2 * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
4 * Copyright (c) 1991 Regents of the University of California.
6 * Copyright (c) 1998 Matthew Dillon. All Rights Reserved.
8 * This code is derived from software contributed to Berkeley by
9 * The Mach Operating System project at Carnegie-Mellon University.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
14 * 1. Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
16 * 2. Redistributions in binary form must reproduce the above copyright
17 * notice, this list of conditions and the following disclaimer in the
18 * documentation and/or other materials provided with the distribution.
19 * 3. Neither the name of the University nor the names of its contributors
20 * may be used to endorse or promote products derived from this software
21 * without specific prior written permission.
23 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
24 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
27 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91
39 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40 * All rights reserved.
42 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
44 * Permission to use, copy, modify and distribute this software and
45 * its documentation is hereby granted, provided that both the copyright
46 * notice and this permission notice appear in all copies of the
47 * software, derivative works or modified versions, and any portions
48 * thereof, and that both notices appear in supporting documentation.
50 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
54 * Carnegie Mellon requests users of this software to return to
56 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
57 * School of Computer Science
58 * Carnegie Mellon University
59 * Pittsburgh PA 15213-3890
61 * any improvements or extensions that they make and grant Carnegie the
62 * rights to redistribute these changes.
66 * Resident memory management module.
69 #include <sys/cdefs.h>
70 __FBSDID("$FreeBSD$");
74 #include <sys/param.h>
75 #include <sys/systm.h>
77 #include <sys/domainset.h>
78 #include <sys/kernel.h>
79 #include <sys/limits.h>
80 #include <sys/linker.h>
81 #include <sys/malloc.h>
83 #include <sys/msgbuf.h>
84 #include <sys/mutex.h>
86 #include <sys/rwlock.h>
88 #include <sys/sched.h>
90 #include <sys/sysctl.h>
91 #include <sys/vmmeter.h>
92 #include <sys/vnode.h>
96 #include <vm/vm_param.h>
97 #include <vm/vm_domainset.h>
98 #include <vm/vm_kern.h>
99 #include <vm/vm_map.h>
100 #include <vm/vm_object.h>
101 #include <vm/vm_page.h>
102 #include <vm/vm_pageout.h>
103 #include <vm/vm_phys.h>
104 #include <vm/vm_pagequeue.h>
105 #include <vm/vm_pager.h>
106 #include <vm/vm_radix.h>
107 #include <vm/vm_reserv.h>
108 #include <vm/vm_extern.h>
110 #include <vm/uma_int.h>
112 #include <machine/md_var.h>
114 extern int uma_startup_count(int);
115 extern void uma_startup(void *, int);
116 extern int vmem_startup_count(void);
118 struct vm_domain vm_dom[MAXMEMDOM];
120 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]);
122 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
124 struct mtx_padalign __exclusive_cache_line vm_domainset_lock;
125 /* The following fields are protected by the domainset lock. */
126 domainset_t __exclusive_cache_line vm_min_domains;
127 domainset_t __exclusive_cache_line vm_severe_domains;
128 static int vm_min_waiters;
129 static int vm_severe_waiters;
130 static int vm_pageproc_waiters;
133 * bogus page -- for I/O to/from partially complete buffers,
134 * or for paging into sparsely invalid regions.
136 vm_page_t bogus_page;
138 vm_page_t vm_page_array;
139 long vm_page_array_size;
142 static int boot_pages;
143 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
145 "number of pages allocated for bootstrapping the VM system");
147 static int pa_tryrelock_restart;
148 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
149 &pa_tryrelock_restart, 0, "Number of tryrelock restarts");
151 static TAILQ_HEAD(, vm_page) blacklist_head;
152 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
153 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
154 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
156 static uma_zone_t fakepg_zone;
158 static void vm_page_alloc_check(vm_page_t m);
159 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
160 static void vm_page_dequeue_complete(vm_page_t m);
161 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
162 static void vm_page_init(void *dummy);
163 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
164 vm_pindex_t pindex, vm_page_t mpred);
165 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
167 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
168 vm_page_t m_run, vm_paddr_t high);
169 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
171 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain,
173 static void vm_page_zone_release(void *arg, void **store, int cnt);
175 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
178 vm_page_init(void *dummy)
181 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
182 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
183 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
184 VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
188 * The cache page zone is initialized later since we need to be able to allocate
189 * pages before UMA is fully initialized.
192 vm_page_init_cache_zones(void *dummy __unused)
194 struct vm_domain *vmd;
195 struct vm_pgcache *pgcache;
198 for (domain = 0; domain < vm_ndomains; domain++) {
199 vmd = VM_DOMAIN(domain);
202 * Don't allow the page caches to take up more than .1875% of
203 * memory. A UMA bucket contains at most 256 free pages, and we
204 * have two buckets per CPU per free pool.
206 if (vmd->vmd_page_count / 600 < 2 * 256 * mp_ncpus *
209 for (pool = 0; pool < VM_NFREEPOOL; pool++) {
210 pgcache = &vmd->vmd_pgcache[pool];
211 pgcache->domain = domain;
212 pgcache->pool = pool;
213 pgcache->zone = uma_zcache_create("vm pgcache",
214 sizeof(struct vm_page), NULL, NULL, NULL, NULL,
215 vm_page_zone_import, vm_page_zone_release, pgcache,
216 UMA_ZONE_NOBUCKETCACHE | UMA_ZONE_MAXBUCKET |
221 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
223 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
224 #if PAGE_SIZE == 32768
226 CTASSERT(sizeof(u_long) >= 8);
231 * Try to acquire a physical address lock while a pmap is locked. If we
232 * fail to trylock we unlock and lock the pmap directly and cache the
233 * locked pa in *locked. The caller should then restart their loop in case
234 * the virtual to physical mapping has changed.
237 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
244 PA_LOCK_ASSERT(lockpa, MA_OWNED);
245 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
252 atomic_add_int(&pa_tryrelock_restart, 1);
261 * Sets the page size, perhaps based upon the memory
262 * size. Must be called before any use of page-size
263 * dependent functions.
266 vm_set_page_size(void)
268 if (vm_cnt.v_page_size == 0)
269 vm_cnt.v_page_size = PAGE_SIZE;
270 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
271 panic("vm_set_page_size: page size not a power of two");
275 * vm_page_blacklist_next:
277 * Find the next entry in the provided string of blacklist
278 * addresses. Entries are separated by space, comma, or newline.
279 * If an invalid integer is encountered then the rest of the
280 * string is skipped. Updates the list pointer to the next
281 * character, or NULL if the string is exhausted or invalid.
284 vm_page_blacklist_next(char **list, char *end)
289 if (list == NULL || *list == NULL)
297 * If there's no end pointer then the buffer is coming from
298 * the kenv and we know it's null-terminated.
301 end = *list + strlen(*list);
303 /* Ensure that strtoq() won't walk off the end */
305 if (*end == '\n' || *end == ' ' || *end == ',')
308 printf("Blacklist not terminated, skipping\n");
314 for (pos = *list; *pos != '\0'; pos = cp) {
315 bad = strtoq(pos, &cp, 0);
316 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
325 if (*cp == '\0' || ++cp >= end)
329 return (trunc_page(bad));
331 printf("Garbage in RAM blacklist, skipping\n");
337 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
339 struct vm_domain *vmd;
343 m = vm_phys_paddr_to_vm_page(pa);
345 return (true); /* page does not exist, no failure */
347 vmd = vm_pagequeue_domain(m);
348 vm_domain_free_lock(vmd);
349 ret = vm_phys_unfree_page(m);
350 vm_domain_free_unlock(vmd);
352 vm_domain_freecnt_inc(vmd, -1);
353 TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
355 printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
361 * vm_page_blacklist_check:
363 * Iterate through the provided string of blacklist addresses, pulling
364 * each entry out of the physical allocator free list and putting it
365 * onto a list for reporting via the vm.page_blacklist sysctl.
368 vm_page_blacklist_check(char *list, char *end)
374 while (next != NULL) {
375 if ((pa = vm_page_blacklist_next(&next, end)) == 0)
377 vm_page_blacklist_add(pa, bootverbose);
382 * vm_page_blacklist_load:
384 * Search for a special module named "ram_blacklist". It'll be a
385 * plain text file provided by the user via the loader directive
389 vm_page_blacklist_load(char **list, char **end)
398 mod = preload_search_by_type("ram_blacklist");
400 ptr = preload_fetch_addr(mod);
401 len = preload_fetch_size(mod);
412 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
419 error = sysctl_wire_old_buffer(req, 0);
422 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
423 TAILQ_FOREACH(m, &blacklist_head, listq) {
424 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
425 (uintmax_t)m->phys_addr);
428 error = sbuf_finish(&sbuf);
434 * Initialize a dummy page for use in scans of the specified paging queue.
435 * In principle, this function only needs to set the flag PG_MARKER.
436 * Nonetheless, it write busies and initializes the hold count to one as
437 * safety precautions.
440 vm_page_init_marker(vm_page_t marker, int queue, uint8_t aflags)
443 bzero(marker, sizeof(*marker));
444 marker->flags = PG_MARKER;
445 marker->aflags = aflags;
446 marker->busy_lock = VPB_SINGLE_EXCLUSIVER;
447 marker->queue = queue;
448 marker->hold_count = 1;
452 vm_page_domain_init(int domain)
454 struct vm_domain *vmd;
455 struct vm_pagequeue *pq;
458 vmd = VM_DOMAIN(domain);
459 bzero(vmd, sizeof(*vmd));
460 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
461 "vm inactive pagequeue";
462 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
463 "vm active pagequeue";
464 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
465 "vm laundry pagequeue";
466 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
467 "vm unswappable pagequeue";
468 vmd->vmd_domain = domain;
469 vmd->vmd_page_count = 0;
470 vmd->vmd_free_count = 0;
472 vmd->vmd_oom = FALSE;
473 for (i = 0; i < PQ_COUNT; i++) {
474 pq = &vmd->vmd_pagequeues[i];
475 TAILQ_INIT(&pq->pq_pl);
476 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
477 MTX_DEF | MTX_DUPOK);
479 vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
481 mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
482 mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
483 snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
486 * inacthead is used to provide FIFO ordering for LRU-bypassing
489 vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
490 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
491 &vmd->vmd_inacthead, plinks.q);
494 * The clock pages are used to implement active queue scanning without
495 * requeues. Scans start at clock[0], which is advanced after the scan
496 * ends. When the two clock hands meet, they are reset and scanning
497 * resumes from the head of the queue.
499 vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
500 vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
501 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
502 &vmd->vmd_clock[0], plinks.q);
503 TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
504 &vmd->vmd_clock[1], plinks.q);
508 * Initialize a physical page in preparation for adding it to the free
512 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
517 m->busy_lock = VPB_UNBUSIED;
519 m->flags = m->aflags = 0;
524 m->order = VM_NFREEORDER;
525 m->pool = VM_FREEPOOL_DEFAULT;
526 m->valid = m->dirty = 0;
533 * Initializes the resident memory module. Allocates physical memory for
534 * bootstrapping UMA and some data structures that are used to manage
535 * physical pages. Initializes these structures, and populates the free
539 vm_page_startup(vm_offset_t vaddr)
541 struct vm_phys_seg *seg;
543 char *list, *listend;
545 vm_paddr_t end, high_avail, low_avail, new_end, page_range, size;
546 vm_paddr_t biggestsize, last_pa, pa;
548 int biggestone, i, segind;
549 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
555 vaddr = round_page(vaddr);
557 for (i = 0; phys_avail[i + 1]; i += 2) {
558 phys_avail[i] = round_page(phys_avail[i]);
559 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
561 for (i = 0; phys_avail[i + 1]; i += 2) {
562 size = phys_avail[i + 1] - phys_avail[i];
563 if (size > biggestsize) {
569 end = phys_avail[biggestone+1];
572 * Initialize the page and queue locks.
574 mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
575 for (i = 0; i < PA_LOCK_COUNT; i++)
576 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
577 for (i = 0; i < vm_ndomains; i++)
578 vm_page_domain_init(i);
581 * Allocate memory for use when boot strapping the kernel memory
582 * allocator. Tell UMA how many zones we are going to create
583 * before going fully functional. UMA will add its zones.
585 * VM startup zones: vmem, vmem_btag, VM OBJECT, RADIX NODE, MAP,
586 * KMAP ENTRY, MAP ENTRY, VMSPACE.
588 boot_pages = uma_startup_count(8);
590 #ifndef UMA_MD_SMALL_ALLOC
591 /* vmem_startup() calls uma_prealloc(). */
592 boot_pages += vmem_startup_count();
593 /* vm_map_startup() calls uma_prealloc(). */
594 boot_pages += howmany(MAX_KMAP,
595 UMA_SLAB_SPACE / sizeof(struct vm_map));
598 * Before going fully functional kmem_init() does allocation
599 * from "KMAP ENTRY" and vmem_create() does allocation from "vmem".
604 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
605 * manually fetch the value.
607 TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
608 new_end = end - (boot_pages * UMA_SLAB_SIZE);
609 new_end = trunc_page(new_end);
610 mapped = pmap_map(&vaddr, new_end, end,
611 VM_PROT_READ | VM_PROT_WRITE);
612 bzero((void *)mapped, end - new_end);
613 uma_startup((void *)mapped, boot_pages);
617 new_end = end - round_page(witness_startup_count());
618 mapped = pmap_map(&vaddr, new_end, end,
619 VM_PROT_READ | VM_PROT_WRITE);
620 bzero((void *)mapped, end - new_end);
621 witness_startup((void *)mapped);
624 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
625 defined(__i386__) || defined(__mips__) || defined(__riscv)
627 * Allocate a bitmap to indicate that a random physical page
628 * needs to be included in a minidump.
