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 * Each page queue has its own spin lock, which is fairly optimal for
533 * allocating and freeing pages at least.
535 * The caller must hold the vm_page_spin_lock() before locking a vm_page's
536 * queue spinlock via this function. Also note that m->queue cannot change
537 * unless both the page and queue are locked.
541 _vm_page_queue_spin_lock(vm_page_t m)
546 if (queue != PQ_NONE) {
547 spin_lock(&vm_page_queues[queue].spin);
548 KKASSERT(queue == m->queue);
554 _vm_page_queue_spin_unlock(vm_page_t m)
560 if (queue != PQ_NONE)
561 spin_unlock(&vm_page_queues[queue].spin);
566 _vm_page_queues_spin_lock(u_short queue)
569 if (queue != PQ_NONE)
570 spin_lock(&vm_page_queues[queue].spin);
576 _vm_page_queues_spin_unlock(u_short queue)
579 if (queue != PQ_NONE)
580 spin_unlock(&vm_page_queues[queue].spin);
584 vm_page_queue_spin_lock(vm_page_t m)
586 _vm_page_queue_spin_lock(m);
590 vm_page_queues_spin_lock(u_short queue)
592 _vm_page_queues_spin_lock(queue);
596 vm_page_queue_spin_unlock(vm_page_t m)
598 _vm_page_queue_spin_unlock(m);
602 vm_page_queues_spin_unlock(u_short queue)
604 _vm_page_queues_spin_unlock(queue);
608 * This locks the specified vm_page and its queue in the proper order
609 * (page first, then queue). The queue may change so the caller must
614 _vm_page_and_queue_spin_lock(vm_page_t m)
616 vm_page_spin_lock(m);
617 _vm_page_queue_spin_lock(m);
622 _vm_page_and_queue_spin_unlock(vm_page_t m)
624 _vm_page_queues_spin_unlock(m->queue);
625 vm_page_spin_unlock(m);
629 vm_page_and_queue_spin_unlock(vm_page_t m)
631 _vm_page_and_queue_spin_unlock(m);
635 vm_page_and_queue_spin_lock(vm_page_t m)
637 _vm_page_and_queue_spin_lock(m);
641 * Helper function removes vm_page from its current queue.
642 * Returns the base queue the page used to be on.
644 * The vm_page and the queue must be spinlocked.
645 * This function will unlock the queue but leave the page spinlocked.
647 static __inline u_short
648 _vm_page_rem_queue_spinlocked(vm_page_t m)
650 struct vpgqueues *pq;
654 if (queue != PQ_NONE) {
655 pq = &vm_page_queues[queue];
656 TAILQ_REMOVE(&pq->pl, m, pageq);
657 atomic_add_int(pq->cnt, -1);
660 vm_page_queues_spin_unlock(queue);
661 if ((queue - m->pc) == PQ_FREE && (m->flags & PG_ZERO))
662 atomic_subtract_int(&vm_page_zero_count, 1);
663 if ((queue - m->pc) == PQ_CACHE || (queue - m->pc) == PQ_FREE)
664 return (queue - m->pc);
670 * Helper function places the vm_page on the specified queue.
672 * The vm_page must be spinlocked.
673 * This function will return with both the page and the queue locked.
676 _vm_page_add_queue_spinlocked(vm_page_t m, u_short queue, int athead)
678 struct vpgqueues *pq;
680 KKASSERT(m->queue == PQ_NONE);
682 if (queue != PQ_NONE) {
683 vm_page_queues_spin_lock(queue);
684 pq = &vm_page_queues[queue];
686 atomic_add_int(pq->cnt, 1);
690 * Put zero'd pages on the end ( where we look for zero'd pages
691 * first ) and non-zerod pages at the head.
693 if (queue - m->pc == PQ_FREE) {
694 if (m->flags & PG_ZERO) {
695 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
696 atomic_add_int(&vm_page_zero_count, 1);
698 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
701 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
703 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
705 /* leave the queue spinlocked */
710 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE)
711 * m->busy is zero. Returns TRUE if it had to sleep, FALSE if we
712 * did not. Only one sleep call will be made before returning.
714 * This function does NOT busy the page and on return the page is not
715 * guaranteed to be available.
718 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg)
726 if ((flags & PG_BUSY) == 0 &&
727 (also_m_busy == 0 || (flags & PG_SBUSY) == 0)) {
730 tsleep_interlock(m, 0);
731 if (atomic_cmpset_int(&m->flags, flags,
732 flags | PG_WANTED | PG_REFERENCED)) {
733 tsleep(m, PINTERLOCKED, msg, 0);
740 * Wait until PG_BUSY can be set, then set it. If also_m_busy is TRUE we
741 * also wait for m->busy to become 0 before setting PG_BUSY.
744 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m,
745 int also_m_busy, const char *msg
753 if (flags & PG_BUSY) {
754 tsleep_interlock(m, 0);
755 if (atomic_cmpset_int(&m->flags, flags,
756 flags | PG_WANTED | PG_REFERENCED)) {
757 tsleep(m, PINTERLOCKED, msg, 0);
759 } else if (also_m_busy && (flags & PG_SBUSY)) {
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);
766 if (atomic_cmpset_int(&m->flags, flags,
770 m->busy_line = lineno;
779 * Attempt to set PG_BUSY. If also_m_busy is TRUE we only succeed if m->busy
782 * Returns non-zero on failure.
785 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy
795 if (also_m_busy && (flags & PG_SBUSY))
797 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
800 m->busy_line = lineno;
808 * Clear the PG_BUSY flag and return non-zero to indicate to the caller
809 * that a wakeup() should be performed.
811 * The vm_page must be spinlocked and will remain spinlocked on return.
812 * The related queue must NOT be spinlocked (which could deadlock us).
818 _vm_page_wakeup(vm_page_t m)
825 if (atomic_cmpset_int(&m->flags, flags,
826 flags & ~(PG_BUSY | PG_WANTED))) {
830 return(flags & PG_WANTED);
834 * Clear the PG_BUSY flag and wakeup anyone waiting for the page. This
835 * is typically the last call you make on a page before moving onto
839 vm_page_wakeup(vm_page_t m)
841 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!"));
842 vm_page_spin_lock(m);
843 if (_vm_page_wakeup(m)) {
844 vm_page_spin_unlock(m);
847 vm_page_spin_unlock(m);
852 * Holding a page keeps it from being reused. Other parts of the system
853 * can still disassociate the page from its current object and free it, or
854 * perform read or write I/O on it and/or otherwise manipulate the page,
855 * but if the page is held the VM system will leave the page and its data
856 * intact and not reuse the page for other purposes until the last hold
857 * reference is released. (see vm_page_wire() if you want to prevent the
858 * page from being disassociated from its object too).
860 * The caller must still validate the contents of the page and, if necessary,
861 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete
862 * before manipulating the page.
864 * XXX get vm_page_spin_lock() here and move FREE->HOLD if necessary
867 vm_page_hold(vm_page_t m)
869 vm_page_spin_lock(m);
870 atomic_add_int(&m->hold_count, 1);
871 if (m->queue - m->pc == PQ_FREE) {
872 _vm_page_queue_spin_lock(m);
873 _vm_page_rem_queue_spinlocked(m);
874 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
875 _vm_page_queue_spin_unlock(m);
877 vm_page_spin_unlock(m);
881 * The opposite of vm_page_hold(). A page can be freed while being held,
882 * which places it on the PQ_HOLD queue. If we are able to busy the page
883 * after the hold count drops to zero we will move the page to the
884 * appropriate PQ_FREE queue by calling vm_page_free_toq().
