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 * 4. 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>
95 #include <vm/vm_page2.h>
96 #include <sys/spinlock2.h>
98 #define VMACTION_HSIZE 256
99 #define VMACTION_HMASK (VMACTION_HSIZE - 1)
101 static void vm_page_queue_init(void);
102 static void vm_page_free_wakeup(void);
103 static vm_page_t vm_page_select_cache(u_short pg_color);
104 static vm_page_t _vm_page_list_find2(int basequeue, int index);
105 static void _vm_page_deactivate_locked(vm_page_t m, int athead);
108 * Array of tailq lists
110 __cachealign struct vpgqueues vm_page_queues[PQ_COUNT];
112 LIST_HEAD(vm_page_action_list, vm_page_action);
113 struct vm_page_action_list action_list[VMACTION_HSIZE];
114 static volatile int vm_pages_waiting;
116 static struct alist vm_contig_alist;
117 static struct almeta vm_contig_ameta[ALIST_RECORDS_65536];
118 static struct spinlock vm_contig_spin = SPINLOCK_INITIALIZER(&vm_contig_spin);
120 static u_long vm_dma_reserved = 0;
121 TUNABLE_ULONG("vm.dma_reserved", &vm_dma_reserved);
122 SYSCTL_ULONG(_vm, OID_AUTO, dma_reserved, CTLFLAG_RD, &vm_dma_reserved, 0,
123 "Memory reserved for DMA");
124 SYSCTL_UINT(_vm, OID_AUTO, dma_free_pages, CTLFLAG_RD,
125 &vm_contig_alist.bl_free, 0, "Memory reserved for DMA");
127 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare,
128 vm_pindex_t, pindex);
131 vm_page_queue_init(void)
135 for (i = 0; i < PQ_L2_SIZE; i++)
136 vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count;
137 for (i = 0; i < PQ_L2_SIZE; i++)
138 vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count;
139 for (i = 0; i < PQ_L2_SIZE; i++)
140 vm_page_queues[PQ_INACTIVE+i].cnt = &vmstats.v_inactive_count;
141 for (i = 0; i < PQ_L2_SIZE; i++)
142 vm_page_queues[PQ_ACTIVE+i].cnt = &vmstats.v_active_count;
143 for (i = 0; i < PQ_L2_SIZE; i++)
144 vm_page_queues[PQ_HOLD+i].cnt = &vmstats.v_active_count;
145 /* PQ_NONE has no queue */
147 for (i = 0; i < PQ_COUNT; i++) {
148 TAILQ_INIT(&vm_page_queues[i].pl);
149 spin_init(&vm_page_queues[i].spin);
152 for (i = 0; i < VMACTION_HSIZE; i++)
153 LIST_INIT(&action_list[i]);
157 * note: place in initialized data section? Is this necessary?
160 int vm_page_array_size = 0;
161 int vm_page_zero_count = 0;
162 vm_page_t vm_page_array = NULL;
163 vm_paddr_t vm_low_phys_reserved;
168 * Sets the page size, perhaps based upon the memory size.
169 * Must be called before any use of page-size dependent functions.
172 vm_set_page_size(void)
174 if (vmstats.v_page_size == 0)
175 vmstats.v_page_size = PAGE_SIZE;
176 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0)
177 panic("vm_set_page_size: page size not a power of two");
183 * Add a new page to the freelist for use by the system. New pages
184 * are added to both the head and tail of the associated free page
185 * queue in a bottom-up fashion, so both zero'd and non-zero'd page
186 * requests pull 'recent' adds (higher physical addresses) first.
188 * Beware that the page zeroing daemon will also be running soon after
189 * boot, moving pages from the head to the tail of the PQ_FREE queues.
191 * Must be called in a critical section.
194 vm_add_new_page(vm_paddr_t pa)
196 struct vpgqueues *vpq;
199 m = PHYS_TO_VM_PAGE(pa);
202 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK;
205 * Twist for cpu localization in addition to page coloring, so
206 * different cpus selecting by m->queue get different page colors.
208 m->pc ^= ((pa >> PAGE_SHIFT) / PQ_L2_SIZE) & PQ_L2_MASK;
209 m->pc ^= ((pa >> PAGE_SHIFT) / (PQ_L2_SIZE * PQ_L2_SIZE)) & PQ_L2_MASK;
212 * Reserve a certain number of contiguous low memory pages for
213 * contigmalloc() to use.
215 if (pa < vm_low_phys_reserved) {
216 atomic_add_int(&vmstats.v_page_count, 1);
217 atomic_add_int(&vmstats.v_dma_pages, 1);
220 alist_free(&vm_contig_alist, pa >> PAGE_SHIFT, 1);
227 m->queue = m->pc + PQ_FREE;
228 KKASSERT(m->dirty == 0);
230 atomic_add_int(&vmstats.v_page_count, 1);
231 atomic_add_int(&vmstats.v_free_count, 1);
232 vpq = &vm_page_queues[m->queue];
233 if ((vpq->flipflop & 15) == 0) {
234 pmap_zero_page(VM_PAGE_TO_PHYS(m));
236 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
237 atomic_add_int(&vm_page_zero_count, 1);
239 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
248 * Initializes the resident memory module.
250 * Preallocates memory for critical VM structures and arrays prior to
251 * kernel_map becoming available.
253 * Memory is allocated from (virtual2_start, virtual2_end) if available,
254 * otherwise memory is allocated from (virtual_start, virtual_end).
256 * On x86-64 (virtual_start, virtual_end) is only 2GB and may not be
257 * large enough to hold vm_page_array & other structures for machines with
258 * large amounts of ram, so we want to use virtual2* when available.
261 vm_page_startup(void)
263 vm_offset_t vaddr = virtual2_start ? virtual2_start : virtual_start;
266 vm_paddr_t page_range;
273 vm_paddr_t biggestone, biggestsize;
280 vaddr = round_page(vaddr);
282 for (i = 0; phys_avail[i + 1]; i += 2) {
283 phys_avail[i] = round_page64(phys_avail[i]);
284 phys_avail[i + 1] = trunc_page64(phys_avail[i + 1]);
287 for (i = 0; phys_avail[i + 1]; i += 2) {
288 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
290 if (size > biggestsize) {
298 end = phys_avail[biggestone+1];
299 end = trunc_page(end);
302 * Initialize the queue headers for the free queue, the active queue
303 * and the inactive queue.
305 vm_page_queue_init();
307 #if !defined(_KERNEL_VIRTUAL)
309 * VKERNELs don't support minidumps and as such don't need
312 * Allocate a bitmap to indicate that a random physical page
313 * needs to be included in a minidump.
315 * The amd64 port needs this to indicate which direct map pages
316 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
318 * However, i386 still needs this workspace internally within the
319 * minidump code. In theory, they are not needed on i386, but are
320 * included should the sf_buf code decide to use them.
322 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE;
323 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
324 end -= vm_page_dump_size;
325 vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size,
326 VM_PROT_READ | VM_PROT_WRITE);
327 bzero((void *)vm_page_dump, vm_page_dump_size);
330 * Compute the number of pages of memory that will be available for
331 * use (taking into account the overhead of a page structure per
334 first_page = phys_avail[0] / PAGE_SIZE;
335 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
336 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE;
338 #ifndef _KERNEL_VIRTUAL
340 * (only applies to real kernels)
342 * Initialize the contiguous reserve map. We initially reserve up
343 * to 1/4 available physical memory or 65536 pages (~256MB), whichever
346 * Once device initialization is complete we return most of the
347 * reserved memory back to the normal page queues but leave some
348 * in reserve for things like usb attachments.
350 vm_low_phys_reserved = (vm_paddr_t)65536 << PAGE_SHIFT;
351 if (vm_low_phys_reserved > total / 4)
352 vm_low_phys_reserved = total / 4;
353 if (vm_dma_reserved == 0) {
354 vm_dma_reserved = 16 * 1024 * 1024; /* 16MB */
355 if (vm_dma_reserved > total / 16)
356 vm_dma_reserved = total / 16;
359 alist_init(&vm_contig_alist, 65536, vm_contig_ameta,
360 ALIST_RECORDS_65536);
363 * Initialize the mem entry structures now, and put them in the free
366 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
367 mapped = pmap_map(&vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE);
368 vm_page_array = (vm_page_t)mapped;
370 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL)
372 * since pmap_map on amd64 returns stuff out of a direct-map region,
373 * we have to manually add these pages to the minidump tracking so
374 * that they can be dumped, including the vm_page_array.
376 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
381 * Clear all of the page structures
383 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
384 vm_page_array_size = page_range;
387 * Construct the free queue(s) in ascending order (by physical
388 * address) so that the first 16MB of physical memory is allocated
389 * last rather than first. On large-memory machines, this avoids
390 * the exhaustion of low physical memory before isa_dmainit has run.
