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 "opt_kdtrace.h"
40 #include <sys/param.h>
41 #include <sys/systm.h>
44 #include <sys/kernel.h>
46 #include <sys/mount.h>
48 #include <sys/vmmeter.h>
49 #include <sys/vnode.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 <nfsclient/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_LOCK(object);
131 if (pages[ap->a_reqpage]->valid != 0) {
132 vm_page_lock_queues();
133 for (i = 0; i < npages; ++i) {
134 if (i != ap->a_reqpage)
135 vm_page_free(pages[i]);
137 vm_page_unlock_queues();
138 VM_OBJECT_UNLOCK(object);
141 VM_OBJECT_UNLOCK(object);
144 * We use only the kva address for the buffer, but this is extremely
145 * convienient and fast.
147 bp = getpbuf(&nfs_pbuf_freecnt);
149 kva = (vm_offset_t) bp->b_data;
150 pmap_qenter(kva, pages, npages);
151 PCPU_INC(cnt.v_vnodein);
152 PCPU_ADD(cnt.v_vnodepgsin, npages);
154 iov.iov_base = (caddr_t) kva;
158 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
159 uio.uio_resid = count;
160 uio.uio_segflg = UIO_SYSSPACE;
161 uio.uio_rw = UIO_READ;
164 error = (nmp->nm_rpcops->nr_readrpc)(vp, &uio, cred);
165 pmap_qremove(kva, npages);
167 relpbuf(bp, &nfs_pbuf_freecnt);
169 if (error && (uio.uio_resid == count)) {
170 nfs_printf("nfs_getpages: error %d\n", error);
171 VM_OBJECT_LOCK(object);
172 vm_page_lock_queues();
173 for (i = 0; i < npages; ++i) {
174 if (i != ap->a_reqpage)
175 vm_page_free(pages[i]);
177 vm_page_unlock_queues();
178 VM_OBJECT_UNLOCK(object);
179 return (VM_PAGER_ERROR);
183 * Calculate the number of bytes read and validate only that number
184 * of bytes. Note that due to pending writes, size may be 0. This
185 * does not mean that the remaining data is invalid!
188 size = count - uio.uio_resid;
189 VM_OBJECT_LOCK(object);
190 vm_page_lock_queues();
191 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
193 nextoff = toff + PAGE_SIZE;
196 if (nextoff <= size) {
198 * Read operation filled an entire page
200 m->valid = VM_PAGE_BITS_ALL;
201 KASSERT(m->dirty == 0,
202 ("nfs_getpages: page %p is dirty", m));
203 } else if (size > toff) {
205 * Read operation filled a partial page.
208 vm_page_set_valid(m, 0, size - toff);
209 KASSERT(m->dirty == 0,
210 ("nfs_getpages: page %p is dirty", m));
213 * Read operation was short. If no error occured
214 * we may have hit a zero-fill section. We simply
215 * leave valid set to 0.
219 if (i != ap->a_reqpage) {
221 * Whether or not to leave the page activated is up in
222 * the air, but we should put the page on a page queue
223 * somewhere (it already is in the object). Result:
224 * It appears that emperical results show that
225 * deactivating pages is best.
229 * Just in case someone was asking for this page we
230 * now tell them that it is ok to use.
233 if (m->oflags & VPO_WANTED)
236 vm_page_deactivate(m);
243 vm_page_unlock_queues();
244 VM_OBJECT_UNLOCK(object);
249 * Vnode op for VM putpages.
252 nfs_putpages(struct vop_putpages_args *ap)
258 int iomode, must_commit, i, error, npages, count;
264 struct nfsmount *nmp;
270 td = curthread; /* XXX */
271 /* Set the cred to n_writecred for the write rpcs. */
272 if (np->n_writecred != NULL)
273 cred = crhold(np->n_writecred);
275 cred = crhold(curthread->td_ucred); /* XXX */
276 nmp = VFSTONFS(vp->v_mount);
279 rtvals = ap->a_rtvals;
280 npages = btoc(count);
281 offset = IDX_TO_OFF(pages[0]->pindex);
283 mtx_lock(&nmp->nm_mtx);
284 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
285 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
286 mtx_unlock(&nmp->nm_mtx);
287 (void)nfs_fsinfo(nmp, vp, cred, td);
289 mtx_unlock(&nmp->nm_mtx);
291 mtx_lock(&np->n_mtx);
292 if (nfs_directio_enable && !nfs_directio_allow_mmap &&
293 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
294 mtx_unlock(&np->n_mtx);
295 nfs_printf("nfs_putpages: called on noncache-able vnode??\n");
296 mtx_lock(&np->n_mtx);
299 for (i = 0; i < npages; i++)
300 rtvals[i] = VM_PAGER_ERROR;
303 * When putting pages, do not extend file past EOF.
305 if (offset + count > np->n_size) {
306 count = np->n_size - offset;
310 mtx_unlock(&np->n_mtx);
313 * We use only the kva address for the buffer, but this is extremely
314 * convienient and fast.
316 bp = getpbuf(&nfs_pbuf_freecnt);
318 kva = (vm_offset_t) bp->b_data;
319 pmap_qenter(kva, pages, npages);
320 PCPU_INC(cnt.v_vnodeout);
321 PCPU_ADD(cnt.v_vnodepgsout, count);
323 iov.iov_base = (caddr_t) kva;
327 uio.uio_offset = offset;
328 uio.uio_resid = count;
329 uio.uio_segflg = UIO_SYSSPACE;
330 uio.uio_rw = UIO_WRITE;
333 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
334 iomode = NFSV3WRITE_UNSTABLE;
336 iomode = NFSV3WRITE_FILESYNC;
338 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, &iomode, &must_commit);
341 pmap_qremove(kva, npages);
342 relpbuf(bp, &nfs_pbuf_freecnt);
345 vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid);
347 nfs_clearcommit(vp->v_mount);
354 * For nfs, cache consistency can only be maintained approximately.
