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>
45 #include <sys/mount.h>
46 #include <sys/vmmeter.h>
47 #include <sys/vnode.h>
50 #include <vm/vm_param.h>
51 #include <vm/vm_extern.h>
52 #include <vm/vm_page.h>
53 #include <vm/vm_object.h>
54 #include <vm/vm_pager.h>
55 #include <vm/vnode_pager.h>
57 #include <fs/nfs/nfsport.h>
58 #include <fs/nfsclient/nfsmount.h>
59 #include <fs/nfsclient/nfs.h>
60 #include <fs/nfsclient/nfsnode.h>
61 #include <fs/nfsclient/nfs_kdtrace.h>
63 extern int newnfs_directio_allow_mmap;
64 extern struct nfsstats newnfsstats;
65 extern struct mtx ncl_iod_mutex;
66 extern int ncl_numasync;
67 extern enum nfsiod_state ncl_iodwant[NFS_MAXASYNCDAEMON];
68 extern struct nfsmount *ncl_iodmount[NFS_MAXASYNCDAEMON];
69 extern int newnfs_directio_enable;
70 extern int nfs_keep_dirty_on_error;
72 int ncl_pbuf_freecnt = -1; /* start out unlimited */
74 static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size,
76 static int nfs_directio_write(struct vnode *vp, struct uio *uiop,
77 struct ucred *cred, int ioflag);
80 * Vnode op for VM getpages.
83 ncl_getpages(struct vop_getpages_args *ap)
85 int i, error, nextoff, size, toff, count, npages;
100 td = curthread; /* XXX */
101 cred = curthread->td_ucred; /* XXX */
102 nmp = VFSTONFS(vp->v_mount);
106 if ((object = vp->v_object) == NULL) {
107 ncl_printf("nfs_getpages: called with non-merged cache vnode??\n");
108 return (VM_PAGER_ERROR);
111 if (newnfs_directio_enable && !newnfs_directio_allow_mmap) {
112 mtx_lock(&np->n_mtx);
113 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
114 mtx_unlock(&np->n_mtx);
115 ncl_printf("nfs_getpages: called on non-cacheable vnode??\n");
116 return (VM_PAGER_ERROR);
118 mtx_unlock(&np->n_mtx);
121 mtx_lock(&nmp->nm_mtx);
122 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
123 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
124 mtx_unlock(&nmp->nm_mtx);
125 /* We'll never get here for v4, because we always have fsinfo */
126 (void)ncl_fsinfo(nmp, vp, cred, td);
128 mtx_unlock(&nmp->nm_mtx);
130 npages = btoc(count);
133 * If the requested page is partially valid, just return it and
134 * allow the pager to zero-out the blanks. Partially valid pages
135 * can only occur at the file EOF.
137 VM_OBJECT_LOCK(object);
138 if (pages[ap->a_reqpage]->valid != 0) {
139 for (i = 0; i < npages; ++i) {
140 if (i != ap->a_reqpage) {
141 vm_page_lock(pages[i]);
142 vm_page_free(pages[i]);
143 vm_page_unlock(pages[i]);
146 VM_OBJECT_UNLOCK(object);
149 VM_OBJECT_UNLOCK(object);
152 * We use only the kva address for the buffer, but this is extremely
153 * convienient and fast.
155 bp = getpbuf(&ncl_pbuf_freecnt);
157 kva = (vm_offset_t) bp->b_data;
158 pmap_qenter(kva, pages, npages);
159 PCPU_INC(cnt.v_vnodein);
160 PCPU_ADD(cnt.v_vnodepgsin, npages);
162 iov.iov_base = (caddr_t) kva;
166 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
167 uio.uio_resid = count;
168 uio.uio_segflg = UIO_SYSSPACE;
169 uio.uio_rw = UIO_READ;
172 error = ncl_readrpc(vp, &uio, cred);
173 pmap_qremove(kva, npages);
175 relpbuf(bp, &ncl_pbuf_freecnt);
177 if (error && (uio.uio_resid == count)) {
178 ncl_printf("nfs_getpages: error %d\n", error);
179 VM_OBJECT_LOCK(object);
180 for (i = 0; i < npages; ++i) {
181 if (i != ap->a_reqpage) {
182 vm_page_lock(pages[i]);
183 vm_page_free(pages[i]);
184 vm_page_unlock(pages[i]);
187 VM_OBJECT_UNLOCK(object);
188 return (VM_PAGER_ERROR);
192 * Calculate the number of bytes read and validate only that number
193 * of bytes. Note that due to pending writes, size may be 0. This
194 * does not mean that the remaining data is invalid!
197 size = count - uio.uio_resid;
198 VM_OBJECT_LOCK(object);
199 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
201 nextoff = toff + PAGE_SIZE;
204 if (nextoff <= size) {
206 * Read operation filled an entire page
208 m->valid = VM_PAGE_BITS_ALL;
209 KASSERT(m->dirty == 0,
210 ("nfs_getpages: page %p is dirty", m));
211 } else if (size > toff) {
213 * Read operation filled a partial page.
216 vm_page_set_valid(m, 0, size - toff);
217 KASSERT(m->dirty == 0,
218 ("nfs_getpages: page %p is dirty", m));
221 * Read operation was short. If no error
222 * occured we may have hit a zero-fill
223 * section. We leave valid set to 0, and page
224 * is freed by vm_page_readahead_finish() if
225 * its index is not equal to requested, or
226 * page is zeroed and set valid by
227 * vm_pager_get_pages() for requested page.