887 vm_page_unhold(vm_page_t m)
889 vm_page_spin_lock(m);
890 atomic_add_int(&m->hold_count, -1);
891 if (m->hold_count == 0 && m->queue - m->pc == PQ_HOLD) {
892 _vm_page_queue_spin_lock(m);
893 _vm_page_rem_queue_spinlocked(m);
894 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0);
895 _vm_page_queue_spin_unlock(m);
897 vm_page_spin_unlock(m);
901 * Inserts the given vm_page into the object and object list.
903 * The pagetables are not updated but will presumably fault the page
904 * in if necessary, or if a kernel page the caller will at some point
905 * enter the page into the kernel's pmap. We are not allowed to block
906 * here so we *can't* do this anyway.
908 * This routine may not block.
909 * This routine must be called with the vm_object held.
910 * This routine must be called with a critical section held.
912 * This routine returns TRUE if the page was inserted into the object
913 * successfully, and FALSE if the page already exists in the object.
916 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
918 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(object));
919 if (m->object != NULL)
920 panic("vm_page_insert: already inserted");
922 object->generation++;
925 * Record the object/offset pair in this page and add the
926 * pv_list_count of the page to the object.
928 * The vm_page spin lock is required for interactions with the pmap.
930 vm_page_spin_lock(m);
933 if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) {
936 vm_page_spin_unlock(m);
939 object->resident_page_count++;
940 /* atomic_add_int(&object->agg_pv_list_count, m->md.pv_list_count); */
941 vm_page_spin_unlock(m);
944 * Since we are inserting a new and possibly dirty page,
945 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
947 if ((m->valid & m->dirty) ||
948 (m->flags & (PG_WRITEABLE | PG_NEED_COMMIT)))
949 vm_object_set_writeable_dirty(object);
952 * Checks for a swap assignment and sets PG_SWAPPED if appropriate.
954 swap_pager_page_inserted(m);
959 * Removes the given vm_page_t from the (object,index) table
961 * The underlying pmap entry (if any) is NOT removed here.
962 * This routine may not block.
964 * The page must be BUSY and will remain BUSY on return.
965 * No other requirements.
967 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave
971 vm_page_remove(vm_page_t m)
975 if (m->object == NULL) {
979 if ((m->flags & PG_BUSY) == 0)
980 panic("vm_page_remove: page not busy");
984 vm_object_hold(object);
987 * Remove the page from the object and update the object.
989 * The vm_page spin lock is required for interactions with the pmap.
991 vm_page_spin_lock(m);
992 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m);
993 object->resident_page_count--;
994 /* atomic_add_int(&object->agg_pv_list_count, -m->md.pv_list_count); */
996 vm_page_spin_unlock(m);
998 object->generation++;
1000 vm_object_drop(object);
1004 * Locate and return the page at (object, pindex), or NULL if the
1005 * page could not be found.
1007 * The caller must hold the vm_object token.
1010 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1015 * Search the hash table for this object/offset pair
1017 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1018 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1019 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex));
1024 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object,
1026 int also_m_busy, const char *msg
1032 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1033 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1035 KKASSERT(m->object == object && m->pindex == pindex);
1038 if (flags & PG_BUSY) {
1039 tsleep_interlock(m, 0);
1040 if (atomic_cmpset_int(&m->flags, flags,
1041 flags | PG_WANTED | PG_REFERENCED)) {
1042 tsleep(m, PINTERLOCKED, msg, 0);
1043 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1046 } else if (also_m_busy && (flags & PG_SBUSY)) {
1047 tsleep_interlock(m, 0);
1048 if (atomic_cmpset_int(&m->flags, flags,
1049 flags | PG_WANTED | PG_REFERENCED)) {
1050 tsleep(m, PINTERLOCKED, msg, 0);
1051 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1054 } else if (atomic_cmpset_int(&m->flags, flags,
1056 #ifdef VM_PAGE_DEBUG
1057 m->busy_func = func;
1058 m->busy_line = lineno;
1067 * Attempt to lookup and busy a page.
1069 * Returns NULL if the page could not be found
1071 * Returns a vm_page and error == TRUE if the page exists but could not
1074 * Returns a vm_page and error == FALSE on success.
1077 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object,
1079 int also_m_busy, int *errorp
1085 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1086 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1089 KKASSERT(m->object == object && m->pindex == pindex);
1092 if (flags & PG_BUSY) {
1096 if (also_m_busy && (flags & PG_SBUSY)) {
1100 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
1101 #ifdef VM_PAGE_DEBUG
1102 m->busy_func = func;
1103 m->busy_line = lineno;
1112 * Caller must hold the related vm_object
1115 vm_page_next(vm_page_t m)
1119 next = vm_page_rb_tree_RB_NEXT(m);
1120 if (next && next->pindex != m->pindex + 1)
1128 * Move the given vm_page from its current object to the specified
1129 * target object/offset. The page must be busy and will remain so
1132 * new_object must be held.
1133 * This routine might block. XXX ?
1135 * NOTE: Swap associated with the page must be invalidated by the move. We
1136 * have to do this for several reasons: (1) we aren't freeing the
1137 * page, (2) we are dirtying the page, (3) the VM system is probably
1138 * moving the page from object A to B, and will then later move
1139 * the backing store from A to B and we can't have a conflict.
1141 * NOTE: We *always* dirty the page. It is necessary both for the
1142 * fact that we moved it, and because we may be invalidating
1143 * swap. If the page is on the cache, we have to deactivate it
1144 * or vm_page_dirty() will panic. Dirty pages are not allowed
1148 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1150 KKASSERT(m->flags & PG_BUSY);
1151 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(new_object));
1153 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(m->object));
1156 if (vm_page_insert(m, new_object, new_pindex) == FALSE) {
1157 panic("vm_page_rename: target exists (%p,%"PRIu64")",
1158 new_object, new_pindex);
1160 if (m->queue - m->pc == PQ_CACHE)
1161 vm_page_deactivate(m);
1166 * vm_page_unqueue() without any wakeup. This routine is used when a page
1167 * is being moved between queues or otherwise is to remain BUSYied by the
1170 * This routine may not block.
1173 vm_page_unqueue_nowakeup(vm_page_t m)
1175 vm_page_and_queue_spin_lock(m);
1176 (void)_vm_page_rem_queue_spinlocked(m);
1177 vm_page_spin_unlock(m);
1181 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
1184 * This routine may not block.
1187 vm_page_unqueue(vm_page_t m)
1191 vm_page_and_queue_spin_lock(m);
1192 queue = _vm_page_rem_queue_spinlocked(m);
1193 if (queue == PQ_FREE || queue == PQ_CACHE) {
1194 vm_page_spin_unlock(m);
1195 pagedaemon_wakeup();
1197 vm_page_spin_unlock(m);
1202 * vm_page_list_find()
1204 * Find a page on the specified queue with color optimization.
1206 * The page coloring optimization attempts to locate a page that does
1207 * not overload other nearby pages in the object in the cpu's L1 or L2
1208 * caches. We need this optimization because cpu caches tend to be
1209 * physical caches, while object spaces tend to be virtual.
1211 * On MP systems each PQ_FREE and PQ_CACHE color queue has its own spinlock
1212 * and the algorithm is adjusted to localize allocations on a per-core basis.
1213 * This is done by 'twisting' the colors.
1215 * The page is returned spinlocked and removed from its queue (it will
1216 * be on PQ_NONE), or NULL. The page is not PG_BUSY'd. The caller
1217 * is responsible for dealing with the busy-page case (usually by
1218 * deactivating the page and looping).
1220 * NOTE: This routine is carefully inlined. A non-inlined version
1221 * is available for outside callers but the only critical path is
1222 * from within this source file.
1224 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE
1225 * represent stable storage, allowing us to order our locks vm_page
1226 * first, then queue.
