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 1.99 2008/04/22 18:46:51 dillon 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>
58 #include <sys/thread2.h>
59 #include <sys/spinlock2.h>
60 #include <vm/vm_page2.h>
71 BQUEUE_NONE
, /* not on any queue */
72 BQUEUE_LOCKED
, /* locked buffers */
73 BQUEUE_CLEAN
, /* non-B_DELWRI buffers */
74 BQUEUE_DIRTY
, /* B_DELWRI buffers */
75 BQUEUE_DIRTY_HW
, /* B_DELWRI buffers - heavy weight */
76 BQUEUE_EMPTYKVA
, /* empty buffer headers with KVA assignment */
77 BQUEUE_EMPTY
, /* empty buffer headers */
79 BUFFER_QUEUES
/* number of buffer queues */
82 typedef enum bufq_type bufq_type_t
;
84 TAILQ_HEAD(bqueues
, buf
) bufqueues
[BUFFER_QUEUES
];
86 static MALLOC_DEFINE(M_BIOBUF
, "BIO buffer", "BIO buffer");
88 struct buf
*buf
; /* buffer header pool */
90 static void vm_hold_free_pages(struct buf
*bp
, vm_offset_t from
,
92 static void vm_hold_load_pages(struct buf
*bp
, vm_offset_t from
,
94 static void vfs_page_set_valid(struct buf
*bp
, vm_ooffset_t off
,
95 int pageno
, vm_page_t m
);
96 static void vfs_clean_pages(struct buf
*bp
);
97 static void vfs_setdirty(struct buf
*bp
);
98 static void vfs_vmio_release(struct buf
*bp
);
99 static int flushbufqueues(bufq_type_t q
);
101 static void buf_daemon(void);
102 static void buf_daemon_hw(void);
104 * bogus page -- for I/O to/from partially complete buffers
105 * this is a temporary solution to the problem, but it is not
106 * really that bad. it would be better to split the buffer
107 * for input in the case of buffers partially already in memory,
108 * but the code is intricate enough already.
110 vm_page_t bogus_page
;
114 * These are all static, but make the ones we export globals so we do
115 * not need to use compiler magic.
117 int bufspace
, maxbufspace
,
118 bufmallocspace
, maxbufmallocspace
, lobufspace
, hibufspace
;
119 static int bufreusecnt
, bufdefragcnt
, buffreekvacnt
;
120 static int lorunningspace
, hirunningspace
, runningbufreq
;
121 int numdirtybuffers
, numdirtybuffershw
, lodirtybuffers
, hidirtybuffers
;
122 static int numfreebuffers
, lofreebuffers
, hifreebuffers
;
123 static int getnewbufcalls
;
124 static int getnewbufrestarts
;
126 static int needsbuffer
; /* locked by needsbuffer_spin */
127 static int bd_request
; /* locked by needsbuffer_spin */
128 static int bd_request_hw
; /* locked by needsbuffer_spin */
129 static struct spinlock needsbuffer_spin
;
132 * Sysctls for operational control of the buffer cache.
134 SYSCTL_INT(_vfs
, OID_AUTO
, lodirtybuffers
, CTLFLAG_RW
, &lodirtybuffers
, 0,
135 "Number of dirty buffers to flush before bufdaemon becomes inactive");
136 SYSCTL_INT(_vfs
, OID_AUTO
, hidirtybuffers
, CTLFLAG_RW
, &hidirtybuffers
, 0,
137 "High watermark used to trigger explicit flushing of dirty buffers");
138 SYSCTL_INT(_vfs
, OID_AUTO
, lofreebuffers
, CTLFLAG_RW
, &lofreebuffers
, 0,
139 "Low watermark for special reserve in low-memory situations");
140 SYSCTL_INT(_vfs
, OID_AUTO
, hifreebuffers
, CTLFLAG_RW
, &hifreebuffers
, 0,
141 "High watermark for special reserve in low-memory situations");
142 SYSCTL_INT(_vfs
, OID_AUTO
, lorunningspace
, CTLFLAG_RW
, &lorunningspace
, 0,
143 "Minimum amount of buffer space required for active I/O");
144 SYSCTL_INT(_vfs
, OID_AUTO
, hirunningspace
, CTLFLAG_RW
, &hirunningspace
, 0,
145 "Maximum amount of buffer space to usable for active I/O");
147 * Sysctls determining current state of the buffer cache.
149 SYSCTL_INT(_vfs
, OID_AUTO
, numdirtybuffers
, CTLFLAG_RD
, &numdirtybuffers
, 0,
150 "Pending number of dirty buffers (all)");
151 SYSCTL_INT(_vfs
, OID_AUTO
, numdirtybuffershw
, CTLFLAG_RD
, &numdirtybuffershw
, 0,
152 "Pending number of dirty buffers (heavy weight)");
153 SYSCTL_INT(_vfs
, OID_AUTO
, numfreebuffers
, CTLFLAG_RD
, &numfreebuffers
, 0,
154 "Number of free buffers on the buffer cache free list");
155 SYSCTL_INT(_vfs
, OID_AUTO
, runningbufspace
, CTLFLAG_RD
, &runningbufspace
, 0,
156 "I/O bytes currently in progress due to asynchronous writes");
157 SYSCTL_INT(_vfs
, OID_AUTO
, maxbufspace
, CTLFLAG_RD
, &maxbufspace
, 0,
158 "Hard limit on maximum amount of memory usable for buffer space");
159 SYSCTL_INT(_vfs
, OID_AUTO
, hibufspace
, CTLFLAG_RD
, &hibufspace
, 0,
160 "Soft limit on maximum amount of memory usable for buffer space");
161 SYSCTL_INT(_vfs
, OID_AUTO
, lobufspace
, CTLFLAG_RD
, &lobufspace
, 0,
162 "Minimum amount of memory to reserve for system buffer space");
163 SYSCTL_INT(_vfs
, OID_AUTO
, bufspace
, CTLFLAG_RD
, &bufspace
, 0,
164 "Amount of memory available for buffers");
165 SYSCTL_INT(_vfs
, OID_AUTO
, maxmallocbufspace
, CTLFLAG_RD
, &maxbufmallocspace
,
166 0, "Maximum amount of memory reserved for buffers using malloc");
167 SYSCTL_INT(_vfs
, OID_AUTO
, bufmallocspace
, CTLFLAG_RD
, &bufmallocspace
, 0,
168 "Amount of memory left for buffers using malloc-scheme");
169 SYSCTL_INT(_vfs
, OID_AUTO
, getnewbufcalls
, CTLFLAG_RD
, &getnewbufcalls
, 0,
170 "New buffer header acquisition requests");
171 SYSCTL_INT(_vfs
, OID_AUTO
, getnewbufrestarts
, CTLFLAG_RD
, &getnewbufrestarts
,
172 0, "New buffer header acquisition restarts");
173 SYSCTL_INT(_vfs
, OID_AUTO
, bufdefragcnt
, CTLFLAG_RD
, &bufdefragcnt
, 0,
174 "Buffer acquisition restarts due to fragmented buffer map");
175 SYSCTL_INT(_vfs
, OID_AUTO
, buffreekvacnt
, CTLFLAG_RD
, &buffreekvacnt
, 0,
176 "Amount of time KVA space was deallocated in an arbitrary buffer");
177 SYSCTL_INT(_vfs
, OID_AUTO
, bufreusecnt
, CTLFLAG_RD
, &bufreusecnt
, 0,
178 "Amount of time buffer re-use operations were successful");
179 SYSCTL_INT(_debug_sizeof
, OID_AUTO
, buf
, CTLFLAG_RD
, 0, sizeof(struct buf
),
180 "sizeof(struct buf)");
182 char *buf_wmesg
= BUF_WMESG
;
184 extern int vm_swap_size
;
186 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
187 #define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */
188 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */
189 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
194 * If someone is blocked due to there being too many dirty buffers,
195 * and numdirtybuffers is now reasonable, wake them up.
200 if (numdirtybuffers
<= (lodirtybuffers
+ hidirtybuffers
) / 2) {
201 if (needsbuffer
& VFS_BIO_NEED_DIRTYFLUSH
) {
202 spin_lock_wr(&needsbuffer_spin
);
203 needsbuffer
&= ~VFS_BIO_NEED_DIRTYFLUSH
;
204 spin_unlock_wr(&needsbuffer_spin
);
205 wakeup(&needsbuffer
);
213 * Called when buffer space is potentially available for recovery.
214 * getnewbuf() will block on this flag when it is unable to free
215 * sufficient buffer space. Buffer space becomes recoverable when
216 * bp's get placed back in the queues.
223 * If someone is waiting for BUF space, wake them up. Even
224 * though we haven't freed the kva space yet, the waiting
225 * process will be able to now.
227 if (needsbuffer
& VFS_BIO_NEED_BUFSPACE
) {
228 spin_lock_wr(&needsbuffer_spin
);
229 needsbuffer
&= ~VFS_BIO_NEED_BUFSPACE
;
230 spin_unlock_wr(&needsbuffer_spin
);
231 wakeup(&needsbuffer
);
238 * Accounting for I/O in progress.
242 runningbufwakeup(struct buf
*bp
)
244 if (bp
->b_runningbufspace
) {
245 runningbufspace
-= bp
->b_runningbufspace
;
246 bp
->b_runningbufspace
= 0;
247 if (runningbufreq
&& runningbufspace
<= lorunningspace
) {
249 wakeup(&runningbufreq
);
257 * Called when a buffer has been added to one of the free queues to
258 * account for the buffer and to wakeup anyone waiting for free buffers.
259 * This typically occurs when large amounts of metadata are being handled
260 * by the buffer cache ( else buffer space runs out first, usually ).
268 spin_lock_wr(&needsbuffer_spin
);
269 needsbuffer
&= ~VFS_BIO_NEED_ANY
;
270 if (numfreebuffers
>= hifreebuffers
)
271 needsbuffer
&= ~VFS_BIO_NEED_FREE
;
272 spin_unlock_wr(&needsbuffer_spin
);
273 wakeup(&needsbuffer
);
278 * waitrunningbufspace()
280 * runningbufspace is a measure of the amount of I/O currently
281 * running. This routine is used in async-write situations to
282 * prevent creating huge backups of pending writes to a device.
283 * Only asynchronous writes are governed by this function.
285 * Reads will adjust runningbufspace, but will not block based on it.
286 * The read load has a side effect of reducing the allowed write load.
288 * This does NOT turn an async write into a sync write. It waits
289 * for earlier writes to complete and generally returns before the
290 * caller's write has reached the device.
293 waitrunningbufspace(void)
295 if (runningbufspace
> hirunningspace
) {
297 while (runningbufspace
> hirunningspace
) {
299 tsleep(&runningbufreq
, 0, "wdrain", 0);
306 * vfs_buf_test_cache:
308 * Called when a buffer is extended. This function clears the B_CACHE
309 * bit if the newly extended portion of the buffer does not contain
314 vfs_buf_test_cache(struct buf
*bp
,
315 vm_ooffset_t foff
, vm_offset_t off
, vm_offset_t size
,
318 if (bp
->b_flags
& B_CACHE
) {
319 int base
= (foff
+ off
) & PAGE_MASK
;
320 if (vm_page_is_valid(m
, base
, size
) == 0)
321 bp
->b_flags
&= ~B_CACHE
;
328 * Wake up the buffer daemon if the number of outstanding dirty buffers
329 * is above specified threshold 'dirtybuflevel'.
331 * The buffer daemons are explicitly woken up when (a) the pending number
332 * of dirty buffers exceeds the recovery and stall mid-point value,
333 * (b) during bwillwrite() or (c) buf freelist was exhausted.
335 * The buffer daemons will generally not stop flushing until the dirty
336 * buffer count goes below lodirtybuffers.
340 bd_wakeup(int dirtybuflevel
)
342 if (bd_request
== 0 && numdirtybuffers
>= dirtybuflevel
) {
343 spin_lock_wr(&needsbuffer_spin
);
345 spin_unlock_wr(&needsbuffer_spin
);
348 if (bd_request_hw
== 0 && numdirtybuffershw
>= dirtybuflevel
) {
349 spin_lock_wr(&needsbuffer_spin
);
351 spin_unlock_wr(&needsbuffer_spin
);
352 wakeup(&bd_request_hw
);
359 * Speed up the buffer cache flushing process.
372 * Load time initialisation of the buffer cache, called from machine
373 * dependant initialization code.
379 vm_offset_t bogus_offset
;
382 spin_init(&needsbuffer_spin
);
384 /* next, make a null set of free lists */
385 for (i
= 0; i
< BUFFER_QUEUES
; i
++)
386 TAILQ_INIT(&bufqueues
[i
]);
388 /* finally, initialize each buffer header and stick on empty q */
389 for (i
= 0; i
< nbuf
; i
++) {
391 bzero(bp
, sizeof *bp
);
392 bp
->b_flags
= B_INVAL
; /* we're just an empty header */
393 bp
->b_cmd
= BUF_CMD_DONE
;
394 bp
->b_qindex
= BQUEUE_EMPTY
;
396 xio_init(&bp
->b_xio
);
399 TAILQ_INSERT_TAIL(&bufqueues
[BQUEUE_EMPTY
], bp
, b_freelist
);
403 * maxbufspace is the absolute maximum amount of buffer space we are
404 * allowed to reserve in KVM and in real terms. The absolute maximum
405 * is nominally used by buf_daemon. hibufspace is the nominal maximum
406 * used by most other processes. The differential is required to
407 * ensure that buf_daemon is able to run when other processes might
408 * be blocked waiting for buffer space.
410 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
411 * this may result in KVM fragmentation which is not handled optimally
414 maxbufspace
= nbuf
* BKVASIZE
;
415 hibufspace
= imax(3 * maxbufspace
/ 4, maxbufspace
- MAXBSIZE
* 10);
416 lobufspace
= hibufspace
- MAXBSIZE
;
418 lorunningspace
= 512 * 1024;
419 hirunningspace
= 1024 * 1024;
422 * Limit the amount of malloc memory since it is wired permanently into
423 * the kernel space. Even though this is accounted for in the buffer
424 * allocation, we don't want the malloced region to grow uncontrolled.
425 * The malloc scheme improves memory utilization significantly on average
426 * (small) directories.
428 maxbufmallocspace
= hibufspace
/ 20;
431 * Reduce the chance of a deadlock occuring by limiting the number
432 * of delayed-write dirty buffers we allow to stack up.
