2 * Copyright (c) 1989, 1993
3 * The Regents of the University of California. All rights reserved.
5 * This code is derived from software contributed to Berkeley by
6 * Rick Macklem at The University of Guelph.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 3. All advertising materials mentioning features or use of this software
17 * must display the following acknowledgement:
18 * This product includes software developed by the University of
19 * California, Berkeley and its contributors.
20 * 4. Neither the name of the University nor the names of its contributors
21 * may be used to endorse or promote products derived from this software
22 * without specific prior written permission.
24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * @(#)nfs_bio.c 8.9 (Berkeley) 3/30/95
37 * $FreeBSD: /repoman/r/ncvs/src/sys/nfsclient/nfs_bio.c,v 1.130 2004/04/14 23:23:55 peadar Exp $
38 * $DragonFly: src/sys/vfs/nfs/nfs_bio.c,v 1.45 2008/07/18 00:09:39 dillon Exp $
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/resourcevar.h>
45 #include <sys/signalvar.h>
48 #include <sys/vnode.h>
49 #include <sys/mount.h>
50 #include <sys/kernel.h>
52 #include <sys/msfbuf.h>
55 #include <vm/vm_extern.h>
56 #include <vm/vm_page.h>
57 #include <vm/vm_object.h>
58 #include <vm/vm_pager.h>
59 #include <vm/vnode_pager.h>
62 #include <sys/thread2.h>
63 #include <vm/vm_page2.h>
71 #include "nfsm_subs.h"
74 static struct buf
*nfs_getcacheblk(struct vnode
*vp
, off_t loffset
,
75 int size
, struct thread
*td
);
76 static int nfs_check_dirent(struct nfs_dirent
*dp
, int maxlen
);
77 static void nfsiodone_sync(struct bio
*bio
);
78 static void nfs_readrpc_bio_done(nfsm_info_t info
);
79 static void nfs_writerpc_bio_done(nfsm_info_t info
);
80 static void nfs_commitrpc_bio_done(nfsm_info_t info
);
83 * Vnode op for read using bio
86 nfs_bioread(struct vnode
*vp
, struct uio
*uio
, int ioflag
)
88 struct nfsnode
*np
= VTONFS(vp
);
90 struct buf
*bp
, *rabp
;
93 struct nfsmount
*nmp
= VFSTONFS(vp
->v_mount
);
103 if (uio
->uio_rw
!= UIO_READ
)
104 panic("nfs_read mode");
106 if (uio
->uio_resid
== 0)
108 if (uio
->uio_offset
< 0) /* XXX VDIR cookies can be negative */
112 if ((nmp
->nm_flag
& NFSMNT_NFSV3
) != 0 &&
113 (nmp
->nm_state
& NFSSTA_GOTFSINFO
) == 0)
114 (void)nfs_fsinfo(nmp
, vp
, td
);
115 if (vp
->v_type
!= VDIR
&&
116 (uio
->uio_offset
+ uio
->uio_resid
) > nmp
->nm_maxfilesize
)
118 biosize
= vp
->v_mount
->mnt_stat
.f_iosize
;
119 seqcount
= (int)((off_t
)(ioflag
>> IO_SEQSHIFT
) * biosize
/ BKVASIZE
);
122 * For nfs, cache consistency can only be maintained approximately.
123 * Although RFC1094 does not specify the criteria, the following is
124 * believed to be compatible with the reference port.
126 * NFS: If local changes have been made and this is a
127 * directory, the directory must be invalidated and
128 * the attribute cache must be cleared.
130 * GETATTR is called to synchronize the file size.
132 * If remote changes are detected local data is flushed
133 * and the cache is invalidated.
135 * NOTE: In the normal case the attribute cache is not
136 * cleared which means GETATTR may use cached data and
137 * not immediately detect changes made on the server.
139 if ((np
->n_flag
& NLMODIFIED
) && vp
->v_type
== VDIR
) {
141 error
= nfs_vinvalbuf(vp
, V_SAVE
, 1);
146 error
= VOP_GETATTR(vp
, &vattr
);
151 * This can deadlock getpages/putpages for regular
152 * files. Only do it for directories.
154 if (np
->n_flag
& NRMODIFIED
) {
155 if (vp
->v_type
== VDIR
) {
157 error
= nfs_vinvalbuf(vp
, V_SAVE
, 1);
160 np
->n_flag
&= ~NRMODIFIED
;
165 * Loop until uio exhausted or we hit EOF
170 switch (vp
->v_type
) {
172 nfsstats
.biocache_reads
++;
173 lbn
= uio
->uio_offset
/ biosize
;
174 boff
= uio
->uio_offset
& (biosize
- 1);
175 loffset
= (off_t
)lbn
* biosize
;
178 * Start the read ahead(s), as required.
180 if (nmp
->nm_readahead
> 0 && nfs_asyncok(nmp
)) {
181 for (nra
= 0; nra
< nmp
->nm_readahead
&& nra
< seqcount
&&
182 (off_t
)(lbn
+ 1 + nra
) * biosize
< np
->n_size
; nra
++) {
183 rabn
= lbn
+ 1 + nra
;
184 raoffset
= (off_t
)rabn
* biosize
;
185 if (findblk(vp
, raoffset
, FINDBLK_TEST
) == NULL
) {
186 rabp
= nfs_getcacheblk(vp
, raoffset
, biosize
, td
);
189 if ((rabp
->b_flags
& (B_CACHE
|B_DELWRI
)) == 0) {
190 rabp
->b_cmd
= BUF_CMD_READ
;
191 vfs_busy_pages(vp
, rabp
);
192 nfs_asyncio(vp
, &rabp
->b_bio2
);
201 * Obtain the buffer cache block. Figure out the buffer size
202 * when we are at EOF. If we are modifying the size of the
203 * buffer based on an EOF condition we need to hold
204 * nfs_rslock() through obtaining the buffer to prevent
205 * a potential writer-appender from messing with n_size.
206 * Otherwise we may accidently truncate the buffer and
209 * Note that bcount is *not* DEV_BSIZE aligned.
211 if (loffset
+ boff
>= np
->n_size
) {
215 bp
= nfs_getcacheblk(vp
, loffset
, biosize
, td
);
221 * If B_CACHE is not set, we must issue the read. If this
222 * fails, we return an error.
224 if ((bp
->b_flags
& B_CACHE
) == 0) {
225 bp
->b_cmd
= BUF_CMD_READ
;
226 bp
->b_bio2
.bio_done
= nfsiodone_sync
;
227 bp
->b_bio2
.bio_flags
|= BIO_SYNC
;
228 vfs_busy_pages(vp
, bp
);
229 error
= nfs_doio(vp
, &bp
->b_bio2
, td
);
237 * on is the offset into the current bp. Figure out how many
238 * bytes we can copy out of the bp. Note that bcount is
239 * NOT DEV_BSIZE aligned.
241 * Then figure out how many bytes we can copy into the uio.
