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 <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_LOCK(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_UNLOCK(object);
142 VM_OBJECT_UNLOCK(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_LOCK(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_UNLOCK(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_LOCK(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(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 occured
215 * we may have hit a zero-fill section. We simply
216 * leave valid set to 0.
220 if (i != ap->a_reqpage) {
222 * Whether or not to leave the page activated is up in
223 * the air, but we should put the page on a page queue
224 * somewhere (it already is in the object). Result:
225 * It appears that emperical results show that
226 * deactivating pages is best.
230 * Just in case someone was asking for this page we
231 * now tell them that it is ok to use.
234 if (m->oflags & VPO_WANTED) {
240 vm_page_deactivate(m);
251 VM_OBJECT_UNLOCK(object);
256 * Vnode op for VM putpages.
259 nfs_putpages(struct vop_putpages_args *ap)
265 int iomode, must_commit, i, error, npages, count;
271 struct nfsmount *nmp;
277 td = curthread; /* XXX */
278 cred = curthread->td_ucred; /* XXX */
279 nmp = VFSTONFS(vp->v_mount);
282 rtvals = ap->a_rtvals;
283 npages = btoc(count);
284 offset = IDX_TO_OFF(pages[0]->pindex);
286 mtx_lock(&nmp->nm_mtx);
287 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
288 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
289 mtx_unlock(&nmp->nm_mtx);
290 (void)nfs_fsinfo(nmp, vp, cred, td);
292 mtx_unlock(&nmp->nm_mtx);
294 mtx_lock(&np->n_mtx);
295 if (nfs_directio_enable && !nfs_directio_allow_mmap &&
296 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
297 mtx_unlock(&np->n_mtx);
298 nfs_printf("nfs_putpages: called on noncache-able vnode??\n");
299 mtx_lock(&np->n_mtx);
302 for (i = 0; i < npages; i++)
303 rtvals[i] = VM_PAGER_ERROR;
306 * When putting pages, do not extend file past EOF.
308 if (offset + count > np->n_size) {
309 count = np->n_size - offset;
313 mtx_unlock(&np->n_mtx);
316 * We use only the kva address for the buffer, but this is extremely
317 * convienient and fast.
319 bp = getpbuf(&nfs_pbuf_freecnt);
321 kva = (vm_offset_t) bp->b_data;
322 pmap_qenter(kva, pages, npages);
323 PCPU_INC(cnt.v_vnodeout);
324 PCPU_ADD(cnt.v_vnodepgsout, count);
326 iov.iov_base = (caddr_t) kva;
330 uio.uio_offset = offset;
331 uio.uio_resid = count;
332 uio.uio_segflg = UIO_SYSSPACE;
333 uio.uio_rw = UIO_WRITE;
336 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
337 iomode = NFSV3WRITE_UNSTABLE;
339 iomode = NFSV3WRITE_FILESYNC;
341 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, &iomode, &must_commit);
343 pmap_qremove(kva, npages);
344 relpbuf(bp, &nfs_pbuf_freecnt);
347 vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid);
349 nfs_clearcommit(vp->v_mount);
356 * For nfs, cache consistency can only be maintained approximately.
357 * Although RFC1094 does not specify the criteria, the following is
358 * believed to be compatible with the reference port.
360 * If the file's modify time on the server has changed since the
361 * last read rpc or you have written to the file,
362 * you may have lost data cache consistency with the
363 * server, so flush all of the file's data out of the cache.
364 * Then force a getattr rpc to ensure that you have up to date
366 * NB: This implies that cache data can be read when up to
367 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
368 * attributes this could be forced by setting n_attrstamp to 0 before
369 * the VOP_GETATTR() call.
372 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
376 struct nfsnode *np = VTONFS(vp);
378 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
381 * Grab the exclusive lock before checking whether the cache is
383 * XXX - We can make this cheaper later (by acquiring cheaper locks).
384 * But for now, this suffices.
