2 * Copyright (c) 1989, 1993
3 * The Regents of the University of California. All rights reserved.
5 * This code is derived from software contributed to Berkeley by
6 * Rick Macklem at The University of Guelph.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 3. Neither the name of the University nor the names of its contributors
17 * may be used to endorse or promote products derived from this software
18 * without specific prior written permission.
20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * @(#)nfs_bio.c 8.9 (Berkeley) 3/30/95
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
38 #include <sys/param.h>
39 #include <sys/systm.h>
42 #include <sys/kernel.h>
43 #include <sys/mount.h>
44 #include <sys/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 int ncl_pbuf_freecnt = -1; /* start out unlimited */
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)
102 mtx_lock(&np->n_mtx);
104 mtx_unlock(&np->n_mtx);
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;
116 ncl_getpages(struct vop_getpages_args *ap)
118 int i, error, nextoff, size, toff, count, npages;
126 struct nfsmount *nmp;
134 cred = curthread->td_ucred;
135 nmp = VFSTONFS(vp->v_mount);
137 npages = ap->a_count;
139 if ((object = vp->v_object) == NULL) {
140 printf("ncl_getpages: called with non-merged cache vnode\n");
141 return (VM_PAGER_ERROR);
144 if (newnfs_directio_enable && !newnfs_directio_allow_mmap) {
145 mtx_lock(&np->n_mtx);
146 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
147 mtx_unlock(&np->n_mtx);
148 printf("ncl_getpages: called on non-cacheable vnode\n");
149 return (VM_PAGER_ERROR);
151 mtx_unlock(&np->n_mtx);
154 mtx_lock(&nmp->nm_mtx);
155 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
156 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
157 mtx_unlock(&nmp->nm_mtx);
158 /* We'll never get here for v4, because we always have fsinfo */
159 (void)ncl_fsinfo(nmp, vp, cred, td);
161 mtx_unlock(&nmp->nm_mtx);
164 return (vfs_bio_getpages(vp, pages, npages, ap->a_rbehind,
165 ap->a_rahead, ncl_gbp_getblkno, ncl_gbp_getblksz));
168 * If the requested page is partially valid, just return it and
169 * allow the pager to zero-out the blanks. Partially valid pages
170 * can only occur at the file EOF.
172 * XXXGL: is that true for NFS, where short read can occur???
174 VM_OBJECT_WLOCK(object);
175 if (pages[npages - 1]->valid != 0 && --npages == 0)
177 VM_OBJECT_WUNLOCK(object);
180 * We use only the kva address for the buffer, but this is extremely
181 * convenient and fast.
183 bp = getpbuf(&ncl_pbuf_freecnt);
185 kva = (vm_offset_t) bp->b_data;
186 pmap_qenter(kva, pages, npages);
187 VM_CNT_INC(v_vnodein);
188 VM_CNT_ADD(v_vnodepgsin, npages);
190 count = npages << PAGE_SHIFT;
191 iov.iov_base = (caddr_t) kva;
195 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
196 uio.uio_resid = count;
197 uio.uio_segflg = UIO_SYSSPACE;
198 uio.uio_rw = UIO_READ;
201 error = ncl_readrpc(vp, &uio, cred);
202 pmap_qremove(kva, npages);
204 relpbuf(bp, &ncl_pbuf_freecnt);
206 if (error && (uio.uio_resid == count)) {
207 printf("ncl_getpages: error %d\n", error);
208 return (VM_PAGER_ERROR);
212 * Calculate the number of bytes read and validate only that number
213 * of bytes. Note that due to pending writes, size may be 0. This
214 * does not mean that the remaining data is invalid!
217 size = count - uio.uio_resid;
218 VM_OBJECT_WLOCK(object);
219 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
221 nextoff = toff + PAGE_SIZE;
224 if (nextoff <= size) {
226 * Read operation filled an entire page
228 m->valid = VM_PAGE_BITS_ALL;
229 KASSERT(m->dirty == 0,
230 ("nfs_getpages: page %p is dirty", m));
231 } else if (size > toff) {
233 * Read operation filled a partial page.
236 vm_page_set_valid_range(m, 0, size - toff);
237 KASSERT(m->dirty == 0,
238 ("nfs_getpages: page %p is dirty", m));
241 * Read operation was short. If no error
242 * occurred we may have hit a zero-fill
243 * section. We leave valid set to 0, and page
244 * is freed by vm_page_readahead_finish() if
245 * its index is not equal to requested, or
246 * page is zeroed and set valid by
247 * vm_pager_get_pages() for requested page.
253 VM_OBJECT_WUNLOCK(object);
258 return (VM_PAGER_OK);
262 * Vnode op for VM putpages.
265 ncl_putpages(struct vop_putpages_args *ap)
269 int i, error, npages, count;
275 struct nfsmount *nmp;
281 td = curthread; /* XXX */
282 /* Set the cred to n_writecred for the write rpcs. */
283 if (np->n_writecred != NULL)
284 cred = crhold(np->n_writecred);
286 cred = crhold(curthread->td_ucred); /* XXX */
287 nmp = VFSTONFS(vp->v_mount);
290 rtvals = ap->a_rtvals;
291 npages = btoc(count);
292 offset = IDX_TO_OFF(pages[0]->pindex);
294 mtx_lock(&nmp->nm_mtx);
295 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
296 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
297 mtx_unlock(&nmp->nm_mtx);
298 (void)ncl_fsinfo(nmp, vp, cred, td);
300 mtx_unlock(&nmp->nm_mtx);
302 mtx_lock(&np->n_mtx);
303 if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
304 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
305 mtx_unlock(&np->n_mtx);
306 printf("ncl_putpages: called on noncache-able vnode\n");
307 mtx_lock(&np->n_mtx);
310 * When putting pages, do not extend file past EOF.
