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>
54 #include <vm/vm_extern.h>
55 #include <vm/vm_page.h>
56 #include <vm/vm_object.h>
57 #include <vm/vm_pager.h>
58 #include <vm/vnode_pager.h>
61 #include <sys/thread2.h>
62 #include <vm/vm_page2.h>
70 #include "nfsm_subs.h"
73 static struct buf *nfs_getcacheblk(struct vnode *vp, off_t loffset,
74 int size, struct thread *td);
75 static int nfs_check_dirent(struct nfs_dirent *dp, int maxlen);
76 static void nfsiodone_sync(struct bio *bio);
77 static void nfs_readrpc_bio_done(nfsm_info_t info);
78 static void nfs_writerpc_bio_done(nfsm_info_t info);
79 static void nfs_commitrpc_bio_done(nfsm_info_t info);
82 * Vnode op for read using bio
85 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag)
87 struct nfsnode *np = VTONFS(vp);
89 struct buf *bp, *rabp;
92 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
102 if (uio->uio_rw != UIO_READ)
103 panic("nfs_read mode");
105 if (uio->uio_resid == 0)
107 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
111 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
112 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
113 (void)nfs_fsinfo(nmp, vp, td);
114 if (vp->v_type != VDIR &&
115 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
117 biosize = vp->v_mount->mnt_stat.f_iosize;
118 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
121 * For nfs, cache consistency can only be maintained approximately.
122 * Although RFC1094 does not specify the criteria, the following is
123 * believed to be compatible with the reference port.
125 * NFS: If local changes have been made and this is a
126 * directory, the directory must be invalidated and
127 * the attribute cache must be cleared.
129 * GETATTR is called to synchronize the file size.
131 * If remote changes are detected local data is flushed
132 * and the cache is invalidated.
134 * NOTE: In the normal case the attribute cache is not
135 * cleared which means GETATTR may use cached data and
136 * not immediately detect changes made on the server.
138 if ((np->n_flag & NLMODIFIED) && vp->v_type == VDIR) {
140 error = nfs_vinvalbuf(vp, V_SAVE, 1);
145 error = VOP_GETATTR(vp, &vattr);
150 * This can deadlock getpages/putpages for regular
151 * files. Only do it for directories.
153 if (np->n_flag & NRMODIFIED) {
154 if (vp->v_type == VDIR) {
156 error = nfs_vinvalbuf(vp, V_SAVE, 1);
159 np->n_flag &= ~NRMODIFIED;
164 * Loop until uio exhausted or we hit EOF
169 switch (vp->v_type) {
171 nfsstats.biocache_reads++;
172 lbn = uio->uio_offset / biosize;
173 boff = uio->uio_offset & (biosize - 1);
174 loffset = (off_t)lbn * biosize;
177 * Start the read ahead(s), as required.
179 if (nmp->nm_readahead > 0 && nfs_asyncok(nmp)) {
180 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
181 (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) {
182 rabn = lbn + 1 + nra;
183 raoffset = (off_t)rabn * biosize;
184 if (findblk(vp, raoffset, FINDBLK_TEST) == NULL) {
185 rabp = nfs_getcacheblk(vp, raoffset, biosize, td);
188 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
189 rabp->b_cmd = BUF_CMD_READ;
190 vfs_busy_pages(vp, rabp);
191 nfs_asyncio(vp, &rabp->b_bio2);
200 * Obtain the buffer cache block. Figure out the buffer size
201 * when we are at EOF. If we are modifying the size of the
202 * buffer based on an EOF condition we need to hold
203 * nfs_rslock() through obtaining the buffer to prevent
204 * a potential writer-appender from messing with n_size.
205 * Otherwise we may accidently truncate the buffer and
208 * Note that bcount is *not* DEV_BSIZE aligned.
210 if (loffset + boff >= np->n_size) {
214 bp = nfs_getcacheblk(vp, loffset, biosize, td);
220 * If B_CACHE is not set, we must issue the read. If this
221 * fails, we return an error.
223 if ((bp->b_flags & B_CACHE) == 0) {
224 bp->b_cmd = BUF_CMD_READ;
225 bp->b_bio2.bio_done = nfsiodone_sync;
226 bp->b_bio2.bio_flags |= BIO_SYNC;
227 vfs_busy_pages(vp, bp);
228 error = nfs_doio(vp, &bp->b_bio2, td);
236 * on is the offset into the current bp. Figure out how many
237 * bytes we can copy out of the bp. Note that bcount is
238 * NOT DEV_BSIZE aligned.
240 * Then figure out how many bytes we can copy into the uio.
243 if (n > uio->uio_resid)
245 if (loffset + boff + n > np->n_size)
246 n = np->n_size - loffset - boff;
249 biosize = min(NFS_MAXPATHLEN, np->n_size);
250 nfsstats.biocache_readlinks++;
251 bp = nfs_getcacheblk(vp, (off_t)0, biosize, td);
254 if ((bp->b_flags & B_CACHE) == 0) {
255 bp->b_cmd = BUF_CMD_READ;
256 bp->b_bio2.bio_done = nfsiodone_sync;
257 bp->b_bio2.bio_flags |= BIO_SYNC;
258 vfs_busy_pages(vp, bp);
259 error = nfs_doio(vp, &bp->b_bio2, td);
261 bp->b_flags |= B_ERROR | B_INVAL;
266 n = szmin(uio->uio_resid, (size_t)bp->b_bcount - bp->b_resid);
270 nfsstats.biocache_readdirs++;
271 if (np->n_direofoffset &&
272 uio->uio_offset >= np->n_direofoffset
276 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
277 boff = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
278 loffset = uio->uio_offset - boff;
279 bp = nfs_getcacheblk(vp, loffset, NFS_DIRBLKSIZ, td);
283 if ((bp->b_flags & B_CACHE) == 0) {
284 bp->b_cmd = BUF_CMD_READ;
285 bp->b_bio2.bio_done = nfsiodone_sync;
286 bp->b_bio2.bio_flags |= BIO_SYNC;
287 vfs_busy_pages(vp, bp);
288 error = nfs_doio(vp, &bp->b_bio2, td);
291 while (error == NFSERR_BAD_COOKIE) {
292 kprintf("got bad cookie vp %p bp %p\n", vp, bp);
294 error = nfs_vinvalbuf(vp, 0, 1);
296 * Yuck! The directory has been modified on the
297 * server. The only way to get the block is by
298 * reading from the beginning to get all the
301 * Leave the last bp intact unless there is an error.