1230 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
1236 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist);
1238 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl);
1240 m = _vm_page_list_find2(basequeue, index);
1243 vm_page_and_queue_spin_lock(m);
1244 if (m->queue == basequeue + index) {
1245 _vm_page_rem_queue_spinlocked(m);
1246 /* vm_page_t spin held, no queue spin */
1249 vm_page_and_queue_spin_unlock(m);
1255 _vm_page_list_find2(int basequeue, int index)
1259 struct vpgqueues *pq;
1261 pq = &vm_page_queues[basequeue];
1264 * Note that for the first loop, index+i and index-i wind up at the
1265 * same place. Even though this is not totally optimal, we've already
1266 * blown it by missing the cache case so we do not care.
1268 for (i = PQ_L2_SIZE / 2; i > 0; --i) {
1270 m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl);
1272 _vm_page_and_queue_spin_lock(m);
1274 basequeue + ((index + i) & PQ_L2_MASK)) {
1275 _vm_page_rem_queue_spinlocked(m);
1278 _vm_page_and_queue_spin_unlock(m);
1281 m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl);
1283 _vm_page_and_queue_spin_lock(m);
1285 basequeue + ((index - i) & PQ_L2_MASK)) {
1286 _vm_page_rem_queue_spinlocked(m);
1289 _vm_page_and_queue_spin_unlock(m);
1299 * Returns a vm_page candidate for allocation. The page is not busied so
1300 * it can move around. The caller must busy the page (and typically
1301 * deactivate it if it cannot be busied!)
1303 * Returns a spinlocked vm_page that has been removed from its queue.
1306 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
1308 return(_vm_page_list_find(basequeue, index, prefer_zero));
1312 * Find a page on the cache queue with color optimization, remove it
1313 * from the queue, and busy it. The returned page will not be spinlocked.
1315 * A candidate failure will be deactivated. Candidates can fail due to
1316 * being busied by someone else, in which case they will be deactivated.
1318 * This routine may not block.
1322 vm_page_select_cache(u_short pg_color)
1327 m = _vm_page_list_find(PQ_CACHE, pg_color & PQ_L2_MASK, FALSE);
1331 * (m) has been removed from its queue and spinlocked
1333 if (vm_page_busy_try(m, TRUE)) {
1334 _vm_page_deactivate_locked(m, 0);
1335 vm_page_spin_unlock(m);
1338 * We successfully busied the page
1340 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) == 0 &&
1341 m->hold_count == 0 &&
1342 m->wire_count == 0 &&
1343 (m->dirty & m->valid) == 0) {
1344 vm_page_spin_unlock(m);
1345 pagedaemon_wakeup();
1350 * The page cannot be recycled, deactivate it.
1352 _vm_page_deactivate_locked(m, 0);
1353 if (_vm_page_wakeup(m)) {
1354 vm_page_spin_unlock(m);
1357 vm_page_spin_unlock(m);
1365 * Find a free or zero page, with specified preference. We attempt to
1366 * inline the nominal case and fall back to _vm_page_select_free()
1367 * otherwise. A busied page is removed from the queue and returned.
1369 * This routine may not block.
1371 static __inline vm_page_t
1372 vm_page_select_free(u_short pg_color, boolean_t prefer_zero)
1377 m = _vm_page_list_find(PQ_FREE, pg_color & PQ_L2_MASK,
1381 if (vm_page_busy_try(m, TRUE)) {
1383 * Various mechanisms such as a pmap_collect can
1384 * result in a busy page on the free queue. We
1385 * have to move the page out of the way so we can
1386 * retry the allocation. If the other thread is not
1387 * allocating the page then m->valid will remain 0 and
1388 * the pageout daemon will free the page later on.
1390 * Since we could not busy the page, however, we
1391 * cannot make assumptions as to whether the page
1392 * will be allocated by the other thread or not,
1393 * so all we can do is deactivate it to move it out
1394 * of the way. In particular, if the other thread
1395 * wires the page it may wind up on the inactive
1396 * queue and the pageout daemon will have to deal
1397 * with that case too.
1399 _vm_page_deactivate_locked(m, 0);
1400 vm_page_spin_unlock(m);
1403 * Theoretically if we are able to busy the page
1404 * atomic with the queue removal (using the vm_page
1405 * lock) nobody else should be able to mess with the
1408 KKASSERT((m->flags & (PG_UNMANAGED |
1409 PG_NEED_COMMIT)) == 0);
1410 KKASSERT(m->hold_count == 0);
1411 KKASSERT(m->wire_count == 0);
1412 vm_page_spin_unlock(m);
1413 pagedaemon_wakeup();
1415 /* return busied and removed page */
1423 * This implements a per-cpu cache of free, zero'd, ready-to-go pages.
1424 * The idea is to populate this cache prior to acquiring any locks so
1425 * we don't wind up potentially zeroing VM pages (under heavy loads) while
1426 * holding potentialy contending locks.
1428 * Note that we allocate the page uninserted into anything and use a pindex
1429 * of 0, the vm_page_alloc() will effectively add gd_cpuid so these
1430 * allocations should wind up being uncontended. However, we still want
1431 * to rove across PQ_L2_SIZE.
1434 vm_page_pcpu_cache(void)
1437 globaldata_t gd = mycpu;
1440 if (gd->gd_vmpg_count < GD_MINVMPG) {
1442 while (gd->gd_vmpg_count < GD_MAXVMPG) {
1443 m = vm_page_alloc(NULL, ticks & ~ncpus2_mask,
1444 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL |
1445 VM_ALLOC_NULL_OK | VM_ALLOC_ZERO);
1446 if (gd->gd_vmpg_count < GD_MAXVMPG) {
1447 if ((m->flags & PG_ZERO) == 0) {
1448 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1449 vm_page_flag_set(m, PG_ZERO);
1451 gd->gd_vmpg_array[gd->gd_vmpg_count++] = m;
1464 * Allocate and return a memory cell associated with this VM object/offset
1465 * pair. If object is NULL an unassociated page will be allocated.
1467 * The returned page will be busied and removed from its queues. This
1468 * routine can block and may return NULL if a race occurs and the page
1469 * is found to already exist at the specified (object, pindex).
1471 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain
1472 * VM_ALLOC_QUICK like normal but cannot use cache
1473 * VM_ALLOC_SYSTEM greater free drain
1474 * VM_ALLOC_INTERRUPT allow free list to be completely drained
1475 * VM_ALLOC_ZERO advisory request for pre-zero'd page only
1476 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only
1477 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision
1478 * (see vm_page_grab())
1479 * VM_ALLOC_USE_GD ok to use per-gd cache
1481 * The object must be held if not NULL
1482 * This routine may not block
1484 * Additional special handling is required when called from an interrupt
1485 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache
1489 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
1491 globaldata_t gd = mycpu;
1498 * Special per-cpu free VM page cache. The pages are pre-busied
1499 * and pre-zerod for us.
1501 if (gd->gd_vmpg_count && (page_req & VM_ALLOC_USE_GD)) {
1503 if (gd->gd_vmpg_count) {
1504 m = gd->gd_vmpg_array[--gd->gd_vmpg_count];
1514 * Cpu twist - cpu localization algorithm
1517 pg_color = gd->gd_cpuid + (pindex & ~ncpus_fit_mask) +
1518 (object->pg_color & ~ncpus_fit_mask);
1520 pg_color = gd->gd_cpuid + (pindex & ~ncpus_fit_mask);
1523 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK|
1524 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
1527 * Certain system threads (pageout daemon, buf_daemon's) are
1528 * allowed to eat deeper into the free page list.
1530 if (curthread->td_flags & TDF_SYSTHREAD)
1531 page_req |= VM_ALLOC_SYSTEM;
1534 if (vmstats.v_free_count > vmstats.v_free_reserved ||
1535 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) ||
1536 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 &&
1537 vmstats.v_free_count > vmstats.v_interrupt_free_min)
1540 * The free queue has sufficient free pages to take one out.