434 hidirtybuffers
= nbuf
/ 4 + 20;
436 numdirtybuffershw
= 0;
438 * To support extreme low-memory systems, make sure hidirtybuffers cannot
439 * eat up all available buffer space. This occurs when our minimum cannot
440 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming
441 * BKVASIZE'd (8K) buffers.
443 while (hidirtybuffers
* BKVASIZE
> 3 * hibufspace
/ 4) {
444 hidirtybuffers
>>= 1;
446 lodirtybuffers
= hidirtybuffers
/ 2;
449 * Try to keep the number of free buffers in the specified range,
450 * and give special processes (e.g. like buf_daemon) access to an
453 lofreebuffers
= nbuf
/ 18 + 5;
454 hifreebuffers
= 2 * lofreebuffers
;
455 numfreebuffers
= nbuf
;
458 * Maximum number of async ops initiated per buf_daemon loop. This is
459 * somewhat of a hack at the moment, we really need to limit ourselves
460 * based on the number of bytes of I/O in-transit that were initiated
464 bogus_offset
= kmem_alloc_pageable(&kernel_map
, PAGE_SIZE
);
465 bogus_page
= vm_page_alloc(&kernel_object
,
466 (bogus_offset
>> PAGE_SHIFT
),
468 vmstats
.v_wire_count
++;
473 * Initialize the embedded bio structures
476 initbufbio(struct buf
*bp
)
478 bp
->b_bio1
.bio_buf
= bp
;
479 bp
->b_bio1
.bio_prev
= NULL
;
480 bp
->b_bio1
.bio_offset
= NOOFFSET
;
481 bp
->b_bio1
.bio_next
= &bp
->b_bio2
;
482 bp
->b_bio1
.bio_done
= NULL
;
484 bp
->b_bio2
.bio_buf
= bp
;
485 bp
->b_bio2
.bio_prev
= &bp
->b_bio1
;
486 bp
->b_bio2
.bio_offset
= NOOFFSET
;
487 bp
->b_bio2
.bio_next
= NULL
;
488 bp
->b_bio2
.bio_done
= NULL
;
492 * Reinitialize the embedded bio structures as well as any additional
493 * translation cache layers.
496 reinitbufbio(struct buf
*bp
)
500 for (bio
= &bp
->b_bio1
; bio
; bio
= bio
->bio_next
) {
501 bio
->bio_done
= NULL
;
502 bio
->bio_offset
= NOOFFSET
;
507 * Push another BIO layer onto an existing BIO and return it. The new
508 * BIO layer may already exist, holding cached translation data.
511 push_bio(struct bio
*bio
)
515 if ((nbio
= bio
->bio_next
) == NULL
) {
516 int index
= bio
- &bio
->bio_buf
->b_bio_array
[0];
517 if (index
>= NBUF_BIO
- 1) {
518 panic("push_bio: too many layers bp %p\n",
521 nbio
= &bio
->bio_buf
->b_bio_array
[index
+ 1];
522 bio
->bio_next
= nbio
;
523 nbio
->bio_prev
= bio
;
524 nbio
->bio_buf
= bio
->bio_buf
;
525 nbio
->bio_offset
= NOOFFSET
;
526 nbio
->bio_done
= NULL
;
527 nbio
->bio_next
= NULL
;
529 KKASSERT(nbio
->bio_done
== NULL
);
534 pop_bio(struct bio
*bio
)
540 clearbiocache(struct bio
*bio
)
543 bio
->bio_offset
= NOOFFSET
;
551 * Free the KVA allocation for buffer 'bp'.
553 * Must be called from a critical section as this is the only locking for
556 * Since this call frees up buffer space, we call bufspacewakeup().
559 bfreekva(struct buf
*bp
)
565 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
566 vm_map_lock(&buffer_map
);
567 bufspace
-= bp
->b_kvasize
;
568 vm_map_delete(&buffer_map
,
569 (vm_offset_t
) bp
->b_kvabase
,
570 (vm_offset_t
) bp
->b_kvabase
+ bp
->b_kvasize
,
573 vm_map_unlock(&buffer_map
);
574 vm_map_entry_release(count
);
583 * Remove the buffer from the appropriate free list.
586 bremfree(struct buf
*bp
)
591 old_qindex
= bp
->b_qindex
;
593 if (bp
->b_qindex
!= BQUEUE_NONE
) {
594 KASSERT(BUF_REFCNTNB(bp
) == 1,
595 ("bremfree: bp %p not locked",bp
));
596 TAILQ_REMOVE(&bufqueues
[bp
->b_qindex
], bp
, b_freelist
);
597 bp
->b_qindex
= BQUEUE_NONE
;
599 if (BUF_REFCNTNB(bp
) <= 1)
600 panic("bremfree: removing a buffer not on a queue");
604 * Fixup numfreebuffers count. If the buffer is invalid or not
605 * delayed-write, and it was on the EMPTY, LRU, or AGE queues,
606 * the buffer was free and we must decrement numfreebuffers.
608 if ((bp
->b_flags
& B_INVAL
) || (bp
->b_flags
& B_DELWRI
) == 0) {
611 case BQUEUE_DIRTY_HW
:
614 case BQUEUE_EMPTYKVA
:
628 * Get a buffer with the specified data. Look in the cache first. We
629 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
630 * is set, the buffer is valid and we do not have to do anything ( see
634 bread(struct vnode
*vp
, off_t loffset
, int size
, struct buf
**bpp
)
638 bp
= getblk(vp
, loffset
, size
, 0, 0);
641 /* if not found in cache, do some I/O */
642 if ((bp
->b_flags
& B_CACHE
) == 0) {
643 KASSERT(!(bp
->b_flags
& B_ASYNC
), ("bread: illegal async bp %p", bp
));
644 bp
->b_flags
&= ~(B_ERROR
| B_INVAL
);
645 bp
->b_cmd
= BUF_CMD_READ
;
646 vfs_busy_pages(vp
, bp
);
647 vn_strategy(vp
, &bp
->b_bio1
);
648 return (biowait(bp
));
656 * Operates like bread, but also starts asynchronous I/O on
657 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior
658 * to initiating I/O . If B_CACHE is set, the buffer is valid
659 * and we do not have to do anything.
662 breadn(struct vnode
*vp
, off_t loffset
, int size
, off_t
*raoffset
,
663 int *rabsize
, int cnt
, struct buf
**bpp
)
665 struct buf
*bp
, *rabp
;
667 int rv
= 0, readwait
= 0;
669 *bpp
= bp
= getblk(vp
, loffset
, size
, 0, 0);
671 /* if not found in cache, do some I/O */
672 if ((bp
->b_flags
& B_CACHE
) == 0) {
673 bp
->b_flags
&= ~(B_ERROR
| B_INVAL
);
674 bp
->b_cmd
= BUF_CMD_READ
;
675 vfs_busy_pages(vp
, bp
);
676 vn_strategy(vp
, &bp
->b_bio1
);
680 for (i
= 0; i
< cnt
; i
++, raoffset
++, rabsize
++) {
681 if (inmem(vp
, *raoffset
))
683 rabp
= getblk(vp
, *raoffset
, *rabsize
, 0, 0);
685 if ((rabp
->b_flags
& B_CACHE
) == 0) {
686 rabp
->b_flags
|= B_ASYNC
;
687 rabp
->b_flags
&= ~(B_ERROR
| B_INVAL
);
688 rabp
->b_cmd
= BUF_CMD_READ
;
689 vfs_busy_pages(vp
, rabp
);
691 vn_strategy(vp
, &rabp
->b_bio1
);
706 * Write, release buffer on completion. (Done by iodone
707 * if async). Do not bother writing anything if the buffer
710 * Note that we set B_CACHE here, indicating that buffer is
711 * fully valid and thus cacheable. This is true even of NFS
712 * now so we set it generally. This could be set either here
713 * or in biodone() since the I/O is synchronous. We put it
717 bwrite(struct buf
*bp
)
721 if (bp
->b_flags
& B_INVAL
) {
726 oldflags
= bp
->b_flags
;
728 if (BUF_REFCNTNB(bp
) == 0)
729 panic("bwrite: buffer is not busy???");
732 /* Mark the buffer clean */
735 bp
->b_flags
&= ~B_ERROR
;
736 bp
->b_flags
|= B_CACHE
;
737 bp
->b_cmd
= BUF_CMD_WRITE
;
738 vfs_busy_pages(bp
->b_vp
, bp
);
741 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
742 * valid for vnode-backed buffers.
744 bp
->b_runningbufspace
= bp
->b_bufsize
;
745 runningbufspace
+= bp
->b_runningbufspace
;
748 if (oldflags
& B_ASYNC
)
750 vn_strategy(bp
->b_vp
, &bp
->b_bio1
);
752 if ((oldflags
& B_ASYNC
) == 0) {
753 int rtval
= biowait(bp
);
763 * Delayed write. (Buffer is marked dirty). Do not bother writing
764 * anything if the buffer is marked invalid.
766 * Note that since the buffer must be completely valid, we can safely
767 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
768 * biodone() in order to prevent getblk from writing the buffer
772 bdwrite(struct buf
*bp
)
774 if (BUF_REFCNTNB(bp
) == 0)
775 panic("bdwrite: buffer is not busy");
777 if (bp
->b_flags
& B_INVAL
) {
784 * Set B_CACHE, indicating that the buffer is fully valid. This is
785 * true even of NFS now.
787 bp
->b_flags
|= B_CACHE
;
790 * This bmap keeps the system from needing to do the bmap later,
791 * perhaps when the system is attempting to do a sync. Since it
792 * is likely that the indirect block -- or whatever other datastructure
793 * that the filesystem needs is still in memory now, it is a good
794 * thing to do this. Note also, that if the pageout daemon is
795 * requesting a sync -- there might not be enough memory to do
796 * the bmap then... So, this is important to do.
798 if (bp
->b_bio2
.bio_offset
== NOOFFSET
) {
799 VOP_BMAP(bp
->b_vp
, bp
->b_loffset
, &bp
->b_bio2
.bio_offset
,
804 * Set the *dirty* buffer range based upon the VM system dirty pages.
809 * We need to do this here to satisfy the vnode_pager and the
810 * pageout daemon, so that it thinks that the pages have been
811 * "cleaned". Note that since the pages are in a delayed write
812 * buffer -- the VFS layer "will" see that the pages get written
813 * out on the next sync, or perhaps the cluster will be completed.
819 * Wakeup the buffer flushing daemon if we have a lot of dirty
820 * buffers (midpoint between our recovery point and our stall
823 bd_wakeup((lodirtybuffers
+ hidirtybuffers
) / 2);
826 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
827 * due to the softdep code.
834 * Turn buffer into delayed write request by marking it B_DELWRI.
835 * B_RELBUF and B_NOCACHE must be cleared.
837 * We reassign the buffer to itself to properly update it in the
840 * Since the buffer is not on a queue, we do not update the
841 * numfreebuffers count.
843 * Must be called from a critical section.
844 * The buffer must be on BQUEUE_NONE.
847 bdirty(struct buf
*bp
)
849 KASSERT(bp
->b_qindex
== BQUEUE_NONE
, ("bdirty: buffer %p still on queue %d", bp
, bp
->b_qindex
));
850 if (bp
->b_flags
& B_NOCACHE
) {
851 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp
);
852 bp
->b_flags
&= ~B_NOCACHE
;
854 if (bp
->b_flags
& B_INVAL
) {
855 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp
);
857 bp
->b_flags
&= ~B_RELBUF
;
859 if ((bp
->b_flags
& B_DELWRI
) == 0) {
860 bp
->b_flags
|= B_DELWRI
;
863 if (bp
->b_flags
& B_HEAVY
)
865 bd_wakeup((lodirtybuffers
+ hidirtybuffers
) / 2);
870 * Set B_HEAVY, indicating that this is a heavy-weight buffer that
871 * needs to be flushed with a different buf_daemon thread to avoid
872 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf().
875 bheavy(struct buf
*bp
)
877 if ((bp
->b_flags
& B_HEAVY
) == 0) {
878 bp
->b_flags
|= B_HEAVY
;
879 if (bp
->b_flags
& B_DELWRI
)
887 * Clear B_DELWRI for buffer.
889 * Since the buffer is not on a queue, we do not update the numfreebuffers
892 * Must be called from a critical section.
894 * The buffer is typically on BQUEUE_NONE but there is one case in
895 * brelse() that calls this function after placing the buffer on
900 bundirty(struct buf
*bp
)
902 if (bp
->b_flags
& B_DELWRI
) {
903 bp
->b_flags
&= ~B_DELWRI
;
906 if (bp
->b_flags
& B_HEAVY
)
911 * Since it is now being written, we can clear its deferred write flag.
913 bp
->b_flags
&= ~B_DEFERRED
;
919 * Asynchronous write. Start output on a buffer, but do not wait for
920 * it to complete. The buffer is released when the output completes.
922 * bwrite() ( or the VOP routine anyway ) is responsible for handling
923 * B_INVAL buffers. Not us.
926 bawrite(struct buf
*bp
)
928 bp
->b_flags
|= B_ASYNC
;
935 * Ordered write. Start output on a buffer, and flag it so that the
936 * device will write it in the order it was queued. The buffer is
937 * released when the output completes. bwrite() ( or the VOP routine
938 * anyway ) is responsible for handling B_INVAL buffers.
941 bowrite(struct buf
*bp
)
943 bp
->b_flags
|= B_ORDERED
| B_ASYNC
;
950 * Called prior to the locking of any vnodes when we are expecting to
951 * write. We do not want to starve the buffer cache with too many
952 * dirty buffers so we block here. By blocking prior to the locking
953 * of any vnodes we attempt to avoid the situation where a locked vnode
954 * prevents the various system daemons from flushing related buffers.
959 if (numdirtybuffers
>= hidirtybuffers
/ 2) {
961 while (numdirtybuffers
>= hidirtybuffers
) {
963 spin_lock_wr(&needsbuffer_spin
);
964 if (numdirtybuffers
>= hidirtybuffers
) {
965 needsbuffer
|= VFS_BIO_NEED_DIRTYFLUSH
;
966 msleep(&needsbuffer
, &needsbuffer_spin
, 0,
969 spin_unlock_wr(&needsbuffer_spin
);
974 else if (numdirtybuffershw
> hidirtybuffers
/ 2) {
977 while (numdirtybuffershw
> hidirtybuffers
) {
978 needsbuffer
|= VFS_BIO_NEED_DIRTYFLUSH
;
979 tsleep(&needsbuffer
, slpflags
, "newbuf",
987 * buf_dirty_count_severe:
989 * Return true if we have too many dirty buffers.
