2 * Copyright (c) 1991 Regents of the University of California.
4 * Copyright (c) 1998 Matthew Dillon. All Rights Reserved.
6 * This code is derived from software contributed to Berkeley by
7 * The Mach Operating System project at Carnegie-Mellon University.
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
17 * 4. Neither the name of the University nor the names of its contributors
18 * may be used to endorse or promote products derived from this software
19 * without specific prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91
37 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38 * All rights reserved.
40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
42 * Permission to use, copy, modify and distribute this software and
43 * its documentation is hereby granted, provided that both the copyright
44 * notice and this permission notice appear in all copies of the
45 * software, derivative works or modified versions, and any portions
46 * thereof, and that both notices appear in supporting documentation.
48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
52 * Carnegie Mellon requests users of this software to return to
54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
55 * School of Computer Science
56 * Carnegie Mellon University
57 * Pittsburgh PA 15213-3890
59 * any improvements or extensions that they make and grant Carnegie the
60 * rights to redistribute these changes.
64 * GENERAL RULES ON VM_PAGE MANIPULATION
66 * - A page queue lock is required when adding or removing a page from a
67 * page queue (vm_pagequeues[]), regardless of other locks or the
68 * busy state of a page.
70 * * In general, no thread besides the page daemon can acquire or
71 * hold more than one page queue lock at a time.
73 * * The page daemon can acquire and hold any pair of page queue
76 * - The object mutex is held when inserting or removing
77 * pages from an object (vm_page_insert() or vm_page_remove()).
82 * Resident memory management module.
85 #include <sys/cdefs.h>
86 __FBSDID("$FreeBSD$");
90 #include <sys/param.h>
91 #include <sys/systm.h>
93 #include <sys/kernel.h>
94 #include <sys/limits.h>
95 #include <sys/malloc.h>
96 #include <sys/msgbuf.h>
97 #include <sys/mutex.h>
99 #include <sys/sysctl.h>
100 #include <sys/vmmeter.h>
101 #include <sys/vnode.h>
105 #include <vm/vm_param.h>
106 #include <vm/vm_kern.h>
107 #include <vm/vm_object.h>
108 #include <vm/vm_page.h>
109 #include <vm/vm_pageout.h>
110 #include <vm/vm_pager.h>
111 #include <vm/vm_phys.h>
112 #include <vm/vm_reserv.h>
113 #include <vm/vm_extern.h>
115 #include <vm/uma_int.h>
117 #include <machine/md_var.h>
120 * Associated with page of user-allocatable memory is a
124 struct vm_pagequeue vm_pagequeues[PQ_COUNT] = {
126 .pq_pl = TAILQ_HEAD_INITIALIZER(
127 vm_pagequeues[PQ_INACTIVE].pq_pl),
128 .pq_cnt = &cnt.v_inactive_count,
129 .pq_name = "vm inactive pagequeue"
132 .pq_pl = TAILQ_HEAD_INITIALIZER(
133 vm_pagequeues[PQ_ACTIVE].pq_pl),
134 .pq_cnt = &cnt.v_active_count,
135 .pq_name = "vm active pagequeue"
138 struct mtx_padalign vm_page_queue_free_mtx;
140 struct mtx_padalign pa_lock[PA_LOCK_COUNT];
142 vm_page_t vm_page_array;
143 long vm_page_array_size;
145 int vm_page_zero_count;
147 static int boot_pages = UMA_BOOT_PAGES;
148 TUNABLE_INT("vm.boot_pages", &boot_pages);
149 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RD, &boot_pages, 0,
150 "number of pages allocated for bootstrapping the VM system");
152 static int pa_tryrelock_restart;
153 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
154 &pa_tryrelock_restart, 0, "Number of tryrelock restarts");
156 static uma_zone_t fakepg_zone;
158 static struct vnode *vm_page_alloc_init(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_enqueue(int queue, vm_page_t m);
161 static void vm_page_init_fakepg(void *dummy);
163 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init_fakepg, NULL);
166 vm_page_init_fakepg(void *dummy)
169 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
170 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
173 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
174 #if PAGE_SIZE == 32768
176 CTASSERT(sizeof(u_long) >= 8);
181 * Try to acquire a physical address lock while a pmap is locked. If we
182 * fail to trylock we unlock and lock the pmap directly and cache the
183 * locked pa in *locked. The caller should then restart their loop in case
184 * the virtual to physical mapping has changed.
187 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
194 PA_LOCK_ASSERT(lockpa, MA_OWNED);
195 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
202 atomic_add_int(&pa_tryrelock_restart, 1);
211 * Sets the page size, perhaps based upon the memory
212 * size. Must be called before any use of page-size
213 * dependent functions.
216 vm_set_page_size(void)
218 if (cnt.v_page_size == 0)
219 cnt.v_page_size = PAGE_SIZE;
220 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0)
221 panic("vm_set_page_size: page size not a power of two");
225 * vm_page_blacklist_lookup:
227 * See if a physical address in this page has been listed
228 * in the blacklist tunable. Entries in the tunable are
229 * separated by spaces or commas. If an invalid integer is
230 * encountered then the rest of the string is skipped.
233 vm_page_blacklist_lookup(char *list, vm_paddr_t pa)
238 for (pos = list; *pos != '\0'; pos = cp) {
239 bad = strtoq(pos, &cp, 0);
241 if (*cp == ' ' || *cp == ',') {
248 if (pa == trunc_page(bad))
257 * Initializes the resident memory module.
259 * Allocates memory for the page cells, and
260 * for the object/offset-to-page hash table headers.
261 * Each page cell is initialized and placed on the free list.
264 vm_page_startup(vm_offset_t vaddr)
267 vm_paddr_t page_range;
274 /* the biggest memory array is the second group of pages */
276 vm_paddr_t biggestsize;
277 vm_paddr_t low_water, high_water;
282 vaddr = round_page(vaddr);
284 for (i = 0; phys_avail[i + 1]; i += 2) {
285 phys_avail[i] = round_page(phys_avail[i]);
286 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
289 low_water = phys_avail[0];
290 high_water = phys_avail[1];
292 for (i = 0; phys_avail[i + 1]; i += 2) {
293 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
295 if (size > biggestsize) {
299 if (phys_avail[i] < low_water)
300 low_water = phys_avail[i];
301 if (phys_avail[i + 1] > high_water)
302 high_water = phys_avail[i + 1];
309 end = phys_avail[biggestone+1];
312 * Initialize the page and queue locks.
314 mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF);
315 for (i = 0; i < PA_LOCK_COUNT; i++)
316 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
317 for (i = 0; i < PQ_COUNT; i++)
318 vm_pagequeue_init_lock(&vm_pagequeues[i]);
321 * Allocate memory for use when boot strapping the kernel memory
324 new_end = end - (boot_pages * UMA_SLAB_SIZE);
325 new_end = trunc_page(new_end);
326 mapped = pmap_map(&vaddr, new_end, end,
327 VM_PROT_READ | VM_PROT_WRITE);
328 bzero((void *)mapped, end - new_end);
329 uma_startup((void *)mapped, boot_pages);
331 #if defined(__amd64__) || defined(__i386__) || defined(__arm__) || \
334 * Allocate a bitmap to indicate that a random physical page
335 * needs to be included in a minidump.
337 * The amd64 port needs this to indicate which direct map pages
338 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
340 * However, i386 still needs this workspace internally within the
341 * minidump code. In theory, they are not needed on i386, but are
342 * included should the sf_buf code decide to use them.
345 for (i = 0; dump_avail[i + 1] != 0; i += 2)
346 if (dump_avail[i + 1] > last_pa)
347 last_pa = dump_avail[i + 1];
348 page_range = last_pa / PAGE_SIZE;
349 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
350 new_end -= vm_page_dump_size;
351 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
352 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
353 bzero((void *)vm_page_dump, vm_page_dump_size);
357 * Request that the physical pages underlying the message buffer be
358 * included in a crash dump. Since the message buffer is accessed
359 * through the direct map, they are not automatically included.
361 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
362 last_pa = pa + round_page(msgbufsize);
363 while (pa < last_pa) {
369 * Compute the number of pages of memory that will be available for
370 * use (taking into account the overhead of a page structure per
373 first_page = low_water / PAGE_SIZE;
374 #ifdef VM_PHYSSEG_SPARSE
376 for (i = 0; phys_avail[i + 1] != 0; i += 2)
377 page_range += atop(phys_avail[i + 1] - phys_avail[i]);
378 #elif defined(VM_PHYSSEG_DENSE)
379 page_range = high_water / PAGE_SIZE - first_page;
381 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
386 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
391 * Initialize the mem entry structures now, and put them in the free
394 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
395 mapped = pmap_map(&vaddr, new_end, end,
396 VM_PROT_READ | VM_PROT_WRITE);
397 vm_page_array = (vm_page_t) mapped;
398 #if VM_NRESERVLEVEL > 0
400 * Allocate memory for the reservation management system's data
403 new_end = vm_reserv_startup(&vaddr, new_end, high_water);
405 #if defined(__amd64__) || defined(__mips__)
407 * pmap_map on amd64 and mips can come out of the direct-map, not kvm
408 * like i386, so the pages must be tracked for a crashdump to include
409 * this data. This includes the vm_page_array and the early UMA
412 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
415 phys_avail[biggestone + 1] = new_end;
418 * Clear all of the page structures
420 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
421 for (i = 0; i < page_range; i++)
422 vm_page_array[i].order = VM_NFREEORDER;
423 vm_page_array_size = page_range;
426 * Initialize the physical memory allocator.
