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>
78 #include <sys/cpu_topology.h>
81 #include <vm/vm_param.h>
83 #include <vm/vm_kern.h>
85 #include <vm/vm_map.h>
86 #include <vm/vm_object.h>
87 #include <vm/vm_page.h>
88 #include <vm/vm_pageout.h>
89 #include <vm/vm_pager.h>
90 #include <vm/vm_extern.h>
91 #include <vm/swap_pager.h>
93 #include <machine/inttypes.h>
94 #include <machine/md_var.h>
95 #include <machine/specialreg.h>
97 #include <vm/vm_page2.h>
98 #include <sys/spinlock2.h>
100 #define VMACTION_HSIZE 256
101 #define VMACTION_HMASK (VMACTION_HSIZE - 1)
103 static void vm_page_queue_init(void);
104 static void vm_page_free_wakeup(void);
105 static vm_page_t
vm_page_select_cache(u_short pg_color
);
106 static vm_page_t
_vm_page_list_find2(int basequeue
, int index
);
107 static void _vm_page_deactivate_locked(vm_page_t m
, int athead
);
110 * Array of tailq lists
112 __cachealign
struct vpgqueues vm_page_queues
[PQ_COUNT
];
114 LIST_HEAD(vm_page_action_list
, vm_page_action
);
115 struct vm_page_action_list action_list
[VMACTION_HSIZE
];
116 static volatile int vm_pages_waiting
;
118 static struct alist vm_contig_alist
;
119 static struct almeta vm_contig_ameta
[ALIST_RECORDS_65536
];
120 static struct spinlock vm_contig_spin
= SPINLOCK_INITIALIZER(&vm_contig_spin
, "vm_contig_spin");
122 static u_long vm_dma_reserved
= 0;
123 TUNABLE_ULONG("vm.dma_reserved", &vm_dma_reserved
);
124 SYSCTL_ULONG(_vm
, OID_AUTO
, dma_reserved
, CTLFLAG_RD
, &vm_dma_reserved
, 0,
125 "Memory reserved for DMA");
126 SYSCTL_UINT(_vm
, OID_AUTO
, dma_free_pages
, CTLFLAG_RD
,
127 &vm_contig_alist
.bl_free
, 0, "Memory reserved for DMA");
129 static int vm_contig_verbose
= 0;
130 TUNABLE_INT("vm.contig_verbose", &vm_contig_verbose
);
132 RB_GENERATE2(vm_page_rb_tree
, vm_page
, rb_entry
, rb_vm_page_compare
,
133 vm_pindex_t
, pindex
);
136 vm_page_queue_init(void)
140 for (i
= 0; i
< PQ_L2_SIZE
; i
++)
141 vm_page_queues
[PQ_FREE
+i
].cnt
= &vmstats
.v_free_count
;
142 for (i
= 0; i
< PQ_L2_SIZE
; i
++)
143 vm_page_queues
[PQ_CACHE
+i
].cnt
= &vmstats
.v_cache_count
;
144 for (i
= 0; i
< PQ_L2_SIZE
; i
++)
145 vm_page_queues
[PQ_INACTIVE
+i
].cnt
= &vmstats
.v_inactive_count
;
146 for (i
= 0; i
< PQ_L2_SIZE
; i
++)
147 vm_page_queues
[PQ_ACTIVE
+i
].cnt
= &vmstats
.v_active_count
;
148 for (i
= 0; i
< PQ_L2_SIZE
; i
++)
149 vm_page_queues
[PQ_HOLD
+i
].cnt
= &vmstats
.v_active_count
;
150 /* PQ_NONE has no queue */
152 for (i
= 0; i
< PQ_COUNT
; i
++) {
153 TAILQ_INIT(&vm_page_queues
[i
].pl
);
154 spin_init(&vm_page_queues
[i
].spin
, "vm_page_queue_init");
157 for (i
= 0; i
< VMACTION_HSIZE
; i
++)
158 LIST_INIT(&action_list
[i
]);
162 * note: place in initialized data section? Is this necessary?
165 int vm_page_array_size
= 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
];
238 /* too expensive time-wise in large-mem configurations */
239 if ((vpq
->flipflop
& 15) == 0) {
240 pmap_zero_page(VM_PAGE_TO_PHYS(m
));
242 TAILQ_INSERT_TAIL(&vpq
->pl
, m
, pageq
);
246 TAILQ_INSERT_HEAD(&vpq
->pl
, m
, pageq
);
257 * Initializes the resident memory module.
259 * Preallocates memory for critical VM structures and arrays prior to
260 * kernel_map becoming available.
262 * Memory is allocated from (virtual2_start, virtual2_end) if available,
263 * otherwise memory is allocated from (virtual_start, virtual_end).
265 * On x86-64 (virtual_start, virtual_end) is only 2GB and may not be
266 * large enough to hold vm_page_array & other structures for machines with
267 * large amounts of ram, so we want to use virtual2* when available.
270 vm_page_startup(void)
272 vm_offset_t vaddr
= virtual2_start
? virtual2_start
: virtual_start
;
275 vm_paddr_t page_range
;
282 vm_paddr_t biggestone
, biggestsize
;
289 vaddr
= round_page(vaddr
);
291 for (i
= 0; phys_avail
[i
+ 1]; i
+= 2) {
292 phys_avail
[i
] = round_page64(phys_avail
[i
]);
293 phys_avail
[i
+ 1] = trunc_page64(phys_avail
[i
+ 1]);
296 for (i
= 0; phys_avail
[i
+ 1]; i
+= 2) {
297 vm_paddr_t size
= phys_avail
[i
+ 1] - phys_avail
[i
];
299 if (size
> biggestsize
) {
307 end
= phys_avail
[biggestone
+1];
308 end
= trunc_page(end
);
311 * Initialize the queue headers for the free queue, the active queue
312 * and the inactive queue.
314 vm_page_queue_init();
316 #if !defined(_KERNEL_VIRTUAL)
318 * VKERNELs don't support minidumps and as such don't need
321 * Allocate a bitmap to indicate that a random physical page
322 * needs to be included in a minidump.
324 * The amd64 port needs this to indicate which direct map pages
325 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
327 * However, i386 still needs this workspace internally within the
328 * minidump code. In theory, they are not needed on i386, but are
329 * included should the sf_buf code decide to use them.
331 page_range
= phys_avail
[(nblocks
- 1) * 2 + 1] / PAGE_SIZE
;
332 vm_page_dump_size
= round_page(roundup2(page_range
, NBBY
) / NBBY
);
333 end
-= vm_page_dump_size
;
334 vm_page_dump
= (void *)pmap_map(&vaddr
, end
, end
+ vm_page_dump_size
,
335 VM_PROT_READ
| VM_PROT_WRITE
);
336 bzero((void *)vm_page_dump
, vm_page_dump_size
);
339 * Compute the number of pages of memory that will be available for
340 * use (taking into account the overhead of a page structure per
343 first_page
= phys_avail
[0] / PAGE_SIZE
;
344 page_range
= phys_avail
[(nblocks
- 1) * 2 + 1] / PAGE_SIZE
- first_page
;
345 npages
= (total
- (page_range
* sizeof(struct vm_page
))) / PAGE_SIZE
;
347 #ifndef _KERNEL_VIRTUAL
349 * (only applies to real kernels)
351 * Reserve a large amount of low memory for potential 32-bit DMA
352 * space allocations. Once device initialization is complete we
353 * release most of it, but keep (vm_dma_reserved) memory reserved
354 * for later use. Typically for X / graphics. Through trial and
355 * error we find that GPUs usually requires ~60-100MB or so.
357 * By default, 128M is left in reserve on machines with 2G+ of ram.
359 vm_low_phys_reserved
= (vm_paddr_t
)65536 << PAGE_SHIFT
;
360 if (vm_low_phys_reserved
> total
/ 4)
361 vm_low_phys_reserved
= total
/ 4;
362 if (vm_dma_reserved
== 0) {
363 vm_dma_reserved
= 128 * 1024 * 1024; /* 128MB */
364 if (vm_dma_reserved
> total
/ 16)
365 vm_dma_reserved
= total
/ 16;
368 alist_init(&vm_contig_alist
, 65536, vm_contig_ameta
,
369 ALIST_RECORDS_65536
);
372 * Initialize the mem entry structures now, and put them in the free
375 new_end
= trunc_page(end
- page_range
* sizeof(struct vm_page
));
376 mapped
= pmap_map(&vaddr
, new_end
, end
, VM_PROT_READ
| VM_PROT_WRITE
);
377 vm_page_array
= (vm_page_t
)mapped
;
379 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL)
381 * since pmap_map on amd64 returns stuff out of a direct-map region,
382 * we have to manually add these pages to the minidump tracking so
383 * that they can be dumped, including the vm_page_array.
385 for (pa
= new_end
; pa
< phys_avail
[biggestone
+ 1]; pa
+= PAGE_SIZE
)
390 * Clear all of the page structures
392 bzero((caddr_t
) vm_page_array
, page_range
* sizeof(struct vm_page
));
393 vm_page_array_size
= page_range
;
396 * Construct the free queue(s) in ascending order (by physical
397 * address) so that the first 16MB of physical memory is allocated
398 * last rather than first. On large-memory machines, this avoids
399 * the exhaustion of low physical memory before isa_dmainit has run.
401 vmstats
.v_page_count
= 0;
402 vmstats
.v_free_count
= 0;
403 for (i
= 0; phys_avail
[i
+ 1] && npages
> 0; i
+= 2) {
408 last_pa
= phys_avail
[i
+ 1];
409 while (pa
< last_pa
&& npages
-- > 0) {
415 virtual2_start
= vaddr
;
417 virtual_start
= vaddr
;
421 * We tended to reserve a ton of memory for contigmalloc(). Now that most
422 * drivers have initialized we want to return most the remaining free
423 * reserve back to the VM page queues so they can be used for normal
426 * We leave vm_dma_reserved bytes worth of free pages in the reserve pool.
429 vm_page_startup_finish(void *dummy __unused
)
438 spin_lock(&vm_contig_spin
);
440 bfree
= alist_free_info(&vm_contig_alist
, &blk
, &count
);
441 if (bfree
<= vm_dma_reserved
/ PAGE_SIZE
)
447 * Figure out how much of the initial reserve we have to
448 * free in order to reach our target.
450 bfree
-= vm_dma_reserved
/ PAGE_SIZE
;
452 blk
+= count
- bfree
;
457 * Calculate the nearest power of 2 <= count.
459 for (xcount
= 1; xcount
<= count
; xcount
<<= 1)
462 blk
+= count
- xcount
;
466 * Allocate the pages from the alist, then free them to
467 * the normal VM page queues.
469 * Pages allocated from the alist are wired. We have to
470 * busy, unwire, and free them. We must also adjust
471 * vm_low_phys_reserved before freeing any pages to prevent
474 rblk
= alist_alloc(&vm_contig_alist
, blk
, count
);
476 kprintf("vm_page_startup_finish: Unable to return "
477 "dma space @0x%08x/%d -> 0x%08x\n",
481 atomic_add_int(&vmstats
.v_dma_pages
, -count
);
482 spin_unlock(&vm_contig_spin
);
484 m
= PHYS_TO_VM_PAGE((vm_paddr_t
)blk
<< PAGE_SHIFT
);
485 vm_low_phys_reserved
= VM_PAGE_TO_PHYS(m
);
487 vm_page_busy_wait(m
, FALSE
, "cpgfr");
488 vm_page_unwire(m
, 0);
493 spin_lock(&vm_contig_spin
);
495 spin_unlock(&vm_contig_spin
);
498 * Print out how much DMA space drivers have already allocated and
499 * how much is left over.