386 old_lock = nfs_upgrade_vnlock(vp);
387 if (vp->v_iflag & VI_DOOMED) {
388 nfs_downgrade_vnlock(vp, old_lock);
392 mtx_lock(&np->n_mtx);
393 if (np->n_flag & NMODIFIED) {
394 mtx_unlock(&np->n_mtx);
395 if (vp->v_type != VREG) {
396 if (vp->v_type != VDIR)
397 panic("nfs: bioread, not dir");
398 (nmp->nm_rpcops->nr_invaldir)(vp);
399 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
404 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
405 error = VOP_GETATTR(vp, &vattr, cred);
408 mtx_lock(&np->n_mtx);
409 np->n_mtime = vattr.va_mtime;
410 mtx_unlock(&np->n_mtx);
412 mtx_unlock(&np->n_mtx);
413 error = VOP_GETATTR(vp, &vattr, cred);
416 mtx_lock(&np->n_mtx);
417 if ((np->n_flag & NSIZECHANGED)
418 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
419 mtx_unlock(&np->n_mtx);
420 if (vp->v_type == VDIR)
421 (nmp->nm_rpcops->nr_invaldir)(vp);
422 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
425 mtx_lock(&np->n_mtx);
426 np->n_mtime = vattr.va_mtime;
427 np->n_flag &= ~NSIZECHANGED;
429 mtx_unlock(&np->n_mtx);
432 nfs_downgrade_vnlock(vp, old_lock);
437 * Vnode op for read using bio
440 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
442 struct nfsnode *np = VTONFS(vp);
444 struct buf *bp, *rabp;
446 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
451 int nra, error = 0, n = 0, on = 0;
453 KASSERT(uio->uio_rw == UIO_READ, ("nfs_read mode"));
454 if (uio->uio_resid == 0)
456 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
460 mtx_lock(&nmp->nm_mtx);
461 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
462 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
463 mtx_unlock(&nmp->nm_mtx);
464 (void)nfs_fsinfo(nmp, vp, cred, td);
466 mtx_unlock(&nmp->nm_mtx);
468 end = uio->uio_offset + uio->uio_resid;
469 if (vp->v_type != VDIR &&
470 (end > nmp->nm_maxfilesize || end < uio->uio_offset))
473 if (nfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
474 /* No caching/ no readaheads. Just read data into the user buffer */
475 return nfs_readrpc(vp, uio, cred);
477 biosize = vp->v_bufobj.bo_bsize;
478 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
480 error = nfs_bioread_check_cons(vp, td, cred);
487 mtx_lock(&np->n_mtx);
489 mtx_unlock(&np->n_mtx);
491 switch (vp->v_type) {
493 nfsstats.biocache_reads++;
494 lbn = uio->uio_offset / biosize;
495 on = uio->uio_offset & (biosize - 1);
498 * Start the read ahead(s), as required.
500 if (nmp->nm_readahead > 0) {
501 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
502 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
503 rabn = lbn + 1 + nra;
504 if (incore(&vp->v_bufobj, rabn) == NULL) {
505 rabp = nfs_getcacheblk(vp, rabn, biosize, td);
507 error = nfs_sigintr(nmp, td);
508 return (error ? error : EINTR);
510 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
511 rabp->b_flags |= B_ASYNC;
512 rabp->b_iocmd = BIO_READ;
513 vfs_busy_pages(rabp, 0);
514 if (nfs_asyncio(nmp, rabp, cred, td)) {
515 rabp->b_flags |= B_INVAL;
516 rabp->b_ioflags |= BIO_ERROR;
517 vfs_unbusy_pages(rabp);
528 /* Note that bcount is *not* DEV_BSIZE aligned. */
530 if ((off_t)lbn * biosize >= nsize) {
532 } else if ((off_t)(lbn + 1) * biosize > nsize) {
533 bcount = nsize - (off_t)lbn * biosize;
535 bp = nfs_getcacheblk(vp, lbn, bcount, td);
538 error = nfs_sigintr(nmp, td);
539 return (error ? error : EINTR);
543 * If B_CACHE is not set, we must issue the read. If this
544 * fails, we return an error.
547 if ((bp->b_flags & B_CACHE) == 0) {
548 bp->b_iocmd = BIO_READ;
549 vfs_busy_pages(bp, 0);
550 error = nfs_doio(vp, bp, cred, td);
558 * on is the offset into the current bp. Figure out how many
559 * bytes we can copy out of the bp. Note that bcount is
560 * NOT DEV_BSIZE aligned.
562 * Then figure out how many bytes we can copy into the uio.
567 n = MIN((unsigned)(bcount - on), uio->uio_resid);
570 nfsstats.biocache_readlinks++;
571 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
573 error = nfs_sigintr(nmp, td);
574 return (error ? error : EINTR);
576 if ((bp->b_flags & B_CACHE) == 0) {
577 bp->b_iocmd = BIO_READ;
578 vfs_busy_pages(bp, 0);
579 error = nfs_doio(vp, bp, cred, td);
581 bp->b_ioflags |= BIO_ERROR;
586 n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
590 nfsstats.biocache_readdirs++;
591 if (np->n_direofoffset
592 && uio->uio_offset >= np->n_direofoffset) {
595 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
596 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
597 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
599 error = nfs_sigintr(nmp, td);
600 return (error ? error : EINTR);
602 if ((bp->b_flags & B_CACHE) == 0) {
603 bp->b_iocmd = BIO_READ;
604 vfs_busy_pages(bp, 0);
605 error = nfs_doio(vp, bp, cred, td);
609 while (error == NFSERR_BAD_COOKIE) {
610 (nmp->nm_rpcops->nr_invaldir)(vp);
611 error = nfs_vinvalbuf(vp, 0, td, 1);
613 * Yuck! The directory has been modified on the
614 * server. The only way to get the block is by
615 * reading from the beginning to get all the
618 * Leave the last bp intact unless there is an error.
619 * Loop back up to the while if the error is another
620 * NFSERR_BAD_COOKIE (double yuch!).
622 for (i = 0; i <= lbn && !error; i++) {
623 if (np->n_direofoffset
624 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
626 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
628 error = nfs_sigintr(nmp, td);
629 return (error ? error : EINTR);
631 if ((bp->b_flags & B_CACHE) == 0) {
632 bp->b_iocmd = BIO_READ;
633 vfs_busy_pages(bp, 0);
634 error = nfs_doio(vp, bp, cred, td);
636 * no error + B_INVAL == directory EOF,
639 if (error == 0 && (bp->b_flags & B_INVAL))
643 * An error will throw away the block and the
644 * for loop will break out. If no error and this
645 * is not the block we want, we throw away the
646 * block and go for the next one via the for loop.
