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 * 4. Neither the name of the University nor the names of its contributors
17 * may be used to endorse or promote products derived from this software
18 * without specific prior written permission.
20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * @(#)nfs_bio.c 8.9 (Berkeley) 3/30/95
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
38 #include <sys/param.h>
39 #include <sys/systm.h>
42 #include <sys/kernel.h>
44 #include <sys/mount.h>
46 #include <sys/rwlock.h>
47 #include <sys/vmmeter.h>
48 #include <sys/vnode.h>
51 #include <vm/vm_param.h>
52 #include <vm/vm_extern.h>
53 #include <vm/vm_page.h>
54 #include <vm/vm_object.h>
55 #include <vm/vm_pager.h>
56 #include <vm/vnode_pager.h>
58 #include <nfs/nfsproto.h>
59 #include <nfsclient/nfs.h>
60 #include <nfsclient/nfsmount.h>
61 #include <nfsclient/nfsnode.h>
62 #include <nfs/nfs_kdtrace.h>
64 static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size,
66 static int nfs_directio_write(struct vnode *vp, struct uio *uiop,
67 struct ucred *cred, int ioflag);
69 extern int nfs_directio_enable;
70 extern int nfs_directio_allow_mmap;
73 * Vnode op for VM getpages.
76 nfs_getpages(struct vop_getpages_args *ap)
78 int i, error, nextoff, size, toff, count, npages;
93 td = curthread; /* XXX */
94 cred = curthread->td_ucred; /* XXX */
95 nmp = VFSTONFS(vp->v_mount);
99 if ((object = vp->v_object) == NULL) {
100 nfs_printf("nfs_getpages: called with non-merged cache vnode??\n");
101 return (VM_PAGER_ERROR);
104 if (nfs_directio_enable && !nfs_directio_allow_mmap) {
105 mtx_lock(&np->n_mtx);
106 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
107 mtx_unlock(&np->n_mtx);
108 nfs_printf("nfs_getpages: called on non-cacheable vnode??\n");
109 return (VM_PAGER_ERROR);
111 mtx_unlock(&np->n_mtx);
114 mtx_lock(&nmp->nm_mtx);
115 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
116 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
117 mtx_unlock(&nmp->nm_mtx);
118 /* We'll never get here for v4, because we always have fsinfo */
119 (void)nfs_fsinfo(nmp, vp, cred, td);
121 mtx_unlock(&nmp->nm_mtx);
123 npages = btoc(count);
126 * If the requested page is partially valid, just return it and
127 * allow the pager to zero-out the blanks. Partially valid pages
128 * can only occur at the file EOF.
130 VM_OBJECT_WLOCK(object);
131 if (pages[ap->a_reqpage]->valid != 0) {
132 for (i = 0; i < npages; ++i) {
133 if (i != ap->a_reqpage) {
134 vm_page_lock(pages[i]);
135 vm_page_free(pages[i]);
136 vm_page_unlock(pages[i]);
139 VM_OBJECT_WUNLOCK(object);
142 VM_OBJECT_WUNLOCK(object);
145 * We use only the kva address for the buffer, but this is extremely
146 * convienient and fast.
148 bp = getpbuf(&nfs_pbuf_freecnt);
150 kva = (vm_offset_t) bp->b_data;
151 pmap_qenter(kva, pages, npages);
152 PCPU_INC(cnt.v_vnodein);
153 PCPU_ADD(cnt.v_vnodepgsin, npages);
155 iov.iov_base = (caddr_t) kva;
159 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
160 uio.uio_resid = count;
161 uio.uio_segflg = UIO_SYSSPACE;
162 uio.uio_rw = UIO_READ;
165 error = (nmp->nm_rpcops->nr_readrpc)(vp, &uio, cred);
166 pmap_qremove(kva, npages);
168 relpbuf(bp, &nfs_pbuf_freecnt);
170 if (error && (uio.uio_resid == count)) {
171 nfs_printf("nfs_getpages: error %d\n", error);
172 VM_OBJECT_WLOCK(object);
173 for (i = 0; i < npages; ++i) {
174 if (i != ap->a_reqpage) {
175 vm_page_lock(pages[i]);
176 vm_page_free(pages[i]);
177 vm_page_unlock(pages[i]);
180 VM_OBJECT_WUNLOCK(object);
181 return (VM_PAGER_ERROR);
185 * Calculate the number of bytes read and validate only that number
186 * of bytes. Note that due to pending writes, size may be 0. This
187 * does not mean that the remaining data is invalid!
190 size = count - uio.uio_resid;
191 VM_OBJECT_WLOCK(object);
192 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
194 nextoff = toff + PAGE_SIZE;
197 if (nextoff <= size) {
199 * Read operation filled an entire page
201 m->valid = VM_PAGE_BITS_ALL;
202 KASSERT(m->dirty == 0,
203 ("nfs_getpages: page %p is dirty", m));
204 } else if (size > toff) {
206 * Read operation filled a partial page.
209 vm_page_set_valid_range(m, 0, size - toff);
210 KASSERT(m->dirty == 0,
211 ("nfs_getpages: page %p is dirty", m));
214 * Read operation was short. If no error
215 * occured we may have hit a zero-fill
216 * section. We leave valid set to 0, and page
217 * is freed by vm_page_readahead_finish() if
218 * its index is not equal to requested, or
219 * page is zeroed and set valid by
220 * vm_pager_get_pages() for requested page.
224 if (i != ap->a_reqpage)
225 vm_page_readahead_finish(m);
227 VM_OBJECT_WUNLOCK(object);
232 * Vnode op for VM putpages.
