2 * SPDX-License-Identifier: BSD-3-Clause
4 * Copyright (c) 1989, 1993
5 * The Regents of the University of California. All rights reserved.
7 * This code is derived from software contributed to Berkeley by
8 * Rick Macklem at The University of Guelph.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * @(#)nfs_bio.c 8.9 (Berkeley) 3/30/95
37 #include <sys/cdefs.h>
38 __FBSDID("$FreeBSD$");
40 #include <sys/param.h>
41 #include <sys/systm.h>
44 #include <sys/kernel.h>
45 #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 <fs/nfs/nfsport.h>
59 #include <fs/nfsclient/nfsmount.h>
60 #include <fs/nfsclient/nfs.h>
61 #include <fs/nfsclient/nfsnode.h>
62 #include <fs/nfsclient/nfs_kdtrace.h>
64 extern int newnfs_directio_allow_mmap;
65 extern struct nfsstatsv1 nfsstatsv1;
66 extern struct mtx ncl_iod_mutex;
67 extern int ncl_numasync;
68 extern enum nfsiod_state ncl_iodwant[NFS_MAXASYNCDAEMON];
69 extern struct nfsmount *ncl_iodmount[NFS_MAXASYNCDAEMON];
70 extern int newnfs_directio_enable;
71 extern int nfs_keep_dirty_on_error;
73 uma_zone_t ncl_pbuf_zone;
75 static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size,
77 static int nfs_directio_write(struct vnode *vp, struct uio *uiop,
78 struct ucred *cred, int ioflag);
81 * Vnode op for VM getpages.
83 SYSCTL_DECL(_vfs_nfs);
84 static int use_buf_pager = 1;
85 SYSCTL_INT(_vfs_nfs, OID_AUTO, use_buf_pager, CTLFLAG_RWTUN,
87 "Use buffer pager instead of direct readrpc call");
90 ncl_gbp_getblkno(struct vnode *vp, vm_ooffset_t off)
93 return (off / vp->v_bufobj.bo_bsize);
97 ncl_gbp_getblksz(struct vnode *vp, daddr_t lbn)
108 biosize = vp->v_bufobj.bo_bsize;
110 if ((off_t)lbn * biosize >= nsize)
112 else if ((off_t)(lbn + 1) * biosize > nsize)
113 bcount = nsize - (off_t)lbn * biosize;
118 ncl_getpages(struct vop_getpages_args *ap)
120 int i, error, nextoff, size, toff, count, npages;
128 struct nfsmount *nmp;
136 cred = curthread->td_ucred;
137 nmp = VFSTONFS(vp->v_mount);
139 npages = ap->a_count;
141 if ((object = vp->v_object) == NULL) {
142 printf("ncl_getpages: called with non-merged cache vnode\n");
143 return (VM_PAGER_ERROR);
146 if (newnfs_directio_enable && !newnfs_directio_allow_mmap) {
148 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
150 printf("ncl_getpages: called on non-cacheable vnode\n");
151 return (VM_PAGER_ERROR);
156 mtx_lock(&nmp->nm_mtx);
157 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
158 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
159 mtx_unlock(&nmp->nm_mtx);
160 /* We'll never get here for v4, because we always have fsinfo */
161 (void)ncl_fsinfo(nmp, vp, cred, td);
163 mtx_unlock(&nmp->nm_mtx);
166 return (vfs_bio_getpages(vp, pages, npages, ap->a_rbehind,
167 ap->a_rahead, ncl_gbp_getblkno, ncl_gbp_getblksz));
170 * If the requested page is partially valid, just return it and
171 * allow the pager to zero-out the blanks. Partially valid pages
172 * can only occur at the file EOF.
174 * XXXGL: is that true for NFS, where short read can occur???
176 VM_OBJECT_WLOCK(object);
177 if (!vm_page_none_valid(pages[npages - 1]) && --npages == 0)
179 VM_OBJECT_WUNLOCK(object);
182 * We use only the kva address for the buffer, but this is extremely
183 * convenient and fast.
185 bp = uma_zalloc(ncl_pbuf_zone, M_WAITOK);
187 kva = (vm_offset_t) bp->b_data;
188 pmap_qenter(kva, pages, npages);
189 VM_CNT_INC(v_vnodein);
190 VM_CNT_ADD(v_vnodepgsin, npages);
192 count = npages << PAGE_SHIFT;
193 iov.iov_base = (caddr_t) kva;
197 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
198 uio.uio_resid = count;
199 uio.uio_segflg = UIO_SYSSPACE;
200 uio.uio_rw = UIO_READ;
203 error = ncl_readrpc(vp, &uio, cred);
204 pmap_qremove(kva, npages);
206 uma_zfree(ncl_pbuf_zone, bp);
208 if (error && (uio.uio_resid == count)) {
209 printf("ncl_getpages: error %d\n", error);
210 return (VM_PAGER_ERROR);
214 * Calculate the number of bytes read and validate only that number
215 * of bytes. Note that due to pending writes, size may be 0. This
216 * does not mean that the remaining data is invalid!
219 size = count - uio.uio_resid;
220 VM_OBJECT_WLOCK(object);
221 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
223 nextoff = toff + PAGE_SIZE;
226 if (nextoff <= size) {
228 * Read operation filled an entire page
231 KASSERT(m->dirty == 0,
232 ("nfs_getpages: page %p is dirty", m));
233 } else if (size > toff) {
235 * Read operation filled a partial page.
238 vm_page_set_valid_range(m, 0, size - toff);
239 KASSERT(m->dirty == 0,
240 ("nfs_getpages: page %p is dirty", m));
243 * Read operation was short. If no error
244 * occurred we may have hit a zero-fill
245 * section. We leave valid set to 0, and page
246 * is freed by vm_page_readahead_finish() if
247 * its index is not equal to requested, or
248 * page is zeroed and set valid by
249 * vm_pager_get_pages() for requested page.
