locked(m);
919 * Wakeup anyone waiting for the page.
920 * The ownership bits do not change.
922 * The given page must be locked.
925 vm_page_flash(vm_page_t m)
929 vm_page_lock_assert(m, MA_OWNED);
933 if ((x & VPB_BIT_WAITERS) == 0)
935 if (atomic_cmpset_int(&m->busy_lock, x,
936 x & (~VPB_BIT_WAITERS)))
943 * Avoid releasing and reacquiring the same page lock.
946 vm_page_change_lock(vm_page_t m, struct mtx **mtx)
950 mtx1 = vm_page_lockptr(m);
960 * Keep page from being freed by the page daemon
961 * much of the same effect as wiring, except much lower
962 * overhead and should be used only for *very* temporary
963 * holding ("wiring").
966 vm_page_hold(vm_page_t mem)
969 vm_page_lock_assert(mem, MA_OWNED);
974 vm_page_unhold(vm_page_t mem)
977 vm_page_lock_assert(mem, MA_OWNED);
978 KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!"));
980 if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0)
981 vm_page_free_toq(mem);
985 * vm_page_unhold_pages:
987 * Unhold each of the pages that is referenced by the given array.
990 vm_page_unhold_pages(vm_page_t *ma, int count)
995 for (; count != 0; count--) {
996 vm_page_change_lock(*ma, &mtx);
1005 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1009 #ifdef VM_PHYSSEG_SPARSE
1010 m = vm_phys_paddr_to_vm_page(pa);
1012 m = vm_phys_fictitious_to_vm_page(pa);
1014 #elif defined(VM_PHYSSEG_DENSE)
1018 if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1019 m = &vm_page_array[pi - first_page];
1022 return (vm_phys_fictitious_to_vm_page(pa));
1024 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1031 * Create a fictitious page with the specified physical address and
1032 * memory attribute. The memory attribute is the only the machine-
1033 * dependent aspect of a fictitious page that must be initialized.
1036 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1040 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1041 vm_page_initfake(m, paddr, memattr);
1046 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1049 if ((m->flags & PG_FICTITIOUS) != 0) {
1051 * The page's memattr might have changed since the
1052 * previous initialization. Update the pmap to the
1057 m->phys_addr = paddr;
1059 /* Fictitious pages don't use "segind". */
1060 m->flags = PG_FICTITIOUS;
1061 /* Fictitious pages don't use "order" or "pool". */
1062 m->oflags = VPO_UNMANAGED;
1063 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1067 pmap_page_set_memattr(m, memattr);
1073 * Release a fictitious page.
1076 vm_page_putfake(vm_page_t m)
1079 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1080 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1081 ("vm_page_putfake: bad page %p", m));
1082 uma_zfree(fakepg_zone, m);
1086 * vm_page_updatefake:
1088 * Update the given fictitious page to the specified physical address and
1092 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1095 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1096 ("vm_page_updatefake: bad page %p", m));
1097 m->phys_addr = paddr;
1098 pmap_page_set_memattr(m, memattr);
1107 vm_page_free(vm_page_t m)
1110 m->flags &= ~PG_ZERO;
1111 vm_page_free_toq(m);
1115 * vm_page_free_zero:
1117 * Free a page to the zerod-pages queue
1120 vm_page_free_zero(vm_page_t m)
1123 m->flags |= PG_ZERO;
1124 vm_page_free_toq(m);
1128 * Unbusy and handle the page queueing for a page from a getpages request that
1129 * was optionally read ahead or behind.
1132 vm_page_readahead_finish(vm_page_t m)
1135 /* We shouldn't put invalid pages on queues. */
1136 KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m));
1139 * Since the page is not the actually needed one, whether it should
1140 * be activated or deactivated is not obvious. Empirical results
1141 * have shown that deactivating the page is usually the best choice,
1142 * unless the page is wanted by another thread.
1145 if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1146 vm_page_activate(m);
1148 vm_page_deactivate(m);
1154 * vm_page_sleep_if_busy:
1156 * Sleep and release the page queues lock if the page is busied.
1157 * Returns TRUE if the thread slept.
1159 * The given page must be unlocked and object containing it must
1163 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1167 vm_page_lock_assert(m, MA_NOTOWNED);
1168 VM_OBJECT_ASSERT_WLOCKED(m->object);
1170 if (vm_page_busied(m)) {
1172 * The page-specific object must be cached because page
1173 * identity can change during the sleep, causing the
1174 * re-lock of a different object.
1175 * It is assumed that a reference to the object is already
1176 * held by the callers.
1180 VM_OBJECT_WUNLOCK(obj);
1181 vm_page_busy_sleep(m, msg, false);
1182 VM_OBJECT_WLOCK(obj);
1189 * vm_page_dirty_KBI: [ internal use only ]
1191 * Set all bits in the page's dirty field.
1193 * The object containing the specified page must be locked if the
1194 * call is made from the machine-independent layer.
1196 * See vm_page_clear_dirty_mask().
1198 * This function should only be called by vm_page_dirty().
1201 vm_page_dirty_KBI(vm_page_t m)
1204 /* Refer to this operation by its public name. */
1205 KASSERT(m->valid == VM_PAGE_BITS_ALL,
1206 ("vm_page_dirty: page is invalid!"));
1207 m->dirty = VM_PAGE_BITS_ALL;
1211 * vm_page_insert: [ internal use only ]
1213 * Inserts the given mem entry into the object and object list.
1215 * The object must be locked.
1218 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1222 VM_OBJECT_ASSERT_WLOCKED(object);
1223 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1224 return (vm_page_insert_after(m, object, pindex, mpred));
1228 * vm_page_insert_after:
1230 * Inserts the page "m" into the specified object at offset "pindex".
1232 * The page "mpred" must immediately precede the offset "pindex" within
1233 * the specified object.
1235 * The object must be locked.
1238 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1243 VM_OBJECT_ASSERT_WLOCKED(object);
1244 KASSERT(m->object == NULL,
1245 ("vm_page_insert_after: page already inserted"));
1246 if (mpred != NULL) {
1247 KASSERT(mpred->object == object,
1248 ("vm_page_insert_after: object doesn't contain mpred"));
1249 KASSERT(mpred->pindex < pindex,
1250 ("vm_page_insert_after: mpred doesn't precede pindex"));
1251 msucc = TAILQ_NEXT(mpred, listq);
1253 msucc = TAILQ_FIRST(&object->memq);
1255 KASSERT(msucc->pindex > pindex,
1256 ("vm_page_insert_after: msucc doesn't succeed pindex"));
1259 * Record the object/offset pair in this page
1265 * Now link into the object's ordered list of backed pages.
1267 if (vm_radix_insert(&object->rtree, m)) {
1272 vm_page_insert_radixdone(m, object, mpred);
1277 * vm_page_insert_radixdone:
1279 * Complete page "m" insertion into the specified object after the
1280 * radix trie hooking.
1282 * The page "mpred" must precede the offset "m->pindex" within the
1285 * The object must be locked.
1288 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1291 VM_OBJECT_ASSERT_WLOCKED(object);
1292 KASSERT(object != NULL && m->object == object,
1293 ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1294 if (mpred != NULL) {
1295 KASSERT(mpred->object == object,
1296 ("vm_page_insert_after: object doesn't contain mpred"));
1297 KASSERT(mpred->pindex < m->pindex,
1298 ("vm_page_insert_after: mpred doesn't precede pindex"));
1302 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1304 TAILQ_INSERT_HEAD(&object->memq, m, listq);
1307 * Show that the object has one more resident page.
1309 object->resident_page_count++;
1312 * Hold the vnode until the last page is released.
1314 if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1315 vhold(object->handle);
1318 * Since we are inserting a new and possibly dirty page,
1319 * update the object's OBJ_MIGHTBEDIRTY flag.
1321 if (pmap_page_is_write_mapped(m))
1322 vm_object_set_writeable_dirty(object);
1328 * Removes the specified page from its containing object, but does not
1329 * invalidate any backing storage.
1331 * The object must be locked. The page must be locked if it is managed.
1334 vm_page_remove(vm_page_t m)
1339 if ((m->oflags & VPO_UNMANAGED) == 0)
1340 vm_page_assert_locked(m);
1341 if ((object = m->object) == NULL)
1343 VM_OBJECT_ASSERT_WLOCKED(object);
1344 if (vm_page_xbusied(m))
1345 vm_page_xunbusy_maybelocked(m);
1346 mrem = vm_radix_remove(&object->rtree, m->pindex);
1347 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1350 * Now remove from the object's list of backed pages.
1352 TAILQ_REMOVE(&object->memq, m, listq);
1355 * And show that the object has one fewer resident page.
1357 object->resident_page_count--;
1360 * The vnode may now be recycled.
1362 if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1363 vdrop(object->handle);
1371 * Returns the page associated with the object/offset
1372 * pair specified; if none is found, NULL is returned.
1374 * The object must be locked.
1377 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1380 VM_OBJECT_ASSERT_LOCKED(object);
1381 return (vm_radix_lookup(&object->rtree, pindex));
1385 * vm_page_find_least:
1387 * Returns the page associated with the object with least pindex
1388 * greater than or equal to the parameter pindex, or NULL.