312 if (offset + count > np->n_size) {
313 count = np->n_size - offset;
317 mtx_unlock(&np->n_mtx);
319 for (i = 0; i < npages; i++)
320 rtvals[i] = VM_PAGER_ERROR;
322 VM_CNT_INC(v_vnodeout);
323 VM_CNT_ADD(v_vnodepgsout, count);
325 iov.iov_base = unmapped_buf;
329 uio.uio_offset = offset;
330 uio.uio_resid = count;
331 uio.uio_segflg = UIO_NOCOPY;
332 uio.uio_rw = UIO_WRITE;
335 error = VOP_WRITE(vp, &uio, vnode_pager_putpages_ioflags(ap->a_sync),
339 if (error == 0 || !nfs_keep_dirty_on_error) {
340 vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid,
341 np->n_size - offset, npages * PAGE_SIZE);
347 * For nfs, cache consistency can only be maintained approximately.
348 * Although RFC1094 does not specify the criteria, the following is
349 * believed to be compatible with the reference port.
351 * If the file's modify time on the server has changed since the
352 * last read rpc or you have written to the file,
353 * you may have lost data cache consistency with the
354 * server, so flush all of the file's data out of the cache.
355 * Then force a getattr rpc to ensure that you have up to date
357 * NB: This implies that cache data can be read when up to
358 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
359 * attributes this could be forced by setting n_attrstamp to 0 before
360 * the VOP_GETATTR() call.
363 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
367 struct nfsnode *np = VTONFS(vp);
371 * Grab the exclusive lock before checking whether the cache is
373 * XXX - We can make this cheaper later (by acquiring cheaper locks).
374 * But for now, this suffices.
376 old_lock = ncl_upgrade_vnlock(vp);
377 if (vp->v_iflag & VI_DOOMED) {
382 mtx_lock(&np->n_mtx);
383 if (np->n_flag & NMODIFIED) {
384 mtx_unlock(&np->n_mtx);
385 if (vp->v_type != VREG) {
386 if (vp->v_type != VDIR)
387 panic("nfs: bioread, not dir");
389 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
390 if (error == 0 && (vp->v_iflag & VI_DOOMED) != 0)
396 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
397 error = VOP_GETATTR(vp, &vattr, cred);
400 mtx_lock(&np->n_mtx);
401 np->n_mtime = vattr.va_mtime;
402 mtx_unlock(&np->n_mtx);
404 mtx_unlock(&np->n_mtx);
405 error = VOP_GETATTR(vp, &vattr, cred);
408 mtx_lock(&np->n_mtx);
409 if ((np->n_flag & NSIZECHANGED)
410 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
411 mtx_unlock(&np->n_mtx);
412 if (vp->v_type == VDIR)
414 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
415 if (error == 0 && (vp->v_iflag & VI_DOOMED) != 0)
419 mtx_lock(&np->n_mtx);
420 np->n_mtime = vattr.va_mtime;
421 np->n_flag &= ~NSIZECHANGED;
423 mtx_unlock(&np->n_mtx);
426 ncl_downgrade_vnlock(vp, old_lock);
431 * Vnode op for read using bio
434 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
436 struct nfsnode *np = VTONFS(vp);
438 struct buf *bp, *rabp;
440 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
444 int nra, error = 0, n = 0, on = 0;
447 KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
448 if (uio->uio_resid == 0)
450 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
454 mtx_lock(&nmp->nm_mtx);
455 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
456 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
457 mtx_unlock(&nmp->nm_mtx);
458 (void)ncl_fsinfo(nmp, vp, cred, td);
459 mtx_lock(&nmp->nm_mtx);
461 if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
462 (void) newnfs_iosize(nmp);
464 tmp_off = uio->uio_offset + uio->uio_resid;
465 if (vp->v_type != VDIR &&
466 (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) {
467 mtx_unlock(&nmp->nm_mtx);
470 mtx_unlock(&nmp->nm_mtx);
472 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
473 /* No caching/ no readaheads. Just read data into the user buffer */
474 return ncl_readrpc(vp, uio, cred);
476 biosize = vp->v_bufobj.bo_bsize;
477 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
479 error = nfs_bioread_check_cons(vp, td, cred);
486 mtx_lock(&np->n_mtx);
488 mtx_unlock(&np->n_mtx);
490 switch (vp->v_type) {
492 NFSINCRGLOBAL(nfsstatsv1.biocache_reads);
493 lbn = uio->uio_offset / biosize;
494 on = uio->uio_offset - (lbn * biosize);
497 * Start the read ahead(s), as required.
499 if (nmp->nm_readahead > 0) {
500 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
501 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
502 rabn = lbn + 1 + nra;
503 if (incore(&vp->v_bufobj, rabn) == NULL) {
504 rabp = nfs_getcacheblk(vp, rabn, biosize, td);
506 error = newnfs_sigintr(nmp, td);
507 return (error ? error : EINTR);
509 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
510 rabp->b_flags |= B_ASYNC;
511 rabp->b_iocmd = BIO_READ;
512 vfs_busy_pages(rabp, 0);
513 if (ncl_asyncio(nmp, rabp, cred, td)) {
514 rabp->b_flags |= B_INVAL;
515 rabp->b_ioflags |= BIO_ERROR;
516 vfs_unbusy_pages(rabp);
527 /* Note that bcount is *not* DEV_BSIZE aligned. */
529 if ((off_t)lbn * biosize >= nsize) {
531 } else if ((off_t)(lbn + 1) * biosize > nsize) {
532 bcount = nsize - (off_t)lbn * biosize;
534 bp = nfs_getcacheblk(vp, lbn, bcount, td);
537 error = newnfs_sigintr(nmp, td);
538 return (error ? error : EINTR);
542 * If B_CACHE is not set, we must issue the read. If this
543 * fails, we return an error.
546 if ((bp->b_flags & B_CACHE) == 0) {
547 bp->b_iocmd = BIO_READ;
548 vfs_busy_pages(bp, 0);
549 error = ncl_doio(vp, bp, cred, td, 0);
557 * on is the offset into the current bp. Figure out how many
558 * bytes we can copy out of the bp. Note that bcount is
559 * NOT DEV_BSIZE aligned.