992 buf_dirty_count_severe(void)
994 return(numdirtybuffers
>= hidirtybuffers
);
1000 * Release a busy buffer and, if requested, free its resources. The
1001 * buffer will be stashed in the appropriate bufqueue[] allowing it
1002 * to be accessed later as a cache entity or reused for other purposes.
1005 brelse(struct buf
*bp
)
1008 int saved_flags
= bp
->b_flags
;
1011 KASSERT(!(bp
->b_flags
& (B_CLUSTER
|B_PAGING
)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp
));
1016 * If B_NOCACHE is set we are being asked to destroy the buffer and
1017 * its backing store. Clear B_DELWRI.
1019 * B_NOCACHE is set in two cases: (1) when the caller really wants
1020 * to destroy the buffer and backing store and (2) when the caller
1021 * wants to destroy the buffer and backing store after a write
1024 if ((bp
->b_flags
& (B_NOCACHE
|B_DELWRI
)) == (B_NOCACHE
|B_DELWRI
)) {
1028 if (bp
->b_flags
& B_LOCKED
)
1029 bp
->b_flags
&= ~B_ERROR
;
1032 * If a write error occurs and the caller does not want to throw
1033 * away the buffer, redirty the buffer. This will also clear
1036 if (bp
->b_cmd
== BUF_CMD_WRITE
&&
1037 (bp
->b_flags
& (B_ERROR
| B_INVAL
)) == B_ERROR
) {
1039 * Failed write, redirty. Must clear B_ERROR to prevent
1040 * pages from being scrapped. If B_INVAL is set then
1041 * this case is not run and the next case is run to
1042 * destroy the buffer. B_INVAL can occur if the buffer
1043 * is outside the range supported by the underlying device.
1045 bp
->b_flags
&= ~B_ERROR
;
1047 } else if ((bp
->b_flags
& (B_NOCACHE
| B_INVAL
| B_ERROR
)) ||
1048 (bp
->b_bufsize
<= 0) || bp
->b_cmd
== BUF_CMD_FREEBLKS
) {
1050 * Either a failed I/O or we were asked to free or not
1053 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the
1054 * buffer cannot be immediately freed.
1056 bp
->b_flags
|= B_INVAL
;
1057 if (LIST_FIRST(&bp
->b_dep
) != NULL
)
1059 if (bp
->b_flags
& B_DELWRI
) {
1061 if (bp
->b_flags
& B_HEAVY
)
1062 --numdirtybuffershw
;
1065 bp
->b_flags
&= ~(B_DELWRI
| B_CACHE
);
1069 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set.
1070 * If vfs_vmio_release() is called with either bit set, the
1071 * underlying pages may wind up getting freed causing a previous
1072 * write (bdwrite()) to get 'lost' because pages associated with
1073 * a B_DELWRI bp are marked clean. Pages associated with a
1074 * B_LOCKED buffer may be mapped by the filesystem.
1076 * If we want to release the buffer ourselves (rather then the
1077 * originator asking us to release it), give the originator a
1078 * chance to countermand the release by setting B_LOCKED.
1080 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1081 * if B_DELWRI is set.
1083 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1084 * on pages to return pages to the VM page queues.
1086 if (bp
->b_flags
& (B_DELWRI
| B_LOCKED
)) {
1087 bp
->b_flags
&= ~B_RELBUF
;
1088 } else if (vm_page_count_severe()) {
1089 if (LIST_FIRST(&bp
->b_dep
) != NULL
)
1091 if (bp
->b_flags
& (B_DELWRI
| B_LOCKED
))
1092 bp
->b_flags
&= ~B_RELBUF
;
1094 bp
->b_flags
|= B_RELBUF
;
1098 * At this point destroying the buffer is governed by the B_INVAL
1099 * or B_RELBUF flags.
1101 bp
->b_cmd
= BUF_CMD_DONE
;
1104 * VMIO buffer rundown. Make sure the VM page array is restored
1105 * after an I/O may have replaces some of the pages with bogus pages
1106 * in order to not destroy dirty pages in a fill-in read.
1108 * Note that due to the code above, if a buffer is marked B_DELWRI
1109 * then the B_RELBUF and B_NOCACHE bits will always be clear.
1110 * B_INVAL may still be set, however.
1112 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer
1113 * but not the backing store. B_NOCACHE will destroy the backing
1116 * Note that dirty NFS buffers contain byte-granular write ranges
1117 * and should not be destroyed w/ B_INVAL even if the backing store
1120 if (bp
->b_flags
& B_VMIO
) {
1122 * Rundown for VMIO buffers which are not dirty NFS buffers.
1134 * Get the base offset and length of the buffer. Note that
1135 * in the VMIO case if the buffer block size is not
1136 * page-aligned then b_data pointer may not be page-aligned.
1137 * But our b_xio.xio_pages array *IS* page aligned.
1139 * block sizes less then DEV_BSIZE (usually 512) are not
1140 * supported due to the page granularity bits (m->valid,
1141 * m->dirty, etc...).
1143 * See man buf(9) for more information
1146 resid
= bp
->b_bufsize
;
1147 foff
= bp
->b_loffset
;
1149 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
1150 m
= bp
->b_xio
.xio_pages
[i
];
1151 vm_page_flag_clear(m
, PG_ZERO
);
1153 * If we hit a bogus page, fixup *all* of them
1154 * now. Note that we left these pages wired
1155 * when we removed them so they had better exist,
1156 * and they cannot be ripped out from under us so
1157 * no critical section protection is necessary.
1159 if (m
== bogus_page
) {
1161 poff
= OFF_TO_IDX(bp
->b_loffset
);
1163 for (j
= i
; j
< bp
->b_xio
.xio_npages
; j
++) {
1166 mtmp
= bp
->b_xio
.xio_pages
[j
];
1167 if (mtmp
== bogus_page
) {
1168 mtmp
= vm_page_lookup(obj
, poff
+ j
);
1170 panic("brelse: page missing");
1172 bp
->b_xio
.xio_pages
[j
] = mtmp
;
1176 if ((bp
->b_flags
& B_INVAL
) == 0) {
1177 pmap_qenter(trunc_page((vm_offset_t
)bp
->b_data
),
1178 bp
->b_xio
.xio_pages
, bp
->b_xio
.xio_npages
);
1180 m
= bp
->b_xio
.xio_pages
[i
];
1184 * Invalidate the backing store if B_NOCACHE is set
1185 * (e.g. used with vinvalbuf()). If this is NFS
1186 * we impose a requirement that the block size be
1187 * a multiple of PAGE_SIZE and create a temporary
1188 * hack to basically invalidate the whole page. The
1189 * problem is that NFS uses really odd buffer sizes
1190 * especially when tracking piecemeal writes and
1191 * it also vinvalbuf()'s a lot, which would result
1192 * in only partial page validation and invalidation
1193 * here. If the file page is mmap()'d, however,
1194 * all the valid bits get set so after we invalidate
1195 * here we would end up with weird m->valid values
1196 * like 0xfc. nfs_getpages() can't handle this so
1197 * we clear all the valid bits for the NFS case
1198 * instead of just some of them.
1200 * The real bug is the VM system having to set m->valid
1201 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1202 * itself is an artifact of the whole 512-byte
1203 * granular mess that exists to support odd block
1204 * sizes and UFS meta-data block sizes (e.g. 6144).
1205 * A complete rewrite is required.
1207 if (bp
->b_flags
& (B_NOCACHE
|B_ERROR
)) {
1208 int poffset
= foff
& PAGE_MASK
;
1211 presid
= PAGE_SIZE
- poffset
;
1212 if (bp
->b_vp
->v_tag
== VT_NFS
&&
1213 bp
->b_vp
->v_type
== VREG
) {
1215 } else if (presid
> resid
) {
1218 KASSERT(presid
>= 0, ("brelse: extra page"));
1219 vm_page_set_invalid(m
, poffset
, presid
);
1221 resid
-= PAGE_SIZE
- (foff
& PAGE_MASK
);
1222 foff
= (foff
+ PAGE_SIZE
) & ~(off_t
)PAGE_MASK
;
1224 if (bp
->b_flags
& (B_INVAL
| B_RELBUF
))
1225 vfs_vmio_release(bp
);
1228 * Rundown for non-VMIO buffers.
1230 if (bp
->b_flags
& (B_INVAL
| B_RELBUF
)) {
1233 kprintf("brelse bp %p %08x/%08x: Warning, caught and fixed brelvp bug\n", bp
, saved_flags
, bp
->b_flags
);
1242 if (bp
->b_qindex
!= BQUEUE_NONE
)
1243 panic("brelse: free buffer onto another queue???");
1244 if (BUF_REFCNTNB(bp
) > 1) {
1245 /* Temporary panic to verify exclusive locking */
1246 /* This panic goes away when we allow shared refs */
1247 panic("brelse: multiple refs");
1248 /* do not release to free list */
1255 * Figure out the correct queue to place the cleaned up buffer on.
1256 * Buffers placed in the EMPTY or EMPTYKVA had better already be
1257 * disassociated from their vnode.
1259 if (bp
->b_flags
& B_LOCKED
) {
1261 * Buffers that are locked are placed in the locked queue
1262 * immediately, regardless of their state.
1264 bp
->b_qindex
= BQUEUE_LOCKED
;
1265 TAILQ_INSERT_TAIL(&bufqueues
[BQUEUE_LOCKED
], bp
, b_freelist
);
1266 } else if (bp
->b_bufsize
== 0) {
1268 * Buffers with no memory. Due to conditionals near the top
1269 * of brelse() such buffers should probably already be
1270 * marked B_INVAL and disassociated from their vnode.
1272 bp
->b_flags
|= B_INVAL
;
1273 KASSERT(bp
->b_vp
== NULL
, ("bp1 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp
, saved_flags
, bp
->b_flags
, bp
->b_vp
));
1274 KKASSERT((bp
->b_flags
& B_HASHED
) == 0);
1275 if (bp
->b_kvasize
) {
1276 bp
->b_qindex
= BQUEUE_EMPTYKVA
;
1278 bp
->b_qindex
= BQUEUE_EMPTY
;
1280 TAILQ_INSERT_HEAD(&bufqueues
[bp
->b_qindex
], bp
, b_freelist
);
1281 } else if (bp
->b_flags
& (B_ERROR
| B_INVAL
| B_NOCACHE
| B_RELBUF
)) {
1283 * Buffers with junk contents. Again these buffers had better
1284 * already be disassociated from their vnode.
1286 KASSERT(bp
->b_vp
== NULL
, ("bp2 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp
, saved_flags
, bp
->b_flags
, bp
->b_vp
));
1287 KKASSERT((bp
->b_flags
& B_HASHED
) == 0);
1288 bp
->b_flags
|= B_INVAL
;
1289 bp
->b_qindex
= BQUEUE_CLEAN
;
1290 TAILQ_INSERT_HEAD(&bufqueues
[BQUEUE_CLEAN
], bp
, b_freelist
);
1293 * Remaining buffers. These buffers are still associated with
1296 switch(bp
->b_flags
& (B_DELWRI
|B_HEAVY
|B_AGE
)) {
1297 case B_DELWRI
| B_AGE
:
1298 bp
->b_qindex
= BQUEUE_DIRTY
;
1299 TAILQ_INSERT_HEAD(&bufqueues
[BQUEUE_DIRTY
], bp
, b_freelist
);
1302 bp
->b_qindex
= BQUEUE_DIRTY
;
1303 TAILQ_INSERT_TAIL(&bufqueues
[BQUEUE_DIRTY
], bp
, b_freelist
);
1305 case B_DELWRI
| B_HEAVY
| B_AGE
:
1306 bp
->b_qindex
= BQUEUE_DIRTY_HW
;
1307 TAILQ_INSERT_HEAD(&bufqueues
[BQUEUE_DIRTY_HW
], bp
,
1310 case B_DELWRI
| B_HEAVY
:
1311 bp
->b_qindex
= BQUEUE_DIRTY_HW
;
1312 TAILQ_INSERT_TAIL(&bufqueues
[BQUEUE_DIRTY_HW
], bp
,
1315 case B_HEAVY
| B_AGE
:
1317 bp
->b_qindex
= BQUEUE_CLEAN
;
1318 TAILQ_INSERT_HEAD(&bufqueues
[BQUEUE_CLEAN
], bp
, b_freelist
);
1321 bp
->b_qindex
= BQUEUE_CLEAN
;
1322 TAILQ_INSERT_TAIL(&bufqueues
[BQUEUE_CLEAN
], bp
, b_freelist
);
1328 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1329 * on the correct queue.
1331 if ((bp
->b_flags
& (B_INVAL
|B_DELWRI
)) == (B_INVAL
|B_DELWRI
))
1335 * Fixup numfreebuffers count. The bp is on an appropriate queue
1336 * unless locked. We then bump numfreebuffers if it is not B_DELWRI.
1337 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1338 * if B_INVAL is set ).
1340 if ((bp
->b_flags
& (B_LOCKED
|B_DELWRI
)) == 0)
1344 * Something we can maybe free or reuse
1346 if (bp
->b_bufsize
|| bp
->b_kvasize
)
1350 * Clean up temporary flags and unlock the buffer.
1352 bp
->b_flags
&= ~(B_ORDERED
| B_ASYNC
| B_NOCACHE
| B_AGE
| B_RELBUF
|
1361 * Release a buffer back to the appropriate queue but do not try to free
1362 * it. The buffer is expected to be used again soon.
1364 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1365 * biodone() to requeue an async I/O on completion. It is also used when
1366 * known good buffers need to be requeued but we think we may need the data
1369 * XXX we should be able to leave the B_RELBUF hint set on completion.
1372 bqrelse(struct buf
*bp
)
1376 KASSERT(!(bp
->b_flags
& (B_CLUSTER
|B_PAGING
)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp
));
1378 if (bp
->b_qindex
!= BQUEUE_NONE
)
1379 panic("bqrelse: free buffer onto another queue???");
1380 if (BUF_REFCNTNB(bp
) > 1) {
1381 /* do not release to free list */
1382 panic("bqrelse: multiple refs");
1387 if (bp
->b_flags
& B_LOCKED
) {
1389 * Locked buffers are released to the locked queue. However,
1390 * if the buffer is dirty it will first go into the dirty
1391 * queue and later on after the I/O completes successfully it
1392 * will be released to the locked queue.