501 kprintf("DMA space used: %jdk, remaining available: %jdk\n",
502 (intmax_t)(vmstats
.v_dma_pages
- vm_contig_alist
.bl_free
) *
504 (intmax_t)vm_contig_alist
.bl_free
* (PAGE_SIZE
/ 1024));
506 SYSINIT(vm_pgend
, SI_SUB_PROC0_POST
, SI_ORDER_ANY
,
507 vm_page_startup_finish
, NULL
);
511 * Scan comparison function for Red-Black tree scans. An inclusive
512 * (start,end) is expected. Other fields are not used.
515 rb_vm_page_scancmp(struct vm_page
*p
, void *data
)
517 struct rb_vm_page_scan_info
*info
= data
;
519 if (p
->pindex
< info
->start_pindex
)
521 if (p
->pindex
> info
->end_pindex
)
527 rb_vm_page_compare(struct vm_page
*p1
, struct vm_page
*p2
)
529 if (p1
->pindex
< p2
->pindex
)
531 if (p1
->pindex
> p2
->pindex
)
537 vm_page_init(vm_page_t m
)
539 /* do nothing for now. Called from pmap_page_init() */
543 * Each page queue has its own spin lock, which is fairly optimal for
544 * allocating and freeing pages at least.
546 * The caller must hold the vm_page_spin_lock() before locking a vm_page's
547 * queue spinlock via this function. Also note that m->queue cannot change
548 * unless both the page and queue are locked.
552 _vm_page_queue_spin_lock(vm_page_t m
)
557 if (queue
!= PQ_NONE
) {
558 spin_lock(&vm_page_queues
[queue
].spin
);
559 KKASSERT(queue
== m
->queue
);
565 _vm_page_queue_spin_unlock(vm_page_t m
)
571 if (queue
!= PQ_NONE
)
572 spin_unlock(&vm_page_queues
[queue
].spin
);
577 _vm_page_queues_spin_lock(u_short queue
)
580 if (queue
!= PQ_NONE
)
581 spin_lock(&vm_page_queues
[queue
].spin
);
587 _vm_page_queues_spin_unlock(u_short queue
)
590 if (queue
!= PQ_NONE
)
591 spin_unlock(&vm_page_queues
[queue
].spin
);
595 vm_page_queue_spin_lock(vm_page_t m
)
597 _vm_page_queue_spin_lock(m
);
601 vm_page_queues_spin_lock(u_short queue
)
603 _vm_page_queues_spin_lock(queue
);
607 vm_page_queue_spin_unlock(vm_page_t m
)
609 _vm_page_queue_spin_unlock(m
);
613 vm_page_queues_spin_unlock(u_short queue
)
615 _vm_page_queues_spin_unlock(queue
);
619 * This locks the specified vm_page and its queue in the proper order
620 * (page first, then queue). The queue may change so the caller must
625 _vm_page_and_queue_spin_lock(vm_page_t m
)
627 vm_page_spin_lock(m
);
628 _vm_page_queue_spin_lock(m
);
633 _vm_page_and_queue_spin_unlock(vm_page_t m
)
635 _vm_page_queues_spin_unlock(m
->queue
);
636 vm_page_spin_unlock(m
);
640 vm_page_and_queue_spin_unlock(vm_page_t m
)
642 _vm_page_and_queue_spin_unlock(m
);
646 vm_page_and_queue_spin_lock(vm_page_t m
)
648 _vm_page_and_queue_spin_lock(m
);
652 * Helper function removes vm_page from its current queue.
653 * Returns the base queue the page used to be on.
655 * The vm_page and the queue must be spinlocked.
656 * This function will unlock the queue but leave the page spinlocked.
658 static __inline u_short
659 _vm_page_rem_queue_spinlocked(vm_page_t m
)
661 struct vpgqueues
*pq
;
666 if (queue
!= PQ_NONE
) {
667 pq
= &vm_page_queues
[queue
];
668 TAILQ_REMOVE(&pq
->pl
, m
, pageq
);
669 atomic_add_int(pq
->cnt
, -1);
673 if ((queue
- m
->pc
) == PQ_FREE
&& (m
->flags
& PG_ZERO
))
675 if ((queue
- m
->pc
) == PQ_CACHE
|| (queue
- m
->pc
) == PQ_FREE
)
677 vm_page_queues_spin_unlock(oqueue
); /* intended */
683 * Helper function places the vm_page on the specified queue.
685 * The vm_page must be spinlocked.
686 * This function will return with both the page and the queue locked.
689 _vm_page_add_queue_spinlocked(vm_page_t m
, u_short queue
, int athead
)
691 struct vpgqueues
*pq
;
693 KKASSERT(m
->queue
== PQ_NONE
);
695 if (queue
!= PQ_NONE
) {
696 vm_page_queues_spin_lock(queue
);
697 pq
= &vm_page_queues
[queue
];
699 atomic_add_int(pq
->cnt
, 1);
703 * Put zero'd pages on the end ( where we look for zero'd pages
704 * first ) and non-zerod pages at the head.
706 if (queue
- m
->pc
== PQ_FREE
) {
707 if (m
->flags
& PG_ZERO
) {
708 TAILQ_INSERT_TAIL(&pq
->pl
, m
, pageq
);
711 TAILQ_INSERT_HEAD(&pq
->pl
, m
, pageq
);
714 TAILQ_INSERT_HEAD(&pq
->pl
, m
, pageq
);
716 TAILQ_INSERT_TAIL(&pq
->pl
, m
, pageq
);
718 /* leave the queue spinlocked */
723 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE)
724 * m->busy is zero. Returns TRUE if it had to sleep, FALSE if we
725 * did not. Only one sleep call will be made before returning.
727 * This function does NOT busy the page and on return the page is not
728 * guaranteed to be available.
731 vm_page_sleep_busy(vm_page_t m
, int also_m_busy
, const char *msg
)
739 if ((flags
& PG_BUSY
) == 0 &&
740 (also_m_busy
== 0 || (flags
& PG_SBUSY
) == 0)) {
743 tsleep_interlock(m
, 0);
744 if (atomic_cmpset_int(&m
->flags
, flags
,
745 flags
| PG_WANTED
| PG_REFERENCED
)) {
746 tsleep(m
, PINTERLOCKED
, msg
, 0);
753 * Wait until PG_BUSY can be set, then set it. If also_m_busy is TRUE we
754 * also wait for m->busy to become 0 before setting PG_BUSY.
757 VM_PAGE_DEBUG_EXT(vm_page_busy_wait
)(vm_page_t m
,
758 int also_m_busy
, const char *msg
766 if (flags
& PG_BUSY
) {
767 tsleep_interlock(m
, 0);
768 if (atomic_cmpset_int(&m
->flags
, flags
,
769 flags
| PG_WANTED
| PG_REFERENCED
)) {
770 tsleep(m
, PINTERLOCKED
, msg
, 0);
772 } else if (also_m_busy
&& (flags
& PG_SBUSY
)) {
773 tsleep_interlock(m
, 0);
774 if (atomic_cmpset_int(&m
->flags
, flags
,
775 flags
| PG_WANTED
| PG_REFERENCED
)) {
776 tsleep(m
, PINTERLOCKED
, msg
, 0);
779 if (atomic_cmpset_int(&m
->flags
, flags
,
783 m
->busy_line
= lineno
;
792 * Attempt to set PG_BUSY. If also_m_busy is TRUE we only succeed if m->busy
795 * Returns non-zero on failure.
798 VM_PAGE_DEBUG_EXT(vm_page_busy_try
)(vm_page_t m
, int also_m_busy
808 if (also_m_busy
&& (flags
& PG_SBUSY
))
810 if (atomic_cmpset_int(&m
->flags
, flags
, flags
| PG_BUSY
)) {
813 m
->busy_line
= lineno
;
821 * Clear the PG_BUSY flag and return non-zero to indicate to the caller
822 * that a wakeup() should be performed.
824 * The vm_page must be spinlocked and will remain spinlocked on return.
825 * The related queue must NOT be spinlocked (which could deadlock us).
831 _vm_page_wakeup(vm_page_t m
)
838 if (atomic_cmpset_int(&m
->flags
, flags
,
839 flags
& ~(PG_BUSY
| PG_WANTED
))) {
843 return(flags
& PG_WANTED
);
847 * Clear the PG_BUSY flag and wakeup anyone waiting for the page. This
848 * is typically the last call you make on a page before moving onto
852 vm_page_wakeup(vm_page_t m
)
854 KASSERT(m
->flags
& PG_BUSY
, ("vm_page_wakeup: page not busy!!!"));
855 vm_page_spin_lock(m
);
856 if (_vm_page_wakeup(m
)) {
857 vm_page_spin_unlock(m
);
860 vm_page_spin_unlock(m
);
865 * Holding a page keeps it from being reused. Other parts of the system
866 * can still disassociate the page from its current object and free it, or
867 * perform read or write I/O on it and/or otherwise manipulate the page,
868 * but if the page is held the VM system will leave the page and its data
869 * intact and not reuse the page for other purposes until the last hold
870 * reference is released. (see vm_page_wire() if you want to prevent the
871 * page from being disassociated from its object too).
873 * The caller must still validate the contents of the page and, if necessary,
874 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete
875 * before manipulating the page.
877 * XXX get vm_page_spin_lock() here and move FREE->HOLD if necessary
880 vm_page_hold(vm_page_t m
)
882 vm_page_spin_lock(m
);
883 atomic_add_int(&m
->hold_count
, 1);
884 if (m
->queue
- m
->pc
== PQ_FREE
) {
885 _vm_page_queue_spin_lock(m
);
886 _vm_page_rem_queue_spinlocked(m
);
887 _vm_page_add_queue_spinlocked(m
, PQ_HOLD
+ m
->pc
, 0);
888 _vm_page_queue_spin_unlock(m
);
890 vm_page_spin_unlock(m
);
894 * The opposite of vm_page_hold(). If the page is on the HOLD queue
895 * it was freed while held and must be moved back to the FREE queue.
898 vm_page_unhold(vm_page_t m
)
900 KASSERT(m
->hold_count
> 0 && m
->queue
- m
->pc
!= PQ_FREE
,
901 ("vm_page_unhold: pg %p illegal hold_count (%d) or on FREE queue (%d)",
902 m
, m
->hold_count
, m
->queue
- m
->pc
));
903 vm_page_spin_lock(m
);
904 atomic_add_int(&m
->hold_count
, -1);
905 if (m
->hold_count
== 0 && m
->queue
- m
->pc
== PQ_HOLD
) {
906 _vm_page_queue_spin_lock(m
);
907 _vm_page_rem_queue_spinlocked(m
);
908 _vm_page_add_queue_spinlocked(m
, PQ_FREE
+ m
->pc
, 0);
909 _vm_page_queue_spin_unlock(m
);
911 vm_page_spin_unlock(m
);
917 * Create a fictitious page with the specified physical address and
918 * memory attribute. The memory attribute is the only the machine-
919 * dependent aspect of a fictitious page that must be initialized.