355 * Although RFC1094 does not specify the criteria, the following is
356 * believed to be compatible with the reference port.
358 * If the file's modify time on the server has changed since the
359 * last read rpc or you have written to the file,
360 * you may have lost data cache consistency with the
361 * server, so flush all of the file's data out of the cache.
362 * Then force a getattr rpc to ensure that you have up to date
364 * NB: This implies that cache data can be read when up to
365 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
366 * attributes this could be forced by setting n_attrstamp to 0 before
367 * the VOP_GETATTR() call.
370 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
374 struct nfsnode *np = VTONFS(vp);
376 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
379 * Grab the exclusive lock before checking whether the cache is
381 * XXX - We can make this cheaper later (by acquiring cheaper locks).
382 * But for now, this suffices.
384 old_lock = nfs_upgrade_vnlock(vp);
385 if (vp->v_iflag & VI_DOOMED) {
386 nfs_downgrade_vnlock(vp, old_lock);
390 mtx_lock(&np->n_mtx);
391 if (np->n_flag & NMODIFIED) {
392 mtx_unlock(&np->n_mtx);
393 if (vp->v_type != VREG) {
394 if (vp->v_type != VDIR)
395 panic("nfs: bioread, not dir");
396 (nmp->nm_rpcops->nr_invaldir)(vp);
397 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
402 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
403 error = VOP_GETATTR(vp, &vattr, cred);
406 mtx_lock(&np->n_mtx);
407 np->n_mtime = vattr.va_mtime;
408 mtx_unlock(&np->n_mtx);
410 mtx_unlock(&np->n_mtx);
411 error = VOP_GETATTR(vp, &vattr, cred);
414 mtx_lock(&np->n_mtx);
415 if ((np->n_flag & NSIZECHANGED)
416 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
417 mtx_unlock(&np->n_mtx);
418 if (vp->v_type == VDIR)
419 (nmp->nm_rpcops->nr_invaldir)(vp);
420 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
423 mtx_lock(&np->n_mtx);
424 np->n_mtime = vattr.va_mtime;
425 np->n_flag &= ~NSIZECHANGED;
427 mtx_unlock(&np->n_mtx);
430 nfs_downgrade_vnlock(vp, old_lock);
435 * Vnode op for read using bio
438 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
440 struct nfsnode *np = VTONFS(vp);
442 struct buf *bp, *rabp;
444 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
449 int nra, error = 0, n = 0, on = 0;
451 KASSERT(uio->uio_rw == UIO_READ, ("nfs_read mode"));
452 if (uio->uio_resid == 0)
454 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
458 mtx_lock(&nmp->nm_mtx);
459 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
460 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
461 mtx_unlock(&nmp->nm_mtx);
462 (void)nfs_fsinfo(nmp, vp, cred, td);
464 mtx_unlock(&nmp->nm_mtx);
466 end = uio->uio_offset + uio->uio_resid;
467 if (vp->v_type != VDIR &&
468 (end > nmp->nm_maxfilesize || end < uio->uio_offset))
471 if (nfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
472 /* No caching/ no readaheads. Just read data into the user buffer */
473 return nfs_readrpc(vp, uio, cred);
475 biosize = vp->v_bufobj.bo_bsize;
476 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
478 error = nfs_bioread_check_cons(vp, td, cred);
485 mtx_lock(&np->n_mtx);
487 mtx_unlock(&np->n_mtx);
489 switch (vp->v_type) {
491 nfsstats.biocache_reads++;
492 lbn = uio->uio_offset / biosize;
493 on = uio->uio_offset & (biosize - 1);
496 * Start the read ahead(s), as required.
498 if (nmp->nm_readahead > 0) {
499 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
500 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
501 rabn = lbn + 1 + nra;
502 if (incore(&vp->v_bufobj, rabn) == NULL) {
503 rabp = nfs_getcacheblk(vp, rabn, biosize, td);
505 error = nfs_sigintr(nmp, td);
506 return (error ? error : EINTR);
508 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
509 rabp->b_flags |= B_ASYNC;
510 rabp->b_iocmd = BIO_READ;
511 vfs_busy_pages(rabp, 0);
512 if (nfs_asyncio(nmp, rabp, cred, td)) {
513 rabp->b_flags |= B_INVAL;
514 rabp->b_ioflags |= BIO_ERROR;
515 vfs_unbusy_pages(rabp);
526 /* Note that bcount is *not* DEV_BSIZE aligned. */
528 if ((off_t)lbn * biosize >= nsize) {
530 } else if ((off_t)(lbn + 1) * biosize > nsize) {
531 bcount = nsize - (off_t)lbn * biosize;
533 bp = nfs_getcacheblk(vp, lbn, bcount, td);
536 error = nfs_sigintr(nmp, td);
537 return (error ? error : EINTR);
541 * If B_CACHE is not set, we must issue the read. If this
542 * fails, we return an error.
545 if ((bp->b_flags & B_CACHE) == 0) {
546 bp->b_iocmd = BIO_READ;
547 vfs_busy_pages(bp, 0);
548 error = nfs_doio(vp, bp, cred, td);
556 * on is the offset into the current bp. Figure out how many
557 * bytes we can copy out of the bp. Note that bcount is
558 * NOT DEV_BSIZE aligned.
560 * Then figure out how many bytes we can copy into the uio.
