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>
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);
79 * Vnode op for VM getpages.
81 * nfs_getpages(struct vnode *a_vp, vm_page_t *a_m, int a_count,
82 * int a_reqpage, vm_ooffset_t a_offset)
85 nfs_getpages(struct vop_getpages_args *ap)
87 struct thread *td = curthread; /* XXX */
88 int i, error, nextoff, size, toff, count, npages;
99 nmp = VFSTONFS(vp->v_mount);
103 if (vp->v_object == NULL) {
104 kprintf("nfs_getpages: called with non-merged cache vnode??\n");
105 return VM_PAGER_ERROR;
108 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
109 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
110 (void)nfs_fsinfo(nmp, vp, td);
112 npages = btoc(count);
115 * NOTE that partially valid pages may occur in cases other
116 * then file EOF, such as when a file is partially written and
117 * ftruncate()-extended to a larger size. It is also possible
118 * for the valid bits to be set on garbage beyond the file EOF and
119 * clear in the area before EOF (e.g. m->valid == 0xfc), which can
120 * occur due to vtruncbuf() and the buffer cache's handling of
121 * pages which 'straddle' buffers or when b_bufsize is not a
122 * multiple of PAGE_SIZE.... the buffer cache cannot normally
123 * clear the extra bits. This kind of situation occurs when you
124 * make a small write() (m->valid == 0x03) and then mmap() and
125 * fault in the buffer(m->valid = 0xFF). When NFS flushes the
126 * buffer (vinvalbuf() m->valid = 0xFC) we are left with a mess.
128 * This is combined with the possibility that the pages are partially
129 * dirty or that there is a buffer backing the pages that is dirty
130 * (even if m->dirty is 0).
132 * To solve this problem several hacks have been made: (1) NFS
133 * guarentees that the IO block size is a multiple of PAGE_SIZE and
134 * (2) The buffer cache, when invalidating an NFS buffer, will
135 * disregard the buffer's fragmentory b_bufsize and invalidate
136 * the whole page rather then just the piece the buffer owns.
138 * This allows us to assume that a partially valid page found here
139 * is fully valid (vm_fault will zero'd out areas of the page not
142 m = pages[ap->a_reqpage];
144 for (i = 0; i < npages; ++i) {
145 if (i != ap->a_reqpage)
146 vnode_pager_freepage(pages[i]);
152 * Use an MSF_BUF as a medium to retrieve data from the pages.
154 msf_map_pagelist(&msf, pages, npages, 0);
156 kva = msf_buf_kva(msf);
162 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
163 uio.uio_resid = count;
164 uio.uio_segflg = UIO_SYSSPACE;
165 uio.uio_rw = UIO_READ;
168 error = nfs_readrpc_uio(vp, &uio);
171 if (error && (uio.uio_resid == count)) {
172 kprintf("nfs_getpages: error %d\n", error);
173 for (i = 0; i < npages; ++i) {
174 if (i != ap->a_reqpage)
175 vnode_pager_freepage(pages[i]);
177 return VM_PAGER_ERROR;
181 * Calculate the number of bytes read and validate only that number
182 * of bytes. Note that due to pending writes, size may be 0. This
183 * does not mean that the remaining data is invalid!
186 size = count - uio.uio_resid;
188 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
189 nextoff = toff + PAGE_SIZE;
192 m->flags &= ~PG_ZERO;
194 if (nextoff <= size) {
196 * Read operation filled an entire page
198 m->valid = VM_PAGE_BITS_ALL;
200 } else if (size > toff) {
202 * Read operation filled a partial page.
205 vm_page_set_validclean(m, 0, size - toff);
206 /* handled by vm_fault now */
207 /* vm_page_zero_invalid(m, TRUE); */
210 * Read operation was short. If no error occured
211 * we may have hit a zero-fill section. We simply
212 * leave valid set to 0.
216 if (i != ap->a_reqpage) {
218 * Whether or not to leave the page activated is up in
219 * the air, but we should put the page on a page queue
220 * somewhere (it already is in the object). Result:
221 * It appears that emperical results show that
222 * deactivating pages is best.
226 * Just in case someone was asking for this page we
227 * now tell them that it is ok to use.
230 if (m->flags & PG_WANTED)
233 vm_page_deactivate(m);
236 vnode_pager_freepage(m);
244 * Vnode op for VM putpages.
246 * nfs_putpages(struct vnode *a_vp, vm_page_t *a_m, int a_count, int a_sync,
247 * int *a_rtvals, vm_ooffset_t a_offset)
250 nfs_putpages(struct vop_putpages_args *ap)
252 struct thread *td = curthread;
256 int iomode, must_commit, i, error, npages, count;
260 struct nfsmount *nmp;
267 nmp = VFSTONFS(vp->v_mount);
270 rtvals = ap->a_rtvals;
271 npages = btoc(count);
272 offset = IDX_TO_OFF(pages[0]->pindex);
274 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
275 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
276 (void)nfs_fsinfo(nmp, vp, td);
278 for (i = 0; i < npages; i++) {
279 rtvals[i] = VM_PAGER_AGAIN;
283 * When putting pages, do not extend file past EOF.
286 if (offset + count > np->n_size) {
287 count = np->n_size - offset;
293 * Use an MSF_BUF as a medium to retrieve data from the pages.