1390 * The object must be locked.
1393 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1397 VM_OBJECT_ASSERT_LOCKED(object);
1398 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1399 m = vm_radix_lookup_ge(&object->rtree, pindex);
1404 * Returns the given page's successor (by pindex) within the object if it is
1405 * resident; if none is found, NULL is returned.
1407 * The object must be locked.
1410 vm_page_next(vm_page_t m)
1414 VM_OBJECT_ASSERT_LOCKED(m->object);
1415 if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1416 MPASS(next->object == m->object);
1417 if (next->pindex != m->pindex + 1)
1424 * Returns the given page's predecessor (by pindex) within the object if it is
1425 * resident; if none is found, NULL is returned.
1427 * The object must be locked.
1430 vm_page_prev(vm_page_t m)
1434 VM_OBJECT_ASSERT_LOCKED(m->object);
1435 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1436 MPASS(prev->object == m->object);
1437 if (prev->pindex != m->pindex - 1)
1444 * Uses the page mnew as a replacement for an existing page at index
1445 * pindex which must be already present in the object.
1447 * The existing page must not be on a paging queue.
1450 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1454 VM_OBJECT_ASSERT_WLOCKED(object);
1455 KASSERT(mnew->object == NULL,
1456 ("vm_page_replace: page already in object"));
1459 * This function mostly follows vm_page_insert() and
1460 * vm_page_remove() without the radix, object count and vnode
1461 * dance. Double check such functions for more comments.
1464 mnew->object = object;
1465 mnew->pindex = pindex;
1466 mold = vm_radix_replace(&object->rtree, mnew);
1467 KASSERT(mold->queue == PQ_NONE,
1468 ("vm_page_replace: mold is on a paging queue"));
1470 /* Keep the resident page list in sorted order. */
1471 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1472 TAILQ_REMOVE(&object->memq, mold, listq);
1474 mold->object = NULL;
1475 vm_page_xunbusy_maybelocked(mold);
1478 * The object's resident_page_count does not change because we have
1479 * swapped one page for another, but OBJ_MIGHTBEDIRTY.
1481 if (pmap_page_is_write_mapped(mnew))
1482 vm_object_set_writeable_dirty(object);
1489 * Move the given memory entry from its
1490 * current object to the specified target object/offset.
1492 * Note: swap associated with the page must be invalidated by the move. We
1493 * have to do this for several reasons: (1) we aren't freeing the
1494 * page, (2) we are dirtying the page, (3) the VM system is probably
1495 * moving the page from object A to B, and will then later move
1496 * the backing store from A to B and we can't have a conflict.
1498 * Note: we *always* dirty the page. It is necessary both for the
1499 * fact that we moved it, and because we may be invalidating
1502 * The objects must be locked.
1505 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1510 VM_OBJECT_ASSERT_WLOCKED(new_object);
1512 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1513 KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1514 ("vm_page_rename: pindex already renamed"));
1517 * Create a custom version of vm_page_insert() which does not depend
1518 * by m_prev and can cheat on the implementation aspects of the
1522 m->pindex = new_pindex;
1523 if (vm_radix_insert(&new_object->rtree, m)) {
1529 * The operation cannot fail anymore. The removal must happen before
1530 * the listq iterator is tainted.
1536 /* Return back to the new pindex to complete vm_page_insert(). */
1537 m->pindex = new_pindex;
1538 m->object = new_object;
1540 vm_page_insert_radixdone(m, new_object, mpred);
1548 * Allocate and return a page that is associated with the specified
1549 * object and offset pair. By default, this page is exclusive busied.
1551 * The caller must always specify an allocation class.
1553 * allocation classes:
1554 * VM_ALLOC_NORMAL normal process request
1555 * VM_ALLOC_SYSTEM system *really* needs a page
1556 * VM_ALLOC_INTERRUPT interrupt time request
1558 * optional allocation flags:
1559 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1560 * intends to allocate
1561 * VM_ALLOC_NOBUSY do not exclusive busy the page
1562 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1563 * VM_ALLOC_NOOBJ page is not associated with an object and
1564 * should not be exclusive busy
1565 * VM_ALLOC_SBUSY shared busy the allocated page
1566 * VM_ALLOC_WIRED wire the allocated page
1567 * VM_ALLOC_ZERO prefer a zeroed page
1569 * This routine may not sleep.
1572 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1575 return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1576 vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1580 * Allocate a page in the specified object with the given page index. To
1581 * optimize insertion of the page into the object, the caller must also specifiy
1582 * the resident page in the object with largest index smaller than the given
1583 * page index, or NULL if no such page exists.
1586 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex, int req,
1590 int flags, req_class;
1592 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1593 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1594 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1595 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1596 ("inconsistent object(%p)/req(%x)", object, req));
1597 KASSERT(mpred == NULL || mpred->pindex < pindex,
1598 ("mpred %p doesn't precede pindex 0x%jx", mpred,
1599 (uintmax_t)pindex));
1601 VM_OBJECT_ASSERT_WLOCKED(object);
1603 req_class = req & VM_ALLOC_CLASS_MASK;
1606 * The page daemon is allowed to dig deeper into the free page list.
1608 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1609 req_class = VM_ALLOC_SYSTEM;
1612 * Allocate a page if the number of free pages exceeds the minimum
1613 * for the request class.
1615 mtx_lock(&vm_page_queue_free_mtx);
1616 if (vm_cnt.v_free_count > vm_cnt.v_free_reserved ||
1617 (req_class == VM_ALLOC_SYSTEM &&
1618 vm_cnt.v_free_count > vm_cnt.v_interrupt_free_min) ||
1619 (req_class == VM_ALLOC_INTERRUPT &&
1620 vm_cnt.v_free_count > 0)) {
1622 * Can we allocate the page from a reservation?
1624 #if VM_NRESERVLEVEL > 0
1625 if (object == NULL || (object->flags & (OBJ_COLORED |
1626 OBJ_FICTITIOUS)) != OBJ_COLORED || (m =
1627 vm_reserv_alloc_page(object, pindex, mpred)) == NULL)
1631 * If not, allocate it from the free page queues.
1633 m = vm_phys_alloc_pages(object != NULL ?
1634 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1635 #if VM_NRESERVLEVEL > 0
1636 if (m == NULL && vm_reserv_reclaim_inactive()) {
1637 m = vm_phys_alloc_pages(object != NULL ?
1638 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1645 * Not allocatable, give up.
1647 mtx_unlock(&vm_page_queue_free_mtx);
1648 atomic_add_int(&vm_pageout_deficit,
1649 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1650 pagedaemon_wakeup();
1655 * At this point we had better have found a good page.
1657 KASSERT(m != NULL, ("missing page"));
1658 vm_phys_freecnt_adj(m, -1);
1659 mtx_unlock(&vm_page_queue_free_mtx);
1660 vm_page_alloc_check(m);
1663 * Initialize the page. Only the PG_ZERO flag is inherited.
1666 if ((req & VM_ALLOC_ZERO) != 0)
1669 if ((req & VM_ALLOC_NODUMP) != 0)
1673 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1675 m->busy_lock = VPB_UNBUSIED;
1676 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1677 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1678 if ((req & VM_ALLOC_SBUSY) != 0)
1679 m->busy_lock = VPB_SHARERS_WORD(1);
1680 if (req & VM_ALLOC_WIRED) {
1682 * The page lock is not required for wiring a page until that
1683 * page is inserted into the object.
1685 atomic_add_int(&vm_cnt.v_wire_count, 1);
1690 if (object != NULL) {
1691 if (vm_page_insert_after(m, object, pindex, mpred)) {
1692 pagedaemon_wakeup();
1693 if (req & VM_ALLOC_WIRED) {
1694 atomic_subtract_int(&vm_cnt.v_wire_count, 1);
1697 KASSERT(m->object == NULL, ("page %p has object", m));
1698 m->oflags = VPO_UNMANAGED;
1699 m->busy_lock = VPB_UNBUSIED;
1700 /* Don't change PG_ZERO. */
1701 vm_page_free_toq(m);
1705 /* Ignore device objects; the pager sets "memattr" for them. */
1706 if (object->memattr != VM_MEMATTR_DEFAULT &&
1707 (object->flags & OBJ_FICTITIOUS) == 0)
1708 pmap_page_set_memattr(m, object->memattr);
1713 * Don't wakeup too often - wakeup the pageout daemon when
1714 * we would be nearly out of memory.
1716 if (vm_paging_needed())
1717 pagedaemon_wakeup();
1723 * vm_page_alloc_contig:
1725 * Allocate a contiguous set of physical pages of the given size "npages"
1726 * from the free lists. All of the physical pages must be at or above
1727 * the given physical address "low" and below the given physical address
1728 * "high". The given value "alignment" determines the alignment of the
1729 * first physical page in the set. If the given value "boundary" is
1730 * non-zero, then the set of physical pages cannot cross any physical
1731 * address boundary that is a multiple of that value. Both "alignment"
1732 * and "boundary" must be a power of two.
