2 * Copyright (c) 1989, 1993
3 * The Regents of the University of California. All rights reserved.
5 * This code is derived from software contributed to Berkeley by
6 * Rick Macklem at The University of Guelph.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 4. Neither the name of the University nor the names of its contributors
17 * may be used to endorse or promote products derived from this software
18 * without specific prior written permission.
20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * @(#)nfs_bio.c 8.9 (Berkeley) 3/30/95
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
38 #include <sys/param.h>
39 #include <sys/systm.h>
42 #include <sys/kernel.h>
43 #include <sys/mount.h>
45 #include <sys/resourcevar.h>
46 #include <sys/signalvar.h>
47 #include <sys/vmmeter.h>
48 #include <sys/vnode.h>
51 #include <vm/vm_extern.h>
52 #include <vm/vm_page.h>
53 #include <vm/vm_object.h>
54 #include <vm/vm_pager.h>
55 #include <vm/vnode_pager.h>
57 #include <fs/nfs/nfsport.h>
58 #include <fs/nfsclient/nfsmount.h>
59 #include <fs/nfsclient/nfs.h>
60 #include <fs/nfsclient/nfsnode.h>
62 extern int newnfs_directio_allow_mmap;
63 extern struct nfsstats newnfsstats;
64 extern struct mtx ncl_iod_mutex;
65 extern int ncl_numasync;
66 extern enum nfsiod_state ncl_iodwant[NFS_MAXRAHEAD];
67 extern struct nfsmount *ncl_iodmount[NFS_MAXRAHEAD];
68 extern int newnfs_directio_enable;
70 int ncl_pbuf_freecnt = -1; /* start out unlimited */
72 static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size,
74 static int nfs_directio_write(struct vnode *vp, struct uio *uiop,
75 struct ucred *cred, int ioflag);
78 * Vnode op for VM getpages.
81 ncl_getpages(struct vop_getpages_args *ap)
83 int i, error, nextoff, size, toff, count, npages;
98 td = curthread; /* XXX */
99 cred = curthread->td_ucred; /* XXX */
100 nmp = VFSTONFS(vp->v_mount);
104 if ((object = vp->v_object) == NULL) {
105 ncl_printf("nfs_getpages: called with non-merged cache vnode??\n");
106 return (VM_PAGER_ERROR);
109 if (newnfs_directio_enable && !newnfs_directio_allow_mmap) {
110 mtx_lock(&np->n_mtx);
111 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
112 mtx_unlock(&np->n_mtx);
113 ncl_printf("nfs_getpages: called on non-cacheable vnode??\n");
114 return (VM_PAGER_ERROR);
116 mtx_unlock(&np->n_mtx);
119 mtx_lock(&nmp->nm_mtx);
120 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
121 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
122 mtx_unlock(&nmp->nm_mtx);
123 /* We'll never get here for v4, because we always have fsinfo */
124 (void)ncl_fsinfo(nmp, vp, cred, td);
126 mtx_unlock(&nmp->nm_mtx);
128 npages = btoc(count);
131 * If the requested page is partially valid, just return it and
132 * allow the pager to zero-out the blanks. Partially valid pages
133 * can only occur at the file EOF.
135 VM_OBJECT_LOCK(object);
136 if (pages[ap->a_reqpage]->valid != 0) {
137 vm_page_lock_queues();
138 for (i = 0; i < npages; ++i) {
139 if (i != ap->a_reqpage)
140 vm_page_free(pages[i]);
142 vm_page_unlock_queues();
143 VM_OBJECT_UNLOCK(object);
146 VM_OBJECT_UNLOCK(object);
149 * We use only the kva address for the buffer, but this is extremely
150 * convienient and fast.