431 * Add every available physical page that is not blacklisted to
434 cnt.v_page_count = 0;
435 cnt.v_free_count = 0;
436 list = getenv("vm.blacklist");
437 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
439 last_pa = phys_avail[i + 1];
440 while (pa < last_pa) {
442 vm_page_blacklist_lookup(list, pa))
443 printf("Skipping page with pa 0x%jx\n",
446 vm_phys_add_page(pa);
451 #if VM_NRESERVLEVEL > 0
453 * Initialize the reservation management system.
461 vm_page_reference(vm_page_t m)
464 vm_page_aflag_set(m, PGA_REFERENCED);
468 vm_page_busy(vm_page_t m)
471 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
472 KASSERT((m->oflags & VPO_BUSY) == 0,
473 ("vm_page_busy: page already busy!!!"));
474 m->oflags |= VPO_BUSY;
480 * wakeup anyone waiting for the page.
483 vm_page_flash(vm_page_t m)
486 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
487 if (m->oflags & VPO_WANTED) {
488 m->oflags &= ~VPO_WANTED;
496 * clear the VPO_BUSY flag and wakeup anyone waiting for the
501 vm_page_wakeup(vm_page_t m)
504 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
505 KASSERT(m->oflags & VPO_BUSY, ("vm_page_wakeup: page not busy!!!"));
506 m->oflags &= ~VPO_BUSY;
511 vm_page_io_start(vm_page_t m)
514 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
519 vm_page_io_finish(vm_page_t m)
522 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
523 KASSERT(m->busy > 0, ("vm_page_io_finish: page %p is not busy", m));
530 * Keep page from being freed by the page daemon
531 * much of the same effect as wiring, except much lower
532 * overhead and should be used only for *very* temporary
533 * holding ("wiring").
536 vm_page_hold(vm_page_t mem)
539 vm_page_lock_assert(mem, MA_OWNED);
544 vm_page_unhold(vm_page_t mem)
547 vm_page_lock_assert(mem, MA_OWNED);
549 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
550 if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0)
551 vm_page_free_toq(mem);
555 * vm_page_unhold_pages:
557 * Unhold each of the pages that is referenced by the given array.
560 vm_page_unhold_pages(vm_page_t *ma, int count)
562 struct mtx *mtx, *new_mtx;
565 for (; count != 0; count--) {
567 * Avoid releasing and reacquiring the same page lock.
569 new_mtx = vm_page_lockptr(*ma);
570 if (mtx != new_mtx) {
584 PHYS_TO_VM_PAGE(vm_paddr_t pa)
588 #ifdef VM_PHYSSEG_SPARSE
589 m = vm_phys_paddr_to_vm_page(pa);
591 m = vm_phys_fictitious_to_vm_page(pa);
593 #elif defined(VM_PHYSSEG_DENSE)
597 if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
598 m = &vm_page_array[pi - first_page];
601 return (vm_phys_fictitious_to_vm_page(pa));
603 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
610 * Create a fictitious page with the specified physical address and
611 * memory attribute. The memory attribute is the only the machine-
612 * dependent aspect of a fictitious page that must be initialized.
615 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
619 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
620 vm_page_initfake(m, paddr, memattr);
625 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
628 if ((m->flags & PG_FICTITIOUS) != 0) {
630 * The page's memattr might have changed since the
631 * previous initialization. Update the pmap to the
636 m->phys_addr = paddr;
638 /* Fictitious pages don't use "segind". */
639 m->flags = PG_FICTITIOUS;
640 /* Fictitious pages don't use "order" or "pool". */
641 m->oflags = VPO_BUSY | VPO_UNMANAGED;
644 pmap_page_set_memattr(m, memattr);
650 * Release a fictitious page.
653 vm_page_putfake(vm_page_t m)
656 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
657 KASSERT((m->flags & PG_FICTITIOUS) != 0,
658 ("vm_page_putfake: bad page %p", m));
659 uma_zfree(fakepg_zone, m);
663 * vm_page_updatefake:
665 * Update the given fictitious page to the specified physical address and
669 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
672 KASSERT((m->flags & PG_FICTITIOUS) != 0,
673 ("vm_page_updatefake: bad page %p", m));
674 m->phys_addr = paddr;
675 pmap_page_set_memattr(m, memattr);
684 vm_page_free(vm_page_t m)
687 m->flags &= ~PG_ZERO;
694 * Free a page to the zerod-pages queue
697 vm_page_free_zero(vm_page_t m)
705 * Unbusy and handle the page queueing for a page from the VOP_GETPAGES()
706 * array which is not the request page.
709 vm_page_readahead_finish(vm_page_t m)
714 * Since the page is not the requested page, whether
715 * it should be activated or deactivated is not
716 * obvious. Empirical results have shown that
717 * deactivating the page is usually the best choice,
718 * unless the page is wanted by another thread.
720 if (m->oflags & VPO_WANTED) {
726 vm_page_deactivate(m);
732 * Free the completely invalid page. Such page state
733 * occurs due to the short read operation which did
734 * not covered our page at all, or in case when a read
746 * Sleep and release the page lock.
748 * The object containing the given page must be locked.
751 vm_page_sleep(vm_page_t m, const char *msg)
754 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
755 if (mtx_owned(vm_page_lockptr(m)))
759 * It's possible that while we sleep, the page will get
760 * unbusied and freed. If we are holding the object
761 * lock, we will assume we hold a reference to the object
762 * such that even if m->object changes, we can re-lock
765 m->oflags |= VPO_WANTED;
766 VM_OBJECT_SLEEP(m->object, m, PVM, msg, 0);
770 * vm_page_dirty_KBI: [ internal use only ]
772 * Set all bits in the page's dirty field.
774 * The object containing the specified page must be locked if the
775 * call is made from the machine-independent layer.
777 * See vm_page_clear_dirty_mask().
779 * This function should only be called by vm_page_dirty().
782 vm_page_dirty_KBI(vm_page_t m)
785 /* These assertions refer to this operation by its public name. */
786 KASSERT((m->flags & PG_CACHED) == 0,
787 ("vm_page_dirty: page in cache!"));
788 KASSERT(!VM_PAGE_IS_FREE(m),
789 ("vm_page_dirty: page is free!"));
790 KASSERT(m->valid == VM_PAGE_BITS_ALL,
791 ("vm_page_dirty: page is invalid!"));
792 m->dirty = VM_PAGE_BITS_ALL;
798 * Implements Sleator and Tarjan's top-down splay algorithm. Returns
799 * the vm_page containing the given pindex. If, however, that
800 * pindex is not found in the vm_object, returns a vm_page that is
801 * adjacent to the pindex, coming before or after it.
804 vm_page_splay(vm_pindex_t pindex, vm_page_t root)
806 struct vm_page dummy;
807 vm_page_t lefttreemax, righttreemin, y;
811 lefttreemax = righttreemin = &dummy;
813 if (pindex < root->pindex) {
814 if ((y = root->left) == NULL)
816 if (pindex < y->pindex) {
818 root->left = y->right;
821 if ((y = root->left) == NULL)
824 /* Link into the new root's right tree. */
825 righttreemin->left = root;
827 } else if (pindex > root->pindex) {
828 if ((y = root->right) == NULL)
830 if (pindex > y->pindex) {
832 root->right = y->left;
835 if ((y = root->right) == NULL)
838 /* Link into the new root's left tree. */
839 lefttreemax->right = root;
844 /* Assemble the new root. */
845 lefttreemax->right = root->left;
846 righttreemin->left = root->right;
847 root->left = dummy.right;
848 root->right = dummy.left;
853 * vm_page_insert: [ internal use only ]
855 * Inserts the given mem entry into the object and object list.
857 * The pagetables are not updated but will presumably fault the page
858 * in if necessary, or if a kernel page the caller will at some point
859 * enter the page into the kernel's pmap. We are not allowed to sleep
860 * here so we *can't* do this anyway.
862 * The object must be locked.
865 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
869 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
870 if (m->object != NULL)
871 panic("vm_page_insert: page already inserted");
874 * Record the object/offset pair in this page
880 * Now link into the object's ordered list of backed pages.
886 TAILQ_INSERT_TAIL(&object->memq, m, listq);
888 root = vm_page_splay(pindex, root);
889 if (pindex < root->pindex) {
890 m->left = root->left;
893 TAILQ_INSERT_BEFORE(root, m, listq);
894 } else if (pindex == root->pindex)
895 panic("vm_page_insert: offset already allocated");
897 m->right = root->right;
900 TAILQ_INSERT_AFTER(&object->memq, root, m, listq);
906 * Show that the object has one more resident page.
908 object->resident_page_count++;
911 * Hold the vnode until the last page is released.
913 if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
914 vhold(object->handle);
917 * Since we are inserting a new and possibly dirty page,
918 * update the object's OBJ_MIGHTBEDIRTY flag.
920 if (pmap_page_is_write_mapped(m))
921 vm_object_set_writeable_dirty(object);
927 * Removes the given mem entry from the object/offset-page
928 * table and the object page list, but do not invalidate/terminate
931 * The underlying pmap entry (if any) is NOT removed here.
933 * The object must be locked. The page must be locked if it is managed.