302 * Loop back up to the while if the error is another
303 * NFSERR_BAD_COOKIE (double yuch!).
305 for (i = 0; i <= lbn && !error; i++) {
306 if (np->n_direofoffset
307 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
309 bp = nfs_getcacheblk(vp, (off_t)i * NFS_DIRBLKSIZ,
313 if ((bp->b_flags & B_CACHE) == 0) {
314 bp->b_cmd = BUF_CMD_READ;
315 bp->b_bio2.bio_done = nfsiodone_sync;
316 bp->b_bio2.bio_flags |= BIO_SYNC;
317 vfs_busy_pages(vp, bp);
318 error = nfs_doio(vp, &bp->b_bio2, td);
320 * no error + B_INVAL == directory EOF,
323 if (error == 0 && (bp->b_flags & B_INVAL))
327 * An error will throw away the block and the
328 * for loop will break out. If no error and this
329 * is not the block we want, we throw away the
330 * block and go for the next one via the for loop.
332 if (error || i < lbn)
337 * The above while is repeated if we hit another cookie
338 * error. If we hit an error and it wasn't a cookie error,
346 * If not eof and read aheads are enabled, start one.
347 * (You need the current block first, so that you have the
348 * directory offset cookie of the next block.)
350 if (nmp->nm_readahead > 0 && nfs_asyncok(nmp) &&
351 (bp->b_flags & B_INVAL) == 0 &&
352 (np->n_direofoffset == 0 ||
353 loffset + NFS_DIRBLKSIZ < np->n_direofoffset) &&
354 findblk(vp, loffset + NFS_DIRBLKSIZ, FINDBLK_TEST) == NULL
356 rabp = nfs_getcacheblk(vp, loffset + NFS_DIRBLKSIZ,
359 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
360 rabp->b_cmd = BUF_CMD_READ;
361 vfs_busy_pages(vp, rabp);
362 nfs_asyncio(vp, &rabp->b_bio2);
369 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
370 * chopped for the EOF condition, we cannot tell how large
371 * NFS directories are going to be until we hit EOF. So
372 * an NFS directory buffer is *not* chopped to its EOF. Now,
373 * it just so happens that b_resid will effectively chop it
374 * to EOF. *BUT* this information is lost if the buffer goes
375 * away and is reconstituted into a B_CACHE state ( due to
376 * being VMIO ) later. So we keep track of the directory eof
377 * in np->n_direofoffset and chop it off as an extra step
380 * NOTE: boff could already be beyond EOF.
382 if ((size_t)boff > NFS_DIRBLKSIZ - bp->b_resid) {
385 n = szmin(uio->uio_resid,
386 NFS_DIRBLKSIZ - bp->b_resid - (size_t)boff);
388 if (np->n_direofoffset &&
389 n > (size_t)(np->n_direofoffset - uio->uio_offset)) {
390 n = (size_t)(np->n_direofoffset - uio->uio_offset);
394 kprintf(" nfs_bioread: type %x unexpected\n",vp->v_type);
399 switch (vp->v_type) {
402 error = uiomove(bp->b_data + boff, n, uio);
406 error = uiomove(bp->b_data + boff, n, uio);
411 off_t old_off = uio->uio_offset;
413 struct nfs_dirent *dp;
416 * We are casting cpos to nfs_dirent, it must be
424 cpos = bp->b_data + boff;
425 epos = bp->b_data + boff + n;
426 while (cpos < epos && error == 0 && uio->uio_resid > 0) {
427 dp = (struct nfs_dirent *)cpos;
428 error = nfs_check_dirent(dp, (int)(epos - cpos));
431 if (vop_write_dirent(&error, uio, dp->nfs_ino,
432 dp->nfs_type, dp->nfs_namlen, dp->nfs_name)) {
435 cpos += dp->nfs_reclen;
439 uio->uio_offset = old_off + cpos -
445 kprintf(" nfs_bioread: type %x unexpected\n",vp->v_type);
449 } while (error == 0 && uio->uio_resid > 0 && n > 0);
454 * Userland can supply any 'seek' offset when reading a NFS directory.
455 * Validate the structure so we don't panic the kernel. Note that
456 * the element name is nul terminated and the nul is not included
461 nfs_check_dirent(struct nfs_dirent *dp, int maxlen)
463 int nfs_name_off = offsetof(struct nfs_dirent, nfs_name[0]);
465 if (nfs_name_off >= maxlen)
467 if (dp->nfs_reclen < nfs_name_off || dp->nfs_reclen > maxlen)
469 if (nfs_name_off + dp->nfs_namlen >= dp->nfs_reclen)
471 if (dp->nfs_reclen & 3)
477 * Vnode op for write using bio
479 * nfs_write(struct vnode *a_vp, struct uio *a_uio, int a_ioflag,
480 * struct ucred *a_cred)
483 nfs_write(struct vop_write_args *ap)
485 struct uio *uio = ap->a_uio;
486 struct thread *td = uio->uio_td;
487 struct vnode *vp = ap->a_vp;
488 struct nfsnode *np = VTONFS(vp);
489 int ioflag = ap->a_ioflag;
492 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
502 if (uio->uio_rw != UIO_WRITE)
503 panic("nfs_write mode");
504 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread)
505 panic("nfs_write proc");
507 if (vp->v_type != VREG)
510 lwkt_gettoken(&nmp->nm_token);
512 if (np->n_flag & NWRITEERR) {
513 np->n_flag &= ~NWRITEERR;
514 lwkt_reltoken(&nmp->nm_token);
515 return (np->n_error);
517 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
518 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
519 (void)nfs_fsinfo(nmp, vp, td);
523 * Synchronously flush pending buffers if we are in synchronous
524 * mode or if we are appending.