295 msf_map_pagelist(&msf, pages, npages, 0);
297 kva = msf_buf_kva(msf);
303 uio.uio_offset = offset;
304 uio.uio_resid = count;
305 uio.uio_segflg = UIO_SYSSPACE;
306 uio.uio_rw = UIO_WRITE;
309 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
310 iomode = NFSV3WRITE_UNSTABLE;
312 iomode = NFSV3WRITE_FILESYNC;
314 error = nfs_writerpc(vp, &uio, &iomode, &must_commit);
319 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE;
320 for (i = 0; i < nwritten; i++) {
321 rtvals[i] = VM_PAGER_OK;
322 vm_page_undirty(pages[i]);
325 nfs_clearcommit(vp->v_mount);
331 * Vnode op for read using bio
334 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag)
336 struct nfsnode *np = VTONFS(vp);
338 struct buf *bp = 0, *rabp;
341 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
347 int nra, error = 0, n = 0, on = 0;
350 if (uio->uio_rw != UIO_READ)
351 panic("nfs_read mode");
353 if (uio->uio_resid == 0)
355 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
359 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
360 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
361 (void)nfs_fsinfo(nmp, vp, td);
362 if (vp->v_type != VDIR &&
363 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
365 biosize = vp->v_mount->mnt_stat.f_iosize;
366 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
369 * For nfs, cache consistency can only be maintained approximately.
370 * Although RFC1094 does not specify the criteria, the following is
371 * believed to be compatible with the reference port.
373 * NFS: If local changes have been made and this is a
374 * directory, the directory must be invalidated and
375 * the attribute cache must be cleared.
377 * GETATTR is called to synchronize the file size.
379 * If remote changes are detected local data is flushed
380 * and the cache is invalidated.
382 * NOTE: In the normal case the attribute cache is not
383 * cleared which means GETATTR may use cached data and
384 * not immediately detect changes made on the server.
386 if ((np->n_flag & NLMODIFIED) && vp->v_type == VDIR) {
388 error = nfs_vinvalbuf(vp, V_SAVE, 1);
393 error = VOP_GETATTR(vp, &vattr);
396 if (np->n_flag & NRMODIFIED) {
397 if (vp->v_type == VDIR)
399 error = nfs_vinvalbuf(vp, V_SAVE, 1);
402 np->n_flag &= ~NRMODIFIED;
405 if (np->n_flag & NDONTCACHE) {
406 switch (vp->v_type) {
408 return (nfs_readrpc_uio(vp, uio));
410 return (nfs_readlinkrpc(vp, uio));
414 kprintf(" NDONTCACHE: type %x unexpected\n", vp->v_type);
418 switch (vp->v_type) {
420 nfsstats.biocache_reads++;
421 lbn = uio->uio_offset / biosize;
422 on = uio->uio_offset & (biosize - 1);
423 loffset = (off_t)lbn * biosize;
426 * Start the read ahead(s), as required.
428 if (nmp->nm_readahead > 0 && nfs_asyncok(nmp)) {
429 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
430 (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) {
431 rabn = lbn + 1 + nra;
432 raoffset = (off_t)rabn * biosize;
433 if (findblk(vp, raoffset, FINDBLK_TEST) == NULL) {
434 rabp = nfs_getcacheblk(vp, raoffset, biosize, td);
437 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
438 rabp->b_cmd = BUF_CMD_READ;
439 vfs_busy_pages(vp, rabp);
440 nfs_asyncio(vp, &rabp->b_bio2);
442 if (nfs_startio(vp, &rabp->b_bio2, td)) {
443 rabp->b_flags |= B_INVAL|B_ERROR;
444 vfs_unbusy_pages(rabp);
457 * Obtain the buffer cache block. Figure out the buffer size
458 * when we are at EOF. If we are modifying the size of the
459 * buffer based on an EOF condition we need to hold
460 * nfs_rslock() through obtaining the buffer to prevent
461 * a potential writer-appender from messing with n_size.
462 * Otherwise we may accidently truncate the buffer and
465 * Note that bcount is *not* DEV_BSIZE aligned.
470 if (loffset >= np->n_size) {
472 } else if (loffset + biosize > np->n_size) {
473 bcount = np->n_size - loffset;
475 if (bcount != biosize) {
476 switch(nfs_rslock(np)) {
489 bp = nfs_getcacheblk(vp, loffset, bcount, td);
491 if (bcount != biosize)
497 * If B_CACHE is not set, we must issue the read. If this
498 * fails, we return an error.
501 if ((bp->b_flags & B_CACHE) == 0) {
502 bp->b_cmd = BUF_CMD_READ;
503 bp->b_bio2.bio_done = nfsiodone_sync;
504 bp->b_bio2.bio_flags |= BIO_SYNC;
505 vfs_busy_pages(vp, bp);
506 nfs_startio(vp, &bp->b_bio2, td);
507 error = nfs_iowait(&bp->b_bio2);
515 * on is the offset into the current bp. Figure out how many
516 * bytes we can copy out of the bp. Note that bcount is
517 * NOT DEV_BSIZE aligned.
519 * Then figure out how many bytes we can copy into the uio.
524 n = min((unsigned)(bcount - on), uio->uio_resid);
527 biosize = min(NFS_MAXPATHLEN, np->n_size);
528 nfsstats.biocache_readlinks++;
529 bp = nfs_getcacheblk(vp, (off_t)0, biosize, td);
532 if ((bp->b_flags & B_CACHE) == 0) {
533 bp->b_cmd = BUF_CMD_READ;
534 bp->b_bio2.bio_done = nfsiodone_sync;
535 bp->b_bio2.bio_flags |= BIO_SYNC;
536 vfs_busy_pages(vp, bp);
537 nfs_startio(vp, &bp->b_bio2, td);
538 error = nfs_iowait(&bp->b_bio2);
540 bp->b_flags |= B_ERROR | B_INVAL;
545 n = min(uio->uio_resid, bp->b_bcount - bp->b_resid);
549 nfsstats.biocache_readdirs++;
550 if (np->n_direofoffset
551 && uio->uio_offset >= np->n_direofoffset) {
554 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
555 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
556 loffset = uio->uio_offset - on;
557 bp = nfs_getcacheblk(vp, loffset, NFS_DIRBLKSIZ, td);
561 if ((bp->b_flags & B_CACHE) == 0) {
562 bp->b_cmd = BUF_CMD_READ;
563 bp->b_bio2.bio_done = nfsiodone_sync;
564 bp->b_bio2.bio_flags |= BIO_SYNC;
565 vfs_busy_pages(vp, bp);
566 nfs_startio(vp, &bp->b_bio2, td);
567 error = nfs_iowait(&bp->b_bio2);
571 while (error == NFSERR_BAD_COOKIE) {
572 kprintf("got bad cookie vp %p bp %p\n", vp, bp);
574 error = nfs_vinvalbuf(vp, 0, 1);
576 * Yuck! The directory has been modified on the
577 * server. The only way to get the block is by
578 * reading from the beginning to get all the
581 * Leave the last bp intact unless there is an error.
