2 * Copyright (c) 1994,1997 John S. Dyson
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice immediately at the beginning of the file, without modification,
10 * this list of conditions, and the following disclaimer.
11 * 2. Absolutely no warranty of function or purpose is made by the author
14 * $FreeBSD: src/sys/kern/vfs_bio.c,v 1.242.2.20 2003/05/28 18:38:10 alc Exp $
15 * $DragonFly: src/sys/kern/vfs_bio.c,v 1.115 2008/08/13 11:02:31 swildner Exp $
19 * this file contains a new buffer I/O scheme implementing a coherent
20 * VM object and buffer cache scheme. Pains have been taken to make
21 * sure that the performance degradation associated with schemes such
22 * as this is not realized.
24 * Author: John S. Dyson
25 * Significant help during the development and debugging phases
26 * had been provided by David Greenman, also of the FreeBSD core team.
28 * see man buf(9) for more info.
31 #include <sys/param.h>
32 #include <sys/systm.h>
35 #include <sys/eventhandler.h>
37 #include <sys/malloc.h>
38 #include <sys/mount.h>
39 #include <sys/kernel.h>
40 #include <sys/kthread.h>
42 #include <sys/reboot.h>
43 #include <sys/resourcevar.h>
44 #include <sys/sysctl.h>
45 #include <sys/vmmeter.h>
46 #include <sys/vnode.h>
49 #include <vm/vm_param.h>
50 #include <vm/vm_kern.h>
51 #include <vm/vm_pageout.h>
52 #include <vm/vm_page.h>
53 #include <vm/vm_object.h>
54 #include <vm/vm_extern.h>
55 #include <vm/vm_map.h>
56 #include <vm/vm_pager.h>
57 #include <vm/swap_pager.h>
60 #include <sys/thread2.h>
61 #include <sys/spinlock2.h>
62 #include <sys/mplock2.h>
63 #include <vm/vm_page2.h>
74 BQUEUE_NONE
, /* not on any queue */
75 BQUEUE_LOCKED
, /* locked buffers */
76 BQUEUE_CLEAN
, /* non-B_DELWRI buffers */
77 BQUEUE_DIRTY
, /* B_DELWRI buffers */
78 BQUEUE_DIRTY_HW
, /* B_DELWRI buffers - heavy weight */
79 BQUEUE_EMPTYKVA
, /* empty buffer headers with KVA assignment */
80 BQUEUE_EMPTY
, /* empty buffer headers */
82 BUFFER_QUEUES
/* number of buffer queues */
85 typedef enum bufq_type bufq_type_t
;
87 #define BD_WAKE_SIZE 16384
88 #define BD_WAKE_MASK (BD_WAKE_SIZE - 1)
90 TAILQ_HEAD(bqueues
, buf
) bufqueues
[BUFFER_QUEUES
];
91 struct spinlock bufspin
= SPINLOCK_INITIALIZER(&bufspin
);
93 static MALLOC_DEFINE(M_BIOBUF
, "BIO buffer", "BIO buffer");
95 struct buf
*buf
; /* buffer header pool */
97 static void vfs_clean_pages(struct buf
*bp
);
98 static void vfs_clean_one_page(struct buf
*bp
, int pageno
, vm_page_t m
);
99 static void vfs_dirty_one_page(struct buf
*bp
, int pageno
, vm_page_t m
);
100 static void vfs_vmio_release(struct buf
*bp
);
101 static int flushbufqueues(bufq_type_t q
);
102 static vm_page_t
bio_page_alloc(vm_object_t obj
, vm_pindex_t pg
, int deficit
);
104 static void bd_signal(int totalspace
);
105 static void buf_daemon(void);
106 static void buf_daemon_hw(void);
109 * bogus page -- for I/O to/from partially complete buffers
110 * this is a temporary solution to the problem, but it is not
111 * really that bad. it would be better to split the buffer
112 * for input in the case of buffers partially already in memory,
113 * but the code is intricate enough already.
115 vm_page_t bogus_page
;
118 * These are all static, but make the ones we export globals so we do
119 * not need to use compiler magic.
121 int bufspace
, maxbufspace
,
122 bufmallocspace
, maxbufmallocspace
, lobufspace
, hibufspace
;
123 static int bufreusecnt
, bufdefragcnt
, buffreekvacnt
;
124 static int lorunningspace
, hirunningspace
, runningbufreq
;
125 int dirtybufspace
, dirtybufspacehw
, lodirtybufspace
, hidirtybufspace
;
126 int dirtybufcount
, dirtybufcounthw
;
127 int runningbufspace
, runningbufcount
;
128 static int getnewbufcalls
;
129 static int getnewbufrestarts
;
130 static int recoverbufcalls
;
131 static int needsbuffer
; /* locked by needsbuffer_spin */
132 static int bd_request
; /* locked by needsbuffer_spin */
133 static int bd_request_hw
; /* locked by needsbuffer_spin */
134 static u_int bd_wake_ary
[BD_WAKE_SIZE
];
135 static u_int bd_wake_index
;
136 static u_int vm_cycle_point
= ACT_INIT
+ ACT_ADVANCE
* 6;
137 static struct spinlock needsbuffer_spin
;
138 static int debug_commit
;
140 static struct thread
*bufdaemon_td
;
141 static struct thread
*bufdaemonhw_td
;
145 * Sysctls for operational control of the buffer cache.
147 SYSCTL_INT(_vfs
, OID_AUTO
, lodirtybufspace
, CTLFLAG_RW
, &lodirtybufspace
, 0,
148 "Number of dirty buffers to flush before bufdaemon becomes inactive");
149 SYSCTL_INT(_vfs
, OID_AUTO
, hidirtybufspace
, CTLFLAG_RW
, &hidirtybufspace
, 0,
150 "High watermark used to trigger explicit flushing of dirty buffers");
151 SYSCTL_INT(_vfs
, OID_AUTO
, lorunningspace
, CTLFLAG_RW
, &lorunningspace
, 0,
152 "Minimum amount of buffer space required for active I/O");
153 SYSCTL_INT(_vfs
, OID_AUTO
, hirunningspace
, CTLFLAG_RW
, &hirunningspace
, 0,
154 "Maximum amount of buffer space to usable for active I/O");
155 SYSCTL_UINT(_vfs
, OID_AUTO
, vm_cycle_point
, CTLFLAG_RW
, &vm_cycle_point
, 0,
156 "Recycle pages to active or inactive queue transition pt 0-64");
158 * Sysctls determining current state of the buffer cache.
160 SYSCTL_INT(_vfs
, OID_AUTO
, nbuf
, CTLFLAG_RD
, &nbuf
, 0,
161 "Total number of buffers in buffer cache");
162 SYSCTL_INT(_vfs
, OID_AUTO
, dirtybufspace
, CTLFLAG_RD
, &dirtybufspace
, 0,
163 "Pending bytes of dirty buffers (all)");
164 SYSCTL_INT(_vfs
, OID_AUTO
, dirtybufspacehw
, CTLFLAG_RD
, &dirtybufspacehw
, 0,
165 "Pending bytes of dirty buffers (heavy weight)");
166 SYSCTL_INT(_vfs
, OID_AUTO
, dirtybufcount
, CTLFLAG_RD
, &dirtybufcount
, 0,
167 "Pending number of dirty buffers");
168 SYSCTL_INT(_vfs
, OID_AUTO
, dirtybufcounthw
, CTLFLAG_RD
, &dirtybufcounthw
, 0,
169 "Pending number of dirty buffers (heavy weight)");
170 SYSCTL_INT(_vfs
, OID_AUTO
, runningbufspace
, CTLFLAG_RD
, &runningbufspace
, 0,
171 "I/O bytes currently in progress due to asynchronous writes");
172 SYSCTL_INT(_vfs
, OID_AUTO
, runningbufcount
, CTLFLAG_RD
, &runningbufcount
, 0,
173 "I/O buffers currently in progress due to asynchronous writes");
174 SYSCTL_INT(_vfs
, OID_AUTO
, maxbufspace
, CTLFLAG_RD
, &maxbufspace
, 0,
175 "Hard limit on maximum amount of memory usable for buffer space");
176 SYSCTL_INT(_vfs
, OID_AUTO
, hibufspace
, CTLFLAG_RD
, &hibufspace
, 0,
177 "Soft limit on maximum amount of memory usable for buffer space");
178 SYSCTL_INT(_vfs
, OID_AUTO
, lobufspace
, CTLFLAG_RD
, &lobufspace
, 0,
179 "Minimum amount of memory to reserve for system buffer space");
180 SYSCTL_INT(_vfs
, OID_AUTO
, bufspace
, CTLFLAG_RD
, &bufspace
, 0,
181 "Amount of memory available for buffers");
182 SYSCTL_INT(_vfs
, OID_AUTO
, maxmallocbufspace
, CTLFLAG_RD
, &maxbufmallocspace
,
183 0, "Maximum amount of memory reserved for buffers using malloc");
184 SYSCTL_INT(_vfs
, OID_AUTO
, bufmallocspace
, CTLFLAG_RD
, &bufmallocspace
, 0,
185 "Amount of memory left for buffers using malloc-scheme");
186 SYSCTL_INT(_vfs
, OID_AUTO
, getnewbufcalls
, CTLFLAG_RD
, &getnewbufcalls
, 0,
187 "New buffer header acquisition requests");
188 SYSCTL_INT(_vfs
, OID_AUTO
, getnewbufrestarts
, CTLFLAG_RD
, &getnewbufrestarts
,
189 0, "New buffer header acquisition restarts");
190 SYSCTL_INT(_vfs
, OID_AUTO
, recoverbufcalls
, CTLFLAG_RD
, &recoverbufcalls
, 0,
191 "Recover VM space in an emergency");
192 SYSCTL_INT(_vfs
, OID_AUTO
, bufdefragcnt
, CTLFLAG_RD
, &bufdefragcnt
, 0,
193 "Buffer acquisition restarts due to fragmented buffer map");
194 SYSCTL_INT(_vfs
, OID_AUTO
, buffreekvacnt
, CTLFLAG_RD
, &buffreekvacnt
, 0,
195 "Amount of time KVA space was deallocated in an arbitrary buffer");
196 SYSCTL_INT(_vfs
, OID_AUTO
, bufreusecnt
, CTLFLAG_RD
, &bufreusecnt
, 0,
197 "Amount of time buffer re-use operations were successful");
198 SYSCTL_INT(_vfs
, OID_AUTO
, debug_commit
, CTLFLAG_RW
, &debug_commit
, 0, "");
199 SYSCTL_INT(_debug_sizeof
, OID_AUTO
, buf
, CTLFLAG_RD
, 0, sizeof(struct buf
),
200 "sizeof(struct buf)");
202 char *buf_wmesg
= BUF_WMESG
;
204 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
205 #define VFS_BIO_NEED_UNUSED02 0x02
206 #define VFS_BIO_NEED_UNUSED04 0x04
207 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
212 * Called when buffer space is potentially available for recovery.
213 * getnewbuf() will block on this flag when it is unable to free
214 * sufficient buffer space. Buffer space becomes recoverable when
215 * bp's get placed back in the queues.
222 * If someone is waiting for BUF space, wake them up. Even
223 * though we haven't freed the kva space yet, the waiting
224 * process will be able to now.
226 if (needsbuffer
& VFS_BIO_NEED_BUFSPACE
) {
227 spin_lock_wr(&needsbuffer_spin
);
228 needsbuffer
&= ~VFS_BIO_NEED_BUFSPACE
;
229 spin_unlock_wr(&needsbuffer_spin
);
230 wakeup(&needsbuffer
);
237 * Accounting for I/O in progress.
241 runningbufwakeup(struct buf
*bp
)
246 if ((totalspace
= bp
->b_runningbufspace
) != 0) {
247 atomic_subtract_int(&runningbufspace
, totalspace
);
248 atomic_subtract_int(&runningbufcount
, 1);
249 bp
->b_runningbufspace
= 0;
252 * see waitrunningbufspace() for limit test.
254 limit
= hirunningspace
* 2 / 3;
255 if (runningbufreq
&& runningbufspace
<= limit
) {
257 wakeup(&runningbufreq
);
259 bd_signal(totalspace
);
266 * Called when a buffer has been added to one of the free queues to
267 * account for the buffer and to wakeup anyone waiting for free buffers.
268 * This typically occurs when large amounts of metadata are being handled
269 * by the buffer cache ( else buffer space runs out first, usually ).
277 spin_lock_wr(&needsbuffer_spin
);
278 needsbuffer
&= ~VFS_BIO_NEED_ANY
;
279 spin_unlock_wr(&needsbuffer_spin
);
280 wakeup(&needsbuffer
);
285 * waitrunningbufspace()
287 * Wait for the amount of running I/O to drop to hirunningspace * 2 / 3.
288 * This is the point where write bursting stops so we don't want to wait
289 * for the running amount to drop below it (at least if we still want bioq
292 * The caller may be using this function to block in a tight loop, we
293 * must block while runningbufspace is greater then or equal to
294 * hirunningspace * 2 / 3.
296 * And even with that it may not be enough, due to the presence of
297 * B_LOCKED dirty buffers, so also wait for at least one running buffer
301 waitrunningbufspace(void)
303 int limit
= hirunningspace
* 2 / 3;
306 if (runningbufspace
> limit
) {
307 while (runningbufspace
> limit
) {
309 tsleep(&runningbufreq
, 0, "wdrn1", 0);
311 } else if (runningbufspace
) {
313 tsleep(&runningbufreq
, 0, "wdrn2", 1);
319 * buf_dirty_count_severe:
321 * Return true if we have too many dirty buffers.
324 buf_dirty_count_severe(void)
326 return (runningbufspace
+ dirtybufspace
>= hidirtybufspace
||
327 dirtybufcount
>= nbuf
/ 2);
331 * Return true if the amount of running I/O is severe and BIOQ should
335 buf_runningbufspace_severe(void)
337 return (runningbufspace
>= hirunningspace
* 2 / 3);
341 * vfs_buf_test_cache:
343 * Called when a buffer is extended. This function clears the B_CACHE
344 * bit if the newly extended portion of the buffer does not contain
347 * NOTE! Dirty VM pages are not processed into dirty (B_DELWRI) buffer
348 * cache buffers. The VM pages remain dirty, as someone had mmap()'d
349 * them while a clean buffer was present.
353 vfs_buf_test_cache(struct buf
*bp
,
354 vm_ooffset_t foff
, vm_offset_t off
, vm_offset_t size
,
357 if (bp
->b_flags
& B_CACHE
) {
358 int base
= (foff
+ off
) & PAGE_MASK
;
359 if (vm_page_is_valid(m
, base
, size
) == 0)
360 bp
->b_flags
&= ~B_CACHE
;
367 * Spank the buf_daemon[_hw] if the total dirty buffer space exceeds the
376 if (dirtybufspace
< lodirtybufspace
&& dirtybufcount
< nbuf
/ 2)
379 if (bd_request
== 0 &&
380 (dirtybufspace
- dirtybufspacehw
> lodirtybufspace
/ 2 ||
381 dirtybufcount
- dirtybufcounthw
>= nbuf
/ 2)) {
382 spin_lock_wr(&needsbuffer_spin
);
384 spin_unlock_wr(&needsbuffer_spin
);
387 if (bd_request_hw
== 0 &&
388 (dirtybufspacehw
> lodirtybufspace
/ 2 ||
389 dirtybufcounthw
>= nbuf
/ 2)) {
390 spin_lock_wr(&needsbuffer_spin
);
392 spin_unlock_wr(&needsbuffer_spin
);
393 wakeup(&bd_request_hw
);
400 * Get the buf_daemon heated up when the number of running and dirty
401 * buffers exceeds the mid-point.
403 * Return the total number of dirty bytes past the second mid point
404 * as a measure of how much excess dirty data there is in the system.
415 mid1
= lodirtybufspace
+ (hidirtybufspace
- lodirtybufspace
) / 2;
417 totalspace
= runningbufspace
+ dirtybufspace
;
418 if (totalspace
>= mid1
|| dirtybufcount
>= nbuf
/ 2) {
420 mid2
= mid1
+ (hidirtybufspace
- mid1
) / 2;
421 if (totalspace
>= mid2
)
422 return(totalspace
- mid2
);
430 * Wait for the buffer cache to flush (totalspace) bytes worth of
431 * buffers, then return.
433 * Regardless this function blocks while the number of dirty buffers
434 * exceeds hidirtybufspace.
439 bd_wait(int totalspace
)
444 if (curthread
== bufdaemonhw_td
|| curthread
== bufdaemon_td
)
447 while (totalspace
> 0) {
449 if (totalspace
> runningbufspace
+ dirtybufspace
)
450 totalspace
= runningbufspace
+ dirtybufspace
;
451 count
= totalspace
/ BKVASIZE
;
452 if (count
>= BD_WAKE_SIZE
)
453 count
= BD_WAKE_SIZE
- 1;
455 spin_lock_wr(&needsbuffer_spin
);
456 i
= (bd_wake_index
+ count
) & BD_WAKE_MASK
;
458 tsleep_interlock(&bd_wake_ary
[i
], 0);
459 spin_unlock_wr(&needsbuffer_spin
);
460 tsleep(&bd_wake_ary
[i
], PINTERLOCKED
, "flstik", hz
);
462 totalspace
= runningbufspace
+ dirtybufspace
- hidirtybufspace
;
469 * This function is called whenever runningbufspace or dirtybufspace
470 * is reduced. Track threads waiting for run+dirty buffer I/O
476 bd_signal(int totalspace
)
480 if (totalspace
> 0) {
481 if (totalspace
> BKVASIZE
* BD_WAKE_SIZE
)
482 totalspace
= BKVASIZE
* BD_WAKE_SIZE
;
483 spin_lock_wr(&needsbuffer_spin
);
484 while (totalspace
> 0) {
487 if (bd_wake_ary
[i
]) {
489 spin_unlock_wr(&needsbuffer_spin
);
490 wakeup(&bd_wake_ary
[i
]);
491 spin_lock_wr(&needsbuffer_spin
);
493 totalspace
-= BKVASIZE
;
495 spin_unlock_wr(&needsbuffer_spin
);
500 * BIO tracking support routines.