565 n = min((unsigned)(bcount - on), uio->uio_resid);
568 nfsstats.biocache_readlinks++;
569 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
571 error = nfs_sigintr(nmp, td);
572 return (error ? error : EINTR);
574 if ((bp->b_flags & B_CACHE) == 0) {
575 bp->b_iocmd = BIO_READ;
576 vfs_busy_pages(bp, 0);
577 error = nfs_doio(vp, bp, cred, td);
579 bp->b_ioflags |= BIO_ERROR;
584 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
588 nfsstats.biocache_readdirs++;
589 if (np->n_direofoffset
590 && uio->uio_offset >= np->n_direofoffset) {
593 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
594 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
595 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
597 error = nfs_sigintr(nmp, td);
598 return (error ? error : EINTR);
600 if ((bp->b_flags & B_CACHE) == 0) {
601 bp->b_iocmd = BIO_READ;
602 vfs_busy_pages(bp, 0);
603 error = nfs_doio(vp, bp, cred, td);
607 while (error == NFSERR_BAD_COOKIE) {
608 (nmp->nm_rpcops->nr_invaldir)(vp);
609 error = nfs_vinvalbuf(vp, 0, td, 1);
611 * Yuck! The directory has been modified on the
612 * server. The only way to get the block is by
613 * reading from the beginning to get all the
616 * Leave the last bp intact unless there is an error.
617 * Loop back up to the while if the error is another
618 * NFSERR_BAD_COOKIE (double yuch!).
620 for (i = 0; i <= lbn && !error; i++) {
621 if (np->n_direofoffset
622 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
624 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
626 error = nfs_sigintr(nmp, td);
627 return (error ? error : EINTR);
629 if ((bp->b_flags & B_CACHE) == 0) {
630 bp->b_iocmd = BIO_READ;
631 vfs_busy_pages(bp, 0);
632 error = nfs_doio(vp, bp, cred, td);
634 * no error + B_INVAL == directory EOF,
637 if (error == 0 && (bp->b_flags & B_INVAL))
641 * An error will throw away the block and the
642 * for loop will break out. If no error and this
643 * is not the block we want, we throw away the
644 * block and go for the next one via the for loop.
646 if (error || i < lbn)
651 * The above while is repeated if we hit another cookie
652 * error. If we hit an error and it wasn't a cookie error,
660 * If not eof and read aheads are enabled, start one.
661 * (You need the current block first, so that you have the
662 * directory offset cookie of the next block.)
664 if (nmp->nm_readahead > 0 &&
665 (bp->b_flags & B_INVAL) == 0 &&
666 (np->n_direofoffset == 0 ||
667 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
668 incore(&vp->v_bufobj, lbn + 1) == NULL) {
669 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
671 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
672 rabp->b_flags |= B_ASYNC;
673 rabp->b_iocmd = BIO_READ;
674 vfs_busy_pages(rabp, 0);
675 if (nfs_asyncio(nmp, rabp, cred, td)) {
676 rabp->b_flags |= B_INVAL;
677 rabp->b_ioflags |= BIO_ERROR;
678 vfs_unbusy_pages(rabp);
687 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
688 * chopped for the EOF condition, we cannot tell how large
689 * NFS directories are going to be until we hit EOF. So
690 * an NFS directory buffer is *not* chopped to its EOF. Now,
691 * it just so happens that b_resid will effectively chop it
692 * to EOF. *BUT* this information is lost if the buffer goes
693 * away and is reconstituted into a B_CACHE state ( due to
694 * being VMIO ) later. So we keep track of the directory eof
695 * in np->n_direofoffset and chop it off as an extra step
698 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
699 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
700 n = np->n_direofoffset - uio->uio_offset;
703 nfs_printf(" nfs_bioread: type %x unexpected\n", vp->v_type);
709 error = uiomove(bp->b_data + on, (int)n, uio);
711 if (vp->v_type == VLNK)
715 } while (error == 0 && uio->uio_resid > 0 && n > 0);
720 * The NFS write path cannot handle iovecs with len > 1. So we need to
721 * break up iovecs accordingly (restricting them to wsize).
722 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
723 * For the ASYNC case, 2 copies are needed. The first a copy from the
724 * user buffer to a staging buffer and then a second copy from the staging
725 * buffer to mbufs. This can be optimized by copying from the user buffer
726 * directly into mbufs and passing the chain down, but that requires a
727 * fair amount of re-working of the relevant codepaths (and can be done
731 nfs_directio_write(vp, uiop, cred, ioflag)
738 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
739 struct thread *td = uiop->uio_td;
743 mtx_lock(&nmp->nm_mtx);
744 wsize = nmp->nm_wsize;
745 mtx_unlock(&nmp->nm_mtx);
746 if (ioflag & IO_SYNC) {
747 int iomode, must_commit;
751 while (uiop->uio_resid > 0) {
752 size = min(uiop->uio_resid, wsize);
753 size = min(uiop->uio_iov->iov_len, size);
754 iov.iov_base = uiop->uio_iov->iov_base;
758 uio.uio_offset = uiop->uio_offset;
759 uio.uio_resid = size;
760 uio.uio_segflg = UIO_USERSPACE;
761 uio.uio_rw = UIO_WRITE;
763 iomode = NFSV3WRITE_FILESYNC;
764 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred,
765 &iomode, &must_commit);
766 KASSERT((must_commit == 0),
767 ("nfs_directio_write: Did not commit write"));
770 uiop->uio_offset += size;
771 uiop->uio_resid -= size;
772 if (uiop->uio_iov->iov_len <= size) {
776 uiop->uio_iov->iov_base =
777 (char *)uiop->uio_iov->iov_base + size;
778 uiop->uio_iov->iov_len -= size;
787 * Break up the write into blocksize chunks and hand these
788 * over to nfsiod's for write back.