392 vmstats.v_page_count = 0;
393 vmstats.v_free_count = 0;
394 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
399 last_pa = phys_avail[i + 1];
400 while (pa < last_pa && npages-- > 0) {
406 virtual2_start = vaddr;
408 virtual_start = vaddr;
412 * We tended to reserve a ton of memory for contigmalloc(). Now that most
413 * drivers have initialized we want to return most the remaining free
414 * reserve back to the VM page queues so they can be used for normal
417 * We leave vm_dma_reserved bytes worth of free pages in the reserve pool.
420 vm_page_startup_finish(void *dummy __unused)
429 spin_lock(&vm_contig_spin);
431 bfree = alist_free_info(&vm_contig_alist, &blk, &count);
432 if (bfree <= vm_dma_reserved / PAGE_SIZE)
438 * Figure out how much of the initial reserve we have to
439 * free in order to reach our target.
441 bfree -= vm_dma_reserved / PAGE_SIZE;
443 blk += count - bfree;
448 * Calculate the nearest power of 2 <= count.
450 for (xcount = 1; xcount <= count; xcount <<= 1)
453 blk += count - xcount;
457 * Allocate the pages from the alist, then free them to
458 * the normal VM page queues.
460 * Pages allocated from the alist are wired. We have to
461 * busy, unwire, and free them. We must also adjust
462 * vm_low_phys_reserved before freeing any pages to prevent
465 rblk = alist_alloc(&vm_contig_alist, blk, count);
467 kprintf("vm_page_startup_finish: Unable to return "
468 "dma space @0x%08x/%d -> 0x%08x\n",
472 atomic_add_int(&vmstats.v_dma_pages, -count);
473 spin_unlock(&vm_contig_spin);
475 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT);
476 vm_low_phys_reserved = VM_PAGE_TO_PHYS(m);
478 vm_page_busy_wait(m, FALSE, "cpgfr");
479 vm_page_unwire(m, 0);
484 spin_lock(&vm_contig_spin);
486 spin_unlock(&vm_contig_spin);
489 * Print out how much DMA space drivers have already allocated and
490 * how much is left over.
492 kprintf("DMA space used: %jdk, remaining available: %jdk\n",
493 (intmax_t)(vmstats.v_dma_pages - vm_contig_alist.bl_free) *
495 (intmax_t)vm_contig_alist.bl_free * (PAGE_SIZE / 1024));
497 SYSINIT(vm_pgend, SI_SUB_PROC0_POST, SI_ORDER_ANY,
498 vm_page_startup_finish, NULL)
502 * Scan comparison function for Red-Black tree scans. An inclusive
503 * (start,end) is expected. Other fields are not used.
506 rb_vm_page_scancmp(struct vm_page *p, void *data)
508 struct rb_vm_page_scan_info *info = data;
510 if (p->pindex < info->start_pindex)
512 if (p->pindex > info->end_pindex)
518 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2)
520 if (p1->pindex < p2->pindex)
522 if (p1->pindex > p2->pindex)
528 * Each page queue has its own spin lock, which is fairly optimal for
529 * allocating and freeing pages at least.
531 * The caller must hold the vm_page_spin_lock() before locking a vm_page's
532 * queue spinlock via this function. Also note that m->queue cannot change
533 * unless both the page and queue are locked.
537 _vm_page_queue_spin_lock(vm_page_t m)
542 if (queue != PQ_NONE) {
543 spin_lock(&vm_page_queues[queue].spin);
544 KKASSERT(queue == m->queue);
550 _vm_page_queue_spin_unlock(vm_page_t m)
556 if (queue != PQ_NONE)
557 spin_unlock(&vm_page_queues[queue].spin);
562 _vm_page_queues_spin_lock(u_short queue)
565 if (queue != PQ_NONE)
566 spin_lock(&vm_page_queues[queue].spin);
572 _vm_page_queues_spin_unlock(u_short queue)
575 if (queue != PQ_NONE)
576 spin_unlock(&vm_page_queues[queue].spin);
580 vm_page_queue_spin_lock(vm_page_t m)
582 _vm_page_queue_spin_lock(m);
586 vm_page_queues_spin_lock(u_short queue)
588 _vm_page_queues_spin_lock(queue);
592 vm_page_queue_spin_unlock(vm_page_t m)
594 _vm_page_queue_spin_unlock(m);
598 vm_page_queues_spin_unlock(u_short queue)
600 _vm_page_queues_spin_unlock(queue);
604 * This locks the specified vm_page and its queue in the proper order
605 * (page first, then queue). The queue may change so the caller must
610 _vm_page_and_queue_spin_lock(vm_page_t m)
612 vm_page_spin_lock(m);
613 _vm_page_queue_spin_lock(m);
618 _vm_page_and_queue_spin_unlock(vm_page_t m)
620 _vm_page_queues_spin_unlock(m->queue);
621 vm_page_spin_unlock(m);
625 vm_page_and_queue_spin_unlock(vm_page_t m)
627 _vm_page_and_queue_spin_unlock(m);
631 vm_page_and_queue_spin_lock(vm_page_t m)
633 _vm_page_and_queue_spin_lock(m);
637 * Helper function removes vm_page from its current queue.
638 * Returns the base queue the page used to be on.
640 * The vm_page and the queue must be spinlocked.
641 * This function will unlock the queue but leave the page spinlocked.
643 static __inline u_short
644 _vm_page_rem_queue_spinlocked(vm_page_t m)
646 struct vpgqueues *pq;
650 if (queue != PQ_NONE) {
651 pq = &vm_page_queues[queue];
652 TAILQ_REMOVE(&pq->pl, m, pageq);
653 atomic_add_int(pq->cnt, -1);
656 vm_page_queues_spin_unlock(queue);
657 if ((queue - m->pc) == PQ_FREE && (m->flags & PG_ZERO))
658 atomic_subtract_int(&vm_page_zero_count, 1);
659 if ((queue - m->pc) == PQ_CACHE || (queue - m->pc) == PQ_FREE)
660 return (queue - m->pc);
666 * Helper function places the vm_page on the specified queue.
668 * The vm_page must be spinlocked.
669 * This function will return with both the page and the queue locked.
672 _vm_page_add_queue_spinlocked(vm_page_t m, u_short queue, int athead)
674 struct vpgqueues *pq;
676 KKASSERT(m->queue == PQ_NONE);
678 if (queue != PQ_NONE) {
679 vm_page_queues_spin_lock(queue);
680 pq = &vm_page_queues[queue];
682 atomic_add_int(pq->cnt, 1);
686 * Put zero'd pages on the end ( where we look for zero'd pages
687 * first ) and non-zerod pages at the head.
689 if (queue - m->pc == PQ_FREE) {
690 if (m->flags & PG_ZERO) {
691 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
692 atomic_add_int(&vm_page_zero_count, 1);
694 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
697 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
699 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
701 /* leave the queue spinlocked */
706 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE)
707 * m->busy is zero. Returns TRUE if it had to sleep, FALSE if we
708 * did not. Only one sleep call will be made before returning.
710 * This function does NOT busy the page and on return the page is not
711 * guaranteed to be available.
714 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg)
722 if ((flags & PG_BUSY) == 0 &&
723 (also_m_busy == 0 || (flags & PG_SBUSY) == 0)) {
726 tsleep_interlock(m, 0);
727 if (atomic_cmpset_int(&m->flags, flags,
728 flags | PG_WANTED | PG_REFERENCED)) {
729 tsleep(m, PINTERLOCKED, msg, 0);
736 * Wait until PG_BUSY can be set, then set it. If also_m_busy is TRUE we
737 * also wait for m->busy to become 0 before setting PG_BUSY.
740 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m,
741 int also_m_busy, const char *msg
749 if (flags & PG_BUSY) {
750 tsleep_interlock(m, 0);
751 if (atomic_cmpset_int(&m->flags, flags,
752 flags | PG_WANTED | PG_REFERENCED)) {
753 tsleep(m, PINTERLOCKED, msg, 0);
755 } else if (also_m_busy && (flags & PG_SBUSY)) {
756 tsleep_interlock(m, 0);
757 if (atomic_cmpset_int(&m->flags, flags,
758 flags | PG_WANTED | PG_REFERENCED)) {
759 tsleep(m, PINTERLOCKED, msg, 0);
762 if (atomic_cmpset_int(&m->flags, flags,
766 m->busy_line = lineno;
775 * Attempt to set PG_BUSY. If also_m_busy is TRUE we only succeed if m->busy
778 * Returns non-zero on failure.
781 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy
791 if (also_m_busy && (flags & PG_SBUSY))
793 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
796 m->busy_line = lineno;
804 * Clear the PG_BUSY flag and return non-zero to indicate to the caller
805 * that a wakeup() should be performed.