936 vm_page_remove(vm_page_t m)
939 vm_page_t next, prev, root;
941 if ((m->oflags & VPO_UNMANAGED) == 0)
942 vm_page_lock_assert(m, MA_OWNED);
943 if ((object = m->object) == NULL)
945 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
946 if (m->oflags & VPO_BUSY) {
947 m->oflags &= ~VPO_BUSY;
952 * Now remove from the object's list of backed pages.
954 if ((next = TAILQ_NEXT(m, listq)) != NULL && next->left == m) {
956 * Since the page's successor in the list is also its parent
957 * in the tree, its right subtree must be empty.
959 next->left = m->left;
960 KASSERT(m->right == NULL,
961 ("vm_page_remove: page %p has right child", m));
962 } else if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
965 * Since the page's predecessor in the list is also its parent
966 * in the tree, its left subtree must be empty.
968 KASSERT(m->left == NULL,
969 ("vm_page_remove: page %p has left child", m));
970 prev->right = m->right;
972 if (m != object->root)
973 vm_page_splay(m->pindex, object->root);
976 else if (m->right == NULL)
980 * Move the page's successor to the root, because
981 * pages are usually removed in ascending order.
983 if (m->right != next)
984 vm_page_splay(m->pindex, m->right);
985 next->left = m->left;
990 TAILQ_REMOVE(&object->memq, m, listq);
993 * And show that the object has one fewer resident page.
995 object->resident_page_count--;
998 * The vnode may now be recycled.
1000 if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1001 vdrop(object->handle);
1009 * Returns the page associated with the object/offset
1010 * pair specified; if none is found, NULL is returned.
1012 * The object must be locked.
1015 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1019 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1020 if ((m = object->root) != NULL && m->pindex != pindex) {
1021 m = vm_page_splay(pindex, m);
1022 if ((object->root = m)->pindex != pindex)
1029 * vm_page_find_least:
1031 * Returns the page associated with the object with least pindex
1032 * greater than or equal to the parameter pindex, or NULL.
1034 * The object must be locked.
1037 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1041 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1042 if ((m = TAILQ_FIRST(&object->memq)) != NULL) {
1043 if (m->pindex < pindex) {
1044 m = vm_page_splay(pindex, object->root);
1045 if ((object->root = m)->pindex < pindex)
1046 m = TAILQ_NEXT(m, listq);
1053 * Returns the given page's successor (by pindex) within the object if it is
1054 * resident; if none is found, NULL is returned.
1056 * The object must be locked.
1059 vm_page_next(vm_page_t m)
1063 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1064 if ((next = TAILQ_NEXT(m, listq)) != NULL &&
1065 next->pindex != m->pindex + 1)
1071 * Returns the given page's predecessor (by pindex) within the object if it is
1072 * resident; if none is found, NULL is returned.
1074 * The object must be locked.
1077 vm_page_prev(vm_page_t m)
1081 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1082 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
1083 prev->pindex != m->pindex - 1)
1091 * Move the given memory entry from its
1092 * current object to the specified target object/offset.
1094 * Note: swap associated with the page must be invalidated by the move. We
1095 * have to do this for several reasons: (1) we aren't freeing the
1096 * page, (2) we are dirtying the page, (3) the VM system is probably
1097 * moving the page from object A to B, and will then later move
1098 * the backing store from A to B and we can't have a conflict.
1100 * Note: we *always* dirty the page. It is necessary both for the
1101 * fact that we moved it, and because we may be invalidating
1102 * swap. If the page is on the cache, we have to deactivate it
1103 * or vm_page_dirty() will panic. Dirty pages are not allowed
1106 * The objects must be locked. The page must be locked if it is managed.
1109 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1113 vm_page_insert(m, new_object, new_pindex);
1118 * Convert all of the given object's cached pages that have a
1119 * pindex within the given range into free pages. If the value
1120 * zero is given for "end", then the range's upper bound is
1121 * infinity. If the given object is backed by a vnode and it
1122 * transitions from having one or more cached pages to none, the
1123 * vnode's hold count is reduced.
1126 vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1128 vm_page_t m, m_next;
1131 mtx_lock(&vm_page_queue_free_mtx);
1132 if (__predict_false(object->cache == NULL)) {
1133 mtx_unlock(&vm_page_queue_free_mtx);
1136 m = object->cache = vm_page_splay(start, object->cache);
1137 if (m->pindex < start) {
1138 if (m->right == NULL)
1141 m_next = vm_page_splay(start, m->right);
1144 m = object->cache = m_next;
1149 * At this point, "m" is either (1) a reference to the page
1150 * with the least pindex that is greater than or equal to
1151 * "start" or (2) NULL.
1153 for (; m != NULL && (m->pindex < end || end == 0); m = m_next) {
1155 * Find "m"'s successor and remove "m" from the
1158 if (m->right == NULL) {
1159 object->cache = m->left;
1162 m_next = vm_page_splay(start, m->right);
1163 m_next->left = m->left;
1164 object->cache = m_next;
1166 /* Convert "m" to a free page. */
1169 /* Clear PG_CACHED and set PG_FREE. */
1170 m->flags ^= PG_CACHED | PG_FREE;
1171 KASSERT((m->flags & (PG_CACHED | PG_FREE)) == PG_FREE,
1172 ("vm_page_cache_free: page %p has inconsistent flags", m));
1173 cnt.v_cache_count--;
1176 empty = object->cache == NULL;
1177 mtx_unlock(&vm_page_queue_free_mtx);
1178 if (object->type == OBJT_VNODE && empty)
1179 vdrop(object->handle);
1183 * Returns the cached page that is associated with the given
1184 * object and offset. If, however, none exists, returns NULL.
1186 * The free page queue must be locked.
1188 static inline vm_page_t
1189 vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex)
1193 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1194 if ((m = object->cache) != NULL && m->pindex != pindex) {
1195 m = vm_page_splay(pindex, m);
1196 if ((object->cache = m)->pindex != pindex)
1203 * Remove the given cached page from its containing object's
1204 * collection of cached pages.
1206 * The free page queue must be locked.
1209 vm_page_cache_remove(vm_page_t m)
1214 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1215 KASSERT((m->flags & PG_CACHED) != 0,
1216 ("vm_page_cache_remove: page %p is not cached", m));
1218 if (m != object->cache) {
1219 root = vm_page_splay(m->pindex, object->cache);
1221 ("vm_page_cache_remove: page %p is not cached in object %p",
1224 if (m->left == NULL)
1226 else if (m->right == NULL)
1229 root = vm_page_splay(m->pindex, m->left);
1230 root->right = m->right;
1232 object->cache = root;
1234 cnt.v_cache_count--;
1238 * Transfer all of the cached pages with offset greater than or
1239 * equal to 'offidxstart' from the original object's cache to the
1240 * new object's cache. However, any cached pages with offset
1241 * greater than or equal to the new object's size are kept in the
1242 * original object. Initially, the new object's cache must be
1243 * empty. Offset 'offidxstart' in the original object must
1244 * correspond to offset zero in the new object.
1246 * The new object must be locked.
1249 vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart,
1250 vm_object_t new_object)
1252 vm_page_t m, m_next;
1255 * Insertion into an object's collection of cached pages
1256 * requires the object to be locked. In contrast, removal does
1259 VM_OBJECT_LOCK_ASSERT(new_object, MA_OWNED);
1260 KASSERT(new_object->cache == NULL,
1261 ("vm_page_cache_transfer: object %p has cached pages",
1263 mtx_lock(&vm_page_queue_free_mtx);
1264 if ((m = orig_object->cache) != NULL) {
1266 * Transfer all of the pages with offset greater than or
1267 * equal to 'offidxstart' from the original object's
1268 * cache to the new object's cache.
1270 m = vm_page_splay(offidxstart, m);
1271 if (m->pindex < offidxstart) {
1272 orig_object->cache = m;
1273 new_object->cache = m->right;
1276 orig_object->cache = m->left;
1277 new_object->cache = m;
1280 while ((m = new_object->cache) != NULL) {
1281 if ((m->pindex - offidxstart) >= new_object->size) {
1283 * Return all of the cached pages with
1284 * offset greater than or equal to the
1285 * new object's size to the original
1288 new_object->cache = m->left;
1289 m->left = orig_object->cache;
1290 orig_object->cache = m;
1293 m_next = vm_page_splay(m->pindex, m->right);
1294 /* Update the page's object and offset. */
1295 m->object = new_object;
1296 m->pindex -= offidxstart;
1301 new_object->cache = m_next;
1303 KASSERT(new_object->cache == NULL ||
1304 new_object->type == OBJT_SWAP,
1305 ("vm_page_cache_transfer: object %p's type is incompatible"
1306 " with cached pages", new_object));
1308 mtx_unlock(&vm_page_queue_free_mtx);
1312 * Returns TRUE if a cached page is associated with the given object and
1313 * offset, and FALSE otherwise.
1315 * The object must be locked.
1318 vm_page_is_cached(vm_object_t object, vm_pindex_t pindex)
1323 * Insertion into an object's collection of cached pages requires the
1324 * object to be locked. Therefore, if the object is locked and the
1325 * object's collection is empty, there is no need to acquire the free
1326 * page queues lock in order to prove that the specified page doesn't
1329 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1330 if (__predict_true(object->cache == NULL))
1332 mtx_lock(&vm_page_queue_free_mtx);
1333 m = vm_page_cache_lookup(object, pindex);
1334 mtx_unlock(&vm_page_queue_free_mtx);
1341 * Allocate and return a page that is associated with the specified
1342 * object and offset pair. By default, this page has the flag VPO_BUSY
1345 * The caller must always specify an allocation class.