152 bp = getpbuf(&ncl_pbuf_freecnt);
154 kva = (vm_offset_t) bp->b_data;
155 pmap_qenter(kva, pages, npages);
156 PCPU_INC(cnt.v_vnodein);
157 PCPU_ADD(cnt.v_vnodepgsin, npages);
159 iov.iov_base = (caddr_t) kva;
163 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
164 uio.uio_resid = count;
165 uio.uio_segflg = UIO_SYSSPACE;
166 uio.uio_rw = UIO_READ;
169 error = ncl_readrpc(vp, &uio, cred);
170 pmap_qremove(kva, npages);
172 relpbuf(bp, &ncl_pbuf_freecnt);
174 if (error && (uio.uio_resid == count)) {
175 ncl_printf("nfs_getpages: error %d\n", error);
176 VM_OBJECT_LOCK(object);
177 vm_page_lock_queues();
178 for (i = 0; i < npages; ++i) {
179 if (i != ap->a_reqpage)
180 vm_page_free(pages[i]);
182 vm_page_unlock_queues();
183 VM_OBJECT_UNLOCK(object);
184 return (VM_PAGER_ERROR);
188 * Calculate the number of bytes read and validate only that number
189 * of bytes. Note that due to pending writes, size may be 0. This
190 * does not mean that the remaining data is invalid!
193 size = count - uio.uio_resid;
194 VM_OBJECT_LOCK(object);
195 vm_page_lock_queues();
196 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
198 nextoff = toff + PAGE_SIZE;
201 if (nextoff <= size) {
203 * Read operation filled an entire page
205 m->valid = VM_PAGE_BITS_ALL;
206 KASSERT(m->dirty == 0,
207 ("nfs_getpages: page %p is dirty", m));
208 } else if (size > toff) {
210 * Read operation filled a partial page.
213 vm_page_set_valid(m, 0, size - toff);
214 KASSERT(m->dirty == 0,
215 ("nfs_getpages: page %p is dirty", m));
218 * Read operation was short. If no error occured
219 * we may have hit a zero-fill section. We simply
220 * leave valid set to 0.
224 if (i != ap->a_reqpage) {
226 * Whether or not to leave the page activated is up in
227 * the air, but we should put the page on a page queue
228 * somewhere (it already is in the object). Result:
229 * It appears that emperical results show that
230 * deactivating pages is best.
234 * Just in case someone was asking for this page we
235 * now tell them that it is ok to use.
238 if (m->oflags & VPO_WANTED)
241 vm_page_deactivate(m);
248 vm_page_unlock_queues();
249 VM_OBJECT_UNLOCK(object);
254 * Vnode op for VM putpages.
257 ncl_putpages(struct vop_putpages_args *ap)
263 int iomode, must_commit, i, error, npages, count;
269 struct nfsmount *nmp;
275 td = curthread; /* XXX */
276 cred = curthread->td_ucred; /* XXX */
277 nmp = VFSTONFS(vp->v_mount);
280 rtvals = ap->a_rtvals;
281 npages = btoc(count);
282 offset = IDX_TO_OFF(pages[0]->pindex);
284 mtx_lock(&nmp->nm_mtx);
285 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
286 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
287 mtx_unlock(&nmp->nm_mtx);
288 (void)ncl_fsinfo(nmp, vp, cred, td);
290 mtx_unlock(&nmp->nm_mtx);
292 mtx_lock(&np->n_mtx);
293 if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
294 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
295 mtx_unlock(&np->n_mtx);
296 ncl_printf("ncl_putpages: called on noncache-able vnode??\n");
297 mtx_lock(&np->n_mtx);
300 for (i = 0; i < npages; i++)
301 rtvals[i] = VM_PAGER_AGAIN;
304 * When putting pages, do not extend file past EOF.
306 if (offset + count > np->n_size) {
307 count = np->n_size - offset;
311 mtx_unlock(&np->n_mtx);
314 * We use only the kva address for the buffer, but this is extremely
315 * convienient and fast.
317 bp = getpbuf(&ncl_pbuf_freecnt);
319 kva = (vm_offset_t) bp->b_data;
320 pmap_qenter(kva, pages, npages);
321 PCPU_INC(cnt.v_vnodeout);
322 PCPU_ADD(cnt.v_vnodepgsout, count);
324 iov.iov_base = (caddr_t) kva;
328 uio.uio_offset = offset;
329 uio.uio_resid = count;
330 uio.uio_segflg = UIO_SYSSPACE;
331 uio.uio_rw = UIO_WRITE;
334 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
335 iomode = NFSWRITE_UNSTABLE;
337 iomode = NFSWRITE_FILESYNC;
339 error = ncl_writerpc(vp, &uio, cred, &iomode, &must_commit, 0);
341 pmap_qremove(kva, npages);
342 relpbuf(bp, &ncl_pbuf_freecnt);
345 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE;
346 for (i = 0; i < nwritten; i++) {
347 rtvals[i] = VM_PAGER_OK;
348 vm_page_undirty(pages[i]);
351 ncl_clearcommit(vp->v_mount);
358 * For nfs, cache consistency can only be maintained approximately.