807 * The vm_page must be spinlocked and will remain spinlocked on return.
808 * The related queue must NOT be spinlocked (which could deadlock us).
814 _vm_page_wakeup(vm_page_t m)
821 if (atomic_cmpset_int(&m->flags, flags,
822 flags & ~(PG_BUSY | PG_WANTED))) {
826 return(flags & PG_WANTED);
830 * Clear the PG_BUSY flag and wakeup anyone waiting for the page. This
831 * is typically the last call you make on a page before moving onto
835 vm_page_wakeup(vm_page_t m)
837 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!"));
838 vm_page_spin_lock(m);
839 if (_vm_page_wakeup(m)) {
840 vm_page_spin_unlock(m);
843 vm_page_spin_unlock(m);
848 * Holding a page keeps it from being reused. Other parts of the system
849 * can still disassociate the page from its current object and free it, or
850 * perform read or write I/O on it and/or otherwise manipulate the page,
851 * but if the page is held the VM system will leave the page and its data
852 * intact and not reuse the page for other purposes until the last hold
853 * reference is released. (see vm_page_wire() if you want to prevent the
854 * page from being disassociated from its object too).
856 * The caller must still validate the contents of the page and, if necessary,
857 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete
858 * before manipulating the page.
860 * XXX get vm_page_spin_lock() here and move FREE->HOLD if necessary
863 vm_page_hold(vm_page_t m)
865 vm_page_spin_lock(m);
866 atomic_add_int(&m->hold_count, 1);
867 if (m->queue - m->pc == PQ_FREE) {
868 _vm_page_queue_spin_lock(m);
869 _vm_page_rem_queue_spinlocked(m);
870 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
871 _vm_page_queue_spin_unlock(m);
873 vm_page_spin_unlock(m);
877 * The opposite of vm_page_hold(). A page can be freed while being held,
878 * which places it on the PQ_HOLD queue. If we are able to busy the page
879 * after the hold count drops to zero we will move the page to the
880 * appropriate PQ_FREE queue by calling vm_page_free_toq().
883 vm_page_unhold(vm_page_t m)
885 vm_page_spin_lock(m);
886 atomic_add_int(&m->hold_count, -1);
887 if (m->hold_count == 0 && m->queue - m->pc == PQ_HOLD) {
888 _vm_page_queue_spin_lock(m);
889 _vm_page_rem_queue_spinlocked(m);
890 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0);
891 _vm_page_queue_spin_unlock(m);
893 vm_page_spin_unlock(m);
897 * Inserts the given vm_page into the object and object list.
899 * The pagetables are not updated but will presumably fault the page
900 * in if necessary, or if a kernel page the caller will at some point
901 * enter the page into the kernel's pmap. We are not allowed to block
902 * here so we *can't* do this anyway.
904 * This routine may not block.
905 * This routine must be called with the vm_object held.
906 * This routine must be called with a critical section held.
908 * This routine returns TRUE if the page was inserted into the object
909 * successfully, and FALSE if the page already exists in the object.
912 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
914 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
915 if (m->object != NULL)
916 panic("vm_page_insert: already inserted");
918 object->generation++;
921 * Record the object/offset pair in this page and add the
922 * pv_list_count of the page to the object.
924 * The vm_page spin lock is required for interactions with the pmap.
926 vm_page_spin_lock(m);
929 if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) {
932 vm_page_spin_unlock(m);
935 object->resident_page_count++;
936 /* atomic_add_int(&object->agg_pv_list_count, m->md.pv_list_count); */
937 vm_page_spin_unlock(m);
940 * Since we are inserting a new and possibly dirty page,
941 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
943 if ((m->valid & m->dirty) || (m->flags & PG_WRITEABLE))
944 vm_object_set_writeable_dirty(object);
947 * Checks for a swap assignment and sets PG_SWAPPED if appropriate.
949 swap_pager_page_inserted(m);
954 * Removes the given vm_page_t from the (object,index) table
956 * The underlying pmap entry (if any) is NOT removed here.
957 * This routine may not block.
959 * The page must be BUSY and will remain BUSY on return.
960 * No other requirements.
962 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave
966 vm_page_remove(vm_page_t m)
970 if (m->object == NULL) {
974 if ((m->flags & PG_BUSY) == 0)
975 panic("vm_page_remove: page not busy");
979 vm_object_hold(object);
982 * Remove the page from the object and update the object.
984 * The vm_page spin lock is required for interactions with the pmap.
986 vm_page_spin_lock(m);
987 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m);
988 object->resident_page_count--;
989 /* atomic_add_int(&object->agg_pv_list_count, -m->md.pv_list_count); */
991 vm_page_spin_unlock(m);
993 object->generation++;
995 vm_object_drop(object);
999 * Locate and return the page at (object, pindex), or NULL if the
1000 * page could not be found.
1002 * The caller must hold the vm_object token.
1005 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1010 * Search the hash table for this object/offset pair
1012 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1013 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1014 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex));
1019 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object,
1021 int also_m_busy, const char *msg
1027 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1028 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1030 KKASSERT(m->object == object && m->pindex == pindex);
1033 if (flags & PG_BUSY) {
1034 tsleep_interlock(m, 0);
1035 if (atomic_cmpset_int(&m->flags, flags,
1036 flags | PG_WANTED | PG_REFERENCED)) {
1037 tsleep(m, PINTERLOCKED, msg, 0);
1038 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1041 } else if (also_m_busy && (flags & PG_SBUSY)) {
1042 tsleep_interlock(m, 0);
1043 if (atomic_cmpset_int(&m->flags, flags,
1044 flags | PG_WANTED | PG_REFERENCED)) {
1045 tsleep(m, PINTERLOCKED, msg, 0);
1046 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1049 } else if (atomic_cmpset_int(&m->flags, flags,
1051 #ifdef VM_PAGE_DEBUG
1052 m->busy_func = func;
1053 m->busy_line = lineno;
1062 * Attempt to lookup and busy a page.
1064 * Returns NULL if the page could not be found
1066 * Returns a vm_page and error == TRUE if the page exists but could not
1069 * Returns a vm_page and error == FALSE on success.
1072 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object,
1074 int also_m_busy, int *errorp
1080 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1081 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1084 KKASSERT(m->object == object && m->pindex == pindex);
1087 if (flags & PG_BUSY) {
1091 if (also_m_busy && (flags & PG_SBUSY)) {
1095 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
1096 #ifdef VM_PAGE_DEBUG
1097 m->busy_func = func;
1098 m->busy_line = lineno;
1107 * Caller must hold the related vm_object
1110 vm_page_next(vm_page_t m)
1114 next = vm_page_rb_tree_RB_NEXT(m);
1115 if (next && next->pindex != m->pindex + 1)
1123 * Move the given vm_page from its current object to the specified
1124 * target object/offset. The page must be busy and will remain so
1127 * new_object must be held.
1128 * This routine might block. XXX ?
1130 * NOTE: Swap associated with the page must be invalidated by the move. We
1131 * have to do this for several reasons: (1) we aren't freeing the
1132 * page, (2) we are dirtying the page, (3) the VM system is probably
1133 * moving the page from object A to B, and will then later move
1134 * the backing store from A to B and we can't have a conflict.
1136 * NOTE: We *always* dirty the page. It is necessary both for the
1137 * fact that we moved it, and because we may be invalidating
1138 * swap. If the page is on the cache, we have to deactivate it
1139 * or vm_page_dirty() will panic. Dirty pages are not allowed
1143 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1145 KKASSERT(m->flags & PG_BUSY);
1146 ASSERT_LWKT_TOKEN_HELD(vm_object_token(new_object));
1148 ASSERT_LWKT_TOKEN_HELD(vm_object_token(m->object));
1151 if (vm_page_insert(m, new_object, new_pindex) == FALSE) {
1152 panic("vm_page_rename: target exists (%p,%"PRIu64")",
1153 new_object, new_pindex);
1155 if (m->queue - m->pc == PQ_CACHE)
1156 vm_page_deactivate(m);
1161 * vm_page_unqueue() without any wakeup. This routine is used when a page
1162 * is being moved between queues or otherwise is to remain BUSYied by the
1165 * This routine may not block.
1168 vm_page_unqueue_nowakeup(vm_page_t m)
1170 vm_page_and_queue_spin_lock(m);
1171 (void)_vm_page_rem_queue_spinlocked(m);
1172 vm_page_spin_unlock(m);
1176 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
1179 * This routine may not block.
1182 vm_page_unqueue(vm_page_t m)
1186 vm_page_and_queue_spin_lock(m);
1187 queue = _vm_page_rem_queue_spinlocked(m);
1188 if (queue == PQ_FREE || queue == PQ_CACHE) {
1189 vm_page_spin_unlock(m);
1190 pagedaemon_wakeup();
1192 vm_page_spin_unlock(m);
1197 * vm_page_list_find()
1199 * Find a page on the specified queue with color optimization.