526 if (ioflag & (IO_APPEND | IO_SYNC)) {
527 if (np->n_flag & NLMODIFIED) {
529 error = nfs_flush(vp, MNT_WAIT, td, 0);
530 /* error = nfs_vinvalbuf(vp, V_SAVE, 1); */
537 * If IO_APPEND then load uio_offset. We restart here if we cannot
538 * get the append lock.
541 if (ioflag & IO_APPEND) {
543 error = VOP_GETATTR(vp, &vattr);
546 uio->uio_offset = np->n_size;
549 if (uio->uio_offset < 0) {
553 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) {
557 if (uio->uio_resid == 0) {
563 * We need to obtain the rslock if we intend to modify np->n_size
564 * in order to guarentee the append point with multiple contending
565 * writers, to guarentee that no other appenders modify n_size
566 * while we are trying to obtain a truncated buffer (i.e. to avoid
567 * accidently truncating data written by another appender due to
568 * the race), and to ensure that the buffer is populated prior to
569 * our extending of the file. We hold rslock through the entire
572 * Note that we do not synchronize the case where someone truncates
573 * the file while we are appending to it because attempting to lock
574 * this case may deadlock other parts of the system unexpectedly.
576 if ((ioflag & IO_APPEND) ||
577 uio->uio_offset + uio->uio_resid > np->n_size) {
578 switch(nfs_rslock(np)) {
594 * Maybe this should be above the vnode op call, but so long as
595 * file servers have no limits, i don't think it matters
597 if (td && td->td_proc && uio->uio_offset + uio->uio_resid >
598 td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
599 lwpsignal(td->td_proc, td->td_lwp, SIGXFSZ);
606 biosize = vp->v_mount->mnt_stat.f_iosize;
609 nfsstats.biocache_writes++;
610 boff = uio->uio_offset & (biosize-1);
611 loffset = uio->uio_offset - boff;
612 bytes = (int)szmin((unsigned)(biosize - boff), uio->uio_resid);
615 * Handle direct append and file extension cases, calculate
616 * unaligned buffer size. When extending B_CACHE will be
617 * set if possible. See UIO_NOCOPY note below.
619 if (uio->uio_offset + bytes > np->n_size) {
620 np->n_flag |= NLMODIFIED;
621 trivial = (uio->uio_segflg != UIO_NOCOPY &&
622 uio->uio_offset <= np->n_size);
623 nfs_meta_setsize(vp, td, uio->uio_offset + bytes,
626 bp = nfs_getcacheblk(vp, loffset, biosize, td);
633 * Actual bytes in buffer which we care about
635 if (loffset + biosize < np->n_size)
638 bcount = (int)(np->n_size - loffset);
641 * Avoid a read by setting B_CACHE where the data we
642 * intend to write covers the entire buffer. Note
643 * that the buffer may have been set to B_CACHE by
644 * nfs_meta_setsize() above or otherwise inherited the
645 * flag, but if B_CACHE isn't set the buffer may be
646 * uninitialized and must be zero'd to accomodate
647 * future seek+write's.
649 * See the comments in kern/vfs_bio.c's getblk() for
652 * When doing a UIO_NOCOPY write the buffer is not
653 * overwritten and we cannot just set B_CACHE unconditionally
654 * for full-block writes.
656 if (boff == 0 && bytes == biosize &&
657 uio->uio_segflg != UIO_NOCOPY) {
658 bp->b_flags |= B_CACHE;
659 bp->b_flags &= ~(B_ERROR | B_INVAL);
663 * b_resid may be set due to file EOF if we extended out.
664 * The NFS bio code will zero the difference anyway so
665 * just acknowledged the fact and set b_resid to 0.
667 if ((bp->b_flags & B_CACHE) == 0) {
668 bp->b_cmd = BUF_CMD_READ;
669 bp->b_bio2.bio_done = nfsiodone_sync;
670 bp->b_bio2.bio_flags |= BIO_SYNC;
671 vfs_busy_pages(vp, bp);
672 error = nfs_doio(vp, &bp->b_bio2, td);
679 np->n_flag |= NLMODIFIED;
682 * If dirtyend exceeds file size, chop it down. This should
683 * not normally occur but there is an append race where it
684 * might occur XXX, so we log it.
686 * If the chopping creates a reverse-indexed or degenerate
687 * situation with dirtyoff/end, we 0 both of them.
689 if (bp->b_dirtyend > bcount) {
690 kprintf("NFS append race @%08llx:%d\n",
691 (long long)bp->b_bio2.bio_offset,
692 bp->b_dirtyend - bcount);
693 bp->b_dirtyend = bcount;
696 if (bp->b_dirtyoff >= bp->b_dirtyend)
697 bp->b_dirtyoff = bp->b_dirtyend = 0;
700 * If the new write will leave a contiguous dirty
701 * area, just update the b_dirtyoff and b_dirtyend,
702 * otherwise force a write rpc of the old dirty area.
704 * While it is possible to merge discontiguous writes due to
705 * our having a B_CACHE buffer ( and thus valid read data
706 * for the hole), we don't because it could lead to
707 * significant cache coherency problems with multiple clients,
708 * especially if locking is implemented later on.
710 * as an optimization we could theoretically maintain
711 * a linked list of discontinuous areas, but we would still
712 * have to commit them separately so there isn't much
713 * advantage to it except perhaps a bit of asynchronization.
715 if (bp->b_dirtyend > 0 &&
716 (boff > bp->b_dirtyend ||
717 (boff + bytes) < bp->b_dirtyoff)
719 if (bwrite(bp) == EINTR) {
726 error = uiomove(bp->b_data + boff, bytes, uio);
729 * Since this block is being modified, it must be written
730 * again and not just committed. Since write clustering does
731 * not work for the stage 1 data write, only the stage 2
732 * commit rpc, we have to clear B_CLUSTEROK as well.