231 if (i != ap->a_reqpage)
232 vm_page_readahead_finish(m);
234 VM_OBJECT_UNLOCK(object);
239 * Vnode op for VM putpages.
242 ncl_putpages(struct vop_putpages_args *ap)
248 int iomode, must_commit, i, error, npages, count;
254 struct nfsmount *nmp;
260 td = curthread; /* XXX */
261 /* Set the cred to n_writecred for the write rpcs. */
262 if (np->n_writecred != NULL)
263 cred = crhold(np->n_writecred);
265 cred = crhold(curthread->td_ucred); /* XXX */
266 nmp = VFSTONFS(vp->v_mount);
269 rtvals = ap->a_rtvals;
270 npages = btoc(count);
271 offset = IDX_TO_OFF(pages[0]->pindex);
273 mtx_lock(&nmp->nm_mtx);
274 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
275 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
276 mtx_unlock(&nmp->nm_mtx);
277 (void)ncl_fsinfo(nmp, vp, cred, td);
279 mtx_unlock(&nmp->nm_mtx);
281 mtx_lock(&np->n_mtx);
282 if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
283 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
284 mtx_unlock(&np->n_mtx);
285 ncl_printf("ncl_putpages: called on noncache-able vnode??\n");
286 mtx_lock(&np->n_mtx);
289 for (i = 0; i < npages; i++)
290 rtvals[i] = VM_PAGER_ERROR;
293 * When putting pages, do not extend file past EOF.
295 if (offset + count > np->n_size) {
296 count = np->n_size - offset;
300 mtx_unlock(&np->n_mtx);
303 * We use only the kva address for the buffer, but this is extremely
304 * convienient and fast.
306 bp = getpbuf(&ncl_pbuf_freecnt);
308 kva = (vm_offset_t) bp->b_data;
309 pmap_qenter(kva, pages, npages);
310 PCPU_INC(cnt.v_vnodeout);
311 PCPU_ADD(cnt.v_vnodepgsout, count);
313 iov.iov_base = (caddr_t) kva;
317 uio.uio_offset = offset;
318 uio.uio_resid = count;
319 uio.uio_segflg = UIO_SYSSPACE;
320 uio.uio_rw = UIO_WRITE;
323 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
324 iomode = NFSWRITE_UNSTABLE;
326 iomode = NFSWRITE_FILESYNC;
328 error = ncl_writerpc(vp, &uio, cred, &iomode, &must_commit, 0);
331 pmap_qremove(kva, npages);
332 relpbuf(bp, &ncl_pbuf_freecnt);
334 if (error == 0 || !nfs_keep_dirty_on_error) {
335 vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid);
337 ncl_clearcommit(vp->v_mount);
343 * For nfs, cache consistency can only be maintained approximately.
344 * Although RFC1094 does not specify the criteria, the following is
345 * believed to be compatible with the reference port.
347 * If the file's modify time on the server has changed since the
348 * last read rpc or you have written to the file,
349 * you may have lost data cache consistency with the
350 * server, so flush all of the file's data out of the cache.
351 * Then force a getattr rpc to ensure that you have up to date
353 * NB: This implies that cache data can be read when up to
354 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
355 * attributes this could be forced by setting n_attrstamp to 0 before
356 * the VOP_GETATTR() call.
359 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
363 struct nfsnode *np = VTONFS(vp);
367 * Grab the exclusive lock before checking whether the cache is
369 * XXX - We can make this cheaper later (by acquiring cheaper locks).
370 * But for now, this suffices.
372 old_lock = ncl_upgrade_vnlock(vp);
373 if (vp->v_iflag & VI_DOOMED) {
374 ncl_downgrade_vnlock(vp, old_lock);
378 mtx_lock(&np->n_mtx);
379 if (np->n_flag & NMODIFIED) {
380 mtx_unlock(&np->n_mtx);
381 if (vp->v_type != VREG) {
382 if (vp->v_type != VDIR)
383 panic("nfs: bioread, not dir");
385 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
390 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
391 error = VOP_GETATTR(vp, &vattr, cred);
394 mtx_lock(&np->n_mtx);
395 np->n_mtime = vattr.va_mtime;
396 mtx_unlock(&np->n_mtx);
398 mtx_unlock(&np->n_mtx);
399 error = VOP_GETATTR(vp, &vattr, cred);
402 mtx_lock(&np->n_mtx);
403 if ((np->n_flag & NSIZECHANGED)
404 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
405 mtx_unlock(&np->n_mtx);
406 if (vp->v_type == VDIR)
408 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
411 mtx_lock(&np->n_mtx);
412 np->n_mtime = vattr.va_mtime;
413 np->n_flag &= ~NSIZECHANGED;
415 mtx_unlock(&np->n_mtx);
418 ncl_downgrade_vnlock(vp, old_lock);
423 * Vnode op for read using bio
426 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
428 struct nfsnode *np = VTONFS(vp);
430 struct buf *bp, *rabp;
432 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
436 int nra, error = 0, n = 0, on = 0;
439 KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
440 if (uio->uio_resid == 0)
442 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
446 mtx_lock(&nmp->nm_mtx);
447 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
448 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
449 mtx_unlock(&nmp->nm_mtx);
450 (void)ncl_fsinfo(nmp, vp, cred, td);
451 mtx_lock(&nmp->nm_mtx);
453 if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
454 (void) newnfs_iosize(nmp);
456 tmp_off = uio->uio_offset + uio->uio_resid;
457 if (vp->v_type != VDIR &&
458 (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) {
459 mtx_unlock(&nmp->nm_mtx);
462 mtx_unlock(&nmp->nm_mtx);
464 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
465 /* No caching/ no readaheads. Just read data into the user buffer */
466 return ncl_readrpc(vp, uio, cred);
468 biosize = vp->v_bufobj.bo_bsize;
469 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
471 error = nfs_bioread_check_cons(vp, td, cred);
478 mtx_lock(&np->n_mtx);
480 mtx_unlock(&np->n_mtx);
482 switch (vp->v_type) {
484 NFSINCRGLOBAL(newnfsstats.biocache_reads);
485 lbn = uio->uio_offset / biosize;
486 on = uio->uio_offset & (biosize - 1);
489 * Start the read ahead(s), as required.
