2 * SPDX-License-Identifier: BSD-3-Clause
4 * Copyright (c) 1989, 1993
5 * The Regents of the University of California. All rights reserved.
7 * This code is derived from software contributed to Berkeley by
8 * Rick Macklem at The University of Guelph.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * @(#)nfs_bio.c 8.9 (Berkeley) 3/30/95
37 #include <sys/cdefs.h>
38 __FBSDID("$FreeBSD$");
40 #include <sys/param.h>
41 #include <sys/systm.h>
44 #include <sys/kernel.h>
45 #include <sys/mount.h>
46 #include <sys/rwlock.h>
47 #include <sys/vmmeter.h>
48 #include <sys/vnode.h>
51 #include <vm/vm_param.h>
52 #include <vm/vm_extern.h>
53 #include <vm/vm_page.h>
54 #include <vm/vm_object.h>
55 #include <vm/vm_pager.h>
56 #include <vm/vnode_pager.h>
58 #include <fs/nfs/nfsport.h>
59 #include <fs/nfsclient/nfsmount.h>
60 #include <fs/nfsclient/nfs.h>
61 #include <fs/nfsclient/nfsnode.h>
62 #include <fs/nfsclient/nfs_kdtrace.h>
64 extern int newnfs_directio_allow_mmap;
65 extern struct nfsstatsv1 nfsstatsv1;
66 extern struct mtx ncl_iod_mutex;
67 extern int ncl_numasync;
68 extern enum nfsiod_state ncl_iodwant[NFS_MAXASYNCDAEMON];
69 extern struct nfsmount *ncl_iodmount[NFS_MAXASYNCDAEMON];
70 extern int newnfs_directio_enable;
71 extern int nfs_keep_dirty_on_error;
73 uma_zone_t ncl_pbuf_zone;
75 static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size,
77 static int nfs_directio_write(struct vnode *vp, struct uio *uiop,
78 struct ucred *cred, int ioflag);
81 * Vnode op for VM getpages.
83 SYSCTL_DECL(_vfs_nfs);
84 static int use_buf_pager = 1;
85 SYSCTL_INT(_vfs_nfs, OID_AUTO, use_buf_pager, CTLFLAG_RWTUN,
87 "Use buffer pager instead of direct readrpc call");
90 ncl_gbp_getblkno(struct vnode *vp, vm_ooffset_t off)
93 return (off / vp->v_bufobj.bo_bsize);
97 ncl_gbp_getblksz(struct vnode *vp, daddr_t lbn, long *sz)
108 biosize = vp->v_bufobj.bo_bsize;
110 if ((off_t)lbn * biosize >= nsize)
112 else if ((off_t)(lbn + 1) * biosize > nsize)
113 bcount = nsize - (off_t)lbn * biosize;
119 ncl_getpages(struct vop_getpages_args *ap)
121 int i, error, nextoff, size, toff, count, npages;
129 struct nfsmount *nmp;
137 cred = curthread->td_ucred;
138 nmp = VFSTONFS(vp->v_mount);
140 npages = ap->a_count;
142 if ((object = vp->v_object) == NULL) {
143 printf("ncl_getpages: called with non-merged cache vnode\n");
144 return (VM_PAGER_ERROR);
147 if (newnfs_directio_enable && !newnfs_directio_allow_mmap) {
149 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
151 printf("ncl_getpages: called on non-cacheable vnode\n");
152 return (VM_PAGER_ERROR);
157 mtx_lock(&nmp->nm_mtx);
158 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
159 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
160 mtx_unlock(&nmp->nm_mtx);
161 /* We'll never get here for v4, because we always have fsinfo */
162 (void)ncl_fsinfo(nmp, vp, cred, td);
164 mtx_unlock(&nmp->nm_mtx);
167 return (vfs_bio_getpages(vp, pages, npages, ap->a_rbehind,
168 ap->a_rahead, ncl_gbp_getblkno, ncl_gbp_getblksz));
171 * If the requested page is partially valid, just return it and
172 * allow the pager to zero-out the blanks. Partially valid pages
173 * can only occur at the file EOF.
175 * XXXGL: is that true for NFS, where short read can occur???
177 VM_OBJECT_WLOCK(object);
178 if (!vm_page_none_valid(pages[npages - 1]) && --npages == 0)
180 VM_OBJECT_WUNLOCK(object);
183 * We use only the kva address for the buffer, but this is extremely
184 * convenient and fast.
186 bp = uma_zalloc(ncl_pbuf_zone, M_WAITOK);
188 kva = (vm_offset_t) bp->b_data;
189 pmap_qenter(kva, pages, npages);
190 VM_CNT_INC(v_vnodein);
191 VM_CNT_ADD(v_vnodepgsin, npages);
193 count = npages << PAGE_SHIFT;
194 iov.iov_base = (caddr_t) kva;
198 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
199 uio.uio_resid = count;
200 uio.uio_segflg = UIO_SYSSPACE;
201 uio.uio_rw = UIO_READ;
204 error = ncl_readrpc(vp, &uio, cred);
205 pmap_qremove(kva, npages);
207 uma_zfree(ncl_pbuf_zone, bp);
209 if (error && (uio.uio_resid == count)) {
210 printf("ncl_getpages: error %d\n", error);
211 return (VM_PAGER_ERROR);
215 * Calculate the number of bytes read and validate only that number
216 * of bytes. Note that due to pending writes, size may be 0. This
217 * does not mean that the remaining data is invalid!
220 size = count - uio.uio_resid;
221 VM_OBJECT_WLOCK(object);
222 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
224 nextoff = toff + PAGE_SIZE;
227 if (nextoff <= size) {
229 * Read operation filled an entire page
232 KASSERT(m->dirty == 0,
233 ("nfs_getpages: page %p is dirty", m));
234 } else if (size > toff) {
236 * Read operation filled a partial page.
239 vm_page_set_valid_range(m, 0, size - toff);
240 KASSERT(m->dirty == 0,
241 ("nfs_getpages: page %p is dirty", m));
244 * Read operation was short. If no error
245 * occurred we may have hit a zero-fill
246 * section. We leave valid set to 0, and page
247 * is freed by vm_page_readahead_finish() if
248 * its index is not equal to requested, or
249 * page is zeroed and set valid by
250 * vm_pager_get_pages() for requested page.