502 * Release a ref on a bio_track. Wakeup requests are atomically released
503 * along with the last reference so bk_active will never wind up set to
510 bio_track_rel(struct bio_track
*track
)
518 active
= track
->bk_active
;
519 if (active
== 1 && atomic_cmpset_int(&track
->bk_active
, 1, 0))
523 * Full-on. Note that the wait flag is only atomically released on
524 * the 1->0 count transition.
526 * We check for a negative count transition using bit 30 since bit 31
527 * has a different meaning.
530 desired
= (active
& 0x7FFFFFFF) - 1;
532 desired
|= active
& 0x80000000;
533 if (atomic_cmpset_int(&track
->bk_active
, active
, desired
)) {
534 if (desired
& 0x40000000)
535 panic("bio_track_rel: bad count: %p\n", track
);
536 if (active
& 0x80000000)
540 active
= track
->bk_active
;
545 * Wait for the tracking count to reach 0.
547 * Use atomic ops such that the wait flag is only set atomically when
548 * bk_active is non-zero.
553 bio_track_wait(struct bio_track
*track
, int slp_flags
, int slp_timo
)
562 if (track
->bk_active
== 0)
566 * Full-on. Note that the wait flag may only be atomically set if
567 * the active count is non-zero.
570 while ((active
= track
->bk_active
) != 0) {
571 desired
= active
| 0x80000000;
572 tsleep_interlock(track
, slp_flags
);
573 if (active
== desired
||
574 atomic_cmpset_int(&track
->bk_active
, active
, desired
)) {
575 error
= tsleep(track
, slp_flags
| PINTERLOCKED
,
587 * Load time initialisation of the buffer cache, called from machine
588 * dependant initialization code.
594 vm_offset_t bogus_offset
;
597 spin_init(&needsbuffer_spin
);
599 /* next, make a null set of free lists */
600 for (i
= 0; i
< BUFFER_QUEUES
; i
++)
601 TAILQ_INIT(&bufqueues
[i
]);
603 /* finally, initialize each buffer header and stick on empty q */
604 for (i
= 0; i
< nbuf
; i
++) {
606 bzero(bp
, sizeof *bp
);
607 bp
->b_flags
= B_INVAL
; /* we're just an empty header */
608 bp
->b_cmd
= BUF_CMD_DONE
;
609 bp
->b_qindex
= BQUEUE_EMPTY
;
611 xio_init(&bp
->b_xio
);
614 TAILQ_INSERT_TAIL(&bufqueues
[BQUEUE_EMPTY
], bp
, b_freelist
);
618 * maxbufspace is the absolute maximum amount of buffer space we are
619 * allowed to reserve in KVM and in real terms. The absolute maximum
620 * is nominally used by buf_daemon. hibufspace is the nominal maximum
621 * used by most other processes. The differential is required to
622 * ensure that buf_daemon is able to run when other processes might
623 * be blocked waiting for buffer space.
625 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
626 * this may result in KVM fragmentation which is not handled optimally
629 maxbufspace
= nbuf
* BKVASIZE
;
630 hibufspace
= imax(3 * maxbufspace
/ 4, maxbufspace
- MAXBSIZE
* 10);
631 lobufspace
= hibufspace
- MAXBSIZE
;
633 lorunningspace
= 512 * 1024;
634 /* hirunningspace -- see below */
637 * Limit the amount of malloc memory since it is wired permanently
638 * into the kernel space. Even though this is accounted for in
639 * the buffer allocation, we don't want the malloced region to grow
640 * uncontrolled. The malloc scheme improves memory utilization
641 * significantly on average (small) directories.
643 maxbufmallocspace
= hibufspace
/ 20;
646 * Reduce the chance of a deadlock occuring by limiting the number
647 * of delayed-write dirty buffers we allow to stack up.
649 * We don't want too much actually queued to the device at once
650 * (XXX this needs to be per-mount!), because the buffers will
651 * wind up locked for a very long period of time while the I/O
654 hidirtybufspace
= hibufspace
/ 2; /* dirty + running */
655 hirunningspace
= hibufspace
/ 16; /* locked & queued to device */
656 if (hirunningspace
< 1024 * 1024)
657 hirunningspace
= 1024 * 1024;
662 lodirtybufspace
= hidirtybufspace
/ 2;
665 * Maximum number of async ops initiated per buf_daemon loop. This is
666 * somewhat of a hack at the moment, we really need to limit ourselves
667 * based on the number of bytes of I/O in-transit that were initiated
671 bogus_offset
= kmem_alloc_pageable(&kernel_map
, PAGE_SIZE
);
672 bogus_page
= vm_page_alloc(&kernel_object
,
673 (bogus_offset
>> PAGE_SHIFT
),
675 vmstats
.v_wire_count
++;
680 * Initialize the embedded bio structures
683 initbufbio(struct buf
*bp
)
685 bp
->b_bio1
.bio_buf
= bp
;
686 bp
->b_bio1
.bio_prev
= NULL
;
687 bp
->b_bio1
.bio_offset
= NOOFFSET
;
688 bp
->b_bio1
.bio_next
= &bp
->b_bio2
;
689 bp
->b_bio1
.bio_done
= NULL
;
690 bp
->b_bio1
.bio_flags
= 0;
692 bp
->b_bio2
.bio_buf
= bp
;
693 bp
->b_bio2
.bio_prev
= &bp
->b_bio1
;
694 bp
->b_bio2
.bio_offset
= NOOFFSET
;
695 bp
->b_bio2
.bio_next
= NULL
;
696 bp
->b_bio2
.bio_done
= NULL
;
697 bp
->b_bio2
.bio_flags
= 0;
701 * Reinitialize the embedded bio structures as well as any additional
702 * translation cache layers.
705 reinitbufbio(struct buf
*bp
)
709 for (bio
= &bp
->b_bio1
; bio
; bio
= bio
->bio_next
) {
710 bio
->bio_done
= NULL
;
711 bio
->bio_offset
= NOOFFSET
;
716 * Push another BIO layer onto an existing BIO and return it. The new
717 * BIO layer may already exist, holding cached translation data.
720 push_bio(struct bio
*bio
)
724 if ((nbio
= bio
->bio_next
) == NULL
) {
725 int index
= bio
- &bio
->bio_buf
->b_bio_array
[0];
726 if (index
>= NBUF_BIO
- 1) {
727 panic("push_bio: too many layers bp %p\n",
730 nbio
= &bio
->bio_buf
->b_bio_array
[index
+ 1];
731 bio
->bio_next
= nbio
;
732 nbio
->bio_prev
= bio
;
733 nbio
->bio_buf
= bio
->bio_buf
;
734 nbio
->bio_offset
= NOOFFSET
;
735 nbio
->bio_done
= NULL
;
736 nbio
->bio_next
= NULL
;
738 KKASSERT(nbio
->bio_done
== NULL
);
743 * Pop a BIO translation layer, returning the previous layer. The
744 * must have been previously pushed.
747 pop_bio(struct bio
*bio
)
749 return(bio
->bio_prev
);
753 clearbiocache(struct bio
*bio
)
756 bio
->bio_offset
= NOOFFSET
;
764 * Free the KVA allocation for buffer 'bp'.
766 * Must be called from a critical section as this is the only locking for
769 * Since this call frees up buffer space, we call bufspacewakeup().
774 bfreekva(struct buf
*bp
)
781 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
782 vm_map_lock(&buffer_map
);
783 bufspace
-= bp
->b_kvasize
;
784 vm_map_delete(&buffer_map
,
785 (vm_offset_t
) bp
->b_kvabase
,
786 (vm_offset_t
) bp
->b_kvabase
+ bp
->b_kvasize
,
789 vm_map_unlock(&buffer_map
);
790 vm_map_entry_release(count
);
800 * Remove the buffer from the appropriate free list.
803 _bremfree(struct buf
*bp
)
805 if (bp
->b_qindex
!= BQUEUE_NONE
) {
806 KASSERT(BUF_REFCNTNB(bp
) == 1,
807 ("bremfree: bp %p not locked",bp
));
808 TAILQ_REMOVE(&bufqueues
[bp
->b_qindex
], bp
, b_freelist
);
809 bp
->b_qindex
= BQUEUE_NONE
;
811 if (BUF_REFCNTNB(bp
) <= 1)
812 panic("bremfree: removing a buffer not on a queue");
817 bremfree(struct buf
*bp
)
819 spin_lock_wr(&bufspin
);
821 spin_unlock_wr(&bufspin
);
825 bremfree_locked(struct buf
*bp
)
833 * Get a buffer with the specified data. Look in the cache first. We
834 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
835 * is set, the buffer is valid and we do not have to do anything ( see
841 bread(struct vnode
*vp
, off_t loffset
, int size
, struct buf
**bpp
)
845 bp
= getblk(vp
, loffset
, size
, 0, 0);
848 /* if not found in cache, do some I/O */
849 if ((bp
->b_flags
& B_CACHE
) == 0) {
851 bp
->b_flags
&= ~(B_ERROR
| B_EINTR
| B_INVAL
);
852 bp
->b_cmd
= BUF_CMD_READ
;
853 bp
->b_bio1
.bio_done
= biodone_sync
;
854 bp
->b_bio1
.bio_flags
|= BIO_SYNC
;
855 vfs_busy_pages(vp
, bp
);
856 vn_strategy(vp
, &bp
->b_bio1
);
858 return (biowait(&bp
->b_bio1
, "biord"));
866 * Operates like bread, but also starts asynchronous I/O on
867 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior
868 * to initiating I/O . If B_CACHE is set, the buffer is valid
869 * and we do not have to do anything.
874 breadn(struct vnode
*vp
, off_t loffset
, int size
, off_t
*raoffset
,
875 int *rabsize
, int cnt
, struct buf
**bpp
)
877 struct buf
*bp
, *rabp
;
879 int rv
= 0, readwait
= 0;
881 *bpp
= bp
= getblk(vp
, loffset
, size
, 0, 0);
883 /* if not found in cache, do some I/O */
884 if ((bp
->b_flags
& B_CACHE
) == 0) {
886 bp
->b_flags
&= ~(B_ERROR
| B_EINTR
| B_INVAL
);
887 bp
->b_cmd
= BUF_CMD_READ
;
888 bp
->b_bio1
.bio_done
= biodone_sync
;
889 bp
->b_bio1
.bio_flags
|= BIO_SYNC
;
890 vfs_busy_pages(vp
, bp
);
891 vn_strategy(vp
, &bp
->b_bio1
);
896 for (i
= 0; i
< cnt
; i
++, raoffset
++, rabsize
++) {
897 if (inmem(vp
, *raoffset
))
899 rabp
= getblk(vp
, *raoffset
, *rabsize
, 0, 0);
901 if ((rabp
->b_flags
& B_CACHE
) == 0) {
903 rabp
->b_flags
&= ~(B_ERROR
| B_EINTR
| B_INVAL
);
904 rabp
->b_cmd
= BUF_CMD_READ
;
905 vfs_busy_pages(vp
, rabp
);
907 vn_strategy(vp
, &rabp
->b_bio1
);
914 rv
= biowait(&bp
->b_bio1
, "biord");
921 * Synchronous write, waits for completion.
923 * Write, release buffer on completion. (Done by iodone
924 * if async). Do not bother writing anything if the buffer
927 * Note that we set B_CACHE here, indicating that buffer is
928 * fully valid and thus cacheable. This is true even of NFS
929 * now so we set it generally. This could be set either here
930 * or in biodone() since the I/O is synchronous. We put it
934 bwrite(struct buf
*bp
)
938 if (bp
->b_flags
& B_INVAL
) {
942 if (BUF_REFCNTNB(bp
) == 0)
943 panic("bwrite: buffer is not busy???");
945 /* Mark the buffer clean */
948 bp
->b_flags
&= ~(B_ERROR
| B_EINTR
);
949 bp
->b_flags
|= B_CACHE
;
950 bp
->b_cmd
= BUF_CMD_WRITE
;
951 bp
->b_bio1
.bio_done
= biodone_sync
;
952 bp
->b_bio1
.bio_flags
|= BIO_SYNC
;
953 vfs_busy_pages(bp
->b_vp
, bp
);
956 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
957 * valid for vnode-backed buffers.
959 bp
->b_runningbufspace
= bp
->b_bufsize
;
960 if (bp
->b_runningbufspace
) {
961 runningbufspace
+= bp
->b_runningbufspace
;
965 vn_strategy(bp
->b_vp
, &bp
->b_bio1
);
966 error
= biowait(&bp
->b_bio1
, "biows");
974 * Asynchronous write. Start output on a buffer, but do not wait for
975 * it to complete. The buffer is released when the output completes.
977 * bwrite() ( or the VOP routine anyway ) is responsible for handling
978 * B_INVAL buffers. Not us.
981 bawrite(struct buf
*bp
)
983 if (bp
->b_flags
& B_INVAL
) {
987 if (BUF_REFCNTNB(bp
) == 0)
988 panic("bwrite: buffer is not busy???");
990 /* Mark the buffer clean */
993 bp
->b_flags
&= ~(B_ERROR
| B_EINTR
);
994 bp
->b_flags
|= B_CACHE
;
995 bp
->b_cmd
= BUF_CMD_WRITE
;
996 KKASSERT(bp
->b_bio1
.bio_done
== NULL
);
997 vfs_busy_pages(bp
->b_vp
, bp
);
1000 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
1001 * valid for vnode-backed buffers.
1003 bp
->b_runningbufspace
= bp
->b_bufsize
;
1004 if (bp
->b_runningbufspace
) {
1005 runningbufspace
+= bp
->b_runningbufspace
;
1010 vn_strategy(bp
->b_vp
, &bp
->b_bio1
);
1016 * Ordered write. Start output on a buffer, and flag it so that the
1017 * device will write it in the order it was queued. The buffer is
1018 * released when the output completes. bwrite() ( or the VOP routine
1019 * anyway ) is responsible for handling B_INVAL buffers.
1022 bowrite(struct buf
*bp
)
1024 bp
->b_flags
|= B_ORDERED
;
1032 * Delayed write. (Buffer is marked dirty). Do not bother writing
1033 * anything if the buffer is marked invalid.
1035 * Note that since the buffer must be completely valid, we can safely
1036 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
1037 * biodone() in order to prevent getblk from writing the buffer
1038 * out synchronously.
1041 bdwrite(struct buf
*bp
)
1043 if (BUF_REFCNTNB(bp
) == 0)
1044 panic("bdwrite: buffer is not busy");
1046 if (bp
->b_flags
& B_INVAL
) {
1053 * Set B_CACHE, indicating that the buffer is fully valid. This is
1054 * true even of NFS now.
1056 bp
->b_flags
|= B_CACHE
;
1059 * This bmap keeps the system from needing to do the bmap later,
1060 * perhaps when the system is attempting to do a sync. Since it
1061 * is likely that the indirect block -- or whatever other datastructure
1062 * that the filesystem needs is still in memory now, it is a good
1063 * thing to do this. Note also, that if the pageout daemon is
1064 * requesting a sync -- there might not be enough memory to do
1065 * the bmap then... So, this is important to do.
1067 if (bp
->b_bio2
.bio_offset
== NOOFFSET
) {
1068 VOP_BMAP(bp
->b_vp
, bp
->b_loffset
, &bp
->b_bio2
.bio_offset
,
1069 NULL
, NULL
, BUF_CMD_WRITE
);
1073 * Because the underlying pages may still be mapped and
1074 * writable trying to set the dirty buffer (b_dirtyoff/end)
1075 * range here will be inaccurate.
1077 * However, we must still clean the pages to satisfy the
1078 * vnode_pager and pageout daemon, so theythink the pages
1079 * have been "cleaned". What has really occured is that
1080 * they've been earmarked for later writing by the buffer
1083 * So we get the b_dirtyoff/end update but will not actually
1084 * depend on it (NFS that is) until the pages are busied for
1087 vfs_clean_pages(bp
);
1091 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1092 * due to the softdep code.
1097 * Fake write - return pages to VM system as dirty, leave the buffer clean.
1098 * This is used by tmpfs.
1100 * It is important for any VFS using this routine to NOT use it for
1101 * IO_SYNC or IO_ASYNC operations which occur when the system really
1102 * wants to flush VM pages to backing store.
1105 buwrite(struct buf
*bp
)
1111 * Only works for VMIO buffers. If the buffer is already
1112 * marked for delayed-write we can't avoid the bdwrite().
1114 if ((bp
->b_flags
& B_VMIO
) == 0 || (bp
->b_flags
& B_DELWRI
)) {
1120 * Set valid & dirty.
1122 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
1123 m
= bp
->b_xio
.xio_pages
[i
];
1124 vfs_dirty_one_page(bp
, i
, m
);
1132 * Turn buffer into delayed write request by marking it B_DELWRI.
1133 * B_RELBUF and B_NOCACHE must be cleared.
1135 * We reassign the buffer to itself to properly update it in the
1136 * dirty/clean lists.
1138 * Must be called from a critical section.
1139 * The buffer must be on BQUEUE_NONE.