1542 if (page_req & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO))
1543 m = vm_page_select_free(pg_color, TRUE);
1545 m = vm_page_select_free(pg_color, FALSE);
1546 } else if (page_req & VM_ALLOC_NORMAL) {
1548 * Allocatable from the cache (non-interrupt only). On
1549 * success, we must free the page and try again, thus
1550 * ensuring that vmstats.v_*_free_min counters are replenished.
1553 if (curthread->td_preempted) {
1554 kprintf("vm_page_alloc(): warning, attempt to allocate"
1555 " cache page from preempting interrupt\n");
1558 m = vm_page_select_cache(pg_color);
1561 m = vm_page_select_cache(pg_color);
1564 * On success move the page into the free queue and loop.
1566 * Only do this if we can safely acquire the vm_object lock,
1567 * because this is effectively a random page and the caller
1568 * might be holding the lock shared, we don't want to
1572 KASSERT(m->dirty == 0,
1573 ("Found dirty cache page %p", m));
1574 if ((obj = m->object) != NULL) {
1575 if (vm_object_hold_try(obj)) {
1576 vm_page_protect(m, VM_PROT_NONE);
1578 /* m->object NULL here */
1579 vm_object_drop(obj);
1581 vm_page_deactivate(m);
1585 vm_page_protect(m, VM_PROT_NONE);
1592 * On failure return NULL
1594 #if defined(DIAGNOSTIC)
1595 if (vmstats.v_cache_count > 0)
1596 kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count);
1598 vm_pageout_deficit++;
1599 pagedaemon_wakeup();
1603 * No pages available, wakeup the pageout daemon and give up.
1605 vm_pageout_deficit++;
1606 pagedaemon_wakeup();
1611 * v_free_count can race so loop if we don't find the expected
1618 * Good page found. The page has already been busied for us and
1619 * removed from its queues.
1621 KASSERT(m->dirty == 0,
1622 ("vm_page_alloc: free/cache page %p was dirty", m));
1623 KKASSERT(m->queue == PQ_NONE);
1629 * Initialize the structure, inheriting some flags but clearing
1630 * all the rest. The page has already been busied for us.
1632 vm_page_flag_clear(m, ~(PG_ZERO | PG_BUSY | PG_SBUSY));
1633 KKASSERT(m->wire_count == 0);
1634 KKASSERT(m->busy == 0);
1639 * Caller must be holding the object lock (asserted by
1640 * vm_page_insert()).
1642 * NOTE: Inserting a page here does not insert it into any pmaps
1643 * (which could cause us to block allocating memory).
1645 * NOTE: If no object an unassociated page is allocated, m->pindex
1646 * can be used by the caller for any purpose.
1649 if (vm_page_insert(m, object, pindex) == FALSE) {
1651 if ((page_req & VM_ALLOC_NULL_OK) == 0)
1652 panic("PAGE RACE %p[%ld]/%p",
1653 object, (long)pindex, m);
1661 * Don't wakeup too often - wakeup the pageout daemon when
1662 * we would be nearly out of memory.
1664 pagedaemon_wakeup();
1667 * A PG_BUSY page is returned.
1673 * Attempt to allocate contiguous physical memory with the specified
1677 vm_page_alloc_contig(vm_paddr_t low, vm_paddr_t high,
1678 unsigned long alignment, unsigned long boundary,
1683 alignment >>= PAGE_SHIFT;
1686 boundary >>= PAGE_SHIFT;
1689 size = (size + PAGE_MASK) >> PAGE_SHIFT;
1691 spin_lock(&vm_contig_spin);
1692 blk = alist_alloc(&vm_contig_alist, 0, size);
1693 if (blk == ALIST_BLOCK_NONE) {
1694 spin_unlock(&vm_contig_spin);
1696 kprintf("vm_page_alloc_contig: %ldk nospace\n",
1697 (size + PAGE_MASK) * (PAGE_SIZE / 1024));
1701 if (high && ((vm_paddr_t)(blk + size) << PAGE_SHIFT) > high) {
1702 alist_free(&vm_contig_alist, blk, size);
1703 spin_unlock(&vm_contig_spin);
1705 kprintf("vm_page_alloc_contig: %ldk high "
1707 (size + PAGE_MASK) * (PAGE_SIZE / 1024),
1712 spin_unlock(&vm_contig_spin);
1713 if (vm_contig_verbose) {
1714 kprintf("vm_page_alloc_contig: %016jx/%ldk\n",
1715 (intmax_t)(vm_paddr_t)blk << PAGE_SHIFT,
1716 (size + PAGE_MASK) * (PAGE_SIZE / 1024));
1718 return (PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT));
1722 * Free contiguously allocated pages. The pages will be wired but not busy.
1723 * When freeing to the alist we leave them wired and not busy.
1726 vm_page_free_contig(vm_page_t m, unsigned long size)
1728 vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1729 vm_pindex_t start = pa >> PAGE_SHIFT;
1730 vm_pindex_t pages = (size + PAGE_MASK) >> PAGE_SHIFT;
1732 if (vm_contig_verbose) {
1733 kprintf("vm_page_free_contig: %016jx/%ldk\n",
1734 (intmax_t)pa, size / 1024);
1736 if (pa < vm_low_phys_reserved) {
1737 KKASSERT(pa + size <= vm_low_phys_reserved);
1738 spin_lock(&vm_contig_spin);
1739 alist_free(&vm_contig_alist, start, pages);
1740 spin_unlock(&vm_contig_spin);
1743 vm_page_busy_wait(m, FALSE, "cpgfr");
1744 vm_page_unwire(m, 0);
1755 * Wait for sufficient free memory for nominal heavy memory use kernel
1758 * WARNING! Be sure never to call this in any vm_pageout code path, which
1759 * will trivially deadlock the system.
1762 vm_wait_nominal(void)
1764 while (vm_page_count_min(0))
1769 * Test if vm_wait_nominal() would block.
1772 vm_test_nominal(void)
1774 if (vm_page_count_min(0))
1780 * Block until free pages are available for allocation, called in various
1781 * places before memory allocations.
1783 * The caller may loop if vm_page_count_min() == FALSE so we cannot be
1784 * more generous then that.
1790 * never wait forever
1794 lwkt_gettoken(&vm_token);
1796 if (curthread == pagethread) {
1798 * The pageout daemon itself needs pages, this is bad.
1800 if (vm_page_count_min(0)) {
1801 vm_pageout_pages_needed = 1;
1802 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo);
1806 * Wakeup the pageout daemon if necessary and wait.
1808 * Do not wait indefinitely for the target to be reached,
1809 * as load might prevent it from being reached any time soon.
1810 * But wait a little to try to slow down page allocations
1811 * and to give more important threads (the pagedaemon)
1812 * allocation priority.
1814 if (vm_page_count_target()) {
1815 if (vm_pages_needed == 0) {
1816 vm_pages_needed = 1;
1817 wakeup(&vm_pages_needed);
1819 ++vm_pages_waiting; /* SMP race ok */
1820 tsleep(&vmstats.v_free_count, 0, "vmwait", timo);
1823 lwkt_reltoken(&vm_token);
1827 * Block until free pages are available for allocation
1829 * Called only from vm_fault so that processes page faulting can be
1833 vm_wait_pfault(void)
1836 * Wakeup the pageout daemon if necessary and wait.
1838 * Do not wait indefinitely for the target to be reached,
1839 * as load might prevent it from being reached any time soon.
1840 * But wait a little to try to slow down page allocations
1841 * and to give more important threads (the pagedaemon)
1842 * allocation priority.