1394 bp
->b_flags
&= ~B_ERROR
;
1395 bp
->b_qindex
= BQUEUE_LOCKED
;
1396 TAILQ_INSERT_TAIL(&bufqueues
[BQUEUE_LOCKED
], bp
, b_freelist
);
1397 } else if (bp
->b_flags
& B_DELWRI
) {
1398 bp
->b_qindex
= (bp
->b_flags
& B_HEAVY
) ?
1399 BQUEUE_DIRTY_HW
: BQUEUE_DIRTY
;
1400 TAILQ_INSERT_TAIL(&bufqueues
[bp
->b_qindex
], bp
, b_freelist
);
1401 } else if (vm_page_count_severe()) {
1403 * We are too low on memory, we have to try to free the
1404 * buffer (most importantly: the wired pages making up its
1405 * backing store) *now*.
1411 bp
->b_qindex
= BQUEUE_CLEAN
;
1412 TAILQ_INSERT_TAIL(&bufqueues
[BQUEUE_CLEAN
], bp
, b_freelist
);
1415 if ((bp
->b_flags
& B_LOCKED
) == 0 &&
1416 ((bp
->b_flags
& B_INVAL
) || (bp
->b_flags
& B_DELWRI
) == 0)) {
1421 * Something we can maybe free or reuse.
1423 if (bp
->b_bufsize
&& !(bp
->b_flags
& B_DELWRI
))
1427 * Final cleanup and unlock. Clear bits that are only used while a
1428 * buffer is actively locked.
1430 bp
->b_flags
&= ~(B_ORDERED
| B_ASYNC
| B_NOCACHE
| B_AGE
| B_RELBUF
);
1438 * Return backing pages held by the buffer 'bp' back to the VM system
1439 * if possible. The pages are freed if they are no longer valid or
1440 * attempt to free if it was used for direct I/O otherwise they are
1441 * sent to the page cache.
1443 * Pages that were marked busy are left alone and skipped.
1445 * The KVA mapping (b_data) for the underlying pages is removed by
1449 vfs_vmio_release(struct buf
*bp
)
1455 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
1456 m
= bp
->b_xio
.xio_pages
[i
];
1457 bp
->b_xio
.xio_pages
[i
] = NULL
;
1459 * In order to keep page LRU ordering consistent, put
1460 * everything on the inactive queue.
1462 vm_page_unwire(m
, 0);
1464 * We don't mess with busy pages, it is
1465 * the responsibility of the process that
1466 * busied the pages to deal with them.
1468 if ((m
->flags
& PG_BUSY
) || (m
->busy
!= 0))
1471 if (m
->wire_count
== 0) {
1472 vm_page_flag_clear(m
, PG_ZERO
);
1474 * Might as well free the page if we can and it has
1475 * no valid data. We also free the page if the
1476 * buffer was used for direct I/O.
1478 if ((bp
->b_flags
& B_ASYNC
) == 0 && !m
->valid
&&
1479 m
->hold_count
== 0) {
1481 vm_page_protect(m
, VM_PROT_NONE
);
1483 } else if (bp
->b_flags
& B_DIRECT
) {
1484 vm_page_try_to_free(m
);
1485 } else if (vm_page_count_severe()) {
1486 vm_page_try_to_cache(m
);
1491 pmap_qremove(trunc_page((vm_offset_t
) bp
->b_data
), bp
->b_xio
.xio_npages
);
1492 if (bp
->b_bufsize
) {
1496 bp
->b_xio
.xio_npages
= 0;
1497 bp
->b_flags
&= ~B_VMIO
;
1505 * Implement clustered async writes for clearing out B_DELWRI buffers.
1506 * This is much better then the old way of writing only one buffer at
1507 * a time. Note that we may not be presented with the buffers in the
1508 * correct order, so we search for the cluster in both directions.
1510 * The buffer is locked on call.
1513 vfs_bio_awrite(struct buf
*bp
)
1517 off_t loffset
= bp
->b_loffset
;
1518 struct vnode
*vp
= bp
->b_vp
;
1526 * right now we support clustered writing only to regular files. If
1527 * we find a clusterable block we could be in the middle of a cluster
1528 * rather then at the beginning.
1530 * NOTE: b_bio1 contains the logical loffset and is aliased
1531 * to b_loffset. b_bio2 contains the translated block number.
1533 if ((vp
->v_type
== VREG
) &&
1534 (vp
->v_mount
!= 0) && /* Only on nodes that have the size info */
1535 (bp
->b_flags
& (B_CLUSTEROK
| B_INVAL
)) == B_CLUSTEROK
) {
1537 size
= vp
->v_mount
->mnt_stat
.f_iosize
;
1539 for (i
= size
; i
< MAXPHYS
; i
+= size
) {
1540 if ((bpa
= findblk(vp
, loffset
+ i
)) &&
1541 BUF_REFCNT(bpa
) == 0 &&
1542 ((bpa
->b_flags
& (B_DELWRI
| B_CLUSTEROK
| B_INVAL
)) ==
1543 (B_DELWRI
| B_CLUSTEROK
)) &&
1544 (bpa
->b_bufsize
== size
)) {
1545 if ((bpa
->b_bio2
.bio_offset
== NOOFFSET
) ||
1546 (bpa
->b_bio2
.bio_offset
!=
1547 bp
->b_bio2
.bio_offset
+ i
))
1553 for (j
= size
; i
+ j
<= MAXPHYS
&& j
<= loffset
; j
+= size
) {
1554 if ((bpa
= findblk(vp
, loffset
- j
)) &&
1555 BUF_REFCNT(bpa
) == 0 &&
1556 ((bpa
->b_flags
& (B_DELWRI
| B_CLUSTEROK
| B_INVAL
)) ==
1557 (B_DELWRI
| B_CLUSTEROK
)) &&
1558 (bpa
->b_bufsize
== size
)) {
1559 if ((bpa
->b_bio2
.bio_offset
== NOOFFSET
) ||
1560 (bpa
->b_bio2
.bio_offset
!=
1561 bp
->b_bio2
.bio_offset
- j
))
1570 * this is a possible cluster write
1572 if (nbytes
!= size
) {
1574 nwritten
= cluster_wbuild(vp
, size
,
1575 loffset
- j
, nbytes
);
1582 bp
->b_flags
|= B_ASYNC
;
1586 * default (old) behavior, writing out only one block
1588 * XXX returns b_bufsize instead of b_bcount for nwritten?
1590 nwritten
= bp
->b_bufsize
;
1599 * Find and initialize a new buffer header, freeing up existing buffers
1600 * in the bufqueues as necessary. The new buffer is returned locked.
1602 * Important: B_INVAL is not set. If the caller wishes to throw the
1603 * buffer away, the caller must set B_INVAL prior to calling brelse().
1606 * We have insufficient buffer headers
1607 * We have insufficient buffer space
1608 * buffer_map is too fragmented ( space reservation fails )
1609 * If we have to flush dirty buffers ( but we try to avoid this )
1611 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1612 * Instead we ask the buf daemon to do it for us. We attempt to
1613 * avoid piecemeal wakeups of the pageout daemon.
1617 getnewbuf(int blkflags
, int slptimeo
, int size
, int maxsize
)
1623 int slpflags
= (blkflags
& GETBLK_PCATCH
) ? PCATCH
: 0;
1624 static int flushingbufs
;
1627 * We can't afford to block since we might be holding a vnode lock,
1628 * which may prevent system daemons from running. We deal with
1629 * low-memory situations by proactively returning memory and running
1630 * async I/O rather then sync I/O.
1634 --getnewbufrestarts
;
1636 ++getnewbufrestarts
;
1639 * Setup for scan. If we do not have enough free buffers,
1640 * we setup a degenerate case that immediately fails. Note
1641 * that if we are specially marked process, we are allowed to
1642 * dip into our reserves.
1644 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1646 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1647 * However, there are a number of cases (defragging, reusing, ...)
1648 * where we cannot backup.
1650 nqindex
= BQUEUE_EMPTYKVA
;
1651 nbp
= TAILQ_FIRST(&bufqueues
[BQUEUE_EMPTYKVA
]);
1655 * If no EMPTYKVA buffers and we are either
1656 * defragging or reusing, locate a CLEAN buffer
1657 * to free or reuse. If bufspace useage is low
1658 * skip this step so we can allocate a new buffer.
1660 if (defrag
|| bufspace
>= lobufspace
) {
1661 nqindex
= BQUEUE_CLEAN
;
1662 nbp
= TAILQ_FIRST(&bufqueues
[BQUEUE_CLEAN
]);
1666 * If we could not find or were not allowed to reuse a
1667 * CLEAN buffer, check to see if it is ok to use an EMPTY
1668 * buffer. We can only use an EMPTY buffer if allocating
1669 * its KVA would not otherwise run us out of buffer space.
1671 if (nbp
== NULL
&& defrag
== 0 &&
1672 bufspace
+ maxsize
< hibufspace
) {
1673 nqindex
= BQUEUE_EMPTY
;
1674 nbp
= TAILQ_FIRST(&bufqueues
[BQUEUE_EMPTY
]);
1679 * Run scan, possibly freeing data and/or kva mappings on the fly
1683 while ((bp
= nbp
) != NULL
) {
1684 int qindex
= nqindex
;
1687 * Calculate next bp ( we can only use it if we do not block
1688 * or do other fancy things ).
1690 if ((nbp
= TAILQ_NEXT(bp
, b_freelist
)) == NULL
) {
1693 nqindex
= BQUEUE_EMPTYKVA
;
1694 if ((nbp
= TAILQ_FIRST(&bufqueues
[BQUEUE_EMPTYKVA
])))
1697 case BQUEUE_EMPTYKVA
:
1698 nqindex
= BQUEUE_CLEAN
;
1699 if ((nbp
= TAILQ_FIRST(&bufqueues
[BQUEUE_CLEAN
])))
1713 KASSERT(bp
->b_qindex
== qindex
, ("getnewbuf: inconsistent queue %d bp %p", qindex
, bp
));
1716 * Note: we no longer distinguish between VMIO and non-VMIO
1720 KASSERT((bp
->b_flags
& B_DELWRI
) == 0, ("delwri buffer %p found in queue %d", bp
, qindex
));
1723 * If we are defragging then we need a buffer with
1724 * b_kvasize != 0. XXX this situation should no longer
1725 * occur, if defrag is non-zero the buffer's b_kvasize
1726 * should also be non-zero at this point. XXX
1728 if (defrag
&& bp
->b_kvasize
== 0) {
1729 kprintf("Warning: defrag empty buffer %p\n", bp
);
1734 * Start freeing the bp. This is somewhat involved. nbp
1735 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers
1736 * on the clean list must be disassociated from their
1737 * current vnode. Buffers on the empty[kva] lists have
1738 * already been disassociated.
1741 if (BUF_LOCK(bp
, LK_EXCLUSIVE
| LK_NOWAIT
) != 0) {
1742 kprintf("getnewbuf: warning, locked buf %p, race corrected\n", bp
);
1743 tsleep(&bd_request
, 0, "gnbxxx", hz
/ 100);
1746 if (bp
->b_qindex
!= qindex
) {
1747 kprintf("getnewbuf: warning, BUF_LOCK blocked unexpectedly on buf %p index %d->%d, race corrected\n", bp
, qindex
, bp
->b_qindex
);
1754 * Dependancies must be handled before we disassociate the
1757 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
1758 * be immediately disassociated. HAMMER then becomes
1759 * responsible for releasing the buffer.
1761 if (LIST_FIRST(&bp
->b_dep
) != NULL
) {
1763 if (bp
->b_flags
& B_LOCKED
) {
1767 KKASSERT(LIST_FIRST(&bp
->b_dep
) == NULL
);
1770 if (qindex
== BQUEUE_CLEAN
) {
1771 if (bp
->b_flags
& B_VMIO
) {
1772 bp
->b_flags
&= ~B_ASYNC
;
1773 vfs_vmio_release(bp
);
1780 * NOTE: nbp is now entirely invalid. We can only restart
1781 * the scan from this point on.
1783 * Get the rest of the buffer freed up. b_kva* is still
1784 * valid after this operation.
1787 KASSERT(bp
->b_vp
== NULL
, ("bp3 %p flags %08x vnode %p qindex %d unexpectededly still associated!", bp
, bp
->b_flags
, bp
->b_vp
, qindex
));
1788 KKASSERT((bp
->b_flags
& B_HASHED
) == 0);
1791 * critical section protection is not required when
1792 * scrapping a buffer's contents because it is already
1798 bp
->b_flags
= B_BNOCLIP
;
1799 bp
->b_cmd
= BUF_CMD_DONE
;
1804 bp
->b_xio
.xio_npages
= 0;
1805 bp
->b_dirtyoff
= bp
->b_dirtyend
= 0;
1808 if (blkflags
& GETBLK_BHEAVY
)
1809 bp
->b_flags
|= B_HEAVY
;
1812 * If we are defragging then free the buffer.
1815 bp
->b_flags
|= B_INVAL
;
1823 * If we are overcomitted then recover the buffer and its
1824 * KVM space. This occurs in rare situations when multiple
1825 * processes are blocked in getnewbuf() or allocbuf().
1827 if (bufspace
>= hibufspace
)
1829 if (flushingbufs
&& bp
->b_kvasize
!= 0) {
1830 bp
->b_flags
|= B_INVAL
;
1835 if (bufspace
< lobufspace
)
1841 * If we exhausted our list, sleep as appropriate. We may have to
1842 * wakeup various daemons and write out some dirty buffers.
1844 * Generally we are sleeping due to insufficient buffer space.
1852 flags
= VFS_BIO_NEED_BUFSPACE
;
1854 } else if (bufspace
>= hibufspace
) {
1856 flags
= VFS_BIO_NEED_BUFSPACE
;
1859 flags
= VFS_BIO_NEED_ANY
;
1862 needsbuffer
|= flags
;
1863 bd_speedup(); /* heeeelp */
1864 while (needsbuffer
& flags
) {
1865 if (tsleep(&needsbuffer
, slpflags
, waitmsg
, slptimeo
))
1870 * We finally have a valid bp. We aren't quite out of the
1871 * woods, we still have to reserve kva space. In order
1872 * to keep fragmentation sane we only allocate kva in
1875 maxsize
= (maxsize
+ BKVAMASK
) & ~BKVAMASK
;
1877 if (maxsize
!= bp
->b_kvasize
) {
1878 vm_offset_t addr
= 0;
1883 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
1884 vm_map_lock(&buffer_map
);
1886 if (vm_map_findspace(&buffer_map
,
1887 vm_map_min(&buffer_map
), maxsize
,
1890 * Uh oh. Buffer map is too fragmented. We
1891 * must defragment the map.
