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 static int vm_contig_verbose = 0;
128 TUNABLE_INT("vm.contig_verbose", &vm_contig_verbose);
130 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare,
131 vm_pindex_t, pindex);
134 vm_page_queue_init(void)
138 for (i = 0; i < PQ_L2_SIZE; i++)
139 vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count;
140 for (i = 0; i < PQ_L2_SIZE; i++)
141 vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count;
142 for (i = 0; i < PQ_L2_SIZE; i++)
143 vm_page_queues[PQ_INACTIVE+i].cnt = &vmstats.v_inactive_count;
144 for (i = 0; i < PQ_L2_SIZE; i++)
145 vm_page_queues[PQ_ACTIVE+i].cnt = &vmstats.v_active_count;
146 for (i = 0; i < PQ_L2_SIZE; i++)
147 vm_page_queues[PQ_HOLD+i].cnt = &vmstats.v_active_count;
148 /* PQ_NONE has no queue */
150 for (i = 0; i < PQ_COUNT; i++) {
151 TAILQ_INIT(&vm_page_queues[i].pl);
152 spin_init(&vm_page_queues[i].spin);
155 for (i = 0; i < VMACTION_HSIZE; i++)
156 LIST_INIT(&action_list[i]);
160 * note: place in initialized data section? Is this necessary?
163 int vm_page_array_size = 0;
164 int vm_page_zero_count = 0;
165 vm_page_t vm_page_array = NULL;
166 vm_paddr_t vm_low_phys_reserved;
171 * Sets the page size, perhaps based upon the memory size.
172 * Must be called before any use of page-size dependent functions.
175 vm_set_page_size(void)
177 if (vmstats.v_page_size == 0)
178 vmstats.v_page_size = PAGE_SIZE;
179 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0)
180 panic("vm_set_page_size: page size not a power of two");
186 * Add a new page to the freelist for use by the system. New pages
187 * are added to both the head and tail of the associated free page
188 * queue in a bottom-up fashion, so both zero'd and non-zero'd page
189 * requests pull 'recent' adds (higher physical addresses) first.
191 * Beware that the page zeroing daemon will also be running soon after
192 * boot, moving pages from the head to the tail of the PQ_FREE queues.
194 * Must be called in a critical section.
197 vm_add_new_page(vm_paddr_t pa)
199 struct vpgqueues *vpq;
202 m = PHYS_TO_VM_PAGE(pa);
205 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK;
207 * Twist for cpu localization in addition to page coloring, so
208 * different cpus selecting by m->queue get different page colors.
210 m->pc ^= ((pa >> PAGE_SHIFT) / PQ_L2_SIZE) & PQ_L2_MASK;
211 m->pc ^= ((pa >> PAGE_SHIFT) / (PQ_L2_SIZE * PQ_L2_SIZE)) & PQ_L2_MASK;
213 * Reserve a certain number of contiguous low memory pages for
214 * contigmalloc() to use.
216 if (pa < vm_low_phys_reserved) {
217 atomic_add_int(&vmstats.v_page_count, 1);
218 atomic_add_int(&vmstats.v_dma_pages, 1);
221 atomic_add_int(&vmstats.v_wire_count, 1);
222 alist_free(&vm_contig_alist, pa >> PAGE_SHIFT, 1);
229 m->queue = m->pc + PQ_FREE;
230 KKASSERT(m->dirty == 0);
232 atomic_add_int(&vmstats.v_page_count, 1);
233 atomic_add_int(&vmstats.v_free_count, 1);
234 vpq = &vm_page_queues[m->queue];
235 if ((vpq->flipflop & 15) == 0) {
236 pmap_zero_page(VM_PAGE_TO_PHYS(m));
238 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
239 atomic_add_int(&vm_page_zero_count, 1);
241 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
250 * Initializes the resident memory module.
252 * Preallocates memory for critical VM structures and arrays prior to
253 * kernel_map becoming available.
255 * Memory is allocated from (virtual2_start, virtual2_end) if available,
256 * otherwise memory is allocated from (virtual_start, virtual_end).
258 * On x86-64 (virtual_start, virtual_end) is only 2GB and may not be
259 * large enough to hold vm_page_array & other structures for machines with
260 * large amounts of ram, so we want to use virtual2* when available.
263 vm_page_startup(void)
265 vm_offset_t vaddr = virtual2_start ? virtual2_start : virtual_start;
268 vm_paddr_t page_range;
275 vm_paddr_t biggestone, biggestsize;
282 vaddr = round_page(vaddr);
284 for (i = 0; phys_avail[i + 1]; i += 2) {
285 phys_avail[i] = round_page64(phys_avail[i]);
286 phys_avail[i + 1] = trunc_page64(phys_avail[i + 1]);
289 for (i = 0; phys_avail[i + 1]; i += 2) {
290 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
292 if (size > biggestsize) {
300 end = phys_avail[biggestone+1];
301 end = trunc_page(end);
304 * Initialize the queue headers for the free queue, the active queue
305 * and the inactive queue.
307 vm_page_queue_init();
309 #if !defined(_KERNEL_VIRTUAL)
311 * VKERNELs don't support minidumps and as such don't need
314 * Allocate a bitmap to indicate that a random physical page
315 * needs to be included in a minidump.
317 * The amd64 port needs this to indicate which direct map pages
318 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
320 * However, i386 still needs this workspace internally within the
321 * minidump code. In theory, they are not needed on i386, but are
322 * included should the sf_buf code decide to use them.
324 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE;
325 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
326 end -= vm_page_dump_size;
327 vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size,
328 VM_PROT_READ | VM_PROT_WRITE);
329 bzero((void *)vm_page_dump, vm_page_dump_size);
332 * Compute the number of pages of memory that will be available for
333 * use (taking into account the overhead of a page structure per
336 first_page = phys_avail[0] / PAGE_SIZE;
337 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
338 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE;
340 #ifndef _KERNEL_VIRTUAL
342 * (only applies to real kernels)
344 * Initialize the contiguous reserve map. We initially reserve up
345 * to 1/4 available physical memory or 65536 pages (~256MB), whichever
348 * Once device initialization is complete we return most of the
349 * reserved memory back to the normal page queues but leave some
350 * in reserve for things like usb attachments.
352 vm_low_phys_reserved = (vm_paddr_t)65536 << PAGE_SHIFT;
353 if (vm_low_phys_reserved > total / 4)
354 vm_low_phys_reserved = total / 4;
355 if (vm_dma_reserved == 0) {
356 vm_dma_reserved = 16 * 1024 * 1024; /* 16MB */
357 if (vm_dma_reserved > total / 16)
358 vm_dma_reserved = total / 16;
361 alist_init(&vm_contig_alist, 65536, vm_contig_ameta,
362 ALIST_RECORDS_65536);
365 * Initialize the mem entry structures now, and put them in the free
368 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
369 mapped = pmap_map(&vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE);
370 vm_page_array = (vm_page_t)mapped;
372 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL)
374 * since pmap_map on amd64 returns stuff out of a direct-map region,
375 * we have to manually add these pages to the minidump tracking so
376 * that they can be dumped, including the vm_page_array.
378 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
383 * Clear all of the page structures
385 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
386 vm_page_array_size = page_range;
389 * Construct the free queue(s) in ascending order (by physical
390 * address) so that the first 16MB of physical memory is allocated
391 * last rather than first. On large-memory machines, this avoids
392 * the exhaustion of low physical memory before isa_dmainit has run.
394 vmstats.v_page_count = 0;
395 vmstats.v_free_count = 0;
396 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
401 last_pa = phys_avail[i + 1];
402 while (pa < last_pa && npages-- > 0) {
408 virtual2_start = vaddr;
410 virtual_start = vaddr;
414 * We tended to reserve a ton of memory for contigmalloc(). Now that most
415 * drivers have initialized we want to return most the remaining free
416 * reserve back to the VM page queues so they can be used for normal
419 * We leave vm_dma_reserved bytes worth of free pages in the reserve pool.
