2 * Copyright (c) 1989, 1993
3 * The Regents of the University of California. All rights reserved.
5 * This code is derived from software contributed to Berkeley by
6 * Rick Macklem at The University of Guelph.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 4. Neither the name of the University nor the names of its contributors
17 * may be used to endorse or promote products derived from this software
18 * without specific prior written permission.
20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * @(#)nfs_bio.c 8.9 (Berkeley) 3/30/95
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
38 #include <sys/param.h>
39 #include <sys/systm.h>
42 #include <sys/kernel.h>
43 #include <sys/mount.h>
44 #include <sys/vmmeter.h>
45 #include <sys/vnode.h>
48 #include <vm/vm_extern.h>
49 #include <vm/vm_page.h>
50 #include <vm/vm_object.h>
51 #include <vm/vm_pager.h>
52 #include <vm/vnode_pager.h>
54 #include <fs/nfs/nfsport.h>
55 #include <fs/nfsclient/nfsmount.h>
56 #include <fs/nfsclient/nfs.h>
57 #include <fs/nfsclient/nfsnode.h>
59 extern int newnfs_directio_allow_mmap;
60 extern struct nfsstats newnfsstats;
61 extern struct mtx ncl_iod_mutex;
62 extern int ncl_numasync;
63 extern enum nfsiod_state ncl_iodwant[NFS_MAXASYNCDAEMON];
64 extern struct nfsmount *ncl_iodmount[NFS_MAXASYNCDAEMON];
65 extern int newnfs_directio_enable;
66 extern int newnfs_keep_dirty_on_error;
68 int ncl_pbuf_freecnt = -1; /* start out unlimited */
70 static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size,
72 static int nfs_directio_write(struct vnode *vp, struct uio *uiop,
73 struct ucred *cred, int ioflag);
76 * Vnode op for VM getpages.
79 ncl_getpages(struct vop_getpages_args *ap)
81 int i, error, nextoff, size, toff, count, npages;
96 td = curthread; /* XXX */
97 cred = curthread->td_ucred; /* XXX */
98 nmp = VFSTONFS(vp->v_mount);
102 if ((object = vp->v_object) == NULL) {
103 ncl_printf("nfs_getpages: called with non-merged cache vnode??\n");
104 return (VM_PAGER_ERROR);
107 if (newnfs_directio_enable && !newnfs_directio_allow_mmap) {
108 mtx_lock(&np->n_mtx);
109 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
110 mtx_unlock(&np->n_mtx);
111 ncl_printf("nfs_getpages: called on non-cacheable vnode??\n");
112 return (VM_PAGER_ERROR);
114 mtx_unlock(&np->n_mtx);
117 mtx_lock(&nmp->nm_mtx);
118 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
119 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
120 mtx_unlock(&nmp->nm_mtx);
121 /* We'll never get here for v4, because we always have fsinfo */
122 (void)ncl_fsinfo(nmp, vp, cred, td);
124 mtx_unlock(&nmp->nm_mtx);
126 npages = btoc(count);
129 * If the requested page is partially valid, just return it and
130 * allow the pager to zero-out the blanks. Partially valid pages
131 * can only occur at the file EOF.
133 VM_OBJECT_LOCK(object);
134 if (pages[ap->a_reqpage]->valid != 0) {
135 vm_page_lock_queues();
136 for (i = 0; i < npages; ++i) {
137 if (i != ap->a_reqpage)
138 vm_page_free(pages[i]);
140 vm_page_unlock_queues();
141 VM_OBJECT_UNLOCK(object);
144 VM_OBJECT_UNLOCK(object);
147 * We use only the kva address for the buffer, but this is extremely
148 * convienient and fast.
150 bp = getpbuf(&ncl_pbuf_freecnt);
152 kva = (vm_offset_t) bp->b_data;
153 pmap_qenter(kva, pages, npages);
154 PCPU_INC(cnt.v_vnodein);
155 PCPU_ADD(cnt.v_vnodepgsin, npages);
157 iov.iov_base = (caddr_t) kva;
161 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
162 uio.uio_resid = count;
163 uio.uio_segflg = UIO_SYSSPACE;
164 uio.uio_rw = UIO_READ;
167 error = ncl_readrpc(vp, &uio, cred);
168 pmap_qremove(kva, npages);
170 relpbuf(bp, &ncl_pbuf_freecnt);
172 if (error && (uio.uio_resid == count)) {
173 ncl_printf("nfs_getpages: error %d\n", error);
174 VM_OBJECT_LOCK(object);
175 vm_page_lock_queues();
176 for (i = 0; i < npages; ++i) {
177 if (i != ap->a_reqpage)
178 vm_page_free(pages[i]);
180 vm_page_unlock_queues();
181 VM_OBJECT_UNLOCK(object);
182 return (VM_PAGER_ERROR);
186 * Calculate the number of bytes read and validate only that number
187 * of bytes. Note that due to pending writes, size may be 0. This
188 * does not mean that the remaining data is invalid!