255 VM_OBJECT_WUNLOCK(object);
260 return (VM_PAGER_OK);
264 * Vnode op for VM putpages.
267 ncl_putpages(struct vop_putpages_args *ap)
271 int i, error, npages, count;
277 struct nfsmount *nmp;
283 td = curthread; /* XXX */
284 /* Set the cred to n_writecred for the write rpcs. */
285 if (np->n_writecred != NULL)
286 cred = crhold(np->n_writecred);
288 cred = crhold(curthread->td_ucred); /* XXX */
289 nmp = VFSTONFS(vp->v_mount);
292 rtvals = ap->a_rtvals;
293 npages = btoc(count);
294 offset = IDX_TO_OFF(pages[0]->pindex);
296 mtx_lock(&nmp->nm_mtx);
297 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
298 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
299 mtx_unlock(&nmp->nm_mtx);
300 (void)ncl_fsinfo(nmp, vp, cred, td);
302 mtx_unlock(&nmp->nm_mtx);
305 if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
306 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
308 printf("ncl_putpages: called on noncache-able vnode\n");
312 * When putting pages, do not extend file past EOF.
314 if (offset + count > np->n_size) {
315 count = np->n_size - offset;
321 for (i = 0; i < npages; i++)
322 rtvals[i] = VM_PAGER_ERROR;
324 VM_CNT_INC(v_vnodeout);
325 VM_CNT_ADD(v_vnodepgsout, count);
327 iov.iov_base = unmapped_buf;
331 uio.uio_offset = offset;
332 uio.uio_resid = count;
333 uio.uio_segflg = UIO_NOCOPY;
334 uio.uio_rw = UIO_WRITE;
337 error = VOP_WRITE(vp, &uio, vnode_pager_putpages_ioflags(ap->a_sync),
341 if (error == 0 || !nfs_keep_dirty_on_error) {
342 vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid,
343 np->n_size - offset, npages * PAGE_SIZE);
349 * For nfs, cache consistency can only be maintained approximately.
350 * Although RFC1094 does not specify the criteria, the following is
351 * believed to be compatible with the reference port.
353 * If the file's modify time on the server has changed since the
354 * last read rpc or you have written to the file,
355 * you may have lost data cache consistency with the
356 * server, so flush all of the file's data out of the cache.
357 * Then force a getattr rpc to ensure that you have up to date
359 * NB: This implies that cache data can be read when up to
360 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
361 * attributes this could be forced by setting n_attrstamp to 0 before
362 * the VOP_GETATTR() call.
365 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
369 struct nfsnode *np = VTONFS(vp);
373 * Ensure the exclusove access to the node before checking
374 * whether the cache is consistent.
376 old_lock = ncl_excl_start(vp);
378 if (np->n_flag & NMODIFIED) {
380 if (vp->v_type != VREG) {
381 if (vp->v_type != VDIR)
382 panic("nfs: bioread, not dir");
384 error = ncl_vinvalbuf(vp, V_SAVE | V_ALLOWCLEAN, td, 1);
389 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
390 error = VOP_GETATTR(vp, &vattr, cred);
394 np->n_mtime = vattr.va_mtime;
398 error = VOP_GETATTR(vp, &vattr, cred);
402 if ((np->n_flag & NSIZECHANGED)
403 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
405 if (vp->v_type == VDIR)
407 error = ncl_vinvalbuf(vp, V_SAVE | V_ALLOWCLEAN, td, 1);
411 np->n_mtime = vattr.va_mtime;
412 np->n_flag &= ~NSIZECHANGED;
417 ncl_excl_finish(vp, old_lock);
422 * Vnode op for read using bio
425 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
427 struct nfsnode *np = VTONFS(vp);
428 struct buf *bp, *rabp;
430 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
432 int biosize, bcount, error, i, n, nra, on, save2, seqcount;
435 KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
436 if (uio->uio_resid == 0)
438 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
442 mtx_lock(&nmp->nm_mtx);
443 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
444 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
445 mtx_unlock(&nmp->nm_mtx);
446 (void)ncl_fsinfo(nmp, vp, cred, td);
447 mtx_lock(&nmp->nm_mtx);
449 if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
450 (void) newnfs_iosize(nmp);
452 tmp_off = uio->uio_offset + uio->uio_resid;
453 if (vp->v_type != VDIR &&
454 (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) {
455 mtx_unlock(&nmp->nm_mtx);
458 mtx_unlock(&nmp->nm_mtx);
460 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
461 /* No caching/ no readaheads. Just read data into the user buffer */
462 return ncl_readrpc(vp, uio, cred);
466 biosize = vp->v_bufobj.bo_bsize;
467 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
469 error = nfs_bioread_check_cons(vp, td, cred);
473 save2 = curthread_pflags2_set(TDP2_SBPAGES);
481 switch (vp->v_type) {
483 NFSINCRGLOBAL(nfsstatsv1.biocache_reads);
484 lbn = uio->uio_offset / biosize;
485 on = uio->uio_offset - (lbn * biosize);
488 * Start the read ahead(s), as required.