1142 bdirty(struct buf
*bp
)
1144 KASSERT(bp
->b_qindex
== BQUEUE_NONE
, ("bdirty: buffer %p still on queue %d", bp
, bp
->b_qindex
));
1145 if (bp
->b_flags
& B_NOCACHE
) {
1146 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp
);
1147 bp
->b_flags
&= ~B_NOCACHE
;
1149 if (bp
->b_flags
& B_INVAL
) {
1150 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp
);
1152 bp
->b_flags
&= ~B_RELBUF
;
1154 if ((bp
->b_flags
& B_DELWRI
) == 0) {
1155 bp
->b_flags
|= B_DELWRI
;
1157 atomic_add_int(&dirtybufcount
, 1);
1158 dirtybufspace
+= bp
->b_bufsize
;
1159 if (bp
->b_flags
& B_HEAVY
) {
1160 atomic_add_int(&dirtybufcounthw
, 1);
1161 atomic_add_int(&dirtybufspacehw
, bp
->b_bufsize
);
1168 * Set B_HEAVY, indicating that this is a heavy-weight buffer that
1169 * needs to be flushed with a different buf_daemon thread to avoid
1170 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf().
1173 bheavy(struct buf
*bp
)
1175 if ((bp
->b_flags
& B_HEAVY
) == 0) {
1176 bp
->b_flags
|= B_HEAVY
;
1177 if (bp
->b_flags
& B_DELWRI
) {
1178 atomic_add_int(&dirtybufcounthw
, 1);
1179 atomic_add_int(&dirtybufspacehw
, bp
->b_bufsize
);
1187 * Clear B_DELWRI for buffer.
1189 * Must be called from a critical section.
1191 * The buffer is typically on BQUEUE_NONE but there is one case in
1192 * brelse() that calls this function after placing the buffer on
1193 * a different queue.
1198 bundirty(struct buf
*bp
)
1200 if (bp
->b_flags
& B_DELWRI
) {
1201 bp
->b_flags
&= ~B_DELWRI
;
1203 atomic_subtract_int(&dirtybufcount
, 1);
1204 atomic_subtract_int(&dirtybufspace
, bp
->b_bufsize
);
1205 if (bp
->b_flags
& B_HEAVY
) {
1206 atomic_subtract_int(&dirtybufcounthw
, 1);
1207 atomic_subtract_int(&dirtybufspacehw
, bp
->b_bufsize
);
1209 bd_signal(bp
->b_bufsize
);
1212 * Since it is now being written, we can clear its deferred write flag.
1214 bp
->b_flags
&= ~B_DEFERRED
;
1220 * Release a busy buffer and, if requested, free its resources. The
1221 * buffer will be stashed in the appropriate bufqueue[] allowing it
1222 * to be accessed later as a cache entity or reused for other purposes.
1227 brelse(struct buf
*bp
)
1230 int saved_flags
= bp
->b_flags
;
1233 KASSERT(!(bp
->b_flags
& (B_CLUSTER
|B_PAGING
)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp
));
1236 * If B_NOCACHE is set we are being asked to destroy the buffer and
1237 * its backing store. Clear B_DELWRI.
1239 * B_NOCACHE is set in two cases: (1) when the caller really wants
1240 * to destroy the buffer and backing store and (2) when the caller
1241 * wants to destroy the buffer and backing store after a write
1244 if ((bp
->b_flags
& (B_NOCACHE
|B_DELWRI
)) == (B_NOCACHE
|B_DELWRI
)) {
1248 if ((bp
->b_flags
& (B_INVAL
| B_DELWRI
)) == B_DELWRI
) {
1250 * A re-dirtied buffer is only subject to destruction
1251 * by B_INVAL. B_ERROR and B_NOCACHE are ignored.
1253 /* leave buffer intact */
1254 } else if ((bp
->b_flags
& (B_NOCACHE
| B_INVAL
| B_ERROR
)) ||
1255 (bp
->b_bufsize
<= 0)) {
1257 * Either a failed read or we were asked to free or not
1258 * cache the buffer. This path is reached with B_DELWRI
1259 * set only if B_INVAL is already set. B_NOCACHE governs
1260 * backing store destruction.
1262 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the
1263 * buffer cannot be immediately freed.
1265 bp
->b_flags
|= B_INVAL
;
1266 if (LIST_FIRST(&bp
->b_dep
) != NULL
) {
1271 if (bp
->b_flags
& B_DELWRI
) {
1272 atomic_subtract_int(&dirtybufcount
, 1);
1273 atomic_subtract_int(&dirtybufspace
, bp
->b_bufsize
);
1274 if (bp
->b_flags
& B_HEAVY
) {
1275 atomic_subtract_int(&dirtybufcounthw
, 1);
1276 atomic_subtract_int(&dirtybufspacehw
, bp
->b_bufsize
);
1278 bd_signal(bp
->b_bufsize
);
1280 bp
->b_flags
&= ~(B_DELWRI
| B_CACHE
);
1284 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set.
1285 * If vfs_vmio_release() is called with either bit set, the
1286 * underlying pages may wind up getting freed causing a previous
1287 * write (bdwrite()) to get 'lost' because pages associated with
1288 * a B_DELWRI bp are marked clean. Pages associated with a
1289 * B_LOCKED buffer may be mapped by the filesystem.
1291 * If we want to release the buffer ourselves (rather then the
1292 * originator asking us to release it), give the originator a
1293 * chance to countermand the release by setting B_LOCKED.
1295 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1296 * if B_DELWRI is set.
1298 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1299 * on pages to return pages to the VM page queues.
1301 if (bp
->b_flags
& (B_DELWRI
| B_LOCKED
)) {
1302 bp
->b_flags
&= ~B_RELBUF
;
1303 } else if (vm_page_count_severe()) {
1304 if (LIST_FIRST(&bp
->b_dep
) != NULL
) {
1306 buf_deallocate(bp
); /* can set B_LOCKED */
1309 if (bp
->b_flags
& (B_DELWRI
| B_LOCKED
))
1310 bp
->b_flags
&= ~B_RELBUF
;
1312 bp
->b_flags
|= B_RELBUF
;
1316 * Make sure b_cmd is clear. It may have already been cleared by
1319 * At this point destroying the buffer is governed by the B_INVAL
1320 * or B_RELBUF flags.
1322 bp
->b_cmd
= BUF_CMD_DONE
;
1325 * VMIO buffer rundown. Make sure the VM page array is restored
1326 * after an I/O may have replaces some of the pages with bogus pages
1327 * in order to not destroy dirty pages in a fill-in read.
1329 * Note that due to the code above, if a buffer is marked B_DELWRI
1330 * then the B_RELBUF and B_NOCACHE bits will always be clear.
1331 * B_INVAL may still be set, however.
1333 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer
1334 * but not the backing store. B_NOCACHE will destroy the backing
1337 * Note that dirty NFS buffers contain byte-granular write ranges
1338 * and should not be destroyed w/ B_INVAL even if the backing store
1341 if (bp
->b_flags
& B_VMIO
) {
1343 * Rundown for VMIO buffers which are not dirty NFS buffers.
1355 * Get the base offset and length of the buffer. Note that
1356 * in the VMIO case if the buffer block size is not
1357 * page-aligned then b_data pointer may not be page-aligned.
1358 * But our b_xio.xio_pages array *IS* page aligned.
1360 * block sizes less then DEV_BSIZE (usually 512) are not
1361 * supported due to the page granularity bits (m->valid,
1362 * m->dirty, etc...).
1364 * See man buf(9) for more information
1367 resid
= bp
->b_bufsize
;
1368 foff
= bp
->b_loffset
;
1371 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
1372 m
= bp
->b_xio
.xio_pages
[i
];
1373 vm_page_flag_clear(m
, PG_ZERO
);
1375 * If we hit a bogus page, fixup *all* of them
1376 * now. Note that we left these pages wired
1377 * when we removed them so they had better exist,
1378 * and they cannot be ripped out from under us so
1379 * no critical section protection is necessary.
1381 if (m
== bogus_page
) {
1383 poff
= OFF_TO_IDX(bp
->b_loffset
);
1385 for (j
= i
; j
< bp
->b_xio
.xio_npages
; j
++) {
1388 mtmp
= bp
->b_xio
.xio_pages
[j
];
1389 if (mtmp
== bogus_page
) {
1390 mtmp
= vm_page_lookup(obj
, poff
+ j
);
1392 panic("brelse: page missing");
1394 bp
->b_xio
.xio_pages
[j
] = mtmp
;
1398 if ((bp
->b_flags
& B_INVAL
) == 0) {
1399 pmap_qenter(trunc_page((vm_offset_t
)bp
->b_data
),
1400 bp
->b_xio
.xio_pages
, bp
->b_xio
.xio_npages
);
1402 m
= bp
->b_xio
.xio_pages
[i
];
1406 * Invalidate the backing store if B_NOCACHE is set
1407 * (e.g. used with vinvalbuf()). If this is NFS
1408 * we impose a requirement that the block size be
1409 * a multiple of PAGE_SIZE and create a temporary
1410 * hack to basically invalidate the whole page. The
1411 * problem is that NFS uses really odd buffer sizes
1412 * especially when tracking piecemeal writes and
1413 * it also vinvalbuf()'s a lot, which would result
1414 * in only partial page validation and invalidation
1415 * here. If the file page is mmap()'d, however,
1416 * all the valid bits get set so after we invalidate
1417 * here we would end up with weird m->valid values
1418 * like 0xfc. nfs_getpages() can't handle this so
1419 * we clear all the valid bits for the NFS case
1420 * instead of just some of them.
1422 * The real bug is the VM system having to set m->valid
1423 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1424 * itself is an artifact of the whole 512-byte
1425 * granular mess that exists to support odd block
1426 * sizes and UFS meta-data block sizes (e.g. 6144).
1427 * A complete rewrite is required.
1431 if (bp
->b_flags
& (B_NOCACHE
|B_ERROR
)) {
1432 int poffset
= foff
& PAGE_MASK
;
1435 presid
= PAGE_SIZE
- poffset
;
1436 if (bp
->b_vp
->v_tag
== VT_NFS
&&
1437 bp
->b_vp
->v_type
== VREG
) {
1439 } else if (presid
> resid
) {
1442 KASSERT(presid
>= 0, ("brelse: extra page"));
1443 vm_page_set_invalid(m
, poffset
, presid
);
1446 * Also make sure any swap cache is removed
1447 * as it is now stale (HAMMER in particular
1448 * uses B_NOCACHE to deal with buffer
1451 swap_pager_unswapped(m
);
1453 resid
-= PAGE_SIZE
- (foff
& PAGE_MASK
);
1454 foff
= (foff
+ PAGE_SIZE
) & ~(off_t
)PAGE_MASK
;
1456 if (bp
->b_flags
& (B_INVAL
| B_RELBUF
))
1457 vfs_vmio_release(bp
);
1461 * Rundown for non-VMIO buffers.
1463 if (bp
->b_flags
& (B_INVAL
| B_RELBUF
)) {
1467 KKASSERT (LIST_FIRST(&bp
->b_dep
) == NULL
);
1474 if (bp
->b_qindex
!= BQUEUE_NONE
)
1475 panic("brelse: free buffer onto another queue???");
1476 if (BUF_REFCNTNB(bp
) > 1) {
1477 /* Temporary panic to verify exclusive locking */
1478 /* This panic goes away when we allow shared refs */
1479 panic("brelse: multiple refs");
1485 * Figure out the correct queue to place the cleaned up buffer on.
1486 * Buffers placed in the EMPTY or EMPTYKVA had better already be
1487 * disassociated from their vnode.
1489 spin_lock_wr(&bufspin
);
1490 if (bp
->b_flags
& B_LOCKED
) {
1492 * Buffers that are locked are placed in the locked queue
1493 * immediately, regardless of their state.
1495 bp
->b_qindex
= BQUEUE_LOCKED
;
1496 TAILQ_INSERT_TAIL(&bufqueues
[BQUEUE_LOCKED
], bp
, b_freelist
);
1497 } else if (bp
->b_bufsize
== 0) {
1499 * Buffers with no memory. Due to conditionals near the top
1500 * of brelse() such buffers should probably already be
1501 * marked B_INVAL and disassociated from their vnode.
1503 bp
->b_flags
|= B_INVAL
;
1504 KASSERT(bp
->b_vp
== NULL
, ("bp1 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp
, saved_flags
, bp
->b_flags
, bp
->b_vp
));
1505 KKASSERT((bp
->b_flags
& B_HASHED
) == 0);
1506 if (bp
->b_kvasize
) {
1507 bp
->b_qindex
= BQUEUE_EMPTYKVA
;
1509 bp
->b_qindex
= BQUEUE_EMPTY
;
1511 TAILQ_INSERT_HEAD(&bufqueues
[bp
->b_qindex
], bp
, b_freelist
);
1512 } else if (bp
->b_flags
& (B_INVAL
| B_NOCACHE
| B_RELBUF
)) {
1514 * Buffers with junk contents. Again these buffers had better
1515 * already be disassociated from their vnode.
1517 KASSERT(bp
->b_vp
== NULL
, ("bp2 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp
, saved_flags
, bp
->b_flags
, bp
->b_vp
));
1518 KKASSERT((bp
->b_flags
& B_HASHED
) == 0);
1519 bp
->b_flags
|= B_INVAL
;
1520 bp
->b_qindex
= BQUEUE_CLEAN
;
1521 TAILQ_INSERT_HEAD(&bufqueues
[BQUEUE_CLEAN
], bp
, b_freelist
);
1524 * Remaining buffers. These buffers are still associated with
1527 switch(bp
->b_flags
& (B_DELWRI
|B_HEAVY
)) {
1529 bp
->b_qindex
= BQUEUE_DIRTY
;
1530 TAILQ_INSERT_TAIL(&bufqueues
[BQUEUE_DIRTY
], bp
, b_freelist
);
1532 case B_DELWRI
| B_HEAVY
:
1533 bp
->b_qindex
= BQUEUE_DIRTY_HW
;
1534 TAILQ_INSERT_TAIL(&bufqueues
[BQUEUE_DIRTY_HW
], bp
,
1539 * NOTE: Buffers are always placed at the end of the
1540 * queue. If B_AGE is not set the buffer will cycle
1541 * through the queue twice.
1543 bp
->b_qindex
= BQUEUE_CLEAN
;
1544 TAILQ_INSERT_TAIL(&bufqueues
[BQUEUE_CLEAN
], bp
, b_freelist
);
1548 spin_unlock_wr(&bufspin
);
1551 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1552 * on the correct queue.
1554 if ((bp
->b_flags
& (B_INVAL
|B_DELWRI
)) == (B_INVAL
|B_DELWRI
))
1558 * The bp is on an appropriate queue unless locked. If it is not
1559 * locked or dirty we can wakeup threads waiting for buffer space.
1561 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1562 * if B_INVAL is set ).
1564 if ((bp
->b_flags
& (B_LOCKED
|B_DELWRI
)) == 0)
1568 * Something we can maybe free or reuse
1570 if (bp
->b_bufsize
|| bp
->b_kvasize
)
1574 * Clean up temporary flags and unlock the buffer.
1576 bp
->b_flags
&= ~(B_ORDERED
| B_NOCACHE
| B_RELBUF
| B_DIRECT
);
1583 * Release a buffer back to the appropriate queue but do not try to free
1584 * it. The buffer is expected to be used again soon.
1586 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1587 * biodone() to requeue an async I/O on completion. It is also used when
1588 * known good buffers need to be requeued but we think we may need the data
1591 * XXX we should be able to leave the B_RELBUF hint set on completion.
1596 bqrelse(struct buf
*bp
)
1598 KASSERT(!(bp
->b_flags
& (B_CLUSTER
|B_PAGING
)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp
));
1600 if (bp
->b_qindex
!= BQUEUE_NONE
)
1601 panic("bqrelse: free buffer onto another queue???");
1602 if (BUF_REFCNTNB(bp
) > 1) {
1603 /* do not release to free list */
1604 panic("bqrelse: multiple refs");
1608 buf_act_advance(bp
);
1610 spin_lock_wr(&bufspin
);
1611 if (bp
->b_flags
& B_LOCKED
) {
1613 * Locked buffers are released to the locked queue. However,
1614 * if the buffer is dirty it will first go into the dirty
1615 * queue and later on after the I/O completes successfully it
1616 * will be released to the locked queue.
1618 bp
->b_qindex
= BQUEUE_LOCKED
;
1619 TAILQ_INSERT_TAIL(&bufqueues
[BQUEUE_LOCKED
], bp
, b_freelist
);
1620 } else if (bp
->b_flags
& B_DELWRI
) {
1621 bp
->b_qindex
= (bp
->b_flags
& B_HEAVY
) ?
1622 BQUEUE_DIRTY_HW
: BQUEUE_DIRTY
;
1623 TAILQ_INSERT_TAIL(&bufqueues
[bp
->b_qindex
], bp
, b_freelist
);
1624 } else if (vm_page_count_severe()) {
1626 * We are too low on memory, we have to try to free the
1627 * buffer (most importantly: the wired pages making up its
1628 * backing store) *now*.
1630 spin_unlock_wr(&bufspin
);
1634 bp
->b_qindex
= BQUEUE_CLEAN
;
1635 TAILQ_INSERT_TAIL(&bufqueues
[BQUEUE_CLEAN
], bp
, b_freelist
);
1637 spin_unlock_wr(&bufspin
);
1639 if ((bp
->b_flags
& B_LOCKED
) == 0 &&
1640 ((bp
->b_flags
& B_INVAL
) || (bp
->b_flags
& B_DELWRI
) == 0)) {
1645 * Something we can maybe free or reuse.