422 vm_page_startup_finish(void *dummy __unused)
431 spin_lock(&vm_contig_spin);
433 bfree = alist_free_info(&vm_contig_alist, &blk, &count);
434 if (bfree <= vm_dma_reserved / PAGE_SIZE)
440 * Figure out how much of the initial reserve we have to
441 * free in order to reach our target.
443 bfree -= vm_dma_reserved / PAGE_SIZE;
445 blk += count - bfree;
450 * Calculate the nearest power of 2 <= count.
452 for (xcount = 1; xcount <= count; xcount <<= 1)
455 blk += count - xcount;
459 * Allocate the pages from the alist, then free them to
460 * the normal VM page queues.
462 * Pages allocated from the alist are wired. We have to
463 * busy, unwire, and free them. We must also adjust
464 * vm_low_phys_reserved before freeing any pages to prevent
467 rblk = alist_alloc(&vm_contig_alist, blk, count);
469 kprintf("vm_page_startup_finish: Unable to return "
470 "dma space @0x%08x/%d -> 0x%08x\n",
474 atomic_add_int(&vmstats.v_dma_pages, -count);
475 spin_unlock(&vm_contig_spin);
477 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT);
478 vm_low_phys_reserved = VM_PAGE_TO_PHYS(m);
480 vm_page_busy_wait(m, FALSE, "cpgfr");
481 vm_page_unwire(m, 0);
486 spin_lock(&vm_contig_spin);
488 spin_unlock(&vm_contig_spin);
491 * Print out how much DMA space drivers have already allocated and
492 * how much is left over.
494 kprintf("DMA space used: %jdk, remaining available: %jdk\n",
495 (intmax_t)(vmstats.v_dma_pages - vm_contig_alist.bl_free) *
497 (intmax_t)vm_contig_alist.bl_free * (PAGE_SIZE / 1024));
499 SYSINIT(vm_pgend, SI_SUB_PROC0_POST, SI_ORDER_ANY,
500 vm_page_startup_finish, NULL)
504 * Scan comparison function for Red-Black tree scans. An inclusive
505 * (start,end) is expected. Other fields are not used.
508 rb_vm_page_scancmp(struct vm_page *p, void *data)
510 struct rb_vm_page_scan_info *info = data;
512 if (p->pindex < info->start_pindex)
514 if (p->pindex > info->end_pindex)
520 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2)
522 if (p1->pindex < p2->pindex)
524 if (p1->pindex > p2->pindex)
530 * Each page queue has its own spin lock, which is fairly optimal for
531 * allocating and freeing pages at least.
533 * The caller must hold the vm_page_spin_lock() before locking a vm_page's
534 * queue spinlock via this function. Also note that m->queue cannot change
535 * unless both the page and queue are locked.
539 _vm_page_queue_spin_lock(vm_page_t m)
544 if (queue != PQ_NONE) {
545 spin_lock(&vm_page_queues[queue].spin);
546 KKASSERT(queue == m->queue);
552 _vm_page_queue_spin_unlock(vm_page_t m)
558 if (queue != PQ_NONE)
559 spin_unlock(&vm_page_queues[queue].spin);
564 _vm_page_queues_spin_lock(u_short queue)
567 if (queue != PQ_NONE)
568 spin_lock(&vm_page_queues[queue].spin);
574 _vm_page_queues_spin_unlock(u_short queue)
577 if (queue != PQ_NONE)
578 spin_unlock(&vm_page_queues[queue].spin);
582 vm_page_queue_spin_lock(vm_page_t m)
584 _vm_page_queue_spin_lock(m);
588 vm_page_queues_spin_lock(u_short queue)
590 _vm_page_queues_spin_lock(queue);
594 vm_page_queue_spin_unlock(vm_page_t m)
596 _vm_page_queue_spin_unlock(m);
600 vm_page_queues_spin_unlock(u_short queue)
602 _vm_page_queues_spin_unlock(queue);
606 * This locks the specified vm_page and its queue in the proper order
607 * (page first, then queue). The queue may change so the caller must
612 _vm_page_and_queue_spin_lock(vm_page_t m)
614 vm_page_spin_lock(m);
615 _vm_page_queue_spin_lock(m);
620 _vm_page_and_queue_spin_unlock(vm_page_t m)
622 _vm_page_queues_spin_unlock(m->queue);
623 vm_page_spin_unlock(m);
627 vm_page_and_queue_spin_unlock(vm_page_t m)
629 _vm_page_and_queue_spin_unlock(m);
633 vm_page_and_queue_spin_lock(vm_page_t m)
635 _vm_page_and_queue_spin_lock(m);
639 * Helper function removes vm_page from its current queue.
640 * Returns the base queue the page used to be on.
642 * The vm_page and the queue must be spinlocked.
643 * This function will unlock the queue but leave the page spinlocked.
645 static __inline u_short
646 _vm_page_rem_queue_spinlocked(vm_page_t m)
648 struct vpgqueues *pq;
652 if (queue != PQ_NONE) {
653 pq = &vm_page_queues[queue];
654 TAILQ_REMOVE(&pq->pl, m, pageq);
655 atomic_add_int(pq->cnt, -1);
658 vm_page_queues_spin_unlock(queue);
659 if ((queue - m->pc) == PQ_FREE && (m->flags & PG_ZERO))
660 atomic_subtract_int(&vm_page_zero_count, 1);
661 if ((queue - m->pc) == PQ_CACHE || (queue - m->pc) == PQ_FREE)
662 return (queue - m->pc);
668 * Helper function places the vm_page on the specified queue.
670 * The vm_page must be spinlocked.
671 * This function will return with both the page and the queue locked.
674 _vm_page_add_queue_spinlocked(vm_page_t m, u_short queue, int athead)
676 struct vpgqueues *pq;
678 KKASSERT(m->queue == PQ_NONE);
680 if (queue != PQ_NONE) {
681 vm_page_queues_spin_lock(queue);
682 pq = &vm_page_queues[queue];
684 atomic_add_int(pq->cnt, 1);
688 * Put zero'd pages on the end ( where we look for zero'd pages
689 * first ) and non-zerod pages at the head.
691 if (queue - m->pc == PQ_FREE) {
692 if (m->flags & PG_ZERO) {
693 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
694 atomic_add_int(&vm_page_zero_count, 1);
696 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
699 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
701 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
703 /* leave the queue spinlocked */
708 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE)
709 * m->busy is zero. Returns TRUE if it had to sleep, FALSE if we
710 * did not. Only one sleep call will be made before returning.
712 * This function does NOT busy the page and on return the page is not
713 * guaranteed to be available.
716 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg)
724 if ((flags & PG_BUSY) == 0 &&
725 (also_m_busy == 0 || (flags & PG_SBUSY) == 0)) {
728 tsleep_interlock(m, 0);
729 if (atomic_cmpset_int(&m->flags, flags,
730 flags | PG_WANTED | PG_REFERENCED)) {
731 tsleep(m, PINTERLOCKED, msg, 0);
738 * Wait until PG_BUSY can be set, then set it. If also_m_busy is TRUE we
739 * also wait for m->busy to become 0 before setting PG_BUSY.
742 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m,
743 int also_m_busy, const char *msg
751 if (flags & PG_BUSY) {
752 tsleep_interlock(m, 0);
753 if (atomic_cmpset_int(&m->flags, flags,
754 flags | PG_WANTED | PG_REFERENCED)) {
755 tsleep(m, PINTERLOCKED, msg, 0);
757 } else if (also_m_busy && (flags & PG_SBUSY)) {
758 tsleep_interlock(m, 0);
759 if (atomic_cmpset_int(&m->flags, flags,
760 flags | PG_WANTED | PG_REFERENCED)) {
761 tsleep(m, PINTERLOCKED, msg, 0);
764 if (atomic_cmpset_int(&m->flags, flags,
768 m->busy_line = lineno;
777 * Attempt to set PG_BUSY. If also_m_busy is TRUE we only succeed if m->busy
780 * Returns non-zero on failure.
783 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy
793 if (also_m_busy && (flags & PG_SBUSY))
795 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
798 m->busy_line = lineno;
806 * Clear the PG_BUSY flag and return non-zero to indicate to the caller
807 * that a wakeup() should be performed.