734 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
742 * Only update dirtyoff/dirtyend if not a degenerate
745 * The underlying VM pages have been marked valid by
746 * virtue of acquiring the bp. Because the entire buffer
747 * is marked dirty we do not have to worry about cleaning
748 * out the related dirty bits (and wouldn't really know
749 * how to deal with byte ranges anyway)
752 if (bp->b_dirtyend > 0) {
753 bp->b_dirtyoff = imin(boff, bp->b_dirtyoff);
754 bp->b_dirtyend = imax(boff + bytes,
757 bp->b_dirtyoff = boff;
758 bp->b_dirtyend = boff + bytes;
763 * If the lease is non-cachable or IO_SYNC do bwrite().
765 * IO_INVAL appears to be unused. The idea appears to be
766 * to turn off caching in this case. Very odd. XXX
768 * If nfs_async is set bawrite() will use an unstable write
769 * (build dirty bufs on the server), so we might as well
770 * push it out with bawrite(). If nfs_async is not set we
771 * use bdwrite() to cache dirty bufs on the client.
773 if (ioflag & IO_SYNC) {
774 if (ioflag & IO_INVAL)
775 bp->b_flags |= B_NOCACHE;
779 } else if (boff + bytes == biosize && nfs_async) {
784 } while (uio->uio_resid > 0 && bytes > 0);
790 lwkt_reltoken(&nmp->nm_token);
795 * Get an nfs cache block.
797 * Allocate a new one if the block isn't currently in the cache
798 * and return the block marked busy. If the calling process is
799 * interrupted by a signal for an interruptible mount point, return
802 * The caller must carefully deal with the possible B_INVAL state of
803 * the buffer. nfs_startio() clears B_INVAL (and nfs_asyncio() clears it
804 * indirectly), so synchronous reads can be issued without worrying about
805 * the B_INVAL state. We have to be a little more careful when dealing
806 * with writes (see comments in nfs_write()) when extending a file past
810 nfs_getcacheblk(struct vnode *vp, off_t loffset, int size, struct thread *td)
814 struct nfsmount *nmp;
819 if (nmp->nm_flag & NFSMNT_INT) {
820 bp = getblk(vp, loffset, size, GETBLK_PCATCH, 0);
822 if (nfs_sigintr(nmp, NULL, td))
824 bp = getblk(vp, loffset, size, 0, 2 * hz);
827 bp = getblk(vp, loffset, size, 0, 0);
831 * bio2, the 'device' layer. Since BIOs use 64 bit byte offsets
832 * now, no translation is necessary.
834 bp->b_bio2.bio_offset = loffset;
839 * Flush and invalidate all dirty buffers. If another process is already
840 * doing the flush, just wait for completion.
843 nfs_vinvalbuf(struct vnode *vp, int flags, int intrflg)
845 struct nfsnode *np = VTONFS(vp);
846 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
847 int error = 0, slpflag, slptimeo;
848 thread_t td = curthread;
850 if (vp->v_flag & VRECLAIMED)
853 if ((nmp->nm_flag & NFSMNT_INT) == 0)
863 * First wait for any other process doing a flush to complete.
865 while (np->n_flag & NFLUSHINPROG) {
866 np->n_flag |= NFLUSHWANT;
867 error = tsleep((caddr_t)&np->n_flag, 0, "nfsvinval", slptimeo);
868 if (error && intrflg && nfs_sigintr(nmp, NULL, td))
873 * Now, flush as required.
875 np->n_flag |= NFLUSHINPROG;
876 error = vinvalbuf(vp, flags, slpflag, 0);
878 if (intrflg && nfs_sigintr(nmp, NULL, td)) {
879 np->n_flag &= ~NFLUSHINPROG;
880 if (np->n_flag & NFLUSHWANT) {
881 np->n_flag &= ~NFLUSHWANT;
882 wakeup((caddr_t)&np->n_flag);
886 error = vinvalbuf(vp, flags, 0, slptimeo);
888 np->n_flag &= ~(NLMODIFIED | NFLUSHINPROG);
889 if (np->n_flag & NFLUSHWANT) {
890 np->n_flag &= ~NFLUSHWANT;
891 wakeup((caddr_t)&np->n_flag);
897 * Return true (non-zero) if the txthread and rxthread are operational
898 * and we do not already have too many not-yet-started BIO's built up.
901 nfs_asyncok(struct nfsmount *nmp)
903 return (nmp->nm_bioqlen < nfs_maxasyncbio &&
904 nmp->nm_bioqlen < nmp->nm_maxasync_scaled / NFS_ASYSCALE &&
905 nmp->nm_rxstate <= NFSSVC_PENDING &&
906 nmp->nm_txstate <= NFSSVC_PENDING);
910 * The read-ahead code calls this to queue a bio to the txthread.
912 * We don't touch the bio otherwise... that is, we do not even
913 * construct or send the initial rpc. The txthread will do it
916 * NOTE! nm_bioqlen is not decremented until the request completes,
917 * so it does not reflect the number of bio's on bioq.
920 nfs_asyncio(struct vnode *vp, struct bio *bio)
922 struct buf *bp = bio->bio_buf;
923 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
925 KKASSERT(vp->v_tag == VT_NFS);
929 * Shortcut swap cache (not done automatically because we are not
932 if (vn_cache_strategy(vp, bio))
935 bio->bio_driver_info = vp;
937 TAILQ_INSERT_TAIL(&nmp->nm_bioq, bio, bio_act);
938 atomic_add_int(&nmp->nm_bioqlen, 1);
940 nfssvc_iod_writer_wakeup(nmp);
944 * nfs_dio() - Execute a BIO operation synchronously. The BIO will be
945 * completed and its error returned. The caller is responsible
946 * for brelse()ing it. ONLY USE FOR BIO_SYNC IOs! Otherwise
947 * our error probe will be against an invalid pointer.
949 * nfs_startio()- Execute a BIO operation assynchronously.
951 * NOTE: nfs_asyncio() is used to initiate an asynchronous BIO operation,
952 * which basically just queues it to the txthread. nfs_startio()
953 * actually initiates the I/O AFTER it has gotten to the txthread.
955 * NOTE: td might be NULL.
957 * NOTE: Caller has already busied the I/O.