1647 if (bp
->b_bufsize
&& !(bp
->b_flags
& B_DELWRI
))
1651 * Final cleanup and unlock. Clear bits that are only used while a
1652 * buffer is actively locked.
1654 bp
->b_flags
&= ~(B_ORDERED
| B_NOCACHE
| B_RELBUF
);
1661 * Return backing pages held by the buffer 'bp' back to the VM system
1662 * if possible. The pages are freed if they are no longer valid or
1663 * attempt to free if it was used for direct I/O otherwise they are
1664 * sent to the page cache.
1666 * Pages that were marked busy are left alone and skipped.
1668 * The KVA mapping (b_data) for the underlying pages is removed by
1672 vfs_vmio_release(struct buf
*bp
)
1678 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
1679 m
= bp
->b_xio
.xio_pages
[i
];
1680 bp
->b_xio
.xio_pages
[i
] = NULL
;
1683 * The VFS is telling us this is not a meta-data buffer
1684 * even if it is backed by a block device.
1686 if (bp
->b_flags
& B_NOTMETA
)
1687 vm_page_flag_set(m
, PG_NOTMETA
);
1690 * This is a very important bit of code. We try to track
1691 * VM page use whether the pages are wired into the buffer
1692 * cache or not. While wired into the buffer cache the
1693 * bp tracks the act_count.
1695 * We can choose to place unwired pages on the inactive
1696 * queue (0) or active queue (1). If we place too many
1697 * on the active queue the queue will cycle the act_count
1698 * on pages we'd like to keep, just from single-use pages
1699 * (such as when doing a tar-up or file scan).
1701 if (bp
->b_act_count
< vm_cycle_point
)
1702 vm_page_unwire(m
, 0);
1704 vm_page_unwire(m
, 1);
1707 * We don't mess with busy pages, it is
1708 * the responsibility of the process that
1709 * busied the pages to deal with them.
1711 if ((m
->flags
& PG_BUSY
) || (m
->busy
!= 0))
1714 if (m
->wire_count
== 0) {
1715 vm_page_flag_clear(m
, PG_ZERO
);
1717 * Might as well free the page if we can and it has
1718 * no valid data. We also free the page if the
1719 * buffer was used for direct I/O.
1722 if ((bp
->b_flags
& B_ASYNC
) == 0 && !m
->valid
&&
1723 m
->hold_count
== 0) {
1725 vm_page_protect(m
, VM_PROT_NONE
);
1729 if (bp
->b_flags
& B_DIRECT
) {
1730 vm_page_try_to_free(m
);
1731 } else if (vm_page_count_severe()) {
1732 m
->act_count
= bp
->b_act_count
;
1733 vm_page_try_to_cache(m
);
1735 m
->act_count
= bp
->b_act_count
;
1740 pmap_qremove(trunc_page((vm_offset_t
) bp
->b_data
), bp
->b_xio
.xio_npages
);
1741 if (bp
->b_bufsize
) {
1745 bp
->b_xio
.xio_npages
= 0;
1746 bp
->b_flags
&= ~B_VMIO
;
1747 KKASSERT (LIST_FIRST(&bp
->b_dep
) == NULL
);
1758 * Implement clustered async writes for clearing out B_DELWRI buffers.
1759 * This is much better then the old way of writing only one buffer at
1760 * a time. Note that we may not be presented with the buffers in the
1761 * correct order, so we search for the cluster in both directions.
1763 * The buffer is locked on call.
1766 vfs_bio_awrite(struct buf
*bp
)
1770 off_t loffset
= bp
->b_loffset
;
1771 struct vnode
*vp
= bp
->b_vp
;
1778 * right now we support clustered writing only to regular files. If
1779 * we find a clusterable block we could be in the middle of a cluster
1780 * rather then at the beginning.
1782 * NOTE: b_bio1 contains the logical loffset and is aliased
1783 * to b_loffset. b_bio2 contains the translated block number.
1785 if ((vp
->v_type
== VREG
) &&
1786 (vp
->v_mount
!= 0) && /* Only on nodes that have the size info */
1787 (bp
->b_flags
& (B_CLUSTEROK
| B_INVAL
)) == B_CLUSTEROK
) {
1789 size
= vp
->v_mount
->mnt_stat
.f_iosize
;
1791 for (i
= size
; i
< MAXPHYS
; i
+= size
) {
1792 if ((bpa
= findblk(vp
, loffset
+ i
, FINDBLK_TEST
)) &&
1793 BUF_REFCNT(bpa
) == 0 &&
1794 ((bpa
->b_flags
& (B_DELWRI
| B_CLUSTEROK
| B_INVAL
)) ==
1795 (B_DELWRI
| B_CLUSTEROK
)) &&
1796 (bpa
->b_bufsize
== size
)) {
1797 if ((bpa
->b_bio2
.bio_offset
== NOOFFSET
) ||
1798 (bpa
->b_bio2
.bio_offset
!=
1799 bp
->b_bio2
.bio_offset
+ i
))
1805 for (j
= size
; i
+ j
<= MAXPHYS
&& j
<= loffset
; j
+= size
) {
1806 if ((bpa
= findblk(vp
, loffset
- j
, FINDBLK_TEST
)) &&
1807 BUF_REFCNT(bpa
) == 0 &&
1808 ((bpa
->b_flags
& (B_DELWRI
| B_CLUSTEROK
| B_INVAL
)) ==
1809 (B_DELWRI
| B_CLUSTEROK
)) &&
1810 (bpa
->b_bufsize
== size
)) {
1811 if ((bpa
->b_bio2
.bio_offset
== NOOFFSET
) ||
1812 (bpa
->b_bio2
.bio_offset
!=
1813 bp
->b_bio2
.bio_offset
- j
))
1823 * this is a possible cluster write
1825 if (nbytes
!= size
) {
1827 nwritten
= cluster_wbuild(vp
, size
,
1828 loffset
- j
, nbytes
);
1834 * default (old) behavior, writing out only one block
1836 * XXX returns b_bufsize instead of b_bcount for nwritten?
1838 nwritten
= bp
->b_bufsize
;
1848 * Find and initialize a new buffer header, freeing up existing buffers
1849 * in the bufqueues as necessary. The new buffer is returned locked.
1851 * Important: B_INVAL is not set. If the caller wishes to throw the
1852 * buffer away, the caller must set B_INVAL prior to calling brelse().
1855 * We have insufficient buffer headers
1856 * We have insufficient buffer space
1857 * buffer_map is too fragmented ( space reservation fails )
1858 * If we have to flush dirty buffers ( but we try to avoid this )
1860 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1861 * Instead we ask the buf daemon to do it for us. We attempt to
1862 * avoid piecemeal wakeups of the pageout daemon.
1867 getnewbuf(int blkflags
, int slptimeo
, int size
, int maxsize
)
1873 int slpflags
= (blkflags
& GETBLK_PCATCH
) ? PCATCH
: 0;
1874 static int flushingbufs
;
1877 * We can't afford to block since we might be holding a vnode lock,
1878 * which may prevent system daemons from running. We deal with
1879 * low-memory situations by proactively returning memory and running
1880 * async I/O rather then sync I/O.
1884 --getnewbufrestarts
;
1886 ++getnewbufrestarts
;
1889 * Setup for scan. If we do not have enough free buffers,
1890 * we setup a degenerate case that immediately fails. Note
1891 * that if we are specially marked process, we are allowed to
1892 * dip into our reserves.
1894 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1896 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1897 * However, there are a number of cases (defragging, reusing, ...)
1898 * where we cannot backup.
1900 nqindex
= BQUEUE_EMPTYKVA
;
1901 spin_lock_wr(&bufspin
);
1902 nbp
= TAILQ_FIRST(&bufqueues
[BQUEUE_EMPTYKVA
]);
1906 * If no EMPTYKVA buffers and we are either
1907 * defragging or reusing, locate a CLEAN buffer
1908 * to free or reuse. If bufspace useage is low
1909 * skip this step so we can allocate a new buffer.
1911 if (defrag
|| bufspace
>= lobufspace
) {
1912 nqindex
= BQUEUE_CLEAN
;
1913 nbp
= TAILQ_FIRST(&bufqueues
[BQUEUE_CLEAN
]);
1917 * If we could not find or were not allowed to reuse a
1918 * CLEAN buffer, check to see if it is ok to use an EMPTY
1919 * buffer. We can only use an EMPTY buffer if allocating
1920 * its KVA would not otherwise run us out of buffer space.
1922 if (nbp
== NULL
&& defrag
== 0 &&
1923 bufspace
+ maxsize
< hibufspace
) {
1924 nqindex
= BQUEUE_EMPTY
;
1925 nbp
= TAILQ_FIRST(&bufqueues
[BQUEUE_EMPTY
]);
1930 * Run scan, possibly freeing data and/or kva mappings on the fly
1933 * WARNING! bufspin is held!
1935 while ((bp
= nbp
) != NULL
) {
1936 int qindex
= nqindex
;
1938 nbp
= TAILQ_NEXT(bp
, b_freelist
);
1941 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
1942 * cycles through the queue twice before being selected.
1944 if (qindex
== BQUEUE_CLEAN
&&
1945 (bp
->b_flags
& B_AGE
) == 0 && nbp
) {
1946 bp
->b_flags
|= B_AGE
;
1947 TAILQ_REMOVE(&bufqueues
[qindex
], bp
, b_freelist
);
1948 TAILQ_INSERT_TAIL(&bufqueues
[qindex
], bp
, b_freelist
);
1953 * Calculate next bp ( we can only use it if we do not block
1954 * or do other fancy things ).
1959 nqindex
= BQUEUE_EMPTYKVA
;
1960 if ((nbp
= TAILQ_FIRST(&bufqueues
[BQUEUE_EMPTYKVA
])))
1963 case BQUEUE_EMPTYKVA
:
1964 nqindex
= BQUEUE_CLEAN
;
1965 if ((nbp
= TAILQ_FIRST(&bufqueues
[BQUEUE_CLEAN
])))
1979 KASSERT(bp
->b_qindex
== qindex
, ("getnewbuf: inconsistent queue %d bp %p", qindex
, bp
));
1982 * Note: we no longer distinguish between VMIO and non-VMIO
1986 KASSERT((bp
->b_flags
& B_DELWRI
) == 0, ("delwri buffer %p found in queue %d", bp
, qindex
));
1989 * If we are defragging then we need a buffer with
1990 * b_kvasize != 0. XXX this situation should no longer
1991 * occur, if defrag is non-zero the buffer's b_kvasize
1992 * should also be non-zero at this point. XXX
1994 if (defrag
&& bp
->b_kvasize
== 0) {
1995 kprintf("Warning: defrag empty buffer %p\n", bp
);
2000 * Start freeing the bp. This is somewhat involved. nbp
2001 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers
2002 * on the clean list must be disassociated from their
2003 * current vnode. Buffers on the empty[kva] lists have
2004 * already been disassociated.
2007 if (BUF_LOCK(bp
, LK_EXCLUSIVE
| LK_NOWAIT
) != 0) {
2008 spin_unlock_wr(&bufspin
);
2009 tsleep(&bd_request
, 0, "gnbxxx", hz
/ 100);
2012 if (bp
->b_qindex
!= qindex
) {
2013 spin_unlock_wr(&bufspin
);
2014 kprintf("getnewbuf: warning, BUF_LOCK blocked unexpectedly on buf %p index %d->%d, race corrected\n", bp
, qindex
, bp
->b_qindex
);
2018 bremfree_locked(bp
);
2019 spin_unlock_wr(&bufspin
);
2022 * Dependancies must be handled before we disassociate the
2025 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
2026 * be immediately disassociated. HAMMER then becomes
2027 * responsible for releasing the buffer.
2029 * NOTE: bufspin is UNLOCKED now.
2031 if (LIST_FIRST(&bp
->b_dep
) != NULL
) {
2035 if (bp
->b_flags
& B_LOCKED
) {
2039 KKASSERT(LIST_FIRST(&bp
->b_dep
) == NULL
);
2042 if (qindex
== BQUEUE_CLEAN
) {
2044 if (bp
->b_flags
& B_VMIO
) {
2046 vfs_vmio_release(bp
);
2055 * NOTE: nbp is now entirely invalid. We can only restart
2056 * the scan from this point on.
2058 * Get the rest of the buffer freed up. b_kva* is still
2059 * valid after this operation.
2062 KASSERT(bp
->b_vp
== NULL
, ("bp3 %p flags %08x vnode %p qindex %d unexpectededly still associated!", bp
, bp
->b_flags
, bp
->b_vp
, qindex
));
2063 KKASSERT((bp
->b_flags
& B_HASHED
) == 0);
2066 * critical section protection is not required when
2067 * scrapping a buffer's contents because it is already
2070 if (bp
->b_bufsize
) {
2076 bp
->b_flags
= B_BNOCLIP
;
2077 bp
->b_cmd
= BUF_CMD_DONE
;
2082 bp
->b_xio
.xio_npages
= 0;
2083 bp
->b_dirtyoff
= bp
->b_dirtyend
= 0;
2084 bp
->b_act_count
= ACT_INIT
;
2086 KKASSERT(LIST_FIRST(&bp
->b_dep
) == NULL
);
2088 if (blkflags
& GETBLK_BHEAVY
)
2089 bp
->b_flags
|= B_HEAVY
;
2092 * If we are defragging then free the buffer.
2095 bp
->b_flags
|= B_INVAL
;
2103 * If we are overcomitted then recover the buffer and its
2104 * KVM space. This occurs in rare situations when multiple
2105 * processes are blocked in getnewbuf() or allocbuf().
2107 if (bufspace
>= hibufspace
)
2109 if (flushingbufs
&& bp
->b_kvasize
!= 0) {
2110 bp
->b_flags
|= B_INVAL
;
2115 if (bufspace
< lobufspace
)
2118 /* NOT REACHED, bufspin not held */
2122 * If we exhausted our list, sleep as appropriate. We may have to
2123 * wakeup various daemons and write out some dirty buffers.
2125 * Generally we are sleeping due to insufficient buffer space.
2127 * NOTE: bufspin is held if bp is NULL, else it is not held.
2133 spin_unlock_wr(&bufspin
);
2135 flags
= VFS_BIO_NEED_BUFSPACE
;
2137 } else if (bufspace
>= hibufspace
) {
2139 flags
= VFS_BIO_NEED_BUFSPACE
;
2142 flags
= VFS_BIO_NEED_ANY
;
2145 needsbuffer
|= flags
;
2146 bd_speedup(); /* heeeelp */
2147 while (needsbuffer
& flags
) {
2148 if (tsleep(&needsbuffer
, slpflags
, waitmsg
, slptimeo
))
2153 * We finally have a valid bp. We aren't quite out of the
2154 * woods, we still have to reserve kva space. In order
2155 * to keep fragmentation sane we only allocate kva in
2158 * (bufspin is not held)
2160 maxsize
= (maxsize
+ BKVAMASK
) & ~BKVAMASK
;
2162 if (maxsize
!= bp
->b_kvasize
) {
2163 vm_offset_t addr
= 0;
2169 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
2170 vm_map_lock(&buffer_map
);
2172 if (vm_map_findspace(&buffer_map
,
2173 vm_map_min(&buffer_map
), maxsize
,
2174 maxsize
, 0, &addr
)) {
2176 * Uh oh. Buffer map is too fragmented. We
2177 * must defragment the map.
2179 vm_map_unlock(&buffer_map
);
2180 vm_map_entry_release(count
);
2183 bp
->b_flags
|= B_INVAL
;
2189 vm_map_insert(&buffer_map
, &count
,
2191 addr
, addr
+ maxsize
,
2193 VM_PROT_ALL
, VM_PROT_ALL
,
2196 bp
->b_kvabase
= (caddr_t
) addr
;
2197 bp
->b_kvasize
= maxsize
;
2198 bufspace
+= bp
->b_kvasize
;
2201 vm_map_unlock(&buffer_map
);
2202 vm_map_entry_release(count
);
2205 bp
->b_data
= bp
->b_kvabase
;
2211 * This routine is called in an emergency to recover VM pages from the
2212 * buffer cache by cashing in clean buffers. The idea is to recover
2213 * enough pages to be able to satisfy a stuck bio_page_alloc().
2216 recoverbufpages(void)
2223 spin_lock_wr(&bufspin
);
2224 while (bytes
< MAXBSIZE
) {
2225 bp
= TAILQ_FIRST(&bufqueues
[BQUEUE_CLEAN
]);
2230 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
2231 * cycles through the queue twice before being selected.
2233 if ((bp
->b_flags
& B_AGE
) == 0 && TAILQ_NEXT(bp
, b_freelist
)) {
2234 bp
->b_flags
|= B_AGE
;
2235 TAILQ_REMOVE(&bufqueues
[BQUEUE_CLEAN
], bp
, b_freelist
);
2236 TAILQ_INSERT_TAIL(&bufqueues
[BQUEUE_CLEAN
],
2244 KKASSERT(bp
->b_qindex
== BQUEUE_CLEAN
);
2245 KKASSERT((bp
->b_flags
& B_DELWRI
) == 0);
2248 * Start freeing the bp. This is somewhat involved.
2250 * Buffers on the clean list must be disassociated from
2251 * their current vnode
2254 if (BUF_LOCK(bp
, LK_EXCLUSIVE
| LK_NOWAIT
) != 0) {
2255 kprintf("recoverbufpages: warning, locked buf %p, race corrected\n", bp
);
2256 tsleep(&bd_request
, 0, "gnbxxx", hz
/ 100);
2259 if (bp
->b_qindex
!= BQUEUE_CLEAN
) {
2260 kprintf("recoverbufpages: warning, BUF_LOCK blocked unexpectedly on buf %p index %d, race corrected\n", bp
, bp
->b_qindex
);
2264 bremfree_locked(bp
);
2265 spin_unlock_wr(&bufspin
);
2268 * Dependancies must be handled before we disassociate the
2271 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
2272 * be immediately disassociated. HAMMER then becomes
2273 * responsible for releasing the buffer.