256 VM_OBJECT_WUNLOCK(object);
261 return (VM_PAGER_OK);
265 * Vnode op for VM putpages.
268 ncl_putpages(struct vop_putpages_args *ap)
272 int i, error, npages, count;
278 struct nfsmount *nmp;
284 td = curthread; /* XXX */
285 /* Set the cred to n_writecred for the write rpcs. */
286 if (np->n_writecred != NULL)
287 cred = crhold(np->n_writecred);
289 cred = crhold(curthread->td_ucred); /* XXX */
290 nmp = VFSTONFS(vp->v_mount);
293 rtvals = ap->a_rtvals;
294 npages = btoc(count);
295 offset = IDX_TO_OFF(pages[0]->pindex);
297 mtx_lock(&nmp->nm_mtx);
298 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
299 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
300 mtx_unlock(&nmp->nm_mtx);
301 (void)ncl_fsinfo(nmp, vp, cred, td);
303 mtx_unlock(&nmp->nm_mtx);
306 if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
307 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
309 printf("ncl_putpages: called on noncache-able vnode\n");
313 * When putting pages, do not extend file past EOF.
315 if (offset + count > np->n_size) {
316 count = np->n_size - offset;
322 for (i = 0; i < npages; i++)
323 rtvals[i] = VM_PAGER_ERROR;
325 VM_CNT_INC(v_vnodeout);
326 VM_CNT_ADD(v_vnodepgsout, count);
328 iov.iov_base = unmapped_buf;
332 uio.uio_offset = offset;
333 uio.uio_resid = count;
334 uio.uio_segflg = UIO_NOCOPY;
335 uio.uio_rw = UIO_WRITE;
338 error = VOP_WRITE(vp, &uio, vnode_pager_putpages_ioflags(ap->a_sync),
342 if (error == 0 || !nfs_keep_dirty_on_error) {
343 vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid,
344 np->n_size - offset, npages * PAGE_SIZE);
350 * For nfs, cache consistency can only be maintained approximately.
351 * Although RFC1094 does not specify the criteria, the following is
352 * believed to be compatible with the reference port.
354 * If the file's modify time on the server has changed since the
355 * last read rpc or you have written to the file,
356 * you may have lost data cache consistency with the
357 * server, so flush all of the file's data out of the cache.
358 * Then force a getattr rpc to ensure that you have up to date
360 * NB: This implies that cache data can be read when up to
361 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
362 * attributes this could be forced by setting n_attrstamp to 0 before
363 * the VOP_GETATTR() call.
366 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
370 struct nfsnode *np = VTONFS(vp);
374 * Ensure the exclusove access to the node before checking
375 * whether the cache is consistent.
377 old_lock = ncl_excl_start(vp);
379 if (np->n_flag & NMODIFIED) {
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 | V_ALLOWCLEAN, td, 1);
390 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
391 error = VOP_GETATTR(vp, &vattr, cred);
395 np->n_mtime = vattr.va_mtime;
399 error = VOP_GETATTR(vp, &vattr, cred);
403 if ((np->n_flag & NSIZECHANGED)
404 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
406 if (vp->v_type == VDIR)
408 error = ncl_vinvalbuf(vp, V_SAVE | V_ALLOWCLEAN, td, 1);
412 np->n_mtime = vattr.va_mtime;
413 np->n_flag &= ~NSIZECHANGED;
418 ncl_excl_finish(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);
429 struct buf *bp, *rabp;
431 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
433 int biosize, bcount, error, i, n, nra, on, save2, seqcount;
436 KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
437 if (uio->uio_resid == 0)
439 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
443 mtx_lock(&nmp->nm_mtx);
444 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
445 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
446 mtx_unlock(&nmp->nm_mtx);
447 (void)ncl_fsinfo(nmp, vp, cred, td);
448 mtx_lock(&nmp->nm_mtx);
450 if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
451 (void) newnfs_iosize(nmp);
453 tmp_off = uio->uio_offset + uio->uio_resid;
454 if (vp->v_type != VDIR &&
455 (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) {
456 mtx_unlock(&nmp->nm_mtx);
459 mtx_unlock(&nmp->nm_mtx);
461 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
462 /* No caching/ no readaheads. Just read data into the user buffer */
463 return ncl_readrpc(vp, uio, cred);
467 biosize = vp->v_bufobj.bo_bsize;
468 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
470 error = nfs_bioread_check_cons(vp, td, cred);
474 save2 = curthread_pflags2_set(TDP2_SBPAGES);
482 switch (vp->v_type) {
484 NFSINCRGLOBAL(nfsstatsv1.biocache_reads);
485 lbn = uio->uio_offset / biosize;
486 on = uio->uio_offset - (lbn * biosize);
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);
503 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
504 rabp->b_flags |= B_ASYNC;
505 rabp->b_iocmd = BIO_READ;
506 vfs_busy_pages(rabp, 0);
507 if (ncl_asyncio(nmp, rabp, cred, td)) {
508 rabp->b_flags |= B_INVAL;
509 rabp->b_ioflags |= BIO_ERROR;
510 vfs_unbusy_pages(rabp);
521 /* Note that bcount is *not* DEV_BSIZE aligned. */
523 if ((off_t)lbn * biosize >= nsize) {
525 } else if ((off_t)(lbn + 1) * biosize > nsize) {
526 bcount = nsize - (off_t)lbn * biosize;
528 bp = nfs_getcacheblk(vp, lbn, bcount, td);
531 error = newnfs_sigintr(nmp, td);
538 * If B_CACHE is not set, we must issue the read. If this
539 * fails, we return an error.
542 if ((bp->b_flags & B_CACHE) == 0) {
543 bp->b_iocmd = BIO_READ;
544 vfs_busy_pages(bp, 0);
545 error = ncl_doio(vp, bp, cred, td, 0);
553 * on is the offset into the current bp. Figure out how many
554 * bytes we can copy out of the bp. Note that bcount is
555 * NOT DEV_BSIZE aligned.