648 if (error || i < lbn)
653 * The above while is repeated if we hit another cookie
654 * error. If we hit an error and it wasn't a cookie error,
662 * If not eof and read aheads are enabled, start one.
663 * (You need the current block first, so that you have the
664 * directory offset cookie of the next block.)
666 if (nmp->nm_readahead > 0 &&
667 (bp->b_flags & B_INVAL) == 0 &&
668 (np->n_direofoffset == 0 ||
669 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
670 incore(&vp->v_bufobj, lbn + 1) == NULL) {
671 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
673 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
674 rabp->b_flags |= B_ASYNC;
675 rabp->b_iocmd = BIO_READ;
676 vfs_busy_pages(rabp, 0);
677 if (nfs_asyncio(nmp, rabp, cred, td)) {
678 rabp->b_flags |= B_INVAL;
679 rabp->b_ioflags |= BIO_ERROR;
680 vfs_unbusy_pages(rabp);
689 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
690 * chopped for the EOF condition, we cannot tell how large
691 * NFS directories are going to be until we hit EOF. So
692 * an NFS directory buffer is *not* chopped to its EOF. Now,
693 * it just so happens that b_resid will effectively chop it
694 * to EOF. *BUT* this information is lost if the buffer goes
695 * away and is reconstituted into a B_CACHE state ( due to
696 * being VMIO ) later. So we keep track of the directory eof
697 * in np->n_direofoffset and chop it off as an extra step
700 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
701 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
702 n = np->n_direofoffset - uio->uio_offset;
705 nfs_printf(" nfs_bioread: type %x unexpected\n", vp->v_type);
711 error = uiomove(bp->b_data + on, (int)n, uio);
713 if (vp->v_type == VLNK)
717 } while (error == 0 && uio->uio_resid > 0 && n > 0);
722 * The NFS write path cannot handle iovecs with len > 1. So we need to
723 * break up iovecs accordingly (restricting them to wsize).
724 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
725 * For the ASYNC case, 2 copies are needed. The first a copy from the
726 * user buffer to a staging buffer and then a second copy from the staging
727 * buffer to mbufs. This can be optimized by copying from the user buffer
728 * directly into mbufs and passing the chain down, but that requires a
729 * fair amount of re-working of the relevant codepaths (and can be done
733 nfs_directio_write(vp, uiop, cred, ioflag)
740 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
741 struct thread *td = uiop->uio_td;
745 mtx_lock(&nmp->nm_mtx);
746 wsize = nmp->nm_wsize;
747 mtx_unlock(&nmp->nm_mtx);
748 if (ioflag & IO_SYNC) {
749 int iomode, must_commit;
753 while (uiop->uio_resid > 0) {
754 size = MIN(uiop->uio_resid, wsize);
755 size = MIN(uiop->uio_iov->iov_len, size);
756 iov.iov_base = uiop->uio_iov->iov_base;
760 uio.uio_offset = uiop->uio_offset;
761 uio.uio_resid = size;
762 uio.uio_segflg = UIO_USERSPACE;
763 uio.uio_rw = UIO_WRITE;
765 iomode = NFSV3WRITE_FILESYNC;
766 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred,
767 &iomode, &must_commit);
768 KASSERT((must_commit == 0),
769 ("nfs_directio_write: Did not commit write"));
772 uiop->uio_offset += size;
773 uiop->uio_resid -= size;
774 if (uiop->uio_iov->iov_len <= size) {
778 uiop->uio_iov->iov_base =
779 (char *)uiop->uio_iov->iov_base + size;
780 uiop->uio_iov->iov_len -= size;
789 * Break up the write into blocksize chunks and hand these
790 * over to nfsiod's for write back.
791 * Unfortunately, this incurs a copy of the data. Since
792 * the user could modify the buffer before the write is
795 * The obvious optimization here is that one of the 2 copies
796 * in the async write path can be eliminated by copying the
797 * data here directly into mbufs and passing the mbuf chain
798 * down. But that will require a fair amount of re-working
799 * of the code and can be done if there's enough interest
800 * in NFS directio access.
802 while (uiop->uio_resid > 0) {
803 size = MIN(uiop->uio_resid, wsize);
804 size = MIN(uiop->uio_iov->iov_len, size);
805 bp = getpbuf(&nfs_pbuf_freecnt);
806 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
807 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
808 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
809 t_iov->iov_len = size;
810 t_uio->uio_iov = t_iov;
811 t_uio->uio_iovcnt = 1;
812 t_uio->uio_offset = uiop->uio_offset;
813 t_uio->uio_resid = size;
814 t_uio->uio_segflg = UIO_SYSSPACE;
815 t_uio->uio_rw = UIO_WRITE;
817 KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
818 uiop->uio_segflg == UIO_SYSSPACE,
819 ("nfs_directio_write: Bad uio_segflg"));
820 if (uiop->uio_segflg == UIO_USERSPACE) {
821 error = copyin(uiop->uio_iov->iov_base,
822 t_iov->iov_base, size);
827 * UIO_SYSSPACE may never happen, but handle
828 * it just in case it does.