1893 vm_map_unlock(&buffer_map
);
1894 vm_map_entry_release(count
);
1897 bp
->b_flags
|= B_INVAL
;
1902 vm_map_insert(&buffer_map
, &count
,
1904 addr
, addr
+ maxsize
,
1906 VM_PROT_ALL
, VM_PROT_ALL
,
1909 bp
->b_kvabase
= (caddr_t
) addr
;
1910 bp
->b_kvasize
= maxsize
;
1911 bufspace
+= bp
->b_kvasize
;
1914 vm_map_unlock(&buffer_map
);
1915 vm_map_entry_release(count
);
1917 bp
->b_data
= bp
->b_kvabase
;
1925 * Buffer flushing daemon. Buffers are normally flushed by the
1926 * update daemon but if it cannot keep up this process starts to
1927 * take the load in an attempt to prevent getnewbuf() from blocking.
1929 * Once a flush is initiated it does not stop until the number
1930 * of buffers falls below lodirtybuffers, but we will wake up anyone
1931 * waiting at the mid-point.
1934 static struct thread
*bufdaemon_td
;
1935 static struct thread
*bufdaemonhw_td
;
1937 static struct kproc_desc buf_kp
= {
1942 SYSINIT(bufdaemon
, SI_SUB_KTHREAD_BUF
, SI_ORDER_FIRST
,
1943 kproc_start
, &buf_kp
)
1945 static struct kproc_desc bufhw_kp
= {
1950 SYSINIT(bufdaemon_hw
, SI_SUB_KTHREAD_BUF
, SI_ORDER_FIRST
,
1951 kproc_start
, &bufhw_kp
)
1957 * This process needs to be suspended prior to shutdown sync.
1959 EVENTHANDLER_REGISTER(shutdown_pre_sync
, shutdown_kproc
,
1960 bufdaemon_td
, SHUTDOWN_PRI_LAST
);
1963 * This process is allowed to take the buffer cache to the limit
1968 kproc_suspend_loop();
1971 * Do the flush. Limit the amount of in-transit I/O we
1972 * allow to build up, otherwise we would completely saturate
1973 * the I/O system. Wakeup any waiting processes before we
1974 * normally would so they can run in parallel with our drain.
1976 while (numdirtybuffers
> lodirtybuffers
) {
1977 if (flushbufqueues(BQUEUE_DIRTY
) == 0)
1979 waitrunningbufspace();
1985 * Only clear bd_request if we have reached our low water
1986 * mark. The buf_daemon normally waits 5 seconds and
1987 * then incrementally flushes any dirty buffers that have
1988 * built up, within reason.
1990 * If we were unable to hit our low water mark and couldn't
1991 * find any flushable buffers, we sleep half a second.
1992 * Otherwise we loop immediately.
1994 if (numdirtybuffers
<= lodirtybuffers
) {
1996 * We reached our low water mark, reset the
1997 * request and sleep until we are needed again.
1998 * The sleep is just so the suspend code works.
2000 spin_lock_wr(&needsbuffer_spin
);
2002 msleep(&bd_request
, &needsbuffer_spin
, 0,
2004 spin_unlock_wr(&needsbuffer_spin
);
2007 * We couldn't find any flushable dirty buffers but
2008 * still have too many dirty buffers, we
2009 * have to sleep and try again. (rare)
2011 tsleep(&bd_request
, 0, "qsleep", hz
/ 2);
2020 * This process needs to be suspended prior to shutdown sync.
2022 EVENTHANDLER_REGISTER(shutdown_pre_sync
, shutdown_kproc
,
2023 bufdaemonhw_td
, SHUTDOWN_PRI_LAST
);
2026 * This process is allowed to take the buffer cache to the limit
2031 kproc_suspend_loop();
2034 * Do the flush. Limit the amount of in-transit I/O we
2035 * allow to build up, otherwise we would completely saturate
2036 * the I/O system. Wakeup any waiting processes before we
2037 * normally would so they can run in parallel with our drain.
2039 while (numdirtybuffershw
> lodirtybuffers
) {
2040 if (flushbufqueues(BQUEUE_DIRTY_HW
) == 0)
2042 waitrunningbufspace();
2047 * Only clear bd_request if we have reached our low water
2048 * mark. The buf_daemon normally waits 5 seconds and
2049 * then incrementally flushes any dirty buffers that have
2050 * built up, within reason.
2052 * If we were unable to hit our low water mark and couldn't
2053 * find any flushable buffers, we sleep half a second.
2054 * Otherwise we loop immediately.
2056 if (numdirtybuffershw
<= lodirtybuffers
) {
2058 * We reached our low water mark, reset the
2059 * request and sleep until we are needed again.
2060 * The sleep is just so the suspend code works.
2062 spin_lock_wr(&needsbuffer_spin
);
2064 msleep(&bd_request_hw
, &needsbuffer_spin
, 0,
2066 spin_unlock_wr(&needsbuffer_spin
);
2069 * We couldn't find any flushable dirty buffers but
2070 * still have too many dirty buffers, we
2071 * have to sleep and try again. (rare)
2073 tsleep(&bd_request_hw
, 0, "qsleep", hz
/ 2);
2081 * Try to flush a buffer in the dirty queue. We must be careful to
2082 * free up B_INVAL buffers instead of write them, which NFS is
2083 * particularly sensitive to.
2087 flushbufqueues(bufq_type_t q
)
2092 bp
= TAILQ_FIRST(&bufqueues
[q
]);
2095 KASSERT((bp
->b_flags
& B_DELWRI
),
2096 ("unexpected clean buffer %p", bp
));
2097 if (bp
->b_flags
& B_DELWRI
) {
2098 if (bp
->b_flags
& B_INVAL
) {
2099 if (BUF_LOCK(bp
, LK_EXCLUSIVE
| LK_NOWAIT
) != 0)
2100 panic("flushbufqueues: locked buf");
2106 if (LIST_FIRST(&bp
->b_dep
) != NULL
&&
2107 (bp
->b_flags
& B_DEFERRED
) == 0 &&
2108 buf_countdeps(bp
, 0)) {
2109 TAILQ_REMOVE(&bufqueues
[q
], bp
, b_freelist
);
2110 TAILQ_INSERT_TAIL(&bufqueues
[q
], bp
,
2112 bp
->b_flags
|= B_DEFERRED
;
2113 bp
= TAILQ_FIRST(&bufqueues
[q
]);
2118 * Only write it out if we can successfully lock
2119 * it. If the buffer has a dependancy,
2120 * buf_checkwrite must also return 0.
2122 if (BUF_LOCK(bp
, LK_EXCLUSIVE
| LK_NOWAIT
) == 0) {
2123 if (LIST_FIRST(&bp
->b_dep
) != NULL
&&
2124 buf_checkwrite(bp
)) {
2134 bp
= TAILQ_NEXT(bp
, b_freelist
);
2142 * Returns true if no I/O is needed to access the associated VM object.
2143 * This is like findblk except it also hunts around in the VM system for
2146 * Note that we ignore vm_page_free() races from interrupts against our
2147 * lookup, since if the caller is not protected our return value will not
2148 * be any more valid then otherwise once we exit the critical section.
2151 inmem(struct vnode
*vp
, off_t loffset
)
2154 vm_offset_t toff
, tinc
, size
;
2157 if (findblk(vp
, loffset
))
2159 if (vp
->v_mount
== NULL
)
2161 if ((obj
= vp
->v_object
) == NULL
)
2165 if (size
> vp
->v_mount
->mnt_stat
.f_iosize
)
2166 size
= vp
->v_mount
->mnt_stat
.f_iosize
;
2168 for (toff
= 0; toff
< vp
->v_mount
->mnt_stat
.f_iosize
; toff
+= tinc
) {
2169 m
= vm_page_lookup(obj
, OFF_TO_IDX(loffset
+ toff
));
2173 if (tinc
> PAGE_SIZE
- ((toff
+ loffset
) & PAGE_MASK
))
2174 tinc
= PAGE_SIZE
- ((toff
+ loffset
) & PAGE_MASK
);
2175 if (vm_page_is_valid(m
,
2176 (vm_offset_t
) ((toff
+ loffset
) & PAGE_MASK
), tinc
) == 0)
2185 * Sets the dirty range for a buffer based on the status of the dirty
2186 * bits in the pages comprising the buffer.
2188 * The range is limited to the size of the buffer.
2190 * This routine is primarily used by NFS, but is generalized for the
2194 vfs_setdirty(struct buf
*bp
)
2200 * Degenerate case - empty buffer
2203 if (bp
->b_bufsize
== 0)
2207 * We qualify the scan for modified pages on whether the
2208 * object has been flushed yet. The OBJ_WRITEABLE flag
2209 * is not cleared simply by protecting pages off.
2212 if ((bp
->b_flags
& B_VMIO
) == 0)
2215 object
= bp
->b_xio
.xio_pages
[0]->object
;
2217 if ((object
->flags
& OBJ_WRITEABLE
) && !(object
->flags
& OBJ_MIGHTBEDIRTY
))
2218 kprintf("Warning: object %p writeable but not mightbedirty\n", object
);
2219 if (!(object
->flags
& OBJ_WRITEABLE
) && (object
->flags
& OBJ_MIGHTBEDIRTY
))
2220 kprintf("Warning: object %p mightbedirty but not writeable\n", object
);
2222 if (object
->flags
& (OBJ_MIGHTBEDIRTY
|OBJ_CLEANING
)) {
2223 vm_offset_t boffset
;
2224 vm_offset_t eoffset
;
2227 * test the pages to see if they have been modified directly
2228 * by users through the VM system.
2230 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
2231 vm_page_flag_clear(bp
->b_xio
.xio_pages
[i
], PG_ZERO
);
2232 vm_page_test_dirty(bp
->b_xio
.xio_pages
[i
]);
2236 * Calculate the encompassing dirty range, boffset and eoffset,
2237 * (eoffset - boffset) bytes.
2240 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
2241 if (bp
->b_xio
.xio_pages
[i
]->dirty
)
2244 boffset
= (i
<< PAGE_SHIFT
) - (bp
->b_loffset
& PAGE_MASK
);
2246 for (i
= bp
->b_xio
.xio_npages
- 1; i
>= 0; --i
) {
2247 if (bp
->b_xio
.xio_pages
[i
]->dirty
) {
2251 eoffset
= ((i
+ 1) << PAGE_SHIFT
) - (bp
->b_loffset
& PAGE_MASK
);
2254 * Fit it to the buffer.
2257 if (eoffset
> bp
->b_bcount
)
2258 eoffset
= bp
->b_bcount
;
2261 * If we have a good dirty range, merge with the existing
2265 if (boffset
< eoffset
) {
2266 if (bp
->b_dirtyoff
> boffset
)
2267 bp
->b_dirtyoff
= boffset
;
2268 if (bp
->b_dirtyend
< eoffset
)
2269 bp
->b_dirtyend
= eoffset
;
2277 * Locate and return the specified buffer, or NULL if the buffer does
2278 * not exist. Do not attempt to lock the buffer or manipulate it in
2279 * any way. The caller must validate that the correct buffer has been
2280 * obtain after locking it.
2283 findblk(struct vnode
*vp
, off_t loffset
)
2288 bp
= buf_rb_hash_RB_LOOKUP(&vp
->v_rbhash_tree
, loffset
);
2296 * Get a block given a specified block and offset into a file/device.
2297 * B_INVAL may or may not be set on return. The caller should clear
2298 * B_INVAL prior to initiating a READ.
2300 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2301 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2302 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer
2303 * without doing any of those things the system will likely believe
2304 * the buffer to be valid (especially if it is not B_VMIO), and the
2305 * next getblk() will return the buffer with B_CACHE set.
2307 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2308 * an existing buffer.
2310 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2311 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2312 * and then cleared based on the backing VM. If the previous buffer is
2313 * non-0-sized but invalid, B_CACHE will be cleared.
2315 * If getblk() must create a new buffer, the new buffer is returned with
2316 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2317 * case it is returned with B_INVAL clear and B_CACHE set based on the
2320 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2321 * B_CACHE bit is clear.
2323 * What this means, basically, is that the caller should use B_CACHE to
2324 * determine whether the buffer is fully valid or not and should clear
2325 * B_INVAL prior to issuing a read. If the caller intends to validate
2326 * the buffer by loading its data area with something, the caller needs
2327 * to clear B_INVAL. If the caller does this without issuing an I/O,
2328 * the caller should set B_CACHE ( as an optimization ), else the caller
2329 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2330 * a write attempt or if it was a successfull read. If the caller
2331 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2332 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2336 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return
2337 * GETBLK_BHEAVY - heavy-weight buffer cache buffer
2340 getblk(struct vnode
*vp
, off_t loffset
, int size
, int blkflags
, int slptimeo
)
2343 int slpflags
= (blkflags
& GETBLK_PCATCH
) ? PCATCH
: 0;
2345 if (size
> MAXBSIZE
)
2346 panic("getblk: size(%d) > MAXBSIZE(%d)", size
, MAXBSIZE
);
2347 if (vp
->v_object
== NULL
)
2348 panic("getblk: vnode %p has no object!", vp
);
2352 if ((bp
= findblk(vp
, loffset
))) {
2354 * The buffer was found in the cache, but we need to lock it.
2355 * Even with LK_NOWAIT the lockmgr may break our critical
2356 * section, so double-check the validity of the buffer
2357 * once the lock has been obtained.
2359 if (BUF_LOCK(bp
, LK_EXCLUSIVE
| LK_NOWAIT
)) {
2360 int lkflags
= LK_EXCLUSIVE
| LK_SLEEPFAIL
;
2361 if (blkflags
& GETBLK_PCATCH
)
2362 lkflags
|= LK_PCATCH
;
2363 if (BUF_TIMELOCK(bp
, lkflags
, "getblk", slptimeo
) ==
2372 * Once the buffer has been locked, make sure we didn't race
2373 * a buffer recyclement. Buffers that are no longer hashed
2374 * will have b_vp == NULL, so this takes care of that check
2377 if (bp
->b_vp
!= vp
|| bp
->b_loffset
!= loffset
) {
2378 kprintf("Warning buffer %p (vp %p loffset %lld) was recycled\n", bp
, vp
, loffset
);
2384 * All vnode-based buffers must be backed by a VM object.