2275 if (LIST_FIRST(&bp
->b_dep
) != NULL
) {
2277 if (bp
->b_flags
& B_LOCKED
) {
2279 spin_lock_wr(&bufspin
);
2282 KKASSERT(LIST_FIRST(&bp
->b_dep
) == NULL
);
2285 bytes
+= bp
->b_bufsize
;
2288 if (bp
->b_flags
& B_VMIO
) {
2289 bp
->b_flags
|= B_DIRECT
; /* try to free pages */
2290 vfs_vmio_release(bp
);
2295 KKASSERT(bp
->b_vp
== NULL
);
2296 KKASSERT((bp
->b_flags
& B_HASHED
) == 0);
2299 * critical section protection is not required when
2300 * scrapping a buffer's contents because it is already
2307 bp
->b_flags
= B_BNOCLIP
;
2308 bp
->b_cmd
= BUF_CMD_DONE
;
2313 bp
->b_xio
.xio_npages
= 0;
2314 bp
->b_dirtyoff
= bp
->b_dirtyend
= 0;
2316 KKASSERT(LIST_FIRST(&bp
->b_dep
) == NULL
);
2318 bp
->b_flags
|= B_INVAL
;
2321 spin_lock_wr(&bufspin
);
2323 spin_unlock_wr(&bufspin
);
2330 * Buffer flushing daemon. Buffers are normally flushed by the
2331 * update daemon but if it cannot keep up this process starts to
2332 * take the load in an attempt to prevent getnewbuf() from blocking.
2334 * Once a flush is initiated it does not stop until the number
2335 * of buffers falls below lodirtybuffers, but we will wake up anyone
2336 * waiting at the mid-point.
2339 static struct kproc_desc buf_kp
= {
2344 SYSINIT(bufdaemon
, SI_SUB_KTHREAD_BUF
, SI_ORDER_FIRST
,
2345 kproc_start
, &buf_kp
)
2347 static struct kproc_desc bufhw_kp
= {
2352 SYSINIT(bufdaemon_hw
, SI_SUB_KTHREAD_BUF
, SI_ORDER_FIRST
,
2353 kproc_start
, &bufhw_kp
)
2361 * This process needs to be suspended prior to shutdown sync.
2363 EVENTHANDLER_REGISTER(shutdown_pre_sync
, shutdown_kproc
,
2364 bufdaemon_td
, SHUTDOWN_PRI_LAST
);
2365 curthread
->td_flags
|= TDF_SYSTHREAD
;
2368 * This process is allowed to take the buffer cache to the limit
2373 kproc_suspend_loop();
2376 * Do the flush as long as the number of dirty buffers
2377 * (including those running) exceeds lodirtybufspace.
2379 * When flushing limit running I/O to hirunningspace
2380 * Do the flush. Limit the amount of in-transit I/O we
2381 * allow to build up, otherwise we would completely saturate
2382 * the I/O system. Wakeup any waiting processes before we
2383 * normally would so they can run in parallel with our drain.
2385 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2386 * but because we split the operation into two threads we
2387 * have to cut it in half for each thread.
2389 waitrunningbufspace();
2390 limit
= lodirtybufspace
/ 2;
2391 while (runningbufspace
+ dirtybufspace
> limit
||
2392 dirtybufcount
- dirtybufcounthw
>= nbuf
/ 2) {
2393 if (flushbufqueues(BQUEUE_DIRTY
) == 0)
2395 if (runningbufspace
< hirunningspace
)
2397 waitrunningbufspace();
2401 * We reached our low water mark, reset the
2402 * request and sleep until we are needed again.
2403 * The sleep is just so the suspend code works.
2405 spin_lock_wr(&needsbuffer_spin
);
2406 if (bd_request
== 0) {
2407 ssleep(&bd_request
, &needsbuffer_spin
, 0,
2411 spin_unlock_wr(&needsbuffer_spin
);
2421 * This process needs to be suspended prior to shutdown sync.
2423 EVENTHANDLER_REGISTER(shutdown_pre_sync
, shutdown_kproc
,
2424 bufdaemonhw_td
, SHUTDOWN_PRI_LAST
);
2425 curthread
->td_flags
|= TDF_SYSTHREAD
;
2428 * This process is allowed to take the buffer cache to the limit
2433 kproc_suspend_loop();
2436 * Do the flush. Limit the amount of in-transit I/O we
2437 * allow to build up, otherwise we would completely saturate
2438 * the I/O system. Wakeup any waiting processes before we
2439 * normally would so they can run in parallel with our drain.
2441 * Once we decide to flush push the queued I/O up to
2442 * hirunningspace in order to trigger bursting by the bioq
2445 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2446 * but because we split the operation into two threads we
2447 * have to cut it in half for each thread.
2449 waitrunningbufspace();
2450 limit
= lodirtybufspace
/ 2;
2451 while (runningbufspace
+ dirtybufspacehw
> limit
||
2452 dirtybufcounthw
>= nbuf
/ 2) {
2453 if (flushbufqueues(BQUEUE_DIRTY_HW
) == 0)
2455 if (runningbufspace
< hirunningspace
)
2457 waitrunningbufspace();
2461 * We reached our low water mark, reset the
2462 * request and sleep until we are needed again.
2463 * The sleep is just so the suspend code works.
2465 spin_lock_wr(&needsbuffer_spin
);
2466 if (bd_request_hw
== 0) {
2467 ssleep(&bd_request_hw
, &needsbuffer_spin
, 0,
2471 spin_unlock_wr(&needsbuffer_spin
);
2478 * Try to flush a buffer in the dirty queue. We must be careful to
2479 * free up B_INVAL buffers instead of write them, which NFS is
2480 * particularly sensitive to.
2482 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate
2483 * that we really want to try to get the buffer out and reuse it
2484 * due to the write load on the machine.
2487 flushbufqueues(bufq_type_t q
)
2493 spin_lock_wr(&bufspin
);
2496 bp
= TAILQ_FIRST(&bufqueues
[q
]);
2498 KASSERT((bp
->b_flags
& B_DELWRI
),
2499 ("unexpected clean buffer %p", bp
));
2501 if (bp
->b_flags
& B_DELWRI
) {
2502 if (bp
->b_flags
& B_INVAL
) {
2503 spin_unlock_wr(&bufspin
);
2505 if (BUF_LOCK(bp
, LK_EXCLUSIVE
| LK_NOWAIT
) != 0)
2506 panic("flushbufqueues: locked buf");
2512 if (LIST_FIRST(&bp
->b_dep
) != NULL
&&
2513 (bp
->b_flags
& B_DEFERRED
) == 0 &&
2514 buf_countdeps(bp
, 0)) {
2515 TAILQ_REMOVE(&bufqueues
[q
], bp
, b_freelist
);
2516 TAILQ_INSERT_TAIL(&bufqueues
[q
], bp
,
2518 bp
->b_flags
|= B_DEFERRED
;
2519 bp
= TAILQ_FIRST(&bufqueues
[q
]);
2524 * Only write it out if we can successfully lock
2525 * it. If the buffer has a dependancy,
2526 * buf_checkwrite must also return 0 for us to
2527 * be able to initate the write.
2529 * If the buffer is flagged B_ERROR it may be
2530 * requeued over and over again, we try to
2531 * avoid a live lock.
2533 if (BUF_LOCK(bp
, LK_EXCLUSIVE
| LK_NOWAIT
) == 0) {
2534 spin_unlock_wr(&bufspin
);
2536 if (LIST_FIRST(&bp
->b_dep
) != NULL
&&
2537 buf_checkwrite(bp
)) {
2540 } else if (bp
->b_flags
& B_ERROR
) {
2541 tsleep(bp
, 0, "bioer", 1);
2542 bp
->b_flags
&= ~B_AGE
;
2545 bp
->b_flags
|= B_AGE
;
2552 bp
= TAILQ_NEXT(bp
, b_freelist
);
2555 spin_unlock_wr(&bufspin
);
2562 * Returns true if no I/O is needed to access the associated VM object.
2563 * This is like findblk except it also hunts around in the VM system for
2566 * Note that we ignore vm_page_free() races from interrupts against our
2567 * lookup, since if the caller is not protected our return value will not
2568 * be any more valid then otherwise once we exit the critical section.
2571 inmem(struct vnode
*vp
, off_t loffset
)
2574 vm_offset_t toff
, tinc
, size
;
2577 if (findblk(vp
, loffset
, FINDBLK_TEST
))
2579 if (vp
->v_mount
== NULL
)
2581 if ((obj
= vp
->v_object
) == NULL
)
2585 if (size
> vp
->v_mount
->mnt_stat
.f_iosize
)
2586 size
= vp
->v_mount
->mnt_stat
.f_iosize
;
2588 for (toff
= 0; toff
< vp
->v_mount
->mnt_stat
.f_iosize
; toff
+= tinc
) {
2589 m
= vm_page_lookup(obj
, OFF_TO_IDX(loffset
+ toff
));
2593 if (tinc
> PAGE_SIZE
- ((toff
+ loffset
) & PAGE_MASK
))
2594 tinc
= PAGE_SIZE
- ((toff
+ loffset
) & PAGE_MASK
);
2595 if (vm_page_is_valid(m
,
2596 (vm_offset_t
) ((toff
+ loffset
) & PAGE_MASK
), tinc
) == 0)
2605 * Locate and return the specified buffer. Unless flagged otherwise,
2606 * a locked buffer will be returned if it exists or NULL if it does not.
2608 * findblk()'d buffers are still on the bufqueues and if you intend
2609 * to use your (locked NON-TEST) buffer you need to bremfree(bp)
2610 * and possibly do other stuff to it.
2612 * FINDBLK_TEST - Do not lock the buffer. The caller is responsible
2613 * for locking the buffer and ensuring that it remains
2614 * the desired buffer after locking.
2616 * FINDBLK_NBLOCK - Lock the buffer non-blocking. If we are unable
2617 * to acquire the lock we return NULL, even if the
2620 * (0) - Lock the buffer blocking.
2625 findblk(struct vnode
*vp
, off_t loffset
, int flags
)
2631 lkflags
= LK_EXCLUSIVE
;
2632 if (flags
& FINDBLK_NBLOCK
)
2633 lkflags
|= LK_NOWAIT
;
2636 lwkt_gettoken(&vlock
, &vp
->v_token
);
2637 bp
= buf_rb_hash_RB_LOOKUP(&vp
->v_rbhash_tree
, loffset
);
2638 lwkt_reltoken(&vlock
);
2639 if (bp
== NULL
|| (flags
& FINDBLK_TEST
))
2641 if (BUF_LOCK(bp
, lkflags
)) {
2645 if (bp
->b_vp
== vp
&& bp
->b_loffset
== loffset
)
2655 * Similar to getblk() except only returns the buffer if it is
2656 * B_CACHE and requires no other manipulation. Otherwise NULL
2659 * If B_RAM is set the buffer might be just fine, but we return
2660 * NULL anyway because we want the code to fall through to the
2661 * cluster read. Otherwise read-ahead breaks.
2664 getcacheblk(struct vnode
*vp
, off_t loffset
)
2668 bp
= findblk(vp
, loffset
, 0);
2670 if ((bp
->b_flags
& (B_INVAL
| B_CACHE
| B_RAM
)) == B_CACHE
) {
2671 bp
->b_flags
&= ~B_AGE
;
2684 * Get a block given a specified block and offset into a file/device.
2685 * B_INVAL may or may not be set on return. The caller should clear
2686 * B_INVAL prior to initiating a READ.
2688 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2689 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2690 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer
2691 * without doing any of those things the system will likely believe
2692 * the buffer to be valid (especially if it is not B_VMIO), and the
2693 * next getblk() will return the buffer with B_CACHE set.
2695 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2696 * an existing buffer.
2698 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2699 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2700 * and then cleared based on the backing VM. If the previous buffer is
2701 * non-0-sized but invalid, B_CACHE will be cleared.
2703 * If getblk() must create a new buffer, the new buffer is returned with
2704 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2705 * case it is returned with B_INVAL clear and B_CACHE set based on the
2708 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2709 * B_CACHE bit is clear.
2711 * What this means, basically, is that the caller should use B_CACHE to
2712 * determine whether the buffer is fully valid or not and should clear
2713 * B_INVAL prior to issuing a read. If the caller intends to validate
2714 * the buffer by loading its data area with something, the caller needs
2715 * to clear B_INVAL. If the caller does this without issuing an I/O,
2716 * the caller should set B_CACHE ( as an optimization ), else the caller
2717 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2718 * a write attempt or if it was a successfull read. If the caller
2719 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2720 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2724 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return
2725 * GETBLK_BHEAVY - heavy-weight buffer cache buffer
2730 getblk(struct vnode
*vp
, off_t loffset
, int size
, int blkflags
, int slptimeo
)
2733 int slpflags
= (blkflags
& GETBLK_PCATCH
) ? PCATCH
: 0;
2737 if (size
> MAXBSIZE
)
2738 panic("getblk: size(%d) > MAXBSIZE(%d)", size
, MAXBSIZE
);
2739 if (vp
->v_object
== NULL
)
2740 panic("getblk: vnode %p has no object!", vp
);
2743 if ((bp
= findblk(vp
, loffset
, FINDBLK_TEST
)) != NULL
) {
2745 * The buffer was found in the cache, but we need to lock it.
2746 * Even with LK_NOWAIT the lockmgr may break our critical
2747 * section, so double-check the validity of the buffer
2748 * once the lock has been obtained.
2750 if (BUF_LOCK(bp
, LK_EXCLUSIVE
| LK_NOWAIT
)) {
2751 if (blkflags
& GETBLK_NOWAIT
)
2753 lkflags
= LK_EXCLUSIVE
| LK_SLEEPFAIL
;
2754 if (blkflags
& GETBLK_PCATCH
)
2755 lkflags
|= LK_PCATCH
;
2756 error
= BUF_TIMELOCK(bp
, lkflags
, "getblk", slptimeo
);
2758 if (error
== ENOLCK
)
2762 /* buffer may have changed on us */
2766 * Once the buffer has been locked, make sure we didn't race
2767 * a buffer recyclement. Buffers that are no longer hashed
2768 * will have b_vp == NULL, so this takes care of that check
2771 if (bp
->b_vp
!= vp
|| bp
->b_loffset
!= loffset
) {
2772 kprintf("Warning buffer %p (vp %p loffset %lld) "
2774 bp
, vp
, (long long)loffset
);
2780 * If SZMATCH any pre-existing buffer must be of the requested
2781 * size or NULL is returned. The caller absolutely does not
2782 * want getblk() to bwrite() the buffer on a size mismatch.
2784 if ((blkflags
& GETBLK_SZMATCH
) && size
!= bp
->b_bcount
) {
2790 * All vnode-based buffers must be backed by a VM object.
2792 KKASSERT(bp
->b_flags
& B_VMIO
);
2793 KKASSERT(bp
->b_cmd
== BUF_CMD_DONE
);
2794 bp
->b_flags
&= ~B_AGE
;
2797 * Make sure that B_INVAL buffers do not have a cached
2798 * block number translation.
2800 if ((bp
->b_flags
& B_INVAL
) && (bp
->b_bio2
.bio_offset
!= NOOFFSET
)) {
2801 kprintf("Warning invalid buffer %p (vp %p loffset %lld)"
2802 " did not have cleared bio_offset cache\n",
2803 bp
, vp
, (long long)loffset
);
2804 clearbiocache(&bp
->b_bio2
);
2808 * The buffer is locked. B_CACHE is cleared if the buffer is
2811 if (bp
->b_flags
& B_INVAL
)
2812 bp
->b_flags
&= ~B_CACHE
;
2816 * Any size inconsistancy with a dirty buffer or a buffer
2817 * with a softupdates dependancy must be resolved. Resizing
2818 * the buffer in such circumstances can lead to problems.
2820 * Dirty or dependant buffers are written synchronously.
2821 * Other types of buffers are simply released and
2822 * reconstituted as they may be backed by valid, dirty VM
2823 * pages (but not marked B_DELWRI).
2825 * NFS NOTE: NFS buffers which straddle EOF are oddly-sized
2826 * and may be left over from a prior truncation (and thus
2827 * no longer represent the actual EOF point), so we
2828 * definitely do not want to B_NOCACHE the backing store.
2830 if (size
!= bp
->b_bcount
) {
2832 if (bp
->b_flags
& B_DELWRI
) {
2833 bp
->b_flags
|= B_RELBUF
;
2835 } else if (LIST_FIRST(&bp
->b_dep
)) {
2836 bp
->b_flags
|= B_RELBUF
;
2839 bp
->b_flags
|= B_RELBUF
;
2845 KKASSERT(size
<= bp
->b_kvasize
);
2846 KASSERT(bp
->b_loffset
!= NOOFFSET
,
2847 ("getblk: no buffer offset"));
2850 * A buffer with B_DELWRI set and B_CACHE clear must
2851 * be committed before we can return the buffer in
2852 * order to prevent the caller from issuing a read
2853 * ( due to B_CACHE not being set ) and overwriting
2856 * Most callers, including NFS and FFS, need this to
2857 * operate properly either because they assume they
2858 * can issue a read if B_CACHE is not set, or because
2859 * ( for example ) an uncached B_DELWRI might loop due
2860 * to softupdates re-dirtying the buffer. In the latter
2861 * case, B_CACHE is set after the first write completes,
2862 * preventing further loops.