830 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
832 bp->b_flags |= B_DIRECT;
833 bp->b_iocmd = BIO_WRITE;
834 if (cred != NOCRED) {
838 bp->b_wcred = NOCRED;
839 bp->b_caller1 = (void *)t_uio;
841 error = nfs_asyncio(nmp, bp, NOCRED, td);
844 free(t_iov->iov_base, M_NFSDIRECTIO);
845 free(t_iov, M_NFSDIRECTIO);
846 free(t_uio, M_NFSDIRECTIO);
848 relpbuf(bp, &nfs_pbuf_freecnt);
853 uiop->uio_offset += size;
854 uiop->uio_resid -= size;
855 if (uiop->uio_iov->iov_len <= size) {
859 uiop->uio_iov->iov_base =
860 (char *)uiop->uio_iov->iov_base + size;
861 uiop->uio_iov->iov_len -= size;
869 * Vnode op for write using bio
872 nfs_write(struct vop_write_args *ap)
875 struct uio *uio = ap->a_uio;
876 struct thread *td = uio->uio_td;
877 struct vnode *vp = ap->a_vp;
878 struct nfsnode *np = VTONFS(vp);
879 struct ucred *cred = ap->a_cred;
880 int ioflag = ap->a_ioflag;
883 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
887 int n, on, error = 0;
889 KASSERT(uio->uio_rw == UIO_WRITE, ("nfs_write mode"));
890 KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
892 if (vp->v_type != VREG)
894 mtx_lock(&np->n_mtx);
895 if (np->n_flag & NWRITEERR) {
896 np->n_flag &= ~NWRITEERR;
897 mtx_unlock(&np->n_mtx);
898 return (np->n_error);
900 mtx_unlock(&np->n_mtx);
901 mtx_lock(&nmp->nm_mtx);
902 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
903 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
904 mtx_unlock(&nmp->nm_mtx);
905 (void)nfs_fsinfo(nmp, vp, cred, td);
907 mtx_unlock(&nmp->nm_mtx);
910 * Synchronously flush pending buffers if we are in synchronous
911 * mode or if we are appending.
913 if (ioflag & (IO_APPEND | IO_SYNC)) {
914 mtx_lock(&np->n_mtx);
915 if (np->n_flag & NMODIFIED) {
916 mtx_unlock(&np->n_mtx);
917 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
919 * Require non-blocking, synchronous writes to
920 * dirty files to inform the program it needs
921 * to fsync(2) explicitly.
923 if (ioflag & IO_NDELAY)
928 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
929 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
933 mtx_unlock(&np->n_mtx);
937 * If IO_APPEND then load uio_offset. We restart here if we cannot
938 * get the append lock.
940 if (ioflag & IO_APPEND) {
942 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
943 error = VOP_GETATTR(vp, &vattr, cred);
946 mtx_lock(&np->n_mtx);
947 uio->uio_offset = np->n_size;
948 mtx_unlock(&np->n_mtx);
951 if (uio->uio_offset < 0)
953 end = uio->uio_offset + uio->uio_resid;
954 if (end > nmp->nm_maxfilesize || end < uio->uio_offset)
956 if (uio->uio_resid == 0)
959 if (nfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
960 return nfs_directio_write(vp, uio, cred, ioflag);
963 * Maybe this should be above the vnode op call, but so long as
964 * file servers have no limits, i don't think it matters
966 if (vn_rlimit_fsize(vp, uio, td))
969 biosize = vp->v_bufobj.bo_bsize;
971 * Find all of this file's B_NEEDCOMMIT buffers. If our writes
972 * would exceed the local maximum per-file write commit size when
973 * combined with those, we must decide whether to flush,
974 * go synchronous, or return error. We don't bother checking
975 * IO_UNIT -- we just make all writes atomic anyway, as there's
976 * no point optimizing for something that really won't ever happen.
978 if (!(ioflag & IO_SYNC)) {
981 mtx_lock(&np->n_mtx);
983 mtx_unlock(&np->n_mtx);
985 if (nmp->nm_wcommitsize < uio->uio_resid) {
987 * If this request could not possibly be completed
988 * without exceeding the maximum outstanding write
989 * commit size, see if we can convert it into a
990 * synchronous write operation.
992 if (ioflag & IO_NDELAY)
995 if (nflag & NMODIFIED)
997 } else if (nflag & NMODIFIED) {
999 BO_LOCK(&vp->v_bufobj);
1000 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
1001 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
1003 if (bp->b_flags & B_NEEDCOMMIT)
1004 wouldcommit += bp->b_bcount;
1007 BO_UNLOCK(&vp->v_bufobj);
1009 * Since we're not operating synchronously and
1010 * bypassing the buffer cache, we are in a commit
1011 * and holding all of these buffers whether
1012 * transmitted or not. If not limited, this
1013 * will lead to the buffer cache deadlocking,
1014 * as no one else can flush our uncommitted buffers.
1016 wouldcommit += uio->uio_resid;
1018 * If we would initially exceed the maximum
1019 * outstanding write commit size, flush and restart.
1021 if (wouldcommit > nmp->nm_wcommitsize)
1025 goto flush_and_restart;
1029 nfsstats.biocache_writes++;
1030 lbn = uio->uio_offset / biosize;
1031 on = uio->uio_offset & (biosize-1);
1032 n = MIN((unsigned)(biosize - on), uio->uio_resid);
1035 * Handle direct append and file extension cases, calculate
1036 * unaligned buffer size.
