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 1.111 2008/07/08 03:34:27 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 #define BD_WAKE_SIZE 128
85 #define BD_WAKE_MASK (BD_WAKE_SIZE - 1)
87 TAILQ_HEAD(bqueues
, buf
) bufqueues
[BUFFER_QUEUES
];
89 static MALLOC_DEFINE(M_BIOBUF
, "BIO buffer", "BIO buffer");
91 struct buf
*buf
; /* buffer header pool */
93 static void vm_hold_free_pages(struct buf
*bp
, vm_offset_t from
,
95 static void vm_hold_load_pages(struct buf
*bp
, vm_offset_t from
,
97 static void vfs_page_set_valid(struct buf
*bp
, vm_ooffset_t off
,
98 int pageno
, vm_page_t m
);
99 static void vfs_clean_pages(struct buf
*bp
);
100 static void vfs_setdirty(struct buf
*bp
);
101 static void vfs_vmio_release(struct buf
*bp
);
102 static int flushbufqueues(bufq_type_t q
);
103 static vm_page_t
bio_page_alloc(vm_object_t obj
, vm_pindex_t pg
, int deficit
);
105 static void bd_signal(int totalspace
);
106 static void buf_daemon(void);
107 static void buf_daemon_hw(void);
110 * bogus page -- for I/O to/from partially complete buffers
111 * this is a temporary solution to the problem, but it is not
112 * really that bad. it would be better to split the buffer
113 * for input in the case of buffers partially already in memory,
114 * but the code is intricate enough already.
116 vm_page_t bogus_page
;
119 * These are all static, but make the ones we export globals so we do
120 * not need to use compiler magic.
122 int bufspace
, maxbufspace
,
123 bufmallocspace
, maxbufmallocspace
, lobufspace
, hibufspace
;
124 static int bufreusecnt
, bufdefragcnt
, buffreekvacnt
;
125 static int lorunningspace
, hirunningspace
, runningbufreq
;
126 int dirtybufspace
, dirtybufspacehw
, lodirtybufspace
, hidirtybufspace
;
127 int dirtybufcount
, dirtybufcounthw
;
128 int runningbufspace
, runningbufcount
;
129 static int getnewbufcalls
;
130 static int getnewbufrestarts
;
131 static int recoverbufcalls
;
132 static int needsbuffer
; /* locked by needsbuffer_spin */
133 static int bd_request
; /* locked by needsbuffer_spin */
134 static int bd_request_hw
; /* locked by needsbuffer_spin */
135 static u_int bd_wake_ary
[BD_WAKE_SIZE
];
136 static u_int bd_wake_index
;
137 static struct spinlock needsbuffer_spin
;
139 static struct thread
*bufdaemon_td
;
140 static struct thread
*bufdaemonhw_td
;
144 * Sysctls for operational control of the buffer cache.
146 SYSCTL_INT(_vfs
, OID_AUTO
, lodirtybufspace
, CTLFLAG_RW
, &lodirtybufspace
, 0,
147 "Number of dirty buffers to flush before bufdaemon becomes inactive");
148 SYSCTL_INT(_vfs
, OID_AUTO
, hidirtybufspace
, CTLFLAG_RW
, &hidirtybufspace
, 0,
149 "High watermark used to trigger explicit flushing of dirty buffers");
150 SYSCTL_INT(_vfs
, OID_AUTO
, lorunningspace
, CTLFLAG_RW
, &lorunningspace
, 0,
151 "Minimum amount of buffer space required for active I/O");
152 SYSCTL_INT(_vfs
, OID_AUTO
, hirunningspace
, CTLFLAG_RW
, &hirunningspace
, 0,
153 "Maximum amount of buffer space to usable for active I/O");
155 * Sysctls determining current state of the buffer cache.
157 SYSCTL_INT(_vfs
, OID_AUTO
, nbuf
, CTLFLAG_RD
, &nbuf
, 0,
158 "Total number of buffers in buffer cache");
159 SYSCTL_INT(_vfs
, OID_AUTO
, dirtybufspace
, CTLFLAG_RD
, &dirtybufspace
, 0,
160 "Pending bytes of dirty buffers (all)");
161 SYSCTL_INT(_vfs
, OID_AUTO
, dirtybufspacehw
, CTLFLAG_RD
, &dirtybufspacehw
, 0,
162 "Pending bytes of dirty buffers (heavy weight)");
163 SYSCTL_INT(_vfs
, OID_AUTO
, dirtybufcount
, CTLFLAG_RD
, &dirtybufcount
, 0,
164 "Pending number of dirty buffers");
165 SYSCTL_INT(_vfs
, OID_AUTO
, dirtybufcounthw
, CTLFLAG_RD
, &dirtybufcounthw
, 0,
166 "Pending number of dirty buffers (heavy weight)");
167 SYSCTL_INT(_vfs
, OID_AUTO
, runningbufspace
, CTLFLAG_RD
, &runningbufspace
, 0,
168 "I/O bytes currently in progress due to asynchronous writes");
169 SYSCTL_INT(_vfs
, OID_AUTO
, runningbufcount
, CTLFLAG_RD
, &runningbufcount
, 0,
170 "I/O buffers currently in progress due to asynchronous writes");
171 SYSCTL_INT(_vfs
, OID_AUTO
, maxbufspace
, CTLFLAG_RD
, &maxbufspace
, 0,
172 "Hard limit on maximum amount of memory usable for buffer space");
173 SYSCTL_INT(_vfs
, OID_AUTO
, hibufspace
, CTLFLAG_RD
, &hibufspace
, 0,
174 "Soft limit on maximum amount of memory usable for buffer space");
175 SYSCTL_INT(_vfs
, OID_AUTO
, lobufspace
, CTLFLAG_RD
, &lobufspace
, 0,
176 "Minimum amount of memory to reserve for system buffer space");
177 SYSCTL_INT(_vfs
, OID_AUTO
, bufspace
, CTLFLAG_RD
, &bufspace
, 0,
178 "Amount of memory available for buffers");
179 SYSCTL_INT(_vfs
, OID_AUTO
, maxmallocbufspace
, CTLFLAG_RD
, &maxbufmallocspace
,
180 0, "Maximum amount of memory reserved for buffers using malloc");
181 SYSCTL_INT(_vfs
, OID_AUTO
, bufmallocspace
, CTLFLAG_RD
, &bufmallocspace
, 0,
182 "Amount of memory left for buffers using malloc-scheme");
183 SYSCTL_INT(_vfs
, OID_AUTO
, getnewbufcalls
, CTLFLAG_RD
, &getnewbufcalls
, 0,
184 "New buffer header acquisition requests");
185 SYSCTL_INT(_vfs
, OID_AUTO
, getnewbufrestarts
, CTLFLAG_RD
, &getnewbufrestarts
,
186 0, "New buffer header acquisition restarts");
187 SYSCTL_INT(_vfs
, OID_AUTO
, recoverbufcalls
, CTLFLAG_RD
, &recoverbufcalls
, 0,
188 "Recover VM space in an emergency");
189 SYSCTL_INT(_vfs
, OID_AUTO
, bufdefragcnt
, CTLFLAG_RD
, &bufdefragcnt
, 0,
190 "Buffer acquisition restarts due to fragmented buffer map");
191 SYSCTL_INT(_vfs
, OID_AUTO
, buffreekvacnt
, CTLFLAG_RD
, &buffreekvacnt
, 0,
192 "Amount of time KVA space was deallocated in an arbitrary buffer");
193 SYSCTL_INT(_vfs
, OID_AUTO
, bufreusecnt
, CTLFLAG_RD
, &bufreusecnt
, 0,
194 "Amount of time buffer re-use operations were successful");
195 SYSCTL_INT(_debug_sizeof
, OID_AUTO
, buf
, CTLFLAG_RD
, 0, sizeof(struct buf
),
196 "sizeof(struct buf)");
198 char *buf_wmesg
= BUF_WMESG
;
200 extern int vm_swap_size
;
202 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
203 #define VFS_BIO_NEED_UNUSED02 0x02
204 #define VFS_BIO_NEED_UNUSED04 0x04
205 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
210 * Called when buffer space is potentially available for recovery.
211 * getnewbuf() will block on this flag when it is unable to free
212 * sufficient buffer space. Buffer space becomes recoverable when
213 * bp's get placed back in the queues.
220 * If someone is waiting for BUF space, wake them up. Even
221 * though we haven't freed the kva space yet, the waiting
222 * process will be able to now.
224 if (needsbuffer
& VFS_BIO_NEED_BUFSPACE
) {
225 spin_lock_wr(&needsbuffer_spin
);
226 needsbuffer
&= ~VFS_BIO_NEED_BUFSPACE
;
227 spin_unlock_wr(&needsbuffer_spin
);
228 wakeup(&needsbuffer
);
235 * Accounting for I/O in progress.
239 runningbufwakeup(struct buf
*bp
)
243 if ((totalspace
= bp
->b_runningbufspace
) != 0) {
244 runningbufspace
-= totalspace
;
246 bp
->b_runningbufspace
= 0;
247 if (runningbufreq
&& runningbufspace
<= lorunningspace
) {
249 wakeup(&runningbufreq
);
251 bd_signal(totalspace
);
258 * Called when a buffer has been added to one of the free queues to
259 * account for the buffer and to wakeup anyone waiting for free buffers.
260 * This typically occurs when large amounts of metadata are being handled
261 * by the buffer cache ( else buffer space runs out first, usually ).
268 spin_lock_wr(&needsbuffer_spin
);
269 needsbuffer
&= ~VFS_BIO_NEED_ANY
;
270 spin_unlock_wr(&needsbuffer_spin
);
271 wakeup(&needsbuffer
);
276 * waitrunningbufspace()
278 * Wait for the amount of running I/O to drop to a reasonable level.
280 * The caller may be using this function to block in a tight loop, we
281 * must block of runningbufspace is greater then the passed limit.
282 * And even with that it may not be enough, due to the presence of
283 * B_LOCKED dirty buffers, so also wait for at least one running buffer
287 waitrunningbufspace(int limit
)
291 if (lorunningspace
< limit
)
292 lorun
= lorunningspace
;
297 if (runningbufspace
> lorun
) {
298 while (runningbufspace
> lorun
) {
300 tsleep(&runningbufreq
, 0, "wdrain", 0);
302 } else if (runningbufspace
) {
304 tsleep(&runningbufreq
, 0, "wdrain2", 1);
310 * vfs_buf_test_cache:
312 * Called when a buffer is extended. This function clears the B_CACHE
313 * bit if the newly extended portion of the buffer does not contain
318 vfs_buf_test_cache(struct buf
*bp
,
319 vm_ooffset_t foff
, vm_offset_t off
, vm_offset_t size
,
322 if (bp
->b_flags
& B_CACHE
) {
323 int base
= (foff
+ off
) & PAGE_MASK
;
324 if (vm_page_is_valid(m
, base
, size
) == 0)
325 bp
->b_flags
&= ~B_CACHE
;
332 * Spank the buf_daemon[_hw] if the total dirty buffer space exceeds the
339 if (dirtybufspace
< lodirtybufspace
&& dirtybufcount
< nbuf
/ 2)
342 if (bd_request
== 0 &&
343 (dirtybufspace
- dirtybufspacehw
> lodirtybufspace
/ 2 ||
344 dirtybufcount
- dirtybufcounthw
>= nbuf
/ 2)) {
345 spin_lock_wr(&needsbuffer_spin
);
347 spin_unlock_wr(&needsbuffer_spin
);
350 if (bd_request_hw
== 0 &&
351 (dirtybufspacehw
> lodirtybufspace
/ 2 ||
352 dirtybufcounthw
>= nbuf
/ 2)) {
353 spin_lock_wr(&needsbuffer_spin
);
355 spin_unlock_wr(&needsbuffer_spin
);
356 wakeup(&bd_request_hw
);
363 * Get the buf_daemon heated up when the number of running and dirty
364 * buffers exceeds the mid-point.
373 mid1
= lodirtybufspace
+ (hidirtybufspace
- lodirtybufspace
) / 2;
375 totalspace
= runningbufspace
+ dirtybufspace
;
376 if (totalspace
>= mid1
|| dirtybufcount
>= nbuf
/ 2) {
378 mid2
= mid1
+ (hidirtybufspace
- mid1
) / 2;
379 if (totalspace
>= mid2
)
380 return(totalspace
- mid2
);
388 * Wait for the buffer cache to flush (totalspace) bytes worth of
389 * buffers, then return.
391 * Regardless this function blocks while the number of dirty buffers
392 * exceeds hidirtybufspace.
395 bd_wait(int totalspace
)
400 if (curthread
== bufdaemonhw_td
|| curthread
== bufdaemon_td
)
403 while (totalspace
> 0) {
406 if (totalspace
> runningbufspace
+ dirtybufspace
)
407 totalspace
= runningbufspace
+ dirtybufspace
;
408 count
= totalspace
/ BKVASIZE
;
409 if (count
>= BD_WAKE_SIZE
)
410 count
= BD_WAKE_SIZE
- 1;
411 i
= (bd_wake_index
+ count
) & BD_WAKE_MASK
;
413 tsleep(&bd_wake_ary
[i
], 0, "flstik", hz
);
416 totalspace
= runningbufspace
+ dirtybufspace
- hidirtybufspace
;
423 * This function is called whenever runningbufspace or dirtybufspace
424 * is reduced. Track threads waiting for run+dirty buffer I/O
428 bd_signal(int totalspace
)
432 while (totalspace
> 0) {
433 i
= atomic_fetchadd_int(&bd_wake_index
, 1);
435 if (bd_wake_ary
[i
]) {
437 wakeup(&bd_wake_ary
[i
]);
439 totalspace
-= BKVASIZE
;
446 * Load time initialisation of the buffer cache, called from machine
447 * dependant initialization code.
453 vm_offset_t bogus_offset
;
456 spin_init(&needsbuffer_spin
);
458 /* next, make a null set of free lists */
459 for (i
= 0; i
< BUFFER_QUEUES
; i
++)
460 TAILQ_INIT(&bufqueues
[i
]);
462 /* finally, initialize each buffer header and stick on empty q */
463 for (i
= 0; i
< nbuf
; i
++) {
465 bzero(bp
, sizeof *bp
);
466 bp
->b_flags
= B_INVAL
; /* we're just an empty header */
467 bp
->b_cmd
= BUF_CMD_DONE
;
468 bp
->b_qindex
= BQUEUE_EMPTY
;
470 xio_init(&bp
->b_xio
);
473 TAILQ_INSERT_TAIL(&bufqueues
[BQUEUE_EMPTY
], bp
, b_freelist
);
477 * maxbufspace is the absolute maximum amount of buffer space we are
478 * allowed to reserve in KVM and in real terms. The absolute maximum
479 * is nominally used by buf_daemon. hibufspace is the nominal maximum
480 * used by most other processes. The differential is required to
481 * ensure that buf_daemon is able to run when other processes might
482 * be blocked waiting for buffer space.
484 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
485 * this may result in KVM fragmentation which is not handled optimally
488 maxbufspace
= nbuf
* BKVASIZE
;
489 hibufspace
= imax(3 * maxbufspace
/ 4, maxbufspace
- MAXBSIZE
* 10);
490 lobufspace
= hibufspace
- MAXBSIZE
;
492 lorunningspace
= 512 * 1024;
493 hirunningspace
= 1024 * 1024;
496 * Limit the amount of malloc memory since it is wired permanently
497 * into the kernel space. Even though this is accounted for in
498 * the buffer allocation, we don't want the malloced region to grow
499 * uncontrolled. The malloc scheme improves memory utilization
500 * significantly on average (small) directories.
502 maxbufmallocspace
= hibufspace
/ 20;
505 * Reduce the chance of a deadlock occuring by limiting the number
506 * of delayed-write dirty buffers we allow to stack up.
508 hidirtybufspace
= hibufspace
/ 2;
512 lodirtybufspace
= hidirtybufspace
/ 2;
515 * Maximum number of async ops initiated per buf_daemon loop. This is
516 * somewhat of a hack at the moment, we really need to limit ourselves
517 * based on the number of bytes of I/O in-transit that were initiated
521 bogus_offset
= kmem_alloc_pageable(&kernel_map
, PAGE_SIZE
);
522 bogus_page
= vm_page_alloc(&kernel_object
,
523 (bogus_offset
>> PAGE_SHIFT
),
525 vmstats
.v_wire_count
++;
530 * Initialize the embedded bio structures
533 initbufbio(struct buf
*bp
)
535 bp
->b_bio1
.bio_buf
= bp
;
536 bp
->b_bio1
.bio_prev
= NULL
;
537 bp
->b_bio1
.bio_offset
= NOOFFSET
;
538 bp
->b_bio1
.bio_next
= &bp
->b_bio2
;
539 bp
->b_bio1
.bio_done
= NULL
;
541 bp
->b_bio2
.bio_buf
= bp
;
542 bp
->b_bio2
.bio_prev
= &bp
->b_bio1
;
543 bp
->b_bio2
.bio_offset
= NOOFFSET
;
544 bp
->b_bio2
.bio_next
= NULL
;
545 bp
->b_bio2
.bio_done
= NULL
;
549 * Reinitialize the embedded bio structures as well as any additional
550 * translation cache layers.
553 reinitbufbio(struct buf
*bp
)
557 for (bio
= &bp
->b_bio1
; bio
; bio
= bio
->bio_next
) {
558 bio
->bio_done
= NULL
;
559 bio
->bio_offset
= NOOFFSET
;
564 * Push another BIO layer onto an existing BIO and return it. The new
565 * BIO layer may already exist, holding cached translation data.
568 push_bio(struct bio
*bio
)
572 if ((nbio
= bio
->bio_next
) == NULL
) {
573 int index
= bio
- &bio
->bio_buf
->b_bio_array
[0];
574 if (index
>= NBUF_BIO
- 1) {
575 panic("push_bio: too many layers bp %p\n",
578 nbio
= &bio
->bio_buf
->b_bio_array
[index
+ 1];
579 bio
->bio_next
= nbio
;
580 nbio
->bio_prev
= bio
;
581 nbio
->bio_buf
= bio
->bio_buf
;
582 nbio
->bio_offset
= NOOFFSET
;
583 nbio
->bio_done
= NULL
;
584 nbio
->bio_next
= NULL
;
586 KKASSERT(nbio
->bio_done
== NULL
);
591 pop_bio(struct bio
*bio
)
597 clearbiocache(struct bio
*bio
)
600 bio
->bio_offset
= NOOFFSET
;
608 * Free the KVA allocation for buffer 'bp'.
610 * Must be called from a critical section as this is the only locking for
613 * Since this call frees up buffer space, we call bufspacewakeup().
616 bfreekva(struct buf
*bp
)
622 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
623 vm_map_lock(&buffer_map
);
624 bufspace
-= bp
->b_kvasize
;
625 vm_map_delete(&buffer_map
,
626 (vm_offset_t
) bp
->b_kvabase
,
627 (vm_offset_t
) bp
->b_kvabase
+ bp
->b_kvasize
,
630 vm_map_unlock(&buffer_map
);
631 vm_map_entry_release(count
);
640 * Remove the buffer from the appropriate free list.
643 bremfree(struct buf
*bp
)
648 old_qindex
= bp
->b_qindex
;
650 if (bp
->b_qindex
!= BQUEUE_NONE
) {
651 KASSERT(BUF_REFCNTNB(bp
) == 1,
652 ("bremfree: bp %p not locked",bp
));
653 TAILQ_REMOVE(&bufqueues
[bp
->b_qindex
], bp
, b_freelist
);
654 bp
->b_qindex
= BQUEUE_NONE
;
656 if (BUF_REFCNTNB(bp
) <= 1)
657 panic("bremfree: removing a buffer not on a queue");
667 * Get a buffer with the specified data. Look in the cache first. We
668 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
669 * is set, the buffer is valid and we do not have to do anything ( see
673 bread(struct vnode
*vp
, off_t loffset
, int size
, struct buf
**bpp
)
677 bp
= getblk(vp
, loffset
, size
, 0, 0);
680 /* if not found in cache, do some I/O */
681 if ((bp
->b_flags
& B_CACHE
) == 0) {
682 KASSERT(!(bp
->b_flags
& B_ASYNC
), ("bread: illegal async bp %p", bp
));
683 bp
->b_flags
&= ~(B_ERROR
| B_INVAL
);
684 bp
->b_cmd
= BUF_CMD_READ
;
685 vfs_busy_pages(vp
, bp
);
686 vn_strategy(vp
, &bp
->b_bio1
);
687 return (biowait(bp
));
695 * Operates like bread, but also starts asynchronous I/O on
696 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior
697 * to initiating I/O . If B_CACHE is set, the buffer is valid
698 * and we do not have to do anything.
701 breadn(struct vnode
*vp
, off_t loffset
, int size
, off_t
*raoffset
,
702 int *rabsize
, int cnt
, struct buf
**bpp
)
704 struct buf
*bp
, *rabp
;
706 int rv
= 0, readwait
= 0;
708 *bpp
= bp
= getblk(vp
, loffset
, size
, 0, 0);
710 /* if not found in cache, do some I/O */
711 if ((bp
->b_flags
& B_CACHE
) == 0) {
712 bp
->b_flags
&= ~(B_ERROR
| B_INVAL
);
713 bp
->b_cmd
= BUF_CMD_READ
;
714 vfs_busy_pages(vp
, bp
);
715 vn_strategy(vp
, &bp
->b_bio1
);
719 for (i
= 0; i
< cnt
; i
++, raoffset
++, rabsize
++) {
720 if (inmem(vp
, *raoffset
))
722 rabp
= getblk(vp
, *raoffset
, *rabsize
, 0, 0);
724 if ((rabp
->b_flags
& B_CACHE
) == 0) {
725 rabp
->b_flags
|= B_ASYNC
;
726 rabp
->b_flags
&= ~(B_ERROR
| B_INVAL
);
727 rabp
->b_cmd
= BUF_CMD_READ
;
728 vfs_busy_pages(vp
, rabp
);
730 vn_strategy(vp
, &rabp
->b_bio1
);
745 * Write, release buffer on completion. (Done by iodone
746 * if async). Do not bother writing anything if the buffer
749 * Note that we set B_CACHE here, indicating that buffer is
750 * fully valid and thus cacheable. This is true even of NFS
751 * now so we set it generally. This could be set either here
752 * or in biodone() since the I/O is synchronous. We put it
756 bwrite(struct buf
*bp
)
760 if (bp
->b_flags
& B_INVAL
) {
765 oldflags
= bp
->b_flags
;
767 if (BUF_REFCNTNB(bp
) == 0)
768 panic("bwrite: buffer is not busy???");
771 /* Mark the buffer clean */
774 bp
->b_flags
&= ~B_ERROR
;
775 bp
->b_flags
|= B_CACHE
;
776 bp
->b_cmd
= BUF_CMD_WRITE
;
777 vfs_busy_pages(bp
->b_vp
, bp
);
780 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
781 * valid for vnode-backed buffers.
783 bp
->b_runningbufspace
= bp
->b_bufsize
;
784 if (bp
->b_runningbufspace
) {
785 runningbufspace
+= bp
->b_runningbufspace
;
790 if (oldflags
& B_ASYNC
)
792 vn_strategy(bp
->b_vp
, &bp
->b_bio1
);
794 if ((oldflags
& B_ASYNC
) == 0) {
795 int rtval
= biowait(bp
);
805 * Delayed write. (Buffer is marked dirty). Do not bother writing
806 * anything if the buffer is marked invalid.
808 * Note that since the buffer must be completely valid, we can safely
809 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
810 * biodone() in order to prevent getblk from writing the buffer
814 bdwrite(struct buf
*bp
)
816 if (BUF_REFCNTNB(bp
) == 0)
817 panic("bdwrite: buffer is not busy");
819 if (bp
->b_flags
& B_INVAL
) {
826 * Set B_CACHE, indicating that the buffer is fully valid. This is
827 * true even of NFS now.
829 bp
->b_flags
|= B_CACHE
;
832 * This bmap keeps the system from needing to do the bmap later,
833 * perhaps when the system is attempting to do a sync. Since it
834 * is likely that the indirect block -- or whatever other datastructure
835 * that the filesystem needs is still in memory now, it is a good
836 * thing to do this. Note also, that if the pageout daemon is
837 * requesting a sync -- there might not be enough memory to do
838 * the bmap then... So, this is important to do.
840 if (bp
->b_bio2
.bio_offset
== NOOFFSET
) {
841 VOP_BMAP(bp
->b_vp
, bp
->b_loffset
, &bp
->b_bio2
.bio_offset
,
842 NULL
, NULL
, BUF_CMD_WRITE
);
846 * Set the *dirty* buffer range based upon the VM system dirty pages.
851 * We need to do this here to satisfy the vnode_pager and the
852 * pageout daemon, so that it thinks that the pages have been
853 * "cleaned". Note that since the pages are in a delayed write
854 * buffer -- the VFS layer "will" see that the pages get written
855 * out on the next sync, or perhaps the cluster will be completed.
861 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
862 * due to the softdep code.
869 * Turn buffer into delayed write request by marking it B_DELWRI.
870 * B_RELBUF and B_NOCACHE must be cleared.
872 * We reassign the buffer to itself to properly update it in the
875 * Must be called from a critical section.
876 * The buffer must be on BQUEUE_NONE.
879 bdirty(struct buf
*bp
)
881 KASSERT(bp
->b_qindex
== BQUEUE_NONE
, ("bdirty: buffer %p still on queue %d", bp
, bp
->b_qindex
));
882 if (bp
->b_flags
& B_NOCACHE
) {
883 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp
);
884 bp
->b_flags
&= ~B_NOCACHE
;
886 if (bp
->b_flags
& B_INVAL
) {
887 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp
);
889 bp
->b_flags
&= ~B_RELBUF
;
891 if ((bp
->b_flags
& B_DELWRI
) == 0) {
892 bp
->b_flags
|= B_DELWRI
;
895 dirtybufspace
+= bp
->b_bufsize
;
896 if (bp
->b_flags
& B_HEAVY
) {
898 dirtybufspacehw
+= bp
->b_bufsize
;
905 * Set B_HEAVY, indicating that this is a heavy-weight buffer that
906 * needs to be flushed with a different buf_daemon thread to avoid
907 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf().
910 bheavy(struct buf
*bp
)
912 if ((bp
->b_flags
& B_HEAVY
) == 0) {
913 bp
->b_flags
|= B_HEAVY
;
914 if (bp
->b_flags
& B_DELWRI
) {
916 dirtybufspacehw
+= bp
->b_bufsize
;
924 * Clear B_DELWRI for buffer.
926 * Must be called from a critical section.
928 * The buffer is typically on BQUEUE_NONE but there is one case in
929 * brelse() that calls this function after placing the buffer on
934 bundirty(struct buf
*bp
)
936 if (bp
->b_flags
& B_DELWRI
) {
937 bp
->b_flags
&= ~B_DELWRI
;
940 dirtybufspace
-= bp
->b_bufsize
;
941 if (bp
->b_flags
& B_HEAVY
) {
943 dirtybufspacehw
-= bp
->b_bufsize
;
945 bd_signal(bp
->b_bufsize
);
948 * Since it is now being written, we can clear its deferred write flag.
950 bp
->b_flags
&= ~B_DEFERRED
;
956 * Asynchronous write. Start output on a buffer, but do not wait for
957 * it to complete. The buffer is released when the output completes.
959 * bwrite() ( or the VOP routine anyway ) is responsible for handling
960 * B_INVAL buffers. Not us.
963 bawrite(struct buf
*bp
)
965 bp
->b_flags
|= B_ASYNC
;
972 * Ordered write. Start output on a buffer, and flag it so that the
973 * device will write it in the order it was queued. The buffer is
974 * released when the output completes. bwrite() ( or the VOP routine
975 * anyway ) is responsible for handling B_INVAL buffers.
978 bowrite(struct buf
*bp
)
980 bp
->b_flags
|= B_ORDERED
| B_ASYNC
;
985 * buf_dirty_count_severe:
987 * Return true if we have too many dirty buffers.
990 buf_dirty_count_severe(void)
992 return (runningbufspace
+ dirtybufspace
>= hidirtybufspace
||
993 dirtybufcount
>= nbuf
/ 2);
999 * Release a busy buffer and, if requested, free its resources. The
1000 * buffer will be stashed in the appropriate bufqueue[] allowing it
1001 * to be accessed later as a cache entity or reused for other purposes.
1004 brelse(struct buf
*bp
)
1007 int saved_flags
= bp
->b_flags
;
1010 KASSERT(!(bp
->b_flags
& (B_CLUSTER
|B_PAGING
)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp
));
1015 * If B_NOCACHE is set we are being asked to destroy the buffer and
1016 * its backing store. Clear B_DELWRI.
1018 * B_NOCACHE is set in two cases: (1) when the caller really wants
1019 * to destroy the buffer and backing store and (2) when the caller
1020 * wants to destroy the buffer and backing store after a write
1023 if ((bp
->b_flags
& (B_NOCACHE
|B_DELWRI
)) == (B_NOCACHE
|B_DELWRI
)) {
1027 if (bp
->b_flags
& B_LOCKED
)
1028 bp
->b_flags
&= ~B_ERROR
;
1031 * If a write error occurs and the caller does not want to throw
1032 * away the buffer, redirty the buffer. This will also clear
1035 if (bp
->b_cmd
== BUF_CMD_WRITE
&&
1036 (bp
->b_flags
& (B_ERROR
| B_INVAL
)) == B_ERROR
) {
1038 * Failed write, redirty. Must clear B_ERROR to prevent
1039 * pages from being scrapped. If B_INVAL is set then
1040 * this case is not run and the next case is run to
1041 * destroy the buffer. B_INVAL can occur if the buffer
1042 * is outside the range supported by the underlying device.
1044 bp
->b_flags
&= ~B_ERROR
;
1046 } else if ((bp
->b_flags
& (B_NOCACHE
| B_INVAL
| B_ERROR
)) ||
1047 (bp
->b_bufsize
<= 0) || bp
->b_cmd
== BUF_CMD_FREEBLKS
) {
1049 * Either a failed I/O or we were asked to free or not
1052 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the
1053 * buffer cannot be immediately freed.
1055 bp
->b_flags
|= B_INVAL
;
1056 if (LIST_FIRST(&bp
->b_dep
) != NULL
)
1058 if (bp
->b_flags
& B_DELWRI
) {
1060 dirtybufspace
-= bp
->b_bufsize
;
1061 if (bp
->b_flags
& B_HEAVY
) {
1063 dirtybufspacehw
-= bp
->b_bufsize
;
1065 bd_signal(bp
->b_bufsize
);
1067 bp
->b_flags
&= ~(B_DELWRI
| B_CACHE
);
1071 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set.
1072 * If vfs_vmio_release() is called with either bit set, the
1073 * underlying pages may wind up getting freed causing a previous
1074 * write (bdwrite()) to get 'lost' because pages associated with
1075 * a B_DELWRI bp are marked clean. Pages associated with a
1076 * B_LOCKED buffer may be mapped by the filesystem.
1078 * If we want to release the buffer ourselves (rather then the
1079 * originator asking us to release it), give the originator a
1080 * chance to countermand the release by setting B_LOCKED.
1082 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1083 * if B_DELWRI is set.
1085 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1086 * on pages to return pages to the VM page queues.
1088 if (bp
->b_flags
& (B_DELWRI
| B_LOCKED
)) {
1089 bp
->b_flags
&= ~B_RELBUF
;
1090 } else if (vm_page_count_severe()) {
1091 if (LIST_FIRST(&bp
->b_dep
) != NULL
)
1093 if (bp
->b_flags
& (B_DELWRI
| B_LOCKED
))
1094 bp
->b_flags
&= ~B_RELBUF
;
1096 bp
->b_flags
|= B_RELBUF
;
1100 * At this point destroying the buffer is governed by the B_INVAL
1101 * or B_RELBUF flags.
1103 bp
->b_cmd
= BUF_CMD_DONE
;
1106 * VMIO buffer rundown. Make sure the VM page array is restored
1107 * after an I/O may have replaces some of the pages with bogus pages
1108 * in order to not destroy dirty pages in a fill-in read.
1110 * Note that due to the code above, if a buffer is marked B_DELWRI
1111 * then the B_RELBUF and B_NOCACHE bits will always be clear.
1112 * B_INVAL may still be set, however.
1114 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer
1115 * but not the backing store. B_NOCACHE will destroy the backing
1118 * Note that dirty NFS buffers contain byte-granular write ranges
1119 * and should not be destroyed w/ B_INVAL even if the backing store
1122 if (bp
->b_flags
& B_VMIO
) {
1124 * Rundown for VMIO buffers which are not dirty NFS buffers.
1136 * Get the base offset and length of the buffer. Note that
1137 * in the VMIO case if the buffer block size is not
1138 * page-aligned then b_data pointer may not be page-aligned.
1139 * But our b_xio.xio_pages array *IS* page aligned.
1141 * block sizes less then DEV_BSIZE (usually 512) are not
1142 * supported due to the page granularity bits (m->valid,
1143 * m->dirty, etc...).
1145 * See man buf(9) for more information
1148 resid
= bp
->b_bufsize
;
1149 foff
= bp
->b_loffset
;
1151 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
1152 m
= bp
->b_xio
.xio_pages
[i
];
1153 vm_page_flag_clear(m
, PG_ZERO
);
1155 * If we hit a bogus page, fixup *all* of them
1156 * now. Note that we left these pages wired
1157 * when we removed them so they had better exist,
1158 * and they cannot be ripped out from under us so
1159 * no critical section protection is necessary.
1161 if (m
== bogus_page
) {
1163 poff
= OFF_TO_IDX(bp
->b_loffset
);
1165 for (j
= i
; j
< bp
->b_xio
.xio_npages
; j
++) {
1168 mtmp
= bp
->b_xio
.xio_pages
[j
];
1169 if (mtmp
== bogus_page
) {
1170 mtmp
= vm_page_lookup(obj
, poff
+ j
);
1172 panic("brelse: page missing");
1174 bp
->b_xio
.xio_pages
[j
] = mtmp
;
1178 if ((bp
->b_flags
& B_INVAL
) == 0) {
1179 pmap_qenter(trunc_page((vm_offset_t
)bp
->b_data
),
1180 bp
->b_xio
.xio_pages
, bp
->b_xio
.xio_npages
);
1182 m
= bp
->b_xio
.xio_pages
[i
];
1186 * Invalidate the backing store if B_NOCACHE is set
1187 * (e.g. used with vinvalbuf()). If this is NFS
1188 * we impose a requirement that the block size be
1189 * a multiple of PAGE_SIZE and create a temporary
1190 * hack to basically invalidate the whole page. The
1191 * problem is that NFS uses really odd buffer sizes
1192 * especially when tracking piecemeal writes and
1193 * it also vinvalbuf()'s a lot, which would result
1194 * in only partial page validation and invalidation
1195 * here. If the file page is mmap()'d, however,
1196 * all the valid bits get set so after we invalidate
1197 * here we would end up with weird m->valid values
1198 * like 0xfc. nfs_getpages() can't handle this so
1199 * we clear all the valid bits for the NFS case
1200 * instead of just some of them.
1202 * The real bug is the VM system having to set m->valid
1203 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1204 * itself is an artifact of the whole 512-byte
1205 * granular mess that exists to support odd block
1206 * sizes and UFS meta-data block sizes (e.g. 6144).
1207 * A complete rewrite is required.
1209 if (bp
->b_flags
& (B_NOCACHE
|B_ERROR
)) {
1210 int poffset
= foff
& PAGE_MASK
;
1213 presid
= PAGE_SIZE
- poffset
;
1214 if (bp
->b_vp
->v_tag
== VT_NFS
&&
1215 bp
->b_vp
->v_type
== VREG
) {
1217 } else if (presid
> resid
) {
1220 KASSERT(presid
>= 0, ("brelse: extra page"));
1221 vm_page_set_invalid(m
, poffset
, presid
);
1223 resid
-= PAGE_SIZE
- (foff
& PAGE_MASK
);
1224 foff
= (foff
+ PAGE_SIZE
) & ~(off_t
)PAGE_MASK
;
1226 if (bp
->b_flags
& (B_INVAL
| B_RELBUF
))
1227 vfs_vmio_release(bp
);
1230 * Rundown for non-VMIO buffers.
1232 if (bp
->b_flags
& (B_INVAL
| B_RELBUF
)) {
1235 kprintf("brelse bp %p %08x/%08x: Warning, caught and fixed brelvp bug\n", bp
, saved_flags
, bp
->b_flags
);
1239 KKASSERT (LIST_FIRST(&bp
->b_dep
) == NULL
);
1245 if (bp
->b_qindex
!= BQUEUE_NONE
)
1246 panic("brelse: free buffer onto another queue???");
1247 if (BUF_REFCNTNB(bp
) > 1) {
1248 /* Temporary panic to verify exclusive locking */
1249 /* This panic goes away when we allow shared refs */
1250 panic("brelse: multiple refs");
1251 /* do not release to free list */
1258 * Figure out the correct queue to place the cleaned up buffer on.
1259 * Buffers placed in the EMPTY or EMPTYKVA had better already be
1260 * disassociated from their vnode.
1262 if (bp
->b_flags
& B_LOCKED
) {
1264 * Buffers that are locked are placed in the locked queue
1265 * immediately, regardless of their state.
1267 bp
->b_qindex
= BQUEUE_LOCKED
;
1268 TAILQ_INSERT_TAIL(&bufqueues
[BQUEUE_LOCKED
], bp
, b_freelist
);
1269 } else if (bp
->b_bufsize
== 0) {
1271 * Buffers with no memory. Due to conditionals near the top
1272 * of brelse() such buffers should probably already be
1273 * marked B_INVAL and disassociated from their vnode.
1275 bp
->b_flags
|= B_INVAL
;
1276 KASSERT(bp
->b_vp
== NULL
, ("bp1 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp
, saved_flags
, bp
->b_flags
, bp
->b_vp
));
1277 KKASSERT((bp
->b_flags
& B_HASHED
) == 0);
1278 if (bp
->b_kvasize
) {
1279 bp
->b_qindex
= BQUEUE_EMPTYKVA
;
1281 bp
->b_qindex
= BQUEUE_EMPTY
;
1283 TAILQ_INSERT_HEAD(&bufqueues
[bp
->b_qindex
], bp
, b_freelist
);
1284 } else if (bp
->b_flags
& (B_ERROR
| B_INVAL
| B_NOCACHE
| B_RELBUF
)) {
1286 * Buffers with junk contents. Again these buffers had better
1287 * already be disassociated from their vnode.
1289 KASSERT(bp
->b_vp
== NULL
, ("bp2 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp
, saved_flags
, bp
->b_flags
, bp
->b_vp
));
1290 KKASSERT((bp
->b_flags
& B_HASHED
) == 0);
1291 bp
->b_flags
|= B_INVAL
;
1292 bp
->b_qindex
= BQUEUE_CLEAN
;
1293 TAILQ_INSERT_HEAD(&bufqueues
[BQUEUE_CLEAN
], bp
, b_freelist
);
1296 * Remaining buffers. These buffers are still associated with
1299 switch(bp
->b_flags
& (B_DELWRI
|B_HEAVY
)) {
1301 bp
->b_qindex
= BQUEUE_DIRTY
;
1302 TAILQ_INSERT_TAIL(&bufqueues
[BQUEUE_DIRTY
], bp
, b_freelist
);
1304 case B_DELWRI
| B_HEAVY
:
1305 bp
->b_qindex
= BQUEUE_DIRTY_HW
;
1306 TAILQ_INSERT_TAIL(&bufqueues
[BQUEUE_DIRTY_HW
], bp
,
1311 * NOTE: Buffers are always placed at the end of the
1312 * queue. If B_AGE is not set the buffer will cycle
1313 * through the queue twice.
1315 bp
->b_qindex
= BQUEUE_CLEAN
;
1316 TAILQ_INSERT_TAIL(&bufqueues
[BQUEUE_CLEAN
], bp
, b_freelist
);
1322 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1323 * on the correct queue.
1325 if ((bp
->b_flags
& (B_INVAL
|B_DELWRI
)) == (B_INVAL
|B_DELWRI
))
1329 * The bp is on an appropriate queue unless locked. If it is not
1330 * locked or dirty we can wakeup threads waiting for buffer space.
1332 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1333 * if B_INVAL is set ).
1335 if ((bp
->b_flags
& (B_LOCKED
|B_DELWRI
)) == 0)
1339 * Something we can maybe free or reuse
1341 if (bp
->b_bufsize
|| bp
->b_kvasize
)
1345 * Clean up temporary flags and unlock the buffer.
1347 bp
->b_flags
&= ~(B_ORDERED
| B_ASYNC
| B_NOCACHE
| B_RELBUF
| B_DIRECT
);
1355 * Release a buffer back to the appropriate queue but do not try to free
1356 * it. The buffer is expected to be used again soon.
1358 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1359 * biodone() to requeue an async I/O on completion. It is also used when
1360 * known good buffers need to be requeued but we think we may need the data
1363 * XXX we should be able to leave the B_RELBUF hint set on completion.
1366 bqrelse(struct buf
*bp
)
1370 KASSERT(!(bp
->b_flags
& (B_CLUSTER
|B_PAGING
)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp
));
1372 if (bp
->b_qindex
!= BQUEUE_NONE
)
1373 panic("bqrelse: free buffer onto another queue???");
1374 if (BUF_REFCNTNB(bp
) > 1) {
1375 /* do not release to free list */
1376 panic("bqrelse: multiple refs");
1381 if (bp
->b_flags
& B_LOCKED
) {
1383 * Locked buffers are released to the locked queue. However,
1384 * if the buffer is dirty it will first go into the dirty
1385 * queue and later on after the I/O completes successfully it
1386 * will be released to the locked queue.
1388 bp
->b_flags
&= ~B_ERROR
;
1389 bp
->b_qindex
= BQUEUE_LOCKED
;
1390 TAILQ_INSERT_TAIL(&bufqueues
[BQUEUE_LOCKED
], bp
, b_freelist
);
1391 } else if (bp
->b_flags
& B_DELWRI
) {
1392 bp
->b_qindex
= (bp
->b_flags
& B_HEAVY
) ?
1393 BQUEUE_DIRTY_HW
: BQUEUE_DIRTY
;
1394 TAILQ_INSERT_TAIL(&bufqueues
[bp
->b_qindex
], bp
, b_freelist
);
1395 } else if (vm_page_count_severe()) {
1397 * We are too low on memory, we have to try to free the
1398 * buffer (most importantly: the wired pages making up its
1399 * backing store) *now*.
1405 bp
->b_qindex
= BQUEUE_CLEAN
;
1406 TAILQ_INSERT_TAIL(&bufqueues
[BQUEUE_CLEAN
], bp
, b_freelist
);
1409 if ((bp
->b_flags
& B_LOCKED
) == 0 &&
1410 ((bp
->b_flags
& B_INVAL
) || (bp
->b_flags
& B_DELWRI
) == 0)) {
1415 * Something we can maybe free or reuse.
1417 if (bp
->b_bufsize
&& !(bp
->b_flags
& B_DELWRI
))
1421 * Final cleanup and unlock. Clear bits that are only used while a
1422 * buffer is actively locked.
1424 bp
->b_flags
&= ~(B_ORDERED
| B_ASYNC
| B_NOCACHE
| B_RELBUF
);
1432 * Return backing pages held by the buffer 'bp' back to the VM system
1433 * if possible. The pages are freed if they are no longer valid or
1434 * attempt to free if it was used for direct I/O otherwise they are
1435 * sent to the page cache.
1437 * Pages that were marked busy are left alone and skipped.
1439 * The KVA mapping (b_data) for the underlying pages is removed by
1443 vfs_vmio_release(struct buf
*bp
)
1449 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
1450 m
= bp
->b_xio
.xio_pages
[i
];
1451 bp
->b_xio
.xio_pages
[i
] = NULL
;
1453 * In order to keep page LRU ordering consistent, put
1454 * everything on the inactive queue.
1456 vm_page_unwire(m
, 0);
1458 * We don't mess with busy pages, it is
1459 * the responsibility of the process that
1460 * busied the pages to deal with them.
1462 if ((m
->flags
& PG_BUSY
) || (m
->busy
!= 0))
1465 if (m
->wire_count
== 0) {
1466 vm_page_flag_clear(m
, PG_ZERO
);
1468 * Might as well free the page if we can and it has
1469 * no valid data. We also free the page if the
1470 * buffer was used for direct I/O.
1472 if ((bp
->b_flags
& B_ASYNC
) == 0 && !m
->valid
&&
1473 m
->hold_count
== 0) {
1475 vm_page_protect(m
, VM_PROT_NONE
);
1477 } else if (bp
->b_flags
& B_DIRECT
) {
1478 vm_page_try_to_free(m
);
1479 } else if (vm_page_count_severe()) {
1480 vm_page_try_to_cache(m
);
1485 pmap_qremove(trunc_page((vm_offset_t
) bp
->b_data
), bp
->b_xio
.xio_npages
);
1486 if (bp
->b_bufsize
) {
1490 bp
->b_xio
.xio_npages
= 0;
1491 bp
->b_flags
&= ~B_VMIO
;
1492 KKASSERT (LIST_FIRST(&bp
->b_dep
) == NULL
);
1500 * Implement clustered async writes for clearing out B_DELWRI buffers.
1501 * This is much better then the old way of writing only one buffer at
1502 * a time. Note that we may not be presented with the buffers in the
1503 * correct order, so we search for the cluster in both directions.
1505 * The buffer is locked on call.
1508 vfs_bio_awrite(struct buf
*bp
)
1512 off_t loffset
= bp
->b_loffset
;
1513 struct vnode
*vp
= bp
->b_vp
;
1521 * right now we support clustered writing only to regular files. If
1522 * we find a clusterable block we could be in the middle of a cluster
1523 * rather then at the beginning.
1525 * NOTE: b_bio1 contains the logical loffset and is aliased
1526 * to b_loffset. b_bio2 contains the translated block number.
1528 if ((vp
->v_type
== VREG
) &&
1529 (vp
->v_mount
!= 0) && /* Only on nodes that have the size info */
1530 (bp
->b_flags
& (B_CLUSTEROK
| B_INVAL
)) == B_CLUSTEROK
) {
1532 size
= vp
->v_mount
->mnt_stat
.f_iosize
;
1534 for (i
= size
; i
< MAXPHYS
; i
+= size
) {
1535 if ((bpa
= findblk(vp
, loffset
+ i
)) &&
1536 BUF_REFCNT(bpa
) == 0 &&
1537 ((bpa
->b_flags
& (B_DELWRI
| B_CLUSTEROK
| B_INVAL
)) ==
1538 (B_DELWRI
| B_CLUSTEROK
)) &&
1539 (bpa
->b_bufsize
== size
)) {
1540 if ((bpa
->b_bio2
.bio_offset
== NOOFFSET
) ||
1541 (bpa
->b_bio2
.bio_offset
!=
1542 bp
->b_bio2
.bio_offset
+ i
))
1548 for (j
= size
; i
+ j
<= MAXPHYS
&& j
<= loffset
; j
+= size
) {
1549 if ((bpa
= findblk(vp
, loffset
- j
)) &&
1550 BUF_REFCNT(bpa
) == 0 &&
1551 ((bpa
->b_flags
& (B_DELWRI
| B_CLUSTEROK
| B_INVAL
)) ==
1552 (B_DELWRI
| B_CLUSTEROK
)) &&
1553 (bpa
->b_bufsize
== size
)) {
1554 if ((bpa
->b_bio2
.bio_offset
== NOOFFSET
) ||
1555 (bpa
->b_bio2
.bio_offset
!=
1556 bp
->b_bio2
.bio_offset
- j
))
1565 * this is a possible cluster write
1567 if (nbytes
!= size
) {
1569 nwritten
= cluster_wbuild(vp
, size
,
1570 loffset
- j
, nbytes
);
1577 bp
->b_flags
|= B_ASYNC
;
1581 * default (old) behavior, writing out only one block
1583 * XXX returns b_bufsize instead of b_bcount for nwritten?
1585 nwritten
= bp
->b_bufsize
;
1594 * Find and initialize a new buffer header, freeing up existing buffers
1595 * in the bufqueues as necessary. The new buffer is returned locked.
1597 * Important: B_INVAL is not set. If the caller wishes to throw the
1598 * buffer away, the caller must set B_INVAL prior to calling brelse().
1601 * We have insufficient buffer headers
1602 * We have insufficient buffer space
1603 * buffer_map is too fragmented ( space reservation fails )
1604 * If we have to flush dirty buffers ( but we try to avoid this )
1606 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1607 * Instead we ask the buf daemon to do it for us. We attempt to
1608 * avoid piecemeal wakeups of the pageout daemon.
1612 getnewbuf(int blkflags
, int slptimeo
, int size
, int maxsize
)
1618 int slpflags
= (blkflags
& GETBLK_PCATCH
) ? PCATCH
: 0;
1619 static int flushingbufs
;
1622 * We can't afford to block since we might be holding a vnode lock,
1623 * which may prevent system daemons from running. We deal with
1624 * low-memory situations by proactively returning memory and running
1625 * async I/O rather then sync I/O.
1629 --getnewbufrestarts
;
1631 ++getnewbufrestarts
;
1634 * Setup for scan. If we do not have enough free buffers,
1635 * we setup a degenerate case that immediately fails. Note
1636 * that if we are specially marked process, we are allowed to
1637 * dip into our reserves.
1639 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1641 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1642 * However, there are a number of cases (defragging, reusing, ...)
1643 * where we cannot backup.
1645 nqindex
= BQUEUE_EMPTYKVA
;
1646 nbp
= TAILQ_FIRST(&bufqueues
[BQUEUE_EMPTYKVA
]);
1650 * If no EMPTYKVA buffers and we are either
1651 * defragging or reusing, locate a CLEAN buffer
1652 * to free or reuse. If bufspace useage is low
1653 * skip this step so we can allocate a new buffer.
1655 if (defrag
|| bufspace
>= lobufspace
) {
1656 nqindex
= BQUEUE_CLEAN
;
1657 nbp
= TAILQ_FIRST(&bufqueues
[BQUEUE_CLEAN
]);
1661 * If we could not find or were not allowed to reuse a
1662 * CLEAN buffer, check to see if it is ok to use an EMPTY
1663 * buffer. We can only use an EMPTY buffer if allocating
1664 * its KVA would not otherwise run us out of buffer space.
1666 if (nbp
== NULL
&& defrag
== 0 &&
1667 bufspace
+ maxsize
< hibufspace
) {
1668 nqindex
= BQUEUE_EMPTY
;
1669 nbp
= TAILQ_FIRST(&bufqueues
[BQUEUE_EMPTY
]);
1674 * Run scan, possibly freeing data and/or kva mappings on the fly
1678 while ((bp
= nbp
) != NULL
) {
1679 int qindex
= nqindex
;
1681 nbp
= TAILQ_NEXT(bp
, b_freelist
);
1684 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
1685 * cycles through the queue twice before being selected.
1687 if (qindex
== BQUEUE_CLEAN
&&
1688 (bp
->b_flags
& B_AGE
) == 0 && nbp
) {
1689 bp
->b_flags
|= B_AGE
;
1690 TAILQ_REMOVE(&bufqueues
[qindex
], bp
, b_freelist
);
1691 TAILQ_INSERT_TAIL(&bufqueues
[qindex
], bp
, b_freelist
);
1696 * Calculate next bp ( we can only use it if we do not block
1697 * or do other fancy things ).
1702 nqindex
= BQUEUE_EMPTYKVA
;
1703 if ((nbp
= TAILQ_FIRST(&bufqueues
[BQUEUE_EMPTYKVA
])))
1706 case BQUEUE_EMPTYKVA
:
1707 nqindex
= BQUEUE_CLEAN
;
1708 if ((nbp
= TAILQ_FIRST(&bufqueues
[BQUEUE_CLEAN
])))
1722 KASSERT(bp
->b_qindex
== qindex
, ("getnewbuf: inconsistent queue %d bp %p", qindex
, bp
));
1725 * Note: we no longer distinguish between VMIO and non-VMIO
1729 KASSERT((bp
->b_flags
& B_DELWRI
) == 0, ("delwri buffer %p found in queue %d", bp
, qindex
));
1732 * If we are defragging then we need a buffer with
1733 * b_kvasize != 0. XXX this situation should no longer
1734 * occur, if defrag is non-zero the buffer's b_kvasize
1735 * should also be non-zero at this point. XXX
1737 if (defrag
&& bp
->b_kvasize
== 0) {
1738 kprintf("Warning: defrag empty buffer %p\n", bp
);
1743 * Start freeing the bp. This is somewhat involved. nbp
1744 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers
1745 * on the clean list must be disassociated from their
1746 * current vnode. Buffers on the empty[kva] lists have
1747 * already been disassociated.
1750 if (BUF_LOCK(bp
, LK_EXCLUSIVE
| LK_NOWAIT
) != 0) {
1751 kprintf("getnewbuf: warning, locked buf %p, race corrected\n", bp
);
1752 tsleep(&bd_request
, 0, "gnbxxx", hz
/ 100);
1755 if (bp
->b_qindex
!= qindex
) {
1756 kprintf("getnewbuf: warning, BUF_LOCK blocked unexpectedly on buf %p index %d->%d, race corrected\n", bp
, qindex
, bp
->b_qindex
);
1763 * Dependancies must be handled before we disassociate the
1766 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
1767 * be immediately disassociated. HAMMER then becomes
1768 * responsible for releasing the buffer.
1770 if (LIST_FIRST(&bp
->b_dep
) != NULL
) {
1772 if (bp
->b_flags
& B_LOCKED
) {
1776 KKASSERT(LIST_FIRST(&bp
->b_dep
) == NULL
);
1779 if (qindex
== BQUEUE_CLEAN
) {
1780 if (bp
->b_flags
& B_VMIO
) {
1781 bp
->b_flags
&= ~B_ASYNC
;
1782 vfs_vmio_release(bp
);
1789 * NOTE: nbp is now entirely invalid. We can only restart
1790 * the scan from this point on.
1792 * Get the rest of the buffer freed up. b_kva* is still
1793 * valid after this operation.
1796 KASSERT(bp
->b_vp
== NULL
, ("bp3 %p flags %08x vnode %p qindex %d unexpectededly still associated!", bp
, bp
->b_flags
, bp
->b_vp
, qindex
));
1797 KKASSERT((bp
->b_flags
& B_HASHED
) == 0);
1800 * critical section protection is not required when
1801 * scrapping a buffer's contents because it is already
1807 bp
->b_flags
= B_BNOCLIP
;
1808 bp
->b_cmd
= BUF_CMD_DONE
;
1813 bp
->b_xio
.xio_npages
= 0;
1814 bp
->b_dirtyoff
= bp
->b_dirtyend
= 0;
1816 KKASSERT(LIST_FIRST(&bp
->b_dep
) == NULL
);
1818 if (blkflags
& GETBLK_BHEAVY
)
1819 bp
->b_flags
|= B_HEAVY
;
1822 * If we are defragging then free the buffer.
1825 bp
->b_flags
|= B_INVAL
;
1833 * If we are overcomitted then recover the buffer and its
1834 * KVM space. This occurs in rare situations when multiple
1835 * processes are blocked in getnewbuf() or allocbuf().
1837 if (bufspace
>= hibufspace
)
1839 if (flushingbufs
&& bp
->b_kvasize
!= 0) {
1840 bp
->b_flags
|= B_INVAL
;
1845 if (bufspace
< lobufspace
)
1851 * If we exhausted our list, sleep as appropriate. We may have to
1852 * wakeup various daemons and write out some dirty buffers.
1854 * Generally we are sleeping due to insufficient buffer space.
1862 flags
= VFS_BIO_NEED_BUFSPACE
;
1864 } else if (bufspace
>= hibufspace
) {
1866 flags
= VFS_BIO_NEED_BUFSPACE
;
1869 flags
= VFS_BIO_NEED_ANY
;
1872 needsbuffer
|= flags
;
1873 bd_speedup(); /* heeeelp */
1874 while (needsbuffer
& flags
) {
1875 if (tsleep(&needsbuffer
, slpflags
, waitmsg
, slptimeo
))
1880 * We finally have a valid bp. We aren't quite out of the
1881 * woods, we still have to reserve kva space. In order
1882 * to keep fragmentation sane we only allocate kva in
1885 maxsize
= (maxsize
+ BKVAMASK
) & ~BKVAMASK
;
1887 if (maxsize
!= bp
->b_kvasize
) {
1888 vm_offset_t addr
= 0;
1893 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
1894 vm_map_lock(&buffer_map
);
1896 if (vm_map_findspace(&buffer_map
,
1897 vm_map_min(&buffer_map
), maxsize
,
1900 * Uh oh. Buffer map is too fragmented. We
1901 * must defragment the map.
1903 vm_map_unlock(&buffer_map
);
1904 vm_map_entry_release(count
);
1907 bp
->b_flags
|= B_INVAL
;
1912 vm_map_insert(&buffer_map
, &count
,
1914 addr
, addr
+ maxsize
,
1916 VM_PROT_ALL
, VM_PROT_ALL
,
1919 bp
->b_kvabase
= (caddr_t
) addr
;
1920 bp
->b_kvasize
= maxsize
;
1921 bufspace
+= bp
->b_kvasize
;
1924 vm_map_unlock(&buffer_map
);
1925 vm_map_entry_release(count
);
1927 bp
->b_data
= bp
->b_kvabase
;
1933 * This routine is called in an emergency to recover VM pages from the
1934 * buffer cache by cashing in clean buffers. The idea is to recover
1935 * enough pages to be able to satisfy a stuck bio_page_alloc().
1938 recoverbufpages(void)
1945 while (bytes
< MAXBSIZE
) {
1946 bp
= TAILQ_FIRST(&bufqueues
[BQUEUE_CLEAN
]);
1951 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
1952 * cycles through the queue twice before being selected.
1954 if ((bp
->b_flags
& B_AGE
) == 0 && TAILQ_NEXT(bp
, b_freelist
)) {
1955 bp
->b_flags
|= B_AGE
;
1956 TAILQ_REMOVE(&bufqueues
[BQUEUE_CLEAN
], bp
, b_freelist
);
1957 TAILQ_INSERT_TAIL(&bufqueues
[BQUEUE_CLEAN
],
1965 KKASSERT(bp
->b_qindex
== BQUEUE_CLEAN
);
1966 KKASSERT((bp
->b_flags
& B_DELWRI
) == 0);
1969 * Start freeing the bp. This is somewhat involved.
1971 * Buffers on the clean list must be disassociated from
1972 * their current vnode
1975 if (BUF_LOCK(bp
, LK_EXCLUSIVE
| LK_NOWAIT
) != 0) {
1976 kprintf("recoverbufpages: warning, locked buf %p, race corrected\n", bp
);
1977 tsleep(&bd_request
, 0, "gnbxxx", hz
/ 100);
1980 if (bp
->b_qindex
!= BQUEUE_CLEAN
) {
1981 kprintf("recoverbufpages: warning, BUF_LOCK blocked unexpectedly on buf %p index %d, race corrected\n", bp
, bp
->b_qindex
);
1988 * Dependancies must be handled before we disassociate the
1991 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
1992 * be immediately disassociated. HAMMER then becomes
1993 * responsible for releasing the buffer.
1995 if (LIST_FIRST(&bp
->b_dep
) != NULL
) {
1997 if (bp
->b_flags
& B_LOCKED
) {
2001 KKASSERT(LIST_FIRST(&bp
->b_dep
) == NULL
);
2004 bytes
+= bp
->b_bufsize
;
2006 if (bp
->b_flags
& B_VMIO
) {
2007 bp
->b_flags
&= ~B_ASYNC
;
2008 bp
->b_flags
|= B_DIRECT
; /* try to free pages */
2009 vfs_vmio_release(bp
);
2014 KKASSERT(bp
->b_vp
== NULL
);
2015 KKASSERT((bp
->b_flags
& B_HASHED
) == 0);
2018 * critical section protection is not required when
2019 * scrapping a buffer's contents because it is already
2025 bp
->b_flags
= B_BNOCLIP
;
2026 bp
->b_cmd
= BUF_CMD_DONE
;
2031 bp
->b_xio
.xio_npages
= 0;
2032 bp
->b_dirtyoff
= bp
->b_dirtyend
= 0;
2034 KKASSERT(LIST_FIRST(&bp
->b_dep
) == NULL
);
2036 bp
->b_flags
|= B_INVAL
;
2046 * Buffer flushing daemon. Buffers are normally flushed by the
2047 * update daemon but if it cannot keep up this process starts to
2048 * take the load in an attempt to prevent getnewbuf() from blocking.
2050 * Once a flush is initiated it does not stop until the number
2051 * of buffers falls below lodirtybuffers, but we will wake up anyone
2052 * waiting at the mid-point.
2055 static struct kproc_desc buf_kp
= {
2060 SYSINIT(bufdaemon
, SI_SUB_KTHREAD_BUF
, SI_ORDER_FIRST
,
2061 kproc_start
, &buf_kp
)
2063 static struct kproc_desc bufhw_kp
= {
2068 SYSINIT(bufdaemon_hw
, SI_SUB_KTHREAD_BUF
, SI_ORDER_FIRST
,
2069 kproc_start
, &bufhw_kp
)
2077 * This process needs to be suspended prior to shutdown sync.
2079 EVENTHANDLER_REGISTER(shutdown_pre_sync
, shutdown_kproc
,
2080 bufdaemon_td
, SHUTDOWN_PRI_LAST
);
2081 curthread
->td_flags
|= TDF_SYSTHREAD
;
2084 * This process is allowed to take the buffer cache to the limit
2089 kproc_suspend_loop();
2092 * Do the flush. Limit the amount of in-transit I/O we
2093 * allow to build up, otherwise we would completely saturate
2094 * the I/O system. Wakeup any waiting processes before we
2095 * normally would so they can run in parallel with our drain.
2097 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2098 * but because we split the operation into two threads we
2099 * have to cut it in half for each thread.
2101 limit
= lodirtybufspace
/ 2;
2102 waitrunningbufspace(limit
);
2103 while (runningbufspace
+ dirtybufspace
> limit
||
2104 dirtybufcount
- dirtybufcounthw
>= nbuf
/ 2) {
2105 if (flushbufqueues(BQUEUE_DIRTY
) == 0)
2107 waitrunningbufspace(limit
);
2111 * We reached our low water mark, reset the
2112 * request and sleep until we are needed again.
2113 * The sleep is just so the suspend code works.
2115 spin_lock_wr(&needsbuffer_spin
);
2116 if (bd_request
== 0) {
2117 msleep(&bd_request
, &needsbuffer_spin
, 0,
2121 spin_unlock_wr(&needsbuffer_spin
);
2131 * This process needs to be suspended prior to shutdown sync.
2133 EVENTHANDLER_REGISTER(shutdown_pre_sync
, shutdown_kproc
,
2134 bufdaemonhw_td
, SHUTDOWN_PRI_LAST
);
2135 curthread
->td_flags
|= TDF_SYSTHREAD
;
2138 * This process is allowed to take the buffer cache to the limit
2143 kproc_suspend_loop();
2146 * Do the flush. Limit the amount of in-transit I/O we
2147 * allow to build up, otherwise we would completely saturate
2148 * the I/O system. Wakeup any waiting processes before we
2149 * normally would so they can run in parallel with our drain.
2151 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2152 * but because we split the operation into two threads we
2153 * have to cut it in half for each thread.
2155 limit
= lodirtybufspace
/ 2;
2156 waitrunningbufspace(limit
);
2157 while (runningbufspace
+ dirtybufspacehw
> limit
||
2158 dirtybufcounthw
>= nbuf
/ 2) {
2159 if (flushbufqueues(BQUEUE_DIRTY_HW
) == 0)
2161 waitrunningbufspace(limit
);
2165 * We reached our low water mark, reset the
2166 * request and sleep until we are needed again.
2167 * The sleep is just so the suspend code works.
2169 spin_lock_wr(&needsbuffer_spin
);
2170 if (bd_request_hw
== 0) {
2171 msleep(&bd_request_hw
, &needsbuffer_spin
, 0,
2175 spin_unlock_wr(&needsbuffer_spin
);
2182 * Try to flush a buffer in the dirty queue. We must be careful to
2183 * free up B_INVAL buffers instead of write them, which NFS is
2184 * particularly sensitive to.
2186 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate
2187 * that we really want to try to get the buffer out and reuse it
2188 * due to the write load on the machine.
2192 flushbufqueues(bufq_type_t q
)
2197 bp
= TAILQ_FIRST(&bufqueues
[q
]);
2200 KASSERT((bp
->b_flags
& B_DELWRI
),
2201 ("unexpected clean buffer %p", bp
));
2203 if (bp
->b_flags
& B_DELWRI
) {
2204 if (bp
->b_flags
& B_INVAL
) {
2205 if (BUF_LOCK(bp
, LK_EXCLUSIVE
| LK_NOWAIT
) != 0)
2206 panic("flushbufqueues: locked buf");
2212 if (LIST_FIRST(&bp
->b_dep
) != NULL
&&
2213 (bp
->b_flags
& B_DEFERRED
) == 0 &&
2214 buf_countdeps(bp
, 0)) {
2215 TAILQ_REMOVE(&bufqueues
[q
], bp
, b_freelist
);
2216 TAILQ_INSERT_TAIL(&bufqueues
[q
], bp
,
2218 bp
->b_flags
|= B_DEFERRED
;
2219 bp
= TAILQ_FIRST(&bufqueues
[q
]);
2224 * Only write it out if we can successfully lock
2225 * it. If the buffer has a dependancy,
2226 * buf_checkwrite must also return 0.
2228 if (BUF_LOCK(bp
, LK_EXCLUSIVE
| LK_NOWAIT
) == 0) {
2229 if (LIST_FIRST(&bp
->b_dep
) != NULL
&&
2230 buf_checkwrite(bp
)) {
2234 bp
->b_flags
|= B_AGE
;
2241 bp
= TAILQ_NEXT(bp
, b_freelist
);
2249 * Returns true if no I/O is needed to access the associated VM object.
2250 * This is like findblk except it also hunts around in the VM system for
2253 * Note that we ignore vm_page_free() races from interrupts against our
2254 * lookup, since if the caller is not protected our return value will not
2255 * be any more valid then otherwise once we exit the critical section.
2258 inmem(struct vnode
*vp
, off_t loffset
)
2261 vm_offset_t toff
, tinc
, size
;
2264 if (findblk(vp
, loffset
))
2266 if (vp
->v_mount
== NULL
)
2268 if ((obj
= vp
->v_object
) == NULL
)
2272 if (size
> vp
->v_mount
->mnt_stat
.f_iosize
)
2273 size
= vp
->v_mount
->mnt_stat
.f_iosize
;
2275 for (toff
= 0; toff
< vp
->v_mount
->mnt_stat
.f_iosize
; toff
+= tinc
) {
2276 m
= vm_page_lookup(obj
, OFF_TO_IDX(loffset
+ toff
));
2280 if (tinc
> PAGE_SIZE
- ((toff
+ loffset
) & PAGE_MASK
))
2281 tinc
= PAGE_SIZE
- ((toff
+ loffset
) & PAGE_MASK
);
2282 if (vm_page_is_valid(m
,
2283 (vm_offset_t
) ((toff
+ loffset
) & PAGE_MASK
), tinc
) == 0)
2292 * Sets the dirty range for a buffer based on the status of the dirty
2293 * bits in the pages comprising the buffer.
2295 * The range is limited to the size of the buffer.
2297 * This routine is primarily used by NFS, but is generalized for the
2301 vfs_setdirty(struct buf
*bp
)
2307 * Degenerate case - empty buffer
2310 if (bp
->b_bufsize
== 0)
2314 * We qualify the scan for modified pages on whether the
2315 * object has been flushed yet. The OBJ_WRITEABLE flag
2316 * is not cleared simply by protecting pages off.
2319 if ((bp
->b_flags
& B_VMIO
) == 0)
2322 object
= bp
->b_xio
.xio_pages
[0]->object
;
2324 if ((object
->flags
& OBJ_WRITEABLE
) && !(object
->flags
& OBJ_MIGHTBEDIRTY
))
2325 kprintf("Warning: object %p writeable but not mightbedirty\n", object
);
2326 if (!(object
->flags
& OBJ_WRITEABLE
) && (object
->flags
& OBJ_MIGHTBEDIRTY
))
2327 kprintf("Warning: object %p mightbedirty but not writeable\n", object
);
2329 if (object
->flags
& (OBJ_MIGHTBEDIRTY
|OBJ_CLEANING
)) {
2330 vm_offset_t boffset
;
2331 vm_offset_t eoffset
;
2334 * test the pages to see if they have been modified directly
2335 * by users through the VM system.
2337 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
2338 vm_page_flag_clear(bp
->b_xio
.xio_pages
[i
], PG_ZERO
);
2339 vm_page_test_dirty(bp
->b_xio
.xio_pages
[i
]);
2343 * Calculate the encompassing dirty range, boffset and eoffset,
2344 * (eoffset - boffset) bytes.
2347 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
2348 if (bp
->b_xio
.xio_pages
[i
]->dirty
)
2351 boffset
= (i
<< PAGE_SHIFT
) - (bp
->b_loffset
& PAGE_MASK
);
2353 for (i
= bp
->b_xio
.xio_npages
- 1; i
>= 0; --i
) {
2354 if (bp
->b_xio
.xio_pages
[i
]->dirty
) {
2358 eoffset
= ((i
+ 1) << PAGE_SHIFT
) - (bp
->b_loffset
& PAGE_MASK
);
2361 * Fit it to the buffer.
2364 if (eoffset
> bp
->b_bcount
)
2365 eoffset
= bp
->b_bcount
;
2368 * If we have a good dirty range, merge with the existing
2372 if (boffset
< eoffset
) {
2373 if (bp
->b_dirtyoff
> boffset
)
2374 bp
->b_dirtyoff
= boffset
;
2375 if (bp
->b_dirtyend
< eoffset
)
2376 bp
->b_dirtyend
= eoffset
;
2384 * Locate and return the specified buffer, or NULL if the buffer does
2385 * not exist. Do not attempt to lock the buffer or manipulate it in
2386 * any way. The caller must validate that the correct buffer has been
2387 * obtain after locking it.
2390 findblk(struct vnode
*vp
, off_t loffset
)
2395 bp
= buf_rb_hash_RB_LOOKUP(&vp
->v_rbhash_tree
, loffset
);
2403 * Get a block given a specified block and offset into a file/device.
2404 * B_INVAL may or may not be set on return. The caller should clear
2405 * B_INVAL prior to initiating a READ.
2407 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2408 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2409 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer
2410 * without doing any of those things the system will likely believe
2411 * the buffer to be valid (especially if it is not B_VMIO), and the
2412 * next getblk() will return the buffer with B_CACHE set.
2414 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2415 * an existing buffer.
2417 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2418 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2419 * and then cleared based on the backing VM. If the previous buffer is
2420 * non-0-sized but invalid, B_CACHE will be cleared.
2422 * If getblk() must create a new buffer, the new buffer is returned with
2423 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2424 * case it is returned with B_INVAL clear and B_CACHE set based on the
2427 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2428 * B_CACHE bit is clear.
2430 * What this means, basically, is that the caller should use B_CACHE to
2431 * determine whether the buffer is fully valid or not and should clear
2432 * B_INVAL prior to issuing a read. If the caller intends to validate
2433 * the buffer by loading its data area with something, the caller needs
2434 * to clear B_INVAL. If the caller does this without issuing an I/O,
2435 * the caller should set B_CACHE ( as an optimization ), else the caller
2436 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2437 * a write attempt or if it was a successfull read. If the caller
2438 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2439 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2443 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return
2444 * GETBLK_BHEAVY - heavy-weight buffer cache buffer
2447 getblk(struct vnode
*vp
, off_t loffset
, int size
, int blkflags
, int slptimeo
)
2450 int slpflags
= (blkflags
& GETBLK_PCATCH
) ? PCATCH
: 0;
2453 if (size
> MAXBSIZE
)
2454 panic("getblk: size(%d) > MAXBSIZE(%d)", size
, MAXBSIZE
);
2455 if (vp
->v_object
== NULL
)
2456 panic("getblk: vnode %p has no object!", vp
);
2460 if ((bp
= findblk(vp
, loffset
))) {
2462 * The buffer was found in the cache, but we need to lock it.
2463 * Even with LK_NOWAIT the lockmgr may break our critical
2464 * section, so double-check the validity of the buffer
2465 * once the lock has been obtained.
2467 if (BUF_LOCK(bp
, LK_EXCLUSIVE
| LK_NOWAIT
)) {
2468 int lkflags
= LK_EXCLUSIVE
| LK_SLEEPFAIL
;
2469 if (blkflags
& GETBLK_PCATCH
)
2470 lkflags
|= LK_PCATCH
;
2471 error
= BUF_TIMELOCK(bp
, lkflags
, "getblk", slptimeo
);
2473 if (error
== ENOLCK
)
2481 * Once the buffer has been locked, make sure we didn't race
2482 * a buffer recyclement. Buffers that are no longer hashed
2483 * will have b_vp == NULL, so this takes care of that check
2486 if (bp
->b_vp
!= vp
|| bp
->b_loffset
!= loffset
) {
2487 kprintf("Warning buffer %p (vp %p loffset %lld) was recycled\n", bp
, vp
, loffset
);
2493 * All vnode-based buffers must be backed by a VM object.
2495 KKASSERT(bp
->b_flags
& B_VMIO
);
2496 KKASSERT(bp
->b_cmd
== BUF_CMD_DONE
);
2497 bp
->b_flags
&= ~B_AGE
;
2500 * Make sure that B_INVAL buffers do not have a cached
2501 * block number translation.
2503 if ((bp
->b_flags
& B_INVAL
) && (bp
->b_bio2
.bio_offset
!= NOOFFSET
)) {
2504 kprintf("Warning invalid buffer %p (vp %p loffset %lld) did not have cleared bio_offset cache\n", bp
, vp
, loffset
);
2505 clearbiocache(&bp
->b_bio2
);
2509 * The buffer is locked. B_CACHE is cleared if the buffer is
2512 if (bp
->b_flags
& B_INVAL
)
2513 bp
->b_flags
&= ~B_CACHE
;
2517 * Any size inconsistancy with a dirty buffer or a buffer
2518 * with a softupdates dependancy must be resolved. Resizing
2519 * the buffer in such circumstances can lead to problems.
2521 if (size
!= bp
->b_bcount
) {
2522 if (bp
->b_flags
& B_DELWRI
) {
2523 bp
->b_flags
|= B_NOCACHE
;
2525 } else if (LIST_FIRST(&bp
->b_dep
)) {
2526 bp
->b_flags
|= B_NOCACHE
;
2529 bp
->b_flags
|= B_RELBUF
;
2534 KKASSERT(size
<= bp
->b_kvasize
);
2535 KASSERT(bp
->b_loffset
!= NOOFFSET
,
2536 ("getblk: no buffer offset"));
2539 * A buffer with B_DELWRI set and B_CACHE clear must
2540 * be committed before we can return the buffer in
2541 * order to prevent the caller from issuing a read
2542 * ( due to B_CACHE not being set ) and overwriting
2545 * Most callers, including NFS and FFS, need this to
2546 * operate properly either because they assume they
2547 * can issue a read if B_CACHE is not set, or because
2548 * ( for example ) an uncached B_DELWRI might loop due
2549 * to softupdates re-dirtying the buffer. In the latter
2550 * case, B_CACHE is set after the first write completes,
2551 * preventing further loops.
2553 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2554 * above while extending the buffer, we cannot allow the
2555 * buffer to remain with B_CACHE set after the write
2556 * completes or it will represent a corrupt state. To
2557 * deal with this we set B_NOCACHE to scrap the buffer
2560 * We might be able to do something fancy, like setting
2561 * B_CACHE in bwrite() except if B_DELWRI is already set,
2562 * so the below call doesn't set B_CACHE, but that gets real
2563 * confusing. This is much easier.
2566 if ((bp
->b_flags
& (B_CACHE
|B_DELWRI
)) == B_DELWRI
) {
2567 bp
->b_flags
|= B_NOCACHE
;
2574 * Buffer is not in-core, create new buffer. The buffer
2575 * returned by getnewbuf() is locked. Note that the returned
2576 * buffer is also considered valid (not marked B_INVAL).
2578 * Calculating the offset for the I/O requires figuring out
2579 * the block size. We use DEV_BSIZE for VBLK or VCHR and
2580 * the mount's f_iosize otherwise. If the vnode does not
2581 * have an associated mount we assume that the passed size is
2584 * Note that vn_isdisk() cannot be used here since it may
2585 * return a failure for numerous reasons. Note that the
2586 * buffer size may be larger then the block size (the caller
2587 * will use block numbers with the proper multiple). Beware
2588 * of using any v_* fields which are part of unions. In
2589 * particular, in DragonFly the mount point overloading
2590 * mechanism uses the namecache only and the underlying
2591 * directory vnode is not a special case.
2595 if (vp
->v_type
== VBLK
|| vp
->v_type
== VCHR
)
2597 else if (vp
->v_mount
)
2598 bsize
= vp
->v_mount
->mnt_stat
.f_iosize
;
2602 maxsize
= size
+ (loffset
& PAGE_MASK
);
2603 maxsize
= imax(maxsize
, bsize
);
2605 if ((bp
= getnewbuf(blkflags
, slptimeo
, size
, maxsize
)) == NULL
) {
2606 if (slpflags
|| slptimeo
) {
2614 * This code is used to make sure that a buffer is not
2615 * created while the getnewbuf routine is blocked.
2616 * This can be a problem whether the vnode is locked or not.
2617 * If the buffer is created out from under us, we have to
2618 * throw away the one we just created. There is no window
2619 * race because we are safely running in a critical section
2620 * from the point of the duplicate buffer creation through
2621 * to here, and we've locked the buffer.
2623 if (findblk(vp
, loffset
)) {
2624 bp
->b_flags
|= B_INVAL
;
2630 * Insert the buffer into the hash, so that it can
2631 * be found by findblk().
2633 * Make sure the translation layer has been cleared.
2635 bp
->b_loffset
= loffset
;
2636 bp
->b_bio2
.bio_offset
= NOOFFSET
;
2637 /* bp->b_bio2.bio_next = NULL; */
2642 * All vnode-based buffers must be backed by a VM object.
2644 KKASSERT(vp
->v_object
!= NULL
);
2645 bp
->b_flags
|= B_VMIO
;
2646 KKASSERT(bp
->b_cmd
== BUF_CMD_DONE
);
2658 * Reacquire a buffer that was previously released to the locked queue,
2659 * or reacquire a buffer which is interlocked by having bioops->io_deallocate
2660 * set B_LOCKED (which handles the acquisition race).
2662 * To this end, either B_LOCKED must be set or the dependancy list must be
2666 regetblk(struct buf
*bp
)
2668 KKASSERT((bp
->b_flags
& B_LOCKED
) || LIST_FIRST(&bp
->b_dep
) != NULL
);
2669 BUF_LOCK(bp
, LK_EXCLUSIVE
| LK_RETRY
);
2678 * Get an empty, disassociated buffer of given size. The buffer is
2679 * initially set to B_INVAL.
2681 * critical section protection is not required for the allocbuf()
2682 * call because races are impossible here.
2690 maxsize
= (size
+ BKVAMASK
) & ~BKVAMASK
;
2693 while ((bp
= getnewbuf(0, 0, size
, maxsize
)) == 0)
2697 bp
->b_flags
|= B_INVAL
; /* b_dep cleared by getnewbuf() */
2705 * This code constitutes the buffer memory from either anonymous system
2706 * memory (in the case of non-VMIO operations) or from an associated
2707 * VM object (in the case of VMIO operations). This code is able to
2708 * resize a buffer up or down.
2710 * Note that this code is tricky, and has many complications to resolve
2711 * deadlock or inconsistant data situations. Tread lightly!!!
2712 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2713 * the caller. Calling this code willy nilly can result in the loss of data.
2715 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
2716 * B_CACHE for the non-VMIO case.
2718 * This routine does not need to be called from a critical section but you
2719 * must own the buffer.
2722 allocbuf(struct buf
*bp
, int size
)
2724 int newbsize
, mbsize
;
2727 if (BUF_REFCNT(bp
) == 0)
2728 panic("allocbuf: buffer not busy");
2730 if (bp
->b_kvasize
< size
)
2731 panic("allocbuf: buffer too small");
2733 if ((bp
->b_flags
& B_VMIO
) == 0) {
2737 * Just get anonymous memory from the kernel. Don't
2738 * mess with B_CACHE.
2740 mbsize
= (size
+ DEV_BSIZE
- 1) & ~(DEV_BSIZE
- 1);
2741 if (bp
->b_flags
& B_MALLOC
)
2744 newbsize
= round_page(size
);
2746 if (newbsize
< bp
->b_bufsize
) {
2748 * Malloced buffers are not shrunk
2750 if (bp
->b_flags
& B_MALLOC
) {
2752 bp
->b_bcount
= size
;
2754 kfree(bp
->b_data
, M_BIOBUF
);
2755 if (bp
->b_bufsize
) {
2756 bufmallocspace
-= bp
->b_bufsize
;
2760 bp
->b_data
= bp
->b_kvabase
;
2762 bp
->b_flags
&= ~B_MALLOC
;
2768 (vm_offset_t
) bp
->b_data
+ newbsize
,
2769 (vm_offset_t
) bp
->b_data
+ bp
->b_bufsize
);
2770 } else if (newbsize
> bp
->b_bufsize
) {
2772 * We only use malloced memory on the first allocation.
2773 * and revert to page-allocated memory when the buffer
2776 if ((bufmallocspace
< maxbufmallocspace
) &&
2777 (bp
->b_bufsize
== 0) &&
2778 (mbsize
<= PAGE_SIZE
/2)) {
2780 bp
->b_data
= kmalloc(mbsize
, M_BIOBUF
, M_WAITOK
);
2781 bp
->b_bufsize
= mbsize
;
2782 bp
->b_bcount
= size
;
2783 bp
->b_flags
|= B_MALLOC
;
2784 bufmallocspace
+= mbsize
;
2790 * If the buffer is growing on its other-than-first
2791 * allocation, then we revert to the page-allocation
2794 if (bp
->b_flags
& B_MALLOC
) {
2795 origbuf
= bp
->b_data
;
2796 origbufsize
= bp
->b_bufsize
;
2797 bp
->b_data
= bp
->b_kvabase
;
2798 if (bp
->b_bufsize
) {
2799 bufmallocspace
-= bp
->b_bufsize
;
2803 bp
->b_flags
&= ~B_MALLOC
;
2804 newbsize
= round_page(newbsize
);
2808 (vm_offset_t
) bp
->b_data
+ bp
->b_bufsize
,
2809 (vm_offset_t
) bp
->b_data
+ newbsize
);
2811 bcopy(origbuf
, bp
->b_data
, origbufsize
);
2812 kfree(origbuf
, M_BIOBUF
);
2819 newbsize
= (size
+ DEV_BSIZE
- 1) & ~(DEV_BSIZE
- 1);
2820 desiredpages
= ((int)(bp
->b_loffset
& PAGE_MASK
) +
2821 newbsize
+ PAGE_MASK
) >> PAGE_SHIFT
;
2822 KKASSERT(desiredpages
<= XIO_INTERNAL_PAGES
);
2824 if (bp
->b_flags
& B_MALLOC
)
2825 panic("allocbuf: VMIO buffer can't be malloced");
2827 * Set B_CACHE initially if buffer is 0 length or will become
2830 if (size
== 0 || bp
->b_bufsize
== 0)
2831 bp
->b_flags
|= B_CACHE
;
2833 if (newbsize
< bp
->b_bufsize
) {
2835 * DEV_BSIZE aligned new buffer size is less then the
2836 * DEV_BSIZE aligned existing buffer size. Figure out
2837 * if we have to remove any pages.
2839 if (desiredpages
< bp
->b_xio
.xio_npages
) {
2840 for (i
= desiredpages
; i
< bp
->b_xio
.xio_npages
; i
++) {
2842 * the page is not freed here -- it
2843 * is the responsibility of
2844 * vnode_pager_setsize
2846 m
= bp
->b_xio
.xio_pages
[i
];
2847 KASSERT(m
!= bogus_page
,
2848 ("allocbuf: bogus page found"));
2849 while (vm_page_sleep_busy(m
, TRUE
, "biodep"))
2852 bp
->b_xio
.xio_pages
[i
] = NULL
;
2853 vm_page_unwire(m
, 0);
2855 pmap_qremove((vm_offset_t
) trunc_page((vm_offset_t
)bp
->b_data
) +
2856 (desiredpages
<< PAGE_SHIFT
), (bp
->b_xio
.xio_npages
- desiredpages
));
2857 bp
->b_xio
.xio_npages
= desiredpages
;
2859 } else if (size
> bp
->b_bcount
) {
2861 * We are growing the buffer, possibly in a
2862 * byte-granular fashion.
2870 * Step 1, bring in the VM pages from the object,
2871 * allocating them if necessary. We must clear
2872 * B_CACHE if these pages are not valid for the
2873 * range covered by the buffer.
2875 * critical section protection is required to protect
2876 * against interrupts unbusying and freeing pages
2877 * between our vm_page_lookup() and our
2878 * busycheck/wiring call.
2884 while (bp
->b_xio
.xio_npages
< desiredpages
) {
2888 pi
= OFF_TO_IDX(bp
->b_loffset
) + bp
->b_xio
.xio_npages
;
2889 if ((m
= vm_page_lookup(obj
, pi
)) == NULL
) {
2891 * note: must allocate system pages
2892 * since blocking here could intefere
2893 * with paging I/O, no matter which
2896 m
= bio_page_alloc(obj
, pi
, desiredpages
- bp
->b_xio
.xio_npages
);
2900 bp
->b_flags
&= ~B_CACHE
;
2901 bp
->b_xio
.xio_pages
[bp
->b_xio
.xio_npages
] = m
;
2902 ++bp
->b_xio
.xio_npages
;
2908 * We found a page. If we have to sleep on it,
2909 * retry because it might have gotten freed out
2912 * We can only test PG_BUSY here. Blocking on
2913 * m->busy might lead to a deadlock:
2915 * vm_fault->getpages->cluster_read->allocbuf
2919 if (vm_page_sleep_busy(m
, FALSE
, "pgtblk"))
2923 * We have a good page. Should we wakeup the
2926 if ((curthread
!= pagethread
) &&
2927 ((m
->queue
- m
->pc
) == PQ_CACHE
) &&
2928 ((vmstats
.v_free_count
+ vmstats
.v_cache_count
) <
2929 (vmstats
.v_free_min
+ vmstats
.v_cache_min
))) {
2930 pagedaemon_wakeup();
2932 vm_page_flag_clear(m
, PG_ZERO
);
2934 bp
->b_xio
.xio_pages
[bp
->b_xio
.xio_npages
] = m
;
2935 ++bp
->b_xio
.xio_npages
;
2940 * Step 2. We've loaded the pages into the buffer,
2941 * we have to figure out if we can still have B_CACHE
2942 * set. Note that B_CACHE is set according to the
2943 * byte-granular range ( bcount and size ), not the
2944 * aligned range ( newbsize ).
2946 * The VM test is against m->valid, which is DEV_BSIZE
2947 * aligned. Needless to say, the validity of the data
2948 * needs to also be DEV_BSIZE aligned. Note that this
2949 * fails with NFS if the server or some other client
2950 * extends the file's EOF. If our buffer is resized,
2951 * B_CACHE may remain set! XXX
2954 toff
= bp
->b_bcount
;
2955 tinc
= PAGE_SIZE
- ((bp
->b_loffset
+ toff
) & PAGE_MASK
);
2957 while ((bp
->b_flags
& B_CACHE
) && toff
< size
) {
2960 if (tinc
> (size
- toff
))
2963 pi
= ((bp
->b_loffset
& PAGE_MASK
) + toff
) >>
2971 bp
->b_xio
.xio_pages
[pi
]
2978 * Step 3, fixup the KVM pmap. Remember that
2979 * bp->b_data is relative to bp->b_loffset, but
2980 * bp->b_loffset may be offset into the first page.
2983 bp
->b_data
= (caddr_t
)
2984 trunc_page((vm_offset_t
)bp
->b_data
);
2986 (vm_offset_t
)bp
->b_data
,
2987 bp
->b_xio
.xio_pages
,
2988 bp
->b_xio
.xio_npages
2990 bp
->b_data
= (caddr_t
)((vm_offset_t
)bp
->b_data
|
2991 (vm_offset_t
)(bp
->b_loffset
& PAGE_MASK
));
2995 /* adjust space use on already-dirty buffer */
2996 if (bp
->b_flags
& B_DELWRI
) {
2997 dirtybufspace
+= newbsize
- bp
->b_bufsize
;
2998 if (bp
->b_flags
& B_HEAVY
)
2999 dirtybufspacehw
+= newbsize
- bp
->b_bufsize
;
3001 if (newbsize
< bp
->b_bufsize
)
3003 bp
->b_bufsize
= newbsize
; /* actual buffer allocation */
3004 bp
->b_bcount
= size
; /* requested buffer size */
3011 * Wait for buffer I/O completion, returning error status. The buffer
3012 * is left locked on return. B_EINTR is converted into an EINTR error
3015 * NOTE! The original b_cmd is lost on return, since b_cmd will be
3016 * set to BUF_CMD_DONE.
3019 biowait(struct buf
*bp
)
3022 while (bp
->b_cmd
!= BUF_CMD_DONE
) {
3023 if (bp
->b_cmd
== BUF_CMD_READ
)
3024 tsleep(bp
, 0, "biord", 0);
3026 tsleep(bp
, 0, "biowr", 0);
3029 if (bp
->b_flags
& B_EINTR
) {
3030 bp
->b_flags
&= ~B_EINTR
;
3033 if (bp
->b_flags
& B_ERROR
) {
3034 return (bp
->b_error
? bp
->b_error
: EIO
);
3041 * This associates a tracking count with an I/O. vn_strategy() and
3042 * dev_dstrategy() do this automatically but there are a few cases
3043 * where a vnode or device layer is bypassed when a block translation
3044 * is cached. In such cases bio_start_transaction() may be called on
3045 * the bypassed layers so the system gets an I/O in progress indication
3046 * for those higher layers.
3049 bio_start_transaction(struct bio
*bio
, struct bio_track
*track
)
3051 bio
->bio_track
= track
;
3052 atomic_add_int(&track
->bk_active
, 1);
3056 * Initiate I/O on a vnode.
3059 vn_strategy(struct vnode
*vp
, struct bio
*bio
)
3061 struct bio_track
*track
;
3063 KKASSERT(bio
->bio_buf
->b_cmd
!= BUF_CMD_DONE
);
3064 if (bio
->bio_buf
->b_cmd
== BUF_CMD_READ
)
3065 track
= &vp
->v_track_read
;
3067 track
= &vp
->v_track_write
;
3068 bio
->bio_track
= track
;
3069 atomic_add_int(&track
->bk_active
, 1);
3070 vop_strategy(*vp
->v_ops
, vp
, bio
);
3077 * Finish I/O on a buffer, optionally calling a completion function.
3078 * This is usually called from an interrupt so process blocking is
3081 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3082 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3083 * assuming B_INVAL is clear.
3085 * For the VMIO case, we set B_CACHE if the op was a read and no
3086 * read error occured, or if the op was a write. B_CACHE is never
3087 * set if the buffer is invalid or otherwise uncacheable.
3089 * biodone does not mess with B_INVAL, allowing the I/O routine or the
3090 * initiator to leave B_INVAL set to brelse the buffer out of existance
3091 * in the biodone routine.
3094 biodone(struct bio
*bio
)
3096 struct buf
*bp
= bio
->bio_buf
;
3101 KASSERT(BUF_REFCNTNB(bp
) > 0,
3102 ("biodone: bp %p not busy %d", bp
, BUF_REFCNTNB(bp
)));
3103 KASSERT(bp
->b_cmd
!= BUF_CMD_DONE
,
3104 ("biodone: bp %p already done!", bp
));
3106 runningbufwakeup(bp
);
3109 * Run up the chain of BIO's. Leave b_cmd intact for the duration.
3112 biodone_t
*done_func
;
3113 struct bio_track
*track
;
3116 * BIO tracking. Most but not all BIOs are tracked.
3118 if ((track
= bio
->bio_track
) != NULL
) {
3119 atomic_subtract_int(&track
->bk_active
, 1);
3120 if (track
->bk_active
< 0) {
3121 panic("biodone: bad active count bio %p\n",
3124 if (track
->bk_waitflag
) {
3125 track
->bk_waitflag
= 0;
3128 bio
->bio_track
= NULL
;
3132 * A bio_done function terminates the loop. The function
3133 * will be responsible for any further chaining and/or
3134 * buffer management.
3136 * WARNING! The done function can deallocate the buffer!
3138 if ((done_func
= bio
->bio_done
) != NULL
) {
3139 bio
->bio_done
= NULL
;
3144 bio
= bio
->bio_prev
;
3148 bp
->b_cmd
= BUF_CMD_DONE
;
3151 * Only reads and writes are processed past this point.
3153 if (cmd
!= BUF_CMD_READ
&& cmd
!= BUF_CMD_WRITE
) {
3160 * Warning: softupdates may re-dirty the buffer.
3162 if (LIST_FIRST(&bp
->b_dep
) != NULL
)
3165 if (bp
->b_flags
& B_VMIO
) {
3171 struct vnode
*vp
= bp
->b_vp
;
3175 #if defined(VFS_BIO_DEBUG)
3176 if (vp
->v_auxrefs
== 0)
3177 panic("biodone: zero vnode hold count");
3178 if ((vp
->v_flag
& VOBJBUF
) == 0)
3179 panic("biodone: vnode is not setup for merged cache");
3182 foff
= bp
->b_loffset
;
3183 KASSERT(foff
!= NOOFFSET
, ("biodone: no buffer offset"));
3184 KASSERT(obj
!= NULL
, ("biodone: missing VM object"));
3186 #if defined(VFS_BIO_DEBUG)
3187 if (obj
->paging_in_progress
< bp
->b_xio
.xio_npages
) {
3188 kprintf("biodone: paging in progress(%d) < bp->b_xio.xio_npages(%d)\n",
3189 obj
->paging_in_progress
, bp
->b_xio
.xio_npages
);
3194 * Set B_CACHE if the op was a normal read and no error
3195 * occured. B_CACHE is set for writes in the b*write()
3198 iosize
= bp
->b_bcount
- bp
->b_resid
;
3199 if (cmd
== BUF_CMD_READ
&& (bp
->b_flags
& (B_INVAL
|B_NOCACHE
|B_ERROR
)) == 0) {
3200 bp
->b_flags
|= B_CACHE
;
3203 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
3207 resid
= ((foff
+ PAGE_SIZE
) & ~(off_t
)PAGE_MASK
) - foff
;
3212 * cleanup bogus pages, restoring the originals. Since
3213 * the originals should still be wired, we don't have
3214 * to worry about interrupt/freeing races destroying
3215 * the VM object association.
3217 m
= bp
->b_xio
.xio_pages
[i
];
3218 if (m
== bogus_page
) {
3220 m
= vm_page_lookup(obj
, OFF_TO_IDX(foff
));
3222 panic("biodone: page disappeared");
3223 bp
->b_xio
.xio_pages
[i
] = m
;
3224 pmap_qenter(trunc_page((vm_offset_t
)bp
->b_data
),
3225 bp
->b_xio
.xio_pages
, bp
->b_xio
.xio_npages
);
3227 #if defined(VFS_BIO_DEBUG)
3228 if (OFF_TO_IDX(foff
) != m
->pindex
) {
3230 "biodone: foff(%lu)/m->pindex(%d) mismatch\n",
3231 (unsigned long)foff
, m
->pindex
);
3236 * In the write case, the valid and clean bits are
3237 * already changed correctly ( see bdwrite() ), so we
3238 * only need to do this here in the read case.
3240 if (cmd
== BUF_CMD_READ
&& !bogusflag
&& resid
> 0) {
3241 vfs_page_set_valid(bp
, foff
, i
, m
);
3243 vm_page_flag_clear(m
, PG_ZERO
);
3246 * when debugging new filesystems or buffer I/O methods, this
3247 * is the most common error that pops up. if you see this, you
3248 * have not set the page busy flag correctly!!!
3251 kprintf("biodone: page busy < 0, "
3252 "pindex: %d, foff: 0x(%x,%x), "
3253 "resid: %d, index: %d\n",
3254 (int) m
->pindex
, (int)(foff
>> 32),
3255 (int) foff
& 0xffffffff, resid
, i
);
3256 if (!vn_isdisk(vp
, NULL
))
3257 kprintf(" iosize: %ld, loffset: %lld, flags: 0x%08x, npages: %d\n",
3258 bp
->b_vp
->v_mount
->mnt_stat
.f_iosize
,
3260 bp
->b_flags
, bp
->b_xio
.xio_npages
);
3262 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n",
3264 bp
->b_flags
, bp
->b_xio
.xio_npages
);
3265 kprintf(" valid: 0x%x, dirty: 0x%x, wired: %d\n",
3266 m
->valid
, m
->dirty
, m
->wire_count
);
3267 panic("biodone: page busy < 0");
3269 vm_page_io_finish(m
);
3270 vm_object_pip_subtract(obj
, 1);
3271 foff
= (foff
+ PAGE_SIZE
) & ~(off_t
)PAGE_MASK
;
3275 vm_object_pip_wakeupn(obj
, 0);
3279 * For asynchronous completions, release the buffer now. The brelse
3280 * will do a wakeup there if necessary - so no need to do a wakeup
3281 * here in the async case. The sync case always needs to do a wakeup.
3284 if (bp
->b_flags
& B_ASYNC
) {
3285 if ((bp
->b_flags
& (B_NOCACHE
| B_INVAL
| B_ERROR
| B_RELBUF
)) != 0)
3298 * This routine is called in lieu of iodone in the case of
3299 * incomplete I/O. This keeps the busy status for pages
3303 vfs_unbusy_pages(struct buf
*bp
)
3307 runningbufwakeup(bp
);
3308 if (bp
->b_flags
& B_VMIO
) {
3309 struct vnode
*vp
= bp
->b_vp
;
3314 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
3315 vm_page_t m
= bp
->b_xio
.xio_pages
[i
];
3318 * When restoring bogus changes the original pages
3319 * should still be wired, so we are in no danger of
3320 * losing the object association and do not need
3321 * critical section protection particularly.
3323 if (m
== bogus_page
) {
3324 m
= vm_page_lookup(obj
, OFF_TO_IDX(bp
->b_loffset
) + i
);
3326 panic("vfs_unbusy_pages: page missing");
3328 bp
->b_xio
.xio_pages
[i
] = m
;
3329 pmap_qenter(trunc_page((vm_offset_t
)bp
->b_data
),
3330 bp
->b_xio
.xio_pages
, bp
->b_xio
.xio_npages
);
3332 vm_object_pip_subtract(obj
, 1);
3333 vm_page_flag_clear(m
, PG_ZERO
);
3334 vm_page_io_finish(m
);
3336 vm_object_pip_wakeupn(obj
, 0);
3341 * vfs_page_set_valid:
3343 * Set the valid bits in a page based on the supplied offset. The
3344 * range is restricted to the buffer's size.
3346 * This routine is typically called after a read completes.
3349 vfs_page_set_valid(struct buf
*bp
, vm_ooffset_t off
, int pageno
, vm_page_t m
)
3351 vm_ooffset_t soff
, eoff
;
3354 * Start and end offsets in buffer. eoff - soff may not cross a
3355 * page boundry or cross the end of the buffer. The end of the
3356 * buffer, in this case, is our file EOF, not the allocation size
3360 eoff
= (off
+ PAGE_SIZE
) & ~(off_t
)PAGE_MASK
;
3361 if (eoff
> bp
->b_loffset
+ bp
->b_bcount
)
3362 eoff
= bp
->b_loffset
+ bp
->b_bcount
;
3365 * Set valid range. This is typically the entire buffer and thus the
3369 vm_page_set_validclean(
3371 (vm_offset_t
) (soff
& PAGE_MASK
),
3372 (vm_offset_t
) (eoff
- soff
)
3380 * This routine is called before a device strategy routine.
3381 * It is used to tell the VM system that paging I/O is in
3382 * progress, and treat the pages associated with the buffer
3383 * almost as being PG_BUSY. Also the object 'paging_in_progress'
3384 * flag is handled to make sure that the object doesn't become
3387 * Since I/O has not been initiated yet, certain buffer flags
3388 * such as B_ERROR or B_INVAL may be in an inconsistant state
3389 * and should be ignored.
3392 vfs_busy_pages(struct vnode
*vp
, struct buf
*bp
)
3395 struct lwp
*lp
= curthread
->td_lwp
;
3398 * The buffer's I/O command must already be set. If reading,
3399 * B_CACHE must be 0 (double check against callers only doing
3400 * I/O when B_CACHE is 0).
3402 KKASSERT(bp
->b_cmd
!= BUF_CMD_DONE
);
3403 KKASSERT(bp
->b_cmd
== BUF_CMD_WRITE
|| (bp
->b_flags
& B_CACHE
) == 0);
3405 if (bp
->b_flags
& B_VMIO
) {
3410 foff
= bp
->b_loffset
;
3411 KASSERT(bp
->b_loffset
!= NOOFFSET
,
3412 ("vfs_busy_pages: no buffer offset"));
3416 * Loop until none of the pages are busy.
3419 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
3420 vm_page_t m
= bp
->b_xio
.xio_pages
[i
];
3422 if (vm_page_sleep_busy(m
, FALSE
, "vbpage"))
3427 * Setup for I/O, soft-busy the page right now because
3428 * the next loop may block.
3430 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
3431 vm_page_t m
= bp
->b_xio
.xio_pages
[i
];
3433 vm_page_flag_clear(m
, PG_ZERO
);
3434 if ((bp
->b_flags
& B_CLUSTER
) == 0) {
3435 vm_object_pip_add(obj
, 1);
3436 vm_page_io_start(m
);
3441 * Adjust protections for I/O and do bogus-page mapping.
3442 * Assume that vm_page_protect() can block (it can block
3443 * if VM_PROT_NONE, don't take any chances regardless).
3446 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
3447 vm_page_t m
= bp
->b_xio
.xio_pages
[i
];
3450 * When readying a vnode-backed buffer for a write
3451 * we must zero-fill any invalid portions of the
3454 * When readying a vnode-backed buffer for a read
3455 * we must replace any dirty pages with a bogus
3456 * page so we do not destroy dirty data when
3457 * filling in gaps. Dirty pages might not
3458 * necessarily be marked dirty yet, so use m->valid
3459 * as a reasonable test.
3461 * Bogus page replacement is, uh, bogus. We need
3462 * to find a better way.
3464 if (bp
->b_cmd
== BUF_CMD_WRITE
) {
3465 vm_page_protect(m
, VM_PROT_READ
);
3466 vfs_page_set_valid(bp
, foff
, i
, m
);
3467 } else if (m
->valid
== VM_PAGE_BITS_ALL
) {
3468 bp
->b_xio
.xio_pages
[i
] = bogus_page
;
3471 vm_page_protect(m
, VM_PROT_NONE
);
3473 foff
= (foff
+ PAGE_SIZE
) & ~(off_t
)PAGE_MASK
;
3476 pmap_qenter(trunc_page((vm_offset_t
)bp
->b_data
),
3477 bp
->b_xio
.xio_pages
, bp
->b_xio
.xio_npages
);
3481 * This is the easiest place to put the process accounting for the I/O
3485 if (bp
->b_cmd
== BUF_CMD_READ
)
3486 lp
->lwp_ru
.ru_inblock
++;
3488 lp
->lwp_ru
.ru_oublock
++;
3495 * Tell the VM system that the pages associated with this buffer
3496 * are clean. This is used for delayed writes where the data is
3497 * going to go to disk eventually without additional VM intevention.
3499 * Note that while we only really need to clean through to b_bcount, we
3500 * just go ahead and clean through to b_bufsize.
3503 vfs_clean_pages(struct buf
*bp
)
3507 if (bp
->b_flags
& B_VMIO
) {
3510 foff
= bp
->b_loffset
;
3511 KASSERT(foff
!= NOOFFSET
, ("vfs_clean_pages: no buffer offset"));
3512 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
3513 vm_page_t m
= bp
->b_xio
.xio_pages
[i
];
3514 vm_ooffset_t noff
= (foff
+ PAGE_SIZE
) & ~(off_t
)PAGE_MASK
;
3515 vm_ooffset_t eoff
= noff
;
3517 if (eoff
> bp
->b_loffset
+ bp
->b_bufsize
)
3518 eoff
= bp
->b_loffset
+ bp
->b_bufsize
;
3519 vfs_page_set_valid(bp
, foff
, i
, m
);
3520 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3527 * vfs_bio_set_validclean:
3529 * Set the range within the buffer to valid and clean. The range is
3530 * relative to the beginning of the buffer, b_loffset. Note that
3531 * b_loffset itself may be offset from the beginning of the first page.
3535 vfs_bio_set_validclean(struct buf
*bp
, int base
, int size
)
3537 if (bp
->b_flags
& B_VMIO
) {
3542 * Fixup base to be relative to beginning of first page.
3543 * Set initial n to be the maximum number of bytes in the
3544 * first page that can be validated.
3547 base
+= (bp
->b_loffset
& PAGE_MASK
);
3548 n
= PAGE_SIZE
- (base
& PAGE_MASK
);
3550 for (i
= base
/ PAGE_SIZE
; size
> 0 && i
< bp
->b_xio
.xio_npages
; ++i
) {
3551 vm_page_t m
= bp
->b_xio
.xio_pages
[i
];
3556 vm_page_set_validclean(m
, base
& PAGE_MASK
, n
);
3567 * Clear a buffer. This routine essentially fakes an I/O, so we need
3568 * to clear B_ERROR and B_INVAL.
3570 * Note that while we only theoretically need to clear through b_bcount,
3571 * we go ahead and clear through b_bufsize.
3575 vfs_bio_clrbuf(struct buf
*bp
)
3579 if ((bp
->b_flags
& (B_VMIO
| B_MALLOC
)) == B_VMIO
) {
3580 bp
->b_flags
&= ~(B_INVAL
|B_ERROR
);
3581 if ((bp
->b_xio
.xio_npages
== 1) && (bp
->b_bufsize
< PAGE_SIZE
) &&
3582 (bp
->b_loffset
& PAGE_MASK
) == 0) {
3583 mask
= (1 << (bp
->b_bufsize
/ DEV_BSIZE
)) - 1;
3584 if ((bp
->b_xio
.xio_pages
[0]->valid
& mask
) == mask
) {
3588 if (((bp
->b_xio
.xio_pages
[0]->flags
& PG_ZERO
) == 0) &&
3589 ((bp
->b_xio
.xio_pages
[0]->valid
& mask
) == 0)) {
3590 bzero(bp
->b_data
, bp
->b_bufsize
);
3591 bp
->b_xio
.xio_pages
[0]->valid
|= mask
;
3596 ea
= sa
= bp
->b_data
;
3597 for(i
=0;i
<bp
->b_xio
.xio_npages
;i
++,sa
=ea
) {
3598 int j
= ((vm_offset_t
)sa
& PAGE_MASK
) / DEV_BSIZE
;
3599 ea
= (caddr_t
)trunc_page((vm_offset_t
)sa
+ PAGE_SIZE
);
3600 ea
= (caddr_t
)(vm_offset_t
)ulmin(
3601 (u_long
)(vm_offset_t
)ea
,
3602 (u_long
)(vm_offset_t
)bp
->b_data
+ bp
->b_bufsize
);
3603 mask
= ((1 << ((ea
- sa
) / DEV_BSIZE
)) - 1) << j
;
3604 if ((bp
->b_xio
.xio_pages
[i
]->valid
& mask
) == mask
)
3606 if ((bp
->b_xio
.xio_pages
[i
]->valid
& mask
) == 0) {
3607 if ((bp
->b_xio
.xio_pages
[i
]->flags
& PG_ZERO
) == 0) {
3611 for (; sa
< ea
; sa
+= DEV_BSIZE
, j
++) {
3612 if (((bp
->b_xio
.xio_pages
[i
]->flags
& PG_ZERO
) == 0) &&
3613 (bp
->b_xio
.xio_pages
[i
]->valid
& (1<<j
)) == 0)
3614 bzero(sa
, DEV_BSIZE
);
3617 bp
->b_xio
.xio_pages
[i
]->valid
|= mask
;
3618 vm_page_flag_clear(bp
->b_xio
.xio_pages
[i
], PG_ZERO
);
3627 * vm_hold_load_pages:
3629 * Load pages into the buffer's address space. The pages are
3630 * allocated from the kernel object in order to reduce interference
3631 * with the any VM paging I/O activity. The range of loaded
3632 * pages will be wired.
3634 * If a page cannot be allocated, the 'pagedaemon' is woken up to
3635 * retrieve the full range (to - from) of pages.
3639 vm_hold_load_pages(struct buf
*bp
, vm_offset_t from
, vm_offset_t to
)
3645 to
= round_page(to
);
3646 from
= round_page(from
);
3647 index
= (from
- trunc_page((vm_offset_t
)bp
->b_data
)) >> PAGE_SHIFT
;
3652 * Note: must allocate system pages since blocking here
3653 * could intefere with paging I/O, no matter which
3656 p
= bio_page_alloc(&kernel_object
, pg
>> PAGE_SHIFT
,
3657 (vm_pindex_t
)((to
- pg
) >> PAGE_SHIFT
));
3660 p
->valid
= VM_PAGE_BITS_ALL
;
3661 vm_page_flag_clear(p
, PG_ZERO
);
3662 pmap_kenter(pg
, VM_PAGE_TO_PHYS(p
));
3663 bp
->b_xio
.xio_pages
[index
] = p
;
3670 bp
->b_xio
.xio_npages
= index
;
3674 * Allocate pages for a buffer cache buffer.
3676 * Under extremely severe memory conditions even allocating out of the
3677 * system reserve can fail. If this occurs we must allocate out of the
3678 * interrupt reserve to avoid a deadlock with the pageout daemon.
3680 * The pageout daemon can run (putpages -> VOP_WRITE -> getblk -> allocbuf).
3681 * If the buffer cache's vm_page_alloc() fails a vm_wait() can deadlock
3682 * against the pageout daemon if pages are not freed from other sources.
3686 bio_page_alloc(vm_object_t obj
, vm_pindex_t pg
, int deficit
)
3691 * Try a normal allocation, allow use of system reserve.
3693 p
= vm_page_alloc(obj
, pg
, VM_ALLOC_NORMAL
| VM_ALLOC_SYSTEM
);
3698 * The normal allocation failed and we clearly have a page
3699 * deficit. Try to reclaim some clean VM pages directly
3700 * from the buffer cache.
3702 vm_pageout_deficit
+= deficit
;
3706 * We may have blocked, the caller will know what to do if the
3709 if (vm_page_lookup(obj
, pg
))
3713 * Allocate and allow use of the interrupt reserve.
3715 * If after all that we still can't allocate a VM page we are
3716 * in real trouble, but we slog on anyway hoping that the system
3719 p
= vm_page_alloc(obj
, pg
, VM_ALLOC_NORMAL
| VM_ALLOC_SYSTEM
|
3720 VM_ALLOC_INTERRUPT
);
3722 if (vm_page_count_severe()) {
3723 kprintf("bio_page_alloc: WARNING emergency page "
3728 kprintf("bio_page_alloc: WARNING emergency page "
3729 "allocation failed\n");
3736 * vm_hold_free_pages:
3738 * Return pages associated with the buffer back to the VM system.
3740 * The range of pages underlying the buffer's address space will
3741 * be unmapped and un-wired.
3744 vm_hold_free_pages(struct buf
*bp
, vm_offset_t from
, vm_offset_t to
)
3748 int index
, newnpages
;
3750 from
= round_page(from
);
3751 to
= round_page(to
);
3752 newnpages
= index
= (from
- trunc_page((vm_offset_t
)bp
->b_data
)) >> PAGE_SHIFT
;
3754 for (pg
= from
; pg
< to
; pg
+= PAGE_SIZE
, index
++) {
3755 p
= bp
->b_xio
.xio_pages
[index
];
3756 if (p
&& (index
< bp
->b_xio
.xio_npages
)) {
3758 kprintf("vm_hold_free_pages: doffset: %lld, loffset: %lld\n",
3759 bp
->b_bio2
.bio_offset
, bp
->b_loffset
);
3761 bp
->b_xio
.xio_pages
[index
] = NULL
;
3764 vm_page_unwire(p
, 0);
3768 bp
->b_xio
.xio_npages
= newnpages
;
3774 * Map a user buffer into KVM via a pbuf. On return the buffer's
3775 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array
3779 vmapbuf(struct buf
*bp
, caddr_t udata
, int bytes
)
3790 * bp had better have a command and it better be a pbuf.
3792 KKASSERT(bp
->b_cmd
!= BUF_CMD_DONE
);
3793 KKASSERT(bp
->b_flags
& B_PAGING
);
3799 * Map the user data into KVM. Mappings have to be page-aligned.
3801 addr
= (caddr_t
)trunc_page((vm_offset_t
)udata
);
3804 vmprot
= VM_PROT_READ
;
3805 if (bp
->b_cmd
== BUF_CMD_READ
)
3806 vmprot
|= VM_PROT_WRITE
;
3808 while (addr
< udata
+ bytes
) {
3810 * Do the vm_fault if needed; do the copy-on-write thing
3811 * when reading stuff off device into memory.
3813 * vm_fault_page*() returns a held VM page.
3815 va
= (addr
>= udata
) ? (vm_offset_t
)addr
: (vm_offset_t
)udata
;
3816 va
= trunc_page(va
);
3818 m
= vm_fault_page_quick(va
, vmprot
, &error
);
3820 for (i
= 0; i
< pidx
; ++i
) {
3821 vm_page_unhold(bp
->b_xio
.xio_pages
[i
]);
3822 bp
->b_xio
.xio_pages
[i
] = NULL
;
3826 bp
->b_xio
.xio_pages
[pidx
] = m
;
3832 * Map the page array and set the buffer fields to point to
3833 * the mapped data buffer.
3835 if (pidx
> btoc(MAXPHYS
))
3836 panic("vmapbuf: mapped more than MAXPHYS");
3837 pmap_qenter((vm_offset_t
)bp
->b_kvabase
, bp
->b_xio
.xio_pages
, pidx
);
3839 bp
->b_xio
.xio_npages
= pidx
;
3840 bp
->b_data
= bp
->b_kvabase
+ ((int)(intptr_t)udata
& PAGE_MASK
);
3841 bp
->b_bcount
= bytes
;
3842 bp
->b_bufsize
= bytes
;
3849 * Free the io map PTEs associated with this IO operation.
3850 * We also invalidate the TLB entries and restore the original b_addr.
3853 vunmapbuf(struct buf
*bp
)
3858 KKASSERT(bp
->b_flags
& B_PAGING
);
3860 npages
= bp
->b_xio
.xio_npages
;
3861 pmap_qremove(trunc_page((vm_offset_t
)bp
->b_data
), npages
);
3862 for (pidx
= 0; pidx
< npages
; ++pidx
) {
3863 vm_page_unhold(bp
->b_xio
.xio_pages
[pidx
]);
3864 bp
->b_xio
.xio_pages
[pidx
] = NULL
;
3866 bp
->b_xio
.xio_npages
= 0;
3867 bp
->b_data
= bp
->b_kvabase
;
3871 * Scan all buffers in the system and issue the callback.
3874 scan_all_buffers(int (*callback
)(struct buf
*, void *), void *info
)
3880 for (n
= 0; n
< nbuf
; ++n
) {
3881 if ((error
= callback(&buf
[n
], info
)) < 0) {
3891 * print out statistics from the current status of the buffer pool
3892 * this can be toggeled by the system control option debug.syncprt
3901 int counts
[(MAXBSIZE
/ PAGE_SIZE
) + 1];
3902 static char *bname
[3] = { "LOCKED", "LRU", "AGE" };
3904 for (dp
= bufqueues
, i
= 0; dp
< &bufqueues
[3]; dp
++, i
++) {
3906 for (j
= 0; j
<= MAXBSIZE
/PAGE_SIZE
; j
++)
3909 TAILQ_FOREACH(bp
, dp
, b_freelist
) {
3910 counts
[bp
->b_bufsize
/PAGE_SIZE
]++;
3914 kprintf("%s: total-%d", bname
[i
], count
);
3915 for (j
= 0; j
<= MAXBSIZE
/PAGE_SIZE
; j
++)
3917 kprintf(", %d-%d", j
* PAGE_SIZE
, counts
[j
]);
3925 DB_SHOW_COMMAND(buffer
, db_show_buffer
)
3928 struct buf
*bp
= (struct buf
*)addr
;
3931 db_printf("usage: show buffer <addr>\n");
3935 db_printf("b_flags = 0x%b\n", (u_int
)bp
->b_flags
, PRINT_BUF_FLAGS
);
3936 db_printf("b_cmd = %d\n", bp
->b_cmd
);
3937 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, "
3938 "b_resid = %d\n, b_data = %p, "
3939 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n",
3940 bp
->b_error
, bp
->b_bufsize
, bp
->b_bcount
, bp
->b_resid
,
3941 bp
->b_data
, bp
->b_bio2
.bio_offset
, (bp
->b_bio2
.bio_next
? bp
->b_bio2
.bio_next
->bio_offset
: (off_t
)-1));
3942 if (bp
->b_xio
.xio_npages
) {
3944 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ",
3945 bp
->b_xio
.xio_npages
);
3946 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
3948 m
= bp
->b_xio
.xio_pages
[i
];
3949 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m
->object
,
3950 (u_long
)m
->pindex
, (u_long
)VM_PAGE_TO_PHYS(m
));
3951 if ((i
+ 1) < bp
->b_xio
.xio_npages
)