630 * The amd64 port needs this to indicate which direct map pages
631 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
633 * However, i386 still needs this workspace internally within the
634 * minidump code. In theory, they are not needed on i386, but are
635 * included should the sf_buf code decide to use them.
638 for (i = 0; dump_avail[i + 1] != 0; i += 2)
639 if (dump_avail[i + 1] > last_pa)
640 last_pa = dump_avail[i + 1];
641 page_range = last_pa / PAGE_SIZE;
642 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
643 new_end -= vm_page_dump_size;
644 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
645 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
646 bzero((void *)vm_page_dump, vm_page_dump_size);
650 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
653 * Include the UMA bootstrap pages and vm_page_dump in a crash dump.
654 * When pmap_map() uses the direct map, they are not automatically
657 for (pa = new_end; pa < end; pa += PAGE_SIZE)
660 phys_avail[biggestone + 1] = new_end;
663 * Request that the physical pages underlying the message buffer be
664 * included in a crash dump. Since the message buffer is accessed
665 * through the direct map, they are not automatically included.
667 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
668 last_pa = pa + round_page(msgbufsize);
669 while (pa < last_pa) {
675 * Compute the number of pages of memory that will be available for
676 * use, taking into account the overhead of a page structure per page.
677 * In other words, solve
678 * "available physical memory" - round_page(page_range *
679 * sizeof(struct vm_page)) = page_range * PAGE_SIZE
682 low_avail = phys_avail[0];
683 high_avail = phys_avail[1];
684 for (i = 0; i < vm_phys_nsegs; i++) {
685 if (vm_phys_segs[i].start < low_avail)
686 low_avail = vm_phys_segs[i].start;
687 if (vm_phys_segs[i].end > high_avail)
688 high_avail = vm_phys_segs[i].end;
690 /* Skip the first chunk. It is already accounted for. */
691 for (i = 2; phys_avail[i + 1] != 0; i += 2) {
692 if (phys_avail[i] < low_avail)
693 low_avail = phys_avail[i];
694 if (phys_avail[i + 1] > high_avail)
695 high_avail = phys_avail[i + 1];
697 first_page = low_avail / PAGE_SIZE;
698 #ifdef VM_PHYSSEG_SPARSE
700 for (i = 0; i < vm_phys_nsegs; i++)
701 size += vm_phys_segs[i].end - vm_phys_segs[i].start;
702 for (i = 0; phys_avail[i + 1] != 0; i += 2)
703 size += phys_avail[i + 1] - phys_avail[i];
704 #elif defined(VM_PHYSSEG_DENSE)
705 size = high_avail - low_avail;
707 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
710 #ifdef VM_PHYSSEG_DENSE
712 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
713 * the overhead of a page structure per page only if vm_page_array is
714 * allocated from the last physical memory chunk. Otherwise, we must
715 * allocate page structures representing the physical memory
716 * underlying vm_page_array, even though they will not be used.
718 if (new_end != high_avail)
719 page_range = size / PAGE_SIZE;
723 page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
726 * If the partial bytes remaining are large enough for
727 * a page (PAGE_SIZE) without a corresponding
728 * 'struct vm_page', then new_end will contain an
729 * extra page after subtracting the length of the VM
730 * page array. Compensate by subtracting an extra
733 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
734 if (new_end == high_avail)
735 high_avail -= PAGE_SIZE;
736 new_end -= PAGE_SIZE;
742 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
743 * However, because this page is allocated from KVM, out-of-bounds
744 * accesses using the direct map will not be trapped.
749 * Allocate physical memory for the page structures, and map it.
751 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
752 mapped = pmap_map(&vaddr, new_end, end,
753 VM_PROT_READ | VM_PROT_WRITE);
754 vm_page_array = (vm_page_t)mapped;
755 vm_page_array_size = page_range;
757 #if VM_NRESERVLEVEL > 0
759 * Allocate physical memory for the reservation management system's
760 * data structures, and map it.
762 if (high_avail == end)
763 high_avail = new_end;
764 new_end = vm_reserv_startup(&vaddr, new_end, high_avail);
766 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
769 * Include vm_page_array and vm_reserv_array in a crash dump.
771 for (pa = new_end; pa < end; pa += PAGE_SIZE)
774 phys_avail[biggestone + 1] = new_end;
777 * Add physical memory segments corresponding to the available
780 for (i = 0; phys_avail[i + 1] != 0; i += 2)
781 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
784 * Initialize the physical memory allocator.
789 * Initialize the page structures and add every available page to the
790 * physical memory allocator's free lists.
792 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
793 for (ii = 0; ii < vm_page_array_size; ii++) {
794 m = &vm_page_array[ii];
795 vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0);
796 m->flags = PG_FICTITIOUS;
799 vm_cnt.v_page_count = 0;
800 for (segind = 0; segind < vm_phys_nsegs; segind++) {
801 seg = &vm_phys_segs[segind];
802 for (m = seg->first_page, pa = seg->start; pa < seg->end;
803 m++, pa += PAGE_SIZE)
804 vm_page_init_page(m, pa, segind);
807 * Add the segment to the free lists only if it is covered by
808 * one of the ranges in phys_avail. Because we've added the
809 * ranges to the vm_phys_segs array, we can assume that each
810 * segment is either entirely contained in one of the ranges,
811 * or doesn't overlap any of them.
813 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
814 struct vm_domain *vmd;
816 if (seg->start < phys_avail[i] ||
817 seg->end > phys_avail[i + 1])
821 pagecount = (u_long)atop(seg->end - seg->start);
823 vmd = VM_DOMAIN(seg->domain);
824 vm_domain_free_lock(vmd);
825 vm_phys_free_contig(m, pagecount);
826 vm_domain_free_unlock(vmd);
827 vm_domain_freecnt_inc(vmd, pagecount);
828 vm_cnt.v_page_count += (u_int)pagecount;
830 vmd = VM_DOMAIN(seg->domain);
831 vmd->vmd_page_count += (u_int)pagecount;
832 vmd->vmd_segs |= 1UL << m->segind;
838 * Remove blacklisted pages from the physical memory allocator.
840 TAILQ_INIT(&blacklist_head);
841 vm_page_blacklist_load(&list, &listend);
842 vm_page_blacklist_check(list, listend);
844 list = kern_getenv("vm.blacklist");
845 vm_page_blacklist_check(list, NULL);
848 #if VM_NRESERVLEVEL > 0
850 * Initialize the reservation management system.
859 vm_page_reference(vm_page_t m)
862 vm_page_aflag_set(m, PGA_REFERENCED);
866 * vm_page_busy_downgrade:
868 * Downgrade an exclusive busy page into a single shared busy page.
871 vm_page_busy_downgrade(vm_page_t m)
876 vm_page_assert_xbusied(m);
877 locked = mtx_owned(vm_page_lockptr(m));
881 x &= VPB_BIT_WAITERS;
882 if (x != 0 && !locked)
884 if (atomic_cmpset_rel_int(&m->busy_lock,
885 VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1)))
887 if (x != 0 && !locked)
900 * Return a positive value if the page is shared busied, 0 otherwise.
903 vm_page_sbusied(vm_page_t m)
908 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
914 * Shared unbusy a page.
917 vm_page_sunbusy(vm_page_t m)
921 vm_page_lock_assert(m, MA_NOTOWNED);
922 vm_page_assert_sbusied(m);
926 if (VPB_SHARERS(x) > 1) {
927 if (atomic_cmpset_int(&m->busy_lock, x,
932 if ((x & VPB_BIT_WAITERS) == 0) {
933 KASSERT(x == VPB_SHARERS_WORD(1),
934 ("vm_page_sunbusy: invalid lock state"));
935 if (atomic_cmpset_int(&m->busy_lock,
936 VPB_SHARERS_WORD(1), VPB_UNBUSIED))
940 KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS),
941 ("vm_page_sunbusy: invalid lock state for waiters"));
944 if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) {
955 * vm_page_busy_sleep:
957 * Sleep and release the page lock, using the page pointer as wchan.
958 * This is used to implement the hard-path of busying mechanism.
960 * The given page must be locked.
962 * If nonshared is true, sleep only if the page is xbusy.
965 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
969 vm_page_assert_locked(m);
972 if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) ||
973 ((x & VPB_BIT_WAITERS) == 0 &&
974 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) {
978 msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0);
984 * Try to shared busy a page.
985 * If the operation succeeds 1 is returned otherwise 0.
986 * The operation never sleeps.
989 vm_page_trysbusy(vm_page_t m)
995 if ((x & VPB_BIT_SHARED) == 0)
997 if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER))
1003 vm_page_xunbusy_locked(vm_page_t m)
1006 vm_page_assert_xbusied(m);
1007 vm_page_assert_locked(m);
1009 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1010 /* There is a waiter, do wakeup() instead of vm_page_flash(). */
1015 vm_page_xunbusy_maybelocked(vm_page_t m)
1019 vm_page_assert_xbusied(m);
1022 * Fast path for unbusy. If it succeeds, we know that there
1023 * are no waiters, so we do not need a wakeup.
1025 if (atomic_cmpset_rel_int(&m->busy_lock, VPB_SINGLE_EXCLUSIVER,
1029 lockacq = !mtx_owned(vm_page_lockptr(m));
1032 vm_page_xunbusy_locked(m);
1038 * vm_page_xunbusy_hard:
1040 * Called after the first try the exclusive unbusy of a page failed.
1041 * It is assumed that the waiters bit is on.
1044 vm_page_xunbusy_hard(vm_page_t m)
1047 vm_page_assert_xbusied(m);
1050 vm_page_xunbusy_locked(m);
1057 * Wakeup anyone waiting for the page.
1058 * The ownership bits do not change.
1060 * The given page must be locked.
1063 vm_page_flash(vm_page_t m)
1067 vm_page_lock_assert(m, MA_OWNED);
1071 if ((x & VPB_BIT_WAITERS) == 0)
1073 if (atomic_cmpset_int(&m->busy_lock, x,
1074 x & (~VPB_BIT_WAITERS)))
1081 * Avoid releasing and reacquiring the same page lock.
1084 vm_page_change_lock(vm_page_t m, struct mtx **mtx)
1088 mtx1 = vm_page_lockptr(m);
1098 * Keep page from being freed by the page daemon
1099 * much of the same effect as wiring, except much lower
1100 * overhead and should be used only for *very* temporary
1101 * holding ("wiring").
1104 vm_page_hold(vm_page_t mem)
1107 vm_page_lock_assert(mem, MA_OWNED);
1112 vm_page_unhold(vm_page_t mem)
1115 vm_page_lock_assert(mem, MA_OWNED);
1116 KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!"));
1118 if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0)
1119 vm_page_free_toq(mem);
1123 * vm_page_unhold_pages:
1125 * Unhold each of the pages that is referenced by the given array.
1128 vm_page_unhold_pages(vm_page_t *ma, int count)
1133 for (; count != 0; count--) {
1134 vm_page_change_lock(*ma, &mtx);
1135 vm_page_unhold(*ma);
1143 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1147 #ifdef VM_PHYSSEG_SPARSE
1148 m = vm_phys_paddr_to_vm_page(pa);
1150 m = vm_phys_fictitious_to_vm_page(pa);
1152 #elif defined(VM_PHYSSEG_DENSE)
1156 if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1157 m = &vm_page_array[pi - first_page];
1160 return (vm_phys_fictitious_to_vm_page(pa));
1162 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1169 * Create a fictitious page with the specified physical address and
1170 * memory attribute. The memory attribute is the only the machine-
1171 * dependent aspect of a fictitious page that must be initialized.
1174 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1178 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1179 vm_page_initfake(m, paddr, memattr);
1184 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1187 if ((m->flags & PG_FICTITIOUS) != 0) {
1189 * The page's memattr might have changed since the
1190 * previous initialization. Update the pmap to the
1195 m->phys_addr = paddr;
1197 /* Fictitious pages don't use "segind". */
1198 m->flags = PG_FICTITIOUS;
1199 /* Fictitious pages don't use "order" or "pool". */
1200 m->oflags = VPO_UNMANAGED;
1201 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1205 pmap_page_set_memattr(m, memattr);
1211 * Release a fictitious page.
1214 vm_page_putfake(vm_page_t m)
1217 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1218 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1219 ("vm_page_putfake: bad page %p", m));
1220 uma_zfree(fakepg_zone, m);
1224 * vm_page_updatefake:
1226 * Update the given fictitious page to the specified physical address and
1230 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1233 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1234 ("vm_page_updatefake: bad page %p", m));
1235 m->phys_addr = paddr;
1236 pmap_page_set_memattr(m, memattr);
1245 vm_page_free(vm_page_t m)
1248 m->flags &= ~PG_ZERO;
1249 vm_page_free_toq(m);
1253 * vm_page_free_zero:
1255 * Free a page to the zerod-pages queue
1258 vm_page_free_zero(vm_page_t m)
1261 m->flags |= PG_ZERO;
1262 vm_page_free_toq(m);
1266 * Unbusy and handle the page queueing for a page from a getpages request that
1267 * was optionally read ahead or behind.
1270 vm_page_readahead_finish(vm_page_t m)
1273 /* We shouldn't put invalid pages on queues. */
1274 KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m));
1277 * Since the page is not the actually needed one, whether it should
1278 * be activated or deactivated is not obvious. Empirical results
1279 * have shown that deactivating the page is usually the best choice,
1280 * unless the page is wanted by another thread.
1283 if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1284 vm_page_activate(m);
1286 vm_page_deactivate(m);
1292 * vm_page_sleep_if_busy:
1294 * Sleep and release the page queues lock if the page is busied.