923 vm_page_initfake(vm_page_t m
, vm_paddr_t paddr
, vm_memattr_t memattr
)
926 if ((m
->flags
& PG_FICTITIOUS
) != 0) {
928 * The page's memattr might have changed since the
929 * previous initialization. Update the pmap to the
934 m
->phys_addr
= paddr
;
936 /* Fictitious pages don't use "segind". */
937 /* Fictitious pages don't use "order" or "pool". */
938 m
->flags
= PG_FICTITIOUS
| PG_UNMANAGED
| PG_BUSY
;
942 pmap_page_set_memattr(m
, memattr
);
946 * Inserts the given vm_page into the object and object list.
948 * The pagetables are not updated but will presumably fault the page
949 * in if necessary, or if a kernel page the caller will at some point
950 * enter the page into the kernel's pmap. We are not allowed to block
951 * here so we *can't* do this anyway.
953 * This routine may not block.
954 * This routine must be called with the vm_object held.
955 * This routine must be called with a critical section held.
957 * This routine returns TRUE if the page was inserted into the object
958 * successfully, and FALSE if the page already exists in the object.
961 vm_page_insert(vm_page_t m
, vm_object_t object
, vm_pindex_t pindex
)
963 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(object
));
964 if (m
->object
!= NULL
)
965 panic("vm_page_insert: already inserted");
967 object
->generation
++;
970 * Record the object/offset pair in this page and add the
971 * pv_list_count of the page to the object.
973 * The vm_page spin lock is required for interactions with the pmap.
975 vm_page_spin_lock(m
);
978 if (vm_page_rb_tree_RB_INSERT(&object
->rb_memq
, m
)) {
981 vm_page_spin_unlock(m
);
984 ++object
->resident_page_count
;
985 ++mycpu
->gd_vmtotal
.t_rm
;
986 /* atomic_add_int(&object->agg_pv_list_count, m->md.pv_list_count); */
987 vm_page_spin_unlock(m
);
990 * Since we are inserting a new and possibly dirty page,
991 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
993 if ((m
->valid
& m
->dirty
) ||
994 (m
->flags
& (PG_WRITEABLE
| PG_NEED_COMMIT
)))
995 vm_object_set_writeable_dirty(object
);
998 * Checks for a swap assignment and sets PG_SWAPPED if appropriate.
1000 swap_pager_page_inserted(m
);
1005 * Removes the given vm_page_t from the (object,index) table
1007 * The underlying pmap entry (if any) is NOT removed here.
1008 * This routine may not block.
1010 * The page must be BUSY and will remain BUSY on return.
1011 * No other requirements.
1013 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave
1017 vm_page_remove(vm_page_t m
)
1021 if (m
->object
== NULL
) {
1025 if ((m
->flags
& PG_BUSY
) == 0)
1026 panic("vm_page_remove: page not busy");
1030 vm_object_hold(object
);
1033 * Remove the page from the object and update the object.
1035 * The vm_page spin lock is required for interactions with the pmap.
1037 vm_page_spin_lock(m
);
1038 vm_page_rb_tree_RB_REMOVE(&object
->rb_memq
, m
);
1039 --object
->resident_page_count
;
1040 --mycpu
->gd_vmtotal
.t_rm
;
1041 /* atomic_add_int(&object->agg_pv_list_count, -m->md.pv_list_count); */
1043 vm_page_spin_unlock(m
);
1045 object
->generation
++;
1047 vm_object_drop(object
);
1051 * Locate and return the page at (object, pindex), or NULL if the
1052 * page could not be found.
1054 * The caller must hold the vm_object token.
1057 vm_page_lookup(vm_object_t object
, vm_pindex_t pindex
)
1062 * Search the hash table for this object/offset pair
1064 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
1065 m
= vm_page_rb_tree_RB_LOOKUP(&object
->rb_memq
, pindex
);
1066 KKASSERT(m
== NULL
|| (m
->object
== object
&& m
->pindex
== pindex
));
1071 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait
)(struct vm_object
*object
,
1073 int also_m_busy
, const char *msg
1079 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
1080 m
= vm_page_rb_tree_RB_LOOKUP(&object
->rb_memq
, pindex
);
1082 KKASSERT(m
->object
== object
&& m
->pindex
== pindex
);
1085 if (flags
& PG_BUSY
) {
1086 tsleep_interlock(m
, 0);
1087 if (atomic_cmpset_int(&m
->flags
, flags
,
1088 flags
| PG_WANTED
| PG_REFERENCED
)) {
1089 tsleep(m
, PINTERLOCKED
, msg
, 0);
1090 m
= vm_page_rb_tree_RB_LOOKUP(&object
->rb_memq
,
1093 } else if (also_m_busy
&& (flags
& PG_SBUSY
)) {
1094 tsleep_interlock(m
, 0);
1095 if (atomic_cmpset_int(&m
->flags
, flags
,
1096 flags
| PG_WANTED
| PG_REFERENCED
)) {
1097 tsleep(m
, PINTERLOCKED
, msg
, 0);
1098 m
= vm_page_rb_tree_RB_LOOKUP(&object
->rb_memq
,
1101 } else if (atomic_cmpset_int(&m
->flags
, flags
,
1103 #ifdef VM_PAGE_DEBUG
1104 m
->busy_func
= func
;
1105 m
->busy_line
= lineno
;
1114 * Attempt to lookup and busy a page.
1116 * Returns NULL if the page could not be found
1118 * Returns a vm_page and error == TRUE if the page exists but could not
1121 * Returns a vm_page and error == FALSE on success.
1124 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try
)(struct vm_object
*object
,
1126 int also_m_busy
, int *errorp
1132 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
1133 m
= vm_page_rb_tree_RB_LOOKUP(&object
->rb_memq
, pindex
);
1136 KKASSERT(m
->object
== object
&& m
->pindex
== pindex
);
1139 if (flags
& PG_BUSY
) {
1143 if (also_m_busy
&& (flags
& PG_SBUSY
)) {
1147 if (atomic_cmpset_int(&m
->flags
, flags
, flags
| PG_BUSY
)) {
1148 #ifdef VM_PAGE_DEBUG
1149 m
->busy_func
= func
;
1150 m
->busy_line
= lineno
;
1159 * Attempt to repurpose the passed-in page. If the passed-in page cannot
1160 * be repurposed it will be released, *must_reenter will be set to 1, and
1161 * this function will fall-through to vm_page_lookup_busy_try().
1163 * The passed-in page must be wired and not busy. The returned page will
1164 * be busied and not wired.
1166 * A different page may be returned. The returned page will be busied and
1169 * NULL can be returned. If so, the required page could not be busied.
1170 * The passed-in page will be unwired.
1173 vm_page_repurpose(struct vm_object
*object
, vm_pindex_t pindex
,
1174 int also_m_busy
, int *errorp
, vm_page_t m
,
1175 int *must_reenter
, int *iswired
)
1178 vm_page_busy_wait(m
, TRUE
, "biodep");
1179 if ((m
->flags
& (PG_UNMANAGED
| PG_MAPPED
| PG_FICTITIOUS
)) ||
1180 m
->busy
|| m
->wire_count
!= 1 || m
->hold_count
) {
1181 vm_page_unwire(m
, 0);
1183 /* fall through to normal lookup */
1184 } else if (m
->dirty
|| (m
->flags
& PG_NEED_COMMIT
)) {
1185 vm_page_unwire(m
, 0);
1186 vm_page_deactivate(m
);
1188 /* fall through to normal lookup */
1191 * We can safely repurpose the page. It should
1192 * already be unqueued.
1194 KKASSERT(m
->queue
== PQ_NONE
&& m
->dirty
== 0);
1198 if (vm_page_insert(m
, object
, pindex
)) {
1204 vm_page_unwire(m
, 0);
1206 /* fall through to normal lookup */
1211 m
= vm_page_lookup_busy_try(object
, pindex
, also_m_busy
, errorp
);
1217 * Caller must hold the related vm_object
1220 vm_page_next(vm_page_t m
)
1224 next
= vm_page_rb_tree_RB_NEXT(m
);
1225 if (next
&& next
->pindex
!= m
->pindex
+ 1)
1233 * Move the given vm_page from its current object to the specified
1234 * target object/offset. The page must be busy and will remain so
1237 * new_object must be held.
1238 * This routine might block. XXX ?
1240 * NOTE: Swap associated with the page must be invalidated by the move. We
1241 * have to do this for several reasons: (1) we aren't freeing the
1242 * page, (2) we are dirtying the page, (3) the VM system is probably
1243 * moving the page from object A to B, and will then later move
1244 * the backing store from A to B and we can't have a conflict.
1246 * NOTE: We *always* dirty the page. It is necessary both for the
1247 * fact that we moved it, and because we may be invalidating
1248 * swap. If the page is on the cache, we have to deactivate it
1249 * or vm_page_dirty() will panic. Dirty pages are not allowed
1253 vm_page_rename(vm_page_t m
, vm_object_t new_object
, vm_pindex_t new_pindex
)
1255 KKASSERT(m
->flags
& PG_BUSY
);
1256 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(new_object
));
1258 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(m
->object
));
1261 if (vm_page_insert(m
, new_object
, new_pindex
) == FALSE
) {
1262 panic("vm_page_rename: target exists (%p,%"PRIu64
")",
1263 new_object
, new_pindex
);
1265 if (m
->queue
- m
->pc
== PQ_CACHE
)
1266 vm_page_deactivate(m
);
1271 * vm_page_unqueue() without any wakeup. This routine is used when a page
1272 * is to remain BUSYied by the caller.
1274 * This routine may not block.
1277 vm_page_unqueue_nowakeup(vm_page_t m
)
1279 vm_page_and_queue_spin_lock(m
);
1280 (void)_vm_page_rem_queue_spinlocked(m
);
1281 vm_page_spin_unlock(m
);
1285 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
1288 * This routine may not block.
1291 vm_page_unqueue(vm_page_t m
)
1295 vm_page_and_queue_spin_lock(m
);
1296 queue
= _vm_page_rem_queue_spinlocked(m
);
1297 if (queue
== PQ_FREE
|| queue
== PQ_CACHE
) {
1298 vm_page_spin_unlock(m
);
1299 pagedaemon_wakeup();
1301 vm_page_spin_unlock(m
);
1306 * vm_page_list_find()
1308 * Find a page on the specified queue with color optimization.
1310 * The page coloring optimization attempts to locate a page that does
1311 * not overload other nearby pages in the object in the cpu's L1 or L2
1312 * caches. We need this optimization because cpu caches tend to be
1313 * physical caches, while object spaces tend to be virtual.
1315 * On MP systems each PQ_FREE and PQ_CACHE color queue has its own spinlock
1316 * and the algorithm is adjusted to localize allocations on a per-core basis.
1317 * This is done by 'twisting' the colors.
1319 * The page is returned spinlocked and removed from its queue (it will
1320 * be on PQ_NONE), or NULL. The page is not PG_BUSY'd. The caller
1321 * is responsible for dealing with the busy-page case (usually by
1322 * deactivating the page and looping).
1324 * NOTE: This routine is carefully inlined. A non-inlined version
1325 * is available for outside callers but the only critical path is
1326 * from within this source file.
1328 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE
1329 * represent stable storage, allowing us to order our locks vm_page
1330 * first, then queue.