789 * Unfortunately, this incurs a copy of the data. Since
790 * the user could modify the buffer before the write is
793 * The obvious optimization here is that one of the 2 copies
794 * in the async write path can be eliminated by copying the
795 * data here directly into mbufs and passing the mbuf chain
796 * down. But that will require a fair amount of re-working
797 * of the code and can be done if there's enough interest
798 * in NFS directio access.
800 while (uiop->uio_resid > 0) {
801 size = min(uiop->uio_resid, wsize);
802 size = min(uiop->uio_iov->iov_len, size);
803 bp = getpbuf(&nfs_pbuf_freecnt);
804 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
805 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
806 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
807 t_iov->iov_len = size;
808 t_uio->uio_iov = t_iov;
809 t_uio->uio_iovcnt = 1;
810 t_uio->uio_offset = uiop->uio_offset;
811 t_uio->uio_resid = size;
812 t_uio->uio_segflg = UIO_SYSSPACE;
813 t_uio->uio_rw = UIO_WRITE;
815 KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
816 uiop->uio_segflg == UIO_SYSSPACE,
817 ("nfs_directio_write: Bad uio_segflg"));
818 if (uiop->uio_segflg == UIO_USERSPACE) {
819 error = copyin(uiop->uio_iov->iov_base,
820 t_iov->iov_base, size);
825 * UIO_SYSSPACE may never happen, but handle
826 * it just in case it does.
828 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
830 bp->b_flags |= B_DIRECT;
831 bp->b_iocmd = BIO_WRITE;
832 if (cred != NOCRED) {
836 bp->b_wcred = NOCRED;
837 bp->b_caller1 = (void *)t_uio;
839 error = nfs_asyncio(nmp, bp, NOCRED, td);
842 free(t_iov->iov_base, M_NFSDIRECTIO);
843 free(t_iov, M_NFSDIRECTIO);
844 free(t_uio, M_NFSDIRECTIO);
846 relpbuf(bp, &nfs_pbuf_freecnt);
851 uiop->uio_offset += size;
852 uiop->uio_resid -= size;
853 if (uiop->uio_iov->iov_len <= size) {
857 uiop->uio_iov->iov_base =
858 (char *)uiop->uio_iov->iov_base + size;
859 uiop->uio_iov->iov_len -= size;
867 * Vnode op for write using bio
870 nfs_write(struct vop_write_args *ap)
873 struct uio *uio = ap->a_uio;
874 struct thread *td = uio->uio_td;
875 struct vnode *vp = ap->a_vp;
876 struct nfsnode *np = VTONFS(vp);
877 struct ucred *cred = ap->a_cred;
878 int ioflag = ap->a_ioflag;
881 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
885 int n, on, error = 0;
887 KASSERT(uio->uio_rw == UIO_WRITE, ("nfs_write mode"));
888 KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
890 if (vp->v_type != VREG)
892 mtx_lock(&np->n_mtx);
893 if (np->n_flag & NWRITEERR) {
894 np->n_flag &= ~NWRITEERR;
895 mtx_unlock(&np->n_mtx);
896 return (np->n_error);
898 mtx_unlock(&np->n_mtx);
899 mtx_lock(&nmp->nm_mtx);
900 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
901 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
902 mtx_unlock(&nmp->nm_mtx);
903 (void)nfs_fsinfo(nmp, vp, cred, td);
905 mtx_unlock(&nmp->nm_mtx);
908 * Synchronously flush pending buffers if we are in synchronous
909 * mode or if we are appending.
911 if (ioflag & (IO_APPEND | IO_SYNC)) {
912 mtx_lock(&np->n_mtx);
913 if (np->n_flag & NMODIFIED) {
914 mtx_unlock(&np->n_mtx);
915 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
917 * Require non-blocking, synchronous writes to
918 * dirty files to inform the program it needs
919 * to fsync(2) explicitly.
921 if (ioflag & IO_NDELAY)
926 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
927 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
931 mtx_unlock(&np->n_mtx);
935 * If IO_APPEND then load uio_offset. We restart here if we cannot
936 * get the append lock.
938 if (ioflag & IO_APPEND) {
940 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
941 error = VOP_GETATTR(vp, &vattr, cred);
944 mtx_lock(&np->n_mtx);
945 uio->uio_offset = np->n_size;
946 mtx_unlock(&np->n_mtx);
949 if (uio->uio_offset < 0)
951 end = uio->uio_offset + uio->uio_resid;
952 if (end > nmp->nm_maxfilesize || end < uio->uio_offset)
954 if (uio->uio_resid == 0)
957 if (nfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
958 return nfs_directio_write(vp, uio, cred, ioflag);
961 * Maybe this should be above the vnode op call, but so long as
962 * file servers have no limits, i don't think it matters
964 if (vn_rlimit_fsize(vp, uio, td))
967 biosize = vp->v_bufobj.bo_bsize;
969 * Find all of this file's B_NEEDCOMMIT buffers. If our writes
970 * would exceed the local maximum per-file write commit size when
971 * combined with those, we must decide whether to flush,
972 * go synchronous, or return error. We don't bother checking
973 * IO_UNIT -- we just make all writes atomic anyway, as there's
974 * no point optimizing for something that really won't ever happen.
976 if (!(ioflag & IO_SYNC)) {
979 mtx_lock(&np->n_mtx);
981 mtx_unlock(&np->n_mtx);
983 if (nmp->nm_wcommitsize < uio->uio_resid) {
985 * If this request could not possibly be completed
986 * without exceeding the maximum outstanding write
987 * commit size, see if we can convert it into a
988 * synchronous write operation.