490 if (nmp->nm_readahead > 0) {
491 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
492 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
493 rabn = lbn + 1 + nra;
494 if (incore(&vp->v_bufobj, rabn) == NULL) {
495 rabp = nfs_getcacheblk(vp, rabn, biosize, td);
497 error = newnfs_sigintr(nmp, td);
502 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
503 rabp->b_flags |= B_ASYNC;
504 rabp->b_iocmd = BIO_READ;
505 vfs_busy_pages(rabp, 0);
506 if (ncl_asyncio(nmp, rabp, cred, td)) {
507 rabp->b_flags |= B_INVAL;
508 rabp->b_ioflags |= BIO_ERROR;
509 vfs_unbusy_pages(rabp);
520 /* Note that bcount is *not* DEV_BSIZE aligned. */
522 if ((off_t)lbn * biosize >= nsize) {
524 } else if ((off_t)(lbn + 1) * biosize > nsize) {
525 bcount = nsize - (off_t)lbn * biosize;
527 bp = nfs_getcacheblk(vp, lbn, bcount, td);
530 error = newnfs_sigintr(nmp, td);
537 * If B_CACHE is not set, we must issue the read. If this
538 * fails, we return an error.
541 if ((bp->b_flags & B_CACHE) == 0) {
542 bp->b_iocmd = BIO_READ;
543 vfs_busy_pages(bp, 0);
544 error = ncl_doio(vp, bp, cred, td, 0);
552 * on is the offset into the current bp. Figure out how many
553 * bytes we can copy out of the bp. Note that bcount is
554 * NOT DEV_BSIZE aligned.
556 * Then figure out how many bytes we can copy into the uio.
561 n = MIN((unsigned)(bcount - on), uio->uio_resid);
564 NFSINCRGLOBAL(nfsstatsv1.biocache_readlinks);
565 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
567 error = newnfs_sigintr(nmp, td);
572 if ((bp->b_flags & B_CACHE) == 0) {
573 bp->b_iocmd = BIO_READ;
574 vfs_busy_pages(bp, 0);
575 error = ncl_doio(vp, bp, cred, td, 0);
577 bp->b_ioflags |= BIO_ERROR;
582 n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
586 NFSINCRGLOBAL(nfsstatsv1.biocache_readdirs);
587 if (np->n_direofoffset
588 && uio->uio_offset >= np->n_direofoffset) {
592 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
593 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
594 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
596 error = newnfs_sigintr(nmp, td);
601 if ((bp->b_flags & B_CACHE) == 0) {
602 bp->b_iocmd = BIO_READ;
603 vfs_busy_pages(bp, 0);
604 error = ncl_doio(vp, bp, cred, td, 0);
608 while (error == NFSERR_BAD_COOKIE) {
610 error = ncl_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) {
628 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
630 error = newnfs_sigintr(nmp, td);
635 if ((bp->b_flags & B_CACHE) == 0) {
636 bp->b_iocmd = BIO_READ;
637 vfs_busy_pages(bp, 0);
638 error = ncl_doio(vp, bp, cred, td, 0);
640 * no error + B_INVAL == directory EOF,
643 if (error == 0 && (bp->b_flags & B_INVAL))
647 * An error will throw away the block and the
648 * for loop will break out. If no error and this
649 * is not the block we want, we throw away the
650 * block and go for the next one via the for loop.
652 if (error || i < lbn)
657 * The above while is repeated if we hit another cookie
658 * error. If we hit an error and it wasn't a cookie error,
666 * If not eof and read aheads are enabled, start one.
667 * (You need the current block first, so that you have the
668 * directory offset cookie of the next block.)
670 if (nmp->nm_readahead > 0 &&
671 (bp->b_flags & B_INVAL) == 0 &&
672 (np->n_direofoffset == 0 ||
673 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
674 incore(&vp->v_bufobj, lbn + 1) == NULL) {
675 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
677 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
678 rabp->b_flags |= B_ASYNC;
679 rabp->b_iocmd = BIO_READ;
680 vfs_busy_pages(rabp, 0);
681 if (ncl_asyncio(nmp, rabp, cred, td)) {
682 rabp->b_flags |= B_INVAL;
683 rabp->b_ioflags |= BIO_ERROR;
684 vfs_unbusy_pages(rabp);
693 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
694 * chopped for the EOF condition, we cannot tell how large
695 * NFS directories are going to be until we hit EOF. So
696 * an NFS directory buffer is *not* chopped to its EOF. Now,
697 * it just so happens that b_resid will effectively chop it
698 * to EOF. *BUT* this information is lost if the buffer goes
699 * away and is reconstituted into a B_CACHE state ( due to
700 * being VMIO ) later. So we keep track of the directory eof
701 * in np->n_direofoffset and chop it off as an extra step
704 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
705 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
706 n = np->n_direofoffset - uio->uio_offset;
709 printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
715 error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio);
717 if (vp->v_type == VLNK)
721 } while (error == 0 && uio->uio_resid > 0 && n > 0);
723 curthread_pflags2_restore(save2);
724 if ((curthread->td_pflags2 & TDP2_SBPAGES) == 0) {
726 ncl_pager_setsize(vp, NULL);
732 * The NFS write path cannot handle iovecs with len > 1. So we need to
733 * break up iovecs accordingly (restricting them to wsize).
734 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
735 * For the ASYNC case, 2 copies are needed. The first a copy from the
736 * user buffer to a staging buffer and then a second copy from the staging
737 * buffer to mbufs. This can be optimized by copying from the user buffer
738 * directly into mbufs and passing the chain down, but that requires a
739 * fair amount of re-working of the relevant codepaths (and can be done
743 nfs_directio_write(vp, uiop, cred, ioflag)
750 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
751 struct thread *td = uiop->uio_td;
755 mtx_lock(&nmp->nm_mtx);
756 wsize = nmp->nm_wsize;
757 mtx_unlock(&nmp->nm_mtx);
758 if (ioflag & IO_SYNC) {
759 int iomode, must_commit;
763 while (uiop->uio_resid > 0) {
764 size = MIN(uiop->uio_resid, wsize);
765 size = MIN(uiop->uio_iov->iov_len, size);
766 iov.iov_base = uiop->uio_iov->iov_base;
770 uio.uio_offset = uiop->uio_offset;
771 uio.uio_resid = size;
772 uio.uio_segflg = UIO_USERSPACE;
773 uio.uio_rw = UIO_WRITE;
775 iomode = NFSWRITE_FILESYNC;
776 error = ncl_writerpc(vp, &uio, cred, &iomode,
778 KASSERT((must_commit == 0),
779 ("ncl_directio_write: Did not commit write"));
782 uiop->uio_offset += size;
783 uiop->uio_resid -= size;
784 if (uiop->uio_iov->iov_len <= size) {
788 uiop->uio_iov->iov_base =
789 (char *)uiop->uio_iov->iov_base + size;
790 uiop->uio_iov->iov_len -= size;
799 * Break up the write into blocksize chunks and hand these
800 * over to nfsiod's for write back.