1201 * The page coloring optimization attempts to locate a page that does
1202 * not overload other nearby pages in the object in the cpu's L1 or L2
1203 * caches. We need this optimization because cpu caches tend to be
1204 * physical caches, while object spaces tend to be virtual.
1206 * On MP systems each PQ_FREE and PQ_CACHE color queue has its own spinlock
1207 * and the algorithm is adjusted to localize allocations on a per-core basis.
1208 * This is done by 'twisting' the colors.
1210 * The page is returned spinlocked and removed from its queue (it will
1211 * be on PQ_NONE), or NULL. The page is not PG_BUSY'd. The caller
1212 * is responsible for dealing with the busy-page case (usually by
1213 * deactivating the page and looping).
1215 * NOTE: This routine is carefully inlined. A non-inlined version
1216 * is available for outside callers but the only critical path is
1217 * from within this source file.
1219 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE
1220 * represent stable storage, allowing us to order our locks vm_page
1221 * first, then queue.
1225 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
1231 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist);
1233 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl);
1235 m = _vm_page_list_find2(basequeue, index);
1238 vm_page_and_queue_spin_lock(m);
1239 if (m->queue == basequeue + index) {
1240 _vm_page_rem_queue_spinlocked(m);
1241 /* vm_page_t spin held, no queue spin */
1244 vm_page_and_queue_spin_unlock(m);
1250 _vm_page_list_find2(int basequeue, int index)
1254 struct vpgqueues *pq;
1256 pq = &vm_page_queues[basequeue];
1259 * Note that for the first loop, index+i and index-i wind up at the
1260 * same place. Even though this is not totally optimal, we've already
1261 * blown it by missing the cache case so we do not care.
1263 for (i = PQ_L2_SIZE / 2; i > 0; --i) {
1265 m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl);
1267 _vm_page_and_queue_spin_lock(m);
1269 basequeue + ((index + i) & PQ_L2_MASK)) {
1270 _vm_page_rem_queue_spinlocked(m);
1273 _vm_page_and_queue_spin_unlock(m);
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);
1294 * Returns a vm_page candidate for allocation. The page is not busied so
1295 * it can move around. The caller must busy the page (and typically
1296 * deactivate it if it cannot be busied!)
1298 * Returns a spinlocked vm_page that has been removed from its queue.
1301 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
1303 return(_vm_page_list_find(basequeue, index, prefer_zero));
1307 * Find a page on the cache queue with color optimization, remove it
1308 * from the queue, and busy it. The returned page will not be spinlocked.
1310 * A candidate failure will be deactivated. Candidates can fail due to
1311 * being busied by someone else, in which case they will be deactivated.
1313 * This routine may not block.
1317 vm_page_select_cache(u_short pg_color)
1322 m = _vm_page_list_find(PQ_CACHE, pg_color & PQ_L2_MASK, FALSE);
1326 * (m) has been removed from its queue and spinlocked
1328 if (vm_page_busy_try(m, TRUE)) {
1329 _vm_page_deactivate_locked(m, 0);
1330 vm_page_spin_unlock(m);
1332 kprintf("Warning: busy page %p found in cache\n", m);
1336 * We successfully busied the page
1338 if ((m->flags & PG_UNMANAGED) == 0 &&
1339 m->hold_count == 0 &&
1340 m->wire_count == 0) {
1341 vm_page_spin_unlock(m);
1342 pagedaemon_wakeup();
1345 _vm_page_deactivate_locked(m, 0);
1346 if (_vm_page_wakeup(m)) {
1347 vm_page_spin_unlock(m);
1350 vm_page_spin_unlock(m);
1358 * Find a free or zero page, with specified preference. We attempt to
1359 * inline the nominal case and fall back to _vm_page_select_free()
1360 * otherwise. A busied page is removed from the queue and returned.
1362 * This routine may not block.
1364 static __inline vm_page_t
1365 vm_page_select_free(u_short pg_color, boolean_t prefer_zero)
1370 m = _vm_page_list_find(PQ_FREE, pg_color & PQ_L2_MASK,
1374 if (vm_page_busy_try(m, TRUE)) {
1376 * Various mechanisms such as a pmap_collect can
1377 * result in a busy page on the free queue. We
1378 * have to move the page out of the way so we can
1379 * retry the allocation. If the other thread is not
1380 * allocating the page then m->valid will remain 0 and
1381 * the pageout daemon will free the page later on.
1383 * Since we could not busy the page, however, we
1384 * cannot make assumptions as to whether the page
1385 * will be allocated by the other thread or not,
1386 * so all we can do is deactivate it to move it out
1387 * of the way. In particular, if the other thread
1388 * wires the page it may wind up on the inactive
1389 * queue and the pageout daemon will have to deal
1390 * with that case too.
1392 _vm_page_deactivate_locked(m, 0);
1393 vm_page_spin_unlock(m);
1395 kprintf("Warning: busy page %p found in cache\n", m);
1399 * Theoretically if we are able to busy the page
1400 * atomic with the queue removal (using the vm_page
1401 * lock) nobody else should be able to mess with the
1404 KKASSERT((m->flags & PG_UNMANAGED) == 0);
1405 KKASSERT(m->hold_count == 0);
1406 KKASSERT(m->wire_count == 0);
1407 vm_page_spin_unlock(m);
1408 pagedaemon_wakeup();
1410 /* return busied and removed page */
1418 * This implements a per-cpu cache of free, zero'd, ready-to-go pages.
1419 * The idea is to populate this cache prior to acquiring any locks so
1420 * we don't wind up potentially zeroing VM pages (under heavy loads) while
1421 * holding potentialy contending locks.
1423 * Note that we allocate the page uninserted into anything and use a pindex
1424 * of 0, the vm_page_alloc() will effectively add gd_cpuid so these
1425 * allocations should wind up being uncontended. However, we still want
1426 * to rove across PQ_L2_SIZE.
1429 vm_page_pcpu_cache(void)
1432 globaldata_t gd = mycpu;
1435 if (gd->gd_vmpg_count < GD_MINVMPG) {
1437 while (gd->gd_vmpg_count < GD_MAXVMPG) {
1438 m = vm_page_alloc(NULL, ticks & ~ncpus2_mask,
1439 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL |
1440 VM_ALLOC_NULL_OK | VM_ALLOC_ZERO);
1441 if (gd->gd_vmpg_count < GD_MAXVMPG) {
1442 if ((m->flags & PG_ZERO) == 0) {
1443 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1444 vm_page_flag_set(m, PG_ZERO);
1446 gd->gd_vmpg_array[gd->gd_vmpg_count++] = m;
1459 * Allocate and return a memory cell associated with this VM object/offset
1460 * pair. If object is NULL an unassociated page will be allocated.
1462 * The returned page will be busied and removed from its queues. This
1463 * routine can block and may return NULL if a race occurs and the page
1464 * is found to already exist at the specified (object, pindex).
1466 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain
1467 * VM_ALLOC_QUICK like normal but cannot use cache
1468 * VM_ALLOC_SYSTEM greater free drain
1469 * VM_ALLOC_INTERRUPT allow free list to be completely drained
1470 * VM_ALLOC_ZERO advisory request for pre-zero'd page only
1471 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only
1472 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision
1473 * (see vm_page_grab())
1474 * VM_ALLOC_USE_GD ok to use per-gd cache
1476 * The object must be held if not NULL
1477 * This routine may not block
1479 * Additional special handling is required when called from an interrupt
1480 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache
1484 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
1487 globaldata_t gd = mycpu;
1494 * Special per-cpu free VM page cache. The pages are pre-busied
1495 * and pre-zerod for us.
1497 if (gd->gd_vmpg_count && (page_req & VM_ALLOC_USE_GD)) {
1499 if (gd->gd_vmpg_count) {
1500 m = gd->gd_vmpg_array[--gd->gd_vmpg_count];
1511 * Cpu twist - cpu localization algorithm
1514 pg_color = gd->gd_cpuid + (pindex & ~ncpus_fit_mask) +
1515 (object->pg_color & ~ncpus_fit_mask);
1517 pg_color = gd->gd_cpuid + (pindex & ~ncpus_fit_mask);
1521 * Normal page coloring algorithm
1524 pg_color = object->pg_color + pindex;
1530 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK|
1531 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
1534 * Certain system threads (pageout daemon, buf_daemon's) are
1535 * allowed to eat deeper into the free page list.
1537 if (curthread->td_flags & TDF_SYSTHREAD)
1538 page_req |= VM_ALLOC_SYSTEM;
1541 if (vmstats.v_free_count > vmstats.v_free_reserved ||
1542 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) ||
1543 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 &&
1544 vmstats.v_free_count > vmstats.v_interrupt_free_min)
1547 * The free queue has sufficient free pages to take one out.