582 * Loop back up to the while if the error is another
583 * NFSERR_BAD_COOKIE (double yuch!).
585 for (i = 0; i <= lbn && !error; i++) {
586 if (np->n_direofoffset
587 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
589 bp = nfs_getcacheblk(vp, (off_t)i * NFS_DIRBLKSIZ,
593 if ((bp->b_flags & B_CACHE) == 0) {
594 bp->b_cmd = BUF_CMD_READ;
595 bp->b_bio2.bio_done = nfsiodone_sync;
596 bp->b_bio2.bio_flags |= BIO_SYNC;
597 vfs_busy_pages(vp, bp);
598 nfs_startio(vp, &bp->b_bio2, td);
599 error = nfs_iowait(&bp->b_bio2);
601 * no error + B_INVAL == directory EOF,
604 if (error == 0 && (bp->b_flags & B_INVAL))
608 * An error will throw away the block and the
609 * for loop will break out. If no error and this
610 * is not the block we want, we throw away the
611 * block and go for the next one via the for loop.
613 if (error || i < lbn)
618 * The above while is repeated if we hit another cookie
619 * error. If we hit an error and it wasn't a cookie error,
627 * If not eof and read aheads are enabled, start one.
628 * (You need the current block first, so that you have the
629 * directory offset cookie of the next block.)
631 if (nmp->nm_readahead > 0 && nfs_asyncok(nmp) &&
632 (bp->b_flags & B_INVAL) == 0 &&
633 (np->n_direofoffset == 0 ||
634 loffset + NFS_DIRBLKSIZ < np->n_direofoffset) &&
635 (np->n_flag & NDONTCACHE) == 0 &&
636 findblk(vp, loffset + NFS_DIRBLKSIZ, FINDBLK_TEST) == NULL
638 rabp = nfs_getcacheblk(vp, loffset + NFS_DIRBLKSIZ,
641 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
642 rabp->b_cmd = BUF_CMD_READ;
643 vfs_busy_pages(vp, rabp);
644 nfs_asyncio(vp, &rabp->b_bio2);
646 if (nfs_startio(vp, &rabp->b_bio2, td)) {
647 rabp->b_flags |= B_INVAL|B_ERROR;
648 vfs_unbusy_pages(rabp);
658 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
659 * chopped for the EOF condition, we cannot tell how large
660 * NFS directories are going to be until we hit EOF. So
661 * an NFS directory buffer is *not* chopped to its EOF. Now,
662 * it just so happens that b_resid will effectively chop it
663 * to EOF. *BUT* this information is lost if the buffer goes
664 * away and is reconstituted into a B_CACHE state ( due to
665 * being VMIO ) later. So we keep track of the directory eof
666 * in np->n_direofoffset and chop it off as an extra step
669 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
670 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
671 n = np->n_direofoffset - uio->uio_offset;
674 kprintf(" nfs_bioread: type %x unexpected\n",vp->v_type);
678 switch (vp->v_type) {
681 error = uiomove(bp->b_data + on, (int)n, uio);
685 error = uiomove(bp->b_data + on, (int)n, uio);
690 off_t old_off = uio->uio_offset;
692 struct nfs_dirent *dp;
695 * We are casting cpos to nfs_dirent, it must be
703 cpos = bp->b_data + on;
704 epos = bp->b_data + on + n;
705 while (cpos < epos && error == 0 && uio->uio_resid > 0) {
706 dp = (struct nfs_dirent *)cpos;
707 error = nfs_check_dirent(dp, (int)(epos - cpos));
710 if (vop_write_dirent(&error, uio, dp->nfs_ino,
711 dp->nfs_type, dp->nfs_namlen, dp->nfs_name)) {
714 cpos += dp->nfs_reclen;
718 uio->uio_offset = old_off + cpos - bp->b_data - on;
721 * Invalidate buffer if caching is disabled, forcing a
722 * re-read from the remote later.
724 if (np->n_flag & NDONTCACHE)
725 bp->b_flags |= B_INVAL;
728 kprintf(" nfs_bioread: type %x unexpected\n",vp->v_type);
731 } while (error == 0 && uio->uio_resid > 0 && n > 0);
736 * Userland can supply any 'seek' offset when reading a NFS directory.
737 * Validate the structure so we don't panic the kernel. Note that
738 * the element name is nul terminated and the nul is not included
743 nfs_check_dirent(struct nfs_dirent *dp, int maxlen)
745 int nfs_name_off = offsetof(struct nfs_dirent, nfs_name[0]);
747 if (nfs_name_off >= maxlen)
749 if (dp->nfs_reclen < nfs_name_off || dp->nfs_reclen > maxlen)
751 if (nfs_name_off + dp->nfs_namlen >= dp->nfs_reclen)
753 if (dp->nfs_reclen & 3)
759 * Vnode op for write using bio
761 * nfs_write(struct vnode *a_vp, struct uio *a_uio, int a_ioflag,
762 * struct ucred *a_cred)
765 nfs_write(struct vop_write_args *ap)
767 struct uio *uio = ap->a_uio;
768 struct thread *td = uio->uio_td;
769 struct vnode *vp = ap->a_vp;
770 struct nfsnode *np = VTONFS(vp);
771 int ioflag = ap->a_ioflag;
774 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
777 int n, on, error = 0, iomode, must_commit;
783 if (uio->uio_rw != UIO_WRITE)
784 panic("nfs_write mode");
785 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread)
786 panic("nfs_write proc");
788 if (vp->v_type != VREG)
790 if (np->n_flag & NWRITEERR) {
791 np->n_flag &= ~NWRITEERR;
792 return (np->n_error);
794 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
795 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
796 (void)nfs_fsinfo(nmp, vp, td);
799 * Synchronously flush pending buffers if we are in synchronous
800 * mode or if we are appending.