809 * The vm_page must be spinlocked and will remain spinlocked on return.
810 * The related queue must NOT be spinlocked (which could deadlock us).
816 _vm_page_wakeup(vm_page_t m)
823 if (atomic_cmpset_int(&m->flags, flags,
824 flags & ~(PG_BUSY | PG_WANTED))) {
828 return(flags & PG_WANTED);
832 * Clear the PG_BUSY flag and wakeup anyone waiting for the page. This
833 * is typically the last call you make on a page before moving onto
837 vm_page_wakeup(vm_page_t m)
839 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!"));
840 vm_page_spin_lock(m);
841 if (_vm_page_wakeup(m)) {
842 vm_page_spin_unlock(m);
845 vm_page_spin_unlock(m);
850 * Holding a page keeps it from being reused. Other parts of the system
851 * can still disassociate the page from its current object and free it, or
852 * perform read or write I/O on it and/or otherwise manipulate the page,
853 * but if the page is held the VM system will leave the page and its data
854 * intact and not reuse the page for other purposes until the last hold
855 * reference is released. (see vm_page_wire() if you want to prevent the
856 * page from being disassociated from its object too).
858 * The caller must still validate the contents of the page and, if necessary,
859 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete
860 * before manipulating the page.
862 * XXX get vm_page_spin_lock() here and move FREE->HOLD if necessary
865 vm_page_hold(vm_page_t m)
867 vm_page_spin_lock(m);
868 atomic_add_int(&m->hold_count, 1);
869 if (m->queue - m->pc == PQ_FREE) {
870 _vm_page_queue_spin_lock(m);
871 _vm_page_rem_queue_spinlocked(m);
872 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
873 _vm_page_queue_spin_unlock(m);
875 vm_page_spin_unlock(m);
879 * The opposite of vm_page_hold(). A page can be freed while being held,
880 * which places it on the PQ_HOLD queue. If we are able to busy the page
881 * after the hold count drops to zero we will move the page to the
882 * appropriate PQ_FREE queue by calling vm_page_free_toq().
885 vm_page_unhold(vm_page_t m)
887 vm_page_spin_lock(m);
888 atomic_add_int(&m->hold_count, -1);
889 if (m->hold_count == 0 && m->queue - m->pc == PQ_HOLD) {
890 _vm_page_queue_spin_lock(m);
891 _vm_page_rem_queue_spinlocked(m);
892 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0);
893 _vm_page_queue_spin_unlock(m);
895 vm_page_spin_unlock(m);
899 * Inserts the given vm_page into the object and object list.
901 * The pagetables are not updated but will presumably fault the page
902 * in if necessary, or if a kernel page the caller will at some point
903 * enter the page into the kernel's pmap. We are not allowed to block
904 * here so we *can't* do this anyway.
906 * This routine may not block.
907 * This routine must be called with the vm_object held.
908 * This routine must be called with a critical section held.
910 * This routine returns TRUE if the page was inserted into the object
911 * successfully, and FALSE if the page already exists in the object.
914 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
916 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
917 if (m->object != NULL)
918 panic("vm_page_insert: already inserted");
920 object->generation++;
923 * Record the object/offset pair in this page and add the
924 * pv_list_count of the page to the object.
926 * The vm_page spin lock is required for interactions with the pmap.
928 vm_page_spin_lock(m);
931 if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) {
934 vm_page_spin_unlock(m);
937 object->resident_page_count++;
938 /* atomic_add_int(&object->agg_pv_list_count, m->md.pv_list_count); */
939 vm_page_spin_unlock(m);
942 * Since we are inserting a new and possibly dirty page,
943 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
945 if ((m->valid & m->dirty) || (m->flags & PG_WRITEABLE))
946 vm_object_set_writeable_dirty(object);
949 * Checks for a swap assignment and sets PG_SWAPPED if appropriate.
951 swap_pager_page_inserted(m);
956 * Removes the given vm_page_t from the (object,index) table
958 * The underlying pmap entry (if any) is NOT removed here.
959 * This routine may not block.
961 * The page must be BUSY and will remain BUSY on return.
962 * No other requirements.
964 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave
968 vm_page_remove(vm_page_t m)
972 if (m->object == NULL) {
976 if ((m->flags & PG_BUSY) == 0)
977 panic("vm_page_remove: page not busy");
981 vm_object_hold(object);
984 * Remove the page from the object and update the object.
986 * The vm_page spin lock is required for interactions with the pmap.
988 vm_page_spin_lock(m);
989 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m);
990 object->resident_page_count--;
991 /* atomic_add_int(&object->agg_pv_list_count, -m->md.pv_list_count); */
993 vm_page_spin_unlock(m);
995 object->generation++;
997 vm_object_drop(object);
1001 * Locate and return the page at (object, pindex), or NULL if the
1002 * page could not be found.
1004 * The caller must hold the vm_object token.
1007 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1012 * Search the hash table for this object/offset pair
1014 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1015 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1016 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex));
1021 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object,
1023 int also_m_busy, const char *msg
1029 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1030 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1032 KKASSERT(m->object == object && m->pindex == pindex);
1035 if (flags & PG_BUSY) {
1036 tsleep_interlock(m, 0);
1037 if (atomic_cmpset_int(&m->flags, flags,
1038 flags | PG_WANTED | PG_REFERENCED)) {
1039 tsleep(m, PINTERLOCKED, msg, 0);
1040 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1043 } else if (also_m_busy && (flags & PG_SBUSY)) {
1044 tsleep_interlock(m, 0);
1045 if (atomic_cmpset_int(&m->flags, flags,
1046 flags | PG_WANTED | PG_REFERENCED)) {
1047 tsleep(m, PINTERLOCKED, msg, 0);
1048 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1051 } else if (atomic_cmpset_int(&m->flags, flags,
1053 #ifdef VM_PAGE_DEBUG
1054 m->busy_func = func;
1055 m->busy_line = lineno;
1064 * Attempt to lookup and busy a page.
1066 * Returns NULL if the page could not be found
1068 * Returns a vm_page and error == TRUE if the page exists but could not
1071 * Returns a vm_page and error == FALSE on success.
1074 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object,
1076 int also_m_busy, int *errorp
1082 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1083 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1086 KKASSERT(m->object == object && m->pindex == pindex);
1089 if (flags & PG_BUSY) {
1093 if (also_m_busy && (flags & PG_SBUSY)) {
1097 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
1098 #ifdef VM_PAGE_DEBUG
1099 m->busy_func = func;
1100 m->busy_line = lineno;
1109 * Caller must hold the related vm_object
1112 vm_page_next(vm_page_t m)
1116 next = vm_page_rb_tree_RB_NEXT(m);
1117 if (next && next->pindex != m->pindex + 1)
1125 * Move the given vm_page from its current object to the specified
1126 * target object/offset. The page must be busy and will remain so
1129 * new_object must be held.
1130 * This routine might block. XXX ?
1132 * NOTE: Swap associated with the page must be invalidated by the move. We
1133 * have to do this for several reasons: (1) we aren't freeing the
1134 * page, (2) we are dirtying the page, (3) the VM system is probably
1135 * moving the page from object A to B, and will then later move
1136 * the backing store from A to B and we can't have a conflict.
1138 * NOTE: We *always* dirty the page. It is necessary both for the
1139 * fact that we moved it, and because we may be invalidating
1140 * swap. If the page is on the cache, we have to deactivate it
1141 * or vm_page_dirty() will panic. Dirty pages are not allowed
1145 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1147 KKASSERT(m->flags & PG_BUSY);
1148 ASSERT_LWKT_TOKEN_HELD(vm_object_token(new_object));
1150 ASSERT_LWKT_TOKEN_HELD(vm_object_token(m->object));
1153 if (vm_page_insert(m, new_object, new_pindex) == FALSE) {
1154 panic("vm_page_rename: target exists (%p,%"PRIu64")",
1155 new_object, new_pindex);
1157 if (m->queue - m->pc == PQ_CACHE)
1158 vm_page_deactivate(m);
1163 * vm_page_unqueue() without any wakeup. This routine is used when a page
1164 * is being moved between queues or otherwise is to remain BUSYied by the
1167 * This routine may not block.