2864 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2865 * above while extending the buffer, we cannot allow the
2866 * buffer to remain with B_CACHE set after the write
2867 * completes or it will represent a corrupt state. To
2868 * deal with this we set B_NOCACHE to scrap the buffer
2871 * XXX Should this be B_RELBUF instead of B_NOCACHE?
2872 * I'm not even sure this state is still possible
2873 * now that getblk() writes out any dirty buffers
2876 * We might be able to do something fancy, like setting
2877 * B_CACHE in bwrite() except if B_DELWRI is already set,
2878 * so the below call doesn't set B_CACHE, but that gets real
2879 * confusing. This is much easier.
2882 if ((bp
->b_flags
& (B_CACHE
|B_DELWRI
)) == B_DELWRI
) {
2884 kprintf("getblk: Warning, bp %p loff=%jx DELWRI set "
2885 "and CACHE clear, b_flags %08x\n",
2886 bp
, (intmax_t)bp
->b_loffset
, bp
->b_flags
);
2887 bp
->b_flags
|= B_NOCACHE
;
2894 * Buffer is not in-core, create new buffer. The buffer
2895 * returned by getnewbuf() is locked. Note that the returned
2896 * buffer is also considered valid (not marked B_INVAL).
2898 * Calculating the offset for the I/O requires figuring out
2899 * the block size. We use DEV_BSIZE for VBLK or VCHR and
2900 * the mount's f_iosize otherwise. If the vnode does not
2901 * have an associated mount we assume that the passed size is
2904 * Note that vn_isdisk() cannot be used here since it may
2905 * return a failure for numerous reasons. Note that the
2906 * buffer size may be larger then the block size (the caller
2907 * will use block numbers with the proper multiple). Beware
2908 * of using any v_* fields which are part of unions. In
2909 * particular, in DragonFly the mount point overloading
2910 * mechanism uses the namecache only and the underlying
2911 * directory vnode is not a special case.
2915 if (vp
->v_type
== VBLK
|| vp
->v_type
== VCHR
)
2917 else if (vp
->v_mount
)
2918 bsize
= vp
->v_mount
->mnt_stat
.f_iosize
;
2922 maxsize
= size
+ (loffset
& PAGE_MASK
);
2923 maxsize
= imax(maxsize
, bsize
);
2925 bp
= getnewbuf(blkflags
, slptimeo
, size
, maxsize
);
2927 if (slpflags
|| slptimeo
)
2933 * Atomically insert the buffer into the hash, so that it can
2934 * be found by findblk().
2936 * If bgetvp() returns non-zero a collision occured, and the
2937 * bp will not be associated with the vnode.
2939 * Make sure the translation layer has been cleared.
2941 bp
->b_loffset
= loffset
;
2942 bp
->b_bio2
.bio_offset
= NOOFFSET
;
2943 /* bp->b_bio2.bio_next = NULL; */
2945 if (bgetvp(vp
, bp
)) {
2946 bp
->b_flags
|= B_INVAL
;
2952 * All vnode-based buffers must be backed by a VM object.
2954 KKASSERT(vp
->v_object
!= NULL
);
2955 bp
->b_flags
|= B_VMIO
;
2956 KKASSERT(bp
->b_cmd
== BUF_CMD_DONE
);
2968 * Reacquire a buffer that was previously released to the locked queue,
2969 * or reacquire a buffer which is interlocked by having bioops->io_deallocate
2970 * set B_LOCKED (which handles the acquisition race).
2972 * To this end, either B_LOCKED must be set or the dependancy list must be
2978 regetblk(struct buf
*bp
)
2980 KKASSERT((bp
->b_flags
& B_LOCKED
) || LIST_FIRST(&bp
->b_dep
) != NULL
);
2981 BUF_LOCK(bp
, LK_EXCLUSIVE
| LK_RETRY
);
2988 * Get an empty, disassociated buffer of given size. The buffer is
2989 * initially set to B_INVAL.
2991 * critical section protection is not required for the allocbuf()
2992 * call because races are impossible here.
3002 maxsize
= (size
+ BKVAMASK
) & ~BKVAMASK
;
3004 while ((bp
= getnewbuf(0, 0, size
, maxsize
)) == 0)
3009 bp
->b_flags
|= B_INVAL
; /* b_dep cleared by getnewbuf() */
3017 * This code constitutes the buffer memory from either anonymous system
3018 * memory (in the case of non-VMIO operations) or from an associated
3019 * VM object (in the case of VMIO operations). This code is able to
3020 * resize a buffer up or down.
3022 * Note that this code is tricky, and has many complications to resolve
3023 * deadlock or inconsistant data situations. Tread lightly!!!
3024 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
3025 * the caller. Calling this code willy nilly can result in the loss of data.
3027 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
3028 * B_CACHE for the non-VMIO case.
3030 * This routine does not need to be called from a critical section but you
3031 * must own the buffer.
3036 allocbuf(struct buf
*bp
, int size
)
3038 int newbsize
, mbsize
;
3041 if (BUF_REFCNT(bp
) == 0)
3042 panic("allocbuf: buffer not busy");
3044 if (bp
->b_kvasize
< size
)
3045 panic("allocbuf: buffer too small");
3047 if ((bp
->b_flags
& B_VMIO
) == 0) {
3051 * Just get anonymous memory from the kernel. Don't
3052 * mess with B_CACHE.
3054 mbsize
= (size
+ DEV_BSIZE
- 1) & ~(DEV_BSIZE
- 1);
3055 if (bp
->b_flags
& B_MALLOC
)
3058 newbsize
= round_page(size
);
3060 if (newbsize
< bp
->b_bufsize
) {
3062 * Malloced buffers are not shrunk
3064 if (bp
->b_flags
& B_MALLOC
) {
3066 bp
->b_bcount
= size
;
3068 kfree(bp
->b_data
, M_BIOBUF
);
3069 if (bp
->b_bufsize
) {
3070 bufmallocspace
-= bp
->b_bufsize
;
3074 bp
->b_data
= bp
->b_kvabase
;
3076 bp
->b_flags
&= ~B_MALLOC
;
3082 (vm_offset_t
) bp
->b_data
+ newbsize
,
3083 (vm_offset_t
) bp
->b_data
+ bp
->b_bufsize
);
3084 } else if (newbsize
> bp
->b_bufsize
) {
3086 * We only use malloced memory on the first allocation.
3087 * and revert to page-allocated memory when the buffer
3090 if ((bufmallocspace
< maxbufmallocspace
) &&
3091 (bp
->b_bufsize
== 0) &&
3092 (mbsize
<= PAGE_SIZE
/2)) {
3094 bp
->b_data
= kmalloc(mbsize
, M_BIOBUF
, M_WAITOK
);
3095 bp
->b_bufsize
= mbsize
;
3096 bp
->b_bcount
= size
;
3097 bp
->b_flags
|= B_MALLOC
;
3098 bufmallocspace
+= mbsize
;
3104 * If the buffer is growing on its other-than-first
3105 * allocation, then we revert to the page-allocation
3108 if (bp
->b_flags
& B_MALLOC
) {
3109 origbuf
= bp
->b_data
;
3110 origbufsize
= bp
->b_bufsize
;
3111 bp
->b_data
= bp
->b_kvabase
;
3112 if (bp
->b_bufsize
) {
3113 bufmallocspace
-= bp
->b_bufsize
;
3117 bp
->b_flags
&= ~B_MALLOC
;
3118 newbsize
= round_page(newbsize
);
3122 (vm_offset_t
) bp
->b_data
+ bp
->b_bufsize
,
3123 (vm_offset_t
) bp
->b_data
+ newbsize
);
3125 bcopy(origbuf
, bp
->b_data
, origbufsize
);
3126 kfree(origbuf
, M_BIOBUF
);
3133 newbsize
= (size
+ DEV_BSIZE
- 1) & ~(DEV_BSIZE
- 1);
3134 desiredpages
= ((int)(bp
->b_loffset
& PAGE_MASK
) +
3135 newbsize
+ PAGE_MASK
) >> PAGE_SHIFT
;
3136 KKASSERT(desiredpages
<= XIO_INTERNAL_PAGES
);
3138 if (bp
->b_flags
& B_MALLOC
)
3139 panic("allocbuf: VMIO buffer can't be malloced");
3141 * Set B_CACHE initially if buffer is 0 length or will become
3144 if (size
== 0 || bp
->b_bufsize
== 0)
3145 bp
->b_flags
|= B_CACHE
;
3147 if (newbsize
< bp
->b_bufsize
) {
3149 * DEV_BSIZE aligned new buffer size is less then the
3150 * DEV_BSIZE aligned existing buffer size. Figure out
3151 * if we have to remove any pages.
3153 if (desiredpages
< bp
->b_xio
.xio_npages
) {
3154 for (i
= desiredpages
; i
< bp
->b_xio
.xio_npages
; i
++) {
3156 * the page is not freed here -- it
3157 * is the responsibility of
3158 * vnode_pager_setsize
3160 m
= bp
->b_xio
.xio_pages
[i
];
3161 KASSERT(m
!= bogus_page
,
3162 ("allocbuf: bogus page found"));
3163 while (vm_page_sleep_busy(m
, TRUE
, "biodep"))
3166 bp
->b_xio
.xio_pages
[i
] = NULL
;
3167 vm_page_unwire(m
, 0);
3169 pmap_qremove((vm_offset_t
) trunc_page((vm_offset_t
)bp
->b_data
) +
3170 (desiredpages
<< PAGE_SHIFT
), (bp
->b_xio
.xio_npages
- desiredpages
));
3171 bp
->b_xio
.xio_npages
= desiredpages
;
3173 } else if (size
> bp
->b_bcount
) {
3175 * We are growing the buffer, possibly in a
3176 * byte-granular fashion.
3184 * Step 1, bring in the VM pages from the object,
3185 * allocating them if necessary. We must clear
3186 * B_CACHE if these pages are not valid for the
3187 * range covered by the buffer.
3189 * critical section protection is required to protect
3190 * against interrupts unbusying and freeing pages
3191 * between our vm_page_lookup() and our
3192 * busycheck/wiring call.
3198 while (bp
->b_xio
.xio_npages
< desiredpages
) {
3202 pi
= OFF_TO_IDX(bp
->b_loffset
) + bp
->b_xio
.xio_npages
;
3203 if ((m
= vm_page_lookup(obj
, pi
)) == NULL
) {
3205 * note: must allocate system pages
3206 * since blocking here could intefere
3207 * with paging I/O, no matter which
3210 m
= bio_page_alloc(obj
, pi
, desiredpages
- bp
->b_xio
.xio_npages
);
3214 vm_page_flag_clear(m
, PG_ZERO
);
3215 bp
->b_flags
&= ~B_CACHE
;
3216 bp
->b_xio
.xio_pages
[bp
->b_xio
.xio_npages
] = m
;
3217 ++bp
->b_xio
.xio_npages
;
3223 * We found a page. If we have to sleep on it,
3224 * retry because it might have gotten freed out
3227 * We can only test PG_BUSY here. Blocking on
3228 * m->busy might lead to a deadlock:
3230 * vm_fault->getpages->cluster_read->allocbuf
3234 if (vm_page_sleep_busy(m
, FALSE
, "pgtblk"))
3236 vm_page_flag_clear(m
, PG_ZERO
);
3238 bp
->b_xio
.xio_pages
[bp
->b_xio
.xio_npages
] = m
;
3239 ++bp
->b_xio
.xio_npages
;
3240 if (bp
->b_act_count
< m
->act_count
)
3241 bp
->b_act_count
= m
->act_count
;
3246 * Step 2. We've loaded the pages into the buffer,
3247 * we have to figure out if we can still have B_CACHE
3248 * set. Note that B_CACHE is set according to the
3249 * byte-granular range ( bcount and size ), not the
3250 * aligned range ( newbsize ).
3252 * The VM test is against m->valid, which is DEV_BSIZE
3253 * aligned. Needless to say, the validity of the data
3254 * needs to also be DEV_BSIZE aligned. Note that this
3255 * fails with NFS if the server or some other client
3256 * extends the file's EOF. If our buffer is resized,
3257 * B_CACHE may remain set! XXX
3260 toff
= bp
->b_bcount
;
3261 tinc
= PAGE_SIZE
- ((bp
->b_loffset
+ toff
) & PAGE_MASK
);
3263 while ((bp
->b_flags
& B_CACHE
) && toff
< size
) {
3266 if (tinc
> (size
- toff
))
3269 pi
= ((bp
->b_loffset
& PAGE_MASK
) + toff
) >>
3277 bp
->b_xio
.xio_pages
[pi
]
3284 * Step 3, fixup the KVM pmap. Remember that
3285 * bp->b_data is relative to bp->b_loffset, but
3286 * bp->b_loffset may be offset into the first page.
3289 bp
->b_data
= (caddr_t
)
3290 trunc_page((vm_offset_t
)bp
->b_data
);
3292 (vm_offset_t
)bp
->b_data
,
3293 bp
->b_xio
.xio_pages
,
3294 bp
->b_xio
.xio_npages
3296 bp
->b_data
= (caddr_t
)((vm_offset_t
)bp
->b_data
|
3297 (vm_offset_t
)(bp
->b_loffset
& PAGE_MASK
));
3301 /* adjust space use on already-dirty buffer */
3302 if (bp
->b_flags
& B_DELWRI
) {
3303 dirtybufspace
+= newbsize
- bp
->b_bufsize
;
3304 if (bp
->b_flags
& B_HEAVY
)
3305 dirtybufspacehw
+= newbsize
- bp
->b_bufsize
;
3307 if (newbsize
< bp
->b_bufsize
)
3309 bp
->b_bufsize
= newbsize
; /* actual buffer allocation */
3310 bp
->b_bcount
= size
; /* requested buffer size */
3317 * Wait for buffer I/O completion, returning error status. B_EINTR
3318 * is converted into an EINTR error but not cleared (since a chain
3319 * of biowait() calls may occur).
3321 * On return bpdone() will have been called but the buffer will remain
3322 * locked and will not have been brelse()'d.
3324 * NOTE! If a timeout is specified and ETIMEDOUT occurs the I/O is
3325 * likely still in progress on return.
3327 * NOTE! This operation is on a BIO, not a BUF.
3329 * NOTE! BIO_DONE is cleared by vn_strategy()
3334 _biowait(struct bio
*bio
, const char *wmesg
, int to
)
3336 struct buf
*bp
= bio
->bio_buf
;
3341 KKASSERT(bio
== &bp
->b_bio1
);
3343 flags
= bio
->bio_flags
;
3344 if (flags
& BIO_DONE
)
3346 tsleep_interlock(bio
, 0);
3347 nflags
= flags
| BIO_WANT
;
3348 tsleep_interlock(bio
, 0);
3349 if (atomic_cmpset_int(&bio
->bio_flags
, flags
, nflags
)) {
3351 error
= tsleep(bio
, PINTERLOCKED
, wmesg
, to
);
3352 else if (bp
->b_cmd
== BUF_CMD_READ
)
3353 error
= tsleep(bio
, PINTERLOCKED
, "biord", to
);
3355 error
= tsleep(bio
, PINTERLOCKED
, "biowr", to
);
3357 kprintf("tsleep error biowait %d\n", error
);
3367 KKASSERT(bp
->b_cmd
== BUF_CMD_DONE
);
3368 bio
->bio_flags
&= ~(BIO_DONE
| BIO_SYNC
);
3369 if (bp
->b_flags
& B_EINTR
)
3371 if (bp
->b_flags
& B_ERROR
)
3372 return (bp
->b_error
? bp
->b_error
: EIO
);
3377 biowait(struct bio
*bio
, const char *wmesg
)
3379 return(_biowait(bio
, wmesg
, 0));
3383 biowait_timeout(struct bio
*bio
, const char *wmesg
, int to
)
3385 return(_biowait(bio
, wmesg
, to
));
3389 * This associates a tracking count with an I/O. vn_strategy() and
3390 * dev_dstrategy() do this automatically but there are a few cases
3391 * where a vnode or device layer is bypassed when a block translation
3392 * is cached. In such cases bio_start_transaction() may be called on
3393 * the bypassed layers so the system gets an I/O in progress indication
3394 * for those higher layers.
3397 bio_start_transaction(struct bio
*bio
, struct bio_track
*track
)
3399 bio
->bio_track
= track
;
3400 bio_track_ref(track
);
3404 * Initiate I/O on a vnode.
3406 * SWAPCACHE OPERATION:
3408 * Real buffer cache buffers have a non-NULL bp->b_vp. Unfortunately
3409 * devfs also uses b_vp for fake buffers so we also have to check
3410 * that B_PAGING is 0. In this case the passed 'vp' is probably the
3411 * underlying block device. The swap assignments are related to the
3412 * buffer cache buffer's b_vp, not the passed vp.
3414 * The passed vp == bp->b_vp only in the case where the strategy call
3415 * is made on the vp itself for its own buffers (a regular file or
3416 * block device vp). The filesystem usually then re-calls vn_strategy()
3417 * after translating the request to an underlying device.
3419 * Cluster buffers set B_CLUSTER and the passed vp is the vp of the
3420 * underlying buffer cache buffers.
3422 * We can only deal with page-aligned buffers at the moment, because
3423 * we can't tell what the real dirty state for pages straddling a buffer
3426 * In order to call swap_pager_strategy() we must provide the VM object
3427 * and base offset for the underlying buffer cache pages so it can find
3431 vn_strategy(struct vnode
*vp
, struct bio
*bio
)
3433 struct bio_track
*track
;
3434 struct buf
*bp
= bio
->bio_buf
;
3436 KKASSERT(bp
->b_cmd
!= BUF_CMD_DONE
);
3439 * Handle the swap cache intercept.
