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 $
18 * this file contains a new buffer I/O scheme implementing a coherent
19 * VM object and buffer cache scheme. Pains have been taken to make
20 * sure that the performance degradation associated with schemes such
21 * as this is not realized.
23 * Author: John S. Dyson
24 * Significant help during the development and debugging phases
25 * had been provided by David Greenman, also of the FreeBSD core team.
27 * see man buf(9) for more info. Note that man buf(9) doesn't reflect
28 * the actual buf/bio implementation in DragonFly.
31 #include <sys/param.h>
32 #include <sys/systm.h>
35 #include <sys/devicestat.h>
36 #include <sys/eventhandler.h>
38 #include <sys/malloc.h>
39 #include <sys/mount.h>
40 #include <sys/kernel.h>
41 #include <sys/kthread.h>
43 #include <sys/reboot.h>
44 #include <sys/resourcevar.h>
45 #include <sys/sysctl.h>
46 #include <sys/vmmeter.h>
47 #include <sys/vnode.h>
48 #include <sys/dsched.h>
50 #include <vm/vm_param.h>
51 #include <vm/vm_kern.h>
52 #include <vm/vm_pageout.h>
53 #include <vm/vm_page.h>
54 #include <vm/vm_object.h>
55 #include <vm/vm_extern.h>
56 #include <vm/vm_map.h>
57 #include <vm/vm_pager.h>
58 #include <vm/swap_pager.h>
61 #include <sys/thread2.h>
62 #include <sys/spinlock2.h>
63 #include <vm/vm_page2.h>
74 BQUEUE_NONE
, /* not on any queue */
75 BQUEUE_LOCKED
, /* locked buffers */
76 BQUEUE_CLEAN
, /* non-B_DELWRI buffers */
77 BQUEUE_DIRTY
, /* B_DELWRI buffers */
78 BQUEUE_DIRTY_HW
, /* B_DELWRI buffers - heavy weight */
79 BQUEUE_EMPTY
, /* empty buffer headers */
81 BUFFER_QUEUES
/* number of buffer queues */
84 typedef enum bufq_type bufq_type_t
;
86 #define BD_WAKE_SIZE 16384
87 #define BD_WAKE_MASK (BD_WAKE_SIZE - 1)
89 TAILQ_HEAD(bqueues
, buf
);
93 struct bqueues bufqueues
[BUFFER_QUEUES
];
96 struct bufpcpu bufpcpu
[MAXCPU
];
98 static MALLOC_DEFINE(M_BIOBUF
, "BIO buffer", "BIO buffer");
100 struct buf
*buf
; /* buffer header pool */
102 static void vfs_clean_pages(struct buf
*bp
);
103 static void vfs_clean_one_page(struct buf
*bp
, int pageno
, vm_page_t m
);
105 static void vfs_dirty_one_page(struct buf
*bp
, int pageno
, vm_page_t m
);
107 static void vfs_vmio_release(struct buf
*bp
);
108 static int flushbufqueues(struct buf
*marker
, bufq_type_t q
);
109 static vm_page_t
bio_page_alloc(struct buf
*bp
, vm_object_t obj
,
110 vm_pindex_t pg
, int deficit
);
112 static void bd_signal(long totalspace
);
113 static void buf_daemon(void);
114 static void buf_daemon_hw(void);
117 * bogus page -- for I/O to/from partially complete buffers
118 * this is a temporary solution to the problem, but it is not
119 * really that bad. it would be better to split the buffer
120 * for input in the case of buffers partially already in memory,
121 * but the code is intricate enough already.
123 vm_page_t bogus_page
;
126 * These are all static, but make the ones we export globals so we do
127 * not need to use compiler magic.
129 long bufspace
; /* atomic ops */
131 long maxbufmallocspace
, lobufspace
, hibufspace
;
132 static long lorunningspace
;
133 static long hirunningspace
;
134 static long dirtykvaspace
; /* atomic */
135 long dirtybufspace
; /* atomic (global for systat) */
136 static long dirtybufcount
; /* atomic */
137 static long dirtybufspacehw
; /* atomic */
138 static long dirtybufcounthw
; /* atomic */
139 static long runningbufspace
; /* atomic */
140 static long runningbufcount
; /* atomic */
141 long lodirtybufspace
;
142 long hidirtybufspace
;
143 static int getnewbufcalls
;
144 static int needsbuffer
; /* atomic */
145 static int runningbufreq
; /* atomic */
146 static int bd_request
; /* atomic */
147 static int bd_request_hw
; /* atomic */
148 static u_int bd_wake_ary
[BD_WAKE_SIZE
];
149 static u_int bd_wake_index
;
150 static u_int vm_cycle_point
= 40; /* 23-36 will migrate more act->inact */
151 static int debug_commit
;
152 static int debug_bufbio
;
153 static int debug_kvabio
;
154 static long bufcache_bw
= 200 * 1024 * 1024;
156 static struct thread
*bufdaemon_td
;
157 static struct thread
*bufdaemonhw_td
;
158 static u_int lowmempgallocs
;
159 static u_int flushperqueue
= 1024;
162 * Sysctls for operational control of the buffer cache.
164 SYSCTL_UINT(_vfs
, OID_AUTO
, flushperqueue
, CTLFLAG_RW
, &flushperqueue
, 0,
165 "Number of buffers to flush from each per-cpu queue");
166 SYSCTL_LONG(_vfs
, OID_AUTO
, lodirtybufspace
, CTLFLAG_RW
, &lodirtybufspace
, 0,
167 "Number of dirty buffers to flush before bufdaemon becomes inactive");
168 SYSCTL_LONG(_vfs
, OID_AUTO
, hidirtybufspace
, CTLFLAG_RW
, &hidirtybufspace
, 0,
169 "High watermark used to trigger explicit flushing of dirty buffers");
170 SYSCTL_LONG(_vfs
, OID_AUTO
, lorunningspace
, CTLFLAG_RW
, &lorunningspace
, 0,
171 "Minimum amount of buffer space required for active I/O");
172 SYSCTL_LONG(_vfs
, OID_AUTO
, hirunningspace
, CTLFLAG_RW
, &hirunningspace
, 0,
173 "Maximum amount of buffer space to usable for active I/O");
174 SYSCTL_LONG(_vfs
, OID_AUTO
, bufcache_bw
, CTLFLAG_RW
, &bufcache_bw
, 0,
175 "Buffer-cache -> VM page cache transfer bandwidth");
176 SYSCTL_UINT(_vfs
, OID_AUTO
, lowmempgallocs
, CTLFLAG_RW
, &lowmempgallocs
, 0,
177 "Page allocations done during periods of very low free memory");
178 SYSCTL_UINT(_vfs
, OID_AUTO
, vm_cycle_point
, CTLFLAG_RW
, &vm_cycle_point
, 0,
179 "Recycle pages to active or inactive queue transition pt 0-64");
181 * Sysctls determining current state of the buffer cache.
183 SYSCTL_LONG(_vfs
, OID_AUTO
, nbuf
, CTLFLAG_RD
, &nbuf
, 0,
184 "Total number of buffers in buffer cache");
185 SYSCTL_LONG(_vfs
, OID_AUTO
, dirtykvaspace
, CTLFLAG_RD
, &dirtykvaspace
, 0,
186 "KVA reserved by dirty buffers (all)");
187 SYSCTL_LONG(_vfs
, OID_AUTO
, dirtybufspace
, CTLFLAG_RD
, &dirtybufspace
, 0,
188 "Pending bytes of dirty buffers (all)");
189 SYSCTL_LONG(_vfs
, OID_AUTO
, dirtybufspacehw
, CTLFLAG_RD
, &dirtybufspacehw
, 0,
190 "Pending bytes of dirty buffers (heavy weight)");
191 SYSCTL_LONG(_vfs
, OID_AUTO
, dirtybufcount
, CTLFLAG_RD
, &dirtybufcount
, 0,
192 "Pending number of dirty buffers");
193 SYSCTL_LONG(_vfs
, OID_AUTO
, dirtybufcounthw
, CTLFLAG_RD
, &dirtybufcounthw
, 0,
194 "Pending number of dirty buffers (heavy weight)");
195 SYSCTL_LONG(_vfs
, OID_AUTO
, runningbufspace
, CTLFLAG_RD
, &runningbufspace
, 0,
196 "I/O bytes currently in progress due to asynchronous writes");
197 SYSCTL_LONG(_vfs
, OID_AUTO
, runningbufcount
, CTLFLAG_RD
, &runningbufcount
, 0,
198 "I/O buffers currently in progress due to asynchronous writes");
199 SYSCTL_LONG(_vfs
, OID_AUTO
, maxbufspace
, CTLFLAG_RD
, &maxbufspace
, 0,
200 "Hard limit on maximum amount of memory usable for buffer space");
201 SYSCTL_LONG(_vfs
, OID_AUTO
, hibufspace
, CTLFLAG_RD
, &hibufspace
, 0,
202 "Soft limit on maximum amount of memory usable for buffer space");
203 SYSCTL_LONG(_vfs
, OID_AUTO
, lobufspace
, CTLFLAG_RD
, &lobufspace
, 0,
204 "Minimum amount of memory to reserve for system buffer space");
205 SYSCTL_LONG(_vfs
, OID_AUTO
, bufspace
, CTLFLAG_RD
, &bufspace
, 0,
206 "Amount of memory available for buffers");
207 SYSCTL_LONG(_vfs
, OID_AUTO
, maxmallocbufspace
, CTLFLAG_RD
, &maxbufmallocspace
,
208 0, "Maximum amount of memory reserved for buffers using malloc");
209 SYSCTL_INT(_vfs
, OID_AUTO
, getnewbufcalls
, CTLFLAG_RD
, &getnewbufcalls
, 0,
210 "New buffer header acquisition requests");
211 SYSCTL_INT(_vfs
, OID_AUTO
, debug_commit
, CTLFLAG_RW
, &debug_commit
, 0, "");
212 SYSCTL_INT(_vfs
, OID_AUTO
, debug_bufbio
, CTLFLAG_RW
, &debug_bufbio
, 0, "");
213 SYSCTL_INT(_vfs
, OID_AUTO
, debug_kvabio
, CTLFLAG_RW
, &debug_kvabio
, 0, "");
214 SYSCTL_INT(_debug_sizeof
, OID_AUTO
, buf
, CTLFLAG_RD
, 0, sizeof(struct buf
),
215 "sizeof(struct buf)");
217 char *buf_wmesg
= BUF_WMESG
;
219 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
220 #define VFS_BIO_NEED_UNUSED02 0x02
221 #define VFS_BIO_NEED_UNUSED04 0x04
222 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
225 * Called when buffer space is potentially available for recovery.
226 * getnewbuf() will block on this flag when it is unable to free
227 * sufficient buffer space. Buffer space becomes recoverable when
228 * bp's get placed back in the queues.
234 * If someone is waiting for BUF space, wake them up. Even
235 * though we haven't freed the kva space yet, the waiting
236 * process will be able to now.
239 int flags
= needsbuffer
;
241 if ((flags
& VFS_BIO_NEED_BUFSPACE
) == 0)
243 if (atomic_cmpset_int(&needsbuffer
, flags
,
244 flags
& ~VFS_BIO_NEED_BUFSPACE
)) {
245 wakeup(&needsbuffer
);
255 * Accounting for I/O in progress.
259 runningbufwakeup(struct buf
*bp
)
264 if ((totalspace
= bp
->b_runningbufspace
) != 0) {
265 atomic_add_long(&runningbufspace
, -totalspace
);
266 atomic_add_long(&runningbufcount
, -1);
267 bp
->b_runningbufspace
= 0;
270 * see waitrunningbufspace() for limit test.
273 flags
= runningbufreq
;
277 if (atomic_cmpset_int(&runningbufreq
, flags
, 0)) {
278 wakeup(&runningbufreq
);
283 bd_signal(totalspace
);
290 * Called when a buffer has been added to one of the free queues to
291 * account for the buffer and to wakeup anyone waiting for free buffers.
292 * This typically occurs when large amounts of metadata are being handled
293 * by the buffer cache ( else buffer space runs out first, usually ).
304 if (atomic_cmpset_int(&needsbuffer
, flags
,
305 (flags
& ~VFS_BIO_NEED_ANY
))) {
306 wakeup(&needsbuffer
);
314 * waitrunningbufspace()
316 * If runningbufspace exceeds 4/6 hirunningspace we block until
317 * runningbufspace drops to 3/6 hirunningspace. We also block if another
318 * thread blocked here in order to be fair, even if runningbufspace
319 * is now lower than the limit.
321 * The caller may be using this function to block in a tight loop, we
322 * must block while runningbufspace is greater than at least
323 * hirunningspace * 3 / 6.
326 waitrunningbufspace(void)
328 long limit
= hirunningspace
* 4 / 6;
331 while (runningbufspace
> limit
|| runningbufreq
) {
332 tsleep_interlock(&runningbufreq
, 0);
333 flags
= atomic_fetchadd_int(&runningbufreq
, 1);
334 if (runningbufspace
> limit
|| flags
)
335 tsleep(&runningbufreq
, PINTERLOCKED
, "wdrn1", hz
);
340 * buf_dirty_count_severe:
342 * Return true if we have too many dirty buffers.
345 buf_dirty_count_severe(void)
347 return (runningbufspace
+ dirtykvaspace
>= hidirtybufspace
||
348 dirtybufcount
>= nbuf
/ 2);
352 * Return true if the amount of running I/O is severe and BIOQ should
356 buf_runningbufspace_severe(void)
358 return (runningbufspace
>= hirunningspace
* 4 / 6);
362 * vfs_buf_test_cache:
364 * Called when a buffer is extended. This function clears the B_CACHE
365 * bit if the newly extended portion of the buffer does not contain
368 * NOTE! Dirty VM pages are not processed into dirty (B_DELWRI) buffer
369 * cache buffers. The VM pages remain dirty, as someone had mmap()'d
370 * them while a clean buffer was present.
374 vfs_buf_test_cache(struct buf
*bp
,
375 vm_ooffset_t foff
, vm_offset_t off
, vm_offset_t size
,
378 if (bp
->b_flags
& B_CACHE
) {
379 int base
= (foff
+ off
) & PAGE_MASK
;
380 if (vm_page_is_valid(m
, base
, size
) == 0)
381 bp
->b_flags
&= ~B_CACHE
;
388 * Spank the buf_daemon[_hw] if the total dirty buffer space exceeds the
395 if (dirtykvaspace
< lodirtybufspace
&& dirtybufcount
< nbuf
/ 2)
398 if (bd_request
== 0 &&
399 (dirtykvaspace
> lodirtybufspace
/ 2 ||
400 dirtybufcount
- dirtybufcounthw
>= nbuf
/ 2)) {
401 if (atomic_fetchadd_int(&bd_request
, 1) == 0)
404 if (bd_request_hw
== 0 &&
405 (dirtykvaspace
> lodirtybufspace
/ 2 ||
406 dirtybufcounthw
>= nbuf
/ 2)) {
407 if (atomic_fetchadd_int(&bd_request_hw
, 1) == 0)
408 wakeup(&bd_request_hw
);
415 * Get the buf_daemon heated up when the number of running and dirty
416 * buffers exceeds the mid-point.
418 * Return the total number of dirty bytes past the second mid point
419 * as a measure of how much excess dirty data there is in the system.
428 mid1
= lodirtybufspace
+ (hidirtybufspace
- lodirtybufspace
) / 2;
430 totalspace
= runningbufspace
+ dirtykvaspace
;
431 if (totalspace
>= mid1
|| dirtybufcount
>= nbuf
/ 2) {
433 mid2
= mid1
+ (hidirtybufspace
- mid1
) / 2;
434 if (totalspace
>= mid2
)
435 return(totalspace
- mid2
);
443 * Wait for the buffer cache to flush (totalspace) bytes worth of
444 * buffers, then return.
446 * Regardless this function blocks while the number of dirty buffers
447 * exceeds hidirtybufspace.
450 bd_wait(long totalspace
)
457 if (curthread
== bufdaemonhw_td
|| curthread
== bufdaemon_td
)
460 while (totalspace
> 0) {
464 * Order is important. Suppliers adjust bd_wake_index after
465 * updating runningbufspace/dirtykvaspace. We want to fetch
466 * bd_wake_index before accessing. Any error should thus
469 i
= atomic_fetchadd_int(&bd_wake_index
, 0);
470 if (totalspace
> runningbufspace
+ dirtykvaspace
)
471 totalspace
= runningbufspace
+ dirtykvaspace
;
472 count
= totalspace
/ MAXBSIZE
;
473 if (count
>= BD_WAKE_SIZE
/ 2)
474 count
= BD_WAKE_SIZE
/ 2;
476 mi
= i
& BD_WAKE_MASK
;
479 * This is not a strict interlock, so we play a bit loose
480 * with locking access to dirtybufspace*. We have to re-check
481 * bd_wake_index to ensure that it hasn't passed us.
483 tsleep_interlock(&bd_wake_ary
[mi
], 0);
484 atomic_add_int(&bd_wake_ary
[mi
], 1);
485 j
= atomic_fetchadd_int(&bd_wake_index
, 0);
486 if ((int)(i
- j
) >= 0)
487 tsleep(&bd_wake_ary
[mi
], PINTERLOCKED
, "flstik", hz
);
489 totalspace
= runningbufspace
+ dirtykvaspace
- hidirtybufspace
;
496 * This function is called whenever runningbufspace or dirtykvaspace
497 * is reduced. Track threads waiting for run+dirty buffer I/O
501 bd_signal(long totalspace
)
505 if (totalspace
> 0) {
506 if (totalspace
> MAXBSIZE
* BD_WAKE_SIZE
)
507 totalspace
= MAXBSIZE
* BD_WAKE_SIZE
;
508 while (totalspace
> 0) {
509 i
= atomic_fetchadd_int(&bd_wake_index
, 1);
511 if (atomic_readandclear_int(&bd_wake_ary
[i
]))
512 wakeup(&bd_wake_ary
[i
]);
513 totalspace
-= MAXBSIZE
;
519 * BIO tracking support routines.