235 nfs_putpages(struct vop_putpages_args *ap)
241 int iomode, must_commit, i, error, npages, count;
247 struct nfsmount *nmp;
253 td = curthread; /* XXX */
254 /* Set the cred to n_writecred for the write rpcs. */
255 if (np->n_writecred != NULL)
256 cred = crhold(np->n_writecred);
258 cred = crhold(curthread->td_ucred); /* XXX */
259 nmp = VFSTONFS(vp->v_mount);
262 rtvals = ap->a_rtvals;
263 npages = btoc(count);
264 offset = IDX_TO_OFF(pages[0]->pindex);
266 mtx_lock(&nmp->nm_mtx);
267 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
268 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
269 mtx_unlock(&nmp->nm_mtx);
270 (void)nfs_fsinfo(nmp, vp, cred, td);
272 mtx_unlock(&nmp->nm_mtx);
274 mtx_lock(&np->n_mtx);
275 if (nfs_directio_enable && !nfs_directio_allow_mmap &&
276 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
277 mtx_unlock(&np->n_mtx);
278 nfs_printf("nfs_putpages: called on noncache-able vnode??\n");
279 mtx_lock(&np->n_mtx);
282 for (i = 0; i < npages; i++)
283 rtvals[i] = VM_PAGER_ERROR;
286 * When putting pages, do not extend file past EOF.
288 if (offset + count > np->n_size) {
289 count = np->n_size - offset;
293 mtx_unlock(&np->n_mtx);
296 * We use only the kva address for the buffer, but this is extremely
297 * convienient and fast.
299 bp = getpbuf(&nfs_pbuf_freecnt);
301 kva = (vm_offset_t) bp->b_data;
302 pmap_qenter(kva, pages, npages);
303 PCPU_INC(cnt.v_vnodeout);
304 PCPU_ADD(cnt.v_vnodepgsout, count);
306 iov.iov_base = (caddr_t) kva;
310 uio.uio_offset = offset;
311 uio.uio_resid = count;
312 uio.uio_segflg = UIO_SYSSPACE;
313 uio.uio_rw = UIO_WRITE;
316 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
317 iomode = NFSV3WRITE_UNSTABLE;
319 iomode = NFSV3WRITE_FILESYNC;
321 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, &iomode, &must_commit);
324 pmap_qremove(kva, npages);
325 relpbuf(bp, &nfs_pbuf_freecnt);
328 vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid);
330 nfs_clearcommit(vp->v_mount);
337 * For nfs, cache consistency can only be maintained approximately.
338 * Although RFC1094 does not specify the criteria, the following is
339 * believed to be compatible with the reference port.
341 * If the file's modify time on the server has changed since the
342 * last read rpc or you have written to the file,
343 * you may have lost data cache consistency with the
344 * server, so flush all of the file's data out of the cache.
345 * Then force a getattr rpc to ensure that you have up to date
347 * NB: This implies that cache data can be read when up to
348 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
349 * attributes this could be forced by setting n_attrstamp to 0 before
350 * the VOP_GETATTR() call.
353 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
357 struct nfsnode *np = VTONFS(vp);
359 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
362 * Grab the exclusive lock before checking whether the cache is
364 * XXX - We can make this cheaper later (by acquiring cheaper locks).
365 * But for now, this suffices.
367 old_lock = nfs_upgrade_vnlock(vp);
368 if (vp->v_iflag & VI_DOOMED) {
369 nfs_downgrade_vnlock(vp, old_lock);
373 mtx_lock(&np->n_mtx);
374 if (np->n_flag & NMODIFIED) {
375 mtx_unlock(&np->n_mtx);
376 if (vp->v_type != VREG) {
377 if (vp->v_type != VDIR)
378 panic("nfs: bioread, not dir");
379 (nmp->nm_rpcops->nr_invaldir)(vp);
380 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
385 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
386 error = VOP_GETATTR(vp, &vattr, cred);
389 mtx_lock(&np->n_mtx);
390 np->n_mtime = vattr.va_mtime;
391 mtx_unlock(&np->n_mtx);
393 mtx_unlock(&np->n_mtx);
394 error = VOP_GETATTR(vp, &vattr, cred);
397 mtx_lock(&np->n_mtx);
398 if ((np->n_flag & NSIZECHANGED)
399 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
400 mtx_unlock(&np->n_mtx);
401 if (vp->v_type == VDIR)
402 (nmp->nm_rpcops->nr_invaldir)(vp);
403 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
406 mtx_lock(&np->n_mtx);
407 np->n_mtime = vattr.va_mtime;
408 np->n_flag &= ~NSIZECHANGED;
410 mtx_unlock(&np->n_mtx);
413 nfs_downgrade_vnlock(vp, old_lock);
418 * Vnode op for read using bio
421 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
423 struct nfsnode *np = VTONFS(vp);
425 struct buf *bp, *rabp;
427 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
432 int nra, error = 0, n = 0, on = 0;
434 KASSERT(uio->uio_rw == UIO_READ, ("nfs_read mode"));
435 if (uio->uio_resid == 0)
437 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
441 mtx_lock(&nmp->nm_mtx);
442 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
443 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
444 mtx_unlock(&nmp->nm_mtx);
445 (void)nfs_fsinfo(nmp, vp, cred, td);
447 mtx_unlock(&nmp->nm_mtx);
449 end = uio->uio_offset + uio->uio_resid;
450 if (vp->v_type != VDIR &&
451 (end > nmp->nm_maxfilesize || end < uio->uio_offset))
454 if (nfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
455 /* No caching/ no readaheads. Just read data into the user buffer */
456 return nfs_readrpc(vp, uio, cred);
458 biosize = vp->v_bufobj.bo_bsize;
459 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
461 error = nfs_bioread_check_cons(vp, td, cred);
468 mtx_lock(&np->n_mtx);
470 mtx_unlock(&np->n_mtx);
472 switch (vp->v_type) {
474 nfsstats.biocache_reads++;
475 lbn = uio->uio_offset / biosize;
476 on = uio->uio_offset - (lbn * biosize);
479 * Start the read ahead(s), as required.