561 * Then figure out how many bytes we can copy into the uio.
566 n = MIN((unsigned)(bcount - on), uio->uio_resid);
569 NFSINCRGLOBAL(nfsstatsv1.biocache_readlinks);
570 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
572 error = newnfs_sigintr(nmp, td);
573 return (error ? error : EINTR);
575 if ((bp->b_flags & B_CACHE) == 0) {
576 bp->b_iocmd = BIO_READ;
577 vfs_busy_pages(bp, 0);
578 error = ncl_doio(vp, bp, cred, td, 0);
580 bp->b_ioflags |= BIO_ERROR;
585 n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
589 NFSINCRGLOBAL(nfsstatsv1.biocache_readdirs);
590 if (np->n_direofoffset
591 && uio->uio_offset >= np->n_direofoffset) {
594 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
595 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
596 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
598 error = newnfs_sigintr(nmp, td);
599 return (error ? error : EINTR);
601 if ((bp->b_flags & B_CACHE) == 0) {
602 bp->b_iocmd = BIO_READ;
603 vfs_busy_pages(bp, 0);
604 error = ncl_doio(vp, bp, cred, td, 0);
608 while (error == NFSERR_BAD_COOKIE) {
610 error = ncl_vinvalbuf(vp, 0, td, 1);
611 if (error == 0 && (vp->v_iflag & VI_DOOMED) != 0)
615 * Yuck! The directory has been modified on the
616 * server. The only way to get the block is by
617 * reading from the beginning to get all the
620 * Leave the last bp intact unless there is an error.
621 * Loop back up to the while if the error is another
622 * NFSERR_BAD_COOKIE (double yuch!).
624 for (i = 0; i <= lbn && !error; i++) {
625 if (np->n_direofoffset
626 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
628 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
630 error = newnfs_sigintr(nmp, td);
631 return (error ? error : EINTR);
633 if ((bp->b_flags & B_CACHE) == 0) {
634 bp->b_iocmd = BIO_READ;
635 vfs_busy_pages(bp, 0);
636 error = ncl_doio(vp, bp, cred, td, 0);
638 * no error + B_INVAL == directory EOF,
641 if (error == 0 && (bp->b_flags & B_INVAL))
645 * An error will throw away the block and the
646 * for loop will break out. If no error and this
647 * is not the block we want, we throw away the
648 * block and go for the next one via the for loop.
650 if (error || i < lbn)
655 * The above while is repeated if we hit another cookie
656 * error. If we hit an error and it wasn't a cookie error,
664 * If not eof and read aheads are enabled, start one.
665 * (You need the current block first, so that you have the
666 * directory offset cookie of the next block.)
668 if (nmp->nm_readahead > 0 &&
669 (bp->b_flags & B_INVAL) == 0 &&
670 (np->n_direofoffset == 0 ||
671 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
672 incore(&vp->v_bufobj, lbn + 1) == NULL) {
673 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
675 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
676 rabp->b_flags |= B_ASYNC;
677 rabp->b_iocmd = BIO_READ;
678 vfs_busy_pages(rabp, 0);
679 if (ncl_asyncio(nmp, rabp, cred, td)) {
680 rabp->b_flags |= B_INVAL;
681 rabp->b_ioflags |= BIO_ERROR;
682 vfs_unbusy_pages(rabp);
691 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
692 * chopped for the EOF condition, we cannot tell how large
693 * NFS directories are going to be until we hit EOF. So
694 * an NFS directory buffer is *not* chopped to its EOF. Now,
695 * it just so happens that b_resid will effectively chop it
696 * to EOF. *BUT* this information is lost if the buffer goes
697 * away and is reconstituted into a B_CACHE state ( due to
698 * being VMIO ) later. So we keep track of the directory eof
699 * in np->n_direofoffset and chop it off as an extra step
702 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
703 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
704 n = np->n_direofoffset - uio->uio_offset;
707 printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
713 error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio);
715 if (vp->v_type == VLNK)
719 } while (error == 0 && uio->uio_resid > 0 && n > 0);
724 * The NFS write path cannot handle iovecs with len > 1. So we need to
725 * break up iovecs accordingly (restricting them to wsize).
726 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
727 * For the ASYNC case, 2 copies are needed. The first a copy from the
728 * user buffer to a staging buffer and then a second copy from the staging
729 * buffer to mbufs. This can be optimized by copying from the user buffer
730 * directly into mbufs and passing the chain down, but that requires a
731 * fair amount of re-working of the relevant codepaths (and can be done
735 nfs_directio_write(vp, uiop, cred, ioflag)
742 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
743 struct thread *td = uiop->uio_td;
747 mtx_lock(&nmp->nm_mtx);
748 wsize = nmp->nm_wsize;
749 mtx_unlock(&nmp->nm_mtx);
750 if (ioflag & IO_SYNC) {
751 int iomode, must_commit;
755 while (uiop->uio_resid > 0) {
756 size = MIN(uiop->uio_resid, wsize);
757 size = MIN(uiop->uio_iov->iov_len, size);
758 iov.iov_base = uiop->uio_iov->iov_base;
762 uio.uio_offset = uiop->uio_offset;
763 uio.uio_resid = size;
764 uio.uio_segflg = UIO_USERSPACE;
765 uio.uio_rw = UIO_WRITE;
767 iomode = NFSWRITE_FILESYNC;
768 error = ncl_writerpc(vp, &uio, cred, &iomode,
770 KASSERT((must_commit == 0),
771 ("ncl_directio_write: Did not commit write"));
774 uiop->uio_offset += size;
775 uiop->uio_resid -= size;
776 if (uiop->uio_iov->iov_len <= size) {
780 uiop->uio_iov->iov_base =
781 (char *)uiop->uio_iov->iov_base + size;
782 uiop->uio_iov->iov_len -= size;
791 * Break up the write into blocksize chunks and hand these
792 * over to nfsiod's for write back.