1734 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1735 * then the memory attribute setting for the physical pages is configured
1736 * to the object's memory attribute setting. Otherwise, the memory
1737 * attribute setting for the physical pages is configured to "memattr",
1738 * overriding the object's memory attribute setting. However, if the
1739 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1740 * memory attribute setting for the physical pages cannot be configured
1741 * to VM_MEMATTR_DEFAULT.
1743 * The specified object may not contain fictitious pages.
1745 * The caller must always specify an allocation class.
1747 * allocation classes:
1748 * VM_ALLOC_NORMAL normal process request
1749 * VM_ALLOC_SYSTEM system *really* needs a page
1750 * VM_ALLOC_INTERRUPT interrupt time request
1752 * optional allocation flags:
1753 * VM_ALLOC_NOBUSY do not exclusive busy the page
1754 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1755 * VM_ALLOC_NOOBJ page is not associated with an object and
1756 * should not be exclusive busy
1757 * VM_ALLOC_SBUSY shared busy the allocated page
1758 * VM_ALLOC_WIRED wire the allocated page
1759 * VM_ALLOC_ZERO prefer a zeroed page
1761 * This routine may not sleep.
1764 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1765 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1766 vm_paddr_t boundary, vm_memattr_t memattr)
1768 vm_page_t m, m_ret, mpred;
1769 u_int busy_lock, flags, oflags;
1772 mpred = NULL; /* XXX: pacify gcc */
1773 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1774 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1775 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1776 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1777 ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
1779 if (object != NULL) {
1780 VM_OBJECT_ASSERT_WLOCKED(object);
1781 KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
1782 ("vm_page_alloc_contig: object %p has fictitious pages",
1785 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
1786 req_class = req & VM_ALLOC_CLASS_MASK;
1789 * The page daemon is allowed to dig deeper into the free page list.
1791 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1792 req_class = VM_ALLOC_SYSTEM;
1794 if (object != NULL) {
1795 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1796 KASSERT(mpred == NULL || mpred->pindex != pindex,
1797 ("vm_page_alloc_contig: pindex already allocated"));
1801 * Can we allocate the pages without the number of free pages falling
1802 * below the lower bound for the allocation class?
1804 mtx_lock(&vm_page_queue_free_mtx);
1805 if (vm_cnt.v_free_count >= npages + vm_cnt.v_free_reserved ||
1806 (req_class == VM_ALLOC_SYSTEM &&
1807 vm_cnt.v_free_count >= npages + vm_cnt.v_interrupt_free_min) ||
1808 (req_class == VM_ALLOC_INTERRUPT &&
1809 vm_cnt.v_free_count >= npages)) {
1811 * Can we allocate the pages from a reservation?
1813 #if VM_NRESERVLEVEL > 0
1815 if (object == NULL || (object->flags & OBJ_COLORED) == 0 ||
1816 (m_ret = vm_reserv_alloc_contig(object, pindex, npages,
1817 low, high, alignment, boundary, mpred)) == NULL)
1820 * If not, allocate them from the free page queues.
1822 m_ret = vm_phys_alloc_contig(npages, low, high,
1823 alignment, boundary);
1825 mtx_unlock(&vm_page_queue_free_mtx);
1826 atomic_add_int(&vm_pageout_deficit, npages);
1827 pagedaemon_wakeup();
1831 vm_phys_freecnt_adj(m_ret, -npages);
1833 #if VM_NRESERVLEVEL > 0
1834 if (vm_reserv_reclaim_contig(npages, low, high, alignment,
1839 mtx_unlock(&vm_page_queue_free_mtx);
1842 for (m = m_ret; m < &m_ret[npages]; m++)
1843 vm_page_alloc_check(m);
1846 * Initialize the pages. Only the PG_ZERO flag is inherited.
1849 if ((req & VM_ALLOC_ZERO) != 0)
1851 if ((req & VM_ALLOC_NODUMP) != 0)
1853 oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1855 busy_lock = VPB_UNBUSIED;
1856 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1857 busy_lock = VPB_SINGLE_EXCLUSIVER;
1858 if ((req & VM_ALLOC_SBUSY) != 0)
1859 busy_lock = VPB_SHARERS_WORD(1);
1860 if ((req & VM_ALLOC_WIRED) != 0)
1861 atomic_add_int(&vm_cnt.v_wire_count, npages);
1862 if (object != NULL) {
1863 if (object->memattr != VM_MEMATTR_DEFAULT &&
1864 memattr == VM_MEMATTR_DEFAULT)
1865 memattr = object->memattr;
1867 for (m = m_ret; m < &m_ret[npages]; m++) {
1869 m->flags = (m->flags | PG_NODUMP) & flags;
1870 m->busy_lock = busy_lock;
1871 if ((req & VM_ALLOC_WIRED) != 0)
1875 if (object != NULL) {
1876 if (vm_page_insert_after(m, object, pindex, mpred)) {
1877 pagedaemon_wakeup();
1878 if ((req & VM_ALLOC_WIRED) != 0)
1879 atomic_subtract_int(
1880 &vm_cnt.v_wire_count, npages);
1881 KASSERT(m->object == NULL,
1882 ("page %p has object", m));
1884 for (m = m_ret; m < &m_ret[npages]; m++) {
1886 (req & VM_ALLOC_WIRED) != 0)
1888 m->oflags = VPO_UNMANAGED;
1889 m->busy_lock = VPB_UNBUSIED;
1890 /* Don't change PG_ZERO. */
1891 vm_page_free_toq(m);
1898 if (memattr != VM_MEMATTR_DEFAULT)
1899 pmap_page_set_memattr(m, memattr);
1902 if (vm_paging_needed())
1903 pagedaemon_wakeup();
1908 * Check a page that has been freshly dequeued from a freelist.
1911 vm_page_alloc_check(vm_page_t m)
1914 KASSERT(m->object == NULL, ("page %p has object", m));
1915 KASSERT(m->queue == PQ_NONE,
1916 ("page %p has unexpected queue %d", m, m->queue));
1917 KASSERT(m->wire_count == 0, ("page %p is wired", m));
1918 KASSERT(m->hold_count == 0, ("page %p is held", m));
1919 KASSERT(!vm_page_busied(m), ("page %p is busy", m));
1920 KASSERT(m->dirty == 0, ("page %p is dirty", m));
1921 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1922 ("page %p has unexpected memattr %d",
1923 m, pmap_page_get_memattr(m)));
1924 KASSERT(m->valid == 0, ("free page %p is valid", m));
1928 * vm_page_alloc_freelist:
1930 * Allocate a physical page from the specified free page list.
1932 * The caller must always specify an allocation class.
1934 * allocation classes:
1935 * VM_ALLOC_NORMAL normal process request
1936 * VM_ALLOC_SYSTEM system *really* needs a page
1937 * VM_ALLOC_INTERRUPT interrupt time request
1939 * optional allocation flags:
1940 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1941 * intends to allocate
1942 * VM_ALLOC_WIRED wire the allocated page
1943 * VM_ALLOC_ZERO prefer a zeroed page
1945 * This routine may not sleep.
1948 vm_page_alloc_freelist(int flind, int req)
1954 req_class = req & VM_ALLOC_CLASS_MASK;
1957 * The page daemon is allowed to dig deeper into the free page list.
1959 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1960 req_class = VM_ALLOC_SYSTEM;
1963 * Do not allocate reserved pages unless the req has asked for it.
1965 mtx_lock(&vm_page_queue_free_mtx);
1966 if (vm_cnt.v_free_count > vm_cnt.v_free_reserved ||
1967 (req_class == VM_ALLOC_SYSTEM &&
1968 vm_cnt.v_free_count > vm_cnt.v_interrupt_free_min) ||
1969 (req_class == VM_ALLOC_INTERRUPT &&
1970 vm_cnt.v_free_count > 0))
1971 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0);
1973 mtx_unlock(&vm_page_queue_free_mtx);
1974 atomic_add_int(&vm_pageout_deficit,
1975 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1976 pagedaemon_wakeup();
1980 mtx_unlock(&vm_page_queue_free_mtx);
1983 vm_phys_freecnt_adj(m, -1);
1984 mtx_unlock(&vm_page_queue_free_mtx);
1985 vm_page_alloc_check(m);
1988 * Initialize the page. Only the PG_ZERO flag is inherited.
1992 if ((req & VM_ALLOC_ZERO) != 0)
1995 if ((req & VM_ALLOC_WIRED) != 0) {
1997 * The page lock is not required for wiring a page that does
1998 * not belong to an object.
2000 atomic_add_int(&vm_cnt.v_wire_count, 1);
2003 /* Unmanaged pages don't use "act_count". */
2004 m->oflags = VPO_UNMANAGED;
2005 if (vm_paging_needed())
2006 pagedaemon_wakeup();
2010 #define VPSC_ANY 0 /* No restrictions. */
2011 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */
2012 #define VPSC_NOSUPER 2 /* Skip superpages. */
2015 * vm_page_scan_contig:
2017 * Scan vm_page_array[] between the specified entries "m_start" and
2018 * "m_end" for a run of contiguous physical pages that satisfy the
2019 * specified conditions, and return the lowest page in the run. The
2020 * specified "alignment" determines the alignment of the lowest physical
2021 * page in the run. If the specified "boundary" is non-zero, then the
2022 * run of physical pages cannot span a physical address that is a
2023 * multiple of "boundary".