244 if (n
> uio
->uio_resid
)
246 if (loffset
+ boff
+ n
> np
->n_size
)
247 n
= np
->n_size
- loffset
- boff
;
250 biosize
= min(NFS_MAXPATHLEN
, np
->n_size
);
251 nfsstats
.biocache_readlinks
++;
252 bp
= nfs_getcacheblk(vp
, (off_t
)0, biosize
, td
);
255 if ((bp
->b_flags
& B_CACHE
) == 0) {
256 bp
->b_cmd
= BUF_CMD_READ
;
257 bp
->b_bio2
.bio_done
= nfsiodone_sync
;
258 bp
->b_bio2
.bio_flags
|= BIO_SYNC
;
259 vfs_busy_pages(vp
, bp
);
260 error
= nfs_doio(vp
, &bp
->b_bio2
, td
);
262 bp
->b_flags
|= B_ERROR
| B_INVAL
;
267 n
= szmin(uio
->uio_resid
, (size_t)bp
->b_bcount
- bp
->b_resid
);
271 nfsstats
.biocache_readdirs
++;
272 if (np
->n_direofoffset
&&
273 uio
->uio_offset
>= np
->n_direofoffset
277 lbn
= (uoff_t
)uio
->uio_offset
/ NFS_DIRBLKSIZ
;
278 boff
= uio
->uio_offset
& (NFS_DIRBLKSIZ
- 1);
279 loffset
= uio
->uio_offset
- boff
;
280 bp
= nfs_getcacheblk(vp
, loffset
, NFS_DIRBLKSIZ
, td
);
284 if ((bp
->b_flags
& B_CACHE
) == 0) {
285 bp
->b_cmd
= BUF_CMD_READ
;
286 bp
->b_bio2
.bio_done
= nfsiodone_sync
;
287 bp
->b_bio2
.bio_flags
|= BIO_SYNC
;
288 vfs_busy_pages(vp
, bp
);
289 error
= nfs_doio(vp
, &bp
->b_bio2
, td
);
292 while (error
== NFSERR_BAD_COOKIE
) {
293 kprintf("got bad cookie vp %p bp %p\n", vp
, bp
);
295 error
= nfs_vinvalbuf(vp
, 0, 1);
297 * Yuck! The directory has been modified on the
298 * server. The only way to get the block is by
299 * reading from the beginning to get all the
302 * Leave the last bp intact unless there is an error.
303 * Loop back up to the while if the error is another
304 * NFSERR_BAD_COOKIE (double yuch!).
306 for (i
= 0; i
<= lbn
&& !error
; i
++) {
307 if (np
->n_direofoffset
308 && (i
* NFS_DIRBLKSIZ
) >= np
->n_direofoffset
)
310 bp
= nfs_getcacheblk(vp
, (off_t
)i
* NFS_DIRBLKSIZ
,
314 if ((bp
->b_flags
& B_CACHE
) == 0) {
315 bp
->b_cmd
= BUF_CMD_READ
;
316 bp
->b_bio2
.bio_done
= nfsiodone_sync
;
317 bp
->b_bio2
.bio_flags
|= BIO_SYNC
;
318 vfs_busy_pages(vp
, bp
);
319 error
= nfs_doio(vp
, &bp
->b_bio2
, td
);
321 * no error + B_INVAL == directory EOF,
324 if (error
== 0 && (bp
->b_flags
& B_INVAL
))
328 * An error will throw away the block and the
329 * for loop will break out. If no error and this
330 * is not the block we want, we throw away the
331 * block and go for the next one via the for loop.
333 if (error
|| i
< lbn
)
338 * The above while is repeated if we hit another cookie
339 * error. If we hit an error and it wasn't a cookie error,
347 * If not eof and read aheads are enabled, start one.
348 * (You need the current block first, so that you have the
349 * directory offset cookie of the next block.)
351 if (nmp
->nm_readahead
> 0 && nfs_asyncok(nmp
) &&
352 (bp
->b_flags
& B_INVAL
) == 0 &&
353 (np
->n_direofoffset
== 0 ||
354 loffset
+ NFS_DIRBLKSIZ
< np
->n_direofoffset
) &&
355 findblk(vp
, loffset
+ NFS_DIRBLKSIZ
, FINDBLK_TEST
) == NULL
357 rabp
= nfs_getcacheblk(vp
, loffset
+ NFS_DIRBLKSIZ
,
360 if ((rabp
->b_flags
& (B_CACHE
|B_DELWRI
)) == 0) {
361 rabp
->b_cmd
= BUF_CMD_READ
;
362 vfs_busy_pages(vp
, rabp
);
363 nfs_asyncio(vp
, &rabp
->b_bio2
);
370 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
371 * chopped for the EOF condition, we cannot tell how large
372 * NFS directories are going to be until we hit EOF. So
373 * an NFS directory buffer is *not* chopped to its EOF. Now,
374 * it just so happens that b_resid will effectively chop it
375 * to EOF. *BUT* this information is lost if the buffer goes
376 * away and is reconstituted into a B_CACHE state ( due to
377 * being VMIO ) later. So we keep track of the directory eof
378 * in np->n_direofoffset and chop it off as an extra step
381 * NOTE: boff could already be beyond EOF.
383 if ((size_t)boff
> NFS_DIRBLKSIZ
- bp
->b_resid
) {
386 n
= szmin(uio
->uio_resid
,
387 NFS_DIRBLKSIZ
- bp
->b_resid
- (size_t)boff
);
389 if (np
->n_direofoffset
&&
390 n
> (size_t)(np
->n_direofoffset
- uio
->uio_offset
)) {
391 n
= (size_t)(np
->n_direofoffset
- uio
->uio_offset
);
395 kprintf(" nfs_bioread: type %x unexpected\n",vp
->v_type
);
400 switch (vp
->v_type
) {
403 error
= uiomove(bp
->b_data
+ boff
, n
, uio
);
407 error
= uiomove(bp
->b_data
+ boff
, n
, uio
);
412 off_t old_off
= uio
->uio_offset
;
414 struct nfs_dirent
*dp
;
417 * We are casting cpos to nfs_dirent, it must be
425 cpos
= bp
->b_data
+ boff
;
426 epos
= bp
->b_data
+ boff
+ n
;
427 while (cpos
< epos
&& error
== 0 && uio
->uio_resid
> 0) {
428 dp
= (struct nfs_dirent
*)cpos
;
429 error
= nfs_check_dirent(dp
, (int)(epos
- cpos
));
432 if (vop_write_dirent(&error
, uio
, dp
->nfs_ino
,
433 dp
->nfs_type
, dp
->nfs_namlen
, dp
->nfs_name
)) {
436 cpos
+= dp
->nfs_reclen
;
440 uio
->uio_offset
= old_off
+ cpos
-
446 kprintf(" nfs_bioread: type %x unexpected\n",vp
->v_type
);
450 } while (error
== 0 && uio
->uio_resid
> 0 && n
> 0);
455 * Userland can supply any 'seek' offset when reading a NFS directory.