557 * Then figure out how many bytes we can copy into the uio.
562 n = MIN((unsigned)(bcount - on), uio->uio_resid);
565 NFSINCRGLOBAL(nfsstatsv1.biocache_readlinks);
566 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
568 error = newnfs_sigintr(nmp, td);
573 if ((bp->b_flags & B_CACHE) == 0) {
574 bp->b_iocmd = BIO_READ;
575 vfs_busy_pages(bp, 0);
576 error = ncl_doio(vp, bp, cred, td, 0);
578 bp->b_ioflags |= BIO_ERROR;
583 n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
587 NFSINCRGLOBAL(nfsstatsv1.biocache_readdirs);
589 if (np->n_direofoffset
590 && uio->uio_offset >= np->n_direofoffset) {
596 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
597 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
598 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
600 error = newnfs_sigintr(nmp, td);
605 if ((bp->b_flags & B_CACHE) == 0) {
606 bp->b_iocmd = BIO_READ;
607 vfs_busy_pages(bp, 0);
608 error = ncl_doio(vp, bp, cred, td, 0);
612 while (error == NFSERR_BAD_COOKIE) {
614 error = ncl_vinvalbuf(vp, 0, td, 1);
617 * Yuck! The directory has been modified on the
618 * server. The only way to get the block is by
619 * reading from the beginning to get all the
622 * Leave the last bp intact unless there is an error.
623 * Loop back up to the while if the error is another
624 * NFSERR_BAD_COOKIE (double yuch!).
626 for (i = 0; i <= lbn && !error; i++) {
628 if (np->n_direofoffset
629 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) {
635 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
637 error = newnfs_sigintr(nmp, td);
642 if ((bp->b_flags & B_CACHE) == 0) {
643 bp->b_iocmd = BIO_READ;
644 vfs_busy_pages(bp, 0);
645 error = ncl_doio(vp, bp, cred, td, 0);
647 * no error + B_INVAL == directory EOF,
650 if (error == 0 && (bp->b_flags & B_INVAL))
654 * An error will throw away the block and the
655 * for loop will break out. If no error and this
656 * is not the block we want, we throw away the
657 * block and go for the next one via the for loop.
659 if (error || i < lbn)
664 * The above while is repeated if we hit another cookie
665 * error. If we hit an error and it wasn't a cookie error,
673 * If not eof and read aheads are enabled, start one.
674 * (You need the current block first, so that you have the
675 * directory offset cookie of the next block.)
678 if (nmp->nm_readahead > 0 &&
679 (bp->b_flags & B_INVAL) == 0 &&
680 (np->n_direofoffset == 0 ||
681 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
682 incore(&vp->v_bufobj, lbn + 1) == NULL) {
684 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
686 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
687 rabp->b_flags |= B_ASYNC;
688 rabp->b_iocmd = BIO_READ;
689 vfs_busy_pages(rabp, 0);
690 if (ncl_asyncio(nmp, rabp, cred, td)) {
691 rabp->b_flags |= B_INVAL;
692 rabp->b_ioflags |= BIO_ERROR;
693 vfs_unbusy_pages(rabp);
703 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
704 * chopped for the EOF condition, we cannot tell how large
705 * NFS directories are going to be until we hit EOF. So
706 * an NFS directory buffer is *not* chopped to its EOF. Now,
707 * it just so happens that b_resid will effectively chop it
708 * to EOF. *BUT* this information is lost if the buffer goes
709 * away and is reconstituted into a B_CACHE state ( due to
710 * being VMIO ) later. So we keep track of the directory eof
711 * in np->n_direofoffset and chop it off as an extra step
714 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
715 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
716 n = np->n_direofoffset - uio->uio_offset;
720 printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
726 error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio);
728 if (vp->v_type == VLNK)
732 } while (error == 0 && uio->uio_resid > 0 && n > 0);
734 curthread_pflags2_restore(save2);
735 if ((curthread->td_pflags2 & TDP2_SBPAGES) == 0) {
737 ncl_pager_setsize(vp, NULL);
743 * The NFS write path cannot handle iovecs with len > 1. So we need to
744 * break up iovecs accordingly (restricting them to wsize).
745 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
746 * For the ASYNC case, 2 copies are needed. The first a copy from the
747 * user buffer to a staging buffer and then a second copy from the staging
748 * buffer to mbufs. This can be optimized by copying from the user buffer
749 * directly into mbufs and passing the chain down, but that requires a
750 * fair amount of re-working of the relevant codepaths (and can be done
754 nfs_directio_write(struct vnode *vp, struct uio *uiop, struct ucred *cred,
758 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
759 struct thread *td = uiop->uio_td;
763 mtx_lock(&nmp->nm_mtx);
764 wsize = nmp->nm_wsize;
765 mtx_unlock(&nmp->nm_mtx);
766 if (ioflag & IO_SYNC) {
767 int iomode, must_commit;
771 while (uiop->uio_resid > 0) {
772 size = MIN(uiop->uio_resid, wsize);
773 size = MIN(uiop->uio_iov->iov_len, size);
774 iov.iov_base = uiop->uio_iov->iov_base;
778 uio.uio_offset = uiop->uio_offset;
779 uio.uio_resid = size;
780 uio.uio_segflg = uiop->uio_segflg;
781 uio.uio_rw = UIO_WRITE;
783 iomode = NFSWRITE_FILESYNC;
785 * When doing direct I/O we do not care if the
786 * server's write verifier has changed, but we
787 * do not want to update the verifier if it has
788 * changed, since that hides the change from
789 * writes being done through the buffer cache.
790 * By passing must_commit in set to two, the code
791 * in nfsrpc_writerpc() will not update the
792 * verifier on the mount point.
795 error = ncl_writerpc(vp, &uio, cred, &iomode,
796 &must_commit, 0, ioflag);
797 KASSERT((must_commit == 2),
798 ("ncl_directio_write: Updated write verifier"));
801 if (iomode != NFSWRITE_FILESYNC)
802 printf("nfs_directio_write: Broken server "
803 "did not reply FILE_SYNC\n");
804 uiop->uio_offset += size;
805 uiop->uio_resid -= size;
806 if (uiop->uio_iov->iov_len <= size) {
810 uiop->uio_iov->iov_base =
811 (char *)uiop->uio_iov->iov_base + size;
812 uiop->uio_iov->iov_len -= size;
821 * Break up the write into blocksize chunks and hand these
822 * over to nfsiod's for write back.