1347 * allocation classes:
1348 * VM_ALLOC_NORMAL normal process request
1349 * VM_ALLOC_SYSTEM system *really* needs a page
1350 * VM_ALLOC_INTERRUPT interrupt time request
1352 * optional allocation flags:
1353 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1354 * intends to allocate
1355 * VM_ALLOC_IFCACHED return page only if it is cached
1356 * VM_ALLOC_IFNOTCACHED return NULL, do not reactivate if the page
1358 * VM_ALLOC_NOBUSY do not set the flag VPO_BUSY on the page
1359 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1360 * VM_ALLOC_NOOBJ page is not associated with an object and
1361 * should not have the flag VPO_BUSY set
1362 * VM_ALLOC_WIRED wire the allocated page
1363 * VM_ALLOC_ZERO prefer a zeroed page
1365 * This routine may not sleep.
1368 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1370 struct vnode *vp = NULL;
1371 vm_object_t m_object;
1373 int flags, req_class;
1375 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0),
1376 ("vm_page_alloc: inconsistent object/req"));
1378 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1380 req_class = req & VM_ALLOC_CLASS_MASK;
1383 * The page daemon is allowed to dig deeper into the free page list.
1385 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1386 req_class = VM_ALLOC_SYSTEM;
1388 mtx_lock(&vm_page_queue_free_mtx);
1389 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved ||
1390 (req_class == VM_ALLOC_SYSTEM &&
1391 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) ||
1392 (req_class == VM_ALLOC_INTERRUPT &&
1393 cnt.v_free_count + cnt.v_cache_count > 0)) {
1395 * Allocate from the free queue if the number of free pages
1396 * exceeds the minimum for the request class.
1398 if (object != NULL &&
1399 (m = vm_page_cache_lookup(object, pindex)) != NULL) {
1400 if ((req & VM_ALLOC_IFNOTCACHED) != 0) {
1401 mtx_unlock(&vm_page_queue_free_mtx);
1404 if (vm_phys_unfree_page(m))
1405 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0);
1406 #if VM_NRESERVLEVEL > 0
1407 else if (!vm_reserv_reactivate_page(m))
1411 panic("vm_page_alloc: cache page %p is missing"
1412 " from the free queue", m);
1413 } else if ((req & VM_ALLOC_IFCACHED) != 0) {
1414 mtx_unlock(&vm_page_queue_free_mtx);
1416 #if VM_NRESERVLEVEL > 0
1417 } else if (object == NULL || (object->flags & (OBJ_COLORED |
1418 OBJ_FICTITIOUS)) != OBJ_COLORED ||
1419 (m = vm_reserv_alloc_page(object, pindex)) == NULL) {
1423 m = vm_phys_alloc_pages(object != NULL ?
1424 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1425 #if VM_NRESERVLEVEL > 0
1426 if (m == NULL && vm_reserv_reclaim_inactive()) {
1427 m = vm_phys_alloc_pages(object != NULL ?
1428 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1435 * Not allocatable, give up.
1437 mtx_unlock(&vm_page_queue_free_mtx);
1438 atomic_add_int(&vm_pageout_deficit,
1439 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1440 pagedaemon_wakeup();
1445 * At this point we had better have found a good page.
1447 KASSERT(m != NULL, ("vm_page_alloc: missing page"));
1448 KASSERT(m->queue == PQ_NONE,
1449 ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue));
1450 KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m));
1451 KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m));
1452 KASSERT(m->busy == 0, ("vm_page_alloc: page %p is busy", m));
1453 KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m));
1454 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1455 ("vm_page_alloc: page %p has unexpected memattr %d", m,
1456 pmap_page_get_memattr(m)));
1457 if ((m->flags & PG_CACHED) != 0) {
1458 KASSERT((m->flags & PG_ZERO) == 0,
1459 ("vm_page_alloc: cached page %p is PG_ZERO", m));
1460 KASSERT(m->valid != 0,
1461 ("vm_page_alloc: cached page %p is invalid", m));
1462 if (m->object == object && m->pindex == pindex)
1463 cnt.v_reactivated++;
1466 m_object = m->object;
1467 vm_page_cache_remove(m);
1468 if (m_object->type == OBJT_VNODE && m_object->cache == NULL)
1469 vp = m_object->handle;
1471 KASSERT(VM_PAGE_IS_FREE(m),
1472 ("vm_page_alloc: page %p is not free", m));
1473 KASSERT(m->valid == 0,
1474 ("vm_page_alloc: free page %p is valid", m));
1479 * Only the PG_ZERO flag is inherited. The PG_CACHED or PG_FREE flag
1480 * must be cleared before the free page queues lock is released.
1483 if (m->flags & PG_ZERO) {
1484 vm_page_zero_count--;
1485 if (req & VM_ALLOC_ZERO)
1488 if (req & VM_ALLOC_NODUMP)
1491 mtx_unlock(&vm_page_queue_free_mtx);
1493 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1495 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ)) == 0)
1496 m->oflags |= VPO_BUSY;
1497 if (req & VM_ALLOC_WIRED) {
1499 * The page lock is not required for wiring a page until that
1500 * page is inserted into the object.
1502 atomic_add_int(&cnt.v_wire_count, 1);
1507 if (object != NULL) {
1508 /* Ignore device objects; the pager sets "memattr" for them. */
1509 if (object->memattr != VM_MEMATTR_DEFAULT &&
1510 (object->flags & OBJ_FICTITIOUS) == 0)
1511 pmap_page_set_memattr(m, object->memattr);
1512 vm_page_insert(m, object, pindex);
1517 * The following call to vdrop() must come after the above call
1518 * to vm_page_insert() in case both affect the same object and
1519 * vnode. Otherwise, the affected vnode's hold count could
1520 * temporarily become zero.
1526 * Don't wakeup too often - wakeup the pageout daemon when
1527 * we would be nearly out of memory.
1529 if (vm_paging_needed())
1530 pagedaemon_wakeup();
1536 * vm_page_alloc_contig:
1538 * Allocate a contiguous set of physical pages of the given size "npages"
1539 * from the free lists. All of the physical pages must be at or above
1540 * the given physical address "low" and below the given physical address
1541 * "high". The given value "alignment" determines the alignment of the
1542 * first physical page in the set. If the given value "boundary" is
1543 * non-zero, then the set of physical pages cannot cross any physical
1544 * address boundary that is a multiple of that value. Both "alignment"
1545 * and "boundary" must be a power of two.
1547 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1548 * then the memory attribute setting for the physical pages is configured
1549 * to the object's memory attribute setting. Otherwise, the memory
1550 * attribute setting for the physical pages is configured to "memattr",
1551 * overriding the object's memory attribute setting. However, if the
1552 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1553 * memory attribute setting for the physical pages cannot be configured
1554 * to VM_MEMATTR_DEFAULT.
1556 * The caller must always specify an allocation class.
1558 * allocation classes:
1559 * VM_ALLOC_NORMAL normal process request
1560 * VM_ALLOC_SYSTEM system *really* needs a page
1561 * VM_ALLOC_INTERRUPT interrupt time request
1563 * optional allocation flags:
1564 * VM_ALLOC_NOBUSY do not set the flag VPO_BUSY on the page
1565 * VM_ALLOC_NOOBJ page is not associated with an object and
1566 * should not have the flag VPO_BUSY set
1567 * VM_ALLOC_WIRED wire the allocated page
1568 * VM_ALLOC_ZERO prefer a zeroed page
1570 * This routine may not sleep.
1573 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1574 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1575 vm_paddr_t boundary, vm_memattr_t memattr)
1578 vm_page_t deferred_vdrop_list, m, m_ret;
1579 u_int flags, oflags;
1582 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0),
1583 ("vm_page_alloc_contig: inconsistent object/req"));
1584 if (object != NULL) {
1585 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1586 KASSERT(object->type == OBJT_PHYS,
1587 ("vm_page_alloc_contig: object %p isn't OBJT_PHYS",
1590 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
1591 req_class = req & VM_ALLOC_CLASS_MASK;
1594 * The page daemon is allowed to dig deeper into the free page list.
1596 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1597 req_class = VM_ALLOC_SYSTEM;
1599 deferred_vdrop_list = NULL;
1600 mtx_lock(&vm_page_queue_free_mtx);
1601 if (cnt.v_free_count + cnt.v_cache_count >= npages +
1602 cnt.v_free_reserved || (req_class == VM_ALLOC_SYSTEM &&
1603 cnt.v_free_count + cnt.v_cache_count >= npages +
1604 cnt.v_interrupt_free_min) || (req_class == VM_ALLOC_INTERRUPT &&
1605 cnt.v_free_count + cnt.v_cache_count >= npages)) {
1606 #if VM_NRESERVLEVEL > 0
1608 if (object == NULL || (object->flags & OBJ_COLORED) == 0 ||
1609 (m_ret = vm_reserv_alloc_contig(object, pindex, npages,
1610 low, high, alignment, boundary)) == NULL)
1612 m_ret = vm_phys_alloc_contig(npages, low, high,
1613 alignment, boundary);
1615 mtx_unlock(&vm_page_queue_free_mtx);
1616 atomic_add_int(&vm_pageout_deficit, npages);
1617 pagedaemon_wakeup();
1621 for (m = m_ret; m < &m_ret[npages]; m++) {
1622 drop = vm_page_alloc_init(m);
1625 * Enqueue the vnode for deferred vdrop().