1334 _vm_page_list_find(int basequeue
, int index
, boolean_t prefer_zero
)
1340 m
= TAILQ_LAST(&vm_page_queues
[basequeue
+index
].pl
, pglist
);
1342 m
= TAILQ_FIRST(&vm_page_queues
[basequeue
+index
].pl
);
1344 m
= _vm_page_list_find2(basequeue
, index
);
1347 vm_page_and_queue_spin_lock(m
);
1348 if (m
->queue
== basequeue
+ index
) {
1349 _vm_page_rem_queue_spinlocked(m
);
1350 /* vm_page_t spin held, no queue spin */
1353 vm_page_and_queue_spin_unlock(m
);
1359 _vm_page_list_find2(int basequeue
, int index
)
1363 struct vpgqueues
*pq
;
1365 pq
= &vm_page_queues
[basequeue
];
1368 * Note that for the first loop, index+i and index-i wind up at the
1369 * same place. Even though this is not totally optimal, we've already
1370 * blown it by missing the cache case so we do not care.
1372 * NOTE: Fan out from our starting index for localization purposes.
1374 for (i
= 1; i
<= PQ_L2_SIZE
/ 2; ++i
) {
1376 m
= TAILQ_FIRST(&pq
[(index
+ i
) & PQ_L2_MASK
].pl
);
1378 _vm_page_and_queue_spin_lock(m
);
1380 basequeue
+ ((index
+ i
) & PQ_L2_MASK
)) {
1381 _vm_page_rem_queue_spinlocked(m
);
1384 _vm_page_and_queue_spin_unlock(m
);
1387 m
= TAILQ_FIRST(&pq
[(index
- i
) & PQ_L2_MASK
].pl
);
1389 _vm_page_and_queue_spin_lock(m
);
1391 basequeue
+ ((index
- i
) & PQ_L2_MASK
)) {
1392 _vm_page_rem_queue_spinlocked(m
);
1395 _vm_page_and_queue_spin_unlock(m
);
1405 * Returns a vm_page candidate for allocation. The page is not busied so
1406 * it can move around. The caller must busy the page (and typically
1407 * deactivate it if it cannot be busied!)
1409 * Returns a spinlocked vm_page that has been removed from its queue.
1412 vm_page_list_find(int basequeue
, int index
, boolean_t prefer_zero
)
1414 return(_vm_page_list_find(basequeue
, index
, prefer_zero
));
1418 * Find a page on the cache queue with color optimization, remove it
1419 * from the queue, and busy it. The returned page will not be spinlocked.
1421 * A candidate failure will be deactivated. Candidates can fail due to
1422 * being busied by someone else, in which case they will be deactivated.
1424 * This routine may not block.
1428 vm_page_select_cache(u_short pg_color
)
1433 m
= _vm_page_list_find(PQ_CACHE
, pg_color
& PQ_L2_MASK
, FALSE
);
1437 * (m) has been removed from its queue and spinlocked
1439 if (vm_page_busy_try(m
, TRUE
)) {
1440 _vm_page_deactivate_locked(m
, 0);
1441 vm_page_spin_unlock(m
);
1444 * We successfully busied the page
1446 if ((m
->flags
& (PG_UNMANAGED
| PG_NEED_COMMIT
)) == 0 &&
1447 m
->hold_count
== 0 &&
1448 m
->wire_count
== 0 &&
1449 (m
->dirty
& m
->valid
) == 0) {
1450 vm_page_spin_unlock(m
);
1451 pagedaemon_wakeup();
1456 * The page cannot be recycled, deactivate it.
1458 _vm_page_deactivate_locked(m
, 0);
1459 if (_vm_page_wakeup(m
)) {
1460 vm_page_spin_unlock(m
);
1463 vm_page_spin_unlock(m
);
1471 * Find a free or zero page, with specified preference. We attempt to
1472 * inline the nominal case and fall back to _vm_page_select_free()
1473 * otherwise. A busied page is removed from the queue and returned.
1475 * This routine may not block.
1477 static __inline vm_page_t
1478 vm_page_select_free(u_short pg_color
, boolean_t prefer_zero
)
1483 m
= _vm_page_list_find(PQ_FREE
, pg_color
& PQ_L2_MASK
,
1487 if (vm_page_busy_try(m
, TRUE
)) {
1489 * Various mechanisms such as a pmap_collect can
1490 * result in a busy page on the free queue. We
1491 * have to move the page out of the way so we can
1492 * retry the allocation. If the other thread is not
1493 * allocating the page then m->valid will remain 0 and
1494 * the pageout daemon will free the page later on.
1496 * Since we could not busy the page, however, we
1497 * cannot make assumptions as to whether the page
1498 * will be allocated by the other thread or not,
1499 * so all we can do is deactivate it to move it out
1500 * of the way. In particular, if the other thread
1501 * wires the page it may wind up on the inactive
1502 * queue and the pageout daemon will have to deal
1503 * with that case too.
1505 _vm_page_deactivate_locked(m
, 0);
1506 vm_page_spin_unlock(m
);
1509 * Theoretically if we are able to busy the page
1510 * atomic with the queue removal (using the vm_page
1511 * lock) nobody else should be able to mess with the
1514 KKASSERT((m
->flags
& (PG_UNMANAGED
|
1515 PG_NEED_COMMIT
)) == 0);
1516 KASSERT(m
->hold_count
== 0, ("m->hold_count is not zero "
1517 "pg %p q=%d flags=%08x hold=%d wire=%d",
1518 m
, m
->queue
, m
->flags
, m
->hold_count
, m
->wire_count
));
1519 KKASSERT(m
->wire_count
== 0);
1520 vm_page_spin_unlock(m
);
1521 pagedaemon_wakeup();
1523 /* return busied and removed page */
1531 * This implements a per-cpu cache of free, zero'd, ready-to-go pages.
1532 * The idea is to populate this cache prior to acquiring any locks so
1533 * we don't wind up potentially zeroing VM pages (under heavy loads) while
1534 * holding potentialy contending locks.
1536 * Note that we allocate the page uninserted into anything and use a pindex
1537 * of 0, the vm_page_alloc() will effectively add gd_cpuid so these
1538 * allocations should wind up being uncontended. However, we still want
1539 * to rove across PQ_L2_SIZE.
1542 vm_page_pcpu_cache(void)
1545 globaldata_t gd
= mycpu
;
1548 if (gd
->gd_vmpg_count
< GD_MINVMPG
) {
1550 while (gd
->gd_vmpg_count
< GD_MAXVMPG
) {
1551 m
= vm_page_alloc(NULL
, ticks
& ~ncpus2_mask
,
1552 VM_ALLOC_NULL_OK
| VM_ALLOC_NORMAL
|
1553 VM_ALLOC_NULL_OK
| VM_ALLOC_ZERO
);
1554 if (gd
->gd_vmpg_count
< GD_MAXVMPG
) {
1555 if ((m
->flags
& PG_ZERO
) == 0) {
1556 pmap_zero_page(VM_PAGE_TO_PHYS(m
));
1557 vm_page_flag_set(m
, PG_ZERO
);
1559 gd
->gd_vmpg_array
[gd
->gd_vmpg_count
++] = m
;
1572 * Allocate and return a memory cell associated with this VM object/offset
1573 * pair. If object is NULL an unassociated page will be allocated.
1575 * The returned page will be busied and removed from its queues. This
1576 * routine can block and may return NULL if a race occurs and the page
1577 * is found to already exist at the specified (object, pindex).
1579 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain
1580 * VM_ALLOC_QUICK like normal but cannot use cache
1581 * VM_ALLOC_SYSTEM greater free drain
1582 * VM_ALLOC_INTERRUPT allow free list to be completely drained
1583 * VM_ALLOC_ZERO advisory request for pre-zero'd page only
1584 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only
1585 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision
1586 * (see vm_page_grab())
1587 * VM_ALLOC_USE_GD ok to use per-gd cache
1589 * The object must be held if not NULL
1590 * This routine may not block
1592 * Additional special handling is required when called from an interrupt
1593 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache
1597 vm_page_alloc(vm_object_t object
, vm_pindex_t pindex
, int page_req
)
1599 globaldata_t gd
= mycpu
;
1605 int object_pg_color
;
1609 * Special per-cpu free VM page cache. The pages are pre-busied
1610 * and pre-zerod for us.
1612 if (gd
->gd_vmpg_count
&& (page_req
& VM_ALLOC_USE_GD
)) {
1614 if (gd
->gd_vmpg_count
) {
1615 m
= gd
->gd_vmpg_array
[--gd
->gd_vmpg_count
];
1627 * CPU localization algorithm. Break the page queues up by physical
1628 * id and core id (note that two cpu threads will have the same core
1629 * id, and core_id != gd_cpuid).
1631 * This is nowhere near perfect, for example the last pindex in a
1632 * subgroup will overflow into the next cpu or package. But this
1633 * should get us good page reuse locality in heavy mixed loads.
1635 phys_id
= get_cpu_phys_id(gd
->gd_cpuid
);
1636 core_id
= get_cpu_core_id(gd
->gd_cpuid
);
1637 object_pg_color
= object
? object
->pg_color
: 0;
1639 if (cpu_topology_phys_ids
&& cpu_topology_core_ids
) {
1640 if (PQ_L2_SIZE
/ ncpus
>= 16) {
1642 * Enough space for a full break-down.
1644 pg_color
= PQ_L2_SIZE
* core_id
/
1645 cpu_topology_core_ids
;
1646 pg_color
+= PQ_L2_SIZE
* phys_id
*
1647 cpu_topology_core_ids
/
1648 cpu_topology_phys_ids
;
1649 pg_color
+= (pindex
+ object_pg_color
) %
1650 (PQ_L2_SIZE
/ (cpu_topology_core_ids
*
1651 cpu_topology_phys_ids
));
1654 * Hopefully enough space to at least break the
1655 * queues down by package id.
1657 pg_color
= PQ_L2_SIZE
* phys_id
/ cpu_topology_phys_ids
;
1658 pg_color
+= (pindex
+ object_pg_color
) %
1659 (PQ_L2_SIZE
/ cpu_topology_phys_ids
);
1663 * Unknown topology, distribute things evenly.
1665 pg_color
= gd
->gd_cpuid
* PQ_L2_SIZE
/ ncpus
;
1666 pg_color
+= pindex
+ object_pg_color
;
1670 (VM_ALLOC_NORMAL
|VM_ALLOC_QUICK
|
1671 VM_ALLOC_INTERRUPT
|VM_ALLOC_SYSTEM
));
1674 * Certain system threads (pageout daemon, buf_daemon's) are
1675 * allowed to eat deeper into the free page list.
1677 if (curthread
->td_flags
& TDF_SYSTHREAD
)
1678 page_req
|= VM_ALLOC_SYSTEM
;
1681 * Impose various limitations. Note that the v_free_reserved test
1682 * must match the opposite of vm_page_count_target() to avoid
1683 * livelocks, be careful.
1686 if (vmstats
.v_free_count
>= vmstats
.v_free_reserved
||
1687 ((page_req
& VM_ALLOC_INTERRUPT
) && vmstats
.v_free_count
> 0) ||
1688 ((page_req
& VM_ALLOC_SYSTEM
) && vmstats
.v_cache_count
== 0 &&
1689 vmstats
.v_free_count
> vmstats
.v_interrupt_free_min
)
1692 * The free queue has sufficient free pages to take one out.
1694 if (page_req
& (VM_ALLOC_ZERO
| VM_ALLOC_FORCE_ZERO
))
1695 m
= vm_page_select_free(pg_color
, TRUE
);
1697 m
= vm_page_select_free(pg_color
, FALSE
);
1698 } else if (page_req
& VM_ALLOC_NORMAL
) {
1700 * Allocatable from the cache (non-interrupt only). On
1701 * success, we must free the page and try again, thus
1702 * ensuring that vmstats.v_*_free_min counters are replenished.