802 if (ioflag & (IO_APPEND | IO_SYNC)) {
803 if (np->n_flag & NLMODIFIED) {
805 error = nfs_flush(vp, MNT_WAIT, td, 0);
806 /* error = nfs_vinvalbuf(vp, V_SAVE, 1); */
813 * If IO_APPEND then load uio_offset. We restart here if we cannot
814 * get the append lock.
817 if (ioflag & IO_APPEND) {
819 error = VOP_GETATTR(vp, &vattr);
822 uio->uio_offset = np->n_size;
825 if (uio->uio_offset < 0)
827 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
829 if (uio->uio_resid == 0)
833 * We need to obtain the rslock if we intend to modify np->n_size
834 * in order to guarentee the append point with multiple contending
835 * writers, to guarentee that no other appenders modify n_size
836 * while we are trying to obtain a truncated buffer (i.e. to avoid
837 * accidently truncating data written by another appender due to
838 * the race), and to ensure that the buffer is populated prior to
839 * our extending of the file. We hold rslock through the entire
842 * Note that we do not synchronize the case where someone truncates
843 * the file while we are appending to it because attempting to lock
844 * this case may deadlock other parts of the system unexpectedly.
846 if ((ioflag & IO_APPEND) ||
847 uio->uio_offset + uio->uio_resid > np->n_size) {
848 switch(nfs_rslock(np)) {
863 * Maybe this should be above the vnode op call, but so long as
864 * file servers have no limits, i don't think it matters
866 if (td->td_proc && uio->uio_offset + uio->uio_resid >
867 td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
868 lwpsignal(td->td_proc, td->td_lwp, SIGXFSZ);
874 biosize = vp->v_mount->mnt_stat.f_iosize;
877 if ((np->n_flag & NDONTCACHE) && uio->uio_iovcnt == 1) {
878 iomode = NFSV3WRITE_FILESYNC;
879 error = nfs_writerpc(vp, uio, &iomode, &must_commit);
881 nfs_clearcommit(vp->v_mount);
884 nfsstats.biocache_writes++;
885 lbn = uio->uio_offset / biosize;
886 on = uio->uio_offset & (biosize-1);
887 loffset = uio->uio_offset - on;
888 n = min((unsigned)(biosize - on), uio->uio_resid);
891 * Handle direct append and file extension cases, calculate
892 * unaligned buffer size.
895 if (uio->uio_offset == np->n_size && n) {
897 * Get the buffer (in its pre-append state to maintain
898 * B_CACHE if it was previously set). Resize the
899 * nfsnode after we have locked the buffer to prevent
900 * readers from reading garbage.
903 bp = nfs_getcacheblk(vp, loffset, bcount, td);
908 np->n_size = uio->uio_offset + n;
909 np->n_flag |= NLMODIFIED;
910 vnode_pager_setsize(vp, np->n_size);
912 save = bp->b_flags & B_CACHE;
914 allocbuf(bp, bcount);
919 * Obtain the locked cache block first, and then
920 * adjust the file's size as appropriate.
923 if (loffset + bcount < np->n_size) {
924 if (loffset + biosize < np->n_size)
927 bcount = np->n_size - loffset;
929 bp = nfs_getcacheblk(vp, loffset, bcount, td);
930 if (uio->uio_offset + n > np->n_size) {
931 np->n_size = uio->uio_offset + n;
932 np->n_flag |= NLMODIFIED;
933 vnode_pager_setsize(vp, np->n_size);
943 * Issue a READ if B_CACHE is not set. In special-append
944 * mode, B_CACHE is based on the buffer prior to the write
945 * op and is typically set, avoiding the read. If a read
946 * is required in special append mode, the server will
947 * probably send us a short-read since we extended the file
948 * on our end, resulting in b_resid == 0 and, thusly,
949 * B_CACHE getting set.
951 * We can also avoid issuing the read if the write covers
952 * the entire buffer. We have to make sure the buffer state
953 * is reasonable in this case since we will not be initiating
954 * I/O. See the comments in kern/vfs_bio.c's getblk() for
957 * B_CACHE may also be set due to the buffer being cached
960 * When doing a UIO_NOCOPY write the buffer is not
961 * overwritten and we cannot just set B_CACHE unconditionally
962 * for full-block writes.
965 if (on == 0 && n == bcount && uio->uio_segflg != UIO_NOCOPY) {
966 bp->b_flags |= B_CACHE;
967 bp->b_flags &= ~(B_ERROR | B_INVAL);
970 if ((bp->b_flags & B_CACHE) == 0) {
971 bp->b_cmd = BUF_CMD_READ;
972 bp->b_bio2.bio_done = nfsiodone_sync;
973 bp->b_bio2.bio_flags |= BIO_SYNC;
974 vfs_busy_pages(vp, bp);
975 nfs_startio(vp, &bp->b_bio2, td);
976 error = nfs_iowait(&bp->b_bio2);
986 np->n_flag |= NLMODIFIED;
989 * If dirtyend exceeds file size, chop it down. This should
990 * not normally occur but there is an append race where it
991 * might occur XXX, so we log it.
993 * If the chopping creates a reverse-indexed or degenerate
994 * situation with dirtyoff/end, we 0 both of them.