1844 if (vm_page_count_min(0)) {
1845 lwkt_gettoken(&vm_token);
1846 while (vm_page_count_severe()) {
1847 if (vm_page_count_target()) {
1848 if (vm_pages_needed == 0) {
1849 vm_pages_needed = 1;
1850 wakeup(&vm_pages_needed);
1852 ++vm_pages_waiting; /* SMP race ok */
1853 tsleep(&vmstats.v_free_count, 0, "pfault", hz);
1856 lwkt_reltoken(&vm_token);
1861 * Put the specified page on the active list (if appropriate). Ensure
1862 * that act_count is at least ACT_INIT but do not otherwise mess with it.
1864 * The caller should be holding the page busied ? XXX
1865 * This routine may not block.
1868 vm_page_activate(vm_page_t m)
1872 vm_page_spin_lock(m);
1873 if (m->queue - m->pc != PQ_ACTIVE) {
1874 _vm_page_queue_spin_lock(m);
1875 oqueue = _vm_page_rem_queue_spinlocked(m);
1876 /* page is left spinlocked, queue is unlocked */
1878 if (oqueue == PQ_CACHE)
1879 mycpu->gd_cnt.v_reactivated++;
1880 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1881 if (m->act_count < ACT_INIT)
1882 m->act_count = ACT_INIT;
1883 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE + m->pc, 0);
1885 _vm_page_and_queue_spin_unlock(m);
1886 if (oqueue == PQ_CACHE || oqueue == PQ_FREE)
1887 pagedaemon_wakeup();
1889 if (m->act_count < ACT_INIT)
1890 m->act_count = ACT_INIT;
1891 vm_page_spin_unlock(m);
1896 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
1897 * routine is called when a page has been added to the cache or free
1900 * This routine may not block.
1902 static __inline void
1903 vm_page_free_wakeup(void)
1906 * If the pageout daemon itself needs pages, then tell it that
1907 * there are some free.
1909 if (vm_pageout_pages_needed &&
1910 vmstats.v_cache_count + vmstats.v_free_count >=
1911 vmstats.v_pageout_free_min
1913 vm_pageout_pages_needed = 0;
1914 wakeup(&vm_pageout_pages_needed);
1918 * Wakeup processes that are waiting on memory.
1920 * Generally speaking we want to wakeup stuck processes as soon as
1921 * possible. !vm_page_count_min(0) is the absolute minimum point
1922 * where we can do this. Wait a bit longer to reduce degenerate
1923 * re-blocking (vm_page_free_hysteresis). The target check is just
1924 * to make sure the min-check w/hysteresis does not exceed the
1927 if (vm_pages_waiting) {
1928 if (!vm_page_count_min(vm_page_free_hysteresis) ||
1929 !vm_page_count_target()) {
1930 vm_pages_waiting = 0;
1931 wakeup(&vmstats.v_free_count);
1932 ++mycpu->gd_cnt.v_ppwakeups;
1935 if (!vm_page_count_target()) {
1937 * Plenty of pages are free, wakeup everyone.
1939 vm_pages_waiting = 0;
1940 wakeup(&vmstats.v_free_count);
1941 ++mycpu->gd_cnt.v_ppwakeups;
1942 } else if (!vm_page_count_min(0)) {
1944 * Some pages are free, wakeup someone.
1946 int wcount = vm_pages_waiting;
1949 vm_pages_waiting = wcount;
1950 wakeup_one(&vmstats.v_free_count);
1951 ++mycpu->gd_cnt.v_ppwakeups;
1958 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates
1959 * it from its VM object.
1961 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on
1962 * return (the page will have been freed).
1965 vm_page_free_toq(vm_page_t m)
1967 mycpu->gd_cnt.v_tfree++;
1968 KKASSERT((m->flags & PG_MAPPED) == 0);
1969 KKASSERT(m->flags & PG_BUSY);
1971 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
1972 kprintf("vm_page_free: pindex(%lu), busy(%d), "
1973 "PG_BUSY(%d), hold(%d)\n",
1974 (u_long)m->pindex, m->busy,
1975 ((m->flags & PG_BUSY) ? 1 : 0), m->hold_count);
1976 if ((m->queue - m->pc) == PQ_FREE)
1977 panic("vm_page_free: freeing free page");
1979 panic("vm_page_free: freeing busy page");
1983 * Remove from object, spinlock the page and its queues and
1984 * remove from any queue. No queue spinlock will be held
1985 * after this section (because the page was removed from any
1989 vm_page_and_queue_spin_lock(m);
1990 _vm_page_rem_queue_spinlocked(m);
1993 * No further management of fictitious pages occurs beyond object
1994 * and queue removal.
1996 if ((m->flags & PG_FICTITIOUS) != 0) {
1997 vm_page_spin_unlock(m);
2005 if (m->wire_count != 0) {
2006 if (m->wire_count > 1) {
2008 "vm_page_free: invalid wire count (%d), pindex: 0x%lx",
2009 m->wire_count, (long)m->pindex);
2011 panic("vm_page_free: freeing wired page");
2015 * Clear the UNMANAGED flag when freeing an unmanaged page.
2016 * Clear the NEED_COMMIT flag
2018 if (m->flags & PG_UNMANAGED)
2019 vm_page_flag_clear(m, PG_UNMANAGED);
2020 if (m->flags & PG_NEED_COMMIT)
2021 vm_page_flag_clear(m, PG_NEED_COMMIT);
2023 if (m->hold_count != 0) {
2024 vm_page_flag_clear(m, PG_ZERO);
2025 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
2027 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0);
2031 * This sequence allows us to clear PG_BUSY while still holding
2032 * its spin lock, which reduces contention vs allocators. We
2033 * must not leave the queue locked or _vm_page_wakeup() may
2036 _vm_page_queue_spin_unlock(m);
2037 if (_vm_page_wakeup(m)) {
2038 vm_page_spin_unlock(m);
2041 vm_page_spin_unlock(m);
2043 vm_page_free_wakeup();
2047 * vm_page_free_fromq_fast()
2049 * Remove a non-zero page from one of the free queues; the page is removed for
2050 * zeroing, so do not issue a wakeup.
2053 vm_page_free_fromq_fast(void)
2059 for (i = 0; i < PQ_L2_SIZE; ++i) {
2060 m = vm_page_list_find(PQ_FREE, qi, FALSE);
2061 /* page is returned spinlocked and removed from its queue */
2063 if (vm_page_busy_try(m, TRUE)) {
2065 * We were unable to busy the page, deactivate
2068 _vm_page_deactivate_locked(m, 0);
2069 vm_page_spin_unlock(m);
2070 } else if (m->flags & PG_ZERO) {
2072 * The page is PG_ZERO, requeue it and loop
2074 _vm_page_add_queue_spinlocked(m,
2077 vm_page_queue_spin_unlock(m);
2078 if (_vm_page_wakeup(m)) {
2079 vm_page_spin_unlock(m);
2082 vm_page_spin_unlock(m);
2086 * The page is not PG_ZERO'd so return it.
2088 vm_page_spin_unlock(m);
2089 KKASSERT((m->flags & (PG_UNMANAGED |
2090 PG_NEED_COMMIT)) == 0);
2091 KKASSERT(m->hold_count == 0);
2092 KKASSERT(m->wire_count == 0);
2097 qi = (qi + PQ_PRIME2) & PQ_L2_MASK;
2103 * vm_page_unmanage()
2105 * Prevent PV management from being done on the page. The page is
2106 * removed from the paging queues as if it were wired, and as a
2107 * consequence of no longer being managed the pageout daemon will not
2108 * touch it (since there is no way to locate the pte mappings for the
2109 * page). madvise() calls that mess with the pmap will also no longer
2110 * operate on the page.
2112 * Beyond that the page is still reasonably 'normal'. Freeing the page
2113 * will clear the flag.
2115 * This routine is used by OBJT_PHYS objects - objects using unswappable
2116 * physical memory as backing store rather then swap-backed memory and
2117 * will eventually be extended to support 4MB unmanaged physical
2120 * Caller must be holding the page busy.