456 * Validate the structure so we don't panic the kernel. Note that
457 * the element name is nul terminated and the nul is not included
462 nfs_check_dirent(struct nfs_dirent
*dp
, int maxlen
)
464 int nfs_name_off
= offsetof(struct nfs_dirent
, nfs_name
[0]);
466 if (nfs_name_off
>= maxlen
)
468 if (dp
->nfs_reclen
< nfs_name_off
|| dp
->nfs_reclen
> maxlen
)
470 if (nfs_name_off
+ dp
->nfs_namlen
>= dp
->nfs_reclen
)
472 if (dp
->nfs_reclen
& 3)
478 * Vnode op for write using bio
480 * nfs_write(struct vnode *a_vp, struct uio *a_uio, int a_ioflag,
481 * struct ucred *a_cred)
484 nfs_write(struct vop_write_args
*ap
)
486 struct uio
*uio
= ap
->a_uio
;
487 struct thread
*td
= uio
->uio_td
;
488 struct vnode
*vp
= ap
->a_vp
;
489 struct nfsnode
*np
= VTONFS(vp
);
490 int ioflag
= ap
->a_ioflag
;
493 struct nfsmount
*nmp
= VFSTONFS(vp
->v_mount
);
503 if (uio
->uio_rw
!= UIO_WRITE
)
504 panic("nfs_write mode");
505 if (uio
->uio_segflg
== UIO_USERSPACE
&& uio
->uio_td
!= curthread
)
506 panic("nfs_write proc");
508 if (vp
->v_type
!= VREG
)
510 if (np
->n_flag
& NWRITEERR
) {
511 np
->n_flag
&= ~NWRITEERR
;
512 return (np
->n_error
);
514 if ((nmp
->nm_flag
& NFSMNT_NFSV3
) != 0 &&
515 (nmp
->nm_state
& NFSSTA_GOTFSINFO
) == 0)
516 (void)nfs_fsinfo(nmp
, vp
, td
);
519 * Synchronously flush pending buffers if we are in synchronous
520 * mode or if we are appending.
522 if (ioflag
& (IO_APPEND
| IO_SYNC
)) {
523 if (np
->n_flag
& NLMODIFIED
) {
525 error
= nfs_flush(vp
, MNT_WAIT
, td
, 0);
526 /* error = nfs_vinvalbuf(vp, V_SAVE, 1); */
533 * If IO_APPEND then load uio_offset. We restart here if we cannot
534 * get the append lock.
537 if (ioflag
& IO_APPEND
) {
539 error
= VOP_GETATTR(vp
, &vattr
);
542 uio
->uio_offset
= np
->n_size
;
545 if (uio
->uio_offset
< 0)
547 if ((uio
->uio_offset
+ uio
->uio_resid
) > nmp
->nm_maxfilesize
)
549 if (uio
->uio_resid
== 0)
553 * We need to obtain the rslock if we intend to modify np->n_size
554 * in order to guarentee the append point with multiple contending
555 * writers, to guarentee that no other appenders modify n_size
556 * while we are trying to obtain a truncated buffer (i.e. to avoid
557 * accidently truncating data written by another appender due to
558 * the race), and to ensure that the buffer is populated prior to
559 * our extending of the file. We hold rslock through the entire
562 * Note that we do not synchronize the case where someone truncates
563 * the file while we are appending to it because attempting to lock
564 * this case may deadlock other parts of the system unexpectedly.
566 if ((ioflag
& IO_APPEND
) ||
567 uio
->uio_offset
+ uio
->uio_resid
> np
->n_size
) {
568 switch(nfs_rslock(np
)) {
583 * Maybe this should be above the vnode op call, but so long as
584 * file servers have no limits, i don't think it matters
586 if (td
&& td
->td_proc
&& uio
->uio_offset
+ uio
->uio_resid
>
587 td
->td_proc
->p_rlimit
[RLIMIT_FSIZE
].rlim_cur
) {
588 lwpsignal(td
->td_proc
, td
->td_lwp
, SIGXFSZ
);
594 biosize
= vp
->v_mount
->mnt_stat
.f_iosize
;
597 nfsstats
.biocache_writes
++;
598 boff
= uio
->uio_offset
& (biosize
-1);
599 loffset
= uio
->uio_offset
- boff
;
600 bytes
= (int)szmin((unsigned)(biosize
- boff
), uio
->uio_resid
);
603 * Handle direct append and file extension cases, calculate
604 * unaligned buffer size. When extending B_CACHE will be
605 * set if possible. See UIO_NOCOPY note below.
607 if (uio
->uio_offset
+ bytes
> np
->n_size
) {
608 np
->n_flag
|= NLMODIFIED
;
609 trivial
= (uio
->uio_segflg
!= UIO_NOCOPY
&&
610 uio
->uio_offset
<= np
->n_size
);
611 nfs_meta_setsize(vp
, td
, uio
->uio_offset
+ bytes
,
614 bp
= nfs_getcacheblk(vp
, loffset
, biosize
, td
);
621 * Actual bytes in buffer which we care about
623 if (loffset
+ biosize
< np
->n_size
)
626 bcount
= (int)(np
->n_size
- loffset
);
629 * Avoid a read by setting B_CACHE where the data we
630 * intend to write covers the entire buffer. Note
631 * that the buffer may have been set to B_CACHE by
632 * nfs_meta_setsize() above or otherwise inherited the
633 * flag, but if B_CACHE isn't set the buffer may be
634 * uninitialized and must be zero'd to accomodate
635 * future seek+write's.
637 * See the comments in kern/vfs_bio.c's getblk() for
640 * When doing a UIO_NOCOPY write the buffer is not
641 * overwritten and we cannot just set B_CACHE unconditionally
642 * for full-block writes.
644 if (boff
== 0 && bytes
== biosize
&&
645 uio
->uio_segflg
!= UIO_NOCOPY
) {
646 bp
->b_flags
|= B_CACHE
;
647 bp
->b_flags
&= ~(B_ERROR
| B_INVAL
);
651 * b_resid may be set due to file EOF if we extended out.
652 * The NFS bio code will zero the difference anyway so
653 * just acknowledged the fact and set b_resid to 0.
655 if ((bp
->b_flags
& B_CACHE
) == 0) {
656 bp
->b_cmd
= BUF_CMD_READ
;
657 bp
->b_bio2
.bio_done
= nfsiodone_sync
;
658 bp
->b_bio2
.bio_flags
|= BIO_SYNC
;
659 vfs_busy_pages(vp
, bp
);
660 error
= nfs_doio(vp
, &bp
->b_bio2
, td
);
667 np
->n_flag
|= NLMODIFIED
;
670 * If dirtyend exceeds file size, chop it down. This should
671 * not normally occur but there is an append race where it
672 * might occur XXX, so we log it.
674 * If the chopping creates a reverse-indexed or degenerate
675 * situation with dirtyoff/end, we 0 both of them.
677 if (bp
->b_dirtyend
> bcount
) {
678 kprintf("NFS append race @%08llx:%d\n",
679 (long long)bp
->b_bio2
.bio_offset
,
680 bp
->b_dirtyend
- bcount
);
681 bp
->b_dirtyend
= bcount
;
684 if (bp
->b_dirtyoff
>= bp
->b_dirtyend
)
685 bp
->b_dirtyoff
= bp
->b_dirtyend
= 0;
688 * If the new write will leave a contiguous dirty
689 * area, just update the b_dirtyoff and b_dirtyend,
690 * otherwise force a write rpc of the old dirty area.