1549 if (page_req & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO))
1550 m = vm_page_select_free(pg_color, TRUE);
1552 m = vm_page_select_free(pg_color, FALSE);
1553 } else if (page_req & VM_ALLOC_NORMAL) {
1555 * Allocatable from the cache (non-interrupt only). On
1556 * success, we must free the page and try again, thus
1557 * ensuring that vmstats.v_*_free_min counters are replenished.
1560 if (curthread->td_preempted) {
1561 kprintf("vm_page_alloc(): warning, attempt to allocate"
1562 " cache page from preempting interrupt\n");
1565 m = vm_page_select_cache(pg_color);
1568 m = vm_page_select_cache(pg_color);
1571 * On success move the page into the free queue and loop.
1573 * Only do this if we can safely acquire the vm_object lock,
1574 * because this is effectively a random page and the caller
1575 * might be holding the lock shared, we don't want to
1579 KASSERT(m->dirty == 0,
1580 ("Found dirty cache page %p", m));
1582 if (vm_object_hold_try(m->object)) {
1583 vm_page_protect(m, VM_PROT_NONE);
1585 vm_object_drop(m->object);
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) {
1656 kprintf("PAGE RACE (%p:%d,%"PRIu64")\n",
1657 object, object->type, pindex);
1660 if ((page_req & VM_ALLOC_NULL_OK) == 0)
1668 * Don't wakeup too often - wakeup the pageout daemon when
1669 * we would be nearly out of memory.
1671 pagedaemon_wakeup();
1674 * A PG_BUSY page is returned.
1680 * Attempt to allocate contiguous physical memory with the specified
1684 vm_page_alloc_contig(vm_paddr_t low, vm_paddr_t high,
1685 unsigned long alignment, unsigned long boundary,
1690 alignment >>= PAGE_SHIFT;
1693 boundary >>= PAGE_SHIFT;
1696 size = (size + PAGE_MASK) >> PAGE_SHIFT;
1698 spin_lock(&vm_contig_spin);
1699 blk = alist_alloc(&vm_contig_alist, 0, size);
1700 if (blk == ALIST_BLOCK_NONE) {
1701 spin_unlock(&vm_contig_spin);
1703 kprintf("vm_page_alloc_contig: %ldk nospace\n",
1704 (size + PAGE_MASK) * (PAGE_SIZE / 1024));
1708 if (high && ((vm_paddr_t)(blk + size) << PAGE_SHIFT) > high) {
1709 alist_free(&vm_contig_alist, blk, size);
1710 spin_unlock(&vm_contig_spin);
1712 kprintf("vm_page_alloc_contig: %ldk high "
1714 (size + PAGE_MASK) * (PAGE_SIZE / 1024),
1719 spin_unlock(&vm_contig_spin);
1721 kprintf("vm_page_alloc_contig: %016jx/%ldk\n",
1722 (intmax_t)(vm_paddr_t)blk << PAGE_SHIFT,
1723 (size + PAGE_MASK) * (PAGE_SIZE / 1024));
1725 return (PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT));
1729 * Free contiguously allocated pages. The pages will be wired but not busy.
1730 * When freeing to the alist we leave them wired and not busy.
1733 vm_page_free_contig(vm_page_t m, unsigned long size)
1735 vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1736 vm_pindex_t start = pa >> PAGE_SHIFT;
1737 vm_pindex_t pages = (size + PAGE_MASK) >> PAGE_SHIFT;
1740 kprintf("vm_page_free_contig: %016jx/%ldk\n",
1741 (intmax_t)pa, size / 1024);
1743 if (pa < vm_low_phys_reserved) {
1744 KKASSERT(pa + size <= vm_low_phys_reserved);
1745 spin_lock(&vm_contig_spin);
1746 alist_free(&vm_contig_alist, start, pages);
1747 spin_unlock(&vm_contig_spin);
1750 vm_page_busy_wait(m, FALSE, "cpgfr");
1751 vm_page_unwire(m, 0);
1762 * Wait for sufficient free memory for nominal heavy memory use kernel
1766 vm_wait_nominal(void)
1768 while (vm_page_count_min(0))
1773 * Test if vm_wait_nominal() would block.
1776 vm_test_nominal(void)
1778 if (vm_page_count_min(0))
1784 * Block until free pages are available for allocation, called in various
1785 * places before memory allocations.
1787 * The caller may loop if vm_page_count_min() == FALSE so we cannot be
1788 * more generous then that.
1794 * never wait forever
1798 lwkt_gettoken(&vm_token);
1800 if (curthread == pagethread) {
1802 * The pageout daemon itself needs pages, this is bad.
1804 if (vm_page_count_min(0)) {
1805 vm_pageout_pages_needed = 1;
1806 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo);
1810 * Wakeup the pageout daemon if necessary and wait.
1812 if (vm_page_count_target()) {
1813 if (vm_pages_needed == 0) {
1814 vm_pages_needed = 1;
1815 wakeup(&vm_pages_needed);
1817 ++vm_pages_waiting; /* SMP race ok */
1818 tsleep(&vmstats.v_free_count, 0, "vmwait", timo);
1821 lwkt_reltoken(&vm_token);
1825 * Block until free pages are available for allocation
1827 * Called only from vm_fault so that processes page faulting can be
1834 * Wakeup the pageout daemon if necessary and wait.
1836 if (vm_page_count_target()) {
1837 lwkt_gettoken(&vm_token);
1838 if (vm_page_count_target()) {
1839 if (vm_pages_needed == 0) {
1840 vm_pages_needed = 1;
1841 wakeup(&vm_pages_needed);
1843 ++vm_pages_waiting; /* SMP race ok */
1844 tsleep(&vmstats.v_free_count, 0, "pfault", hz);
1846 lwkt_reltoken(&vm_token);
1851 * Put the specified page on the active list (if appropriate). Ensure
1852 * that act_count is at least ACT_INIT but do not otherwise mess with it.
1854 * The caller should be holding the page busied ? XXX
1855 * This routine may not block.
1858 vm_page_activate(vm_page_t m)
1862 vm_page_spin_lock(m);
1863 if (m->queue - m->pc != PQ_ACTIVE) {
1864 _vm_page_queue_spin_lock(m);
1865 oqueue = _vm_page_rem_queue_spinlocked(m);
1866 /* page is left spinlocked, queue is unlocked */
1868 if (oqueue == PQ_CACHE)
1869 mycpu->gd_cnt.v_reactivated++;
1870 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1871 if (m->act_count < ACT_INIT)
1872 m->act_count = ACT_INIT;
1873 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE + m->pc, 0);
1875 _vm_page_and_queue_spin_unlock(m);
1876 if (oqueue == PQ_CACHE || oqueue == PQ_FREE)
1877 pagedaemon_wakeup();
1879 if (m->act_count < ACT_INIT)
1880 m->act_count = ACT_INIT;
1881 vm_page_spin_unlock(m);
1886 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
1887 * routine is called when a page has been added to the cache or free
1890 * This routine may not block.
1892 static __inline void
1893 vm_page_free_wakeup(void)
1896 * If the pageout daemon itself needs pages, then tell it that
1897 * there are some free.
1899 if (vm_pageout_pages_needed &&
1900 vmstats.v_cache_count + vmstats.v_free_count >=
1901 vmstats.v_pageout_free_min
1903 wakeup(&vm_pageout_pages_needed);
1904 vm_pageout_pages_needed = 0;
1908 * Wakeup processes that are waiting on memory.
1910 * NOTE: vm_paging_target() is the pageout daemon's target, while
1911 * vm_page_count_target() is somewhere inbetween. We want
1912 * to wake processes up prior to the pageout daemon reaching
1913 * its target to provide some hysteresis.
1915 if (vm_pages_waiting) {
1916 if (!vm_page_count_target()) {
1918 * Plenty of pages are free, wakeup everyone.
1920 vm_pages_waiting = 0;
1921 wakeup(&vmstats.v_free_count);
1922 ++mycpu->gd_cnt.v_ppwakeups;
1923 } else if (!vm_page_count_min(0)) {
1925 * Some pages are free, wakeup someone.
1927 int wcount = vm_pages_waiting;
1930 vm_pages_waiting = wcount;
1931 wakeup_one(&vmstats.v_free_count);
1932 ++mycpu->gd_cnt.v_ppwakeups;
1938 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates
1939 * it from its VM object.
1941 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on
1942 * return (the page will have been freed).