1038 mtx_lock(&np->n_mtx);
1039 if (uio->uio_offset == np->n_size && n) {
1040 mtx_unlock(&np->n_mtx);
1042 * Get the buffer (in its pre-append state to maintain
1043 * B_CACHE if it was previously set). Resize the
1044 * nfsnode after we have locked the buffer to prevent
1045 * readers from reading garbage.
1048 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1053 mtx_lock(&np->n_mtx);
1054 np->n_size = uio->uio_offset + n;
1055 np->n_flag |= NMODIFIED;
1056 vnode_pager_setsize(vp, np->n_size);
1057 mtx_unlock(&np->n_mtx);
1059 save = bp->b_flags & B_CACHE;
1061 allocbuf(bp, bcount);
1062 bp->b_flags |= save;
1066 * Obtain the locked cache block first, and then
1067 * adjust the file's size as appropriate.
1070 if ((off_t)lbn * biosize + bcount < np->n_size) {
1071 if ((off_t)(lbn + 1) * biosize < np->n_size)
1074 bcount = np->n_size - (off_t)lbn * biosize;
1076 mtx_unlock(&np->n_mtx);
1077 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1078 mtx_lock(&np->n_mtx);
1079 if (uio->uio_offset + n > np->n_size) {
1080 np->n_size = uio->uio_offset + n;
1081 np->n_flag |= NMODIFIED;
1082 vnode_pager_setsize(vp, np->n_size);
1084 mtx_unlock(&np->n_mtx);
1088 error = nfs_sigintr(nmp, td);
1095 * Issue a READ if B_CACHE is not set. In special-append
1096 * mode, B_CACHE is based on the buffer prior to the write
1097 * op and is typically set, avoiding the read. If a read
1098 * is required in special append mode, the server will
1099 * probably send us a short-read since we extended the file
1100 * on our end, resulting in b_resid == 0 and, thusly,
1101 * B_CACHE getting set.
1103 * We can also avoid issuing the read if the write covers
1104 * the entire buffer. We have to make sure the buffer state
1105 * is reasonable in this case since we will not be initiating
1106 * I/O. See the comments in kern/vfs_bio.c's getblk() for
1109 * B_CACHE may also be set due to the buffer being cached
1113 if (on == 0 && n == bcount) {
1114 bp->b_flags |= B_CACHE;
1115 bp->b_flags &= ~B_INVAL;
1116 bp->b_ioflags &= ~BIO_ERROR;
1119 if ((bp->b_flags & B_CACHE) == 0) {
1120 bp->b_iocmd = BIO_READ;
1121 vfs_busy_pages(bp, 0);
1122 error = nfs_doio(vp, bp, cred, td);
1128 if (bp->b_wcred == NOCRED)
1129 bp->b_wcred = crhold(cred);
1130 mtx_lock(&np->n_mtx);
1131 np->n_flag |= NMODIFIED;
1132 mtx_unlock(&np->n_mtx);
1135 * If dirtyend exceeds file size, chop it down. This should
1136 * not normally occur but there is an append race where it
1137 * might occur XXX, so we log it.
1139 * If the chopping creates a reverse-indexed or degenerate
1140 * situation with dirtyoff/end, we 0 both of them.
1143 if (bp->b_dirtyend > bcount) {
1144 nfs_printf("NFS append race @%lx:%d\n",
1145 (long)bp->b_blkno * DEV_BSIZE,
1146 bp->b_dirtyend - bcount);
1147 bp->b_dirtyend = bcount;
1150 if (bp->b_dirtyoff >= bp->b_dirtyend)
1151 bp->b_dirtyoff = bp->b_dirtyend = 0;
1154 * If the new write will leave a contiguous dirty
1155 * area, just update the b_dirtyoff and b_dirtyend,
1156 * otherwise force a write rpc of the old dirty area.
1158 * While it is possible to merge discontiguous writes due to
1159 * our having a B_CACHE buffer ( and thus valid read data
1160 * for the hole), we don't because it could lead to
1161 * significant cache coherency problems with multiple clients,
1162 * especially if locking is implemented later on.
1164 * as an optimization we could theoretically maintain
1165 * a linked list of discontinuous areas, but we would still
1166 * have to commit them separately so there isn't much
1167 * advantage to it except perhaps a bit of asynchronization.
1170 if (bp->b_dirtyend > 0 &&
1171 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1172 if (bwrite(bp) == EINTR) {
1179 error = uiomove((char *)bp->b_data + on, n, uio);
1182 * Since this block is being modified, it must be written
1183 * again and not just committed. Since write clustering does
1184 * not work for the stage 1 data write, only the stage 2
1185 * commit rpc, we have to clear B_CLUSTEROK as well.
1187 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1190 bp->b_ioflags |= BIO_ERROR;
1196 * Only update dirtyoff/dirtyend if not a degenerate
1200 if (bp->b_dirtyend > 0) {
1201 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1202 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1204 bp->b_dirtyoff = on;
1205 bp->b_dirtyend = on + n;
1207 vfs_bio_set_valid(bp, on, n);
1211 * If IO_SYNC do bwrite().