990 if (ioflag & IO_NDELAY)
993 if (nflag & NMODIFIED)
995 } else if (nflag & NMODIFIED) {
997 BO_LOCK(&vp->v_bufobj);
998 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
999 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
1001 if (bp->b_flags & B_NEEDCOMMIT)
1002 wouldcommit += bp->b_bcount;
1005 BO_UNLOCK(&vp->v_bufobj);
1007 * Since we're not operating synchronously and
1008 * bypassing the buffer cache, we are in a commit
1009 * and holding all of these buffers whether
1010 * transmitted or not. If not limited, this
1011 * will lead to the buffer cache deadlocking,
1012 * as no one else can flush our uncommitted buffers.
1014 wouldcommit += uio->uio_resid;
1016 * If we would initially exceed the maximum
1017 * outstanding write commit size, flush and restart.
1019 if (wouldcommit > nmp->nm_wcommitsize)
1023 goto flush_and_restart;
1027 nfsstats.biocache_writes++;
1028 lbn = uio->uio_offset / biosize;
1029 on = uio->uio_offset & (biosize-1);
1030 n = min((unsigned)(biosize - on), uio->uio_resid);
1033 * Handle direct append and file extension cases, calculate
1034 * unaligned buffer size.
1036 mtx_lock(&np->n_mtx);
1037 if (uio->uio_offset == np->n_size && n) {
1038 mtx_unlock(&np->n_mtx);
1040 * Get the buffer (in its pre-append state to maintain
1041 * B_CACHE if it was previously set). Resize the
1042 * nfsnode after we have locked the buffer to prevent
1043 * readers from reading garbage.
1046 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1051 mtx_lock(&np->n_mtx);
1052 np->n_size = uio->uio_offset + n;
1053 np->n_flag |= NMODIFIED;
1054 vnode_pager_setsize(vp, np->n_size);
1055 mtx_unlock(&np->n_mtx);
1057 save = bp->b_flags & B_CACHE;
1059 allocbuf(bp, bcount);
1060 bp->b_flags |= save;
1064 * Obtain the locked cache block first, and then
1065 * adjust the file's size as appropriate.
1068 if ((off_t)lbn * biosize + bcount < np->n_size) {
1069 if ((off_t)(lbn + 1) * biosize < np->n_size)
1072 bcount = np->n_size - (off_t)lbn * biosize;
1074 mtx_unlock(&np->n_mtx);
1075 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1076 mtx_lock(&np->n_mtx);
1077 if (uio->uio_offset + n > np->n_size) {
1078 np->n_size = uio->uio_offset + n;
1079 np->n_flag |= NMODIFIED;
1080 vnode_pager_setsize(vp, np->n_size);
1082 mtx_unlock(&np->n_mtx);
1086 error = nfs_sigintr(nmp, td);
1093 * Issue a READ if B_CACHE is not set. In special-append
1094 * mode, B_CACHE is based on the buffer prior to the write
1095 * op and is typically set, avoiding the read. If a read
1096 * is required in special append mode, the server will
1097 * probably send us a short-read since we extended the file
1098 * on our end, resulting in b_resid == 0 and, thusly,
1099 * B_CACHE getting set.
1101 * We can also avoid issuing the read if the write covers
1102 * the entire buffer. We have to make sure the buffer state
1103 * is reasonable in this case since we will not be initiating
1104 * I/O. See the comments in kern/vfs_bio.c's getblk() for
1107 * B_CACHE may also be set due to the buffer being cached
1111 if (on == 0 && n == bcount) {
1112 bp->b_flags |= B_CACHE;
1113 bp->b_flags &= ~B_INVAL;
1114 bp->b_ioflags &= ~BIO_ERROR;
1117 if ((bp->b_flags & B_CACHE) == 0) {
1118 bp->b_iocmd = BIO_READ;
1119 vfs_busy_pages(bp, 0);
1120 error = nfs_doio(vp, bp, cred, td);
1126 if (bp->b_wcred == NOCRED)
1127 bp->b_wcred = crhold(cred);
1128 mtx_lock(&np->n_mtx);
1129 np->n_flag |= NMODIFIED;
1130 mtx_unlock(&np->n_mtx);
1133 * If dirtyend exceeds file size, chop it down. This should
1134 * not normally occur but there is an append race where it
1135 * might occur XXX, so we log it.
1137 * If the chopping creates a reverse-indexed or degenerate
1138 * situation with dirtyoff/end, we 0 both of them.
1141 if (bp->b_dirtyend > bcount) {
1142 nfs_printf("NFS append race @%lx:%d\n",
1143 (long)bp->b_blkno * DEV_BSIZE,
1144 bp->b_dirtyend - bcount);
1145 bp->b_dirtyend = bcount;
1148 if (bp->b_dirtyoff >= bp->b_dirtyend)
1149 bp->b_dirtyoff = bp->b_dirtyend = 0;
1152 * If the new write will leave a contiguous dirty
1153 * area, just update the b_dirtyoff and b_dirtyend,
1154 * otherwise force a write rpc of the old dirty area.
1156 * While it is possible to merge discontiguous writes due to
1157 * our having a B_CACHE buffer ( and thus valid read data
1158 * for the hole), we don't because it could lead to
1159 * significant cache coherency problems with multiple clients,
1160 * especially if locking is implemented later on.
1162 * as an optimization we could theoretically maintain
1163 * a linked list of discontinuous areas, but we would still
1164 * have to commit them separately so there isn't much
1165 * advantage to it except perhaps a bit of asynchronization.