359 * Although RFC1094 does not specify the criteria, the following is
360 * believed to be compatible with the reference port.
362 * If the file's modify time on the server has changed since the
363 * last read rpc or you have written to the file,
364 * you may have lost data cache consistency with the
365 * server, so flush all of the file's data out of the cache.
366 * Then force a getattr rpc to ensure that you have up to date
368 * NB: This implies that cache data can be read when up to
369 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
370 * attributes this could be forced by setting n_attrstamp to 0 before
371 * the VOP_GETATTR() call.
374 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
378 struct nfsnode *np = VTONFS(vp);
382 * Grab the exclusive lock before checking whether the cache is
384 * XXX - We can make this cheaper later (by acquiring cheaper locks).
385 * But for now, this suffices.
387 old_lock = ncl_upgrade_vnlock(vp);
388 if (vp->v_iflag & VI_DOOMED) {
389 ncl_downgrade_vnlock(vp, old_lock);
393 mtx_lock(&np->n_mtx);
394 if (np->n_flag & NMODIFIED) {
395 mtx_unlock(&np->n_mtx);
396 if (vp->v_type != VREG) {
397 if (vp->v_type != VDIR)
398 panic("nfs: bioread, not dir");
400 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
405 error = VOP_GETATTR(vp, &vattr, cred);
408 mtx_lock(&np->n_mtx);
409 np->n_mtime = vattr.va_mtime;
410 mtx_unlock(&np->n_mtx);
412 mtx_unlock(&np->n_mtx);
413 error = VOP_GETATTR(vp, &vattr, cred);
416 mtx_lock(&np->n_mtx);
417 if ((np->n_flag & NSIZECHANGED)
418 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
419 mtx_unlock(&np->n_mtx);
420 if (vp->v_type == VDIR)
422 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
425 mtx_lock(&np->n_mtx);
426 np->n_mtime = vattr.va_mtime;
427 np->n_flag &= ~NSIZECHANGED;
429 mtx_unlock(&np->n_mtx);
432 ncl_downgrade_vnlock(vp, old_lock);
437 * Vnode op for read using bio
440 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
442 struct nfsnode *np = VTONFS(vp);
444 struct buf *bp, *rabp;
446 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
450 int nra, error = 0, n = 0, on = 0;
452 KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
453 if (uio->uio_resid == 0)
455 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
459 mtx_lock(&nmp->nm_mtx);
460 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
461 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
462 mtx_unlock(&nmp->nm_mtx);
463 (void)ncl_fsinfo(nmp, vp, cred, td);
464 mtx_lock(&nmp->nm_mtx);
466 if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
467 (void) newnfs_iosize(nmp);
468 mtx_unlock(&nmp->nm_mtx);
470 if (vp->v_type != VDIR &&
471 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
474 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
475 /* No caching/ no readaheads. Just read data into the user buffer */
476 return ncl_readrpc(vp, uio, cred);
478 biosize = vp->v_mount->mnt_stat.f_iosize;
479 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
481 error = nfs_bioread_check_cons(vp, td, cred);
488 mtx_lock(&np->n_mtx);
490 mtx_unlock(&np->n_mtx);
492 switch (vp->v_type) {
494 NFSINCRGLOBAL(newnfsstats.biocache_reads);
495 lbn = uio->uio_offset / biosize;
496 on = uio->uio_offset & (biosize - 1);
499 * Start the read ahead(s), as required.
501 if (nmp->nm_readahead > 0) {
502 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
503 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
504 rabn = lbn + 1 + nra;
505 if (incore(&vp->v_bufobj, rabn) == NULL) {
506 rabp = nfs_getcacheblk(vp, rabn, biosize, td);
508 error = newnfs_sigintr(nmp, td);
514 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
515 rabp->b_flags |= B_ASYNC;
516 rabp->b_iocmd = BIO_READ;
517 vfs_busy_pages(rabp, 0);
518 if (ncl_asyncio(nmp, rabp, cred, td)) {
519 rabp->b_flags |= B_INVAL;
520 rabp->b_ioflags |= BIO_ERROR;
521 vfs_unbusy_pages(rabp);
532 /* Note that bcount is *not* DEV_BSIZE aligned. */
534 if ((off_t)lbn * biosize >= nsize) {
536 } else if ((off_t)(lbn + 1) * biosize > nsize) {
537 bcount = nsize - (off_t)lbn * biosize;
539 bp = nfs_getcacheblk(vp, lbn, bcount, td);
542 error = newnfs_sigintr(nmp, td);
543 return (error ? error : EINTR);
547 * If B_CACHE is not set, we must issue the read. If this
548 * fails, we return an error.