801 * Unfortunately, this incurs a copy of the data. Since
802 * the user could modify the buffer before the write is
805 * The obvious optimization here is that one of the 2 copies
806 * in the async write path can be eliminated by copying the
807 * data here directly into mbufs and passing the mbuf chain
808 * down. But that will require a fair amount of re-working
809 * of the code and can be done if there's enough interest
810 * in NFS directio access.
812 while (uiop->uio_resid > 0) {
813 size = MIN(uiop->uio_resid, wsize);
814 size = MIN(uiop->uio_iov->iov_len, size);
815 bp = uma_zalloc(ncl_pbuf_zone, M_WAITOK);
816 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
817 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
818 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
819 t_iov->iov_len = size;
820 t_uio->uio_iov = t_iov;
821 t_uio->uio_iovcnt = 1;
822 t_uio->uio_offset = uiop->uio_offset;
823 t_uio->uio_resid = size;
824 t_uio->uio_segflg = UIO_SYSSPACE;
825 t_uio->uio_rw = UIO_WRITE;
827 KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
828 uiop->uio_segflg == UIO_SYSSPACE,
829 ("nfs_directio_write: Bad uio_segflg"));
830 if (uiop->uio_segflg == UIO_USERSPACE) {
831 error = copyin(uiop->uio_iov->iov_base,
832 t_iov->iov_base, size);
837 * UIO_SYSSPACE may never happen, but handle
838 * it just in case it does.
840 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
842 bp->b_flags |= B_DIRECT;
843 bp->b_iocmd = BIO_WRITE;
844 if (cred != NOCRED) {
848 bp->b_wcred = NOCRED;
849 bp->b_caller1 = (void *)t_uio;
851 error = ncl_asyncio(nmp, bp, NOCRED, td);
854 free(t_iov->iov_base, M_NFSDIRECTIO);
855 free(t_iov, M_NFSDIRECTIO);
856 free(t_uio, M_NFSDIRECTIO);
858 uma_zfree(ncl_pbuf_zone, bp);
863 uiop->uio_offset += size;
864 uiop->uio_resid -= size;
865 if (uiop->uio_iov->iov_len <= size) {
869 uiop->uio_iov->iov_base =
870 (char *)uiop->uio_iov->iov_base + size;
871 uiop->uio_iov->iov_len -= size;
879 * Vnode op for write using bio
882 ncl_write(struct vop_write_args *ap)
885 struct uio *uio = ap->a_uio;
886 struct thread *td = uio->uio_td;
887 struct vnode *vp = ap->a_vp;
888 struct nfsnode *np = VTONFS(vp);
889 struct ucred *cred = ap->a_cred;
890 int ioflag = ap->a_ioflag;
893 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
895 int bcount, noncontig_write, obcount;
896 int bp_cached, n, on, error = 0, error1, save2, wouldcommit;
897 size_t orig_resid, local_resid;
898 off_t orig_size, tmp_off;
900 KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
901 KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
903 if (vp->v_type != VREG)
906 if (np->n_flag & NWRITEERR) {
907 np->n_flag &= ~NWRITEERR;
909 return (np->n_error);
912 mtx_lock(&nmp->nm_mtx);
913 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
914 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
915 mtx_unlock(&nmp->nm_mtx);
916 (void)ncl_fsinfo(nmp, vp, cred, td);
917 mtx_lock(&nmp->nm_mtx);
919 if (nmp->nm_wsize == 0)
920 (void) newnfs_iosize(nmp);
921 mtx_unlock(&nmp->nm_mtx);
924 * Synchronously flush pending buffers if we are in synchronous
925 * mode or if we are appending.
927 if (ioflag & (IO_APPEND | IO_SYNC)) {
929 if (np->n_flag & NMODIFIED) {
931 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
933 * Require non-blocking, synchronous writes to
934 * dirty files to inform the program it needs
935 * to fsync(2) explicitly.
937 if (ioflag & IO_NDELAY)
941 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
942 error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
943 IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
950 orig_resid = uio->uio_resid;
952 orig_size = np->n_size;
956 * If IO_APPEND then load uio_offset. We restart here if we cannot
957 * get the append lock.
959 if (ioflag & IO_APPEND) {
961 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
962 error = VOP_GETATTR(vp, &vattr, cred);
966 uio->uio_offset = np->n_size;
970 if (uio->uio_offset < 0)
972 tmp_off = uio->uio_offset + uio->uio_resid;
973 if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
975 if (uio->uio_resid == 0)
978 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
979 return nfs_directio_write(vp, uio, cred, ioflag);
982 * Maybe this should be above the vnode op call, but so long as
983 * file servers have no limits, i don't think it matters
985 if (vn_rlimit_fsize(vp, uio, td))
988 save2 = curthread_pflags2_set(TDP2_SBPAGES);
989 biosize = vp->v_bufobj.bo_bsize;
991 * Find all of this file's B_NEEDCOMMIT buffers. If our writes
992 * would exceed the local maximum per-file write commit size when
993 * combined with those, we must decide whether to flush,
994 * go synchronous, or return error. We don't bother checking
995 * IO_UNIT -- we just make all writes atomic anyway, as there's
996 * no point optimizing for something that really won't ever happen.