3441 if (vn_cache_strategy(vp
, bio
))
3445 * Otherwise do the operation through the filesystem
3447 if (bp
->b_cmd
== BUF_CMD_READ
)
3448 track
= &vp
->v_track_read
;
3450 track
= &vp
->v_track_write
;
3451 KKASSERT((bio
->bio_flags
& BIO_DONE
) == 0);
3452 bio
->bio_track
= track
;
3453 bio_track_ref(track
);
3454 vop_strategy(*vp
->v_ops
, vp
, bio
);
3458 vn_cache_strategy(struct vnode
*vp
, struct bio
*bio
)
3460 struct buf
*bp
= bio
->bio_buf
;
3467 * Is this buffer cache buffer suitable for reading from
3470 if (vm_swapcache_read_enable
== 0 ||
3471 bp
->b_cmd
!= BUF_CMD_READ
||
3472 ((bp
->b_flags
& B_CLUSTER
) == 0 &&
3473 (bp
->b_vp
== NULL
|| (bp
->b_flags
& B_PAGING
))) ||
3474 ((int)bp
->b_loffset
& PAGE_MASK
) != 0 ||
3475 (bp
->b_bcount
& PAGE_MASK
) != 0) {
3480 * Figure out the original VM object (it will match the underlying
3481 * VM pages). Note that swap cached data uses page indices relative
3482 * to that object, not relative to bio->bio_offset.
3484 if (bp
->b_flags
& B_CLUSTER
)
3485 object
= vp
->v_object
;
3487 object
= bp
->b_vp
->v_object
;
3490 * In order to be able to use the swap cache all underlying VM
3491 * pages must be marked as such, and we can't have any bogus pages.
3493 for (i
= 0; i
< bp
->b_xio
.xio_npages
; ++i
) {
3494 m
= bp
->b_xio
.xio_pages
[i
];
3495 if ((m
->flags
& PG_SWAPPED
) == 0)
3497 if (m
== bogus_page
)
3502 * If we are good then issue the I/O using swap_pager_strategy()
3504 if (i
== bp
->b_xio
.xio_npages
) {
3505 m
= bp
->b_xio
.xio_pages
[0];
3506 nbio
= push_bio(bio
);
3507 nbio
->bio_offset
= ptoa(m
->pindex
);
3508 KKASSERT(m
->object
== object
);
3509 swap_pager_strategy(object
, nbio
);
3518 * Finish I/O on a buffer after all BIOs have been processed.
3519 * Called when the bio chain is exhausted or by biowait. If called
3520 * by biowait, elseit is typically 0.
3522 * bpdone is also responsible for setting B_CACHE in a B_VMIO bp.
3523 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3524 * assuming B_INVAL is clear.
3526 * For the VMIO case, we set B_CACHE if the op was a read and no
3527 * read error occured, or if the op was a write. B_CACHE is never
3528 * set if the buffer is invalid or otherwise uncacheable.
3530 * bpdone does not mess with B_INVAL, allowing the I/O routine or the
3531 * initiator to leave B_INVAL set to brelse the buffer out of existance
3532 * in the biodone routine.
3535 bpdone(struct buf
*bp
, int elseit
)
3539 KASSERT(BUF_REFCNTNB(bp
) > 0,
3540 ("biodone: bp %p not busy %d", bp
, BUF_REFCNTNB(bp
)));
3541 KASSERT(bp
->b_cmd
!= BUF_CMD_DONE
,
3542 ("biodone: bp %p already done!", bp
));
3545 * No more BIOs are left. All completion functions have been dealt
3546 * with, now we clean up the buffer.
3549 bp
->b_cmd
= BUF_CMD_DONE
;
3552 * Only reads and writes are processed past this point.
3554 if (cmd
!= BUF_CMD_READ
&& cmd
!= BUF_CMD_WRITE
) {
3555 if (cmd
== BUF_CMD_FREEBLKS
)
3556 bp
->b_flags
|= B_NOCACHE
;
3563 * Warning: softupdates may re-dirty the buffer, and HAMMER can do
3564 * a lot worse. XXX - move this above the clearing of b_cmd
3566 if (LIST_FIRST(&bp
->b_dep
) != NULL
)
3570 * A failed write must re-dirty the buffer unless B_INVAL
3571 * was set. Only applicable to normal buffers (with VPs).
3572 * vinum buffers may not have a vp.
3574 if (cmd
== BUF_CMD_WRITE
&&
3575 (bp
->b_flags
& (B_ERROR
| B_INVAL
)) == B_ERROR
) {
3576 bp
->b_flags
&= ~B_NOCACHE
;
3581 if (bp
->b_flags
& B_VMIO
) {
3587 struct vnode
*vp
= bp
->b_vp
;
3591 #if defined(VFS_BIO_DEBUG)
3592 if (vp
->v_auxrefs
== 0)
3593 panic("biodone: zero vnode hold count");
3594 if ((vp
->v_flag
& VOBJBUF
) == 0)
3595 panic("biodone: vnode is not setup for merged cache");
3598 foff
= bp
->b_loffset
;
3599 KASSERT(foff
!= NOOFFSET
, ("biodone: no buffer offset"));
3600 KASSERT(obj
!= NULL
, ("biodone: missing VM object"));
3602 #if defined(VFS_BIO_DEBUG)
3603 if (obj
->paging_in_progress
< bp
->b_xio
.xio_npages
) {
3604 kprintf("biodone: paging in progress(%d) < bp->b_xio.xio_npages(%d)\n",
3605 obj
->paging_in_progress
, bp
->b_xio
.xio_npages
);
3610 * Set B_CACHE if the op was a normal read and no error
3611 * occured. B_CACHE is set for writes in the b*write()
3614 iosize
= bp
->b_bcount
- bp
->b_resid
;
3615 if (cmd
== BUF_CMD_READ
&&
3616 (bp
->b_flags
& (B_INVAL
|B_NOCACHE
|B_ERROR
)) == 0) {
3617 bp
->b_flags
|= B_CACHE
;
3622 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
3626 resid
= ((foff
+ PAGE_SIZE
) & ~(off_t
)PAGE_MASK
) - foff
;
3631 * cleanup bogus pages, restoring the originals. Since
3632 * the originals should still be wired, we don't have
3633 * to worry about interrupt/freeing races destroying
3634 * the VM object association.
3636 m
= bp
->b_xio
.xio_pages
[i
];
3637 if (m
== bogus_page
) {
3639 m
= vm_page_lookup(obj
, OFF_TO_IDX(foff
));
3641 panic("biodone: page disappeared");
3642 bp
->b_xio
.xio_pages
[i
] = m
;
3643 pmap_qenter(trunc_page((vm_offset_t
)bp
->b_data
),
3644 bp
->b_xio
.xio_pages
, bp
->b_xio
.xio_npages
);
3646 #if defined(VFS_BIO_DEBUG)
3647 if (OFF_TO_IDX(foff
) != m
->pindex
) {
3648 kprintf("biodone: foff(%lu)/m->pindex(%ld) "
3650 (unsigned long)foff
, (long)m
->pindex
);
3655 * In the write case, the valid and clean bits are
3656 * already changed correctly (see bdwrite()), so we
3657 * only need to do this here in the read case.
3659 if (cmd
== BUF_CMD_READ
&& !bogusflag
&& resid
> 0) {
3660 vfs_clean_one_page(bp
, i
, m
);
3662 vm_page_flag_clear(m
, PG_ZERO
);
3665 * when debugging new filesystems or buffer I/O
3666 * methods, this is the most common error that pops
3667 * up. if you see this, you have not set the page
3668 * busy flag correctly!!!
3671 kprintf("biodone: page busy < 0, "
3672 "pindex: %d, foff: 0x(%x,%x), "
3673 "resid: %d, index: %d\n",
3674 (int) m
->pindex
, (int)(foff
>> 32),
3675 (int) foff
& 0xffffffff, resid
, i
);
3676 if (!vn_isdisk(vp
, NULL
))
3677 kprintf(" iosize: %ld, loffset: %lld, "
3678 "flags: 0x%08x, npages: %d\n",
3679 bp
->b_vp
->v_mount
->mnt_stat
.f_iosize
,
3680 (long long)bp
->b_loffset
,
3681 bp
->b_flags
, bp
->b_xio
.xio_npages
);
3683 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n",
3684 (long long)bp
->b_loffset
,
3685 bp
->b_flags
, bp
->b_xio
.xio_npages
);
3686 kprintf(" valid: 0x%x, dirty: 0x%x, wired: %d\n",
3687 m
->valid
, m
->dirty
, m
->wire_count
);
3688 panic("biodone: page busy < 0");
3690 vm_page_io_finish(m
);
3691 vm_object_pip_subtract(obj
, 1);
3692 foff
= (foff
+ PAGE_SIZE
) & ~(off_t
)PAGE_MASK
;
3696 vm_object_pip_wakeupn(obj
, 0);
3702 * Finish up by releasing the buffer. There are no more synchronous
3703 * or asynchronous completions, those were handled by bio_done
3707 if (bp
->b_flags
& (B_NOCACHE
|B_INVAL
|B_ERROR
|B_RELBUF
))
3718 biodone(struct bio
*bio
)
3720 struct buf
*bp
= bio
->bio_buf
;
3722 runningbufwakeup(bp
);
3725 * Run up the chain of BIO's. Leave b_cmd intact for the duration.
3728 biodone_t
*done_func
;
3729 struct bio_track
*track
;
3732 * BIO tracking. Most but not all BIOs are tracked.
3734 if ((track
= bio
->bio_track
) != NULL
) {
3735 bio_track_rel(track
);
3736 bio
->bio_track
= NULL
;
3740 * A bio_done function terminates the loop. The function
3741 * will be responsible for any further chaining and/or
3742 * buffer management.
3744 * WARNING! The done function can deallocate the buffer!
3746 if ((done_func
= bio
->bio_done
) != NULL
) {
3747 bio
->bio_done
= NULL
;
3751 bio
= bio
->bio_prev
;
3755 * If we've run out of bio's do normal [a]synchronous completion.
3761 * Synchronous biodone - this terminates a synchronous BIO.
3763 * bpdone() is called with elseit=FALSE, leaving the buffer completed
3764 * but still locked. The caller must brelse() the buffer after waiting
3768 biodone_sync(struct bio
*bio
)
3770 struct buf
*bp
= bio
->bio_buf
;
3774 KKASSERT(bio
== &bp
->b_bio1
);
3778 flags
= bio
->bio_flags
;
3779 nflags
= (flags
| BIO_DONE
) & ~BIO_WANT
;
3781 if (atomic_cmpset_int(&bio
->bio_flags
, flags
, nflags
)) {
3782 if (flags
& BIO_WANT
)
3792 * This routine is called in lieu of iodone in the case of
3793 * incomplete I/O. This keeps the busy status for pages
3797 vfs_unbusy_pages(struct buf
*bp
)
3801 runningbufwakeup(bp
);
3802 if (bp
->b_flags
& B_VMIO
) {
3803 struct vnode
*vp
= bp
->b_vp
;
3808 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
3809 vm_page_t m
= bp
->b_xio
.xio_pages
[i
];
3812 * When restoring bogus changes the original pages
3813 * should still be wired, so we are in no danger of
3814 * losing the object association and do not need
3815 * critical section protection particularly.
3817 if (m
== bogus_page
) {
3818 m
= vm_page_lookup(obj
, OFF_TO_IDX(bp
->b_loffset
) + i
);
3820 panic("vfs_unbusy_pages: page missing");
3822 bp
->b_xio
.xio_pages
[i
] = m
;
3823 pmap_qenter(trunc_page((vm_offset_t
)bp
->b_data
),
3824 bp
->b_xio
.xio_pages
, bp
->b_xio
.xio_npages
);
3826 vm_object_pip_subtract(obj
, 1);
3827 vm_page_flag_clear(m
, PG_ZERO
);
3828 vm_page_io_finish(m
);
3830 vm_object_pip_wakeupn(obj
, 0);
3837 * This routine is called before a device strategy routine.
3838 * It is used to tell the VM system that paging I/O is in
3839 * progress, and treat the pages associated with the buffer
3840 * almost as being PG_BUSY. Also the object 'paging_in_progress'
3841 * flag is handled to make sure that the object doesn't become
3844 * Since I/O has not been initiated yet, certain buffer flags
3845 * such as B_ERROR or B_INVAL may be in an inconsistant state
3846 * and should be ignored.
3849 vfs_busy_pages(struct vnode
*vp
, struct buf
*bp
)
3852 struct lwp
*lp
= curthread
->td_lwp
;
3855 * The buffer's I/O command must already be set. If reading,
3856 * B_CACHE must be 0 (double check against callers only doing
3857 * I/O when B_CACHE is 0).
3859 KKASSERT(bp
->b_cmd
!= BUF_CMD_DONE
);
3860 KKASSERT(bp
->b_cmd
== BUF_CMD_WRITE
|| (bp
->b_flags
& B_CACHE
) == 0);
3862 if (bp
->b_flags
& B_VMIO
) {
3866 KASSERT(bp
->b_loffset
!= NOOFFSET
,
3867 ("vfs_busy_pages: no buffer offset"));
3870 * Loop until none of the pages are busy.
3873 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
3874 vm_page_t m
= bp
->b_xio
.xio_pages
[i
];
3876 if (vm_page_sleep_busy(m
, FALSE
, "vbpage"))
3881 * Setup for I/O, soft-busy the page right now because
3882 * the next loop may block.
3884 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
3885 vm_page_t m
= bp
->b_xio
.xio_pages
[i
];
3887 vm_page_flag_clear(m
, PG_ZERO
);
3888 if ((bp
->b_flags
& B_CLUSTER
) == 0) {
3889 vm_object_pip_add(obj
, 1);
3890 vm_page_io_start(m
);
3895 * Adjust protections for I/O and do bogus-page mapping.
3896 * Assume that vm_page_protect() can block (it can block
3897 * if VM_PROT_NONE, don't take any chances regardless).
3899 * In particular note that for writes we must incorporate
3900 * page dirtyness from the VM system into the buffer's
3903 * For reads we theoretically must incorporate page dirtyness
3904 * from the VM system to determine if the page needs bogus
3905 * replacement, but we shortcut the test by simply checking
3906 * that all m->valid bits are set, indicating that the page
3907 * is fully valid and does not need to be re-read. For any
3908 * VM system dirtyness the page will also be fully valid
3909 * since it was mapped at one point.
3912 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
3913 vm_page_t m
= bp
->b_xio
.xio_pages
[i
];
3915 vm_page_flag_clear(m
, PG_ZERO
); /* XXX */
3916 if (bp
->b_cmd
== BUF_CMD_WRITE
) {
3918 * When readying a vnode-backed buffer for
3919 * a write we must zero-fill any invalid
3920 * portions of the backing VM pages, mark
3921 * it valid and clear related dirty bits.
3923 * vfs_clean_one_page() incorporates any
3924 * VM dirtyness and updates the b_dirtyoff
3925 * range (after we've made the page RO).
3927 * It is also expected that the pmap modified
3928 * bit has already been cleared by the
3929 * vm_page_protect(). We may not be able
3930 * to clear all dirty bits for a page if it
3931 * was also memory mapped (NFS).
3933 * Finally be sure to unassign any swap-cache
3934 * backing store as it is now stale.
3936 vm_page_protect(m
, VM_PROT_READ
);
3937 vfs_clean_one_page(bp
, i
, m
);
3938 swap_pager_unswapped(m
);
3939 } else if (m
->valid
== VM_PAGE_BITS_ALL
) {
3941 * When readying a vnode-backed buffer for
3942 * read we must replace any dirty pages with
3943 * a bogus page so dirty data is not destroyed
3944 * when filling gaps.
3946 * To avoid testing whether the page is
3947 * dirty we instead test that the page was
3948 * at some point mapped (m->valid fully
3949 * valid) with the understanding that
3950 * this also covers the dirty case.
3952 bp
->b_xio
.xio_pages
[i
] = bogus_page
;
3954 } else if (m
->valid
& m
->dirty
) {
3956 * This case should not occur as partial
3957 * dirtyment can only happen if the buffer
3958 * is B_CACHE, and this code is not entered
3959 * if the buffer is B_CACHE.
3961 kprintf("Warning: vfs_busy_pages - page not "
3962 "fully valid! loff=%jx bpf=%08x "
3963 "idx=%d val=%02x dir=%02x\n",
3964 (intmax_t)bp
->b_loffset
, bp
->b_flags
,
3965 i
, m
->valid
, m
->dirty
);
3966 vm_page_protect(m
, VM_PROT_NONE
);
3969 * The page is not valid and can be made
3972 vm_page_protect(m
, VM_PROT_NONE
);
3976 pmap_qenter(trunc_page((vm_offset_t
)bp
->b_data
),
3977 bp
->b_xio
.xio_pages
, bp
->b_xio
.xio_npages
);
3982 * This is the easiest place to put the process accounting for the I/O
3986 if (bp
->b_cmd
== BUF_CMD_READ
)
3987 lp
->lwp_ru
.ru_inblock
++;
3989 lp
->lwp_ru
.ru_oublock
++;
3996 * Tell the VM system that the pages associated with this buffer
3997 * are clean. This is used for delayed writes where the data is
3998 * going to go to disk eventually without additional VM intevention.
4000 * Note that while we only really need to clean through to b_bcount, we
4001 * just go ahead and clean through to b_bufsize.