793 * Unfortunately, this incurs a copy of the data. Since
794 * the user could modify the buffer before the write is
797 * The obvious optimization here is that one of the 2 copies
798 * in the async write path can be eliminated by copying the
799 * data here directly into mbufs and passing the mbuf chain
800 * down. But that will require a fair amount of re-working
801 * of the code and can be done if there's enough interest
802 * in NFS directio access.
804 while (uiop->uio_resid > 0) {
805 size = MIN(uiop->uio_resid, wsize);
806 size = MIN(uiop->uio_iov->iov_len, size);
807 bp = getpbuf(&ncl_pbuf_freecnt);
808 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
809 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
810 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
811 t_iov->iov_len = size;
812 t_uio->uio_iov = t_iov;
813 t_uio->uio_iovcnt = 1;
814 t_uio->uio_offset = uiop->uio_offset;
815 t_uio->uio_resid = size;
816 t_uio->uio_segflg = UIO_SYSSPACE;
817 t_uio->uio_rw = UIO_WRITE;
819 KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
820 uiop->uio_segflg == UIO_SYSSPACE,
821 ("nfs_directio_write: Bad uio_segflg"));
822 if (uiop->uio_segflg == UIO_USERSPACE) {
823 error = copyin(uiop->uio_iov->iov_base,
824 t_iov->iov_base, size);
829 * UIO_SYSSPACE may never happen, but handle
830 * it just in case it does.
832 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
834 bp->b_flags |= B_DIRECT;
835 bp->b_iocmd = BIO_WRITE;
836 if (cred != NOCRED) {
840 bp->b_wcred = NOCRED;
841 bp->b_caller1 = (void *)t_uio;
843 error = ncl_asyncio(nmp, bp, NOCRED, td);
846 free(t_iov->iov_base, M_NFSDIRECTIO);
847 free(t_iov, M_NFSDIRECTIO);
848 free(t_uio, M_NFSDIRECTIO);
850 relpbuf(bp, &ncl_pbuf_freecnt);
855 uiop->uio_offset += size;
856 uiop->uio_resid -= size;
857 if (uiop->uio_iov->iov_len <= size) {
861 uiop->uio_iov->iov_base =
862 (char *)uiop->uio_iov->iov_base + size;
863 uiop->uio_iov->iov_len -= size;
871 * Vnode op for write using bio
874 ncl_write(struct vop_write_args *ap)
877 struct uio *uio = ap->a_uio;
878 struct thread *td = uio->uio_td;
879 struct vnode *vp = ap->a_vp;
880 struct nfsnode *np = VTONFS(vp);
881 struct ucred *cred = ap->a_cred;
882 int ioflag = ap->a_ioflag;
885 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
887 int bcount, noncontig_write, obcount;
888 int bp_cached, n, on, error = 0, error1, wouldcommit;
889 size_t orig_resid, local_resid;
890 off_t orig_size, tmp_off;
892 KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
893 KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
895 if (vp->v_type != VREG)
897 mtx_lock(&np->n_mtx);
898 if (np->n_flag & NWRITEERR) {
899 np->n_flag &= ~NWRITEERR;
900 mtx_unlock(&np->n_mtx);
901 return (np->n_error);
903 mtx_unlock(&np->n_mtx);
904 mtx_lock(&nmp->nm_mtx);
905 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
906 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
907 mtx_unlock(&nmp->nm_mtx);
908 (void)ncl_fsinfo(nmp, vp, cred, td);
909 mtx_lock(&nmp->nm_mtx);
911 if (nmp->nm_wsize == 0)
912 (void) newnfs_iosize(nmp);
913 mtx_unlock(&nmp->nm_mtx);
916 * Synchronously flush pending buffers if we are in synchronous
917 * mode or if we are appending.
919 if (ioflag & (IO_APPEND | IO_SYNC)) {
920 mtx_lock(&np->n_mtx);
921 if (np->n_flag & NMODIFIED) {
922 mtx_unlock(&np->n_mtx);
923 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
925 * Require non-blocking, synchronous writes to
926 * dirty files to inform the program it needs
927 * to fsync(2) explicitly.
929 if (ioflag & IO_NDELAY)
933 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
934 error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
935 IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
936 if (error == 0 && (vp->v_iflag & VI_DOOMED) != 0)
941 mtx_unlock(&np->n_mtx);
944 orig_resid = uio->uio_resid;
945 mtx_lock(&np->n_mtx);
946 orig_size = np->n_size;
947 mtx_unlock(&np->n_mtx);
950 * If IO_APPEND then load uio_offset. We restart here if we cannot
951 * get the append lock.
953 if (ioflag & IO_APPEND) {
955 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
956 error = VOP_GETATTR(vp, &vattr, cred);
959 mtx_lock(&np->n_mtx);
960 uio->uio_offset = np->n_size;
961 mtx_unlock(&np->n_mtx);
964 if (uio->uio_offset < 0)
966 tmp_off = uio->uio_offset + uio->uio_resid;
967 if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
969 if (uio->uio_resid == 0)
972 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
973 return nfs_directio_write(vp, uio, cred, ioflag);
976 * Maybe this should be above the vnode op call, but so long as
977 * file servers have no limits, i don't think it matters
979 if (vn_rlimit_fsize(vp, uio, td))
982 biosize = vp->v_bufobj.bo_bsize;
984 * Find all of this file's B_NEEDCOMMIT buffers. If our writes
985 * would exceed the local maximum per-file write commit size when
986 * combined with those, we must decide whether to flush,
987 * go synchronous, or return error. We don't bother checking
988 * IO_UNIT -- we just make all writes atomic anyway, as there's
989 * no point optimizing for something that really won't ever happen.
992 if (!(ioflag & IO_SYNC)) {
995 mtx_lock(&np->n_mtx);
997 mtx_unlock(&np->n_mtx);
998 if (nflag & NMODIFIED) {
999 BO_LOCK(&vp->v_bufobj);
1000 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
1001 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
1003 if (bp->b_flags & B_NEEDCOMMIT)
1004 wouldcommit += bp->b_bcount;
1007 BO_UNLOCK(&vp->v_bufobj);
1012 if (!(ioflag & IO_SYNC)) {
1013 wouldcommit += biosize;
1014 if (wouldcommit > nmp->nm_wcommitsize) {
1015 np->n_attrstamp = 0;
1016 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1017 error = ncl_vinvalbuf(vp, V_SAVE | ((ioflag &
1018 IO_VMIO) != 0 ? V_VMIO : 0), td, 1);
1020 (vp->v_iflag & VI_DOOMED) != 0)
1024 wouldcommit = biosize;
1028 NFSINCRGLOBAL(nfsstatsv1.biocache_writes);
1029 lbn = uio->uio_offset / biosize;
1030 on = uio->uio_offset - (lbn * biosize);
1031 n = MIN((unsigned)(biosize - on), uio->uio_resid);
1034 * Handle direct append and file extension cases, calculate
1035 * unaligned buffer size.
1037 mtx_lock(&np->n_mtx);
1038 if ((np->n_flag & NHASBEENLOCKED) == 0 &&
1039 (nmp->nm_flag & NFSMNT_NONCONTIGWR) != 0)
1040 noncontig_write = 1;
1042 noncontig_write = 0;
1043 if ((uio->uio_offset == np->n_size ||
1044 (noncontig_write != 0 &&
1045 lbn == (np->n_size / biosize) &&
1046 uio->uio_offset + n > np->n_size)) && n) {
1047 mtx_unlock(&np->n_mtx);
1049 * Get the buffer (in its pre-append state to maintain
1050 * B_CACHE if it was previously set). Resize the
1051 * nfsnode after we have locked the buffer to prevent
1052 * readers from reading garbage.
1054 obcount = np->n_size - (lbn * biosize);
1055 bp = nfs_getcacheblk(vp, lbn, obcount, td);
1060 mtx_lock(&np->n_mtx);
1061 np->n_size = uio->uio_offset + n;
1062 np->n_flag |= NMODIFIED;
1063 vnode_pager_setsize(vp, np->n_size);
1064 mtx_unlock(&np->n_mtx);
1066 save = bp->b_flags & B_CACHE;
1068 allocbuf(bp, bcount);
1069 bp->b_flags |= save;
1070 if (noncontig_write != 0 && on > obcount)
1071 vfs_bio_bzero_buf(bp, obcount, on -
1076 * Obtain the locked cache block first, and then
1077 * adjust the file's size as appropriate.
1080 if ((off_t)lbn * biosize + bcount < np->n_size) {
1081 if ((off_t)(lbn + 1) * biosize < np->n_size)
1084 bcount = np->n_size - (off_t)lbn * biosize;
1086 mtx_unlock(&np->n_mtx);
1087 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1088 mtx_lock(&np->n_mtx);
1089 if (uio->uio_offset + n > np->n_size) {
1090 np->n_size = uio->uio_offset + n;
1091 np->n_flag |= NMODIFIED;
1092 vnode_pager_setsize(vp, np->n_size);
1094 mtx_unlock(&np->n_mtx);
1098 error = newnfs_sigintr(nmp, td);
1105 * Issue a READ if B_CACHE is not set. In special-append
1106 * mode, B_CACHE is based on the buffer prior to the write
1107 * op and is typically set, avoiding the read. If a read
1108 * is required in special append mode, the server will
1109 * probably send us a short-read since we extended the file
1110 * on our end, resulting in b_resid == 0 and, thusly,
1111 * B_CACHE getting set.
1113 * We can also avoid issuing the read if the write covers
1114 * the entire buffer. We have to make sure the buffer state
1115 * is reasonable in this case since we will not be initiating
1116 * I/O. See the comments in kern/vfs_bio.c's getblk() for
1119 * B_CACHE may also be set due to the buffer being cached
1124 if (on == 0 && n == bcount) {
1125 if ((bp->b_flags & B_CACHE) == 0)
1127 bp->b_flags |= B_CACHE;
1128 bp->b_flags &= ~B_INVAL;
1129 bp->b_ioflags &= ~BIO_ERROR;
1132 if ((bp->b_flags & B_CACHE) == 0) {
1133 bp->b_iocmd = BIO_READ;
1134 vfs_busy_pages(bp, 0);
1135 error = ncl_doio(vp, bp, cred, td, 0);
1141 if (bp->b_wcred == NOCRED)
1142 bp->b_wcred = crhold(cred);
1143 mtx_lock(&np->n_mtx);
1144 np->n_flag |= NMODIFIED;
1145 mtx_unlock(&np->n_mtx);
1148 * If dirtyend exceeds file size, chop it down. This should
1149 * not normally occur but there is an append race where it
1150 * might occur XXX, so we log it.
1152 * If the chopping creates a reverse-indexed or degenerate
1153 * situation with dirtyoff/end, we 0 both of them.
1156 if (bp->b_dirtyend > bcount) {
1157 printf("NFS append race @%lx:%d\n",
1158 (long)bp->b_blkno * DEV_BSIZE,
1159 bp->b_dirtyend - bcount);
1160 bp->b_dirtyend = bcount;
1163 if (bp->b_dirtyoff >= bp->b_dirtyend)
1164 bp->b_dirtyoff = bp->b_dirtyend = 0;
1167 * If the new write will leave a contiguous dirty
1168 * area, just update the b_dirtyoff and b_dirtyend,
1169 * otherwise force a write rpc of the old dirty area.
1171 * If there has been a file lock applied to this file
1172 * or vfs.nfs.old_noncontig_writing is set, do the following:
1173 * While it is possible to merge discontiguous writes due to
1174 * our having a B_CACHE buffer ( and thus valid read data
1175 * for the hole), we don't because it could lead to
1176 * significant cache coherency problems with multiple clients,
1177 * especially if locking is implemented later on.
1179 * If vfs.nfs.old_noncontig_writing is not set and there has
1180 * not been file locking done on this file:
1181 * Relax coherency a bit for the sake of performance and
1182 * expand the current dirty region to contain the new
1183 * write even if it means we mark some non-dirty data as
1187 if (noncontig_write == 0 && bp->b_dirtyend > 0 &&
1188 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1189 if (bwrite(bp) == EINTR) {
1196 local_resid = uio->uio_resid;
1197 error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio);
1199 if (error != 0 && !bp_cached) {
1201 * This block has no other content then what
1202 * possibly was written by the faulty uiomove.
1203 * Release it, forgetting the data pages, to
1204 * prevent the leak of uninitialized data to
1207 bp->b_ioflags |= BIO_ERROR;
1209 uio->uio_offset -= local_resid - uio->uio_resid;
1210 uio->uio_resid = local_resid;
1215 * Since this block is being modified, it must be written
1216 * again and not just committed. Since write clustering does
1217 * not work for the stage 1 data write, only the stage 2
1218 * commit rpc, we have to clear B_CLUSTEROK as well.
1220 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1223 * Get the partial update on the progress made from
1224 * uiomove, if an error occurred.
1227 n = local_resid - uio->uio_resid;
1230 * Only update dirtyoff/dirtyend if not a degenerate
1234 if (bp->b_dirtyend > 0) {
1235 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1236 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1238 bp->b_dirtyoff = on;
1239 bp->b_dirtyend = on + n;
1241 vfs_bio_set_valid(bp, on, n);
1245 * If IO_SYNC do bwrite().
1247 * IO_INVAL appears to be unused. The idea appears to be
1248 * to turn off caching in this case. Very odd. XXX
1250 if ((ioflag & IO_SYNC)) {
1251 if (ioflag & IO_INVAL)
1252 bp->b_flags |= B_NOCACHE;
1253 error1 = bwrite(bp);
1259 } else if ((n + on) == biosize || (ioflag & IO_ASYNC) != 0) {
1260 bp->b_flags |= B_ASYNC;
1261 (void) ncl_writebp(bp, 0, NULL);
1268 } while (uio->uio_resid > 0 && n > 0);
1271 if (ioflag & IO_UNIT) {
1273 vattr.va_size = orig_size;
1274 /* IO_SYNC is handled implicitely */
1275 (void)VOP_SETATTR(vp, &vattr, cred);
1276 uio->uio_offset -= orig_resid - uio->uio_resid;
1277 uio->uio_resid = orig_resid;
1285 * Get an nfs cache block.
1287 * Allocate a new one if the block isn't currently in the cache
1288 * and return the block marked busy. If the calling process is
1289 * interrupted by a signal for an interruptible mount point, return
1292 * The caller must carefully deal with the possible B_INVAL state of
1293 * the buffer. ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
1294 * indirectly), so synchronous reads can be issued without worrying about
1295 * the B_INVAL state. We have to be a little more careful when dealing
1296 * with writes (see comments in nfs_write()) when extending a file past
1300 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1304 struct nfsmount *nmp;
1309 if (nmp->nm_flag & NFSMNT_INT) {
1312 newnfs_set_sigmask(td, &oldset);
1313 bp = getblk(vp, bn, size, PCATCH, 0, 0);
1314 newnfs_restore_sigmask(td, &oldset);
1315 while (bp == NULL) {
1316 if (newnfs_sigintr(nmp, td))
1318 bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1321 bp = getblk(vp, bn, size, 0, 0, 0);
1324 if (vp->v_type == VREG)
1325 bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1330 * Flush and invalidate all dirty buffers. If another process is already
1331 * doing the flush, just wait for completion.
1334 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1336 struct nfsnode *np = VTONFS(vp);
1337 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1338 int error = 0, slpflag, slptimeo;
1341 ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
1343 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1345 if (NFSCL_FORCEDISM(nmp->nm_mountp))
1355 old_lock = ncl_upgrade_vnlock(vp);
1356 if (vp->v_iflag & VI_DOOMED) {
1358 * Since vgonel() uses the generic vinvalbuf() to flush
1359 * dirty buffers and it does not call this function, it
1360 * is safe to just return OK when VI_DOOMED is set.
1362 ncl_downgrade_vnlock(vp, old_lock);
1367 * Now, flush as required.
1369 if ((flags & (V_SAVE | V_VMIO)) == V_SAVE &&
1370 vp->v_bufobj.bo_object != NULL) {
1371 VM_OBJECT_WLOCK(vp->v_bufobj.bo_object);
1372 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1373 VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object);
1375 * If the page clean was interrupted, fail the invalidation.
1376 * Not doing so, we run the risk of losing dirty pages in the
1377 * vinvalbuf() call below.
1379 if (intrflg && (error = newnfs_sigintr(nmp, td)))
1383 error = vinvalbuf(vp, flags, slpflag, 0);
1385 if (intrflg && (error = newnfs_sigintr(nmp, td)))
1387 error = vinvalbuf(vp, flags, 0, slptimeo);
1389 if (NFSHASPNFS(nmp)) {
1390 nfscl_layoutcommit(vp, td);
1392 * Invalidate the attribute cache, since writes to a DS
1393 * won't update the size attribute.
1395 mtx_lock(&np->n_mtx);
1396 np->n_attrstamp = 0;
1398 mtx_lock(&np->n_mtx);
1399 if (np->n_directio_asyncwr == 0)
1400 np->n_flag &= ~NMODIFIED;
1401 mtx_unlock(&np->n_mtx);
1403 ncl_downgrade_vnlock(vp, old_lock);
1408 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1409 * This is mainly to avoid queueing async I/O requests when the nfsiods
1410 * are all hung on a dead server.
1412 * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1413 * is eventually dequeued by the async daemon, ncl_doio() *will*.
1416 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1425 * Commits are usually short and sweet so lets save some cpu and
1426 * leave the async daemons for more important rpc's (such as reads
1429 * Readdirplus RPCs do vget()s to acquire the vnodes for entries
1430 * in the directory in order to update attributes. This can deadlock
1431 * with another thread that is waiting for async I/O to be done by
1432 * an nfsiod thread while holding a lock on one of these vnodes.
1433 * To avoid this deadlock, don't allow the async nfsiod threads to
1434 * perform Readdirplus RPCs.
1436 mtx_lock(&ncl_iod_mutex);
1437 if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1438 (nmp->nm_bufqiods > ncl_numasync / 2)) ||
1439 (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
1440 mtx_unlock(&ncl_iod_mutex);
1444 if (nmp->nm_flag & NFSMNT_INT)
1449 * Find a free iod to process this request.
1451 for (iod = 0; iod < ncl_numasync; iod++)
1452 if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1458 * Try to create one if none are free.
1464 * Found one, so wake it up and tell it which
1467 NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1469 ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1470 ncl_iodmount[iod] = nmp;
1472 wakeup(&ncl_iodwant[iod]);
1476 * If none are free, we may already have an iod working on this mount
1477 * point. If so, it will process our request.
1480 if (nmp->nm_bufqiods > 0) {
1482 ("ncl_asyncio: %d iods are already processing mount %p\n",
1483 nmp->nm_bufqiods, nmp));
1489 * If we have an iod which can process the request, then queue
1494 * Ensure that the queue never grows too large. We still want
1495 * to asynchronize so we block rather then return EIO.
1497 while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1499 ("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1500 nmp->nm_bufqwant = TRUE;
1501 error = newnfs_msleep(td, &nmp->nm_bufq,
1502 &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1505 error2 = newnfs_sigintr(nmp, td);
1507 mtx_unlock(&ncl_iod_mutex);
1510 if (slpflag == PCATCH) {
1516 * We might have lost our iod while sleeping,
1517 * so check and loop if necessary.
1522 /* We might have lost our nfsiod */
1523 if (nmp->nm_bufqiods == 0) {
1525 ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1529 if (bp->b_iocmd == BIO_READ) {
1530 if (bp->b_rcred == NOCRED && cred != NOCRED)
1531 bp->b_rcred = crhold(cred);
1533 if (bp->b_wcred == NOCRED && cred != NOCRED)
1534 bp->b_wcred = crhold(cred);
1537 if (bp->b_flags & B_REMFREE)
1540 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1542 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1543 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx);
1544 VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1545 VTONFS(bp->b_vp)->n_directio_asyncwr++;
1546 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx);
1548 mtx_unlock(&ncl_iod_mutex);
1552 mtx_unlock(&ncl_iod_mutex);
1555 * All the iods are busy on other mounts, so return EIO to
1556 * force the caller to process the i/o synchronously.
1558 NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1563 ncl_doio_directwrite(struct buf *bp)
1565 int iomode, must_commit;
1566 struct uio *uiop = (struct uio *)bp->b_caller1;
1567 char *iov_base = uiop->uio_iov->iov_base;
1569 iomode = NFSWRITE_FILESYNC;
1570 uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1571 ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0);
1572 KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write"));
1573 free(iov_base, M_NFSDIRECTIO);
1574 free(uiop->uio_iov, M_NFSDIRECTIO);
1575 free(uiop, M_NFSDIRECTIO);
1576 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1577 struct nfsnode *np = VTONFS(bp->b_vp);
1578 mtx_lock(&np->n_mtx);
1579 if (NFSHASPNFS(VFSTONFS(vnode_mount(bp->b_vp)))) {
1581 * Invalidate the attribute cache, since writes to a DS
1582 * won't update the size attribute.
1584 np->n_attrstamp = 0;
1586 np->n_directio_asyncwr--;
1587 if (np->n_directio_asyncwr == 0) {
1588 np->n_flag &= ~NMODIFIED;
1589 if ((np->n_flag & NFSYNCWAIT)) {
1590 np->n_flag &= ~NFSYNCWAIT;
1591 wakeup((caddr_t)&np->n_directio_asyncwr);
1594 mtx_unlock(&np->n_mtx);
1597 relpbuf(bp, &ncl_pbuf_freecnt);
1601 * Do an I/O operation to/from a cache block. This may be called
1602 * synchronously or from an nfsiod.
1605 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1606 int called_from_strategy)
1610 struct nfsmount *nmp;
1611 int error = 0, iomode, must_commit = 0;
1614 struct proc *p = td ? td->td_proc : NULL;
1618 nmp = VFSTONFS(vp->v_mount);
1620 uiop->uio_iov = &io;
1621 uiop->uio_iovcnt = 1;
1622 uiop->uio_segflg = UIO_SYSSPACE;
1626 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We
1627 * do this here so we do not have to do it in all the code that
1630 bp->b_flags &= ~B_INVAL;
1631 bp->b_ioflags &= ~BIO_ERROR;
1633 KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1634 iocmd = bp->b_iocmd;
1635 if (iocmd == BIO_READ) {
1636 io.iov_len = uiop->uio_resid = bp->b_bcount;
1637 io.iov_base = bp->b_data;
1638 uiop->uio_rw = UIO_READ;
1640 switch (vp->v_type) {
1642 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1643 NFSINCRGLOBAL(nfsstatsv1.read_bios);
1644 error = ncl_readrpc(vp, uiop, cr);
1647 if (uiop->uio_resid) {
1649 * If we had a short read with no error, we must have
1650 * hit a file hole. We should zero-fill the remainder.
1651 * This can also occur if the server hits the file EOF.
1653 * Holes used to be able to occur due to pending
1654 * writes, but that is not possible any longer.
1656 int nread = bp->b_bcount - uiop->uio_resid;
1657 ssize_t left = uiop->uio_resid;
1660 bzero((char *)bp->b_data + nread, left);
1661 uiop->uio_resid = 0;
1664 /* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1665 if (p && (vp->v_vflag & VV_TEXT)) {
1666 mtx_lock(&np->n_mtx);
1667 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1668 mtx_unlock(&np->n_mtx);
1670 killproc(p, "text file modification");
1673 mtx_unlock(&np->n_mtx);
1677 uiop->uio_offset = (off_t)0;
1678 NFSINCRGLOBAL(nfsstatsv1.readlink_bios);
1679 error = ncl_readlinkrpc(vp, uiop, cr);
1682 NFSINCRGLOBAL(nfsstatsv1.readdir_bios);
1683 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1684 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1685 error = ncl_readdirplusrpc(vp, uiop, cr, td);
1686 if (error == NFSERR_NOTSUPP)
1687 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1689 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1690 error = ncl_readdirrpc(vp, uiop, cr, td);
1692 * end-of-directory sets B_INVAL but does not generate an
1695 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1696 bp->b_flags |= B_INVAL;
1699 printf("ncl_doio: type %x unexpected\n", vp->v_type);
1703 bp->b_ioflags |= BIO_ERROR;
1704 bp->b_error = error;
1708 * If we only need to commit, try to commit
1710 if (bp->b_flags & B_NEEDCOMMIT) {
1714 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1715 retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1718 bp->b_dirtyoff = bp->b_dirtyend = 0;
1719 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1724 if (retv == NFSERR_STALEWRITEVERF) {
1725 ncl_clearcommit(vp->v_mount);
1730 * Setup for actual write
1732 mtx_lock(&np->n_mtx);
1733 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1734 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1735 mtx_unlock(&np->n_mtx);
1737 if (bp->b_dirtyend > bp->b_dirtyoff) {
1738 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1740 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1742 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1743 uiop->uio_rw = UIO_WRITE;
1744 NFSINCRGLOBAL(nfsstatsv1.write_bios);
1746 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1747 iomode = NFSWRITE_UNSTABLE;
1749 iomode = NFSWRITE_FILESYNC;
1751 error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1752 called_from_strategy);
1755 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1756 * to cluster the buffers needing commit. This will allow
1757 * the system to submit a single commit rpc for the whole
1758 * cluster. We can do this even if the buffer is not 100%
1759 * dirty (relative to the NFS blocksize), so we optimize the
1760 * append-to-file-case.
1762 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1763 * cleared because write clustering only works for commit
1764 * rpc's, not for the data portion of the write).
1767 if (!error && iomode == NFSWRITE_UNSTABLE) {
1768 bp->b_flags |= B_NEEDCOMMIT;
1769 if (bp->b_dirtyoff == 0
1770 && bp->b_dirtyend == bp->b_bcount)
1771 bp->b_flags |= B_CLUSTEROK;
1773 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1777 * For an interrupted write, the buffer is still valid
1778 * and the write hasn't been pushed to the server yet,
1779 * so we can't set BIO_ERROR and report the interruption
1780 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1781 * is not relevant, so the rpc attempt is essentially
1782 * a noop. For the case of a V3 write rpc not being
1783 * committed to stable storage, the block is still
1784 * dirty and requires either a commit rpc or another
1785 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1786 * the block is reused. This is indicated by setting
1787 * the B_DELWRI and B_NEEDCOMMIT flags.
1789 * EIO is returned by ncl_writerpc() to indicate a recoverable
1790 * write error and is handled as above, except that
1791 * B_EINTR isn't set. One cause of this is a stale stateid
1792 * error for the RPC that indicates recovery is required,
1793 * when called with called_from_strategy != 0.
1795 * If the buffer is marked B_PAGING, it does not reside on
1796 * the vp's paging queues so we cannot call bdirty(). The
1797 * bp in this case is not an NFS cache block so we should
1800 * The logic below breaks up errors into recoverable and
1801 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1802 * and keep the buffer around for potential write retries.
1803 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1804 * and save the error in the nfsnode. This is less than ideal
1805 * but necessary. Keeping such buffers around could potentially
1806 * cause buffer exhaustion eventually (they can never be written
1807 * out, so will get constantly be re-dirtied). It also causes
1808 * all sorts of vfs panics. For non-recoverable write errors,
1809 * also invalidate the attrcache, so we'll be forced to go over
1810 * the wire for this object, returning an error to user on next
1811 * call (most of the time).
1813 if (error == EINTR || error == EIO || error == ETIMEDOUT
1814 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1815 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1816 if ((bp->b_flags & B_PAGING) == 0) {
1818 bp->b_flags &= ~B_DONE;
1820 if ((error == EINTR || error == ETIMEDOUT) &&
1821 (bp->b_flags & B_ASYNC) == 0)
1822 bp->b_flags |= B_EINTR;
1825 bp->b_ioflags |= BIO_ERROR;
1826 bp->b_flags |= B_INVAL;
1827 bp->b_error = np->n_error = error;
1828 mtx_lock(&np->n_mtx);
1829 np->n_flag |= NWRITEERR;
1830 np->n_attrstamp = 0;
1831 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1832 mtx_unlock(&np->n_mtx);
1834 bp->b_dirtyoff = bp->b_dirtyend = 0;
1842 bp->b_resid = uiop->uio_resid;
1844 ncl_clearcommit(vp->v_mount);
1850 * Used to aid in handling ftruncate() operations on the NFS client side.
1851 * Truncation creates a number of special problems for NFS. We have to
1852 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1853 * we have to properly handle VM pages or (potentially dirty) buffers
1854 * that straddle the truncation point.
1858 ncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1860 struct nfsnode *np = VTONFS(vp);
1862 int biosize = vp->v_bufobj.bo_bsize;
1865 mtx_lock(&np->n_mtx);
1868 mtx_unlock(&np->n_mtx);
1870 if (nsize < tsize) {
1876 * vtruncbuf() doesn't get the buffer overlapping the
1877 * truncation point. We may have a B_DELWRI and/or B_CACHE
1878 * buffer that now needs to be truncated.
1880 error = vtruncbuf(vp, cred, nsize, biosize);
1881 lbn = nsize / biosize;
1882 bufsize = nsize - (lbn * biosize);
1883 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1886 if (bp->b_dirtyoff > bp->b_bcount)
1887 bp->b_dirtyoff = bp->b_bcount;
1888 if (bp->b_dirtyend > bp->b_bcount)
1889 bp->b_dirtyend = bp->b_bcount;
1890 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1893 vnode_pager_setsize(vp, nsize);