999 if (!(ioflag & IO_SYNC)) {
1005 if (nflag & NMODIFIED) {
1006 BO_LOCK(&vp->v_bufobj);
1007 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
1008 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
1010 if (bp->b_flags & B_NEEDCOMMIT)
1011 wouldcommit += bp->b_bcount;
1014 BO_UNLOCK(&vp->v_bufobj);
1019 if (!(ioflag & IO_SYNC)) {
1020 wouldcommit += biosize;
1021 if (wouldcommit > nmp->nm_wcommitsize) {
1022 np->n_attrstamp = 0;
1023 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1024 error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
1025 IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
1028 wouldcommit = biosize;
1032 NFSINCRGLOBAL(nfsstatsv1.biocache_writes);
1033 lbn = uio->uio_offset / biosize;
1034 on = uio->uio_offset - (lbn * biosize);
1035 n = MIN((unsigned)(biosize - on), uio->uio_resid);
1038 * Handle direct append and file extension cases, calculate
1039 * unaligned buffer size.
1042 if ((np->n_flag & NHASBEENLOCKED) == 0 &&
1043 (nmp->nm_flag & NFSMNT_NONCONTIGWR) != 0)
1044 noncontig_write = 1;
1046 noncontig_write = 0;
1047 if ((uio->uio_offset == np->n_size ||
1048 (noncontig_write != 0 &&
1049 lbn == (np->n_size / biosize) &&
1050 uio->uio_offset + n > np->n_size)) && n) {
1053 * Get the buffer (in its pre-append state to maintain
1054 * B_CACHE if it was previously set). Resize the
1055 * nfsnode after we have locked the buffer to prevent
1056 * readers from reading garbage.
1058 obcount = np->n_size - (lbn * biosize);
1059 bp = nfs_getcacheblk(vp, lbn, obcount, td);
1065 np->n_size = uio->uio_offset + n;
1066 np->n_flag |= NMODIFIED;
1067 np->n_flag &= ~NVNSETSZSKIP;
1068 vnode_pager_setsize(vp, np->n_size);
1071 save = bp->b_flags & B_CACHE;
1073 allocbuf(bp, bcount);
1074 bp->b_flags |= save;
1075 if (noncontig_write != 0 && on > obcount)
1076 vfs_bio_bzero_buf(bp, obcount, on -
1081 * Obtain the locked cache block first, and then
1082 * adjust the file's size as appropriate.
1085 if ((off_t)lbn * biosize + bcount < np->n_size) {
1086 if ((off_t)(lbn + 1) * biosize < np->n_size)
1089 bcount = np->n_size - (off_t)lbn * biosize;
1092 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1094 if (uio->uio_offset + n > np->n_size) {
1095 np->n_size = uio->uio_offset + n;
1096 np->n_flag |= NMODIFIED;
1097 np->n_flag &= ~NVNSETSZSKIP;
1098 vnode_pager_setsize(vp, np->n_size);
1104 error = newnfs_sigintr(nmp, td);
1111 * Issue a READ if B_CACHE is not set. In special-append
1112 * mode, B_CACHE is based on the buffer prior to the write
1113 * op and is typically set, avoiding the read. If a read
1114 * is required in special append mode, the server will
1115 * probably send us a short-read since we extended the file
1116 * on our end, resulting in b_resid == 0 and, thusly,
1117 * B_CACHE getting set.
1119 * We can also avoid issuing the read if the write covers
1120 * the entire buffer. We have to make sure the buffer state
1121 * is reasonable in this case since we will not be initiating
1122 * I/O. See the comments in kern/vfs_bio.c's getblk() for
1125 * B_CACHE may also be set due to the buffer being cached
1130 if (on == 0 && n == bcount) {
1131 if ((bp->b_flags & B_CACHE) == 0)
1133 bp->b_flags |= B_CACHE;
1134 bp->b_flags &= ~B_INVAL;
1135 bp->b_ioflags &= ~BIO_ERROR;
1138 if ((bp->b_flags & B_CACHE) == 0) {
1139 bp->b_iocmd = BIO_READ;
1140 vfs_busy_pages(bp, 0);
1141 error = ncl_doio(vp, bp, cred, td, 0);
1147 if (bp->b_wcred == NOCRED)
1148 bp->b_wcred = crhold(cred);
1150 np->n_flag |= NMODIFIED;
1154 * If dirtyend exceeds file size, chop it down. This should
1155 * not normally occur but there is an append race where it
1156 * might occur XXX, so we log it.
1158 * If the chopping creates a reverse-indexed or degenerate
1159 * situation with dirtyoff/end, we 0 both of them.
1162 if (bp->b_dirtyend > bcount) {
1163 printf("NFS append race @%lx:%d\n",
1164 (long)bp->b_blkno * DEV_BSIZE,
1165 bp->b_dirtyend - bcount);
1166 bp->b_dirtyend = bcount;
1169 if (bp->b_dirtyoff >= bp->b_dirtyend)
1170 bp->b_dirtyoff = bp->b_dirtyend = 0;
1173 * If the new write will leave a contiguous dirty
1174 * area, just update the b_dirtyoff and b_dirtyend,
1175 * otherwise force a write rpc of the old dirty area.
1177 * If there has been a file lock applied to this file
1178 * or vfs.nfs.old_noncontig_writing is set, do the following:
1179 * While it is possible to merge discontiguous writes due to
1180 * our having a B_CACHE buffer ( and thus valid read data
1181 * for the hole), we don't because it could lead to
1182 * significant cache coherency problems with multiple clients,
1183 * especially if locking is implemented later on.