960 nfs_startio(struct vnode *vp, struct bio *bio, struct thread *td)
962 struct buf *bp = bio->bio_buf;
964 struct nfsmount *nmp;
966 KKASSERT(vp->v_tag == VT_NFS);
968 nmp = VFSTONFS(vp->v_mount);
971 * clear B_ERROR and B_INVAL state prior to initiating the I/O. We
972 * do this here so we do not have to do it in all the code that
975 bp->b_flags &= ~(B_ERROR | B_INVAL);
977 KASSERT(bp->b_cmd != BUF_CMD_DONE,
978 ("nfs_doio: bp %p already marked done!", bp));
980 if (bp->b_cmd == BUF_CMD_READ) {
981 switch (vp->v_type) {
983 nfsstats.read_bios++;
984 nfs_readrpc_bio(vp, bio);
989 nfsstats.readlink_bios++;
990 nfs_readlinkrpc_bio(vp, bio);
992 nfs_doio(vp, bio, td);
997 * NOTE: If nfs_readdirplusrpc_bio() is requested but
998 * not supported, it will chain to
999 * nfs_readdirrpc_bio().
1002 nfsstats.readdir_bios++;
1003 uiop->uio_offset = bio->bio_offset;
1004 if (nmp->nm_flag & NFSMNT_RDIRPLUS)
1005 nfs_readdirplusrpc_bio(vp, bio);
1007 nfs_readdirrpc_bio(vp, bio);
1009 nfs_doio(vp, bio, td);
1013 kprintf("nfs_doio: type %x unexpected\n",vp->v_type);
1014 bp->b_flags |= B_ERROR;
1015 bp->b_error = EINVAL;
1021 * If we only need to commit, try to commit. If this fails
1022 * it will chain through to the write. Basically all the logic
1023 * in nfs_doio() is replicated.
1025 KKASSERT(bp->b_cmd == BUF_CMD_WRITE);
1026 if (bp->b_flags & B_NEEDCOMMIT)
1027 nfs_commitrpc_bio(vp, bio);
1029 nfs_writerpc_bio(vp, bio);
1034 nfs_doio(struct vnode *vp, struct bio *bio, struct thread *td)
1036 struct buf *bp = bio->bio_buf;
1039 struct nfsmount *nmp;
1041 int iomode, must_commit;
1048 * Shortcut swap cache (not done automatically because we are not
1051 * XXX The biowait is a hack until we can figure out how to stop a
1052 * biodone chain when a middle element is BIO_SYNC. BIO_SYNC is
1053 * set so the bp shouldn't get ripped out from under us. The only
1054 * use-cases are fully synchronous I/O cases.
1056 * XXX This is having problems, give up for now.
1058 if (vn_cache_strategy(vp, bio)) {
1060 error = biowait(&bio->bio_buf->b_bio1, "nfsrsw");
1065 KKASSERT(vp->v_tag == VT_NFS);
1067 nmp = VFSTONFS(vp->v_mount);
1069 uiop->uio_iov = &io;
1070 uiop->uio_iovcnt = 1;
1071 uiop->uio_segflg = UIO_SYSSPACE;
1075 * clear B_ERROR and B_INVAL state prior to initiating the I/O. We
1076 * do this here so we do not have to do it in all the code that
1079 bp->b_flags &= ~(B_ERROR | B_INVAL);
1081 KASSERT(bp->b_cmd != BUF_CMD_DONE,
1082 ("nfs_doio: bp %p already marked done!", bp));
1084 if (bp->b_cmd == BUF_CMD_READ) {
1085 io.iov_len = uiop->uio_resid = (size_t)bp->b_bcount;
1086 io.iov_base = bp->b_data;
1087 uiop->uio_rw = UIO_READ;
1089 switch (vp->v_type) {
1092 * When reading from a regular file zero-fill any residual.
1093 * Note that this residual has nothing to do with NFS short
1094 * reads, which nfs_readrpc_uio() will handle for us.
1096 * We have to do this because when we are write extending
1097 * a file the server may not have the same notion of
1098 * filesize as we do. Our BIOs should already be sized
1099 * (b_bcount) to account for the file EOF.
1101 nfsstats.read_bios++;
1102 uiop->uio_offset = bio->bio_offset;
1103 error = nfs_readrpc_uio(vp, uiop);
1104 if (error == 0 && uiop->uio_resid) {
1105 n = (size_t)bp->b_bcount - uiop->uio_resid;
1106 bzero(bp->b_data + n, bp->b_bcount - n);
1107 uiop->uio_resid = 0;
1109 if (td && td->td_proc && (vp->v_flag & VTEXT) &&
1110 np->n_mtime != np->n_vattr.va_mtime.tv_sec) {
1111 uprintf("Process killed due to text file modification\n");
1112 ksignal(td->td_proc, SIGKILL);
1116 uiop->uio_offset = 0;
1117 nfsstats.readlink_bios++;
1118 error = nfs_readlinkrpc_uio(vp, uiop);
1121 nfsstats.readdir_bios++;
1122 uiop->uio_offset = bio->bio_offset;
1123 if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
1124 error = nfs_readdirplusrpc_uio(vp, uiop);
1125 if (error == NFSERR_NOTSUPP)
1126 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1128 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1129 error = nfs_readdirrpc_uio(vp, uiop);
1131 * end-of-directory sets B_INVAL but does not generate an
1134 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1135 bp->b_flags |= B_INVAL;
1138 kprintf("nfs_doio: type %x unexpected\n",vp->v_type);
1142 bp->b_flags |= B_ERROR;
1143 bp->b_error = error;
1145 bp->b_resid = uiop->uio_resid;
1148 * If we only need to commit, try to commit.
1150 * NOTE: The I/O has already been staged for the write and
1151 * its pages busied, so b_dirtyoff/end is valid.