1170 vm_page_unqueue_nowakeup(vm_page_t m)
1172 vm_page_and_queue_spin_lock(m);
1173 (void)_vm_page_rem_queue_spinlocked(m);
1174 vm_page_spin_unlock(m);
1178 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
1181 * This routine may not block.
1184 vm_page_unqueue(vm_page_t m)
1188 vm_page_and_queue_spin_lock(m);
1189 queue = _vm_page_rem_queue_spinlocked(m);
1190 if (queue == PQ_FREE || queue == PQ_CACHE) {
1191 vm_page_spin_unlock(m);
1192 pagedaemon_wakeup();
1194 vm_page_spin_unlock(m);
1199 * vm_page_list_find()
1201 * Find a page on the specified queue with color optimization.
1203 * The page coloring optimization attempts to locate a page that does
1204 * not overload other nearby pages in the object in the cpu's L1 or L2
1205 * caches. We need this optimization because cpu caches tend to be
1206 * physical caches, while object spaces tend to be virtual.
1208 * On MP systems each PQ_FREE and PQ_CACHE color queue has its own spinlock
1209 * and the algorithm is adjusted to localize allocations on a per-core basis.
1210 * This is done by 'twisting' the colors.
1212 * The page is returned spinlocked and removed from its queue (it will
1213 * be on PQ_NONE), or NULL. The page is not PG_BUSY'd. The caller
1214 * is responsible for dealing with the busy-page case (usually by
1215 * deactivating the page and looping).
1217 * NOTE: This routine is carefully inlined. A non-inlined version
1218 * is available for outside callers but the only critical path is
1219 * from within this source file.
1221 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE
1222 * represent stable storage, allowing us to order our locks vm_page
1223 * first, then queue.
1227 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
1233 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist);
1235 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl);
1237 m = _vm_page_list_find2(basequeue, index);
1240 vm_page_and_queue_spin_lock(m);
1241 if (m->queue == basequeue + index) {
1242 _vm_page_rem_queue_spinlocked(m);
1243 /* vm_page_t spin held, no queue spin */
1246 vm_page_and_queue_spin_unlock(m);
1252 _vm_page_list_find2(int basequeue, int index)
1256 struct vpgqueues *pq;
1258 pq = &vm_page_queues[basequeue];
1261 * Note that for the first loop, index+i and index-i wind up at the
1262 * same place. Even though this is not totally optimal, we've already
1263 * blown it by missing the cache case so we do not care.
1265 for (i = PQ_L2_SIZE / 2; i > 0; --i) {
1267 m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl);
1269 _vm_page_and_queue_spin_lock(m);
1271 basequeue + ((index + i) & PQ_L2_MASK)) {
1272 _vm_page_rem_queue_spinlocked(m);
1275 _vm_page_and_queue_spin_unlock(m);
1278 m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl);
1280 _vm_page_and_queue_spin_lock(m);
1282 basequeue + ((index - i) & PQ_L2_MASK)) {
1283 _vm_page_rem_queue_spinlocked(m);
1286 _vm_page_and_queue_spin_unlock(m);
1296 * Returns a vm_page candidate for allocation. The page is not busied so
1297 * it can move around. The caller must busy the page (and typically
1298 * deactivate it if it cannot be busied!)
1300 * Returns a spinlocked vm_page that has been removed from its queue.
1303 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
1305 return(_vm_page_list_find(basequeue, index, prefer_zero));
1309 * Find a page on the cache queue with color optimization, remove it
1310 * from the queue, and busy it. The returned page will not be spinlocked.
1312 * A candidate failure will be deactivated. Candidates can fail due to
1313 * being busied by someone else, in which case they will be deactivated.
1315 * This routine may not block.
1319 vm_page_select_cache(u_short pg_color)
1324 m = _vm_page_list_find(PQ_CACHE, pg_color & PQ_L2_MASK, FALSE);
1328 * (m) has been removed from its queue and spinlocked
1330 if (vm_page_busy_try(m, TRUE)) {
1331 _vm_page_deactivate_locked(m, 0);
1332 vm_page_spin_unlock(m);
1334 kprintf("Warning: busy page %p found in cache\n", 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);
1402 kprintf("Warning: busy page %p found in cache\n", m);
1406 * Theoretically if we are able to busy the page
1407 * atomic with the queue removal (using the vm_page
1408 * lock) nobody else should be able to mess with the
1411 KKASSERT((m->flags & (PG_UNMANAGED |
1412 PG_NEED_COMMIT)) == 0);
1413 KKASSERT(m->hold_count == 0);
1414 KKASSERT(m->wire_count == 0);
1415 vm_page_spin_unlock(m);
1416 pagedaemon_wakeup();
1418 /* return busied and removed page */
1426 * This implements a per-cpu cache of free, zero'd, ready-to-go pages.
1427 * The idea is to populate this cache prior to acquiring any locks so
1428 * we don't wind up potentially zeroing VM pages (under heavy loads) while
1429 * holding potentialy contending locks.
1431 * Note that we allocate the page uninserted into anything and use a pindex
1432 * of 0, the vm_page_alloc() will effectively add gd_cpuid so these
1433 * allocations should wind up being uncontended. However, we still want
1434 * to rove across PQ_L2_SIZE.
1437 vm_page_pcpu_cache(void)
1440 globaldata_t gd = mycpu;
1443 if (gd->gd_vmpg_count < GD_MINVMPG) {
1445 while (gd->gd_vmpg_count < GD_MAXVMPG) {
1446 m = vm_page_alloc(NULL, ticks & ~ncpus2_mask,
1447 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL |
1448 VM_ALLOC_NULL_OK | VM_ALLOC_ZERO);
1449 if (gd->gd_vmpg_count < GD_MAXVMPG) {
1450 if ((m->flags & PG_ZERO) == 0) {
1451 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1452 vm_page_flag_set(m, PG_ZERO);
1454 gd->gd_vmpg_array[gd->gd_vmpg_count++] = m;
1467 * Allocate and return a memory cell associated with this VM object/offset
1468 * pair. If object is NULL an unassociated page will be allocated.
1470 * The returned page will be busied and removed from its queues. This
1471 * routine can block and may return NULL if a race occurs and the page
1472 * is found to already exist at the specified (object, pindex).
1474 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain
1475 * VM_ALLOC_QUICK like normal but cannot use cache
1476 * VM_ALLOC_SYSTEM greater free drain
1477 * VM_ALLOC_INTERRUPT allow free list to be completely drained
1478 * VM_ALLOC_ZERO advisory request for pre-zero'd page only
1479 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only
1480 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision
1481 * (see vm_page_grab())
1482 * VM_ALLOC_USE_GD ok to use per-gd cache
1484 * The object must be held if not NULL
1485 * This routine may not block
1487 * Additional special handling is required when called from an interrupt
1488 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache
1492 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
1494 globaldata_t gd = mycpu;
1501 * Special per-cpu free VM page cache. The pages are pre-busied
1502 * and pre-zerod for us.
1504 if (gd->gd_vmpg_count && (page_req & VM_ALLOC_USE_GD)) {
1506 if (gd->gd_vmpg_count) {
1507 m = gd->gd_vmpg_array[--gd->gd_vmpg_count];
1517 * Cpu twist - cpu localization algorithm
1520 pg_color = gd->gd_cpuid + (pindex & ~ncpus_fit_mask) +
1521 (object->pg_color & ~ncpus_fit_mask);
1523 pg_color = gd->gd_cpuid + (pindex & ~ncpus_fit_mask);
1526 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK|
1527 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
1530 * Certain system threads (pageout daemon, buf_daemon's) are
1531 * allowed to eat deeper into the free page list.
1533 if (curthread->td_flags & TDF_SYSTHREAD)
1534 page_req |= VM_ALLOC_SYSTEM;
1537 if (vmstats.v_free_count > vmstats.v_free_reserved ||
1538 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) ||
1539 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 &&
1540 vmstats.v_free_count > vmstats.v_interrupt_free_min)
1543 * The free queue has sufficient free pages to take one out.