1185 * If vfs.nfs.old_noncontig_writing is not set and there has
1186 * not been file locking done on this file:
1187 * Relax coherency a bit for the sake of performance and
1188 * expand the current dirty region to contain the new
1189 * write even if it means we mark some non-dirty data as
1193 if (noncontig_write == 0 && bp->b_dirtyend > 0 &&
1194 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1195 if (bwrite(bp) == EINTR) {
1202 local_resid = uio->uio_resid;
1203 error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio);
1205 if (error != 0 && !bp_cached) {
1207 * This block has no other content then what
1208 * possibly was written by the faulty uiomove.
1209 * Release it, forgetting the data pages, to
1210 * prevent the leak of uninitialized data to
1213 bp->b_ioflags |= BIO_ERROR;
1215 uio->uio_offset -= local_resid - uio->uio_resid;
1216 uio->uio_resid = local_resid;
1221 * Since this block is being modified, it must be written
1222 * again and not just committed. Since write clustering does
1223 * not work for the stage 1 data write, only the stage 2
1224 * commit rpc, we have to clear B_CLUSTEROK as well.
1226 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1229 * Get the partial update on the progress made from
1230 * uiomove, if an error occurred.
1233 n = local_resid - uio->uio_resid;
1236 * Only update dirtyoff/dirtyend if not a degenerate
1240 if (bp->b_dirtyend > 0) {
1241 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1242 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1244 bp->b_dirtyoff = on;
1245 bp->b_dirtyend = on + n;
1247 vfs_bio_set_valid(bp, on, n);
1251 * If IO_SYNC do bwrite().
1253 * IO_INVAL appears to be unused. The idea appears to be
1254 * to turn off caching in this case. Very odd. XXX
1256 if ((ioflag & IO_SYNC)) {
1257 if (ioflag & IO_INVAL)
1258 bp->b_flags |= B_NOCACHE;
1259 error1 = bwrite(bp);
1265 } else if ((n + on) == biosize || (ioflag & IO_ASYNC) != 0) {
1266 bp->b_flags |= B_ASYNC;
1267 (void) ncl_writebp(bp, 0, NULL);
1274 } while (uio->uio_resid > 0 && n > 0);
1277 if (ioflag & IO_UNIT) {
1279 vattr.va_size = orig_size;
1280 /* IO_SYNC is handled implicitely */
1281 (void)VOP_SETATTR(vp, &vattr, cred);
1282 uio->uio_offset -= orig_resid - uio->uio_resid;
1283 uio->uio_resid = orig_resid;
1288 curthread_pflags2_restore(save2);
1293 * Get an nfs cache block.
1295 * Allocate a new one if the block isn't currently in the cache
1296 * and return the block marked busy. If the calling process is
1297 * interrupted by a signal for an interruptible mount point, return
1300 * The caller must carefully deal with the possible B_INVAL state of
1301 * the buffer. ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
1302 * indirectly), so synchronous reads can be issued without worrying about
1303 * the B_INVAL state. We have to be a little more careful when dealing
1304 * with writes (see comments in nfs_write()) when extending a file past
1308 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1312 struct nfsmount *nmp;
1317 if (nmp->nm_flag & NFSMNT_INT) {
1320 newnfs_set_sigmask(td, &oldset);
1321 bp = getblk(vp, bn, size, PCATCH, 0, 0);
1322 newnfs_restore_sigmask(td, &oldset);
1323 while (bp == NULL) {
1324 if (newnfs_sigintr(nmp, td))
1326 bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1329 bp = getblk(vp, bn, size, 0, 0, 0);
1332 if (vp->v_type == VREG)
1333 bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1338 * Flush and invalidate all dirty buffers. If another process is already
1339 * doing the flush, just wait for completion.
1342 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1344 struct nfsnode *np = VTONFS(vp);
1345 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1346 int error = 0, slpflag, slptimeo;
1349 ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
1351 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1353 if (NFSCL_FORCEDISM(nmp->nm_mountp))
1363 old_lock = ncl_excl_start(vp);
1365 flags |= V_ALLOWCLEAN;
1368 * Now, flush as required.
1370 if ((flags & (V_SAVE | V_VMIO)) == V_SAVE &&
1371 vp->v_bufobj.bo_object != NULL) {
1372 VM_OBJECT_WLOCK(vp->v_bufobj.bo_object);
1373 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1374 VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object);
1376 * If the page clean was interrupted, fail the invalidation.
1377 * Not doing so, we run the risk of losing dirty pages in the
1378 * vinvalbuf() call below.
1380 if (intrflg && (error = newnfs_sigintr(nmp, td)))
1384 error = vinvalbuf(vp, flags, slpflag, 0);
1386 if (intrflg && (error = newnfs_sigintr(nmp, td)))
1388 error = vinvalbuf(vp, flags, 0, slptimeo);
1390 if (NFSHASPNFS(nmp)) {
1391 nfscl_layoutcommit(vp, td);
1393 * Invalidate the attribute cache, since writes to a DS
1394 * won't update the size attribute.
1397 np->n_attrstamp = 0;
1400 if (np->n_directio_asyncwr == 0)
1401 np->n_flag &= ~NMODIFIED;
1404 ncl_excl_finish(vp, old_lock);
1409 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1410 * This is mainly to avoid queueing async I/O requests when the nfsiods
1411 * are all hung on a dead server.
1413 * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1414 * is eventually dequeued by the async daemon, ncl_doio() *will*.
1417 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1426 * Commits are usually short and sweet so lets save some cpu and
1427 * leave the async daemons for more important rpc's (such as reads
1430 * Readdirplus RPCs do vget()s to acquire the vnodes for entries
1431 * in the directory in order to update attributes. This can deadlock
1432 * with another thread that is waiting for async I/O to be done by
1433 * an nfsiod thread while holding a lock on one of these vnodes.