1627 * Once the pages are removed from the free
1628 * page list, "pageq" can be safely abused to
1629 * construct a short-lived list of vnodes.
1631 m->pageq.tqe_prev = (void *)drop;
1632 m->pageq.tqe_next = deferred_vdrop_list;
1633 deferred_vdrop_list = m;
1637 #if VM_NRESERVLEVEL > 0
1638 if (vm_reserv_reclaim_contig(npages, low, high, alignment,
1643 mtx_unlock(&vm_page_queue_free_mtx);
1648 * Initialize the pages. Only the PG_ZERO flag is inherited.
1651 if ((req & VM_ALLOC_ZERO) != 0)
1653 if ((req & VM_ALLOC_NODUMP) != 0)
1655 if ((req & VM_ALLOC_WIRED) != 0)
1656 atomic_add_int(&cnt.v_wire_count, npages);
1657 oflags = VPO_UNMANAGED;
1658 if (object != NULL) {
1659 if ((req & VM_ALLOC_NOBUSY) == 0)
1661 if (object->memattr != VM_MEMATTR_DEFAULT &&
1662 memattr == VM_MEMATTR_DEFAULT)
1663 memattr = object->memattr;
1665 for (m = m_ret; m < &m_ret[npages]; m++) {
1667 m->flags = (m->flags | PG_NODUMP) & flags;
1668 if ((req & VM_ALLOC_WIRED) != 0)
1670 /* Unmanaged pages don't use "act_count". */
1672 if (memattr != VM_MEMATTR_DEFAULT)
1673 pmap_page_set_memattr(m, memattr);
1675 vm_page_insert(m, object, pindex);
1680 while (deferred_vdrop_list != NULL) {
1681 vdrop((struct vnode *)deferred_vdrop_list->pageq.tqe_prev);
1682 deferred_vdrop_list = deferred_vdrop_list->pageq.tqe_next;
1684 if (vm_paging_needed())
1685 pagedaemon_wakeup();
1690 * Initialize a page that has been freshly dequeued from a freelist.
1691 * The caller has to drop the vnode returned, if it is not NULL.
1693 * This function may only be used to initialize unmanaged pages.
1695 * To be called with vm_page_queue_free_mtx held.
1697 static struct vnode *
1698 vm_page_alloc_init(vm_page_t m)
1701 vm_object_t m_object;
1703 KASSERT(m->queue == PQ_NONE,
1704 ("vm_page_alloc_init: page %p has unexpected queue %d",
1706 KASSERT(m->wire_count == 0,
1707 ("vm_page_alloc_init: page %p is wired", m));
1708 KASSERT(m->hold_count == 0,
1709 ("vm_page_alloc_init: page %p is held", m));
1710 KASSERT(m->busy == 0,
1711 ("vm_page_alloc_init: page %p is busy", m));
1712 KASSERT(m->dirty == 0,
1713 ("vm_page_alloc_init: page %p is dirty", m));
1714 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1715 ("vm_page_alloc_init: page %p has unexpected memattr %d",
1716 m, pmap_page_get_memattr(m)));
1717 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1719 if ((m->flags & PG_CACHED) != 0) {
1720 KASSERT((m->flags & PG_ZERO) == 0,
1721 ("vm_page_alloc_init: cached page %p is PG_ZERO", m));
1723 m_object = m->object;
1724 vm_page_cache_remove(m);
1725 if (m_object->type == OBJT_VNODE && m_object->cache == NULL)
1726 drop = m_object->handle;
1728 KASSERT(VM_PAGE_IS_FREE(m),
1729 ("vm_page_alloc_init: page %p is not free", m));
1730 KASSERT(m->valid == 0,
1731 ("vm_page_alloc_init: free page %p is valid", m));
1733 if ((m->flags & PG_ZERO) != 0)
1734 vm_page_zero_count--;
1736 /* Don't clear the PG_ZERO flag; we'll need it later. */
1737 m->flags &= PG_ZERO;
1742 * vm_page_alloc_freelist:
1744 * Allocate a physical page from the specified free page list.
1746 * The caller must always specify an allocation class.
1748 * allocation classes:
1749 * VM_ALLOC_NORMAL normal process request
1750 * VM_ALLOC_SYSTEM system *really* needs a page
1751 * VM_ALLOC_INTERRUPT interrupt time request
1753 * optional allocation flags:
1754 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1755 * intends to allocate
1756 * VM_ALLOC_WIRED wire the allocated page
1757 * VM_ALLOC_ZERO prefer a zeroed page
1759 * This routine may not sleep.
1762 vm_page_alloc_freelist(int flind, int req)
1769 req_class = req & VM_ALLOC_CLASS_MASK;
1772 * The page daemon is allowed to dig deeper into the free page list.
1774 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1775 req_class = VM_ALLOC_SYSTEM;
1778 * Do not allocate reserved pages unless the req has asked for it.
1780 mtx_lock(&vm_page_queue_free_mtx);
1781 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved ||
1782 (req_class == VM_ALLOC_SYSTEM &&
1783 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) ||
1784 (req_class == VM_ALLOC_INTERRUPT &&
1785 cnt.v_free_count + cnt.v_cache_count > 0))
1786 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0);
1788 mtx_unlock(&vm_page_queue_free_mtx);
1789 atomic_add_int(&vm_pageout_deficit,
1790 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1791 pagedaemon_wakeup();
1795 mtx_unlock(&vm_page_queue_free_mtx);
1798 drop = vm_page_alloc_init(m);
1799 mtx_unlock(&vm_page_queue_free_mtx);
1802 * Initialize the page. Only the PG_ZERO flag is inherited.
1806 if ((req & VM_ALLOC_ZERO) != 0)
1809 if ((req & VM_ALLOC_WIRED) != 0) {
1811 * The page lock is not required for wiring a page that does
1812 * not belong to an object.
1814 atomic_add_int(&cnt.v_wire_count, 1);
1817 /* Unmanaged pages don't use "act_count". */
1818 m->oflags = VPO_UNMANAGED;
1821 if (vm_paging_needed())
1822 pagedaemon_wakeup();
1827 * vm_wait: (also see VM_WAIT macro)
1829 * Sleep until free pages are available for allocation.
1830 * - Called in various places before memory allocations.
1836 mtx_lock(&vm_page_queue_free_mtx);
1837 if (curproc == pageproc) {
1838 vm_pageout_pages_needed = 1;
1839 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx,
1840 PDROP | PSWP, "VMWait", 0);
1842 if (!vm_pages_needed) {
1843 vm_pages_needed = 1;
1844 wakeup(&vm_pages_needed);
1846 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM,
1852 * vm_waitpfault: (also see VM_WAITPFAULT macro)
1854 * Sleep until free pages are available for allocation.
1855 * - Called only in vm_fault so that processes page faulting
1856 * can be easily tracked.
1857 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
1858 * processes will be able to grab memory first. Do not change
1859 * this balance without careful testing first.
1865 mtx_lock(&vm_page_queue_free_mtx);
1866 if (!vm_pages_needed) {
1867 vm_pages_needed = 1;
1868 wakeup(&vm_pages_needed);
1870 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER,
1877 * Remove the given page from its current page queue.
1879 * The page must be locked.
1882 vm_page_dequeue(vm_page_t m)
1884 struct vm_pagequeue *pq;
1886 vm_page_lock_assert(m, MA_OWNED);
1887 KASSERT(m->queue != PQ_NONE,
1888 ("vm_page_dequeue: page %p is not queued", m));
1889 pq = &vm_pagequeues[m->queue];
1890 vm_pagequeue_lock(pq);
1892 TAILQ_REMOVE(&pq->pq_pl, m, pageq);
1894 vm_pagequeue_unlock(pq);
1898 * vm_page_dequeue_locked:
1900 * Remove the given page from its current page queue.
1902 * The page and page queue must be locked.
1905 vm_page_dequeue_locked(vm_page_t m)
1907 struct vm_pagequeue *pq;
1909 vm_page_lock_assert(m, MA_OWNED);
1910 pq = &vm_pagequeues[m->queue];
1911 vm_pagequeue_assert_locked(pq);
1913 TAILQ_REMOVE(&pq->pq_pl, m, pageq);
1920 * Add the given page to the specified page queue.
1922 * The page must be locked.
1925 vm_page_enqueue(int queue, vm_page_t m)
1927 struct vm_pagequeue *pq;
1929 vm_page_lock_assert(m, MA_OWNED);
1930 pq = &vm_pagequeues[queue];
1931 vm_pagequeue_lock(pq);
1933 TAILQ_INSERT_TAIL(&pq->pq_pl, m, pageq);
1935 vm_pagequeue_unlock(pq);
1941 * Move the given page to the tail of its current page queue.
1943 * The page must be locked.
1946 vm_page_requeue(vm_page_t m)
1948 struct vm_pagequeue *pq;
1950 vm_page_lock_assert(m, MA_OWNED);
1951 KASSERT(m->queue != PQ_NONE,
1952 ("vm_page_requeue: page %p is not queued", m));
1953 pq = &vm_pagequeues[m->queue];
1954 vm_pagequeue_lock(pq);
1955 TAILQ_REMOVE(&pq->pq_pl, m, pageq);
1956 TAILQ_INSERT_TAIL(&pq->pq_pl, m, pageq);
1957 vm_pagequeue_unlock(pq);
1961 * vm_page_requeue_locked:
1963 * Move the given page to the tail of its current page queue.