997 if (bp->b_dirtyend > bcount) {
998 kprintf("NFS append race @%08llx:%d\n",
999 (long long)bp->b_bio2.bio_offset,
1000 bp->b_dirtyend - bcount);
1001 bp->b_dirtyend = bcount;
1004 if (bp->b_dirtyoff >= bp->b_dirtyend)
1005 bp->b_dirtyoff = bp->b_dirtyend = 0;
1008 * If the new write will leave a contiguous dirty
1009 * area, just update the b_dirtyoff and b_dirtyend,
1010 * otherwise force a write rpc of the old dirty area.
1012 * While it is possible to merge discontiguous writes due to
1013 * our having a B_CACHE buffer ( and thus valid read data
1014 * for the hole), we don't because it could lead to
1015 * significant cache coherency problems with multiple clients,
1016 * especially if locking is implemented later on.
1018 * as an optimization we could theoretically maintain
1019 * a linked list of discontinuous areas, but we would still
1020 * have to commit them separately so there isn't much
1021 * advantage to it except perhaps a bit of asynchronization.
1024 if (bp->b_dirtyend > 0 &&
1025 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1026 if (bwrite(bp) == EINTR) {
1033 error = uiomove((char *)bp->b_data + on, n, uio);
1036 * Since this block is being modified, it must be written
1037 * again and not just committed. Since write clustering does
1038 * not work for the stage 1 data write, only the stage 2
1039 * commit rpc, we have to clear B_CLUSTEROK as well.
1041 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1044 bp->b_flags |= B_ERROR;
1050 * Only update dirtyoff/dirtyend if not a degenerate
1054 if (bp->b_dirtyend > 0) {
1055 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1056 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1058 bp->b_dirtyoff = on;
1059 bp->b_dirtyend = on + n;
1061 vfs_bio_set_validclean(bp, on, n);
1065 * If the lease is non-cachable or IO_SYNC do bwrite().
1067 * IO_INVAL appears to be unused. The idea appears to be
1068 * to turn off caching in this case. Very odd. XXX
1070 * If nfs_async is set bawrite() will use an unstable write
1071 * (build dirty bufs on the server), so we might as well
1072 * push it out with bawrite(). If nfs_async is not set we
1073 * use bdwrite() to cache dirty bufs on the client.
1075 if ((np->n_flag & NDONTCACHE) || (ioflag & IO_SYNC)) {
1076 if (ioflag & IO_INVAL)
1077 bp->b_flags |= B_NOCACHE;
1081 if (np->n_flag & NDONTCACHE) {
1082 error = nfs_vinvalbuf(vp, V_SAVE, 1);
1086 } else if ((n + on) == biosize && nfs_async) {
1091 } while (uio->uio_resid > 0 && n > 0);
1100 * Get an nfs cache block.
1102 * Allocate a new one if the block isn't currently in the cache
1103 * and return the block marked busy. If the calling process is
1104 * interrupted by a signal for an interruptible mount point, return
1107 * The caller must carefully deal with the possible B_INVAL state of
1108 * the buffer. nfs_startio() clears B_INVAL (and nfs_asyncio() clears it
1109 * indirectly), so synchronous reads can be issued without worrying about
1110 * the B_INVAL state. We have to be a little more careful when dealing
1111 * with writes (see comments in nfs_write()) when extending a file past
1115 nfs_getcacheblk(struct vnode *vp, off_t loffset, int size, struct thread *td)
1119 struct nfsmount *nmp;
1124 if (nmp->nm_flag & NFSMNT_INT) {
1125 bp = getblk(vp, loffset, size, GETBLK_PCATCH, 0);
1126 while (bp == NULL) {
1127 if (nfs_sigintr(nmp, NULL, td))
1129 bp = getblk(vp, loffset, size, 0, 2 * hz);
1132 bp = getblk(vp, loffset, size, 0, 0);
1136 * bio2, the 'device' layer. Since BIOs use 64 bit byte offsets
1137 * now, no translation is necessary.
1139 bp->b_bio2.bio_offset = loffset;
1144 * Flush and invalidate all dirty buffers. If another process is already
1145 * doing the flush, just wait for completion.
1148 nfs_vinvalbuf(struct vnode *vp, int flags, int intrflg)
1150 struct nfsnode *np = VTONFS(vp);
1151 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1152 int error = 0, slpflag, slptimeo;
1153 thread_t td = curthread;
1155 if (vp->v_flag & VRECLAIMED)
1158 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1168 * First wait for any other process doing a flush to complete.
1170 while (np->n_flag & NFLUSHINPROG) {
1171 np->n_flag |= NFLUSHWANT;
1172 error = tsleep((caddr_t)&np->n_flag, 0, "nfsvinval", slptimeo);
1173 if (error && intrflg && nfs_sigintr(nmp, NULL, td))
1178 * Now, flush as required.
1180 np->n_flag |= NFLUSHINPROG;
1181 error = vinvalbuf(vp, flags, slpflag, 0);
1183 if (intrflg && nfs_sigintr(nmp, NULL, td)) {
1184 np->n_flag &= ~NFLUSHINPROG;
1185 if (np->n_flag & NFLUSHWANT) {
1186 np->n_flag &= ~NFLUSHWANT;
1187 wakeup((caddr_t)&np->n_flag);
1191 error = vinvalbuf(vp, flags, 0, slptimeo);
1193 np->n_flag &= ~(NLMODIFIED | NFLUSHINPROG);
1194 if (np->n_flag & NFLUSHWANT) {
1195 np->n_flag &= ~NFLUSHWANT;
1196 wakeup((caddr_t)&np->n_flag);
1202 * Return true (non-zero) if the txthread and rxthread are operational
1203 * and we do not already have too many not-yet-started BIO's built up.