1213 * IO_INVAL appears to be unused. The idea appears to be
1214 * to turn off caching in this case. Very odd. XXX
1216 if ((ioflag & IO_SYNC)) {
1217 if (ioflag & IO_INVAL)
1218 bp->b_flags |= B_NOCACHE;
1222 } else if ((n + on) == biosize) {
1223 bp->b_flags |= B_ASYNC;
1224 (void) (nmp->nm_rpcops->nr_writebp)(bp, 0, NULL);
1228 } while (uio->uio_resid > 0 && n > 0);
1234 * Get an nfs cache block.
1236 * Allocate a new one if the block isn't currently in the cache
1237 * and return the block marked busy. If the calling process is
1238 * interrupted by a signal for an interruptible mount point, return
1241 * The caller must carefully deal with the possible B_INVAL state of
1242 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it
1243 * indirectly), so synchronous reads can be issued without worrying about
1244 * the B_INVAL state. We have to be a little more careful when dealing
1245 * with writes (see comments in nfs_write()) when extending a file past
1249 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1253 struct nfsmount *nmp;
1258 if (nmp->nm_flag & NFSMNT_INT) {
1261 nfs_set_sigmask(td, &oldset);
1262 bp = getblk(vp, bn, size, NFS_PCATCH, 0, 0);
1263 nfs_restore_sigmask(td, &oldset);
1264 while (bp == NULL) {
1265 if (nfs_sigintr(nmp, td))
1267 bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1270 bp = getblk(vp, bn, size, 0, 0, 0);
1273 if (vp->v_type == VREG)
1274 bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1279 * Flush and invalidate all dirty buffers. If another process is already
1280 * doing the flush, just wait for completion.
1283 nfs_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1285 struct nfsnode *np = VTONFS(vp);
1286 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1287 int error = 0, slpflag, slptimeo;
1290 ASSERT_VOP_LOCKED(vp, "nfs_vinvalbuf");
1292 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1295 slpflag = NFS_PCATCH;
1302 old_lock = nfs_upgrade_vnlock(vp);
1303 if (vp->v_iflag & VI_DOOMED) {
1305 * Since vgonel() uses the generic vinvalbuf() to flush
1306 * dirty buffers and it does not call this function, it
1307 * is safe to just return OK when VI_DOOMED is set.
1309 nfs_downgrade_vnlock(vp, old_lock);
1314 * Now, flush as required.
1316 if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) {
1317 VM_OBJECT_LOCK(vp->v_bufobj.bo_object);
1318 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1319 VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object);
1321 * If the page clean was interrupted, fail the invalidation.
1322 * Not doing so, we run the risk of losing dirty pages in the
1323 * vinvalbuf() call below.
1325 if (intrflg && (error = nfs_sigintr(nmp, td)))
1329 error = vinvalbuf(vp, flags, slpflag, 0);
1331 if (intrflg && (error = nfs_sigintr(nmp, td)))
1333 error = vinvalbuf(vp, flags, 0, slptimeo);
1335 mtx_lock(&np->n_mtx);
1336 if (np->n_directio_asyncwr == 0)
1337 np->n_flag &= ~NMODIFIED;
1338 mtx_unlock(&np->n_mtx);
1340 nfs_downgrade_vnlock(vp, old_lock);
1345 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1346 * This is mainly to avoid queueing async I/O requests when the nfsiods
1347 * are all hung on a dead server.
1349 * Note: nfs_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1350 * is eventually dequeued by the async daemon, nfs_doio() *will*.
1353 nfs_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1362 * Commits are usually short and sweet so lets save some cpu and
1363 * leave the async daemons for more important rpc's (such as reads
1366 mtx_lock(&nfs_iod_mtx);
1367 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1368 (nmp->nm_bufqiods > nfs_numasync / 2)) {
1369 mtx_unlock(&nfs_iod_mtx);
1373 if (nmp->nm_flag & NFSMNT_INT)
1374 slpflag = NFS_PCATCH;
1378 * Find a free iod to process this request.
1380 for (iod = 0; iod < nfs_numasync; iod++)
1381 if (nfs_iodwant[iod] == NFSIOD_AVAILABLE) {
1387 * Try to create one if none are free.
1393 * Found one, so wake it up and tell it which
1396 NFS_DPF(ASYNCIO, ("nfs_asyncio: waking iod %d for mount %p\n",
1398 nfs_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1399 nfs_iodmount[iod] = nmp;
1401 wakeup(&nfs_iodwant[iod]);
1405 * If none are free, we may already have an iod working on this mount
1406 * point. If so, it will process our request.
1409 if (nmp->nm_bufqiods > 0) {
1411 ("nfs_asyncio: %d iods are already processing mount %p\n",
1412 nmp->nm_bufqiods, nmp));
1418 * If we have an iod which can process the request, then queue
1423 * Ensure that the queue never grows too large. We still want
1424 * to asynchronize so we block rather then return EIO.
1426 while (nmp->nm_bufqlen >= 2 * nfs_numasync) {
1428 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp));
1429 nmp->nm_bufqwant = TRUE;
1430 error = nfs_msleep(td, &nmp->nm_bufq, &nfs_iod_mtx,
1432 "nfsaio", slptimeo);
1434 error2 = nfs_sigintr(nmp, td);
1436 mtx_unlock(&nfs_iod_mtx);
1439 if (slpflag == NFS_PCATCH) {
1445 * We might have lost our iod while sleeping,
1446 * so check and loop if nescessary.