2123 vm_page_unmanage(vm_page_t m)
2125 KKASSERT(m->flags & PG_BUSY);
2126 if ((m->flags & PG_UNMANAGED) == 0) {
2127 if (m->wire_count == 0)
2130 vm_page_flag_set(m, PG_UNMANAGED);
2134 * Mark this page as wired down by yet another map, removing it from
2135 * paging queues as necessary.
2137 * Caller must be holding the page busy.
2140 vm_page_wire(vm_page_t m)
2143 * Only bump the wire statistics if the page is not already wired,
2144 * and only unqueue the page if it is on some queue (if it is unmanaged
2145 * it is already off the queues). Don't do anything with fictitious
2146 * pages because they are always wired.
2148 KKASSERT(m->flags & PG_BUSY);
2149 if ((m->flags & PG_FICTITIOUS) == 0) {
2150 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) {
2151 if ((m->flags & PG_UNMANAGED) == 0)
2153 atomic_add_int(&vmstats.v_wire_count, 1);
2155 KASSERT(m->wire_count != 0,
2156 ("vm_page_wire: wire_count overflow m=%p", m));
2161 * Release one wiring of this page, potentially enabling it to be paged again.
2163 * Many pages placed on the inactive queue should actually go
2164 * into the cache, but it is difficult to figure out which. What
2165 * we do instead, if the inactive target is well met, is to put
2166 * clean pages at the head of the inactive queue instead of the tail.
2167 * This will cause them to be moved to the cache more quickly and
2168 * if not actively re-referenced, freed more quickly. If we just
2169 * stick these pages at the end of the inactive queue, heavy filesystem
2170 * meta-data accesses can cause an unnecessary paging load on memory bound
2171 * processes. This optimization causes one-time-use metadata to be
2172 * reused more quickly.
2174 * Pages marked PG_NEED_COMMIT are always activated and never placed on
2175 * the inactive queue. This helps the pageout daemon determine memory
2176 * pressure and act on out-of-memory situations more quickly.
2178 * BUT, if we are in a low-memory situation we have no choice but to
2179 * put clean pages on the cache queue.
2181 * A number of routines use vm_page_unwire() to guarantee that the page
2182 * will go into either the inactive or active queues, and will NEVER
2183 * be placed in the cache - for example, just after dirtying a page.
2184 * dirty pages in the cache are not allowed.
2186 * The page queues must be locked.
2187 * This routine may not block.
2190 vm_page_unwire(vm_page_t m, int activate)
2192 KKASSERT(m->flags & PG_BUSY);
2193 if (m->flags & PG_FICTITIOUS) {
2195 } else if (m->wire_count <= 0) {
2196 panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
2198 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) {
2199 atomic_add_int(&vmstats.v_wire_count, -1);
2200 if (m->flags & PG_UNMANAGED) {
2202 } else if (activate || (m->flags & PG_NEED_COMMIT)) {
2203 vm_page_spin_lock(m);
2204 _vm_page_add_queue_spinlocked(m,
2205 PQ_ACTIVE + m->pc, 0);
2206 _vm_page_and_queue_spin_unlock(m);
2208 vm_page_spin_lock(m);
2209 vm_page_flag_clear(m, PG_WINATCFLS);
2210 _vm_page_add_queue_spinlocked(m,
2211 PQ_INACTIVE + m->pc, 0);
2212 ++vm_swapcache_inactive_heuristic;
2213 _vm_page_and_queue_spin_unlock(m);
2220 * Move the specified page to the inactive queue. If the page has
2221 * any associated swap, the swap is deallocated.
2223 * Normally athead is 0 resulting in LRU operation. athead is set
2224 * to 1 if we want this page to be 'as if it were placed in the cache',
2225 * except without unmapping it from the process address space.
2227 * vm_page's spinlock must be held on entry and will remain held on return.
2228 * This routine may not block.
2231 _vm_page_deactivate_locked(vm_page_t m, int athead)
2236 * Ignore if already inactive.
2238 if (m->queue - m->pc == PQ_INACTIVE)
2240 _vm_page_queue_spin_lock(m);
2241 oqueue = _vm_page_rem_queue_spinlocked(m);
2243 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
2244 if (oqueue == PQ_CACHE)
2245 mycpu->gd_cnt.v_reactivated++;
2246 vm_page_flag_clear(m, PG_WINATCFLS);
2247 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead);
2249 ++vm_swapcache_inactive_heuristic;
2251 _vm_page_queue_spin_unlock(m);
2252 /* leaves vm_page spinlocked */
2256 * Attempt to deactivate a page.
2261 vm_page_deactivate(vm_page_t m)
2263 vm_page_spin_lock(m);
2264 _vm_page_deactivate_locked(m, 0);
2265 vm_page_spin_unlock(m);
2269 vm_page_deactivate_locked(vm_page_t m)
2271 _vm_page_deactivate_locked(m, 0);
2275 * Attempt to move a page to PQ_CACHE.
2277 * Returns 0 on failure, 1 on success
2279 * The page should NOT be busied by the caller. This function will validate
2280 * whether the page can be safely moved to the cache.
2283 vm_page_try_to_cache(vm_page_t m)
2285 vm_page_spin_lock(m);
2286 if (vm_page_busy_try(m, TRUE)) {
2287 vm_page_spin_unlock(m);
2290 if (m->dirty || m->hold_count || m->wire_count ||
2291 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT))) {
2292 if (_vm_page_wakeup(m)) {
2293 vm_page_spin_unlock(m);
2296 vm_page_spin_unlock(m);
2300 vm_page_spin_unlock(m);
2303 * Page busied by us and no longer spinlocked. Dirty pages cannot
2304 * be moved to the cache.
2306 vm_page_test_dirty(m);
2307 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2316 * Attempt to free the page. If we cannot free it, we do nothing.
2317 * 1 is returned on success, 0 on failure.
2322 vm_page_try_to_free(vm_page_t m)
2324 vm_page_spin_lock(m);
2325 if (vm_page_busy_try(m, TRUE)) {
2326 vm_page_spin_unlock(m);
2331 * The page can be in any state, including already being on the free
2332 * queue. Check to see if it really can be freed.
2334 if (m->dirty || /* can't free if it is dirty */
2335 m->hold_count || /* or held (XXX may be wrong) */
2336 m->wire_count || /* or wired */
2337 (m->flags & (PG_UNMANAGED | /* or unmanaged */
2338 PG_NEED_COMMIT)) || /* or needs a commit */
2339 m->queue - m->pc == PQ_FREE || /* already on PQ_FREE */
2340 m->queue - m->pc == PQ_HOLD) { /* already on PQ_HOLD */
2341 if (_vm_page_wakeup(m)) {
2342 vm_page_spin_unlock(m);
2345 vm_page_spin_unlock(m);
2349 vm_page_spin_unlock(m);
2352 * We can probably free the page.
2354 * Page busied by us and no longer spinlocked. Dirty pages will
2355 * not be freed by this function. We have to re-test the
2356 * dirty bit after cleaning out the pmaps.
2358 vm_page_test_dirty(m);
2359 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2363 vm_page_protect(m, VM_PROT_NONE);
2364 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2375 * Put the specified page onto the page cache queue (if appropriate).
2377 * The page must be busy, and this routine will release the busy and
2378 * possibly even free the page.
2381 vm_page_cache(vm_page_t m)
2383 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) ||
2384 m->busy || m->wire_count || m->hold_count) {
2385 kprintf("vm_page_cache: attempting to cache busy/held page\n");
2391 * Already in the cache (and thus not mapped)
2393 if ((m->queue - m->pc) == PQ_CACHE) {
2394 KKASSERT((m->flags & PG_MAPPED) == 0);
2400 * Caller is required to test m->dirty, but note that the act of
2401 * removing the page from its maps can cause it to become dirty
2402 * on an SMP system due to another cpu running in usermode.