1206 nfs_asyncok(struct nfsmount *nmp)
1208 return (nmp->nm_bioqlen < NFS_MAXASYNCBIO &&
1209 nmp->nm_bioqlen < nmp->nm_maxasync_scaled / NFS_ASYSCALE &&
1210 nmp->nm_rxstate <= NFSSVC_PENDING &&
1211 nmp->nm_txstate <= NFSSVC_PENDING);
1215 * The read-ahead code calls this to queue a bio to the txthread.
1217 * We don't touch the bio otherwise... that is, we do not even
1218 * construct or send the initial rpc. The txthread will do it
1221 * NOTE! nm_bioqlen is not decremented until the request completes,
1222 * so it does not reflect the number of bio's on bioq.
1225 nfs_asyncio(struct vnode *vp, struct bio *bio)
1227 struct buf *bp = bio->bio_buf;
1228 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1230 KKASSERT(vp->v_tag == VT_NFS);
1232 bio->bio_driver_info = vp;
1234 TAILQ_INSERT_TAIL(&nmp->nm_bioq, bio, bio_act);
1235 atomic_add_int(&nmp->nm_bioqlen, 1);
1237 nfssvc_iod_writer_wakeup(nmp);
1241 * Initiate an I/O operation to/from a cache block. If the BIO is
1242 * flagged BIO_SYNC, or if the async thread is not running, the
1243 * operation will be executed synchronously.
1245 * Typically for BIO_SYNC the caller set up the completion and will
1246 * call nfs_iowait() to obtain the error code, then brelse().
1247 * iowait is a degenerate routine.
1249 * For async operation we set up a request and queue it the transmit
1250 * thread along with a done function to deal with cleanup after
1251 * the RPC completes. The presence of a done function causes the
1252 * state machine to automatically move the req onto the reqrxq when
1253 * a reponse is received.
1255 * NOTE! TD MIGHT BE NULL
1258 nfs_startio(struct vnode *vp, struct bio *bio, struct thread *td)
1260 struct buf *bp = bio->bio_buf;
1263 struct nfsmount *nmp;
1264 int error = 0, iomode, must_commit = 0;
1268 KKASSERT(vp->v_tag == VT_NFS);
1270 nmp = VFSTONFS(vp->v_mount);
1272 uiop->uio_iov = &io;
1273 uiop->uio_iovcnt = 1;
1274 uiop->uio_segflg = UIO_SYSSPACE;
1278 * clear B_ERROR and B_INVAL state prior to initiating the I/O. We
1279 * do this here so we do not have to do it in all the code that
1282 * NOTE: An EINPROGRESS response can be returned if the bio was
1285 bp->b_flags &= ~(B_ERROR | B_INVAL);
1287 KASSERT(bp->b_cmd != BUF_CMD_DONE,
1288 ("nfs_doio: bp %p already marked done!", bp));
1290 if (bp->b_cmd == BUF_CMD_READ) {
1291 io.iov_len = uiop->uio_resid = bp->b_bcount;
1292 io.iov_base = bp->b_data;
1293 uiop->uio_rw = UIO_READ;
1295 switch (vp->v_type) {
1298 * Note: NFS assumes BIO_SYNC is run synchronously, so
1299 * be sure to do that.
1301 nfsstats.read_bios++;
1302 if ((bio->bio_flags & BIO_SYNC) == 0) {
1303 nfs_readrpc_bio(vp, bio);
1306 uiop->uio_offset = bio->bio_offset;
1307 error = nfs_readrpc_uio(vp, uiop);
1309 if (uiop->uio_resid) {
1311 * If we had a short read with no error, we must have
1312 * hit a file hole. We should zero-fill the remainder.
1313 * This can also occur if the server hits the file EOF.
1315 * Holes used to be able to occur due to pending
1316 * writes, but that is not possible any longer.
1318 int nread = bp->b_bcount - bp->b_resid;
1319 int left = bp->b_resid;
1322 bzero((char *)bp->b_data + nread, left);
1326 if (td && td->td_proc && (vp->v_flag & VTEXT) &&
1327 np->n_mtime != np->n_vattr.va_mtime.tv_sec) {
1328 uprintf("Process killed due to text file modification\n");
1329 ksignal(td->td_proc, SIGKILL);
1333 uiop->uio_offset = 0;
1334 nfsstats.readlink_bios++;
1335 error = nfs_readlinkrpc(vp, uiop);
1338 nfsstats.readdir_bios++;
1339 uiop->uio_offset = bio->bio_offset;
1340 if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
1341 error = nfs_readdirplusrpc(vp, uiop);
1342 if (error == NFSERR_NOTSUPP)
1343 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1345 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1346 error = nfs_readdirrpc(vp, uiop);
1348 * end-of-directory sets B_INVAL but does not generate an
1351 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1352 bp->b_flags |= B_INVAL;
1355 kprintf("nfs_doio: type %x unexpected\n",vp->v_type);
1359 bp->b_flags |= B_ERROR;
1360 bp->b_error = error;
1364 * If we only need to commit, try to commit
1366 KKASSERT(bp->b_cmd == BUF_CMD_WRITE);
1367 if (bp->b_flags & B_NEEDCOMMIT) {
1371 off = bio->bio_offset + bp->b_dirtyoff;
1372 retv = nfs_commit(vp, off,
1373 bp->b_dirtyend - bp->b_dirtyoff, td);
1375 bp->b_dirtyoff = bp->b_dirtyend = 0;
1376 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1381 if (retv == NFSERR_STALEWRITEVERF) {
1382 nfs_clearcommit(vp->v_mount);
1387 * Setup for actual write
1390 if (bio->bio_offset + bp->b_dirtyend > np->n_size)
1391 bp->b_dirtyend = np->n_size - bio->bio_offset;
1393 if (bp->b_dirtyend > bp->b_dirtyoff) {
1394 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1396 uiop->uio_offset = bio->bio_offset + bp->b_dirtyoff;
1397 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1398 uiop->uio_rw = UIO_WRITE;
1399 nfsstats.write_bios++;
1401 if ((bp->b_flags & (B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == 0)
1402 iomode = NFSV3WRITE_UNSTABLE;
1404 iomode = NFSV3WRITE_FILESYNC;
1406 error = nfs_writerpc(vp, uiop, &iomode, &must_commit);
1409 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1410 * to cluster the buffers needing commit. This will allow
1411 * the system to submit a single commit rpc for the whole
1412 * cluster. We can do this even if the buffer is not 100%
1413 * dirty (relative to the NFS blocksize), so we optimize the
1414 * append-to-file-case.