1965 * The page queue must be locked.
1968 vm_page_requeue_locked(vm_page_t m)
1970 struct vm_pagequeue *pq;
1972 KASSERT(m->queue != PQ_NONE,
1973 ("vm_page_requeue_locked: page %p is not queued", m));
1974 pq = &vm_pagequeues[m->queue];
1975 vm_pagequeue_assert_locked(pq);
1976 TAILQ_REMOVE(&pq->pq_pl, m, pageq);
1977 TAILQ_INSERT_TAIL(&pq->pq_pl, m, pageq);
1983 * Put the specified page on the active list (if appropriate).
1984 * Ensure that act_count is at least ACT_INIT but do not otherwise
1987 * The page must be locked.
1990 vm_page_activate(vm_page_t m)
1994 vm_page_lock_assert(m, MA_OWNED);
1995 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1996 if ((queue = m->queue) != PQ_ACTIVE) {
1997 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
1998 if (m->act_count < ACT_INIT)
1999 m->act_count = ACT_INIT;
2000 if (queue != PQ_NONE)
2002 vm_page_enqueue(PQ_ACTIVE, m);
2004 KASSERT(queue == PQ_NONE,
2005 ("vm_page_activate: wired page %p is queued", m));
2007 if (m->act_count < ACT_INIT)
2008 m->act_count = ACT_INIT;
2013 * vm_page_free_wakeup:
2015 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
2016 * routine is called when a page has been added to the cache or free
2019 * The page queues must be locked.
2022 vm_page_free_wakeup(void)
2025 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2027 * if pageout daemon needs pages, then tell it that there are
2030 if (vm_pageout_pages_needed &&
2031 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) {
2032 wakeup(&vm_pageout_pages_needed);
2033 vm_pageout_pages_needed = 0;
2036 * wakeup processes that are waiting on memory if we hit a
2037 * high water mark. And wakeup scheduler process if we have
2038 * lots of memory. this process will swapin processes.
2040 if (vm_pages_needed && !vm_page_count_min()) {
2041 vm_pages_needed = 0;
2042 wakeup(&cnt.v_free_count);
2049 * Returns the given page to the free list,
2050 * disassociating it with any VM object.
2052 * The object must be locked. The page must be locked if it is managed.
2055 vm_page_free_toq(vm_page_t m)
2058 if ((m->oflags & VPO_UNMANAGED) == 0) {
2059 vm_page_lock_assert(m, MA_OWNED);
2060 KASSERT(!pmap_page_is_mapped(m),
2061 ("vm_page_free_toq: freeing mapped page %p", m));
2063 KASSERT(m->queue == PQ_NONE,
2064 ("vm_page_free_toq: unmanaged page %p is queued", m));
2065 PCPU_INC(cnt.v_tfree);
2067 if (VM_PAGE_IS_FREE(m))
2068 panic("vm_page_free: freeing free page %p", m);
2069 else if (m->busy != 0)
2070 panic("vm_page_free: freeing busy page %p", m);
2073 * Unqueue, then remove page. Note that we cannot destroy
2074 * the page here because we do not want to call the pager's
2075 * callback routine until after we've put the page on the
2076 * appropriate free queue.
2082 * If fictitious remove object association and
2083 * return, otherwise delay object association removal.
2085 if ((m->flags & PG_FICTITIOUS) != 0) {
2092 if (m->wire_count != 0)
2093 panic("vm_page_free: freeing wired page %p", m);
2094 if (m->hold_count != 0) {
2095 m->flags &= ~PG_ZERO;
2096 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2097 ("vm_page_free: freeing PG_UNHOLDFREE page %p", m));
2098 m->flags |= PG_UNHOLDFREE;
2101 * Restore the default memory attribute to the page.
2103 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2104 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2107 * Insert the page into the physical memory allocator's
2108 * cache/free page queues.
2110 mtx_lock(&vm_page_queue_free_mtx);
2111 m->flags |= PG_FREE;
2113 #if VM_NRESERVLEVEL > 0
2114 if (!vm_reserv_free_page(m))
2118 vm_phys_free_pages(m, 0);
2119 if ((m->flags & PG_ZERO) != 0)
2120 ++vm_page_zero_count;
2122 vm_page_zero_idle_wakeup();
2123 vm_page_free_wakeup();
2124 mtx_unlock(&vm_page_queue_free_mtx);
2131 * Mark this page as wired down by yet
2132 * another map, removing it from paging queues
2135 * If the page is fictitious, then its wire count must remain one.
2137 * The page must be locked.
2140 vm_page_wire(vm_page_t m)
2144 * Only bump the wire statistics if the page is not already wired,
2145 * and only unqueue the page if it is on some queue (if it is unmanaged
2146 * it is already off the queues).
2148 vm_page_lock_assert(m, MA_OWNED);
2149 if ((m->flags & PG_FICTITIOUS) != 0) {
2150 KASSERT(m->wire_count == 1,
2151 ("vm_page_wire: fictitious page %p's wire count isn't one",
2155 if (m->wire_count == 0) {
2156 KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
2157 m->queue == PQ_NONE,
2158 ("vm_page_wire: unmanaged page %p is queued", m));
2160 atomic_add_int(&cnt.v_wire_count, 1);
2163 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
2169 * Release one wiring of the specified page, potentially enabling it to be
2170 * paged again. If paging is enabled, then the value of the parameter
2171 * "activate" determines to which queue the page is added. If "activate" is
2172 * non-zero, then the page is added to the active queue. Otherwise, it is
2173 * added to the inactive queue.
2175 * However, unless the page belongs to an object, it is not enqueued because
2176 * it cannot be paged out.
2178 * If a page is fictitious, then its wire count must alway be one.
2180 * A managed page must be locked.
2183 vm_page_unwire(vm_page_t m, int activate)
2186 if ((m->oflags & VPO_UNMANAGED) == 0)
2187 vm_page_lock_assert(m, MA_OWNED);
2188 if ((m->flags & PG_FICTITIOUS) != 0) {
2189 KASSERT(m->wire_count == 1,
2190 ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
2193 if (m->wire_count > 0) {
2195 if (m->wire_count == 0) {
2196 atomic_subtract_int(&cnt.v_wire_count, 1);
2197 if ((m->oflags & VPO_UNMANAGED) != 0 ||
2201 m->flags &= ~PG_WINATCFLS;
2202 vm_page_enqueue(activate ? PQ_ACTIVE : PQ_INACTIVE, m);
2205 panic("vm_page_unwire: page %p's wire count is zero", m);
2209 * Move the specified page to the inactive queue.
2211 * Many pages placed on the inactive queue should actually go
2212 * into the cache, but it is difficult to figure out which. What
2213 * we do instead, if the inactive target is well met, is to put
2214 * clean pages at the head of the inactive queue instead of the tail.
2215 * This will cause them to be moved to the cache more quickly and
2216 * if not actively re-referenced, reclaimed more quickly. If we just
2217 * stick these pages at the end of the inactive queue, heavy filesystem
2218 * meta-data accesses can cause an unnecessary paging load on memory bound
2219 * processes. This optimization causes one-time-use metadata to be
2220 * reused more quickly.
2222 * Normally athead is 0 resulting in LRU operation. athead is set
2223 * to 1 if we want this page to be 'as if it were placed in the cache',
2224 * except without unmapping it from the process address space.
2226 * The page must be locked.
2229 _vm_page_deactivate(vm_page_t m, int athead)
2231 struct vm_pagequeue *pq;
2234 vm_page_lock_assert(m, MA_OWNED);
2237 * Ignore if already inactive.
2239 if ((queue = m->queue) == PQ_INACTIVE)
2241 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2242 if (queue != PQ_NONE)
2244 m->flags &= ~PG_WINATCFLS;
2245 pq = &vm_pagequeues[PQ_INACTIVE];
2246 vm_pagequeue_lock(pq);
2247 m->queue = PQ_INACTIVE;
2249 TAILQ_INSERT_HEAD(&pq->pq_pl, m, pageq);
2251 TAILQ_INSERT_TAIL(&pq->pq_pl, m, pageq);
2252 cnt.v_inactive_count++;
2253 vm_pagequeue_unlock(pq);
2258 * Move the specified page to the inactive queue.
2260 * The page must be locked.
2263 vm_page_deactivate(vm_page_t m)
2266 _vm_page_deactivate(m, 0);
2270 * vm_page_try_to_cache:
2272 * Returns 0 on failure, 1 on success
2275 vm_page_try_to_cache(vm_page_t m)
2278 vm_page_lock_assert(m, MA_OWNED);
2279 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2280 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
2281 (m->oflags & (VPO_BUSY | VPO_UNMANAGED)) != 0)
2291 * vm_page_try_to_free()
2293 * Attempt to free the page. If we cannot free it, we do nothing.
2294 * 1 is returned on success, 0 on failure.
2297 vm_page_try_to_free(vm_page_t m)
2300 vm_page_lock_assert(m, MA_OWNED);
2301 if (m->object != NULL)
2302 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2303 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
2304 (m->oflags & (VPO_BUSY | VPO_UNMANAGED)) != 0)
2316 * Put the specified page onto the page cache queue (if appropriate).
2318 * The object and page must be locked.