481 if (nmp->nm_readahead > 0) {
482 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
483 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
484 rabn = lbn + 1 + nra;
485 if (incore(&vp->v_bufobj, rabn) == NULL) {
486 rabp = nfs_getcacheblk(vp, rabn, biosize, td);
488 error = nfs_sigintr(nmp, td);
489 return (error ? error : EINTR);
491 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
492 rabp->b_flags |= B_ASYNC;
493 rabp->b_iocmd = BIO_READ;
494 vfs_busy_pages(rabp, 0);
495 if (nfs_asyncio(nmp, rabp, cred, td)) {
496 rabp->b_flags |= B_INVAL;
497 rabp->b_ioflags |= BIO_ERROR;
498 vfs_unbusy_pages(rabp);
509 /* Note that bcount is *not* DEV_BSIZE aligned. */
511 if ((off_t)lbn * biosize >= nsize) {
513 } else if ((off_t)(lbn + 1) * biosize > nsize) {
514 bcount = nsize - (off_t)lbn * biosize;
516 bp = nfs_getcacheblk(vp, lbn, bcount, td);
519 error = nfs_sigintr(nmp, td);
520 return (error ? error : EINTR);
524 * If B_CACHE is not set, we must issue the read. If this
525 * fails, we return an error.
528 if ((bp->b_flags & B_CACHE) == 0) {
529 bp->b_iocmd = BIO_READ;
530 vfs_busy_pages(bp, 0);
531 error = nfs_doio(vp, bp, cred, td);
539 * on is the offset into the current bp. Figure out how many
540 * bytes we can copy out of the bp. Note that bcount is
541 * NOT DEV_BSIZE aligned.
543 * Then figure out how many bytes we can copy into the uio.
548 n = MIN((unsigned)(bcount - on), uio->uio_resid);
551 nfsstats.biocache_readlinks++;
552 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
554 error = nfs_sigintr(nmp, td);
555 return (error ? error : EINTR);
557 if ((bp->b_flags & B_CACHE) == 0) {
558 bp->b_iocmd = BIO_READ;
559 vfs_busy_pages(bp, 0);
560 error = nfs_doio(vp, bp, cred, td);
562 bp->b_ioflags |= BIO_ERROR;
567 n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
571 nfsstats.biocache_readdirs++;
572 if (np->n_direofoffset
573 && uio->uio_offset >= np->n_direofoffset) {
576 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
577 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
578 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
580 error = nfs_sigintr(nmp, td);
581 return (error ? error : EINTR);
583 if ((bp->b_flags & B_CACHE) == 0) {
584 bp->b_iocmd = BIO_READ;
585 vfs_busy_pages(bp, 0);
586 error = nfs_doio(vp, bp, cred, td);
590 while (error == NFSERR_BAD_COOKIE) {
591 (nmp->nm_rpcops->nr_invaldir)(vp);
592 error = nfs_vinvalbuf(vp, 0, td, 1);
594 * Yuck! The directory has been modified on the
595 * server. The only way to get the block is by
596 * reading from the beginning to get all the
599 * Leave the last bp intact unless there is an error.
600 * Loop back up to the while if the error is another
601 * NFSERR_BAD_COOKIE (double yuch!).
603 for (i = 0; i <= lbn && !error; i++) {
604 if (np->n_direofoffset
605 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
607 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
609 error = nfs_sigintr(nmp, td);
610 return (error ? error : EINTR);
612 if ((bp->b_flags & B_CACHE) == 0) {
613 bp->b_iocmd = BIO_READ;
614 vfs_busy_pages(bp, 0);
615 error = nfs_doio(vp, bp, cred, td);
617 * no error + B_INVAL == directory EOF,
620 if (error == 0 && (bp->b_flags & B_INVAL))
624 * An error will throw away the block and the
625 * for loop will break out. If no error and this
626 * is not the block we want, we throw away the
627 * block and go for the next one via the for loop.
629 if (error || i < lbn)
634 * The above while is repeated if we hit another cookie
635 * error. If we hit an error and it wasn't a cookie error,
643 * If not eof and read aheads are enabled, start one.
644 * (You need the current block first, so that you have the
645 * directory offset cookie of the next block.)
647 if (nmp->nm_readahead > 0 &&
648 (bp->b_flags & B_INVAL) == 0 &&
649 (np->n_direofoffset == 0 ||
650 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
651 incore(&vp->v_bufobj, lbn + 1) == NULL) {
652 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
654 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
655 rabp->b_flags |= B_ASYNC;
656 rabp->b_iocmd = BIO_READ;
657 vfs_busy_pages(rabp, 0);
658 if (nfs_asyncio(nmp, rabp, cred, td)) {
659 rabp->b_flags |= B_INVAL;
660 rabp->b_ioflags |= BIO_ERROR;
661 vfs_unbusy_pages(rabp);
670 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
671 * chopped for the EOF condition, we cannot tell how large
672 * NFS directories are going to be until we hit EOF. So
673 * an NFS directory buffer is *not* chopped to its EOF. Now,
674 * it just so happens that b_resid will effectively chop it
675 * to EOF. *BUT* this information is lost if the buffer goes
676 * away and is reconstituted into a B_CACHE state ( due to
677 * being VMIO ) later. So we keep track of the directory eof
678 * in np->n_direofoffset and chop it off as an extra step
681 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
682 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
683 n = np->n_direofoffset - uio->uio_offset;
686 nfs_printf(" nfs_bioread: type %x unexpected\n", vp->v_type);
692 error = uiomove(bp->b_data + on, (int)n, uio);
694 if (vp->v_type == VLNK)
698 } while (error == 0 && uio->uio_resid > 0 && n > 0);
703 * The NFS write path cannot handle iovecs with len > 1. So we need to
704 * break up iovecs accordingly (restricting them to wsize).