1945 vm_page_free_toq(vm_page_t m)
1947 mycpu->gd_cnt.v_tfree++;
1948 KKASSERT((m->flags & PG_MAPPED) == 0);
1949 KKASSERT(m->flags & PG_BUSY);
1951 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
1952 kprintf("vm_page_free: pindex(%lu), busy(%d), "
1953 "PG_BUSY(%d), hold(%d)\n",
1954 (u_long)m->pindex, m->busy,
1955 ((m->flags & PG_BUSY) ? 1 : 0), m->hold_count);
1956 if ((m->queue - m->pc) == PQ_FREE)
1957 panic("vm_page_free: freeing free page");
1959 panic("vm_page_free: freeing busy page");
1963 * Remove from object, spinlock the page and its queues and
1964 * remove from any queue. No queue spinlock will be held
1965 * after this section (because the page was removed from any
1969 vm_page_and_queue_spin_lock(m);
1970 _vm_page_rem_queue_spinlocked(m);
1973 * No further management of fictitious pages occurs beyond object
1974 * and queue removal.
1976 if ((m->flags & PG_FICTITIOUS) != 0) {
1977 vm_page_spin_unlock(m);
1985 if (m->wire_count != 0) {
1986 if (m->wire_count > 1) {
1988 "vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1989 m->wire_count, (long)m->pindex);
1991 panic("vm_page_free: freeing wired page");
1995 * Clear the UNMANAGED flag when freeing an unmanaged page.
1997 if (m->flags & PG_UNMANAGED) {
1998 vm_page_flag_clear(m, PG_UNMANAGED);
2001 if (m->hold_count != 0) {
2002 vm_page_flag_clear(m, PG_ZERO);
2003 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
2005 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0);
2009 * This sequence allows us to clear PG_BUSY while still holding
2010 * its spin lock, which reduces contention vs allocators. We
2011 * must not leave the queue locked or _vm_page_wakeup() may
2014 _vm_page_queue_spin_unlock(m);
2015 if (_vm_page_wakeup(m)) {
2016 vm_page_spin_unlock(m);
2019 vm_page_spin_unlock(m);
2021 vm_page_free_wakeup();
2025 * vm_page_free_fromq_fast()
2027 * Remove a non-zero page from one of the free queues; the page is removed for
2028 * zeroing, so do not issue a wakeup.
2031 vm_page_free_fromq_fast(void)
2037 for (i = 0; i < PQ_L2_SIZE; ++i) {
2038 m = vm_page_list_find(PQ_FREE, qi, FALSE);
2039 /* page is returned spinlocked and removed from its queue */
2041 if (vm_page_busy_try(m, TRUE)) {
2043 * We were unable to busy the page, deactivate
2046 _vm_page_deactivate_locked(m, 0);
2047 vm_page_spin_unlock(m);
2048 } else if (m->flags & PG_ZERO) {
2050 * The page is PG_ZERO, requeue it and loop
2052 _vm_page_add_queue_spinlocked(m,
2055 vm_page_queue_spin_unlock(m);
2056 if (_vm_page_wakeup(m)) {
2057 vm_page_spin_unlock(m);
2060 vm_page_spin_unlock(m);
2064 * The page is not PG_ZERO'd so return it.
2066 vm_page_spin_unlock(m);
2067 KKASSERT((m->flags & PG_UNMANAGED) == 0);
2068 KKASSERT(m->hold_count == 0);
2069 KKASSERT(m->wire_count == 0);
2074 qi = (qi + PQ_PRIME2) & PQ_L2_MASK;
2080 * vm_page_unmanage()
2082 * Prevent PV management from being done on the page. The page is
2083 * removed from the paging queues as if it were wired, and as a
2084 * consequence of no longer being managed the pageout daemon will not
2085 * touch it (since there is no way to locate the pte mappings for the
2086 * page). madvise() calls that mess with the pmap will also no longer
2087 * operate on the page.
2089 * Beyond that the page is still reasonably 'normal'. Freeing the page
2090 * will clear the flag.
2092 * This routine is used by OBJT_PHYS objects - objects using unswappable
2093 * physical memory as backing store rather then swap-backed memory and
2094 * will eventually be extended to support 4MB unmanaged physical
2097 * Caller must be holding the page busy.
2100 vm_page_unmanage(vm_page_t m)
2102 KKASSERT(m->flags & PG_BUSY);
2103 if ((m->flags & PG_UNMANAGED) == 0) {
2104 if (m->wire_count == 0)
2107 vm_page_flag_set(m, PG_UNMANAGED);
2111 * Mark this page as wired down by yet another map, removing it from
2112 * paging queues as necessary.
2114 * Caller must be holding the page busy.
2117 vm_page_wire(vm_page_t m)
2120 * Only bump the wire statistics if the page is not already wired,
2121 * and only unqueue the page if it is on some queue (if it is unmanaged
2122 * it is already off the queues). Don't do anything with fictitious
2123 * pages because they are always wired.
2125 KKASSERT(m->flags & PG_BUSY);
2126 if ((m->flags & PG_FICTITIOUS) == 0) {
2127 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) {
2128 if ((m->flags & PG_UNMANAGED) == 0)
2130 atomic_add_int(&vmstats.v_wire_count, 1);
2132 KASSERT(m->wire_count != 0,
2133 ("vm_page_wire: wire_count overflow m=%p", m));
2138 * Release one wiring of this page, potentially enabling it to be paged again.
2140 * Many pages placed on the inactive queue should actually go
2141 * into the cache, but it is difficult to figure out which. What
2142 * we do instead, if the inactive target is well met, is to put
2143 * clean pages at the head of the inactive queue instead of the tail.
2144 * This will cause them to be moved to the cache more quickly and
2145 * if not actively re-referenced, freed more quickly. If we just
2146 * stick these pages at the end of the inactive queue, heavy filesystem
2147 * meta-data accesses can cause an unnecessary paging load on memory bound
2148 * processes. This optimization causes one-time-use metadata to be
2149 * reused more quickly.
2151 * BUT, if we are in a low-memory situation we have no choice but to
2152 * put clean pages on the cache queue.
2154 * A number of routines use vm_page_unwire() to guarantee that the page
2155 * will go into either the inactive or active queues, and will NEVER
2156 * be placed in the cache - for example, just after dirtying a page.
2157 * dirty pages in the cache are not allowed.
2159 * The page queues must be locked.
2160 * This routine may not block.
2163 vm_page_unwire(vm_page_t m, int activate)
2165 KKASSERT(m->flags & PG_BUSY);
2166 if (m->flags & PG_FICTITIOUS) {
2168 } else if (m->wire_count <= 0) {
2169 panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
2171 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) {
2172 atomic_add_int(&vmstats.v_wire_count, -1);
2173 if (m->flags & PG_UNMANAGED) {
2175 } else if (activate) {
2176 vm_page_spin_lock(m);
2177 _vm_page_add_queue_spinlocked(m,
2178 PQ_ACTIVE + m->pc, 0);
2179 _vm_page_and_queue_spin_unlock(m);
2181 vm_page_spin_lock(m);
2182 vm_page_flag_clear(m, PG_WINATCFLS);
2183 _vm_page_add_queue_spinlocked(m,
2184 PQ_INACTIVE + m->pc, 0);
2185 ++vm_swapcache_inactive_heuristic;
2186 _vm_page_and_queue_spin_unlock(m);
2193 * Move the specified page to the inactive queue. If the page has
2194 * any associated swap, the swap is deallocated.
2196 * Normally athead is 0 resulting in LRU operation. athead is set
2197 * to 1 if we want this page to be 'as if it were placed in the cache',
2198 * except without unmapping it from the process address space.
2200 * vm_page's spinlock must be held on entry and will remain held on return.
2201 * This routine may not block.
2204 _vm_page_deactivate_locked(vm_page_t m, int athead)
2209 * Ignore if already inactive.
2211 if (m->queue - m->pc == PQ_INACTIVE)
2213 _vm_page_queue_spin_lock(m);
2214 oqueue = _vm_page_rem_queue_spinlocked(m);
2216 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
2217 if (oqueue == PQ_CACHE)
2218 mycpu->gd_cnt.v_reactivated++;
2219 vm_page_flag_clear(m, PG_WINATCFLS);
2220 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead);
2222 ++vm_swapcache_inactive_heuristic;
2224 _vm_page_queue_spin_unlock(m);
2225 /* leaves vm_page spinlocked */
2229 * Attempt to deactivate a page.
2234 vm_page_deactivate(vm_page_t m)
2236 vm_page_spin_lock(m);
2237 _vm_page_deactivate_locked(m, 0);
2238 vm_page_spin_unlock(m);
2242 vm_page_deactivate_locked(vm_page_t m)
2244 _vm_page_deactivate_locked(m, 0);
2248 * Attempt to move a page to PQ_CACHE.
2250 * Returns 0 on failure, 1 on success
2252 * The page should NOT be busied by the caller. This function will validate
2253 * whether the page can be safely moved to the cache.