1705 if (curthread
->td_preempted
) {
1706 kprintf("vm_page_alloc(): warning, attempt to allocate"
1707 " cache page from preempting interrupt\n");
1710 m
= vm_page_select_cache(pg_color
);
1713 m
= vm_page_select_cache(pg_color
);
1716 * On success move the page into the free queue and loop.
1718 * Only do this if we can safely acquire the vm_object lock,
1719 * because this is effectively a random page and the caller
1720 * might be holding the lock shared, we don't want to
1724 KASSERT(m
->dirty
== 0,
1725 ("Found dirty cache page %p", m
));
1726 if ((obj
= m
->object
) != NULL
) {
1727 if (vm_object_hold_try(obj
)) {
1728 vm_page_protect(m
, VM_PROT_NONE
);
1730 /* m->object NULL here */
1731 vm_object_drop(obj
);
1733 vm_page_deactivate(m
);
1737 vm_page_protect(m
, VM_PROT_NONE
);
1744 * On failure return NULL
1746 #if defined(DIAGNOSTIC)
1747 if (vmstats
.v_cache_count
> 0)
1748 kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats
.v_cache_count
);
1750 vm_pageout_deficit
++;
1751 pagedaemon_wakeup();
1755 * No pages available, wakeup the pageout daemon and give up.
1757 vm_pageout_deficit
++;
1758 pagedaemon_wakeup();
1763 * v_free_count can race so loop if we don't find the expected
1770 * Good page found. The page has already been busied for us and
1771 * removed from its queues.
1773 KASSERT(m
->dirty
== 0,
1774 ("vm_page_alloc: free/cache page %p was dirty", m
));
1775 KKASSERT(m
->queue
== PQ_NONE
);
1781 * Initialize the structure, inheriting some flags but clearing
1782 * all the rest. The page has already been busied for us.
1784 vm_page_flag_clear(m
, ~(PG_ZERO
| PG_BUSY
| PG_SBUSY
));
1785 KKASSERT(m
->wire_count
== 0);
1786 KKASSERT(m
->busy
== 0);
1791 * Caller must be holding the object lock (asserted by
1792 * vm_page_insert()).
1794 * NOTE: Inserting a page here does not insert it into any pmaps
1795 * (which could cause us to block allocating memory).
1797 * NOTE: If no object an unassociated page is allocated, m->pindex
1798 * can be used by the caller for any purpose.
1801 if (vm_page_insert(m
, object
, pindex
) == FALSE
) {
1803 if ((page_req
& VM_ALLOC_NULL_OK
) == 0)
1804 panic("PAGE RACE %p[%ld]/%p",
1805 object
, (long)pindex
, m
);
1813 * Don't wakeup too often - wakeup the pageout daemon when
1814 * we would be nearly out of memory.
1816 pagedaemon_wakeup();
1819 * A PG_BUSY page is returned.
1825 * Returns number of pages available in our DMA memory reserve
1826 * (adjusted with vm.dma_reserved=<value>m in /boot/loader.conf)
1829 vm_contig_avail_pages(void)
1834 spin_lock(&vm_contig_spin
);
1835 bfree
= alist_free_info(&vm_contig_alist
, &blk
, &count
);
1836 spin_unlock(&vm_contig_spin
);
1842 * Attempt to allocate contiguous physical memory with the specified
1846 vm_page_alloc_contig(vm_paddr_t low
, vm_paddr_t high
,
1847 unsigned long alignment
, unsigned long boundary
,
1848 unsigned long size
, vm_memattr_t memattr
)
1854 alignment
>>= PAGE_SHIFT
;
1857 boundary
>>= PAGE_SHIFT
;
1860 size
= (size
+ PAGE_MASK
) >> PAGE_SHIFT
;
1862 spin_lock(&vm_contig_spin
);
1863 blk
= alist_alloc(&vm_contig_alist
, 0, size
);
1864 if (blk
== ALIST_BLOCK_NONE
) {
1865 spin_unlock(&vm_contig_spin
);
1867 kprintf("vm_page_alloc_contig: %ldk nospace\n",
1868 (size
+ PAGE_MASK
) * (PAGE_SIZE
/ 1024));
1872 if (high
&& ((vm_paddr_t
)(blk
+ size
) << PAGE_SHIFT
) > high
) {
1873 alist_free(&vm_contig_alist
, blk
, size
);
1874 spin_unlock(&vm_contig_spin
);
1876 kprintf("vm_page_alloc_contig: %ldk high "
1878 (size
+ PAGE_MASK
) * (PAGE_SIZE
/ 1024),
1883 spin_unlock(&vm_contig_spin
);
1884 if (vm_contig_verbose
) {
1885 kprintf("vm_page_alloc_contig: %016jx/%ldk\n",
1886 (intmax_t)(vm_paddr_t
)blk
<< PAGE_SHIFT
,
1887 (size
+ PAGE_MASK
) * (PAGE_SIZE
/ 1024));
1890 m
= PHYS_TO_VM_PAGE((vm_paddr_t
)blk
<< PAGE_SHIFT
);
1891 if (memattr
!= VM_MEMATTR_DEFAULT
)
1892 for (i
= 0;i
< size
;i
++)
1893 pmap_page_set_memattr(&m
[i
], memattr
);
1898 * Free contiguously allocated pages. The pages will be wired but not busy.
1899 * When freeing to the alist we leave them wired and not busy.
1902 vm_page_free_contig(vm_page_t m
, unsigned long size
)
1904 vm_paddr_t pa
= VM_PAGE_TO_PHYS(m
);
1905 vm_pindex_t start
= pa
>> PAGE_SHIFT
;
1906 vm_pindex_t pages
= (size
+ PAGE_MASK
) >> PAGE_SHIFT
;
1908 if (vm_contig_verbose
) {
1909 kprintf("vm_page_free_contig: %016jx/%ldk\n",
1910 (intmax_t)pa
, size
/ 1024);
1912 if (pa
< vm_low_phys_reserved
) {
1913 KKASSERT(pa
+ size
<= vm_low_phys_reserved
);
1914 spin_lock(&vm_contig_spin
);
1915 alist_free(&vm_contig_alist
, start
, pages
);
1916 spin_unlock(&vm_contig_spin
);
1919 vm_page_busy_wait(m
, FALSE
, "cpgfr");
1920 vm_page_unwire(m
, 0);
1931 * Wait for sufficient free memory for nominal heavy memory use kernel
1934 * WARNING! Be sure never to call this in any vm_pageout code path, which
1935 * will trivially deadlock the system.
1938 vm_wait_nominal(void)
1940 while (vm_page_count_min(0))
1945 * Test if vm_wait_nominal() would block.
1948 vm_test_nominal(void)
1950 if (vm_page_count_min(0))
1956 * Block until free pages are available for allocation, called in various
1957 * places before memory allocations.
1959 * The caller may loop if vm_page_count_min() == FALSE so we cannot be
1960 * more generous then that.
1966 * never wait forever
1970 lwkt_gettoken(&vm_token
);
1972 if (curthread
== pagethread
) {
1974 * The pageout daemon itself needs pages, this is bad.
1976 if (vm_page_count_min(0)) {
1977 vm_pageout_pages_needed
= 1;
1978 tsleep(&vm_pageout_pages_needed
, 0, "VMWait", timo
);
1982 * Wakeup the pageout daemon if necessary and wait.
1984 * Do not wait indefinitely for the target to be reached,
1985 * as load might prevent it from being reached any time soon.
1986 * But wait a little to try to slow down page allocations
1987 * and to give more important threads (the pagedaemon)
1988 * allocation priority.
1990 if (vm_page_count_target()) {
1991 if (vm_pages_needed
== 0) {
1992 vm_pages_needed
= 1;
1993 wakeup(&vm_pages_needed
);
1995 ++vm_pages_waiting
; /* SMP race ok */
1996 tsleep(&vmstats
.v_free_count
, 0, "vmwait", timo
);
1999 lwkt_reltoken(&vm_token
);
2003 * Block until free pages are available for allocation
2005 * Called only from vm_fault so that processes page faulting can be
2009 vm_wait_pfault(void)
2012 * Wakeup the pageout daemon if necessary and wait.
2014 * Do not wait indefinitely for the target to be reached,
2015 * as load might prevent it from being reached any time soon.
2016 * But wait a little to try to slow down page allocations
2017 * and to give more important threads (the pagedaemon)
2018 * allocation priority.
2020 if (vm_page_count_min(0)) {
2021 lwkt_gettoken(&vm_token
);
2022 while (vm_page_count_severe()) {
2023 if (vm_page_count_target()) {
2024 if (vm_pages_needed
== 0) {
2025 vm_pages_needed
= 1;
2026 wakeup(&vm_pages_needed
);
2028 ++vm_pages_waiting
; /* SMP race ok */
2029 tsleep(&vmstats
.v_free_count
, 0, "pfault", hz
);
2032 lwkt_reltoken(&vm_token
);
2037 * Put the specified page on the active list (if appropriate). Ensure
2038 * that act_count is at least ACT_INIT but do not otherwise mess with it.
2040 * The caller should be holding the page busied ? XXX
2041 * This routine may not block.
2044 vm_page_activate(vm_page_t m
)
2048 vm_page_spin_lock(m
);
2049 if (m
->queue
- m
->pc
!= PQ_ACTIVE
) {
2050 _vm_page_queue_spin_lock(m
);
2051 oqueue
= _vm_page_rem_queue_spinlocked(m
);
2052 /* page is left spinlocked, queue is unlocked */
2054 if (oqueue
== PQ_CACHE
)
2055 mycpu
->gd_cnt
.v_reactivated
++;
2056 if (m
->wire_count
== 0 && (m
->flags
& PG_UNMANAGED
) == 0) {
2057 if (m
->act_count
< ACT_INIT
)
2058 m
->act_count
= ACT_INIT
;
2059 _vm_page_add_queue_spinlocked(m
, PQ_ACTIVE
+ m
->pc
, 0);
2061 _vm_page_and_queue_spin_unlock(m
);
2062 if (oqueue
== PQ_CACHE
|| oqueue
== PQ_FREE
)
2063 pagedaemon_wakeup();
2065 if (m
->act_count
< ACT_INIT
)
2066 m
->act_count
= ACT_INIT
;
2067 vm_page_spin_unlock(m
);
2072 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
2073 * routine is called when a page has been added to the cache or free
2076 * This routine may not block.
2078 static __inline
void
2079 vm_page_free_wakeup(void)
2082 * If the pageout daemon itself needs pages, then tell it that
2083 * there are some free.
2085 if (vm_pageout_pages_needed
&&
2086 vmstats
.v_cache_count
+ vmstats
.v_free_count
>=
2087 vmstats
.v_pageout_free_min
2089 vm_pageout_pages_needed
= 0;
2090 wakeup(&vm_pageout_pages_needed
);
2094 * Wakeup processes that are waiting on memory.