2386 KKASSERT(bp
->b_flags
& B_VMIO
);
2387 KKASSERT(bp
->b_cmd
== BUF_CMD_DONE
);
2390 * Make sure that B_INVAL buffers do not have a cached
2391 * block number translation.
2393 if ((bp
->b_flags
& B_INVAL
) && (bp
->b_bio2
.bio_offset
!= NOOFFSET
)) {
2394 kprintf("Warning invalid buffer %p (vp %p loffset %lld) did not have cleared bio_offset cache\n", bp
, vp
, loffset
);
2395 clearbiocache(&bp
->b_bio2
);
2399 * The buffer is locked. B_CACHE is cleared if the buffer is
2402 if (bp
->b_flags
& B_INVAL
)
2403 bp
->b_flags
&= ~B_CACHE
;
2407 * Any size inconsistancy with a dirty buffer or a buffer
2408 * with a softupdates dependancy must be resolved. Resizing
2409 * the buffer in such circumstances can lead to problems.
2411 if (size
!= bp
->b_bcount
) {
2412 if (bp
->b_flags
& B_DELWRI
) {
2413 bp
->b_flags
|= B_NOCACHE
;
2415 } else if (LIST_FIRST(&bp
->b_dep
)) {
2416 bp
->b_flags
|= B_NOCACHE
;
2419 bp
->b_flags
|= B_RELBUF
;
2424 KKASSERT(size
<= bp
->b_kvasize
);
2425 KASSERT(bp
->b_loffset
!= NOOFFSET
,
2426 ("getblk: no buffer offset"));
2429 * A buffer with B_DELWRI set and B_CACHE clear must
2430 * be committed before we can return the buffer in
2431 * order to prevent the caller from issuing a read
2432 * ( due to B_CACHE not being set ) and overwriting
2435 * Most callers, including NFS and FFS, need this to
2436 * operate properly either because they assume they
2437 * can issue a read if B_CACHE is not set, or because
2438 * ( for example ) an uncached B_DELWRI might loop due
2439 * to softupdates re-dirtying the buffer. In the latter
2440 * case, B_CACHE is set after the first write completes,
2441 * preventing further loops.
2443 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2444 * above while extending the buffer, we cannot allow the
2445 * buffer to remain with B_CACHE set after the write
2446 * completes or it will represent a corrupt state. To
2447 * deal with this we set B_NOCACHE to scrap the buffer
2450 * We might be able to do something fancy, like setting
2451 * B_CACHE in bwrite() except if B_DELWRI is already set,
2452 * so the below call doesn't set B_CACHE, but that gets real
2453 * confusing. This is much easier.
2456 if ((bp
->b_flags
& (B_CACHE
|B_DELWRI
)) == B_DELWRI
) {
2457 bp
->b_flags
|= B_NOCACHE
;
2464 * Buffer is not in-core, create new buffer. The buffer
2465 * returned by getnewbuf() is locked. Note that the returned
2466 * buffer is also considered valid (not marked B_INVAL).
2468 * Calculating the offset for the I/O requires figuring out
2469 * the block size. We use DEV_BSIZE for VBLK or VCHR and
2470 * the mount's f_iosize otherwise. If the vnode does not
2471 * have an associated mount we assume that the passed size is
2474 * Note that vn_isdisk() cannot be used here since it may
2475 * return a failure for numerous reasons. Note that the
2476 * buffer size may be larger then the block size (the caller
2477 * will use block numbers with the proper multiple). Beware
2478 * of using any v_* fields which are part of unions. In
2479 * particular, in DragonFly the mount point overloading
2480 * mechanism uses the namecache only and the underlying
2481 * directory vnode is not a special case.
2485 if (vp
->v_type
== VBLK
|| vp
->v_type
== VCHR
)
2487 else if (vp
->v_mount
)
2488 bsize
= vp
->v_mount
->mnt_stat
.f_iosize
;
2492 maxsize
= size
+ (loffset
& PAGE_MASK
);
2493 maxsize
= imax(maxsize
, bsize
);
2495 if ((bp
= getnewbuf(blkflags
, slptimeo
, size
, maxsize
)) == NULL
) {
2496 if (slpflags
|| slptimeo
) {
2504 * This code is used to make sure that a buffer is not
2505 * created while the getnewbuf routine is blocked.
2506 * This can be a problem whether the vnode is locked or not.
2507 * If the buffer is created out from under us, we have to
2508 * throw away the one we just created. There is no window
2509 * race because we are safely running in a critical section
2510 * from the point of the duplicate buffer creation through
2511 * to here, and we've locked the buffer.
2513 if (findblk(vp
, loffset
)) {
2514 bp
->b_flags
|= B_INVAL
;
2520 * Insert the buffer into the hash, so that it can
2521 * be found by findblk().
2523 * Make sure the translation layer has been cleared.
2525 bp
->b_loffset
= loffset
;
2526 bp
->b_bio2
.bio_offset
= NOOFFSET
;
2527 /* bp->b_bio2.bio_next = NULL; */
2532 * All vnode-based buffers must be backed by a VM object.
2534 KKASSERT(vp
->v_object
!= NULL
);
2535 bp
->b_flags
|= B_VMIO
;
2536 KKASSERT(bp
->b_cmd
== BUF_CMD_DONE
);
2548 * Reacquire a buffer that was previously released to the locked queue,
2549 * or reacquire a buffer which is interlocked by having bioops->io_deallocate
2550 * set B_LOCKED (which handles the acquisition race).
2552 * To this end, either B_LOCKED must be set or the dependancy list must be
2556 regetblk(struct buf
*bp
)
2558 KKASSERT((bp
->b_flags
& B_LOCKED
) || LIST_FIRST(&bp
->b_dep
) != NULL
);
2559 BUF_LOCK(bp
, LK_EXCLUSIVE
| LK_RETRY
);
2568 * Get an empty, disassociated buffer of given size. The buffer is
2569 * initially set to B_INVAL.
2571 * critical section protection is not required for the allocbuf()
2572 * call because races are impossible here.
2580 maxsize
= (size
+ BKVAMASK
) & ~BKVAMASK
;
2583 while ((bp
= getnewbuf(0, 0, size
, maxsize
)) == 0)
2587 bp
->b_flags
|= B_INVAL
; /* b_dep cleared by getnewbuf() */
2595 * This code constitutes the buffer memory from either anonymous system
2596 * memory (in the case of non-VMIO operations) or from an associated
2597 * VM object (in the case of VMIO operations). This code is able to
2598 * resize a buffer up or down.
2600 * Note that this code is tricky, and has many complications to resolve
2601 * deadlock or inconsistant data situations. Tread lightly!!!
2602 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2603 * the caller. Calling this code willy nilly can result in the loss of data.
2605 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
2606 * B_CACHE for the non-VMIO case.
2608 * This routine does not need to be called from a critical section but you
2609 * must own the buffer.
2612 allocbuf(struct buf
*bp
, int size
)
2614 int newbsize
, mbsize
;
2617 if (BUF_REFCNT(bp
) == 0)
2618 panic("allocbuf: buffer not busy");
2620 if (bp
->b_kvasize
< size
)
2621 panic("allocbuf: buffer too small");
2623 if ((bp
->b_flags
& B_VMIO
) == 0) {
2627 * Just get anonymous memory from the kernel. Don't
2628 * mess with B_CACHE.
2630 mbsize
= (size
+ DEV_BSIZE
- 1) & ~(DEV_BSIZE
- 1);
2631 if (bp
->b_flags
& B_MALLOC
)
2634 newbsize
= round_page(size
);
2636 if (newbsize
< bp
->b_bufsize
) {
2638 * Malloced buffers are not shrunk
2640 if (bp
->b_flags
& B_MALLOC
) {
2642 bp
->b_bcount
= size
;
2644 kfree(bp
->b_data
, M_BIOBUF
);
2645 if (bp
->b_bufsize
) {
2646 bufmallocspace
-= bp
->b_bufsize
;
2650 bp
->b_data
= bp
->b_kvabase
;
2652 bp
->b_flags
&= ~B_MALLOC
;
2658 (vm_offset_t
) bp
->b_data
+ newbsize
,
2659 (vm_offset_t
) bp
->b_data
+ bp
->b_bufsize
);
2660 } else if (newbsize
> bp
->b_bufsize
) {
2662 * We only use malloced memory on the first allocation.
2663 * and revert to page-allocated memory when the buffer
2666 if ((bufmallocspace
< maxbufmallocspace
) &&
2667 (bp
->b_bufsize
== 0) &&
2668 (mbsize
<= PAGE_SIZE
/2)) {
2670 bp
->b_data
= kmalloc(mbsize
, M_BIOBUF
, M_WAITOK
);
2671 bp
->b_bufsize
= mbsize
;
2672 bp
->b_bcount
= size
;
2673 bp
->b_flags
|= B_MALLOC
;
2674 bufmallocspace
+= mbsize
;
2680 * If the buffer is growing on its other-than-first
2681 * allocation, then we revert to the page-allocation
2684 if (bp
->b_flags
& B_MALLOC
) {
2685 origbuf
= bp
->b_data
;
2686 origbufsize
= bp
->b_bufsize
;
2687 bp
->b_data
= bp
->b_kvabase
;
2688 if (bp
->b_bufsize
) {
2689 bufmallocspace
-= bp
->b_bufsize
;
2693 bp
->b_flags
&= ~B_MALLOC
;
2694 newbsize
= round_page(newbsize
);
2698 (vm_offset_t
) bp
->b_data
+ bp
->b_bufsize
,
2699 (vm_offset_t
) bp
->b_data
+ newbsize
);
2701 bcopy(origbuf
, bp
->b_data
, origbufsize
);
2702 kfree(origbuf
, M_BIOBUF
);
2709 newbsize
= (size
+ DEV_BSIZE
- 1) & ~(DEV_BSIZE
- 1);
2710 desiredpages
= ((int)(bp
->b_loffset
& PAGE_MASK
) +
2711 newbsize
+ PAGE_MASK
) >> PAGE_SHIFT
;
2712 KKASSERT(desiredpages
<= XIO_INTERNAL_PAGES
);
2714 if (bp
->b_flags
& B_MALLOC
)
2715 panic("allocbuf: VMIO buffer can't be malloced");
2717 * Set B_CACHE initially if buffer is 0 length or will become
2720 if (size
== 0 || bp
->b_bufsize
== 0)
2721 bp
->b_flags
|= B_CACHE
;
2723 if (newbsize
< bp
->b_bufsize
) {
2725 * DEV_BSIZE aligned new buffer size is less then the
2726 * DEV_BSIZE aligned existing buffer size. Figure out
2727 * if we have to remove any pages.
2729 if (desiredpages
< bp
->b_xio
.xio_npages
) {
2730 for (i
= desiredpages
; i
< bp
->b_xio
.xio_npages
; i
++) {
2732 * the page is not freed here -- it
2733 * is the responsibility of
2734 * vnode_pager_setsize
2736 m
= bp
->b_xio
.xio_pages
[i
];
2737 KASSERT(m
!= bogus_page
,
2738 ("allocbuf: bogus page found"));
2739 while (vm_page_sleep_busy(m
, TRUE
, "biodep"))
2742 bp
->b_xio
.xio_pages
[i
] = NULL
;
2743 vm_page_unwire(m
, 0);
2745 pmap_qremove((vm_offset_t
) trunc_page((vm_offset_t
)bp
->b_data
) +
2746 (desiredpages
<< PAGE_SHIFT
), (bp
->b_xio
.xio_npages
- desiredpages
));
2747 bp
->b_xio
.xio_npages
= desiredpages
;
2749 } else if (size
> bp
->b_bcount
) {
2751 * We are growing the buffer, possibly in a
2752 * byte-granular fashion.
2760 * Step 1, bring in the VM pages from the object,
2761 * allocating them if necessary. We must clear
2762 * B_CACHE if these pages are not valid for the
2763 * range covered by the buffer.
2765 * critical section protection is required to protect
2766 * against interrupts unbusying and freeing pages
2767 * between our vm_page_lookup() and our
2768 * busycheck/wiring call.
2774 while (bp
->b_xio
.xio_npages
< desiredpages
) {
2778 pi
= OFF_TO_IDX(bp
->b_loffset
) + bp
->b_xio
.xio_npages
;
2779 if ((m
= vm_page_lookup(obj
, pi
)) == NULL
) {
2781 * note: must allocate system pages
2782 * since blocking here could intefere
2783 * with paging I/O, no matter which
2786 m
= vm_page_alloc(obj
, pi
, VM_ALLOC_NORMAL
| VM_ALLOC_SYSTEM
);
2789 vm_pageout_deficit
+= desiredpages
-
2790 bp
->b_xio
.xio_npages
;
2794 bp
->b_flags
&= ~B_CACHE
;
2795 bp
->b_xio
.xio_pages
[bp
->b_xio
.xio_npages
] = m
;
2796 ++bp
->b_xio
.xio_npages
;
2802 * We found a page. If we have to sleep on it,
2803 * retry because it might have gotten freed out
2806 * We can only test PG_BUSY here. Blocking on
2807 * m->busy might lead to a deadlock:
2809 * vm_fault->getpages->cluster_read->allocbuf
2813 if (vm_page_sleep_busy(m
, FALSE
, "pgtblk"))
2817 * We have a good page. Should we wakeup the
2820 if ((curthread
!= pagethread
) &&
2821 ((m
->queue
- m
->pc
) == PQ_CACHE
) &&
2822 ((vmstats
.v_free_count
+ vmstats
.v_cache_count
) <
2823 (vmstats
.v_free_min
+ vmstats
.v_cache_min
))) {
2824 pagedaemon_wakeup();
2826 vm_page_flag_clear(m
, PG_ZERO
);
2828 bp
->b_xio
.xio_pages
[bp
->b_xio
.xio_npages
] = m
;
2829 ++bp
->b_xio
.xio_npages
;
2834 * Step 2. We've loaded the pages into the buffer,
2835 * we have to figure out if we can still have B_CACHE
2836 * set. Note that B_CACHE is set according to the
2837 * byte-granular range ( bcount and size ), not the
2838 * aligned range ( newbsize ).