705 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
706 * For the ASYNC case, 2 copies are needed. The first a copy from the
707 * user buffer to a staging buffer and then a second copy from the staging
708 * buffer to mbufs. This can be optimized by copying from the user buffer
709 * directly into mbufs and passing the chain down, but that requires a
710 * fair amount of re-working of the relevant codepaths (and can be done
714 nfs_directio_write(vp, uiop, cred, ioflag)
721 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
722 struct thread *td = uiop->uio_td;
726 mtx_lock(&nmp->nm_mtx);
727 wsize = nmp->nm_wsize;
728 mtx_unlock(&nmp->nm_mtx);
729 if (ioflag & IO_SYNC) {
730 int iomode, must_commit;
734 while (uiop->uio_resid > 0) {
735 size = MIN(uiop->uio_resid, wsize);
736 size = MIN(uiop->uio_iov->iov_len, size);
737 iov.iov_base = uiop->uio_iov->iov_base;
741 uio.uio_offset = uiop->uio_offset;
742 uio.uio_resid = size;
743 uio.uio_segflg = UIO_USERSPACE;
744 uio.uio_rw = UIO_WRITE;
746 iomode = NFSV3WRITE_FILESYNC;
747 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred,
748 &iomode, &must_commit);
749 KASSERT((must_commit == 0),
750 ("nfs_directio_write: Did not commit write"));
753 uiop->uio_offset += size;
754 uiop->uio_resid -= size;
755 if (uiop->uio_iov->iov_len <= size) {
759 uiop->uio_iov->iov_base =
760 (char *)uiop->uio_iov->iov_base + size;
761 uiop->uio_iov->iov_len -= size;
770 * Break up the write into blocksize chunks and hand these
771 * over to nfsiod's for write back.
772 * Unfortunately, this incurs a copy of the data. Since
773 * the user could modify the buffer before the write is
776 * The obvious optimization here is that one of the 2 copies
777 * in the async write path can be eliminated by copying the
778 * data here directly into mbufs and passing the mbuf chain
779 * down. But that will require a fair amount of re-working
780 * of the code and can be done if there's enough interest
781 * in NFS directio access.
783 while (uiop->uio_resid > 0) {
784 size = MIN(uiop->uio_resid, wsize);
785 size = MIN(uiop->uio_iov->iov_len, size);
786 bp = getpbuf(&nfs_pbuf_freecnt);
787 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
788 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
789 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
790 t_iov->iov_len = size;
791 t_uio->uio_iov = t_iov;
792 t_uio->uio_iovcnt = 1;
793 t_uio->uio_offset = uiop->uio_offset;
794 t_uio->uio_resid = size;
795 t_uio->uio_segflg = UIO_SYSSPACE;
796 t_uio->uio_rw = UIO_WRITE;
798 KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
799 uiop->uio_segflg == UIO_SYSSPACE,
800 ("nfs_directio_write: Bad uio_segflg"));
801 if (uiop->uio_segflg == UIO_USERSPACE) {
802 error = copyin(uiop->uio_iov->iov_base,
803 t_iov->iov_base, size);
808 * UIO_SYSSPACE may never happen, but handle
809 * it just in case it does.
811 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
813 bp->b_flags |= B_DIRECT;
814 bp->b_iocmd = BIO_WRITE;
815 if (cred != NOCRED) {
819 bp->b_wcred = NOCRED;
820 bp->b_caller1 = (void *)t_uio;
822 error = nfs_asyncio(nmp, bp, NOCRED, td);
825 free(t_iov->iov_base, M_NFSDIRECTIO);
826 free(t_iov, M_NFSDIRECTIO);
827 free(t_uio, M_NFSDIRECTIO);
829 relpbuf(bp, &nfs_pbuf_freecnt);
834 uiop->uio_offset += size;
835 uiop->uio_resid -= size;
836 if (uiop->uio_iov->iov_len <= size) {
840 uiop->uio_iov->iov_base =
841 (char *)uiop->uio_iov->iov_base + size;
842 uiop->uio_iov->iov_len -= size;
850 * Vnode op for write using bio
853 nfs_write(struct vop_write_args *ap)
856 struct uio *uio = ap->a_uio;
857 struct thread *td = uio->uio_td;
858 struct vnode *vp = ap->a_vp;
859 struct nfsnode *np = VTONFS(vp);
860 struct ucred *cred = ap->a_cred;
861 int ioflag = ap->a_ioflag;
864 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
868 int n, on, error = 0;
870 KASSERT(uio->uio_rw == UIO_WRITE, ("nfs_write mode"));
871 KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
873 if (vp->v_type != VREG)
875 mtx_lock(&np->n_mtx);
876 if (np->n_flag & NWRITEERR) {
877 np->n_flag &= ~NWRITEERR;
878 mtx_unlock(&np->n_mtx);
879 return (np->n_error);
881 mtx_unlock(&np->n_mtx);
882 mtx_lock(&nmp->nm_mtx);
883 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
884 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
885 mtx_unlock(&nmp->nm_mtx);
886 (void)nfs_fsinfo(nmp, vp, cred, td);
888 mtx_unlock(&nmp->nm_mtx);
891 * Synchronously flush pending buffers if we are in synchronous
892 * mode or if we are appending.
894 if (ioflag & (IO_APPEND | IO_SYNC)) {
895 mtx_lock(&np->n_mtx);
896 if (np->n_flag & NMODIFIED) {
897 mtx_unlock(&np->n_mtx);
898 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
900 * Require non-blocking, synchronous writes to
901 * dirty files to inform the program it needs
902 * to fsync(2) explicitly.
904 if (ioflag & IO_NDELAY)
909 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
910 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
914 mtx_unlock(&np->n_mtx);
918 * If IO_APPEND then load uio_offset. We restart here if we cannot
919 * get the append lock.