1416 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1417 * cleared because write clustering only works for commit
1418 * rpc's, not for the data portion of the write).
1421 if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1422 bp->b_flags |= B_NEEDCOMMIT;
1423 if (bp->b_dirtyoff == 0
1424 && bp->b_dirtyend == bp->b_bcount)
1425 bp->b_flags |= B_CLUSTEROK;
1427 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1431 * For an interrupted write, the buffer is still valid
1432 * and the write hasn't been pushed to the server yet,
1433 * so we can't set B_ERROR and report the interruption
1434 * by setting B_EINTR. For the async case, B_EINTR
1435 * is not relevant, so the rpc attempt is essentially
1436 * a noop. For the case of a V3 write rpc not being
1437 * committed to stable storage, the block is still
1438 * dirty and requires either a commit rpc or another
1439 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1440 * the block is reused. This is indicated by setting
1441 * the B_DELWRI and B_NEEDCOMMIT flags.
1443 * If the buffer is marked B_PAGING, it does not reside on
1444 * the vp's paging queues so we cannot call bdirty(). The
1445 * bp in this case is not an NFS cache block so we should
1449 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1451 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1452 if ((bp->b_flags & B_PAGING) == 0)
1455 bp->b_flags |= B_EINTR;
1459 bp->b_flags |= B_ERROR;
1460 bp->b_error = np->n_error = error;
1461 np->n_flag |= NWRITEERR;
1463 bp->b_dirtyoff = bp->b_dirtyend = 0;
1471 bp->b_resid = uiop->uio_resid;
1473 nfs_clearcommit(vp->v_mount);
1478 * Used to aid in handling ftruncate() operations on the NFS client side.
1479 * Truncation creates a number of special problems for NFS. We have to
1480 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1481 * we have to properly handle VM pages or (potentially dirty) buffers
1482 * that straddle the truncation point.
1486 nfs_meta_setsize(struct vnode *vp, struct thread *td, u_quad_t nsize)
1488 struct nfsnode *np = VTONFS(vp);
1489 u_quad_t tsize = np->n_size;
1490 int biosize = vp->v_mount->mnt_stat.f_iosize;
1495 if (np->n_size < tsize) {
1502 * vtruncbuf() doesn't get the buffer overlapping the
1503 * truncation point. We may have a B_DELWRI and/or B_CACHE
1504 * buffer that now needs to be truncated.
1506 error = vtruncbuf(vp, nsize, biosize);
1507 lbn = nsize / biosize;
1508 bufsize = nsize & (biosize - 1);
1509 loffset = nsize - bufsize;
1510 bp = nfs_getcacheblk(vp, loffset, bufsize, td);
1511 if (bp->b_dirtyoff > bp->b_bcount)
1512 bp->b_dirtyoff = bp->b_bcount;
1513 if (bp->b_dirtyend > bp->b_bcount)
1514 bp->b_dirtyend = bp->b_bcount;
1515 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1518 vnode_pager_setsize(vp, nsize);
1524 * Synchronous completion for nfs_doio. Call bpdone() with elseit=FALSE.
1525 * Caller is responsible for brelse()'ing the bp.
1528 nfsiodone_sync(struct bio *bio)
1531 bpdone(bio->bio_buf, 0);
1535 * If nfs_startio() was told to do the request BIO_SYNC it will
1536 * complete the request before returning, so assert that the
1537 * request is in-fact complete.
1540 nfs_iowait(struct bio *bio)
1542 struct buf *bp = bio->bio_buf;
1544 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
1545 if (bp->b_flags & B_EINTR)
1547 if (bp->b_flags & B_ERROR)
1548 return (bp->b_error ? bp->b_error : EIO);
1553 * nfs read rpc - BIO version
1555 static void nfs_readrpc_bio_done(nfsm_info_t info);
1558 nfs_readrpc_bio(struct vnode *vp, struct bio *bio)
1560 struct buf *bp = bio->bio_buf;
1562 struct nfsmount *nmp;
1563 int error = 0, len, tsiz;
1564 struct nfsm_info *info;
1566 info = kmalloc(sizeof(*info), M_NFSREQ, M_WAITOK);
1568 info->v3 = NFS_ISV3(vp);
1570 nmp = VFSTONFS(vp->v_mount);
1571 tsiz = bp->b_bcount;
1572 if (bio->bio_offset + tsiz > nmp->nm_maxfilesize) {
1576 nfsstats.rpccnt[NFSPROC_READ]++;
1577 len = (tsiz > nmp->nm_rsize) ? nmp->nm_rsize : tsiz;
1578 nfsm_reqhead(info, vp, NFSPROC_READ,
1579 NFSX_FH(info->v3) + NFSX_UNSIGNED * 3);
1580 ERROROUT(nfsm_fhtom(info, vp));
1581 tl = nfsm_build(info, NFSX_UNSIGNED * 3);