692 * While it is possible to merge discontiguous writes due to
693 * our having a B_CACHE buffer ( and thus valid read data
694 * for the hole), we don't because it could lead to
695 * significant cache coherency problems with multiple clients,
696 * especially if locking is implemented later on.
698 * as an optimization we could theoretically maintain
699 * a linked list of discontinuous areas, but we would still
700 * have to commit them separately so there isn't much
701 * advantage to it except perhaps a bit of asynchronization.
703 if (bp
->b_dirtyend
> 0 &&
704 (boff
> bp
->b_dirtyend
||
705 (boff
+ bytes
) < bp
->b_dirtyoff
)
707 if (bwrite(bp
) == EINTR
) {
714 error
= uiomove(bp
->b_data
+ boff
, bytes
, uio
);
717 * Since this block is being modified, it must be written
718 * again and not just committed. Since write clustering does
719 * not work for the stage 1 data write, only the stage 2
720 * commit rpc, we have to clear B_CLUSTEROK as well.
722 bp
->b_flags
&= ~(B_NEEDCOMMIT
| B_CLUSTEROK
);
730 * Only update dirtyoff/dirtyend if not a degenerate
733 * The underlying VM pages have been marked valid by
734 * virtue of acquiring the bp. Because the entire buffer
735 * is marked dirty we do not have to worry about cleaning
736 * out the related dirty bits (and wouldn't really know
737 * how to deal with byte ranges anyway)
740 if (bp
->b_dirtyend
> 0) {
741 bp
->b_dirtyoff
= imin(boff
, bp
->b_dirtyoff
);
742 bp
->b_dirtyend
= imax(boff
+ bytes
,
745 bp
->b_dirtyoff
= boff
;
746 bp
->b_dirtyend
= boff
+ bytes
;
751 * If the lease is non-cachable or IO_SYNC do bwrite().
753 * IO_INVAL appears to be unused. The idea appears to be
754 * to turn off caching in this case. Very odd. XXX
756 * If nfs_async is set bawrite() will use an unstable write
757 * (build dirty bufs on the server), so we might as well
758 * push it out with bawrite(). If nfs_async is not set we
759 * use bdwrite() to cache dirty bufs on the client.
761 if (ioflag
& IO_SYNC
) {
762 if (ioflag
& IO_INVAL
)
763 bp
->b_flags
|= B_NOCACHE
;
767 } else if (boff
+ bytes
== biosize
&& nfs_async
) {
772 } while (uio
->uio_resid
> 0 && bytes
> 0);
781 * Get an nfs cache block.
783 * Allocate a new one if the block isn't currently in the cache
784 * and return the block marked busy. If the calling process is
785 * interrupted by a signal for an interruptible mount point, return
788 * The caller must carefully deal with the possible B_INVAL state of
789 * the buffer. nfs_startio() clears B_INVAL (and nfs_asyncio() clears it
790 * indirectly), so synchronous reads can be issued without worrying about
791 * the B_INVAL state. We have to be a little more careful when dealing
792 * with writes (see comments in nfs_write()) when extending a file past
796 nfs_getcacheblk(struct vnode
*vp
, off_t loffset
, int size
, struct thread
*td
)
800 struct nfsmount
*nmp
;
805 if (nmp
->nm_flag
& NFSMNT_INT
) {
806 bp
= getblk(vp
, loffset
, size
, GETBLK_PCATCH
, 0);
808 if (nfs_sigintr(nmp
, NULL
, td
))
810 bp
= getblk(vp
, loffset
, size
, 0, 2 * hz
);
813 bp
= getblk(vp
, loffset
, size
, 0, 0);
817 * bio2, the 'device' layer. Since BIOs use 64 bit byte offsets
818 * now, no translation is necessary.
820 bp
->b_bio2
.bio_offset
= loffset
;
825 * Flush and invalidate all dirty buffers. If another process is already
826 * doing the flush, just wait for completion.
829 nfs_vinvalbuf(struct vnode
*vp
, int flags
, int intrflg
)
831 struct nfsnode
*np
= VTONFS(vp
);
832 struct nfsmount
*nmp
= VFSTONFS(vp
->v_mount
);
833 int error
= 0, slpflag
, slptimeo
;
834 thread_t td
= curthread
;
836 if (vp
->v_flag
& VRECLAIMED
)
839 if ((nmp
->nm_flag
& NFSMNT_INT
) == 0)
849 * First wait for any other process doing a flush to complete.
851 while (np
->n_flag
& NFLUSHINPROG
) {
852 np
->n_flag
|= NFLUSHWANT
;
853 error
= tsleep((caddr_t
)&np
->n_flag
, 0, "nfsvinval", slptimeo
);
854 if (error
&& intrflg
&& nfs_sigintr(nmp
, NULL
, td
))
859 * Now, flush as required.
861 np
->n_flag
|= NFLUSHINPROG
;
862 error
= vinvalbuf(vp
, flags
, slpflag
, 0);
864 if (intrflg
&& nfs_sigintr(nmp
, NULL
, td
)) {
865 np
->n_flag
&= ~NFLUSHINPROG
;
866 if (np
->n_flag
& NFLUSHWANT
) {
867 np
->n_flag
&= ~NFLUSHWANT
;
868 wakeup((caddr_t
)&np
->n_flag
);
872 error
= vinvalbuf(vp
, flags
, 0, slptimeo
);
874 np
->n_flag
&= ~(NLMODIFIED
| NFLUSHINPROG
);
875 if (np
->n_flag
& NFLUSHWANT
) {
876 np
->n_flag
&= ~NFLUSHWANT
;
877 wakeup((caddr_t
)&np
->n_flag
);
883 * Return true (non-zero) if the txthread and rxthread are operational
884 * and we do not already have too many not-yet-started BIO's built up.
887 nfs_asyncok(struct nfsmount
*nmp
)
889 return (nmp
->nm_bioqlen
< nfs_maxasyncbio
&&
890 nmp
->nm_bioqlen
< nmp
->nm_maxasync_scaled
/ NFS_ASYSCALE
&&
891 nmp
->nm_rxstate
<= NFSSVC_PENDING
&&
892 nmp
->nm_txstate
<= NFSSVC_PENDING
);
896 * The read-ahead code calls this to queue a bio to the txthread.
898 * We don't touch the bio otherwise... that is, we do not even
899 * construct or send the initial rpc. The txthread will do it
902 * NOTE! nm_bioqlen is not decremented until the request completes,
903 * so it does not reflect the number of bio's on bioq.