2256 vm_page_try_to_cache(vm_page_t m)
2258 vm_page_spin_lock(m);
2259 if (vm_page_busy_try(m, TRUE)) {
2260 vm_page_spin_unlock(m);
2263 if (m->dirty || m->hold_count || m->wire_count ||
2264 (m->flags & PG_UNMANAGED)) {
2265 if (_vm_page_wakeup(m)) {
2266 vm_page_spin_unlock(m);
2269 vm_page_spin_unlock(m);
2273 vm_page_spin_unlock(m);
2276 * Page busied by us and no longer spinlocked. Dirty pages cannot
2277 * be moved to the cache.
2279 vm_page_test_dirty(m);
2289 * Attempt to free the page. If we cannot free it, we do nothing.
2290 * 1 is returned on success, 0 on failure.
2295 vm_page_try_to_free(vm_page_t m)
2297 vm_page_spin_lock(m);
2298 if (vm_page_busy_try(m, TRUE)) {
2299 vm_page_spin_unlock(m);
2302 if (m->dirty || m->hold_count || m->wire_count ||
2303 (m->flags & PG_UNMANAGED)) {
2304 if (_vm_page_wakeup(m)) {
2305 vm_page_spin_unlock(m);
2308 vm_page_spin_unlock(m);
2312 vm_page_spin_unlock(m);
2315 * Page busied by us and no longer spinlocked. Dirty pages will
2316 * not be freed by this function. We have to re-test the
2317 * dirty bit after cleaning out the pmaps.
2319 vm_page_test_dirty(m);
2324 vm_page_protect(m, VM_PROT_NONE);
2336 * Put the specified page onto the page cache queue (if appropriate).
2338 * The page must be busy, and this routine will release the busy and
2339 * possibly even free the page.
2342 vm_page_cache(vm_page_t m)
2344 if ((m->flags & PG_UNMANAGED) || m->busy ||
2345 m->wire_count || m->hold_count) {
2346 kprintf("vm_page_cache: attempting to cache busy/held page\n");
2352 * Already in the cache (and thus not mapped)
2354 if ((m->queue - m->pc) == PQ_CACHE) {
2355 KKASSERT((m->flags & PG_MAPPED) == 0);
2361 * Caller is required to test m->dirty, but note that the act of
2362 * removing the page from its maps can cause it to become dirty
2363 * on an SMP system due to another cpu running in usermode.
2366 panic("vm_page_cache: caching a dirty page, pindex: %ld",
2371 * Remove all pmaps and indicate that the page is not
2372 * writeable or mapped. Our vm_page_protect() call may
2373 * have blocked (especially w/ VM_PROT_NONE), so recheck
2376 vm_page_protect(m, VM_PROT_NONE);
2377 if ((m->flags & (PG_UNMANAGED|PG_MAPPED)) || m->busy ||
2378 m->wire_count || m->hold_count) {
2380 } else if (m->dirty) {
2381 vm_page_deactivate(m);
2384 _vm_page_and_queue_spin_lock(m);
2385 _vm_page_rem_queue_spinlocked(m);
2386 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0);
2387 _vm_page_queue_spin_unlock(m);
2388 if (_vm_page_wakeup(m)) {
2389 vm_page_spin_unlock(m);
2392 vm_page_spin_unlock(m);
2394 vm_page_free_wakeup();
2399 * vm_page_dontneed()
2401 * Cache, deactivate, or do nothing as appropriate. This routine
2402 * is typically used by madvise() MADV_DONTNEED.
2404 * Generally speaking we want to move the page into the cache so
2405 * it gets reused quickly. However, this can result in a silly syndrome
2406 * due to the page recycling too quickly. Small objects will not be
2407 * fully cached. On the otherhand, if we move the page to the inactive
2408 * queue we wind up with a problem whereby very large objects
2409 * unnecessarily blow away our inactive and cache queues.
2411 * The solution is to move the pages based on a fixed weighting. We
2412 * either leave them alone, deactivate them, or move them to the cache,
2413 * where moving them to the cache has the highest weighting.
2414 * By forcing some pages into other queues we eventually force the
2415 * system to balance the queues, potentially recovering other unrelated
2416 * space from active. The idea is to not force this to happen too
2419 * The page must be busied.
2422 vm_page_dontneed(vm_page_t m)
2424 static int dnweight;
2431 * occassionally leave the page alone
2433 if ((dnw & 0x01F0) == 0 ||
2434 m->queue - m->pc == PQ_INACTIVE ||
2435 m->queue - m->pc == PQ_CACHE
2437 if (m->act_count >= ACT_INIT)
2443 * If vm_page_dontneed() is inactivating a page, it must clear
2444 * the referenced flag; otherwise the pagedaemon will see references
2445 * on the page in the inactive queue and reactivate it. Until the
2446 * page can move to the cache queue, madvise's job is not done.
2448 vm_page_flag_clear(m, PG_REFERENCED);
2449 pmap_clear_reference(m);
2452 vm_page_test_dirty(m);
2454 if (m->dirty || (dnw & 0x0070) == 0) {
2456 * Deactivate the page 3 times out of 32.
2461 * Cache the page 28 times out of every 32. Note that
2462 * the page is deactivated instead of cached, but placed
2463 * at the head of the queue instead of the tail.
2467 vm_page_spin_lock(m);
2468 _vm_page_deactivate_locked(m, head);
2469 vm_page_spin_unlock(m);
2473 * These routines manipulate the 'soft busy' count for a page. A soft busy
2474 * is almost like PG_BUSY except that it allows certain compatible operations
2475 * to occur on the page while it is busy. For example, a page undergoing a
2476 * write can still be mapped read-only.
2478 * Because vm_pages can overlap buffers m->busy can be > 1. m->busy is only
2479 * adjusted while the vm_page is PG_BUSY so the flash will occur when the
2480 * busy bit is cleared.
2483 vm_page_io_start(vm_page_t m)
2485 KASSERT(m->flags & PG_BUSY, ("vm_page_io_start: page not busy!!!"));
2486 atomic_add_char(&m->busy, 1);
2487 vm_page_flag_set(m, PG_SBUSY);
2491 vm_page_io_finish(vm_page_t m)
2493 KASSERT(m->flags & PG_BUSY, ("vm_page_io_finish: page not busy!!!"));
2494 atomic_subtract_char(&m->busy, 1);
2496 vm_page_flag_clear(m, PG_SBUSY);
2500 * Grab a page, blocking if it is busy and allocating a page if necessary.
2501 * A busy page is returned or NULL. The page may or may not be valid and
2502 * might not be on a queue (the caller is responsible for the disposition of
2505 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the
2506 * page will be zero'd and marked valid.
2508 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked
2509 * valid even if it already exists.
2511 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also
2512 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified.
2513 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified.
2515 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
2516 * always returned if we had blocked.
2518 * This routine may not be called from an interrupt.
2520 * PG_ZERO is *ALWAYS* cleared by this routine.
2522 * No other requirements.
2525 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2530 KKASSERT(allocflags &
2531 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
2532 vm_object_hold(object);
2534 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error);
2536 vm_page_sleep_busy(m, TRUE, "pgrbwt");
2537 if ((allocflags & VM_ALLOC_RETRY) == 0) {
2542 } else if (m == NULL) {
2543 if (allocflags & VM_ALLOC_RETRY)
2544 allocflags |= VM_ALLOC_NULL_OK;
2545 m = vm_page_alloc(object, pindex,
2546 allocflags & ~VM_ALLOC_RETRY);
2550 if ((allocflags & VM_ALLOC_RETRY) == 0)
2559 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid.
2561 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set
2562 * valid even if already valid.
2564 if (m->valid == 0) {
2565 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) {
2566 if ((m->flags & PG_ZERO) == 0)
2567 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2568 m->valid = VM_PAGE_BITS_ALL;
2570 } else if (allocflags & VM_ALLOC_FORCE_ZERO) {
2571 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2572 m->valid = VM_PAGE_BITS_ALL;
2574 vm_page_flag_clear(m, PG_ZERO);
2576 vm_object_drop(object);
2581 * Mapping function for valid bits or for dirty bits in
2582 * a page. May not block.
2584 * Inputs are required to range within a page.
2590 vm_page_bits(int base, int size)
2596 base + size <= PAGE_SIZE,
2597 ("vm_page_bits: illegal base/size %d/%d", base, size)
2600 if (size == 0) /* handle degenerate case */
2603 first_bit = base >> DEV_BSHIFT;
2604 last_bit = (base + size - 1) >> DEV_BSHIFT;
2606 return ((2 << last_bit) - (1 << first_bit));
2610 * Sets portions of a page valid and clean. The arguments are expected
2611 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2612 * of any partial chunks touched by the range. The invalid portion of
2613 * such chunks will be zero'd.