1168 if (bp->b_dirtyend > 0 &&
1169 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1170 if (bwrite(bp) == EINTR) {
1177 error = uiomove((char *)bp->b_data + on, n, uio);
1180 * Since this block is being modified, it must be written
1181 * again and not just committed. Since write clustering does
1182 * not work for the stage 1 data write, only the stage 2
1183 * commit rpc, we have to clear B_CLUSTEROK as well.
1185 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1188 bp->b_ioflags |= BIO_ERROR;
1194 * Only update dirtyoff/dirtyend if not a degenerate
1198 if (bp->b_dirtyend > 0) {
1199 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1200 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1202 bp->b_dirtyoff = on;
1203 bp->b_dirtyend = on + n;
1205 vfs_bio_set_valid(bp, on, n);
1209 * If IO_SYNC do bwrite().
1211 * IO_INVAL appears to be unused. The idea appears to be
1212 * to turn off caching in this case. Very odd. XXX
1214 if ((ioflag & IO_SYNC)) {
1215 if (ioflag & IO_INVAL)
1216 bp->b_flags |= B_NOCACHE;
1220 } else if ((n + on) == biosize) {
1221 bp->b_flags |= B_ASYNC;
1222 (void) (nmp->nm_rpcops->nr_writebp)(bp, 0, NULL);
1226 } while (uio->uio_resid > 0 && n > 0);
1232 * Get an nfs cache block.
1234 * Allocate a new one if the block isn't currently in the cache
1235 * and return the block marked busy. If the calling process is
1236 * interrupted by a signal for an interruptible mount point, return
1239 * The caller must carefully deal with the possible B_INVAL state of
1240 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it
1241 * indirectly), so synchronous reads can be issued without worrying about
1242 * the B_INVAL state. We have to be a little more careful when dealing
1243 * with writes (see comments in nfs_write()) when extending a file past
1247 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1251 struct nfsmount *nmp;
1256 if (nmp->nm_flag & NFSMNT_INT) {
1259 nfs_set_sigmask(td, &oldset);
1260 bp = getblk(vp, bn, size, NFS_PCATCH, 0, 0);
1261 nfs_restore_sigmask(td, &oldset);
1262 while (bp == NULL) {
1263 if (nfs_sigintr(nmp, td))
1265 bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1268 bp = getblk(vp, bn, size, 0, 0, 0);
1271 if (vp->v_type == VREG)
1272 bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1277 * Flush and invalidate all dirty buffers. If another process is already
1278 * doing the flush, just wait for completion.
1281 nfs_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1283 struct nfsnode *np = VTONFS(vp);
1284 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1285 int error = 0, slpflag, slptimeo;
1288 ASSERT_VOP_LOCKED(vp, "nfs_vinvalbuf");
1290 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1293 slpflag = NFS_PCATCH;
1300 old_lock = nfs_upgrade_vnlock(vp);
1301 if (vp->v_iflag & VI_DOOMED) {
1303 * Since vgonel() uses the generic vinvalbuf() to flush
1304 * dirty buffers and it does not call this function, it
1305 * is safe to just return OK when VI_DOOMED is set.
1307 nfs_downgrade_vnlock(vp, old_lock);
1312 * Now, flush as required.
1314 if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) {
1315 VM_OBJECT_LOCK(vp->v_bufobj.bo_object);
1316 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1317 VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object);
1319 * If the page clean was interrupted, fail the invalidation.
1320 * Not doing so, we run the risk of losing dirty pages in the
1321 * vinvalbuf() call below.
1323 if (intrflg && (error = nfs_sigintr(nmp, td)))
1327 error = vinvalbuf(vp, flags, slpflag, 0);
1329 if (intrflg && (error = nfs_sigintr(nmp, td)))
1331 error = vinvalbuf(vp, flags, 0, slptimeo);
1333 mtx_lock(&np->n_mtx);
1334 if (np->n_directio_asyncwr == 0)
1335 np->n_flag &= ~NMODIFIED;
1336 mtx_unlock(&np->n_mtx);
1338 nfs_downgrade_vnlock(vp, old_lock);
1343 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1344 * This is mainly to avoid queueing async I/O requests when the nfsiods
1345 * are all hung on a dead server.
1347 * Note: nfs_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1348 * is eventually dequeued by the async daemon, nfs_doio() *will*.
1351 nfs_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1360 * Commits are usually short and sweet so lets save some cpu and
1361 * leave the async daemons for more important rpc's (such as reads
1364 * Readdirplus RPCs do vget()s to acquire the vnodes for entries
1365 * in the directory in order to update attributes. This can deadlock
1366 * with another thread that is waiting for async I/O to be done by
1367 * an nfsiod thread while holding a lock on one of these vnodes.
1368 * To avoid this deadlock, don't allow the async nfsiod threads to
1369 * perform Readdirplus RPCs.
1371 mtx_lock(&nfs_iod_mtx);
1372 if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1373 (nmp->nm_bufqiods > nfs_numasync / 2)) ||
1374 (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
1375 mtx_unlock(&nfs_iod_mtx);
1379 if (nmp->nm_flag & NFSMNT_INT)
1380 slpflag = NFS_PCATCH;
1384 * Find a free iod to process this request.
1386 for (iod = 0; iod < nfs_numasync; iod++)
1387 if (nfs_iodwant[iod] == NFSIOD_AVAILABLE) {
1393 * Try to create one if none are free.
1399 * Found one, so wake it up and tell it which
1402 NFS_DPF(ASYNCIO, ("nfs_asyncio: waking iod %d for mount %p\n",
1404 nfs_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1405 nfs_iodmount[iod] = nmp;
1407 wakeup(&nfs_iodwant[iod]);
1411 * If none are free, we may already have an iod working on this mount
1412 * point. If so, it will process our request.