2096 * Generally speaking we want to wakeup stuck processes as soon as
2097 * possible. !vm_page_count_min(0) is the absolute minimum point
2098 * where we can do this. Wait a bit longer to reduce degenerate
2099 * re-blocking (vm_page_free_hysteresis). The target check is just
2100 * to make sure the min-check w/hysteresis does not exceed the
2103 if (vm_pages_waiting
) {
2104 if (!vm_page_count_min(vm_page_free_hysteresis
) ||
2105 !vm_page_count_target()) {
2106 vm_pages_waiting
= 0;
2107 wakeup(&vmstats
.v_free_count
);
2108 ++mycpu
->gd_cnt
.v_ppwakeups
;
2111 if (!vm_page_count_target()) {
2113 * Plenty of pages are free, wakeup everyone.
2115 vm_pages_waiting
= 0;
2116 wakeup(&vmstats
.v_free_count
);
2117 ++mycpu
->gd_cnt
.v_ppwakeups
;
2118 } else if (!vm_page_count_min(0)) {
2120 * Some pages are free, wakeup someone.
2122 int wcount
= vm_pages_waiting
;
2125 vm_pages_waiting
= wcount
;
2126 wakeup_one(&vmstats
.v_free_count
);
2127 ++mycpu
->gd_cnt
.v_ppwakeups
;
2134 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates
2135 * it from its VM object.
2137 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on
2138 * return (the page will have been freed).
2141 vm_page_free_toq(vm_page_t m
)
2143 mycpu
->gd_cnt
.v_tfree
++;
2144 KKASSERT((m
->flags
& PG_MAPPED
) == 0);
2145 KKASSERT(m
->flags
& PG_BUSY
);
2147 if (m
->busy
|| ((m
->queue
- m
->pc
) == PQ_FREE
)) {
2148 kprintf("vm_page_free: pindex(%lu), busy(%d), "
2149 "PG_BUSY(%d), hold(%d)\n",
2150 (u_long
)m
->pindex
, m
->busy
,
2151 ((m
->flags
& PG_BUSY
) ? 1 : 0), m
->hold_count
);
2152 if ((m
->queue
- m
->pc
) == PQ_FREE
)
2153 panic("vm_page_free: freeing free page");
2155 panic("vm_page_free: freeing busy page");
2159 * Remove from object, spinlock the page and its queues and
2160 * remove from any queue. No queue spinlock will be held
2161 * after this section (because the page was removed from any
2165 vm_page_and_queue_spin_lock(m
);
2166 _vm_page_rem_queue_spinlocked(m
);
2169 * No further management of fictitious pages occurs beyond object
2170 * and queue removal.
2172 if ((m
->flags
& PG_FICTITIOUS
) != 0) {
2173 vm_page_spin_unlock(m
);
2181 if (m
->wire_count
!= 0) {
2182 if (m
->wire_count
> 1) {
2184 "vm_page_free: invalid wire count (%d), pindex: 0x%lx",
2185 m
->wire_count
, (long)m
->pindex
);
2187 panic("vm_page_free: freeing wired page");
2191 * Clear the UNMANAGED flag when freeing an unmanaged page.
2192 * Clear the NEED_COMMIT flag
2194 if (m
->flags
& PG_UNMANAGED
)
2195 vm_page_flag_clear(m
, PG_UNMANAGED
);
2196 if (m
->flags
& PG_NEED_COMMIT
)
2197 vm_page_flag_clear(m
, PG_NEED_COMMIT
);
2199 if (m
->hold_count
!= 0) {
2200 vm_page_flag_clear(m
, PG_ZERO
);
2201 _vm_page_add_queue_spinlocked(m
, PQ_HOLD
+ m
->pc
, 0);
2203 _vm_page_add_queue_spinlocked(m
, PQ_FREE
+ m
->pc
, 0);
2207 * This sequence allows us to clear PG_BUSY while still holding
2208 * its spin lock, which reduces contention vs allocators. We
2209 * must not leave the queue locked or _vm_page_wakeup() may
2212 _vm_page_queue_spin_unlock(m
);
2213 if (_vm_page_wakeup(m
)) {
2214 vm_page_spin_unlock(m
);
2217 vm_page_spin_unlock(m
);
2219 vm_page_free_wakeup();
2223 * vm_page_free_fromq_fast()
2225 * Remove a non-zero page from one of the free queues; the page is removed for
2226 * zeroing, so do not issue a wakeup.
2229 vm_page_free_fromq_fast(void)
2235 for (i
= 0; i
< PQ_L2_SIZE
; ++i
) {
2236 m
= vm_page_list_find(PQ_FREE
, qi
, FALSE
);
2237 /* page is returned spinlocked and removed from its queue */
2239 if (vm_page_busy_try(m
, TRUE
)) {
2241 * We were unable to busy the page, deactivate
2244 _vm_page_deactivate_locked(m
, 0);
2245 vm_page_spin_unlock(m
);
2246 } else if (m
->flags
& PG_ZERO
) {
2248 * The page is already PG_ZERO, requeue it
2251 _vm_page_add_queue_spinlocked(m
,
2254 vm_page_queue_spin_unlock(m
);
2255 if (_vm_page_wakeup(m
)) {
2256 vm_page_spin_unlock(m
);
2259 vm_page_spin_unlock(m
);
2263 * The page is not PG_ZERO'd so return it.
2265 KKASSERT((m
->flags
& (PG_UNMANAGED
|
2266 PG_NEED_COMMIT
)) == 0);
2267 KKASSERT(m
->hold_count
== 0);
2268 KKASSERT(m
->wire_count
== 0);
2269 vm_page_spin_unlock(m
);
2274 qi
= (qi
+ PQ_PRIME2
) & PQ_L2_MASK
;
2280 * vm_page_unmanage()
2282 * Prevent PV management from being done on the page. The page is
2283 * removed from the paging queues as if it were wired, and as a
2284 * consequence of no longer being managed the pageout daemon will not
2285 * touch it (since there is no way to locate the pte mappings for the
2286 * page). madvise() calls that mess with the pmap will also no longer
2287 * operate on the page.
2289 * Beyond that the page is still reasonably 'normal'. Freeing the page
2290 * will clear the flag.
2292 * This routine is used by OBJT_PHYS objects - objects using unswappable
2293 * physical memory as backing store rather then swap-backed memory and
2294 * will eventually be extended to support 4MB unmanaged physical
2297 * Caller must be holding the page busy.
2300 vm_page_unmanage(vm_page_t m
)
2302 KKASSERT(m
->flags
& PG_BUSY
);
2303 if ((m
->flags
& PG_UNMANAGED
) == 0) {
2304 if (m
->wire_count
== 0)
2307 vm_page_flag_set(m
, PG_UNMANAGED
);
2311 * Mark this page as wired down by yet another map, removing it from
2312 * paging queues as necessary.
2314 * Caller must be holding the page busy.
2317 vm_page_wire(vm_page_t m
)
2320 * Only bump the wire statistics if the page is not already wired,
2321 * and only unqueue the page if it is on some queue (if it is unmanaged
2322 * it is already off the queues). Don't do anything with fictitious
2323 * pages because they are always wired.
2325 KKASSERT(m
->flags
& PG_BUSY
);
2326 if ((m
->flags
& PG_FICTITIOUS
) == 0) {
2327 if (atomic_fetchadd_int(&m
->wire_count
, 1) == 0) {
2328 if ((m
->flags
& PG_UNMANAGED
) == 0)
2330 atomic_add_int(&vmstats
.v_wire_count
, 1);
2332 KASSERT(m
->wire_count
!= 0,
2333 ("vm_page_wire: wire_count overflow m=%p", m
));
2338 * Release one wiring of this page, potentially enabling it to be paged again.
2340 * Many pages placed on the inactive queue should actually go
2341 * into the cache, but it is difficult to figure out which. What
2342 * we do instead, if the inactive target is well met, is to put
2343 * clean pages at the head of the inactive queue instead of the tail.
2344 * This will cause them to be moved to the cache more quickly and
2345 * if not actively re-referenced, freed more quickly. If we just
2346 * stick these pages at the end of the inactive queue, heavy filesystem
2347 * meta-data accesses can cause an unnecessary paging load on memory bound
2348 * processes. This optimization causes one-time-use metadata to be
2349 * reused more quickly.
2351 * Pages marked PG_NEED_COMMIT are always activated and never placed on
2352 * the inactive queue. This helps the pageout daemon determine memory
2353 * pressure and act on out-of-memory situations more quickly.
2355 * BUT, if we are in a low-memory situation we have no choice but to
2356 * put clean pages on the cache queue.
2358 * A number of routines use vm_page_unwire() to guarantee that the page
2359 * will go into either the inactive or active queues, and will NEVER
2360 * be placed in the cache - for example, just after dirtying a page.
2361 * dirty pages in the cache are not allowed.
2363 * This routine may not block.
2366 vm_page_unwire(vm_page_t m
, int activate
)
2368 KKASSERT(m
->flags
& PG_BUSY
);
2369 if (m
->flags
& PG_FICTITIOUS
) {
2371 } else if (m
->wire_count
<= 0) {
2372 panic("vm_page_unwire: invalid wire count: %d", m
->wire_count
);
2374 if (atomic_fetchadd_int(&m
->wire_count
, -1) == 1) {
2375 atomic_add_int(&vmstats
.v_wire_count
, -1);
2376 if (m
->flags
& PG_UNMANAGED
) {
2378 } else if (activate
|| (m
->flags
& PG_NEED_COMMIT
)) {
2379 vm_page_spin_lock(m
);
2380 _vm_page_add_queue_spinlocked(m
,
2381 PQ_ACTIVE
+ m
->pc
, 0);
2382 _vm_page_and_queue_spin_unlock(m
);
2384 vm_page_spin_lock(m
);
2385 vm_page_flag_clear(m
, PG_WINATCFLS
);
2386 _vm_page_add_queue_spinlocked(m
,
2387 PQ_INACTIVE
+ m
->pc
, 0);
2388 ++vm_swapcache_inactive_heuristic
;
2389 _vm_page_and_queue_spin_unlock(m
);
2396 * Move the specified page to the inactive queue. If the page has
2397 * any associated swap, the swap is deallocated.
2399 * Normally athead is 0 resulting in LRU operation. athead is set
2400 * to 1 if we want this page to be 'as if it were placed in the cache',
2401 * except without unmapping it from the process address space.
2403 * vm_page's spinlock must be held on entry and will remain held on return.
2404 * This routine may not block.
2407 _vm_page_deactivate_locked(vm_page_t m
, int athead
)
2412 * Ignore if already inactive.
2414 if (m
->queue
- m
->pc
== PQ_INACTIVE
)
2416 _vm_page_queue_spin_lock(m
);
2417 oqueue
= _vm_page_rem_queue_spinlocked(m
);
2419 if (m
->wire_count
== 0 && (m
->flags
& PG_UNMANAGED
) == 0) {
2420 if (oqueue
== PQ_CACHE
)
2421 mycpu
->gd_cnt
.v_reactivated
++;
2422 vm_page_flag_clear(m
, PG_WINATCFLS
);
2423 _vm_page_add_queue_spinlocked(m
, PQ_INACTIVE
+ m
->pc
, athead
);
2425 ++vm_swapcache_inactive_heuristic
;
2427 /* NOTE: PQ_NONE if condition not taken */
2428 _vm_page_queue_spin_unlock(m
);
2429 /* leaves vm_page spinlocked */
2433 * Attempt to deactivate a page.
2438 vm_page_deactivate(vm_page_t m
)
2440 vm_page_spin_lock(m
);
2441 _vm_page_deactivate_locked(m
, 0);
2442 vm_page_spin_unlock(m
);
2446 vm_page_deactivate_locked(vm_page_t m
)
2448 _vm_page_deactivate_locked(m
, 0);
2452 * Attempt to move a page to PQ_CACHE.
2454 * Returns 0 on failure, 1 on success
2456 * The page should NOT be busied by the caller. This function will validate
2457 * whether the page can be safely moved to the cache.
2460 vm_page_try_to_cache(vm_page_t m
)
2462 vm_page_spin_lock(m
);
2463 if (vm_page_busy_try(m
, TRUE
)) {
2464 vm_page_spin_unlock(m
);
2467 if (m
->dirty
|| m
->hold_count
|| m
->wire_count
||
2468 (m
->flags
& (PG_UNMANAGED
| PG_NEED_COMMIT
))) {
2469 if (_vm_page_wakeup(m
)) {
2470 vm_page_spin_unlock(m
);
2473 vm_page_spin_unlock(m
);
2477 vm_page_spin_unlock(m
);
2480 * Page busied by us and no longer spinlocked. Dirty pages cannot
2481 * be moved to the cache.
2483 vm_page_test_dirty(m
);
2484 if (m
->dirty
|| (m
->flags
& PG_NEED_COMMIT
)) {
2493 * Attempt to free the page. If we cannot free it, we do nothing.
2494 * 1 is returned on success, 0 on failure.
2499 vm_page_try_to_free(vm_page_t m
)
2501 vm_page_spin_lock(m
);
2502 if (vm_page_busy_try(m
, TRUE
)) {
2503 vm_page_spin_unlock(m
);
2508 * The page can be in any state, including already being on the free
2509 * queue. Check to see if it really can be freed.
2511 if (m
->dirty
|| /* can't free if it is dirty */
2512 m
->hold_count
|| /* or held (XXX may be wrong) */
2513 m
->wire_count
|| /* or wired */
2514 (m
->flags
& (PG_UNMANAGED
| /* or unmanaged */
2515 PG_NEED_COMMIT
)) || /* or needs a commit */
2516 m
->queue
- m
->pc
== PQ_FREE
|| /* already on PQ_FREE */
2517 m
->queue
- m
->pc
== PQ_HOLD
) { /* already on PQ_HOLD */
2518 if (_vm_page_wakeup(m
)) {
2519 vm_page_spin_unlock(m
);
2522 vm_page_spin_unlock(m
);
2526 vm_page_spin_unlock(m
);
2529 * We can probably free the page.
2531 * Page busied by us and no longer spinlocked. Dirty pages will
2532 * not be freed by this function. We have to re-test the
2533 * dirty bit after cleaning out the pmaps.
2535 vm_page_test_dirty(m
);
2536 if (m
->dirty
|| (m
->flags
& PG_NEED_COMMIT
)) {
2540 vm_page_protect(m
, VM_PROT_NONE
);
2541 if (m
->dirty
|| (m
->flags
& PG_NEED_COMMIT
)) {
2552 * Put the specified page onto the page cache queue (if appropriate).
2554 * The page must be busy, and this routine will release the busy and
2555 * possibly even free the page.
2558 vm_page_cache(vm_page_t m
)
2560 if ((m
->flags
& (PG_UNMANAGED
| PG_NEED_COMMIT
)) ||
2561 m
->busy
|| m
->wire_count
|| m
->hold_count
) {
2562 kprintf("vm_page_cache: attempting to cache busy/held page\n");
2568 * Already in the cache (and thus not mapped)
2570 if ((m
->queue
- m
->pc
) == PQ_CACHE
) {
2571 KKASSERT((m
->flags
& PG_MAPPED
) == 0);
2577 * Caller is required to test m->dirty, but note that the act of
2578 * removing the page from its maps can cause it to become dirty
2579 * on an SMP system due to another cpu running in usermode.
2582 panic("vm_page_cache: caching a dirty page, pindex: %ld",
2587 * Remove all pmaps and indicate that the page is not
2588 * writeable or mapped. Our vm_page_protect() call may
2589 * have blocked (especially w/ VM_PROT_NONE), so recheck
2592 vm_page_protect(m
, VM_PROT_NONE
);
2593 if ((m
->flags
& (PG_UNMANAGED
| PG_MAPPED
)) ||
2594 m
->busy
|| m
->wire_count
|| m
->hold_count
) {
2596 } else if (m
->dirty
|| (m
->flags
& PG_NEED_COMMIT
)) {
2597 vm_page_deactivate(m
);
2600 _vm_page_and_queue_spin_lock(m
);
2601 _vm_page_rem_queue_spinlocked(m
);
2602 _vm_page_add_queue_spinlocked(m
, PQ_CACHE
+ m
->pc
, 0);
2603 _vm_page_queue_spin_unlock(m
);
2604 if (_vm_page_wakeup(m
)) {
2605 vm_page_spin_unlock(m
);
2608 vm_page_spin_unlock(m
);
2610 vm_page_free_wakeup();
2615 * vm_page_dontneed()
2617 * Cache, deactivate, or do nothing as appropriate. This routine
2618 * is typically used by madvise() MADV_DONTNEED.
2620 * Generally speaking we want to move the page into the cache so
2621 * it gets reused quickly. However, this can result in a silly syndrome
2622 * due to the page recycling too quickly. Small objects will not be
2623 * fully cached. On the otherhand, if we move the page to the inactive
2624 * queue we wind up with a problem whereby very large objects
2625 * unnecessarily blow away our inactive and cache queues.
2627 * The solution is to move the pages based on a fixed weighting. We
2628 * either leave them alone, deactivate them, or move them to the cache,
2629 * where moving them to the cache has the highest weighting.
2630 * By forcing some pages into other queues we eventually force the
2631 * system to balance the queues, potentially recovering other unrelated
2632 * space from active. The idea is to not force this to happen too
2635 * The page must be busied.
2638 vm_page_dontneed(vm_page_t m
)
2640 static int dnweight
;
2647 * occassionally leave the page alone
2649 if ((dnw
& 0x01F0) == 0 ||
2650 m
->queue
- m
->pc
== PQ_INACTIVE
||
2651 m
->queue
- m
->pc
== PQ_CACHE
2653 if (m
->act_count
>= ACT_INIT
)
2659 * If vm_page_dontneed() is inactivating a page, it must clear
2660 * the referenced flag; otherwise the pagedaemon will see references
2661 * on the page in the inactive queue and reactivate it. Until the
2662 * page can move to the cache queue, madvise's job is not done.
2664 vm_page_flag_clear(m
, PG_REFERENCED
);
2665 pmap_clear_reference(m
);
2668 vm_page_test_dirty(m
);
2670 if (m
->dirty
|| (dnw
& 0x0070) == 0) {
2672 * Deactivate the page 3 times out of 32.
2677 * Cache the page 28 times out of every 32. Note that
2678 * the page is deactivated instead of cached, but placed
2679 * at the head of the queue instead of the tail.
2683 vm_page_spin_lock(m
);
2684 _vm_page_deactivate_locked(m
, head
);
2685 vm_page_spin_unlock(m
);
2689 * These routines manipulate the 'soft busy' count for a page. A soft busy
2690 * is almost like PG_BUSY except that it allows certain compatible operations
2691 * to occur on the page while it is busy. For example, a page undergoing a
2692 * write can still be mapped read-only.
2694 * Because vm_pages can overlap buffers m->busy can be > 1. m->busy is only
2695 * adjusted while the vm_page is PG_BUSY so the flash will occur when the
2696 * busy bit is cleared.
2699 vm_page_io_start(vm_page_t m
)
2701 KASSERT(m
->flags
& PG_BUSY
, ("vm_page_io_start: page not busy!!!"));
2702 atomic_add_char(&m
->busy
, 1);
2703 vm_page_flag_set(m
, PG_SBUSY
);
2707 vm_page_io_finish(vm_page_t m
)
2709 KASSERT(m
->flags
& PG_BUSY
, ("vm_page_io_finish: page not busy!!!"));
2710 atomic_subtract_char(&m
->busy
, 1);
2712 vm_page_flag_clear(m
, PG_SBUSY
);
2716 * Indicate that a clean VM page requires a filesystem commit and cannot
2717 * be reused. Used by tmpfs.
2720 vm_page_need_commit(vm_page_t m
)
2722 vm_page_flag_set(m
, PG_NEED_COMMIT
);
2723 vm_object_set_writeable_dirty(m
->object
);
2727 vm_page_clear_commit(vm_page_t m
)
2729 vm_page_flag_clear(m
, PG_NEED_COMMIT
);
2733 * Grab a page, blocking if it is busy and allocating a page if necessary.
2734 * A busy page is returned or NULL. The page may or may not be valid and
2735 * might not be on a queue (the caller is responsible for the disposition of
2738 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the
2739 * page will be zero'd and marked valid.
2741 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked
2742 * valid even if it already exists.
2744 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also
2745 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified.
2746 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified.
2748 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
2749 * always returned if we had blocked.
2751 * This routine may not be called from an interrupt.
2753 * PG_ZERO is *ALWAYS* cleared by this routine.
2755 * No other requirements.
2758 vm_page_grab(vm_object_t object
, vm_pindex_t pindex
, int allocflags
)
2764 KKASSERT(allocflags
&
2765 (VM_ALLOC_NORMAL
|VM_ALLOC_INTERRUPT
|VM_ALLOC_SYSTEM
));
2766 vm_object_hold_shared(object
);
2768 m
= vm_page_lookup_busy_try(object
, pindex
, TRUE
, &error
);
2770 vm_page_sleep_busy(m
, TRUE
, "pgrbwt");
2771 if ((allocflags
& VM_ALLOC_RETRY
) == 0) {
2776 } else if (m
== NULL
) {
2778 vm_object_upgrade(object
);
2781 if (allocflags
& VM_ALLOC_RETRY
)
2782 allocflags
|= VM_ALLOC_NULL_OK
;
2783 m
= vm_page_alloc(object
, pindex
,
2784 allocflags
& ~VM_ALLOC_RETRY
);
2788 if ((allocflags
& VM_ALLOC_RETRY
) == 0)
2797 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid.
2799 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set
2800 * valid even if already valid.
2802 if (m
->valid
== 0) {
2803 if (allocflags
& (VM_ALLOC_ZERO
| VM_ALLOC_FORCE_ZERO
)) {
2804 if ((m
->flags
& PG_ZERO
) == 0)
2805 pmap_zero_page(VM_PAGE_TO_PHYS(m
));
2806 m
->valid
= VM_PAGE_BITS_ALL
;
2808 } else if (allocflags
& VM_ALLOC_FORCE_ZERO
) {
2809 pmap_zero_page(VM_PAGE_TO_PHYS(m
));
2810 m
->valid
= VM_PAGE_BITS_ALL
;
2812 vm_page_flag_clear(m
, PG_ZERO
);
2814 vm_object_drop(object
);
2819 * Mapping function for valid bits or for dirty bits in
2820 * a page. May not block.
2822 * Inputs are required to range within a page.
2828 vm_page_bits(int base
, int size
)
2834 base
+ size
<= PAGE_SIZE
,
2835 ("vm_page_bits: illegal base/size %d/%d", base
, size
)
2838 if (size
== 0) /* handle degenerate case */
2841 first_bit
= base
>> DEV_BSHIFT
;
2842 last_bit
= (base
+ size
- 1) >> DEV_BSHIFT
;
2844 return ((2 << last_bit
) - (1 << first_bit
));
2848 * Sets portions of a page valid and clean. The arguments are expected
2849 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2850 * of any partial chunks touched by the range. The invalid portion of
2851 * such chunks will be zero'd.
2853 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically
2854 * align base to DEV_BSIZE so as not to mark clean a partially
2855 * truncated device block. Otherwise the dirty page status might be
2858 * This routine may not block.
2860 * (base + size) must be less then or equal to PAGE_SIZE.
2863 _vm_page_zero_valid(vm_page_t m
, int base
, int size
)
2868 if (size
== 0) /* handle degenerate case */
2872 * If the base is not DEV_BSIZE aligned and the valid
2873 * bit is clear, we have to zero out a portion of the
2877 if ((frag
= base
& ~(DEV_BSIZE
- 1)) != base
&&
2878 (m
->valid
& (1 << (base
>> DEV_BSHIFT
))) == 0
2880 pmap_zero_page_area(
2888 * If the ending offset is not DEV_BSIZE aligned and the
2889 * valid bit is clear, we have to zero out a portion of
2893 endoff
= base
+ size
;
2895 if ((frag
= endoff
& ~(DEV_BSIZE
- 1)) != endoff
&&
2896 (m
->valid
& (1 << (endoff
>> DEV_BSHIFT
))) == 0
2898 pmap_zero_page_area(
2901 DEV_BSIZE
- (endoff
& (DEV_BSIZE
- 1))
2907 * Set valid, clear dirty bits. If validating the entire
2908 * page we can safely clear the pmap modify bit. We also
2909 * use this opportunity to clear the PG_NOSYNC flag. If a process
2910 * takes a write fault on a MAP_NOSYNC memory area the flag will
2913 * We set valid bits inclusive of any overlap, but we can only
2914 * clear dirty bits for DEV_BSIZE chunks that are fully within
2917 * Page must be busied?
2918 * No other requirements.
2921 vm_page_set_valid(vm_page_t m
, int base
, int size
)
2923 _vm_page_zero_valid(m
, base
, size
);
2924 m
->valid
|= vm_page_bits(base
, size
);
2929 * Set valid bits and clear dirty bits.
2931 * NOTE: This function does not clear the pmap modified bit.
2932 * Also note that e.g. NFS may use a byte-granular base
2935 * WARNING: Page must be busied? But vfs_clean_one_page() will call
2936 * this without necessarily busying the page (via bdwrite()).
2937 * So for now vm_token must also be held.
2939 * No other requirements.
2942 vm_page_set_validclean(vm_page_t m
, int base
, int size
)
2946 _vm_page_zero_valid(m
, base
, size
);
2947 pagebits
= vm_page_bits(base
, size
);
2948 m
->valid
|= pagebits
;
2949 m
->dirty
&= ~pagebits
;
2950 if (base
== 0 && size
== PAGE_SIZE
) {
2951 /*pmap_clear_modify(m);*/
2952 vm_page_flag_clear(m
, PG_NOSYNC
);
2957 * Set valid & dirty. Used by buwrite()
2959 * WARNING: Page must be busied? But vfs_dirty_one_page() will
2960 * call this function in buwrite() so for now vm_token must
2963 * No other requirements.
2966 vm_page_set_validdirty(vm_page_t m
, int base
, int size
)
2970 pagebits
= vm_page_bits(base
, size
);
2971 m
->valid
|= pagebits
;
2972 m
->dirty
|= pagebits
;
2974 vm_object_set_writeable_dirty(m
->object
);
2980 * NOTE: This function does not clear the pmap modified bit.
2981 * Also note that e.g. NFS may use a byte-granular base
2984 * Page must be busied?
2985 * No other requirements.
2988 vm_page_clear_dirty(vm_page_t m
, int base
, int size
)
2990 m
->dirty
&= ~vm_page_bits(base
, size
);
2991 if (base
== 0 && size
== PAGE_SIZE
) {
2992 /*pmap_clear_modify(m);*/
2993 vm_page_flag_clear(m
, PG_NOSYNC
);
2998 * Make the page all-dirty.
3000 * Also make sure the related object and vnode reflect the fact that the
3001 * object may now contain a dirty page.
3003 * Page must be busied?
3004 * No other requirements.
3007 vm_page_dirty(vm_page_t m
)
3010 int pqtype
= m
->queue
- m
->pc
;
3012 KASSERT(pqtype
!= PQ_CACHE
&& pqtype
!= PQ_FREE
,
3013 ("vm_page_dirty: page in free/cache queue!"));
3014 if (m
->dirty
!= VM_PAGE_BITS_ALL
) {
3015 m
->dirty
= VM_PAGE_BITS_ALL
;
3017 vm_object_set_writeable_dirty(m
->object
);
3022 * Invalidates DEV_BSIZE'd chunks within a page. Both the
3023 * valid and dirty bits for the effected areas are cleared.
3025 * Page must be busied?
3027 * No other requirements.
3030 vm_page_set_invalid(vm_page_t m
, int base
, int size
)
3034 bits
= vm_page_bits(base
, size
);
3037 m
->object
->generation
++;
3041 * The kernel assumes that the invalid portions of a page contain
3042 * garbage, but such pages can be mapped into memory by user code.
3043 * When this occurs, we must zero out the non-valid portions of the
3044 * page so user code sees what it expects.
3046 * Pages are most often semi-valid when the end of a file is mapped
3047 * into memory and the file's size is not page aligned.
3049 * Page must be busied?
3050 * No other requirements.
3053 vm_page_zero_invalid(vm_page_t m
, boolean_t setvalid
)
3059 * Scan the valid bits looking for invalid sections that
3060 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
3061 * valid bit may be set ) have already been zerod by
3062 * vm_page_set_validclean().
3064 for (b
= i
= 0; i
<= PAGE_SIZE
/ DEV_BSIZE
; ++i
) {
3065 if (i
== (PAGE_SIZE
/ DEV_BSIZE
) ||
3066 (m
->valid
& (1 << i
))
3069 pmap_zero_page_area(
3072 (i
- b
) << DEV_BSHIFT
3080 * setvalid is TRUE when we can safely set the zero'd areas
3081 * as being valid. We can do this if there are no cache consistency
3082 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
3085 m
->valid
= VM_PAGE_BITS_ALL
;
3089 * Is a (partial) page valid? Note that the case where size == 0
3090 * will return FALSE in the degenerate case where the page is entirely
3091 * invalid, and TRUE otherwise.
3094 * No other requirements.
3097 vm_page_is_valid(vm_page_t m
, int base
, int size
)
3099 int bits
= vm_page_bits(base
, size
);
3101 if (m
->valid
&& ((m
->valid
& bits
) == bits
))
3108 * update dirty bits from pmap/mmu. May not block.
3110 * Caller must hold the page busy
3113 vm_page_test_dirty(vm_page_t m
)
3115 if ((m
->dirty
!= VM_PAGE_BITS_ALL
) && pmap_is_modified(m
)) {
3121 * Register an action, associating it with its vm_page
3124 vm_page_register_action(vm_page_action_t action
, vm_page_event_t event
)
3126 struct vm_page_action_list
*list
;
3129 hv
= (int)((intptr_t)action
->m
>> 8) & VMACTION_HMASK
;
3130 list
= &action_list
[hv
];
3132 lwkt_gettoken(&vm_token
);
3133 vm_page_flag_set(action
->m
, PG_ACTIONLIST
);
3134 action
->event
= event
;
3135 LIST_INSERT_HEAD(list
, action
, entry
);
3136 lwkt_reltoken(&vm_token
);
3140 * Unregister an action, disassociating it from its related vm_page
3143 vm_page_unregister_action(vm_page_action_t action
)
3145 struct vm_page_action_list
*list
;
3148 lwkt_gettoken(&vm_token
);
3149 if (action
->event
!= VMEVENT_NONE
) {
3150 action
->event
= VMEVENT_NONE
;
3151 LIST_REMOVE(action
, entry
);
3153 hv
= (int)((intptr_t)action
->m
>> 8) & VMACTION_HMASK
;
3154 list
= &action_list
[hv
];
3155 if (LIST_EMPTY(list
))
3156 vm_page_flag_clear(action
->m
, PG_ACTIONLIST
);
3158 lwkt_reltoken(&vm_token
);
3162 * Issue an event on a VM page. Corresponding action structures are
3163 * removed from the page's list and called.
3165 * If the vm_page has no more pending action events we clear its
3166 * PG_ACTIONLIST flag.
3169 vm_page_event_internal(vm_page_t m
, vm_page_event_t event
)
3171 struct vm_page_action_list
*list
;
3172 struct vm_page_action
*scan
;
3173 struct vm_page_action
*next
;
3177 hv
= (int)((intptr_t)m
>> 8) & VMACTION_HMASK
;
3178 list
= &action_list
[hv
];
3181 lwkt_gettoken(&vm_token
);
3182 LIST_FOREACH_MUTABLE(scan
, list
, entry
, next
) {
3184 if (scan
->event
== event
) {
3185 scan
->event
= VMEVENT_NONE
;
3186 LIST_REMOVE(scan
, entry
);
3187 scan
->func(m
, scan
);
3195 vm_page_flag_clear(m
, PG_ACTIONLIST
);
3196 lwkt_reltoken(&vm_token
);
3199 #include "opt_ddb.h"
3201 #include <sys/kernel.h>
3203 #include <ddb/ddb.h>
3205 DB_SHOW_COMMAND(page
, vm_page_print_page_info
)
3207 db_printf("vmstats.v_free_count: %d\n", vmstats
.v_free_count
);
3208 db_printf("vmstats.v_cache_count: %d\n", vmstats
.v_cache_count
);
3209 db_printf("vmstats.v_inactive_count: %d\n", vmstats
.v_inactive_count
);
3210 db_printf("vmstats.v_active_count: %d\n", vmstats
.v_active_count
);
3211 db_printf("vmstats.v_wire_count: %d\n", vmstats
.v_wire_count
);
3212 db_printf("vmstats.v_free_reserved: %d\n", vmstats
.v_free_reserved
);
3213 db_printf("vmstats.v_free_min: %d\n", vmstats
.v_free_min
);
3214 db_printf("vmstats.v_free_target: %d\n", vmstats
.v_free_target
);
3215 db_printf("vmstats.v_cache_min: %d\n", vmstats
.v_cache_min
);
3216 db_printf("vmstats.v_inactive_target: %d\n", vmstats
.v_inactive_target
);
3219 DB_SHOW_COMMAND(pageq
, vm_page_print_pageq_info
)
3222 db_printf("PQ_FREE:");
3223 for(i
=0;i
<PQ_L2_SIZE
;i
++) {
3224 db_printf(" %d", vm_page_queues
[PQ_FREE
+ i
].lcnt
);
3228 db_printf("PQ_CACHE:");
3229 for(i
=0;i
<PQ_L2_SIZE
;i
++) {
3230 db_printf(" %d", vm_page_queues
[PQ_CACHE
+ i
].lcnt
);
3234 db_printf("PQ_ACTIVE:");
3235 for(i
=0;i
<PQ_L2_SIZE
;i
++) {
3236 db_printf(" %d", vm_page_queues
[PQ_ACTIVE
+ i
].lcnt
);
3240 db_printf("PQ_INACTIVE:");
3241 for(i
=0;i
<PQ_L2_SIZE
;i
++) {
3242 db_printf(" %d", vm_page_queues
[PQ_INACTIVE
+ i
].lcnt
);