491 if (nmp->nm_readahead > 0) {
492 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
493 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
494 rabn = lbn + 1 + nra;
495 if (incore(&vp->v_bufobj, rabn) == NULL) {
496 rabp = nfs_getcacheblk(vp, rabn, biosize, td);
498 error = newnfs_sigintr(nmp, td);
499 return (error ? error : EINTR);
501 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
502 rabp->b_flags |= B_ASYNC;
503 rabp->b_iocmd = BIO_READ;
504 vfs_busy_pages(rabp, 0);
505 if (ncl_asyncio(nmp, rabp, cred, td)) {
506 rabp->b_flags |= B_INVAL;
507 rabp->b_ioflags |= BIO_ERROR;
508 vfs_unbusy_pages(rabp);
519 /* Note that bcount is *not* DEV_BSIZE aligned. */
521 if ((off_t)lbn * biosize >= nsize) {
523 } else if ((off_t)(lbn + 1) * biosize > nsize) {
524 bcount = nsize - (off_t)lbn * biosize;
526 bp = nfs_getcacheblk(vp, lbn, bcount, td);
529 error = newnfs_sigintr(nmp, td);
530 return (error ? error : EINTR);
534 * If B_CACHE is not set, we must issue the read. If this
535 * fails, we return an error.
538 if ((bp->b_flags & B_CACHE) == 0) {
539 bp->b_iocmd = BIO_READ;
540 vfs_busy_pages(bp, 0);
541 error = ncl_doio(vp, bp, cred, td, 0);
549 * on is the offset into the current bp. Figure out how many
550 * bytes we can copy out of the bp. Note that bcount is
551 * NOT DEV_BSIZE aligned.
553 * Then figure out how many bytes we can copy into the uio.
558 n = MIN((unsigned)(bcount - on), uio->uio_resid);
561 NFSINCRGLOBAL(newnfsstats.biocache_readlinks);
562 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
564 error = newnfs_sigintr(nmp, td);
565 return (error ? error : EINTR);
567 if ((bp->b_flags & B_CACHE) == 0) {
568 bp->b_iocmd = BIO_READ;
569 vfs_busy_pages(bp, 0);
570 error = ncl_doio(vp, bp, cred, td, 0);
572 bp->b_ioflags |= BIO_ERROR;
577 n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
581 NFSINCRGLOBAL(newnfsstats.biocache_readdirs);
582 if (np->n_direofoffset
583 && uio->uio_offset >= np->n_direofoffset) {
586 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
587 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
588 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
590 error = newnfs_sigintr(nmp, td);
591 return (error ? error : EINTR);
593 if ((bp->b_flags & B_CACHE) == 0) {
594 bp->b_iocmd = BIO_READ;
595 vfs_busy_pages(bp, 0);
596 error = ncl_doio(vp, bp, cred, td, 0);
600 while (error == NFSERR_BAD_COOKIE) {
602 error = ncl_vinvalbuf(vp, 0, td, 1);
604 * Yuck! The directory has been modified on the
605 * server. The only way to get the block is by
606 * reading from the beginning to get all the
609 * Leave the last bp intact unless there is an error.
610 * Loop back up to the while if the error is another
611 * NFSERR_BAD_COOKIE (double yuch!).
613 for (i = 0; i <= lbn && !error; i++) {
614 if (np->n_direofoffset
615 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
617 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
619 error = newnfs_sigintr(nmp, td);
620 return (error ? error : EINTR);
622 if ((bp->b_flags & B_CACHE) == 0) {
623 bp->b_iocmd = BIO_READ;
624 vfs_busy_pages(bp, 0);
625 error = ncl_doio(vp, bp, cred, td, 0);
627 * no error + B_INVAL == directory EOF,
630 if (error == 0 && (bp->b_flags & B_INVAL))
634 * An error will throw away the block and the
635 * for loop will break out. If no error and this
636 * is not the block we want, we throw away the
637 * block and go for the next one via the for loop.
639 if (error || i < lbn)
644 * The above while is repeated if we hit another cookie
645 * error. If we hit an error and it wasn't a cookie error,
653 * If not eof and read aheads are enabled, start one.
654 * (You need the current block first, so that you have the
655 * directory offset cookie of the next block.)
657 if (nmp->nm_readahead > 0 &&
658 (bp->b_flags & B_INVAL) == 0 &&
659 (np->n_direofoffset == 0 ||
660 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
661 incore(&vp->v_bufobj, lbn + 1) == NULL) {
662 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
664 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
665 rabp->b_flags |= B_ASYNC;
666 rabp->b_iocmd = BIO_READ;
667 vfs_busy_pages(rabp, 0);
668 if (ncl_asyncio(nmp, rabp, cred, td)) {
669 rabp->b_flags |= B_INVAL;
670 rabp->b_ioflags |= BIO_ERROR;
671 vfs_unbusy_pages(rabp);
680 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
681 * chopped for the EOF condition, we cannot tell how large
682 * NFS directories are going to be until we hit EOF. So
683 * an NFS directory buffer is *not* chopped to its EOF. Now,
684 * it just so happens that b_resid will effectively chop it
685 * to EOF. *BUT* this information is lost if the buffer goes
686 * away and is reconstituted into a B_CACHE state ( due to
687 * being VMIO ) later. So we keep track of the directory eof
688 * in np->n_direofoffset and chop it off as an extra step
691 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
692 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
693 n = np->n_direofoffset - uio->uio_offset;
696 ncl_printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
702 error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio);
704 if (vp->v_type == VLNK)
708 } while (error == 0 && uio->uio_resid > 0 && n > 0);
713 * The NFS write path cannot handle iovecs with len > 1. So we need to
714 * break up iovecs accordingly (restricting them to wsize).
715 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
716 * For the ASYNC case, 2 copies are needed. The first a copy from the
717 * user buffer to a staging buffer and then a second copy from the staging
718 * buffer to mbufs. This can be optimized by copying from the user buffer
719 * directly into mbufs and passing the chain down, but that requires a
720 * fair amount of re-working of the relevant codepaths (and can be done
724 nfs_directio_write(vp, uiop, cred, ioflag)
731 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
732 struct thread *td = uiop->uio_td;
736 mtx_lock(&nmp->nm_mtx);
737 wsize = nmp->nm_wsize;
738 mtx_unlock(&nmp->nm_mtx);
739 if (ioflag & IO_SYNC) {
740 int iomode, must_commit;
744 while (uiop->uio_resid > 0) {
745 size = MIN(uiop->uio_resid, wsize);
746 size = MIN(uiop->uio_iov->iov_len, size);
747 iov.iov_base = uiop->uio_iov->iov_base;
751 uio.uio_offset = uiop->uio_offset;
752 uio.uio_resid = size;
753 uio.uio_segflg = UIO_USERSPACE;
754 uio.uio_rw = UIO_WRITE;
756 iomode = NFSWRITE_FILESYNC;
757 error = ncl_writerpc(vp, &uio, cred, &iomode,
759 KASSERT((must_commit == 0),
760 ("ncl_directio_write: Did not commit write"));
763 uiop->uio_offset += size;
764 uiop->uio_resid -= size;
765 if (uiop->uio_iov->iov_len <= size) {
769 uiop->uio_iov->iov_base =
770 (char *)uiop->uio_iov->iov_base + size;
771 uiop->uio_iov->iov_len -= size;
780 * Break up the write into blocksize chunks and hand these
781 * over to nfsiod's for write back.
782 * Unfortunately, this incurs a copy of the data. Since
783 * the user could modify the buffer before the write is
786 * The obvious optimization here is that one of the 2 copies
787 * in the async write path can be eliminated by copying the
788 * data here directly into mbufs and passing the mbuf chain
789 * down. But that will require a fair amount of re-working
790 * of the code and can be done if there's enough interest
791 * in NFS directio access.
793 while (uiop->uio_resid > 0) {
794 size = MIN(uiop->uio_resid, wsize);
795 size = MIN(uiop->uio_iov->iov_len, size);
796 bp = getpbuf(&ncl_pbuf_freecnt);
797 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
798 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
799 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
800 t_iov->iov_len = size;
801 t_uio->uio_iov = t_iov;
802 t_uio->uio_iovcnt = 1;
803 t_uio->uio_offset = uiop->uio_offset;
804 t_uio->uio_resid = size;
805 t_uio->uio_segflg = UIO_SYSSPACE;
806 t_uio->uio_rw = UIO_WRITE;
808 KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
809 uiop->uio_segflg == UIO_SYSSPACE,
810 ("nfs_directio_write: Bad uio_segflg"));
811 if (uiop->uio_segflg == UIO_USERSPACE) {
812 error = copyin(uiop->uio_iov->iov_base,
813 t_iov->iov_base, size);
818 * UIO_SYSSPACE may never happen, but handle
819 * it just in case it does.