191 size = count - uio.uio_resid;
192 VM_OBJECT_LOCK(object);
193 vm_page_lock_queues();
194 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
196 nextoff = toff + PAGE_SIZE;
199 if (nextoff <= size) {
201 * Read operation filled an entire page
203 m->valid = VM_PAGE_BITS_ALL;
204 KASSERT(m->dirty == 0,
205 ("nfs_getpages: page %p is dirty", m));
206 } else if (size > toff) {
208 * Read operation filled a partial page.
211 vm_page_set_valid(m, 0, size - toff);
212 KASSERT(m->dirty == 0,
213 ("nfs_getpages: page %p is dirty", m));
216 * Read operation was short. If no error occured
217 * we may have hit a zero-fill section. We simply
218 * leave valid set to 0.
222 if (i != ap->a_reqpage) {
224 * Whether or not to leave the page activated is up in
225 * the air, but we should put the page on a page queue
226 * somewhere (it already is in the object). Result:
227 * It appears that emperical results show that
228 * deactivating pages is best.
232 * Just in case someone was asking for this page we
233 * now tell them that it is ok to use.
236 if (m->oflags & VPO_WANTED)
239 vm_page_deactivate(m);
246 vm_page_unlock_queues();
247 VM_OBJECT_UNLOCK(object);
252 * Vnode op for VM putpages.
255 ncl_putpages(struct vop_putpages_args *ap)
261 int iomode, must_commit, i, error, npages, count;
267 struct nfsmount *nmp;
273 td = curthread; /* XXX */
274 cred = curthread->td_ucred; /* XXX */
275 nmp = VFSTONFS(vp->v_mount);
278 rtvals = ap->a_rtvals;
279 npages = btoc(count);
280 offset = IDX_TO_OFF(pages[0]->pindex);
282 mtx_lock(&nmp->nm_mtx);
283 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
284 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
285 mtx_unlock(&nmp->nm_mtx);
286 (void)ncl_fsinfo(nmp, vp, cred, td);
288 mtx_unlock(&nmp->nm_mtx);
290 mtx_lock(&np->n_mtx);
291 if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
292 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
293 mtx_unlock(&np->n_mtx);
294 ncl_printf("ncl_putpages: called on noncache-able vnode??\n");
295 mtx_lock(&np->n_mtx);
298 for (i = 0; i < npages; i++)
299 rtvals[i] = VM_PAGER_ERROR;
302 * When putting pages, do not extend file past EOF.
304 if (offset + count > np->n_size) {
305 count = np->n_size - offset;
309 mtx_unlock(&np->n_mtx);
312 * We use only the kva address for the buffer, but this is extremely
313 * convienient and fast.
315 bp = getpbuf(&ncl_pbuf_freecnt);
317 kva = (vm_offset_t) bp->b_data;
318 pmap_qenter(kva, pages, npages);
319 PCPU_INC(cnt.v_vnodeout);
320 PCPU_ADD(cnt.v_vnodepgsout, count);
322 iov.iov_base = (caddr_t) kva;
326 uio.uio_offset = offset;
327 uio.uio_resid = count;
328 uio.uio_segflg = UIO_SYSSPACE;
329 uio.uio_rw = UIO_WRITE;
332 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
333 iomode = NFSWRITE_UNSTABLE;
335 iomode = NFSWRITE_FILESYNC;
337 error = ncl_writerpc(vp, &uio, cred, &iomode, &must_commit, 0);
339 pmap_qremove(kva, npages);
340 relpbuf(bp, &ncl_pbuf_freecnt);
342 if (error == 0 || !newnfs_keep_dirty_on_error) {
343 vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid);
345 ncl_clearcommit(vp->v_mount);
351 * For nfs, cache consistency can only be maintained approximately.
352 * Although RFC1094 does not specify the criteria, the following is
353 * believed to be compatible with the reference port.