921 if (ioflag & IO_APPEND) {
923 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
924 error = VOP_GETATTR(vp, &vattr, cred);
927 mtx_lock(&np->n_mtx);
928 uio->uio_offset = np->n_size;
929 mtx_unlock(&np->n_mtx);
932 if (uio->uio_offset < 0)
934 end = uio->uio_offset + uio->uio_resid;
935 if (end > nmp->nm_maxfilesize || end < uio->uio_offset)
937 if (uio->uio_resid == 0)
940 if (nfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
941 return nfs_directio_write(vp, uio, cred, ioflag);
944 * Maybe this should be above the vnode op call, but so long as
945 * file servers have no limits, i don't think it matters
947 if (vn_rlimit_fsize(vp, uio, td))
950 biosize = vp->v_bufobj.bo_bsize;
952 * Find all of this file's B_NEEDCOMMIT buffers. If our writes
953 * would exceed the local maximum per-file write commit size when
954 * combined with those, we must decide whether to flush,
955 * go synchronous, or return error. We don't bother checking
956 * IO_UNIT -- we just make all writes atomic anyway, as there's
957 * no point optimizing for something that really won't ever happen.
959 if (!(ioflag & IO_SYNC)) {
962 mtx_lock(&np->n_mtx);
964 mtx_unlock(&np->n_mtx);
966 if (nmp->nm_wcommitsize < uio->uio_resid) {
968 * If this request could not possibly be completed
969 * without exceeding the maximum outstanding write
970 * commit size, see if we can convert it into a
971 * synchronous write operation.
973 if (ioflag & IO_NDELAY)
976 if (nflag & NMODIFIED)
978 } else if (nflag & NMODIFIED) {
980 BO_LOCK(&vp->v_bufobj);
981 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
982 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
984 if (bp->b_flags & B_NEEDCOMMIT)
985 wouldcommit += bp->b_bcount;
988 BO_UNLOCK(&vp->v_bufobj);
990 * Since we're not operating synchronously and
991 * bypassing the buffer cache, we are in a commit
992 * and holding all of these buffers whether
993 * transmitted or not. If not limited, this
994 * will lead to the buffer cache deadlocking,
995 * as no one else can flush our uncommitted buffers.
997 wouldcommit += uio->uio_resid;
999 * If we would initially exceed the maximum
1000 * outstanding write commit size, flush and restart.
1002 if (wouldcommit > nmp->nm_wcommitsize)
1006 goto flush_and_restart;
1010 nfsstats.biocache_writes++;
1011 lbn = uio->uio_offset / biosize;
1012 on = uio->uio_offset - (lbn * biosize);
1013 n = MIN((unsigned)(biosize - on), uio->uio_resid);
1016 * Handle direct append and file extension cases, calculate
1017 * unaligned buffer size.
1019 mtx_lock(&np->n_mtx);
1020 if (uio->uio_offset == np->n_size && n) {
1021 mtx_unlock(&np->n_mtx);
1023 * Get the buffer (in its pre-append state to maintain
1024 * B_CACHE if it was previously set). Resize the
1025 * nfsnode after we have locked the buffer to prevent
1026 * readers from reading garbage.
1029 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1034 mtx_lock(&np->n_mtx);
1035 np->n_size = uio->uio_offset + n;
1036 np->n_flag |= NMODIFIED;
1037 vnode_pager_setsize(vp, np->n_size);
1038 mtx_unlock(&np->n_mtx);
1040 save = bp->b_flags & B_CACHE;
1042 allocbuf(bp, bcount);
1043 bp->b_flags |= save;
1047 * Obtain the locked cache block first, and then
1048 * adjust the file's size as appropriate.
1051 if ((off_t)lbn * biosize + bcount < np->n_size) {
1052 if ((off_t)(lbn + 1) * biosize < np->n_size)
1055 bcount = np->n_size - (off_t)lbn * biosize;
1057 mtx_unlock(&np->n_mtx);
1058 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1059 mtx_lock(&np->n_mtx);
1060 if (uio->uio_offset + n > np->n_size) {
1061 np->n_size = uio->uio_offset + n;
1062 np->n_flag |= NMODIFIED;
1063 vnode_pager_setsize(vp, np->n_size);
1065 mtx_unlock(&np->n_mtx);
1069 error = nfs_sigintr(nmp, td);
1076 * Issue a READ if B_CACHE is not set. In special-append
1077 * mode, B_CACHE is based on the buffer prior to the write
1078 * op and is typically set, avoiding the read. If a read
1079 * is required in special append mode, the server will
1080 * probably send us a short-read since we extended the file
1081 * on our end, resulting in b_resid == 0 and, thusly,
1082 * B_CACHE getting set.
1084 * We can also avoid issuing the read if the write covers
1085 * the entire buffer. We have to make sure the buffer state
1086 * is reasonable in this case since we will not be initiating
1087 * I/O. See the comments in kern/vfs_bio.c's getblk() for
1090 * B_CACHE may also be set due to the buffer being cached
1094 if (on == 0 && n == bcount) {
1095 bp->b_flags |= B_CACHE;
1096 bp->b_flags &= ~B_INVAL;
1097 bp->b_ioflags &= ~BIO_ERROR;
1100 if ((bp->b_flags & B_CACHE) == 0) {
1101 bp->b_iocmd = BIO_READ;
1102 vfs_busy_pages(bp, 0);
1103 error = nfs_doio(vp, bp, cred, td);
1109 if (bp->b_wcred == NOCRED)
1110 bp->b_wcred = crhold(cred);
1111 mtx_lock(&np->n_mtx);
1112 np->n_flag |= NMODIFIED;
1113 mtx_unlock(&np->n_mtx);
1116 * If dirtyend exceeds file size, chop it down. This should
1117 * not normally occur but there is an append race where it
1118 * might occur XXX, so we log it.
1120 * If the chopping creates a reverse-indexed or degenerate
1121 * situation with dirtyoff/end, we 0 both of them.