906 nfs_asyncio(struct vnode
*vp
, struct bio
*bio
)
908 struct buf
*bp
= bio
->bio_buf
;
909 struct nfsmount
*nmp
= VFSTONFS(vp
->v_mount
);
911 KKASSERT(vp
->v_tag
== VT_NFS
);
913 bio
->bio_driver_info
= vp
;
915 TAILQ_INSERT_TAIL(&nmp
->nm_bioq
, bio
, bio_act
);
916 atomic_add_int(&nmp
->nm_bioqlen
, 1);
918 nfssvc_iod_writer_wakeup(nmp
);
922 * nfs_dio() - Execute a BIO operation synchronously. The BIO will be
923 * completed and its error returned. The caller is responsible
924 * for brelse()ing it. ONLY USE FOR BIO_SYNC IOs! Otherwise
925 * our error probe will be against an invalid pointer.
927 * nfs_startio()- Execute a BIO operation assynchronously.
929 * NOTE: nfs_asyncio() is used to initiate an asynchronous BIO operation,
930 * which basically just queues it to the txthread. nfs_startio()
931 * actually initiates the I/O AFTER it has gotten to the txthread.
933 * NOTE: td might be NULL.
935 * NOTE: Caller has already busied the I/O.
938 nfs_startio(struct vnode
*vp
, struct bio
*bio
, struct thread
*td
)
940 struct buf
*bp
= bio
->bio_buf
;
942 struct nfsmount
*nmp
;
944 KKASSERT(vp
->v_tag
== VT_NFS
);
946 nmp
= VFSTONFS(vp
->v_mount
);
949 * clear B_ERROR and B_INVAL state prior to initiating the I/O. We
950 * do this here so we do not have to do it in all the code that
953 bp
->b_flags
&= ~(B_ERROR
| B_INVAL
);
955 KASSERT(bp
->b_cmd
!= BUF_CMD_DONE
,
956 ("nfs_doio: bp %p already marked done!", bp
));
958 if (bp
->b_cmd
== BUF_CMD_READ
) {
959 switch (vp
->v_type
) {
961 nfsstats
.read_bios
++;
962 nfs_readrpc_bio(vp
, bio
);
967 nfsstats
.readlink_bios
++;
968 nfs_readlinkrpc_bio(vp
, bio
);
970 nfs_doio(vp
, bio
, td
);
975 * NOTE: If nfs_readdirplusrpc_bio() is requested but
976 * not supported, it will chain to
977 * nfs_readdirrpc_bio().
980 nfsstats
.readdir_bios
++;
981 uiop
->uio_offset
= bio
->bio_offset
;
982 if (nmp
->nm_flag
& NFSMNT_RDIRPLUS
)
983 nfs_readdirplusrpc_bio(vp
, bio
);
985 nfs_readdirrpc_bio(vp
, bio
);
987 nfs_doio(vp
, bio
, td
);
991 kprintf("nfs_doio: type %x unexpected\n",vp
->v_type
);
992 bp
->b_flags
|= B_ERROR
;
993 bp
->b_error
= EINVAL
;
999 * If we only need to commit, try to commit. If this fails
1000 * it will chain through to the write. Basically all the logic
1001 * in nfs_doio() is replicated.
1003 KKASSERT(bp
->b_cmd
== BUF_CMD_WRITE
);
1004 if (bp
->b_flags
& B_NEEDCOMMIT
)
1005 nfs_commitrpc_bio(vp
, bio
);
1007 nfs_writerpc_bio(vp
, bio
);
1012 nfs_doio(struct vnode
*vp
, struct bio
*bio
, struct thread
*td
)
1014 struct buf
*bp
= bio
->bio_buf
;
1017 struct nfsmount
*nmp
;
1019 int iomode
, must_commit
;
1024 KKASSERT(vp
->v_tag
== VT_NFS
);
1026 nmp
= VFSTONFS(vp
->v_mount
);
1028 uiop
->uio_iov
= &io
;
1029 uiop
->uio_iovcnt
= 1;
1030 uiop
->uio_segflg
= UIO_SYSSPACE
;
1034 * clear B_ERROR and B_INVAL state prior to initiating the I/O. We
1035 * do this here so we do not have to do it in all the code that
1038 bp
->b_flags
&= ~(B_ERROR
| B_INVAL
);
1040 KASSERT(bp
->b_cmd
!= BUF_CMD_DONE
,
1041 ("nfs_doio: bp %p already marked done!", bp
));
1043 if (bp
->b_cmd
== BUF_CMD_READ
) {
1044 io
.iov_len
= uiop
->uio_resid
= (size_t)bp
->b_bcount
;
1045 io
.iov_base
= bp
->b_data
;
1046 uiop
->uio_rw
= UIO_READ
;
1048 switch (vp
->v_type
) {
1051 * When reading from a regular file zero-fill any residual.
1052 * Note that this residual has nothing to do with NFS short
1053 * reads, which nfs_readrpc_uio() will handle for us.
1055 * We have to do this because when we are write extending
1056 * a file the server may not have the same notion of
1057 * filesize as we do. Our BIOs should already be sized
1058 * (b_bcount) to account for the file EOF.
1060 nfsstats
.read_bios
++;
1061 uiop
->uio_offset
= bio
->bio_offset
;
1062 error
= nfs_readrpc_uio(vp
, uiop
);
1063 if (error
== 0 && uiop
->uio_resid
) {
1064 n
= (size_t)bp
->b_bcount
- uiop
->uio_resid
;
1065 bzero(bp
->b_data
+ n
, bp
->b_bcount
- n
);
1066 uiop
->uio_resid
= 0;
1068 if (td
&& td
->td_proc
&& (vp
->v_flag
& VTEXT
) &&
1069 np
->n_mtime
!= np
->n_vattr
.va_mtime
.tv_sec
) {
1070 uprintf("Process killed due to text file modification\n");
1071 ksignal(td
->td_proc
, SIGKILL
);
1075 uiop
->uio_offset
= 0;
1076 nfsstats
.readlink_bios
++;
1077 error
= nfs_readlinkrpc_uio(vp
, uiop
);
1080 nfsstats
.readdir_bios
++;
1081 uiop
->uio_offset
= bio
->bio_offset
;
1082 if (nmp
->nm_flag
& NFSMNT_RDIRPLUS
) {
1083 error
= nfs_readdirplusrpc_uio(vp
, uiop
);
1084 if (error
== NFSERR_NOTSUPP
)
1085 nmp
->nm_flag
&= ~NFSMNT_RDIRPLUS
;
1087 if ((nmp
->nm_flag
& NFSMNT_RDIRPLUS
) == 0)
1088 error
= nfs_readdirrpc_uio(vp
, uiop
);
1090 * end-of-directory sets B_INVAL but does not generate an
1093 if (error
== 0 && uiop
->uio_resid
== bp
->b_bcount
)
1094 bp
->b_flags
|= B_INVAL
;
1097 kprintf("nfs_doio: type %x unexpected\n",vp
->v_type
);
1101 bp
->b_flags
|= B_ERROR
;
1102 bp
->b_error
= error
;
1104 bp
->b_resid
= uiop
->uio_resid
;
1107 * If we only need to commit, try to commit.
1109 * NOTE: The I/O has already been staged for the write and
1110 * its pages busied, so b_dirtyoff/end is valid.
