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 #include <sys/param.h>
39 #include <sys/systm.h>
42 #include <sys/kernel.h>
43 #include <sys/mount.h>
44 #include <sys/rwlock.h>
45 #include <sys/vmmeter.h>
46 #include <sys/vnode.h>
49 #include <vm/vm_param.h>
50 #include <vm/vm_extern.h>
51 #include <vm/vm_page.h>
52 #include <vm/vm_object.h>
53 #include <vm/vm_pager.h>
54 #include <vm/vnode_pager.h>
56 #include <fs/nfs/nfsport.h>
57 #include <fs/nfsclient/nfsmount.h>
58 #include <fs/nfsclient/nfs.h>
59 #include <fs/nfsclient/nfsnode.h>
60 #include <fs/nfsclient/nfs_kdtrace.h>
62 extern int newnfs_directio_allow_mmap;
63 extern struct nfsstatsv1 nfsstatsv1;
64 extern struct mtx ncl_iod_mutex;
65 extern int ncl_numasync;
66 extern enum nfsiod_state ncl_iodwant[NFS_MAXASYNCDAEMON];
67 extern struct nfsmount *ncl_iodmount[NFS_MAXASYNCDAEMON];
68 extern int newnfs_directio_enable;
69 extern int nfs_keep_dirty_on_error;
71 uma_zone_t ncl_pbuf_zone;
73 static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size,
75 static int nfs_directio_write(struct vnode *vp, struct uio *uiop,
76 struct ucred *cred, int ioflag);
79 * Vnode op for VM getpages.
81 SYSCTL_DECL(_vfs_nfs);
82 static int use_buf_pager = 1;
83 SYSCTL_INT(_vfs_nfs, OID_AUTO, use_buf_pager, CTLFLAG_RWTUN,
85 "Use buffer pager instead of direct readrpc call");
88 ncl_gbp_getblkno(struct vnode *vp, vm_ooffset_t off)
91 return (off / vp->v_bufobj.bo_bsize);
95 ncl_gbp_getblksz(struct vnode *vp, daddr_t lbn, long *sz)
106 biosize = vp->v_bufobj.bo_bsize;
108 if ((off_t)lbn * biosize >= nsize)
110 else if ((off_t)(lbn + 1) * biosize > nsize)
111 bcount = nsize - (off_t)lbn * biosize;
117 ncl_getpages(struct vop_getpages_args *ap)
119 int i, error, nextoff, size, toff, count, npages;
127 struct nfsmount *nmp;
135 cred = curthread->td_ucred;
136 nmp = VFSTONFS(vp->v_mount);
138 npages = ap->a_count;
140 if ((object = vp->v_object) == NULL) {
141 printf("ncl_getpages: called with non-merged cache vnode\n");
142 return (VM_PAGER_ERROR);
145 if (newnfs_directio_enable && !newnfs_directio_allow_mmap) {
147 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
149 printf("ncl_getpages: called on non-cacheable vnode\n");
150 return (VM_PAGER_ERROR);
155 mtx_lock(&nmp->nm_mtx);
156 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
157 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
158 mtx_unlock(&nmp->nm_mtx);
159 /* We'll never get here for v4, because we always have fsinfo */
160 (void)ncl_fsinfo(nmp, vp, cred, td);
162 mtx_unlock(&nmp->nm_mtx);
165 return (vfs_bio_getpages(vp, pages, npages, ap->a_rbehind,
166 ap->a_rahead, ncl_gbp_getblkno, ncl_gbp_getblksz));
169 * If the requested page is partially valid, just return it and
170 * allow the pager to zero-out the blanks. Partially valid pages
171 * can only occur at the file EOF.
173 * XXXGL: is that true for NFS, where short read can occur???
175 VM_OBJECT_WLOCK(object);
176 if (!vm_page_none_valid(pages[npages - 1]) && --npages == 0)
178 VM_OBJECT_WUNLOCK(object);
181 * We use only the kva address for the buffer, but this is extremely
182 * convenient and fast.
184 bp = uma_zalloc(ncl_pbuf_zone, M_WAITOK);
186 kva = (vm_offset_t) bp->b_data;
187 pmap_qenter(kva, pages, npages);
188 VM_CNT_INC(v_vnodein);
189 VM_CNT_ADD(v_vnodepgsin, npages);
191 count = npages << PAGE_SHIFT;
192 iov.iov_base = (caddr_t) kva;
196 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
197 uio.uio_resid = count;
198 uio.uio_segflg = UIO_SYSSPACE;
199 uio.uio_rw = UIO_READ;
202 error = ncl_readrpc(vp, &uio, cred);
203 pmap_qremove(kva, npages);
205 uma_zfree(ncl_pbuf_zone, bp);
207 if (error && (uio.uio_resid == count)) {
208 printf("ncl_getpages: error %d\n", error);
209 return (VM_PAGER_ERROR);
213 * Calculate the number of bytes read and validate only that number
214 * of bytes. Note that due to pending writes, size may be 0. This
215 * does not mean that the remaining data is invalid!
218 size = count - uio.uio_resid;
219 VM_OBJECT_WLOCK(object);
220 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
222 nextoff = toff + PAGE_SIZE;
225 if (nextoff <= size) {
227 * Read operation filled an entire page
230 KASSERT(m->dirty == 0,
231 ("nfs_getpages: page %p is dirty", m));
232 } else if (size > toff) {
234 * Read operation filled a partial page.
237 vm_page_set_valid_range(m, 0, size - toff);
238 KASSERT(m->dirty == 0,
239 ("nfs_getpages: page %p is dirty", m));
242 * Read operation was short. If no error
243 * occurred we may have hit a zero-fill
244 * section. We leave valid set to 0, and page
245 * is freed by vm_page_readahead_finish() if
246 * its index is not equal to requested, or
247 * page is zeroed and set valid by
248 * vm_pager_get_pages() for requested page.
254 VM_OBJECT_WUNLOCK(object);
259 return (VM_PAGER_OK);
263 * Vnode op for VM putpages.
266 ncl_putpages(struct vop_putpages_args *ap)
270 int i, error, npages, count;
276 struct nfsmount *nmp;
282 td = curthread; /* XXX */
283 /* Set the cred to n_writecred for the write rpcs. */
284 if (np->n_writecred != NULL)
285 cred = crhold(np->n_writecred);
287 cred = crhold(curthread->td_ucred); /* XXX */
288 nmp = VFSTONFS(vp->v_mount);
291 rtvals = ap->a_rtvals;
292 npages = btoc(count);
293 offset = IDX_TO_OFF(pages[0]->pindex);
295 mtx_lock(&nmp->nm_mtx);
296 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
297 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
298 mtx_unlock(&nmp->nm_mtx);
299 (void)ncl_fsinfo(nmp, vp, cred, td);
301 mtx_unlock(&nmp->nm_mtx);
304 if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
305 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
307 printf("ncl_putpages: called on noncache-able vnode\n");
311 * When putting pages, do not extend file past EOF.
313 if (offset + count > np->n_size) {
314 count = np->n_size - offset;
320 for (i = 0; i < npages; i++)
321 rtvals[i] = VM_PAGER_ERROR;
323 VM_CNT_INC(v_vnodeout);
324 VM_CNT_ADD(v_vnodepgsout, count);
326 iov.iov_base = unmapped_buf;
330 uio.uio_offset = offset;
331 uio.uio_resid = count;
332 uio.uio_segflg = UIO_NOCOPY;
333 uio.uio_rw = UIO_WRITE;
336 error = VOP_WRITE(vp, &uio, vnode_pager_putpages_ioflags(ap->a_sync),
340 if (error == 0 || !nfs_keep_dirty_on_error) {
341 vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid,
342 np->n_size - offset, npages * PAGE_SIZE);
348 * For nfs, cache consistency can only be maintained approximately.
349 * Although RFC1094 does not specify the criteria, the following is
350 * believed to be compatible with the reference port.
352 * If the file's modify time on the server has changed since the
353 * last read rpc or you have written to the file,
354 * you may have lost data cache consistency with the
355 * server, so flush all of the file's data out of the cache.
356 * Then force a getattr rpc to ensure that you have up to date
358 * NB: This implies that cache data can be read when up to
359 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
360 * attributes this could be forced by setting n_attrstamp to 0 before
361 * the VOP_GETATTR() call.
364 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
368 struct nfsnode *np = VTONFS(vp);
372 * Ensure the exclusove access to the node before checking
373 * whether the cache is consistent.
375 old_lock = ncl_excl_start(vp);
377 if (np->n_flag & NMODIFIED) {
379 if (vp->v_type != VREG) {
380 if (vp->v_type != VDIR)
381 panic("nfs: bioread, not dir");
383 error = ncl_vinvalbuf(vp, V_SAVE | V_ALLOWCLEAN, td, 1);
388 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
389 error = VOP_GETATTR(vp, &vattr, cred);
393 np->n_mtime = vattr.va_mtime;
397 error = VOP_GETATTR(vp, &vattr, cred);
401 if ((np->n_flag & NSIZECHANGED)
402 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
404 if (vp->v_type == VDIR)
406 error = ncl_vinvalbuf(vp, V_SAVE | V_ALLOWCLEAN, td, 1);
410 np->n_mtime = vattr.va_mtime;
411 np->n_flag &= ~NSIZECHANGED;
416 ncl_excl_finish(vp, old_lock);
421 * Vnode op for read using bio
424 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
426 struct nfsnode *np = VTONFS(vp);
427 struct buf *bp, *rabp;
429 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
431 int biosize, bcount, error, i, n, nra, on, save2, seqcount;
434 KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
435 if (uio->uio_resid == 0)
437 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
441 mtx_lock(&nmp->nm_mtx);
442 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
443 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
444 mtx_unlock(&nmp->nm_mtx);
445 (void)ncl_fsinfo(nmp, vp, cred, td);
446 mtx_lock(&nmp->nm_mtx);
448 if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
449 (void) newnfs_iosize(nmp);
451 tmp_off = uio->uio_offset + uio->uio_resid;
452 if (vp->v_type != VDIR &&
453 (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) {
454 mtx_unlock(&nmp->nm_mtx);
457 mtx_unlock(&nmp->nm_mtx);
459 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
460 /* No caching/ no readaheads. Just read data into the user buffer */
461 return ncl_readrpc(vp, uio, cred);
465 biosize = vp->v_bufobj.bo_bsize;
466 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
468 error = nfs_bioread_check_cons(vp, td, cred);
472 save2 = curthread_pflags2_set(TDP2_SBPAGES);
480 switch (vp->v_type) {
482 NFSINCRGLOBAL(nfsstatsv1.biocache_reads);
483 lbn = uio->uio_offset / biosize;
484 on = uio->uio_offset - (lbn * biosize);
487 * Start the read ahead(s), as required. Do not do
488 * read-ahead if there are writeable mappings, since
489 * unlocked read by nfsiod could obliterate changes
492 if (nmp->nm_readahead > 0 &&
493 !vm_object_mightbedirty(vp->v_object) &&
494 vp->v_object->un_pager.vnp.writemappings == 0) {
495 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
496 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
497 rabn = lbn + 1 + nra;
498 if (incore(&vp->v_bufobj, rabn) == NULL) {
499 rabp = nfs_getcacheblk(vp, rabn, biosize, td);
501 error = newnfs_sigintr(nmp, td);
506 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
507 rabp->b_flags |= B_ASYNC;
508 rabp->b_iocmd = BIO_READ;
509 vfs_busy_pages(rabp, 0);
510 if (ncl_asyncio(nmp, rabp, cred, td)) {
511 rabp->b_flags |= B_INVAL;
512 rabp->b_ioflags |= BIO_ERROR;
513 vfs_unbusy_pages(rabp);
524 /* Note that bcount is *not* DEV_BSIZE aligned. */
526 if ((off_t)lbn * biosize >= nsize) {
528 } else if ((off_t)(lbn + 1) * biosize > nsize) {
529 bcount = nsize - (off_t)lbn * biosize;
531 bp = nfs_getcacheblk(vp, lbn, bcount, td);
534 error = newnfs_sigintr(nmp, td);
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(nfsstatsv1.biocache_readlinks);
569 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
571 error = newnfs_sigintr(nmp, td);
576 if ((bp->b_flags & B_CACHE) == 0) {
577 bp->b_iocmd = BIO_READ;
578 vfs_busy_pages(bp, 0);
579 error = ncl_doio(vp, bp, cred, td, 0);
581 bp->b_ioflags |= BIO_ERROR;
586 n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
590 NFSINCRGLOBAL(nfsstatsv1.biocache_readdirs);
592 if (np->n_direofoffset
593 && uio->uio_offset >= np->n_direofoffset) {
599 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
600 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
601 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
603 error = newnfs_sigintr(nmp, td);
608 if ((bp->b_flags & B_CACHE) == 0) {
609 bp->b_iocmd = BIO_READ;
610 vfs_busy_pages(bp, 0);
611 error = ncl_doio(vp, bp, cred, td, 0);
615 while (error == NFSERR_BAD_COOKIE) {
617 error = ncl_vinvalbuf(vp, 0, td, 1);
620 * Yuck! The directory has been modified on the
621 * server. The only way to get the block is by
622 * reading from the beginning to get all the
625 * Leave the last bp intact unless there is an error.
626 * Loop back up to the while if the error is another
627 * NFSERR_BAD_COOKIE (double yuch!).
629 for (i = 0; i <= lbn && !error; i++) {
631 if (np->n_direofoffset
632 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) {
638 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
640 error = newnfs_sigintr(nmp, td);
645 if ((bp->b_flags & B_CACHE) == 0) {
646 bp->b_iocmd = BIO_READ;
647 vfs_busy_pages(bp, 0);
648 error = ncl_doio(vp, bp, cred, td, 0);
650 * no error + B_INVAL == directory EOF,
653 if (error == 0 && (bp->b_flags & B_INVAL))
657 * An error will throw away the block and the
658 * for loop will break out. If no error and this
659 * is not the block we want, we throw away the
660 * block and go for the next one via the for loop.
662 if (error || i < lbn)
667 * The above while is repeated if we hit another cookie
668 * error. If we hit an error and it wasn't a cookie error,
676 * If not eof and read aheads are enabled, start one.
677 * (You need the current block first, so that you have the
678 * directory offset cookie of the next block.)
681 if (nmp->nm_readahead > 0 &&
682 !vm_object_mightbedirty(vp->v_object) &&
683 vp->v_object->un_pager.vnp.writemappings == 0 &&
684 (bp->b_flags & B_INVAL) == 0 &&
685 (np->n_direofoffset == 0 ||
686 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
687 incore(&vp->v_bufobj, lbn + 1) == NULL) {
689 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
691 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
692 rabp->b_flags |= B_ASYNC;
693 rabp->b_iocmd = BIO_READ;
694 vfs_busy_pages(rabp, 0);
695 if (ncl_asyncio(nmp, rabp, cred, td)) {
696 rabp->b_flags |= B_INVAL;
697 rabp->b_ioflags |= BIO_ERROR;
698 vfs_unbusy_pages(rabp);
708 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
709 * chopped for the EOF condition, we cannot tell how large
710 * NFS directories are going to be until we hit EOF. So
711 * an NFS directory buffer is *not* chopped to its EOF. Now,
712 * it just so happens that b_resid will effectively chop it
713 * to EOF. *BUT* this information is lost if the buffer goes
714 * away and is reconstituted into a B_CACHE state ( due to
715 * being VMIO ) later. So we keep track of the directory eof
716 * in np->n_direofoffset and chop it off as an extra step
719 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
720 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
721 n = np->n_direofoffset - uio->uio_offset;
725 printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
731 error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio);
733 if (vp->v_type == VLNK)
737 } while (error == 0 && uio->uio_resid > 0 && n > 0);
739 curthread_pflags2_restore(save2);
740 if ((curthread->td_pflags2 & TDP2_SBPAGES) == 0) {
742 ncl_pager_setsize(vp, NULL);
748 * The NFS write path cannot handle iovecs with len > 1. So we need to
749 * break up iovecs accordingly (restricting them to wsize).
750 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
751 * For the ASYNC case, 2 copies are needed. The first a copy from the
752 * user buffer to a staging buffer and then a second copy from the staging
753 * buffer to mbufs. This can be optimized by copying from the user buffer
754 * directly into mbufs and passing the chain down, but that requires a
755 * fair amount of re-working of the relevant codepaths (and can be done
759 nfs_directio_write(struct vnode *vp, struct uio *uiop, struct ucred *cred,
763 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
764 struct thread *td = uiop->uio_td;
768 mtx_lock(&nmp->nm_mtx);
769 wsize = nmp->nm_wsize;
770 mtx_unlock(&nmp->nm_mtx);
771 if (ioflag & IO_SYNC) {
772 int iomode, must_commit;
776 while (uiop->uio_resid > 0) {
777 size = MIN(uiop->uio_resid, wsize);
778 size = MIN(uiop->uio_iov->iov_len, size);
779 iov.iov_base = uiop->uio_iov->iov_base;
783 uio.uio_offset = uiop->uio_offset;
784 uio.uio_resid = size;
785 uio.uio_segflg = uiop->uio_segflg;
786 uio.uio_rw = UIO_WRITE;
788 iomode = NFSWRITE_FILESYNC;
790 * When doing direct I/O we do not care if the
791 * server's write verifier has changed, but we
792 * do not want to update the verifier if it has
793 * changed, since that hides the change from
794 * writes being done through the buffer cache.
795 * By passing must_commit in set to two, the code
796 * in nfsrpc_writerpc() will not update the
797 * verifier on the mount point.
800 error = ncl_writerpc(vp, &uio, cred, &iomode,
801 &must_commit, 0, ioflag);
802 KASSERT((must_commit == 2),
803 ("ncl_directio_write: Updated write verifier"));
806 if (iomode != NFSWRITE_FILESYNC)
807 printf("nfs_directio_write: Broken server "
808 "did not reply FILE_SYNC\n");
809 uiop->uio_offset += size;
810 uiop->uio_resid -= size;
811 if (uiop->uio_iov->iov_len <= size) {
815 uiop->uio_iov->iov_base =
816 (char *)uiop->uio_iov->iov_base + size;
817 uiop->uio_iov->iov_len -= size;
826 * Break up the write into blocksize chunks and hand these
827 * over to nfsiod's for write back.
828 * Unfortunately, this incurs a copy of the data. Since
829 * the user could modify the buffer before the write is
832 * The obvious optimization here is that one of the 2 copies
833 * in the async write path can be eliminated by copying the
834 * data here directly into mbufs and passing the mbuf chain
835 * down. But that will require a fair amount of re-working
836 * of the code and can be done if there's enough interest
837 * in NFS directio access.
839 while (uiop->uio_resid > 0) {
840 size = MIN(uiop->uio_resid, wsize);
841 size = MIN(uiop->uio_iov->iov_len, size);
842 bp = uma_zalloc(ncl_pbuf_zone, M_WAITOK);
843 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
844 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
845 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
846 t_iov->iov_len = size;
847 t_uio->uio_iov = t_iov;
848 t_uio->uio_iovcnt = 1;
849 t_uio->uio_offset = uiop->uio_offset;
850 t_uio->uio_resid = size;
851 t_uio->uio_segflg = UIO_SYSSPACE;
852 t_uio->uio_rw = UIO_WRITE;
854 KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
855 uiop->uio_segflg == UIO_SYSSPACE,
856 ("nfs_directio_write: Bad uio_segflg"));
857 if (uiop->uio_segflg == UIO_USERSPACE) {
858 error = copyin(uiop->uio_iov->iov_base,
859 t_iov->iov_base, size);
864 * UIO_SYSSPACE may never happen, but handle
865 * it just in case it does.
867 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
869 bp->b_flags |= B_DIRECT;
870 bp->b_iocmd = BIO_WRITE;
871 if (cred != NOCRED) {
875 bp->b_wcred = NOCRED;
876 bp->b_caller1 = (void *)t_uio;
878 error = ncl_asyncio(nmp, bp, NOCRED, td);
881 free(t_iov->iov_base, M_NFSDIRECTIO);
882 free(t_iov, M_NFSDIRECTIO);
883 free(t_uio, M_NFSDIRECTIO);
885 uma_zfree(ncl_pbuf_zone, bp);
890 uiop->uio_offset += size;
891 uiop->uio_resid -= size;
892 if (uiop->uio_iov->iov_len <= size) {
896 uiop->uio_iov->iov_base =
897 (char *)uiop->uio_iov->iov_base + size;
898 uiop->uio_iov->iov_len -= size;
906 * Vnode op for write using bio
909 ncl_write(struct vop_write_args *ap)
912 struct uio *uio = ap->a_uio;
913 struct thread *td = uio->uio_td;
914 struct vnode *vp = ap->a_vp;
915 struct nfsnode *np = VTONFS(vp);
916 struct ucred *cred = ap->a_cred;
917 int ioflag = ap->a_ioflag;
920 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
922 int bcount, noncontig_write, obcount;
923 int bp_cached, n, on, error = 0, error1, save2, wouldcommit;
924 size_t orig_resid, local_resid;
925 off_t orig_size, tmp_off;
928 KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
929 KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
931 if (vp->v_type != VREG)
934 if (np->n_flag & NWRITEERR) {
935 np->n_flag &= ~NWRITEERR;
937 return (np->n_error);
940 mtx_lock(&nmp->nm_mtx);
941 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
942 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
943 mtx_unlock(&nmp->nm_mtx);
944 (void)ncl_fsinfo(nmp, vp, cred, td);
945 mtx_lock(&nmp->nm_mtx);
947 if (nmp->nm_wsize == 0)
948 (void) newnfs_iosize(nmp);
949 mtx_unlock(&nmp->nm_mtx);
952 * Synchronously flush pending buffers if we are in synchronous
953 * mode or if we are appending.
955 if ((ioflag & IO_APPEND) || ((ioflag & IO_SYNC) && (np->n_flag &
958 * For the case where IO_APPEND is being done using a
959 * direct output (to the NFS server) RPC and
960 * newnfs_directio_enable is 0, all buffer cache buffers,
961 * including ones not modified, must be invalidated.
962 * This ensures that stale data is not read out of the
963 * buffer cache. The call also invalidates all mapped
964 * pages and, since the exclusive lock is held on the vnode,
965 * new pages cannot be faulted in.
967 * For the case where newnfs_directio_enable is set
968 * (which is not the default), it is not obvious that
969 * stale data should be left in the buffer cache, but
970 * the code has been this way for over a decade without
971 * complaints. Note that, unlike doing IO_APPEND via
972 * a direct write RPC when newnfs_directio_enable is not set,
973 * when newnfs_directio_enable is set, reading is done via
974 * direct to NFS server RPCs as well.
977 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
978 error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
979 IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
984 orig_resid = uio->uio_resid;
986 orig_size = np->n_size;
990 * If IO_APPEND then load uio_offset. We restart here if we cannot
991 * get the append lock.
993 if (ioflag & IO_APPEND) {
995 * For NFSv4, the AppendWrite will Verify the size against
996 * the file's size on the server. If not the same, the
997 * write will then be retried, using the file size returned
998 * by the AppendWrite. However, for NFSv2 and NFSv3, the
999 * size must be acquired here via a Getattr RPC.
1000 * The AppendWrite is not done for a pNFS mount.
1002 if (!NFSHASNFSV4(nmp) || NFSHASPNFS(nmp)) {
1003 np->n_attrstamp = 0;
1004 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1005 error = VOP_GETATTR(vp, &vattr, cred);
1010 uio->uio_offset = np->n_size;
1014 if (uio->uio_offset < 0)
1016 tmp_off = uio->uio_offset + uio->uio_resid;
1017 if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
1019 if (uio->uio_resid == 0)
1023 * Do IO_APPEND writing via a synchronous direct write.
1024 * This can result in a significant performance improvement.
1026 if ((newnfs_directio_enable && (ioflag & IO_DIRECT)) ||
1027 (ioflag & IO_APPEND)) {
1029 * Direct writes to the server must be done NFSWRITE_FILESYNC,
1030 * because the write data is not cached and, therefore, the
1031 * write cannot be redone after a server reboot.
1032 * Set IO_SYNC to make this happen.
1035 return (nfs_directio_write(vp, uio, cred, ioflag));
1039 * Maybe this should be above the vnode op call, but so long as
1040 * file servers have no limits, i don't think it matters
1042 error = vn_rlimit_fsize(vp, uio, td);
1046 save2 = curthread_pflags2_set(TDP2_SBPAGES);
1047 biosize = vp->v_bufobj.bo_bsize;
1049 * Find all of this file's B_NEEDCOMMIT buffers. If our writes
1050 * would exceed the local maximum per-file write commit size when
1051 * combined with those, we must decide whether to flush,
1052 * go synchronous, or return error. We don't bother checking
1053 * IO_UNIT -- we just make all writes atomic anyway, as there's
1054 * no point optimizing for something that really won't ever happen.
1057 if (!(ioflag & IO_SYNC)) {
1063 if (nflag & NMODIFIED) {
1064 BO_LOCK(&vp->v_bufobj);
1065 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
1066 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
1068 if (bp->b_flags & B_NEEDCOMMIT)
1069 wouldcommit += bp->b_bcount;
1072 BO_UNLOCK(&vp->v_bufobj);
1077 if (!(ioflag & IO_SYNC)) {
1078 wouldcommit += biosize;
1079 if (wouldcommit > nmp->nm_wcommitsize) {
1080 np->n_attrstamp = 0;
1081 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1082 error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
1083 IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
1086 wouldcommit = biosize;
1090 NFSINCRGLOBAL(nfsstatsv1.biocache_writes);
1091 lbn = uio->uio_offset / biosize;
1092 on = uio->uio_offset - (lbn * biosize);
1093 n = MIN((unsigned)(biosize - on), uio->uio_resid);
1096 * Handle direct append and file extension cases, calculate
1097 * unaligned buffer size.
1100 if ((np->n_flag & NHASBEENLOCKED) == 0 &&
1101 (nmp->nm_flag & NFSMNT_NONCONTIGWR) != 0)
1102 noncontig_write = 1;
1104 noncontig_write = 0;
1105 if ((uio->uio_offset == np->n_size ||
1106 (noncontig_write != 0 &&
1107 lbn == (np->n_size / biosize) &&
1108 uio->uio_offset + n > np->n_size)) && n) {
1111 * Get the buffer (in its pre-append state to maintain
1112 * B_CACHE if it was previously set). Resize the
1113 * nfsnode after we have locked the buffer to prevent
1114 * readers from reading garbage.
1116 obcount = np->n_size - (lbn * biosize);
1117 bp = nfs_getcacheblk(vp, lbn, obcount, td);
1123 np->n_size = uio->uio_offset + n;
1124 np->n_flag |= NMODIFIED;
1125 np->n_flag &= ~NVNSETSZSKIP;
1126 vnode_pager_setsize(vp, np->n_size);
1129 save = bp->b_flags & B_CACHE;
1131 allocbuf(bp, bcount);
1132 bp->b_flags |= save;
1133 if (noncontig_write != 0 && on > obcount)
1134 vfs_bio_bzero_buf(bp, obcount, on -
1139 * Obtain the locked cache block first, and then
1140 * adjust the file's size as appropriate.
1143 if ((off_t)lbn * biosize + bcount < np->n_size) {
1144 if ((off_t)(lbn + 1) * biosize < np->n_size)
1147 bcount = np->n_size - (off_t)lbn * biosize;
1150 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1152 if (uio->uio_offset + n > np->n_size) {
1153 np->n_size = uio->uio_offset + n;
1154 np->n_flag |= NMODIFIED;
1155 np->n_flag &= ~NVNSETSZSKIP;
1156 vnode_pager_setsize(vp, np->n_size);
1162 error = newnfs_sigintr(nmp, td);
1169 * Issue a READ if B_CACHE is not set. In special-append
1170 * mode, B_CACHE is based on the buffer prior to the write
1171 * op and is typically set, avoiding the read. If a read
1172 * is required in special append mode, the server will
1173 * probably send us a short-read since we extended the file
1174 * on our end, resulting in b_resid == 0 and, thusly,
1175 * B_CACHE getting set.
1177 * We can also avoid issuing the read if the write covers
1178 * the entire buffer. We have to make sure the buffer state
1179 * is reasonable in this case since we will not be initiating
1180 * I/O. See the comments in kern/vfs_bio.c's getblk() for
1183 * B_CACHE may also be set due to the buffer being cached
1188 if (on == 0 && n == bcount) {
1189 if ((bp->b_flags & B_CACHE) == 0)
1191 bp->b_flags |= B_CACHE;
1192 bp->b_flags &= ~B_INVAL;
1193 bp->b_ioflags &= ~BIO_ERROR;
1196 if ((bp->b_flags & B_CACHE) == 0) {
1197 bp->b_iocmd = BIO_READ;
1198 vfs_busy_pages(bp, 0);
1199 error = ncl_doio(vp, bp, cred, td, 0);
1205 if (bp->b_wcred == NOCRED)
1206 bp->b_wcred = crhold(cred);
1208 np->n_flag |= NMODIFIED;
1212 * If dirtyend exceeds file size, chop it down. This should
1213 * not normally occur but there is an append race where it
1214 * might occur XXX, so we log it.
1216 * If the chopping creates a reverse-indexed or degenerate
1217 * situation with dirtyoff/end, we 0 both of them.
1220 if (bp->b_dirtyend > bcount) {
1221 printf("NFS append race @%lx:%d\n",
1222 (long)bp->b_blkno * DEV_BSIZE,
1223 bp->b_dirtyend - bcount);
1224 bp->b_dirtyend = bcount;
1227 if (bp->b_dirtyoff >= bp->b_dirtyend)
1228 bp->b_dirtyoff = bp->b_dirtyend = 0;
1231 * If the new write will leave a contiguous dirty
1232 * area, just update the b_dirtyoff and b_dirtyend,
1233 * otherwise force a write rpc of the old dirty area.
1235 * If there has been a file lock applied to this file
1236 * or vfs.nfs.old_noncontig_writing is set, do the following:
1237 * While it is possible to merge discontiguous writes due to
1238 * our having a B_CACHE buffer ( and thus valid read data
1239 * for the hole), we don't because it could lead to
1240 * significant cache coherency problems with multiple clients,
1241 * especially if locking is implemented later on.
1243 * If vfs.nfs.old_noncontig_writing is not set and there has
1244 * not been file locking done on this file:
1245 * Relax coherency a bit for the sake of performance and
1246 * expand the current dirty region to contain the new
1247 * write even if it means we mark some non-dirty data as
1251 if (noncontig_write == 0 && bp->b_dirtyend > 0 &&
1252 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1253 if (bwrite(bp) == EINTR) {
1260 local_resid = uio->uio_resid;
1261 error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio);
1263 if (error != 0 && !bp_cached) {
1265 * This block has no other content then what
1266 * possibly was written by the faulty uiomove.
1267 * Release it, forgetting the data pages, to
1268 * prevent the leak of uninitialized data to
1271 bp->b_ioflags |= BIO_ERROR;
1273 uio->uio_offset -= local_resid - uio->uio_resid;
1274 uio->uio_resid = local_resid;
1279 * Since this block is being modified, it must be written
1280 * again and not just committed. Since write clustering does
1281 * not work for the stage 1 data write, only the stage 2
1282 * commit rpc, we have to clear B_CLUSTEROK as well.
1284 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1287 * Get the partial update on the progress made from
1288 * uiomove, if an error occurred.
1291 n = local_resid - uio->uio_resid;
1294 * Only update dirtyoff/dirtyend if not a degenerate
1298 if (bp->b_dirtyend > 0) {
1299 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1300 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1302 bp->b_dirtyoff = on;
1303 bp->b_dirtyend = on + n;
1305 vfs_bio_set_valid(bp, on, n);
1309 * If IO_SYNC do bwrite().
1311 * IO_INVAL appears to be unused. The idea appears to be
1312 * to turn off caching in this case. Very odd. XXX
1314 if ((ioflag & IO_SYNC)) {
1315 if (ioflag & IO_INVAL)
1316 bp->b_flags |= B_NOCACHE;
1317 error1 = bwrite(bp);
1323 } else if ((n + on) == biosize || (ioflag & IO_ASYNC) != 0) {
1324 bp->b_flags |= B_ASYNC;
1332 } while (uio->uio_resid > 0 && n > 0);
1337 np->n_localmodtime = ts;
1340 if (ioflag & IO_UNIT) {
1342 vattr.va_size = orig_size;
1343 /* IO_SYNC is handled implicitely */
1344 (void)VOP_SETATTR(vp, &vattr, cred);
1345 uio->uio_offset -= orig_resid - uio->uio_resid;
1346 uio->uio_resid = orig_resid;
1351 curthread_pflags2_restore(save2);
1356 * Get an nfs cache block.
1358 * Allocate a new one if the block isn't currently in the cache
1359 * and return the block marked busy. If the calling process is
1360 * interrupted by a signal for an interruptible mount point, return
1363 * The caller must carefully deal with the possible B_INVAL state of
1364 * the buffer. ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
1365 * indirectly), so synchronous reads can be issued without worrying about
1366 * the B_INVAL state. We have to be a little more careful when dealing
1367 * with writes (see comments in nfs_write()) when extending a file past
1371 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1375 struct nfsmount *nmp;
1380 if (nmp->nm_flag & NFSMNT_INT) {
1383 newnfs_set_sigmask(td, &oldset);
1384 bp = getblk(vp, bn, size, PCATCH, 0, 0);
1385 newnfs_restore_sigmask(td, &oldset);
1386 while (bp == NULL) {
1387 if (newnfs_sigintr(nmp, td))
1389 bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1392 bp = getblk(vp, bn, size, 0, 0, 0);
1395 if (vp->v_type == VREG)
1396 bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1401 * Flush and invalidate all dirty buffers. If another process is already
1402 * doing the flush, just wait for completion.
1405 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1407 struct nfsnode *np = VTONFS(vp);
1408 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1409 int error = 0, slpflag, slptimeo;
1413 ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
1415 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1417 if (NFSCL_FORCEDISM(nmp->nm_mountp))
1427 old_lock = ncl_excl_start(vp);
1429 flags |= V_ALLOWCLEAN;
1432 * Now, flush as required.
1434 if ((flags & (V_SAVE | V_VMIO)) == V_SAVE &&
1435 vp->v_bufobj.bo_object != NULL) {
1436 VM_OBJECT_WLOCK(vp->v_bufobj.bo_object);
1437 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1438 VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object);
1440 * If the page clean was interrupted, fail the invalidation.
1441 * Not doing so, we run the risk of losing dirty pages in the
1442 * vinvalbuf() call below.
1444 if (intrflg && (error = newnfs_sigintr(nmp, td)))
1448 error = vinvalbuf(vp, flags, slpflag, 0);
1450 if (intrflg && (error = newnfs_sigintr(nmp, td)))
1452 error = vinvalbuf(vp, flags, 0, slptimeo);
1454 if (NFSHASPNFS(nmp)) {
1455 nfscl_layoutcommit(vp, td);
1458 * Invalidate the attribute cache, since writes to a DS
1459 * won't update the size attribute.
1462 np->n_attrstamp = 0;
1467 if (np->n_directio_asyncwr == 0 && (np->n_flag & NMODIFIED) != 0) {
1468 np->n_localmodtime = ts;
1469 np->n_flag &= ~NMODIFIED;
1473 ncl_excl_finish(vp, old_lock);
1478 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1479 * This is mainly to avoid queueing async I/O requests when the nfsiods
1480 * are all hung on a dead server.
1482 * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1483 * is eventually dequeued by the async daemon, ncl_doio() *will*.
1486 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1495 * Commits are usually short and sweet so lets save some cpu and
1496 * leave the async daemons for more important rpc's (such as reads
1499 * Readdirplus RPCs do vget()s to acquire the vnodes for entries
1500 * in the directory in order to update attributes. This can deadlock
1501 * with another thread that is waiting for async I/O to be done by
1502 * an nfsiod thread while holding a lock on one of these vnodes.
1503 * To avoid this deadlock, don't allow the async nfsiod threads to
1504 * perform Readdirplus RPCs.
1507 if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1508 (nmp->nm_bufqiods > ncl_numasync / 2)) ||
1509 (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
1514 if (nmp->nm_flag & NFSMNT_INT)
1519 * Find a free iod to process this request.
1521 for (iod = 0; iod < ncl_numasync; iod++)
1522 if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1528 * Try to create one if none are free.
1534 * Found one, so wake it up and tell it which
1537 NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1539 ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1540 ncl_iodmount[iod] = nmp;
1542 wakeup(&ncl_iodwant[iod]);
1546 * If none are free, we may already have an iod working on this mount
1547 * point. If so, it will process our request.
1550 if (nmp->nm_bufqiods > 0) {
1552 ("ncl_asyncio: %d iods are already processing mount %p\n",
1553 nmp->nm_bufqiods, nmp));
1559 * If we have an iod which can process the request, then queue
1564 * Ensure that the queue never grows too large. We still want
1565 * to asynchronize so we block rather then return EIO.
1567 while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1569 ("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1570 nmp->nm_bufqwant = TRUE;
1571 error = newnfs_msleep(td, &nmp->nm_bufq,
1572 &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1575 error2 = newnfs_sigintr(nmp, td);
1580 if (slpflag == PCATCH) {
1586 * We might have lost our iod while sleeping,
1587 * so check and loop if necessary.
1592 /* We might have lost our nfsiod */
1593 if (nmp->nm_bufqiods == 0) {
1595 ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1599 if (bp->b_iocmd == BIO_READ) {
1600 if (bp->b_rcred == NOCRED && cred != NOCRED)
1601 bp->b_rcred = crhold(cred);
1603 if (bp->b_wcred == NOCRED && cred != NOCRED)
1604 bp->b_wcred = crhold(cred);
1607 if (bp->b_flags & B_REMFREE)
1610 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1612 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1613 NFSLOCKNODE(VTONFS(bp->b_vp));
1614 VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1615 VTONFS(bp->b_vp)->n_directio_asyncwr++;
1616 NFSUNLOCKNODE(VTONFS(bp->b_vp));
1625 * All the iods are busy on other mounts, so return EIO to
1626 * force the caller to process the i/o synchronously.
1628 NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1633 ncl_doio_directwrite(struct buf *bp)
1635 int iomode, must_commit;
1636 struct uio *uiop = (struct uio *)bp->b_caller1;
1637 char *iov_base = uiop->uio_iov->iov_base;
1639 iomode = NFSWRITE_FILESYNC;
1640 uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1642 * When doing direct I/O we do not care if the
1643 * server's write verifier has changed, but we
1644 * do not want to update the verifier if it has
1645 * changed, since that hides the change from
1646 * writes being done through the buffer cache.
1647 * By passing must_commit in set to two, the code
1648 * in nfsrpc_writerpc() will not update the
1649 * verifier on the mount point.
1652 ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0, 0);
1653 KASSERT((must_commit == 2), ("ncl_doio_directwrite: Updated write"
1655 if (iomode != NFSWRITE_FILESYNC)
1656 printf("ncl_doio_directwrite: Broken server "
1657 "did not reply FILE_SYNC\n");
1658 free(iov_base, M_NFSDIRECTIO);
1659 free(uiop->uio_iov, M_NFSDIRECTIO);
1660 free(uiop, M_NFSDIRECTIO);
1661 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1662 struct nfsnode *np = VTONFS(bp->b_vp);
1664 if (NFSHASPNFS(VFSTONFS(bp->b_vp->v_mount))) {
1666 * Invalidate the attribute cache, since writes to a DS
1667 * won't update the size attribute.
1669 np->n_attrstamp = 0;
1671 np->n_directio_asyncwr--;
1672 if (np->n_directio_asyncwr == 0) {
1673 np->n_flag &= ~NMODIFIED;
1674 if ((np->n_flag & NFSYNCWAIT)) {
1675 np->n_flag &= ~NFSYNCWAIT;
1676 wakeup((caddr_t)&np->n_directio_asyncwr);
1682 uma_zfree(ncl_pbuf_zone, bp);
1686 * Do an I/O operation to/from a cache block. This may be called
1687 * synchronously or from an nfsiod.
1690 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1691 int called_from_strategy)
1695 struct nfsmount *nmp;
1696 int error = 0, iomode, must_commit = 0;
1699 struct proc *p = td ? td->td_proc : NULL;
1703 nmp = VFSTONFS(vp->v_mount);
1705 uiop->uio_iov = &io;
1706 uiop->uio_iovcnt = 1;
1707 uiop->uio_segflg = UIO_SYSSPACE;
1711 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We
1712 * do this here so we do not have to do it in all the code that
1715 bp->b_flags &= ~B_INVAL;
1716 bp->b_ioflags &= ~BIO_ERROR;
1718 KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1719 iocmd = bp->b_iocmd;
1720 if (iocmd == BIO_READ) {
1721 io.iov_len = uiop->uio_resid = bp->b_bcount;
1722 io.iov_base = bp->b_data;
1723 uiop->uio_rw = UIO_READ;
1725 switch (vp->v_type) {
1727 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1728 NFSINCRGLOBAL(nfsstatsv1.read_bios);
1729 error = ncl_readrpc(vp, uiop, cr);
1732 if (uiop->uio_resid) {
1734 * If we had a short read with no error, we must have
1735 * hit a file hole. We should zero-fill the remainder.
1736 * This can also occur if the server hits the file EOF.
1738 * Holes used to be able to occur due to pending
1739 * writes, but that is not possible any longer.
1741 int nread = bp->b_bcount - uiop->uio_resid;
1742 ssize_t left = uiop->uio_resid;
1745 bzero((char *)bp->b_data + nread, left);
1746 uiop->uio_resid = 0;
1749 /* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1750 if (p && vp->v_writecount <= -1) {
1752 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1755 killproc(p, "text file modification");
1762 uiop->uio_offset = (off_t)0;
1763 NFSINCRGLOBAL(nfsstatsv1.readlink_bios);
1764 error = ncl_readlinkrpc(vp, uiop, cr);
1767 NFSINCRGLOBAL(nfsstatsv1.readdir_bios);
1768 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1769 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1770 error = ncl_readdirplusrpc(vp, uiop, cr, td);
1771 if (error == NFSERR_NOTSUPP)
1772 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1774 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1775 error = ncl_readdirrpc(vp, uiop, cr, td);
1777 * end-of-directory sets B_INVAL but does not generate an
1780 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1781 bp->b_flags |= B_INVAL;
1784 printf("ncl_doio: type %x unexpected\n", vp->v_type);
1788 bp->b_ioflags |= BIO_ERROR;
1789 bp->b_error = error;
1793 * If we only need to commit, try to commit
1795 if (bp->b_flags & B_NEEDCOMMIT) {
1799 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1800 retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1802 if (NFSCL_FORCEDISM(vp->v_mount) || retv == 0) {
1803 bp->b_dirtyoff = bp->b_dirtyend = 0;
1804 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1809 if (retv == NFSERR_STALEWRITEVERF) {
1810 ncl_clearcommit(vp->v_mount);
1815 * Setup for actual write
1818 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1819 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1822 if (bp->b_dirtyend > bp->b_dirtyoff) {
1823 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1825 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1827 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1828 uiop->uio_rw = UIO_WRITE;
1829 NFSINCRGLOBAL(nfsstatsv1.write_bios);
1831 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1832 iomode = NFSWRITE_UNSTABLE;
1834 iomode = NFSWRITE_FILESYNC;
1836 error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1837 called_from_strategy, 0);
1840 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1841 * to cluster the buffers needing commit. This will allow
1842 * the system to submit a single commit rpc for the whole
1843 * cluster. We can do this even if the buffer is not 100%
1844 * dirty (relative to the NFS blocksize), so we optimize the
1845 * append-to-file-case.
1847 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1848 * cleared because write clustering only works for commit
1849 * rpc's, not for the data portion of the write).
1852 if (!error && iomode == NFSWRITE_UNSTABLE) {
1853 bp->b_flags |= B_NEEDCOMMIT;
1854 if (bp->b_dirtyoff == 0
1855 && bp->b_dirtyend == bp->b_bcount)
1856 bp->b_flags |= B_CLUSTEROK;
1858 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1862 * For an interrupted write, the buffer is still valid
1863 * and the write hasn't been pushed to the server yet,
1864 * so we can't set BIO_ERROR and report the interruption
1865 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1866 * is not relevant, so the rpc attempt is essentially
1867 * a noop. For the case of a V3 write rpc not being
1868 * committed to stable storage, the block is still
1869 * dirty and requires either a commit rpc or another
1870 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1871 * the block is reused. This is indicated by setting
1872 * the B_DELWRI and B_NEEDCOMMIT flags.
1874 * EIO is returned by ncl_writerpc() to indicate a recoverable
1875 * write error and is handled as above, except that
1876 * B_EINTR isn't set. One cause of this is a stale stateid
1877 * error for the RPC that indicates recovery is required,
1878 * when called with called_from_strategy != 0.
1880 * If the buffer is marked B_PAGING, it does not reside on
1881 * the vp's paging queues so we cannot call bdirty(). The
1882 * bp in this case is not an NFS cache block so we should
1885 * The logic below breaks up errors into recoverable and
1886 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1887 * and keep the buffer around for potential write retries.
1888 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1889 * and save the error in the nfsnode. This is less than ideal
1890 * but necessary. Keeping such buffers around could potentially
1891 * cause buffer exhaustion eventually (they can never be written
1892 * out, so will get constantly be re-dirtied). It also causes
1893 * all sorts of vfs panics. For non-recoverable write errors,
1894 * also invalidate the attrcache, so we'll be forced to go over
1895 * the wire for this object, returning an error to user on next
1896 * call (most of the time).
1898 if (error == EINTR || error == EIO || error == ETIMEDOUT
1899 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1900 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1901 if ((bp->b_flags & B_PAGING) == 0) {
1903 bp->b_flags &= ~B_DONE;
1905 if ((error == EINTR || error == ETIMEDOUT) &&
1906 (bp->b_flags & B_ASYNC) == 0)
1907 bp->b_flags |= B_EINTR;
1910 bp->b_ioflags |= BIO_ERROR;
1911 bp->b_flags |= B_INVAL;
1912 bp->b_error = np->n_error = error;
1914 np->n_flag |= NWRITEERR;
1915 np->n_attrstamp = 0;
1916 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1919 bp->b_dirtyoff = bp->b_dirtyend = 0;
1927 bp->b_resid = uiop->uio_resid;
1928 if (must_commit == 1)
1929 ncl_clearcommit(vp->v_mount);
1935 * Used to aid in handling ftruncate() operations on the NFS client side.
1936 * Truncation creates a number of special problems for NFS. We have to
1937 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1938 * we have to properly handle VM pages or (potentially dirty) buffers
1939 * that straddle the truncation point.
1943 ncl_meta_setsize(struct vnode *vp, struct thread *td, u_quad_t nsize)
1945 struct nfsnode *np = VTONFS(vp);
1947 int biosize = vp->v_bufobj.bo_bsize;
1955 if (nsize < tsize) {
1961 * vtruncbuf() doesn't get the buffer overlapping the
1962 * truncation point. We may have a B_DELWRI and/or B_CACHE
1963 * buffer that now needs to be truncated.
1965 error = vtruncbuf(vp, nsize, biosize);
1966 lbn = nsize / biosize;
1967 bufsize = nsize - (lbn * biosize);
1968 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1971 if (bp->b_dirtyoff > bp->b_bcount)
1972 bp->b_dirtyoff = bp->b_bcount;
1973 if (bp->b_dirtyend > bp->b_bcount)
1974 bp->b_dirtyend = bp->b_bcount;
1975 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1978 vnode_pager_setsize(vp, nsize);