1451 /* We might have lost our nfsiod */
1452 if (nmp->nm_bufqiods == 0) {
1454 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1458 if (bp->b_iocmd == BIO_READ) {
1459 if (bp->b_rcred == NOCRED && cred != NOCRED)
1460 bp->b_rcred = crhold(cred);
1462 if (bp->b_wcred == NOCRED && cred != NOCRED)
1463 bp->b_wcred = crhold(cred);
1466 if (bp->b_flags & B_REMFREE)
1469 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1471 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1472 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx);
1473 VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1474 VTONFS(bp->b_vp)->n_directio_asyncwr++;
1475 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx);
1477 mtx_unlock(&nfs_iod_mtx);
1481 mtx_unlock(&nfs_iod_mtx);
1484 * All the iods are busy on other mounts, so return EIO to
1485 * force the caller to process the i/o synchronously.
1487 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n"));
1492 nfs_doio_directwrite(struct buf *bp)
1494 int iomode, must_commit;
1495 struct uio *uiop = (struct uio *)bp->b_caller1;
1496 char *iov_base = uiop->uio_iov->iov_base;
1497 struct nfsmount *nmp = VFSTONFS(bp->b_vp->v_mount);
1499 iomode = NFSV3WRITE_FILESYNC;
1500 uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1501 (nmp->nm_rpcops->nr_writerpc)(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit);
1502 KASSERT((must_commit == 0), ("nfs_doio_directwrite: Did not commit write"));
1503 free(iov_base, M_NFSDIRECTIO);
1504 free(uiop->uio_iov, M_NFSDIRECTIO);
1505 free(uiop, M_NFSDIRECTIO);
1506 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1507 struct nfsnode *np = VTONFS(bp->b_vp);
1508 mtx_lock(&np->n_mtx);
1509 np->n_directio_asyncwr--;
1510 if (np->n_directio_asyncwr == 0) {
1511 VTONFS(bp->b_vp)->n_flag &= ~NMODIFIED;
1512 if ((np->n_flag & NFSYNCWAIT)) {
1513 np->n_flag &= ~NFSYNCWAIT;
1514 wakeup((caddr_t)&np->n_directio_asyncwr);
1517 mtx_unlock(&np->n_mtx);
1520 relpbuf(bp, &nfs_pbuf_freecnt);
1524 * Do an I/O operation to/from a cache block. This may be called
1525 * synchronously or from an nfsiod.
1528 nfs_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td)
1532 struct nfsmount *nmp;
1533 int error = 0, iomode, must_commit = 0;
1536 struct proc *p = td ? td->td_proc : NULL;
1540 nmp = VFSTONFS(vp->v_mount);
1542 uiop->uio_iov = &io;
1543 uiop->uio_iovcnt = 1;
1544 uiop->uio_segflg = UIO_SYSSPACE;
1548 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We
1549 * do this here so we do not have to do it in all the code that
1552 bp->b_flags &= ~B_INVAL;
1553 bp->b_ioflags &= ~BIO_ERROR;
1555 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp));
1556 iocmd = bp->b_iocmd;
1557 if (iocmd == BIO_READ) {
1558 io.iov_len = uiop->uio_resid = bp->b_bcount;
1559 io.iov_base = bp->b_data;
1560 uiop->uio_rw = UIO_READ;
1562 switch (vp->v_type) {
1564 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1565 nfsstats.read_bios++;
1566 error = (nmp->nm_rpcops->nr_readrpc)(vp, uiop, cr);
1569 if (uiop->uio_resid) {
1571 * If we had a short read with no error, we must have
1572 * hit a file hole. We should zero-fill the remainder.
1573 * This can also occur if the server hits the file EOF.
1575 * Holes used to be able to occur due to pending
1576 * writes, but that is not possible any longer.