355 * If the file's modify time on the server has changed since the
356 * last read rpc or you have written to the file,
357 * you may have lost data cache consistency with the
358 * server, so flush all of the file's data out of the cache.
359 * Then force a getattr rpc to ensure that you have up to date
361 * NB: This implies that cache data can be read when up to
362 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
363 * attributes this could be forced by setting n_attrstamp to 0 before
364 * the VOP_GETATTR() call.
367 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
371 struct nfsnode *np = VTONFS(vp);
375 * Grab the exclusive lock before checking whether the cache is
377 * XXX - We can make this cheaper later (by acquiring cheaper locks).
378 * But for now, this suffices.
380 old_lock = ncl_upgrade_vnlock(vp);
381 if (vp->v_iflag & VI_DOOMED) {
382 ncl_downgrade_vnlock(vp, old_lock);
386 mtx_lock(&np->n_mtx);
387 if (np->n_flag & NMODIFIED) {
388 mtx_unlock(&np->n_mtx);
389 if (vp->v_type != VREG) {
390 if (vp->v_type != VDIR)
391 panic("nfs: bioread, not dir");
393 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
398 error = VOP_GETATTR(vp, &vattr, cred);
401 mtx_lock(&np->n_mtx);
402 np->n_mtime = vattr.va_mtime;
403 mtx_unlock(&np->n_mtx);
405 mtx_unlock(&np->n_mtx);
406 error = VOP_GETATTR(vp, &vattr, cred);
409 mtx_lock(&np->n_mtx);
410 if ((np->n_flag & NSIZECHANGED)
411 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
412 mtx_unlock(&np->n_mtx);
413 if (vp->v_type == VDIR)
415 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
418 mtx_lock(&np->n_mtx);
419 np->n_mtime = vattr.va_mtime;
420 np->n_flag &= ~NSIZECHANGED;
422 mtx_unlock(&np->n_mtx);
425 ncl_downgrade_vnlock(vp, old_lock);
430 * Vnode op for read using bio
433 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
435 struct nfsnode *np = VTONFS(vp);
437 struct buf *bp, *rabp;
439 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
443 int nra, error = 0, n = 0, on = 0;
446 KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
447 if (uio->uio_resid == 0)
449 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
453 mtx_lock(&nmp->nm_mtx);
454 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
455 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
456 mtx_unlock(&nmp->nm_mtx);
457 (void)ncl_fsinfo(nmp, vp, cred, td);
458 mtx_lock(&nmp->nm_mtx);
460 if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
461 (void) newnfs_iosize(nmp);
463 tmp_off = uio->uio_offset + uio->uio_resid;
464 if (vp->v_type != VDIR &&
465 (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) {
466 mtx_unlock(&nmp->nm_mtx);
469 mtx_unlock(&nmp->nm_mtx);
471 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
472 /* No caching/ no readaheads. Just read data into the user buffer */
473 return ncl_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 NFSINCRGLOBAL(newnfsstats.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 = newnfs_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 (ncl_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 = newnfs_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 = ncl_doio(vp, bp, cred, td, 0);
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 NFSINCRGLOBAL(newnfsstats.biocache_readlinks);
569 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
571 error = newnfs_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 = ncl_doio(vp, bp, cred, td, 0);
579 bp->b_ioflags |= BIO_ERROR;
584 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
588 NFSINCRGLOBAL(newnfsstats.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 = newnfs_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 = ncl_doio(vp, bp, cred, td, 0);
607 while (error == NFSERR_BAD_COOKIE) {
609 error = ncl_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 = newnfs_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 = ncl_doio(vp, bp, cred, td, 0);
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 (ncl_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 ncl_printf(" ncl_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 = NFSWRITE_FILESYNC;
764 error = ncl_writerpc(vp, &uio, cred, &iomode,
766 KASSERT((must_commit == 0),
767 ("ncl_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(&ncl_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 = ncl_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, &ncl_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 ncl_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);
884 int n, on, error = 0;
887 KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_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)ncl_fsinfo(nmp, vp, cred, td);
904 mtx_lock(&nmp->nm_mtx);
906 if (nmp->nm_wsize == 0)
907 (void) newnfs_iosize(nmp);
908 mtx_unlock(&nmp->nm_mtx);
911 * Synchronously flush pending buffers if we are in synchronous
912 * mode or if we are appending.