823 * Unfortunately, this incurs a copy of the data. Since
824 * the user could modify the buffer before the write is
827 * The obvious optimization here is that one of the 2 copies
828 * in the async write path can be eliminated by copying the
829 * data here directly into mbufs and passing the mbuf chain
830 * down. But that will require a fair amount of re-working
831 * of the code and can be done if there's enough interest
832 * in NFS directio access.
834 while (uiop->uio_resid > 0) {
835 size = MIN(uiop->uio_resid, wsize);
836 size = MIN(uiop->uio_iov->iov_len, size);
837 bp = uma_zalloc(ncl_pbuf_zone, M_WAITOK);
838 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
839 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
840 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
841 t_iov->iov_len = size;
842 t_uio->uio_iov = t_iov;
843 t_uio->uio_iovcnt = 1;
844 t_uio->uio_offset = uiop->uio_offset;
845 t_uio->uio_resid = size;
846 t_uio->uio_segflg = UIO_SYSSPACE;
847 t_uio->uio_rw = UIO_WRITE;
849 KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
850 uiop->uio_segflg == UIO_SYSSPACE,
851 ("nfs_directio_write: Bad uio_segflg"));
852 if (uiop->uio_segflg == UIO_USERSPACE) {
853 error = copyin(uiop->uio_iov->iov_base,
854 t_iov->iov_base, size);
859 * UIO_SYSSPACE may never happen, but handle
860 * it just in case it does.
862 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
864 bp->b_flags |= B_DIRECT;
865 bp->b_iocmd = BIO_WRITE;
866 if (cred != NOCRED) {
870 bp->b_wcred = NOCRED;
871 bp->b_caller1 = (void *)t_uio;
873 error = ncl_asyncio(nmp, bp, NOCRED, td);
876 free(t_iov->iov_base, M_NFSDIRECTIO);
877 free(t_iov, M_NFSDIRECTIO);
878 free(t_uio, M_NFSDIRECTIO);
880 uma_zfree(ncl_pbuf_zone, bp);
885 uiop->uio_offset += size;
886 uiop->uio_resid -= size;
887 if (uiop->uio_iov->iov_len <= size) {
891 uiop->uio_iov->iov_base =
892 (char *)uiop->uio_iov->iov_base + size;
893 uiop->uio_iov->iov_len -= size;
901 * Vnode op for write using bio
904 ncl_write(struct vop_write_args *ap)
907 struct uio *uio = ap->a_uio;
908 struct thread *td = uio->uio_td;
909 struct vnode *vp = ap->a_vp;
910 struct nfsnode *np = VTONFS(vp);
911 struct ucred *cred = ap->a_cred;
912 int ioflag = ap->a_ioflag;
915 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
917 int bcount, noncontig_write, obcount;
918 int bp_cached, n, on, error = 0, error1, save2, wouldcommit;
919 size_t orig_resid, local_resid;
920 off_t orig_size, tmp_off;
923 KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
924 KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
926 if (vp->v_type != VREG)
929 if (np->n_flag & NWRITEERR) {
930 np->n_flag &= ~NWRITEERR;
932 return (np->n_error);
935 mtx_lock(&nmp->nm_mtx);
936 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
937 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
938 mtx_unlock(&nmp->nm_mtx);
939 (void)ncl_fsinfo(nmp, vp, cred, td);
940 mtx_lock(&nmp->nm_mtx);
942 if (nmp->nm_wsize == 0)
943 (void) newnfs_iosize(nmp);
944 mtx_unlock(&nmp->nm_mtx);
947 * Synchronously flush pending buffers if we are in synchronous
948 * mode or if we are appending.
950 if ((ioflag & IO_APPEND) || ((ioflag & IO_SYNC) && (np->n_flag &
953 * For the case where IO_APPEND is being done using a
954 * direct output (to the NFS server) RPC and
955 * newnfs_directio_enable is 0, all buffer cache buffers,
956 * including ones not modified, must be invalidated.
957 * This ensures that stale data is not read out of the
958 * buffer cache. The call also invalidates all mapped
959 * pages and, since the exclusive lock is held on the vnode,
960 * new pages cannot be faulted in.
962 * For the case where newnfs_directio_enable is set
963 * (which is not the default), it is not obvious that
964 * stale data should be left in the buffer cache, but
965 * the code has been this way for over a decade without
966 * complaints. Note that, unlike doing IO_APPEND via
967 * a direct write RPC when newnfs_directio_enable is not set,
968 * when newnfs_directio_enable is set, reading is done via
969 * direct to NFS server RPCs as well.
972 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
973 error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
974 IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
979 orig_resid = uio->uio_resid;
981 orig_size = np->n_size;
985 * If IO_APPEND then load uio_offset. We restart here if we cannot
986 * get the append lock.
988 if (ioflag & IO_APPEND) {
990 * For NFSv4, the AppendWrite will Verify the size against
991 * the file's size on the server. If not the same, the
992 * write will then be retried, using the file size returned
993 * by the AppendWrite. However, for NFSv2 and NFSv3, the
994 * size must be acquired here via a Getattr RPC.
995 * The AppendWrite is not done for a pNFS mount.
997 if (!NFSHASNFSV4(nmp) || NFSHASPNFS(nmp)) {
999 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1000 error = VOP_GETATTR(vp, &vattr, cred);
1005 uio->uio_offset = np->n_size;
1009 if (uio->uio_offset < 0)
1011 tmp_off = uio->uio_offset + uio->uio_resid;
1012 if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
1014 if (uio->uio_resid == 0)
1018 * Do IO_APPEND writing via a synchronous direct write.
1019 * This can result in a significant performance improvement.
1021 if ((newnfs_directio_enable && (ioflag & IO_DIRECT)) ||
1022 (ioflag & IO_APPEND)) {
1024 * Direct writes to the server must be done NFSWRITE_FILESYNC,
1025 * because the write data is not cached and, therefore, the
1026 * write cannot be redone after a server reboot.