1578 int nread = bp->b_bcount - uiop->uio_resid;
1579 int left = uiop->uio_resid;
1582 bzero((char *)bp->b_data + nread, left);
1583 uiop->uio_resid = 0;
1586 /* ASSERT_VOP_LOCKED(vp, "nfs_doio"); */
1587 if (p && (vp->v_vflag & VV_TEXT)) {
1588 mtx_lock(&np->n_mtx);
1589 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.va_mtime)) {
1590 mtx_unlock(&np->n_mtx);
1592 killproc(p, "text file modification");
1595 mtx_unlock(&np->n_mtx);
1599 uiop->uio_offset = (off_t)0;
1600 nfsstats.readlink_bios++;
1601 error = (nmp->nm_rpcops->nr_readlinkrpc)(vp, uiop, cr);
1604 nfsstats.readdir_bios++;
1605 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1606 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1607 error = nfs_readdirplusrpc(vp, uiop, cr);
1608 if (error == NFSERR_NOTSUPP)
1609 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1611 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1612 error = nfs_readdirrpc(vp, uiop, cr);
1614 * end-of-directory sets B_INVAL but does not generate an
1617 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1618 bp->b_flags |= B_INVAL;
1621 nfs_printf("nfs_doio: type %x unexpected\n", vp->v_type);
1625 bp->b_ioflags |= BIO_ERROR;
1626 bp->b_error = error;
1630 * If we only need to commit, try to commit
1632 if (bp->b_flags & B_NEEDCOMMIT) {
1636 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1637 retv = (nmp->nm_rpcops->nr_commit)(
1638 vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1641 bp->b_dirtyoff = bp->b_dirtyend = 0;
1642 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1647 if (retv == NFSERR_STALEWRITEVERF) {
1648 nfs_clearcommit(vp->v_mount);
1653 * Setup for actual write
1655 mtx_lock(&np->n_mtx);
1656 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1657 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1658 mtx_unlock(&np->n_mtx);
1660 if (bp->b_dirtyend > bp->b_dirtyoff) {
1661 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1663 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1665 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1666 uiop->uio_rw = UIO_WRITE;
1667 nfsstats.write_bios++;
1669 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1670 iomode = NFSV3WRITE_UNSTABLE;
1672 iomode = NFSV3WRITE_FILESYNC;
1674 error = (nmp->nm_rpcops->nr_writerpc)(vp, uiop, cr, &iomode, &must_commit);
1677 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1678 * to cluster the buffers needing commit. This will allow
1679 * the system to submit a single commit rpc for the whole
1680 * cluster. We can do this even if the buffer is not 100%
1681 * dirty (relative to the NFS blocksize), so we optimize the
1682 * append-to-file-case.
1684 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1685 * cleared because write clustering only works for commit
1686 * rpc's, not for the data portion of the write).
1689 if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1690 bp->b_flags |= B_NEEDCOMMIT;
1691 if (bp->b_dirtyoff == 0
1692 && bp->b_dirtyend == bp->b_bcount)
1693 bp->b_flags |= B_CLUSTEROK;
1695 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1699 * For an interrupted write, the buffer is still valid
1700 * and the write hasn't been pushed to the server yet,
1701 * so we can't set BIO_ERROR and report the interruption
1702 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1703 * is not relevant, so the rpc attempt is essentially
1704 * a noop. For the case of a V3 write rpc not being
1705 * committed to stable storage, the block is still
1706 * dirty and requires either a commit rpc or another
1707 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1708 * the block is reused. This is indicated by setting
1709 * the B_DELWRI and B_NEEDCOMMIT flags.
1711 * If the buffer is marked B_PAGING, it does not reside on
1712 * the vp's paging queues so we cannot call bdirty(). The
1713 * bp in this case is not an NFS cache block so we should
1716 * The logic below breaks up errors into recoverable and
1717 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1718 * and keep the buffer around for potential write retries.
1719 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1720 * and save the error in the nfsnode. This is less than ideal
1721 * but necessary. Keeping such buffers around could potentially
1722 * cause buffer exhaustion eventually (they can never be written
1723 * out, so will get constantly be re-dirtied). It also causes
1724 * all sorts of vfs panics. For non-recoverable write errors,
1725 * also invalidate the attrcache, so we'll be forced to go over
1726 * the wire for this object, returning an error to user on next
1727 * call (most of the time).
1729 if (error == EINTR || error == EIO || error == ETIMEDOUT
1730 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1734 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1735 if ((bp->b_flags & B_PAGING) == 0) {
1737 bp->b_flags &= ~B_DONE;
1739 if (error && (bp->b_flags & B_ASYNC) == 0)
1740 bp->b_flags |= B_EINTR;
1744 bp->b_ioflags |= BIO_ERROR;
1745 bp->b_flags |= B_INVAL;
1746 bp->b_error = np->n_error = error;
1747 mtx_lock(&np->n_mtx);
1748 np->n_flag |= NWRITEERR;
1749 np->n_attrstamp = 0;
1750 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1751 mtx_unlock(&np->n_mtx);
1753 bp->b_dirtyoff = bp->b_dirtyend = 0;
1761 bp->b_resid = uiop->uio_resid;
1763 nfs_clearcommit(vp->v_mount);
1769 * Used to aid in handling ftruncate() operations on the NFS client side.
1770 * Truncation creates a number of special problems for NFS. We have to
1771 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1772 * we have to properly handle VM pages or (potentially dirty) buffers
1773 * that straddle the truncation point.
1777 nfs_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1779 struct nfsnode *np = VTONFS(vp);
1781 int biosize = vp->v_bufobj.bo_bsize;
1784 mtx_lock(&np->n_mtx);
1787 mtx_unlock(&np->n_mtx);
1789 if (nsize < tsize) {
1795 * vtruncbuf() doesn't get the buffer overlapping the
1796 * truncation point. We may have a B_DELWRI and/or B_CACHE
1797 * buffer that now needs to be truncated.
1799 error = vtruncbuf(vp, cred, td, nsize, biosize);
1800 lbn = nsize / biosize;
1801 bufsize = nsize & (biosize - 1);
1802 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1805 if (bp->b_dirtyoff > bp->b_bcount)
1806 bp->b_dirtyoff = bp->b_bcount;
1807 if (bp->b_dirtyend > bp->b_bcount)
1808 bp->b_dirtyend = bp->b_bcount;
1809 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1812 vnode_pager_setsize(vp, nsize);