1415 if (nmp->nm_bufqiods > 0) {
1417 ("nfs_asyncio: %d iods are already processing mount %p\n",
1418 nmp->nm_bufqiods, nmp));
1424 * If we have an iod which can process the request, then queue
1429 * Ensure that the queue never grows too large. We still want
1430 * to asynchronize so we block rather then return EIO.
1432 while (nmp->nm_bufqlen >= 2 * nfs_numasync) {
1434 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp));
1435 nmp->nm_bufqwant = TRUE;
1436 error = nfs_msleep(td, &nmp->nm_bufq, &nfs_iod_mtx,
1438 "nfsaio", slptimeo);
1440 error2 = nfs_sigintr(nmp, td);
1442 mtx_unlock(&nfs_iod_mtx);
1445 if (slpflag == NFS_PCATCH) {
1451 * We might have lost our iod while sleeping,
1452 * so check and loop if nescessary.
1457 /* We might have lost our nfsiod */
1458 if (nmp->nm_bufqiods == 0) {
1460 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1464 if (bp->b_iocmd == BIO_READ) {
1465 if (bp->b_rcred == NOCRED && cred != NOCRED)
1466 bp->b_rcred = crhold(cred);
1468 if (bp->b_wcred == NOCRED && cred != NOCRED)
1469 bp->b_wcred = crhold(cred);
1472 if (bp->b_flags & B_REMFREE)
1475 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1477 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1478 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx);
1479 VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1480 VTONFS(bp->b_vp)->n_directio_asyncwr++;
1481 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx);
1483 mtx_unlock(&nfs_iod_mtx);
1487 mtx_unlock(&nfs_iod_mtx);
1490 * All the iods are busy on other mounts, so return EIO to
1491 * force the caller to process the i/o synchronously.
1493 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n"));
1498 nfs_doio_directwrite(struct buf *bp)
1500 int iomode, must_commit;
1501 struct uio *uiop = (struct uio *)bp->b_caller1;
1502 char *iov_base = uiop->uio_iov->iov_base;
1503 struct nfsmount *nmp = VFSTONFS(bp->b_vp->v_mount);
1505 iomode = NFSV3WRITE_FILESYNC;
1506 uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1507 (nmp->nm_rpcops->nr_writerpc)(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit);
1508 KASSERT((must_commit == 0), ("nfs_doio_directwrite: Did not commit write"));
1509 free(iov_base, M_NFSDIRECTIO);
1510 free(uiop->uio_iov, M_NFSDIRECTIO);
1511 free(uiop, M_NFSDIRECTIO);
1512 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1513 struct nfsnode *np = VTONFS(bp->b_vp);
1514 mtx_lock(&np->n_mtx);
1515 np->n_directio_asyncwr--;
1516 if (np->n_directio_asyncwr == 0) {
1517 VTONFS(bp->b_vp)->n_flag &= ~NMODIFIED;
1518 if ((np->n_flag & NFSYNCWAIT)) {
1519 np->n_flag &= ~NFSYNCWAIT;
1520 wakeup((caddr_t)&np->n_directio_asyncwr);
1523 mtx_unlock(&np->n_mtx);
1526 relpbuf(bp, &nfs_pbuf_freecnt);
1530 * Do an I/O operation to/from a cache block. This may be called
1531 * synchronously or from an nfsiod.
1534 nfs_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td)
1538 struct nfsmount *nmp;
1539 int error = 0, iomode, must_commit = 0;
1542 struct proc *p = td ? td->td_proc : NULL;
1546 nmp = VFSTONFS(vp->v_mount);
1548 uiop->uio_iov = &io;
1549 uiop->uio_iovcnt = 1;
1550 uiop->uio_segflg = UIO_SYSSPACE;
1554 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We
1555 * do this here so we do not have to do it in all the code that
1558 bp->b_flags &= ~B_INVAL;
1559 bp->b_ioflags &= ~BIO_ERROR;
1561 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp));
1562 iocmd = bp->b_iocmd;
1563 if (iocmd == BIO_READ) {
1564 io.iov_len = uiop->uio_resid = bp->b_bcount;
1565 io.iov_base = bp->b_data;
1566 uiop->uio_rw = UIO_READ;
1568 switch (vp->v_type) {
1570 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1571 nfsstats.read_bios++;
1572 error = (nmp->nm_rpcops->nr_readrpc)(vp, uiop, cr);
1575 if (uiop->uio_resid) {
1577 * If we had a short read with no error, we must have
1578 * hit a file hole. We should zero-fill the remainder.
1579 * This can also occur if the server hits the file EOF.
1581 * Holes used to be able to occur due to pending
1582 * writes, but that is not possible any longer.