1295 * Returns TRUE if the thread slept.
1297 * The given page must be unlocked and object containing it must
1301 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1305 vm_page_lock_assert(m, MA_NOTOWNED);
1306 VM_OBJECT_ASSERT_WLOCKED(m->object);
1308 if (vm_page_busied(m)) {
1310 * The page-specific object must be cached because page
1311 * identity can change during the sleep, causing the
1312 * re-lock of a different object.
1313 * It is assumed that a reference to the object is already
1314 * held by the callers.
1318 VM_OBJECT_WUNLOCK(obj);
1319 vm_page_busy_sleep(m, msg, false);
1320 VM_OBJECT_WLOCK(obj);
1327 * vm_page_dirty_KBI: [ internal use only ]
1329 * Set all bits in the page's dirty field.
1331 * The object containing the specified page must be locked if the
1332 * call is made from the machine-independent layer.
1334 * See vm_page_clear_dirty_mask().
1336 * This function should only be called by vm_page_dirty().
1339 vm_page_dirty_KBI(vm_page_t m)
1342 /* Refer to this operation by its public name. */
1343 KASSERT(m->valid == VM_PAGE_BITS_ALL,
1344 ("vm_page_dirty: page is invalid!"));
1345 m->dirty = VM_PAGE_BITS_ALL;
1349 * vm_page_insert: [ internal use only ]
1351 * Inserts the given mem entry into the object and object list.
1353 * The object must be locked.
1356 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1360 VM_OBJECT_ASSERT_WLOCKED(object);
1361 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1362 return (vm_page_insert_after(m, object, pindex, mpred));
1366 * vm_page_insert_after:
1368 * Inserts the page "m" into the specified object at offset "pindex".
1370 * The page "mpred" must immediately precede the offset "pindex" within
1371 * the specified object.
1373 * The object must be locked.
1376 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1381 VM_OBJECT_ASSERT_WLOCKED(object);
1382 KASSERT(m->object == NULL,
1383 ("vm_page_insert_after: page already inserted"));
1384 if (mpred != NULL) {
1385 KASSERT(mpred->object == object,
1386 ("vm_page_insert_after: object doesn't contain mpred"));
1387 KASSERT(mpred->pindex < pindex,
1388 ("vm_page_insert_after: mpred doesn't precede pindex"));
1389 msucc = TAILQ_NEXT(mpred, listq);
1391 msucc = TAILQ_FIRST(&object->memq);
1393 KASSERT(msucc->pindex > pindex,
1394 ("vm_page_insert_after: msucc doesn't succeed pindex"));
1397 * Record the object/offset pair in this page
1403 * Now link into the object's ordered list of backed pages.
1405 if (vm_radix_insert(&object->rtree, m)) {
1410 vm_page_insert_radixdone(m, object, mpred);
1415 * vm_page_insert_radixdone:
1417 * Complete page "m" insertion into the specified object after the
1418 * radix trie hooking.
1420 * The page "mpred" must precede the offset "m->pindex" within the
1423 * The object must be locked.
1426 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1429 VM_OBJECT_ASSERT_WLOCKED(object);
1430 KASSERT(object != NULL && m->object == object,
1431 ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1432 if (mpred != NULL) {
1433 KASSERT(mpred->object == object,
1434 ("vm_page_insert_after: object doesn't contain mpred"));
1435 KASSERT(mpred->pindex < m->pindex,
1436 ("vm_page_insert_after: mpred doesn't precede pindex"));
1440 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1442 TAILQ_INSERT_HEAD(&object->memq, m, listq);
1445 * Show that the object has one more resident page.
1447 object->resident_page_count++;
1450 * Hold the vnode until the last page is released.
1452 if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1453 vhold(object->handle);
1456 * Since we are inserting a new and possibly dirty page,
1457 * update the object's OBJ_MIGHTBEDIRTY flag.
1459 if (pmap_page_is_write_mapped(m))
1460 vm_object_set_writeable_dirty(object);
1466 * Removes the specified page from its containing object, but does not
1467 * invalidate any backing storage. Return true if the page may be safely
1468 * freed and false otherwise.
1470 * The object must be locked. The page must be locked if it is managed.
1473 vm_page_remove(vm_page_t m)
1480 if ((m->oflags & VPO_UNMANAGED) == 0)
1481 vm_page_assert_locked(m);
1482 VM_OBJECT_ASSERT_WLOCKED(object);
1483 if (vm_page_xbusied(m))
1484 vm_page_xunbusy_maybelocked(m);
1485 mrem = vm_radix_remove(&object->rtree, m->pindex);
1486 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1489 * Now remove from the object's list of backed pages.
1491 TAILQ_REMOVE(&object->memq, m, listq);
1494 * And show that the object has one fewer resident page.
1496 object->resident_page_count--;
1499 * The vnode may now be recycled.
1501 if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1502 vdrop(object->handle);
1505 return (!vm_page_wired(m));
1511 * Returns the page associated with the object/offset
1512 * pair specified; if none is found, NULL is returned.
1514 * The object must be locked.
1517 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1520 VM_OBJECT_ASSERT_LOCKED(object);
1521 return (vm_radix_lookup(&object->rtree, pindex));
1525 * vm_page_find_least:
1527 * Returns the page associated with the object with least pindex
1528 * greater than or equal to the parameter pindex, or NULL.
1530 * The object must be locked.
1533 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1537 VM_OBJECT_ASSERT_LOCKED(object);
1538 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1539 m = vm_radix_lookup_ge(&object->rtree, pindex);
1544 * Returns the given page's successor (by pindex) within the object if it is
1545 * resident; if none is found, NULL is returned.
1547 * The object must be locked.
1550 vm_page_next(vm_page_t m)
1554 VM_OBJECT_ASSERT_LOCKED(m->object);
1555 if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1556 MPASS(next->object == m->object);
1557 if (next->pindex != m->pindex + 1)
1564 * Returns the given page's predecessor (by pindex) within the object if it is
1565 * resident; if none is found, NULL is returned.
1567 * The object must be locked.
1570 vm_page_prev(vm_page_t m)
1574 VM_OBJECT_ASSERT_LOCKED(m->object);
1575 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1576 MPASS(prev->object == m->object);
1577 if (prev->pindex != m->pindex - 1)
1584 * Uses the page mnew as a replacement for an existing page at index
1585 * pindex which must be already present in the object.
1587 * The existing page must not be on a paging queue.
1590 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1594 VM_OBJECT_ASSERT_WLOCKED(object);
1595 KASSERT(mnew->object == NULL,
1596 ("vm_page_replace: page %p already in object", mnew));
1597 KASSERT(mnew->queue == PQ_NONE,
1598 ("vm_page_replace: new page %p is on a paging queue", mnew));
1601 * This function mostly follows vm_page_insert() and
1602 * vm_page_remove() without the radix, object count and vnode
1603 * dance. Double check such functions for more comments.
1606 mnew->object = object;
1607 mnew->pindex = pindex;
1608 mold = vm_radix_replace(&object->rtree, mnew);
1609 KASSERT(mold->queue == PQ_NONE,
1610 ("vm_page_replace: old page %p is on a paging queue", mold));
1612 /* Keep the resident page list in sorted order. */
1613 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1614 TAILQ_REMOVE(&object->memq, mold, listq);
1616 mold->object = NULL;
1617 vm_page_xunbusy_maybelocked(mold);
1620 * The object's resident_page_count does not change because we have
1621 * swapped one page for another, but OBJ_MIGHTBEDIRTY.
1623 if (pmap_page_is_write_mapped(mnew))
1624 vm_object_set_writeable_dirty(object);
1631 * Move the given memory entry from its
1632 * current object to the specified target object/offset.
1634 * Note: swap associated with the page must be invalidated by the move. We
1635 * have to do this for several reasons: (1) we aren't freeing the
1636 * page, (2) we are dirtying the page, (3) the VM system is probably
1637 * moving the page from object A to B, and will then later move
1638 * the backing store from A to B and we can't have a conflict.
1640 * Note: we *always* dirty the page. It is necessary both for the
1641 * fact that we moved it, and because we may be invalidating
1644 * The objects must be locked.
1647 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1652 VM_OBJECT_ASSERT_WLOCKED(new_object);
1654 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1655 KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1656 ("vm_page_rename: pindex already renamed"));
1659 * Create a custom version of vm_page_insert() which does not depend
1660 * by m_prev and can cheat on the implementation aspects of the
1664 m->pindex = new_pindex;
1665 if (vm_radix_insert(&new_object->rtree, m)) {
1671 * The operation cannot fail anymore. The removal must happen before
1672 * the listq iterator is tainted.
1676 (void)vm_page_remove(m);
1678 /* Return back to the new pindex to complete vm_page_insert(). */
1679 m->pindex = new_pindex;
1680 m->object = new_object;
1682 vm_page_insert_radixdone(m, new_object, mpred);
1690 * Allocate and return a page that is associated with the specified
1691 * object and offset pair. By default, this page is exclusive busied.
1693 * The caller must always specify an allocation class.
1695 * allocation classes:
1696 * VM_ALLOC_NORMAL normal process request
1697 * VM_ALLOC_SYSTEM system *really* needs a page
1698 * VM_ALLOC_INTERRUPT interrupt time request
1700 * optional allocation flags:
1701 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1702 * intends to allocate
1703 * VM_ALLOC_NOBUSY do not exclusive busy the page
1704 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1705 * VM_ALLOC_NOOBJ page is not associated with an object and
1706 * should not be exclusive busy
1707 * VM_ALLOC_SBUSY shared busy the allocated page
1708 * VM_ALLOC_WIRED wire the allocated page
1709 * VM_ALLOC_ZERO prefer a zeroed page
1712 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1715 return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1716 vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1720 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1724 return (vm_page_alloc_domain_after(object, pindex, domain, req,
1725 object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) :
1730 * Allocate a page in the specified object with the given page index. To
1731 * optimize insertion of the page into the object, the caller must also specifiy
1732 * the resident page in the object with largest index smaller than the given
1733 * page index, or NULL if no such page exists.
1736 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1737 int req, vm_page_t mpred)
1739 struct vm_domainset_iter di;
1743 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1745 m = vm_page_alloc_domain_after(object, pindex, domain, req,
1749 } while (vm_domainset_iter_page(&di, object, &domain) == 0);
1755 * Returns true if the number of free pages exceeds the minimum
1756 * for the request class and false otherwise.
1759 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
1761 u_int limit, old, new;
1763 if (req_class == VM_ALLOC_INTERRUPT)
1765 else if (req_class == VM_ALLOC_SYSTEM)
1766 limit = vmd->vmd_interrupt_free_min;
1768 limit = vmd->vmd_free_reserved;
1771 * Attempt to reserve the pages. Fail if we're below the limit.
1774 old = vmd->vmd_free_count;
1779 } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
1781 /* Wake the page daemon if we've crossed the threshold. */
1782 if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
1783 pagedaemon_wakeup(vmd->vmd_domain);
1785 /* Only update bitsets on transitions. */
1786 if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
1787 (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
1794 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
1799 * The page daemon is allowed to dig deeper into the free page list.
1801 req_class = req & VM_ALLOC_CLASS_MASK;
1802 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1803 req_class = VM_ALLOC_SYSTEM;
1804 return (_vm_domain_allocate(vmd, req_class, npages));
1808 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
1809 int req, vm_page_t mpred)
1811 struct vm_domain *vmd;
1815 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1816 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1817 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1818 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1819 ("inconsistent object(%p)/req(%x)", object, req));
1820 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
1821 ("Can't sleep and retry object insertion."));
1822 KASSERT(mpred == NULL || mpred->pindex < pindex,
1823 ("mpred %p doesn't precede pindex 0x%jx", mpred,
1824 (uintmax_t)pindex));
1826 VM_OBJECT_ASSERT_WLOCKED(object);
1830 pool = object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT;
1832 #if VM_NRESERVLEVEL > 0
1834 * Can we allocate the page from a reservation?
1836 if (vm_object_reserv(object) &&
1837 ((m = vm_reserv_extend(req, object, pindex, domain, mpred)) != NULL ||
1838 (m = vm_reserv_alloc_page(req, object, pindex, domain, mpred)) != NULL)) {
1839 domain = vm_phys_domain(m);
1840 vmd = VM_DOMAIN(domain);
1844 vmd = VM_DOMAIN(domain);
1845 if (vmd->vmd_pgcache[pool].zone != NULL) {
1846 m = uma_zalloc(vmd->vmd_pgcache[pool].zone, M_NOWAIT);
1848 flags |= PG_PCPU_CACHE;
1852 if (vm_domain_allocate(vmd, req, 1)) {
1854 * If not, allocate it from the free page queues.
1856 vm_domain_free_lock(vmd);
1857 m = vm_phys_alloc_pages(domain, pool, 0);
1858 vm_domain_free_unlock(vmd);
1860 vm_domain_freecnt_inc(vmd, 1);
1861 #if VM_NRESERVLEVEL > 0
1862 if (vm_reserv_reclaim_inactive(domain))
1869 * Not allocatable, give up.
1871 if (vm_domain_alloc_fail(vmd, object, req))
1877 * At this point we had better have found a good page.
1881 vm_page_alloc_check(m);
1884 * Initialize the page. Only the PG_ZERO flag is inherited.
1886 if ((req & VM_ALLOC_ZERO) != 0)
1887 flags |= (m->flags & PG_ZERO);
1888 if ((req & VM_ALLOC_NODUMP) != 0)
1892 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1894 m->busy_lock = VPB_UNBUSIED;
1895 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1896 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1897 if ((req & VM_ALLOC_SBUSY) != 0)
1898 m->busy_lock = VPB_SHARERS_WORD(1);
1899 if (req & VM_ALLOC_WIRED) {
1901 * The page lock is not required for wiring a page until that
1902 * page is inserted into the object.