1545 if (page_req & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO))
1546 m = vm_page_select_free(pg_color, TRUE);
1548 m = vm_page_select_free(pg_color, FALSE);
1549 } else if (page_req & VM_ALLOC_NORMAL) {
1551 * Allocatable from the cache (non-interrupt only). On
1552 * success, we must free the page and try again, thus
1553 * ensuring that vmstats.v_*_free_min counters are replenished.
1556 if (curthread->td_preempted) {
1557 kprintf("vm_page_alloc(): warning, attempt to allocate"
1558 " cache page from preempting interrupt\n");
1561 m = vm_page_select_cache(pg_color);
1564 m = vm_page_select_cache(pg_color);
1567 * On success move the page into the free queue and loop.
1569 * Only do this if we can safely acquire the vm_object lock,
1570 * because this is effectively a random page and the caller
1571 * might be holding the lock shared, we don't want to
1575 KASSERT(m->dirty == 0,
1576 ("Found dirty cache page %p", m));
1577 if ((obj = m->object) != NULL) {
1578 if (vm_object_hold_try(obj)) {
1579 vm_page_protect(m, VM_PROT_NONE);
1581 /* m->object NULL here */
1582 vm_object_drop(obj);
1584 vm_page_deactivate(m);
1588 vm_page_protect(m, VM_PROT_NONE);
1595 * On failure return NULL
1597 #if defined(DIAGNOSTIC)
1598 if (vmstats.v_cache_count > 0)
1599 kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count);
1601 vm_pageout_deficit++;
1602 pagedaemon_wakeup();
1606 * No pages available, wakeup the pageout daemon and give up.
1608 vm_pageout_deficit++;
1609 pagedaemon_wakeup();
1614 * v_free_count can race so loop if we don't find the expected
1621 * Good page found. The page has already been busied for us and
1622 * removed from its queues.
1624 KASSERT(m->dirty == 0,
1625 ("vm_page_alloc: free/cache page %p was dirty", m));
1626 KKASSERT(m->queue == PQ_NONE);
1632 * Initialize the structure, inheriting some flags but clearing
1633 * all the rest. The page has already been busied for us.
1635 vm_page_flag_clear(m, ~(PG_ZERO | PG_BUSY | PG_SBUSY));
1636 KKASSERT(m->wire_count == 0);
1637 KKASSERT(m->busy == 0);
1642 * Caller must be holding the object lock (asserted by
1643 * vm_page_insert()).
1645 * NOTE: Inserting a page here does not insert it into any pmaps
1646 * (which could cause us to block allocating memory).
1648 * NOTE: If no object an unassociated page is allocated, m->pindex
1649 * can be used by the caller for any purpose.
1652 if (vm_page_insert(m, object, pindex) == FALSE) {
1653 kprintf("PAGE RACE (%p:%d,%"PRIu64")\n",
1654 object, object->type, pindex);
1657 if ((page_req & VM_ALLOC_NULL_OK) == 0)
1665 * Don't wakeup too often - wakeup the pageout daemon when
1666 * we would be nearly out of memory.
1668 pagedaemon_wakeup();
1671 * A PG_BUSY page is returned.
1677 * Attempt to allocate contiguous physical memory with the specified
1681 vm_page_alloc_contig(vm_paddr_t low, vm_paddr_t high,
1682 unsigned long alignment, unsigned long boundary,
1687 alignment >>= PAGE_SHIFT;
1690 boundary >>= PAGE_SHIFT;
1693 size = (size + PAGE_MASK) >> PAGE_SHIFT;
1695 spin_lock(&vm_contig_spin);
1696 blk = alist_alloc(&vm_contig_alist, 0, size);
1697 if (blk == ALIST_BLOCK_NONE) {
1698 spin_unlock(&vm_contig_spin);
1700 kprintf("vm_page_alloc_contig: %ldk nospace\n",
1701 (size + PAGE_MASK) * (PAGE_SIZE / 1024));
1705 if (high && ((vm_paddr_t)(blk + size) << PAGE_SHIFT) > high) {
1706 alist_free(&vm_contig_alist, blk, size);
1707 spin_unlock(&vm_contig_spin);
1709 kprintf("vm_page_alloc_contig: %ldk high "
1711 (size + PAGE_MASK) * (PAGE_SIZE / 1024),
1716 spin_unlock(&vm_contig_spin);
1717 if (vm_contig_verbose) {
1718 kprintf("vm_page_alloc_contig: %016jx/%ldk\n",
1719 (intmax_t)(vm_paddr_t)blk << PAGE_SHIFT,
1720 (size + PAGE_MASK) * (PAGE_SIZE / 1024));
1722 return (PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT));
1726 * Free contiguously allocated pages. The pages will be wired but not busy.
1727 * When freeing to the alist we leave them wired and not busy.
1730 vm_page_free_contig(vm_page_t m, unsigned long size)
1732 vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1733 vm_pindex_t start = pa >> PAGE_SHIFT;
1734 vm_pindex_t pages = (size + PAGE_MASK) >> PAGE_SHIFT;
1736 if (vm_contig_verbose) {
1737 kprintf("vm_page_free_contig: %016jx/%ldk\n",
1738 (intmax_t)pa, size / 1024);
1740 if (pa < vm_low_phys_reserved) {
1741 KKASSERT(pa + size <= vm_low_phys_reserved);
1742 spin_lock(&vm_contig_spin);
1743 alist_free(&vm_contig_alist, start, pages);
1744 spin_unlock(&vm_contig_spin);
1747 vm_page_busy_wait(m, FALSE, "cpgfr");
1748 vm_page_unwire(m, 0);
1759 * Wait for sufficient free memory for nominal heavy memory use kernel
1762 * WARNING! Be sure never to call this in any vm_pageout code path, which
1763 * will trivially deadlock the system.
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.
1996 * Clear the NEED_COMMIT flag
1998 if (m->flags & PG_UNMANAGED)
1999 vm_page_flag_clear(m, PG_UNMANAGED);
2000 if (m->flags & PG_NEED_COMMIT)
2001 vm_page_flag_clear(m, PG_NEED_COMMIT);
2003 if (m->hold_count != 0) {
2004 vm_page_flag_clear(m, PG_ZERO);
2005 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
2007 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0);
2011 * This sequence allows us to clear PG_BUSY while still holding
2012 * its spin lock, which reduces contention vs allocators. We
2013 * must not leave the queue locked or _vm_page_wakeup() may
2016 _vm_page_queue_spin_unlock(m);
2017 if (_vm_page_wakeup(m)) {
2018 vm_page_spin_unlock(m);
2021 vm_page_spin_unlock(m);
2023 vm_page_free_wakeup();
2027 * vm_page_free_fromq_fast()
2029 * Remove a non-zero page from one of the free queues; the page is removed for
2030 * zeroing, so do not issue a wakeup.
2033 vm_page_free_fromq_fast(void)
2039 for (i = 0; i < PQ_L2_SIZE; ++i) {
2040 m = vm_page_list_find(PQ_FREE, qi, FALSE);
2041 /* page is returned spinlocked and removed from its queue */
2043 if (vm_page_busy_try(m, TRUE)) {
2045 * We were unable to busy the page, deactivate
2048 _vm_page_deactivate_locked(m, 0);
2049 vm_page_spin_unlock(m);
2050 } else if (m->flags & PG_ZERO) {
2052 * The page is PG_ZERO, requeue it and loop
2054 _vm_page_add_queue_spinlocked(m,
2057 vm_page_queue_spin_unlock(m);
2058 if (_vm_page_wakeup(m)) {
2059 vm_page_spin_unlock(m);
2062 vm_page_spin_unlock(m);
2066 * The page is not PG_ZERO'd so return it.
2068 vm_page_spin_unlock(m);
2069 KKASSERT((m->flags & (PG_UNMANAGED |
2070 PG_NEED_COMMIT)) == 0);
2071 KKASSERT(m->hold_count == 0);
2072 KKASSERT(m->wire_count == 0);
2077 qi = (qi + PQ_PRIME2) & PQ_L2_MASK;
2083 * vm_page_unmanage()
2085 * Prevent PV management from being done on the page. The page is
2086 * removed from the paging queues as if it were wired, and as a
2087 * consequence of no longer being managed the pageout daemon will not
2088 * touch it (since there is no way to locate the pte mappings for the
2089 * page). madvise() calls that mess with the pmap will also no longer
2090 * operate on the page.
2092 * Beyond that the page is still reasonably 'normal'. Freeing the page
2093 * will clear the flag.
2095 * This routine is used by OBJT_PHYS objects - objects using unswappable
2096 * physical memory as backing store rather then swap-backed memory and
2097 * will eventually be extended to support 4MB unmanaged physical
2100 * Caller must be holding the page busy.
2103 vm_page_unmanage(vm_page_t m)
2105 KKASSERT(m->flags & PG_BUSY);
2106 if ((m->flags & PG_UNMANAGED) == 0) {
2107 if (m->wire_count == 0)
2110 vm_page_flag_set(m, PG_UNMANAGED);
2114 * Mark this page as wired down by yet another map, removing it from
2115 * paging queues as necessary.
2117 * Caller must be holding the page busy.
2120 vm_page_wire(vm_page_t m)
2123 * Only bump the wire statistics if the page is not already wired,
2124 * and only unqueue the page if it is on some queue (if it is unmanaged
2125 * it is already off the queues). Don't do anything with fictitious
2126 * pages because they are always wired.
2128 KKASSERT(m->flags & PG_BUSY);
2129 if ((m->flags & PG_FICTITIOUS) == 0) {
2130 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) {
2131 if ((m->flags & PG_UNMANAGED) == 0)
2133 atomic_add_int(&vmstats.v_wire_count, 1);
2135 KASSERT(m->wire_count != 0,
2136 ("vm_page_wire: wire_count overflow m=%p", m));
2141 * Release one wiring of this page, potentially enabling it to be paged again.
2143 * Many pages placed on the inactive queue should actually go
2144 * into the cache, but it is difficult to figure out which. What
2145 * we do instead, if the inactive target is well met, is to put
2146 * clean pages at the head of the inactive queue instead of the tail.
2147 * This will cause them to be moved to the cache more quickly and
2148 * if not actively re-referenced, freed more quickly. If we just
2149 * stick these pages at the end of the inactive queue, heavy filesystem
2150 * meta-data accesses can cause an unnecessary paging load on memory bound
2151 * processes. This optimization causes one-time-use metadata to be
2152 * reused more quickly.
2154 * Pages marked PG_NEED_COMMIT are always activated and never placed on
2155 * the inactive queue. This helps the pageout daemon determine memory
2156 * pressure and act on out-of-memory situations more quickly.
2158 * BUT, if we are in a low-memory situation we have no choice but to
2159 * put clean pages on the cache queue.
2161 * A number of routines use vm_page_unwire() to guarantee that the page
2162 * will go into either the inactive or active queues, and will NEVER
2163 * be placed in the cache - for example, just after dirtying a page.
2164 * dirty pages in the cache are not allowed.
2166 * The page queues must be locked.
2167 * This routine may not block.
2170 vm_page_unwire(vm_page_t m, int activate)
2172 KKASSERT(m->flags & PG_BUSY);
2173 if (m->flags & PG_FICTITIOUS) {
2175 } else if (m->wire_count <= 0) {
2176 panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
2178 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) {
2179 atomic_add_int(&vmstats.v_wire_count, -1);
2180 if (m->flags & PG_UNMANAGED) {
2182 } else if (activate || (m->flags & PG_NEED_COMMIT)) {
2183 vm_page_spin_lock(m);
2184 _vm_page_add_queue_spinlocked(m,
2185 PQ_ACTIVE + m->pc, 0);
2186 _vm_page_and_queue_spin_unlock(m);
2188 vm_page_spin_lock(m);
2189 vm_page_flag_clear(m, PG_WINATCFLS);
2190 _vm_page_add_queue_spinlocked(m,
2191 PQ_INACTIVE + m->pc, 0);
2192 ++vm_swapcache_inactive_heuristic;
2193 _vm_page_and_queue_spin_unlock(m);
2200 * Move the specified page to the inactive queue. If the page has
2201 * any associated swap, the swap is deallocated.
2203 * Normally athead is 0 resulting in LRU operation. athead is set
2204 * to 1 if we want this page to be 'as if it were placed in the cache',
2205 * except without unmapping it from the process address space.
2207 * vm_page's spinlock must be held on entry and will remain held on return.
2208 * This routine may not block.
2211 _vm_page_deactivate_locked(vm_page_t m, int athead)
2216 * Ignore if already inactive.
2218 if (m->queue - m->pc == PQ_INACTIVE)
2220 _vm_page_queue_spin_lock(m);
2221 oqueue = _vm_page_rem_queue_spinlocked(m);
2223 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
2224 if (oqueue == PQ_CACHE)
2225 mycpu->gd_cnt.v_reactivated++;
2226 vm_page_flag_clear(m, PG_WINATCFLS);
2227 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead);
2229 ++vm_swapcache_inactive_heuristic;
2231 _vm_page_queue_spin_unlock(m);
2232 /* leaves vm_page spinlocked */
2236 * Attempt to deactivate a page.
2241 vm_page_deactivate(vm_page_t m)
2243 vm_page_spin_lock(m);
2244 _vm_page_deactivate_locked(m, 0);
2245 vm_page_spin_unlock(m);
2249 vm_page_deactivate_locked(vm_page_t m)
2251 _vm_page_deactivate_locked(m, 0);
2255 * Attempt to move a page to PQ_CACHE.
2257 * Returns 0 on failure, 1 on success
2259 * The page should NOT be busied by the caller. This function will validate
2260 * whether the page can be safely moved to the cache.
2263 vm_page_try_to_cache(vm_page_t m)
2265 vm_page_spin_lock(m);
2266 if (vm_page_busy_try(m, TRUE)) {
2267 vm_page_spin_unlock(m);
2270 if (m->dirty || m->hold_count || m->wire_count ||
2271 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT))) {
2272 if (_vm_page_wakeup(m)) {
2273 vm_page_spin_unlock(m);
2276 vm_page_spin_unlock(m);
2280 vm_page_spin_unlock(m);
2283 * Page busied by us and no longer spinlocked. Dirty pages cannot
2284 * be moved to the cache.
2286 vm_page_test_dirty(m);
2296 * Attempt to free the page. If we cannot free it, we do nothing.
2297 * 1 is returned on success, 0 on failure.
2302 vm_page_try_to_free(vm_page_t m)
2304 vm_page_spin_lock(m);
2305 if (vm_page_busy_try(m, TRUE)) {
2306 vm_page_spin_unlock(m);
2311 * The page can be in any state, including already being on the free
2312 * queue. Check to see if it really can be freed.
2314 if (m->dirty || /* can't free if it is dirty */
2315 m->hold_count || /* or held (XXX may be wrong) */
2316 m->wire_count || /* or wired */
2317 (m->flags & (PG_UNMANAGED | /* or unmanaged */
2318 PG_NEED_COMMIT)) || /* or needs a commit */
2319 m->queue - m->pc == PQ_FREE || /* already on PQ_FREE */
2320 m->queue - m->pc == PQ_HOLD) { /* already on PQ_HOLD */
2321 if (_vm_page_wakeup(m)) {
2322 vm_page_spin_unlock(m);
2325 vm_page_spin_unlock(m);
2329 vm_page_spin_unlock(m);
2332 * We can probably free the page.
2334 * Page busied by us and no longer spinlocked. Dirty pages will
2335 * not be freed by this function. We have to re-test the
2336 * dirty bit after cleaning out the pmaps.
2338 vm_page_test_dirty(m);
2343 vm_page_protect(m, VM_PROT_NONE);
2355 * Put the specified page onto the page cache queue (if appropriate).
2357 * The page must be busy, and this routine will release the busy and
2358 * possibly even free the page.
2361 vm_page_cache(vm_page_t m)
2363 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) ||
2364 m->busy || m->wire_count || m->hold_count) {
2365 kprintf("vm_page_cache: attempting to cache busy/held page\n");
2371 * Already in the cache (and thus not mapped)
2373 if ((m->queue - m->pc) == PQ_CACHE) {
2374 KKASSERT((m->flags & PG_MAPPED) == 0);
2380 * Caller is required to test m->dirty, but note that the act of
2381 * removing the page from its maps can cause it to become dirty
2382 * on an SMP system due to another cpu running in usermode.
2385 panic("vm_page_cache: caching a dirty page, pindex: %ld",
2390 * Remove all pmaps and indicate that the page is not
2391 * writeable or mapped. Our vm_page_protect() call may
2392 * have blocked (especially w/ VM_PROT_NONE), so recheck
2395 vm_page_protect(m, VM_PROT_NONE);
2396 if ((m->flags & (PG_UNMANAGED | PG_MAPPED)) ||
2397 m->busy || m->wire_count || m->hold_count) {
2399 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2400 vm_page_deactivate(m);
2403 _vm_page_and_queue_spin_lock(m);
2404 _vm_page_rem_queue_spinlocked(m);
2405 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0);
2406 _vm_page_queue_spin_unlock(m);
2407 if (_vm_page_wakeup(m)) {
2408 vm_page_spin_unlock(m);
2411 vm_page_spin_unlock(m);
2413 vm_page_free_wakeup();
2418 * vm_page_dontneed()
2420 * Cache, deactivate, or do nothing as appropriate. This routine
2421 * is typically used by madvise() MADV_DONTNEED.
2423 * Generally speaking we want to move the page into the cache so
2424 * it gets reused quickly. However, this can result in a silly syndrome
2425 * due to the page recycling too quickly. Small objects will not be
2426 * fully cached. On the otherhand, if we move the page to the inactive
2427 * queue we wind up with a problem whereby very large objects
2428 * unnecessarily blow away our inactive and cache queues.
2430 * The solution is to move the pages based on a fixed weighting. We
2431 * either leave them alone, deactivate them, or move them to the cache,
2432 * where moving them to the cache has the highest weighting.
2433 * By forcing some pages into other queues we eventually force the
2434 * system to balance the queues, potentially recovering other unrelated
2435 * space from active. The idea is to not force this to happen too
2438 * The page must be busied.
2441 vm_page_dontneed(vm_page_t m)
2443 static int dnweight;
2450 * occassionally leave the page alone
2452 if ((dnw & 0x01F0) == 0 ||
2453 m->queue - m->pc == PQ_INACTIVE ||
2454 m->queue - m->pc == PQ_CACHE
2456 if (m->act_count >= ACT_INIT)
2462 * If vm_page_dontneed() is inactivating a page, it must clear
2463 * the referenced flag; otherwise the pagedaemon will see references
2464 * on the page in the inactive queue and reactivate it. Until the
2465 * page can move to the cache queue, madvise's job is not done.
2467 vm_page_flag_clear(m, PG_REFERENCED);
2468 pmap_clear_reference(m);
2471 vm_page_test_dirty(m);
2473 if (m->dirty || (dnw & 0x0070) == 0) {
2475 * Deactivate the page 3 times out of 32.
2480 * Cache the page 28 times out of every 32. Note that
2481 * the page is deactivated instead of cached, but placed
2482 * at the head of the queue instead of the tail.
2486 vm_page_spin_lock(m);
2487 _vm_page_deactivate_locked(m, head);
2488 vm_page_spin_unlock(m);
2492 * These routines manipulate the 'soft busy' count for a page. A soft busy
2493 * is almost like PG_BUSY except that it allows certain compatible operations
2494 * to occur on the page while it is busy. For example, a page undergoing a
2495 * write can still be mapped read-only.
2497 * Because vm_pages can overlap buffers m->busy can be > 1. m->busy is only
2498 * adjusted while the vm_page is PG_BUSY so the flash will occur when the
2499 * busy bit is cleared.
2502 vm_page_io_start(vm_page_t m)
2504 KASSERT(m->flags & PG_BUSY, ("vm_page_io_start: page not busy!!!"));
2505 atomic_add_char(&m->busy, 1);
2506 vm_page_flag_set(m, PG_SBUSY);
2510 vm_page_io_finish(vm_page_t m)
2512 KASSERT(m->flags & PG_BUSY, ("vm_page_io_finish: page not busy!!!"));
2513 atomic_subtract_char(&m->busy, 1);
2515 vm_page_flag_clear(m, PG_SBUSY);
2519 * Indicate that a clean VM page requires a filesystem commit and cannot
2520 * be reused. Used by tmpfs.
2523 vm_page_need_commit(vm_page_t m)
2525 vm_page_flag_set(m, PG_NEED_COMMIT);
2529 vm_page_clear_commit(vm_page_t m)
2531 vm_page_flag_clear(m, PG_NEED_COMMIT);
2535 * Grab a page, blocking if it is busy and allocating a page if necessary.
2536 * A busy page is returned or NULL. The page may or may not be valid and
2537 * might not be on a queue (the caller is responsible for the disposition of
2540 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the
2541 * page will be zero'd and marked valid.
2543 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked
2544 * valid even if it already exists.
2546 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also
2547 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified.
2548 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified.
2550 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
2551 * always returned if we had blocked.
2553 * This routine may not be called from an interrupt.
2555 * PG_ZERO is *ALWAYS* cleared by this routine.
2557 * No other requirements.
2560 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2565 KKASSERT(allocflags &
2566 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
2567 vm_object_hold(object);
2569 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error);
2571 vm_page_sleep_busy(m, TRUE, "pgrbwt");
2572 if ((allocflags & VM_ALLOC_RETRY) == 0) {
2577 } else if (m == NULL) {
2578 if (allocflags & VM_ALLOC_RETRY)
2579 allocflags |= VM_ALLOC_NULL_OK;
2580 m = vm_page_alloc(object, pindex,
2581 allocflags & ~VM_ALLOC_RETRY);
2585 if ((allocflags & VM_ALLOC_RETRY) == 0)
2594 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid.
2596 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set
2597 * valid even if already valid.
2599 if (m->valid == 0) {
2600 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) {
2601 if ((m->flags & PG_ZERO) == 0)
2602 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2603 m->valid = VM_PAGE_BITS_ALL;
2605 } else if (allocflags & VM_ALLOC_FORCE_ZERO) {
2606 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2607 m->valid = VM_PAGE_BITS_ALL;
2609 vm_page_flag_clear(m, PG_ZERO);
2611 vm_object_drop(object);
2616 * Mapping function for valid bits or for dirty bits in
2617 * a page. May not block.
2619 * Inputs are required to range within a page.
2625 vm_page_bits(int base, int size)
2631 base + size <= PAGE_SIZE,
2632 ("vm_page_bits: illegal base/size %d/%d", base, size)
2635 if (size == 0) /* handle degenerate case */
2638 first_bit = base >> DEV_BSHIFT;
2639 last_bit = (base + size - 1) >> DEV_BSHIFT;
2641 return ((2 << last_bit) - (1 << first_bit));
2645 * Sets portions of a page valid and clean. The arguments are expected
2646 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2647 * of any partial chunks touched by the range. The invalid portion of
2648 * such chunks will be zero'd.
2650 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically
2651 * align base to DEV_BSIZE so as not to mark clean a partially
2652 * truncated device block. Otherwise the dirty page status might be
2655 * This routine may not block.
2657 * (base + size) must be less then or equal to PAGE_SIZE.
2660 _vm_page_zero_valid(vm_page_t m, int base, int size)
2665 if (size == 0) /* handle degenerate case */
2669 * If the base is not DEV_BSIZE aligned and the valid
2670 * bit is clear, we have to zero out a portion of the
2674 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2675 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
2677 pmap_zero_page_area(
2685 * If the ending offset is not DEV_BSIZE aligned and the
2686 * valid bit is clear, we have to zero out a portion of
2690 endoff = base + size;
2692 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2693 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
2695 pmap_zero_page_area(
2698 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
2704 * Set valid, clear dirty bits. If validating the entire
2705 * page we can safely clear the pmap modify bit. We also
2706 * use this opportunity to clear the PG_NOSYNC flag. If a process
2707 * takes a write fault on a MAP_NOSYNC memory area the flag will
2710 * We set valid bits inclusive of any overlap, but we can only
2711 * clear dirty bits for DEV_BSIZE chunks that are fully within
2714 * Page must be busied?
2715 * No other requirements.
2718 vm_page_set_valid(vm_page_t m, int base, int size)
2720 _vm_page_zero_valid(m, base, size);
2721 m->valid |= vm_page_bits(base, size);
2726 * Set valid bits and clear dirty bits.
2728 * NOTE: This function does not clear the pmap modified bit.
2729 * Also note that e.g. NFS may use a byte-granular base
2732 * WARNING: Page must be busied? But vfs_clean_one_page() will call
2733 * this without necessarily busying the page (via bdwrite()).
2734 * So for now vm_token must also be held.
2736 * No other requirements.
2739 vm_page_set_validclean(vm_page_t m, int base, int size)
2743 _vm_page_zero_valid(m, base, size);
2744 pagebits = vm_page_bits(base, size);
2745 m->valid |= pagebits;
2746 m->dirty &= ~pagebits;
2747 if (base == 0 && size == PAGE_SIZE) {
2748 /*pmap_clear_modify(m);*/
2749 vm_page_flag_clear(m, PG_NOSYNC);
2754 * Set valid & dirty. Used by buwrite()
2756 * WARNING: Page must be busied? But vfs_dirty_one_page() will
2757 * call this function in buwrite() so for now vm_token must
2760 * No other requirements.
2763 vm_page_set_validdirty(vm_page_t m, int base, int size)
2767 pagebits = vm_page_bits(base, size);
2768 m->valid |= pagebits;
2769 m->dirty |= pagebits;
2771 vm_object_set_writeable_dirty(m->object);
2777 * NOTE: This function does not clear the pmap modified bit.
2778 * Also note that e.g. NFS may use a byte-granular base
2781 * Page must be busied?
2782 * No other requirements.
2785 vm_page_clear_dirty(vm_page_t m, int base, int size)
2787 m->dirty &= ~vm_page_bits(base, size);
2788 if (base == 0 && size == PAGE_SIZE) {
2789 /*pmap_clear_modify(m);*/
2790 vm_page_flag_clear(m, PG_NOSYNC);
2795 * Make the page all-dirty.
2797 * Also make sure the related object and vnode reflect the fact that the
2798 * object may now contain a dirty page.
2800 * Page must be busied?
2801 * No other requirements.
2804 vm_page_dirty(vm_page_t m)
2807 int pqtype = m->queue - m->pc;
2809 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE,
2810 ("vm_page_dirty: page in free/cache queue!"));
2811 if (m->dirty != VM_PAGE_BITS_ALL) {
2812 m->dirty = VM_PAGE_BITS_ALL;
2814 vm_object_set_writeable_dirty(m->object);
2819 * Invalidates DEV_BSIZE'd chunks within a page. Both the
2820 * valid and dirty bits for the effected areas are cleared.
2822 * Page must be busied?
2824 * No other requirements.
2827 vm_page_set_invalid(vm_page_t m, int base, int size)
2831 bits = vm_page_bits(base, size);
2834 m->object->generation++;
2838 * The kernel assumes that the invalid portions of a page contain
2839 * garbage, but such pages can be mapped into memory by user code.
2840 * When this occurs, we must zero out the non-valid portions of the
2841 * page so user code sees what it expects.
2843 * Pages are most often semi-valid when the end of a file is mapped
2844 * into memory and the file's size is not page aligned.
2846 * Page must be busied?
2847 * No other requirements.
2850 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
2856 * Scan the valid bits looking for invalid sections that
2857 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
2858 * valid bit may be set ) have already been zerod by
2859 * vm_page_set_validclean().
2861 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
2862 if (i == (PAGE_SIZE / DEV_BSIZE) ||
2863 (m->valid & (1 << i))
2866 pmap_zero_page_area(
2869 (i - b) << DEV_BSHIFT
2877 * setvalid is TRUE when we can safely set the zero'd areas
2878 * as being valid. We can do this if there are no cache consistency
2879 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
2882 m->valid = VM_PAGE_BITS_ALL;
2886 * Is a (partial) page valid? Note that the case where size == 0
2887 * will return FALSE in the degenerate case where the page is entirely
2888 * invalid, and TRUE otherwise.
2891 * No other requirements.
2894 vm_page_is_valid(vm_page_t m, int base, int size)
2896 int bits = vm_page_bits(base, size);
2898 if (m->valid && ((m->valid & bits) == bits))
2905 * update dirty bits from pmap/mmu. May not block.
2907 * Caller must hold the page busy
2910 vm_page_test_dirty(vm_page_t m)
2912 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
2918 * Register an action, associating it with its vm_page
2921 vm_page_register_action(vm_page_action_t action, vm_page_event_t event)
2923 struct vm_page_action_list *list;
2926 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
2927 list = &action_list[hv];
2929 lwkt_gettoken(&vm_token);
2930 vm_page_flag_set(action->m, PG_ACTIONLIST);
2931 action->event = event;
2932 LIST_INSERT_HEAD(list, action, entry);
2933 lwkt_reltoken(&vm_token);
2937 * Unregister an action, disassociating it from its related vm_page
2940 vm_page_unregister_action(vm_page_action_t action)
2942 struct vm_page_action_list *list;
2945 lwkt_gettoken(&vm_token);
2946 if (action->event != VMEVENT_NONE) {
2947 action->event = VMEVENT_NONE;
2948 LIST_REMOVE(action, entry);
2950 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
2951 list = &action_list[hv];
2952 if (LIST_EMPTY(list))
2953 vm_page_flag_clear(action->m, PG_ACTIONLIST);
2955 lwkt_reltoken(&vm_token);
2959 * Issue an event on a VM page. Corresponding action structures are
2960 * removed from the page's list and called.
2962 * If the vm_page has no more pending action events we clear its
2963 * PG_ACTIONLIST flag.
2966 vm_page_event_internal(vm_page_t m, vm_page_event_t event)
2968 struct vm_page_action_list *list;
2969 struct vm_page_action *scan;
2970 struct vm_page_action *next;
2974 hv = (int)((intptr_t)m >> 8) & VMACTION_HMASK;
2975 list = &action_list[hv];
2978 lwkt_gettoken(&vm_token);
2979 LIST_FOREACH_MUTABLE(scan, list, entry, next) {
2981 if (scan->event == event) {
2982 scan->event = VMEVENT_NONE;
2983 LIST_REMOVE(scan, entry);
2984 scan->func(m, scan);
2992 vm_page_flag_clear(m, PG_ACTIONLIST);
2993 lwkt_reltoken(&vm_token);
2996 #include "opt_ddb.h"
2998 #include <sys/kernel.h>
3000 #include <ddb/ddb.h>
3002 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3004 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count);
3005 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count);
3006 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count);
3007 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count);
3008 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count);
3009 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved);
3010 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min);
3011 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target);
3012 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min);
3013 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target);
3016 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3019 db_printf("PQ_FREE:");
3020 for(i=0;i<PQ_L2_SIZE;i++) {
3021 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
3025 db_printf("PQ_CACHE:");
3026 for(i=0;i<PQ_L2_SIZE;i++) {
3027 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
3031 db_printf("PQ_ACTIVE:");
3032 for(i=0;i<PQ_L2_SIZE;i++) {
3033 db_printf(" %d", vm_page_queues[PQ_ACTIVE + i].lcnt);
3037 db_printf("PQ_INACTIVE:");
3038 for(i=0;i<PQ_L2_SIZE;i++) {
3039 db_printf(" %d", vm_page_queues[PQ_INACTIVE + i].lcnt);