1112 KKASSERT(bp
->b_cmd
== BUF_CMD_WRITE
);
1113 if (bp
->b_flags
& B_NEEDCOMMIT
) {
1117 off
= bio
->bio_offset
+ bp
->b_dirtyoff
;
1118 retv
= nfs_commitrpc_uio(vp
, off
,
1119 bp
->b_dirtyend
- bp
->b_dirtyoff
,
1122 bp
->b_dirtyoff
= bp
->b_dirtyend
= 0;
1123 bp
->b_flags
&= ~(B_NEEDCOMMIT
| B_CLUSTEROK
);
1128 if (retv
== NFSERR_STALEWRITEVERF
) {
1129 nfs_clearcommit(vp
->v_mount
);
1134 * Setup for actual write
1136 if (bio
->bio_offset
+ bp
->b_dirtyend
> np
->n_size
)
1137 bp
->b_dirtyend
= np
->n_size
- bio
->bio_offset
;
1139 if (bp
->b_dirtyend
> bp
->b_dirtyoff
) {
1140 io
.iov_len
= uiop
->uio_resid
= bp
->b_dirtyend
1142 uiop
->uio_offset
= bio
->bio_offset
+ bp
->b_dirtyoff
;
1143 io
.iov_base
= (char *)bp
->b_data
+ bp
->b_dirtyoff
;
1144 uiop
->uio_rw
= UIO_WRITE
;
1145 nfsstats
.write_bios
++;
1147 if ((bp
->b_flags
& (B_NEEDCOMMIT
| B_NOCACHE
| B_CLUSTER
)) == 0)
1148 iomode
= NFSV3WRITE_UNSTABLE
;
1150 iomode
= NFSV3WRITE_FILESYNC
;
1153 error
= nfs_writerpc_uio(vp
, uiop
, &iomode
, &must_commit
);
1156 * We no longer try to use kern/vfs_bio's cluster code to
1157 * cluster commits, so B_CLUSTEROK is no longer set with
1158 * B_NEEDCOMMIT. The problem is that a vfs_busy_pages()
1159 * may have to clear B_NEEDCOMMIT if it finds underlying
1160 * pages have been redirtied through a memory mapping
1161 * and doing this on a clustered bp will probably cause
1162 * a panic, plus the flag in the underlying NFS bufs
1163 * making up the cluster bp will not be properly cleared.
1165 if (!error
&& iomode
== NFSV3WRITE_UNSTABLE
) {
1166 bp
->b_flags
|= B_NEEDCOMMIT
;
1168 /* XXX do not enable commit clustering */
1169 if (bp
->b_dirtyoff
== 0
1170 && bp
->b_dirtyend
== bp
->b_bcount
)
1171 bp
->b_flags
|= B_CLUSTEROK
;
1174 bp
->b_flags
&= ~(B_NEEDCOMMIT
| B_CLUSTEROK
);
1178 * For an interrupted write, the buffer is still valid
1179 * and the write hasn't been pushed to the server yet,
1180 * so we can't set B_ERROR and report the interruption
1181 * by setting B_EINTR. For the async case, B_EINTR
1182 * is not relevant, so the rpc attempt is essentially
1183 * a noop. For the case of a V3 write rpc not being
1184 * committed to stable storage, the block is still
1185 * dirty and requires either a commit rpc or another
1186 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1187 * the block is reused. This is indicated by setting
1188 * the B_DELWRI and B_NEEDCOMMIT flags.
1190 * If the buffer is marked B_PAGING, it does not reside on
1191 * the vp's paging queues so we cannot call bdirty(). The
1192 * bp in this case is not an NFS cache block so we should
1196 || (!error
&& (bp
->b_flags
& B_NEEDCOMMIT
))) {
1198 bp
->b_flags
&= ~(B_INVAL
|B_NOCACHE
);
1199 if ((bp
->b_flags
& B_PAGING
) == 0)
1202 bp
->b_flags
|= B_EINTR
;
1206 bp
->b_flags
|= B_ERROR
;
1207 bp
->b_error
= np
->n_error
= error
;
1208 np
->n_flag
|= NWRITEERR
;
1210 bp
->b_dirtyoff
= bp
->b_dirtyend
= 0;
1213 nfs_clearcommit(vp
->v_mount
);
1214 bp
->b_resid
= uiop
->uio_resid
;
1221 * I/O was run synchronously, biodone() it and calculate the
1225 KKASSERT(bp
->b_cmd
== BUF_CMD_DONE
);
1226 if (bp
->b_flags
& B_EINTR
)
1228 if (bp
->b_flags
& B_ERROR
)
1229 return (bp
->b_error
? bp
->b_error
: EIO
);
1234 * Handle all truncation, write-extend, and ftruncate()-extend operations
1235 * on the NFS lcient side.
1237 * We use the new API in kern/vfs_vm.c to perform these operations in a
1238 * VM-friendly way. With this API VM pages are properly zerod and pages
1239 * still mapped into the buffer straddling EOF are not invalidated.
1242 nfs_meta_setsize(struct vnode
*vp
, struct thread
*td
, off_t nsize
, int trivial
)
1244 struct nfsnode
*np
= VTONFS(vp
);
1246 int biosize
= vp
->v_mount
->mnt_stat
.f_iosize
;
1252 if (nsize
< osize
) {
1253 error
= nvtruncbuf(vp
, nsize
, biosize
, -1);
1255 error
= nvextendbuf(vp
, osize
, nsize
,
1256 biosize
, biosize
, -1, -1,
1263 * Synchronous completion for nfs_doio. Call bpdone() with elseit=FALSE.
1264 * Caller is responsible for brelse()'ing the bp.
1267 nfsiodone_sync(struct bio
*bio
)
1270 bpdone(bio
->bio_buf
, 0);
1274 * nfs read rpc - BIO version
1277 nfs_readrpc_bio(struct vnode
*vp
, struct bio
*bio
)
1279 struct buf
*bp
= bio
->bio_buf
;
1281 struct nfsmount
*nmp
;
1282 int error
= 0, len
, tsiz
;
1283 struct nfsm_info
*info
;
1285 info
= kmalloc(sizeof(*info
), M_NFSREQ
, M_WAITOK
);
1287 info
->v3
= NFS_ISV3(vp
);
1289 nmp
= VFSTONFS(vp
->v_mount
);
1290 tsiz
= bp
->b_bcount
;
1291 KKASSERT(tsiz
<= nmp
->nm_rsize
);
1292 if (bio
->bio_offset
+ tsiz
> nmp
->nm_maxfilesize
) {
1296 nfsstats
.rpccnt
[NFSPROC_READ
]++;
1298 nfsm_reqhead(info
, vp
, NFSPROC_READ
,
1299 NFSX_FH(info
->v3
) + NFSX_UNSIGNED
* 3);
1300 ERROROUT(nfsm_fhtom(info
, vp
));
1301 tl
= nfsm_build(info
, NFSX_UNSIGNED
* 3);
1303 txdr_hyper(bio
->bio_offset
, tl
);
1304 *(tl
+ 2) = txdr_unsigned(len
);
1306 *tl
++ = txdr_unsigned(bio
->bio_offset
);
1307 *tl
++ = txdr_unsigned(len
);
1311 info
->done
= nfs_readrpc_bio_done
;
1312 nfsm_request_bio(info
, vp
, NFSPROC_READ
, NULL
,
1313 nfs_vpcred(vp
, ND_READ
));
1316 kfree(info
, M_NFSREQ
);
1317 bp
->b_error
= error
;
1318 bp
->b_flags
|= B_ERROR
;
1323 nfs_readrpc_bio_done(nfsm_info_t info
)
1325 struct nfsmount
*nmp
= VFSTONFS(info
->vp
->v_mount
);
1326 struct bio
*bio
= info
->bio
;
1327 struct buf
*bp
= bio
->bio_buf
;
1334 KKASSERT(info
->state
== NFSM_STATE_DONE
);
1337 ERROROUT(nfsm_postop_attr(info
, info
->vp
, &attrflag
,
1338 NFS_LATTR_NOSHRINK
));
1339 NULLOUT(tl
= nfsm_dissect(info
, 2 * NFSX_UNSIGNED
));
1340 eof
= fxdr_unsigned(int, *(tl
+ 1));
1342 ERROROUT(nfsm_loadattr(info
, info
->vp
, NULL
));
1345 NEGATIVEOUT(retlen
= nfsm_strsiz(info
, nmp
->nm_rsize
));
1346 ERROROUT(nfsm_mtobio(info
, bio
, retlen
));
1347 m_freem(info
->mrep
);
1351 * No error occured, if retlen is less then bcount and no EOF
1352 * and NFSv3 a zero-fill short read occured.