2405 panic("vm_page_cache: caching a dirty page, pindex: %ld",
2410 * Remove all pmaps and indicate that the page is not
2411 * writeable or mapped. Our vm_page_protect() call may
2412 * have blocked (especially w/ VM_PROT_NONE), so recheck
2415 vm_page_protect(m, VM_PROT_NONE);
2416 if ((m->flags & (PG_UNMANAGED | PG_MAPPED)) ||
2417 m->busy || m->wire_count || m->hold_count) {
2419 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2420 vm_page_deactivate(m);
2423 _vm_page_and_queue_spin_lock(m);
2424 _vm_page_rem_queue_spinlocked(m);
2425 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0);
2426 _vm_page_queue_spin_unlock(m);
2427 if (_vm_page_wakeup(m)) {
2428 vm_page_spin_unlock(m);
2431 vm_page_spin_unlock(m);
2433 vm_page_free_wakeup();
2438 * vm_page_dontneed()
2440 * Cache, deactivate, or do nothing as appropriate. This routine
2441 * is typically used by madvise() MADV_DONTNEED.
2443 * Generally speaking we want to move the page into the cache so
2444 * it gets reused quickly. However, this can result in a silly syndrome
2445 * due to the page recycling too quickly. Small objects will not be
2446 * fully cached. On the otherhand, if we move the page to the inactive
2447 * queue we wind up with a problem whereby very large objects
2448 * unnecessarily blow away our inactive and cache queues.
2450 * The solution is to move the pages based on a fixed weighting. We
2451 * either leave them alone, deactivate them, or move them to the cache,
2452 * where moving them to the cache has the highest weighting.
2453 * By forcing some pages into other queues we eventually force the
2454 * system to balance the queues, potentially recovering other unrelated
2455 * space from active. The idea is to not force this to happen too
2458 * The page must be busied.
2461 vm_page_dontneed(vm_page_t m)
2463 static int dnweight;
2470 * occassionally leave the page alone
2472 if ((dnw & 0x01F0) == 0 ||
2473 m->queue - m->pc == PQ_INACTIVE ||
2474 m->queue - m->pc == PQ_CACHE
2476 if (m->act_count >= ACT_INIT)
2482 * If vm_page_dontneed() is inactivating a page, it must clear
2483 * the referenced flag; otherwise the pagedaemon will see references
2484 * on the page in the inactive queue and reactivate it. Until the
2485 * page can move to the cache queue, madvise's job is not done.
2487 vm_page_flag_clear(m, PG_REFERENCED);
2488 pmap_clear_reference(m);
2491 vm_page_test_dirty(m);
2493 if (m->dirty || (dnw & 0x0070) == 0) {
2495 * Deactivate the page 3 times out of 32.
2500 * Cache the page 28 times out of every 32. Note that
2501 * the page is deactivated instead of cached, but placed
2502 * at the head of the queue instead of the tail.
2506 vm_page_spin_lock(m);
2507 _vm_page_deactivate_locked(m, head);
2508 vm_page_spin_unlock(m);
2512 * These routines manipulate the 'soft busy' count for a page. A soft busy
2513 * is almost like PG_BUSY except that it allows certain compatible operations
2514 * to occur on the page while it is busy. For example, a page undergoing a
2515 * write can still be mapped read-only.
2517 * Because vm_pages can overlap buffers m->busy can be > 1. m->busy is only
2518 * adjusted while the vm_page is PG_BUSY so the flash will occur when the
2519 * busy bit is cleared.
2522 vm_page_io_start(vm_page_t m)
2524 KASSERT(m->flags & PG_BUSY, ("vm_page_io_start: page not busy!!!"));
2525 atomic_add_char(&m->busy, 1);
2526 vm_page_flag_set(m, PG_SBUSY);
2530 vm_page_io_finish(vm_page_t m)
2532 KASSERT(m->flags & PG_BUSY, ("vm_page_io_finish: page not busy!!!"));
2533 atomic_subtract_char(&m->busy, 1);
2535 vm_page_flag_clear(m, PG_SBUSY);
2539 * Indicate that a clean VM page requires a filesystem commit and cannot
2540 * be reused. Used by tmpfs.
2543 vm_page_need_commit(vm_page_t m)
2545 vm_page_flag_set(m, PG_NEED_COMMIT);
2546 vm_object_set_writeable_dirty(m->object);
2550 vm_page_clear_commit(vm_page_t m)
2552 vm_page_flag_clear(m, PG_NEED_COMMIT);
2556 * Grab a page, blocking if it is busy and allocating a page if necessary.
2557 * A busy page is returned or NULL. The page may or may not be valid and
2558 * might not be on a queue (the caller is responsible for the disposition of
2561 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the
2562 * page will be zero'd and marked valid.
2564 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked
2565 * valid even if it already exists.
2567 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also
2568 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified.
2569 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified.
2571 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
2572 * always returned if we had blocked.
2574 * This routine may not be called from an interrupt.
2576 * PG_ZERO is *ALWAYS* cleared by this routine.
2578 * No other requirements.
2581 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2586 KKASSERT(allocflags &
2587 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
2588 vm_object_hold(object);
2590 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error);
2592 vm_page_sleep_busy(m, TRUE, "pgrbwt");
2593 if ((allocflags & VM_ALLOC_RETRY) == 0) {
2598 } else if (m == NULL) {
2599 if (allocflags & VM_ALLOC_RETRY)
2600 allocflags |= VM_ALLOC_NULL_OK;
2601 m = vm_page_alloc(object, pindex,
2602 allocflags & ~VM_ALLOC_RETRY);
2606 if ((allocflags & VM_ALLOC_RETRY) == 0)
2615 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid.
2617 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set
2618 * valid even if already valid.
2620 if (m->valid == 0) {
2621 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) {
2622 if ((m->flags & PG_ZERO) == 0)
2623 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2624 m->valid = VM_PAGE_BITS_ALL;
2626 } else if (allocflags & VM_ALLOC_FORCE_ZERO) {
2627 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2628 m->valid = VM_PAGE_BITS_ALL;
2630 vm_page_flag_clear(m, PG_ZERO);
2632 vm_object_drop(object);
2637 * Mapping function for valid bits or for dirty bits in
2638 * a page. May not block.
2640 * Inputs are required to range within a page.
2646 vm_page_bits(int base, int size)
2652 base + size <= PAGE_SIZE,
2653 ("vm_page_bits: illegal base/size %d/%d", base, size)
2656 if (size == 0) /* handle degenerate case */
2659 first_bit = base >> DEV_BSHIFT;
2660 last_bit = (base + size - 1) >> DEV_BSHIFT;
2662 return ((2 << last_bit) - (1 << first_bit));
2666 * Sets portions of a page valid and clean. The arguments are expected
2667 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2668 * of any partial chunks touched by the range. The invalid portion of
2669 * such chunks will be zero'd.
2671 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically
2672 * align base to DEV_BSIZE so as not to mark clean a partially
2673 * truncated device block. Otherwise the dirty page status might be
2676 * This routine may not block.
2678 * (base + size) must be less then or equal to PAGE_SIZE.
2681 _vm_page_zero_valid(vm_page_t m, int base, int size)
2686 if (size == 0) /* handle degenerate case */
2690 * If the base is not DEV_BSIZE aligned and the valid
2691 * bit is clear, we have to zero out a portion of the
2695 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2696 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
2698 pmap_zero_page_area(
2706 * If the ending offset is not DEV_BSIZE aligned and the
2707 * valid bit is clear, we have to zero out a portion of
2711 endoff = base + size;
2713 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2714 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
2716 pmap_zero_page_area(
2719 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
2725 * Set valid, clear dirty bits. If validating the entire
2726 * page we can safely clear the pmap modify bit. We also
2727 * use this opportunity to clear the PG_NOSYNC flag. If a process
2728 * takes a write fault on a MAP_NOSYNC memory area the flag will
2731 * We set valid bits inclusive of any overlap, but we can only
2732 * clear dirty bits for DEV_BSIZE chunks that are fully within
2735 * Page must be busied?