1027 * Set IO_SYNC to make this happen.
1030 return (nfs_directio_write(vp, uio, cred, ioflag));
1034 * Maybe this should be above the vnode op call, but so long as
1035 * file servers have no limits, i don't think it matters
1037 if (vn_rlimit_fsize(vp, uio, td))
1040 save2 = curthread_pflags2_set(TDP2_SBPAGES);
1041 biosize = vp->v_bufobj.bo_bsize;
1043 * Find all of this file's B_NEEDCOMMIT buffers. If our writes
1044 * would exceed the local maximum per-file write commit size when
1045 * combined with those, we must decide whether to flush,
1046 * go synchronous, or return error. We don't bother checking
1047 * IO_UNIT -- we just make all writes atomic anyway, as there's
1048 * no point optimizing for something that really won't ever happen.
1051 if (!(ioflag & IO_SYNC)) {
1057 if (nflag & NMODIFIED) {
1058 BO_LOCK(&vp->v_bufobj);
1059 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
1060 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
1062 if (bp->b_flags & B_NEEDCOMMIT)
1063 wouldcommit += bp->b_bcount;
1066 BO_UNLOCK(&vp->v_bufobj);
1071 if (!(ioflag & IO_SYNC)) {
1072 wouldcommit += biosize;
1073 if (wouldcommit > nmp->nm_wcommitsize) {
1074 np->n_attrstamp = 0;
1075 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1076 error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
1077 IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
1080 wouldcommit = biosize;
1084 NFSINCRGLOBAL(nfsstatsv1.biocache_writes);
1085 lbn = uio->uio_offset / biosize;
1086 on = uio->uio_offset - (lbn * biosize);
1087 n = MIN((unsigned)(biosize - on), uio->uio_resid);
1090 * Handle direct append and file extension cases, calculate
1091 * unaligned buffer size.
1094 if ((np->n_flag & NHASBEENLOCKED) == 0 &&
1095 (nmp->nm_flag & NFSMNT_NONCONTIGWR) != 0)
1096 noncontig_write = 1;
1098 noncontig_write = 0;
1099 if ((uio->uio_offset == np->n_size ||
1100 (noncontig_write != 0 &&
1101 lbn == (np->n_size / biosize) &&
1102 uio->uio_offset + n > np->n_size)) && n) {
1105 * Get the buffer (in its pre-append state to maintain
1106 * B_CACHE if it was previously set). Resize the
1107 * nfsnode after we have locked the buffer to prevent
1108 * readers from reading garbage.
1110 obcount = np->n_size - (lbn * biosize);
1111 bp = nfs_getcacheblk(vp, lbn, obcount, td);
1117 np->n_size = uio->uio_offset + n;
1118 np->n_flag |= NMODIFIED;
1119 np->n_flag &= ~NVNSETSZSKIP;
1120 vnode_pager_setsize(vp, np->n_size);
1123 save = bp->b_flags & B_CACHE;
1125 allocbuf(bp, bcount);
1126 bp->b_flags |= save;
1127 if (noncontig_write != 0 && on > obcount)
1128 vfs_bio_bzero_buf(bp, obcount, on -
1133 * Obtain the locked cache block first, and then
1134 * adjust the file's size as appropriate.
1137 if ((off_t)lbn * biosize + bcount < np->n_size) {
1138 if ((off_t)(lbn + 1) * biosize < np->n_size)
1141 bcount = np->n_size - (off_t)lbn * biosize;
1144 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1146 if (uio->uio_offset + n > np->n_size) {
1147 np->n_size = uio->uio_offset + n;
1148 np->n_flag |= NMODIFIED;
1149 np->n_flag &= ~NVNSETSZSKIP;
1150 vnode_pager_setsize(vp, np->n_size);
1156 error = newnfs_sigintr(nmp, td);
1163 * Issue a READ if B_CACHE is not set. In special-append
1164 * mode, B_CACHE is based on the buffer prior to the write
1165 * op and is typically set, avoiding the read. If a read
1166 * is required in special append mode, the server will
1167 * probably send us a short-read since we extended the file
1168 * on our end, resulting in b_resid == 0 and, thusly,
1169 * B_CACHE getting set.
1171 * We can also avoid issuing the read if the write covers
1172 * the entire buffer. We have to make sure the buffer state
1173 * is reasonable in this case since we will not be initiating
1174 * I/O. See the comments in kern/vfs_bio.c's getblk() for
1177 * B_CACHE may also be set due to the buffer being cached
1182 if (on == 0 && n == bcount) {
1183 if ((bp->b_flags & B_CACHE) == 0)
1185 bp->b_flags |= B_CACHE;
1186 bp->b_flags &= ~B_INVAL;
1187 bp->b_ioflags &= ~BIO_ERROR;
1190 if ((bp->b_flags & B_CACHE) == 0) {
1191 bp->b_iocmd = BIO_READ;
1192 vfs_busy_pages(bp, 0);
1193 error = ncl_doio(vp, bp, cred, td, 0);
1199 if (bp->b_wcred == NOCRED)
1200 bp->b_wcred = crhold(cred);
1202 np->n_flag |= NMODIFIED;
1206 * If dirtyend exceeds file size, chop it down. This should
1207 * not normally occur but there is an append race where it
1208 * might occur XXX, so we log it.
1210 * If the chopping creates a reverse-indexed or degenerate
1211 * situation with dirtyoff/end, we 0 both of them.
1214 if (bp->b_dirtyend > bcount) {
1215 printf("NFS append race @%lx:%d\n",
1216 (long)bp->b_blkno * DEV_BSIZE,
1217 bp->b_dirtyend - bcount);
1218 bp->b_dirtyend = bcount;
1221 if (bp->b_dirtyoff >= bp->b_dirtyend)
1222 bp->b_dirtyoff = bp->b_dirtyend = 0;
1225 * If the new write will leave a contiguous dirty
1226 * area, just update the b_dirtyoff and b_dirtyend,
1227 * otherwise force a write rpc of the old dirty area.
1229 * If there has been a file lock applied to this file
1230 * or vfs.nfs.old_noncontig_writing is set, do the following:
1231 * While it is possible to merge discontiguous writes due to
1232 * our having a B_CACHE buffer ( and thus valid read data
1233 * for the hole), we don't because it could lead to
1234 * significant cache coherency problems with multiple clients,
1235 * especially if locking is implemented later on.
1237 * If vfs.nfs.old_noncontig_writing is not set and there has
1238 * not been file locking done on this file:
1239 * Relax coherency a bit for the sake of performance and
1240 * expand the current dirty region to contain the new
1241 * write even if it means we mark some non-dirty data as
1245 if (noncontig_write == 0 && bp->b_dirtyend > 0 &&
1246 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1247 if (bwrite(bp) == EINTR) {
1254 local_resid = uio->uio_resid;
1255 error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio);
1257 if (error != 0 && !bp_cached) {
1259 * This block has no other content then what
1260 * possibly was written by the faulty uiomove.
1261 * Release it, forgetting the data pages, to
1262 * prevent the leak of uninitialized data to
1265 bp->b_ioflags |= BIO_ERROR;
1267 uio->uio_offset -= local_resid - uio->uio_resid;
1268 uio->uio_resid = local_resid;
1273 * Since this block is being modified, it must be written
1274 * again and not just committed. Since write clustering does
1275 * not work for the stage 1 data write, only the stage 2
1276 * commit rpc, we have to clear B_CLUSTEROK as well.
1278 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1281 * Get the partial update on the progress made from
1282 * uiomove, if an error occurred.
1285 n = local_resid - uio->uio_resid;
1288 * Only update dirtyoff/dirtyend if not a degenerate
1292 if (bp->b_dirtyend > 0) {
1293 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1294 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1296 bp->b_dirtyoff = on;
1297 bp->b_dirtyend = on + n;
1299 vfs_bio_set_valid(bp, on, n);
1303 * If IO_SYNC do bwrite().
1305 * IO_INVAL appears to be unused. The idea appears to be
1306 * to turn off caching in this case. Very odd. XXX
1308 if ((ioflag & IO_SYNC)) {
1309 if (ioflag & IO_INVAL)
1310 bp->b_flags |= B_NOCACHE;
1311 error1 = bwrite(bp);
1317 } else if ((n + on) == biosize || (ioflag & IO_ASYNC) != 0) {
1318 bp->b_flags |= B_ASYNC;
1319 (void) ncl_writebp(bp, 0, NULL);
1326 } while (uio->uio_resid > 0 && n > 0);
1331 np->n_localmodtime = ts;
1334 if (ioflag & IO_UNIT) {
1336 vattr.va_size = orig_size;
1337 /* IO_SYNC is handled implicitely */
1338 (void)VOP_SETATTR(vp, &vattr, cred);
1339 uio->uio_offset -= orig_resid - uio->uio_resid;
1340 uio->uio_resid = orig_resid;
1345 curthread_pflags2_restore(save2);
1350 * Get an nfs cache block.
1352 * Allocate a new one if the block isn't currently in the cache
1353 * and return the block marked busy. If the calling process is
1354 * interrupted by a signal for an interruptible mount point, return
1357 * The caller must carefully deal with the possible B_INVAL state of
1358 * the buffer. ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
1359 * indirectly), so synchronous reads can be issued without worrying about
1360 * the B_INVAL state. We have to be a little more careful when dealing
1361 * with writes (see comments in nfs_write()) when extending a file past
1365 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1369 struct nfsmount *nmp;
1374 if (nmp->nm_flag & NFSMNT_INT) {
1377 newnfs_set_sigmask(td, &oldset);
1378 bp = getblk(vp, bn, size, PCATCH, 0, 0);
1379 newnfs_restore_sigmask(td, &oldset);
1380 while (bp == NULL) {
1381 if (newnfs_sigintr(nmp, td))
1383 bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1386 bp = getblk(vp, bn, size, 0, 0, 0);
1389 if (vp->v_type == VREG)
1390 bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1395 * Flush and invalidate all dirty buffers. If another process is already
1396 * doing the flush, just wait for completion.
1399 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1401 struct nfsnode *np = VTONFS(vp);
1402 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1403 int error = 0, slpflag, slptimeo;
1407 ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
1409 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1411 if (NFSCL_FORCEDISM(nmp->nm_mountp))
1421 old_lock = ncl_excl_start(vp);
1423 flags |= V_ALLOWCLEAN;
1426 * Now, flush as required.
1428 if ((flags & (V_SAVE | V_VMIO)) == V_SAVE &&
1429 vp->v_bufobj.bo_object != NULL) {
1430 VM_OBJECT_WLOCK(vp->v_bufobj.bo_object);
1431 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1432 VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object);
1434 * If the page clean was interrupted, fail the invalidation.
1435 * Not doing so, we run the risk of losing dirty pages in the
1436 * vinvalbuf() call below.
1438 if (intrflg && (error = newnfs_sigintr(nmp, td)))
1442 error = vinvalbuf(vp, flags, slpflag, 0);
1444 if (intrflg && (error = newnfs_sigintr(nmp, td)))
1446 error = vinvalbuf(vp, flags, 0, slptimeo);
1448 if (NFSHASPNFS(nmp)) {
1449 nfscl_layoutcommit(vp, td);
1452 * Invalidate the attribute cache, since writes to a DS
1453 * won't update the size attribute.
1456 np->n_attrstamp = 0;
1461 if (np->n_directio_asyncwr == 0 && (np->n_flag & NMODIFIED) != 0) {
1462 np->n_localmodtime = ts;
1463 np->n_flag &= ~NMODIFIED;
1467 ncl_excl_finish(vp, old_lock);
1472 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1473 * This is mainly to avoid queueing async I/O requests when the nfsiods
1474 * are all hung on a dead server.
1476 * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1477 * is eventually dequeued by the async daemon, ncl_doio() *will*.
1480 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1489 * Commits are usually short and sweet so lets save some cpu and
1490 * leave the async daemons for more important rpc's (such as reads
1493 * Readdirplus RPCs do vget()s to acquire the vnodes for entries
1494 * in the directory in order to update attributes. This can deadlock
1495 * with another thread that is waiting for async I/O to be done by
1496 * an nfsiod thread while holding a lock on one of these vnodes.
1497 * To avoid this deadlock, don't allow the async nfsiod threads to
1498 * perform Readdirplus RPCs.
1501 if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1502 (nmp->nm_bufqiods > ncl_numasync / 2)) ||
1503 (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
1508 if (nmp->nm_flag & NFSMNT_INT)
1513 * Find a free iod to process this request.
1515 for (iod = 0; iod < ncl_numasync; iod++)
1516 if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1522 * Try to create one if none are free.
1528 * Found one, so wake it up and tell it which
1531 NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1533 ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1534 ncl_iodmount[iod] = nmp;
1536 wakeup(&ncl_iodwant[iod]);
1540 * If none are free, we may already have an iod working on this mount
1541 * point. If so, it will process our request.
1544 if (nmp->nm_bufqiods > 0) {
1546 ("ncl_asyncio: %d iods are already processing mount %p\n",
1547 nmp->nm_bufqiods, nmp));
1553 * If we have an iod which can process the request, then queue
1558 * Ensure that the queue never grows too large. We still want
1559 * to asynchronize so we block rather then return EIO.
1561 while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1563 ("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1564 nmp->nm_bufqwant = TRUE;
1565 error = newnfs_msleep(td, &nmp->nm_bufq,
1566 &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1569 error2 = newnfs_sigintr(nmp, td);
1574 if (slpflag == PCATCH) {
1580 * We might have lost our iod while sleeping,
1581 * so check and loop if necessary.
1586 /* We might have lost our nfsiod */
1587 if (nmp->nm_bufqiods == 0) {
1589 ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1593 if (bp->b_iocmd == BIO_READ) {
1594 if (bp->b_rcred == NOCRED && cred != NOCRED)
1595 bp->b_rcred = crhold(cred);
1597 if (bp->b_wcred == NOCRED && cred != NOCRED)
1598 bp->b_wcred = crhold(cred);
1601 if (bp->b_flags & B_REMFREE)
1604 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1606 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1607 NFSLOCKNODE(VTONFS(bp->b_vp));
1608 VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1609 VTONFS(bp->b_vp)->n_directio_asyncwr++;
1610 NFSUNLOCKNODE(VTONFS(bp->b_vp));
1619 * All the iods are busy on other mounts, so return EIO to
1620 * force the caller to process the i/o synchronously.
1622 NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1627 ncl_doio_directwrite(struct buf *bp)
1629 int iomode, must_commit;
1630 struct uio *uiop = (struct uio *)bp->b_caller1;
1631 char *iov_base = uiop->uio_iov->iov_base;
1633 iomode = NFSWRITE_FILESYNC;
1634 uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1636 * When doing direct I/O we do not care if the
1637 * server's write verifier has changed, but we
1638 * do not want to update the verifier if it has
1639 * changed, since that hides the change from
1640 * writes being done through the buffer cache.
1641 * By passing must_commit in set to two, the code
1642 * in nfsrpc_writerpc() will not update the
1643 * verifier on the mount point.
1646 ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0, 0);
1647 KASSERT((must_commit == 2), ("ncl_doio_directwrite: Updated write"
1649 if (iomode != NFSWRITE_FILESYNC)
1650 printf("ncl_doio_directwrite: Broken server "
1651 "did not reply FILE_SYNC\n");
1652 free(iov_base, M_NFSDIRECTIO);
1653 free(uiop->uio_iov, M_NFSDIRECTIO);
1654 free(uiop, M_NFSDIRECTIO);
1655 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1656 struct nfsnode *np = VTONFS(bp->b_vp);
1658 if (NFSHASPNFS(VFSTONFS(bp->b_vp->v_mount))) {
1660 * Invalidate the attribute cache, since writes to a DS
1661 * won't update the size attribute.
1663 np->n_attrstamp = 0;
1665 np->n_directio_asyncwr--;
1666 if (np->n_directio_asyncwr == 0) {
1667 np->n_flag &= ~NMODIFIED;
1668 if ((np->n_flag & NFSYNCWAIT)) {
1669 np->n_flag &= ~NFSYNCWAIT;
1670 wakeup((caddr_t)&np->n_directio_asyncwr);
1676 uma_zfree(ncl_pbuf_zone, bp);
1680 * Do an I/O operation to/from a cache block. This may be called
1681 * synchronously or from an nfsiod.
1684 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1685 int called_from_strategy)
1689 struct nfsmount *nmp;
1690 int error = 0, iomode, must_commit = 0;
1693 struct proc *p = td ? td->td_proc : NULL;
1697 nmp = VFSTONFS(vp->v_mount);
1699 uiop->uio_iov = &io;
1700 uiop->uio_iovcnt = 1;
1701 uiop->uio_segflg = UIO_SYSSPACE;
1705 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We
1706 * do this here so we do not have to do it in all the code that
1709 bp->b_flags &= ~B_INVAL;
1710 bp->b_ioflags &= ~BIO_ERROR;
1712 KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1713 iocmd = bp->b_iocmd;
1714 if (iocmd == BIO_READ) {
1715 io.iov_len = uiop->uio_resid = bp->b_bcount;
1716 io.iov_base = bp->b_data;
1717 uiop->uio_rw = UIO_READ;
1719 switch (vp->v_type) {
1721 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1722 NFSINCRGLOBAL(nfsstatsv1.read_bios);
1723 error = ncl_readrpc(vp, uiop, cr);
1726 if (uiop->uio_resid) {
1728 * If we had a short read with no error, we must have
1729 * hit a file hole. We should zero-fill the remainder.
1730 * This can also occur if the server hits the file EOF.
1732 * Holes used to be able to occur due to pending
1733 * writes, but that is not possible any longer.
1735 int nread = bp->b_bcount - uiop->uio_resid;
1736 ssize_t left = uiop->uio_resid;
1739 bzero((char *)bp->b_data + nread, left);
1740 uiop->uio_resid = 0;
1743 /* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1744 if (p && vp->v_writecount <= -1) {
1746 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1749 killproc(p, "text file modification");
1756 uiop->uio_offset = (off_t)0;
1757 NFSINCRGLOBAL(nfsstatsv1.readlink_bios);
1758 error = ncl_readlinkrpc(vp, uiop, cr);
1761 NFSINCRGLOBAL(nfsstatsv1.readdir_bios);
1762 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1763 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1764 error = ncl_readdirplusrpc(vp, uiop, cr, td);
1765 if (error == NFSERR_NOTSUPP)
1766 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1768 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1769 error = ncl_readdirrpc(vp, uiop, cr, td);
1771 * end-of-directory sets B_INVAL but does not generate an
1774 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1775 bp->b_flags |= B_INVAL;
1778 printf("ncl_doio: type %x unexpected\n", vp->v_type);
1782 bp->b_ioflags |= BIO_ERROR;
1783 bp->b_error = error;
1787 * If we only need to commit, try to commit
1789 if (bp->b_flags & B_NEEDCOMMIT) {
1793 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1794 retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1796 if (NFSCL_FORCEDISM(vp->v_mount) || retv == 0) {
1797 bp->b_dirtyoff = bp->b_dirtyend = 0;
1798 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1803 if (retv == NFSERR_STALEWRITEVERF) {
1804 ncl_clearcommit(vp->v_mount);
1809 * Setup for actual write
1812 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1813 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1816 if (bp->b_dirtyend > bp->b_dirtyoff) {
1817 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1819 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1821 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1822 uiop->uio_rw = UIO_WRITE;
1823 NFSINCRGLOBAL(nfsstatsv1.write_bios);
1825 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1826 iomode = NFSWRITE_UNSTABLE;
1828 iomode = NFSWRITE_FILESYNC;
1830 error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1831 called_from_strategy, 0);
1834 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1835 * to cluster the buffers needing commit. This will allow
1836 * the system to submit a single commit rpc for the whole
1837 * cluster. We can do this even if the buffer is not 100%
1838 * dirty (relative to the NFS blocksize), so we optimize the
1839 * append-to-file-case.
1841 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1842 * cleared because write clustering only works for commit
1843 * rpc's, not for the data portion of the write).
1846 if (!error && iomode == NFSWRITE_UNSTABLE) {
1847 bp->b_flags |= B_NEEDCOMMIT;
1848 if (bp->b_dirtyoff == 0
1849 && bp->b_dirtyend == bp->b_bcount)
1850 bp->b_flags |= B_CLUSTEROK;
1852 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1856 * For an interrupted write, the buffer is still valid
1857 * and the write hasn't been pushed to the server yet,
1858 * so we can't set BIO_ERROR and report the interruption
1859 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1860 * is not relevant, so the rpc attempt is essentially
1861 * a noop. For the case of a V3 write rpc not being
1862 * committed to stable storage, the block is still
1863 * dirty and requires either a commit rpc or another
1864 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1865 * the block is reused. This is indicated by setting
1866 * the B_DELWRI and B_NEEDCOMMIT flags.
1868 * EIO is returned by ncl_writerpc() to indicate a recoverable
1869 * write error and is handled as above, except that
1870 * B_EINTR isn't set. One cause of this is a stale stateid
1871 * error for the RPC that indicates recovery is required,
1872 * when called with called_from_strategy != 0.
1874 * If the buffer is marked B_PAGING, it does not reside on
1875 * the vp's paging queues so we cannot call bdirty(). The
1876 * bp in this case is not an NFS cache block so we should
1879 * The logic below breaks up errors into recoverable and
1880 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1881 * and keep the buffer around for potential write retries.
1882 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1883 * and save the error in the nfsnode. This is less than ideal
1884 * but necessary. Keeping such buffers around could potentially
1885 * cause buffer exhaustion eventually (they can never be written
1886 * out, so will get constantly be re-dirtied). It also causes
1887 * all sorts of vfs panics. For non-recoverable write errors,
1888 * also invalidate the attrcache, so we'll be forced to go over
1889 * the wire for this object, returning an error to user on next
1890 * call (most of the time).
1892 if (error == EINTR || error == EIO || error == ETIMEDOUT
1893 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1894 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1895 if ((bp->b_flags & B_PAGING) == 0) {
1897 bp->b_flags &= ~B_DONE;
1899 if ((error == EINTR || error == ETIMEDOUT) &&
1900 (bp->b_flags & B_ASYNC) == 0)
1901 bp->b_flags |= B_EINTR;
1904 bp->b_ioflags |= BIO_ERROR;
1905 bp->b_flags |= B_INVAL;
1906 bp->b_error = np->n_error = error;
1908 np->n_flag |= NWRITEERR;
1909 np->n_attrstamp = 0;
1910 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1913 bp->b_dirtyoff = bp->b_dirtyend = 0;
1921 bp->b_resid = uiop->uio_resid;
1922 if (must_commit == 1)
1923 ncl_clearcommit(vp->v_mount);
1929 * Used to aid in handling ftruncate() operations on the NFS client side.
1930 * Truncation creates a number of special problems for NFS. We have to
1931 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1932 * we have to properly handle VM pages or (potentially dirty) buffers
1933 * that straddle the truncation point.
1937 ncl_meta_setsize(struct vnode *vp, struct thread *td, u_quad_t nsize)
1939 struct nfsnode *np = VTONFS(vp);
1941 int biosize = vp->v_bufobj.bo_bsize;
1949 if (nsize < tsize) {
1955 * vtruncbuf() doesn't get the buffer overlapping the
1956 * truncation point. We may have a B_DELWRI and/or B_CACHE
1957 * buffer that now needs to be truncated.
1959 error = vtruncbuf(vp, nsize, biosize);
1960 lbn = nsize / biosize;
1961 bufsize = nsize - (lbn * biosize);
1962 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1965 if (bp->b_dirtyoff > bp->b_bcount)
1966 bp->b_dirtyoff = bp->b_bcount;
1967 if (bp->b_dirtyend > bp->b_bcount)
1968 bp->b_dirtyend = bp->b_bcount;
1969 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1972 vnode_pager_setsize(vp, nsize);