1909 if (object != NULL) {
1910 if (vm_page_insert_after(m, object, pindex, mpred)) {
1911 if (req & VM_ALLOC_WIRED) {
1915 KASSERT(m->object == NULL, ("page %p has object", m));
1916 m->oflags = VPO_UNMANAGED;
1917 m->busy_lock = VPB_UNBUSIED;
1918 /* Don't change PG_ZERO. */
1919 vm_page_free_toq(m);
1920 if (req & VM_ALLOC_WAITFAIL) {
1921 VM_OBJECT_WUNLOCK(object);
1923 VM_OBJECT_WLOCK(object);
1928 /* Ignore device objects; the pager sets "memattr" for them. */
1929 if (object->memattr != VM_MEMATTR_DEFAULT &&
1930 (object->flags & OBJ_FICTITIOUS) == 0)
1931 pmap_page_set_memattr(m, object->memattr);
1939 * vm_page_alloc_contig:
1941 * Allocate a contiguous set of physical pages of the given size "npages"
1942 * from the free lists. All of the physical pages must be at or above
1943 * the given physical address "low" and below the given physical address
1944 * "high". The given value "alignment" determines the alignment of the
1945 * first physical page in the set. If the given value "boundary" is
1946 * non-zero, then the set of physical pages cannot cross any physical
1947 * address boundary that is a multiple of that value. Both "alignment"
1948 * and "boundary" must be a power of two.
1950 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1951 * then the memory attribute setting for the physical pages is configured
1952 * to the object's memory attribute setting. Otherwise, the memory
1953 * attribute setting for the physical pages is configured to "memattr",
1954 * overriding the object's memory attribute setting. However, if the
1955 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1956 * memory attribute setting for the physical pages cannot be configured
1957 * to VM_MEMATTR_DEFAULT.
1959 * The specified object may not contain fictitious pages.
1961 * The caller must always specify an allocation class.
1963 * allocation classes:
1964 * VM_ALLOC_NORMAL normal process request
1965 * VM_ALLOC_SYSTEM system *really* needs a page
1966 * VM_ALLOC_INTERRUPT interrupt time request
1968 * optional allocation flags:
1969 * VM_ALLOC_NOBUSY do not exclusive busy the page
1970 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1971 * VM_ALLOC_NOOBJ page is not associated with an object and
1972 * should not be exclusive busy
1973 * VM_ALLOC_SBUSY shared busy the allocated page
1974 * VM_ALLOC_WIRED wire the allocated page
1975 * VM_ALLOC_ZERO prefer a zeroed page
1978 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1979 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1980 vm_paddr_t boundary, vm_memattr_t memattr)
1982 struct vm_domainset_iter di;
1986 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1988 m = vm_page_alloc_contig_domain(object, pindex, domain, req,
1989 npages, low, high, alignment, boundary, memattr);
1992 } while (vm_domainset_iter_page(&di, object, &domain) == 0);
1998 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1999 int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2000 vm_paddr_t boundary, vm_memattr_t memattr)
2002 struct vm_domain *vmd;
2003 vm_page_t m, m_ret, mpred;
2004 u_int busy_lock, flags, oflags;
2006 mpred = NULL; /* XXX: pacify gcc */
2007 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
2008 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
2009 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2010 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2011 ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
2013 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2014 ("Can't sleep and retry object insertion."));
2015 if (object != NULL) {
2016 VM_OBJECT_ASSERT_WLOCKED(object);
2017 KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2018 ("vm_page_alloc_contig: object %p has fictitious pages",
2021 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2023 if (object != NULL) {
2024 mpred = vm_radix_lookup_le(&object->rtree, pindex);
2025 KASSERT(mpred == NULL || mpred->pindex != pindex,
2026 ("vm_page_alloc_contig: pindex already allocated"));
2030 * Can we allocate the pages without the number of free pages falling
2031 * below the lower bound for the allocation class?
2035 #if VM_NRESERVLEVEL > 0
2037 * Can we allocate the pages from a reservation?
2039 if (vm_object_reserv(object) &&
2040 ((m_ret = vm_reserv_extend_contig(req, object, pindex, domain,
2041 npages, low, high, alignment, boundary, mpred)) != NULL ||
2042 (m_ret = vm_reserv_alloc_contig(req, object, pindex, domain,
2043 npages, low, high, alignment, boundary, mpred)) != NULL)) {
2044 domain = vm_phys_domain(m_ret);
2045 vmd = VM_DOMAIN(domain);
2049 vmd = VM_DOMAIN(domain);
2050 if (vm_domain_allocate(vmd, req, npages)) {
2052 * allocate them from the free page queues.
2054 vm_domain_free_lock(vmd);
2055 m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2056 alignment, boundary);
2057 vm_domain_free_unlock(vmd);
2058 if (m_ret == NULL) {
2059 vm_domain_freecnt_inc(vmd, npages);
2060 #if VM_NRESERVLEVEL > 0
2061 if (vm_reserv_reclaim_contig(domain, npages, low,
2062 high, alignment, boundary))
2067 if (m_ret == NULL) {
2068 if (vm_domain_alloc_fail(vmd, object, req))
2072 #if VM_NRESERVLEVEL > 0
2075 for (m = m_ret; m < &m_ret[npages]; m++) {
2077 vm_page_alloc_check(m);
2081 * Initialize the pages. Only the PG_ZERO flag is inherited.
2084 if ((req & VM_ALLOC_ZERO) != 0)
2086 if ((req & VM_ALLOC_NODUMP) != 0)
2088 oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2090 busy_lock = VPB_UNBUSIED;
2091 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2092 busy_lock = VPB_SINGLE_EXCLUSIVER;
2093 if ((req & VM_ALLOC_SBUSY) != 0)
2094 busy_lock = VPB_SHARERS_WORD(1);
2095 if ((req & VM_ALLOC_WIRED) != 0)
2096 vm_wire_add(npages);
2097 if (object != NULL) {
2098 if (object->memattr != VM_MEMATTR_DEFAULT &&
2099 memattr == VM_MEMATTR_DEFAULT)
2100 memattr = object->memattr;
2102 for (m = m_ret; m < &m_ret[npages]; m++) {
2104 m->flags = (m->flags | PG_NODUMP) & flags;
2105 m->busy_lock = busy_lock;
2106 if ((req & VM_ALLOC_WIRED) != 0)
2110 if (object != NULL) {
2111 if (vm_page_insert_after(m, object, pindex, mpred)) {
2112 if ((req & VM_ALLOC_WIRED) != 0)
2113 vm_wire_sub(npages);
2114 KASSERT(m->object == NULL,
2115 ("page %p has object", m));
2117 for (m = m_ret; m < &m_ret[npages]; m++) {
2119 (req & VM_ALLOC_WIRED) != 0)
2121 m->oflags = VPO_UNMANAGED;
2122 m->busy_lock = VPB_UNBUSIED;
2123 /* Don't change PG_ZERO. */
2124 vm_page_free_toq(m);
2126 if (req & VM_ALLOC_WAITFAIL) {
2127 VM_OBJECT_WUNLOCK(object);
2129 VM_OBJECT_WLOCK(object);
2136 if (memattr != VM_MEMATTR_DEFAULT)
2137 pmap_page_set_memattr(m, memattr);
2144 * Check a page that has been freshly dequeued from a freelist.
2147 vm_page_alloc_check(vm_page_t m)
2150 KASSERT(m->object == NULL, ("page %p has object", m));
2151 KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0,
2152 ("page %p has unexpected queue %d, flags %#x",
2153 m, m->queue, (m->aflags & PGA_QUEUE_STATE_MASK)));
2154 KASSERT(!vm_page_held(m), ("page %p is held", m));
2155 KASSERT(!vm_page_busied(m), ("page %p is busy", m));
2156 KASSERT(m->dirty == 0, ("page %p is dirty", m));
2157 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2158 ("page %p has unexpected memattr %d",
2159 m, pmap_page_get_memattr(m)));
2160 KASSERT(m->valid == 0, ("free page %p is valid", m));
2164 * vm_page_alloc_freelist:
2166 * Allocate a physical page from the specified free page list.
2168 * The caller must always specify an allocation class.
2170 * allocation classes:
2171 * VM_ALLOC_NORMAL normal process request
2172 * VM_ALLOC_SYSTEM system *really* needs a page
2173 * VM_ALLOC_INTERRUPT interrupt time request
2175 * optional allocation flags:
2176 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
2177 * intends to allocate
2178 * VM_ALLOC_WIRED wire the allocated page
2179 * VM_ALLOC_ZERO prefer a zeroed page
2182 vm_page_alloc_freelist(int freelist, int req)
2184 struct vm_domainset_iter di;
2188 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2190 m = vm_page_alloc_freelist_domain(domain, freelist, req);
2193 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2199 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2201 struct vm_domain *vmd;
2206 vmd = VM_DOMAIN(domain);
2208 if (vm_domain_allocate(vmd, req, 1)) {
2209 vm_domain_free_lock(vmd);
2210 m = vm_phys_alloc_freelist_pages(domain, freelist,
2211 VM_FREEPOOL_DIRECT, 0);
2212 vm_domain_free_unlock(vmd);
2214 vm_domain_freecnt_inc(vmd, 1);
2217 if (vm_domain_alloc_fail(vmd, NULL, req))
2222 vm_page_alloc_check(m);
2225 * Initialize the page. Only the PG_ZERO flag is inherited.
2229 if ((req & VM_ALLOC_ZERO) != 0)
2232 if ((req & VM_ALLOC_WIRED) != 0) {
2234 * The page lock is not required for wiring a page that does
2235 * not belong to an object.
2240 /* Unmanaged pages don't use "act_count". */
2241 m->oflags = VPO_UNMANAGED;
2246 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2248 struct vm_domain *vmd;
2249 struct vm_pgcache *pgcache;
2253 vmd = VM_DOMAIN(pgcache->domain);
2256 * The page daemon should avoid creating extra memory pressure since its
2257 * main purpose is to replenish the store of free pages.
2259 if (vmd->vmd_severeset || curproc == pageproc ||
2260 !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2262 domain = vmd->vmd_domain;
2263 vm_domain_free_lock(vmd);
2264 i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2265 (vm_page_t *)store);
2266 vm_domain_free_unlock(vmd);
2268 vm_domain_freecnt_inc(vmd, cnt - i);
2274 vm_page_zone_release(void *arg, void **store, int cnt)
2276 struct vm_domain *vmd;
2277 struct vm_pgcache *pgcache;
2282 vmd = VM_DOMAIN(pgcache->domain);
2283 vm_domain_free_lock(vmd);
2284 for (i = 0; i < cnt; i++) {
2285 m = (vm_page_t)store[i];
2286 vm_phys_free_pages(m, 0);
2288 vm_domain_free_unlock(vmd);
2289 vm_domain_freecnt_inc(vmd, cnt);
2292 #define VPSC_ANY 0 /* No restrictions. */
2293 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */
2294 #define VPSC_NOSUPER 2 /* Skip superpages. */
2297 * vm_page_scan_contig:
2299 * Scan vm_page_array[] between the specified entries "m_start" and
2300 * "m_end" for a run of contiguous physical pages that satisfy the
2301 * specified conditions, and return the lowest page in the run. The
2302 * specified "alignment" determines the alignment of the lowest physical
2303 * page in the run. If the specified "boundary" is non-zero, then the
2304 * run of physical pages cannot span a physical address that is a
2305 * multiple of "boundary".
2307 * "m_end" is never dereferenced, so it need not point to a vm_page
2308 * structure within vm_page_array[].
2310 * "npages" must be greater than zero. "m_start" and "m_end" must not
2311 * span a hole (or discontiguity) in the physical address space. Both
2312 * "alignment" and "boundary" must be a power of two.
2315 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2316 u_long alignment, vm_paddr_t boundary, int options)
2322 #if VM_NRESERVLEVEL > 0
2325 int m_inc, order, run_ext, run_len;
2327 KASSERT(npages > 0, ("npages is 0"));
2328 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2329 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2333 for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2334 KASSERT((m->flags & PG_MARKER) == 0,
2335 ("page %p is PG_MARKER", m));
2336 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->wire_count == 1,
2337 ("fictitious page %p has invalid wire count", m));
2340 * If the current page would be the start of a run, check its
2341 * physical address against the end, alignment, and boundary
2342 * conditions. If it doesn't satisfy these conditions, either
2343 * terminate the scan or advance to the next page that
2344 * satisfies the failed condition.
2347 KASSERT(m_run == NULL, ("m_run != NULL"));
2348 if (m + npages > m_end)
2350 pa = VM_PAGE_TO_PHYS(m);
2351 if ((pa & (alignment - 1)) != 0) {
2352 m_inc = atop(roundup2(pa, alignment) - pa);
2355 if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2357 m_inc = atop(roundup2(pa, boundary) - pa);
2361 KASSERT(m_run != NULL, ("m_run == NULL"));
2363 vm_page_change_lock(m, &m_mtx);
2366 if (vm_page_held(m))
2368 #if VM_NRESERVLEVEL > 0
2369 else if ((level = vm_reserv_level(m)) >= 0 &&
2370 (options & VPSC_NORESERV) != 0) {
2372 /* Advance to the end of the reservation. */
2373 pa = VM_PAGE_TO_PHYS(m);
2374 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2378 else if ((object = m->object) != NULL) {
2380 * The page is considered eligible for relocation if
2381 * and only if it could be laundered or reclaimed by
2384 if (!VM_OBJECT_TRYRLOCK(object)) {
2386 VM_OBJECT_RLOCK(object);
2388 if (m->object != object) {
2390 * The page may have been freed.