2736 * No other requirements.
2739 vm_page_set_valid(vm_page_t m, int base, int size)
2741 _vm_page_zero_valid(m, base, size);
2742 m->valid |= vm_page_bits(base, size);
2747 * Set valid bits and clear dirty bits.
2749 * NOTE: This function does not clear the pmap modified bit.
2750 * Also note that e.g. NFS may use a byte-granular base
2753 * WARNING: Page must be busied? But vfs_clean_one_page() will call
2754 * this without necessarily busying the page (via bdwrite()).
2755 * So for now vm_token must also be held.
2757 * No other requirements.
2760 vm_page_set_validclean(vm_page_t m, int base, int size)
2764 _vm_page_zero_valid(m, base, size);
2765 pagebits = vm_page_bits(base, size);
2766 m->valid |= pagebits;
2767 m->dirty &= ~pagebits;
2768 if (base == 0 && size == PAGE_SIZE) {
2769 /*pmap_clear_modify(m);*/
2770 vm_page_flag_clear(m, PG_NOSYNC);
2775 * Set valid & dirty. Used by buwrite()
2777 * WARNING: Page must be busied? But vfs_dirty_one_page() will
2778 * call this function in buwrite() so for now vm_token must
2781 * No other requirements.
2784 vm_page_set_validdirty(vm_page_t m, int base, int size)
2788 pagebits = vm_page_bits(base, size);
2789 m->valid |= pagebits;
2790 m->dirty |= pagebits;
2792 vm_object_set_writeable_dirty(m->object);
2798 * NOTE: This function does not clear the pmap modified bit.
2799 * Also note that e.g. NFS may use a byte-granular base
2802 * Page must be busied?
2803 * No other requirements.
2806 vm_page_clear_dirty(vm_page_t m, int base, int size)
2808 m->dirty &= ~vm_page_bits(base, size);
2809 if (base == 0 && size == PAGE_SIZE) {
2810 /*pmap_clear_modify(m);*/
2811 vm_page_flag_clear(m, PG_NOSYNC);
2816 * Make the page all-dirty.
2818 * Also make sure the related object and vnode reflect the fact that the
2819 * object may now contain a dirty page.
2821 * Page must be busied?
2822 * No other requirements.
2825 vm_page_dirty(vm_page_t m)
2828 int pqtype = m->queue - m->pc;
2830 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE,
2831 ("vm_page_dirty: page in free/cache queue!"));
2832 if (m->dirty != VM_PAGE_BITS_ALL) {
2833 m->dirty = VM_PAGE_BITS_ALL;
2835 vm_object_set_writeable_dirty(m->object);
2840 * Invalidates DEV_BSIZE'd chunks within a page. Both the
2841 * valid and dirty bits for the effected areas are cleared.
2843 * Page must be busied?
2845 * No other requirements.
2848 vm_page_set_invalid(vm_page_t m, int base, int size)
2852 bits = vm_page_bits(base, size);
2855 m->object->generation++;
2859 * The kernel assumes that the invalid portions of a page contain
2860 * garbage, but such pages can be mapped into memory by user code.
2861 * When this occurs, we must zero out the non-valid portions of the
2862 * page so user code sees what it expects.
2864 * Pages are most often semi-valid when the end of a file is mapped
2865 * into memory and the file's size is not page aligned.
2867 * Page must be busied?
2868 * No other requirements.
2871 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
2877 * Scan the valid bits looking for invalid sections that
2878 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
2879 * valid bit may be set ) have already been zerod by
2880 * vm_page_set_validclean().
2882 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
2883 if (i == (PAGE_SIZE / DEV_BSIZE) ||
2884 (m->valid & (1 << i))
2887 pmap_zero_page_area(
2890 (i - b) << DEV_BSHIFT
2898 * setvalid is TRUE when we can safely set the zero'd areas
2899 * as being valid. We can do this if there are no cache consistency
2900 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
2903 m->valid = VM_PAGE_BITS_ALL;
2907 * Is a (partial) page valid? Note that the case where size == 0
2908 * will return FALSE in the degenerate case where the page is entirely
2909 * invalid, and TRUE otherwise.
2912 * No other requirements.
2915 vm_page_is_valid(vm_page_t m, int base, int size)
2917 int bits = vm_page_bits(base, size);
2919 if (m->valid && ((m->valid & bits) == bits))
2926 * update dirty bits from pmap/mmu. May not block.
2928 * Caller must hold the page busy
2931 vm_page_test_dirty(vm_page_t m)
2933 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
2939 * Register an action, associating it with its vm_page
2942 vm_page_register_action(vm_page_action_t action, vm_page_event_t event)
2944 struct vm_page_action_list *list;
2947 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
2948 list = &action_list[hv];
2950 lwkt_gettoken(&vm_token);
2951 vm_page_flag_set(action->m, PG_ACTIONLIST);
2952 action->event = event;
2953 LIST_INSERT_HEAD(list, action, entry);
2954 lwkt_reltoken(&vm_token);
2958 * Unregister an action, disassociating it from its related vm_page
2961 vm_page_unregister_action(vm_page_action_t action)
2963 struct vm_page_action_list *list;
2966 lwkt_gettoken(&vm_token);
2967 if (action->event != VMEVENT_NONE) {
2968 action->event = VMEVENT_NONE;
2969 LIST_REMOVE(action, entry);
2971 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
2972 list = &action_list[hv];
2973 if (LIST_EMPTY(list))
2974 vm_page_flag_clear(action->m, PG_ACTIONLIST);
2976 lwkt_reltoken(&vm_token);
2980 * Issue an event on a VM page. Corresponding action structures are
2981 * removed from the page's list and called.
2983 * If the vm_page has no more pending action events we clear its
2984 * PG_ACTIONLIST flag.
2987 vm_page_event_internal(vm_page_t m, vm_page_event_t event)
2989 struct vm_page_action_list *list;
2990 struct vm_page_action *scan;
2991 struct vm_page_action *next;
2995 hv = (int)((intptr_t)m >> 8) & VMACTION_HMASK;
2996 list = &action_list[hv];
2999 lwkt_gettoken(&vm_token);
3000 LIST_FOREACH_MUTABLE(scan, list, entry, next) {
3002 if (scan->event == event) {
3003 scan->event = VMEVENT_NONE;
3004 LIST_REMOVE(scan, entry);
3005 scan->func(m, scan);
3013 vm_page_flag_clear(m, PG_ACTIONLIST);
3014 lwkt_reltoken(&vm_token);
3017 #include "opt_ddb.h"
3019 #include <sys/kernel.h>
3021 #include <ddb/ddb.h>
3023 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3025 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count);
3026 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count);
3027 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count);
3028 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count);
3029 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count);
3030 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved);
3031 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min);
3032 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target);
3033 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min);
3034 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target);
3037 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3040 db_printf("PQ_FREE:");
3041 for(i=0;i<PQ_L2_SIZE;i++) {
3042 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
3046 db_printf("PQ_CACHE:");
3047 for(i=0;i<PQ_L2_SIZE;i++) {
3048 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
3052 db_printf("PQ_ACTIVE:");
3053 for(i=0;i<PQ_L2_SIZE;i++) {
3054 db_printf(" %d", vm_page_queues[PQ_ACTIVE + i].lcnt);
3058 db_printf("PQ_INACTIVE:");
3059 for(i=0;i<PQ_L2_SIZE;i++) {
3060 db_printf(" %d", vm_page_queues[PQ_INACTIVE + i].lcnt);