1153 KKASSERT(bp->b_cmd == BUF_CMD_WRITE);
1154 if (bp->b_flags & B_NEEDCOMMIT) {
1158 off = bio->bio_offset + bp->b_dirtyoff;
1159 retv = nfs_commitrpc_uio(vp, off,
1160 bp->b_dirtyend - bp->b_dirtyoff,
1163 bp->b_dirtyoff = bp->b_dirtyend = 0;
1164 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1169 if (retv == NFSERR_STALEWRITEVERF) {
1170 nfs_clearcommit(vp->v_mount);
1175 * Setup for actual write
1177 if (bio->bio_offset + bp->b_dirtyend > np->n_size)
1178 bp->b_dirtyend = np->n_size - bio->bio_offset;
1180 if (bp->b_dirtyend > bp->b_dirtyoff) {
1181 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1183 uiop->uio_offset = bio->bio_offset + bp->b_dirtyoff;
1184 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1185 uiop->uio_rw = UIO_WRITE;
1186 nfsstats.write_bios++;
1188 if ((bp->b_flags & (B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == 0)
1189 iomode = NFSV3WRITE_UNSTABLE;
1191 iomode = NFSV3WRITE_FILESYNC;
1194 error = nfs_writerpc_uio(vp, uiop, &iomode, &must_commit);
1197 * We no longer try to use kern/vfs_bio's cluster code to
1198 * cluster commits, so B_CLUSTEROK is no longer set with
1199 * B_NEEDCOMMIT. The problem is that a vfs_busy_pages()
1200 * may have to clear B_NEEDCOMMIT if it finds underlying
1201 * pages have been redirtied through a memory mapping
1202 * and doing this on a clustered bp will probably cause
1203 * a panic, plus the flag in the underlying NFS bufs
1204 * making up the cluster bp will not be properly cleared.
1206 if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1207 bp->b_flags |= B_NEEDCOMMIT;
1209 /* XXX do not enable commit clustering */
1210 if (bp->b_dirtyoff == 0
1211 && bp->b_dirtyend == bp->b_bcount)
1212 bp->b_flags |= B_CLUSTEROK;
1215 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1219 * For an interrupted write, the buffer is still valid
1220 * and the write hasn't been pushed to the server yet,
1221 * so we can't set B_ERROR and report the interruption
1222 * by setting B_EINTR. For the async case, B_EINTR
1223 * is not relevant, so the rpc attempt is essentially
1224 * a noop. For the case of a V3 write rpc not being
1225 * committed to stable storage, the block is still
1226 * dirty and requires either a commit rpc or another
1227 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1228 * the block is reused. This is indicated by setting
1229 * the B_DELWRI and B_NEEDCOMMIT flags.
1231 * If the buffer is marked B_PAGING, it does not reside on
1232 * the vp's paging queues so we cannot call bdirty(). The
1233 * bp in this case is not an NFS cache block so we should
1237 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1239 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1240 if ((bp->b_flags & B_PAGING) == 0)
1243 bp->b_flags |= B_EINTR;
1247 bp->b_flags |= B_ERROR;
1248 bp->b_error = np->n_error = error;
1249 np->n_flag |= NWRITEERR;
1251 bp->b_dirtyoff = bp->b_dirtyend = 0;
1254 nfs_clearcommit(vp->v_mount);
1255 bp->b_resid = uiop->uio_resid;
1262 * I/O was run synchronously, biodone() it and calculate the
1266 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
1267 if (bp->b_flags & B_EINTR)
1269 if (bp->b_flags & B_ERROR)
1270 return (bp->b_error ? bp->b_error : EIO);
1275 * Handle all truncation, write-extend, and ftruncate()-extend operations
1276 * on the NFS lcient side.
1278 * We use the new API in kern/vfs_vm.c to perform these operations in a
1279 * VM-friendly way. With this API VM pages are properly zerod and pages
1280 * still mapped into the buffer straddling EOF are not invalidated.
1283 nfs_meta_setsize(struct vnode *vp, struct thread *td, off_t nsize, int trivial)
1285 struct nfsnode *np = VTONFS(vp);
1287 int biosize = vp->v_mount->mnt_stat.f_iosize;
1293 if (nsize < osize) {
1294 error = nvtruncbuf(vp, nsize, biosize, -1);
1296 error = nvextendbuf(vp, osize, nsize,
1297 biosize, biosize, -1, -1,
1304 * Synchronous completion for nfs_doio. Call bpdone() with elseit=FALSE.
1305 * Caller is responsible for brelse()'ing the bp.
1308 nfsiodone_sync(struct bio *bio)
1311 bpdone(bio->bio_buf, 0);
1315 * nfs read rpc - BIO version
1318 nfs_readrpc_bio(struct vnode *vp, struct bio *bio)
1320 struct buf *bp = bio->bio_buf;
1322 struct nfsmount *nmp;
1323 int error = 0, len, tsiz;
1324 struct nfsm_info *info;
1326 info = kmalloc(sizeof(*info), M_NFSREQ, M_WAITOK);
1328 info->v3 = NFS_ISV3(vp);
1330 nmp = VFSTONFS(vp->v_mount);
1331 tsiz = bp->b_bcount;
1332 KKASSERT(tsiz <= nmp->nm_rsize);
1333 if (bio->bio_offset + tsiz > nmp->nm_maxfilesize) {
1337 nfsstats.rpccnt[NFSPROC_READ]++;
1339 nfsm_reqhead(info, vp, NFSPROC_READ,
1340 NFSX_FH(info->v3) + NFSX_UNSIGNED * 3);
1341 ERROROUT(nfsm_fhtom(info, vp));
1342 tl = nfsm_build(info, NFSX_UNSIGNED * 3);
1344 txdr_hyper(bio->bio_offset, tl);
1345 *(tl + 2) = txdr_unsigned(len);
1347 *tl++ = txdr_unsigned(bio->bio_offset);
1348 *tl++ = txdr_unsigned(len);
1352 info->done = nfs_readrpc_bio_done;
1353 nfsm_request_bio(info, vp, NFSPROC_READ, NULL,
1354 nfs_vpcred(vp, ND_READ));
1357 kfree(info, M_NFSREQ);
1358 bp->b_error = error;
1359 bp->b_flags |= B_ERROR;
1364 nfs_readrpc_bio_done(nfsm_info_t info)
1366 struct nfsmount *nmp = VFSTONFS(info->vp->v_mount);
1367 struct bio *bio = info->bio;
1368 struct buf *bp = bio->bio_buf;
1375 KKASSERT(info->state == NFSM_STATE_DONE);
1377 lwkt_gettoken(&nmp->nm_token);
1380 ERROROUT(nfsm_postop_attr(info, info->vp, &attrflag,
1381 NFS_LATTR_NOSHRINK));
1382 NULLOUT(tl = nfsm_dissect(info, 2 * NFSX_UNSIGNED));
1383 eof = fxdr_unsigned(int, *(tl + 1));
1385 ERROROUT(nfsm_loadattr(info, info->vp, NULL));
1388 NEGATIVEOUT(retlen = nfsm_strsiz(info, nmp->nm_rsize));
1389 ERROROUT(nfsm_mtobio(info, bio, retlen));
1390 m_freem(info->mrep);
1394 * No error occured, if retlen is less then bcount and no EOF
1395 * and NFSv3 a zero-fill short read occured.