2392 VM_OBJECT_RUNLOCK(object);
2394 } else if (vm_page_held(m)) {
2399 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2400 ("page %p is PG_UNHOLDFREE", m));
2401 /* Don't care: PG_NODUMP, PG_ZERO. */
2402 if (object->type != OBJT_DEFAULT &&
2403 object->type != OBJT_SWAP &&
2404 object->type != OBJT_VNODE) {
2406 #if VM_NRESERVLEVEL > 0
2407 } else if ((options & VPSC_NOSUPER) != 0 &&
2408 (level = vm_reserv_level_iffullpop(m)) >= 0) {
2410 /* Advance to the end of the superpage. */
2411 pa = VM_PAGE_TO_PHYS(m);
2412 m_inc = atop(roundup2(pa + 1,
2413 vm_reserv_size(level)) - pa);
2415 } else if (object->memattr == VM_MEMATTR_DEFAULT &&
2416 vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) {
2418 * The page is allocated but eligible for
2419 * relocation. Extend the current run by one
2422 KASSERT(pmap_page_get_memattr(m) ==
2424 ("page %p has an unexpected memattr", m));
2425 KASSERT((m->oflags & (VPO_SWAPINPROG |
2426 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2427 ("page %p has unexpected oflags", m));
2428 /* Don't care: VPO_NOSYNC. */
2433 VM_OBJECT_RUNLOCK(object);
2434 #if VM_NRESERVLEVEL > 0
2435 } else if (level >= 0) {
2437 * The page is reserved but not yet allocated. In
2438 * other words, it is still free. Extend the current
2443 } else if ((order = m->order) < VM_NFREEORDER) {
2445 * The page is enqueued in the physical memory
2446 * allocator's free page queues. Moreover, it is the
2447 * first page in a power-of-two-sized run of
2448 * contiguous free pages. Add these pages to the end
2449 * of the current run, and jump ahead.
2451 run_ext = 1 << order;
2455 * Skip the page for one of the following reasons: (1)
2456 * It is enqueued in the physical memory allocator's
2457 * free page queues. However, it is not the first
2458 * page in a run of contiguous free pages. (This case
2459 * rarely occurs because the scan is performed in
2460 * ascending order.) (2) It is not reserved, and it is
2461 * transitioning from free to allocated. (Conversely,
2462 * the transition from allocated to free for managed
2463 * pages is blocked by the page lock.) (3) It is
2464 * allocated but not contained by an object and not
2465 * wired, e.g., allocated by Xen's balloon driver.
2471 * Extend or reset the current run of pages.
2486 if (run_len >= npages)
2492 * vm_page_reclaim_run:
2494 * Try to relocate each of the allocated virtual pages within the
2495 * specified run of physical pages to a new physical address. Free the
2496 * physical pages underlying the relocated virtual pages. A virtual page
2497 * is relocatable if and only if it could be laundered or reclaimed by
2498 * the page daemon. Whenever possible, a virtual page is relocated to a
2499 * physical address above "high".
2501 * Returns 0 if every physical page within the run was already free or
2502 * just freed by a successful relocation. Otherwise, returns a non-zero
2503 * value indicating why the last attempt to relocate a virtual page was
2506 * "req_class" must be an allocation class.
2509 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2512 struct vm_domain *vmd;
2514 struct spglist free;
2517 vm_page_t m, m_end, m_new;
2518 int error, order, req;
2520 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2521 ("req_class is not an allocation class"));
2525 m_end = m_run + npages;
2527 for (; error == 0 && m < m_end; m++) {
2528 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2529 ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2532 * Avoid releasing and reacquiring the same page lock.
2534 vm_page_change_lock(m, &m_mtx);
2536 if (vm_page_held(m))
2538 else if ((object = m->object) != NULL) {
2540 * The page is relocated if and only if it could be
2541 * laundered or reclaimed by the page daemon.
2543 if (!VM_OBJECT_TRYWLOCK(object)) {
2545 VM_OBJECT_WLOCK(object);
2547 if (m->object != object) {
2549 * The page may have been freed.
2551 VM_OBJECT_WUNLOCK(object);
2553 } else if (vm_page_held(m)) {
2558 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2559 ("page %p is PG_UNHOLDFREE", m));
2560 /* Don't care: PG_NODUMP, PG_ZERO. */
2561 if (object->type != OBJT_DEFAULT &&
2562 object->type != OBJT_SWAP &&
2563 object->type != OBJT_VNODE)
2565 else if (object->memattr != VM_MEMATTR_DEFAULT)
2567 else if (vm_page_queue(m) != PQ_NONE &&
2568 !vm_page_busied(m)) {
2569 KASSERT(pmap_page_get_memattr(m) ==
2571 ("page %p has an unexpected memattr", m));
2572 KASSERT((m->oflags & (VPO_SWAPINPROG |
2573 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2574 ("page %p has unexpected oflags", m));
2575 /* Don't care: VPO_NOSYNC. */
2576 if (m->valid != 0) {
2578 * First, try to allocate a new page
2579 * that is above "high". Failing
2580 * that, try to allocate a new page
2581 * that is below "m_run". Allocate
2582 * the new page between the end of
2583 * "m_run" and "high" only as a last
2586 req = req_class | VM_ALLOC_NOOBJ;
2587 if ((m->flags & PG_NODUMP) != 0)
2588 req |= VM_ALLOC_NODUMP;
2589 if (trunc_page(high) !=
2590 ~(vm_paddr_t)PAGE_MASK) {
2591 m_new = vm_page_alloc_contig(
2596 VM_MEMATTR_DEFAULT);
2599 if (m_new == NULL) {
2600 pa = VM_PAGE_TO_PHYS(m_run);
2601 m_new = vm_page_alloc_contig(
2603 0, pa - 1, PAGE_SIZE, 0,
2604 VM_MEMATTR_DEFAULT);
2606 if (m_new == NULL) {
2608 m_new = vm_page_alloc_contig(
2610 pa, high, PAGE_SIZE, 0,
2611 VM_MEMATTR_DEFAULT);
2613 if (m_new == NULL) {
2617 KASSERT(!vm_page_wired(m_new),
2618 ("page %p is wired", m_new));
2621 * Replace "m" with the new page. For
2622 * vm_page_replace(), "m" must be busy
2623 * and dequeued. Finally, change "m"
2624 * as if vm_page_free() was called.
2626 if (object->ref_count != 0)
2628 m_new->aflags = m->aflags &
2629 ~PGA_QUEUE_STATE_MASK;
2630 KASSERT(m_new->oflags == VPO_UNMANAGED,
2631 ("page %p is managed", m_new));
2632 m_new->oflags = m->oflags & VPO_NOSYNC;
2633 pmap_copy_page(m, m_new);
2634 m_new->valid = m->valid;
2635 m_new->dirty = m->dirty;
2636 m->flags &= ~PG_ZERO;
2639 vm_page_replace_checked(m_new, object,
2641 if (vm_page_free_prep(m))
2642 SLIST_INSERT_HEAD(&free, m,
2646 * The new page must be deactivated
2647 * before the object is unlocked.
2649 vm_page_change_lock(m_new, &m_mtx);
2650 vm_page_deactivate(m_new);
2652 m->flags &= ~PG_ZERO;
2654 if (vm_page_free_prep(m))
2655 SLIST_INSERT_HEAD(&free, m,
2657 KASSERT(m->dirty == 0,
2658 ("page %p is dirty", m));
2663 VM_OBJECT_WUNLOCK(object);
2665 MPASS(vm_phys_domain(m) == domain);
2666 vmd = VM_DOMAIN(domain);
2667 vm_domain_free_lock(vmd);
2669 if (order < VM_NFREEORDER) {
2671 * The page is enqueued in the physical memory
2672 * allocator's free page queues. Moreover, it
2673 * is the first page in a power-of-two-sized
2674 * run of contiguous free pages. Jump ahead
2675 * to the last page within that run, and
2676 * continue from there.
2678 m += (1 << order) - 1;
2680 #if VM_NRESERVLEVEL > 0
2681 else if (vm_reserv_is_page_free(m))
2684 vm_domain_free_unlock(vmd);
2685 if (order == VM_NFREEORDER)
2691 if ((m = SLIST_FIRST(&free)) != NULL) {
2694 vmd = VM_DOMAIN(domain);
2696 vm_domain_free_lock(vmd);
2698 MPASS(vm_phys_domain(m) == domain);
2699 SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2700 vm_phys_free_pages(m, 0);
2702 } while ((m = SLIST_FIRST(&free)) != NULL);
2703 vm_domain_free_unlock(vmd);
2704 vm_domain_freecnt_inc(vmd, cnt);
2711 CTASSERT(powerof2(NRUNS));
2713 #define RUN_INDEX(count) ((count) & (NRUNS - 1))
2715 #define MIN_RECLAIM 8
2718 * vm_page_reclaim_contig:
2720 * Reclaim allocated, contiguous physical memory satisfying the specified
2721 * conditions by relocating the virtual pages using that physical memory.
2722 * Returns true if reclamation is successful and false otherwise. Since
2723 * relocation requires the allocation of physical pages, reclamation may
2724 * fail due to a shortage of free pages. When reclamation fails, callers
2725 * are expected to perform vm_wait() before retrying a failed allocation
2726 * operation, e.g., vm_page_alloc_contig().
2728 * The caller must always specify an allocation class through "req".
2730 * allocation classes:
2731 * VM_ALLOC_NORMAL normal process request
2732 * VM_ALLOC_SYSTEM system *really* needs a page
2733 * VM_ALLOC_INTERRUPT interrupt time request
2735 * The optional allocation flags are ignored.
2737 * "npages" must be greater than zero. Both "alignment" and "boundary"
2738 * must be a power of two.
2741 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
2742 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2744 struct vm_domain *vmd;
2745 vm_paddr_t curr_low;
2746 vm_page_t m_run, m_runs[NRUNS];
2747 u_long count, reclaimed;
2748 int error, i, options, req_class;
2750 KASSERT(npages > 0, ("npages is 0"));
2751 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2752 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2753 req_class = req & VM_ALLOC_CLASS_MASK;
2756 * The page daemon is allowed to dig deeper into the free page list.
2758 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2759 req_class = VM_ALLOC_SYSTEM;
2762 * Return if the number of free pages cannot satisfy the requested
2765 vmd = VM_DOMAIN(domain);
2766 count = vmd->vmd_free_count;
2767 if (count < npages + vmd->vmd_free_reserved || (count < npages +
2768 vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2769 (count < npages && req_class == VM_ALLOC_INTERRUPT))
2773 * Scan up to three times, relaxing the restrictions ("options") on
2774 * the reclamation of reservations and superpages each time.
2776 for (options = VPSC_NORESERV;;) {
2778 * Find the highest runs that satisfy the given constraints
2779 * and restrictions, and record them in "m_runs".
2784 m_run = vm_phys_scan_contig(domain, npages, curr_low,
2785 high, alignment, boundary, options);
2788 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2789 m_runs[RUN_INDEX(count)] = m_run;
2794 * Reclaim the highest runs in LIFO (descending) order until
2795 * the number of reclaimed pages, "reclaimed", is at least
2796 * MIN_RECLAIM. Reset "reclaimed" each time because each
2797 * reclamation is idempotent, and runs will (likely) recur
2798 * from one scan to the next as restrictions are relaxed.
2801 for (i = 0; count > 0 && i < NRUNS; i++) {
2803 m_run = m_runs[RUN_INDEX(count)];
2804 error = vm_page_reclaim_run(req_class, domain, npages,
2807 reclaimed += npages;
2808 if (reclaimed >= MIN_RECLAIM)
2814 * Either relax the restrictions on the next scan or return if
2815 * the last scan had no restrictions.
2817 if (options == VPSC_NORESERV)
2818 options = VPSC_NOSUPER;
2819 else if (options == VPSC_NOSUPER)
2821 else if (options == VPSC_ANY)
2822 return (reclaimed != 0);
2827 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
2828 u_long alignment, vm_paddr_t boundary)
2830 struct vm_domainset_iter di;
2834 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2836 ret = vm_page_reclaim_contig_domain(domain, req, npages, low,
2837 high, alignment, boundary);
2840 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2846 * Set the domain in the appropriate page level domainset.
2849 vm_domain_set(struct vm_domain *vmd)
2852 mtx_lock(&vm_domainset_lock);
2853 if (!vmd->vmd_minset && vm_paging_min(vmd)) {
2854 vmd->vmd_minset = 1;
2855 DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
2857 if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
2858 vmd->vmd_severeset = 1;
2859 DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
2861 mtx_unlock(&vm_domainset_lock);
2865 * Clear the domain from the appropriate page level domainset.
2868 vm_domain_clear(struct vm_domain *vmd)
2871 mtx_lock(&vm_domainset_lock);
2872 if (vmd->vmd_minset && !vm_paging_min(vmd)) {
2873 vmd->vmd_minset = 0;
2874 DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
2875 if (vm_min_waiters != 0) {
2877 wakeup(&vm_min_domains);
2880 if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
2881 vmd->vmd_severeset = 0;
2882 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
2883 if (vm_severe_waiters != 0) {
2884 vm_severe_waiters = 0;
2885 wakeup(&vm_severe_domains);
2890 * If pageout daemon needs pages, then tell it that there are
2893 if (vmd->vmd_pageout_pages_needed &&
2894 vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
2895 wakeup(&vmd->vmd_pageout_pages_needed);
2896 vmd->vmd_pageout_pages_needed = 0;
2899 /* See comments in vm_wait_doms(). */
2900 if (vm_pageproc_waiters) {
2901 vm_pageproc_waiters = 0;
2902 wakeup(&vm_pageproc_waiters);
2904 mtx_unlock(&vm_domainset_lock);
2908 * Wait for free pages to exceed the min threshold globally.