2025 * "m_end" is never dereferenced, so it need not point to a vm_page
2026 * structure within vm_page_array[].
2028 * "npages" must be greater than zero. "m_start" and "m_end" must not
2029 * span a hole (or discontiguity) in the physical address space. Both
2030 * "alignment" and "boundary" must be a power of two.
2033 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2034 u_long alignment, vm_paddr_t boundary, int options)
2040 #if VM_NRESERVLEVEL > 0
2043 int m_inc, order, run_ext, run_len;
2045 KASSERT(npages > 0, ("npages is 0"));
2046 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2047 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2051 for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2052 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2053 ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2056 * If the current page would be the start of a run, check its
2057 * physical address against the end, alignment, and boundary
2058 * conditions. If it doesn't satisfy these conditions, either
2059 * terminate the scan or advance to the next page that
2060 * satisfies the failed condition.
2063 KASSERT(m_run == NULL, ("m_run != NULL"));
2064 if (m + npages > m_end)
2066 pa = VM_PAGE_TO_PHYS(m);
2067 if ((pa & (alignment - 1)) != 0) {
2068 m_inc = atop(roundup2(pa, alignment) - pa);
2071 if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2073 m_inc = atop(roundup2(pa, boundary) - pa);
2077 KASSERT(m_run != NULL, ("m_run == NULL"));
2079 vm_page_change_lock(m, &m_mtx);
2082 if (m->wire_count != 0 || m->hold_count != 0)
2084 #if VM_NRESERVLEVEL > 0
2085 else if ((level = vm_reserv_level(m)) >= 0 &&
2086 (options & VPSC_NORESERV) != 0) {
2088 /* Advance to the end of the reservation. */
2089 pa = VM_PAGE_TO_PHYS(m);
2090 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2094 else if ((object = m->object) != NULL) {
2096 * The page is considered eligible for relocation if
2097 * and only if it could be laundered or reclaimed by
2100 if (!VM_OBJECT_TRYRLOCK(object)) {
2102 VM_OBJECT_RLOCK(object);
2104 if (m->object != object) {
2106 * The page may have been freed.
2108 VM_OBJECT_RUNLOCK(object);
2110 } else if (m->wire_count != 0 ||
2111 m->hold_count != 0) {
2116 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2117 ("page %p is PG_UNHOLDFREE", m));
2118 /* Don't care: PG_NODUMP, PG_ZERO. */
2119 if (object->type != OBJT_DEFAULT &&
2120 object->type != OBJT_SWAP &&
2121 object->type != OBJT_VNODE) {
2123 #if VM_NRESERVLEVEL > 0
2124 } else if ((options & VPSC_NOSUPER) != 0 &&
2125 (level = vm_reserv_level_iffullpop(m)) >= 0) {
2127 /* Advance to the end of the superpage. */
2128 pa = VM_PAGE_TO_PHYS(m);
2129 m_inc = atop(roundup2(pa + 1,
2130 vm_reserv_size(level)) - pa);
2132 } else if (object->memattr == VM_MEMATTR_DEFAULT &&
2133 m->queue != PQ_NONE && !vm_page_busied(m)) {
2135 * The page is allocated but eligible for
2136 * relocation. Extend the current run by one
2139 KASSERT(pmap_page_get_memattr(m) ==
2141 ("page %p has an unexpected memattr", m));
2142 KASSERT((m->oflags & (VPO_SWAPINPROG |
2143 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2144 ("page %p has unexpected oflags", m));
2145 /* Don't care: VPO_NOSYNC. */
2150 VM_OBJECT_RUNLOCK(object);
2151 #if VM_NRESERVLEVEL > 0
2152 } else if (level >= 0) {
2154 * The page is reserved but not yet allocated. In
2155 * other words, it is still free. Extend the current
2160 } else if ((order = m->order) < VM_NFREEORDER) {
2162 * The page is enqueued in the physical memory
2163 * allocator's free page queues. Moreover, it is the
2164 * first page in a power-of-two-sized run of
2165 * contiguous free pages. Add these pages to the end
2166 * of the current run, and jump ahead.
2168 run_ext = 1 << order;
2172 * Skip the page for one of the following reasons: (1)
2173 * It is enqueued in the physical memory allocator's
2174 * free page queues. However, it is not the first
2175 * page in a run of contiguous free pages. (This case
2176 * rarely occurs because the scan is performed in
2177 * ascending order.) (2) It is not reserved, and it is
2178 * transitioning from free to allocated. (Conversely,
2179 * the transition from allocated to free for managed
2180 * pages is blocked by the page lock.) (3) It is
2181 * allocated but not contained by an object and not
2182 * wired, e.g., allocated by Xen's balloon driver.
2188 * Extend or reset the current run of pages.
2203 if (run_len >= npages)
2209 * vm_page_reclaim_run:
2211 * Try to relocate each of the allocated virtual pages within the
2212 * specified run of physical pages to a new physical address. Free the
2213 * physical pages underlying the relocated virtual pages. A virtual page
2214 * is relocatable if and only if it could be laundered or reclaimed by
2215 * the page daemon. Whenever possible, a virtual page is relocated to a
2216 * physical address above "high".
2218 * Returns 0 if every physical page within the run was already free or
2219 * just freed by a successful relocation. Otherwise, returns a non-zero
2220 * value indicating why the last attempt to relocate a virtual page was
2223 * "req_class" must be an allocation class.
2226 vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run,
2230 struct spglist free;
2233 vm_page_t m, m_end, m_new;
2234 int error, order, req;
2236 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2237 ("req_class is not an allocation class"));
2241 m_end = m_run + npages;
2243 for (; error == 0 && m < m_end; m++) {
2244 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2245 ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2248 * Avoid releasing and reacquiring the same page lock.
2250 vm_page_change_lock(m, &m_mtx);
2252 if (m->wire_count != 0 || m->hold_count != 0)
2254 else if ((object = m->object) != NULL) {
2256 * The page is relocated if and only if it could be
2257 * laundered or reclaimed by the page daemon.
2259 if (!VM_OBJECT_TRYWLOCK(object)) {
2261 VM_OBJECT_WLOCK(object);
2263 if (m->object != object) {
2265 * The page may have been freed.
2267 VM_OBJECT_WUNLOCK(object);
2269 } else if (m->wire_count != 0 ||
2270 m->hold_count != 0) {
2275 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2276 ("page %p is PG_UNHOLDFREE", m));
2277 /* Don't care: PG_NODUMP, PG_ZERO. */
2278 if (object->type != OBJT_DEFAULT &&
2279 object->type != OBJT_SWAP &&
2280 object->type != OBJT_VNODE)
2282 else if (object->memattr != VM_MEMATTR_DEFAULT)
2284 else if (m->queue != PQ_NONE && !vm_page_busied(m)) {
2285 KASSERT(pmap_page_get_memattr(m) ==
2287 ("page %p has an unexpected memattr", m));
2288 KASSERT((m->oflags & (VPO_SWAPINPROG |
2289 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2290 ("page %p has unexpected oflags", m));
2291 /* Don't care: VPO_NOSYNC. */
2292 if (m->valid != 0) {
2294 * First, try to allocate a new page
2295 * that is above "high". Failing
2296 * that, try to allocate a new page
2297 * that is below "m_run". Allocate
2298 * the new page between the end of
2299 * "m_run" and "high" only as a last
2302 req = req_class | VM_ALLOC_NOOBJ;
2303 if ((m->flags & PG_NODUMP) != 0)
2304 req |= VM_ALLOC_NODUMP;
2305 if (trunc_page(high) !=
2306 ~(vm_paddr_t)PAGE_MASK) {
2307 m_new = vm_page_alloc_contig(
2312 VM_MEMATTR_DEFAULT);
2315 if (m_new == NULL) {
2316 pa = VM_PAGE_TO_PHYS(m_run);
2317 m_new = vm_page_alloc_contig(
2319 0, pa - 1, PAGE_SIZE, 0,
2320 VM_MEMATTR_DEFAULT);
2322 if (m_new == NULL) {
2324 m_new = vm_page_alloc_contig(
2326 pa, high, PAGE_SIZE, 0,
2327 VM_MEMATTR_DEFAULT);
2329 if (m_new == NULL) {
2333 KASSERT(m_new->wire_count == 0,
2334 ("page %p is wired", m));
2337 * Replace "m" with the new page. For
2338 * vm_page_replace(), "m" must be busy
2339 * and dequeued. Finally, change "m"
2340 * as if vm_page_free() was called.
2342 if (object->ref_count != 0)
2344 m_new->aflags = m->aflags;
2345 KASSERT(m_new->oflags == VPO_UNMANAGED,
2346 ("page %p is managed", m));
2347 m_new->oflags = m->oflags & VPO_NOSYNC;
2348 pmap_copy_page(m, m_new);
2349 m_new->valid = m->valid;
2350 m_new->dirty = m->dirty;
2351 m->flags &= ~PG_ZERO;
2354 vm_page_replace_checked(m_new, object,
2360 * The new page must be deactivated
2361 * before the object is unlocked.