2321 vm_page_cache(vm_page_t m)
2324 vm_page_t next, prev, root;
2326 vm_page_lock_assert(m, MA_OWNED);
2328 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2329 if ((m->oflags & (VPO_UNMANAGED | VPO_BUSY)) || m->busy ||
2330 m->hold_count || m->wire_count)
2331 panic("vm_page_cache: attempting to cache busy page");
2332 KASSERT(!pmap_page_is_mapped(m),
2333 ("vm_page_cache: page %p is mapped", m));
2334 KASSERT(m->dirty == 0, ("vm_page_cache: page %p is dirty", m));
2335 if (m->valid == 0 || object->type == OBJT_DEFAULT ||
2336 (object->type == OBJT_SWAP &&
2337 !vm_pager_has_page(object, m->pindex, NULL, NULL))) {
2339 * Hypothesis: A cache-elgible page belonging to a
2340 * default object or swap object but without a backing
2341 * store must be zero filled.
2346 KASSERT((m->flags & PG_CACHED) == 0,
2347 ("vm_page_cache: page %p is already cached", m));
2348 PCPU_INC(cnt.v_tcached);
2351 * Remove the page from the paging queues.
2356 * Remove the page from the object's collection of resident
2359 if ((next = TAILQ_NEXT(m, listq)) != NULL && next->left == m) {
2361 * Since the page's successor in the list is also its parent
2362 * in the tree, its right subtree must be empty.
2364 next->left = m->left;
2365 KASSERT(m->right == NULL,
2366 ("vm_page_cache: page %p has right child", m));
2367 } else if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
2370 * Since the page's predecessor in the list is also its parent
2371 * in the tree, its left subtree must be empty.
2373 KASSERT(m->left == NULL,
2374 ("vm_page_cache: page %p has left child", m));
2375 prev->right = m->right;
2377 if (m != object->root)
2378 vm_page_splay(m->pindex, object->root);
2379 if (m->left == NULL)
2381 else if (m->right == NULL)
2385 * Move the page's successor to the root, because
2386 * pages are usually removed in ascending order.
2388 if (m->right != next)
2389 vm_page_splay(m->pindex, m->right);
2390 next->left = m->left;
2393 object->root = root;
2395 TAILQ_REMOVE(&object->memq, m, listq);
2396 object->resident_page_count--;
2399 * Restore the default memory attribute to the page.
2401 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2402 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2405 * Insert the page into the object's collection of cached pages
2406 * and the physical memory allocator's cache/free page queues.
2408 m->flags &= ~PG_ZERO;
2409 mtx_lock(&vm_page_queue_free_mtx);
2410 m->flags |= PG_CACHED;
2411 cnt.v_cache_count++;
2412 root = object->cache;
2417 root = vm_page_splay(m->pindex, root);
2418 if (m->pindex < root->pindex) {
2419 m->left = root->left;
2422 } else if (__predict_false(m->pindex == root->pindex))
2423 panic("vm_page_cache: offset already cached");
2425 m->right = root->right;
2431 #if VM_NRESERVLEVEL > 0
2432 if (!vm_reserv_free_page(m)) {
2436 vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0);
2437 vm_phys_free_pages(m, 0);
2439 vm_page_free_wakeup();
2440 mtx_unlock(&vm_page_queue_free_mtx);
2443 * Increment the vnode's hold count if this is the object's only
2444 * cached page. Decrement the vnode's hold count if this was
2445 * the object's only resident page.
2447 if (object->type == OBJT_VNODE) {
2448 if (root == NULL && object->resident_page_count != 0)
2449 vhold(object->handle);
2450 else if (root != NULL && object->resident_page_count == 0)
2451 vdrop(object->handle);
2458 * Cache, deactivate, or do nothing as appropriate. This routine
2459 * is typically used by madvise() MADV_DONTNEED.
2461 * Generally speaking we want to move the page into the cache so
2462 * it gets reused quickly. However, this can result in a silly syndrome
2463 * due to the page recycling too quickly. Small objects will not be
2464 * fully cached. On the otherhand, if we move the page to the inactive
2465 * queue we wind up with a problem whereby very large objects
2466 * unnecessarily blow away our inactive and cache queues.
2468 * The solution is to move the pages based on a fixed weighting. We
2469 * either leave them alone, deactivate them, or move them to the cache,
2470 * where moving them to the cache has the highest weighting.
2471 * By forcing some pages into other queues we eventually force the
2472 * system to balance the queues, potentially recovering other unrelated
2473 * space from active. The idea is to not force this to happen too
2476 * The object and page must be locked.
2479 vm_page_dontneed(vm_page_t m)
2484 vm_page_lock_assert(m, MA_OWNED);
2485 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2486 dnw = PCPU_GET(dnweight);
2490 * Occasionally leave the page alone.
2492 if ((dnw & 0x01F0) == 0 || m->queue == PQ_INACTIVE) {
2493 if (m->act_count >= ACT_INIT)
2499 * Clear any references to the page. Otherwise, the page daemon will
2500 * immediately reactivate the page.
2502 * Perform the pmap_clear_reference() first. Otherwise, a concurrent
2503 * pmap operation, such as pmap_remove(), could clear a reference in
2504 * the pmap and set PGA_REFERENCED on the page before the
2505 * pmap_clear_reference() had completed. Consequently, the page would
2506 * appear referenced based upon an old reference that occurred before
2507 * this function ran.
2509 pmap_clear_reference(m);
2510 vm_page_aflag_clear(m, PGA_REFERENCED);
2512 if (m->dirty == 0 && pmap_is_modified(m))
2515 if (m->dirty || (dnw & 0x0070) == 0) {
2517 * Deactivate the page 3 times out of 32.
2522 * Cache the page 28 times out of every 32. Note that
2523 * the page is deactivated instead of cached, but placed
2524 * at the head of the queue instead of the tail.
2528 _vm_page_deactivate(m, head);
2532 * Grab a page, waiting until we are waken up due to the page
2533 * changing state. We keep on waiting, if the page continues
2534 * to be in the object. If the page doesn't exist, first allocate it
2535 * and then conditionally zero it.
2537 * The caller must always specify the VM_ALLOC_RETRY flag. This is intended
2538 * to facilitate its eventual removal.
2540 * This routine may sleep.
2542 * The object must be locked on entry. The lock will, however, be released
2543 * and reacquired if the routine sleeps.
2546 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2550 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2551 KASSERT((allocflags & VM_ALLOC_RETRY) != 0,
2552 ("vm_page_grab: VM_ALLOC_RETRY is required"));
2554 if ((m = vm_page_lookup(object, pindex)) != NULL) {
2555 if ((m->oflags & VPO_BUSY) != 0 ||
2556 ((allocflags & VM_ALLOC_IGN_SBUSY) == 0 && m->busy != 0)) {
2558 * Reference the page before unlocking and
2559 * sleeping so that the page daemon is less
2560 * likely to reclaim it.
2562 vm_page_aflag_set(m, PGA_REFERENCED);
2563 vm_page_sleep(m, "pgrbwt");
2566 if ((allocflags & VM_ALLOC_WIRED) != 0) {
2571 if ((allocflags & VM_ALLOC_NOBUSY) == 0)
2576 m = vm_page_alloc(object, pindex, allocflags & ~(VM_ALLOC_RETRY |
2577 VM_ALLOC_IGN_SBUSY));
2579 VM_OBJECT_UNLOCK(object);
2581 VM_OBJECT_LOCK(object);
2583 } else if (m->valid != 0)
2585 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
2591 * Mapping function for valid or dirty bits in a page.
2593 * Inputs are required to range within a page.
2596 vm_page_bits(int base, int size)
2602 base + size <= PAGE_SIZE,
2603 ("vm_page_bits: illegal base/size %d/%d", base, size)
2606 if (size == 0) /* handle degenerate case */
2609 first_bit = base >> DEV_BSHIFT;
2610 last_bit = (base + size - 1) >> DEV_BSHIFT;
2612 return (((vm_page_bits_t)2 << last_bit) -
2613 ((vm_page_bits_t)1 << first_bit));
2617 * vm_page_set_valid_range:
2619 * Sets portions of a page valid. The arguments are expected
2620 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2621 * of any partial chunks touched by the range. The invalid portion of
2622 * such chunks will be zeroed.
2624 * (base + size) must be less then or equal to PAGE_SIZE.
2627 vm_page_set_valid_range(vm_page_t m, int base, int size)
2631 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2632 if (size == 0) /* handle degenerate case */
2636 * If the base is not DEV_BSIZE aligned and the valid
2637 * bit is clear, we have to zero out a portion of the
2640 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2641 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
2642 pmap_zero_page_area(m, frag, base - frag);
2645 * If the ending offset is not DEV_BSIZE aligned and the
2646 * valid bit is clear, we have to zero out a portion of
2649 endoff = base + size;
2650 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2651 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
2652 pmap_zero_page_area(m, endoff,
2653 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
2656 * Assert that no previously invalid block that is now being validated
2659 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
2660 ("vm_page_set_valid_range: page %p is dirty", m));
2663 * Set valid bits inclusive of any overlap.
2665 m->valid |= vm_page_bits(base, size);
2669 * Clear the given bits from the specified page's dirty field.
2671 static __inline void
2672 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
2675 #if PAGE_SIZE < 16384
2680 * If the object is locked and the page is neither VPO_BUSY nor
2681 * write mapped, then the page's dirty field cannot possibly be
2682 * set by a concurrent pmap operation.