1124 if (bp->b_dirtyend > bcount) {
1125 nfs_printf("NFS append race @%lx:%d\n",
1126 (long)bp->b_blkno * DEV_BSIZE,
1127 bp->b_dirtyend - bcount);
1128 bp->b_dirtyend = bcount;
1131 if (bp->b_dirtyoff >= bp->b_dirtyend)
1132 bp->b_dirtyoff = bp->b_dirtyend = 0;
1135 * If the new write will leave a contiguous dirty
1136 * area, just update the b_dirtyoff and b_dirtyend,
1137 * otherwise force a write rpc of the old dirty area.
1139 * While it is possible to merge discontiguous writes due to
1140 * our having a B_CACHE buffer ( and thus valid read data
1141 * for the hole), we don't because it could lead to
1142 * significant cache coherency problems with multiple clients,
1143 * especially if locking is implemented later on.
1145 * as an optimization we could theoretically maintain
1146 * a linked list of discontinuous areas, but we would still
1147 * have to commit them separately so there isn't much
1148 * advantage to it except perhaps a bit of asynchronization.
1151 if (bp->b_dirtyend > 0 &&
1152 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1153 if (bwrite(bp) == EINTR) {
1160 error = uiomove((char *)bp->b_data + on, n, uio);
1163 * Since this block is being modified, it must be written
1164 * again and not just committed. Since write clustering does
1165 * not work for the stage 1 data write, only the stage 2
1166 * commit rpc, we have to clear B_CLUSTEROK as well.
1168 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1171 bp->b_ioflags |= BIO_ERROR;
1177 * Only update dirtyoff/dirtyend if not a degenerate
1181 if (bp->b_dirtyend > 0) {
1182 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1183 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1185 bp->b_dirtyoff = on;
1186 bp->b_dirtyend = on + n;
1188 vfs_bio_set_valid(bp, on, n);
1192 * If IO_SYNC do bwrite().
1194 * IO_INVAL appears to be unused. The idea appears to be
1195 * to turn off caching in this case. Very odd. XXX
1197 if ((ioflag & IO_SYNC)) {
1198 if (ioflag & IO_INVAL)
1199 bp->b_flags |= B_NOCACHE;
1203 } else if ((n + on) == biosize) {
1204 bp->b_flags |= B_ASYNC;
1205 (void) (nmp->nm_rpcops->nr_writebp)(bp, 0, NULL);
1209 } while (uio->uio_resid > 0 && n > 0);
1215 * Get an nfs cache block.
1217 * Allocate a new one if the block isn't currently in the cache
1218 * and return the block marked busy. If the calling process is
1219 * interrupted by a signal for an interruptible mount point, return
1222 * The caller must carefully deal with the possible B_INVAL state of
1223 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it
1224 * indirectly), so synchronous reads can be issued without worrying about
1225 * the B_INVAL state. We have to be a little more careful when dealing
1226 * with writes (see comments in nfs_write()) when extending a file past
1230 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1234 struct nfsmount *nmp;
1239 if (nmp->nm_flag & NFSMNT_INT) {
1242 nfs_set_sigmask(td, &oldset);
1243 bp = getblk(vp, bn, size, PCATCH, 0, 0);
1244 nfs_restore_sigmask(td, &oldset);
1245 while (bp == NULL) {
1246 if (nfs_sigintr(nmp, td))
1248 bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1251 bp = getblk(vp, bn, size, 0, 0, 0);
1254 if (vp->v_type == VREG)
1255 bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1260 * Flush and invalidate all dirty buffers. If another process is already
1261 * doing the flush, just wait for completion.
1264 nfs_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1266 struct nfsnode *np = VTONFS(vp);
1267 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1268 int error = 0, slpflag, slptimeo;
1271 ASSERT_VOP_LOCKED(vp, "nfs_vinvalbuf");
1273 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1283 old_lock = nfs_upgrade_vnlock(vp);
1284 if (vp->v_iflag & VI_DOOMED) {
1286 * Since vgonel() uses the generic vinvalbuf() to flush
1287 * dirty buffers and it does not call this function, it
1288 * is safe to just return OK when VI_DOOMED is set.
1290 nfs_downgrade_vnlock(vp, old_lock);
1295 * Now, flush as required.
1297 if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) {
1298 VM_OBJECT_WLOCK(vp->v_bufobj.bo_object);
1299 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1300 VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object);
1302 * If the page clean was interrupted, fail the invalidation.
1303 * Not doing so, we run the risk of losing dirty pages in the
1304 * vinvalbuf() call below.
1306 if (intrflg && (error = nfs_sigintr(nmp, td)))
1310 error = vinvalbuf(vp, flags, slpflag, 0);
1312 if (intrflg && (error = nfs_sigintr(nmp, td)))
1314 error = vinvalbuf(vp, flags, 0, slptimeo);
1316 mtx_lock(&np->n_mtx);
1317 if (np->n_directio_asyncwr == 0)
1318 np->n_flag &= ~NMODIFIED;
1319 mtx_unlock(&np->n_mtx);
1321 nfs_downgrade_vnlock(vp, old_lock);
1326 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1327 * This is mainly to avoid queueing async I/O requests when the nfsiods
1328 * are all hung on a dead server.
1330 * Note: nfs_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1331 * is eventually dequeued by the async daemon, nfs_doio() *will*.
1334 nfs_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1343 * Commits are usually short and sweet so lets save some cpu and
1344 * leave the async daemons for more important rpc's (such as reads
1347 * Readdirplus RPCs do vget()s to acquire the vnodes for entries
1348 * in the directory in order to update attributes. This can deadlock
1349 * with another thread that is waiting for async I/O to be done by
1350 * an nfsiod thread while holding a lock on one of these vnodes.
1351 * To avoid this deadlock, don't allow the async nfsiod threads to
1352 * perform Readdirplus RPCs.
1354 mtx_lock(&nfs_iod_mtx);
1355 if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1356 (nmp->nm_bufqiods > nfs_numasync / 2)) ||
1357 (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
1358 mtx_unlock(&nfs_iod_mtx);
1362 if (nmp->nm_flag & NFSMNT_INT)
1367 * Find a free iod to process this request.