914 if (ioflag & (IO_APPEND | IO_SYNC)) {
915 mtx_lock(&np->n_mtx);
916 if (np->n_flag & NMODIFIED) {
917 mtx_unlock(&np->n_mtx);
918 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
920 * Require non-blocking, synchronous writes to
921 * dirty files to inform the program it needs
922 * to fsync(2) explicitly.
924 if (ioflag & IO_NDELAY)
929 error = ncl_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 error = VOP_GETATTR(vp, &vattr, cred);
945 mtx_lock(&np->n_mtx);
946 uio->uio_offset = np->n_size;
947 mtx_unlock(&np->n_mtx);
950 if (uio->uio_offset < 0)
952 tmp_off = uio->uio_offset + uio->uio_resid;
953 if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
955 if (uio->uio_resid == 0)
958 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
959 return nfs_directio_write(vp, uio, cred, ioflag);
962 * Maybe this should be above the vnode op call, but so long as
963 * file servers have no limits, i don't think it matters
965 if (vn_rlimit_fsize(vp, uio, td))
968 biosize = vp->v_bufobj.bo_bsize;
970 * Find all of this file's B_NEEDCOMMIT buffers. If our writes
971 * would exceed the local maximum per-file write commit size when
972 * combined with those, we must decide whether to flush,
973 * go synchronous, or return error. We don't bother checking
974 * IO_UNIT -- we just make all writes atomic anyway, as there's
975 * no point optimizing for something that really won't ever happen.
977 if (!(ioflag & IO_SYNC)) {
980 mtx_lock(&np->n_mtx);
982 mtx_unlock(&np->n_mtx);
984 if (nmp->nm_wcommitsize < uio->uio_resid) {
986 * If this request could not possibly be completed
987 * without exceeding the maximum outstanding write
988 * commit size, see if we can convert it into a
989 * synchronous write operation.
991 if (ioflag & IO_NDELAY)
994 if (nflag & NMODIFIED)
996 } else if (nflag & NMODIFIED) {
998 BO_LOCK(&vp->v_bufobj);
999 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
1000 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
1002 if (bp->b_flags & B_NEEDCOMMIT)
1003 wouldcommit += bp->b_bcount;
1006 BO_UNLOCK(&vp->v_bufobj);
1008 * Since we're not operating synchronously and
1009 * bypassing the buffer cache, we are in a commit
1010 * and holding all of these buffers whether
1011 * transmitted or not. If not limited, this
1012 * will lead to the buffer cache deadlocking,
1013 * as no one else can flush our uncommitted buffers.
1015 wouldcommit += uio->uio_resid;
1017 * If we would initially exceed the maximum
1018 * outstanding write commit size, flush and restart.
1020 if (wouldcommit > nmp->nm_wcommitsize)
1024 goto flush_and_restart;
1028 NFSINCRGLOBAL(newnfsstats.biocache_writes);
1029 lbn = uio->uio_offset / biosize;
1030 on = uio->uio_offset & (biosize-1);
1031 n = min((unsigned)(biosize - on), uio->uio_resid);
1034 * Handle direct append and file extension cases, calculate
1035 * unaligned buffer size.
1037 mtx_lock(&np->n_mtx);
1038 if (uio->uio_offset == np->n_size && n) {
1039 mtx_unlock(&np->n_mtx);
1041 * Get the buffer (in its pre-append state to maintain
1042 * B_CACHE if it was previously set). Resize the
1043 * nfsnode after we have locked the buffer to prevent
1044 * readers from reading garbage.
1047 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1052 mtx_lock(&np->n_mtx);
1053 np->n_size = uio->uio_offset + n;
1054 np->n_flag |= NMODIFIED;
1055 vnode_pager_setsize(vp, np->n_size);
1056 mtx_unlock(&np->n_mtx);
1058 save = bp->b_flags & B_CACHE;
1060 allocbuf(bp, bcount);
1061 bp->b_flags |= save;
1065 * Obtain the locked cache block first, and then
1066 * adjust the file's size as appropriate.
1069 if ((off_t)lbn * biosize + bcount < np->n_size) {
1070 if ((off_t)(lbn + 1) * biosize < np->n_size)
1073 bcount = np->n_size - (off_t)lbn * biosize;
1075 mtx_unlock(&np->n_mtx);
1076 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1077 mtx_lock(&np->n_mtx);
1078 if (uio->uio_offset + n > np->n_size) {
1079 np->n_size = uio->uio_offset + n;
1080 np->n_flag |= NMODIFIED;
1081 vnode_pager_setsize(vp, np->n_size);
1083 mtx_unlock(&np->n_mtx);
1087 error = newnfs_sigintr(nmp, td);
1094 * Issue a READ if B_CACHE is not set. In special-append
1095 * mode, B_CACHE is based on the buffer prior to the write
1096 * op and is typically set, avoiding the read. If a read
1097 * is required in special append mode, the server will
1098 * probably send us a short-read since we extended the file
1099 * on our end, resulting in b_resid == 0 and, thusly,
1100 * B_CACHE getting set.
1102 * We can also avoid issuing the read if the write covers
1103 * the entire buffer. We have to make sure the buffer state
1104 * is reasonable in this case since we will not be initiating
1105 * I/O. See the comments in kern/vfs_bio.c's getblk() for
1108 * B_CACHE may also be set due to the buffer being cached
1112 if (on == 0 && n == bcount) {
1113 bp->b_flags |= B_CACHE;
1114 bp->b_flags &= ~B_INVAL;
1115 bp->b_ioflags &= ~BIO_ERROR;
1118 if ((bp->b_flags & B_CACHE) == 0) {
1119 bp->b_iocmd = BIO_READ;
1120 vfs_busy_pages(bp, 0);
1121 error = ncl_doio(vp, bp, cred, td, 0);
1127 if (bp->b_wcred == NOCRED)
1128 bp->b_wcred = crhold(cred);
1129 mtx_lock(&np->n_mtx);
1130 np->n_flag |= NMODIFIED;
1131 mtx_unlock(&np->n_mtx);
1134 * If dirtyend exceeds file size, chop it down. This should
1135 * not normally occur but there is an append race where it
1136 * might occur XXX, so we log it.
1138 * If the chopping creates a reverse-indexed or degenerate
1139 * situation with dirtyoff/end, we 0 both of them.
1142 if (bp->b_dirtyend > bcount) {
1143 ncl_printf("NFS append race @%lx:%d\n",
1144 (long)bp->b_blkno * DEV_BSIZE,
1145 bp->b_dirtyend - bcount);
1146 bp->b_dirtyend = bcount;
1149 if (bp->b_dirtyoff >= bp->b_dirtyend)
1150 bp->b_dirtyoff = bp->b_dirtyend = 0;
1153 * If the new write will leave a contiguous dirty
1154 * area, just update the b_dirtyoff and b_dirtyend,
1155 * otherwise force a write rpc of the old dirty area.
1157 * While it is possible to merge discontiguous writes due to
1158 * our having a B_CACHE buffer ( and thus valid read data
1159 * for the hole), we don't because it could lead to
1160 * significant cache coherency problems with multiple clients,
1161 * especially if locking is implemented later on.
1163 * as an optimization we could theoretically maintain
1164 * a linked list of discontinuous areas, but we would still
1165 * have to commit them separately so there isn't much
1166 * advantage to it except perhaps a bit of asynchronization.
1169 if (bp->b_dirtyend > 0 &&
1170 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1171 if (bwrite(bp) == EINTR) {
1178 error = uiomove((char *)bp->b_data + on, n, uio);
1181 * Since this block is being modified, it must be written
1182 * again and not just committed. Since write clustering does
1183 * not work for the stage 1 data write, only the stage 2
1184 * commit rpc, we have to clear B_CLUSTEROK as well.
1186 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1189 bp->b_ioflags |= BIO_ERROR;
1195 * Only update dirtyoff/dirtyend if not a degenerate
1199 if (bp->b_dirtyend > 0) {
1200 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1201 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1203 bp->b_dirtyoff = on;
1204 bp->b_dirtyend = on + n;
1206 vfs_bio_set_valid(bp, on, n);
1210 * If IO_SYNC do bwrite().
1212 * IO_INVAL appears to be unused. The idea appears to be
1213 * to turn off caching in this case. Very odd. XXX
1215 if ((ioflag & IO_SYNC)) {
1216 if (ioflag & IO_INVAL)
1217 bp->b_flags |= B_NOCACHE;
1221 } else if ((n + on) == biosize) {
1222 bp->b_flags |= B_ASYNC;
1223 (void) ncl_writebp(bp, 0, NULL);
1227 } while (uio->uio_resid > 0 && n > 0);
1233 * Get an nfs cache block.
1235 * Allocate a new one if the block isn't currently in the cache
1236 * and return the block marked busy. If the calling process is
1237 * interrupted by a signal for an interruptible mount point, return
1240 * The caller must carefully deal with the possible B_INVAL state of
1241 * the buffer. ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
1242 * indirectly), so synchronous reads can be issued without worrying about
1243 * the B_INVAL state. We have to be a little more careful when dealing
1244 * with writes (see comments in nfs_write()) when extending a file past
1248 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1252 struct nfsmount *nmp;
1257 if (nmp->nm_flag & NFSMNT_INT) {
1260 newnfs_set_sigmask(td, &oldset);
1261 bp = getblk(vp, bn, size, NFS_PCATCH, 0, 0);
1262 newnfs_restore_sigmask(td, &oldset);
1263 while (bp == NULL) {
1264 if (newnfs_sigintr(nmp, td))
1266 bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1269 bp = getblk(vp, bn, size, 0, 0, 0);
1272 if (vp->v_type == VREG)
1273 bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1278 * Flush and invalidate all dirty buffers. If another process is already
1279 * doing the flush, just wait for completion.
1282 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1284 struct nfsnode *np = VTONFS(vp);
1285 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1286 int error = 0, slpflag, slptimeo;
1289 ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
1291 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1293 if ((nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF))
1296 slpflag = NFS_PCATCH;
1303 old_lock = ncl_upgrade_vnlock(vp);
1304 if (vp->v_iflag & VI_DOOMED) {
1306 * Since vgonel() uses the generic vinvalbuf() to flush
1307 * dirty buffers and it does not call this function, it
1308 * is safe to just return OK when VI_DOOMED is set.
1310 ncl_downgrade_vnlock(vp, old_lock);
1315 * Now, flush as required.
1317 if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) {
1318 VM_OBJECT_LOCK(vp->v_bufobj.bo_object);
1319 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1320 VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object);
1322 * If the page clean was interrupted, fail the invalidation.
1323 * Not doing so, we run the risk of losing dirty pages in the
1324 * vinvalbuf() call below.
1326 if (intrflg && (error = newnfs_sigintr(nmp, td)))
1330 error = vinvalbuf(vp, flags, slpflag, 0);
1332 if (intrflg && (error = newnfs_sigintr(nmp, td)))
1334 error = vinvalbuf(vp, flags, 0, slptimeo);
1336 mtx_lock(&np->n_mtx);
1337 if (np->n_directio_asyncwr == 0)
1338 np->n_flag &= ~NMODIFIED;
1339 mtx_unlock(&np->n_mtx);
1341 ncl_downgrade_vnlock(vp, old_lock);
1346 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1347 * This is mainly to avoid queueing async I/O requests when the nfsiods
1348 * are all hung on a dead server.
1350 * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1351 * is eventually dequeued by the async daemon, ncl_doio() *will*.
1354 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1363 * Commits are usually short and sweet so lets save some cpu and
1364 * leave the async daemons for more important rpc's (such as reads
1367 mtx_lock(&ncl_iod_mutex);
1368 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1369 (nmp->nm_bufqiods > ncl_numasync / 2)) {
1370 mtx_unlock(&ncl_iod_mutex);
1374 if (nmp->nm_flag & NFSMNT_INT)
1375 slpflag = NFS_PCATCH;
1379 * Find a free iod to process this request.
1381 for (iod = 0; iod < ncl_numasync; iod++)
1382 if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1388 * Try to create one if none are free.
1394 * Found one, so wake it up and tell it which
1397 NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1399 ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1400 ncl_iodmount[iod] = nmp;
1402 wakeup(&ncl_iodwant[iod]);
1406 * If none are free, we may already have an iod working on this mount
1407 * point. If so, it will process our request.
1410 if (nmp->nm_bufqiods > 0) {
1412 ("ncl_asyncio: %d iods are already processing mount %p\n",
1413 nmp->nm_bufqiods, nmp));
1419 * If we have an iod which can process the request, then queue
1424 * Ensure that the queue never grows too large. We still want
1425 * to asynchronize so we block rather then return EIO.
1427 while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1429 ("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1430 nmp->nm_bufqwant = TRUE;
1431 error = newnfs_msleep(td, &nmp->nm_bufq,
1432 &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1435 error2 = newnfs_sigintr(nmp, td);
1437 mtx_unlock(&ncl_iod_mutex);
1440 if (slpflag == NFS_PCATCH) {
1446 * We might have lost our iod while sleeping,
1447 * so check and loop if nescessary.