4004 vfs_clean_pages(struct buf
*bp
)
4009 if ((bp
->b_flags
& B_VMIO
) == 0)
4012 KASSERT(bp
->b_loffset
!= NOOFFSET
,
4013 ("vfs_clean_pages: no buffer offset"));
4015 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
4016 m
= bp
->b_xio
.xio_pages
[i
];
4017 vfs_clean_one_page(bp
, i
, m
);
4022 * vfs_clean_one_page:
4024 * Set the valid bits and clear the dirty bits in a page within a
4025 * buffer. The range is restricted to the buffer's size and the
4026 * buffer's logical offset might index into the first page.
4028 * The caller has busied or soft-busied the page and it is not mapped,
4029 * test and incorporate the dirty bits into b_dirtyoff/end before
4030 * clearing them. Note that we need to clear the pmap modified bits
4031 * after determining the the page was dirty, vm_page_set_validclean()
4032 * does not do it for us.
4034 * This routine is typically called after a read completes (dirty should
4035 * be zero in that case as we are not called on bogus-replace pages),
4036 * or before a write is initiated.
4039 vfs_clean_one_page(struct buf
*bp
, int pageno
, vm_page_t m
)
4047 * Calculate offset range within the page but relative to buffer's
4048 * loffset. loffset might be offset into the first page.
4050 xoff
= (int)bp
->b_loffset
& PAGE_MASK
; /* loffset offset into pg 0 */
4051 bcount
= bp
->b_bcount
+ xoff
; /* offset adjusted */
4057 soff
= (pageno
<< PAGE_SHIFT
);
4058 eoff
= soff
+ PAGE_SIZE
;
4066 * Test dirty bits and adjust b_dirtyoff/end.
4068 * If dirty pages are incorporated into the bp any prior
4069 * B_NEEDCOMMIT state (NFS) must be cleared because the
4070 * caller has not taken into account the new dirty data.
4072 * If the page was memory mapped the dirty bits might go beyond the
4073 * end of the buffer, but we can't really make the assumption that
4074 * a file EOF straddles the buffer (even though this is the case for
4075 * NFS if B_NEEDCOMMIT is also set). So for the purposes of clearing
4076 * B_NEEDCOMMIT we only test the dirty bits covered by the buffer.
4077 * This also saves some console spam.
4079 * When clearing B_NEEDCOMMIT we must also clear B_CLUSTEROK,
4080 * NFS can handle huge commits but not huge writes.
4082 vm_page_test_dirty(m
);
4084 if ((bp
->b_flags
& B_NEEDCOMMIT
) &&
4085 (m
->dirty
& vm_page_bits(soff
& PAGE_MASK
, eoff
- soff
))) {
4087 kprintf("Warning: vfs_clean_one_page: bp %p "
4088 "loff=%jx,%d flgs=%08x clr B_NEEDCOMMIT"
4089 " cmd %d vd %02x/%02x x/s/e %d %d %d "
4091 bp
, (intmax_t)bp
->b_loffset
, bp
->b_bcount
,
4092 bp
->b_flags
, bp
->b_cmd
,
4093 m
->valid
, m
->dirty
, xoff
, soff
, eoff
,
4094 bp
->b_dirtyoff
, bp
->b_dirtyend
);
4095 bp
->b_flags
&= ~(B_NEEDCOMMIT
| B_CLUSTEROK
);
4100 * Only clear the pmap modified bits if ALL the dirty bits
4101 * are set, otherwise the system might mis-clear portions
4104 if (m
->dirty
== VM_PAGE_BITS_ALL
&&
4105 (bp
->b_flags
& B_NEEDCOMMIT
) == 0) {
4106 pmap_clear_modify(m
);
4108 if (bp
->b_dirtyoff
> soff
- xoff
)
4109 bp
->b_dirtyoff
= soff
- xoff
;
4110 if (bp
->b_dirtyend
< eoff
- xoff
)
4111 bp
->b_dirtyend
= eoff
- xoff
;
4115 * Set related valid bits, clear related dirty bits.
4116 * Does not mess with the pmap modified bit.
4118 * WARNING! We cannot just clear all of m->dirty here as the
4119 * buffer cache buffers may use a DEV_BSIZE'd aligned
4120 * block size, or have an odd size (e.g. NFS at file EOF).
4121 * The putpages code can clear m->dirty to 0.
4123 * If a VOP_WRITE generates a buffer cache buffer which
4124 * covers the same space as mapped writable pages the
4125 * buffer flush might not be able to clear all the dirty
4126 * bits and still require a putpages from the VM system
4129 vm_page_set_validclean(m
, soff
& PAGE_MASK
, eoff
- soff
);
4133 * Similar to vfs_clean_one_page() but sets the bits to valid and dirty.
4134 * The page data is assumed to be valid (there is no zeroing here).
4137 vfs_dirty_one_page(struct buf
*bp
, int pageno
, vm_page_t m
)
4145 * Calculate offset range within the page but relative to buffer's
4146 * loffset. loffset might be offset into the first page.
4148 xoff
= (int)bp
->b_loffset
& PAGE_MASK
; /* loffset offset into pg 0 */
4149 bcount
= bp
->b_bcount
+ xoff
; /* offset adjusted */
4155 soff
= (pageno
<< PAGE_SHIFT
);
4156 eoff
= soff
+ PAGE_SIZE
;
4162 vm_page_set_validdirty(m
, soff
& PAGE_MASK
, eoff
- soff
);
4168 * Clear a buffer. This routine essentially fakes an I/O, so we need
4169 * to clear B_ERROR and B_INVAL.
4171 * Note that while we only theoretically need to clear through b_bcount,
4172 * we go ahead and clear through b_bufsize.
4176 vfs_bio_clrbuf(struct buf
*bp
)
4180 if ((bp
->b_flags
& (B_VMIO
| B_MALLOC
)) == B_VMIO
) {
4181 bp
->b_flags
&= ~(B_INVAL
| B_EINTR
| B_ERROR
);
4182 if ((bp
->b_xio
.xio_npages
== 1) && (bp
->b_bufsize
< PAGE_SIZE
) &&
4183 (bp
->b_loffset
& PAGE_MASK
) == 0) {
4184 mask
= (1 << (bp
->b_bufsize
/ DEV_BSIZE
)) - 1;
4185 if ((bp
->b_xio
.xio_pages
[0]->valid
& mask
) == mask
) {
4189 if (((bp
->b_xio
.xio_pages
[0]->flags
& PG_ZERO
) == 0) &&
4190 ((bp
->b_xio
.xio_pages
[0]->valid
& mask
) == 0)) {
4191 bzero(bp
->b_data
, bp
->b_bufsize
);
4192 bp
->b_xio
.xio_pages
[0]->valid
|= mask
;
4198 for(i
=0;i
<bp
->b_xio
.xio_npages
;i
++,sa
=ea
) {
4199 int j
= ((vm_offset_t
)sa
& PAGE_MASK
) / DEV_BSIZE
;
4200 ea
= (caddr_t
)trunc_page((vm_offset_t
)sa
+ PAGE_SIZE
);
4201 ea
= (caddr_t
)(vm_offset_t
)ulmin(
4202 (u_long
)(vm_offset_t
)ea
,
4203 (u_long
)(vm_offset_t
)bp
->b_data
+ bp
->b_bufsize
);
4204 mask
= ((1 << ((ea
- sa
) / DEV_BSIZE
)) - 1) << j
;
4205 if ((bp
->b_xio
.xio_pages
[i
]->valid
& mask
) == mask
)
4207 if ((bp
->b_xio
.xio_pages
[i
]->valid
& mask
) == 0) {
4208 if ((bp
->b_xio
.xio_pages
[i
]->flags
& PG_ZERO
) == 0) {
4212 for (; sa
< ea
; sa
+= DEV_BSIZE
, j
++) {
4213 if (((bp
->b_xio
.xio_pages
[i
]->flags
& PG_ZERO
) == 0) &&
4214 (bp
->b_xio
.xio_pages
[i
]->valid
& (1<<j
)) == 0)
4215 bzero(sa
, DEV_BSIZE
);
4218 bp
->b_xio
.xio_pages
[i
]->valid
|= mask
;
4219 vm_page_flag_clear(bp
->b_xio
.xio_pages
[i
], PG_ZERO
);
4228 * vm_hold_load_pages:
4230 * Load pages into the buffer's address space. The pages are
4231 * allocated from the kernel object in order to reduce interference
4232 * with the any VM paging I/O activity. The range of loaded
4233 * pages will be wired.
4235 * If a page cannot be allocated, the 'pagedaemon' is woken up to
4236 * retrieve the full range (to - from) of pages.
4240 vm_hold_load_pages(struct buf
*bp
, vm_offset_t from
, vm_offset_t to
)
4246 to
= round_page(to
);
4247 from
= round_page(from
);
4248 index
= (from
- trunc_page((vm_offset_t
)bp
->b_data
)) >> PAGE_SHIFT
;
4253 * Note: must allocate system pages since blocking here
4254 * could intefere with paging I/O, no matter which
4257 p
= bio_page_alloc(&kernel_object
, pg
>> PAGE_SHIFT
,
4258 (vm_pindex_t
)((to
- pg
) >> PAGE_SHIFT
));
4261 p
->valid
= VM_PAGE_BITS_ALL
;
4262 vm_page_flag_clear(p
, PG_ZERO
);
4263 pmap_kenter(pg
, VM_PAGE_TO_PHYS(p
));
4264 bp
->b_xio
.xio_pages
[index
] = p
;
4271 bp
->b_xio
.xio_npages
= index
;
4275 * Allocate pages for a buffer cache buffer.
4277 * Under extremely severe memory conditions even allocating out of the
4278 * system reserve can fail. If this occurs we must allocate out of the
4279 * interrupt reserve to avoid a deadlock with the pageout daemon.
4281 * The pageout daemon can run (putpages -> VOP_WRITE -> getblk -> allocbuf).
4282 * If the buffer cache's vm_page_alloc() fails a vm_wait() can deadlock
4283 * against the pageout daemon if pages are not freed from other sources.
4287 bio_page_alloc(vm_object_t obj
, vm_pindex_t pg
, int deficit
)
4292 * Try a normal allocation, allow use of system reserve.
4294 p
= vm_page_alloc(obj
, pg
, VM_ALLOC_NORMAL
| VM_ALLOC_SYSTEM
);
4299 * The normal allocation failed and we clearly have a page
4300 * deficit. Try to reclaim some clean VM pages directly
4301 * from the buffer cache.
4303 vm_pageout_deficit
+= deficit
;
4307 * We may have blocked, the caller will know what to do if the
4310 if (vm_page_lookup(obj
, pg
))
4314 * Allocate and allow use of the interrupt reserve.
4316 * If after all that we still can't allocate a VM page we are
4317 * in real trouble, but we slog on anyway hoping that the system
4320 p
= vm_page_alloc(obj
, pg
, VM_ALLOC_NORMAL
| VM_ALLOC_SYSTEM
|
4321 VM_ALLOC_INTERRUPT
);
4323 if (vm_page_count_severe()) {
4324 kprintf("bio_page_alloc: WARNING emergency page "
4329 kprintf("bio_page_alloc: WARNING emergency page "
4330 "allocation failed\n");
4337 * vm_hold_free_pages:
4339 * Return pages associated with the buffer back to the VM system.
4341 * The range of pages underlying the buffer's address space will
4342 * be unmapped and un-wired.
4345 vm_hold_free_pages(struct buf
*bp
, vm_offset_t from
, vm_offset_t to
)
4349 int index
, newnpages
;
4351 from
= round_page(from
);
4352 to
= round_page(to
);
4353 newnpages
= index
= (from
- trunc_page((vm_offset_t
)bp
->b_data
)) >> PAGE_SHIFT
;
4355 for (pg
= from
; pg
< to
; pg
+= PAGE_SIZE
, index
++) {
4356 p
= bp
->b_xio
.xio_pages
[index
];
4357 if (p
&& (index
< bp
->b_xio
.xio_npages
)) {
4359 kprintf("vm_hold_free_pages: doffset: %lld, "
4361 (long long)bp
->b_bio2
.bio_offset
,
4362 (long long)bp
->b_loffset
);
4364 bp
->b_xio
.xio_pages
[index
] = NULL
;
4367 vm_page_unwire(p
, 0);
4371 bp
->b_xio
.xio_npages
= newnpages
;
4377 * Map a user buffer into KVM via a pbuf. On return the buffer's
4378 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array
4382 vmapbuf(struct buf
*bp
, caddr_t udata
, int bytes
)
4393 * bp had better have a command and it better be a pbuf.
4395 KKASSERT(bp
->b_cmd
!= BUF_CMD_DONE
);
4396 KKASSERT(bp
->b_flags
& B_PAGING
);
4402 * Map the user data into KVM. Mappings have to be page-aligned.
4404 addr
= (caddr_t
)trunc_page((vm_offset_t
)udata
);
4407 vmprot
= VM_PROT_READ
;
4408 if (bp
->b_cmd
== BUF_CMD_READ
)
4409 vmprot
|= VM_PROT_WRITE
;
4411 while (addr
< udata
+ bytes
) {
4413 * Do the vm_fault if needed; do the copy-on-write thing
4414 * when reading stuff off device into memory.
4416 * vm_fault_page*() returns a held VM page.
4418 va
= (addr
>= udata
) ? (vm_offset_t
)addr
: (vm_offset_t
)udata
;
4419 va
= trunc_page(va
);
4421 m
= vm_fault_page_quick(va
, vmprot
, &error
);
4423 for (i
= 0; i
< pidx
; ++i
) {
4424 vm_page_unhold(bp
->b_xio
.xio_pages
[i
]);
4425 bp
->b_xio
.xio_pages
[i
] = NULL
;
4429 bp
->b_xio
.xio_pages
[pidx
] = m
;
4435 * Map the page array and set the buffer fields to point to
4436 * the mapped data buffer.
4438 if (pidx
> btoc(MAXPHYS
))
4439 panic("vmapbuf: mapped more than MAXPHYS");
4440 pmap_qenter((vm_offset_t
)bp
->b_kvabase
, bp
->b_xio
.xio_pages
, pidx
);
4442 bp
->b_xio
.xio_npages
= pidx
;
4443 bp
->b_data
= bp
->b_kvabase
+ ((int)(intptr_t)udata
& PAGE_MASK
);
4444 bp
->b_bcount
= bytes
;
4445 bp
->b_bufsize
= bytes
;
4452 * Free the io map PTEs associated with this IO operation.
4453 * We also invalidate the TLB entries and restore the original b_addr.
4456 vunmapbuf(struct buf
*bp
)
4461 KKASSERT(bp
->b_flags
& B_PAGING
);
4463 npages
= bp
->b_xio
.xio_npages
;
4464 pmap_qremove(trunc_page((vm_offset_t
)bp
->b_data
), npages
);
4465 for (pidx
= 0; pidx
< npages
; ++pidx
) {
4466 vm_page_unhold(bp
->b_xio
.xio_pages
[pidx
]);
4467 bp
->b_xio
.xio_pages
[pidx
] = NULL
;
4469 bp
->b_xio
.xio_npages
= 0;
4470 bp
->b_data
= bp
->b_kvabase
;
4474 * Scan all buffers in the system and issue the callback.
4477 scan_all_buffers(int (*callback
)(struct buf
*, void *), void *info
)
4483 for (n
= 0; n
< nbuf
; ++n
) {
4484 if ((error
= callback(&buf
[n
], info
)) < 0) {
4494 * print out statistics from the current status of the buffer pool
4495 * this can be toggeled by the system control option debug.syncprt
4504 int counts
[(MAXBSIZE
/ PAGE_SIZE
) + 1];
4505 static char *bname
[3] = { "LOCKED", "LRU", "AGE" };
4507 for (dp
= bufqueues
, i
= 0; dp
< &bufqueues
[3]; dp
++, i
++) {
4509 for (j
= 0; j
<= MAXBSIZE
/PAGE_SIZE
; j
++)
4512 TAILQ_FOREACH(bp
, dp
, b_freelist
) {
4513 counts
[bp
->b_bufsize
/PAGE_SIZE
]++;
4517 kprintf("%s: total-%d", bname
[i
], count
);
4518 for (j
= 0; j
<= MAXBSIZE
/PAGE_SIZE
; j
++)
4520 kprintf(", %d-%d", j
* PAGE_SIZE
, counts
[j
]);
4528 DB_SHOW_COMMAND(buffer
, db_show_buffer
)
4531 struct buf
*bp
= (struct buf
*)addr
;
4534 db_printf("usage: show buffer <addr>\n");
4538 db_printf("b_flags = 0x%b\n", (u_int
)bp
->b_flags
, PRINT_BUF_FLAGS
);
4539 db_printf("b_cmd = %d\n", bp
->b_cmd
);
4540 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, "
4541 "b_resid = %d\n, b_data = %p, "
4542 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n",
4543 bp
->b_error
, bp
->b_bufsize
, bp
->b_bcount
, bp
->b_resid
,
4545 (long long)bp
->b_bio2
.bio_offset
,
4546 (long long)(bp
->b_bio2
.bio_next
?
4547 bp
->b_bio2
.bio_next
->bio_offset
: (off_t
)-1));
4548 if (bp
->b_xio
.xio_npages
) {
4550 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ",
4551 bp
->b_xio
.xio_npages
);
4552 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
4554 m
= bp
->b_xio
.xio_pages
[i
];
4555 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m
->object
,
4556 (u_long
)m
->pindex
, (u_long
)VM_PAGE_TO_PHYS(m
));
4557 if ((i
+ 1) < bp
->b_xio
.xio_npages
)