2840 * The VM test is against m->valid, which is DEV_BSIZE
2841 * aligned. Needless to say, the validity of the data
2842 * needs to also be DEV_BSIZE aligned. Note that this
2843 * fails with NFS if the server or some other client
2844 * extends the file's EOF. If our buffer is resized,
2845 * B_CACHE may remain set! XXX
2848 toff
= bp
->b_bcount
;
2849 tinc
= PAGE_SIZE
- ((bp
->b_loffset
+ toff
) & PAGE_MASK
);
2851 while ((bp
->b_flags
& B_CACHE
) && toff
< size
) {
2854 if (tinc
> (size
- toff
))
2857 pi
= ((bp
->b_loffset
& PAGE_MASK
) + toff
) >>
2865 bp
->b_xio
.xio_pages
[pi
]
2872 * Step 3, fixup the KVM pmap. Remember that
2873 * bp->b_data is relative to bp->b_loffset, but
2874 * bp->b_loffset may be offset into the first page.
2877 bp
->b_data
= (caddr_t
)
2878 trunc_page((vm_offset_t
)bp
->b_data
);
2880 (vm_offset_t
)bp
->b_data
,
2881 bp
->b_xio
.xio_pages
,
2882 bp
->b_xio
.xio_npages
2884 bp
->b_data
= (caddr_t
)((vm_offset_t
)bp
->b_data
|
2885 (vm_offset_t
)(bp
->b_loffset
& PAGE_MASK
));
2888 if (newbsize
< bp
->b_bufsize
)
2890 bp
->b_bufsize
= newbsize
; /* actual buffer allocation */
2891 bp
->b_bcount
= size
; /* requested buffer size */
2898 * Wait for buffer I/O completion, returning error status. The buffer
2899 * is left locked on return. B_EINTR is converted into an EINTR error
2902 * NOTE! The original b_cmd is lost on return, since b_cmd will be
2903 * set to BUF_CMD_DONE.
2906 biowait(struct buf
*bp
)
2909 while (bp
->b_cmd
!= BUF_CMD_DONE
) {
2910 if (bp
->b_cmd
== BUF_CMD_READ
)
2911 tsleep(bp
, 0, "biord", 0);
2913 tsleep(bp
, 0, "biowr", 0);
2916 if (bp
->b_flags
& B_EINTR
) {
2917 bp
->b_flags
&= ~B_EINTR
;
2920 if (bp
->b_flags
& B_ERROR
) {
2921 return (bp
->b_error
? bp
->b_error
: EIO
);
2928 * This associates a tracking count with an I/O. vn_strategy() and
2929 * dev_dstrategy() do this automatically but there are a few cases
2930 * where a vnode or device layer is bypassed when a block translation
2931 * is cached. In such cases bio_start_transaction() may be called on
2932 * the bypassed layers so the system gets an I/O in progress indication
2933 * for those higher layers.
2936 bio_start_transaction(struct bio
*bio
, struct bio_track
*track
)
2938 bio
->bio_track
= track
;
2939 atomic_add_int(&track
->bk_active
, 1);
2943 * Initiate I/O on a vnode.
2946 vn_strategy(struct vnode
*vp
, struct bio
*bio
)
2948 struct bio_track
*track
;
2950 KKASSERT(bio
->bio_buf
->b_cmd
!= BUF_CMD_DONE
);
2951 if (bio
->bio_buf
->b_cmd
== BUF_CMD_READ
)
2952 track
= &vp
->v_track_read
;
2954 track
= &vp
->v_track_write
;
2955 bio
->bio_track
= track
;
2956 atomic_add_int(&track
->bk_active
, 1);
2957 vop_strategy(*vp
->v_ops
, vp
, bio
);
2964 * Finish I/O on a buffer, optionally calling a completion function.
2965 * This is usually called from an interrupt so process blocking is
2968 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
2969 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
2970 * assuming B_INVAL is clear.
2972 * For the VMIO case, we set B_CACHE if the op was a read and no
2973 * read error occured, or if the op was a write. B_CACHE is never
2974 * set if the buffer is invalid or otherwise uncacheable.
2976 * biodone does not mess with B_INVAL, allowing the I/O routine or the
2977 * initiator to leave B_INVAL set to brelse the buffer out of existance
2978 * in the biodone routine.
2981 biodone(struct bio
*bio
)
2983 struct buf
*bp
= bio
->bio_buf
;
2988 KASSERT(BUF_REFCNTNB(bp
) > 0,
2989 ("biodone: bp %p not busy %d", bp
, BUF_REFCNTNB(bp
)));
2990 KASSERT(bp
->b_cmd
!= BUF_CMD_DONE
,
2991 ("biodone: bp %p already done!", bp
));
2993 runningbufwakeup(bp
);
2996 * Run up the chain of BIO's. Leave b_cmd intact for the duration.
2999 biodone_t
*done_func
;
3000 struct bio_track
*track
;
3003 * BIO tracking. Most but not all BIOs are tracked.
3005 if ((track
= bio
->bio_track
) != NULL
) {
3006 atomic_subtract_int(&track
->bk_active
, 1);
3007 if (track
->bk_active
< 0) {
3008 panic("biodone: bad active count bio %p\n",
3011 if (track
->bk_waitflag
) {
3012 track
->bk_waitflag
= 0;
3015 bio
->bio_track
= NULL
;
3019 * A bio_done function terminates the loop. The function
3020 * will be responsible for any further chaining and/or
3021 * buffer management.
3023 * WARNING! The done function can deallocate the buffer!
3025 if ((done_func
= bio
->bio_done
) != NULL
) {
3026 bio
->bio_done
= NULL
;
3031 bio
= bio
->bio_prev
;
3035 bp
->b_cmd
= BUF_CMD_DONE
;
3038 * Only reads and writes are processed past this point.
3040 if (cmd
!= BUF_CMD_READ
&& cmd
!= BUF_CMD_WRITE
) {
3047 * Warning: softupdates may re-dirty the buffer.
3049 if (LIST_FIRST(&bp
->b_dep
) != NULL
)
3052 if (bp
->b_flags
& B_VMIO
) {
3058 struct vnode
*vp
= bp
->b_vp
;
3062 #if defined(VFS_BIO_DEBUG)
3063 if (vp
->v_auxrefs
== 0)
3064 panic("biodone: zero vnode hold count");
3065 if ((vp
->v_flag
& VOBJBUF
) == 0)
3066 panic("biodone: vnode is not setup for merged cache");
3069 foff
= bp
->b_loffset
;
3070 KASSERT(foff
!= NOOFFSET
, ("biodone: no buffer offset"));
3071 KASSERT(obj
!= NULL
, ("biodone: missing VM object"));
3073 #if defined(VFS_BIO_DEBUG)
3074 if (obj
->paging_in_progress
< bp
->b_xio
.xio_npages
) {
3075 kprintf("biodone: paging in progress(%d) < bp->b_xio.xio_npages(%d)\n",
3076 obj
->paging_in_progress
, bp
->b_xio
.xio_npages
);
3081 * Set B_CACHE if the op was a normal read and no error
3082 * occured. B_CACHE is set for writes in the b*write()
3085 iosize
= bp
->b_bcount
- bp
->b_resid
;
3086 if (cmd
== BUF_CMD_READ
&& (bp
->b_flags
& (B_INVAL
|B_NOCACHE
|B_ERROR
)) == 0) {
3087 bp
->b_flags
|= B_CACHE
;
3090 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
3094 resid
= ((foff
+ PAGE_SIZE
) & ~(off_t
)PAGE_MASK
) - foff
;
3099 * cleanup bogus pages, restoring the originals. Since
3100 * the originals should still be wired, we don't have
3101 * to worry about interrupt/freeing races destroying
3102 * the VM object association.
3104 m
= bp
->b_xio
.xio_pages
[i
];
3105 if (m
== bogus_page
) {
3107 m
= vm_page_lookup(obj
, OFF_TO_IDX(foff
));
3109 panic("biodone: page disappeared");
3110 bp
->b_xio
.xio_pages
[i
] = m
;
3111 pmap_qenter(trunc_page((vm_offset_t
)bp
->b_data
),
3112 bp
->b_xio
.xio_pages
, bp
->b_xio
.xio_npages
);
3114 #if defined(VFS_BIO_DEBUG)
3115 if (OFF_TO_IDX(foff
) != m
->pindex
) {
3117 "biodone: foff(%lu)/m->pindex(%d) mismatch\n",
3118 (unsigned long)foff
, m
->pindex
);
3123 * In the write case, the valid and clean bits are
3124 * already changed correctly ( see bdwrite() ), so we
3125 * only need to do this here in the read case.
3127 if (cmd
== BUF_CMD_READ
&& !bogusflag
&& resid
> 0) {
3128 vfs_page_set_valid(bp
, foff
, i
, m
);
3130 vm_page_flag_clear(m
, PG_ZERO
);
3133 * when debugging new filesystems or buffer I/O methods, this
3134 * is the most common error that pops up. if you see this, you
3135 * have not set the page busy flag correctly!!!
3138 kprintf("biodone: page busy < 0, "
3139 "pindex: %d, foff: 0x(%x,%x), "
3140 "resid: %d, index: %d\n",
3141 (int) m
->pindex
, (int)(foff
>> 32),
3142 (int) foff
& 0xffffffff, resid
, i
);
3143 if (!vn_isdisk(vp
, NULL
))
3144 kprintf(" iosize: %ld, loffset: %lld, flags: 0x%08x, npages: %d\n",
3145 bp
->b_vp
->v_mount
->mnt_stat
.f_iosize
,
3147 bp
->b_flags
, bp
->b_xio
.xio_npages
);
3149 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n",
3151 bp
->b_flags
, bp
->b_xio
.xio_npages
);
3152 kprintf(" valid: 0x%x, dirty: 0x%x, wired: %d\n",
3153 m
->valid
, m
->dirty
, m
->wire_count
);
3154 panic("biodone: page busy < 0");
3156 vm_page_io_finish(m
);
3157 vm_object_pip_subtract(obj
, 1);
3158 foff
= (foff
+ PAGE_SIZE
) & ~(off_t
)PAGE_MASK
;
3162 vm_object_pip_wakeupn(obj
, 0);
3166 * For asynchronous completions, release the buffer now. The brelse
3167 * will do a wakeup there if necessary - so no need to do a wakeup
3168 * here in the async case. The sync case always needs to do a wakeup.
3171 if (bp
->b_flags
& B_ASYNC
) {
3172 if ((bp
->b_flags
& (B_NOCACHE
| B_INVAL
| B_ERROR
| B_RELBUF
)) != 0)
3185 * This routine is called in lieu of iodone in the case of
3186 * incomplete I/O. This keeps the busy status for pages
3190 vfs_unbusy_pages(struct buf
*bp
)
3194 runningbufwakeup(bp
);
3195 if (bp
->b_flags
& B_VMIO
) {
3196 struct vnode
*vp
= bp
->b_vp
;
3201 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
3202 vm_page_t m
= bp
->b_xio
.xio_pages
[i
];
3205 * When restoring bogus changes the original pages
3206 * should still be wired, so we are in no danger of
3207 * losing the object association and do not need
3208 * critical section protection particularly.
3210 if (m
== bogus_page
) {
3211 m
= vm_page_lookup(obj
, OFF_TO_IDX(bp
->b_loffset
) + i
);
3213 panic("vfs_unbusy_pages: page missing");
3215 bp
->b_xio
.xio_pages
[i
] = m
;
3216 pmap_qenter(trunc_page((vm_offset_t
)bp
->b_data
),
3217 bp
->b_xio
.xio_pages
, bp
->b_xio
.xio_npages
);
3219 vm_object_pip_subtract(obj
, 1);
3220 vm_page_flag_clear(m
, PG_ZERO
);
3221 vm_page_io_finish(m
);
3223 vm_object_pip_wakeupn(obj
, 0);
3228 * vfs_page_set_valid:
3230 * Set the valid bits in a page based on the supplied offset. The
3231 * range is restricted to the buffer's size.
3233 * This routine is typically called after a read completes.
3236 vfs_page_set_valid(struct buf
*bp
, vm_ooffset_t off
, int pageno
, vm_page_t m
)
3238 vm_ooffset_t soff
, eoff
;
3241 * Start and end offsets in buffer. eoff - soff may not cross a
3242 * page boundry or cross the end of the buffer. The end of the
3243 * buffer, in this case, is our file EOF, not the allocation size
3247 eoff
= (off
+ PAGE_SIZE
) & ~(off_t
)PAGE_MASK
;
3248 if (eoff
> bp
->b_loffset
+ bp
->b_bcount
)
3249 eoff
= bp
->b_loffset
+ bp
->b_bcount
;
3252 * Set valid range. This is typically the entire buffer and thus the
3256 vm_page_set_validclean(
3258 (vm_offset_t
) (soff
& PAGE_MASK
),
3259 (vm_offset_t
) (eoff
- soff
)
3267 * This routine is called before a device strategy routine.
3268 * It is used to tell the VM system that paging I/O is in
3269 * progress, and treat the pages associated with the buffer
3270 * almost as being PG_BUSY. Also the object 'paging_in_progress'
3271 * flag is handled to make sure that the object doesn't become
3274 * Since I/O has not been initiated yet, certain buffer flags
3275 * such as B_ERROR or B_INVAL may be in an inconsistant state
3276 * and should be ignored.
3279 vfs_busy_pages(struct vnode
*vp
, struct buf
*bp
)
3282 struct lwp
*lp
= curthread
->td_lwp
;
3285 * The buffer's I/O command must already be set. If reading,
3286 * B_CACHE must be 0 (double check against callers only doing
3287 * I/O when B_CACHE is 0).
3289 KKASSERT(bp
->b_cmd
!= BUF_CMD_DONE
);
3290 KKASSERT(bp
->b_cmd
== BUF_CMD_WRITE
|| (bp
->b_flags
& B_CACHE
) == 0);
3292 if (bp
->b_flags
& B_VMIO
) {
3297 foff
= bp
->b_loffset
;
3298 KASSERT(bp
->b_loffset
!= NOOFFSET
,
3299 ("vfs_busy_pages: no buffer offset"));
3303 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
3304 vm_page_t m
= bp
->b_xio
.xio_pages
[i
];
3305 if (vm_page_sleep_busy(m
, FALSE
, "vbpage"))
3310 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
3311 vm_page_t m
= bp
->b_xio
.xio_pages
[i
];
3313 vm_page_flag_clear(m
, PG_ZERO
);
3314 if ((bp
->b_flags
& B_CLUSTER
) == 0) {
3315 vm_object_pip_add(obj
, 1);
3316 vm_page_io_start(m
);
3320 * When readying a vnode-backed buffer for a write
3321 * we must zero-fill any invalid portions of the
3324 * When readying a vnode-backed buffer for a read
3325 * we must replace any dirty pages with a bogus
3326 * page so we do not destroy dirty data when
3327 * filling in gaps. Dirty pages might not
3328 * necessarily be marked dirty yet, so use m->valid
3329 * as a reasonable test.