1397 * For NFSv2 a short-read indicates EOF.
1399 if (retlen < bp->b_bcount && info->v3 && eof == 0) {
1400 bzero(bp->b_data + retlen, bp->b_bcount - retlen);
1401 retlen = bp->b_bcount;
1405 * If we hit an EOF we still zero-fill, but return the expected
1406 * b_resid anyway. This should normally not occur since async
1407 * BIOs are not used for read-before-write case. Races against
1408 * the server can cause it though and we don't want to leave
1409 * garbage in the buffer.
1411 if (retlen < bp->b_bcount) {
1412 bzero(bp->b_data + retlen, bp->b_bcount - retlen);
1415 /* bp->b_resid = bp->b_bcount - retlen; */
1417 lwkt_reltoken(&nmp->nm_token);
1418 kfree(info, M_NFSREQ);
1420 bp->b_error = error;
1421 bp->b_flags |= B_ERROR;
1427 * nfs write call - BIO version
1429 * NOTE: Caller has already busied the I/O.
1432 nfs_writerpc_bio(struct vnode *vp, struct bio *bio)
1434 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1435 struct nfsnode *np = VTONFS(vp);
1436 struct buf *bp = bio->bio_buf;
1441 struct nfsm_info *info;
1445 * Setup for actual write. Just clean up the bio if there
1446 * is nothing to do. b_dirtyoff/end have already been staged
1447 * by the bp's pages getting busied.
1449 if (bio->bio_offset + bp->b_dirtyend > np->n_size)
1450 bp->b_dirtyend = np->n_size - bio->bio_offset;
1452 if (bp->b_dirtyend <= bp->b_dirtyoff) {
1457 len = bp->b_dirtyend - bp->b_dirtyoff;
1458 offset = bio->bio_offset + bp->b_dirtyoff;
1459 if (offset + len > nmp->nm_maxfilesize) {
1460 bp->b_flags |= B_ERROR;
1461 bp->b_error = EFBIG;
1466 nfsstats.write_bios++;
1468 info = kmalloc(sizeof(*info), M_NFSREQ, M_WAITOK);
1470 info->v3 = NFS_ISV3(vp);
1471 info->info_writerpc.must_commit = 0;
1472 if ((bp->b_flags & (B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == 0)
1473 iomode = NFSV3WRITE_UNSTABLE;
1475 iomode = NFSV3WRITE_FILESYNC;
1477 KKASSERT(len <= nmp->nm_wsize);
1479 nfsstats.rpccnt[NFSPROC_WRITE]++;
1480 nfsm_reqhead(info, vp, NFSPROC_WRITE,
1481 NFSX_FH(info->v3) + 5 * NFSX_UNSIGNED + nfsm_rndup(len));
1482 ERROROUT(nfsm_fhtom(info, vp));
1484 tl = nfsm_build(info, 5 * NFSX_UNSIGNED);
1485 txdr_hyper(offset, tl);
1487 *tl++ = txdr_unsigned(len);
1488 *tl++ = txdr_unsigned(iomode);
1489 *tl = txdr_unsigned(len);
1493 tl = nfsm_build(info, 4 * NFSX_UNSIGNED);
1494 /* Set both "begin" and "current" to non-garbage. */
1495 x = txdr_unsigned((u_int32_t)offset);
1496 *tl++ = x; /* "begin offset" */
1497 *tl++ = x; /* "current offset" */
1498 x = txdr_unsigned(len);
1499 *tl++ = x; /* total to this offset */
1500 *tl = x; /* size of this write */
1502 ERROROUT(nfsm_biotom(info, bio, bp->b_dirtyoff, len));
1504 info->done = nfs_writerpc_bio_done;
1505 nfsm_request_bio(info, vp, NFSPROC_WRITE, NULL,
1506 nfs_vpcred(vp, ND_WRITE));
1509 kfree(info, M_NFSREQ);
1510 bp->b_error = error;
1511 bp->b_flags |= B_ERROR;
1516 nfs_writerpc_bio_done(nfsm_info_t info)
1518 struct nfsmount *nmp = VFSTONFS(info->vp->v_mount);
1519 struct nfsnode *np = VTONFS(info->vp);
1520 struct bio *bio = info->bio;
1521 struct buf *bp = bio->bio_buf;
1522 int wccflag = NFSV3_WCCRATTR;
1523 int iomode = NFSV3WRITE_FILESYNC;
1527 int len = bp->b_resid; /* b_resid was set to shortened length */
1530 lwkt_gettoken(&nmp->nm_token);
1534 * The write RPC returns a before and after mtime. The
1535 * nfsm_wcc_data() macro checks the before n_mtime
1536 * against the before time and stores the after time
1537 * in the nfsnode's cached vattr and n_mtime field.
1538 * The NRMODIFIED bit will be set if the before
1539 * time did not match the original mtime.
1541 wccflag = NFSV3_WCCCHK;
1542 ERROROUT(nfsm_wcc_data(info, info->vp, &wccflag));
1544 NULLOUT(tl = nfsm_dissect(info, 2 * NFSX_UNSIGNED + NFSX_V3WRITEVERF));
1545 rlen = fxdr_unsigned(int, *tl++);
1548 m_freem(info->mrep);
1551 } else if (rlen < len) {
1554 * XXX what do we do here?