1583 txdr_hyper(bio->bio_offset, tl);
1584 *(tl + 2) = txdr_unsigned(len);
1586 *tl++ = txdr_unsigned(bio->bio_offset);
1587 *tl++ = txdr_unsigned(len);
1591 info->done = nfs_readrpc_bio_done;
1592 nfsm_request_bio(info, vp, NFSPROC_READ, NULL,
1593 nfs_vpcred(vp, ND_READ));
1596 kfree(info, M_NFSREQ);
1597 bp->b_error = error;
1598 bp->b_flags |= B_ERROR;
1603 nfs_readrpc_bio_done(nfsm_info_t info)
1605 struct nfsmount *nmp = VFSTONFS(info->vp->v_mount);
1606 struct bio *bio = info->bio;
1607 struct buf *bp = bio->bio_buf;
1614 KKASSERT(info->state == NFSM_STATE_DONE);
1617 ERROROUT(nfsm_postop_attr(info, info->vp, &attrflag,
1618 NFS_LATTR_NOSHRINK));
1619 NULLOUT(tl = nfsm_dissect(info, 2 * NFSX_UNSIGNED));
1620 eof = fxdr_unsigned(int, *(tl + 1));
1622 ERROROUT(nfsm_loadattr(info, info->vp, NULL));
1625 NEGATIVEOUT(retlen = nfsm_strsiz(info, nmp->nm_rsize));
1626 ERROROUT(nfsm_mtobio(info, bio, retlen));
1627 m_freem(info->mrep);
1631 * No error occured, fill the hole if any
1633 if (retlen < bp->b_bcount) {
1634 bzero(bp->b_data + retlen, bp->b_bcount - retlen);
1636 bp->b_resid = bp->b_bcount - retlen;
1641 if (eof || retlen == 0) {
1644 } else if (retlen < len) {
1649 kfree(info, M_NFSREQ);
1651 bp->b_error = error;
1652 bp->b_flags |= B_ERROR;
1660 * nfs write call - BIO version
1663 nfs_writerpc_bio(struct vnode *vp, struct bio *bio, int *iomode, int *must_commit)
1667 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1668 int error = 0, len, tsiz, wccflag = NFSV3_WCCRATTR, rlen, commit;
1669 int committed = NFSV3WRITE_FILESYNC;
1670 struct nfsm_info info;
1673 info.v3 = NFS_ISV3(vp);
1676 if (uiop->uio_iovcnt != 1)
1677 panic("nfs: writerpc iovcnt > 1");
1680 tsiz = uiop->uio_resid;
1681 if (uiop->uio_offset + tsiz > nmp->nm_maxfilesize)
1684 nfsstats.rpccnt[NFSPROC_WRITE]++;
1685 len = (tsiz > nmp->nm_wsize) ? nmp->nm_wsize : tsiz;
1686 nfsm_reqhead(&info, vp, NFSPROC_WRITE,
1687 NFSX_FH(info.v3) + 5 * NFSX_UNSIGNED + nfsm_rndup(len));
1688 ERROROUT(nfsm_fhtom(&info, vp));
1690 tl = nfsm_build(&info, 5 * NFSX_UNSIGNED);
1691 txdr_hyper(uiop->uio_offset, tl);
1693 *tl++ = txdr_unsigned(len);
1694 *tl++ = txdr_unsigned(*iomode);
1695 *tl = txdr_unsigned(len);
1699 tl = nfsm_build(&info, 4 * NFSX_UNSIGNED);
1700 /* Set both "begin" and "current" to non-garbage. */
1701 x = txdr_unsigned((u_int32_t)uiop->uio_offset);
1702 *tl++ = x; /* "begin offset" */
1703 *tl++ = x; /* "current offset" */
1704 x = txdr_unsigned(len);
1705 *tl++ = x; /* total to this offset */
1706 *tl = x; /* size of this write */
1708 ERROROUT(nfsm_uiotom(&info, uiop, len));
1709 NEGKEEPOUT(nfsm_request(&info, vp, NFSPROC_WRITE, uiop->uio_td,
1710 nfs_vpcred(vp, ND_WRITE), &error));
1713 * The write RPC returns a before and after mtime. The
1714 * nfsm_wcc_data() macro checks the before n_mtime
1715 * against the before time and stores the after time
1716 * in the nfsnode's cached vattr and n_mtime field.
1717 * The NRMODIFIED bit will be set if the before
1718 * time did not match the original mtime.
1720 wccflag = NFSV3_WCCCHK;
1721 ERROROUT(nfsm_wcc_data(&info, vp, &wccflag));
1723 NULLOUT(tl = nfsm_dissect(&info, 2 * NFSX_UNSIGNED + NFSX_V3WRITEVERF));
1724 rlen = fxdr_unsigned(int, *tl++);
1730 } else if (rlen < len) {
1731 backup = len - rlen;
1732 uiop->uio_iov->iov_base = (char *)uiop->uio_iov->iov_base - backup;
1733 uiop->uio_iov->iov_len += backup;
1734 uiop->uio_offset -= backup;
1735 uiop->uio_resid += backup;
1738 commit = fxdr_unsigned(int, *tl++);
1741 * Return the lowest committment level
1742 * obtained by any of the RPCs.
1744 if (committed == NFSV3WRITE_FILESYNC)
1746 else if (committed == NFSV3WRITE_DATASYNC &&
1747 commit == NFSV3WRITE_UNSTABLE)
1749 if ((nmp->nm_state & NFSSTA_HASWRITEVERF) == 0){
1750 bcopy((caddr_t)tl, (caddr_t)nmp->nm_verf,
1752 nmp->nm_state |= NFSSTA_HASWRITEVERF;
1753 } else if (bcmp((caddr_t)tl,
1754 (caddr_t)nmp->nm_verf, NFSX_V3WRITEVERF)) {
1756 bcopy((caddr_t)tl, (caddr_t)nmp->nm_verf,
1761 ERROROUT(nfsm_loadattr(&info, vp, NULL));
1770 if (vp->v_mount->mnt_flag & MNT_ASYNC)
1771 committed = NFSV3WRITE_FILESYNC;
1772 *iomode = committed;
1774 uiop->uio_resid = tsiz;