1369 for (iod = 0; iod < nfs_numasync; iod++)
1370 if (nfs_iodwant[iod] == NFSIOD_AVAILABLE) {
1376 * Try to create one if none are free.
1382 * Found one, so wake it up and tell it which
1385 NFS_DPF(ASYNCIO, ("nfs_asyncio: waking iod %d for mount %p\n",
1387 nfs_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1388 nfs_iodmount[iod] = nmp;
1390 wakeup(&nfs_iodwant[iod]);
1394 * If none are free, we may already have an iod working on this mount
1395 * point. If so, it will process our request.
1398 if (nmp->nm_bufqiods > 0) {
1400 ("nfs_asyncio: %d iods are already processing mount %p\n",
1401 nmp->nm_bufqiods, nmp));
1407 * If we have an iod which can process the request, then queue
1412 * Ensure that the queue never grows too large. We still want
1413 * to asynchronize so we block rather then return EIO.
1415 while (nmp->nm_bufqlen >= 2 * nfs_numasync) {
1417 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp));
1418 nmp->nm_bufqwant = TRUE;
1419 error = nfs_msleep(td, &nmp->nm_bufq, &nfs_iod_mtx,
1421 "nfsaio", slptimeo);
1423 error2 = nfs_sigintr(nmp, td);
1425 mtx_unlock(&nfs_iod_mtx);
1428 if (slpflag == PCATCH) {
1434 * We might have lost our iod while sleeping,
1435 * so check and loop if nescessary.
1440 /* We might have lost our nfsiod */
1441 if (nmp->nm_bufqiods == 0) {
1443 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1447 if (bp->b_iocmd == BIO_READ) {
1448 if (bp->b_rcred == NOCRED && cred != NOCRED)
1449 bp->b_rcred = crhold(cred);
1451 if (bp->b_wcred == NOCRED && cred != NOCRED)
1452 bp->b_wcred = crhold(cred);
1455 if (bp->b_flags & B_REMFREE)
1458 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1460 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1461 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx);
1462 VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1463 VTONFS(bp->b_vp)->n_directio_asyncwr++;
1464 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx);
1466 mtx_unlock(&nfs_iod_mtx);
1470 mtx_unlock(&nfs_iod_mtx);
1473 * All the iods are busy on other mounts, so return EIO to
1474 * force the caller to process the i/o synchronously.
1476 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n"));
1481 nfs_doio_directwrite(struct buf *bp)
1483 int iomode, must_commit;
1484 struct uio *uiop = (struct uio *)bp->b_caller1;
1485 char *iov_base = uiop->uio_iov->iov_base;
1486 struct nfsmount *nmp = VFSTONFS(bp->b_vp->v_mount);
1488 iomode = NFSV3WRITE_FILESYNC;
1489 uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1490 (nmp->nm_rpcops->nr_writerpc)(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit);
1491 KASSERT((must_commit == 0), ("nfs_doio_directwrite: Did not commit write"));
1492 free(iov_base, M_NFSDIRECTIO);
1493 free(uiop->uio_iov, M_NFSDIRECTIO);
1494 free(uiop, M_NFSDIRECTIO);
1495 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1496 struct nfsnode *np = VTONFS(bp->b_vp);
1497 mtx_lock(&np->n_mtx);
1498 np->n_directio_asyncwr--;
1499 if (np->n_directio_asyncwr == 0) {
1500 VTONFS(bp->b_vp)->n_flag &= ~NMODIFIED;
1501 if ((np->n_flag & NFSYNCWAIT)) {
1502 np->n_flag &= ~NFSYNCWAIT;
1503 wakeup((caddr_t)&np->n_directio_asyncwr);
1506 mtx_unlock(&np->n_mtx);
1509 relpbuf(bp, &nfs_pbuf_freecnt);
1513 * Do an I/O operation to/from a cache block. This may be called
1514 * synchronously or from an nfsiod.
1517 nfs_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td)
1521 struct nfsmount *nmp;
1522 int error = 0, iomode, must_commit = 0;
1525 struct proc *p = td ? td->td_proc : NULL;
1529 nmp = VFSTONFS(vp->v_mount);
1531 uiop->uio_iov = &io;
1532 uiop->uio_iovcnt = 1;
1533 uiop->uio_segflg = UIO_SYSSPACE;
1537 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We
1538 * do this here so we do not have to do it in all the code that
1541 bp->b_flags &= ~B_INVAL;
1542 bp->b_ioflags &= ~BIO_ERROR;
1544 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp));
1545 iocmd = bp->b_iocmd;
1546 if (iocmd == BIO_READ) {
1547 io.iov_len = uiop->uio_resid = bp->b_bcount;
1548 io.iov_base = bp->b_data;
1549 uiop->uio_rw = UIO_READ;
1551 switch (vp->v_type) {
1553 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1554 nfsstats.read_bios++;
1555 error = (nmp->nm_rpcops->nr_readrpc)(vp, uiop, cr);
1558 if (uiop->uio_resid) {
1560 * If we had a short read with no error, we must have
1561 * hit a file hole. We should zero-fill the remainder.
1562 * This can also occur if the server hits the file EOF.
1564 * Holes used to be able to occur due to pending
1565 * writes, but that is not possible any longer.