2684 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2685 if ((m->oflags & VPO_BUSY) == 0 && !pmap_page_is_write_mapped(m))
2686 m->dirty &= ~pagebits;
2689 * The pmap layer can call vm_page_dirty() without
2690 * holding a distinguished lock. The combination of
2691 * the object's lock and an atomic operation suffice
2692 * to guarantee consistency of the page dirty field.
2694 * For PAGE_SIZE == 32768 case, compiler already
2695 * properly aligns the dirty field, so no forcible
2696 * alignment is needed. Only require existence of
2697 * atomic_clear_64 when page size is 32768.
2699 addr = (uintptr_t)&m->dirty;
2700 #if PAGE_SIZE == 32768
2701 atomic_clear_64((uint64_t *)addr, pagebits);
2702 #elif PAGE_SIZE == 16384
2703 atomic_clear_32((uint32_t *)addr, pagebits);
2704 #else /* PAGE_SIZE <= 8192 */
2706 * Use a trick to perform a 32-bit atomic on the
2707 * containing aligned word, to not depend on the existence
2708 * of atomic_clear_{8, 16}.
2710 shift = addr & (sizeof(uint32_t) - 1);
2711 #if BYTE_ORDER == BIG_ENDIAN
2712 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
2716 addr &= ~(sizeof(uint32_t) - 1);
2717 atomic_clear_32((uint32_t *)addr, pagebits << shift);
2718 #endif /* PAGE_SIZE */
2723 * vm_page_set_validclean:
2725 * Sets portions of a page valid and clean. The arguments are expected
2726 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2727 * of any partial chunks touched by the range. The invalid portion of
2728 * such chunks will be zero'd.
2730 * (base + size) must be less then or equal to PAGE_SIZE.
2733 vm_page_set_validclean(vm_page_t m, int base, int size)
2735 vm_page_bits_t oldvalid, pagebits;
2738 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2739 if (size == 0) /* handle degenerate case */
2743 * If the base is not DEV_BSIZE aligned and the valid
2744 * bit is clear, we have to zero out a portion of the
2747 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2748 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
2749 pmap_zero_page_area(m, frag, base - frag);
2752 * If the ending offset is not DEV_BSIZE aligned and the
2753 * valid bit is clear, we have to zero out a portion of
2756 endoff = base + size;
2757 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2758 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
2759 pmap_zero_page_area(m, endoff,
2760 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
2763 * Set valid, clear dirty bits. If validating the entire
2764 * page we can safely clear the pmap modify bit. We also
2765 * use this opportunity to clear the VPO_NOSYNC flag. If a process
2766 * takes a write fault on a MAP_NOSYNC memory area the flag will
2769 * We set valid bits inclusive of any overlap, but we can only
2770 * clear dirty bits for DEV_BSIZE chunks that are fully within
2773 oldvalid = m->valid;
2774 pagebits = vm_page_bits(base, size);
2775 m->valid |= pagebits;
2777 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
2778 frag = DEV_BSIZE - frag;
2784 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
2786 if (base == 0 && size == PAGE_SIZE) {
2788 * The page can only be modified within the pmap if it is
2789 * mapped, and it can only be mapped if it was previously
2792 if (oldvalid == VM_PAGE_BITS_ALL)
2794 * Perform the pmap_clear_modify() first. Otherwise,
2795 * a concurrent pmap operation, such as
2796 * pmap_protect(), could clear a modification in the
2797 * pmap and set the dirty field on the page before
2798 * pmap_clear_modify() had begun and after the dirty
2799 * field was cleared here.
2801 pmap_clear_modify(m);
2803 m->oflags &= ~VPO_NOSYNC;
2804 } else if (oldvalid != VM_PAGE_BITS_ALL)
2805 m->dirty &= ~pagebits;
2807 vm_page_clear_dirty_mask(m, pagebits);
2811 vm_page_clear_dirty(vm_page_t m, int base, int size)
2814 vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
2818 * vm_page_set_invalid:
2820 * Invalidates DEV_BSIZE'd chunks within a page. Both the
2821 * valid and dirty bits for the effected areas are cleared.
2824 vm_page_set_invalid(vm_page_t m, int base, int size)
2826 vm_page_bits_t bits;
2828 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2829 KASSERT((m->oflags & VPO_BUSY) == 0,
2830 ("vm_page_set_invalid: page %p is busy", m));
2831 bits = vm_page_bits(base, size);
2832 if (m->valid == VM_PAGE_BITS_ALL && bits != 0)
2834 KASSERT(!pmap_page_is_mapped(m),
2835 ("vm_page_set_invalid: page %p is mapped", m));
2841 * vm_page_zero_invalid()
2843 * The kernel assumes that the invalid portions of a page contain
2844 * garbage, but such pages can be mapped into memory by user code.
2845 * When this occurs, we must zero out the non-valid portions of the
2846 * page so user code sees what it expects.
2848 * Pages are most often semi-valid when the end of a file is mapped
2849 * into memory and the file's size is not page aligned.
2852 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
2857 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2859 * Scan the valid bits looking for invalid sections that
2860 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
2861 * valid bit may be set ) have already been zerod by
2862 * vm_page_set_validclean().
2864 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
2865 if (i == (PAGE_SIZE / DEV_BSIZE) ||
2866 (m->valid & ((vm_page_bits_t)1 << i))) {
2868 pmap_zero_page_area(m,
2869 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
2876 * setvalid is TRUE when we can safely set the zero'd areas
2877 * as being valid. We can do this if there are no cache consistancy
2878 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
2881 m->valid = VM_PAGE_BITS_ALL;
2887 * Is (partial) page valid? Note that the case where size == 0
2888 * will return FALSE in the degenerate case where the page is
2889 * entirely invalid, and TRUE otherwise.
2892 vm_page_is_valid(vm_page_t m, int base, int size)
2894 vm_page_bits_t bits;
2896 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2897 bits = vm_page_bits(base, size);
2898 if (m->valid && ((m->valid & bits) == bits))
2905 * Set the page's dirty bits if the page is modified.
2908 vm_page_test_dirty(vm_page_t m)
2911 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2912 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
2917 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
2920 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
2924 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
2927 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
2931 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
2934 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
2937 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
2939 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
2942 mtx_assert_(vm_page_lockptr(m), a, file, line);
2946 int so_zerocp_fullpage = 0;
2949 * Replace the given page with a copy. The copied page assumes
2950 * the portion of the given page's "wire_count" that is not the
2951 * responsibility of this copy-on-write mechanism.
2953 * The object containing the given page must have a non-zero
2954 * paging-in-progress count and be locked.
2957 vm_page_cowfault(vm_page_t m)
2963 vm_page_lock_assert(m, MA_OWNED);
2965 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2966 KASSERT(object->paging_in_progress != 0,
2967 ("vm_page_cowfault: object %p's paging-in-progress count is zero.",
2974 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY);
2976 vm_page_insert(m, object, pindex);
2978 VM_OBJECT_UNLOCK(object);
2980 VM_OBJECT_LOCK(object);
2981 if (m == vm_page_lookup(object, pindex)) {
2986 * Page disappeared during the wait.
2994 * check to see if we raced with an xmit complete when
2995 * waiting to allocate a page. If so, put things back
3001 vm_page_unlock(mnew);
3002 vm_page_insert(m, object, pindex);
3003 } else { /* clear COW & copy page */
3004 if (!so_zerocp_fullpage)
3005 pmap_copy_page(m, mnew);
3006 mnew->valid = VM_PAGE_BITS_ALL;
3007 vm_page_dirty(mnew);
3008 mnew->wire_count = m->wire_count - m->cow;
3009 m->wire_count = m->cow;
3015 vm_page_cowclear(vm_page_t m)
3018 vm_page_lock_assert(m, MA_OWNED);
3022 * let vm_fault add back write permission lazily
3026 * sf_buf_free() will free the page, so we needn't do it here
3031 vm_page_cowsetup(vm_page_t m)
3034 vm_page_lock_assert(m, MA_OWNED);
3035 if ((m->flags & PG_FICTITIOUS) != 0 ||
3036 (m->oflags & VPO_UNMANAGED) != 0 ||
3037 m->cow == USHRT_MAX - 1 || !VM_OBJECT_TRYLOCK(m->object))
3040 pmap_remove_write(m);
3041 VM_OBJECT_UNLOCK(m->object);
3047 vm_page_object_lock_assert(vm_page_t m)
3051 * Certain of the page's fields may only be modified by the
3052 * holder of the containing object's lock or the setter of the
3053 * page's VPO_BUSY flag. Unfortunately, the setter of the
3054 * VPO_BUSY flag is not recorded, and thus cannot be checked
3057 if (m->object != NULL && (m->oflags & VPO_BUSY) == 0)
3058 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
3062 #include "opt_ddb.h"
3064 #include <sys/kernel.h>
3066 #include <ddb/ddb.h>
3068 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3070 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count);
3071 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count);
3072 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count);
3073 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count);
3074 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count);
3075 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved);
3076 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min);
3077 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target);
3078 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min);
3079 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target);
3082 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3085 db_printf("PQ_FREE:");
3086 db_printf(" %d", cnt.v_free_count);
3089 db_printf("PQ_CACHE:");
3090 db_printf(" %d", cnt.v_cache_count);
3093 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
3094 *vm_pagequeues[PQ_ACTIVE].pq_cnt,
3095 *vm_pagequeues[PQ_INACTIVE].pq_cnt);