521 * Release a ref on a bio_track. Wakeup requests are atomically released
522 * along with the last reference so bk_active will never wind up set to
527 bio_track_rel(struct bio_track
*track
)
535 active
= track
->bk_active
;
536 if (active
== 1 && atomic_cmpset_int(&track
->bk_active
, 1, 0))
540 * Full-on. Note that the wait flag is only atomically released on
541 * the 1->0 count transition.
543 * We check for a negative count transition using bit 30 since bit 31
544 * has a different meaning.
547 desired
= (active
& 0x7FFFFFFF) - 1;
549 desired
|= active
& 0x80000000;
550 if (atomic_cmpset_int(&track
->bk_active
, active
, desired
)) {
551 if (desired
& 0x40000000)
552 panic("bio_track_rel: bad count: %p", track
);
553 if (active
& 0x80000000)
557 active
= track
->bk_active
;
562 * Wait for the tracking count to reach 0.
564 * Use atomic ops such that the wait flag is only set atomically when
565 * bk_active is non-zero.
568 bio_track_wait(struct bio_track
*track
, int slp_flags
, int slp_timo
)
577 if (track
->bk_active
== 0)
581 * Full-on. Note that the wait flag may only be atomically set if
582 * the active count is non-zero.
584 * NOTE: We cannot optimize active == desired since a wakeup could
585 * clear active prior to our tsleep_interlock().
588 while ((active
= track
->bk_active
) != 0) {
590 desired
= active
| 0x80000000;
591 tsleep_interlock(track
, slp_flags
);
592 if (atomic_cmpset_int(&track
->bk_active
, active
, desired
)) {
593 error
= tsleep(track
, slp_flags
| PINTERLOCKED
,
605 * Load time initialisation of the buffer cache, called from machine
606 * dependant initialization code.
610 bufinit(void *dummy __unused
)
612 struct bufpcpu
*pcpu
;
614 vm_offset_t bogus_offset
;
619 /* next, make a null set of free lists */
620 for (i
= 0; i
< ncpus
; ++i
) {
622 spin_init(&pcpu
->spin
, "bufinit");
623 for (j
= 0; j
< BUFFER_QUEUES
; j
++)
624 TAILQ_INIT(&pcpu
->bufqueues
[j
]);
628 * Finally, initialize each buffer header and stick on empty q.
629 * Each buffer gets its own KVA reservation.
634 for (n
= 0; n
< nbuf
; n
++) {
636 bzero(bp
, sizeof *bp
);
637 bp
->b_flags
= B_INVAL
; /* we're just an empty header */
638 bp
->b_cmd
= BUF_CMD_DONE
;
639 bp
->b_qindex
= BQUEUE_EMPTY
;
641 bp
->b_kvabase
= (void *)(vm_map_min(&buffer_map
) +
643 bp
->b_kvasize
= MAXBSIZE
;
645 xio_init(&bp
->b_xio
);
647 TAILQ_INSERT_TAIL(&pcpu
->bufqueues
[bp
->b_qindex
],
655 * maxbufspace is the absolute maximum amount of buffer space we are
656 * allowed to reserve in KVM and in real terms. The absolute maximum
657 * is nominally used by buf_daemon. hibufspace is the nominal maximum
658 * used by most other processes. The differential is required to
659 * ensure that buf_daemon is able to run when other processes might
660 * be blocked waiting for buffer space.
662 * Calculate hysteresis (lobufspace, hibufspace). Don't make it
663 * too large or we might lockup a cpu for too long a period of
664 * time in our tight loop.
666 maxbufspace
= nbuf
* NBUFCALCSIZE
;
667 hibufspace
= lmax(3 * maxbufspace
/ 4, maxbufspace
- MAXBSIZE
* 10);
668 lobufspace
= hibufspace
* 7 / 8;
669 if (hibufspace
- lobufspace
> 64 * 1024 * 1024)
670 lobufspace
= hibufspace
- 64 * 1024 * 1024;
671 if (lobufspace
> hibufspace
- MAXBSIZE
)
672 lobufspace
= hibufspace
- MAXBSIZE
;
674 lorunningspace
= 512 * 1024;
675 /* hirunningspace -- see below */
678 * Limit the amount of malloc memory since it is wired permanently
679 * into the kernel space. Even though this is accounted for in
680 * the buffer allocation, we don't want the malloced region to grow
681 * uncontrolled. The malloc scheme improves memory utilization
682 * significantly on average (small) directories.
684 maxbufmallocspace
= hibufspace
/ 20;
687 * Reduce the chance of a deadlock occuring by limiting the number
688 * of delayed-write dirty buffers we allow to stack up.
690 * We don't want too much actually queued to the device at once
691 * (XXX this needs to be per-mount!), because the buffers will
692 * wind up locked for a very long period of time while the I/O
695 hidirtybufspace
= hibufspace
/ 2; /* dirty + running */
696 hirunningspace
= hibufspace
/ 16; /* locked & queued to device */
697 if (hirunningspace
< 1024 * 1024)
698 hirunningspace
= 1024 * 1024;
704 lodirtybufspace
= hidirtybufspace
/ 2;
707 * Maximum number of async ops initiated per buf_daemon loop. This is
708 * somewhat of a hack at the moment, we really need to limit ourselves
709 * based on the number of bytes of I/O in-transit that were initiated
713 bogus_offset
= kmem_alloc_pageable(&kernel_map
, PAGE_SIZE
,
715 vm_object_hold(&kernel_object
);
716 bogus_page
= vm_page_alloc(&kernel_object
,
717 (bogus_offset
>> PAGE_SHIFT
),
719 vm_object_drop(&kernel_object
);
720 vmstats
.v_wire_count
++;
724 SYSINIT(do_bufinit
, SI_BOOT2_MACHDEP
, SI_ORDER_FIRST
, bufinit
, NULL
);
727 * Initialize the embedded bio structures, typically used by
728 * deprecated code which tries to allocate its own struct bufs.
731 initbufbio(struct buf
*bp
)
733 bp
->b_bio1
.bio_buf
= bp
;
734 bp
->b_bio1
.bio_prev
= NULL
;
735 bp
->b_bio1
.bio_offset
= NOOFFSET
;
736 bp
->b_bio1
.bio_next
= &bp
->b_bio2
;
737 bp
->b_bio1
.bio_done
= NULL
;
738 bp
->b_bio1
.bio_flags
= 0;
740 bp
->b_bio2
.bio_buf
= bp
;
741 bp
->b_bio2
.bio_prev
= &bp
->b_bio1
;
742 bp
->b_bio2
.bio_offset
= NOOFFSET
;
743 bp
->b_bio2
.bio_next
= NULL
;
744 bp
->b_bio2
.bio_done
= NULL
;
745 bp
->b_bio2
.bio_flags
= 0;
751 * Reinitialize the embedded bio structures as well as any additional
752 * translation cache layers.
755 reinitbufbio(struct buf
*bp
)
759 for (bio
= &bp
->b_bio1
; bio
; bio
= bio
->bio_next
) {
760 bio
->bio_done
= NULL
;
761 bio
->bio_offset
= NOOFFSET
;
766 * Undo the effects of an initbufbio().
769 uninitbufbio(struct buf
*bp
)
776 * Push another BIO layer onto an existing BIO and return it. The new
777 * BIO layer may already exist, holding cached translation data.
780 push_bio(struct bio
*bio
)
784 if ((nbio
= bio
->bio_next
) == NULL
) {
785 int index
= bio
- &bio
->bio_buf
->b_bio_array
[0];
786 if (index
>= NBUF_BIO
- 1) {
787 panic("push_bio: too many layers %d for bp %p",
788 index
, bio
->bio_buf
);
790 nbio
= &bio
->bio_buf
->b_bio_array
[index
+ 1];
791 bio
->bio_next
= nbio
;
792 nbio
->bio_prev
= bio
;
793 nbio
->bio_buf
= bio
->bio_buf
;
794 nbio
->bio_offset
= NOOFFSET
;
795 nbio
->bio_done
= NULL
;
796 nbio
->bio_next
= NULL
;
798 KKASSERT(nbio
->bio_done
== NULL
);
803 * Pop a BIO translation layer, returning the previous layer. The
804 * must have been previously pushed.
807 pop_bio(struct bio
*bio
)
809 return(bio
->bio_prev
);
813 clearbiocache(struct bio
*bio
)
816 bio
->bio_offset
= NOOFFSET
;
822 * Remove the buffer from the appropriate free list.
823 * (caller must be locked)
826 _bremfree(struct buf
*bp
)
828 struct bufpcpu
*pcpu
= &bufpcpu
[bp
->b_qcpu
];
830 if (bp
->b_qindex
!= BQUEUE_NONE
) {
831 KASSERT(BUF_LOCKINUSE(bp
), ("bremfree: bp %p not locked", bp
));
832 TAILQ_REMOVE(&pcpu
->bufqueues
[bp
->b_qindex
], bp
, b_freelist
);
833 bp
->b_qindex
= BQUEUE_NONE
;
835 if (!BUF_LOCKINUSE(bp
))
836 panic("bremfree: removing a buffer not on a queue");
841 * bremfree() - must be called with a locked buffer
844 bremfree(struct buf
*bp
)
846 struct bufpcpu
*pcpu
= &bufpcpu
[bp
->b_qcpu
];
848 spin_lock(&pcpu
->spin
);
850 spin_unlock(&pcpu
->spin
);
854 * bremfree_locked - must be called with pcpu->spin locked
857 bremfree_locked(struct buf
*bp
)
863 * This version of bread issues any required I/O asyncnronously and
864 * makes a callback on completion.
866 * The callback must check whether BIO_DONE is set in the bio and issue
867 * the bpdone(bp, 0) if it isn't. The callback is responsible for clearing
868 * BIO_DONE and disposing of the I/O (bqrelse()ing it).
871 breadcb(struct vnode
*vp
, off_t loffset
, int size
, int bflags
,
872 void (*func
)(struct bio
*), void *arg
)
876 bp
= getblk(vp
, loffset
, size
, 0, 0);
878 /* if not found in cache, do some I/O */
879 if ((bp
->b_flags
& B_CACHE
) == 0) {
880 bp
->b_flags
&= ~(B_ERROR
| B_EINTR
| B_INVAL
| B_NOTMETA
);
881 bp
->b_flags
|= bflags
;
882 bp
->b_cmd
= BUF_CMD_READ
;
883 bp
->b_bio1
.bio_done
= func
;
884 bp
->b_bio1
.bio_caller_info1
.ptr
= arg
;
885 vfs_busy_pages(vp
, bp
);
887 vn_strategy(vp
, &bp
->b_bio1
);
890 * Since we are issuing the callback synchronously it cannot
891 * race the BIO_DONE, so no need for atomic ops here.
893 /*bp->b_bio1.bio_done = func;*/
894 bp
->b_bio1
.bio_caller_info1
.ptr
= arg
;
895 bp
->b_bio1
.bio_flags
|= BIO_DONE
;
903 * breadnx() - Terminal function for bread() and breadn().
905 * This function will start asynchronous I/O on read-ahead blocks as well
906 * as satisfy the primary request.
908 * We must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE is
909 * set, the buffer is valid and we do not have to do anything.
912 breadnx(struct vnode
*vp
, off_t loffset
, int size
, int bflags
,
913 off_t
*raoffset
, int *rabsize
,
914 int cnt
, struct buf
**bpp
)
916 struct buf
*bp
, *rabp
;
918 int rv
= 0, readwait
= 0;
919 int blkflags
= (bflags
& B_KVABIO
) ? GETBLK_KVABIO
: 0;
924 *bpp
= bp
= getblk(vp
, loffset
, size
, blkflags
, 0);
926 /* if not found in cache, do some I/O */
927 if ((bp
->b_flags
& B_CACHE
) == 0) {
928 bp
->b_flags
&= ~(B_ERROR
| B_EINTR
| B_INVAL
| B_NOTMETA
);
929 bp
->b_flags
|= bflags
;
930 bp
->b_cmd
= BUF_CMD_READ
;
931 bp
->b_bio1
.bio_done
= biodone_sync
;
932 bp
->b_bio1
.bio_flags
|= BIO_SYNC
;
933 vfs_busy_pages(vp
, bp
);
934 vn_strategy(vp
, &bp
->b_bio1
);
938 for (i
= 0; i
< cnt
; i
++, raoffset
++, rabsize
++) {
939 if (inmem(vp
, *raoffset
))
941 rabp
= getblk(vp
, *raoffset
, *rabsize
, GETBLK_KVABIO
, 0);
943 if ((rabp
->b_flags
& B_CACHE
) == 0) {
944 rabp
->b_flags
&= ~(B_ERROR
| B_EINTR
|
945 B_INVAL
| B_NOTMETA
);
946 rabp
->b_flags
|= (bflags
& ~B_KVABIO
);
947 rabp
->b_cmd
= BUF_CMD_READ
;
948 vfs_busy_pages(vp
, rabp
);
950 vn_strategy(vp
, &rabp
->b_bio1
);
956 rv
= biowait(&bp
->b_bio1
, "biord");
963 * Synchronous write, waits for completion.
965 * Write, release buffer on completion. (Done by iodone
966 * if async). Do not bother writing anything if the buffer
969 * Note that we set B_CACHE here, indicating that buffer is
970 * fully valid and thus cacheable. This is true even of NFS
971 * now so we set it generally. This could be set either here
972 * or in biodone() since the I/O is synchronous. We put it
976 bwrite(struct buf
*bp
)
980 if (bp
->b_flags
& B_INVAL
) {
984 if (BUF_LOCKINUSE(bp
) == 0)
985 panic("bwrite: buffer is not busy???");
988 * NOTE: We no longer mark the buffer clear prior to the vn_strategy()
989 * call because it will remove the buffer from the vnode's
990 * dirty buffer list prematurely and possibly cause filesystem
991 * checks to race buffer flushes. This is now handled in
994 * bundirty(bp); REMOVED
997 bp
->b_flags
&= ~(B_ERROR
| B_EINTR
);
998 bp
->b_flags
|= B_CACHE
;
999 bp
->b_cmd
= BUF_CMD_WRITE
;
1001 bp
->b_bio1
.bio_done
= biodone_sync
;
1002 bp
->b_bio1
.bio_flags
|= BIO_SYNC
;
1003 vfs_busy_pages(bp
->b_vp
, bp
);
1006 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
1007 * valid for vnode-backed buffers.
1009 bsetrunningbufspace(bp
, bp
->b_bufsize
);
1010 vn_strategy(bp
->b_vp
, &bp
->b_bio1
);
1011 error
= biowait(&bp
->b_bio1
, "biows");
1020 * Asynchronous write. Start output on a buffer, but do not wait for
1021 * it to complete. The buffer is released when the output completes.
1023 * bwrite() ( or the VOP routine anyway ) is responsible for handling
1024 * B_INVAL buffers. Not us.
1027 bawrite(struct buf
*bp
)
1029 if (bp
->b_flags
& B_INVAL
) {
1033 if (BUF_LOCKINUSE(bp
) == 0)
1034 panic("bawrite: buffer is not busy???");
1037 * NOTE: We no longer mark the buffer clear prior to the vn_strategy()
1038 * call because it will remove the buffer from the vnode's
1039 * dirty buffer list prematurely and possibly cause filesystem
1040 * checks to race buffer flushes. This is now handled in
1043 * bundirty(bp); REMOVED
1045 bp
->b_flags
&= ~(B_ERROR
| B_EINTR
);
1046 bp
->b_flags
|= B_CACHE
;
1047 bp
->b_cmd
= BUF_CMD_WRITE
;
1049 KKASSERT(bp
->b_bio1
.bio_done
== NULL
);
1050 vfs_busy_pages(bp
->b_vp
, bp
);
1053 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
1054 * valid for vnode-backed buffers.
1056 bsetrunningbufspace(bp
, bp
->b_bufsize
);
1058 vn_strategy(bp
->b_vp
, &bp
->b_bio1
);
1064 * Delayed write. (Buffer is marked dirty). Do not bother writing
1065 * anything if the buffer is marked invalid.
1067 * Note that since the buffer must be completely valid, we can safely
1068 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
1069 * biodone() in order to prevent getblk from writing the buffer
1070 * out synchronously.
1073 bdwrite(struct buf
*bp
)
1075 if (BUF_LOCKINUSE(bp
) == 0)
1076 panic("bdwrite: buffer is not busy");
1078 if (bp
->b_flags
& B_INVAL
) {
1084 dsched_buf_enter(bp
); /* might stack */
1087 * Set B_CACHE, indicating that the buffer is fully valid. This is
1088 * true even of NFS now.
1090 bp
->b_flags
|= B_CACHE
;
1093 * This bmap keeps the system from needing to do the bmap later,
1094 * perhaps when the system is attempting to do a sync. Since it
1095 * is likely that the indirect block -- or whatever other datastructure
1096 * that the filesystem needs is still in memory now, it is a good
1097 * thing to do this. Note also, that if the pageout daemon is
1098 * requesting a sync -- there might not be enough memory to do
1099 * the bmap then... So, this is important to do.
1101 if (bp
->b_bio2
.bio_offset
== NOOFFSET
) {
1102 VOP_BMAP(bp
->b_vp
, bp
->b_loffset
, &bp
->b_bio2
.bio_offset
,
1103 NULL
, NULL
, BUF_CMD_WRITE
);
1107 * Because the underlying pages may still be mapped and
1108 * writable trying to set the dirty buffer (b_dirtyoff/end)
1109 * range here will be inaccurate.
1111 * However, we must still clean the pages to satisfy the
1112 * vnode_pager and pageout daemon, so they think the pages
1113 * have been "cleaned". What has really occured is that
1114 * they've been earmarked for later writing by the buffer
1117 * So we get the b_dirtyoff/end update but will not actually
1118 * depend on it (NFS that is) until the pages are busied for
1121 vfs_clean_pages(bp
);
1125 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1126 * due to the softdep code.
1131 * Fake write - return pages to VM system as dirty, leave the buffer clean.
1132 * This is used by tmpfs.
1134 * It is important for any VFS using this routine to NOT use it for
1135 * IO_SYNC or IO_ASYNC operations which occur when the system really
1136 * wants to flush VM pages to backing store.
1139 buwrite(struct buf
*bp
)
1145 * Only works for VMIO buffers. If the buffer is already
1146 * marked for delayed-write we can't avoid the bdwrite().
1148 if ((bp
->b_flags
& B_VMIO
) == 0 || (bp
->b_flags
& B_DELWRI
)) {
1154 * Mark as needing a commit.
1156 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
1157 m
= bp
->b_xio
.xio_pages
[i
];
1158 vm_page_need_commit(m
);
1166 * Turn buffer into delayed write request by marking it B_DELWRI.
1167 * B_RELBUF and B_NOCACHE must be cleared.
1169 * We reassign the buffer to itself to properly update it in the
1170 * dirty/clean lists.
1172 * Must be called from a critical section.
1173 * The buffer must be on BQUEUE_NONE.
1176 bdirty(struct buf
*bp
)
1178 KASSERT(bp
->b_qindex
== BQUEUE_NONE
,
1179 ("bdirty: buffer %p still on queue %d", bp
, bp
->b_qindex
));
1180 if (bp
->b_flags
& B_NOCACHE
) {
1181 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp
);
1182 bp
->b_flags
&= ~B_NOCACHE
;
1184 if (bp
->b_flags
& B_INVAL
) {
1185 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp
);
1187 bp
->b_flags
&= ~B_RELBUF
;
1189 if ((bp
->b_flags
& B_DELWRI
) == 0) {
1190 lwkt_gettoken(&bp
->b_vp
->v_token
);
1191 bp
->b_flags
|= B_DELWRI
;
1193 lwkt_reltoken(&bp
->b_vp
->v_token
);
1195 atomic_add_long(&dirtybufcount
, 1);
1196 atomic_add_long(&dirtykvaspace
, bp
->b_kvasize
);
1197 atomic_add_long(&dirtybufspace
, bp
->b_bufsize
);
1198 if (bp
->b_flags
& B_HEAVY
) {
1199 atomic_add_long(&dirtybufcounthw
, 1);
1200 atomic_add_long(&dirtybufspacehw
, bp
->b_bufsize
);
1207 * Set B_HEAVY, indicating that this is a heavy-weight buffer that
1208 * needs to be flushed with a different buf_daemon thread to avoid
1209 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf().
1212 bheavy(struct buf
*bp
)
1214 if ((bp
->b_flags
& B_HEAVY
) == 0) {
1215 bp
->b_flags
|= B_HEAVY
;
1216 if (bp
->b_flags
& B_DELWRI
) {
1217 atomic_add_long(&dirtybufcounthw
, 1);
1218 atomic_add_long(&dirtybufspacehw
, bp
->b_bufsize
);
1226 * Clear B_DELWRI for buffer.
1228 * Must be called from a critical section.
1230 * The buffer is typically on BQUEUE_NONE but there is one case in
1231 * brelse() that calls this function after placing the buffer on
1232 * a different queue.
1235 bundirty(struct buf
*bp
)
1237 if (bp
->b_flags
& B_DELWRI
) {
1238 lwkt_gettoken(&bp
->b_vp
->v_token
);
1239 bp
->b_flags
&= ~B_DELWRI
;
1241 lwkt_reltoken(&bp
->b_vp
->v_token
);
1243 atomic_add_long(&dirtybufcount
, -1);
1244 atomic_add_long(&dirtykvaspace
, -bp
->b_kvasize
);
1245 atomic_add_long(&dirtybufspace
, -bp
->b_bufsize
);
1246 if (bp
->b_flags
& B_HEAVY
) {
1247 atomic_add_long(&dirtybufcounthw
, -1);
1248 atomic_add_long(&dirtybufspacehw
, -bp
->b_bufsize
);
1250 bd_signal(bp
->b_bufsize
);
1253 * Since it is now being written, we can clear its deferred write flag.
1255 bp
->b_flags
&= ~B_DEFERRED
;
1259 * Set the b_runningbufspace field, used to track how much I/O is
1260 * in progress at any given moment.
1263 bsetrunningbufspace(struct buf
*bp
, int bytes
)
1265 bp
->b_runningbufspace
= bytes
;
1267 atomic_add_long(&runningbufspace
, bytes
);
1268 atomic_add_long(&runningbufcount
, 1);
1275 * Release a busy buffer and, if requested, free its resources. The
1276 * buffer will be stashed in the appropriate bufqueue[] allowing it
1277 * to be accessed later as a cache entity or reused for other purposes.
1280 brelse(struct buf
*bp
)
1282 struct bufpcpu
*pcpu
;
1284 int saved_flags
= bp
->b_flags
;
1287 KASSERT(!(bp
->b_flags
& (B_CLUSTER
|B_PAGING
)),
1288 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp
));
1291 * If B_NOCACHE is set we are being asked to destroy the buffer and
1292 * its backing store. Clear B_DELWRI.
1294 * B_NOCACHE is set in two cases: (1) when the caller really wants
1295 * to destroy the buffer and backing store and (2) when the caller
1296 * wants to destroy the buffer and backing store after a write
1299 if ((bp
->b_flags
& (B_NOCACHE
|B_DELWRI
)) == (B_NOCACHE
|B_DELWRI
)) {
1303 if ((bp
->b_flags
& (B_INVAL
| B_DELWRI
)) == B_DELWRI
) {
1305 * A re-dirtied buffer is only subject to destruction
1306 * by B_INVAL. B_ERROR and B_NOCACHE are ignored.
1308 /* leave buffer intact */
1309 } else if ((bp
->b_flags
& (B_NOCACHE
| B_INVAL
| B_ERROR
)) ||
1310 (bp
->b_bufsize
<= 0)) {
1312 * Either a failed read or we were asked to free or not
1313 * cache the buffer. This path is reached with B_DELWRI
1314 * set only if B_INVAL is already set. B_NOCACHE governs
1315 * backing store destruction.
1317 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the
1318 * buffer cannot be immediately freed.
1320 bp
->b_flags
|= B_INVAL
;
1321 if (LIST_FIRST(&bp
->b_dep
) != NULL
)
1323 if (bp
->b_flags
& B_DELWRI
) {
1324 atomic_add_long(&dirtybufcount
, -1);
1325 atomic_add_long(&dirtykvaspace
, -bp
->b_kvasize
);
1326 atomic_add_long(&dirtybufspace
, -bp
->b_bufsize
);
1327 if (bp
->b_flags
& B_HEAVY
) {
1328 atomic_add_long(&dirtybufcounthw
, -1);
1329 atomic_add_long(&dirtybufspacehw
,
1332 bd_signal(bp
->b_bufsize
);
1334 bp
->b_flags
&= ~(B_DELWRI
| B_CACHE
);
1338 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set,
1339 * or if b_refs is non-zero.
1341 * If vfs_vmio_release() is called with either bit set, the
1342 * underlying pages may wind up getting freed causing a previous
1343 * write (bdwrite()) to get 'lost' because pages associated with
1344 * a B_DELWRI bp are marked clean. Pages associated with a
1345 * B_LOCKED buffer may be mapped by the filesystem.
1347 * If we want to release the buffer ourselves (rather then the
1348 * originator asking us to release it), give the originator a
1349 * chance to countermand the release by setting B_LOCKED.
1351 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1352 * if B_DELWRI is set.
1354 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1355 * on pages to return pages to the VM page queues.
1357 if ((bp
->b_flags
& (B_DELWRI
| B_LOCKED
)) || bp
->b_refs
) {
1358 bp
->b_flags
&= ~B_RELBUF
;
1359 } else if (vm_page_count_min(0)) {
1360 if (LIST_FIRST(&bp
->b_dep
) != NULL
)
1361 buf_deallocate(bp
); /* can set B_LOCKED */
1362 if (bp
->b_flags
& (B_DELWRI
| B_LOCKED
))
1363 bp
->b_flags
&= ~B_RELBUF
;
1365 bp
->b_flags
|= B_RELBUF
;
1369 * Make sure b_cmd is clear. It may have already been cleared by
1372 * At this point destroying the buffer is governed by the B_INVAL
1373 * or B_RELBUF flags.
1375 bp
->b_cmd
= BUF_CMD_DONE
;
1376 dsched_buf_exit(bp
);
1379 * VMIO buffer rundown. Make sure the VM page array is restored
1380 * after an I/O may have replaces some of the pages with bogus pages
1381 * in order to not destroy dirty pages in a fill-in read.
1383 * Note that due to the code above, if a buffer is marked B_DELWRI
1384 * then the B_RELBUF and B_NOCACHE bits will always be clear.
1385 * B_INVAL may still be set, however.
1387 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer
1388 * but not the backing store. B_NOCACHE will destroy the backing
1391 * Note that dirty NFS buffers contain byte-granular write ranges
1392 * and should not be destroyed w/ B_INVAL even if the backing store
1395 if (bp
->b_flags
& B_VMIO
) {
1397 * Rundown for VMIO buffers which are not dirty NFS buffers.
1409 * Get the base offset and length of the buffer. Note that
1410 * in the VMIO case if the buffer block size is not
1411 * page-aligned then b_data pointer may not be page-aligned.
1412 * But our b_xio.xio_pages array *IS* page aligned.
1414 * block sizes less then DEV_BSIZE (usually 512) are not
1415 * supported due to the page granularity bits (m->valid,
1416 * m->dirty, etc...).
1418 * See man buf(9) for more information
1421 resid
= bp
->b_bufsize
;
1422 foff
= bp
->b_loffset
;
1424 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
1425 m
= bp
->b_xio
.xio_pages
[i
];
1428 * If we hit a bogus page, fixup *all* of them
1429 * now. Note that we left these pages wired
1430 * when we removed them so they had better exist,
1431 * and they cannot be ripped out from under us so
1432 * no critical section protection is necessary.
1434 if (m
== bogus_page
) {
1436 poff
= OFF_TO_IDX(bp
->b_loffset
);
1438 vm_object_hold(obj
);
1439 for (j
= i
; j
< bp
->b_xio
.xio_npages
; j
++) {
1442 mtmp
= bp
->b_xio
.xio_pages
[j
];
1443 if (mtmp
== bogus_page
) {
1444 if ((bp
->b_flags
& B_HASBOGUS
) == 0)
1445 panic("brelse: bp %p corrupt bogus", bp
);
1446 mtmp
= vm_page_lookup(obj
, poff
+ j
);
1448 panic("brelse: bp %p page %d missing", bp
, j
);
1449 bp
->b_xio
.xio_pages
[j
] = mtmp
;
1452 vm_object_drop(obj
);
1454 if ((bp
->b_flags
& B_HASBOGUS
) ||
1455 (bp
->b_flags
& B_INVAL
) == 0) {
1456 pmap_qenter_noinval(
1457 trunc_page((vm_offset_t
)bp
->b_data
),
1458 bp
->b_xio
.xio_pages
,
1459 bp
->b_xio
.xio_npages
);
1460 bp
->b_flags
&= ~B_HASBOGUS
;
1461 bp
->b_flags
|= B_KVABIO
;
1464 m
= bp
->b_xio
.xio_pages
[i
];
1468 * Invalidate the backing store if B_NOCACHE is set
1469 * (e.g. used with vinvalbuf()). If this is NFS
1470 * we impose a requirement that the block size be
1471 * a multiple of PAGE_SIZE and create a temporary
1472 * hack to basically invalidate the whole page. The
1473 * problem is that NFS uses really odd buffer sizes
1474 * especially when tracking piecemeal writes and
1475 * it also vinvalbuf()'s a lot, which would result
1476 * in only partial page validation and invalidation
1477 * here. If the file page is mmap()'d, however,
1478 * all the valid bits get set so after we invalidate
1479 * here we would end up with weird m->valid values
1480 * like 0xfc. nfs_getpages() can't handle this so
1481 * we clear all the valid bits for the NFS case
1482 * instead of just some of them.
1484 * The real bug is the VM system having to set m->valid
1485 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1486 * itself is an artifact of the whole 512-byte
1487 * granular mess that exists to support odd block
1488 * sizes and UFS meta-data block sizes (e.g. 6144).
1489 * A complete rewrite is required.
1493 if (bp
->b_flags
& (B_NOCACHE
|B_ERROR
)) {
1494 int poffset
= foff
& PAGE_MASK
;
1497 presid
= PAGE_SIZE
- poffset
;
1498 if (bp
->b_vp
->v_tag
== VT_NFS
&&
1499 bp
->b_vp
->v_type
== VREG
) {
1501 } else if (presid
> resid
) {
1504 KASSERT(presid
>= 0, ("brelse: extra page"));
1505 vm_page_set_invalid(m
, poffset
, presid
);
1508 * Also make sure any swap cache is removed
1509 * as it is now stale (HAMMER in particular
1510 * uses B_NOCACHE to deal with buffer
1513 swap_pager_unswapped(m
);
1515 resid
-= PAGE_SIZE
- (foff
& PAGE_MASK
);
1516 foff
= (foff
+ PAGE_SIZE
) & ~(off_t
)PAGE_MASK
;
1518 if (bp
->b_flags
& (B_INVAL
| B_RELBUF
))
1519 vfs_vmio_release(bp
);
1522 * Rundown for non-VMIO buffers.
1524 if (bp
->b_flags
& (B_INVAL
| B_RELBUF
)) {
1527 KKASSERT (LIST_FIRST(&bp
->b_dep
) == NULL
);
1533 if (bp
->b_qindex
!= BQUEUE_NONE
)
1534 panic("brelse: free buffer onto another queue???");
1537 * Figure out the correct queue to place the cleaned up buffer on.
1538 * Buffers placed in the EMPTY or EMPTYKVA had better already be
1539 * disassociated from their vnode.
1541 * Return the buffer to its original pcpu area
1543 pcpu
= &bufpcpu
[bp
->b_qcpu
];
1544 spin_lock(&pcpu
->spin
);
1546 if (bp
->b_flags
& B_LOCKED
) {
1548 * Buffers that are locked are placed in the locked queue
1549 * immediately, regardless of their state.
1551 bp
->b_qindex
= BQUEUE_LOCKED
;
1552 TAILQ_INSERT_TAIL(&pcpu
->bufqueues
[bp
->b_qindex
],
1554 } else if (bp
->b_bufsize
== 0) {
1556 * Buffers with no memory. Due to conditionals near the top
1557 * of brelse() such buffers should probably already be
1558 * marked B_INVAL and disassociated from their vnode.
1560 bp
->b_flags
|= B_INVAL
;
1561 KASSERT(bp
->b_vp
== NULL
,
1562 ("bp1 %p flags %08x/%08x vnode %p "
1563 "unexpectededly still associated!",
1564 bp
, saved_flags
, bp
->b_flags
, bp
->b_vp
));
1565 KKASSERT((bp
->b_flags
& B_HASHED
) == 0);
1566 bp
->b_qindex
= BQUEUE_EMPTY
;
1567 TAILQ_INSERT_HEAD(&pcpu
->bufqueues
[bp
->b_qindex
],
1569 } else if (bp
->b_flags
& (B_INVAL
| B_NOCACHE
| B_RELBUF
)) {
1571 * Buffers with junk contents. Again these buffers had better
1572 * already be disassociated from their vnode.
1574 KASSERT(bp
->b_vp
== NULL
,
1575 ("bp2 %p flags %08x/%08x vnode %p unexpectededly "
1576 "still associated!",
1577 bp
, saved_flags
, bp
->b_flags
, bp
->b_vp
));
1578 KKASSERT((bp
->b_flags
& B_HASHED
) == 0);
1579 bp
->b_flags
|= B_INVAL
;
1580 bp
->b_qindex
= BQUEUE_CLEAN
;
1581 TAILQ_INSERT_HEAD(&pcpu
->bufqueues
[bp
->b_qindex
],
1585 * Remaining buffers. These buffers are still associated with
1588 switch(bp
->b_flags
& (B_DELWRI
|B_HEAVY
)) {
1590 bp
->b_qindex
= BQUEUE_DIRTY
;
1591 TAILQ_INSERT_TAIL(&pcpu
->bufqueues
[bp
->b_qindex
],
1594 case B_DELWRI
| B_HEAVY
:
1595 bp
->b_qindex
= BQUEUE_DIRTY_HW
;
1596 TAILQ_INSERT_TAIL(&pcpu
->bufqueues
[bp
->b_qindex
],
1601 * NOTE: Buffers are always placed at the end of the
1602 * queue. If B_AGE is not set the buffer will cycle
1603 * through the queue twice.
1605 bp
->b_qindex
= BQUEUE_CLEAN
;
1606 TAILQ_INSERT_TAIL(&pcpu
->bufqueues
[bp
->b_qindex
],
1611 spin_unlock(&pcpu
->spin
);
1614 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1615 * on the correct queue but we have not yet unlocked it.
1617 if ((bp
->b_flags
& (B_INVAL
|B_DELWRI
)) == (B_INVAL
|B_DELWRI
))
1621 * The bp is on an appropriate queue unless locked. If it is not
1622 * locked or dirty we can wakeup threads waiting for buffer space.
1624 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1625 * if B_INVAL is set ).
1627 if ((bp
->b_flags
& (B_LOCKED
|B_DELWRI
)) == 0)
1631 * Something we can maybe free or reuse
1633 if (bp
->b_bufsize
|| bp
->b_kvasize
)
1637 * Clean up temporary flags and unlock the buffer.
1639 bp
->b_flags
&= ~(B_NOCACHE
| B_RELBUF
| B_DIRECT
);
1646 * Release a buffer back to the appropriate queue but do not try to free
1647 * it. The buffer is expected to be used again soon.
1649 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1650 * biodone() to requeue an async I/O on completion. It is also used when
1651 * known good buffers need to be requeued but we think we may need the data
1654 * XXX we should be able to leave the B_RELBUF hint set on completion.
1657 bqrelse(struct buf
*bp
)
1659 struct bufpcpu
*pcpu
;
1661 KASSERT(!(bp
->b_flags
& (B_CLUSTER
|B_PAGING
)),
1662 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp
));
1664 if (bp
->b_qindex
!= BQUEUE_NONE
)
1665 panic("bqrelse: free buffer onto another queue???");
1667 buf_act_advance(bp
);
1669 pcpu
= &bufpcpu
[bp
->b_qcpu
];
1670 spin_lock(&pcpu
->spin
);
1672 if (bp
->b_flags
& B_LOCKED
) {
1674 * Locked buffers are released to the locked queue. However,
1675 * if the buffer is dirty it will first go into the dirty
1676 * queue and later on after the I/O completes successfully it
1677 * will be released to the locked queue.
1679 bp
->b_qindex
= BQUEUE_LOCKED
;
1680 TAILQ_INSERT_TAIL(&pcpu
->bufqueues
[bp
->b_qindex
],
1682 } else if (bp
->b_flags
& B_DELWRI
) {
1683 bp
->b_qindex
= (bp
->b_flags
& B_HEAVY
) ?
1684 BQUEUE_DIRTY_HW
: BQUEUE_DIRTY
;
1685 TAILQ_INSERT_TAIL(&pcpu
->bufqueues
[bp
->b_qindex
],
1687 } else if (vm_page_count_min(0)) {
1689 * We are too low on memory, we have to try to free the
1690 * buffer (most importantly: the wired pages making up its
1691 * backing store) *now*.
1693 spin_unlock(&pcpu
->spin
);
1697 bp
->b_qindex
= BQUEUE_CLEAN
;
1698 TAILQ_INSERT_TAIL(&pcpu
->bufqueues
[bp
->b_qindex
],
1701 spin_unlock(&pcpu
->spin
);
1704 * We have now placed the buffer on the proper queue, but have yet
1707 if ((bp
->b_flags
& B_LOCKED
) == 0 &&
1708 ((bp
->b_flags
& B_INVAL
) || (bp
->b_flags
& B_DELWRI
) == 0)) {
1713 * Something we can maybe free or reuse.
1715 if (bp
->b_bufsize
&& !(bp
->b_flags
& B_DELWRI
))
1719 * Final cleanup and unlock. Clear bits that are only used while a
1720 * buffer is actively locked.
1722 bp
->b_flags
&= ~(B_NOCACHE
| B_RELBUF
);
1723 dsched_buf_exit(bp
);
1728 * Hold a buffer, preventing it from being reused. This will prevent
1729 * normal B_RELBUF operations on the buffer but will not prevent B_INVAL
1730 * operations. If a B_INVAL operation occurs the buffer will remain held
1731 * but the underlying pages may get ripped out.
1733 * These functions are typically used in VOP_READ/VOP_WRITE functions
1734 * to hold a buffer during a copyin or copyout, preventing deadlocks
1735 * or recursive lock panics when read()/write() is used over mmap()'d
1738 * NOTE: bqhold() requires that the buffer be locked at the time of the
1739 * hold. bqdrop() has no requirements other than the buffer having
1740 * previously been held.
1743 bqhold(struct buf
*bp
)
1745 atomic_add_int(&bp
->b_refs
, 1);
1749 bqdrop(struct buf
*bp
)
1751 KKASSERT(bp
->b_refs
> 0);
1752 atomic_add_int(&bp
->b_refs
, -1);
1756 * Return backing pages held by the buffer 'bp' back to the VM system.
1757 * This routine is called when the bp is invalidated, released, or
1760 * The KVA mapping (b_data) for the underlying pages is removed by
1763 * WARNING! This routine is integral to the low memory critical path
1764 * when a buffer is B_RELBUF'd. If the system has a severe page
1765 * deficit we need to get the page(s) onto the PQ_FREE or PQ_CACHE
1766 * queues so they can be reused in the current pageout daemon
1770 vfs_vmio_release(struct buf
*bp
)
1775 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
1776 m
= bp
->b_xio
.xio_pages
[i
];
1777 bp
->b_xio
.xio_pages
[i
] = NULL
;
1780 * We need to own the page in order to safely unwire it.
1782 vm_page_busy_wait(m
, FALSE
, "vmiopg");
1785 * The VFS is telling us this is not a meta-data buffer
1786 * even if it is backed by a block device.
1788 if (bp
->b_flags
& B_NOTMETA
)
1789 vm_page_flag_set(m
, PG_NOTMETA
);
1792 * This is a very important bit of code. We try to track
1793 * VM page use whether the pages are wired into the buffer
1794 * cache or not. While wired into the buffer cache the
1795 * bp tracks the act_count.
1797 * We can choose to place unwired pages on the inactive
1798 * queue (0) or active queue (1). If we place too many
1799 * on the active queue the queue will cycle the act_count
1800 * on pages we'd like to keep, just from single-use pages
1801 * (such as when doing a tar-up or file scan).
1803 if (bp
->b_act_count
< vm_cycle_point
)
1804 vm_page_unwire(m
, 0);
1806 vm_page_unwire(m
, 1);
1809 * If the wire_count has dropped to 0 we may need to take
1810 * further action before unbusying the page.
1812 * WARNING: vm_page_try_*() also checks PG_NEED_COMMIT for us.
1814 if (m
->wire_count
== 0) {
1815 if (bp
->b_flags
& B_DIRECT
) {
1817 * Attempt to free the page if B_DIRECT is
1818 * set, the caller does not desire the page
1822 vm_page_try_to_free(m
);
1823 } else if ((bp
->b_flags
& B_NOTMETA
) ||
1824 vm_page_count_min(0)) {
1826 * Attempt to move the page to PQ_CACHE
1827 * if B_NOTMETA is set. This flag is set
1828 * by HAMMER to remove one of the two pages
1829 * present when double buffering is enabled.
1831 * Attempt to move the page to PQ_CACHE
1832 * If we have a severe page deficit. This
1833 * will cause buffer cache operations related
1834 * to pageouts to recycle the related pages
1835 * in order to avoid a low memory deadlock.
1837 m
->act_count
= bp
->b_act_count
;
1838 vm_page_try_to_cache(m
);
1841 * Nominal case, leave the page on the
1842 * queue the original unwiring placed it on
1843 * (active or inactive).
1845 m
->act_count
= bp
->b_act_count
;
1854 * Zero out the pmap pte's for the mapping, but don't bother
1855 * invalidating the TLB. The range will be properly invalidating
1856 * when new pages are entered into the mapping.
1858 * This in particular reduces tmpfs tear-down overhead and reduces
1859 * buffer cache re-use overhead (one invalidation sequence instead
1860 * of two per re-use).
1862 pmap_qremove_noinval(trunc_page((vm_offset_t
) bp
->b_data
),
1863 bp
->b_xio
.xio_npages
);
1864 CPUMASK_ASSZERO(bp
->b_cpumask
);
1865 if (bp
->b_bufsize
) {
1866 atomic_add_long(&bufspace
, -bp
->b_bufsize
);
1870 bp
->b_xio
.xio_npages
= 0;
1871 bp
->b_flags
&= ~B_VMIO
;
1872 KKASSERT (LIST_FIRST(&bp
->b_dep
) == NULL
);
1878 * Find and initialize a new buffer header, freeing up existing buffers
1879 * in the bufqueues as necessary. The new buffer is returned locked.
1881 * Important: B_INVAL is not set. If the caller wishes to throw the
1882 * buffer away, the caller must set B_INVAL prior to calling brelse().
1885 * We have insufficient buffer headers
1886 * We have insufficient buffer space
1888 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1889 * Instead we ask the buf daemon to do it for us. We attempt to
1890 * avoid piecemeal wakeups of the pageout daemon.
1893 getnewbuf(int blkflags
, int slptimeo
, int size
, int maxsize
)
1895 struct bufpcpu
*pcpu
;
1900 int slpflags
= (blkflags
& GETBLK_PCATCH
) ? PCATCH
: 0;
1901 int maxloops
= 200000;
1902 int restart_reason
= 0;
1903 struct buf
*restart_bp
= NULL
;
1904 static char flushingbufs
[MAXCPU
];
1908 * We can't afford to block since we might be holding a vnode lock,
1909 * which may prevent system daemons from running. We deal with
1910 * low-memory situations by proactively returning memory and running
1911 * async I/O rather then sync I/O.
1915 nqcpu
= mycpu
->gd_cpuid
;
1916 flushingp
= &flushingbufs
[nqcpu
];
1918 if (bufspace
< lobufspace
)
1921 if (debug_bufbio
&& --maxloops
== 0)
1922 panic("getnewbuf, excessive loops on cpu %d restart %d (%p)",
1923 mycpu
->gd_cpuid
, restart_reason
, restart_bp
);
1926 * Setup for scan. If we do not have enough free buffers,
1927 * we setup a degenerate case that immediately fails. Note
1928 * that if we are specially marked process, we are allowed to
1929 * dip into our reserves.
1931 * The scanning sequence is nominally: EMPTY->CLEAN
1933 pcpu
= &bufpcpu
[nqcpu
];
1934 spin_lock(&pcpu
->spin
);
1937 * Prime the scan for this cpu. Locate the first buffer to
1938 * check. If we are flushing buffers we must skip the
1941 nqindex
= BQUEUE_EMPTY
;
1942 nbp
= TAILQ_FIRST(&pcpu
->bufqueues
[BQUEUE_EMPTY
]);
1943 if (nbp
== NULL
|| *flushingp
) {
1944 nqindex
= BQUEUE_CLEAN
;
1945 nbp
= TAILQ_FIRST(&pcpu
->bufqueues
[BQUEUE_CLEAN
]);
1949 * Run scan, possibly freeing data and/or kva mappings on the fly,
1952 * WARNING! spin is held!
1954 while ((bp
= nbp
) != NULL
) {
1955 int qindex
= nqindex
;
1957 nbp
= TAILQ_NEXT(bp
, b_freelist
);
1960 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
1961 * cycles through the queue twice before being selected.
1963 if (qindex
== BQUEUE_CLEAN
&&
1964 (bp
->b_flags
& B_AGE
) == 0 && nbp
) {
1965 bp
->b_flags
|= B_AGE
;
1966 TAILQ_REMOVE(&pcpu
->bufqueues
[qindex
],
1968 TAILQ_INSERT_TAIL(&pcpu
->bufqueues
[qindex
],
1974 * Calculate next bp ( we can only use it if we do not block
1975 * or do other fancy things ).
1980 nqindex
= BQUEUE_CLEAN
;
1981 if ((nbp
= TAILQ_FIRST(&pcpu
->bufqueues
[BQUEUE_CLEAN
])))
1995 KASSERT(bp
->b_qindex
== qindex
,
1996 ("getnewbuf: inconsistent queue %d bp %p", qindex
, bp
));
1999 * Note: we no longer distinguish between VMIO and non-VMIO
2002 KASSERT((bp
->b_flags
& B_DELWRI
) == 0,
2003 ("delwri buffer %p found in queue %d", bp
, qindex
));
2006 * Do not try to reuse a buffer with a non-zero b_refs.
2007 * This is an unsynchronized test. A synchronized test
2008 * is also performed after we lock the buffer.
2014 * Start freeing the bp. This is somewhat involved. nbp
2015 * remains valid only for BQUEUE_EMPTY bp's. Buffers
2016 * on the clean list must be disassociated from their
2017 * current vnode. Buffers on the empty lists have
2018 * already been disassociated.
2020 * b_refs is checked after locking along with queue changes.
2021 * We must check here to deal with zero->nonzero transitions
2022 * made by the owner of the buffer lock, which is used by
2023 * VFS's to hold the buffer while issuing an unlocked
2024 * uiomove()s. We cannot invalidate the buffer's pages
2025 * for this case. Once we successfully lock a buffer the
2026 * only 0->1 transitions of b_refs will occur via findblk().
2028 * We must also check for queue changes after successful
2029 * locking as the current lock holder may dispose of the
2030 * buffer and change its queue.
2032 if (BUF_LOCK(bp
, LK_EXCLUSIVE
| LK_NOWAIT
) != 0) {
2033 spin_unlock(&pcpu
->spin
);
2034 tsleep(&bd_request
, 0, "gnbxxx", (hz
+ 99) / 100);
2039 if (bp
->b_qindex
!= qindex
|| bp
->b_refs
) {
2040 spin_unlock(&pcpu
->spin
);
2046 bremfree_locked(bp
);
2047 spin_unlock(&pcpu
->spin
);
2050 * Dependancies must be handled before we disassociate the
2053 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
2054 * be immediately disassociated. HAMMER then becomes
2055 * responsible for releasing the buffer.
2057 * NOTE: spin is UNLOCKED now.
2059 if (LIST_FIRST(&bp
->b_dep
) != NULL
) {
2061 if (bp
->b_flags
& B_LOCKED
) {
2067 KKASSERT(LIST_FIRST(&bp
->b_dep
) == NULL
);
2071 * CLEAN buffers have content or associations that must be
2072 * cleaned out if not repurposing.
2074 if (qindex
== BQUEUE_CLEAN
) {
2075 if (bp
->b_flags
& B_VMIO
)
2076 vfs_vmio_release(bp
);
2082 * NOTE: nbp is now entirely invalid. We can only restart
2083 * the scan from this point on.
2085 * Get the rest of the buffer freed up. b_kva* is still
2086 * valid after this operation.
2088 KASSERT(bp
->b_vp
== NULL
,
2089 ("bp3 %p flags %08x vnode %p qindex %d "
2090 "unexpectededly still associated!",
2091 bp
, bp
->b_flags
, bp
->b_vp
, qindex
));
2092 KKASSERT((bp
->b_flags
& B_HASHED
) == 0);
2097 if (bp
->b_flags
& (B_VNDIRTY
| B_VNCLEAN
| B_HASHED
)) {
2098 kprintf("getnewbuf: caught bug vp queue "
2099 "%p/%08x qidx %d\n",
2100 bp
, bp
->b_flags
, qindex
);
2103 bp
->b_flags
= B_BNOCLIP
;
2104 bp
->b_cmd
= BUF_CMD_DONE
;
2109 bp
->b_xio
.xio_npages
= 0;
2110 bp
->b_dirtyoff
= bp
->b_dirtyend
= 0;
2111 bp
->b_act_count
= ACT_INIT
;
2113 KKASSERT(LIST_FIRST(&bp
->b_dep
) == NULL
);
2115 if (blkflags
& GETBLK_BHEAVY
)
2116 bp
->b_flags
|= B_HEAVY
;
2118 if (bufspace
>= hibufspace
)
2120 if (bufspace
< lobufspace
)
2123 bp
->b_flags
|= B_INVAL
;
2131 * b_refs can transition to a non-zero value while we hold
2132 * the buffer locked due to a findblk(). Our brelvp() above
2133 * interlocked any future possible transitions due to
2136 * If we find b_refs to be non-zero we can destroy the
2137 * buffer's contents but we cannot yet reuse the buffer.
2140 bp
->b_flags
|= B_INVAL
;
2149 * We found our buffer!
2155 * If we exhausted our list, iterate other cpus. If that fails,
2156 * sleep as appropriate. We may have to wakeup various daemons
2157 * and write out some dirty buffers.
2159 * Generally we are sleeping due to insufficient buffer space.
2161 * NOTE: spin is held if bp is NULL, else it is not held.
2167 spin_unlock(&pcpu
->spin
);
2169 nqcpu
= (nqcpu
+ 1) % ncpus
;
2170 if (nqcpu
!= mycpu
->gd_cpuid
) {
2176 if (bufspace
>= hibufspace
) {
2178 flags
= VFS_BIO_NEED_BUFSPACE
;
2181 flags
= VFS_BIO_NEED_ANY
;
2184 bd_speedup(); /* heeeelp */
2185 atomic_set_int(&needsbuffer
, flags
);
2186 while (needsbuffer
& flags
) {
2189 tsleep_interlock(&needsbuffer
, 0);
2190 value
= atomic_fetchadd_int(&needsbuffer
, 0);
2191 if (value
& flags
) {
2192 if (tsleep(&needsbuffer
, PINTERLOCKED
|slpflags
,
2193 waitmsg
, slptimeo
)) {
2200 * We finally have a valid bp. Reset b_data.
2202 * (spin is not held)
2204 bp
->b_data
= bp
->b_kvabase
;
2212 * Buffer flushing daemon. Buffers are normally flushed by the
2213 * update daemon but if it cannot keep up this process starts to
2214 * take the load in an attempt to prevent getnewbuf() from blocking.
2216 * Once a flush is initiated it does not stop until the number
2217 * of buffers falls below lodirtybuffers, but we will wake up anyone
2218 * waiting at the mid-point.
2220 static struct kproc_desc buf_kp
= {
2225 SYSINIT(bufdaemon
, SI_SUB_KTHREAD_BUF
, SI_ORDER_FIRST
,
2226 kproc_start
, &buf_kp
);
2228 static struct kproc_desc bufhw_kp
= {
2233 SYSINIT(bufdaemon_hw
, SI_SUB_KTHREAD_BUF
, SI_ORDER_FIRST
,
2234 kproc_start
, &bufhw_kp
);
2237 buf_daemon1(struct thread
*td
, int queue
, int (*buf_limit_fn
)(long),
2243 marker
= kmalloc(sizeof(*marker
), M_BIOBUF
, M_WAITOK
| M_ZERO
);
2244 marker
->b_flags
|= B_MARKER
;
2245 marker
->b_qindex
= BQUEUE_NONE
;
2249 * This process needs to be suspended prior to shutdown sync.
2251 EVENTHANDLER_REGISTER(shutdown_pre_sync
, shutdown_kproc
,
2252 td
, SHUTDOWN_PRI_LAST
);
2253 curthread
->td_flags
|= TDF_SYSTHREAD
;
2256 * This process is allowed to take the buffer cache to the limit
2259 kproc_suspend_loop();
2262 * Do the flush as long as the number of dirty buffers
2263 * (including those running) exceeds lodirtybufspace.
2265 * When flushing limit running I/O to hirunningspace
2266 * Do the flush. Limit the amount of in-transit I/O we
2267 * allow to build up, otherwise we would completely saturate
2268 * the I/O system. Wakeup any waiting processes before we
2269 * normally would so they can run in parallel with our drain.
2271 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2272 * but because we split the operation into two threads we
2273 * have to cut it in half for each thread.
2275 waitrunningbufspace();
2276 limit
= lodirtybufspace
/ 2;
2277 while (buf_limit_fn(limit
)) {
2278 if (flushbufqueues(marker
, queue
) == 0)
2280 if (runningbufspace
< hirunningspace
)
2282 waitrunningbufspace();
2286 * We reached our low water mark, reset the
2287 * request and sleep until we are needed again.
2288 * The sleep is just so the suspend code works.
2290 tsleep_interlock(bd_req
, 0);
2291 if (atomic_swap_int(bd_req
, 0) == 0)
2292 tsleep(bd_req
, PINTERLOCKED
, "psleep", hz
);
2295 /*kfree(marker, M_BIOBUF);*/
2299 buf_daemon_limit(long limit
)
2301 return (runningbufspace
+ dirtykvaspace
> limit
||
2302 dirtybufcount
- dirtybufcounthw
>= nbuf
/ 2);
2306 buf_daemon_hw_limit(long limit
)
2308 return (runningbufspace
+ dirtykvaspace
> limit
||
2309 dirtybufcounthw
>= nbuf
/ 2);
2315 buf_daemon1(bufdaemon_td
, BQUEUE_DIRTY
, buf_daemon_limit
,
2322 buf_daemon1(bufdaemonhw_td
, BQUEUE_DIRTY_HW
, buf_daemon_hw_limit
,
2327 * Flush up to (flushperqueue) buffers in the dirty queue. Each cpu has a
2328 * localized version of the queue. Each call made to this function iterates
2329 * to another cpu. It is desireable to flush several buffers from the same
2330 * cpu's queue at once, as these are likely going to be linear.
2332 * We must be careful to free up B_INVAL buffers instead of write them, which
2333 * NFS is particularly sensitive to.
2335 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate that we
2336 * really want to try to get the buffer out and reuse it due to the write
2337 * load on the machine.
2339 * We must lock the buffer in order to check its validity before we can mess
2340 * with its contents. spin isn't enough.
2343 flushbufqueues(struct buf
*marker
, bufq_type_t q
)
2345 struct bufpcpu
*pcpu
;
2348 u_int loops
= flushperqueue
;
2349 int lcpu
= marker
->b_qcpu
;
2351 KKASSERT(marker
->b_qindex
== BQUEUE_NONE
);
2352 KKASSERT(marker
->b_flags
& B_MARKER
);
2356 * Spinlock needed to perform operations on the queue and may be
2357 * held through a non-blocking BUF_LOCK(), but cannot be held when
2358 * BUF_UNLOCK()ing or through any other major operation.
2360 pcpu
= &bufpcpu
[marker
->b_qcpu
];
2361 spin_lock(&pcpu
->spin
);
2362 marker
->b_qindex
= q
;
2363 TAILQ_INSERT_HEAD(&pcpu
->bufqueues
[q
], marker
, b_freelist
);
2366 while ((bp
= TAILQ_NEXT(bp
, b_freelist
)) != NULL
) {
2368 * NOTE: spinlock is always held at the top of the loop
2370 if (bp
->b_flags
& B_MARKER
)
2372 if ((bp
->b_flags
& B_DELWRI
) == 0) {
2373 kprintf("Unexpected clean buffer %p\n", bp
);
2376 if (BUF_LOCK(bp
, LK_EXCLUSIVE
| LK_NOWAIT
))
2378 KKASSERT(bp
->b_qcpu
== marker
->b_qcpu
&& bp
->b_qindex
== q
);
2381 * Once the buffer is locked we will have no choice but to
2382 * unlock the spinlock around a later BUF_UNLOCK and re-set
2383 * bp = marker when looping. Move the marker now to make
2386 TAILQ_REMOVE(&pcpu
->bufqueues
[q
], marker
, b_freelist
);
2387 TAILQ_INSERT_AFTER(&pcpu
->bufqueues
[q
], bp
, marker
, b_freelist
);
2390 * Must recheck B_DELWRI after successfully locking
2393 if ((bp
->b_flags
& B_DELWRI
) == 0) {
2394 spin_unlock(&pcpu
->spin
);
2396 spin_lock(&pcpu
->spin
);
2402 * Remove the buffer from its queue. We still own the
2408 * Disposing of an invalid buffer counts as a flush op
2410 if (bp
->b_flags
& B_INVAL
) {
2411 spin_unlock(&pcpu
->spin
);
2417 * Release the spinlock for the more complex ops we
2418 * are now going to do.
2420 spin_unlock(&pcpu
->spin
);
2424 * This is a bit messy
2426 if (LIST_FIRST(&bp
->b_dep
) != NULL
&&
2427 (bp
->b_flags
& B_DEFERRED
) == 0 &&
2428 buf_countdeps(bp
, 0)) {
2429 spin_lock(&pcpu
->spin
);
2430 TAILQ_INSERT_TAIL(&pcpu
->bufqueues
[q
], bp
, b_freelist
);
2432 bp
->b_flags
|= B_DEFERRED
;
2433 spin_unlock(&pcpu
->spin
);
2435 spin_lock(&pcpu
->spin
);
2441 * spinlock not held here.
2443 * If the buffer has a dependancy, buf_checkwrite() must
2444 * also return 0 for us to be able to initate the write.
2446 * If the buffer is flagged B_ERROR it may be requeued
2447 * over and over again, we try to avoid a live lock.
2449 if (LIST_FIRST(&bp
->b_dep
) != NULL
&& buf_checkwrite(bp
)) {
2451 } else if (bp
->b_flags
& B_ERROR
) {
2452 tsleep(bp
, 0, "bioer", 1);
2453 bp
->b_flags
&= ~B_AGE
;
2456 bp
->b_flags
|= B_AGE
| B_KVABIO
;
2459 /* bp invalid but needs to be NULL-tested if we break out */
2461 spin_lock(&pcpu
->spin
);
2467 /* bp is invalid here but can be NULL-tested to advance */
2469 TAILQ_REMOVE(&pcpu
->bufqueues
[q
], marker
, b_freelist
);
2470 marker
->b_qindex
= BQUEUE_NONE
;
2471 spin_unlock(&pcpu
->spin
);
2474 * Advance the marker to be fair.
2476 marker
->b_qcpu
= (marker
->b_qcpu
+ 1) % ncpus
;
2478 if (marker
->b_qcpu
!= lcpu
)
2488 * Returns true if no I/O is needed to access the associated VM object.
2489 * This is like findblk except it also hunts around in the VM system for
2492 * Note that we ignore vm_page_free() races from interrupts against our
2493 * lookup, since if the caller is not protected our return value will not
2494 * be any more valid then otherwise once we exit the critical section.
2497 inmem(struct vnode
*vp
, off_t loffset
)
2500 vm_offset_t toff
, tinc
, size
;
2504 if (findblk(vp
, loffset
, FINDBLK_TEST
))
2506 if (vp
->v_mount
== NULL
)
2508 if ((obj
= vp
->v_object
) == NULL
)
2512 if (size
> vp
->v_mount
->mnt_stat
.f_iosize
)
2513 size
= vp
->v_mount
->mnt_stat
.f_iosize
;
2515 vm_object_hold(obj
);
2516 for (toff
= 0; toff
< vp
->v_mount
->mnt_stat
.f_iosize
; toff
+= tinc
) {
2517 m
= vm_page_lookup(obj
, OFF_TO_IDX(loffset
+ toff
));
2523 if (tinc
> PAGE_SIZE
- ((toff
+ loffset
) & PAGE_MASK
))
2524 tinc
= PAGE_SIZE
- ((toff
+ loffset
) & PAGE_MASK
);
2525 if (vm_page_is_valid(m
,
2526 (vm_offset_t
) ((toff
+ loffset
) & PAGE_MASK
), tinc
) == 0) {
2531 vm_object_drop(obj
);
2538 * Locate and return the specified buffer. Unless flagged otherwise,
2539 * a locked buffer will be returned if it exists or NULL if it does not.
2541 * findblk()'d buffers are still on the bufqueues and if you intend
2542 * to use your (locked NON-TEST) buffer you need to bremfree(bp)
2543 * and possibly do other stuff to it.
2545 * FINDBLK_TEST - Do not lock the buffer. The caller is responsible
2546 * for locking the buffer and ensuring that it remains
2547 * the desired buffer after locking.
2549 * FINDBLK_NBLOCK - Lock the buffer non-blocking. If we are unable
2550 * to acquire the lock we return NULL, even if the
2553 * FINDBLK_REF - Returns the buffer ref'd, which prevents normal
2554 * reuse by getnewbuf() but does not prevent
2555 * disassociation (B_INVAL). Used to avoid deadlocks
2556 * against random (vp,loffset)s due to reassignment.
2558 * FINDBLK_KVABIO - Only applicable when returning a locked buffer.
2559 * Indicates that the caller supports B_KVABIO.
2561 * (0) - Lock the buffer blocking.
2564 findblk(struct vnode
*vp
, off_t loffset
, int flags
)
2569 lkflags
= LK_EXCLUSIVE
;
2570 if (flags
& FINDBLK_NBLOCK
)
2571 lkflags
|= LK_NOWAIT
;
2575 * Lookup. Ref the buf while holding v_token to prevent
2576 * reuse (but does not prevent diassociation).
2578 lwkt_gettoken_shared(&vp
->v_token
);
2579 bp
= buf_rb_hash_RB_LOOKUP(&vp
->v_rbhash_tree
, loffset
);
2581 lwkt_reltoken(&vp
->v_token
);
2585 lwkt_reltoken(&vp
->v_token
);
2588 * If testing only break and return bp, do not lock.
2590 if (flags
& FINDBLK_TEST
)
2594 * Lock the buffer, return an error if the lock fails.
2595 * (only FINDBLK_NBLOCK can cause the lock to fail).
2597 if (BUF_LOCK(bp
, lkflags
)) {
2598 atomic_subtract_int(&bp
->b_refs
, 1);
2599 /* bp = NULL; not needed */
2604 * Revalidate the locked buf before allowing it to be
2607 * B_KVABIO is only set/cleared when locking. When
2608 * clearing B_KVABIO, we must ensure that the buffer
2609 * is synchronized to all cpus.
2611 if (bp
->b_vp
== vp
&& bp
->b_loffset
== loffset
) {
2612 if (flags
& FINDBLK_KVABIO
)
2613 bp
->b_flags
|= B_KVABIO
;
2618 atomic_subtract_int(&bp
->b_refs
, 1);
2625 if ((flags
& FINDBLK_REF
) == 0)
2626 atomic_subtract_int(&bp
->b_refs
, 1);
2633 * Similar to getblk() except only returns the buffer if it is
2634 * B_CACHE and requires no other manipulation. Otherwise NULL
2635 * is returned. NULL is also returned if GETBLK_NOWAIT is set
2636 * and the getblk() would block.
2638 * If B_RAM is set the buffer might be just fine, but we return
2639 * NULL anyway because we want the code to fall through to the
2640 * cluster read to issue more read-aheads. Otherwise read-ahead breaks.
2642 * If blksize is 0 the buffer cache buffer must already be fully
2645 * If blksize is non-zero getblk() will be used, allowing a buffer
2646 * to be reinstantiated from its VM backing store. The buffer must
2647 * still be fully cached after reinstantiation to be returned.
2650 getcacheblk(struct vnode
*vp
, off_t loffset
, int blksize
, int blkflags
)
2655 if (blkflags
& GETBLK_NOWAIT
)
2656 fndflags
|= FINDBLK_NBLOCK
;
2657 if (blkflags
& GETBLK_KVABIO
)
2658 fndflags
|= FINDBLK_KVABIO
;
2661 bp
= getblk(vp
, loffset
, blksize
, blkflags
, 0);
2663 if ((bp
->b_flags
& (B_INVAL
| B_CACHE
)) == B_CACHE
) {
2664 bp
->b_flags
&= ~B_AGE
;
2665 if (bp
->b_flags
& B_RAM
) {
2675 bp
= findblk(vp
, loffset
, fndflags
);
2677 if ((bp
->b_flags
& (B_INVAL
| B_CACHE
| B_RAM
)) ==
2679 bp
->b_flags
&= ~B_AGE
;
2693 * Get a block given a specified block and offset into a file/device.
2694 * B_INVAL may or may not be set on return. The caller should clear
2695 * B_INVAL prior to initiating a READ.
2697 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2698 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2699 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer
2700 * without doing any of those things the system will likely believe
2701 * the buffer to be valid (especially if it is not B_VMIO), and the
2702 * next getblk() will return the buffer with B_CACHE set.
2704 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2705 * an existing buffer.
2707 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2708 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2709 * and then cleared based on the backing VM. If the previous buffer is
2710 * non-0-sized but invalid, B_CACHE will be cleared.
2712 * If getblk() must create a new buffer, the new buffer is returned with
2713 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2714 * case it is returned with B_INVAL clear and B_CACHE set based on the
2717 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2718 * B_CACHE bit is clear.
2720 * What this means, basically, is that the caller should use B_CACHE to
2721 * determine whether the buffer is fully valid or not and should clear
2722 * B_INVAL prior to issuing a read. If the caller intends to validate
2723 * the buffer by loading its data area with something, the caller needs
2724 * to clear B_INVAL. If the caller does this without issuing an I/O,
2725 * the caller should set B_CACHE ( as an optimization ), else the caller
2726 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2727 * a write attempt or if it was a successfull read. If the caller
2728 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2729 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2733 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return
2734 * GETBLK_BHEAVY - heavy-weight buffer cache buffer
2737 getblk(struct vnode
*vp
, off_t loffset
, int size
, int blkflags
, int slptimeo
)
2740 int slpflags
= (blkflags
& GETBLK_PCATCH
) ? PCATCH
: 0;
2744 if (size
> MAXBSIZE
)
2745 panic("getblk: size(%d) > MAXBSIZE(%d)", size
, MAXBSIZE
);
2746 if (vp
->v_object
== NULL
)
2747 panic("getblk: vnode %p has no object!", vp
);
2750 * NOTE: findblk does not try to resolve KVABIO in REF-only mode.
2751 * we still have to handle that ourselves.
2754 if ((bp
= findblk(vp
, loffset
, FINDBLK_REF
| FINDBLK_TEST
)) != NULL
) {
2756 * The buffer was found in the cache, but we need to lock it.
2757 * We must acquire a ref on the bp to prevent reuse, but
2758 * this will not prevent disassociation (brelvp()) so we
2759 * must recheck (vp,loffset) after acquiring the lock.
2761 * Without the ref the buffer could potentially be reused
2762 * before we acquire the lock and create a deadlock
2763 * situation between the thread trying to reuse the buffer
2764 * and us due to the fact that we would wind up blocking
2765 * on a random (vp,loffset).
2767 if (BUF_LOCK(bp
, LK_EXCLUSIVE
| LK_NOWAIT
)) {
2768 if (blkflags
& GETBLK_NOWAIT
) {
2772 lkflags
= LK_EXCLUSIVE
| LK_SLEEPFAIL
;
2773 if (blkflags
& GETBLK_PCATCH
)
2774 lkflags
|= LK_PCATCH
;
2775 error
= BUF_TIMELOCK(bp
, lkflags
, "getblk", slptimeo
);
2778 if (error
== ENOLCK
)
2782 /* buffer may have changed on us */
2787 * Once the buffer has been locked, make sure we didn't race
2788 * a buffer recyclement. Buffers that are no longer hashed
2789 * will have b_vp == NULL, so this takes care of that check
2792 if (bp
->b_vp
!= vp
|| bp
->b_loffset
!= loffset
) {
2794 kprintf("Warning buffer %p (vp %p loffset %lld) "
2796 bp
, vp
, (long long)loffset
);
2803 * If SZMATCH any pre-existing buffer must be of the requested
2804 * size or NULL is returned. The caller absolutely does not
2805 * want getblk() to bwrite() the buffer on a size mismatch.
2807 if ((blkflags
& GETBLK_SZMATCH
) && size
!= bp
->b_bcount
) {
2813 * All vnode-based buffers must be backed by a VM object.
2815 * Set B_KVABIO for any incidental work, we will fix it
2818 KKASSERT(bp
->b_flags
& B_VMIO
);
2819 KKASSERT(bp
->b_cmd
== BUF_CMD_DONE
);
2820 bp
->b_flags
&= ~B_AGE
;
2821 bp
->b_flags
|= B_KVABIO
;
2824 * Make sure that B_INVAL buffers do not have a cached
2825 * block number translation.
2827 if ((bp
->b_flags
& B_INVAL
) &&
2828 (bp
->b_bio2
.bio_offset
!= NOOFFSET
)) {
2829 kprintf("Warning invalid buffer %p (vp %p loffset %lld)"
2830 " did not have cleared bio_offset cache\n",
2831 bp
, vp
, (long long)loffset
);
2832 clearbiocache(&bp
->b_bio2
);
2836 * The buffer is locked. B_CACHE is cleared if the buffer is
2839 * After the bremfree(), disposals must use b[q]relse().
2841 if (bp
->b_flags
& B_INVAL
)
2842 bp
->b_flags
&= ~B_CACHE
;
2846 * Any size inconsistancy with a dirty buffer or a buffer
2847 * with a softupdates dependancy must be resolved. Resizing
2848 * the buffer in such circumstances can lead to problems.
2850 * Dirty or dependant buffers are written synchronously.
2851 * Other types of buffers are simply released and
2852 * reconstituted as they may be backed by valid, dirty VM
2853 * pages (but not marked B_DELWRI).
2855 * NFS NOTE: NFS buffers which straddle EOF are oddly-sized
2856 * and may be left over from a prior truncation (and thus
2857 * no longer represent the actual EOF point), so we
2858 * definitely do not want to B_NOCACHE the backing store.
2860 if (size
!= bp
->b_bcount
) {
2861 if (bp
->b_flags
& B_DELWRI
) {
2862 bp
->b_flags
|= B_RELBUF
;
2864 } else if (LIST_FIRST(&bp
->b_dep
)) {
2865 bp
->b_flags
|= B_RELBUF
;
2868 bp
->b_flags
|= B_RELBUF
;
2873 KKASSERT(size
<= bp
->b_kvasize
);
2874 KASSERT(bp
->b_loffset
!= NOOFFSET
,
2875 ("getblk: no buffer offset"));
2878 * A buffer with B_DELWRI set and B_CACHE clear must
2879 * be committed before we can return the buffer in
2880 * order to prevent the caller from issuing a read
2881 * ( due to B_CACHE not being set ) and overwriting
2884 * Most callers, including NFS and FFS, need this to
2885 * operate properly either because they assume they
2886 * can issue a read if B_CACHE is not set, or because
2887 * ( for example ) an uncached B_DELWRI might loop due
2888 * to softupdates re-dirtying the buffer. In the latter
2889 * case, B_CACHE is set after the first write completes,
2890 * preventing further loops.
2892 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2893 * above while extending the buffer, we cannot allow the
2894 * buffer to remain with B_CACHE set after the write
2895 * completes or it will represent a corrupt state. To
2896 * deal with this we set B_NOCACHE to scrap the buffer
2899 * XXX Should this be B_RELBUF instead of B_NOCACHE?
2900 * I'm not even sure this state is still possible
2901 * now that getblk() writes out any dirty buffers
2904 * We might be able to do something fancy, like setting
2905 * B_CACHE in bwrite() except if B_DELWRI is already set,
2906 * so the below call doesn't set B_CACHE, but that gets real
2907 * confusing. This is much easier.
2909 if ((bp
->b_flags
& (B_CACHE
|B_DELWRI
)) == B_DELWRI
) {
2910 kprintf("getblk: Warning, bp %p loff=%jx DELWRI set "
2911 "and CACHE clear, b_flags %08x\n",
2912 bp
, (uintmax_t)bp
->b_loffset
, bp
->b_flags
);
2913 bp
->b_flags
|= B_NOCACHE
;
2919 * Buffer is not in-core, create new buffer. The buffer
2920 * returned by getnewbuf() is locked. Note that the returned
2921 * buffer is also considered valid (not marked B_INVAL).
2923 * Calculating the offset for the I/O requires figuring out
2924 * the block size. We use DEV_BSIZE for VBLK or VCHR and
2925 * the mount's f_iosize otherwise. If the vnode does not
2926 * have an associated mount we assume that the passed size is
2929 * Note that vn_isdisk() cannot be used here since it may
2930 * return a failure for numerous reasons. Note that the
2931 * buffer size may be larger then the block size (the caller
2932 * will use block numbers with the proper multiple). Beware
2933 * of using any v_* fields which are part of unions. In
2934 * particular, in DragonFly the mount point overloading
2935 * mechanism uses the namecache only and the underlying
2936 * directory vnode is not a special case.
2940 if (vp
->v_type
== VBLK
|| vp
->v_type
== VCHR
)
2942 else if (vp
->v_mount
)
2943 bsize
= vp
->v_mount
->mnt_stat
.f_iosize
;
2947 maxsize
= size
+ (loffset
& PAGE_MASK
);
2948 maxsize
= imax(maxsize
, bsize
);
2950 bp
= getnewbuf(blkflags
, slptimeo
, size
, maxsize
);
2952 if (slpflags
|| slptimeo
)
2958 * Atomically insert the buffer into the hash, so that it can
2959 * be found by findblk().
2961 * If bgetvp() returns non-zero a collision occured, and the
2962 * bp will not be associated with the vnode.
2964 * Make sure the translation layer has been cleared.
2966 bp
->b_loffset
= loffset
;
2967 bp
->b_bio2
.bio_offset
= NOOFFSET
;
2968 /* bp->b_bio2.bio_next = NULL; */
2970 if (bgetvp(vp
, bp
, size
)) {
2971 bp
->b_flags
|= B_INVAL
;
2977 * All vnode-based buffers must be backed by a VM object.
2979 * Set B_KVABIO for incidental work
2981 KKASSERT(vp
->v_object
!= NULL
);
2982 bp
->b_flags
|= B_VMIO
| B_KVABIO
;
2983 KKASSERT(bp
->b_cmd
== BUF_CMD_DONE
);
2989 * Do the nasty smp broadcast (if the buffer needs it) when KVABIO
2992 if (bp
&& (blkflags
& GETBLK_KVABIO
) == 0) {
3001 * Reacquire a buffer that was previously released to the locked queue,
3002 * or reacquire a buffer which is interlocked by having bioops->io_deallocate
3003 * set B_LOCKED (which handles the acquisition race).
3005 * To this end, either B_LOCKED must be set or the dependancy list must be
3009 regetblk(struct buf
*bp
)
3011 KKASSERT((bp
->b_flags
& B_LOCKED
) || LIST_FIRST(&bp
->b_dep
) != NULL
);
3012 BUF_LOCK(bp
, LK_EXCLUSIVE
| LK_RETRY
);
3019 * This code constitutes the buffer memory from either anonymous system
3020 * memory (in the case of non-VMIO operations) or from an associated
3021 * VM object (in the case of VMIO operations). This code is able to
3022 * resize a buffer up or down.
3024 * Note that this code is tricky, and has many complications to resolve
3025 * deadlock or inconsistant data situations. Tread lightly!!!
3026 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
3027 * the caller. Calling this code willy nilly can result in the loss of
3030 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
3031 * B_CACHE for the non-VMIO case.
3033 * This routine does not need to be called from a critical section but you
3034 * must own the buffer.
3037 allocbuf(struct buf
*bp
, int size
)
3044 if (BUF_LOCKINUSE(bp
) == 0)
3045 panic("allocbuf: buffer not busy");
3047 if (bp
->b_kvasize
< size
)
3048 panic("allocbuf: buffer too small");
3050 KKASSERT(bp
->b_flags
& B_VMIO
);
3052 newbsize
= roundup2(size
, DEV_BSIZE
);
3053 desiredpages
= ((int)(bp
->b_loffset
& PAGE_MASK
) +
3054 newbsize
+ PAGE_MASK
) >> PAGE_SHIFT
;
3055 KKASSERT(desiredpages
<= XIO_INTERNAL_PAGES
);
3058 * Set B_CACHE initially if buffer is 0 length or will become
3061 if (size
== 0 || bp
->b_bufsize
== 0)
3062 bp
->b_flags
|= B_CACHE
;
3064 if (newbsize
< bp
->b_bufsize
) {
3066 * DEV_BSIZE aligned new buffer size is less then the
3067 * DEV_BSIZE aligned existing buffer size. Figure out
3068 * if we have to remove any pages.
3070 if (desiredpages
< bp
->b_xio
.xio_npages
) {
3071 for (i
= desiredpages
; i
< bp
->b_xio
.xio_npages
; i
++) {
3073 * the page is not freed here -- it
3074 * is the responsibility of
3075 * vnode_pager_setsize
3077 m
= bp
->b_xio
.xio_pages
[i
];
3078 KASSERT(m
!= bogus_page
,
3079 ("allocbuf: bogus page found"));
3080 vm_page_busy_wait(m
, TRUE
, "biodep");
3081 bp
->b_xio
.xio_pages
[i
] = NULL
;
3082 vm_page_unwire(m
, 0);
3085 pmap_qremove_noinval((vm_offset_t
)
3086 trunc_page((vm_offset_t
)bp
->b_data
) +
3087 (desiredpages
<< PAGE_SHIFT
),
3088 (bp
->b_xio
.xio_npages
- desiredpages
));
3089 bp
->b_xio
.xio_npages
= desiredpages
;
3092 * Don't bother invalidating the pmap changes
3093 * (which wastes global SMP invalidation IPIs)
3094 * when setting the size to 0. This case occurs
3095 * when called via getnewbuf() during buffer
3098 if (desiredpages
== 0) {
3099 CPUMASK_ASSZERO(bp
->b_cpumask
);
3104 } else if (size
> bp
->b_bcount
) {
3106 * We are growing the buffer, possibly in a
3107 * byte-granular fashion.
3115 * Step 1, bring in the VM pages from the object,
3116 * allocating them if necessary. We must clear
3117 * B_CACHE if these pages are not valid for the
3118 * range covered by the buffer.
3123 vm_object_hold(obj
);
3124 while (bp
->b_xio
.xio_npages
< desiredpages
) {
3129 pi
= OFF_TO_IDX(bp
->b_loffset
) +
3130 bp
->b_xio
.xio_npages
;
3133 * Blocking on m->busy_count might lead to a
3136 * vm_fault->getpages->cluster_read->allocbuf
3138 m
= vm_page_lookup_busy_try(obj
, pi
, FALSE
,
3141 vm_page_sleep_busy(m
, FALSE
, "pgtblk");
3146 * note: must allocate system pages
3147 * since blocking here could intefere
3148 * with paging I/O, no matter which
3151 m
= bio_page_alloc(bp
, obj
, pi
,
3153 bp
->b_xio
.xio_npages
);
3157 bp
->b_flags
&= ~B_CACHE
;
3158 bp
->b_xio
.xio_pages
[bp
->b_xio
.xio_npages
] = m
;
3159 ++bp
->b_xio
.xio_npages
;
3165 * We found a page and were able to busy it.
3169 bp
->b_xio
.xio_pages
[bp
->b_xio
.xio_npages
] = m
;
3170 ++bp
->b_xio
.xio_npages
;
3171 if (bp
->b_act_count
< m
->act_count
)
3172 bp
->b_act_count
= m
->act_count
;
3174 vm_object_drop(obj
);
3177 * Step 2. We've loaded the pages into the buffer,
3178 * we have to figure out if we can still have B_CACHE
3179 * set. Note that B_CACHE is set according to the
3180 * byte-granular range ( bcount and size ), not the
3181 * aligned range ( newbsize ).
3183 * The VM test is against m->valid, which is DEV_BSIZE
3184 * aligned. Needless to say, the validity of the data
3185 * needs to also be DEV_BSIZE aligned. Note that this
3186 * fails with NFS if the server or some other client
3187 * extends the file's EOF. If our buffer is resized,
3188 * B_CACHE may remain set! XXX
3191 toff
= bp
->b_bcount
;
3192 tinc
= PAGE_SIZE
- ((bp
->b_loffset
+ toff
) & PAGE_MASK
);
3194 while ((bp
->b_flags
& B_CACHE
) && toff
< size
) {
3197 if (tinc
> (size
- toff
))
3200 pi
= ((bp
->b_loffset
& PAGE_MASK
) + toff
) >>
3208 bp
->b_xio
.xio_pages
[pi
]
3215 * Step 3, fixup the KVM pmap. Remember that
3216 * bp->b_data is relative to bp->b_loffset, but
3217 * bp->b_loffset may be offset into the first page.
3219 bp
->b_data
= (caddr_t
)trunc_page((vm_offset_t
)bp
->b_data
);
3220 pmap_qenter_noinval((vm_offset_t
)bp
->b_data
,
3221 bp
->b_xio
.xio_pages
, bp
->b_xio
.xio_npages
);
3222 bp
->b_data
= (caddr_t
)((vm_offset_t
)bp
->b_data
|
3223 (vm_offset_t
)(bp
->b_loffset
& PAGE_MASK
));
3226 atomic_add_long(&bufspace
, newbsize
- bp
->b_bufsize
);
3228 /* adjust space use on already-dirty buffer */
3229 if (bp
->b_flags
& B_DELWRI
) {
3230 /* dirtykvaspace unchanged */
3231 atomic_add_long(&dirtybufspace
, newbsize
- bp
->b_bufsize
);
3232 if (bp
->b_flags
& B_HEAVY
) {
3233 atomic_add_long(&dirtybufspacehw
,
3234 newbsize
- bp
->b_bufsize
);
3237 bp
->b_bufsize
= newbsize
; /* actual buffer allocation */
3238 bp
->b_bcount
= size
; /* requested buffer size */
3245 * Wait for buffer I/O completion, returning error status. B_EINTR
3246 * is converted into an EINTR error but not cleared (since a chain
3247 * of biowait() calls may occur).
3249 * On return bpdone() will have been called but the buffer will remain
3250 * locked and will not have been brelse()'d.
3252 * NOTE! If a timeout is specified and ETIMEDOUT occurs the I/O is
3253 * likely still in progress on return.
3255 * NOTE! This operation is on a BIO, not a BUF.
3257 * NOTE! BIO_DONE is cleared by vn_strategy()
3260 _biowait(struct bio
*bio
, const char *wmesg
, int to
)
3262 struct buf
*bp
= bio
->bio_buf
;
3267 KKASSERT(bio
== &bp
->b_bio1
);
3269 flags
= bio
->bio_flags
;
3270 if (flags
& BIO_DONE
)
3272 nflags
= flags
| BIO_WANT
;
3273 tsleep_interlock(bio
, 0);
3274 if (atomic_cmpset_int(&bio
->bio_flags
, flags
, nflags
)) {
3276 error
= tsleep(bio
, PINTERLOCKED
, wmesg
, to
);
3277 else if (bp
->b_cmd
== BUF_CMD_READ
)
3278 error
= tsleep(bio
, PINTERLOCKED
, "biord", to
);
3280 error
= tsleep(bio
, PINTERLOCKED
, "biowr", to
);
3282 kprintf("tsleep error biowait %d\n", error
);
3291 KKASSERT(bp
->b_cmd
== BUF_CMD_DONE
);
3292 bio
->bio_flags
&= ~(BIO_DONE
| BIO_SYNC
);
3293 if (bp
->b_flags
& B_EINTR
)
3295 if (bp
->b_flags
& B_ERROR
)
3296 return (bp
->b_error
? bp
->b_error
: EIO
);
3301 biowait(struct bio
*bio
, const char *wmesg
)
3303 return(_biowait(bio
, wmesg
, 0));
3307 biowait_timeout(struct bio
*bio
, const char *wmesg
, int to
)
3309 return(_biowait(bio
, wmesg
, to
));
3313 * This associates a tracking count with an I/O. vn_strategy() and
3314 * dev_dstrategy() do this automatically but there are a few cases
3315 * where a vnode or device layer is bypassed when a block translation
3316 * is cached. In such cases bio_start_transaction() may be called on
3317 * the bypassed layers so the system gets an I/O in progress indication
3318 * for those higher layers.
3321 bio_start_transaction(struct bio
*bio
, struct bio_track
*track
)
3323 bio
->bio_track
= track
;
3324 bio_track_ref(track
);
3325 dsched_buf_enter(bio
->bio_buf
); /* might stack */
3329 * Initiate I/O on a vnode.
3331 * SWAPCACHE OPERATION:
3333 * Real buffer cache buffers have a non-NULL bp->b_vp. Unfortunately
3334 * devfs also uses b_vp for fake buffers so we also have to check
3335 * that B_PAGING is 0. In this case the passed 'vp' is probably the
3336 * underlying block device. The swap assignments are related to the
3337 * buffer cache buffer's b_vp, not the passed vp.
3339 * The passed vp == bp->b_vp only in the case where the strategy call
3340 * is made on the vp itself for its own buffers (a regular file or
3341 * block device vp). The filesystem usually then re-calls vn_strategy()
3342 * after translating the request to an underlying device.
3344 * Cluster buffers set B_CLUSTER and the passed vp is the vp of the
3345 * underlying buffer cache buffers.
3347 * We can only deal with page-aligned buffers at the moment, because
3348 * we can't tell what the real dirty state for pages straddling a buffer
3351 * In order to call swap_pager_strategy() we must provide the VM object
3352 * and base offset for the underlying buffer cache pages so it can find
3356 vn_strategy(struct vnode
*vp
, struct bio
*bio
)
3358 struct bio_track
*track
;
3359 struct buf
*bp
= bio
->bio_buf
;
3361 KKASSERT(bp
->b_cmd
!= BUF_CMD_DONE
);
3364 * Set when an I/O is issued on the bp. Cleared by consumers
3365 * (aka HAMMER), allowing the consumer to determine if I/O had
3366 * actually occurred.
3368 bp
->b_flags
|= B_IOISSUED
;
3371 * Handle the swapcache intercept.
3373 * NOTE: The swapcache itself always supports KVABIO and will
3374 * do the right thing if its underlying devices do not.
3376 if (vn_cache_strategy(vp
, bio
))
3380 * If the vnode does not support KVABIO and the buffer is using
3381 * KVABIO, we must synchronize b_data to all cpus before dispatching.
3383 if ((vp
->v_flag
& VKVABIO
) == 0 && (bp
->b_flags
& B_KVABIO
))
3387 * Otherwise do the operation through the filesystem
3389 if (bp
->b_cmd
== BUF_CMD_READ
)
3390 track
= &vp
->v_track_read
;
3392 track
= &vp
->v_track_write
;
3393 KKASSERT((bio
->bio_flags
& BIO_DONE
) == 0);
3394 bio
->bio_track
= track
;
3395 bio_track_ref(track
);
3396 dsched_buf_enter(bp
); /* might stack */
3397 vop_strategy(*vp
->v_ops
, vp
, bio
);
3401 * vn_cache_strategy()
3403 * Returns 1 if the interrupt was successful, 0 if not.
3405 * NOTE: This function supports the KVABIO API wherein b_data might not
3406 * be synchronized to the current cpu.
3408 static void vn_cache_strategy_callback(struct bio
*bio
);
3411 vn_cache_strategy(struct vnode
*vp
, struct bio
*bio
)
3413 struct buf
*bp
= bio
->bio_buf
;
3420 * Stop using swapcache if paniced, dumping, or dumped
3422 if (panicstr
|| dumping
)
3426 * Is this buffer cache buffer suitable for reading from
3429 if (vm_swapcache_read_enable
== 0 ||
3430 bp
->b_cmd
!= BUF_CMD_READ
||
3431 ((bp
->b_flags
& B_CLUSTER
) == 0 &&
3432 (bp
->b_vp
== NULL
|| (bp
->b_flags
& B_PAGING
))) ||
3433 ((int)bp
->b_loffset
& PAGE_MASK
) != 0 ||
3434 (bp
->b_bcount
& PAGE_MASK
) != 0) {
3439 * Figure out the original VM object (it will match the underlying
3440 * VM pages). Note that swap cached data uses page indices relative
3441 * to that object, not relative to bio->bio_offset.
3443 if (bp
->b_flags
& B_CLUSTER
)
3444 object
= vp
->v_object
;
3446 object
= bp
->b_vp
->v_object
;
3449 * In order to be able to use the swap cache all underlying VM
3450 * pages must be marked as such, and we can't have any bogus pages.
3452 for (i
= 0; i
< bp
->b_xio
.xio_npages
; ++i
) {
3453 m
= bp
->b_xio
.xio_pages
[i
];
3454 if ((m
->flags
& PG_SWAPPED
) == 0)
3456 if (m
== bogus_page
)
3461 * If we are good then issue the I/O using swap_pager_strategy().
3463 * We can only do this if the buffer actually supports object-backed
3464 * I/O. If it doesn't npages will be 0.
3466 if (i
&& i
== bp
->b_xio
.xio_npages
) {
3467 m
= bp
->b_xio
.xio_pages
[0];
3468 nbio
= push_bio(bio
);
3469 nbio
->bio_done
= vn_cache_strategy_callback
;
3470 nbio
->bio_offset
= ptoa(m
->pindex
);
3471 KKASSERT(m
->object
== object
);
3472 swap_pager_strategy(object
, nbio
);
3479 * This is a bit of a hack but since the vn_cache_strategy() function can
3480 * override a VFS's strategy function we must make sure that the bio, which
3481 * is probably bio2, doesn't leak an unexpected offset value back to the
3482 * filesystem. The filesystem (e.g. UFS) might otherwise assume that the
3483 * bio went through its own file strategy function and the the bio2 offset
3484 * is a cached disk offset when, in fact, it isn't.
3487 vn_cache_strategy_callback(struct bio
*bio
)
3489 bio
->bio_offset
= NOOFFSET
;
3490 biodone(pop_bio(bio
));
3496 * Finish I/O on a buffer after all BIOs have been processed.
3497 * Called when the bio chain is exhausted or by biowait. If called
3498 * by biowait, elseit is typically 0.
3500 * bpdone is also responsible for setting B_CACHE in a B_VMIO bp.
3501 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3502 * assuming B_INVAL is clear.
3504 * For the VMIO case, we set B_CACHE if the op was a read and no
3505 * read error occured, or if the op was a write. B_CACHE is never
3506 * set if the buffer is invalid or otherwise uncacheable.
3508 * bpdone does not mess with B_INVAL, allowing the I/O routine or the
3509 * initiator to leave B_INVAL set to brelse the buffer out of existance
3510 * in the biodone routine.
3512 * bpdone is responsible for calling bundirty() on the buffer after a
3513 * successful write. We previously did this prior to initiating the
3514 * write under the assumption that the buffer might be dirtied again
3515 * while the write was in progress, however doing it before-hand creates
3516 * a race condition prior to the call to vn_strategy() where the
3517 * filesystem may not be aware that a dirty buffer is present.
3518 * It should not be possible for the buffer or its underlying pages to
3519 * be redirtied prior to bpdone()'s unbusying of the underlying VM
3523 bpdone(struct buf
*bp
, int elseit
)
3527 KASSERT(BUF_LOCKINUSE(bp
), ("bpdone: bp %p not busy", bp
));
3528 KASSERT(bp
->b_cmd
!= BUF_CMD_DONE
,
3529 ("bpdone: bp %p already done!", bp
));
3532 * No more BIOs are left. All completion functions have been dealt
3533 * with, now we clean up the buffer.
3536 bp
->b_cmd
= BUF_CMD_DONE
;
3539 * Only reads and writes are processed past this point.
3541 if (cmd
!= BUF_CMD_READ
&& cmd
!= BUF_CMD_WRITE
) {
3542 if (cmd
== BUF_CMD_FREEBLKS
)
3543 bp
->b_flags
|= B_NOCACHE
;
3550 * A failed write must re-dirty the buffer unless B_INVAL
3553 * A successful write must clear the dirty flag. This is done after
3554 * the write to ensure that the buffer remains on the vnode's dirty
3555 * list for filesystem interlocks / checks until the write is actually
3556 * complete. HAMMER2 is sensitive to this issue.
3558 * Only applicable to normal buffers (with VPs). vinum buffers may
3561 * Must be done prior to calling buf_complete() as the callback might
3562 * re-dirty the buffer.
3564 if (cmd
== BUF_CMD_WRITE
) {
3565 if ((bp
->b_flags
& (B_ERROR
| B_INVAL
)) == B_ERROR
) {
3566 bp
->b_flags
&= ~B_NOCACHE
;
3576 * Warning: softupdates may re-dirty the buffer, and HAMMER can do
3577 * a lot worse. XXX - move this above the clearing of b_cmd
3579 if (LIST_FIRST(&bp
->b_dep
) != NULL
)
3582 if (bp
->b_flags
& B_VMIO
) {
3588 struct vnode
*vp
= bp
->b_vp
;
3592 #if defined(VFS_BIO_DEBUG)
3593 if (vp
->v_auxrefs
== 0)
3594 panic("bpdone: zero vnode hold count");
3595 if ((vp
->v_flag
& VOBJBUF
) == 0)
3596 panic("bpdone: vnode is not setup for merged cache");
3599 foff
= bp
->b_loffset
;
3600 KASSERT(foff
!= NOOFFSET
, ("bpdone: no buffer offset"));
3601 KASSERT(obj
!= NULL
, ("bpdone: missing VM object"));
3603 #if defined(VFS_BIO_DEBUG)
3604 if (obj
->paging_in_progress
< bp
->b_xio
.xio_npages
) {
3605 kprintf("bpdone: paging in progress(%d) < "
3606 "bp->b_xio.xio_npages(%d)\n",
3607 obj
->paging_in_progress
,
3608 bp
->b_xio
.xio_npages
);
3613 * Set B_CACHE if the op was a normal read and no error
3614 * occured. B_CACHE is set for writes in the b*write()
3617 iosize
= bp
->b_bcount
- bp
->b_resid
;
3618 if (cmd
== BUF_CMD_READ
&&
3619 (bp
->b_flags
& (B_INVAL
|B_NOCACHE
|B_ERROR
)) == 0) {
3620 bp
->b_flags
|= B_CACHE
;
3623 vm_object_hold(obj
);
3624 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
3628 resid
= ((foff
+ PAGE_SIZE
) & ~(off_t
)PAGE_MASK
) - foff
;
3633 * cleanup bogus pages, restoring the originals. Since
3634 * the originals should still be wired, we don't have
3635 * to worry about interrupt/freeing races destroying
3636 * the VM object association.
3638 m
= bp
->b_xio
.xio_pages
[i
];
3639 if (m
== bogus_page
) {
3640 if ((bp
->b_flags
& B_HASBOGUS
) == 0)
3641 panic("bpdone: bp %p corrupt bogus", bp
);
3642 m
= vm_page_lookup(obj
, OFF_TO_IDX(foff
));
3644 panic("bpdone: page disappeared");
3645 bp
->b_xio
.xio_pages
[i
] = m
;
3650 #if defined(VFS_BIO_DEBUG)
3651 if (OFF_TO_IDX(foff
) != m
->pindex
) {
3652 kprintf("bpdone: foff(%lu)/m->pindex(%ld) "
3654 (unsigned long)foff
, (long)m
->pindex
);
3659 * In the write case, the valid and clean bits are
3660 * already changed correctly (see bdwrite()), so we
3661 * only need to do this here in the read case.
3663 vm_page_busy_wait(m
, FALSE
, "bpdpgw");
3664 if (cmd
== BUF_CMD_READ
&& isbogus
== 0 && resid
> 0)
3665 vfs_clean_one_page(bp
, i
, m
);
3668 * when debugging new filesystems or buffer I/O
3669 * methods, this is the most common error that pops
3670 * up. if you see this, you have not set the page
3671 * busy flag correctly!!!
3673 if ((m
->busy_count
& PBUSY_MASK
) == 0) {
3674 kprintf("bpdone: page busy < 0, "
3675 "pindex: %d, foff: 0x(%x,%x), "
3676 "resid: %d, index: %d\n",
3677 (int) m
->pindex
, (int)(foff
>> 32),
3678 (int) foff
& 0xffffffff, resid
, i
);
3679 if (!vn_isdisk(vp
, NULL
))
3680 kprintf(" iosize: %ld, loffset: %lld, "
3681 "flags: 0x%08x, npages: %d\n",
3682 bp
->b_vp
->v_mount
->mnt_stat
.f_iosize
,
3683 (long long)bp
->b_loffset
,
3684 bp
->b_flags
, bp
->b_xio
.xio_npages
);
3686 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n",
3687 (long long)bp
->b_loffset
,
3688 bp
->b_flags
, bp
->b_xio
.xio_npages
);
3689 kprintf(" valid: 0x%x, dirty: 0x%x, "
3693 panic("bpdone: page busy < 0");
3695 vm_page_io_finish(m
);
3697 vm_object_pip_wakeup(obj
);
3698 foff
= (foff
+ PAGE_SIZE
) & ~(off_t
)PAGE_MASK
;
3701 if (bp
->b_flags
& B_HASBOGUS
) {
3702 pmap_qenter_noinval(trunc_page((vm_offset_t
)bp
->b_data
),
3703 bp
->b_xio
.xio_pages
,
3704 bp
->b_xio
.xio_npages
);
3705 bp
->b_flags
&= ~B_HASBOGUS
;
3708 vm_object_drop(obj
);
3712 * Finish up by releasing the buffer. There are no more synchronous
3713 * or asynchronous completions, those were handled by bio_done
3717 if (bp
->b_flags
& (B_NOCACHE
|B_INVAL
|B_ERROR
|B_RELBUF
))
3728 biodone(struct bio
*bio
)
3730 struct buf
*bp
= bio
->bio_buf
;
3732 runningbufwakeup(bp
);
3735 * Run up the chain of BIO's. Leave b_cmd intact for the duration.
3738 biodone_t
*done_func
;
3739 struct bio_track
*track
;
3742 * BIO tracking. Most but not all BIOs are tracked.
3744 if ((track
= bio
->bio_track
) != NULL
) {
3745 bio_track_rel(track
);
3746 bio
->bio_track
= NULL
;
3750 * A bio_done function terminates the loop. The function
3751 * will be responsible for any further chaining and/or
3752 * buffer management.
3754 * WARNING! The done function can deallocate the buffer!
3756 if ((done_func
= bio
->bio_done
) != NULL
) {
3757 bio
->bio_done
= NULL
;
3761 bio
= bio
->bio_prev
;
3765 * If we've run out of bio's do normal [a]synchronous completion.
3771 * Synchronous biodone - this terminates a synchronous BIO.
3773 * bpdone() is called with elseit=FALSE, leaving the buffer completed
3774 * but still locked. The caller must brelse() the buffer after waiting
3778 biodone_sync(struct bio
*bio
)
3780 struct buf
*bp
= bio
->bio_buf
;
3784 KKASSERT(bio
== &bp
->b_bio1
);
3788 flags
= bio
->bio_flags
;
3789 nflags
= (flags
| BIO_DONE
) & ~BIO_WANT
;
3791 if (atomic_cmpset_int(&bio
->bio_flags
, flags
, nflags
)) {
3792 if (flags
& BIO_WANT
)
3802 * This routine is called in lieu of iodone in the case of
3803 * incomplete I/O. This keeps the busy status for pages
3807 vfs_unbusy_pages(struct buf
*bp
)
3811 runningbufwakeup(bp
);
3813 if (bp
->b_flags
& B_VMIO
) {
3814 struct vnode
*vp
= bp
->b_vp
;
3818 vm_object_hold(obj
);
3820 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
3821 vm_page_t m
= bp
->b_xio
.xio_pages
[i
];
3824 * When restoring bogus changes the original pages
3825 * should still be wired, so we are in no danger of
3826 * losing the object association and do not need
3827 * critical section protection particularly.
3829 if (m
== bogus_page
) {
3830 m
= vm_page_lookup(obj
, OFF_TO_IDX(bp
->b_loffset
) + i
);
3832 panic("vfs_unbusy_pages: page missing");
3834 bp
->b_xio
.xio_pages
[i
] = m
;
3836 vm_page_busy_wait(m
, FALSE
, "bpdpgw");
3837 vm_page_io_finish(m
);
3839 vm_object_pip_wakeup(obj
);
3841 if (bp
->b_flags
& B_HASBOGUS
) {
3842 pmap_qenter_noinval(trunc_page((vm_offset_t
)bp
->b_data
),
3843 bp
->b_xio
.xio_pages
,
3844 bp
->b_xio
.xio_npages
);
3845 bp
->b_flags
&= ~B_HASBOGUS
;
3848 vm_object_drop(obj
);
3855 * This routine is called before a device strategy routine.
3856 * It is used to tell the VM system that paging I/O is in
3857 * progress, and treat the pages associated with the buffer
3858 * almost as being PBUSY_LOCKED. Also the object 'paging_in_progress'
3859 * flag is handled to make sure that the object doesn't become
3862 * Since I/O has not been initiated yet, certain buffer flags
3863 * such as B_ERROR or B_INVAL may be in an inconsistant state
3864 * and should be ignored.
3867 vfs_busy_pages(struct vnode
*vp
, struct buf
*bp
)
3870 struct lwp
*lp
= curthread
->td_lwp
;
3873 * The buffer's I/O command must already be set. If reading,
3874 * B_CACHE must be 0 (double check against callers only doing
3875 * I/O when B_CACHE is 0).
3877 KKASSERT(bp
->b_cmd
!= BUF_CMD_DONE
);
3878 KKASSERT(bp
->b_cmd
== BUF_CMD_WRITE
|| (bp
->b_flags
& B_CACHE
) == 0);
3880 if (bp
->b_flags
& B_VMIO
) {
3884 KASSERT(bp
->b_loffset
!= NOOFFSET
,
3885 ("vfs_busy_pages: no buffer offset"));
3888 * Busy all the pages. We have to busy them all at once
3889 * to avoid deadlocks.
3892 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
3893 vm_page_t m
= bp
->b_xio
.xio_pages
[i
];
3895 if (vm_page_busy_try(m
, FALSE
)) {
3896 vm_page_sleep_busy(m
, FALSE
, "vbpage");
3898 vm_page_wakeup(bp
->b_xio
.xio_pages
[i
]);
3904 * Setup for I/O, soft-busy the page right now because
3905 * the next loop may block.
3907 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
3908 vm_page_t m
= bp
->b_xio
.xio_pages
[i
];
3910 if ((bp
->b_flags
& B_CLUSTER
) == 0) {
3911 vm_object_pip_add(obj
, 1);
3912 vm_page_io_start(m
);
3917 * Adjust protections for I/O and do bogus-page mapping.
3918 * Assume that vm_page_protect() can block (it can block
3919 * if VM_PROT_NONE, don't take any chances regardless).
3921 * In particular note that for writes we must incorporate
3922 * page dirtyness from the VM system into the buffer's
3925 * For reads we theoretically must incorporate page dirtyness
3926 * from the VM system to determine if the page needs bogus
3927 * replacement, but we shortcut the test by simply checking
3928 * that all m->valid bits are set, indicating that the page
3929 * is fully valid and does not need to be re-read. For any
3930 * VM system dirtyness the page will also be fully valid
3931 * since it was mapped at one point.
3934 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
3935 vm_page_t m
= bp
->b_xio
.xio_pages
[i
];
3937 if (bp
->b_cmd
== BUF_CMD_WRITE
) {
3939 * When readying a vnode-backed buffer for
3940 * a write we must zero-fill any invalid
3941 * portions of the backing VM pages, mark
3942 * it valid and clear related dirty bits.
3944 * vfs_clean_one_page() incorporates any
3945 * VM dirtyness and updates the b_dirtyoff
3946 * range (after we've made the page RO).
3948 * It is also expected that the pmap modified
3949 * bit has already been cleared by the
3950 * vm_page_protect(). We may not be able
3951 * to clear all dirty bits for a page if it
3952 * was also memory mapped (NFS).
3954 * Finally be sure to unassign any swap-cache
3955 * backing store as it is now stale.
3957 vm_page_protect(m
, VM_PROT_READ
);
3958 vfs_clean_one_page(bp
, i
, m
);
3959 swap_pager_unswapped(m
);
3960 } else if (m
->valid
== VM_PAGE_BITS_ALL
) {
3962 * When readying a vnode-backed buffer for
3963 * read we must replace any dirty pages with
3964 * a bogus page so dirty data is not destroyed
3965 * when filling gaps.
3967 * To avoid testing whether the page is
3968 * dirty we instead test that the page was
3969 * at some point mapped (m->valid fully
3970 * valid) with the understanding that
3971 * this also covers the dirty case.
3973 bp
->b_xio
.xio_pages
[i
] = bogus_page
;
3974 bp
->b_flags
|= B_HASBOGUS
;
3976 } else if (m
->valid
& m
->dirty
) {
3978 * This case should not occur as partial
3979 * dirtyment can only happen if the buffer
3980 * is B_CACHE, and this code is not entered
3981 * if the buffer is B_CACHE.
3983 kprintf("Warning: vfs_busy_pages - page not "
3984 "fully valid! loff=%jx bpf=%08x "
3985 "idx=%d val=%02x dir=%02x\n",
3986 (uintmax_t)bp
->b_loffset
, bp
->b_flags
,
3987 i
, m
->valid
, m
->dirty
);
3988 vm_page_protect(m
, VM_PROT_NONE
);
3991 * The page is not valid and can be made
3994 vm_page_protect(m
, VM_PROT_NONE
);
3999 pmap_qenter_noinval(trunc_page((vm_offset_t
)bp
->b_data
),
4000 bp
->b_xio
.xio_pages
,
4001 bp
->b_xio
.xio_npages
);
4007 * This is the easiest place to put the process accounting for the I/O
4011 if (bp
->b_cmd
== BUF_CMD_READ
)
4012 lp
->lwp_ru
.ru_inblock
++;
4014 lp
->lwp_ru
.ru_oublock
++;
4019 * Tell the VM system that the pages associated with this buffer
4020 * are clean. This is used for delayed writes where the data is
4021 * going to go to disk eventually without additional VM intevention.
4023 * NOTE: While we only really need to clean through to b_bcount, we
4024 * just go ahead and clean through to b_bufsize.
4027 vfs_clean_pages(struct buf
*bp
)
4032 if ((bp
->b_flags
& B_VMIO
) == 0)
4035 KASSERT(bp
->b_loffset
!= NOOFFSET
,
4036 ("vfs_clean_pages: no buffer offset"));
4038 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
4039 m
= bp
->b_xio
.xio_pages
[i
];
4040 vfs_clean_one_page(bp
, i
, m
);
4045 * vfs_clean_one_page:
4047 * Set the valid bits and clear the dirty bits in a page within a
4048 * buffer. The range is restricted to the buffer's size and the
4049 * buffer's logical offset might index into the first page.
4051 * The caller has busied or soft-busied the page and it is not mapped,
4052 * test and incorporate the dirty bits into b_dirtyoff/end before
4053 * clearing them. Note that we need to clear the pmap modified bits
4054 * after determining the the page was dirty, vm_page_set_validclean()
4055 * does not do it for us.
4057 * This routine is typically called after a read completes (dirty should
4058 * be zero in that case as we are not called on bogus-replace pages),
4059 * or before a write is initiated.
4062 vfs_clean_one_page(struct buf
*bp
, int pageno
, vm_page_t m
)
4070 * Calculate offset range within the page but relative to buffer's
4071 * loffset. loffset might be offset into the first page.
4073 xoff
= (int)bp
->b_loffset
& PAGE_MASK
; /* loffset offset into pg 0 */
4074 bcount
= bp
->b_bcount
+ xoff
; /* offset adjusted */
4080 soff
= (pageno
<< PAGE_SHIFT
);
4081 eoff
= soff
+ PAGE_SIZE
;
4089 * Test dirty bits and adjust b_dirtyoff/end.
4091 * If dirty pages are incorporated into the bp any prior
4092 * B_NEEDCOMMIT state (NFS) must be cleared because the
4093 * caller has not taken into account the new dirty data.
4095 * If the page was memory mapped the dirty bits might go beyond the
4096 * end of the buffer, but we can't really make the assumption that
4097 * a file EOF straddles the buffer (even though this is the case for
4098 * NFS if B_NEEDCOMMIT is also set). So for the purposes of clearing
4099 * B_NEEDCOMMIT we only test the dirty bits covered by the buffer.
4100 * This also saves some console spam.
4102 * When clearing B_NEEDCOMMIT we must also clear B_CLUSTEROK,
4103 * NFS can handle huge commits but not huge writes.
4105 vm_page_test_dirty(m
);
4107 if ((bp
->b_flags
& B_NEEDCOMMIT
) &&
4108 (m
->dirty
& vm_page_bits(soff
& PAGE_MASK
, eoff
- soff
))) {
4110 kprintf("Warning: vfs_clean_one_page: bp %p "
4111 "loff=%jx,%d flgs=%08x clr B_NEEDCOMMIT"
4112 " cmd %d vd %02x/%02x x/s/e %d %d %d "
4114 bp
, (uintmax_t)bp
->b_loffset
, bp
->b_bcount
,
4115 bp
->b_flags
, bp
->b_cmd
,
4116 m
->valid
, m
->dirty
, xoff
, soff
, eoff
,
4117 bp
->b_dirtyoff
, bp
->b_dirtyend
);
4118 bp
->b_flags
&= ~(B_NEEDCOMMIT
| B_CLUSTEROK
);
4120 print_backtrace(-1);
4123 * Only clear the pmap modified bits if ALL the dirty bits
4124 * are set, otherwise the system might mis-clear portions
4127 if (m
->dirty
== VM_PAGE_BITS_ALL
&&
4128 (bp
->b_flags
& B_NEEDCOMMIT
) == 0) {
4129 pmap_clear_modify(m
);
4131 if (bp
->b_dirtyoff
> soff
- xoff
)
4132 bp
->b_dirtyoff
= soff
- xoff
;
4133 if (bp
->b_dirtyend
< eoff
- xoff
)
4134 bp
->b_dirtyend
= eoff
- xoff
;
4138 * Set related valid bits, clear related dirty bits.
4139 * Does not mess with the pmap modified bit.
4141 * WARNING! We cannot just clear all of m->dirty here as the
4142 * buffer cache buffers may use a DEV_BSIZE'd aligned
4143 * block size, or have an odd size (e.g. NFS at file EOF).
4144 * The putpages code can clear m->dirty to 0.
4146 * If a VOP_WRITE generates a buffer cache buffer which
4147 * covers the same space as mapped writable pages the
4148 * buffer flush might not be able to clear all the dirty
4149 * bits and still require a putpages from the VM system
4152 * WARNING! vm_page_set_validclean() currently assumes vm_token
4153 * is held. The page might not be busied (bdwrite() case).
4154 * XXX remove this comment once we've validated that this
4155 * is no longer an issue.
4157 vm_page_set_validclean(m
, soff
& PAGE_MASK
, eoff
- soff
);
4162 * Similar to vfs_clean_one_page() but sets the bits to valid and dirty.
4163 * The page data is assumed to be valid (there is no zeroing here).
4166 vfs_dirty_one_page(struct buf
*bp
, int pageno
, vm_page_t m
)
4174 * Calculate offset range within the page but relative to buffer's
4175 * loffset. loffset might be offset into the first page.
4177 xoff
= (int)bp
->b_loffset
& PAGE_MASK
; /* loffset offset into pg 0 */
4178 bcount
= bp
->b_bcount
+ xoff
; /* offset adjusted */
4184 soff
= (pageno
<< PAGE_SHIFT
);
4185 eoff
= soff
+ PAGE_SIZE
;
4191 vm_page_set_validdirty(m
, soff
& PAGE_MASK
, eoff
- soff
);
4198 * Clear a buffer. This routine essentially fakes an I/O, so we need
4199 * to clear B_ERROR and B_INVAL.
4201 * Note that while we only theoretically need to clear through b_bcount,
4202 * we go ahead and clear through b_bufsize.
4205 vfs_bio_clrbuf(struct buf
*bp
)
4209 KKASSERT(bp
->b_flags
& B_VMIO
);
4211 bp
->b_flags
&= ~(B_INVAL
| B_EINTR
| B_ERROR
);
4214 if ((bp
->b_xio
.xio_npages
== 1) && (bp
->b_bufsize
< PAGE_SIZE
) &&
4215 (bp
->b_loffset
& PAGE_MASK
) == 0) {
4216 mask
= (1 << (bp
->b_bufsize
/ DEV_BSIZE
)) - 1;
4217 if ((bp
->b_xio
.xio_pages
[0]->valid
& mask
) == mask
) {
4221 if ((bp
->b_xio
.xio_pages
[0]->valid
& mask
) == 0) {
4222 bzero(bp
->b_data
, bp
->b_bufsize
);
4223 bp
->b_xio
.xio_pages
[0]->valid
|= mask
;
4229 for(i
= 0; i
< bp
->b_xio
.xio_npages
; i
++, sa
=ea
) {
4230 int j
= ((vm_offset_t
)sa
& PAGE_MASK
) / DEV_BSIZE
;
4231 ea
= (caddr_t
)trunc_page((vm_offset_t
)sa
+ PAGE_SIZE
);
4232 ea
= (caddr_t
)(vm_offset_t
)ulmin(
4233 (u_long
)(vm_offset_t
)ea
,
4234 (u_long
)(vm_offset_t
)bp
->b_data
+ bp
->b_bufsize
);
4235 mask
= ((1 << ((ea
- sa
) / DEV_BSIZE
)) - 1) << j
;
4236 if ((bp
->b_xio
.xio_pages
[i
]->valid
& mask
) == mask
)
4238 if ((bp
->b_xio
.xio_pages
[i
]->valid
& mask
) == 0) {
4241 for (; sa
< ea
; sa
+= DEV_BSIZE
, j
++) {
4242 if ((bp
->b_xio
.xio_pages
[i
]->valid
&
4244 bzero(sa
, DEV_BSIZE
);
4248 bp
->b_xio
.xio_pages
[i
]->valid
|= mask
;
4254 * Allocate a page for a buffer cache buffer.
4256 * If NULL is returned the caller is expected to retry (typically check if
4257 * the page already exists on retry before trying to allocate one).
4259 * NOTE! Low-memory handling is dealt with in b[q]relse(), not here. This
4260 * function will use the system reserve with the hope that the page
4261 * allocations can be returned to PQ_CACHE/PQ_FREE when the caller
4262 * is done with the buffer.
4264 * NOTE! However, TMPFS is a special case because flushing a dirty buffer
4265 * to TMPFS doesn't clean the page. For TMPFS, only the pagedaemon
4266 * is capable of retiring pages (to swap). For TMPFS we don't dig
4267 * into the system reserve because doing so could stall out pretty
4268 * much every process running on the system.
4272 bio_page_alloc(struct buf
*bp
, vm_object_t obj
, vm_pindex_t pg
, int deficit
)
4274 int vmflags
= VM_ALLOC_NORMAL
| VM_ALLOC_NULL_OK
;
4277 ASSERT_LWKT_TOKEN_HELD(vm_object_token(obj
));
4280 * Try a normal allocation first.
4282 p
= vm_page_alloc(obj
, pg
, vmflags
);
4285 if (vm_page_lookup(obj
, pg
))
4287 vm_pageout_deficit
+= deficit
;
4290 * Try again, digging into the system reserve.
4292 * Trying to recover pages from the buffer cache here can deadlock
4293 * against other threads trying to busy underlying pages so we
4294 * depend on the code in brelse() and bqrelse() to free/cache the
4295 * underlying buffer cache pages when memory is low.
4297 if (curthread
->td_flags
& TDF_SYSTHREAD
)
4298 vmflags
|= VM_ALLOC_SYSTEM
| VM_ALLOC_INTERRUPT
;
4299 else if (bp
->b_vp
&& bp
->b_vp
->v_tag
== VT_TMPFS
)
4302 vmflags
|= VM_ALLOC_SYSTEM
;
4304 /*recoverbufpages();*/
4305 p
= vm_page_alloc(obj
, pg
, vmflags
);
4308 if (vm_page_lookup(obj
, pg
))
4312 * Wait for memory to free up and try again
4314 if (vm_page_count_severe())
4316 vm_wait(hz
/ 20 + 1);
4318 p
= vm_page_alloc(obj
, pg
, vmflags
);
4321 if (vm_page_lookup(obj
, pg
))
4325 * Ok, now we are really in trouble.
4328 static struct krate biokrate
= { .freq
= 1 };
4329 krateprintf(&biokrate
,
4330 "Warning: bio_page_alloc: memory exhausted "
4331 "during buffer cache page allocation from %s\n",
4332 curthread
->td_comm
);
4334 if (curthread
->td_flags
& TDF_SYSTHREAD
)
4335 vm_wait(hz
/ 20 + 1);
4337 vm_wait(hz
/ 2 + 1);
4342 * The buffer's mapping has changed. Adjust the buffer's memory
4343 * synchronization. The caller is the exclusive holder of the buffer
4344 * and has set or cleared B_KVABIO according to preference.
4346 * WARNING! If the caller is using B_KVABIO mode, this function will
4347 * not map the data to the current cpu. The caller must also
4348 * call bkvasync(bp).
4351 bkvareset(struct buf
*bp
)
4353 if (bp
->b_flags
& B_KVABIO
) {
4354 CPUMASK_ASSZERO(bp
->b_cpumask
);
4356 CPUMASK_ORMASK(bp
->b_cpumask
, smp_active_mask
);
4363 * The buffer will be used by the caller on the caller's cpu, synchronize
4364 * its data to the current cpu. Caller must control the buffer by holding
4365 * its lock, but calling cpu does not necessarily have to be the owner of
4366 * the lock (i.e. HAMMER2's concurrent I/O accessors).
4368 * If B_KVABIO is not set, the buffer is already fully synchronized.
4371 bkvasync(struct buf
*bp
)
4373 int cpuid
= mycpu
->gd_cpuid
;
4376 if ((bp
->b_flags
& B_KVABIO
) &&
4377 CPUMASK_TESTBIT(bp
->b_cpumask
, cpuid
) == 0) {
4379 while (bdata
< bp
->b_data
+ bp
->b_bufsize
) {
4381 bdata
+= PAGE_SIZE
-
4382 ((intptr_t)bdata
& PAGE_MASK
);
4384 ATOMIC_CPUMASK_ORBIT(bp
->b_cpumask
, cpuid
);
4389 * The buffer will be used by a subsystem that does not understand
4390 * the KVABIO API. Make sure its data is synchronized to all cpus.
4392 * If B_KVABIO is not set, the buffer is already fully synchronized.
4394 * NOTE! This is the only safe way to clear B_KVABIO on a buffer.
4397 bkvasync_all(struct buf
*bp
)
4399 if (debug_kvabio
> 0) {
4401 print_backtrace(10);
4404 if ((bp
->b_flags
& B_KVABIO
) &&
4405 CPUMASK_CMPMASKNEQ(bp
->b_cpumask
, smp_active_mask
)) {
4408 ATOMIC_CPUMASK_ORMASK(bp
->b_cpumask
, smp_active_mask
);
4410 bp
->b_flags
&= ~B_KVABIO
;
4414 * Scan all buffers in the system and issue the callback.
4417 scan_all_buffers(int (*callback
)(struct buf
*, void *), void *info
)
4423 for (n
= 0; n
< nbuf
; ++n
) {
4424 if ((error
= callback(&buf
[n
], info
)) < 0) {
4434 * nestiobuf_iodone: biodone callback for nested buffers and propagate
4435 * completion to the master buffer.
4438 nestiobuf_iodone(struct bio
*bio
)
4441 struct buf
*mbp
, *bp
;
4442 struct devstat
*stats
;
4447 mbio
= bio
->bio_caller_info1
.ptr
;
4448 stats
= bio
->bio_caller_info2
.ptr
;
4449 mbp
= mbio
->bio_buf
;
4451 KKASSERT(bp
->b_bcount
<= bp
->b_bufsize
);
4452 KKASSERT(mbp
!= bp
);
4454 error
= bp
->b_error
;
4455 if (bp
->b_error
== 0 &&
4456 (bp
->b_bcount
< bp
->b_bufsize
|| bp
->b_resid
> 0)) {
4458 * Not all got transfered, raise an error. We have no way to
4459 * propagate these conditions to mbp.
4464 donebytes
= bp
->b_bufsize
;
4468 nestiobuf_done(mbio
, donebytes
, error
, stats
);
4472 nestiobuf_done(struct bio
*mbio
, int donebytes
, int error
, struct devstat
*stats
)
4476 mbp
= mbio
->bio_buf
;
4478 KKASSERT((int)(intptr_t)mbio
->bio_driver_info
> 0);
4481 * If an error occured, propagate it to the master buffer.
4483 * Several biodone()s may wind up running concurrently so
4484 * use an atomic op to adjust b_flags.
4487 mbp
->b_error
= error
;
4488 atomic_set_int(&mbp
->b_flags
, B_ERROR
);
4492 * Decrement the operations in progress counter and terminate the
4493 * I/O if this was the last bit.
4495 if (atomic_fetchadd_int((int *)&mbio
->bio_driver_info
, -1) == 1) {
4498 devstat_end_transaction_buf(stats
, mbp
);
4504 * Initialize a nestiobuf for use. Set an initial count of 1 to prevent
4505 * the mbio from being biodone()'d while we are still adding sub-bios to
4509 nestiobuf_init(struct bio
*bio
)
4511 bio
->bio_driver_info
= (void *)1;
4515 * The BIOs added to the nestedio have already been started, remove the
4516 * count that placeheld our mbio and biodone() it if the count would
4520 nestiobuf_start(struct bio
*mbio
)
4522 struct buf
*mbp
= mbio
->bio_buf
;
4525 * Decrement the operations in progress counter and terminate the
4526 * I/O if this was the last bit.
4528 if (atomic_fetchadd_int((int *)&mbio
->bio_driver_info
, -1) == 1) {
4529 if (mbp
->b_flags
& B_ERROR
)
4530 mbp
->b_resid
= mbp
->b_bcount
;
4538 * Set an intermediate error prior to calling nestiobuf_start()
4541 nestiobuf_error(struct bio
*mbio
, int error
)
4543 struct buf
*mbp
= mbio
->bio_buf
;
4546 mbp
->b_error
= error
;
4547 atomic_set_int(&mbp
->b_flags
, B_ERROR
);
4552 * nestiobuf_add: setup a "nested" buffer.
4554 * => 'mbp' is a "master" buffer which is being divided into sub pieces.
4555 * => 'bp' should be a buffer allocated by getiobuf.
4556 * => 'offset' is a byte offset in the master buffer.
4557 * => 'size' is a size in bytes of this nested buffer.
4560 nestiobuf_add(struct bio
*mbio
, struct buf
*bp
, int offset
, size_t size
, struct devstat
*stats
)
4562 struct buf
*mbp
= mbio
->bio_buf
;
4563 struct vnode
*vp
= mbp
->b_vp
;
4565 KKASSERT(mbp
->b_bcount
>= offset
+ size
);
4567 atomic_add_int((int *)&mbio
->bio_driver_info
, 1);
4569 /* kernel needs to own the lock for it to be released in biodone */
4572 bp
->b_cmd
= mbp
->b_cmd
;
4573 bp
->b_bio1
.bio_done
= nestiobuf_iodone
;
4574 bp
->b_data
= (char *)mbp
->b_data
+ offset
;
4575 bp
->b_resid
= bp
->b_bcount
= size
;
4576 bp
->b_bufsize
= bp
->b_bcount
;
4578 bp
->b_bio1
.bio_track
= NULL
;
4579 bp
->b_bio1
.bio_caller_info1
.ptr
= mbio
;
4580 bp
->b_bio1
.bio_caller_info2
.ptr
= stats
;
4585 DB_SHOW_COMMAND(buffer
, db_show_buffer
)
4588 struct buf
*bp
= (struct buf
*)addr
;
4591 db_printf("usage: show buffer <addr>\n");
4595 db_printf("b_flags = 0x%pb%i\n", PRINT_BUF_FLAGS
, bp
->b_flags
);
4596 db_printf("b_cmd = %d\n", bp
->b_cmd
);
4597 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, "
4598 "b_resid = %d\n, b_data = %p, "
4599 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n",
4600 bp
->b_error
, bp
->b_bufsize
, bp
->b_bcount
, bp
->b_resid
,
4602 (long long)bp
->b_bio2
.bio_offset
,
4603 (long long)(bp
->b_bio2
.bio_next
?
4604 bp
->b_bio2
.bio_next
->bio_offset
: (off_t
)-1));
4605 if (bp
->b_xio
.xio_npages
) {
4607 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ",
4608 bp
->b_xio
.xio_npages
);
4609 for (i
= 0; i
< bp
->b_xio
.xio_npages
; i
++) {
4611 m
= bp
->b_xio
.xio_pages
[i
];
4612 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m
->object
,
4613 (u_long
)m
->pindex
, (u_long
)VM_PAGE_TO_PHYS(m
));
4614 if ((i
+ 1) < bp
->b_xio
.xio_npages
)