1567 int nread = bp->b_bcount - uiop->uio_resid;
1568 int left = uiop->uio_resid;
1571 bzero((char *)bp->b_data + nread, left);
1572 uiop->uio_resid = 0;
1575 /* ASSERT_VOP_LOCKED(vp, "nfs_doio"); */
1576 if (p && (vp->v_vflag & VV_TEXT)) {
1577 mtx_lock(&np->n_mtx);
1578 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.va_mtime)) {
1579 mtx_unlock(&np->n_mtx);
1581 killproc(p, "text file modification");
1584 mtx_unlock(&np->n_mtx);
1588 uiop->uio_offset = (off_t)0;
1589 nfsstats.readlink_bios++;
1590 error = (nmp->nm_rpcops->nr_readlinkrpc)(vp, uiop, cr);
1593 nfsstats.readdir_bios++;
1594 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1595 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1596 error = nfs_readdirplusrpc(vp, uiop, cr);
1597 if (error == NFSERR_NOTSUPP)
1598 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1600 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1601 error = nfs_readdirrpc(vp, uiop, cr);
1603 * end-of-directory sets B_INVAL but does not generate an
1606 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1607 bp->b_flags |= B_INVAL;
1610 nfs_printf("nfs_doio: type %x unexpected\n", vp->v_type);
1614 bp->b_ioflags |= BIO_ERROR;
1615 bp->b_error = error;
1619 * If we only need to commit, try to commit
1621 if (bp->b_flags & B_NEEDCOMMIT) {
1625 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1626 retv = (nmp->nm_rpcops->nr_commit)(
1627 vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1630 bp->b_dirtyoff = bp->b_dirtyend = 0;
1631 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1636 if (retv == NFSERR_STALEWRITEVERF) {
1637 nfs_clearcommit(vp->v_mount);
1642 * Setup for actual write
1644 mtx_lock(&np->n_mtx);
1645 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1646 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1647 mtx_unlock(&np->n_mtx);
1649 if (bp->b_dirtyend > bp->b_dirtyoff) {
1650 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1652 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1654 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1655 uiop->uio_rw = UIO_WRITE;
1656 nfsstats.write_bios++;
1658 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1659 iomode = NFSV3WRITE_UNSTABLE;
1661 iomode = NFSV3WRITE_FILESYNC;
1663 error = (nmp->nm_rpcops->nr_writerpc)(vp, uiop, cr, &iomode, &must_commit);
1666 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1667 * to cluster the buffers needing commit. This will allow
1668 * the system to submit a single commit rpc for the whole
1669 * cluster. We can do this even if the buffer is not 100%
1670 * dirty (relative to the NFS blocksize), so we optimize the
1671 * append-to-file-case.
1673 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1674 * cleared because write clustering only works for commit
1675 * rpc's, not for the data portion of the write).
1678 if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1679 bp->b_flags |= B_NEEDCOMMIT;
1680 if (bp->b_dirtyoff == 0
1681 && bp->b_dirtyend == bp->b_bcount)
1682 bp->b_flags |= B_CLUSTEROK;
1684 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1688 * For an interrupted write, the buffer is still valid
1689 * and the write hasn't been pushed to the server yet,
1690 * so we can't set BIO_ERROR and report the interruption
1691 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1692 * is not relevant, so the rpc attempt is essentially
1693 * a noop. For the case of a V3 write rpc not being
1694 * committed to stable storage, the block is still
1695 * dirty and requires either a commit rpc or another
1696 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1697 * the block is reused. This is indicated by setting
1698 * the B_DELWRI and B_NEEDCOMMIT flags.
1700 * If the buffer is marked B_PAGING, it does not reside on
1701 * the vp's paging queues so we cannot call bdirty(). The
1702 * bp in this case is not an NFS cache block so we should
1705 * The logic below breaks up errors into recoverable and
1706 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1707 * and keep the buffer around for potential write retries.
1708 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1709 * and save the error in the nfsnode. This is less than ideal
1710 * but necessary. Keeping such buffers around could potentially
1711 * cause buffer exhaustion eventually (they can never be written
1712 * out, so will get constantly be re-dirtied). It also causes
1713 * all sorts of vfs panics. For non-recoverable write errors,
1714 * also invalidate the attrcache, so we'll be forced to go over
1715 * the wire for this object, returning an error to user on next
1716 * call (most of the time).
1718 if (error == EINTR || error == EIO || error == ETIMEDOUT
1719 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1723 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1724 if ((bp->b_flags & B_PAGING) == 0) {
1726 bp->b_flags &= ~B_DONE;
1728 if (error && (bp->b_flags & B_ASYNC) == 0)
1729 bp->b_flags |= B_EINTR;
1733 bp->b_ioflags |= BIO_ERROR;
1734 bp->b_flags |= B_INVAL;
1735 bp->b_error = np->n_error = error;
1736 mtx_lock(&np->n_mtx);
1737 np->n_flag |= NWRITEERR;
1738 np->n_attrstamp = 0;
1739 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1740 mtx_unlock(&np->n_mtx);
1742 bp->b_dirtyoff = bp->b_dirtyend = 0;
1750 bp->b_resid = uiop->uio_resid;
1752 nfs_clearcommit(vp->v_mount);
1758 * Used to aid in handling ftruncate() operations on the NFS client side.
1759 * Truncation creates a number of special problems for NFS. We have to
1760 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1761 * we have to properly handle VM pages or (potentially dirty) buffers
1762 * that straddle the truncation point.
1766 nfs_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1768 struct nfsnode *np = VTONFS(vp);
1770 int biosize = vp->v_bufobj.bo_bsize;
1773 mtx_lock(&np->n_mtx);
1776 mtx_unlock(&np->n_mtx);
1778 if (nsize < tsize) {
1784 * vtruncbuf() doesn't get the buffer overlapping the
1785 * truncation point. We may have a B_DELWRI and/or B_CACHE
1786 * buffer that now needs to be truncated.
1788 error = vtruncbuf(vp, cred, nsize, biosize);
1789 lbn = nsize / biosize;
1790 bufsize = nsize - (lbn * biosize);
1791 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1794 if (bp->b_dirtyoff > bp->b_bcount)
1795 bp->b_dirtyoff = bp->b_bcount;
1796 if (bp->b_dirtyend > bp->b_bcount)
1797 bp->b_dirtyend = bp->b_bcount;
1798 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1801 vnode_pager_setsize(vp, nsize);