2914 mtx_lock(&vm_domainset_lock);
2915 while (vm_page_count_min()) {
2917 msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
2919 mtx_unlock(&vm_domainset_lock);
2923 * Wait for free pages to exceed the severe threshold globally.
2926 vm_wait_severe(void)
2929 mtx_lock(&vm_domainset_lock);
2930 while (vm_page_count_severe()) {
2931 vm_severe_waiters++;
2932 msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
2935 mtx_unlock(&vm_domainset_lock);
2942 return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
2946 vm_wait_doms(const domainset_t *wdoms)
2950 * We use racey wakeup synchronization to avoid expensive global
2951 * locking for the pageproc when sleeping with a non-specific vm_wait.
2952 * To handle this, we only sleep for one tick in this instance. It
2953 * is expected that most allocations for the pageproc will come from
2954 * kmem or vm_page_grab* which will use the more specific and
2955 * race-free vm_wait_domain().
2957 if (curproc == pageproc) {
2958 mtx_lock(&vm_domainset_lock);
2959 vm_pageproc_waiters++;
2960 msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP,
2964 * XXX Ideally we would wait only until the allocation could
2965 * be satisfied. This condition can cause new allocators to
2966 * consume all freed pages while old allocators wait.
2968 mtx_lock(&vm_domainset_lock);
2969 if (vm_page_count_min_set(wdoms)) {
2971 msleep(&vm_min_domains, &vm_domainset_lock,
2972 PVM | PDROP, "vmwait", 0);
2974 mtx_unlock(&vm_domainset_lock);
2981 * Sleep until free pages are available for allocation.
2982 * - Called in various places after failed memory allocations.
2985 vm_wait_domain(int domain)
2987 struct vm_domain *vmd;
2990 vmd = VM_DOMAIN(domain);
2991 vm_domain_free_assert_unlocked(vmd);
2993 if (curproc == pageproc) {
2994 mtx_lock(&vm_domainset_lock);
2995 if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
2996 vmd->vmd_pageout_pages_needed = 1;
2997 msleep(&vmd->vmd_pageout_pages_needed,
2998 &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3000 mtx_unlock(&vm_domainset_lock);
3002 if (pageproc == NULL)
3003 panic("vm_wait in early boot");
3004 DOMAINSET_ZERO(&wdom);
3005 DOMAINSET_SET(vmd->vmd_domain, &wdom);
3006 vm_wait_doms(&wdom);
3013 * Sleep until free pages are available for allocation in the
3014 * affinity domains of the obj. If obj is NULL, the domain set
3015 * for the calling thread is used.
3016 * Called in various places after failed memory allocations.
3019 vm_wait(vm_object_t obj)
3021 struct domainset *d;
3026 * Carefully fetch pointers only once: the struct domainset
3027 * itself is ummutable but the pointer might change.
3030 d = obj->domain.dr_policy;
3032 d = curthread->td_domain.dr_policy;
3034 vm_wait_doms(&d->ds_mask);
3038 * vm_domain_alloc_fail:
3040 * Called when a page allocation function fails. Informs the
3041 * pagedaemon and performs the requested wait. Requires the
3042 * domain_free and object lock on entry. Returns with the
3043 * object lock held and free lock released. Returns an error when
3044 * retry is necessary.
3048 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3051 vm_domain_free_assert_unlocked(vmd);
3053 atomic_add_int(&vmd->vmd_pageout_deficit,
3054 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3055 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3057 VM_OBJECT_WUNLOCK(object);
3058 vm_wait_domain(vmd->vmd_domain);
3060 VM_OBJECT_WLOCK(object);
3061 if (req & VM_ALLOC_WAITOK)
3071 * Sleep until free pages are available for allocation.
3072 * - Called only in vm_fault so that processes page faulting
3073 * can be easily tracked.
3074 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3075 * processes will be able to grab memory first. Do not change
3076 * this balance without careful testing first.
3079 vm_waitpfault(struct domainset *dset, int timo)
3083 * XXX Ideally we would wait only until the allocation could
3084 * be satisfied. This condition can cause new allocators to
3085 * consume all freed pages while old allocators wait.
3087 mtx_lock(&vm_domainset_lock);
3088 if (vm_page_count_min_set(&dset->ds_mask)) {
3090 msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3093 mtx_unlock(&vm_domainset_lock);
3096 static struct vm_pagequeue *
3097 vm_page_pagequeue(vm_page_t m)
3102 if ((queue = atomic_load_8(&m->queue)) == PQ_NONE)
3104 return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3108 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m)
3110 struct vm_domain *vmd;
3113 CRITICAL_ASSERT(curthread);
3114 vm_pagequeue_assert_locked(pq);
3117 * The page daemon is allowed to set m->queue = PQ_NONE without
3118 * the page queue lock held. In this case it is about to free the page,
3119 * which must not have any queue state.
3121 qflags = atomic_load_8(&m->aflags);
3122 KASSERT(pq == vm_page_pagequeue(m) ||
3123 (qflags & PGA_QUEUE_STATE_MASK) == 0,
3124 ("page %p doesn't belong to queue %p but has aflags %#x",
3127 if ((qflags & PGA_DEQUEUE) != 0) {
3128 if (__predict_true((qflags & PGA_ENQUEUED) != 0))
3129 vm_pagequeue_remove(pq, m);
3130 vm_page_dequeue_complete(m);
3131 } else if ((qflags & (PGA_REQUEUE | PGA_REQUEUE_HEAD)) != 0) {
3132 if ((qflags & PGA_ENQUEUED) != 0)
3133 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3135 vm_pagequeue_cnt_inc(pq);
3136 vm_page_aflag_set(m, PGA_ENQUEUED);
3140 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE.
3141 * In particular, if both flags are set in close succession,
3142 * only PGA_REQUEUE_HEAD will be applied, even if it was set
3145 if ((qflags & PGA_REQUEUE_HEAD) != 0) {
3146 KASSERT(m->queue == PQ_INACTIVE,
3147 ("head enqueue not supported for page %p", m));
3148 vmd = vm_pagequeue_domain(m);
3149 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3151 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3153 vm_page_aflag_clear(m, qflags & (PGA_REQUEUE |
3159 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3165 for (i = 0; i < bq->bq_cnt; i++) {
3167 if (__predict_false(m->queue != queue))
3169 vm_pqbatch_process_page(pq, m);
3171 vm_batchqueue_init(bq);
3175 vm_pqbatch_submit_page(vm_page_t m, uint8_t queue)
3177 struct vm_batchqueue *bq;
3178 struct vm_pagequeue *pq;
3181 vm_page_assert_locked(m);
3182 KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3184 domain = vm_phys_domain(m);
3185 pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue];
3188 bq = DPCPU_PTR(pqbatch[domain][queue]);
3189 if (vm_batchqueue_insert(bq, m)) {
3193 if (!vm_pagequeue_trylock(pq)) {
3195 vm_pagequeue_lock(pq);
3197 bq = DPCPU_PTR(pqbatch[domain][queue]);
3199 vm_pqbatch_process(pq, bq, queue);
3202 * The page may have been logically dequeued before we acquired the
3203 * page queue lock. In this case, the page lock prevents the page
3204 * from being logically enqueued elsewhere.
3206 if (__predict_true(m->queue == queue))
3207 vm_pqbatch_process_page(pq, m);
3209 KASSERT(m->queue == PQ_NONE,
3210 ("invalid queue transition for page %p", m));
3211 KASSERT((m->aflags & PGA_ENQUEUED) == 0,
3212 ("page %p is enqueued with invalid queue index", m));
3213 vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK);
3215 vm_pagequeue_unlock(pq);
3220 * vm_page_drain_pqbatch: [ internal use only ]
3222 * Force all per-CPU page queue batch queues to be drained. This is
3223 * intended for use in severe memory shortages, to ensure that pages
3224 * do not remain stuck in the batch queues.
3227 vm_page_drain_pqbatch(void)
3230 struct vm_domain *vmd;
3231 struct vm_pagequeue *pq;
3232 int cpu, domain, queue;
3237 sched_bind(td, cpu);
3240 for (domain = 0; domain < vm_ndomains; domain++) {
3241 vmd = VM_DOMAIN(domain);
3242 for (queue = 0; queue < PQ_COUNT; queue++) {
3243 pq = &vmd->vmd_pagequeues[queue];
3244 vm_pagequeue_lock(pq);
3246 vm_pqbatch_process(pq,
3247 DPCPU_PTR(pqbatch[domain][queue]), queue);
3249 vm_pagequeue_unlock(pq);
3259 * Complete the logical removal of a page from a page queue. We must be
3260 * careful to synchronize with the page daemon, which may be concurrently
3261 * examining the page with only the page lock held. The page must not be
3262 * in a state where it appears to be logically enqueued.
3265 vm_page_dequeue_complete(vm_page_t m)
3269 atomic_thread_fence_rel();
3270 vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK);
3274 * vm_page_dequeue_deferred: [ internal use only ]
3276 * Request removal of the given page from its current page
3277 * queue. Physical removal from the queue may be deferred
3280 * The page must be locked.
3283 vm_page_dequeue_deferred(vm_page_t m)
3287 vm_page_assert_locked(m);
3289 if ((queue = vm_page_queue(m)) == PQ_NONE)
3291 vm_page_aflag_set(m, PGA_DEQUEUE);
3292 vm_pqbatch_submit_page(m, queue);
3298 * Remove the page from whichever page queue it's in, if any.
3299 * The page must either be locked or unallocated. This constraint
3300 * ensures that the queue state of the page will remain consistent
3301 * after this function returns.
3304 vm_page_dequeue(vm_page_t m)
3306 struct vm_pagequeue *pq, *pq1;
3309 KASSERT(mtx_owned(vm_page_lockptr(m)) || m->object == NULL,
3310 ("page %p is allocated and unlocked", m));
3312 for (pq = vm_page_pagequeue(m);; pq = pq1) {
3315 * A thread may be concurrently executing
3316 * vm_page_dequeue_complete(). Ensure that all queue
3317 * state is cleared before we return.
3319 aflags = atomic_load_8(&m->aflags);
3320 if ((aflags & PGA_QUEUE_STATE_MASK) == 0)
3322 KASSERT((aflags & PGA_DEQUEUE) != 0,
3323 ("page %p has unexpected queue state flags %#x",
3327 * Busy wait until the thread updating queue state is
3328 * finished. Such a thread must be executing in a
3332 pq1 = vm_page_pagequeue(m);
3335 vm_pagequeue_lock(pq);
3336 if ((pq1 = vm_page_pagequeue(m)) == pq)
3338 vm_pagequeue_unlock(pq);
3340 KASSERT(pq == vm_page_pagequeue(m),
3341 ("%s: page %p migrated directly between queues", __func__, m));
3342 KASSERT((m->aflags & PGA_DEQUEUE) != 0 ||
3343 mtx_owned(vm_page_lockptr(m)),
3344 ("%s: queued unlocked page %p", __func__, m));
3346 if ((m->aflags & PGA_ENQUEUED) != 0)
3347 vm_pagequeue_remove(pq, m);
3348 vm_page_dequeue_complete(m);
3349 vm_pagequeue_unlock(pq);
3353 * Schedule the given page for insertion into the specified page queue.
3354 * Physical insertion of the page may be deferred indefinitely.
3357 vm_page_enqueue(vm_page_t m, uint8_t queue)
3360 vm_page_assert_locked(m);
3361 KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0,
3362 ("%s: page %p is already enqueued", __func__, m));
3365 if ((m->aflags & PGA_REQUEUE) == 0)
3366 vm_page_aflag_set(m, PGA_REQUEUE);
3367 vm_pqbatch_submit_page(m, queue);
3371 * vm_page_requeue: [ internal use only ]
3373 * Schedule a requeue of the given page.
3375 * The page must be locked.
3378 vm_page_requeue(vm_page_t m)
3381 vm_page_assert_locked(m);
3382 KASSERT(vm_page_queue(m) != PQ_NONE,
3383 ("%s: page %p is not logically enqueued", __func__, m));
3385 if ((m->aflags & PGA_REQUEUE) == 0)
3386 vm_page_aflag_set(m, PGA_REQUEUE);
3387 vm_pqbatch_submit_page(m, atomic_load_8(&m->queue));
3391 * vm_page_free_prep:
3393 * Prepares the given page to be put on the free list,
3394 * disassociating it from any VM object. The caller may return
3395 * the page to the free list only if this function returns true.