2615 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically
2616 * align base to DEV_BSIZE so as not to mark clean a partially
2617 * truncated device block. Otherwise the dirty page status might be
2620 * This routine may not block.
2622 * (base + size) must be less then or equal to PAGE_SIZE.
2625 _vm_page_zero_valid(vm_page_t m, int base, int size)
2630 if (size == 0) /* handle degenerate case */
2634 * If the base is not DEV_BSIZE aligned and the valid
2635 * bit is clear, we have to zero out a portion of the
2639 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2640 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
2642 pmap_zero_page_area(
2650 * If the ending offset is not DEV_BSIZE aligned and the
2651 * valid bit is clear, we have to zero out a portion of
2655 endoff = base + size;
2657 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2658 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
2660 pmap_zero_page_area(
2663 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
2669 * Set valid, clear dirty bits. If validating the entire
2670 * page we can safely clear the pmap modify bit. We also
2671 * use this opportunity to clear the PG_NOSYNC flag. If a process
2672 * takes a write fault on a MAP_NOSYNC memory area the flag will
2675 * We set valid bits inclusive of any overlap, but we can only
2676 * clear dirty bits for DEV_BSIZE chunks that are fully within
2679 * Page must be busied?
2680 * No other requirements.
2683 vm_page_set_valid(vm_page_t m, int base, int size)
2685 _vm_page_zero_valid(m, base, size);
2686 m->valid |= vm_page_bits(base, size);
2691 * Set valid bits and clear dirty bits.
2693 * NOTE: This function does not clear the pmap modified bit.
2694 * Also note that e.g. NFS may use a byte-granular base
2697 * WARNING: Page must be busied? But vfs_clean_one_page() will call
2698 * this without necessarily busying the page (via bdwrite()).
2699 * So for now vm_token must also be held.
2701 * No other requirements.
2704 vm_page_set_validclean(vm_page_t m, int base, int size)
2708 _vm_page_zero_valid(m, base, size);
2709 pagebits = vm_page_bits(base, size);
2710 m->valid |= pagebits;
2711 m->dirty &= ~pagebits;
2712 if (base == 0 && size == PAGE_SIZE) {
2713 /*pmap_clear_modify(m);*/
2714 vm_page_flag_clear(m, PG_NOSYNC);
2719 * Set valid & dirty. Used by buwrite()
2721 * WARNING: Page must be busied? But vfs_dirty_one_page() will
2722 * call this function in buwrite() so for now vm_token must
2725 * No other requirements.
2728 vm_page_set_validdirty(vm_page_t m, int base, int size)
2732 pagebits = vm_page_bits(base, size);
2733 m->valid |= pagebits;
2734 m->dirty |= pagebits;
2736 vm_object_set_writeable_dirty(m->object);
2742 * NOTE: This function does not clear the pmap modified bit.
2743 * Also note that e.g. NFS may use a byte-granular base
2746 * Page must be busied?
2747 * No other requirements.
2750 vm_page_clear_dirty(vm_page_t m, int base, int size)
2752 m->dirty &= ~vm_page_bits(base, size);
2753 if (base == 0 && size == PAGE_SIZE) {
2754 /*pmap_clear_modify(m);*/
2755 vm_page_flag_clear(m, PG_NOSYNC);
2760 * Make the page all-dirty.
2762 * Also make sure the related object and vnode reflect the fact that the
2763 * object may now contain a dirty page.
2765 * Page must be busied?
2766 * No other requirements.
2769 vm_page_dirty(vm_page_t m)
2772 int pqtype = m->queue - m->pc;
2774 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE,
2775 ("vm_page_dirty: page in free/cache queue!"));
2776 if (m->dirty != VM_PAGE_BITS_ALL) {
2777 m->dirty = VM_PAGE_BITS_ALL;
2779 vm_object_set_writeable_dirty(m->object);
2784 * Invalidates DEV_BSIZE'd chunks within a page. Both the
2785 * valid and dirty bits for the effected areas are cleared.
2787 * Page must be busied?
2789 * No other requirements.
2792 vm_page_set_invalid(vm_page_t m, int base, int size)
2796 bits = vm_page_bits(base, size);
2799 m->object->generation++;
2803 * The kernel assumes that the invalid portions of a page contain
2804 * garbage, but such pages can be mapped into memory by user code.
2805 * When this occurs, we must zero out the non-valid portions of the
2806 * page so user code sees what it expects.
2808 * Pages are most often semi-valid when the end of a file is mapped
2809 * into memory and the file's size is not page aligned.
2811 * Page must be busied?
2812 * No other requirements.
2815 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
2821 * Scan the valid bits looking for invalid sections that
2822 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
2823 * valid bit may be set ) have already been zerod by
2824 * vm_page_set_validclean().
2826 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
2827 if (i == (PAGE_SIZE / DEV_BSIZE) ||
2828 (m->valid & (1 << i))
2831 pmap_zero_page_area(
2834 (i - b) << DEV_BSHIFT
2842 * setvalid is TRUE when we can safely set the zero'd areas
2843 * as being valid. We can do this if there are no cache consistency
2844 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
2847 m->valid = VM_PAGE_BITS_ALL;
2851 * Is a (partial) page valid? Note that the case where size == 0
2852 * will return FALSE in the degenerate case where the page is entirely
2853 * invalid, and TRUE otherwise.
2856 * No other requirements.
2859 vm_page_is_valid(vm_page_t m, int base, int size)
2861 int bits = vm_page_bits(base, size);
2863 if (m->valid && ((m->valid & bits) == bits))
2870 * update dirty bits from pmap/mmu. May not block.
2872 * Caller must hold the page busy
2875 vm_page_test_dirty(vm_page_t m)
2877 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
2883 * Register an action, associating it with its vm_page
2886 vm_page_register_action(vm_page_action_t action, vm_page_event_t event)
2888 struct vm_page_action_list *list;
2891 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
2892 list = &action_list[hv];
2894 lwkt_gettoken(&vm_token);
2895 vm_page_flag_set(action->m, PG_ACTIONLIST);
2896 action->event = event;
2897 LIST_INSERT_HEAD(list, action, entry);
2898 lwkt_reltoken(&vm_token);
2902 * Unregister an action, disassociating it from its related vm_page
2905 vm_page_unregister_action(vm_page_action_t action)
2907 struct vm_page_action_list *list;
2910 lwkt_gettoken(&vm_token);
2911 if (action->event != VMEVENT_NONE) {
2912 action->event = VMEVENT_NONE;
2913 LIST_REMOVE(action, entry);
2915 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
2916 list = &action_list[hv];
2917 if (LIST_EMPTY(list))
2918 vm_page_flag_clear(action->m, PG_ACTIONLIST);
2920 lwkt_reltoken(&vm_token);
2924 * Issue an event on a VM page. Corresponding action structures are
2925 * removed from the page's list and called.
2927 * If the vm_page has no more pending action events we clear its
2928 * PG_ACTIONLIST flag.
2931 vm_page_event_internal(vm_page_t m, vm_page_event_t event)
2933 struct vm_page_action_list *list;
2934 struct vm_page_action *scan;
2935 struct vm_page_action *next;
2939 hv = (int)((intptr_t)m >> 8) & VMACTION_HMASK;
2940 list = &action_list[hv];
2943 lwkt_gettoken(&vm_token);
2944 LIST_FOREACH_MUTABLE(scan, list, entry, next) {
2946 if (scan->event == event) {
2947 scan->event = VMEVENT_NONE;
2948 LIST_REMOVE(scan, entry);
2949 scan->func(m, scan);
2957 vm_page_flag_clear(m, PG_ACTIONLIST);
2958 lwkt_reltoken(&vm_token);
2961 #include "opt_ddb.h"
2963 #include <sys/kernel.h>
2965 #include <ddb/ddb.h>
2967 DB_SHOW_COMMAND(page, vm_page_print_page_info)
2969 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count);
2970 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count);
2971 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count);
2972 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count);
2973 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count);
2974 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved);
2975 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min);
2976 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target);
2977 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min);
2978 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target);
2981 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
2984 db_printf("PQ_FREE:");
2985 for(i=0;i<PQ_L2_SIZE;i++) {
2986 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
2990 db_printf("PQ_CACHE:");
2991 for(i=0;i<PQ_L2_SIZE;i++) {
2992 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
2996 db_printf("PQ_ACTIVE:");
2997 for(i=0;i<PQ_L2_SIZE;i++) {
2998 db_printf(" %d", vm_page_queues[PQ_ACTIVE + i].lcnt);
3002 db_printf("PQ_INACTIVE:");
3003 for(i=0;i<PQ_L2_SIZE;i++) {
3004 db_printf(" %d", vm_page_queues[PQ_INACTIVE + i].lcnt);