1584 int nread = bp->b_bcount - uiop->uio_resid;
1585 int left = uiop->uio_resid;
1588 bzero((char *)bp->b_data + nread, left);
1589 uiop->uio_resid = 0;
1592 /* ASSERT_VOP_LOCKED(vp, "nfs_doio"); */
1593 if (p && (vp->v_vflag & VV_TEXT)) {
1594 mtx_lock(&np->n_mtx);
1595 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.va_mtime)) {
1596 mtx_unlock(&np->n_mtx);
1598 killproc(p, "text file modification");
1601 mtx_unlock(&np->n_mtx);
1605 uiop->uio_offset = (off_t)0;
1606 nfsstats.readlink_bios++;
1607 error = (nmp->nm_rpcops->nr_readlinkrpc)(vp, uiop, cr);
1610 nfsstats.readdir_bios++;
1611 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1612 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1613 error = nfs_readdirplusrpc(vp, uiop, cr);
1614 if (error == NFSERR_NOTSUPP)
1615 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1617 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1618 error = nfs_readdirrpc(vp, uiop, cr);
1620 * end-of-directory sets B_INVAL but does not generate an
1623 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1624 bp->b_flags |= B_INVAL;
1627 nfs_printf("nfs_doio: type %x unexpected\n", vp->v_type);
1631 bp->b_ioflags |= BIO_ERROR;
1632 bp->b_error = error;
1636 * If we only need to commit, try to commit
1638 if (bp->b_flags & B_NEEDCOMMIT) {
1642 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1643 retv = (nmp->nm_rpcops->nr_commit)(
1644 vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1647 bp->b_dirtyoff = bp->b_dirtyend = 0;
1648 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1653 if (retv == NFSERR_STALEWRITEVERF) {
1654 nfs_clearcommit(vp->v_mount);
1659 * Setup for actual write
1661 mtx_lock(&np->n_mtx);
1662 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1663 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1664 mtx_unlock(&np->n_mtx);
1666 if (bp->b_dirtyend > bp->b_dirtyoff) {
1667 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1669 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1671 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1672 uiop->uio_rw = UIO_WRITE;
1673 nfsstats.write_bios++;
1675 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1676 iomode = NFSV3WRITE_UNSTABLE;
1678 iomode = NFSV3WRITE_FILESYNC;
1680 error = (nmp->nm_rpcops->nr_writerpc)(vp, uiop, cr, &iomode, &must_commit);
1683 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1684 * to cluster the buffers needing commit. This will allow
1685 * the system to submit a single commit rpc for the whole
1686 * cluster. We can do this even if the buffer is not 100%
1687 * dirty (relative to the NFS blocksize), so we optimize the
1688 * append-to-file-case.
1690 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1691 * cleared because write clustering only works for commit
1692 * rpc's, not for the data portion of the write).
1695 if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1696 bp->b_flags |= B_NEEDCOMMIT;
1697 if (bp->b_dirtyoff == 0
1698 && bp->b_dirtyend == bp->b_bcount)
1699 bp->b_flags |= B_CLUSTEROK;
1701 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1705 * For an interrupted write, the buffer is still valid
1706 * and the write hasn't been pushed to the server yet,
1707 * so we can't set BIO_ERROR and report the interruption
1708 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1709 * is not relevant, so the rpc attempt is essentially
1710 * a noop. For the case of a V3 write rpc not being
1711 * committed to stable storage, the block is still
1712 * dirty and requires either a commit rpc or another
1713 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1714 * the block is reused. This is indicated by setting
1715 * the B_DELWRI and B_NEEDCOMMIT flags.
1717 * If the buffer is marked B_PAGING, it does not reside on
1718 * the vp's paging queues so we cannot call bdirty(). The
1719 * bp in this case is not an NFS cache block so we should
1722 * The logic below breaks up errors into recoverable and
1723 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1724 * and keep the buffer around for potential write retries.
1725 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1726 * and save the error in the nfsnode. This is less than ideal
1727 * but necessary. Keeping such buffers around could potentially
1728 * cause buffer exhaustion eventually (they can never be written
1729 * out, so will get constantly be re-dirtied). It also causes
1730 * all sorts of vfs panics. For non-recoverable write errors,
1731 * also invalidate the attrcache, so we'll be forced to go over
1732 * the wire for this object, returning an error to user on next
1733 * call (most of the time).
1735 if (error == EINTR || error == EIO || error == ETIMEDOUT
1736 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1740 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1741 if ((bp->b_flags & B_PAGING) == 0) {
1743 bp->b_flags &= ~B_DONE;
1745 if (error && (bp->b_flags & B_ASYNC) == 0)
1746 bp->b_flags |= B_EINTR;
1750 bp->b_ioflags |= BIO_ERROR;
1751 bp->b_flags |= B_INVAL;
1752 bp->b_error = np->n_error = error;
1753 mtx_lock(&np->n_mtx);
1754 np->n_flag |= NWRITEERR;
1755 np->n_attrstamp = 0;
1756 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1757 mtx_unlock(&np->n_mtx);
1759 bp->b_dirtyoff = bp->b_dirtyend = 0;
1767 bp->b_resid = uiop->uio_resid;
1769 nfs_clearcommit(vp->v_mount);
1775 * Used to aid in handling ftruncate() operations on the NFS client side.
1776 * Truncation creates a number of special problems for NFS. We have to
1777 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1778 * we have to properly handle VM pages or (potentially dirty) buffers
1779 * that straddle the truncation point.
1783 nfs_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1785 struct nfsnode *np = VTONFS(vp);
1787 int biosize = vp->v_bufobj.bo_bsize;
1790 mtx_lock(&np->n_mtx);
1793 mtx_unlock(&np->n_mtx);
1795 if (nsize < tsize) {
1801 * vtruncbuf() doesn't get the buffer overlapping the
1802 * truncation point. We may have a B_DELWRI and/or B_CACHE
1803 * buffer that now needs to be truncated.
1805 error = vtruncbuf(vp, cred, td, nsize, biosize);
1806 lbn = nsize / biosize;
1807 bufsize = nsize & (biosize - 1);
1808 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1811 if (bp->b_dirtyoff > bp->b_bcount)
1812 bp->b_dirtyoff = bp->b_bcount;
1813 if (bp->b_dirtyend > bp->b_bcount)
1814 bp->b_dirtyend = bp->b_bcount;
1815 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1818 vnode_pager_setsize(vp, nsize);