1434 * To avoid this deadlock, don't allow the async nfsiod threads to
1435 * perform Readdirplus RPCs.
1438 if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1439 (nmp->nm_bufqiods > ncl_numasync / 2)) ||
1440 (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
1445 if (nmp->nm_flag & NFSMNT_INT)
1450 * Find a free iod to process this request.
1452 for (iod = 0; iod < ncl_numasync; iod++)
1453 if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1459 * Try to create one if none are free.
1465 * Found one, so wake it up and tell it which
1468 NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1470 ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1471 ncl_iodmount[iod] = nmp;
1473 wakeup(&ncl_iodwant[iod]);
1477 * If none are free, we may already have an iod working on this mount
1478 * point. If so, it will process our request.
1481 if (nmp->nm_bufqiods > 0) {
1483 ("ncl_asyncio: %d iods are already processing mount %p\n",
1484 nmp->nm_bufqiods, nmp));
1490 * If we have an iod which can process the request, then queue
1495 * Ensure that the queue never grows too large. We still want
1496 * to asynchronize so we block rather then return EIO.
1498 while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1500 ("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1501 nmp->nm_bufqwant = TRUE;
1502 error = newnfs_msleep(td, &nmp->nm_bufq,
1503 &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1506 error2 = newnfs_sigintr(nmp, td);
1511 if (slpflag == PCATCH) {
1517 * We might have lost our iod while sleeping,
1518 * so check and loop if necessary.
1523 /* We might have lost our nfsiod */
1524 if (nmp->nm_bufqiods == 0) {
1526 ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1530 if (bp->b_iocmd == BIO_READ) {
1531 if (bp->b_rcred == NOCRED && cred != NOCRED)
1532 bp->b_rcred = crhold(cred);
1534 if (bp->b_wcred == NOCRED && cred != NOCRED)
1535 bp->b_wcred = crhold(cred);
1538 if (bp->b_flags & B_REMFREE)
1541 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1543 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1544 NFSLOCKNODE(VTONFS(bp->b_vp));
1545 VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1546 VTONFS(bp->b_vp)->n_directio_asyncwr++;
1547 NFSUNLOCKNODE(VTONFS(bp->b_vp));
1556 * All the iods are busy on other mounts, so return EIO to
1557 * force the caller to process the i/o synchronously.
1559 NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1564 ncl_doio_directwrite(struct buf *bp)
1566 int iomode, must_commit;
1567 struct uio *uiop = (struct uio *)bp->b_caller1;
1568 char *iov_base = uiop->uio_iov->iov_base;
1570 iomode = NFSWRITE_FILESYNC;
1571 uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1572 ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0);
1573 KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write"));
1574 free(iov_base, M_NFSDIRECTIO);
1575 free(uiop->uio_iov, M_NFSDIRECTIO);
1576 free(uiop, M_NFSDIRECTIO);
1577 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1578 struct nfsnode *np = VTONFS(bp->b_vp);
1580 if (NFSHASPNFS(VFSTONFS(bp->b_vp->v_mount))) {
1582 * Invalidate the attribute cache, since writes to a DS
1583 * won't update the size attribute.
1585 np->n_attrstamp = 0;
1587 np->n_directio_asyncwr--;
1588 if (np->n_directio_asyncwr == 0) {
1589 np->n_flag &= ~NMODIFIED;
1590 if ((np->n_flag & NFSYNCWAIT)) {
1591 np->n_flag &= ~NFSYNCWAIT;
1592 wakeup((caddr_t)&np->n_directio_asyncwr);
1598 uma_zfree(ncl_pbuf_zone, bp);
1602 * Do an I/O operation to/from a cache block. This may be called
1603 * synchronously or from an nfsiod.
1606 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1607 int called_from_strategy)
1611 struct nfsmount *nmp;
1612 int error = 0, iomode, must_commit = 0;
1615 struct proc *p = td ? td->td_proc : NULL;
1619 nmp = VFSTONFS(vp->v_mount);
1621 uiop->uio_iov = &io;
1622 uiop->uio_iovcnt = 1;
1623 uiop->uio_segflg = UIO_SYSSPACE;
1627 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We
1628 * do this here so we do not have to do it in all the code that
1631 bp->b_flags &= ~B_INVAL;
1632 bp->b_ioflags &= ~BIO_ERROR;
1634 KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1635 iocmd = bp->b_iocmd;
1636 if (iocmd == BIO_READ) {
1637 io.iov_len = uiop->uio_resid = bp->b_bcount;
1638 io.iov_base = bp->b_data;
1639 uiop->uio_rw = UIO_READ;
1641 switch (vp->v_type) {
1643 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1644 NFSINCRGLOBAL(nfsstatsv1.read_bios);
1645 error = ncl_readrpc(vp, uiop, cr);
1648 if (uiop->uio_resid) {
1650 * If we had a short read with no error, we must have
1651 * hit a file hole. We should zero-fill the remainder.
1652 * This can also occur if the server hits the file EOF.
1654 * Holes used to be able to occur due to pending
1655 * writes, but that is not possible any longer.
1657 int nread = bp->b_bcount - uiop->uio_resid;
1658 ssize_t left = uiop->uio_resid;
1661 bzero((char *)bp->b_data + nread, left);
1662 uiop->uio_resid = 0;
1665 /* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1666 if (p && vp->v_writecount <= -1) {
1668 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1671 killproc(p, "text file modification");
1678 uiop->uio_offset = (off_t)0;
1679 NFSINCRGLOBAL(nfsstatsv1.readlink_bios);
1680 error = ncl_readlinkrpc(vp, uiop, cr);
1683 NFSINCRGLOBAL(nfsstatsv1.readdir_bios);
1684 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1685 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1686 error = ncl_readdirplusrpc(vp, uiop, cr, td);
1687 if (error == NFSERR_NOTSUPP)
1688 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1690 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1691 error = ncl_readdirrpc(vp, uiop, cr, td);
1693 * end-of-directory sets B_INVAL but does not generate an
1696 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1697 bp->b_flags |= B_INVAL;
1700 printf("ncl_doio: type %x unexpected\n", vp->v_type);
1704 bp->b_ioflags |= BIO_ERROR;
1705 bp->b_error = error;
1709 * If we only need to commit, try to commit
1711 if (bp->b_flags & B_NEEDCOMMIT) {
1715 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1716 retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1719 bp->b_dirtyoff = bp->b_dirtyend = 0;
1720 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1725 if (retv == NFSERR_STALEWRITEVERF) {
1726 ncl_clearcommit(vp->v_mount);
1731 * Setup for actual write
1734 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1735 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1738 if (bp->b_dirtyend > bp->b_dirtyoff) {
1739 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1741 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1743 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1744 uiop->uio_rw = UIO_WRITE;
1745 NFSINCRGLOBAL(nfsstatsv1.write_bios);
1747 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1748 iomode = NFSWRITE_UNSTABLE;
1750 iomode = NFSWRITE_FILESYNC;
1752 error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1753 called_from_strategy);
1756 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1757 * to cluster the buffers needing commit. This will allow
1758 * the system to submit a single commit rpc for the whole
1759 * cluster. We can do this even if the buffer is not 100%
1760 * dirty (relative to the NFS blocksize), so we optimize the
1761 * append-to-file-case.
1763 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1764 * cleared because write clustering only works for commit
1765 * rpc's, not for the data portion of the write).
1768 if (!error && iomode == NFSWRITE_UNSTABLE) {
1769 bp->b_flags |= B_NEEDCOMMIT;
1770 if (bp->b_dirtyoff == 0
1771 && bp->b_dirtyend == bp->b_bcount)
1772 bp->b_flags |= B_CLUSTEROK;
1774 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1778 * For an interrupted write, the buffer is still valid
1779 * and the write hasn't been pushed to the server yet,
1780 * so we can't set BIO_ERROR and report the interruption
1781 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1782 * is not relevant, so the rpc attempt is essentially
1783 * a noop. For the case of a V3 write rpc not being
1784 * committed to stable storage, the block is still
1785 * dirty and requires either a commit rpc or another
1786 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1787 * the block is reused. This is indicated by setting
1788 * the B_DELWRI and B_NEEDCOMMIT flags.
1790 * EIO is returned by ncl_writerpc() to indicate a recoverable
1791 * write error and is handled as above, except that
1792 * B_EINTR isn't set. One cause of this is a stale stateid
1793 * error for the RPC that indicates recovery is required,
1794 * when called with called_from_strategy != 0.
1796 * If the buffer is marked B_PAGING, it does not reside on
1797 * the vp's paging queues so we cannot call bdirty(). The
1798 * bp in this case is not an NFS cache block so we should
1801 * The logic below breaks up errors into recoverable and
1802 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1803 * and keep the buffer around for potential write retries.
1804 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1805 * and save the error in the nfsnode. This is less than ideal
1806 * but necessary. Keeping such buffers around could potentially
1807 * cause buffer exhaustion eventually (they can never be written
1808 * out, so will get constantly be re-dirtied). It also causes
1809 * all sorts of vfs panics. For non-recoverable write errors,
1810 * also invalidate the attrcache, so we'll be forced to go over
1811 * the wire for this object, returning an error to user on next
1812 * call (most of the time).
1814 if (error == EINTR || error == EIO || error == ETIMEDOUT
1815 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1816 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1817 if ((bp->b_flags & B_PAGING) == 0) {
1819 bp->b_flags &= ~B_DONE;
1821 if ((error == EINTR || error == ETIMEDOUT) &&
1822 (bp->b_flags & B_ASYNC) == 0)
1823 bp->b_flags |= B_EINTR;
1826 bp->b_ioflags |= BIO_ERROR;
1827 bp->b_flags |= B_INVAL;
1828 bp->b_error = np->n_error = error;
1830 np->n_flag |= NWRITEERR;
1831 np->n_attrstamp = 0;
1832 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1835 bp->b_dirtyoff = bp->b_dirtyend = 0;
1843 bp->b_resid = uiop->uio_resid;
1845 ncl_clearcommit(vp->v_mount);
1851 * Used to aid in handling ftruncate() operations on the NFS client side.
1852 * Truncation creates a number of special problems for NFS. We have to
1853 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1854 * we have to properly handle VM pages or (potentially dirty) buffers
1855 * that straddle the truncation point.
1859 ncl_meta_setsize(struct vnode *vp, struct thread *td, u_quad_t nsize)
1861 struct nfsnode *np = VTONFS(vp);
1863 int biosize = vp->v_bufobj.bo_bsize;
1871 if (nsize < tsize) {
1877 * vtruncbuf() doesn't get the buffer overlapping the
1878 * truncation point. We may have a B_DELWRI and/or B_CACHE
1879 * buffer that now needs to be truncated.
1881 error = vtruncbuf(vp, nsize, biosize);
1882 lbn = nsize / biosize;
1883 bufsize = nsize - (lbn * biosize);
1884 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1887 if (bp->b_dirtyoff > bp->b_bcount)
1888 bp->b_dirtyoff = bp->b_bcount;
1889 if (bp->b_dirtyend > bp->b_bcount)
1890 bp->b_dirtyend = bp->b_bcount;
1891 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1894 vnode_pager_setsize(vp, nsize);