1556 backup = len - rlen;
1557 uiop->uio_iov->iov_base = (char *)uiop->uio_iov->iov_base - backup;
1558 uiop->uio_iov->iov_len += backup;
1559 uiop->uio_offset -= backup;
1560 uiop->uio_resid += backup;
1564 commit = fxdr_unsigned(int, *tl++);
1567 * Return the lowest committment level
1568 * obtained by any of the RPCs.
1570 if (iomode == NFSV3WRITE_FILESYNC)
1572 else if (iomode == NFSV3WRITE_DATASYNC &&
1573 commit == NFSV3WRITE_UNSTABLE)
1575 if ((nmp->nm_state & NFSSTA_HASWRITEVERF) == 0){
1576 bcopy(tl, (caddr_t)nmp->nm_verf, NFSX_V3WRITEVERF);
1577 nmp->nm_state |= NFSSTA_HASWRITEVERF;
1578 } else if (bcmp(tl, nmp->nm_verf, NFSX_V3WRITEVERF)) {
1579 info->info_writerpc.must_commit = 1;
1580 bcopy(tl, (caddr_t)nmp->nm_verf, NFSX_V3WRITEVERF);
1584 ERROROUT(nfsm_loadattr(info, info->vp, NULL));
1586 m_freem(info->mrep);
1590 if (info->vp->v_mount->mnt_flag & MNT_ASYNC)
1591 iomode = NFSV3WRITE_FILESYNC;
1595 * End of RPC. Now clean up the bp.
1597 * We no longer enable write clustering for commit operations,
1598 * See around line 1157 for a more detailed comment.
1600 if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1601 bp->b_flags |= B_NEEDCOMMIT;
1603 /* XXX do not enable commit clustering */
1604 if (bp->b_dirtyoff == 0 && bp->b_dirtyend == bp->b_bcount)
1605 bp->b_flags |= B_CLUSTEROK;
1608 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1612 * For an interrupted write, the buffer is still valid
1613 * and the write hasn't been pushed to the server yet,
1614 * so we can't set B_ERROR and report the interruption
1615 * by setting B_EINTR. For the async case, B_EINTR
1616 * is not relevant, so the rpc attempt is essentially
1617 * a noop. For the case of a V3 write rpc not being
1618 * committed to stable storage, the block is still
1619 * dirty and requires either a commit rpc or another
1620 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1621 * the block is reused. This is indicated by setting
1622 * the B_DELWRI and B_NEEDCOMMIT flags.
1624 * If the buffer is marked B_PAGING, it does not reside on
1625 * the vp's paging queues so we cannot call bdirty(). The
1626 * bp in this case is not an NFS cache block so we should
1629 if (error == EINTR || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1631 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1632 if ((bp->b_flags & B_PAGING) == 0)
1635 bp->b_flags |= B_EINTR;
1639 bp->b_flags |= B_ERROR;
1640 bp->b_error = np->n_error = error;
1641 np->n_flag |= NWRITEERR;
1643 bp->b_dirtyoff = bp->b_dirtyend = 0;
1645 if (info->info_writerpc.must_commit)
1646 nfs_clearcommit(info->vp->v_mount);
1647 lwkt_reltoken(&nmp->nm_token);
1649 kfree(info, M_NFSREQ);
1651 bp->b_flags |= B_ERROR;
1652 bp->b_error = error;
1658 * Nfs Version 3 commit rpc - BIO version
1660 * This function issues the commit rpc and will chain to a write
1664 nfs_commitrpc_bio(struct vnode *vp, struct bio *bio)
1666 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1667 struct buf *bp = bio->bio_buf;
1668 struct nfsm_info *info;
1672 if ((nmp->nm_state & NFSSTA_HASWRITEVERF) == 0) {
1673 bp->b_dirtyoff = bp->b_dirtyend = 0;
1674 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1680 info = kmalloc(sizeof(*info), M_NFSREQ, M_WAITOK);
1684 nfsstats.rpccnt[NFSPROC_COMMIT]++;
1685 nfsm_reqhead(info, vp, NFSPROC_COMMIT, NFSX_FH(1));
1686 ERROROUT(nfsm_fhtom(info, vp));
1687 tl = nfsm_build(info, 3 * NFSX_UNSIGNED);
1688 txdr_hyper(bio->bio_offset + bp->b_dirtyoff, tl);
1690 *tl = txdr_unsigned(bp->b_dirtyend - bp->b_dirtyoff);
1692 info->done = nfs_commitrpc_bio_done;
1693 nfsm_request_bio(info, vp, NFSPROC_COMMIT, NULL,
1694 nfs_vpcred(vp, ND_WRITE));
1698 * Chain to write RPC on (early) error
1700 kfree(info, M_NFSREQ);
1701 nfs_writerpc_bio(vp, bio);
1705 nfs_commitrpc_bio_done(nfsm_info_t info)
1707 struct nfsmount *nmp = VFSTONFS(info->vp->v_mount);
1708 struct bio *bio = info->bio;
1709 struct buf *bp = bio->bio_buf;
1711 int wccflag = NFSV3_WCCRATTR;
1714 lwkt_gettoken(&nmp->nm_token);
1716 ERROROUT(nfsm_wcc_data(info, info->vp, &wccflag));
1718 NULLOUT(tl = nfsm_dissect(info, NFSX_V3WRITEVERF));
1719 if (bcmp(nmp->nm_verf, tl, NFSX_V3WRITEVERF)) {
1720 bcopy(tl, nmp->nm_verf, NFSX_V3WRITEVERF);
1721 error = NFSERR_STALEWRITEVERF;
1724 m_freem(info->mrep);
1728 * On completion we must chain to a write bio if an
1732 kfree(info, M_NFSREQ);
1734 bp->b_dirtyoff = bp->b_dirtyend = 0;
1735 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1739 nfs_writerpc_bio(info->vp, bio);
1741 lwkt_reltoken(&nmp->nm_token);