1452 /* We might have lost our nfsiod */
1453 if (nmp->nm_bufqiods == 0) {
1455 ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1459 if (bp->b_iocmd == BIO_READ) {
1460 if (bp->b_rcred == NOCRED && cred != NOCRED)
1461 bp->b_rcred = crhold(cred);
1463 if (bp->b_wcred == NOCRED && cred != NOCRED)
1464 bp->b_wcred = crhold(cred);
1467 if (bp->b_flags & B_REMFREE)
1470 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1472 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1473 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx);
1474 VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1475 VTONFS(bp->b_vp)->n_directio_asyncwr++;
1476 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx);
1478 mtx_unlock(&ncl_iod_mutex);
1482 mtx_unlock(&ncl_iod_mutex);
1485 * All the iods are busy on other mounts, so return EIO to
1486 * force the caller to process the i/o synchronously.
1488 NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1493 ncl_doio_directwrite(struct buf *bp)
1495 int iomode, must_commit;
1496 struct uio *uiop = (struct uio *)bp->b_caller1;
1497 char *iov_base = uiop->uio_iov->iov_base;
1499 iomode = NFSWRITE_FILESYNC;
1500 uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1501 ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0);
1502 KASSERT((must_commit == 0), ("ncl_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 np->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, &ncl_pbuf_freecnt);
1524 * Do an I/O operation to/from a cache block. This may be called
1525 * synchronously or from an nfsiod.
1528 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1529 int called_from_strategy)
1533 struct nfsmount *nmp;
1534 int error = 0, iomode, must_commit = 0;
1537 struct proc *p = td ? td->td_proc : NULL;
1541 nmp = VFSTONFS(vp->v_mount);
1543 uiop->uio_iov = &io;
1544 uiop->uio_iovcnt = 1;
1545 uiop->uio_segflg = UIO_SYSSPACE;
1549 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We
1550 * do this here so we do not have to do it in all the code that
1553 bp->b_flags &= ~B_INVAL;
1554 bp->b_ioflags &= ~BIO_ERROR;
1556 KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1557 iocmd = bp->b_iocmd;
1558 if (iocmd == BIO_READ) {
1559 io.iov_len = uiop->uio_resid = bp->b_bcount;
1560 io.iov_base = bp->b_data;
1561 uiop->uio_rw = UIO_READ;
1563 switch (vp->v_type) {
1565 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1566 NFSINCRGLOBAL(newnfsstats.read_bios);
1567 error = ncl_readrpc(vp, uiop, cr);
1570 if (uiop->uio_resid) {
1572 * If we had a short read with no error, we must have
1573 * hit a file hole. We should zero-fill the remainder.
1574 * This can also occur if the server hits the file EOF.
1576 * Holes used to be able to occur due to pending
1577 * writes, but that is not possible any longer.
1579 int nread = bp->b_bcount - uiop->uio_resid;
1580 int left = uiop->uio_resid;
1583 bzero((char *)bp->b_data + nread, left);
1584 uiop->uio_resid = 0;
1587 /* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1588 if (p && (vp->v_vflag & VV_TEXT)) {
1589 mtx_lock(&np->n_mtx);
1590 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1591 mtx_unlock(&np->n_mtx);
1593 killproc(p, "text file modification");
1596 mtx_unlock(&np->n_mtx);
1600 uiop->uio_offset = (off_t)0;
1601 NFSINCRGLOBAL(newnfsstats.readlink_bios);
1602 error = ncl_readlinkrpc(vp, uiop, cr);
1605 NFSINCRGLOBAL(newnfsstats.readdir_bios);
1606 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1607 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1608 error = ncl_readdirplusrpc(vp, uiop, cr, td);
1609 if (error == NFSERR_NOTSUPP)
1610 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1612 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1613 error = ncl_readdirrpc(vp, uiop, cr, td);
1615 * end-of-directory sets B_INVAL but does not generate an
1618 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1619 bp->b_flags |= B_INVAL;
1622 ncl_printf("ncl_doio: type %x unexpected\n", vp->v_type);
1626 bp->b_ioflags |= BIO_ERROR;
1627 bp->b_error = error;
1631 * If we only need to commit, try to commit
1633 if (bp->b_flags & B_NEEDCOMMIT) {
1637 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1638 retv = ncl_commit(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 ncl_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 NFSINCRGLOBAL(newnfsstats.write_bios);
1669 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1670 iomode = NFSWRITE_UNSTABLE;
1672 iomode = NFSWRITE_FILESYNC;
1674 error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1675 called_from_strategy);
1678 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1679 * to cluster the buffers needing commit. This will allow
1680 * the system to submit a single commit rpc for the whole
1681 * cluster. We can do this even if the buffer is not 100%
1682 * dirty (relative to the NFS blocksize), so we optimize the
1683 * append-to-file-case.
1685 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1686 * cleared because write clustering only works for commit
1687 * rpc's, not for the data portion of the write).
1690 if (!error && iomode == NFSWRITE_UNSTABLE) {
1691 bp->b_flags |= B_NEEDCOMMIT;
1692 if (bp->b_dirtyoff == 0
1693 && bp->b_dirtyend == bp->b_bcount)
1694 bp->b_flags |= B_CLUSTEROK;
1696 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1700 * For an interrupted write, the buffer is still valid
1701 * and the write hasn't been pushed to the server yet,
1702 * so we can't set BIO_ERROR and report the interruption
1703 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1704 * is not relevant, so the rpc attempt is essentially
1705 * a noop. For the case of a V3 write rpc not being
1706 * committed to stable storage, the block is still
1707 * dirty and requires either a commit rpc or another
1708 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1709 * the block is reused. This is indicated by setting
1710 * the B_DELWRI and B_NEEDCOMMIT flags.
1712 * EIO is returned by ncl_writerpc() to indicate a recoverable
1713 * write error and is handled as above, except that
1714 * B_EINTR isn't set. One cause of this is a stale stateid
1715 * error for the RPC that indicates recovery is required,
1716 * when called with called_from_strategy != 0.
1718 * If the buffer is marked B_PAGING, it does not reside on
1719 * the vp's paging queues so we cannot call bdirty(). The
1720 * bp in this case is not an NFS cache block so we should
1723 * The logic below breaks up errors into recoverable and
1724 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1725 * and keep the buffer around for potential write retries.
1726 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1727 * and save the error in the nfsnode. This is less than ideal
1728 * but necessary. Keeping such buffers around could potentially
1729 * cause buffer exhaustion eventually (they can never be written
1730 * out, so will get constantly be re-dirtied). It also causes
1731 * all sorts of vfs panics. For non-recoverable write errors,
1732 * also invalidate the attrcache, so we'll be forced to go over
1733 * the wire for this object, returning an error to user on next
1734 * call (most of the time).
1736 if (error == EINTR || error == EIO || error == ETIMEDOUT
1737 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1741 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1742 if ((bp->b_flags & B_PAGING) == 0) {
1744 bp->b_flags &= ~B_DONE;
1746 if ((error == EINTR || error == ETIMEDOUT) &&
1747 (bp->b_flags & B_ASYNC) == 0)
1748 bp->b_flags |= B_EINTR;
1752 bp->b_ioflags |= BIO_ERROR;
1753 bp->b_flags |= B_INVAL;
1754 bp->b_error = np->n_error = error;
1755 mtx_lock(&np->n_mtx);
1756 np->n_flag |= NWRITEERR;
1757 np->n_attrstamp = 0;
1758 mtx_unlock(&np->n_mtx);
1760 bp->b_dirtyoff = bp->b_dirtyend = 0;
1768 bp->b_resid = uiop->uio_resid;
1770 ncl_clearcommit(vp->v_mount);
1776 * Used to aid in handling ftruncate() operations on the NFS client side.
1777 * Truncation creates a number of special problems for NFS. We have to
1778 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1779 * we have to properly handle VM pages or (potentially dirty) buffers
1780 * that straddle the truncation point.
1784 ncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1786 struct nfsnode *np = VTONFS(vp);
1788 int biosize = vp->v_bufobj.bo_bsize;
1791 mtx_lock(&np->n_mtx);
1794 mtx_unlock(&np->n_mtx);
1796 if (nsize < tsize) {
1802 * vtruncbuf() doesn't get the buffer overlapping the
1803 * truncation point. We may have a B_DELWRI and/or B_CACHE
1804 * buffer that now needs to be truncated.
1806 error = vtruncbuf(vp, cred, td, nsize, biosize);
1807 lbn = nsize / biosize;
1808 bufsize = nsize & (biosize - 1);
1809 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1812 if (bp->b_dirtyoff > bp->b_bcount)
1813 bp->b_dirtyoff = bp->b_bcount;
1814 if (bp->b_dirtyend > bp->b_bcount)
1815 bp->b_dirtyend = bp->b_bcount;
1816 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1819 vnode_pager_setsize(vp, nsize);