2363 vm_page_change_lock(m_new, &m_mtx);
2364 vm_page_deactivate(m_new);
2366 m->flags &= ~PG_ZERO;
2369 KASSERT(m->dirty == 0,
2370 ("page %p is dirty", m));
2372 SLIST_INSERT_HEAD(&free, m, plinks.s.ss);
2376 VM_OBJECT_WUNLOCK(object);
2378 mtx_lock(&vm_page_queue_free_mtx);
2380 if (order < VM_NFREEORDER) {
2382 * The page is enqueued in the physical memory
2383 * allocator's free page queues. Moreover, it
2384 * is the first page in a power-of-two-sized
2385 * run of contiguous free pages. Jump ahead
2386 * to the last page within that run, and
2387 * continue from there.
2389 m += (1 << order) - 1;
2391 #if VM_NRESERVLEVEL > 0
2392 else if (vm_reserv_is_page_free(m))
2395 mtx_unlock(&vm_page_queue_free_mtx);
2396 if (order == VM_NFREEORDER)
2402 if ((m = SLIST_FIRST(&free)) != NULL) {
2403 mtx_lock(&vm_page_queue_free_mtx);
2405 SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2406 vm_page_free_phys(m);
2407 } while ((m = SLIST_FIRST(&free)) != NULL);
2408 vm_page_free_wakeup();
2409 mtx_unlock(&vm_page_queue_free_mtx);
2416 CTASSERT(powerof2(NRUNS));
2418 #define RUN_INDEX(count) ((count) & (NRUNS - 1))
2420 #define MIN_RECLAIM 8
2423 * vm_page_reclaim_contig:
2425 * Reclaim allocated, contiguous physical memory satisfying the specified
2426 * conditions by relocating the virtual pages using that physical memory.
2427 * Returns true if reclamation is successful and false otherwise. Since
2428 * relocation requires the allocation of physical pages, reclamation may
2429 * fail due to a shortage of free pages. When reclamation fails, callers
2430 * are expected to perform VM_WAIT before retrying a failed allocation
2431 * operation, e.g., vm_page_alloc_contig().
2433 * The caller must always specify an allocation class through "req".
2435 * allocation classes:
2436 * VM_ALLOC_NORMAL normal process request
2437 * VM_ALLOC_SYSTEM system *really* needs a page
2438 * VM_ALLOC_INTERRUPT interrupt time request
2440 * The optional allocation flags are ignored.
2442 * "npages" must be greater than zero. Both "alignment" and "boundary"
2443 * must be a power of two.
2446 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
2447 u_long alignment, vm_paddr_t boundary)
2449 vm_paddr_t curr_low;
2450 vm_page_t m_run, m_runs[NRUNS];
2451 u_long count, reclaimed;
2452 int error, i, options, req_class;
2454 KASSERT(npages > 0, ("npages is 0"));
2455 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2456 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2457 req_class = req & VM_ALLOC_CLASS_MASK;
2460 * The page daemon is allowed to dig deeper into the free page list.
2462 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2463 req_class = VM_ALLOC_SYSTEM;
2466 * Return if the number of free pages cannot satisfy the requested
2469 count = vm_cnt.v_free_count;
2470 if (count < npages + vm_cnt.v_free_reserved || (count < npages +
2471 vm_cnt.v_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2472 (count < npages && req_class == VM_ALLOC_INTERRUPT))
2476 * Scan up to three times, relaxing the restrictions ("options") on
2477 * the reclamation of reservations and superpages each time.
2479 for (options = VPSC_NORESERV;;) {
2481 * Find the highest runs that satisfy the given constraints
2482 * and restrictions, and record them in "m_runs".
2487 m_run = vm_phys_scan_contig(npages, curr_low, high,
2488 alignment, boundary, options);
2491 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2492 m_runs[RUN_INDEX(count)] = m_run;
2497 * Reclaim the highest runs in LIFO (descending) order until
2498 * the number of reclaimed pages, "reclaimed", is at least
2499 * MIN_RECLAIM. Reset "reclaimed" each time because each
2500 * reclamation is idempotent, and runs will (likely) recur
2501 * from one scan to the next as restrictions are relaxed.
2504 for (i = 0; count > 0 && i < NRUNS; i++) {
2506 m_run = m_runs[RUN_INDEX(count)];
2507 error = vm_page_reclaim_run(req_class, npages, m_run,
2510 reclaimed += npages;
2511 if (reclaimed >= MIN_RECLAIM)
2517 * Either relax the restrictions on the next scan or return if
2518 * the last scan had no restrictions.
2520 if (options == VPSC_NORESERV)
2521 options = VPSC_NOSUPER;
2522 else if (options == VPSC_NOSUPER)
2524 else if (options == VPSC_ANY)
2525 return (reclaimed != 0);
2530 * vm_wait: (also see VM_WAIT macro)
2532 * Sleep until free pages are available for allocation.
2533 * - Called in various places before memory allocations.
2539 mtx_lock(&vm_page_queue_free_mtx);
2540 if (curproc == pageproc) {
2541 vm_pageout_pages_needed = 1;
2542 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx,
2543 PDROP | PSWP, "VMWait", 0);
2545 if (__predict_false(pageproc == NULL))
2546 panic("vm_wait in early boot");
2547 if (!vm_pageout_wanted) {
2548 vm_pageout_wanted = true;
2549 wakeup(&vm_pageout_wanted);
2551 vm_pages_needed = true;
2552 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM,
2558 * vm_waitpfault: (also see VM_WAITPFAULT macro)
2560 * Sleep until free pages are available for allocation.
2561 * - Called only in vm_fault so that processes page faulting
2562 * can be easily tracked.
2563 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
2564 * processes will be able to grab memory first. Do not change
2565 * this balance without careful testing first.
2571 mtx_lock(&vm_page_queue_free_mtx);
2572 if (!vm_pageout_wanted) {
2573 vm_pageout_wanted = true;
2574 wakeup(&vm_pageout_wanted);
2576 vm_pages_needed = true;
2577 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER,
2581 struct vm_pagequeue *
2582 vm_page_pagequeue(vm_page_t m)
2585 if (vm_page_in_laundry(m))
2586 return (&vm_dom[0].vmd_pagequeues[m->queue]);
2588 return (&vm_phys_domain(m)->vmd_pagequeues[m->queue]);
2594 * Remove the given page from its current page queue.
2596 * The page must be locked.
2599 vm_page_dequeue(vm_page_t m)
2601 struct vm_pagequeue *pq;
2603 vm_page_assert_locked(m);
2604 KASSERT(m->queue < PQ_COUNT, ("vm_page_dequeue: page %p is not queued",
2606 pq = vm_page_pagequeue(m);
2607 vm_pagequeue_lock(pq);
2609 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2610 vm_pagequeue_cnt_dec(pq);
2611 vm_pagequeue_unlock(pq);
2615 * vm_page_dequeue_locked:
2617 * Remove the given page from its current page queue.
2619 * The page and page queue must be locked.
2622 vm_page_dequeue_locked(vm_page_t m)
2624 struct vm_pagequeue *pq;
2626 vm_page_lock_assert(m, MA_OWNED);
2627 pq = vm_page_pagequeue(m);
2628 vm_pagequeue_assert_locked(pq);
2630 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2631 vm_pagequeue_cnt_dec(pq);
2637 * Add the given page to the specified page queue.
2639 * The page must be locked.
2642 vm_page_enqueue(uint8_t queue, vm_page_t m)
2644 struct vm_pagequeue *pq;
2646 vm_page_lock_assert(m, MA_OWNED);
2647 KASSERT(queue < PQ_COUNT,
2648 ("vm_page_enqueue: invalid queue %u request for page %p",
2650 if (queue == PQ_LAUNDRY || queue == PQ_UNSWAPPABLE)
2651 pq = &vm_dom[0].vmd_pagequeues[queue];
2653 pq = &vm_phys_domain(m)->vmd_pagequeues[queue];
2654 vm_pagequeue_lock(pq);
2656 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2657 vm_pagequeue_cnt_inc(pq);
2658 vm_pagequeue_unlock(pq);
2664 * Move the given page to the tail of its current page queue.
2666 * The page must be locked.
2669 vm_page_requeue(vm_page_t m)
2671 struct vm_pagequeue *pq;
2673 vm_page_lock_assert(m, MA_OWNED);
2674 KASSERT(m->queue != PQ_NONE,
2675 ("vm_page_requeue: page %p is not queued", m));
2676 pq = vm_page_pagequeue(m);
2677 vm_pagequeue_lock(pq);
2678 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2679 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2680 vm_pagequeue_unlock(pq);
2684 * vm_page_requeue_locked:
2686 * Move the given page to the tail of its current page queue.
2688 * The page queue must be locked.
2691 vm_page_requeue_locked(vm_page_t m)
2693 struct vm_pagequeue *pq;
2695 KASSERT(m->queue != PQ_NONE,
2696 ("vm_page_requeue_locked: page %p is not queued", m));
2697 pq = vm_page_pagequeue(m);
2698 vm_pagequeue_assert_locked(pq);
2699 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2700 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2706 * Put the specified page on the active list (if appropriate).
2707 * Ensure that act_count is at least ACT_INIT but do not otherwise
2710 * The page must be locked.
2713 vm_page_activate(vm_page_t m)
2717 vm_page_lock_assert(m, MA_OWNED);
2718 if ((queue = m->queue) != PQ_ACTIVE) {
2719 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2720 if (m->act_count < ACT_INIT)
2721 m->act_count = ACT_INIT;
2722 if (queue != PQ_NONE)
2724 vm_page_enqueue(PQ_ACTIVE, m);
2726 KASSERT(queue == PQ_NONE,
2727 ("vm_page_activate: wired page %p is queued", m));
2729 if (m->act_count < ACT_INIT)
2730 m->act_count = ACT_INIT;
2735 * vm_page_free_wakeup:
2737 * Helper routine for vm_page_free_toq(). This routine is called
2738 * when a page is added to the free queues.
2740 * The page queues must be locked.
2743 vm_page_free_wakeup(void)
2746 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2748 * if pageout daemon needs pages, then tell it that there are
2751 if (vm_pageout_pages_needed &&
2752 vm_cnt.v_free_count >= vm_cnt.v_pageout_free_min) {
2753 wakeup(&vm_pageout_pages_needed);
2754 vm_pageout_pages_needed = 0;
2757 * wakeup processes that are waiting on memory if we hit a
2758 * high water mark. And wakeup scheduler process if we have
2759 * lots of memory. this process will swapin processes.
2761 if (vm_pages_needed && !vm_page_count_min()) {
2762 vm_pages_needed = false;
2763 wakeup(&vm_cnt.v_free_count);
2768 * vm_page_free_prep:
2770 * Prepares the given page to be put on the free list,
2771 * disassociating it from any VM object. The caller may return
2772 * the page to the free list only if this function returns true.
2774 * The object must be locked. The page must be locked if it is
2775 * managed. For a queued managed page, the pagequeue_locked
2776 * argument specifies whether the page queue is already locked.
2779 vm_page_free_prep(vm_page_t m, bool pagequeue_locked)
2782 if ((m->oflags & VPO_UNMANAGED) == 0) {
2783 vm_page_lock_assert(m, MA_OWNED);
2784 KASSERT(!pmap_page_is_mapped(m),
2785 ("vm_page_free_toq: freeing mapped page %p", m));
2787 KASSERT(m->queue == PQ_NONE,
2788 ("vm_page_free_toq: unmanaged page %p is queued", m));
2789 VM_CNT_INC(v_tfree);
2791 if (vm_page_sbusied(m))
2792 panic("vm_page_free: freeing busy page %p", m);
2795 * Unqueue, then remove page. Note that we cannot destroy
2796 * the page here because we do not want to call the pager's
2797 * callback routine until after we've put the page on the
2798 * appropriate free queue.
2800 if (m->queue != PQ_NONE) {
2801 if (pagequeue_locked)
2802 vm_page_dequeue_locked(m);
2809 * If fictitious remove object association and
2810 * return, otherwise delay object association removal.
2812 if ((m->flags & PG_FICTITIOUS) != 0)
2818 if (m->wire_count != 0)
2819 panic("vm_page_free: freeing wired page %p", m);
2820 if (m->hold_count != 0) {
2821 m->flags &= ~PG_ZERO;
2822 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2823 ("vm_page_free: freeing PG_UNHOLDFREE page %p", m));
2824 m->flags |= PG_UNHOLDFREE;
2829 * Restore the default memory attribute to the page.
2831 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2832 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2838 * Insert the page into the physical memory allocator's free page
2839 * queues. This is the last step to free a page.
2842 vm_page_free_phys(vm_page_t m)
2845 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2847 vm_phys_freecnt_adj(m, 1);
2848 #if VM_NRESERVLEVEL > 0
2849 if (!vm_reserv_free_page(m))
2851 vm_phys_free_pages(m, 0);
2855 vm_page_free_phys_pglist(struct pglist *tq)
2859 mtx_lock(&vm_page_queue_free_mtx);
2860 TAILQ_FOREACH(m, tq, listq)
2861 vm_page_free_phys(m);
2862 vm_page_free_wakeup();
2863 mtx_unlock(&vm_page_queue_free_mtx);
2869 * Returns the given page to the free list, disassociating it
2870 * from any VM object.
2872 * The object must be locked. The page must be locked if it is
2876 vm_page_free_toq(vm_page_t m)
2879 if (!vm_page_free_prep(m, false))
2881 mtx_lock(&vm_page_queue_free_mtx);
2882 vm_page_free_phys(m);
2883 vm_page_free_wakeup();
2884 mtx_unlock(&vm_page_queue_free_mtx);
2890 * Mark this page as wired down by yet
2891 * another map, removing it from paging queues
2894 * If the page is fictitious, then its wire count must remain one.
2896 * The page must be locked.
2899 vm_page_wire(vm_page_t m)
2903 * Only bump the wire statistics if the page is not already wired,
2904 * and only unqueue the page if it is on some queue (if it is unmanaged
2905 * it is already off the queues).
2907 vm_page_lock_assert(m, MA_OWNED);
2908 if ((m->flags & PG_FICTITIOUS) != 0) {
2909 KASSERT(m->wire_count == 1,
2910 ("vm_page_wire: fictitious page %p's wire count isn't one",
2914 if (m->wire_count == 0) {
2915 KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
2916 m->queue == PQ_NONE,
2917 ("vm_page_wire: unmanaged page %p is queued", m));
2919 atomic_add_int(&vm_cnt.v_wire_count, 1);
2922 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
2928 * Release one wiring of the specified page, potentially allowing it to be
2929 * paged out. Returns TRUE if the number of wirings transitions to zero and
2932 * Only managed pages belonging to an object can be paged out. If the number
2933 * of wirings transitions to zero and the page is eligible for page out, then
2934 * the page is added to the specified paging queue (unless PQ_NONE is
2937 * If a page is fictitious, then its wire count must always be one.
2939 * A managed page must be locked.
2942 vm_page_unwire(vm_page_t m, uint8_t queue)
2945 KASSERT(queue < PQ_COUNT || queue == PQ_NONE,
2946 ("vm_page_unwire: invalid queue %u request for page %p",
2948 if ((m->oflags & VPO_UNMANAGED) == 0)
2949 vm_page_assert_locked(m);
2950 if ((m->flags & PG_FICTITIOUS) != 0) {
2951 KASSERT(m->wire_count == 1,
2952 ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
2955 if (m->wire_count > 0) {
2957 if (m->wire_count == 0) {
2958 atomic_subtract_int(&vm_cnt.v_wire_count, 1);
2959 if ((m->oflags & VPO_UNMANAGED) == 0 &&
2960 m->object != NULL && queue != PQ_NONE)
2961 vm_page_enqueue(queue, m);
2966 panic("vm_page_unwire: page %p's wire count is zero", m);
2970 * Move the specified page to the inactive queue.
2972 * Normally, "noreuse" is FALSE, resulting in LRU ordering of the inactive
2973 * queue. However, setting "noreuse" to TRUE will accelerate the specified
2974 * page's reclamation, but it will not unmap the page from any address space.
2975 * This is implemented by inserting the page near the head of the inactive
2976 * queue, using a marker page to guide FIFO insertion ordering.
2978 * The page must be locked.
2981 _vm_page_deactivate(vm_page_t m, boolean_t noreuse)
2983 struct vm_pagequeue *pq;
2986 vm_page_assert_locked(m);
2989 * Ignore if the page is already inactive, unless it is unlikely to be
2992 if ((queue = m->queue) == PQ_INACTIVE && !noreuse)
2994 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2995 pq = &vm_phys_domain(m)->vmd_pagequeues[PQ_INACTIVE];
2996 /* Avoid multiple acquisitions of the inactive queue lock. */
2997 if (queue == PQ_INACTIVE) {
2998 vm_pagequeue_lock(pq);
2999 vm_page_dequeue_locked(m);
3001 if (queue != PQ_NONE)
3003 vm_pagequeue_lock(pq);
3005 m->queue = PQ_INACTIVE;
3007 TAILQ_INSERT_BEFORE(&vm_phys_domain(m)->vmd_inacthead,
3010 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3011 vm_pagequeue_cnt_inc(pq);
3012 vm_pagequeue_unlock(pq);
3017 * Move the specified page to the inactive queue.
3019 * The page must be locked.
3022 vm_page_deactivate(vm_page_t m)
3025 _vm_page_deactivate(m, FALSE);
3029 * Move the specified page to the inactive queue with the expectation
3030 * that it is unlikely to be reused.
3032 * The page must be locked.
3035 vm_page_deactivate_noreuse(vm_page_t m)
3038 _vm_page_deactivate(m, TRUE);
3044 * Put a page in the laundry.
3047 vm_page_launder(vm_page_t m)
3051 vm_page_assert_locked(m);
3052 if ((queue = m->queue) != PQ_LAUNDRY) {
3053 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
3054 if (queue != PQ_NONE)
3056 vm_page_enqueue(PQ_LAUNDRY, m);
3058 KASSERT(queue == PQ_NONE,
3059 ("wired page %p is queued", m));
3064 * vm_page_unswappable
3066 * Put a page in the PQ_UNSWAPPABLE holding queue.
3069 vm_page_unswappable(vm_page_t m)
3072 vm_page_assert_locked(m);
3073 KASSERT(m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0,
3074 ("page %p already unswappable", m));
3075 if (m->queue != PQ_NONE)
3077 vm_page_enqueue(PQ_UNSWAPPABLE, m);
3081 * Attempt to free the page. If it cannot be freed, do nothing. Returns true
3082 * if the page is freed and false otherwise.
3084 * The page must be managed. The page and its containing object must be
3088 vm_page_try_to_free(vm_page_t m)
3091 vm_page_assert_locked(m);
3092 VM_OBJECT_ASSERT_WLOCKED(m->object);
3093 KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("page %p is unmanaged", m));
3094 if (m->dirty != 0 || m->hold_count != 0 || m->wire_count != 0 ||
3097 if (m->object->ref_count != 0) {
3109 * Apply the specified advice to the given page.
3111 * The object and page must be locked.
3114 vm_page_advise(vm_page_t m, int advice)
3117 vm_page_assert_locked(m);
3118 VM_OBJECT_ASSERT_WLOCKED(m->object);
3119 if (advice == MADV_FREE)
3121 * Mark the page clean. This will allow the page to be freed
3122 * without first paging it out. MADV_FREE pages are often
3123 * quickly reused by malloc(3), so we do not do anything that
3124 * would result in a page fault on a later access.
3127 else if (advice != MADV_DONTNEED) {
3128 if (advice == MADV_WILLNEED)
3129 vm_page_activate(m);
3134 * Clear any references to the page. Otherwise, the page daemon will
3135 * immediately reactivate the page.
3137 vm_page_aflag_clear(m, PGA_REFERENCED);
3139 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
3143 * Place clean pages near the head of the inactive queue rather than
3144 * the tail, thus defeating the queue's LRU operation and ensuring that
3145 * the page will be reused quickly. Dirty pages not already in the
3146 * laundry are moved there.
3149 vm_page_deactivate_noreuse(m);
3155 * Grab a page, waiting until we are waken up due to the page
3156 * changing state. We keep on waiting, if the page continues
3157 * to be in the object. If the page doesn't exist, first allocate it
3158 * and then conditionally zero it.
3160 * This routine may sleep.
3162 * The object must be locked on entry. The lock will, however, be released
3163 * and reacquired if the routine sleeps.
3166 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
3171 VM_OBJECT_ASSERT_WLOCKED(object);
3172 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3173 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3174 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
3176 if ((m = vm_page_lookup(object, pindex)) != NULL) {
3177 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3178 vm_page_xbusied(m) : vm_page_busied(m);
3180 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3183 * Reference the page before unlocking and
3184 * sleeping so that the page daemon is less
3185 * likely to reclaim it.
3187 vm_page_aflag_set(m, PGA_REFERENCED);
3189 VM_OBJECT_WUNLOCK(object);
3190 vm_page_busy_sleep(m, "pgrbwt", (allocflags &
3191 VM_ALLOC_IGN_SBUSY) != 0);
3192 VM_OBJECT_WLOCK(object);
3195 if ((allocflags & VM_ALLOC_WIRED) != 0) {
3201 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
3203 if ((allocflags & VM_ALLOC_SBUSY) != 0)
3208 m = vm_page_alloc(object, pindex, allocflags);
3210 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3212 VM_OBJECT_WUNLOCK(object);
3214 VM_OBJECT_WLOCK(object);
3217 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
3223 * Return the specified range of pages from the given object. For each
3224 * page offset within the range, if a page already exists within the object
3225 * at that offset and it is busy, then wait for it to change state. If,
3226 * instead, the page doesn't exist, then allocate it.
3228 * The caller must always specify an allocation class.
3230 * allocation classes:
3231 * VM_ALLOC_NORMAL normal process request
3232 * VM_ALLOC_SYSTEM system *really* needs the pages
3234 * The caller must always specify that the pages are to be busied and/or
3237 * optional allocation flags:
3238 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages
3239 * VM_ALLOC_NOBUSY do not exclusive busy the page
3240 * VM_ALLOC_NOWAIT do not sleep
3241 * VM_ALLOC_SBUSY set page to sbusy state
3242 * VM_ALLOC_WIRED wire the pages
3243 * VM_ALLOC_ZERO zero and validate any invalid pages
3245 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it
3246 * may return a partial prefix of the requested range.
3249 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
3250 vm_page_t *ma, int count)
3256 VM_OBJECT_ASSERT_WLOCKED(object);
3257 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
3258 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
3259 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
3260 (allocflags & VM_ALLOC_WIRED) != 0,
3261 ("vm_page_grab_pages: the pages must be busied or wired"));
3262 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3263 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3264 ("vm_page_grab_pages: VM_ALLOC_SBUSY/IGN_SBUSY mismatch"));
3269 m = vm_radix_lookup_le(&object->rtree, pindex + i);
3270 if (m == NULL || m->pindex != pindex + i) {
3274 mpred = TAILQ_PREV(m, pglist, listq);
3275 for (; i < count; i++) {
3277 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3278 vm_page_xbusied(m) : vm_page_busied(m);
3280 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3283 * Reference the page before unlocking and
3284 * sleeping so that the page daemon is less
3285 * likely to reclaim it.
3287 vm_page_aflag_set(m, PGA_REFERENCED);
3289 VM_OBJECT_WUNLOCK(object);
3290 vm_page_busy_sleep(m, "grbmaw", (allocflags &
3291 VM_ALLOC_IGN_SBUSY) != 0);
3292 VM_OBJECT_WLOCK(object);
3295 if ((allocflags & VM_ALLOC_WIRED) != 0) {
3300 if ((allocflags & (VM_ALLOC_NOBUSY |
3301 VM_ALLOC_SBUSY)) == 0)
3303 if ((allocflags & VM_ALLOC_SBUSY) != 0)
3306 m = vm_page_alloc_after(object, pindex + i,
3307 (allocflags & ~VM_ALLOC_IGN_SBUSY) |
3308 VM_ALLOC_COUNT(count - i), mpred);
3310 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3312 VM_OBJECT_WUNLOCK(object);
3314 VM_OBJECT_WLOCK(object);
3318 if (m->valid == 0 && (allocflags & VM_ALLOC_ZERO) != 0) {
3319 if ((m->flags & PG_ZERO) == 0)
3321 m->valid = VM_PAGE_BITS_ALL;
3324 m = vm_page_next(m);
3330 * Mapping function for valid or dirty bits in a page.
3332 * Inputs are required to range within a page.
3335 vm_page_bits(int base, int size)
3341 base + size <= PAGE_SIZE,
3342 ("vm_page_bits: illegal base/size %d/%d", base, size)
3345 if (size == 0) /* handle degenerate case */
3348 first_bit = base >> DEV_BSHIFT;
3349 last_bit = (base + size - 1) >> DEV_BSHIFT;
3351 return (((vm_page_bits_t)2 << last_bit) -
3352 ((vm_page_bits_t)1 << first_bit));
3356 * vm_page_set_valid_range:
3358 * Sets portions of a page valid. The arguments are expected
3359 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3360 * of any partial chunks touched by the range. The invalid portion of
3361 * such chunks will be zeroed.
3363 * (base + size) must be less then or equal to PAGE_SIZE.
3366 vm_page_set_valid_range(vm_page_t m, int base, int size)
3370 VM_OBJECT_ASSERT_WLOCKED(m->object);
3371 if (size == 0) /* handle degenerate case */
3375 * If the base is not DEV_BSIZE aligned and the valid
3376 * bit is clear, we have to zero out a portion of the
3379 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
3380 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
3381 pmap_zero_page_area(m, frag, base - frag);
3384 * If the ending offset is not DEV_BSIZE aligned and the
3385 * valid bit is clear, we have to zero out a portion of
3388 endoff = base + size;
3389 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
3390 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
3391 pmap_zero_page_area(m, endoff,
3392 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
3395 * Assert that no previously invalid block that is now being validated
3398 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
3399 ("vm_page_set_valid_range: page %p is dirty", m));
3402 * Set valid bits inclusive of any overlap.
3404 m->valid |= vm_page_bits(base, size);
3408 * Clear the given bits from the specified page's dirty field.
3410 static __inline void
3411 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
3414 #if PAGE_SIZE < 16384
3419 * If the object is locked and the page is neither exclusive busy nor
3420 * write mapped, then the page's dirty field cannot possibly be
3421 * set by a concurrent pmap operation.
3423 VM_OBJECT_ASSERT_WLOCKED(m->object);
3424 if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
3425 m->dirty &= ~pagebits;
3428 * The pmap layer can call vm_page_dirty() without
3429 * holding a distinguished lock. The combination of
3430 * the object's lock and an atomic operation suffice
3431 * to guarantee consistency of the page dirty field.
3433 * For PAGE_SIZE == 32768 case, compiler already
3434 * properly aligns the dirty field, so no forcible
3435 * alignment is needed. Only require existence of
3436 * atomic_clear_64 when page size is 32768.
3438 addr = (uintptr_t)&m->dirty;
3439 #if PAGE_SIZE == 32768
3440 atomic_clear_64((uint64_t *)addr, pagebits);
3441 #elif PAGE_SIZE == 16384
3442 atomic_clear_32((uint32_t *)addr, pagebits);
3443 #else /* PAGE_SIZE <= 8192 */
3445 * Use a trick to perform a 32-bit atomic on the
3446 * containing aligned word, to not depend on the existence
3447 * of atomic_clear_{8, 16}.
3449 shift = addr & (sizeof(uint32_t) - 1);
3450 #if BYTE_ORDER == BIG_ENDIAN
3451 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
3455 addr &= ~(sizeof(uint32_t) - 1);
3456 atomic_clear_32((uint32_t *)addr, pagebits << shift);
3457 #endif /* PAGE_SIZE */
3462 * vm_page_set_validclean:
3464 * Sets portions of a page valid and clean. The arguments are expected
3465 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3466 * of any partial chunks touched by the range. The invalid portion of
3467 * such chunks will be zero'd.
3469 * (base + size) must be less then or equal to PAGE_SIZE.
3472 vm_page_set_validclean(vm_page_t m, int base, int size)
3474 vm_page_bits_t oldvalid, pagebits;
3477 VM_OBJECT_ASSERT_WLOCKED(m->object);
3478 if (size == 0) /* handle degenerate case */
3482 * If the base is not DEV_BSIZE aligned and the valid
3483 * bit is clear, we have to zero out a portion of the
3486 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
3487 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
3488 pmap_zero_page_area(m, frag, base - frag);
3491 * If the ending offset is not DEV_BSIZE aligned and the
3492 * valid bit is clear, we have to zero out a portion of
3495 endoff = base + size;
3496 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
3497 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
3498 pmap_zero_page_area(m, endoff,
3499 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
3502 * Set valid, clear dirty bits. If validating the entire
3503 * page we can safely clear the pmap modify bit. We also
3504 * use this opportunity to clear the VPO_NOSYNC flag. If a process
3505 * takes a write fault on a MAP_NOSYNC memory area the flag will
3508 * We set valid bits inclusive of any overlap, but we can only
3509 * clear dirty bits for DEV_BSIZE chunks that are fully within
3512 oldvalid = m->valid;
3513 pagebits = vm_page_bits(base, size);
3514 m->valid |= pagebits;
3516 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
3517 frag = DEV_BSIZE - frag;
3523 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
3525 if (base == 0 && size == PAGE_SIZE) {
3527 * The page can only be modified within the pmap if it is
3528 * mapped, and it can only be mapped if it was previously
3531 if (oldvalid == VM_PAGE_BITS_ALL)
3533 * Perform the pmap_clear_modify() first. Otherwise,
3534 * a concurrent pmap operation, such as
3535 * pmap_protect(), could clear a modification in the
3536 * pmap and set the dirty field on the page before
3537 * pmap_clear_modify() had begun and after the dirty
3538 * field was cleared here.
3540 pmap_clear_modify(m);
3542 m->oflags &= ~VPO_NOSYNC;
3543 } else if (oldvalid != VM_PAGE_BITS_ALL)
3544 m->dirty &= ~pagebits;
3546 vm_page_clear_dirty_mask(m, pagebits);
3550 vm_page_clear_dirty(vm_page_t m, int base, int size)
3553 vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
3557 * vm_page_set_invalid:
3559 * Invalidates DEV_BSIZE'd chunks within a page. Both the
3560 * valid and dirty bits for the effected areas are cleared.
3563 vm_page_set_invalid(vm_page_t m, int base, int size)
3565 vm_page_bits_t bits;
3569 VM_OBJECT_ASSERT_WLOCKED(object);
3570 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
3571 size >= object->un_pager.vnp.vnp_size)
3572 bits = VM_PAGE_BITS_ALL;
3574 bits = vm_page_bits(base, size);
3575 if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL &&
3578 KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) ||
3579 !pmap_page_is_mapped(m),
3580 ("vm_page_set_invalid: page %p is mapped", m));
3586 * vm_page_zero_invalid()
3588 * The kernel assumes that the invalid portions of a page contain
3589 * garbage, but such pages can be mapped into memory by user code.
3590 * When this occurs, we must zero out the non-valid portions of the
3591 * page so user code sees what it expects.
3593 * Pages are most often semi-valid when the end of a file is mapped
3594 * into memory and the file's size is not page aligned.
3597 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
3602 VM_OBJECT_ASSERT_WLOCKED(m->object);
3604 * Scan the valid bits looking for invalid sections that
3605 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the
3606 * valid bit may be set ) have already been zeroed by
3607 * vm_page_set_validclean().
3609 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
3610 if (i == (PAGE_SIZE / DEV_BSIZE) ||
3611 (m->valid & ((vm_page_bits_t)1 << i))) {
3613 pmap_zero_page_area(m,
3614 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
3621 * setvalid is TRUE when we can safely set the zero'd areas
3622 * as being valid. We can do this if there are no cache consistancy
3623 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
3626 m->valid = VM_PAGE_BITS_ALL;
3632 * Is (partial) page valid? Note that the case where size == 0
3633 * will return FALSE in the degenerate case where the page is
3634 * entirely invalid, and TRUE otherwise.
3637 vm_page_is_valid(vm_page_t m, int base, int size)
3639 vm_page_bits_t bits;
3641 VM_OBJECT_ASSERT_LOCKED(m->object);
3642 bits = vm_page_bits(base, size);
3643 return (m->valid != 0 && (m->valid & bits) == bits);
3647 * Returns true if all of the specified predicates are true for the entire
3648 * (super)page and false otherwise.
3651 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
3657 VM_OBJECT_ASSERT_LOCKED(object);
3658 npages = atop(pagesizes[m->psind]);
3661 * The physically contiguous pages that make up a superpage, i.e., a
3662 * page with a page size index ("psind") greater than zero, will
3663 * occupy adjacent entries in vm_page_array[].
3665 for (i = 0; i < npages; i++) {
3666 /* Always test object consistency, including "skip_m". */
3667 if (m[i].object != object)
3669 if (&m[i] == skip_m)
3671 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
3673 if ((flags & PS_ALL_DIRTY) != 0) {
3675 * Calling vm_page_test_dirty() or pmap_is_modified()
3676 * might stop this case from spuriously returning
3677 * "false". However, that would require a write lock
3678 * on the object containing "m[i]".
3680 if (m[i].dirty != VM_PAGE_BITS_ALL)
3683 if ((flags & PS_ALL_VALID) != 0 &&
3684 m[i].valid != VM_PAGE_BITS_ALL)
3691 * Set the page's dirty bits if the page is modified.
3694 vm_page_test_dirty(vm_page_t m)
3697 VM_OBJECT_ASSERT_WLOCKED(m->object);
3698 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
3703 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
3706 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
3710 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
3713 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
3717 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
3720 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
3723 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
3725 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
3728 vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
3732 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
3735 mtx_assert_(vm_page_lockptr(m), a, file, line);
3741 vm_page_object_lock_assert(vm_page_t m)
3745 * Certain of the page's fields may only be modified by the
3746 * holder of the containing object's lock or the exclusive busy.
3747 * holder. Unfortunately, the holder of the write busy is
3748 * not recorded, and thus cannot be checked here.
3750 if (m->object != NULL && !vm_page_xbusied(m))
3751 VM_OBJECT_ASSERT_WLOCKED(m->object);
3755 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits)
3758 if ((bits & PGA_WRITEABLE) == 0)
3762 * The PGA_WRITEABLE flag can only be set if the page is
3763 * managed, is exclusively busied or the object is locked.
3764 * Currently, this flag is only set by pmap_enter().
3766 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3767 ("PGA_WRITEABLE on unmanaged page"));
3768 if (!vm_page_xbusied(m))
3769 VM_OBJECT_ASSERT_LOCKED(m->object);
3773 #include "opt_ddb.h"
3775 #include <sys/kernel.h>
3777 #include <ddb/ddb.h>
3779 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3782 db_printf("vm_cnt.v_free_count: %d\n", vm_cnt.v_free_count);
3783 db_printf("vm_cnt.v_inactive_count: %d\n", vm_cnt.v_inactive_count);
3784 db_printf("vm_cnt.v_active_count: %d\n", vm_cnt.v_active_count);
3785 db_printf("vm_cnt.v_laundry_count: %d\n", vm_cnt.v_laundry_count);
3786 db_printf("vm_cnt.v_wire_count: %d\n", vm_cnt.v_wire_count);
3787 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
3788 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
3789 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
3790 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
3793 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3797 db_printf("pq_free %d\n", vm_cnt.v_free_count);
3798 for (dom = 0; dom < vm_ndomains; dom++) {
3800 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
3802 vm_dom[dom].vmd_page_count,
3803 vm_dom[dom].vmd_free_count,
3804 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
3805 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
3806 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
3807 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
3811 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
3817 db_printf("show pginfo addr\n");
3821 phys = strchr(modif, 'p') != NULL;
3823 m = PHYS_TO_VM_PAGE(addr);
3825 m = (vm_page_t)addr;
3827 "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n"
3828 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
3829 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
3830 m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags,
3831 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty);