4 * Copyright (c) 1991 Regents of the University of California.
7 * This code is derived from software contributed to Berkeley by
8 * The Mach Operating System project at Carnegie-Mellon University.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91
35 * $FreeBSD: src/sys/vm/vm_page.c,v 1.147.2.18 2002/03/10 05:03:19 alc Exp $
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.
65 * Resident memory management module. The module manipulates 'VM pages'.
66 * A VM page is the core building block for memory management.
69 #include <sys/param.h>
70 #include <sys/systm.h>
71 #include <sys/malloc.h>
73 #include <sys/vmmeter.h>
74 #include <sys/vnode.h>
75 #include <sys/kernel.h>
76 #include <sys/alist.h>
77 #include <sys/sysctl.h>
80 #include <vm/vm_param.h>
82 #include <vm/vm_kern.h>
84 #include <vm/vm_map.h>
85 #include <vm/vm_object.h>
86 #include <vm/vm_page.h>
87 #include <vm/vm_pageout.h>
88 #include <vm/vm_pager.h>
89 #include <vm/vm_extern.h>
90 #include <vm/swap_pager.h>
92 #include <machine/inttypes.h>
93 #include <machine/md_var.h>
94 #include <machine/specialreg.h>
96 #include <vm/vm_page2.h>
97 #include <sys/spinlock2.h>
99 #define VMACTION_HSIZE 256
100 #define VMACTION_HMASK (VMACTION_HSIZE - 1)
102 static void vm_page_queue_init(void);
103 static void vm_page_free_wakeup(void);
104 static vm_page_t vm_page_select_cache(u_short pg_color);
105 static vm_page_t _vm_page_list_find2(int basequeue, int index);
106 static void _vm_page_deactivate_locked(vm_page_t m, int athead);
109 * Array of tailq lists
111 __cachealign struct vpgqueues vm_page_queues[PQ_COUNT];
113 LIST_HEAD(vm_page_action_list, vm_page_action);
114 struct vm_page_action_list action_list[VMACTION_HSIZE];
115 static volatile int vm_pages_waiting;
117 static struct alist vm_contig_alist;
118 static struct almeta vm_contig_ameta[ALIST_RECORDS_65536];
119 static struct spinlock vm_contig_spin = SPINLOCK_INITIALIZER(&vm_contig_spin);
121 static u_long vm_dma_reserved = 0;
122 TUNABLE_ULONG("vm.dma_reserved", &vm_dma_reserved);
123 SYSCTL_ULONG(_vm, OID_AUTO, dma_reserved, CTLFLAG_RD, &vm_dma_reserved, 0,
124 "Memory reserved for DMA");
125 SYSCTL_UINT(_vm, OID_AUTO, dma_free_pages, CTLFLAG_RD,
126 &vm_contig_alist.bl_free, 0, "Memory reserved for DMA");
128 static int vm_contig_verbose = 0;
129 TUNABLE_INT("vm.contig_verbose", &vm_contig_verbose);
131 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare,
132 vm_pindex_t, pindex);
135 vm_page_queue_init(void)
139 for (i = 0; i < PQ_L2_SIZE; i++)
140 vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count;
141 for (i = 0; i < PQ_L2_SIZE; i++)
142 vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count;
143 for (i = 0; i < PQ_L2_SIZE; i++)
144 vm_page_queues[PQ_INACTIVE+i].cnt = &vmstats.v_inactive_count;
145 for (i = 0; i < PQ_L2_SIZE; i++)
146 vm_page_queues[PQ_ACTIVE+i].cnt = &vmstats.v_active_count;
147 for (i = 0; i < PQ_L2_SIZE; i++)
148 vm_page_queues[PQ_HOLD+i].cnt = &vmstats.v_active_count;
149 /* PQ_NONE has no queue */
151 for (i = 0; i < PQ_COUNT; i++) {
152 TAILQ_INIT(&vm_page_queues[i].pl);
153 spin_init(&vm_page_queues[i].spin);
156 for (i = 0; i < VMACTION_HSIZE; i++)
157 LIST_INIT(&action_list[i]);
161 * note: place in initialized data section? Is this necessary?
164 int vm_page_array_size = 0;
165 int vm_page_zero_count = 0;
166 vm_page_t vm_page_array = NULL;
167 vm_paddr_t vm_low_phys_reserved;
172 * Sets the page size, perhaps based upon the memory size.
173 * Must be called before any use of page-size dependent functions.
176 vm_set_page_size(void)
178 if (vmstats.v_page_size == 0)
179 vmstats.v_page_size = PAGE_SIZE;
180 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0)
181 panic("vm_set_page_size: page size not a power of two");
187 * Add a new page to the freelist for use by the system. New pages
188 * are added to both the head and tail of the associated free page
189 * queue in a bottom-up fashion, so both zero'd and non-zero'd page
190 * requests pull 'recent' adds (higher physical addresses) first.
192 * Beware that the page zeroing daemon will also be running soon after
193 * boot, moving pages from the head to the tail of the PQ_FREE queues.
195 * Must be called in a critical section.
198 vm_add_new_page(vm_paddr_t pa)
200 struct vpgqueues *vpq;
203 m = PHYS_TO_VM_PAGE(pa);
206 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK;
207 m->pat_mode = PAT_WRITE_BACK;
209 * Twist for cpu localization in addition to page coloring, so
210 * different cpus selecting by m->queue get different page colors.
212 m->pc ^= ((pa >> PAGE_SHIFT) / PQ_L2_SIZE) & PQ_L2_MASK;
213 m->pc ^= ((pa >> PAGE_SHIFT) / (PQ_L2_SIZE * PQ_L2_SIZE)) & PQ_L2_MASK;
215 * Reserve a certain number of contiguous low memory pages for
216 * contigmalloc() to use.
218 if (pa < vm_low_phys_reserved) {
219 atomic_add_int(&vmstats.v_page_count, 1);
220 atomic_add_int(&vmstats.v_dma_pages, 1);
223 atomic_add_int(&vmstats.v_wire_count, 1);
224 alist_free(&vm_contig_alist, pa >> PAGE_SHIFT, 1);
231 m->queue = m->pc + PQ_FREE;
232 KKASSERT(m->dirty == 0);
234 atomic_add_int(&vmstats.v_page_count, 1);
235 atomic_add_int(&vmstats.v_free_count, 1);
236 vpq = &vm_page_queues[m->queue];
237 if ((vpq->flipflop & 15) == 0) {
238 pmap_zero_page(VM_PAGE_TO_PHYS(m));
240 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
241 atomic_add_int(&vm_page_zero_count, 1);
243 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
252 * Initializes the resident memory module.
254 * Preallocates memory for critical VM structures and arrays prior to
255 * kernel_map becoming available.
257 * Memory is allocated from (virtual2_start, virtual2_end) if available,
258 * otherwise memory is allocated from (virtual_start, virtual_end).
260 * On x86-64 (virtual_start, virtual_end) is only 2GB and may not be
261 * large enough to hold vm_page_array & other structures for machines with
262 * large amounts of ram, so we want to use virtual2* when available.
265 vm_page_startup(void)
267 vm_offset_t vaddr = virtual2_start ? virtual2_start : virtual_start;
270 vm_paddr_t page_range;
277 vm_paddr_t biggestone, biggestsize;
284 vaddr = round_page(vaddr);
286 for (i = 0; phys_avail[i + 1]; i += 2) {
287 phys_avail[i] = round_page64(phys_avail[i]);
288 phys_avail[i + 1] = trunc_page64(phys_avail[i + 1]);
291 for (i = 0; phys_avail[i + 1]; i += 2) {
292 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
294 if (size > biggestsize) {
302 end = phys_avail[biggestone+1];
303 end = trunc_page(end);
306 * Initialize the queue headers for the free queue, the active queue
307 * and the inactive queue.
309 vm_page_queue_init();
311 #if !defined(_KERNEL_VIRTUAL)
313 * VKERNELs don't support minidumps and as such don't need
316 * Allocate a bitmap to indicate that a random physical page
317 * needs to be included in a minidump.
319 * The amd64 port needs this to indicate which direct map pages
320 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
322 * However, i386 still needs this workspace internally within the
323 * minidump code. In theory, they are not needed on i386, but are
324 * included should the sf_buf code decide to use them.
326 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE;
327 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
328 end -= vm_page_dump_size;
329 vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size,
330 VM_PROT_READ | VM_PROT_WRITE);
331 bzero((void *)vm_page_dump, vm_page_dump_size);
334 * Compute the number of pages of memory that will be available for
335 * use (taking into account the overhead of a page structure per
338 first_page = phys_avail[0] / PAGE_SIZE;
339 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
340 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE;
342 #ifndef _KERNEL_VIRTUAL
344 * (only applies to real kernels)
346 * Initialize the contiguous reserve map. We initially reserve up
347 * to 1/4 available physical memory or 65536 pages (~256MB), whichever
350 * Once device initialization is complete we return most of the
351 * reserved memory back to the normal page queues but leave some
352 * in reserve for things like usb attachments.
354 vm_low_phys_reserved = (vm_paddr_t)65536 << PAGE_SHIFT;
355 if (vm_low_phys_reserved > total / 4)
356 vm_low_phys_reserved = total / 4;
357 if (vm_dma_reserved == 0) {
358 vm_dma_reserved = 16 * 1024 * 1024; /* 16MB */
359 if (vm_dma_reserved > total / 16)
360 vm_dma_reserved = total / 16;
363 alist_init(&vm_contig_alist, 65536, vm_contig_ameta,
364 ALIST_RECORDS_65536);
367 * Initialize the mem entry structures now, and put them in the free
370 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
371 mapped = pmap_map(&vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE);
372 vm_page_array = (vm_page_t)mapped;
374 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL)
376 * since pmap_map on amd64 returns stuff out of a direct-map region,
377 * we have to manually add these pages to the minidump tracking so
378 * that they can be dumped, including the vm_page_array.
380 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
385 * Clear all of the page structures
387 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
388 vm_page_array_size = page_range;
391 * Construct the free queue(s) in ascending order (by physical
392 * address) so that the first 16MB of physical memory is allocated
393 * last rather than first. On large-memory machines, this avoids
394 * the exhaustion of low physical memory before isa_dmainit has run.
396 vmstats.v_page_count = 0;
397 vmstats.v_free_count = 0;
398 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
403 last_pa = phys_avail[i + 1];
404 while (pa < last_pa && npages-- > 0) {
410 virtual2_start = vaddr;
412 virtual_start = vaddr;
416 * We tended to reserve a ton of memory for contigmalloc(). Now that most
417 * drivers have initialized we want to return most the remaining free
418 * reserve back to the VM page queues so they can be used for normal
421 * We leave vm_dma_reserved bytes worth of free pages in the reserve pool.
424 vm_page_startup_finish(void *dummy __unused)
433 spin_lock(&vm_contig_spin);
435 bfree = alist_free_info(&vm_contig_alist, &blk, &count);
436 if (bfree <= vm_dma_reserved / PAGE_SIZE)
442 * Figure out how much of the initial reserve we have to
443 * free in order to reach our target.
445 bfree -= vm_dma_reserved / PAGE_SIZE;
447 blk += count - bfree;
452 * Calculate the nearest power of 2 <= count.
454 for (xcount = 1; xcount <= count; xcount <<= 1)
457 blk += count - xcount;
461 * Allocate the pages from the alist, then free them to
462 * the normal VM page queues.
464 * Pages allocated from the alist are wired. We have to
465 * busy, unwire, and free them. We must also adjust
466 * vm_low_phys_reserved before freeing any pages to prevent
469 rblk = alist_alloc(&vm_contig_alist, blk, count);
471 kprintf("vm_page_startup_finish: Unable to return "
472 "dma space @0x%08x/%d -> 0x%08x\n",
476 atomic_add_int(&vmstats.v_dma_pages, -count);
477 spin_unlock(&vm_contig_spin);
479 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT);
480 vm_low_phys_reserved = VM_PAGE_TO_PHYS(m);
482 vm_page_busy_wait(m, FALSE, "cpgfr");
483 vm_page_unwire(m, 0);
488 spin_lock(&vm_contig_spin);
490 spin_unlock(&vm_contig_spin);
493 * Print out how much DMA space drivers have already allocated and
494 * how much is left over.
496 kprintf("DMA space used: %jdk, remaining available: %jdk\n",
497 (intmax_t)(vmstats.v_dma_pages - vm_contig_alist.bl_free) *
499 (intmax_t)vm_contig_alist.bl_free * (PAGE_SIZE / 1024));
501 SYSINIT(vm_pgend, SI_SUB_PROC0_POST, SI_ORDER_ANY,
502 vm_page_startup_finish, NULL)
506 * Scan comparison function for Red-Black tree scans. An inclusive
507 * (start,end) is expected. Other fields are not used.
510 rb_vm_page_scancmp(struct vm_page *p, void *data)
512 struct rb_vm_page_scan_info *info = data;
514 if (p->pindex < info->start_pindex)
516 if (p->pindex > info->end_pindex)
522 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2)
524 if (p1->pindex < p2->pindex)
526 if (p1->pindex > p2->pindex)
532 vm_page_init(vm_page_t m)
534 /* do nothing for now. Called from pmap_page_init() */
538 * Each page queue has its own spin lock, which is fairly optimal for
539 * allocating and freeing pages at least.
541 * The caller must hold the vm_page_spin_lock() before locking a vm_page's
542 * queue spinlock via this function. Also note that m->queue cannot change
543 * unless both the page and queue are locked.
547 _vm_page_queue_spin_lock(vm_page_t m)
552 if (queue != PQ_NONE) {
553 spin_lock(&vm_page_queues[queue].spin);
554 KKASSERT(queue == m->queue);
560 _vm_page_queue_spin_unlock(vm_page_t m)
566 if (queue != PQ_NONE)
567 spin_unlock(&vm_page_queues[queue].spin);
572 _vm_page_queues_spin_lock(u_short queue)
575 if (queue != PQ_NONE)
576 spin_lock(&vm_page_queues[queue].spin);
582 _vm_page_queues_spin_unlock(u_short queue)
585 if (queue != PQ_NONE)
586 spin_unlock(&vm_page_queues[queue].spin);
590 vm_page_queue_spin_lock(vm_page_t m)
592 _vm_page_queue_spin_lock(m);
596 vm_page_queues_spin_lock(u_short queue)
598 _vm_page_queues_spin_lock(queue);
602 vm_page_queue_spin_unlock(vm_page_t m)
604 _vm_page_queue_spin_unlock(m);
608 vm_page_queues_spin_unlock(u_short queue)
610 _vm_page_queues_spin_unlock(queue);
614 * This locks the specified vm_page and its queue in the proper order
615 * (page first, then queue). The queue may change so the caller must
620 _vm_page_and_queue_spin_lock(vm_page_t m)
622 vm_page_spin_lock(m);
623 _vm_page_queue_spin_lock(m);
628 _vm_page_and_queue_spin_unlock(vm_page_t m)
630 _vm_page_queues_spin_unlock(m->queue);
631 vm_page_spin_unlock(m);
635 vm_page_and_queue_spin_unlock(vm_page_t m)
637 _vm_page_and_queue_spin_unlock(m);
641 vm_page_and_queue_spin_lock(vm_page_t m)
643 _vm_page_and_queue_spin_lock(m);
647 * Helper function removes vm_page from its current queue.
648 * Returns the base queue the page used to be on.
650 * The vm_page and the queue must be spinlocked.
651 * This function will unlock the queue but leave the page spinlocked.
653 static __inline u_short
654 _vm_page_rem_queue_spinlocked(vm_page_t m)
656 struct vpgqueues *pq;
660 if (queue != PQ_NONE) {
661 pq = &vm_page_queues[queue];
662 TAILQ_REMOVE(&pq->pl, m, pageq);
663 atomic_add_int(pq->cnt, -1);
666 vm_page_queues_spin_unlock(queue);
667 if ((queue - m->pc) == PQ_FREE && (m->flags & PG_ZERO))
668 atomic_subtract_int(&vm_page_zero_count, 1);
669 if ((queue - m->pc) == PQ_CACHE || (queue - m->pc) == PQ_FREE)
670 return (queue - m->pc);
676 * Helper function places the vm_page on the specified queue.
678 * The vm_page must be spinlocked.
679 * This function will return with both the page and the queue locked.
682 _vm_page_add_queue_spinlocked(vm_page_t m, u_short queue, int athead)
684 struct vpgqueues *pq;
686 KKASSERT(m->queue == PQ_NONE);
688 if (queue != PQ_NONE) {
689 vm_page_queues_spin_lock(queue);
690 pq = &vm_page_queues[queue];
692 atomic_add_int(pq->cnt, 1);
696 * Put zero'd pages on the end ( where we look for zero'd pages
697 * first ) and non-zerod pages at the head.
699 if (queue - m->pc == PQ_FREE) {
700 if (m->flags & PG_ZERO) {
701 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
702 atomic_add_int(&vm_page_zero_count, 1);
704 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
707 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
709 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
711 /* leave the queue spinlocked */
716 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE)
717 * m->busy is zero. Returns TRUE if it had to sleep, FALSE if we
718 * did not. Only one sleep call will be made before returning.
720 * This function does NOT busy the page and on return the page is not
721 * guaranteed to be available.
724 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg)
732 if ((flags & PG_BUSY) == 0 &&
733 (also_m_busy == 0 || (flags & PG_SBUSY) == 0)) {
736 tsleep_interlock(m, 0);
737 if (atomic_cmpset_int(&m->flags, flags,
738 flags | PG_WANTED | PG_REFERENCED)) {
739 tsleep(m, PINTERLOCKED, msg, 0);
746 * Wait until PG_BUSY can be set, then set it. If also_m_busy is TRUE we
747 * also wait for m->busy to become 0 before setting PG_BUSY.
750 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m,
751 int also_m_busy, const char *msg
759 if (flags & PG_BUSY) {
760 tsleep_interlock(m, 0);
761 if (atomic_cmpset_int(&m->flags, flags,
762 flags | PG_WANTED | PG_REFERENCED)) {
763 tsleep(m, PINTERLOCKED, msg, 0);
765 } else if (also_m_busy && (flags & PG_SBUSY)) {
766 tsleep_interlock(m, 0);
767 if (atomic_cmpset_int(&m->flags, flags,
768 flags | PG_WANTED | PG_REFERENCED)) {
769 tsleep(m, PINTERLOCKED, msg, 0);
772 if (atomic_cmpset_int(&m->flags, flags,
776 m->busy_line = lineno;
785 * Attempt to set PG_BUSY. If also_m_busy is TRUE we only succeed if m->busy
788 * Returns non-zero on failure.
791 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy
801 if (also_m_busy && (flags & PG_SBUSY))
803 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
806 m->busy_line = lineno;
814 * Clear the PG_BUSY flag and return non-zero to indicate to the caller
815 * that a wakeup() should be performed.
817 * The vm_page must be spinlocked and will remain spinlocked on return.
818 * The related queue must NOT be spinlocked (which could deadlock us).
824 _vm_page_wakeup(vm_page_t m)
831 if (atomic_cmpset_int(&m->flags, flags,
832 flags & ~(PG_BUSY | PG_WANTED))) {
836 return(flags & PG_WANTED);
840 * Clear the PG_BUSY flag and wakeup anyone waiting for the page. This
841 * is typically the last call you make on a page before moving onto
845 vm_page_wakeup(vm_page_t m)
847 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!"));
848 vm_page_spin_lock(m);
849 if (_vm_page_wakeup(m)) {
850 vm_page_spin_unlock(m);
853 vm_page_spin_unlock(m);
858 * Holding a page keeps it from being reused. Other parts of the system
859 * can still disassociate the page from its current object and free it, or
860 * perform read or write I/O on it and/or otherwise manipulate the page,
861 * but if the page is held the VM system will leave the page and its data
862 * intact and not reuse the page for other purposes until the last hold
863 * reference is released. (see vm_page_wire() if you want to prevent the
864 * page from being disassociated from its object too).
866 * The caller must still validate the contents of the page and, if necessary,
867 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete
868 * before manipulating the page.
870 * XXX get vm_page_spin_lock() here and move FREE->HOLD if necessary
873 vm_page_hold(vm_page_t m)
875 vm_page_spin_lock(m);
876 atomic_add_int(&m->hold_count, 1);
877 if (m->queue - m->pc == PQ_FREE) {
878 _vm_page_queue_spin_lock(m);
879 _vm_page_rem_queue_spinlocked(m);
880 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
881 _vm_page_queue_spin_unlock(m);
883 vm_page_spin_unlock(m);
887 * The opposite of vm_page_hold(). A page can be freed while being held,
888 * which places it on the PQ_HOLD queue. If we are able to busy the page
889 * after the hold count drops to zero we will move the page to the
890 * appropriate PQ_FREE queue by calling vm_page_free_toq().
893 vm_page_unhold(vm_page_t m)
895 vm_page_spin_lock(m);
896 atomic_add_int(&m->hold_count, -1);
897 if (m->hold_count == 0 && m->queue - m->pc == PQ_HOLD) {
898 _vm_page_queue_spin_lock(m);
899 _vm_page_rem_queue_spinlocked(m);
900 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0);
901 _vm_page_queue_spin_unlock(m);
903 vm_page_spin_unlock(m);
907 * Inserts the given vm_page into the object and object list.
909 * The pagetables are not updated but will presumably fault the page
910 * in if necessary, or if a kernel page the caller will at some point
911 * enter the page into the kernel's pmap. We are not allowed to block
912 * here so we *can't* do this anyway.
914 * This routine may not block.
915 * This routine must be called with the vm_object held.
916 * This routine must be called with a critical section held.
918 * This routine returns TRUE if the page was inserted into the object
919 * successfully, and FALSE if the page already exists in the object.
922 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
924 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(object));
925 if (m->object != NULL)
926 panic("vm_page_insert: already inserted");
928 object->generation++;
931 * Record the object/offset pair in this page and add the
932 * pv_list_count of the page to the object.
934 * The vm_page spin lock is required for interactions with the pmap.
936 vm_page_spin_lock(m);
939 if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) {
942 vm_page_spin_unlock(m);
945 object->resident_page_count++;
946 /* atomic_add_int(&object->agg_pv_list_count, m->md.pv_list_count); */
947 vm_page_spin_unlock(m);
950 * Since we are inserting a new and possibly dirty page,
951 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
953 if ((m->valid & m->dirty) ||
954 (m->flags & (PG_WRITEABLE | PG_NEED_COMMIT)))
955 vm_object_set_writeable_dirty(object);
958 * Checks for a swap assignment and sets PG_SWAPPED if appropriate.
960 swap_pager_page_inserted(m);
965 * Removes the given vm_page_t from the (object,index) table
967 * The underlying pmap entry (if any) is NOT removed here.
968 * This routine may not block.
970 * The page must be BUSY and will remain BUSY on return.
971 * No other requirements.
973 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave
977 vm_page_remove(vm_page_t m)
981 if (m->object == NULL) {
985 if ((m->flags & PG_BUSY) == 0)
986 panic("vm_page_remove: page not busy");
990 vm_object_hold(object);
993 * Remove the page from the object and update the object.
995 * The vm_page spin lock is required for interactions with the pmap.
997 vm_page_spin_lock(m);
998 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m);
999 object->resident_page_count--;
1000 /* atomic_add_int(&object->agg_pv_list_count, -m->md.pv_list_count); */
1002 vm_page_spin_unlock(m);
1004 object->generation++;
1006 vm_object_drop(object);
1010 * Locate and return the page at (object, pindex), or NULL if the
1011 * page could not be found.
1013 * The caller must hold the vm_object token.
1016 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1021 * Search the hash table for this object/offset pair
1023 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1024 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1025 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex));
1030 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object,
1032 int also_m_busy, const char *msg
1038 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1039 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1041 KKASSERT(m->object == object && m->pindex == pindex);
1044 if (flags & PG_BUSY) {
1045 tsleep_interlock(m, 0);
1046 if (atomic_cmpset_int(&m->flags, flags,
1047 flags | PG_WANTED | PG_REFERENCED)) {
1048 tsleep(m, PINTERLOCKED, msg, 0);
1049 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1052 } else if (also_m_busy && (flags & PG_SBUSY)) {
1053 tsleep_interlock(m, 0);
1054 if (atomic_cmpset_int(&m->flags, flags,
1055 flags | PG_WANTED | PG_REFERENCED)) {
1056 tsleep(m, PINTERLOCKED, msg, 0);
1057 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1060 } else if (atomic_cmpset_int(&m->flags, flags,
1062 #ifdef VM_PAGE_DEBUG
1063 m->busy_func = func;
1064 m->busy_line = lineno;
1073 * Attempt to lookup and busy a page.
1075 * Returns NULL if the page could not be found
1077 * Returns a vm_page and error == TRUE if the page exists but could not
1080 * Returns a vm_page and error == FALSE on success.
1083 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object,
1085 int also_m_busy, int *errorp
1091 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1092 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1095 KKASSERT(m->object == object && m->pindex == pindex);
1098 if (flags & PG_BUSY) {
1102 if (also_m_busy && (flags & PG_SBUSY)) {
1106 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
1107 #ifdef VM_PAGE_DEBUG
1108 m->busy_func = func;
1109 m->busy_line = lineno;
1118 * Caller must hold the related vm_object
1121 vm_page_next(vm_page_t m)
1125 next = vm_page_rb_tree_RB_NEXT(m);
1126 if (next && next->pindex != m->pindex + 1)
1134 * Move the given vm_page from its current object to the specified
1135 * target object/offset. The page must be busy and will remain so
1138 * new_object must be held.
1139 * This routine might block. XXX ?
1141 * NOTE: Swap associated with the page must be invalidated by the move. We
1142 * have to do this for several reasons: (1) we aren't freeing the
1143 * page, (2) we are dirtying the page, (3) the VM system is probably
1144 * moving the page from object A to B, and will then later move
1145 * the backing store from A to B and we can't have a conflict.
1147 * NOTE: We *always* dirty the page. It is necessary both for the
1148 * fact that we moved it, and because we may be invalidating
1149 * swap. If the page is on the cache, we have to deactivate it
1150 * or vm_page_dirty() will panic. Dirty pages are not allowed
1154 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1156 KKASSERT(m->flags & PG_BUSY);
1157 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(new_object));
1159 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(m->object));
1162 if (vm_page_insert(m, new_object, new_pindex) == FALSE) {
1163 panic("vm_page_rename: target exists (%p,%"PRIu64")",
1164 new_object, new_pindex);
1166 if (m->queue - m->pc == PQ_CACHE)
1167 vm_page_deactivate(m);
1172 * vm_page_unqueue() without any wakeup. This routine is used when a page
1173 * is being moved between queues or otherwise is to remain BUSYied by the
1176 * This routine may not block.
1179 vm_page_unqueue_nowakeup(vm_page_t m)
1181 vm_page_and_queue_spin_lock(m);
1182 (void)_vm_page_rem_queue_spinlocked(m);
1183 vm_page_spin_unlock(m);
1187 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
1190 * This routine may not block.
1193 vm_page_unqueue(vm_page_t m)
1197 vm_page_and_queue_spin_lock(m);
1198 queue = _vm_page_rem_queue_spinlocked(m);
1199 if (queue == PQ_FREE || queue == PQ_CACHE) {
1200 vm_page_spin_unlock(m);
1201 pagedaemon_wakeup();
1203 vm_page_spin_unlock(m);
1208 * vm_page_list_find()
1210 * Find a page on the specified queue with color optimization.
1212 * The page coloring optimization attempts to locate a page that does
1213 * not overload other nearby pages in the object in the cpu's L1 or L2
1214 * caches. We need this optimization because cpu caches tend to be
1215 * physical caches, while object spaces tend to be virtual.
1217 * On MP systems each PQ_FREE and PQ_CACHE color queue has its own spinlock
1218 * and the algorithm is adjusted to localize allocations on a per-core basis.
1219 * This is done by 'twisting' the colors.
1221 * The page is returned spinlocked and removed from its queue (it will
1222 * be on PQ_NONE), or NULL. The page is not PG_BUSY'd. The caller
1223 * is responsible for dealing with the busy-page case (usually by
1224 * deactivating the page and looping).
1226 * NOTE: This routine is carefully inlined. A non-inlined version
1227 * is available for outside callers but the only critical path is
1228 * from within this source file.
1230 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE
1231 * represent stable storage, allowing us to order our locks vm_page
1232 * first, then queue.
1236 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
1242 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist);
1244 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl);
1246 m = _vm_page_list_find2(basequeue, index);
1249 vm_page_and_queue_spin_lock(m);
1250 if (m->queue == basequeue + index) {
1251 _vm_page_rem_queue_spinlocked(m);
1252 /* vm_page_t spin held, no queue spin */
1255 vm_page_and_queue_spin_unlock(m);
1261 _vm_page_list_find2(int basequeue, int index)
1265 struct vpgqueues *pq;
1267 pq = &vm_page_queues[basequeue];
1270 * Note that for the first loop, index+i and index-i wind up at the
1271 * same place. Even though this is not totally optimal, we've already
1272 * blown it by missing the cache case so we do not care.
1274 for (i = PQ_L2_SIZE / 2; i > 0; --i) {
1276 m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl);
1278 _vm_page_and_queue_spin_lock(m);
1280 basequeue + ((index + i) & PQ_L2_MASK)) {
1281 _vm_page_rem_queue_spinlocked(m);
1284 _vm_page_and_queue_spin_unlock(m);
1287 m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl);
1289 _vm_page_and_queue_spin_lock(m);
1291 basequeue + ((index - i) & PQ_L2_MASK)) {
1292 _vm_page_rem_queue_spinlocked(m);
1295 _vm_page_and_queue_spin_unlock(m);
1305 * Returns a vm_page candidate for allocation. The page is not busied so
1306 * it can move around. The caller must busy the page (and typically
1307 * deactivate it if it cannot be busied!)
1309 * Returns a spinlocked vm_page that has been removed from its queue.
1312 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
1314 return(_vm_page_list_find(basequeue, index, prefer_zero));
1318 * Find a page on the cache queue with color optimization, remove it
1319 * from the queue, and busy it. The returned page will not be spinlocked.
1321 * A candidate failure will be deactivated. Candidates can fail due to
1322 * being busied by someone else, in which case they will be deactivated.
1324 * This routine may not block.
1328 vm_page_select_cache(u_short pg_color)
1333 m = _vm_page_list_find(PQ_CACHE, pg_color & PQ_L2_MASK, FALSE);
1337 * (m) has been removed from its queue and spinlocked
1339 if (vm_page_busy_try(m, TRUE)) {
1340 _vm_page_deactivate_locked(m, 0);
1341 vm_page_spin_unlock(m);
1344 * We successfully busied the page
1346 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) == 0 &&
1347 m->hold_count == 0 &&
1348 m->wire_count == 0 &&
1349 (m->dirty & m->valid) == 0) {
1350 vm_page_spin_unlock(m);
1351 pagedaemon_wakeup();
1356 * The page cannot be recycled, deactivate it.
1358 _vm_page_deactivate_locked(m, 0);
1359 if (_vm_page_wakeup(m)) {
1360 vm_page_spin_unlock(m);
1363 vm_page_spin_unlock(m);
1371 * Find a free or zero page, with specified preference. We attempt to
1372 * inline the nominal case and fall back to _vm_page_select_free()
1373 * otherwise. A busied page is removed from the queue and returned.
1375 * This routine may not block.
1377 static __inline vm_page_t
1378 vm_page_select_free(u_short pg_color, boolean_t prefer_zero)
1383 m = _vm_page_list_find(PQ_FREE, pg_color & PQ_L2_MASK,
1387 if (vm_page_busy_try(m, TRUE)) {
1389 * Various mechanisms such as a pmap_collect can
1390 * result in a busy page on the free queue. We
1391 * have to move the page out of the way so we can
1392 * retry the allocation. If the other thread is not
1393 * allocating the page then m->valid will remain 0 and
1394 * the pageout daemon will free the page later on.
1396 * Since we could not busy the page, however, we
1397 * cannot make assumptions as to whether the page
1398 * will be allocated by the other thread or not,
1399 * so all we can do is deactivate it to move it out
1400 * of the way. In particular, if the other thread
1401 * wires the page it may wind up on the inactive
1402 * queue and the pageout daemon will have to deal
1403 * with that case too.
1405 _vm_page_deactivate_locked(m, 0);
1406 vm_page_spin_unlock(m);
1409 * Theoretically if we are able to busy the page
1410 * atomic with the queue removal (using the vm_page
1411 * lock) nobody else should be able to mess with the
1414 KKASSERT((m->flags & (PG_UNMANAGED |
1415 PG_NEED_COMMIT)) == 0);
1416 KKASSERT(m->hold_count == 0);
1417 KKASSERT(m->wire_count == 0);
1418 vm_page_spin_unlock(m);
1419 pagedaemon_wakeup();
1421 /* return busied and removed page */
1429 * This implements a per-cpu cache of free, zero'd, ready-to-go pages.
1430 * The idea is to populate this cache prior to acquiring any locks so
1431 * we don't wind up potentially zeroing VM pages (under heavy loads) while
1432 * holding potentialy contending locks.
1434 * Note that we allocate the page uninserted into anything and use a pindex
1435 * of 0, the vm_page_alloc() will effectively add gd_cpuid so these
1436 * allocations should wind up being uncontended. However, we still want
1437 * to rove across PQ_L2_SIZE.
1440 vm_page_pcpu_cache(void)
1443 globaldata_t gd = mycpu;
1446 if (gd->gd_vmpg_count < GD_MINVMPG) {
1448 while (gd->gd_vmpg_count < GD_MAXVMPG) {
1449 m = vm_page_alloc(NULL, ticks & ~ncpus2_mask,
1450 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL |
1451 VM_ALLOC_NULL_OK | VM_ALLOC_ZERO);
1452 if (gd->gd_vmpg_count < GD_MAXVMPG) {
1453 if ((m->flags & PG_ZERO) == 0) {
1454 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1455 vm_page_flag_set(m, PG_ZERO);
1457 gd->gd_vmpg_array[gd->gd_vmpg_count++] = m;
1470 * Allocate and return a memory cell associated with this VM object/offset
1471 * pair. If object is NULL an unassociated page will be allocated.
1473 * The returned page will be busied and removed from its queues. This
1474 * routine can block and may return NULL if a race occurs and the page
1475 * is found to already exist at the specified (object, pindex).
1477 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain
1478 * VM_ALLOC_QUICK like normal but cannot use cache
1479 * VM_ALLOC_SYSTEM greater free drain
1480 * VM_ALLOC_INTERRUPT allow free list to be completely drained
1481 * VM_ALLOC_ZERO advisory request for pre-zero'd page only
1482 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only
1483 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision
1484 * (see vm_page_grab())
1485 * VM_ALLOC_USE_GD ok to use per-gd cache
1487 * The object must be held if not NULL
1488 * This routine may not block
1490 * Additional special handling is required when called from an interrupt
1491 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache
1495 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
1497 globaldata_t gd = mycpu;
1504 * Special per-cpu free VM page cache. The pages are pre-busied
1505 * and pre-zerod for us.
1507 if (gd->gd_vmpg_count && (page_req & VM_ALLOC_USE_GD)) {
1509 if (gd->gd_vmpg_count) {
1510 m = gd->gd_vmpg_array[--gd->gd_vmpg_count];
1520 * Cpu twist - cpu localization algorithm
1523 pg_color = gd->gd_cpuid + (pindex & ~ncpus_fit_mask) +
1524 (object->pg_color & ~ncpus_fit_mask);
1526 pg_color = gd->gd_cpuid + (pindex & ~ncpus_fit_mask);
1529 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK|
1530 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
1533 * Certain system threads (pageout daemon, buf_daemon's) are
1534 * allowed to eat deeper into the free page list.
1536 if (curthread->td_flags & TDF_SYSTHREAD)
1537 page_req |= VM_ALLOC_SYSTEM;
1540 if (vmstats.v_free_count > vmstats.v_free_reserved ||
1541 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) ||
1542 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 &&
1543 vmstats.v_free_count > vmstats.v_interrupt_free_min)
1546 * The free queue has sufficient free pages to take one out.
1548 if (page_req & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO))
1549 m = vm_page_select_free(pg_color, TRUE);
1551 m = vm_page_select_free(pg_color, FALSE);
1552 } else if (page_req & VM_ALLOC_NORMAL) {
1554 * Allocatable from the cache (non-interrupt only). On
1555 * success, we must free the page and try again, thus
1556 * ensuring that vmstats.v_*_free_min counters are replenished.
1559 if (curthread->td_preempted) {
1560 kprintf("vm_page_alloc(): warning, attempt to allocate"
1561 " cache page from preempting interrupt\n");
1564 m = vm_page_select_cache(pg_color);
1567 m = vm_page_select_cache(pg_color);
1570 * On success move the page into the free queue and loop.
1572 * Only do this if we can safely acquire the vm_object lock,
1573 * because this is effectively a random page and the caller
1574 * might be holding the lock shared, we don't want to
1578 KASSERT(m->dirty == 0,
1579 ("Found dirty cache page %p", m));
1580 if ((obj = m->object) != NULL) {
1581 if (vm_object_hold_try(obj)) {
1582 vm_page_protect(m, VM_PROT_NONE);
1584 /* m->object NULL here */
1585 vm_object_drop(obj);
1587 vm_page_deactivate(m);
1591 vm_page_protect(m, VM_PROT_NONE);
1598 * On failure return NULL
1600 #if defined(DIAGNOSTIC)
1601 if (vmstats.v_cache_count > 0)
1602 kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count);
1604 vm_pageout_deficit++;
1605 pagedaemon_wakeup();
1609 * No pages available, wakeup the pageout daemon and give up.
1611 vm_pageout_deficit++;
1612 pagedaemon_wakeup();
1617 * v_free_count can race so loop if we don't find the expected
1624 * Good page found. The page has already been busied for us and
1625 * removed from its queues.
1627 KASSERT(m->dirty == 0,
1628 ("vm_page_alloc: free/cache page %p was dirty", m));
1629 KKASSERT(m->queue == PQ_NONE);
1635 * Initialize the structure, inheriting some flags but clearing
1636 * all the rest. The page has already been busied for us.
1638 vm_page_flag_clear(m, ~(PG_ZERO | PG_BUSY | PG_SBUSY));
1639 KKASSERT(m->wire_count == 0);
1640 KKASSERT(m->busy == 0);
1645 * Caller must be holding the object lock (asserted by
1646 * vm_page_insert()).
1648 * NOTE: Inserting a page here does not insert it into any pmaps
1649 * (which could cause us to block allocating memory).
1651 * NOTE: If no object an unassociated page is allocated, m->pindex
1652 * can be used by the caller for any purpose.
1655 if (vm_page_insert(m, object, pindex) == FALSE) {
1657 if ((page_req & VM_ALLOC_NULL_OK) == 0)
1658 panic("PAGE RACE %p[%ld]/%p",
1659 object, (long)pindex, m);
1667 * Don't wakeup too often - wakeup the pageout daemon when
1668 * we would be nearly out of memory.
1670 pagedaemon_wakeup();
1673 * A PG_BUSY page is returned.
1679 * Attempt to allocate contiguous physical memory with the specified
1683 vm_page_alloc_contig(vm_paddr_t low, vm_paddr_t high,
1684 unsigned long alignment, unsigned long boundary,
1689 alignment >>= PAGE_SHIFT;
1692 boundary >>= PAGE_SHIFT;
1695 size = (size + PAGE_MASK) >> PAGE_SHIFT;
1697 spin_lock(&vm_contig_spin);
1698 blk = alist_alloc(&vm_contig_alist, 0, size);
1699 if (blk == ALIST_BLOCK_NONE) {
1700 spin_unlock(&vm_contig_spin);
1702 kprintf("vm_page_alloc_contig: %ldk nospace\n",
1703 (size + PAGE_MASK) * (PAGE_SIZE / 1024));
1707 if (high && ((vm_paddr_t)(blk + size) << PAGE_SHIFT) > high) {
1708 alist_free(&vm_contig_alist, blk, size);
1709 spin_unlock(&vm_contig_spin);
1711 kprintf("vm_page_alloc_contig: %ldk high "
1713 (size + PAGE_MASK) * (PAGE_SIZE / 1024),
1718 spin_unlock(&vm_contig_spin);
1719 if (vm_contig_verbose) {
1720 kprintf("vm_page_alloc_contig: %016jx/%ldk\n",
1721 (intmax_t)(vm_paddr_t)blk << PAGE_SHIFT,
1722 (size + PAGE_MASK) * (PAGE_SIZE / 1024));
1724 return (PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT));
1728 * Free contiguously allocated pages. The pages will be wired but not busy.
1729 * When freeing to the alist we leave them wired and not busy.
1732 vm_page_free_contig(vm_page_t m, unsigned long size)
1734 vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1735 vm_pindex_t start = pa >> PAGE_SHIFT;
1736 vm_pindex_t pages = (size + PAGE_MASK) >> PAGE_SHIFT;
1738 if (vm_contig_verbose) {
1739 kprintf("vm_page_free_contig: %016jx/%ldk\n",
1740 (intmax_t)pa, size / 1024);
1742 if (pa < vm_low_phys_reserved) {
1743 KKASSERT(pa + size <= vm_low_phys_reserved);
1744 spin_lock(&vm_contig_spin);
1745 alist_free(&vm_contig_alist, start, pages);
1746 spin_unlock(&vm_contig_spin);
1749 vm_page_busy_wait(m, FALSE, "cpgfr");
1750 vm_page_unwire(m, 0);
1761 * Wait for sufficient free memory for nominal heavy memory use kernel
1764 * WARNING! Be sure never to call this in any vm_pageout code path, which
1765 * will trivially deadlock the system.
1768 vm_wait_nominal(void)
1770 while (vm_page_count_min(0))
1775 * Test if vm_wait_nominal() would block.
1778 vm_test_nominal(void)
1780 if (vm_page_count_min(0))
1786 * Block until free pages are available for allocation, called in various
1787 * places before memory allocations.
1789 * The caller may loop if vm_page_count_min() == FALSE so we cannot be
1790 * more generous then that.
1796 * never wait forever
1800 lwkt_gettoken(&vm_token);
1802 if (curthread == pagethread) {
1804 * The pageout daemon itself needs pages, this is bad.
1806 if (vm_page_count_min(0)) {
1807 vm_pageout_pages_needed = 1;
1808 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo);
1812 * Wakeup the pageout daemon if necessary and wait.
1814 * Do not wait indefinitely for the target to be reached,
1815 * as load might prevent it from being reached any time soon.
1816 * But wait a little to try to slow down page allocations
1817 * and to give more important threads (the pagedaemon)
1818 * allocation priority.
1820 if (vm_page_count_target()) {
1821 if (vm_pages_needed == 0) {
1822 vm_pages_needed = 1;
1823 wakeup(&vm_pages_needed);
1825 ++vm_pages_waiting; /* SMP race ok */
1826 tsleep(&vmstats.v_free_count, 0, "vmwait", timo);
1829 lwkt_reltoken(&vm_token);
1833 * Block until free pages are available for allocation
1835 * Called only from vm_fault so that processes page faulting can be
1839 vm_wait_pfault(void)
1842 * Wakeup the pageout daemon if necessary and wait.
1844 * Do not wait indefinitely for the target to be reached,
1845 * as load might prevent it from being reached any time soon.
1846 * But wait a little to try to slow down page allocations
1847 * and to give more important threads (the pagedaemon)
1848 * allocation priority.
1850 if (vm_page_count_min(0)) {
1851 lwkt_gettoken(&vm_token);
1852 while (vm_page_count_severe()) {
1853 if (vm_page_count_target()) {
1854 if (vm_pages_needed == 0) {
1855 vm_pages_needed = 1;
1856 wakeup(&vm_pages_needed);
1858 ++vm_pages_waiting; /* SMP race ok */
1859 tsleep(&vmstats.v_free_count, 0, "pfault", hz);
1862 lwkt_reltoken(&vm_token);
1867 * Put the specified page on the active list (if appropriate). Ensure
1868 * that act_count is at least ACT_INIT but do not otherwise mess with it.
1870 * The caller should be holding the page busied ? XXX
1871 * This routine may not block.
1874 vm_page_activate(vm_page_t m)
1878 vm_page_spin_lock(m);
1879 if (m->queue - m->pc != PQ_ACTIVE) {
1880 _vm_page_queue_spin_lock(m);
1881 oqueue = _vm_page_rem_queue_spinlocked(m);
1882 /* page is left spinlocked, queue is unlocked */
1884 if (oqueue == PQ_CACHE)
1885 mycpu->gd_cnt.v_reactivated++;
1886 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1887 if (m->act_count < ACT_INIT)
1888 m->act_count = ACT_INIT;
1889 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE + m->pc, 0);
1891 _vm_page_and_queue_spin_unlock(m);
1892 if (oqueue == PQ_CACHE || oqueue == PQ_FREE)
1893 pagedaemon_wakeup();
1895 if (m->act_count < ACT_INIT)
1896 m->act_count = ACT_INIT;
1897 vm_page_spin_unlock(m);
1902 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
1903 * routine is called when a page has been added to the cache or free
1906 * This routine may not block.
1908 static __inline void
1909 vm_page_free_wakeup(void)
1912 * If the pageout daemon itself needs pages, then tell it that
1913 * there are some free.
1915 if (vm_pageout_pages_needed &&
1916 vmstats.v_cache_count + vmstats.v_free_count >=
1917 vmstats.v_pageout_free_min
1919 vm_pageout_pages_needed = 0;
1920 wakeup(&vm_pageout_pages_needed);
1924 * Wakeup processes that are waiting on memory.
1926 * Generally speaking we want to wakeup stuck processes as soon as
1927 * possible. !vm_page_count_min(0) is the absolute minimum point
1928 * where we can do this. Wait a bit longer to reduce degenerate
1929 * re-blocking (vm_page_free_hysteresis). The target check is just
1930 * to make sure the min-check w/hysteresis does not exceed the
1933 if (vm_pages_waiting) {
1934 if (!vm_page_count_min(vm_page_free_hysteresis) ||
1935 !vm_page_count_target()) {
1936 vm_pages_waiting = 0;
1937 wakeup(&vmstats.v_free_count);
1938 ++mycpu->gd_cnt.v_ppwakeups;
1941 if (!vm_page_count_target()) {
1943 * Plenty of pages are free, wakeup everyone.
1945 vm_pages_waiting = 0;
1946 wakeup(&vmstats.v_free_count);
1947 ++mycpu->gd_cnt.v_ppwakeups;
1948 } else if (!vm_page_count_min(0)) {
1950 * Some pages are free, wakeup someone.
1952 int wcount = vm_pages_waiting;
1955 vm_pages_waiting = wcount;
1956 wakeup_one(&vmstats.v_free_count);
1957 ++mycpu->gd_cnt.v_ppwakeups;
1964 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates
1965 * it from its VM object.
1967 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on
1968 * return (the page will have been freed).
1971 vm_page_free_toq(vm_page_t m)
1973 mycpu->gd_cnt.v_tfree++;
1974 KKASSERT((m->flags & PG_MAPPED) == 0);
1975 KKASSERT(m->flags & PG_BUSY);
1977 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
1978 kprintf("vm_page_free: pindex(%lu), busy(%d), "
1979 "PG_BUSY(%d), hold(%d)\n",
1980 (u_long)m->pindex, m->busy,
1981 ((m->flags & PG_BUSY) ? 1 : 0), m->hold_count);
1982 if ((m->queue - m->pc) == PQ_FREE)
1983 panic("vm_page_free: freeing free page");
1985 panic("vm_page_free: freeing busy page");
1989 * Remove from object, spinlock the page and its queues and
1990 * remove from any queue. No queue spinlock will be held
1991 * after this section (because the page was removed from any
1995 vm_page_and_queue_spin_lock(m);
1996 _vm_page_rem_queue_spinlocked(m);
1999 * No further management of fictitious pages occurs beyond object
2000 * and queue removal.
2002 if ((m->flags & PG_FICTITIOUS) != 0) {
2003 vm_page_spin_unlock(m);
2011 if (m->wire_count != 0) {
2012 if (m->wire_count > 1) {
2014 "vm_page_free: invalid wire count (%d), pindex: 0x%lx",
2015 m->wire_count, (long)m->pindex);
2017 panic("vm_page_free: freeing wired page");
2021 * Clear the UNMANAGED flag when freeing an unmanaged page.
2022 * Clear the NEED_COMMIT flag
2024 if (m->flags & PG_UNMANAGED)
2025 vm_page_flag_clear(m, PG_UNMANAGED);
2026 if (m->flags & PG_NEED_COMMIT)
2027 vm_page_flag_clear(m, PG_NEED_COMMIT);
2029 if (m->hold_count != 0) {
2030 vm_page_flag_clear(m, PG_ZERO);
2031 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
2033 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0);
2037 * This sequence allows us to clear PG_BUSY while still holding
2038 * its spin lock, which reduces contention vs allocators. We
2039 * must not leave the queue locked or _vm_page_wakeup() may
2042 _vm_page_queue_spin_unlock(m);
2043 if (_vm_page_wakeup(m)) {
2044 vm_page_spin_unlock(m);
2047 vm_page_spin_unlock(m);
2049 vm_page_free_wakeup();
2053 * vm_page_free_fromq_fast()
2055 * Remove a non-zero page from one of the free queues; the page is removed for
2056 * zeroing, so do not issue a wakeup.
2059 vm_page_free_fromq_fast(void)
2065 for (i = 0; i < PQ_L2_SIZE; ++i) {
2066 m = vm_page_list_find(PQ_FREE, qi, FALSE);
2067 /* page is returned spinlocked and removed from its queue */
2069 if (vm_page_busy_try(m, TRUE)) {
2071 * We were unable to busy the page, deactivate
2074 _vm_page_deactivate_locked(m, 0);
2075 vm_page_spin_unlock(m);
2076 } else if (m->flags & PG_ZERO) {
2078 * The page is PG_ZERO, requeue it and loop
2080 _vm_page_add_queue_spinlocked(m,
2083 vm_page_queue_spin_unlock(m);
2084 if (_vm_page_wakeup(m)) {
2085 vm_page_spin_unlock(m);
2088 vm_page_spin_unlock(m);
2092 * The page is not PG_ZERO'd so return it.
2094 vm_page_spin_unlock(m);
2095 KKASSERT((m->flags & (PG_UNMANAGED |
2096 PG_NEED_COMMIT)) == 0);
2097 KKASSERT(m->hold_count == 0);
2098 KKASSERT(m->wire_count == 0);
2103 qi = (qi + PQ_PRIME2) & PQ_L2_MASK;
2109 * vm_page_unmanage()
2111 * Prevent PV management from being done on the page. The page is
2112 * removed from the paging queues as if it were wired, and as a
2113 * consequence of no longer being managed the pageout daemon will not
2114 * touch it (since there is no way to locate the pte mappings for the
2115 * page). madvise() calls that mess with the pmap will also no longer
2116 * operate on the page.
2118 * Beyond that the page is still reasonably 'normal'. Freeing the page
2119 * will clear the flag.
2121 * This routine is used by OBJT_PHYS objects - objects using unswappable
2122 * physical memory as backing store rather then swap-backed memory and
2123 * will eventually be extended to support 4MB unmanaged physical
2126 * Caller must be holding the page busy.
2129 vm_page_unmanage(vm_page_t m)
2131 KKASSERT(m->flags & PG_BUSY);
2132 if ((m->flags & PG_UNMANAGED) == 0) {
2133 if (m->wire_count == 0)
2136 vm_page_flag_set(m, PG_UNMANAGED);
2140 * Mark this page as wired down by yet another map, removing it from
2141 * paging queues as necessary.
2143 * Caller must be holding the page busy.
2146 vm_page_wire(vm_page_t m)
2149 * Only bump the wire statistics if the page is not already wired,
2150 * and only unqueue the page if it is on some queue (if it is unmanaged
2151 * it is already off the queues). Don't do anything with fictitious
2152 * pages because they are always wired.
2154 KKASSERT(m->flags & PG_BUSY);
2155 if ((m->flags & PG_FICTITIOUS) == 0) {
2156 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) {
2157 if ((m->flags & PG_UNMANAGED) == 0)
2159 atomic_add_int(&vmstats.v_wire_count, 1);
2161 KASSERT(m->wire_count != 0,
2162 ("vm_page_wire: wire_count overflow m=%p", m));
2167 * Release one wiring of this page, potentially enabling it to be paged again.
2169 * Many pages placed on the inactive queue should actually go
2170 * into the cache, but it is difficult to figure out which. What
2171 * we do instead, if the inactive target is well met, is to put
2172 * clean pages at the head of the inactive queue instead of the tail.
2173 * This will cause them to be moved to the cache more quickly and
2174 * if not actively re-referenced, freed more quickly. If we just
2175 * stick these pages at the end of the inactive queue, heavy filesystem
2176 * meta-data accesses can cause an unnecessary paging load on memory bound
2177 * processes. This optimization causes one-time-use metadata to be
2178 * reused more quickly.
2180 * Pages marked PG_NEED_COMMIT are always activated and never placed on
2181 * the inactive queue. This helps the pageout daemon determine memory
2182 * pressure and act on out-of-memory situations more quickly.
2184 * BUT, if we are in a low-memory situation we have no choice but to
2185 * put clean pages on the cache queue.
2187 * A number of routines use vm_page_unwire() to guarantee that the page
2188 * will go into either the inactive or active queues, and will NEVER
2189 * be placed in the cache - for example, just after dirtying a page.
2190 * dirty pages in the cache are not allowed.
2192 * The page queues must be locked.
2193 * This routine may not block.
2196 vm_page_unwire(vm_page_t m, int activate)
2198 KKASSERT(m->flags & PG_BUSY);
2199 if (m->flags & PG_FICTITIOUS) {
2201 } else if (m->wire_count <= 0) {
2202 panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
2204 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) {
2205 atomic_add_int(&vmstats.v_wire_count, -1);
2206 if (m->flags & PG_UNMANAGED) {
2208 } else if (activate || (m->flags & PG_NEED_COMMIT)) {
2209 vm_page_spin_lock(m);
2210 _vm_page_add_queue_spinlocked(m,
2211 PQ_ACTIVE + m->pc, 0);
2212 _vm_page_and_queue_spin_unlock(m);
2214 vm_page_spin_lock(m);
2215 vm_page_flag_clear(m, PG_WINATCFLS);
2216 _vm_page_add_queue_spinlocked(m,
2217 PQ_INACTIVE + m->pc, 0);
2218 ++vm_swapcache_inactive_heuristic;
2219 _vm_page_and_queue_spin_unlock(m);
2226 * Move the specified page to the inactive queue. If the page has
2227 * any associated swap, the swap is deallocated.
2229 * Normally athead is 0 resulting in LRU operation. athead is set
2230 * to 1 if we want this page to be 'as if it were placed in the cache',
2231 * except without unmapping it from the process address space.
2233 * vm_page's spinlock must be held on entry and will remain held on return.
2234 * This routine may not block.
2237 _vm_page_deactivate_locked(vm_page_t m, int athead)
2242 * Ignore if already inactive.
2244 if (m->queue - m->pc == PQ_INACTIVE)
2246 _vm_page_queue_spin_lock(m);
2247 oqueue = _vm_page_rem_queue_spinlocked(m);
2249 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
2250 if (oqueue == PQ_CACHE)
2251 mycpu->gd_cnt.v_reactivated++;
2252 vm_page_flag_clear(m, PG_WINATCFLS);
2253 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead);
2255 ++vm_swapcache_inactive_heuristic;
2257 _vm_page_queue_spin_unlock(m);
2258 /* leaves vm_page spinlocked */
2262 * Attempt to deactivate a page.
2267 vm_page_deactivate(vm_page_t m)
2269 vm_page_spin_lock(m);
2270 _vm_page_deactivate_locked(m, 0);
2271 vm_page_spin_unlock(m);
2275 vm_page_deactivate_locked(vm_page_t m)
2277 _vm_page_deactivate_locked(m, 0);
2281 * Attempt to move a page to PQ_CACHE.
2283 * Returns 0 on failure, 1 on success
2285 * The page should NOT be busied by the caller. This function will validate
2286 * whether the page can be safely moved to the cache.
2289 vm_page_try_to_cache(vm_page_t m)
2291 vm_page_spin_lock(m);
2292 if (vm_page_busy_try(m, TRUE)) {
2293 vm_page_spin_unlock(m);
2296 if (m->dirty || m->hold_count || m->wire_count ||
2297 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT))) {
2298 if (_vm_page_wakeup(m)) {
2299 vm_page_spin_unlock(m);
2302 vm_page_spin_unlock(m);
2306 vm_page_spin_unlock(m);
2309 * Page busied by us and no longer spinlocked. Dirty pages cannot
2310 * be moved to the cache.
2312 vm_page_test_dirty(m);
2313 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2322 * Attempt to free the page. If we cannot free it, we do nothing.
2323 * 1 is returned on success, 0 on failure.
2328 vm_page_try_to_free(vm_page_t m)
2330 vm_page_spin_lock(m);
2331 if (vm_page_busy_try(m, TRUE)) {
2332 vm_page_spin_unlock(m);
2337 * The page can be in any state, including already being on the free
2338 * queue. Check to see if it really can be freed.
2340 if (m->dirty || /* can't free if it is dirty */
2341 m->hold_count || /* or held (XXX may be wrong) */
2342 m->wire_count || /* or wired */
2343 (m->flags & (PG_UNMANAGED | /* or unmanaged */
2344 PG_NEED_COMMIT)) || /* or needs a commit */
2345 m->queue - m->pc == PQ_FREE || /* already on PQ_FREE */
2346 m->queue - m->pc == PQ_HOLD) { /* already on PQ_HOLD */
2347 if (_vm_page_wakeup(m)) {
2348 vm_page_spin_unlock(m);
2351 vm_page_spin_unlock(m);
2355 vm_page_spin_unlock(m);
2358 * We can probably free the page.
2360 * Page busied by us and no longer spinlocked. Dirty pages will
2361 * not be freed by this function. We have to re-test the
2362 * dirty bit after cleaning out the pmaps.
2364 vm_page_test_dirty(m);
2365 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2369 vm_page_protect(m, VM_PROT_NONE);
2370 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2381 * Put the specified page onto the page cache queue (if appropriate).
2383 * The page must be busy, and this routine will release the busy and
2384 * possibly even free the page.
2387 vm_page_cache(vm_page_t m)
2389 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) ||
2390 m->busy || m->wire_count || m->hold_count) {
2391 kprintf("vm_page_cache: attempting to cache busy/held page\n");
2397 * Already in the cache (and thus not mapped)
2399 if ((m->queue - m->pc) == PQ_CACHE) {
2400 KKASSERT((m->flags & PG_MAPPED) == 0);
2406 * Caller is required to test m->dirty, but note that the act of
2407 * removing the page from its maps can cause it to become dirty
2408 * on an SMP system due to another cpu running in usermode.
2411 panic("vm_page_cache: caching a dirty page, pindex: %ld",
2416 * Remove all pmaps and indicate that the page is not
2417 * writeable or mapped. Our vm_page_protect() call may
2418 * have blocked (especially w/ VM_PROT_NONE), so recheck
2421 vm_page_protect(m, VM_PROT_NONE);
2422 if ((m->flags & (PG_UNMANAGED | PG_MAPPED)) ||
2423 m->busy || m->wire_count || m->hold_count) {
2425 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2426 vm_page_deactivate(m);
2429 _vm_page_and_queue_spin_lock(m);
2430 _vm_page_rem_queue_spinlocked(m);
2431 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0);
2432 _vm_page_queue_spin_unlock(m);
2433 if (_vm_page_wakeup(m)) {
2434 vm_page_spin_unlock(m);
2437 vm_page_spin_unlock(m);
2439 vm_page_free_wakeup();
2444 * vm_page_dontneed()
2446 * Cache, deactivate, or do nothing as appropriate. This routine
2447 * is typically used by madvise() MADV_DONTNEED.
2449 * Generally speaking we want to move the page into the cache so
2450 * it gets reused quickly. However, this can result in a silly syndrome
2451 * due to the page recycling too quickly. Small objects will not be
2452 * fully cached. On the otherhand, if we move the page to the inactive
2453 * queue we wind up with a problem whereby very large objects
2454 * unnecessarily blow away our inactive and cache queues.
2456 * The solution is to move the pages based on a fixed weighting. We
2457 * either leave them alone, deactivate them, or move them to the cache,
2458 * where moving them to the cache has the highest weighting.
2459 * By forcing some pages into other queues we eventually force the
2460 * system to balance the queues, potentially recovering other unrelated
2461 * space from active. The idea is to not force this to happen too
2464 * The page must be busied.
2467 vm_page_dontneed(vm_page_t m)
2469 static int dnweight;
2476 * occassionally leave the page alone
2478 if ((dnw & 0x01F0) == 0 ||
2479 m->queue - m->pc == PQ_INACTIVE ||
2480 m->queue - m->pc == PQ_CACHE
2482 if (m->act_count >= ACT_INIT)
2488 * If vm_page_dontneed() is inactivating a page, it must clear
2489 * the referenced flag; otherwise the pagedaemon will see references
2490 * on the page in the inactive queue and reactivate it. Until the
2491 * page can move to the cache queue, madvise's job is not done.
2493 vm_page_flag_clear(m, PG_REFERENCED);
2494 pmap_clear_reference(m);
2497 vm_page_test_dirty(m);
2499 if (m->dirty || (dnw & 0x0070) == 0) {
2501 * Deactivate the page 3 times out of 32.
2506 * Cache the page 28 times out of every 32. Note that
2507 * the page is deactivated instead of cached, but placed
2508 * at the head of the queue instead of the tail.
2512 vm_page_spin_lock(m);
2513 _vm_page_deactivate_locked(m, head);
2514 vm_page_spin_unlock(m);
2518 * These routines manipulate the 'soft busy' count for a page. A soft busy
2519 * is almost like PG_BUSY except that it allows certain compatible operations
2520 * to occur on the page while it is busy. For example, a page undergoing a
2521 * write can still be mapped read-only.
2523 * Because vm_pages can overlap buffers m->busy can be > 1. m->busy is only
2524 * adjusted while the vm_page is PG_BUSY so the flash will occur when the
2525 * busy bit is cleared.
2528 vm_page_io_start(vm_page_t m)
2530 KASSERT(m->flags & PG_BUSY, ("vm_page_io_start: page not busy!!!"));
2531 atomic_add_char(&m->busy, 1);
2532 vm_page_flag_set(m, PG_SBUSY);
2536 vm_page_io_finish(vm_page_t m)
2538 KASSERT(m->flags & PG_BUSY, ("vm_page_io_finish: page not busy!!!"));
2539 atomic_subtract_char(&m->busy, 1);
2541 vm_page_flag_clear(m, PG_SBUSY);
2545 * Indicate that a clean VM page requires a filesystem commit and cannot
2546 * be reused. Used by tmpfs.
2549 vm_page_need_commit(vm_page_t m)
2551 vm_page_flag_set(m, PG_NEED_COMMIT);
2552 vm_object_set_writeable_dirty(m->object);
2556 vm_page_clear_commit(vm_page_t m)
2558 vm_page_flag_clear(m, PG_NEED_COMMIT);
2562 * Grab a page, blocking if it is busy and allocating a page if necessary.
2563 * A busy page is returned or NULL. The page may or may not be valid and
2564 * might not be on a queue (the caller is responsible for the disposition of
2567 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the
2568 * page will be zero'd and marked valid.
2570 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked
2571 * valid even if it already exists.
2573 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also
2574 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified.
2575 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified.
2577 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
2578 * always returned if we had blocked.
2580 * This routine may not be called from an interrupt.
2582 * PG_ZERO is *ALWAYS* cleared by this routine.
2584 * No other requirements.
2587 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2592 KKASSERT(allocflags &
2593 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
2594 vm_object_hold(object);
2596 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error);
2598 vm_page_sleep_busy(m, TRUE, "pgrbwt");
2599 if ((allocflags & VM_ALLOC_RETRY) == 0) {
2604 } else if (m == NULL) {
2605 if (allocflags & VM_ALLOC_RETRY)
2606 allocflags |= VM_ALLOC_NULL_OK;
2607 m = vm_page_alloc(object, pindex,
2608 allocflags & ~VM_ALLOC_RETRY);
2612 if ((allocflags & VM_ALLOC_RETRY) == 0)
2621 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid.
2623 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set
2624 * valid even if already valid.
2626 if (m->valid == 0) {
2627 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) {
2628 if ((m->flags & PG_ZERO) == 0)
2629 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2630 m->valid = VM_PAGE_BITS_ALL;
2632 } else if (allocflags & VM_ALLOC_FORCE_ZERO) {
2633 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2634 m->valid = VM_PAGE_BITS_ALL;
2636 vm_page_flag_clear(m, PG_ZERO);
2638 vm_object_drop(object);
2643 * Mapping function for valid bits or for dirty bits in
2644 * a page. May not block.
2646 * Inputs are required to range within a page.
2652 vm_page_bits(int base, int size)
2658 base + size <= PAGE_SIZE,
2659 ("vm_page_bits: illegal base/size %d/%d", base, size)
2662 if (size == 0) /* handle degenerate case */
2665 first_bit = base >> DEV_BSHIFT;
2666 last_bit = (base + size - 1) >> DEV_BSHIFT;
2668 return ((2 << last_bit) - (1 << first_bit));
2672 * Sets portions of a page valid and clean. The arguments are expected
2673 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2674 * of any partial chunks touched by the range. The invalid portion of
2675 * such chunks will be zero'd.
2677 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically
2678 * align base to DEV_BSIZE so as not to mark clean a partially
2679 * truncated device block. Otherwise the dirty page status might be
2682 * This routine may not block.
2684 * (base + size) must be less then or equal to PAGE_SIZE.
2687 _vm_page_zero_valid(vm_page_t m, int base, int size)
2692 if (size == 0) /* handle degenerate case */
2696 * If the base is not DEV_BSIZE aligned and the valid
2697 * bit is clear, we have to zero out a portion of the
2701 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2702 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
2704 pmap_zero_page_area(
2712 * If the ending offset is not DEV_BSIZE aligned and the
2713 * valid bit is clear, we have to zero out a portion of
2717 endoff = base + size;
2719 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2720 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
2722 pmap_zero_page_area(
2725 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
2731 * Set valid, clear dirty bits. If validating the entire
2732 * page we can safely clear the pmap modify bit. We also
2733 * use this opportunity to clear the PG_NOSYNC flag. If a process
2734 * takes a write fault on a MAP_NOSYNC memory area the flag will
2737 * We set valid bits inclusive of any overlap, but we can only
2738 * clear dirty bits for DEV_BSIZE chunks that are fully within
2741 * Page must be busied?
2742 * No other requirements.
2745 vm_page_set_valid(vm_page_t m, int base, int size)
2747 _vm_page_zero_valid(m, base, size);
2748 m->valid |= vm_page_bits(base, size);
2753 * Set valid bits and clear dirty bits.
2755 * NOTE: This function does not clear the pmap modified bit.
2756 * Also note that e.g. NFS may use a byte-granular base
2759 * WARNING: Page must be busied? But vfs_clean_one_page() will call
2760 * this without necessarily busying the page (via bdwrite()).
2761 * So for now vm_token must also be held.
2763 * No other requirements.
2766 vm_page_set_validclean(vm_page_t m, int base, int size)
2770 _vm_page_zero_valid(m, base, size);
2771 pagebits = vm_page_bits(base, size);
2772 m->valid |= pagebits;
2773 m->dirty &= ~pagebits;
2774 if (base == 0 && size == PAGE_SIZE) {
2775 /*pmap_clear_modify(m);*/
2776 vm_page_flag_clear(m, PG_NOSYNC);
2781 * Set valid & dirty. Used by buwrite()
2783 * WARNING: Page must be busied? But vfs_dirty_one_page() will
2784 * call this function in buwrite() so for now vm_token must
2787 * No other requirements.
2790 vm_page_set_validdirty(vm_page_t m, int base, int size)
2794 pagebits = vm_page_bits(base, size);
2795 m->valid |= pagebits;
2796 m->dirty |= pagebits;
2798 vm_object_set_writeable_dirty(m->object);
2804 * NOTE: This function does not clear the pmap modified bit.
2805 * Also note that e.g. NFS may use a byte-granular base
2808 * Page must be busied?
2809 * No other requirements.
2812 vm_page_clear_dirty(vm_page_t m, int base, int size)
2814 m->dirty &= ~vm_page_bits(base, size);
2815 if (base == 0 && size == PAGE_SIZE) {
2816 /*pmap_clear_modify(m);*/
2817 vm_page_flag_clear(m, PG_NOSYNC);
2822 * Make the page all-dirty.
2824 * Also make sure the related object and vnode reflect the fact that the
2825 * object may now contain a dirty page.
2827 * Page must be busied?
2828 * No other requirements.
2831 vm_page_dirty(vm_page_t m)
2834 int pqtype = m->queue - m->pc;
2836 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE,
2837 ("vm_page_dirty: page in free/cache queue!"));
2838 if (m->dirty != VM_PAGE_BITS_ALL) {
2839 m->dirty = VM_PAGE_BITS_ALL;
2841 vm_object_set_writeable_dirty(m->object);
2846 * Invalidates DEV_BSIZE'd chunks within a page. Both the
2847 * valid and dirty bits for the effected areas are cleared.
2849 * Page must be busied?
2851 * No other requirements.
2854 vm_page_set_invalid(vm_page_t m, int base, int size)
2858 bits = vm_page_bits(base, size);
2861 m->object->generation++;
2865 * The kernel assumes that the invalid portions of a page contain
2866 * garbage, but such pages can be mapped into memory by user code.
2867 * When this occurs, we must zero out the non-valid portions of the
2868 * page so user code sees what it expects.
2870 * Pages are most often semi-valid when the end of a file is mapped
2871 * into memory and the file's size is not page aligned.
2873 * Page must be busied?
2874 * No other requirements.
2877 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
2883 * Scan the valid bits looking for invalid sections that
2884 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
2885 * valid bit may be set ) have already been zerod by
2886 * vm_page_set_validclean().
2888 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
2889 if (i == (PAGE_SIZE / DEV_BSIZE) ||
2890 (m->valid & (1 << i))
2893 pmap_zero_page_area(
2896 (i - b) << DEV_BSHIFT
2904 * setvalid is TRUE when we can safely set the zero'd areas
2905 * as being valid. We can do this if there are no cache consistency
2906 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
2909 m->valid = VM_PAGE_BITS_ALL;
2913 * Is a (partial) page valid? Note that the case where size == 0
2914 * will return FALSE in the degenerate case where the page is entirely
2915 * invalid, and TRUE otherwise.
2918 * No other requirements.
2921 vm_page_is_valid(vm_page_t m, int base, int size)
2923 int bits = vm_page_bits(base, size);
2925 if (m->valid && ((m->valid & bits) == bits))
2932 * update dirty bits from pmap/mmu. May not block.
2934 * Caller must hold the page busy
2937 vm_page_test_dirty(vm_page_t m)
2939 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
2945 * Register an action, associating it with its vm_page
2948 vm_page_register_action(vm_page_action_t action, vm_page_event_t event)
2950 struct vm_page_action_list *list;
2953 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
2954 list = &action_list[hv];
2956 lwkt_gettoken(&vm_token);
2957 vm_page_flag_set(action->m, PG_ACTIONLIST);
2958 action->event = event;
2959 LIST_INSERT_HEAD(list, action, entry);
2960 lwkt_reltoken(&vm_token);
2964 * Unregister an action, disassociating it from its related vm_page
2967 vm_page_unregister_action(vm_page_action_t action)
2969 struct vm_page_action_list *list;
2972 lwkt_gettoken(&vm_token);
2973 if (action->event != VMEVENT_NONE) {
2974 action->event = VMEVENT_NONE;
2975 LIST_REMOVE(action, entry);
2977 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
2978 list = &action_list[hv];
2979 if (LIST_EMPTY(list))
2980 vm_page_flag_clear(action->m, PG_ACTIONLIST);
2982 lwkt_reltoken(&vm_token);
2986 * Issue an event on a VM page. Corresponding action structures are
2987 * removed from the page's list and called.
2989 * If the vm_page has no more pending action events we clear its
2990 * PG_ACTIONLIST flag.
2993 vm_page_event_internal(vm_page_t m, vm_page_event_t event)
2995 struct vm_page_action_list *list;
2996 struct vm_page_action *scan;
2997 struct vm_page_action *next;
3001 hv = (int)((intptr_t)m >> 8) & VMACTION_HMASK;
3002 list = &action_list[hv];
3005 lwkt_gettoken(&vm_token);
3006 LIST_FOREACH_MUTABLE(scan, list, entry, next) {
3008 if (scan->event == event) {
3009 scan->event = VMEVENT_NONE;
3010 LIST_REMOVE(scan, entry);
3011 scan->func(m, scan);
3019 vm_page_flag_clear(m, PG_ACTIONLIST);
3020 lwkt_reltoken(&vm_token);
3023 #include "opt_ddb.h"
3025 #include <sys/kernel.h>
3027 #include <ddb/ddb.h>
3029 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3031 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count);
3032 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count);
3033 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count);
3034 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count);
3035 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count);
3036 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved);
3037 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min);
3038 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target);
3039 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min);
3040 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target);
3043 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3046 db_printf("PQ_FREE:");
3047 for(i=0;i<PQ_L2_SIZE;i++) {
3048 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
3052 db_printf("PQ_CACHE:");
3053 for(i=0;i<PQ_L2_SIZE;i++) {
3054 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
3058 db_printf("PQ_ACTIVE:");
3059 for(i=0;i<PQ_L2_SIZE;i++) {
3060 db_printf(" %d", vm_page_queues[PQ_ACTIVE + i].lcnt);
3064 db_printf("PQ_INACTIVE:");
3065 for(i=0;i<PQ_L2_SIZE;i++) {
3066 db_printf(" %d", vm_page_queues[PQ_INACTIVE + i].lcnt);