1354 * For NFSv2 a short-read indicates EOF.
1356 if (retlen
< bp
->b_bcount
&& info
->v3
&& eof
== 0) {
1357 bzero(bp
->b_data
+ retlen
, bp
->b_bcount
- retlen
);
1358 retlen
= bp
->b_bcount
;
1362 * If we hit an EOF we still zero-fill, but return the expected
1363 * b_resid anyway. This should normally not occur since async
1364 * BIOs are not used for read-before-write case. Races against
1365 * the server can cause it though and we don't want to leave
1366 * garbage in the buffer.
1368 if (retlen
< bp
->b_bcount
) {
1369 bzero(bp
->b_data
+ retlen
, bp
->b_bcount
- retlen
);
1372 /* bp->b_resid = bp->b_bcount - retlen; */
1374 kfree(info
, M_NFSREQ
);
1376 bp
->b_error
= error
;
1377 bp
->b_flags
|= B_ERROR
;
1383 * nfs write call - BIO version
1385 * NOTE: Caller has already busied the I/O.
1388 nfs_writerpc_bio(struct vnode
*vp
, struct bio
*bio
)
1390 struct nfsmount
*nmp
= VFSTONFS(vp
->v_mount
);
1391 struct nfsnode
*np
= VTONFS(vp
);
1392 struct buf
*bp
= bio
->bio_buf
;
1397 struct nfsm_info
*info
;
1401 * Setup for actual write. Just clean up the bio if there
1402 * is nothing to do. b_dirtyoff/end have already been staged
1403 * by the bp's pages getting busied.
1405 if (bio
->bio_offset
+ bp
->b_dirtyend
> np
->n_size
)
1406 bp
->b_dirtyend
= np
->n_size
- bio
->bio_offset
;
1408 if (bp
->b_dirtyend
<= bp
->b_dirtyoff
) {
1413 len
= bp
->b_dirtyend
- bp
->b_dirtyoff
;
1414 offset
= bio
->bio_offset
+ bp
->b_dirtyoff
;
1415 if (offset
+ len
> nmp
->nm_maxfilesize
) {
1416 bp
->b_flags
|= B_ERROR
;
1417 bp
->b_error
= EFBIG
;
1422 nfsstats
.write_bios
++;
1424 info
= kmalloc(sizeof(*info
), M_NFSREQ
, M_WAITOK
);
1426 info
->v3
= NFS_ISV3(vp
);
1427 info
->info_writerpc
.must_commit
= 0;
1428 if ((bp
->b_flags
& (B_NEEDCOMMIT
| B_NOCACHE
| B_CLUSTER
)) == 0)
1429 iomode
= NFSV3WRITE_UNSTABLE
;
1431 iomode
= NFSV3WRITE_FILESYNC
;
1433 KKASSERT(len
<= nmp
->nm_wsize
);
1435 nfsstats
.rpccnt
[NFSPROC_WRITE
]++;
1436 nfsm_reqhead(info
, vp
, NFSPROC_WRITE
,
1437 NFSX_FH(info
->v3
) + 5 * NFSX_UNSIGNED
+ nfsm_rndup(len
));
1438 ERROROUT(nfsm_fhtom(info
, vp
));
1440 tl
= nfsm_build(info
, 5 * NFSX_UNSIGNED
);
1441 txdr_hyper(offset
, tl
);
1443 *tl
++ = txdr_unsigned(len
);
1444 *tl
++ = txdr_unsigned(iomode
);
1445 *tl
= txdr_unsigned(len
);
1449 tl
= nfsm_build(info
, 4 * NFSX_UNSIGNED
);
1450 /* Set both "begin" and "current" to non-garbage. */
1451 x
= txdr_unsigned((u_int32_t
)offset
);
1452 *tl
++ = x
; /* "begin offset" */
1453 *tl
++ = x
; /* "current offset" */
1454 x
= txdr_unsigned(len
);
1455 *tl
++ = x
; /* total to this offset */
1456 *tl
= x
; /* size of this write */
1458 ERROROUT(nfsm_biotom(info
, bio
, bp
->b_dirtyoff
, len
));
1460 info
->done
= nfs_writerpc_bio_done
;
1461 nfsm_request_bio(info
, vp
, NFSPROC_WRITE
, NULL
,
1462 nfs_vpcred(vp
, ND_WRITE
));
1465 kfree(info
, M_NFSREQ
);
1466 bp
->b_error
= error
;
1467 bp
->b_flags
|= B_ERROR
;
1472 nfs_writerpc_bio_done(nfsm_info_t info
)
1474 struct nfsmount
*nmp
= VFSTONFS(info
->vp
->v_mount
);
1475 struct nfsnode
*np
= VTONFS(info
->vp
);
1476 struct bio
*bio
= info
->bio
;
1477 struct buf
*bp
= bio
->bio_buf
;
1478 int wccflag
= NFSV3_WCCRATTR
;
1479 int iomode
= NFSV3WRITE_FILESYNC
;
1483 int len
= bp
->b_resid
; /* b_resid was set to shortened length */
1488 * The write RPC returns a before and after mtime. The
1489 * nfsm_wcc_data() macro checks the before n_mtime
1490 * against the before time and stores the after time
1491 * in the nfsnode's cached vattr and n_mtime field.
1492 * The NRMODIFIED bit will be set if the before
1493 * time did not match the original mtime.
1495 wccflag
= NFSV3_WCCCHK
;
1496 ERROROUT(nfsm_wcc_data(info
, info
->vp
, &wccflag
));
1498 NULLOUT(tl
= nfsm_dissect(info
, 2 * NFSX_UNSIGNED
+ NFSX_V3WRITEVERF
));
1499 rlen
= fxdr_unsigned(int, *tl
++);
1502 m_freem(info
->mrep
);
1505 } else if (rlen
< len
) {
1508 * XXX what do we do here?