3397 * The object must be locked. The page must be locked if it is
3401 vm_page_free_prep(vm_page_t m)
3404 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3405 if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3408 p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3409 for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3410 KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3411 m, i, (uintmax_t)*p));
3414 if ((m->oflags & VPO_UNMANAGED) == 0) {
3415 vm_page_lock_assert(m, MA_OWNED);
3416 KASSERT(!pmap_page_is_mapped(m),
3417 ("vm_page_free_prep: freeing mapped page %p", m));
3418 KASSERT((m->aflags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
3419 ("vm_page_free_prep: mapping flags set in page %p", m));
3421 KASSERT(m->queue == PQ_NONE,
3422 ("vm_page_free_prep: unmanaged page %p is queued", m));
3424 VM_CNT_INC(v_tfree);
3426 if (vm_page_sbusied(m))
3427 panic("vm_page_free_prep: freeing busy page %p", m);
3429 if (m->object != NULL)
3430 (void)vm_page_remove(m);
3433 * If fictitious remove object association and
3436 if ((m->flags & PG_FICTITIOUS) != 0) {
3437 KASSERT(m->wire_count == 1,
3438 ("fictitious page %p is not wired", m));
3439 KASSERT(m->queue == PQ_NONE,
3440 ("fictitious page %p is queued", m));
3445 * Pages need not be dequeued before they are returned to the physical
3446 * memory allocator, but they must at least be marked for a deferred
3449 if ((m->oflags & VPO_UNMANAGED) == 0)
3450 vm_page_dequeue_deferred(m);
3455 if (vm_page_wired(m) != 0)
3456 panic("vm_page_free_prep: freeing wired page %p", m);
3457 if (m->hold_count != 0) {
3458 m->flags &= ~PG_ZERO;
3459 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
3460 ("vm_page_free_prep: freeing PG_UNHOLDFREE page %p", m));
3461 m->flags |= PG_UNHOLDFREE;
3466 * Restore the default memory attribute to the page.
3468 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3469 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3471 #if VM_NRESERVLEVEL > 0
3473 * Determine whether the page belongs to a reservation. If the page was
3474 * allocated from a per-CPU cache, it cannot belong to a reservation, so
3475 * as an optimization, we avoid the check in that case.
3477 if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
3487 * Returns the given page to the free list, disassociating it
3488 * from any VM object.
3490 * The object must be locked. The page must be locked if it is
3494 vm_page_free_toq(vm_page_t m)
3496 struct vm_domain *vmd;
3499 if (!vm_page_free_prep(m))
3502 vmd = vm_pagequeue_domain(m);
3503 zone = vmd->vmd_pgcache[m->pool].zone;
3504 if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
3508 vm_domain_free_lock(vmd);
3509 vm_phys_free_pages(m, 0);
3510 vm_domain_free_unlock(vmd);
3511 vm_domain_freecnt_inc(vmd, 1);
3515 * vm_page_free_pages_toq:
3517 * Returns a list of pages to the free list, disassociating it
3518 * from any VM object. In other words, this is equivalent to
3519 * calling vm_page_free_toq() for each page of a list of VM objects.
3521 * The objects must be locked. The pages must be locked if it is
3525 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
3530 if (SLIST_EMPTY(free))
3534 while ((m = SLIST_FIRST(free)) != NULL) {
3536 SLIST_REMOVE_HEAD(free, plinks.s.ss);
3537 vm_page_free_toq(m);
3540 if (update_wire_count)
3547 * Mark this page as wired down. If the page is fictitious, then
3548 * its wire count must remain one.
3550 * The page must be locked.
3553 vm_page_wire(vm_page_t m)
3556 vm_page_assert_locked(m);
3557 if ((m->flags & PG_FICTITIOUS) != 0) {
3558 KASSERT(m->wire_count == 1,
3559 ("vm_page_wire: fictitious page %p's wire count isn't one",
3563 if (!vm_page_wired(m)) {
3564 KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
3565 m->queue == PQ_NONE,
3566 ("vm_page_wire: unmanaged page %p is queued", m));
3570 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
3576 * Release one wiring of the specified page, potentially allowing it to be
3577 * paged out. Returns TRUE if the number of wirings transitions to zero and
3580 * Only managed pages belonging to an object can be paged out. If the number
3581 * of wirings transitions to zero and the page is eligible for page out, then
3582 * the page is added to the specified paging queue (unless PQ_NONE is
3583 * specified, in which case the page is dequeued if it belongs to a paging
3586 * If a page is fictitious, then its wire count must always be one.
3588 * A managed page must be locked.
3591 vm_page_unwire(vm_page_t m, uint8_t queue)
3595 KASSERT(queue < PQ_COUNT || queue == PQ_NONE,
3596 ("vm_page_unwire: invalid queue %u request for page %p",
3598 if ((m->oflags & VPO_UNMANAGED) == 0)
3599 vm_page_assert_locked(m);
3601 unwired = vm_page_unwire_noq(m);
3602 if (!unwired || (m->oflags & VPO_UNMANAGED) != 0 || m->object == NULL)
3605 if (vm_page_queue(m) == queue) {
3606 if (queue == PQ_ACTIVE)
3607 vm_page_reference(m);
3608 else if (queue != PQ_NONE)
3612 if (queue != PQ_NONE) {
3613 vm_page_enqueue(m, queue);
3614 if (queue == PQ_ACTIVE)
3615 /* Initialize act_count. */
3616 vm_page_activate(m);
3624 * vm_page_unwire_noq:
3626 * Unwire a page without (re-)inserting it into a page queue. It is up
3627 * to the caller to enqueue, requeue, or free the page as appropriate.
3628 * In most cases, vm_page_unwire() should be used instead.
3631 vm_page_unwire_noq(vm_page_t m)
3634 if ((m->oflags & VPO_UNMANAGED) == 0)
3635 vm_page_assert_locked(m);
3636 if ((m->flags & PG_FICTITIOUS) != 0) {
3637 KASSERT(m->wire_count == 1,
3638 ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
3641 if (!vm_page_wired(m))
3642 panic("vm_page_unwire: page %p's wire count is zero", m);
3644 if (m->wire_count == 0) {
3654 * Put the specified page on the active list (if appropriate).
3655 * Ensure that act_count is at least ACT_INIT but do not otherwise
3658 * The page must be locked.
3661 vm_page_activate(vm_page_t m)
3664 vm_page_assert_locked(m);
3666 if (vm_page_wired(m) || (m->oflags & VPO_UNMANAGED) != 0)
3668 if (vm_page_queue(m) == PQ_ACTIVE) {
3669 if (m->act_count < ACT_INIT)
3670 m->act_count = ACT_INIT;
3675 if (m->act_count < ACT_INIT)
3676 m->act_count = ACT_INIT;
3677 vm_page_enqueue(m, PQ_ACTIVE);
3681 * Move the specified page to the tail of the inactive queue, or requeue
3682 * the page if it is already in the inactive queue.
3684 * The page must be locked.
3687 vm_page_deactivate(vm_page_t m)
3690 vm_page_assert_locked(m);
3692 if (vm_page_wired(m) || (m->oflags & VPO_UNMANAGED) != 0)
3695 if (!vm_page_inactive(m)) {
3697 vm_page_enqueue(m, PQ_INACTIVE);
3703 * Move the specified page close to the head of the inactive queue,
3704 * bypassing LRU. A marker page is used to maintain FIFO ordering.
3705 * As with regular enqueues, we use a per-CPU batch queue to reduce
3706 * contention on the page queue lock.
3708 * The page must be locked.
3711 vm_page_deactivate_noreuse(vm_page_t m)
3714 vm_page_assert_locked(m);
3716 if (vm_page_wired(m) || (m->oflags & VPO_UNMANAGED) != 0)
3719 if (!vm_page_inactive(m)) {
3721 m->queue = PQ_INACTIVE;
3723 if ((m->aflags & PGA_REQUEUE_HEAD) == 0)
3724 vm_page_aflag_set(m, PGA_REQUEUE_HEAD);
3725 vm_pqbatch_submit_page(m, PQ_INACTIVE);
3731 * Put a page in the laundry, or requeue it if it is already there.
3734 vm_page_launder(vm_page_t m)
3737 vm_page_assert_locked(m);
3738 if (vm_page_wired(m) || (m->oflags & VPO_UNMANAGED) != 0)
3741 if (vm_page_in_laundry(m))
3745 vm_page_enqueue(m, PQ_LAUNDRY);
3750 * vm_page_unswappable
3752 * Put a page in the PQ_UNSWAPPABLE holding queue.
3755 vm_page_unswappable(vm_page_t m)
3758 vm_page_assert_locked(m);
3759 KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0,
3760 ("page %p already unswappable", m));
3763 vm_page_enqueue(m, PQ_UNSWAPPABLE);
3767 vm_page_release_toq(vm_page_t m, int flags)
3771 * Use a check of the valid bits to determine whether we should
3772 * accelerate reclamation of the page. The object lock might not be
3773 * held here, in which case the check is racy. At worst we will either
3774 * accelerate reclamation of a valid page and violate LRU, or
3775 * unnecessarily defer reclamation of an invalid page.
3777 * If we were asked to not cache the page, place it near the head of the
3778 * inactive queue so that is reclaimed sooner.
3780 if ((flags & (VPR_TRYFREE | VPR_NOREUSE)) != 0 || m->valid == 0)
3781 vm_page_deactivate_noreuse(m);
3782 else if (vm_page_active(m))
3783 vm_page_reference(m);
3785 vm_page_deactivate(m);
3789 * Unwire a page and either attempt to free it or re-add it to the page queues.
3792 vm_page_release(vm_page_t m, int flags)
3797 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3798 ("vm_page_release: page %p is unmanaged", m));
3801 if (m->object != NULL)
3802 VM_OBJECT_ASSERT_UNLOCKED(m->object);
3803 if (vm_page_unwire_noq(m)) {
3804 if ((object = m->object) == NULL) {
3808 if ((flags & VPR_TRYFREE) != 0 && !vm_page_busied(m) &&
3809 /* Depends on type stability. */
3810 VM_OBJECT_TRYWLOCK(object)) {
3812 * Only free unmapped pages. The busy test from
3813 * before the object was locked cannot be relied
3816 if ((object->ref_count == 0 ||
3817 !pmap_page_is_mapped(m)) && m->dirty == 0 &&
3818 !vm_page_busied(m)) {
3822 VM_OBJECT_WUNLOCK(object);
3826 vm_page_release_toq(m, flags);
3832 /* See vm_page_release(). */
3834 vm_page_release_locked(vm_page_t m, int flags)
3837 VM_OBJECT_ASSERT_WLOCKED(m->object);
3838 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3839 ("vm_page_release_locked: page %p is unmanaged", m));
3842 if (vm_page_unwire_noq(m)) {
3843 if ((flags & VPR_TRYFREE) != 0 &&
3844 (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
3845 m->dirty == 0 && !vm_page_busied(m)) {
3848 vm_page_release_toq(m, flags);
3857 * Apply the specified advice to the given page.
3859 * The object and page must be locked.
3862 vm_page_advise(vm_page_t m, int advice)
3865 vm_page_assert_locked(m);
3866 VM_OBJECT_ASSERT_WLOCKED(m->object);
3867 if (advice == MADV_FREE)
3869 * Mark the page clean. This will allow the page to be freed
3870 * without first paging it out. MADV_FREE pages are often
3871 * quickly reused by malloc(3), so we do not do anything that
3872 * would result in a page fault on a later access.
3875 else if (advice != MADV_DONTNEED) {
3876 if (advice == MADV_WILLNEED)
3877 vm_page_activate(m);
3882 * Clear any references to the page. Otherwise, the page daemon will
3883 * immediately reactivate the page.
3885 vm_page_aflag_clear(m, PGA_REFERENCED);
3887 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
3891 * Place clean pages near the head of the inactive queue rather than
3892 * the tail, thus defeating the queue's LRU operation and ensuring that
3893 * the page will be reused quickly. Dirty pages not already in the
3894 * laundry are moved there.
3897 vm_page_deactivate_noreuse(m);
3898 else if (!vm_page_in_laundry(m))
3903 * Grab a page, waiting until we are waken up due to the page
3904 * changing state. We keep on waiting, if the page continues
3905 * to be in the object. If the page doesn't exist, first allocate it
3906 * and then conditionally zero it.
3908 * This routine may sleep.
3910 * The object must be locked on entry. The lock will, however, be released
3911 * and reacquired if the routine sleeps.
3914 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
3920 VM_OBJECT_ASSERT_WLOCKED(object);
3921 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3922 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3923 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
3924 pflags = allocflags &
3925 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
3926 if ((allocflags & VM_ALLOC_NOWAIT) == 0)
3927 pflags |= VM_ALLOC_WAITFAIL;
3929 if ((m = vm_page_lookup(object, pindex)) != NULL) {
3930 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3931 vm_page_xbusied(m) : vm_page_busied(m);
3933 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3936 * Reference the page before unlocking and
3937 * sleeping so that the page daemon is less
3938 * likely to reclaim it.
3940 vm_page_aflag_set(m, PGA_REFERENCED);
3942 VM_OBJECT_WUNLOCK(object);
3943 vm_page_busy_sleep(m, "pgrbwt", (allocflags &
3944 VM_ALLOC_IGN_SBUSY) != 0);
3945 VM_OBJECT_WLOCK(object);
3948 if ((allocflags & VM_ALLOC_WIRED) != 0) {
3954 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
3956 if ((allocflags & VM_ALLOC_SBUSY) != 0)
3961 m = vm_page_alloc(object, pindex, pflags);
3963 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3967 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
3973 * Return the specified range of pages from the given object. For each
3974 * page offset within the range, if a page already exists within the object
3975 * at that offset and it is busy, then wait for it to change state. If,
3976 * instead, the page doesn't exist, then allocate it.
3978 * The caller must always specify an allocation class.
3980 * allocation classes:
3981 * VM_ALLOC_NORMAL normal process request
3982 * VM_ALLOC_SYSTEM system *really* needs the pages
3984 * The caller must always specify that the pages are to be busied and/or
3987 * optional allocation flags:
3988 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages
3989 * VM_ALLOC_NOBUSY do not exclusive busy the page
3990 * VM_ALLOC_NOWAIT do not sleep
3991 * VM_ALLOC_SBUSY set page to sbusy state
3992 * VM_ALLOC_WIRED wire the pages
3993 * VM_ALLOC_ZERO zero and validate any invalid pages
3995 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it
3996 * may return a partial prefix of the requested range.