3331 * Bogus page replacement is, uh, bogus. We need
3332 * to find a better way.
3334 vm_page_protect(m
, VM_PROT_NONE
);
3335 if (bp
->b_cmd
== BUF_CMD_WRITE
) {
3336 vfs_page_set_valid(bp
, foff
, i
, m
);
3337 } else if (m
->valid
== VM_PAGE_BITS_ALL
) {
3338 bp
->b_xio
.xio_pages
[i
] = bogus_page
;
3341 foff
= (foff
+ PAGE_SIZE
) & ~(off_t
)PAGE_MASK
;
3344 pmap_qenter(trunc_page((vm_offset_t
)bp
->b_data
),
3345 bp
->b_xio
.xio_pages
, bp
->b_xio
.xio_npages
);
3349 * This is the easiest place to put the process accounting for the I/O
3353 if (bp
->b_cmd
== BUF_CMD_READ
)
3354 lp
->lwp_ru
.ru_inblock
++;
3356 lp
->lwp_ru
.ru_oublock
++;
3363 * Tell the VM system that the pages associated with this buffer
3364 * are clean. This is used for delayed writes where the data is
3365 * going to go to disk eventually without additional VM intevention.
3367 * Note that while we only really need to clean through to b_bcount, we
3368 * just go ahead and clean through to b_bufsize.
3371 vfs_clean_pages(struct buf
*bp
)
3375 if (bp
->b_flags
& B_VMIO
) {
3378 foff
= bp
->b_loffset
;
3379 KASSERT(foff
!= NOOFFSET
, ("vfs_clean_pages: no buffer offset"));
3380 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
3381 vm_page_t m
= bp
->b_xio
.xio_pages
[i
];
3382 vm_ooffset_t noff
= (foff
+ PAGE_SIZE
) & ~(off_t
)PAGE_MASK
;
3383 vm_ooffset_t eoff
= noff
;
3385 if (eoff
> bp
->b_loffset
+ bp
->b_bufsize
)
3386 eoff
= bp
->b_loffset
+ bp
->b_bufsize
;
3387 vfs_page_set_valid(bp
, foff
, i
, m
);
3388 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3395 * vfs_bio_set_validclean:
3397 * Set the range within the buffer to valid and clean. The range is
3398 * relative to the beginning of the buffer, b_loffset. Note that
3399 * b_loffset itself may be offset from the beginning of the first page.
3403 vfs_bio_set_validclean(struct buf
*bp
, int base
, int size
)
3405 if (bp
->b_flags
& B_VMIO
) {
3410 * Fixup base to be relative to beginning of first page.
3411 * Set initial n to be the maximum number of bytes in the
3412 * first page that can be validated.
3415 base
+= (bp
->b_loffset
& PAGE_MASK
);
3416 n
= PAGE_SIZE
- (base
& PAGE_MASK
);
3418 for (i
= base
/ PAGE_SIZE
; size
> 0 && i
< bp
->b_xio
.xio_npages
; ++i
) {
3419 vm_page_t m
= bp
->b_xio
.xio_pages
[i
];
3424 vm_page_set_validclean(m
, base
& PAGE_MASK
, n
);
3435 * Clear a buffer. This routine essentially fakes an I/O, so we need
3436 * to clear B_ERROR and B_INVAL.
3438 * Note that while we only theoretically need to clear through b_bcount,
3439 * we go ahead and clear through b_bufsize.
3443 vfs_bio_clrbuf(struct buf
*bp
)
3447 if ((bp
->b_flags
& (B_VMIO
| B_MALLOC
)) == B_VMIO
) {
3448 bp
->b_flags
&= ~(B_INVAL
|B_ERROR
);
3449 if ((bp
->b_xio
.xio_npages
== 1) && (bp
->b_bufsize
< PAGE_SIZE
) &&
3450 (bp
->b_loffset
& PAGE_MASK
) == 0) {
3451 mask
= (1 << (bp
->b_bufsize
/ DEV_BSIZE
)) - 1;
3452 if ((bp
->b_xio
.xio_pages
[0]->valid
& mask
) == mask
) {
3456 if (((bp
->b_xio
.xio_pages
[0]->flags
& PG_ZERO
) == 0) &&
3457 ((bp
->b_xio
.xio_pages
[0]->valid
& mask
) == 0)) {
3458 bzero(bp
->b_data
, bp
->b_bufsize
);
3459 bp
->b_xio
.xio_pages
[0]->valid
|= mask
;
3464 ea
= sa
= bp
->b_data
;
3465 for(i
=0;i
<bp
->b_xio
.xio_npages
;i
++,sa
=ea
) {
3466 int j
= ((vm_offset_t
)sa
& PAGE_MASK
) / DEV_BSIZE
;
3467 ea
= (caddr_t
)trunc_page((vm_offset_t
)sa
+ PAGE_SIZE
);
3468 ea
= (caddr_t
)(vm_offset_t
)ulmin(
3469 (u_long
)(vm_offset_t
)ea
,
3470 (u_long
)(vm_offset_t
)bp
->b_data
+ bp
->b_bufsize
);
3471 mask
= ((1 << ((ea
- sa
) / DEV_BSIZE
)) - 1) << j
;
3472 if ((bp
->b_xio
.xio_pages
[i
]->valid
& mask
) == mask
)
3474 if ((bp
->b_xio
.xio_pages
[i
]->valid
& mask
) == 0) {
3475 if ((bp
->b_xio
.xio_pages
[i
]->flags
& PG_ZERO
) == 0) {
3479 for (; sa
< ea
; sa
+= DEV_BSIZE
, j
++) {
3480 if (((bp
->b_xio
.xio_pages
[i
]->flags
& PG_ZERO
) == 0) &&
3481 (bp
->b_xio
.xio_pages
[i
]->valid
& (1<<j
)) == 0)
3482 bzero(sa
, DEV_BSIZE
);
3485 bp
->b_xio
.xio_pages
[i
]->valid
|= mask
;
3486 vm_page_flag_clear(bp
->b_xio
.xio_pages
[i
], PG_ZERO
);
3495 * vm_hold_load_pages:
3497 * Load pages into the buffer's address space. The pages are
3498 * allocated from the kernel object in order to reduce interference
3499 * with the any VM paging I/O activity. The range of loaded
3500 * pages will be wired.
3502 * If a page cannot be allocated, the 'pagedaemon' is woken up to
3503 * retrieve the full range (to - from) of pages.
3507 vm_hold_load_pages(struct buf
*bp
, vm_offset_t from
, vm_offset_t to
)
3513 to
= round_page(to
);
3514 from
= round_page(from
);
3515 index
= (from
- trunc_page((vm_offset_t
)bp
->b_data
)) >> PAGE_SHIFT
;
3517 for (pg
= from
; pg
< to
; pg
+= PAGE_SIZE
, index
++) {
3522 * Note: must allocate system pages since blocking here
3523 * could intefere with paging I/O, no matter which
3526 p
= vm_page_alloc(&kernel_object
,
3528 VM_ALLOC_NORMAL
| VM_ALLOC_SYSTEM
);
3530 vm_pageout_deficit
+= (to
- from
) >> PAGE_SHIFT
;
3535 p
->valid
= VM_PAGE_BITS_ALL
;
3536 vm_page_flag_clear(p
, PG_ZERO
);
3537 pmap_kenter(pg
, VM_PAGE_TO_PHYS(p
));
3538 bp
->b_xio
.xio_pages
[index
] = p
;
3541 bp
->b_xio
.xio_npages
= index
;
3545 * vm_hold_free_pages:
3547 * Return pages associated with the buffer back to the VM system.
3549 * The range of pages underlying the buffer's address space will
3550 * be unmapped and un-wired.
3553 vm_hold_free_pages(struct buf
*bp
, vm_offset_t from
, vm_offset_t to
)
3557 int index
, newnpages
;
3559 from
= round_page(from
);
3560 to
= round_page(to
);
3561 newnpages
= index
= (from
- trunc_page((vm_offset_t
)bp
->b_data
)) >> PAGE_SHIFT
;
3563 for (pg
= from
; pg
< to
; pg
+= PAGE_SIZE
, index
++) {
3564 p
= bp
->b_xio
.xio_pages
[index
];
3565 if (p
&& (index
< bp
->b_xio
.xio_npages
)) {
3567 kprintf("vm_hold_free_pages: doffset: %lld, loffset: %lld\n",
3568 bp
->b_bio2
.bio_offset
, bp
->b_loffset
);
3570 bp
->b_xio
.xio_pages
[index
] = NULL
;
3573 vm_page_unwire(p
, 0);
3577 bp
->b_xio
.xio_npages
= newnpages
;
3583 * Map a user buffer into KVM via a pbuf. On return the buffer's
3584 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array
3588 vmapbuf(struct buf
*bp
, caddr_t udata
, int bytes
)
3599 * bp had better have a command and it better be a pbuf.
3601 KKASSERT(bp
->b_cmd
!= BUF_CMD_DONE
);
3602 KKASSERT(bp
->b_flags
& B_PAGING
);
3608 * Map the user data into KVM. Mappings have to be page-aligned.
3610 addr
= (caddr_t
)trunc_page((vm_offset_t
)udata
);
3613 vmprot
= VM_PROT_READ
;
3614 if (bp
->b_cmd
== BUF_CMD_READ
)
3615 vmprot
|= VM_PROT_WRITE
;
3617 while (addr
< udata
+ bytes
) {
3619 * Do the vm_fault if needed; do the copy-on-write thing
3620 * when reading stuff off device into memory.
3622 * vm_fault_page*() returns a held VM page.
3624 va
= (addr
>= udata
) ? (vm_offset_t
)addr
: (vm_offset_t
)udata
;
3625 va
= trunc_page(va
);
3627 m
= vm_fault_page_quick(va
, vmprot
, &error
);
3629 for (i
= 0; i
< pidx
; ++i
) {
3630 vm_page_unhold(bp
->b_xio
.xio_pages
[i
]);
3631 bp
->b_xio
.xio_pages
[i
] = NULL
;
3635 bp
->b_xio
.xio_pages
[pidx
] = m
;
3641 * Map the page array and set the buffer fields to point to
3642 * the mapped data buffer.
3644 if (pidx
> btoc(MAXPHYS
))
3645 panic("vmapbuf: mapped more than MAXPHYS");
3646 pmap_qenter((vm_offset_t
)bp
->b_kvabase
, bp
->b_xio
.xio_pages
, pidx
);
3648 bp
->b_xio
.xio_npages
= pidx
;
3649 bp
->b_data
= bp
->b_kvabase
+ ((int)(intptr_t)udata
& PAGE_MASK
);
3650 bp
->b_bcount
= bytes
;
3651 bp
->b_bufsize
= bytes
;
3658 * Free the io map PTEs associated with this IO operation.
3659 * We also invalidate the TLB entries and restore the original b_addr.
3662 vunmapbuf(struct buf
*bp
)
3667 KKASSERT(bp
->b_flags
& B_PAGING
);
3669 npages
= bp
->b_xio
.xio_npages
;
3670 pmap_qremove(trunc_page((vm_offset_t
)bp
->b_data
), npages
);
3671 for (pidx
= 0; pidx
< npages
; ++pidx
) {
3672 vm_page_unhold(bp
->b_xio
.xio_pages
[pidx
]);
3673 bp
->b_xio
.xio_pages
[pidx
] = NULL
;
3675 bp
->b_xio
.xio_npages
= 0;
3676 bp
->b_data
= bp
->b_kvabase
;
3680 * Scan all buffers in the system and issue the callback.
3683 scan_all_buffers(int (*callback
)(struct buf
*, void *), void *info
)
3689 for (n
= 0; n
< nbuf
; ++n
) {
3690 if ((error
= callback(&buf
[n
], info
)) < 0) {
3700 * print out statistics from the current status of the buffer pool
3701 * this can be toggeled by the system control option debug.syncprt
3710 int counts
[(MAXBSIZE
/ PAGE_SIZE
) + 1];
3711 static char *bname
[3] = { "LOCKED", "LRU", "AGE" };
3713 for (dp
= bufqueues
, i
= 0; dp
< &bufqueues
[3]; dp
++, i
++) {
3715 for (j
= 0; j
<= MAXBSIZE
/PAGE_SIZE
; j
++)
3718 TAILQ_FOREACH(bp
, dp
, b_freelist
) {
3719 counts
[bp
->b_bufsize
/PAGE_SIZE
]++;
3723 kprintf("%s: total-%d", bname
[i
], count
);
3724 for (j
= 0; j
<= MAXBSIZE
/PAGE_SIZE
; j
++)
3726 kprintf(", %d-%d", j
* PAGE_SIZE
, counts
[j
]);
3734 DB_SHOW_COMMAND(buffer
, db_show_buffer
)
3737 struct buf
*bp
= (struct buf
*)addr
;
3740 db_printf("usage: show buffer <addr>\n");
3744 db_printf("b_flags = 0x%b\n", (u_int
)bp
->b_flags
, PRINT_BUF_FLAGS
);
3745 db_printf("b_cmd = %d\n", bp
->b_cmd
);
3746 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, "
3747 "b_resid = %d\n, b_data = %p, "
3748 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n",
3749 bp
->b_error
, bp
->b_bufsize
, bp
->b_bcount
, bp
->b_resid
,
3750 bp
->b_data
, bp
->b_bio2
.bio_offset
, (bp
->b_bio2
.bio_next
? bp
->b_bio2
.bio_next
->bio_offset
: (off_t
)-1));
3751 if (bp
->b_xio
.xio_npages
) {
3753 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ",
3754 bp
->b_xio
.xio_npages
);
3755 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
3757 m
= bp
->b_xio
.xio_pages
[i
];
3758 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m
->object
,
3759 (u_long
)m
->pindex
, (u_long
)VM_PAGE_TO_PHYS(m
));
3760 if ((i
+ 1) < bp
->b_xio
.xio_npages
)