1510 backup
= len
- rlen
;
1511 uiop
->uio_iov
->iov_base
= (char *)uiop
->uio_iov
->iov_base
- backup
;
1512 uiop
->uio_iov
->iov_len
+= backup
;
1513 uiop
->uio_offset
-= backup
;
1514 uiop
->uio_resid
+= backup
;
1518 commit
= fxdr_unsigned(int, *tl
++);
1521 * Return the lowest committment level
1522 * obtained by any of the RPCs.
1524 if (iomode
== NFSV3WRITE_FILESYNC
)
1526 else if (iomode
== NFSV3WRITE_DATASYNC
&&
1527 commit
== NFSV3WRITE_UNSTABLE
)
1529 if ((nmp
->nm_state
& NFSSTA_HASWRITEVERF
) == 0){
1530 bcopy(tl
, (caddr_t
)nmp
->nm_verf
, NFSX_V3WRITEVERF
);
1531 nmp
->nm_state
|= NFSSTA_HASWRITEVERF
;
1532 } else if (bcmp(tl
, nmp
->nm_verf
, NFSX_V3WRITEVERF
)) {
1533 info
->info_writerpc
.must_commit
= 1;
1534 bcopy(tl
, (caddr_t
)nmp
->nm_verf
, NFSX_V3WRITEVERF
);
1538 ERROROUT(nfsm_loadattr(info
, info
->vp
, NULL
));
1540 m_freem(info
->mrep
);
1544 if (info
->vp
->v_mount
->mnt_flag
& MNT_ASYNC
)
1545 iomode
= NFSV3WRITE_FILESYNC
;
1549 * End of RPC. Now clean up the bp.
1551 * We no longer enable write clustering for commit operations,
1552 * See around line 1157 for a more detailed comment.
1554 if (!error
&& iomode
== NFSV3WRITE_UNSTABLE
) {
1555 bp
->b_flags
|= B_NEEDCOMMIT
;
1557 /* XXX do not enable commit clustering */
1558 if (bp
->b_dirtyoff
== 0 && bp
->b_dirtyend
== bp
->b_bcount
)
1559 bp
->b_flags
|= B_CLUSTEROK
;
1562 bp
->b_flags
&= ~(B_NEEDCOMMIT
| B_CLUSTEROK
);
1566 * For an interrupted write, the buffer is still valid
1567 * and the write hasn't been pushed to the server yet,
1568 * so we can't set B_ERROR and report the interruption
1569 * by setting B_EINTR. For the async case, B_EINTR
1570 * is not relevant, so the rpc attempt is essentially
1571 * a noop. For the case of a V3 write rpc not being
1572 * committed to stable storage, the block is still
1573 * dirty and requires either a commit rpc or another
1574 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1575 * the block is reused. This is indicated by setting
1576 * the B_DELWRI and B_NEEDCOMMIT flags.
1578 * If the buffer is marked B_PAGING, it does not reside on
1579 * the vp's paging queues so we cannot call bdirty(). The
1580 * bp in this case is not an NFS cache block so we should
1583 if (error
== EINTR
|| (!error
&& (bp
->b_flags
& B_NEEDCOMMIT
))) {
1585 bp
->b_flags
&= ~(B_INVAL
|B_NOCACHE
);
1586 if ((bp
->b_flags
& B_PAGING
) == 0)
1589 bp
->b_flags
|= B_EINTR
;
1593 bp
->b_flags
|= B_ERROR
;
1594 bp
->b_error
= np
->n_error
= error
;
1595 np
->n_flag
|= NWRITEERR
;
1597 bp
->b_dirtyoff
= bp
->b_dirtyend
= 0;
1599 if (info
->info_writerpc
.must_commit
)
1600 nfs_clearcommit(info
->vp
->v_mount
);
1601 kfree(info
, M_NFSREQ
);
1603 bp
->b_flags
|= B_ERROR
;
1604 bp
->b_error
= error
;
1610 * Nfs Version 3 commit rpc - BIO version
1612 * This function issues the commit rpc and will chain to a write
1616 nfs_commitrpc_bio(struct vnode
*vp
, struct bio
*bio
)
1618 struct nfsmount
*nmp
= VFSTONFS(vp
->v_mount
);
1619 struct buf
*bp
= bio
->bio_buf
;
1620 struct nfsm_info
*info
;
1624 if ((nmp
->nm_state
& NFSSTA_HASWRITEVERF
) == 0) {
1625 bp
->b_dirtyoff
= bp
->b_dirtyend
= 0;
1626 bp
->b_flags
&= ~(B_NEEDCOMMIT
| B_CLUSTEROK
);
1632 info
= kmalloc(sizeof(*info
), M_NFSREQ
, M_WAITOK
);
1636 nfsstats
.rpccnt
[NFSPROC_COMMIT
]++;
1637 nfsm_reqhead(info
, vp
, NFSPROC_COMMIT
, NFSX_FH(1));
1638 ERROROUT(nfsm_fhtom(info
, vp
));
1639 tl
= nfsm_build(info
, 3 * NFSX_UNSIGNED
);
1640 txdr_hyper(bio
->bio_offset
+ bp
->b_dirtyoff
, tl
);
1642 *tl
= txdr_unsigned(bp
->b_dirtyend
- bp
->b_dirtyoff
);
1644 info
->done
= nfs_commitrpc_bio_done
;
1645 nfsm_request_bio(info
, vp
, NFSPROC_COMMIT
, NULL
,
1646 nfs_vpcred(vp
, ND_WRITE
));
1650 * Chain to write RPC on (early) error
1652 kfree(info
, M_NFSREQ
);
1653 nfs_writerpc_bio(vp
, bio
);
1657 nfs_commitrpc_bio_done(nfsm_info_t info
)
1659 struct nfsmount
*nmp
= VFSTONFS(info
->vp
->v_mount
);
1660 struct bio
*bio
= info
->bio
;
1661 struct buf
*bp
= bio
->bio_buf
;
1663 int wccflag
= NFSV3_WCCRATTR
;
1666 ERROROUT(nfsm_wcc_data(info
, info
->vp
, &wccflag
));
1668 NULLOUT(tl
= nfsm_dissect(info
, NFSX_V3WRITEVERF
));
1669 if (bcmp(nmp
->nm_verf
, tl
, NFSX_V3WRITEVERF
)) {
1670 bcopy(tl
, nmp
->nm_verf
, NFSX_V3WRITEVERF
);
1671 error
= NFSERR_STALEWRITEVERF
;
1674 m_freem(info
->mrep
);
1678 * On completion we must chain to a write bio if an
1682 kfree(info
, M_NFSREQ
);
1684 bp
->b_dirtyoff
= bp
->b_dirtyend
= 0;
1685 bp
->b_flags
&= ~(B_NEEDCOMMIT
| B_CLUSTEROK
);
1689 nfs_writerpc_bio(info
->vp
, bio
);