3999 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
4000 vm_page_t *ma, int count)
4007 VM_OBJECT_ASSERT_WLOCKED(object);
4008 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
4009 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
4010 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4011 (allocflags & VM_ALLOC_WIRED) != 0,
4012 ("vm_page_grab_pages: the pages must be busied or wired"));
4013 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4014 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4015 ("vm_page_grab_pages: VM_ALLOC_SBUSY/IGN_SBUSY mismatch"));
4018 pflags = allocflags & ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK |
4019 VM_ALLOC_WAITFAIL | VM_ALLOC_IGN_SBUSY);
4020 if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4021 pflags |= VM_ALLOC_WAITFAIL;
4024 m = vm_radix_lookup_le(&object->rtree, pindex + i);
4025 if (m == NULL || m->pindex != pindex + i) {
4029 mpred = TAILQ_PREV(m, pglist, listq);
4030 for (; i < count; i++) {
4032 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
4033 vm_page_xbusied(m) : vm_page_busied(m);
4035 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4038 * Reference the page before unlocking and
4039 * sleeping so that the page daemon is less
4040 * likely to reclaim it.
4042 vm_page_aflag_set(m, PGA_REFERENCED);
4044 VM_OBJECT_WUNLOCK(object);
4045 vm_page_busy_sleep(m, "grbmaw", (allocflags &
4046 VM_ALLOC_IGN_SBUSY) != 0);
4047 VM_OBJECT_WLOCK(object);
4050 if ((allocflags & VM_ALLOC_WIRED) != 0) {
4055 if ((allocflags & (VM_ALLOC_NOBUSY |
4056 VM_ALLOC_SBUSY)) == 0)
4058 if ((allocflags & VM_ALLOC_SBUSY) != 0)
4061 m = vm_page_alloc_after(object, pindex + i,
4062 pflags | VM_ALLOC_COUNT(count - i), mpred);
4064 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4069 if (m->valid == 0 && (allocflags & VM_ALLOC_ZERO) != 0) {
4070 if ((m->flags & PG_ZERO) == 0)
4072 m->valid = VM_PAGE_BITS_ALL;
4075 m = vm_page_next(m);
4081 * Mapping function for valid or dirty bits in a page.
4083 * Inputs are required to range within a page.
4086 vm_page_bits(int base, int size)
4092 base + size <= PAGE_SIZE,
4093 ("vm_page_bits: illegal base/size %d/%d", base, size)
4096 if (size == 0) /* handle degenerate case */
4099 first_bit = base >> DEV_BSHIFT;
4100 last_bit = (base + size - 1) >> DEV_BSHIFT;
4102 return (((vm_page_bits_t)2 << last_bit) -
4103 ((vm_page_bits_t)1 << first_bit));
4107 * vm_page_set_valid_range:
4109 * Sets portions of a page valid. The arguments are expected
4110 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4111 * of any partial chunks touched by the range. The invalid portion of
4112 * such chunks will be zeroed.
4114 * (base + size) must be less then or equal to PAGE_SIZE.
4117 vm_page_set_valid_range(vm_page_t m, int base, int size)
4121 VM_OBJECT_ASSERT_WLOCKED(m->object);
4122 if (size == 0) /* handle degenerate case */
4126 * If the base is not DEV_BSIZE aligned and the valid
4127 * bit is clear, we have to zero out a portion of the
4130 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4131 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
4132 pmap_zero_page_area(m, frag, base - frag);
4135 * If the ending offset is not DEV_BSIZE aligned and the
4136 * valid bit is clear, we have to zero out a portion of
4139 endoff = base + size;
4140 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4141 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
4142 pmap_zero_page_area(m, endoff,
4143 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4146 * Assert that no previously invalid block that is now being validated
4149 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
4150 ("vm_page_set_valid_range: page %p is dirty", m));
4153 * Set valid bits inclusive of any overlap.
4155 m->valid |= vm_page_bits(base, size);
4159 * Clear the given bits from the specified page's dirty field.
4161 static __inline void
4162 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
4165 #if PAGE_SIZE < 16384
4170 * If the object is locked and the page is neither exclusive busy nor
4171 * write mapped, then the page's dirty field cannot possibly be
4172 * set by a concurrent pmap operation.
4174 VM_OBJECT_ASSERT_WLOCKED(m->object);
4175 if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
4176 m->dirty &= ~pagebits;
4179 * The pmap layer can call vm_page_dirty() without
4180 * holding a distinguished lock. The combination of
4181 * the object's lock and an atomic operation suffice
4182 * to guarantee consistency of the page dirty field.
4184 * For PAGE_SIZE == 32768 case, compiler already
4185 * properly aligns the dirty field, so no forcible
4186 * alignment is needed. Only require existence of
4187 * atomic_clear_64 when page size is 32768.
4189 addr = (uintptr_t)&m->dirty;
4190 #if PAGE_SIZE == 32768
4191 atomic_clear_64((uint64_t *)addr, pagebits);
4192 #elif PAGE_SIZE == 16384
4193 atomic_clear_32((uint32_t *)addr, pagebits);
4194 #else /* PAGE_SIZE <= 8192 */
4196 * Use a trick to perform a 32-bit atomic on the
4197 * containing aligned word, to not depend on the existence
4198 * of atomic_clear_{8, 16}.
4200 shift = addr & (sizeof(uint32_t) - 1);
4201 #if BYTE_ORDER == BIG_ENDIAN
4202 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
4206 addr &= ~(sizeof(uint32_t) - 1);
4207 atomic_clear_32((uint32_t *)addr, pagebits << shift);
4208 #endif /* PAGE_SIZE */
4213 * vm_page_set_validclean:
4215 * Sets portions of a page valid and clean. The arguments are expected
4216 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4217 * of any partial chunks touched by the range. The invalid portion of
4218 * such chunks will be zero'd.
4220 * (base + size) must be less then or equal to PAGE_SIZE.
4223 vm_page_set_validclean(vm_page_t m, int base, int size)
4225 vm_page_bits_t oldvalid, pagebits;
4228 VM_OBJECT_ASSERT_WLOCKED(m->object);
4229 if (size == 0) /* handle degenerate case */
4233 * If the base is not DEV_BSIZE aligned and the valid
4234 * bit is clear, we have to zero out a portion of the
4237 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4238 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
4239 pmap_zero_page_area(m, frag, base - frag);
4242 * If the ending offset is not DEV_BSIZE aligned and the
4243 * valid bit is clear, we have to zero out a portion of
4246 endoff = base + size;
4247 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4248 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
4249 pmap_zero_page_area(m, endoff,
4250 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4253 * Set valid, clear dirty bits. If validating the entire
4254 * page we can safely clear the pmap modify bit. We also
4255 * use this opportunity to clear the VPO_NOSYNC flag. If a process
4256 * takes a write fault on a MAP_NOSYNC memory area the flag will
4259 * We set valid bits inclusive of any overlap, but we can only
4260 * clear dirty bits for DEV_BSIZE chunks that are fully within
4263 oldvalid = m->valid;
4264 pagebits = vm_page_bits(base, size);
4265 m->valid |= pagebits;
4267 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
4268 frag = DEV_BSIZE - frag;
4274 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
4276 if (base == 0 && size == PAGE_SIZE) {
4278 * The page can only be modified within the pmap if it is
4279 * mapped, and it can only be mapped if it was previously
4282 if (oldvalid == VM_PAGE_BITS_ALL)
4284 * Perform the pmap_clear_modify() first. Otherwise,
4285 * a concurrent pmap operation, such as
4286 * pmap_protect(), could clear a modification in the
4287 * pmap and set the dirty field on the page before
4288 * pmap_clear_modify() had begun and after the dirty
4289 * field was cleared here.
4291 pmap_clear_modify(m);
4293 m->oflags &= ~VPO_NOSYNC;
4294 } else if (oldvalid != VM_PAGE_BITS_ALL)
4295 m->dirty &= ~pagebits;
4297 vm_page_clear_dirty_mask(m, pagebits);
4301 vm_page_clear_dirty(vm_page_t m, int base, int size)
4304 vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
4308 * vm_page_set_invalid:
4310 * Invalidates DEV_BSIZE'd chunks within a page. Both the
4311 * valid and dirty bits for the effected areas are cleared.
4314 vm_page_set_invalid(vm_page_t m, int base, int size)
4316 vm_page_bits_t bits;
4320 VM_OBJECT_ASSERT_WLOCKED(object);
4321 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
4322 size >= object->un_pager.vnp.vnp_size)
4323 bits = VM_PAGE_BITS_ALL;
4325 bits = vm_page_bits(base, size);
4326 if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL &&
4329 KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) ||
4330 !pmap_page_is_mapped(m),
4331 ("vm_page_set_invalid: page %p is mapped", m));
4337 * vm_page_zero_invalid()
4339 * The kernel assumes that the invalid portions of a page contain
4340 * garbage, but such pages can be mapped into memory by user code.
4341 * When this occurs, we must zero out the non-valid portions of the
4342 * page so user code sees what it expects.
4344 * Pages are most often semi-valid when the end of a file is mapped
4345 * into memory and the file's size is not page aligned.
4348 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
4353 VM_OBJECT_ASSERT_WLOCKED(m->object);
4355 * Scan the valid bits looking for invalid sections that
4356 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the
4357 * valid bit may be set ) have already been zeroed by
4358 * vm_page_set_validclean().
4360 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
4361 if (i == (PAGE_SIZE / DEV_BSIZE) ||
4362 (m->valid & ((vm_page_bits_t)1 << i))) {
4364 pmap_zero_page_area(m,
4365 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
4372 * setvalid is TRUE when we can safely set the zero'd areas
4373 * as being valid. We can do this if there are no cache consistancy
4374 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
4377 m->valid = VM_PAGE_BITS_ALL;
4383 * Is (partial) page valid? Note that the case where size == 0
4384 * will return FALSE in the degenerate case where the page is
4385 * entirely invalid, and TRUE otherwise.
4388 vm_page_is_valid(vm_page_t m, int base, int size)
4390 vm_page_bits_t bits;
4392 VM_OBJECT_ASSERT_LOCKED(m->object);
4393 bits = vm_page_bits(base, size);
4394 return (m->valid != 0 && (m->valid & bits) == bits);
4398 * Returns true if all of the specified predicates are true for the entire
4399 * (super)page and false otherwise.
4402 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
4408 if (skip_m != NULL && skip_m->object != object)
4410 VM_OBJECT_ASSERT_LOCKED(object);
4411 npages = atop(pagesizes[m->psind]);
4414 * The physically contiguous pages that make up a superpage, i.e., a
4415 * page with a page size index ("psind") greater than zero, will
4416 * occupy adjacent entries in vm_page_array[].
4418 for (i = 0; i < npages; i++) {
4419 /* Always test object consistency, including "skip_m". */
4420 if (m[i].object != object)
4422 if (&m[i] == skip_m)
4424 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
4426 if ((flags & PS_ALL_DIRTY) != 0) {
4428 * Calling vm_page_test_dirty() or pmap_is_modified()
4429 * might stop this case from spuriously returning
4430 * "false". However, that would require a write lock
4431 * on the object containing "m[i]".
4433 if (m[i].dirty != VM_PAGE_BITS_ALL)
4436 if ((flags & PS_ALL_VALID) != 0 &&
4437 m[i].valid != VM_PAGE_BITS_ALL)
4444 * Set the page's dirty bits if the page is modified.
4447 vm_page_test_dirty(vm_page_t m)
4450 VM_OBJECT_ASSERT_WLOCKED(m->object);
4451 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
4456 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
4459 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
4463 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
4466 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
4470 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
4473 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
4476 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
4478 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
4481 vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
4485 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
4488 mtx_assert_(vm_page_lockptr(m), a, file, line);
4494 vm_page_object_lock_assert(vm_page_t m)
4498 * Certain of the page's fields may only be modified by the
4499 * holder of the containing object's lock or the exclusive busy.
4500 * holder. Unfortunately, the holder of the write busy is
4501 * not recorded, and thus cannot be checked here.
4503 if (m->object != NULL && !vm_page_xbusied(m))
4504 VM_OBJECT_ASSERT_WLOCKED(m->object);
4508 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits)
4511 if ((bits & PGA_WRITEABLE) == 0)
4515 * The PGA_WRITEABLE flag can only be set if the page is
4516 * managed, is exclusively busied or the object is locked.
4517 * Currently, this flag is only set by pmap_enter().
4519 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4520 ("PGA_WRITEABLE on unmanaged page"));
4521 if (!vm_page_xbusied(m))
4522 VM_OBJECT_ASSERT_LOCKED(m->object);
4526 #include "opt_ddb.h"
4528 #include <sys/kernel.h>
4530 #include <ddb/ddb.h>
4532 DB_SHOW_COMMAND(page, vm_page_print_page_info)
4535 db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
4536 db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
4537 db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
4538 db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
4539 db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
4540 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
4541 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
4542 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
4543 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
4546 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
4550 db_printf("pq_free %d\n", vm_free_count());
4551 for (dom = 0; dom < vm_ndomains; dom++) {
4553 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
4555 vm_dom[dom].vmd_page_count,
4556 vm_dom[dom].vmd_free_count,
4557 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
4558 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
4559 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
4560 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
4564 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
4567 boolean_t phys, virt;
4570 db_printf("show pginfo addr\n");
4574 phys = strchr(modif, 'p') != NULL;
4575 virt = strchr(modif, 'v') != NULL;
4577 m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
4579 m = PHYS_TO_VM_PAGE(addr);
4581 m = (vm_page_t)addr;
4583 "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n"
4584 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
4585 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
4586 m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags,
4587 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty);