4 * Copyright (c) 1991 Regents of the University of California.
7 * This code is derived from software contributed to Berkeley by
8 * The Mach Operating System project at Carnegie-Mellon University.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91
35 * $FreeBSD: src/sys/vm/vm_page.c,v 1.147.2.18 2002/03/10 05:03:19 alc Exp $
39 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40 * All rights reserved.
42 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
44 * Permission to use, copy, modify and distribute this software and
45 * its documentation is hereby granted, provided that both the copyright
46 * notice and this permission notice appear in all copies of the
47 * software, derivative works or modified versions, and any portions
48 * thereof, and that both notices appear in supporting documentation.
50 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
54 * Carnegie Mellon requests users of this software to return to
56 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
57 * School of Computer Science
58 * Carnegie Mellon University
59 * Pittsburgh PA 15213-3890
61 * any improvements or extensions that they make and grant Carnegie the
62 * rights to redistribute these changes.
65 * Resident memory management module. The module manipulates 'VM pages'.
66 * A VM page is the core building block for memory management.
69 #include <sys/param.h>
70 #include <sys/systm.h>
71 #include <sys/malloc.h>
73 #include <sys/vmmeter.h>
74 #include <sys/vnode.h>
75 #include <sys/kernel.h>
76 #include <sys/alist.h>
77 #include <sys/sysctl.h>
80 #include <vm/vm_param.h>
82 #include <vm/vm_kern.h>
84 #include <vm/vm_map.h>
85 #include <vm/vm_object.h>
86 #include <vm/vm_page.h>
87 #include <vm/vm_pageout.h>
88 #include <vm/vm_pager.h>
89 #include <vm/vm_extern.h>
90 #include <vm/swap_pager.h>
92 #include <machine/inttypes.h>
93 #include <machine/md_var.h>
94 #include <machine/specialreg.h>
96 #include <vm/vm_page2.h>
97 #include <sys/spinlock2.h>
99 #define VMACTION_HSIZE 256
100 #define VMACTION_HMASK (VMACTION_HSIZE - 1)
102 static void vm_page_queue_init(void);
103 static void vm_page_free_wakeup(void);
104 static vm_page_t
vm_page_select_cache(u_short pg_color
);
105 static vm_page_t
_vm_page_list_find2(int basequeue
, int index
);
106 static void _vm_page_deactivate_locked(vm_page_t m
, int athead
);
109 * Array of tailq lists
111 __cachealign
struct vpgqueues vm_page_queues
[PQ_COUNT
];
113 LIST_HEAD(vm_page_action_list
, vm_page_action
);
114 struct vm_page_action_list action_list
[VMACTION_HSIZE
];
115 static volatile int vm_pages_waiting
;
117 static struct alist vm_contig_alist
;
118 static struct almeta vm_contig_ameta
[ALIST_RECORDS_65536
];
119 static struct spinlock vm_contig_spin
= SPINLOCK_INITIALIZER(&vm_contig_spin
, "vm_contig_spin");
121 static u_long vm_dma_reserved
= 0;
122 TUNABLE_ULONG("vm.dma_reserved", &vm_dma_reserved
);
123 SYSCTL_ULONG(_vm
, OID_AUTO
, dma_reserved
, CTLFLAG_RD
, &vm_dma_reserved
, 0,
124 "Memory reserved for DMA");
125 SYSCTL_UINT(_vm
, OID_AUTO
, dma_free_pages
, CTLFLAG_RD
,
126 &vm_contig_alist
.bl_free
, 0, "Memory reserved for DMA");
128 static int vm_contig_verbose
= 0;
129 TUNABLE_INT("vm.contig_verbose", &vm_contig_verbose
);
131 RB_GENERATE2(vm_page_rb_tree
, vm_page
, rb_entry
, rb_vm_page_compare
,
132 vm_pindex_t
, pindex
);
135 vm_page_queue_init(void)
139 for (i
= 0; i
< PQ_L2_SIZE
; i
++)
140 vm_page_queues
[PQ_FREE
+i
].cnt
= &vmstats
.v_free_count
;
141 for (i
= 0; i
< PQ_L2_SIZE
; i
++)
142 vm_page_queues
[PQ_CACHE
+i
].cnt
= &vmstats
.v_cache_count
;
143 for (i
= 0; i
< PQ_L2_SIZE
; i
++)
144 vm_page_queues
[PQ_INACTIVE
+i
].cnt
= &vmstats
.v_inactive_count
;
145 for (i
= 0; i
< PQ_L2_SIZE
; i
++)
146 vm_page_queues
[PQ_ACTIVE
+i
].cnt
= &vmstats
.v_active_count
;
147 for (i
= 0; i
< PQ_L2_SIZE
; i
++)
148 vm_page_queues
[PQ_HOLD
+i
].cnt
= &vmstats
.v_active_count
;
149 /* PQ_NONE has no queue */
151 for (i
= 0; i
< PQ_COUNT
; i
++) {
152 TAILQ_INIT(&vm_page_queues
[i
].pl
);
153 spin_init(&vm_page_queues
[i
].spin
, "vm_page_queue_init");
156 for (i
= 0; i
< VMACTION_HSIZE
; i
++)
157 LIST_INIT(&action_list
[i
]);
161 * note: place in initialized data section? Is this necessary?
164 int vm_page_array_size
= 0;
165 int vm_page_zero_count
= 0;
166 vm_page_t vm_page_array
= NULL
;
167 vm_paddr_t vm_low_phys_reserved
;
172 * Sets the page size, perhaps based upon the memory size.
173 * Must be called before any use of page-size dependent functions.
176 vm_set_page_size(void)
178 if (vmstats
.v_page_size
== 0)
179 vmstats
.v_page_size
= PAGE_SIZE
;
180 if (((vmstats
.v_page_size
- 1) & vmstats
.v_page_size
) != 0)
181 panic("vm_set_page_size: page size not a power of two");
187 * Add a new page to the freelist for use by the system. New pages
188 * are added to both the head and tail of the associated free page
189 * queue in a bottom-up fashion, so both zero'd and non-zero'd page
190 * requests pull 'recent' adds (higher physical addresses) first.
192 * Beware that the page zeroing daemon will also be running soon after
193 * boot, moving pages from the head to the tail of the PQ_FREE queues.
195 * Must be called in a critical section.
198 vm_add_new_page(vm_paddr_t pa
)
200 struct vpgqueues
*vpq
;
203 m
= PHYS_TO_VM_PAGE(pa
);
206 m
->pc
= (pa
>> PAGE_SHIFT
) & PQ_L2_MASK
;
207 m
->pat_mode
= PAT_WRITE_BACK
;
209 * Twist for cpu localization in addition to page coloring, so
210 * different cpus selecting by m->queue get different page colors.
212 m
->pc
^= ((pa
>> PAGE_SHIFT
) / PQ_L2_SIZE
) & PQ_L2_MASK
;
213 m
->pc
^= ((pa
>> PAGE_SHIFT
) / (PQ_L2_SIZE
* PQ_L2_SIZE
)) & PQ_L2_MASK
;
215 * Reserve a certain number of contiguous low memory pages for
216 * contigmalloc() to use.
218 if (pa
< vm_low_phys_reserved
) {
219 atomic_add_int(&vmstats
.v_page_count
, 1);
220 atomic_add_int(&vmstats
.v_dma_pages
, 1);
223 atomic_add_int(&vmstats
.v_wire_count
, 1);
224 alist_free(&vm_contig_alist
, pa
>> PAGE_SHIFT
, 1);
231 m
->queue
= m
->pc
+ PQ_FREE
;
232 KKASSERT(m
->dirty
== 0);
234 atomic_add_int(&vmstats
.v_page_count
, 1);
235 atomic_add_int(&vmstats
.v_free_count
, 1);
236 vpq
= &vm_page_queues
[m
->queue
];
237 if ((vpq
->flipflop
& 15) == 0) {
238 pmap_zero_page(VM_PAGE_TO_PHYS(m
));
240 TAILQ_INSERT_TAIL(&vpq
->pl
, m
, pageq
);
241 atomic_add_int(&vm_page_zero_count
, 1);
243 TAILQ_INSERT_HEAD(&vpq
->pl
, m
, pageq
);
252 * Initializes the resident memory module.
254 * Preallocates memory for critical VM structures and arrays prior to
255 * kernel_map becoming available.
257 * Memory is allocated from (virtual2_start, virtual2_end) if available,
258 * otherwise memory is allocated from (virtual_start, virtual_end).
260 * On x86-64 (virtual_start, virtual_end) is only 2GB and may not be
261 * large enough to hold vm_page_array & other structures for machines with
262 * large amounts of ram, so we want to use virtual2* when available.
265 vm_page_startup(void)
267 vm_offset_t vaddr
= virtual2_start
? virtual2_start
: virtual_start
;
270 vm_paddr_t page_range
;
277 vm_paddr_t biggestone
, biggestsize
;
284 vaddr
= round_page(vaddr
);
286 for (i
= 0; phys_avail
[i
+ 1]; i
+= 2) {
287 phys_avail
[i
] = round_page64(phys_avail
[i
]);
288 phys_avail
[i
+ 1] = trunc_page64(phys_avail
[i
+ 1]);
291 for (i
= 0; phys_avail
[i
+ 1]; i
+= 2) {
292 vm_paddr_t size
= phys_avail
[i
+ 1] - phys_avail
[i
];
294 if (size
> biggestsize
) {
302 end
= phys_avail
[biggestone
+1];
303 end
= trunc_page(end
);
306 * Initialize the queue headers for the free queue, the active queue
307 * and the inactive queue.
309 vm_page_queue_init();
311 #if !defined(_KERNEL_VIRTUAL)
313 * VKERNELs don't support minidumps and as such don't need
316 * Allocate a bitmap to indicate that a random physical page
317 * needs to be included in a minidump.
319 * The amd64 port needs this to indicate which direct map pages
320 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
322 * However, i386 still needs this workspace internally within the
323 * minidump code. In theory, they are not needed on i386, but are
324 * included should the sf_buf code decide to use them.
326 page_range
= phys_avail
[(nblocks
- 1) * 2 + 1] / PAGE_SIZE
;
327 vm_page_dump_size
= round_page(roundup2(page_range
, NBBY
) / NBBY
);
328 end
-= vm_page_dump_size
;
329 vm_page_dump
= (void *)pmap_map(&vaddr
, end
, end
+ vm_page_dump_size
,
330 VM_PROT_READ
| VM_PROT_WRITE
);
331 bzero((void *)vm_page_dump
, vm_page_dump_size
);
334 * Compute the number of pages of memory that will be available for
335 * use (taking into account the overhead of a page structure per
338 first_page
= phys_avail
[0] / PAGE_SIZE
;
339 page_range
= phys_avail
[(nblocks
- 1) * 2 + 1] / PAGE_SIZE
- first_page
;
340 npages
= (total
- (page_range
* sizeof(struct vm_page
))) / PAGE_SIZE
;
342 #ifndef _KERNEL_VIRTUAL
344 * (only applies to real kernels)
346 * Reserve a large amount of low memory for potential 32-bit DMA
347 * space allocations. Once device initialization is complete we
348 * release most of it, but keep (vm_dma_reserved) memory reserved
349 * for later use. Typically for X / graphics. Through trial and
350 * error we find that GPUs usually requires ~60-100MB or so.
352 * By default, 128M is left in reserve on machines with 2G+ of ram.
354 vm_low_phys_reserved
= (vm_paddr_t
)65536 << PAGE_SHIFT
;
355 if (vm_low_phys_reserved
> total
/ 4)
356 vm_low_phys_reserved
= total
/ 4;
357 if (vm_dma_reserved
== 0) {
358 vm_dma_reserved
= 128 * 1024 * 1024; /* 128MB */
359 if (vm_dma_reserved
> total
/ 16)
360 vm_dma_reserved
= total
/ 16;
363 alist_init(&vm_contig_alist
, 65536, vm_contig_ameta
,
364 ALIST_RECORDS_65536
);
367 * Initialize the mem entry structures now, and put them in the free
370 new_end
= trunc_page(end
- page_range
* sizeof(struct vm_page
));
371 mapped
= pmap_map(&vaddr
, new_end
, end
, VM_PROT_READ
| VM_PROT_WRITE
);
372 vm_page_array
= (vm_page_t
)mapped
;
374 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL)
376 * since pmap_map on amd64 returns stuff out of a direct-map region,
377 * we have to manually add these pages to the minidump tracking so
378 * that they can be dumped, including the vm_page_array.
380 for (pa
= new_end
; pa
< phys_avail
[biggestone
+ 1]; pa
+= PAGE_SIZE
)
385 * Clear all of the page structures
387 bzero((caddr_t
) vm_page_array
, page_range
* sizeof(struct vm_page
));
388 vm_page_array_size
= page_range
;
391 * Construct the free queue(s) in ascending order (by physical
392 * address) so that the first 16MB of physical memory is allocated
393 * last rather than first. On large-memory machines, this avoids
394 * the exhaustion of low physical memory before isa_dmainit has run.
396 vmstats
.v_page_count
= 0;
397 vmstats
.v_free_count
= 0;
398 for (i
= 0; phys_avail
[i
+ 1] && npages
> 0; i
+= 2) {
403 last_pa
= phys_avail
[i
+ 1];
404 while (pa
< last_pa
&& npages
-- > 0) {
410 virtual2_start
= vaddr
;
412 virtual_start
= vaddr
;
416 * We tended to reserve a ton of memory for contigmalloc(). Now that most
417 * drivers have initialized we want to return most the remaining free
418 * reserve back to the VM page queues so they can be used for normal
421 * We leave vm_dma_reserved bytes worth of free pages in the reserve pool.
424 vm_page_startup_finish(void *dummy __unused
)
433 spin_lock(&vm_contig_spin
);
435 bfree
= alist_free_info(&vm_contig_alist
, &blk
, &count
);
436 if (bfree
<= vm_dma_reserved
/ PAGE_SIZE
)
442 * Figure out how much of the initial reserve we have to
443 * free in order to reach our target.
445 bfree
-= vm_dma_reserved
/ PAGE_SIZE
;
447 blk
+= count
- bfree
;
452 * Calculate the nearest power of 2 <= count.
454 for (xcount
= 1; xcount
<= count
; xcount
<<= 1)
457 blk
+= count
- xcount
;
461 * Allocate the pages from the alist, then free them to
462 * the normal VM page queues.
464 * Pages allocated from the alist are wired. We have to
465 * busy, unwire, and free them. We must also adjust
466 * vm_low_phys_reserved before freeing any pages to prevent
469 rblk
= alist_alloc(&vm_contig_alist
, blk
, count
);
471 kprintf("vm_page_startup_finish: Unable to return "
472 "dma space @0x%08x/%d -> 0x%08x\n",
476 atomic_add_int(&vmstats
.v_dma_pages
, -count
);
477 spin_unlock(&vm_contig_spin
);
479 m
= PHYS_TO_VM_PAGE((vm_paddr_t
)blk
<< PAGE_SHIFT
);
480 vm_low_phys_reserved
= VM_PAGE_TO_PHYS(m
);
482 vm_page_busy_wait(m
, FALSE
, "cpgfr");
483 vm_page_unwire(m
, 0);
488 spin_lock(&vm_contig_spin
);
490 spin_unlock(&vm_contig_spin
);
493 * Print out how much DMA space drivers have already allocated and
494 * how much is left over.
496 kprintf("DMA space used: %jdk, remaining available: %jdk\n",
497 (intmax_t)(vmstats
.v_dma_pages
- vm_contig_alist
.bl_free
) *
499 (intmax_t)vm_contig_alist
.bl_free
* (PAGE_SIZE
/ 1024));
501 SYSINIT(vm_pgend
, SI_SUB_PROC0_POST
, SI_ORDER_ANY
,
502 vm_page_startup_finish
, NULL
);
506 * Scan comparison function for Red-Black tree scans. An inclusive
507 * (start,end) is expected. Other fields are not used.
510 rb_vm_page_scancmp(struct vm_page
*p
, void *data
)
512 struct rb_vm_page_scan_info
*info
= data
;
514 if (p
->pindex
< info
->start_pindex
)
516 if (p
->pindex
> info
->end_pindex
)
522 rb_vm_page_compare(struct vm_page
*p1
, struct vm_page
*p2
)
524 if (p1
->pindex
< p2
->pindex
)
526 if (p1
->pindex
> p2
->pindex
)
532 vm_page_init(vm_page_t m
)
534 /* do nothing for now. Called from pmap_page_init() */
538 * Each page queue has its own spin lock, which is fairly optimal for
539 * allocating and freeing pages at least.
541 * The caller must hold the vm_page_spin_lock() before locking a vm_page's
542 * queue spinlock via this function. Also note that m->queue cannot change
543 * unless both the page and queue are locked.
547 _vm_page_queue_spin_lock(vm_page_t m
)
552 if (queue
!= PQ_NONE
) {
553 spin_lock(&vm_page_queues
[queue
].spin
);
554 KKASSERT(queue
== m
->queue
);
560 _vm_page_queue_spin_unlock(vm_page_t m
)
566 if (queue
!= PQ_NONE
)
567 spin_unlock(&vm_page_queues
[queue
].spin
);
572 _vm_page_queues_spin_lock(u_short queue
)
575 if (queue
!= PQ_NONE
)
576 spin_lock(&vm_page_queues
[queue
].spin
);
582 _vm_page_queues_spin_unlock(u_short queue
)
585 if (queue
!= PQ_NONE
)
586 spin_unlock(&vm_page_queues
[queue
].spin
);
590 vm_page_queue_spin_lock(vm_page_t m
)
592 _vm_page_queue_spin_lock(m
);
596 vm_page_queues_spin_lock(u_short queue
)
598 _vm_page_queues_spin_lock(queue
);
602 vm_page_queue_spin_unlock(vm_page_t m
)
604 _vm_page_queue_spin_unlock(m
);
608 vm_page_queues_spin_unlock(u_short queue
)
610 _vm_page_queues_spin_unlock(queue
);
614 * This locks the specified vm_page and its queue in the proper order
615 * (page first, then queue). The queue may change so the caller must
620 _vm_page_and_queue_spin_lock(vm_page_t m
)
622 vm_page_spin_lock(m
);
623 _vm_page_queue_spin_lock(m
);
628 _vm_page_and_queue_spin_unlock(vm_page_t m
)
630 _vm_page_queues_spin_unlock(m
->queue
);
631 vm_page_spin_unlock(m
);
635 vm_page_and_queue_spin_unlock(vm_page_t m
)
637 _vm_page_and_queue_spin_unlock(m
);
641 vm_page_and_queue_spin_lock(vm_page_t m
)
643 _vm_page_and_queue_spin_lock(m
);
647 * Helper function removes vm_page from its current queue.
648 * Returns the base queue the page used to be on.
650 * The vm_page and the queue must be spinlocked.
651 * This function will unlock the queue but leave the page spinlocked.
653 static __inline u_short
654 _vm_page_rem_queue_spinlocked(vm_page_t m
)
656 struct vpgqueues
*pq
;
660 if (queue
!= PQ_NONE
) {
661 pq
= &vm_page_queues
[queue
];
662 TAILQ_REMOVE(&pq
->pl
, m
, pageq
);
663 atomic_add_int(pq
->cnt
, -1);
666 vm_page_queues_spin_unlock(queue
);
667 if ((queue
- m
->pc
) == PQ_FREE
&& (m
->flags
& PG_ZERO
))
668 atomic_subtract_int(&vm_page_zero_count
, 1);
669 if ((queue
- m
->pc
) == PQ_CACHE
|| (queue
- m
->pc
) == PQ_FREE
)
670 return (queue
- m
->pc
);
676 * Helper function places the vm_page on the specified queue.
678 * The vm_page must be spinlocked.
679 * This function will return with both the page and the queue locked.
682 _vm_page_add_queue_spinlocked(vm_page_t m
, u_short queue
, int athead
)
684 struct vpgqueues
*pq
;
686 KKASSERT(m
->queue
== PQ_NONE
);
688 if (queue
!= PQ_NONE
) {
689 vm_page_queues_spin_lock(queue
);
690 pq
= &vm_page_queues
[queue
];
692 atomic_add_int(pq
->cnt
, 1);
696 * Put zero'd pages on the end ( where we look for zero'd pages
697 * first ) and non-zerod pages at the head.
699 if (queue
- m
->pc
== PQ_FREE
) {
700 if (m
->flags
& PG_ZERO
) {
701 TAILQ_INSERT_TAIL(&pq
->pl
, m
, pageq
);
702 atomic_add_int(&vm_page_zero_count
, 1);
704 TAILQ_INSERT_HEAD(&pq
->pl
, m
, pageq
);
707 TAILQ_INSERT_HEAD(&pq
->pl
, m
, pageq
);
709 TAILQ_INSERT_TAIL(&pq
->pl
, m
, pageq
);
711 /* leave the queue spinlocked */
716 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE)
717 * m->busy is zero. Returns TRUE if it had to sleep, FALSE if we
718 * did not. Only one sleep call will be made before returning.
720 * This function does NOT busy the page and on return the page is not
721 * guaranteed to be available.
724 vm_page_sleep_busy(vm_page_t m
, int also_m_busy
, const char *msg
)
732 if ((flags
& PG_BUSY
) == 0 &&
733 (also_m_busy
== 0 || (flags
& PG_SBUSY
) == 0)) {
736 tsleep_interlock(m
, 0);
737 if (atomic_cmpset_int(&m
->flags
, flags
,
738 flags
| PG_WANTED
| PG_REFERENCED
)) {
739 tsleep(m
, PINTERLOCKED
, msg
, 0);
746 * Wait until PG_BUSY can be set, then set it. If also_m_busy is TRUE we
747 * also wait for m->busy to become 0 before setting PG_BUSY.
750 VM_PAGE_DEBUG_EXT(vm_page_busy_wait
)(vm_page_t m
,
751 int also_m_busy
, const char *msg
759 if (flags
& PG_BUSY
) {
760 tsleep_interlock(m
, 0);
761 if (atomic_cmpset_int(&m
->flags
, flags
,
762 flags
| PG_WANTED
| PG_REFERENCED
)) {
763 tsleep(m
, PINTERLOCKED
, msg
, 0);
765 } else if (also_m_busy
&& (flags
& PG_SBUSY
)) {
766 tsleep_interlock(m
, 0);
767 if (atomic_cmpset_int(&m
->flags
, flags
,
768 flags
| PG_WANTED
| PG_REFERENCED
)) {
769 tsleep(m
, PINTERLOCKED
, msg
, 0);
772 if (atomic_cmpset_int(&m
->flags
, flags
,
776 m
->busy_line
= lineno
;
785 * Attempt to set PG_BUSY. If also_m_busy is TRUE we only succeed if m->busy
788 * Returns non-zero on failure.
791 VM_PAGE_DEBUG_EXT(vm_page_busy_try
)(vm_page_t m
, int also_m_busy
801 if (also_m_busy
&& (flags
& PG_SBUSY
))
803 if (atomic_cmpset_int(&m
->flags
, flags
, flags
| PG_BUSY
)) {
806 m
->busy_line
= lineno
;
814 * Clear the PG_BUSY flag and return non-zero to indicate to the caller
815 * that a wakeup() should be performed.
817 * The vm_page must be spinlocked and will remain spinlocked on return.
818 * The related queue must NOT be spinlocked (which could deadlock us).
824 _vm_page_wakeup(vm_page_t m
)
831 if (atomic_cmpset_int(&m
->flags
, flags
,
832 flags
& ~(PG_BUSY
| PG_WANTED
))) {
836 return(flags
& PG_WANTED
);
840 * Clear the PG_BUSY flag and wakeup anyone waiting for the page. This
841 * is typically the last call you make on a page before moving onto
845 vm_page_wakeup(vm_page_t m
)
847 KASSERT(m
->flags
& PG_BUSY
, ("vm_page_wakeup: page not busy!!!"));
848 vm_page_spin_lock(m
);
849 if (_vm_page_wakeup(m
)) {
850 vm_page_spin_unlock(m
);
853 vm_page_spin_unlock(m
);
858 * Holding a page keeps it from being reused. Other parts of the system
859 * can still disassociate the page from its current object and free it, or
860 * perform read or write I/O on it and/or otherwise manipulate the page,
861 * but if the page is held the VM system will leave the page and its data
862 * intact and not reuse the page for other purposes until the last hold
863 * reference is released. (see vm_page_wire() if you want to prevent the
864 * page from being disassociated from its object too).
866 * The caller must still validate the contents of the page and, if necessary,
867 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete
868 * before manipulating the page.
870 * XXX get vm_page_spin_lock() here and move FREE->HOLD if necessary
873 vm_page_hold(vm_page_t m
)
875 vm_page_spin_lock(m
);
876 atomic_add_int(&m
->hold_count
, 1);
877 if (m
->queue
- m
->pc
== PQ_FREE
) {
878 _vm_page_queue_spin_lock(m
);
879 _vm_page_rem_queue_spinlocked(m
);
880 _vm_page_add_queue_spinlocked(m
, PQ_HOLD
+ m
->pc
, 0);
881 _vm_page_queue_spin_unlock(m
);
883 vm_page_spin_unlock(m
);
887 * The opposite of vm_page_hold(). If the page is on the HOLD queue
888 * it was freed while held and must be moved back to the FREE queue.
891 vm_page_unhold(vm_page_t m
)
893 KASSERT(m
->hold_count
> 0 && m
->queue
- m
->pc
!= PQ_FREE
,
894 ("vm_page_unhold: pg %p illegal hold_count (%d) or on FREE queue (%d)",
895 m
, m
->hold_count
, m
->queue
- m
->pc
));
896 vm_page_spin_lock(m
);
897 atomic_add_int(&m
->hold_count
, -1);
898 if (m
->hold_count
== 0 && m
->queue
- m
->pc
== PQ_HOLD
) {
899 _vm_page_queue_spin_lock(m
);
900 _vm_page_rem_queue_spinlocked(m
);
901 _vm_page_add_queue_spinlocked(m
, PQ_FREE
+ m
->pc
, 0);
902 _vm_page_queue_spin_unlock(m
);
904 vm_page_spin_unlock(m
);
910 * Create a fictitious page with the specified physical address and
911 * memory attribute. The memory attribute is the only the machine-
912 * dependent aspect of a fictitious page that must be initialized.
916 vm_page_initfake(vm_page_t m
, vm_paddr_t paddr
, vm_memattr_t memattr
)
919 if ((m
->flags
& PG_FICTITIOUS
) != 0) {
921 * The page's memattr might have changed since the
922 * previous initialization. Update the pmap to the
927 m
->phys_addr
= paddr
;
929 /* Fictitious pages don't use "segind". */
930 /* Fictitious pages don't use "order" or "pool". */
931 m
->flags
= PG_FICTITIOUS
| PG_UNMANAGED
| PG_BUSY
;
935 pmap_page_set_memattr(m
, memattr
);
939 * Inserts the given vm_page into the object and object list.
941 * The pagetables are not updated but will presumably fault the page
942 * in if necessary, or if a kernel page the caller will at some point
943 * enter the page into the kernel's pmap. We are not allowed to block
944 * here so we *can't* do this anyway.
946 * This routine may not block.
947 * This routine must be called with the vm_object held.
948 * This routine must be called with a critical section held.
950 * This routine returns TRUE if the page was inserted into the object
951 * successfully, and FALSE if the page already exists in the object.
954 vm_page_insert(vm_page_t m
, vm_object_t object
, vm_pindex_t pindex
)
956 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(object
));
957 if (m
->object
!= NULL
)
958 panic("vm_page_insert: already inserted");
960 object
->generation
++;
963 * Record the object/offset pair in this page and add the
964 * pv_list_count of the page to the object.
966 * The vm_page spin lock is required for interactions with the pmap.
968 vm_page_spin_lock(m
);
971 if (vm_page_rb_tree_RB_INSERT(&object
->rb_memq
, m
)) {
974 vm_page_spin_unlock(m
);
977 ++object
->resident_page_count
;
978 ++mycpu
->gd_vmtotal
.t_rm
;
979 /* atomic_add_int(&object->agg_pv_list_count, m->md.pv_list_count); */
980 vm_page_spin_unlock(m
);
983 * Since we are inserting a new and possibly dirty page,
984 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
986 if ((m
->valid
& m
->dirty
) ||
987 (m
->flags
& (PG_WRITEABLE
| PG_NEED_COMMIT
)))
988 vm_object_set_writeable_dirty(object
);
991 * Checks for a swap assignment and sets PG_SWAPPED if appropriate.
993 swap_pager_page_inserted(m
);
998 * Removes the given vm_page_t from the (object,index) table
1000 * The underlying pmap entry (if any) is NOT removed here.
1001 * This routine may not block.
1003 * The page must be BUSY and will remain BUSY on return.
1004 * No other requirements.
1006 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave
1010 vm_page_remove(vm_page_t m
)
1014 if (m
->object
== NULL
) {
1018 if ((m
->flags
& PG_BUSY
) == 0)
1019 panic("vm_page_remove: page not busy");
1023 vm_object_hold(object
);
1026 * Remove the page from the object and update the object.
1028 * The vm_page spin lock is required for interactions with the pmap.
1030 vm_page_spin_lock(m
);
1031 vm_page_rb_tree_RB_REMOVE(&object
->rb_memq
, m
);
1032 --object
->resident_page_count
;
1033 --mycpu
->gd_vmtotal
.t_rm
;
1034 /* atomic_add_int(&object->agg_pv_list_count, -m->md.pv_list_count); */
1036 vm_page_spin_unlock(m
);
1038 object
->generation
++;
1040 vm_object_drop(object
);
1044 * Locate and return the page at (object, pindex), or NULL if the
1045 * page could not be found.
1047 * The caller must hold the vm_object token.
1050 vm_page_lookup(vm_object_t object
, vm_pindex_t pindex
)
1055 * Search the hash table for this object/offset pair
1057 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
1058 m
= vm_page_rb_tree_RB_LOOKUP(&object
->rb_memq
, pindex
);
1059 KKASSERT(m
== NULL
|| (m
->object
== object
&& m
->pindex
== pindex
));
1064 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait
)(struct vm_object
*object
,
1066 int also_m_busy
, const char *msg
1072 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
1073 m
= vm_page_rb_tree_RB_LOOKUP(&object
->rb_memq
, pindex
);
1075 KKASSERT(m
->object
== object
&& m
->pindex
== pindex
);
1078 if (flags
& PG_BUSY
) {
1079 tsleep_interlock(m
, 0);
1080 if (atomic_cmpset_int(&m
->flags
, flags
,
1081 flags
| PG_WANTED
| PG_REFERENCED
)) {
1082 tsleep(m
, PINTERLOCKED
, msg
, 0);
1083 m
= vm_page_rb_tree_RB_LOOKUP(&object
->rb_memq
,
1086 } else if (also_m_busy
&& (flags
& PG_SBUSY
)) {
1087 tsleep_interlock(m
, 0);
1088 if (atomic_cmpset_int(&m
->flags
, flags
,
1089 flags
| PG_WANTED
| PG_REFERENCED
)) {
1090 tsleep(m
, PINTERLOCKED
, msg
, 0);
1091 m
= vm_page_rb_tree_RB_LOOKUP(&object
->rb_memq
,
1094 } else if (atomic_cmpset_int(&m
->flags
, flags
,
1096 #ifdef VM_PAGE_DEBUG
1097 m
->busy_func
= func
;
1098 m
->busy_line
= lineno
;
1107 * Attempt to lookup and busy a page.
1109 * Returns NULL if the page could not be found
1111 * Returns a vm_page and error == TRUE if the page exists but could not
1114 * Returns a vm_page and error == FALSE on success.
1117 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try
)(struct vm_object
*object
,
1119 int also_m_busy
, int *errorp
1125 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
1126 m
= vm_page_rb_tree_RB_LOOKUP(&object
->rb_memq
, pindex
);
1129 KKASSERT(m
->object
== object
&& m
->pindex
== pindex
);
1132 if (flags
& PG_BUSY
) {
1136 if (also_m_busy
&& (flags
& PG_SBUSY
)) {
1140 if (atomic_cmpset_int(&m
->flags
, flags
, flags
| PG_BUSY
)) {
1141 #ifdef VM_PAGE_DEBUG
1142 m
->busy_func
= func
;
1143 m
->busy_line
= lineno
;
1152 * Caller must hold the related vm_object
1155 vm_page_next(vm_page_t m
)
1159 next
= vm_page_rb_tree_RB_NEXT(m
);
1160 if (next
&& next
->pindex
!= m
->pindex
+ 1)
1168 * Move the given vm_page from its current object to the specified
1169 * target object/offset. The page must be busy and will remain so
1172 * new_object must be held.
1173 * This routine might block. XXX ?
1175 * NOTE: Swap associated with the page must be invalidated by the move. We
1176 * have to do this for several reasons: (1) we aren't freeing the
1177 * page, (2) we are dirtying the page, (3) the VM system is probably
1178 * moving the page from object A to B, and will then later move
1179 * the backing store from A to B and we can't have a conflict.
1181 * NOTE: We *always* dirty the page. It is necessary both for the
1182 * fact that we moved it, and because we may be invalidating
1183 * swap. If the page is on the cache, we have to deactivate it
1184 * or vm_page_dirty() will panic. Dirty pages are not allowed
1188 vm_page_rename(vm_page_t m
, vm_object_t new_object
, vm_pindex_t new_pindex
)
1190 KKASSERT(m
->flags
& PG_BUSY
);
1191 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(new_object
));
1193 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(m
->object
));
1196 if (vm_page_insert(m
, new_object
, new_pindex
) == FALSE
) {
1197 panic("vm_page_rename: target exists (%p,%"PRIu64
")",
1198 new_object
, new_pindex
);
1200 if (m
->queue
- m
->pc
== PQ_CACHE
)
1201 vm_page_deactivate(m
);
1206 * vm_page_unqueue() without any wakeup. This routine is used when a page
1207 * is to remain BUSYied by the caller.
1209 * This routine may not block.
1212 vm_page_unqueue_nowakeup(vm_page_t m
)
1214 vm_page_and_queue_spin_lock(m
);
1215 (void)_vm_page_rem_queue_spinlocked(m
);
1216 vm_page_spin_unlock(m
);
1220 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
1223 * This routine may not block.
1226 vm_page_unqueue(vm_page_t m
)
1230 vm_page_and_queue_spin_lock(m
);
1231 queue
= _vm_page_rem_queue_spinlocked(m
);
1232 if (queue
== PQ_FREE
|| queue
== PQ_CACHE
) {
1233 vm_page_spin_unlock(m
);
1234 pagedaemon_wakeup();
1236 vm_page_spin_unlock(m
);
1241 * vm_page_list_find()
1243 * Find a page on the specified queue with color optimization.
1245 * The page coloring optimization attempts to locate a page that does
1246 * not overload other nearby pages in the object in the cpu's L1 or L2
1247 * caches. We need this optimization because cpu caches tend to be
1248 * physical caches, while object spaces tend to be virtual.
1250 * On MP systems each PQ_FREE and PQ_CACHE color queue has its own spinlock
1251 * and the algorithm is adjusted to localize allocations on a per-core basis.
1252 * This is done by 'twisting' the colors.
1254 * The page is returned spinlocked and removed from its queue (it will
1255 * be on PQ_NONE), or NULL. The page is not PG_BUSY'd. The caller
1256 * is responsible for dealing with the busy-page case (usually by
1257 * deactivating the page and looping).
1259 * NOTE: This routine is carefully inlined. A non-inlined version
1260 * is available for outside callers but the only critical path is
1261 * from within this source file.
1263 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE
1264 * represent stable storage, allowing us to order our locks vm_page
1265 * first, then queue.
1269 _vm_page_list_find(int basequeue
, int index
, boolean_t prefer_zero
)
1275 m
= TAILQ_LAST(&vm_page_queues
[basequeue
+index
].pl
, pglist
);
1277 m
= TAILQ_FIRST(&vm_page_queues
[basequeue
+index
].pl
);
1279 m
= _vm_page_list_find2(basequeue
, index
);
1282 vm_page_and_queue_spin_lock(m
);
1283 if (m
->queue
== basequeue
+ index
) {
1284 _vm_page_rem_queue_spinlocked(m
);
1285 /* vm_page_t spin held, no queue spin */
1288 vm_page_and_queue_spin_unlock(m
);
1294 _vm_page_list_find2(int basequeue
, int index
)
1298 struct vpgqueues
*pq
;
1300 pq
= &vm_page_queues
[basequeue
];
1303 * Note that for the first loop, index+i and index-i wind up at the
1304 * same place. Even though this is not totally optimal, we've already
1305 * blown it by missing the cache case so we do not care.
1307 for (i
= PQ_L2_SIZE
/ 2; i
> 0; --i
) {
1309 m
= TAILQ_FIRST(&pq
[(index
+ i
) & PQ_L2_MASK
].pl
);
1311 _vm_page_and_queue_spin_lock(m
);
1313 basequeue
+ ((index
+ i
) & PQ_L2_MASK
)) {
1314 _vm_page_rem_queue_spinlocked(m
);
1317 _vm_page_and_queue_spin_unlock(m
);
1320 m
= TAILQ_FIRST(&pq
[(index
- i
) & PQ_L2_MASK
].pl
);
1322 _vm_page_and_queue_spin_lock(m
);
1324 basequeue
+ ((index
- i
) & PQ_L2_MASK
)) {
1325 _vm_page_rem_queue_spinlocked(m
);
1328 _vm_page_and_queue_spin_unlock(m
);
1338 * Returns a vm_page candidate for allocation. The page is not busied so
1339 * it can move around. The caller must busy the page (and typically
1340 * deactivate it if it cannot be busied!)
1342 * Returns a spinlocked vm_page that has been removed from its queue.
1345 vm_page_list_find(int basequeue
, int index
, boolean_t prefer_zero
)
1347 return(_vm_page_list_find(basequeue
, index
, prefer_zero
));
1351 * Find a page on the cache queue with color optimization, remove it
1352 * from the queue, and busy it. The returned page will not be spinlocked.
1354 * A candidate failure will be deactivated. Candidates can fail due to
1355 * being busied by someone else, in which case they will be deactivated.
1357 * This routine may not block.
1361 vm_page_select_cache(u_short pg_color
)
1366 m
= _vm_page_list_find(PQ_CACHE
, pg_color
& PQ_L2_MASK
, FALSE
);
1370 * (m) has been removed from its queue and spinlocked
1372 if (vm_page_busy_try(m
, TRUE
)) {
1373 _vm_page_deactivate_locked(m
, 0);
1374 vm_page_spin_unlock(m
);
1377 * We successfully busied the page
1379 if ((m
->flags
& (PG_UNMANAGED
| PG_NEED_COMMIT
)) == 0 &&
1380 m
->hold_count
== 0 &&
1381 m
->wire_count
== 0 &&
1382 (m
->dirty
& m
->valid
) == 0) {
1383 vm_page_spin_unlock(m
);
1384 pagedaemon_wakeup();
1389 * The page cannot be recycled, deactivate it.
1391 _vm_page_deactivate_locked(m
, 0);
1392 if (_vm_page_wakeup(m
)) {
1393 vm_page_spin_unlock(m
);
1396 vm_page_spin_unlock(m
);
1404 * Find a free or zero page, with specified preference. We attempt to
1405 * inline the nominal case and fall back to _vm_page_select_free()
1406 * otherwise. A busied page is removed from the queue and returned.
1408 * This routine may not block.
1410 static __inline vm_page_t
1411 vm_page_select_free(u_short pg_color
, boolean_t prefer_zero
)
1416 m
= _vm_page_list_find(PQ_FREE
, pg_color
& PQ_L2_MASK
,
1420 if (vm_page_busy_try(m
, TRUE
)) {
1422 * Various mechanisms such as a pmap_collect can
1423 * result in a busy page on the free queue. We
1424 * have to move the page out of the way so we can
1425 * retry the allocation. If the other thread is not
1426 * allocating the page then m->valid will remain 0 and
1427 * the pageout daemon will free the page later on.
1429 * Since we could not busy the page, however, we
1430 * cannot make assumptions as to whether the page
1431 * will be allocated by the other thread or not,
1432 * so all we can do is deactivate it to move it out
1433 * of the way. In particular, if the other thread
1434 * wires the page it may wind up on the inactive
1435 * queue and the pageout daemon will have to deal
1436 * with that case too.
1438 _vm_page_deactivate_locked(m
, 0);
1439 vm_page_spin_unlock(m
);
1442 * Theoretically if we are able to busy the page
1443 * atomic with the queue removal (using the vm_page
1444 * lock) nobody else should be able to mess with the
1447 KKASSERT((m
->flags
& (PG_UNMANAGED
|
1448 PG_NEED_COMMIT
)) == 0);
1449 KASSERT(m
->hold_count
== 0, ("m->hold_count is not zero "
1450 "pg %p q=%d flags=%08x hold=%d wire=%d",
1451 m
, m
->queue
, m
->flags
, m
->hold_count
, m
->wire_count
));
1452 KKASSERT(m
->wire_count
== 0);
1453 vm_page_spin_unlock(m
);
1454 pagedaemon_wakeup();
1456 /* return busied and removed page */
1464 * This implements a per-cpu cache of free, zero'd, ready-to-go pages.
1465 * The idea is to populate this cache prior to acquiring any locks so
1466 * we don't wind up potentially zeroing VM pages (under heavy loads) while
1467 * holding potentialy contending locks.
1469 * Note that we allocate the page uninserted into anything and use a pindex
1470 * of 0, the vm_page_alloc() will effectively add gd_cpuid so these
1471 * allocations should wind up being uncontended. However, we still want
1472 * to rove across PQ_L2_SIZE.
1475 vm_page_pcpu_cache(void)
1478 globaldata_t gd
= mycpu
;
1481 if (gd
->gd_vmpg_count
< GD_MINVMPG
) {
1483 while (gd
->gd_vmpg_count
< GD_MAXVMPG
) {
1484 m
= vm_page_alloc(NULL
, ticks
& ~ncpus2_mask
,
1485 VM_ALLOC_NULL_OK
| VM_ALLOC_NORMAL
|
1486 VM_ALLOC_NULL_OK
| VM_ALLOC_ZERO
);
1487 if (gd
->gd_vmpg_count
< GD_MAXVMPG
) {
1488 if ((m
->flags
& PG_ZERO
) == 0) {
1489 pmap_zero_page(VM_PAGE_TO_PHYS(m
));
1490 vm_page_flag_set(m
, PG_ZERO
);
1492 gd
->gd_vmpg_array
[gd
->gd_vmpg_count
++] = m
;
1505 * Allocate and return a memory cell associated with this VM object/offset
1506 * pair. If object is NULL an unassociated page will be allocated.
1508 * The returned page will be busied and removed from its queues. This
1509 * routine can block and may return NULL if a race occurs and the page
1510 * is found to already exist at the specified (object, pindex).
1512 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain
1513 * VM_ALLOC_QUICK like normal but cannot use cache
1514 * VM_ALLOC_SYSTEM greater free drain
1515 * VM_ALLOC_INTERRUPT allow free list to be completely drained
1516 * VM_ALLOC_ZERO advisory request for pre-zero'd page only
1517 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only
1518 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision
1519 * (see vm_page_grab())
1520 * VM_ALLOC_USE_GD ok to use per-gd cache
1522 * The object must be held if not NULL
1523 * This routine may not block
1525 * Additional special handling is required when called from an interrupt
1526 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache
1530 vm_page_alloc(vm_object_t object
, vm_pindex_t pindex
, int page_req
)
1532 globaldata_t gd
= mycpu
;
1539 * Special per-cpu free VM page cache. The pages are pre-busied
1540 * and pre-zerod for us.
1542 if (gd
->gd_vmpg_count
&& (page_req
& VM_ALLOC_USE_GD
)) {
1544 if (gd
->gd_vmpg_count
) {
1545 m
= gd
->gd_vmpg_array
[--gd
->gd_vmpg_count
];
1555 * Cpu twist - cpu localization algorithm
1558 pg_color
= gd
->gd_cpuid
+ (pindex
& ~ncpus_fit_mask
) +
1559 (object
->pg_color
& ~ncpus_fit_mask
);
1561 pg_color
= gd
->gd_cpuid
+ (pindex
& ~ncpus_fit_mask
);
1564 (VM_ALLOC_NORMAL
|VM_ALLOC_QUICK
|
1565 VM_ALLOC_INTERRUPT
|VM_ALLOC_SYSTEM
));
1568 * Certain system threads (pageout daemon, buf_daemon's) are
1569 * allowed to eat deeper into the free page list.
1571 if (curthread
->td_flags
& TDF_SYSTHREAD
)
1572 page_req
|= VM_ALLOC_SYSTEM
;
1575 * Impose various limitations. Note that the v_free_reserved test
1576 * must match the opposite of vm_page_count_target() to avoid
1577 * livelocks, be careful.
1580 if (vmstats
.v_free_count
>= vmstats
.v_free_reserved
||
1581 ((page_req
& VM_ALLOC_INTERRUPT
) && vmstats
.v_free_count
> 0) ||
1582 ((page_req
& VM_ALLOC_SYSTEM
) && vmstats
.v_cache_count
== 0 &&
1583 vmstats
.v_free_count
> vmstats
.v_interrupt_free_min
)
1586 * The free queue has sufficient free pages to take one out.
1588 if (page_req
& (VM_ALLOC_ZERO
| VM_ALLOC_FORCE_ZERO
))
1589 m
= vm_page_select_free(pg_color
, TRUE
);
1591 m
= vm_page_select_free(pg_color
, FALSE
);
1592 } else if (page_req
& VM_ALLOC_NORMAL
) {
1594 * Allocatable from the cache (non-interrupt only). On
1595 * success, we must free the page and try again, thus
1596 * ensuring that vmstats.v_*_free_min counters are replenished.
1599 if (curthread
->td_preempted
) {
1600 kprintf("vm_page_alloc(): warning, attempt to allocate"
1601 " cache page from preempting interrupt\n");
1604 m
= vm_page_select_cache(pg_color
);
1607 m
= vm_page_select_cache(pg_color
);
1610 * On success move the page into the free queue and loop.
1612 * Only do this if we can safely acquire the vm_object lock,
1613 * because this is effectively a random page and the caller
1614 * might be holding the lock shared, we don't want to
1618 KASSERT(m
->dirty
== 0,
1619 ("Found dirty cache page %p", m
));
1620 if ((obj
= m
->object
) != NULL
) {
1621 if (vm_object_hold_try(obj
)) {
1622 vm_page_protect(m
, VM_PROT_NONE
);
1624 /* m->object NULL here */
1625 vm_object_drop(obj
);
1627 vm_page_deactivate(m
);
1631 vm_page_protect(m
, VM_PROT_NONE
);
1638 * On failure return NULL
1640 #if defined(DIAGNOSTIC)
1641 if (vmstats
.v_cache_count
> 0)
1642 kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats
.v_cache_count
);
1644 vm_pageout_deficit
++;
1645 pagedaemon_wakeup();
1649 * No pages available, wakeup the pageout daemon and give up.
1651 vm_pageout_deficit
++;
1652 pagedaemon_wakeup();
1657 * v_free_count can race so loop if we don't find the expected
1664 * Good page found. The page has already been busied for us and
1665 * removed from its queues.
1667 KASSERT(m
->dirty
== 0,
1668 ("vm_page_alloc: free/cache page %p was dirty", m
));
1669 KKASSERT(m
->queue
== PQ_NONE
);
1675 * Initialize the structure, inheriting some flags but clearing
1676 * all the rest. The page has already been busied for us.
1678 vm_page_flag_clear(m
, ~(PG_ZERO
| PG_BUSY
| PG_SBUSY
));
1679 KKASSERT(m
->wire_count
== 0);
1680 KKASSERT(m
->busy
== 0);
1685 * Caller must be holding the object lock (asserted by
1686 * vm_page_insert()).
1688 * NOTE: Inserting a page here does not insert it into any pmaps
1689 * (which could cause us to block allocating memory).
1691 * NOTE: If no object an unassociated page is allocated, m->pindex
1692 * can be used by the caller for any purpose.
1695 if (vm_page_insert(m
, object
, pindex
) == FALSE
) {
1697 if ((page_req
& VM_ALLOC_NULL_OK
) == 0)
1698 panic("PAGE RACE %p[%ld]/%p",
1699 object
, (long)pindex
, m
);
1707 * Don't wakeup too often - wakeup the pageout daemon when
1708 * we would be nearly out of memory.
1710 pagedaemon_wakeup();
1713 * A PG_BUSY page is returned.
1719 * Returns number of pages available in our DMA memory reserve
1720 * (adjusted with vm.dma_reserved=<value>m in /boot/loader.conf)
1723 vm_contig_avail_pages(void)
1728 spin_lock(&vm_contig_spin
);
1729 bfree
= alist_free_info(&vm_contig_alist
, &blk
, &count
);
1730 spin_unlock(&vm_contig_spin
);
1736 * Attempt to allocate contiguous physical memory with the specified
1740 vm_page_alloc_contig(vm_paddr_t low
, vm_paddr_t high
,
1741 unsigned long alignment
, unsigned long boundary
,
1742 unsigned long size
, vm_memattr_t memattr
)
1748 alignment
>>= PAGE_SHIFT
;
1751 boundary
>>= PAGE_SHIFT
;
1754 size
= (size
+ PAGE_MASK
) >> PAGE_SHIFT
;
1756 spin_lock(&vm_contig_spin
);
1757 blk
= alist_alloc(&vm_contig_alist
, 0, size
);
1758 if (blk
== ALIST_BLOCK_NONE
) {
1759 spin_unlock(&vm_contig_spin
);
1761 kprintf("vm_page_alloc_contig: %ldk nospace\n",
1762 (size
+ PAGE_MASK
) * (PAGE_SIZE
/ 1024));
1766 if (high
&& ((vm_paddr_t
)(blk
+ size
) << PAGE_SHIFT
) > high
) {
1767 alist_free(&vm_contig_alist
, blk
, size
);
1768 spin_unlock(&vm_contig_spin
);
1770 kprintf("vm_page_alloc_contig: %ldk high "
1772 (size
+ PAGE_MASK
) * (PAGE_SIZE
/ 1024),
1777 spin_unlock(&vm_contig_spin
);
1778 if (vm_contig_verbose
) {
1779 kprintf("vm_page_alloc_contig: %016jx/%ldk\n",
1780 (intmax_t)(vm_paddr_t
)blk
<< PAGE_SHIFT
,
1781 (size
+ PAGE_MASK
) * (PAGE_SIZE
/ 1024));
1784 m
= PHYS_TO_VM_PAGE((vm_paddr_t
)blk
<< PAGE_SHIFT
);
1785 if (memattr
!= VM_MEMATTR_DEFAULT
)
1786 for (i
= 0;i
< size
;i
++)
1787 pmap_page_set_memattr(&m
[i
], memattr
);
1792 * Free contiguously allocated pages. The pages will be wired but not busy.
1793 * When freeing to the alist we leave them wired and not busy.
1796 vm_page_free_contig(vm_page_t m
, unsigned long size
)
1798 vm_paddr_t pa
= VM_PAGE_TO_PHYS(m
);
1799 vm_pindex_t start
= pa
>> PAGE_SHIFT
;
1800 vm_pindex_t pages
= (size
+ PAGE_MASK
) >> PAGE_SHIFT
;
1802 if (vm_contig_verbose
) {
1803 kprintf("vm_page_free_contig: %016jx/%ldk\n",
1804 (intmax_t)pa
, size
/ 1024);
1806 if (pa
< vm_low_phys_reserved
) {
1807 KKASSERT(pa
+ size
<= vm_low_phys_reserved
);
1808 spin_lock(&vm_contig_spin
);
1809 alist_free(&vm_contig_alist
, start
, pages
);
1810 spin_unlock(&vm_contig_spin
);
1813 vm_page_busy_wait(m
, FALSE
, "cpgfr");
1814 vm_page_unwire(m
, 0);
1825 * Wait for sufficient free memory for nominal heavy memory use kernel
1828 * WARNING! Be sure never to call this in any vm_pageout code path, which
1829 * will trivially deadlock the system.
1832 vm_wait_nominal(void)
1834 while (vm_page_count_min(0))
1839 * Test if vm_wait_nominal() would block.
1842 vm_test_nominal(void)
1844 if (vm_page_count_min(0))
1850 * Block until free pages are available for allocation, called in various
1851 * places before memory allocations.
1853 * The caller may loop if vm_page_count_min() == FALSE so we cannot be
1854 * more generous then that.
1860 * never wait forever
1864 lwkt_gettoken(&vm_token
);
1866 if (curthread
== pagethread
) {
1868 * The pageout daemon itself needs pages, this is bad.
1870 if (vm_page_count_min(0)) {
1871 vm_pageout_pages_needed
= 1;
1872 tsleep(&vm_pageout_pages_needed
, 0, "VMWait", timo
);
1876 * Wakeup the pageout daemon if necessary and wait.
1878 * Do not wait indefinitely for the target to be reached,
1879 * as load might prevent it from being reached any time soon.
1880 * But wait a little to try to slow down page allocations
1881 * and to give more important threads (the pagedaemon)
1882 * allocation priority.
1884 if (vm_page_count_target()) {
1885 if (vm_pages_needed
== 0) {
1886 vm_pages_needed
= 1;
1887 wakeup(&vm_pages_needed
);
1889 ++vm_pages_waiting
; /* SMP race ok */
1890 tsleep(&vmstats
.v_free_count
, 0, "vmwait", timo
);
1893 lwkt_reltoken(&vm_token
);
1897 * Block until free pages are available for allocation
1899 * Called only from vm_fault so that processes page faulting can be
1903 vm_wait_pfault(void)
1906 * Wakeup the pageout daemon if necessary and wait.
1908 * Do not wait indefinitely for the target to be reached,
1909 * as load might prevent it from being reached any time soon.
1910 * But wait a little to try to slow down page allocations
1911 * and to give more important threads (the pagedaemon)
1912 * allocation priority.
1914 if (vm_page_count_min(0)) {
1915 lwkt_gettoken(&vm_token
);
1916 while (vm_page_count_severe()) {
1917 if (vm_page_count_target()) {
1918 if (vm_pages_needed
== 0) {
1919 vm_pages_needed
= 1;
1920 wakeup(&vm_pages_needed
);
1922 ++vm_pages_waiting
; /* SMP race ok */
1923 tsleep(&vmstats
.v_free_count
, 0, "pfault", hz
);
1926 lwkt_reltoken(&vm_token
);
1931 * Put the specified page on the active list (if appropriate). Ensure
1932 * that act_count is at least ACT_INIT but do not otherwise mess with it.
1934 * The caller should be holding the page busied ? XXX
1935 * This routine may not block.
1938 vm_page_activate(vm_page_t m
)
1942 vm_page_spin_lock(m
);
1943 if (m
->queue
- m
->pc
!= PQ_ACTIVE
) {
1944 _vm_page_queue_spin_lock(m
);
1945 oqueue
= _vm_page_rem_queue_spinlocked(m
);
1946 /* page is left spinlocked, queue is unlocked */
1948 if (oqueue
== PQ_CACHE
)
1949 mycpu
->gd_cnt
.v_reactivated
++;
1950 if (m
->wire_count
== 0 && (m
->flags
& PG_UNMANAGED
) == 0) {
1951 if (m
->act_count
< ACT_INIT
)
1952 m
->act_count
= ACT_INIT
;
1953 _vm_page_add_queue_spinlocked(m
, PQ_ACTIVE
+ m
->pc
, 0);
1955 _vm_page_and_queue_spin_unlock(m
);
1956 if (oqueue
== PQ_CACHE
|| oqueue
== PQ_FREE
)
1957 pagedaemon_wakeup();
1959 if (m
->act_count
< ACT_INIT
)
1960 m
->act_count
= ACT_INIT
;
1961 vm_page_spin_unlock(m
);
1966 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
1967 * routine is called when a page has been added to the cache or free
1970 * This routine may not block.
1972 static __inline
void
1973 vm_page_free_wakeup(void)
1976 * If the pageout daemon itself needs pages, then tell it that
1977 * there are some free.
1979 if (vm_pageout_pages_needed
&&
1980 vmstats
.v_cache_count
+ vmstats
.v_free_count
>=
1981 vmstats
.v_pageout_free_min
1983 vm_pageout_pages_needed
= 0;
1984 wakeup(&vm_pageout_pages_needed
);
1988 * Wakeup processes that are waiting on memory.
1990 * Generally speaking we want to wakeup stuck processes as soon as
1991 * possible. !vm_page_count_min(0) is the absolute minimum point
1992 * where we can do this. Wait a bit longer to reduce degenerate
1993 * re-blocking (vm_page_free_hysteresis). The target check is just
1994 * to make sure the min-check w/hysteresis does not exceed the
1997 if (vm_pages_waiting
) {
1998 if (!vm_page_count_min(vm_page_free_hysteresis
) ||
1999 !vm_page_count_target()) {
2000 vm_pages_waiting
= 0;
2001 wakeup(&vmstats
.v_free_count
);
2002 ++mycpu
->gd_cnt
.v_ppwakeups
;
2005 if (!vm_page_count_target()) {
2007 * Plenty of pages are free, wakeup everyone.
2009 vm_pages_waiting
= 0;
2010 wakeup(&vmstats
.v_free_count
);
2011 ++mycpu
->gd_cnt
.v_ppwakeups
;
2012 } else if (!vm_page_count_min(0)) {
2014 * Some pages are free, wakeup someone.
2016 int wcount
= vm_pages_waiting
;
2019 vm_pages_waiting
= wcount
;
2020 wakeup_one(&vmstats
.v_free_count
);
2021 ++mycpu
->gd_cnt
.v_ppwakeups
;
2028 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates
2029 * it from its VM object.
2031 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on
2032 * return (the page will have been freed).
2035 vm_page_free_toq(vm_page_t m
)
2037 mycpu
->gd_cnt
.v_tfree
++;
2038 KKASSERT((m
->flags
& PG_MAPPED
) == 0);
2039 KKASSERT(m
->flags
& PG_BUSY
);
2041 if (m
->busy
|| ((m
->queue
- m
->pc
) == PQ_FREE
)) {
2042 kprintf("vm_page_free: pindex(%lu), busy(%d), "
2043 "PG_BUSY(%d), hold(%d)\n",
2044 (u_long
)m
->pindex
, m
->busy
,
2045 ((m
->flags
& PG_BUSY
) ? 1 : 0), m
->hold_count
);
2046 if ((m
->queue
- m
->pc
) == PQ_FREE
)
2047 panic("vm_page_free: freeing free page");
2049 panic("vm_page_free: freeing busy page");
2053 * Remove from object, spinlock the page and its queues and
2054 * remove from any queue. No queue spinlock will be held
2055 * after this section (because the page was removed from any
2059 vm_page_and_queue_spin_lock(m
);
2060 _vm_page_rem_queue_spinlocked(m
);
2063 * No further management of fictitious pages occurs beyond object
2064 * and queue removal.
2066 if ((m
->flags
& PG_FICTITIOUS
) != 0) {
2067 vm_page_spin_unlock(m
);
2075 if (m
->wire_count
!= 0) {
2076 if (m
->wire_count
> 1) {
2078 "vm_page_free: invalid wire count (%d), pindex: 0x%lx",
2079 m
->wire_count
, (long)m
->pindex
);
2081 panic("vm_page_free: freeing wired page");
2085 * Clear the UNMANAGED flag when freeing an unmanaged page.
2086 * Clear the NEED_COMMIT flag
2088 if (m
->flags
& PG_UNMANAGED
)
2089 vm_page_flag_clear(m
, PG_UNMANAGED
);
2090 if (m
->flags
& PG_NEED_COMMIT
)
2091 vm_page_flag_clear(m
, PG_NEED_COMMIT
);
2093 if (m
->hold_count
!= 0) {
2094 vm_page_flag_clear(m
, PG_ZERO
);
2095 _vm_page_add_queue_spinlocked(m
, PQ_HOLD
+ m
->pc
, 0);
2097 _vm_page_add_queue_spinlocked(m
, PQ_FREE
+ m
->pc
, 0);
2101 * This sequence allows us to clear PG_BUSY while still holding
2102 * its spin lock, which reduces contention vs allocators. We
2103 * must not leave the queue locked or _vm_page_wakeup() may
2106 _vm_page_queue_spin_unlock(m
);
2107 if (_vm_page_wakeup(m
)) {
2108 vm_page_spin_unlock(m
);
2111 vm_page_spin_unlock(m
);
2113 vm_page_free_wakeup();
2117 * vm_page_free_fromq_fast()
2119 * Remove a non-zero page from one of the free queues; the page is removed for
2120 * zeroing, so do not issue a wakeup.
2123 vm_page_free_fromq_fast(void)
2129 for (i
= 0; i
< PQ_L2_SIZE
; ++i
) {
2130 m
= vm_page_list_find(PQ_FREE
, qi
, FALSE
);
2131 /* page is returned spinlocked and removed from its queue */
2133 if (vm_page_busy_try(m
, TRUE
)) {
2135 * We were unable to busy the page, deactivate
2138 _vm_page_deactivate_locked(m
, 0);
2139 vm_page_spin_unlock(m
);
2140 } else if (m
->flags
& PG_ZERO
) {
2142 * The page is already PG_ZERO, requeue it and loop
2144 _vm_page_add_queue_spinlocked(m
,
2147 vm_page_queue_spin_unlock(m
);
2148 if (_vm_page_wakeup(m
)) {
2149 vm_page_spin_unlock(m
);
2152 vm_page_spin_unlock(m
);
2156 * The page is not PG_ZERO'd so return it.
2158 KKASSERT((m
->flags
& (PG_UNMANAGED
|
2159 PG_NEED_COMMIT
)) == 0);
2160 KKASSERT(m
->hold_count
== 0);
2161 KKASSERT(m
->wire_count
== 0);
2162 vm_page_spin_unlock(m
);
2167 qi
= (qi
+ PQ_PRIME2
) & PQ_L2_MASK
;
2173 * vm_page_unmanage()
2175 * Prevent PV management from being done on the page. The page is
2176 * removed from the paging queues as if it were wired, and as a
2177 * consequence of no longer being managed the pageout daemon will not
2178 * touch it (since there is no way to locate the pte mappings for the
2179 * page). madvise() calls that mess with the pmap will also no longer
2180 * operate on the page.
2182 * Beyond that the page is still reasonably 'normal'. Freeing the page
2183 * will clear the flag.
2185 * This routine is used by OBJT_PHYS objects - objects using unswappable
2186 * physical memory as backing store rather then swap-backed memory and
2187 * will eventually be extended to support 4MB unmanaged physical
2190 * Caller must be holding the page busy.
2193 vm_page_unmanage(vm_page_t m
)
2195 KKASSERT(m
->flags
& PG_BUSY
);
2196 if ((m
->flags
& PG_UNMANAGED
) == 0) {
2197 if (m
->wire_count
== 0)
2200 vm_page_flag_set(m
, PG_UNMANAGED
);
2204 * Mark this page as wired down by yet another map, removing it from
2205 * paging queues as necessary.
2207 * Caller must be holding the page busy.
2210 vm_page_wire(vm_page_t m
)
2213 * Only bump the wire statistics if the page is not already wired,
2214 * and only unqueue the page if it is on some queue (if it is unmanaged
2215 * it is already off the queues). Don't do anything with fictitious
2216 * pages because they are always wired.
2218 KKASSERT(m
->flags
& PG_BUSY
);
2219 if ((m
->flags
& PG_FICTITIOUS
) == 0) {
2220 if (atomic_fetchadd_int(&m
->wire_count
, 1) == 0) {
2221 if ((m
->flags
& PG_UNMANAGED
) == 0)
2223 atomic_add_int(&vmstats
.v_wire_count
, 1);
2225 KASSERT(m
->wire_count
!= 0,
2226 ("vm_page_wire: wire_count overflow m=%p", m
));
2231 * Release one wiring of this page, potentially enabling it to be paged again.
2233 * Many pages placed on the inactive queue should actually go
2234 * into the cache, but it is difficult to figure out which. What
2235 * we do instead, if the inactive target is well met, is to put
2236 * clean pages at the head of the inactive queue instead of the tail.
2237 * This will cause them to be moved to the cache more quickly and
2238 * if not actively re-referenced, freed more quickly. If we just
2239 * stick these pages at the end of the inactive queue, heavy filesystem
2240 * meta-data accesses can cause an unnecessary paging load on memory bound
2241 * processes. This optimization causes one-time-use metadata to be
2242 * reused more quickly.
2244 * Pages marked PG_NEED_COMMIT are always activated and never placed on
2245 * the inactive queue. This helps the pageout daemon determine memory
2246 * pressure and act on out-of-memory situations more quickly.
2248 * BUT, if we are in a low-memory situation we have no choice but to
2249 * put clean pages on the cache queue.
2251 * A number of routines use vm_page_unwire() to guarantee that the page
2252 * will go into either the inactive or active queues, and will NEVER
2253 * be placed in the cache - for example, just after dirtying a page.
2254 * dirty pages in the cache are not allowed.
2256 * This routine may not block.
2259 vm_page_unwire(vm_page_t m
, int activate
)
2261 KKASSERT(m
->flags
& PG_BUSY
);
2262 if (m
->flags
& PG_FICTITIOUS
) {
2264 } else if (m
->wire_count
<= 0) {
2265 panic("vm_page_unwire: invalid wire count: %d", m
->wire_count
);
2267 if (atomic_fetchadd_int(&m
->wire_count
, -1) == 1) {
2268 atomic_add_int(&vmstats
.v_wire_count
, -1);
2269 if (m
->flags
& PG_UNMANAGED
) {
2271 } else if (activate
|| (m
->flags
& PG_NEED_COMMIT
)) {
2272 vm_page_spin_lock(m
);
2273 _vm_page_add_queue_spinlocked(m
,
2274 PQ_ACTIVE
+ m
->pc
, 0);
2275 _vm_page_and_queue_spin_unlock(m
);
2277 vm_page_spin_lock(m
);
2278 vm_page_flag_clear(m
, PG_WINATCFLS
);
2279 _vm_page_add_queue_spinlocked(m
,
2280 PQ_INACTIVE
+ m
->pc
, 0);
2281 ++vm_swapcache_inactive_heuristic
;
2282 _vm_page_and_queue_spin_unlock(m
);
2289 * Move the specified page to the inactive queue. If the page has
2290 * any associated swap, the swap is deallocated.
2292 * Normally athead is 0 resulting in LRU operation. athead is set
2293 * to 1 if we want this page to be 'as if it were placed in the cache',
2294 * except without unmapping it from the process address space.
2296 * vm_page's spinlock must be held on entry and will remain held on return.
2297 * This routine may not block.
2300 _vm_page_deactivate_locked(vm_page_t m
, int athead
)
2305 * Ignore if already inactive.
2307 if (m
->queue
- m
->pc
== PQ_INACTIVE
)
2309 _vm_page_queue_spin_lock(m
);
2310 oqueue
= _vm_page_rem_queue_spinlocked(m
);
2312 if (m
->wire_count
== 0 && (m
->flags
& PG_UNMANAGED
) == 0) {
2313 if (oqueue
== PQ_CACHE
)
2314 mycpu
->gd_cnt
.v_reactivated
++;
2315 vm_page_flag_clear(m
, PG_WINATCFLS
);
2316 _vm_page_add_queue_spinlocked(m
, PQ_INACTIVE
+ m
->pc
, athead
);
2318 ++vm_swapcache_inactive_heuristic
;
2320 /* NOTE: PQ_NONE if condition not taken */
2321 _vm_page_queue_spin_unlock(m
);
2322 /* leaves vm_page spinlocked */
2326 * Attempt to deactivate a page.
2331 vm_page_deactivate(vm_page_t m
)
2333 vm_page_spin_lock(m
);
2334 _vm_page_deactivate_locked(m
, 0);
2335 vm_page_spin_unlock(m
);
2339 vm_page_deactivate_locked(vm_page_t m
)
2341 _vm_page_deactivate_locked(m
, 0);
2345 * Attempt to move a page to PQ_CACHE.
2347 * Returns 0 on failure, 1 on success
2349 * The page should NOT be busied by the caller. This function will validate
2350 * whether the page can be safely moved to the cache.
2353 vm_page_try_to_cache(vm_page_t m
)
2355 vm_page_spin_lock(m
);
2356 if (vm_page_busy_try(m
, TRUE
)) {
2357 vm_page_spin_unlock(m
);
2360 if (m
->dirty
|| m
->hold_count
|| m
->wire_count
||
2361 (m
->flags
& (PG_UNMANAGED
| PG_NEED_COMMIT
))) {
2362 if (_vm_page_wakeup(m
)) {
2363 vm_page_spin_unlock(m
);
2366 vm_page_spin_unlock(m
);
2370 vm_page_spin_unlock(m
);
2373 * Page busied by us and no longer spinlocked. Dirty pages cannot
2374 * be moved to the cache.
2376 vm_page_test_dirty(m
);
2377 if (m
->dirty
|| (m
->flags
& PG_NEED_COMMIT
)) {
2386 * Attempt to free the page. If we cannot free it, we do nothing.
2387 * 1 is returned on success, 0 on failure.
2392 vm_page_try_to_free(vm_page_t m
)
2394 vm_page_spin_lock(m
);
2395 if (vm_page_busy_try(m
, TRUE
)) {
2396 vm_page_spin_unlock(m
);
2401 * The page can be in any state, including already being on the free
2402 * queue. Check to see if it really can be freed.
2404 if (m
->dirty
|| /* can't free if it is dirty */
2405 m
->hold_count
|| /* or held (XXX may be wrong) */
2406 m
->wire_count
|| /* or wired */
2407 (m
->flags
& (PG_UNMANAGED
| /* or unmanaged */
2408 PG_NEED_COMMIT
)) || /* or needs a commit */
2409 m
->queue
- m
->pc
== PQ_FREE
|| /* already on PQ_FREE */
2410 m
->queue
- m
->pc
== PQ_HOLD
) { /* already on PQ_HOLD */
2411 if (_vm_page_wakeup(m
)) {
2412 vm_page_spin_unlock(m
);
2415 vm_page_spin_unlock(m
);
2419 vm_page_spin_unlock(m
);
2422 * We can probably free the page.
2424 * Page busied by us and no longer spinlocked. Dirty pages will
2425 * not be freed by this function. We have to re-test the
2426 * dirty bit after cleaning out the pmaps.
2428 vm_page_test_dirty(m
);
2429 if (m
->dirty
|| (m
->flags
& PG_NEED_COMMIT
)) {
2433 vm_page_protect(m
, VM_PROT_NONE
);
2434 if (m
->dirty
|| (m
->flags
& PG_NEED_COMMIT
)) {
2445 * Put the specified page onto the page cache queue (if appropriate).
2447 * The page must be busy, and this routine will release the busy and
2448 * possibly even free the page.
2451 vm_page_cache(vm_page_t m
)
2453 if ((m
->flags
& (PG_UNMANAGED
| PG_NEED_COMMIT
)) ||
2454 m
->busy
|| m
->wire_count
|| m
->hold_count
) {
2455 kprintf("vm_page_cache: attempting to cache busy/held page\n");
2461 * Already in the cache (and thus not mapped)
2463 if ((m
->queue
- m
->pc
) == PQ_CACHE
) {
2464 KKASSERT((m
->flags
& PG_MAPPED
) == 0);
2470 * Caller is required to test m->dirty, but note that the act of
2471 * removing the page from its maps can cause it to become dirty
2472 * on an SMP system due to another cpu running in usermode.
2475 panic("vm_page_cache: caching a dirty page, pindex: %ld",
2480 * Remove all pmaps and indicate that the page is not
2481 * writeable or mapped. Our vm_page_protect() call may
2482 * have blocked (especially w/ VM_PROT_NONE), so recheck
2485 vm_page_protect(m
, VM_PROT_NONE
);
2486 if ((m
->flags
& (PG_UNMANAGED
| PG_MAPPED
)) ||
2487 m
->busy
|| m
->wire_count
|| m
->hold_count
) {
2489 } else if (m
->dirty
|| (m
->flags
& PG_NEED_COMMIT
)) {
2490 vm_page_deactivate(m
);
2493 _vm_page_and_queue_spin_lock(m
);
2494 _vm_page_rem_queue_spinlocked(m
);
2495 _vm_page_add_queue_spinlocked(m
, PQ_CACHE
+ m
->pc
, 0);
2496 _vm_page_queue_spin_unlock(m
);
2497 if (_vm_page_wakeup(m
)) {
2498 vm_page_spin_unlock(m
);
2501 vm_page_spin_unlock(m
);
2503 vm_page_free_wakeup();
2508 * vm_page_dontneed()
2510 * Cache, deactivate, or do nothing as appropriate. This routine
2511 * is typically used by madvise() MADV_DONTNEED.
2513 * Generally speaking we want to move the page into the cache so
2514 * it gets reused quickly. However, this can result in a silly syndrome
2515 * due to the page recycling too quickly. Small objects will not be
2516 * fully cached. On the otherhand, if we move the page to the inactive
2517 * queue we wind up with a problem whereby very large objects
2518 * unnecessarily blow away our inactive and cache queues.
2520 * The solution is to move the pages based on a fixed weighting. We
2521 * either leave them alone, deactivate them, or move them to the cache,
2522 * where moving them to the cache has the highest weighting.
2523 * By forcing some pages into other queues we eventually force the
2524 * system to balance the queues, potentially recovering other unrelated
2525 * space from active. The idea is to not force this to happen too
2528 * The page must be busied.
2531 vm_page_dontneed(vm_page_t m
)
2533 static int dnweight
;
2540 * occassionally leave the page alone
2542 if ((dnw
& 0x01F0) == 0 ||
2543 m
->queue
- m
->pc
== PQ_INACTIVE
||
2544 m
->queue
- m
->pc
== PQ_CACHE
2546 if (m
->act_count
>= ACT_INIT
)
2552 * If vm_page_dontneed() is inactivating a page, it must clear
2553 * the referenced flag; otherwise the pagedaemon will see references
2554 * on the page in the inactive queue and reactivate it. Until the
2555 * page can move to the cache queue, madvise's job is not done.
2557 vm_page_flag_clear(m
, PG_REFERENCED
);
2558 pmap_clear_reference(m
);
2561 vm_page_test_dirty(m
);
2563 if (m
->dirty
|| (dnw
& 0x0070) == 0) {
2565 * Deactivate the page 3 times out of 32.
2570 * Cache the page 28 times out of every 32. Note that
2571 * the page is deactivated instead of cached, but placed
2572 * at the head of the queue instead of the tail.
2576 vm_page_spin_lock(m
);
2577 _vm_page_deactivate_locked(m
, head
);
2578 vm_page_spin_unlock(m
);
2582 * These routines manipulate the 'soft busy' count for a page. A soft busy
2583 * is almost like PG_BUSY except that it allows certain compatible operations
2584 * to occur on the page while it is busy. For example, a page undergoing a
2585 * write can still be mapped read-only.
2587 * Because vm_pages can overlap buffers m->busy can be > 1. m->busy is only
2588 * adjusted while the vm_page is PG_BUSY so the flash will occur when the
2589 * busy bit is cleared.
2592 vm_page_io_start(vm_page_t m
)
2594 KASSERT(m
->flags
& PG_BUSY
, ("vm_page_io_start: page not busy!!!"));
2595 atomic_add_char(&m
->busy
, 1);
2596 vm_page_flag_set(m
, PG_SBUSY
);
2600 vm_page_io_finish(vm_page_t m
)
2602 KASSERT(m
->flags
& PG_BUSY
, ("vm_page_io_finish: page not busy!!!"));
2603 atomic_subtract_char(&m
->busy
, 1);
2605 vm_page_flag_clear(m
, PG_SBUSY
);
2609 * Indicate that a clean VM page requires a filesystem commit and cannot
2610 * be reused. Used by tmpfs.
2613 vm_page_need_commit(vm_page_t m
)
2615 vm_page_flag_set(m
, PG_NEED_COMMIT
);
2616 vm_object_set_writeable_dirty(m
->object
);
2620 vm_page_clear_commit(vm_page_t m
)
2622 vm_page_flag_clear(m
, PG_NEED_COMMIT
);
2626 * Grab a page, blocking if it is busy and allocating a page if necessary.
2627 * A busy page is returned or NULL. The page may or may not be valid and
2628 * might not be on a queue (the caller is responsible for the disposition of
2631 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the
2632 * page will be zero'd and marked valid.
2634 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked
2635 * valid even if it already exists.
2637 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also
2638 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified.
2639 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified.
2641 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
2642 * always returned if we had blocked.
2644 * This routine may not be called from an interrupt.
2646 * PG_ZERO is *ALWAYS* cleared by this routine.
2648 * No other requirements.
2651 vm_page_grab(vm_object_t object
, vm_pindex_t pindex
, int allocflags
)
2657 KKASSERT(allocflags
&
2658 (VM_ALLOC_NORMAL
|VM_ALLOC_INTERRUPT
|VM_ALLOC_SYSTEM
));
2659 vm_object_hold_shared(object
);
2661 m
= vm_page_lookup_busy_try(object
, pindex
, TRUE
, &error
);
2663 vm_page_sleep_busy(m
, TRUE
, "pgrbwt");
2664 if ((allocflags
& VM_ALLOC_RETRY
) == 0) {
2669 } else if (m
== NULL
) {
2671 vm_object_upgrade(object
);
2674 if (allocflags
& VM_ALLOC_RETRY
)
2675 allocflags
|= VM_ALLOC_NULL_OK
;
2676 m
= vm_page_alloc(object
, pindex
,
2677 allocflags
& ~VM_ALLOC_RETRY
);
2681 if ((allocflags
& VM_ALLOC_RETRY
) == 0)
2690 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid.
2692 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set
2693 * valid even if already valid.
2695 if (m
->valid
== 0) {
2696 if (allocflags
& (VM_ALLOC_ZERO
| VM_ALLOC_FORCE_ZERO
)) {
2697 if ((m
->flags
& PG_ZERO
) == 0)
2698 pmap_zero_page(VM_PAGE_TO_PHYS(m
));
2699 m
->valid
= VM_PAGE_BITS_ALL
;
2701 } else if (allocflags
& VM_ALLOC_FORCE_ZERO
) {
2702 pmap_zero_page(VM_PAGE_TO_PHYS(m
));
2703 m
->valid
= VM_PAGE_BITS_ALL
;
2705 vm_page_flag_clear(m
, PG_ZERO
);
2707 vm_object_drop(object
);
2712 * Mapping function for valid bits or for dirty bits in
2713 * a page. May not block.
2715 * Inputs are required to range within a page.
2721 vm_page_bits(int base
, int size
)
2727 base
+ size
<= PAGE_SIZE
,
2728 ("vm_page_bits: illegal base/size %d/%d", base
, size
)
2731 if (size
== 0) /* handle degenerate case */
2734 first_bit
= base
>> DEV_BSHIFT
;
2735 last_bit
= (base
+ size
- 1) >> DEV_BSHIFT
;
2737 return ((2 << last_bit
) - (1 << first_bit
));
2741 * Sets portions of a page valid and clean. The arguments are expected
2742 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2743 * of any partial chunks touched by the range. The invalid portion of
2744 * such chunks will be zero'd.
2746 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically
2747 * align base to DEV_BSIZE so as not to mark clean a partially
2748 * truncated device block. Otherwise the dirty page status might be
2751 * This routine may not block.
2753 * (base + size) must be less then or equal to PAGE_SIZE.
2756 _vm_page_zero_valid(vm_page_t m
, int base
, int size
)
2761 if (size
== 0) /* handle degenerate case */
2765 * If the base is not DEV_BSIZE aligned and the valid
2766 * bit is clear, we have to zero out a portion of the
2770 if ((frag
= base
& ~(DEV_BSIZE
- 1)) != base
&&
2771 (m
->valid
& (1 << (base
>> DEV_BSHIFT
))) == 0
2773 pmap_zero_page_area(
2781 * If the ending offset is not DEV_BSIZE aligned and the
2782 * valid bit is clear, we have to zero out a portion of
2786 endoff
= base
+ size
;
2788 if ((frag
= endoff
& ~(DEV_BSIZE
- 1)) != endoff
&&
2789 (m
->valid
& (1 << (endoff
>> DEV_BSHIFT
))) == 0
2791 pmap_zero_page_area(
2794 DEV_BSIZE
- (endoff
& (DEV_BSIZE
- 1))
2800 * Set valid, clear dirty bits. If validating the entire
2801 * page we can safely clear the pmap modify bit. We also
2802 * use this opportunity to clear the PG_NOSYNC flag. If a process
2803 * takes a write fault on a MAP_NOSYNC memory area the flag will
2806 * We set valid bits inclusive of any overlap, but we can only
2807 * clear dirty bits for DEV_BSIZE chunks that are fully within
2810 * Page must be busied?
2811 * No other requirements.
2814 vm_page_set_valid(vm_page_t m
, int base
, int size
)
2816 _vm_page_zero_valid(m
, base
, size
);
2817 m
->valid
|= vm_page_bits(base
, size
);
2822 * Set valid bits and clear dirty bits.
2824 * NOTE: This function does not clear the pmap modified bit.
2825 * Also note that e.g. NFS may use a byte-granular base
2828 * WARNING: Page must be busied? But vfs_clean_one_page() will call
2829 * this without necessarily busying the page (via bdwrite()).
2830 * So for now vm_token must also be held.
2832 * No other requirements.
2835 vm_page_set_validclean(vm_page_t m
, int base
, int size
)
2839 _vm_page_zero_valid(m
, base
, size
);
2840 pagebits
= vm_page_bits(base
, size
);
2841 m
->valid
|= pagebits
;
2842 m
->dirty
&= ~pagebits
;
2843 if (base
== 0 && size
== PAGE_SIZE
) {
2844 /*pmap_clear_modify(m);*/
2845 vm_page_flag_clear(m
, PG_NOSYNC
);
2850 * Set valid & dirty. Used by buwrite()
2852 * WARNING: Page must be busied? But vfs_dirty_one_page() will
2853 * call this function in buwrite() so for now vm_token must
2856 * No other requirements.
2859 vm_page_set_validdirty(vm_page_t m
, int base
, int size
)
2863 pagebits
= vm_page_bits(base
, size
);
2864 m
->valid
|= pagebits
;
2865 m
->dirty
|= pagebits
;
2867 vm_object_set_writeable_dirty(m
->object
);
2873 * NOTE: This function does not clear the pmap modified bit.
2874 * Also note that e.g. NFS may use a byte-granular base
2877 * Page must be busied?
2878 * No other requirements.
2881 vm_page_clear_dirty(vm_page_t m
, int base
, int size
)
2883 m
->dirty
&= ~vm_page_bits(base
, size
);
2884 if (base
== 0 && size
== PAGE_SIZE
) {
2885 /*pmap_clear_modify(m);*/
2886 vm_page_flag_clear(m
, PG_NOSYNC
);
2891 * Make the page all-dirty.
2893 * Also make sure the related object and vnode reflect the fact that the
2894 * object may now contain a dirty page.
2896 * Page must be busied?
2897 * No other requirements.
2900 vm_page_dirty(vm_page_t m
)
2903 int pqtype
= m
->queue
- m
->pc
;
2905 KASSERT(pqtype
!= PQ_CACHE
&& pqtype
!= PQ_FREE
,
2906 ("vm_page_dirty: page in free/cache queue!"));
2907 if (m
->dirty
!= VM_PAGE_BITS_ALL
) {
2908 m
->dirty
= VM_PAGE_BITS_ALL
;
2910 vm_object_set_writeable_dirty(m
->object
);
2915 * Invalidates DEV_BSIZE'd chunks within a page. Both the
2916 * valid and dirty bits for the effected areas are cleared.
2918 * Page must be busied?
2920 * No other requirements.
2923 vm_page_set_invalid(vm_page_t m
, int base
, int size
)
2927 bits
= vm_page_bits(base
, size
);
2930 m
->object
->generation
++;
2934 * The kernel assumes that the invalid portions of a page contain
2935 * garbage, but such pages can be mapped into memory by user code.
2936 * When this occurs, we must zero out the non-valid portions of the
2937 * page so user code sees what it expects.
2939 * Pages are most often semi-valid when the end of a file is mapped
2940 * into memory and the file's size is not page aligned.
2942 * Page must be busied?
2943 * No other requirements.
2946 vm_page_zero_invalid(vm_page_t m
, boolean_t setvalid
)
2952 * Scan the valid bits looking for invalid sections that
2953 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
2954 * valid bit may be set ) have already been zerod by
2955 * vm_page_set_validclean().
2957 for (b
= i
= 0; i
<= PAGE_SIZE
/ DEV_BSIZE
; ++i
) {
2958 if (i
== (PAGE_SIZE
/ DEV_BSIZE
) ||
2959 (m
->valid
& (1 << i
))
2962 pmap_zero_page_area(
2965 (i
- b
) << DEV_BSHIFT
2973 * setvalid is TRUE when we can safely set the zero'd areas
2974 * as being valid. We can do this if there are no cache consistency
2975 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
2978 m
->valid
= VM_PAGE_BITS_ALL
;
2982 * Is a (partial) page valid? Note that the case where size == 0
2983 * will return FALSE in the degenerate case where the page is entirely
2984 * invalid, and TRUE otherwise.
2987 * No other requirements.
2990 vm_page_is_valid(vm_page_t m
, int base
, int size
)
2992 int bits
= vm_page_bits(base
, size
);
2994 if (m
->valid
&& ((m
->valid
& bits
) == bits
))
3001 * update dirty bits from pmap/mmu. May not block.
3003 * Caller must hold the page busy
3006 vm_page_test_dirty(vm_page_t m
)
3008 if ((m
->dirty
!= VM_PAGE_BITS_ALL
) && pmap_is_modified(m
)) {
3014 * Register an action, associating it with its vm_page
3017 vm_page_register_action(vm_page_action_t action
, vm_page_event_t event
)
3019 struct vm_page_action_list
*list
;
3022 hv
= (int)((intptr_t)action
->m
>> 8) & VMACTION_HMASK
;
3023 list
= &action_list
[hv
];
3025 lwkt_gettoken(&vm_token
);
3026 vm_page_flag_set(action
->m
, PG_ACTIONLIST
);
3027 action
->event
= event
;
3028 LIST_INSERT_HEAD(list
, action
, entry
);
3029 lwkt_reltoken(&vm_token
);
3033 * Unregister an action, disassociating it from its related vm_page
3036 vm_page_unregister_action(vm_page_action_t action
)
3038 struct vm_page_action_list
*list
;
3041 lwkt_gettoken(&vm_token
);
3042 if (action
->event
!= VMEVENT_NONE
) {
3043 action
->event
= VMEVENT_NONE
;
3044 LIST_REMOVE(action
, entry
);
3046 hv
= (int)((intptr_t)action
->m
>> 8) & VMACTION_HMASK
;
3047 list
= &action_list
[hv
];
3048 if (LIST_EMPTY(list
))
3049 vm_page_flag_clear(action
->m
, PG_ACTIONLIST
);
3051 lwkt_reltoken(&vm_token
);
3055 * Issue an event on a VM page. Corresponding action structures are
3056 * removed from the page's list and called.
3058 * If the vm_page has no more pending action events we clear its
3059 * PG_ACTIONLIST flag.
3062 vm_page_event_internal(vm_page_t m
, vm_page_event_t event
)
3064 struct vm_page_action_list
*list
;
3065 struct vm_page_action
*scan
;
3066 struct vm_page_action
*next
;
3070 hv
= (int)((intptr_t)m
>> 8) & VMACTION_HMASK
;
3071 list
= &action_list
[hv
];
3074 lwkt_gettoken(&vm_token
);
3075 LIST_FOREACH_MUTABLE(scan
, list
, entry
, next
) {
3077 if (scan
->event
== event
) {
3078 scan
->event
= VMEVENT_NONE
;
3079 LIST_REMOVE(scan
, entry
);
3080 scan
->func(m
, scan
);
3088 vm_page_flag_clear(m
, PG_ACTIONLIST
);
3089 lwkt_reltoken(&vm_token
);
3092 #include "opt_ddb.h"
3094 #include <sys/kernel.h>
3096 #include <ddb/ddb.h>
3098 DB_SHOW_COMMAND(page
, vm_page_print_page_info
)
3100 db_printf("vmstats.v_free_count: %d\n", vmstats
.v_free_count
);
3101 db_printf("vmstats.v_cache_count: %d\n", vmstats
.v_cache_count
);
3102 db_printf("vmstats.v_inactive_count: %d\n", vmstats
.v_inactive_count
);
3103 db_printf("vmstats.v_active_count: %d\n", vmstats
.v_active_count
);
3104 db_printf("vmstats.v_wire_count: %d\n", vmstats
.v_wire_count
);
3105 db_printf("vmstats.v_free_reserved: %d\n", vmstats
.v_free_reserved
);
3106 db_printf("vmstats.v_free_min: %d\n", vmstats
.v_free_min
);
3107 db_printf("vmstats.v_free_target: %d\n", vmstats
.v_free_target
);
3108 db_printf("vmstats.v_cache_min: %d\n", vmstats
.v_cache_min
);
3109 db_printf("vmstats.v_inactive_target: %d\n", vmstats
.v_inactive_target
);
3112 DB_SHOW_COMMAND(pageq
, vm_page_print_pageq_info
)
3115 db_printf("PQ_FREE:");
3116 for(i
=0;i
<PQ_L2_SIZE
;i
++) {
3117 db_printf(" %d", vm_page_queues
[PQ_FREE
+ i
].lcnt
);
3121 db_printf("PQ_CACHE:");
3122 for(i
=0;i
<PQ_L2_SIZE
;i
++) {
3123 db_printf(" %d", vm_page_queues
[PQ_CACHE
+ i
].lcnt
);
3127 db_printf("PQ_ACTIVE:");
3128 for(i
=0;i
<PQ_L2_SIZE
;i
++) {
3129 db_printf(" %d", vm_page_queues
[PQ_ACTIVE
+ i
].lcnt
);
3133 db_printf("PQ_INACTIVE:");
3134 for(i
=0;i
<PQ_L2_SIZE
;i
++) {
3135 db_printf(" %d", vm_page_queues
[PQ_INACTIVE
+ i
].lcnt
);