821 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
823 bp->b_flags |= B_DIRECT;
824 bp->b_iocmd = BIO_WRITE;
825 if (cred != NOCRED) {
829 bp->b_wcred = NOCRED;
830 bp->b_caller1 = (void *)t_uio;
832 error = ncl_asyncio(nmp, bp, NOCRED, td);
835 free(t_iov->iov_base, M_NFSDIRECTIO);
836 free(t_iov, M_NFSDIRECTIO);
837 free(t_uio, M_NFSDIRECTIO);
839 relpbuf(bp, &ncl_pbuf_freecnt);
844 uiop->uio_offset += size;
845 uiop->uio_resid -= size;
846 if (uiop->uio_iov->iov_len <= size) {
850 uiop->uio_iov->iov_base =
851 (char *)uiop->uio_iov->iov_base + size;
852 uiop->uio_iov->iov_len -= size;
860 * Vnode op for write using bio
863 ncl_write(struct vop_write_args *ap)
866 struct uio *uio = ap->a_uio;
867 struct thread *td = uio->uio_td;
868 struct vnode *vp = ap->a_vp;
869 struct nfsnode *np = VTONFS(vp);
870 struct ucred *cred = ap->a_cred;
871 int ioflag = ap->a_ioflag;
874 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
877 int bp_cached, n, on, error = 0, error1;
878 size_t orig_resid, local_resid;
879 off_t orig_size, tmp_off;
881 KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
882 KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
884 if (vp->v_type != VREG)
886 mtx_lock(&np->n_mtx);
887 if (np->n_flag & NWRITEERR) {
888 np->n_flag &= ~NWRITEERR;
889 mtx_unlock(&np->n_mtx);
890 return (np->n_error);
892 mtx_unlock(&np->n_mtx);
893 mtx_lock(&nmp->nm_mtx);
894 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
895 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
896 mtx_unlock(&nmp->nm_mtx);
897 (void)ncl_fsinfo(nmp, vp, cred, td);
898 mtx_lock(&nmp->nm_mtx);
900 if (nmp->nm_wsize == 0)
901 (void) newnfs_iosize(nmp);
902 mtx_unlock(&nmp->nm_mtx);
905 * Synchronously flush pending buffers if we are in synchronous
906 * mode or if we are appending.
908 if (ioflag & (IO_APPEND | IO_SYNC)) {
909 mtx_lock(&np->n_mtx);
910 if (np->n_flag & NMODIFIED) {
911 mtx_unlock(&np->n_mtx);
912 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
914 * Require non-blocking, synchronous writes to
915 * dirty files to inform the program it needs
916 * to fsync(2) explicitly.
918 if (ioflag & IO_NDELAY)
923 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
924 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
928 mtx_unlock(&np->n_mtx);
931 orig_resid = uio->uio_resid;
932 mtx_lock(&np->n_mtx);
933 orig_size = np->n_size;
934 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 tmp_off = uio->uio_offset + uio->uio_resid;
954 if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
956 if (uio->uio_resid == 0)
959 if (newnfs_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 NFSINCRGLOBAL(newnfsstats.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 = newnfs_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
1114 if (on == 0 && n == bcount) {
1115 if ((bp->b_flags & B_CACHE) == 0)
1117 bp->b_flags |= B_CACHE;
1118 bp->b_flags &= ~B_INVAL;
1119 bp->b_ioflags &= ~BIO_ERROR;
1122 if ((bp->b_flags & B_CACHE) == 0) {
1123 bp->b_iocmd = BIO_READ;
1124 vfs_busy_pages(bp, 0);
1125 error = ncl_doio(vp, bp, cred, td, 0);
1131 if (bp->b_wcred == NOCRED)
1132 bp->b_wcred = crhold(cred);
1133 mtx_lock(&np->n_mtx);
1134 np->n_flag |= NMODIFIED;
1135 mtx_unlock(&np->n_mtx);
1138 * If dirtyend exceeds file size, chop it down. This should
1139 * not normally occur but there is an append race where it
1140 * might occur XXX, so we log it.
1142 * If the chopping creates a reverse-indexed or degenerate
1143 * situation with dirtyoff/end, we 0 both of them.
1146 if (bp->b_dirtyend > bcount) {
1147 ncl_printf("NFS append race @%lx:%d\n",
1148 (long)bp->b_blkno * DEV_BSIZE,
1149 bp->b_dirtyend - bcount);
1150 bp->b_dirtyend = bcount;
1153 if (bp->b_dirtyoff >= bp->b_dirtyend)
1154 bp->b_dirtyoff = bp->b_dirtyend = 0;
1157 * If the new write will leave a contiguous dirty
1158 * area, just update the b_dirtyoff and b_dirtyend,
1159 * otherwise force a write rpc of the old dirty area.
1161 * While it is possible to merge discontiguous writes due to
1162 * our having a B_CACHE buffer ( and thus valid read data
1163 * for the hole), we don't because it could lead to
1164 * significant cache coherency problems with multiple clients,
1165 * especially if locking is implemented later on.
1167 * As an optimization we could theoretically maintain
1168 * a linked list of discontinuous areas, but we would still
1169 * have to commit them separately so there isn't much
1170 * advantage to it except perhaps a bit of asynchronization.
1173 if (bp->b_dirtyend > 0 &&
1174 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1175 if (bwrite(bp) == EINTR) {
1182 local_resid = uio->uio_resid;
1183 error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio);
1185 if (error != 0 && !bp_cached) {
1187 * This block has no other content then what
1188 * possibly was written by the faulty uiomove.
1189 * Release it, forgetting the data pages, to
1190 * prevent the leak of uninitialized data to
1193 bp->b_ioflags |= BIO_ERROR;
1195 uio->uio_offset -= local_resid - uio->uio_resid;
1196 uio->uio_resid = local_resid;
1201 * Since this block is being modified, it must be written
1202 * again and not just committed. Since write clustering does
1203 * not work for the stage 1 data write, only the stage 2
1204 * commit rpc, we have to clear B_CLUSTEROK as well.
1206 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1209 * Get the partial update on the progress made from
1210 * uiomove, if an error occured.
1213 n = local_resid - uio->uio_resid;
1216 * Only update dirtyoff/dirtyend if not a degenerate
1220 if (bp->b_dirtyend > 0) {
1221 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1222 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1224 bp->b_dirtyoff = on;
1225 bp->b_dirtyend = on + n;
1227 vfs_bio_set_valid(bp, on, n);
1231 * If IO_SYNC do bwrite().
1233 * IO_INVAL appears to be unused. The idea appears to be
1234 * to turn off caching in this case. Very odd. XXX
1236 if ((ioflag & IO_SYNC)) {
1237 if (ioflag & IO_INVAL)
1238 bp->b_flags |= B_NOCACHE;
1239 error1 = bwrite(bp);
1245 } else if ((n + on) == biosize) {
1246 bp->b_flags |= B_ASYNC;
1247 (void) ncl_writebp(bp, 0, NULL);
1254 } while (uio->uio_resid > 0 && n > 0);
1257 if (ioflag & IO_UNIT) {
1259 vattr.va_size = orig_size;
1260 /* IO_SYNC is handled implicitely */
1261 (void)VOP_SETATTR(vp, &vattr, cred);
1262 uio->uio_offset -= orig_resid - uio->uio_resid;
1263 uio->uio_resid = orig_resid;
1271 * Get an nfs cache block.
1273 * Allocate a new one if the block isn't currently in the cache
1274 * and return the block marked busy. If the calling process is
1275 * interrupted by a signal for an interruptible mount point, return
1278 * The caller must carefully deal with the possible B_INVAL state of
1279 * the buffer. ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
1280 * indirectly), so synchronous reads can be issued without worrying about
1281 * the B_INVAL state. We have to be a little more careful when dealing
1282 * with writes (see comments in nfs_write()) when extending a file past
1286 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1290 struct nfsmount *nmp;
1295 if (nmp->nm_flag & NFSMNT_INT) {
1298 newnfs_set_sigmask(td, &oldset);
1299 bp = getblk(vp, bn, size, NFS_PCATCH, 0, 0);
1300 newnfs_restore_sigmask(td, &oldset);
1301 while (bp == NULL) {
1302 if (newnfs_sigintr(nmp, td))
1304 bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1307 bp = getblk(vp, bn, size, 0, 0, 0);
1310 if (vp->v_type == VREG)
1311 bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1316 * Flush and invalidate all dirty buffers. If another process is already
1317 * doing the flush, just wait for completion.
1320 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1322 struct nfsnode *np = VTONFS(vp);
1323 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1324 int error = 0, slpflag, slptimeo;
1327 ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
1329 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1331 if ((nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF))
1334 slpflag = NFS_PCATCH;
1341 old_lock = ncl_upgrade_vnlock(vp);
1342 if (vp->v_iflag & VI_DOOMED) {
1344 * Since vgonel() uses the generic vinvalbuf() to flush
1345 * dirty buffers and it does not call this function, it
1346 * is safe to just return OK when VI_DOOMED is set.
1348 ncl_downgrade_vnlock(vp, old_lock);
1353 * Now, flush as required.
1355 if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) {
1356 VM_OBJECT_LOCK(vp->v_bufobj.bo_object);
1357 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1358 VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object);
1360 * If the page clean was interrupted, fail the invalidation.
1361 * Not doing so, we run the risk of losing dirty pages in the
1362 * vinvalbuf() call below.
1364 if (intrflg && (error = newnfs_sigintr(nmp, td)))
1368 error = vinvalbuf(vp, flags, slpflag, 0);
1370 if (intrflg && (error = newnfs_sigintr(nmp, td)))
1372 error = vinvalbuf(vp, flags, 0, slptimeo);
1374 mtx_lock(&np->n_mtx);
1375 if (np->n_directio_asyncwr == 0)
1376 np->n_flag &= ~NMODIFIED;
1377 mtx_unlock(&np->n_mtx);
1379 ncl_downgrade_vnlock(vp, old_lock);
1384 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1385 * This is mainly to avoid queueing async I/O requests when the nfsiods
1386 * are all hung on a dead server.
1388 * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1389 * is eventually dequeued by the async daemon, ncl_doio() *will*.
1392 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1401 * Commits are usually short and sweet so lets save some cpu and
1402 * leave the async daemons for more important rpc's (such as reads
1405 * Readdirplus RPCs do vget()s to acquire the vnodes for entries
1406 * in the directory in order to update attributes. This can deadlock
1407 * with another thread that is waiting for async I/O to be done by
1408 * an nfsiod thread while holding a lock on one of these vnodes.
1409 * To avoid this deadlock, don't allow the async nfsiod threads to
1410 * perform Readdirplus RPCs.
1412 mtx_lock(&ncl_iod_mutex);
1413 if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1414 (nmp->nm_bufqiods > ncl_numasync / 2)) ||
1415 (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
1416 mtx_unlock(&ncl_iod_mutex);
1420 if (nmp->nm_flag & NFSMNT_INT)
1421 slpflag = NFS_PCATCH;
1425 * Find a free iod to process this request.
1427 for (iod = 0; iod < ncl_numasync; iod++)
1428 if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1434 * Try to create one if none are free.
1440 * Found one, so wake it up and tell it which
1443 NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1445 ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1446 ncl_iodmount[iod] = nmp;
1448 wakeup(&ncl_iodwant[iod]);
1452 * If none are free, we may already have an iod working on this mount
1453 * point. If so, it will process our request.
1456 if (nmp->nm_bufqiods > 0) {
1458 ("ncl_asyncio: %d iods are already processing mount %p\n",
1459 nmp->nm_bufqiods, nmp));
1465 * If we have an iod which can process the request, then queue
1470 * Ensure that the queue never grows too large. We still want
1471 * to asynchronize so we block rather then return EIO.
1473 while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1475 ("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1476 nmp->nm_bufqwant = TRUE;
1477 error = newnfs_msleep(td, &nmp->nm_bufq,
1478 &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1481 error2 = newnfs_sigintr(nmp, td);
1483 mtx_unlock(&ncl_iod_mutex);
1486 if (slpflag == NFS_PCATCH) {
1492 * We might have lost our iod while sleeping,
1493 * so check and loop if nescessary.
1498 /* We might have lost our nfsiod */
1499 if (nmp->nm_bufqiods == 0) {
1501 ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1505 if (bp->b_iocmd == BIO_READ) {
1506 if (bp->b_rcred == NOCRED && cred != NOCRED)
1507 bp->b_rcred = crhold(cred);
1509 if (bp->b_wcred == NOCRED && cred != NOCRED)
1510 bp->b_wcred = crhold(cred);
1513 if (bp->b_flags & B_REMFREE)
1516 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1518 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1519 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx);
1520 VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1521 VTONFS(bp->b_vp)->n_directio_asyncwr++;
1522 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx);
1524 mtx_unlock(&ncl_iod_mutex);
1528 mtx_unlock(&ncl_iod_mutex);
1531 * All the iods are busy on other mounts, so return EIO to
1532 * force the caller to process the i/o synchronously.
1534 NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1539 ncl_doio_directwrite(struct buf *bp)
1541 int iomode, must_commit;
1542 struct uio *uiop = (struct uio *)bp->b_caller1;
1543 char *iov_base = uiop->uio_iov->iov_base;
1545 iomode = NFSWRITE_FILESYNC;
1546 uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1547 ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0);
1548 KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write"));
1549 free(iov_base, M_NFSDIRECTIO);
1550 free(uiop->uio_iov, M_NFSDIRECTIO);
1551 free(uiop, M_NFSDIRECTIO);
1552 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1553 struct nfsnode *np = VTONFS(bp->b_vp);
1554 mtx_lock(&np->n_mtx);
1555 np->n_directio_asyncwr--;
1556 if (np->n_directio_asyncwr == 0) {
1557 np->n_flag &= ~NMODIFIED;
1558 if ((np->n_flag & NFSYNCWAIT)) {
1559 np->n_flag &= ~NFSYNCWAIT;
1560 wakeup((caddr_t)&np->n_directio_asyncwr);
1563 mtx_unlock(&np->n_mtx);
1566 relpbuf(bp, &ncl_pbuf_freecnt);
1570 * Do an I/O operation to/from a cache block. This may be called
1571 * synchronously or from an nfsiod.
1574 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1575 int called_from_strategy)
1579 struct nfsmount *nmp;
1580 int error = 0, iomode, must_commit = 0;
1583 struct proc *p = td ? td->td_proc : NULL;
1587 nmp = VFSTONFS(vp->v_mount);
1589 uiop->uio_iov = &io;
1590 uiop->uio_iovcnt = 1;
1591 uiop->uio_segflg = UIO_SYSSPACE;
1595 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We
1596 * do this here so we do not have to do it in all the code that
1599 bp->b_flags &= ~B_INVAL;
1600 bp->b_ioflags &= ~BIO_ERROR;
1602 KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1603 iocmd = bp->b_iocmd;
1604 if (iocmd == BIO_READ) {
1605 io.iov_len = uiop->uio_resid = bp->b_bcount;
1606 io.iov_base = bp->b_data;
1607 uiop->uio_rw = UIO_READ;
1609 switch (vp->v_type) {
1611 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1612 NFSINCRGLOBAL(newnfsstats.read_bios);
1613 error = ncl_readrpc(vp, uiop, cr);
1616 if (uiop->uio_resid) {
1618 * If we had a short read with no error, we must have
1619 * hit a file hole. We should zero-fill the remainder.
1620 * This can also occur if the server hits the file EOF.
1622 * Holes used to be able to occur due to pending
1623 * writes, but that is not possible any longer.
1625 int nread = bp->b_bcount - uiop->uio_resid;
1626 ssize_t left = uiop->uio_resid;
1629 bzero((char *)bp->b_data + nread, left);
1630 uiop->uio_resid = 0;
1633 /* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1634 if (p && (vp->v_vflag & VV_TEXT)) {
1635 mtx_lock(&np->n_mtx);
1636 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1637 mtx_unlock(&np->n_mtx);
1639 killproc(p, "text file modification");
1642 mtx_unlock(&np->n_mtx);
1646 uiop->uio_offset = (off_t)0;
1647 NFSINCRGLOBAL(newnfsstats.readlink_bios);
1648 error = ncl_readlinkrpc(vp, uiop, cr);
1651 NFSINCRGLOBAL(newnfsstats.readdir_bios);
1652 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1653 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1654 error = ncl_readdirplusrpc(vp, uiop, cr, td);
1655 if (error == NFSERR_NOTSUPP)
1656 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1658 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1659 error = ncl_readdirrpc(vp, uiop, cr, td);
1661 * end-of-directory sets B_INVAL but does not generate an
1664 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1665 bp->b_flags |= B_INVAL;
1668 ncl_printf("ncl_doio: type %x unexpected\n", vp->v_type);
1672 bp->b_ioflags |= BIO_ERROR;
1673 bp->b_error = error;
1677 * If we only need to commit, try to commit
1679 if (bp->b_flags & B_NEEDCOMMIT) {
1683 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1684 retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1687 bp->b_dirtyoff = bp->b_dirtyend = 0;
1688 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1693 if (retv == NFSERR_STALEWRITEVERF) {
1694 ncl_clearcommit(vp->v_mount);
1699 * Setup for actual write
1701 mtx_lock(&np->n_mtx);
1702 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1703 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1704 mtx_unlock(&np->n_mtx);
1706 if (bp->b_dirtyend > bp->b_dirtyoff) {
1707 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1709 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1711 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1712 uiop->uio_rw = UIO_WRITE;
1713 NFSINCRGLOBAL(newnfsstats.write_bios);
1715 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1716 iomode = NFSWRITE_UNSTABLE;
1718 iomode = NFSWRITE_FILESYNC;
1720 error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1721 called_from_strategy);
1724 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1725 * to cluster the buffers needing commit. This will allow
1726 * the system to submit a single commit rpc for the whole
1727 * cluster. We can do this even if the buffer is not 100%
1728 * dirty (relative to the NFS blocksize), so we optimize the
1729 * append-to-file-case.
1731 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1732 * cleared because write clustering only works for commit
1733 * rpc's, not for the data portion of the write).
1736 if (!error && iomode == NFSWRITE_UNSTABLE) {
1737 bp->b_flags |= B_NEEDCOMMIT;
1738 if (bp->b_dirtyoff == 0
1739 && bp->b_dirtyend == bp->b_bcount)
1740 bp->b_flags |= B_CLUSTEROK;
1742 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1746 * For an interrupted write, the buffer is still valid
1747 * and the write hasn't been pushed to the server yet,
1748 * so we can't set BIO_ERROR and report the interruption
1749 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1750 * is not relevant, so the rpc attempt is essentially
1751 * a noop. For the case of a V3 write rpc not being
1752 * committed to stable storage, the block is still
1753 * dirty and requires either a commit rpc or another
1754 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1755 * the block is reused. This is indicated by setting
1756 * the B_DELWRI and B_NEEDCOMMIT flags.
1758 * EIO is returned by ncl_writerpc() to indicate a recoverable
1759 * write error and is handled as above, except that
1760 * B_EINTR isn't set. One cause of this is a stale stateid
1761 * error for the RPC that indicates recovery is required,
1762 * when called with called_from_strategy != 0.
1764 * If the buffer is marked B_PAGING, it does not reside on
1765 * the vp's paging queues so we cannot call bdirty(). The
1766 * bp in this case is not an NFS cache block so we should
1769 * The logic below breaks up errors into recoverable and
1770 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1771 * and keep the buffer around for potential write retries.
1772 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1773 * and save the error in the nfsnode. This is less than ideal
1774 * but necessary. Keeping such buffers around could potentially
1775 * cause buffer exhaustion eventually (they can never be written
1776 * out, so will get constantly be re-dirtied). It also causes
1777 * all sorts of vfs panics. For non-recoverable write errors,
1778 * also invalidate the attrcache, so we'll be forced to go over
1779 * the wire for this object, returning an error to user on next
1780 * call (most of the time).
1782 if (error == EINTR || error == EIO || error == ETIMEDOUT
1783 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1787 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1788 if ((bp->b_flags & B_PAGING) == 0) {
1790 bp->b_flags &= ~B_DONE;
1792 if ((error == EINTR || error == ETIMEDOUT) &&
1793 (bp->b_flags & B_ASYNC) == 0)
1794 bp->b_flags |= B_EINTR;
1798 bp->b_ioflags |= BIO_ERROR;
1799 bp->b_flags |= B_INVAL;
1800 bp->b_error = np->n_error = error;
1801 mtx_lock(&np->n_mtx);
1802 np->n_flag |= NWRITEERR;
1803 np->n_attrstamp = 0;
1804 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1805 mtx_unlock(&np->n_mtx);
1807 bp->b_dirtyoff = bp->b_dirtyend = 0;
1815 bp->b_resid = uiop->uio_resid;
1817 ncl_clearcommit(vp->v_mount);
1823 * Used to aid in handling ftruncate() operations on the NFS client side.
1824 * Truncation creates a number of special problems for NFS. We have to
1825 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1826 * we have to properly handle VM pages or (potentially dirty) buffers
1827 * that straddle the truncation point.
1831 ncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1833 struct nfsnode *np = VTONFS(vp);
1835 int biosize = vp->v_bufobj.bo_bsize;
1838 mtx_lock(&np->n_mtx);
1841 mtx_unlock(&np->n_mtx);
1843 if (nsize < tsize) {
1849 * vtruncbuf() doesn't get the buffer overlapping the
1850 * truncation point. We may have a B_DELWRI and/or B_CACHE
1851 * buffer that now needs to be truncated.
1853 error = vtruncbuf(vp, cred, td, nsize, biosize);
1854 lbn = nsize / biosize;
1855 bufsize = nsize & (biosize - 1);
1856 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1859 if (bp->b_dirtyoff > bp->b_bcount)
1860 bp->b_dirtyoff = bp->b_bcount;
1861 if (bp->b_dirtyend > bp->b_bcount)
1862 bp->b_dirtyend = bp->b_bcount;
1863 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1866 vnode_pager_setsize(vp, nsize);