551 if ((bp->b_flags & B_CACHE) == 0) {
552 bp->b_iocmd = BIO_READ;
553 vfs_busy_pages(bp, 0);
554 error = ncl_doio(vp, bp, cred, td, 0);
562 * on is the offset into the current bp. Figure out how many
563 * bytes we can copy out of the bp. Note that bcount is
564 * NOT DEV_BSIZE aligned.
566 * Then figure out how many bytes we can copy into the uio.
571 n = min((unsigned)(bcount - on), uio->uio_resid);
574 NFSINCRGLOBAL(newnfsstats.biocache_readlinks);
575 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
577 error = newnfs_sigintr(nmp, td);
578 return (error ? error : EINTR);
580 if ((bp->b_flags & B_CACHE) == 0) {
581 bp->b_iocmd = BIO_READ;
582 vfs_busy_pages(bp, 0);
583 error = ncl_doio(vp, bp, cred, td, 0);
585 bp->b_ioflags |= BIO_ERROR;
590 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
594 NFSINCRGLOBAL(newnfsstats.biocache_readdirs);
595 if (np->n_direofoffset
596 && uio->uio_offset >= np->n_direofoffset) {
599 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
600 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
601 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
603 error = newnfs_sigintr(nmp, td);
604 return (error ? error : EINTR);
606 if ((bp->b_flags & B_CACHE) == 0) {
607 bp->b_iocmd = BIO_READ;
608 vfs_busy_pages(bp, 0);
609 error = ncl_doio(vp, bp, cred, td, 0);
613 while (error == NFSERR_BAD_COOKIE) {
615 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++) {
627 if (np->n_direofoffset
628 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
630 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
632 error = newnfs_sigintr(nmp, td);
633 return (error ? error : EINTR);
635 if ((bp->b_flags & B_CACHE) == 0) {
636 bp->b_iocmd = BIO_READ;
637 vfs_busy_pages(bp, 0);
638 error = ncl_doio(vp, bp, cred, td, 0);
640 * no error + B_INVAL == directory EOF,
643 if (error == 0 && (bp->b_flags & B_INVAL))
647 * An error will throw away the block and the
648 * for loop will break out. If no error and this
649 * is not the block we want, we throw away the
650 * block and go for the next one via the for loop.
652 if (error || i < lbn)
657 * The above while is repeated if we hit another cookie
658 * error. If we hit an error and it wasn't a cookie error,
666 * If not eof and read aheads are enabled, start one.
667 * (You need the current block first, so that you have the
668 * directory offset cookie of the next block.)
670 if (nmp->nm_readahead > 0 &&
671 (bp->b_flags & B_INVAL) == 0 &&
672 (np->n_direofoffset == 0 ||
673 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
674 incore(&vp->v_bufobj, lbn + 1) == NULL) {
675 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
677 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
678 rabp->b_flags |= B_ASYNC;
679 rabp->b_iocmd = BIO_READ;
680 vfs_busy_pages(rabp, 0);
681 if (ncl_asyncio(nmp, rabp, cred, td)) {
682 rabp->b_flags |= B_INVAL;
683 rabp->b_ioflags |= BIO_ERROR;
684 vfs_unbusy_pages(rabp);
693 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
694 * chopped for the EOF condition, we cannot tell how large
695 * NFS directories are going to be until we hit EOF. So
696 * an NFS directory buffer is *not* chopped to its EOF. Now,
697 * it just so happens that b_resid will effectively chop it
698 * to EOF. *BUT* this information is lost if the buffer goes
699 * away and is reconstituted into a B_CACHE state ( due to
700 * being VMIO ) later. So we keep track of the directory eof
701 * in np->n_direofoffset and chop it off as an extra step
704 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
705 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
706 n = np->n_direofoffset - uio->uio_offset;
709 ncl_printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
715 error = uiomove(bp->b_data + on, (int)n, uio);
717 if (vp->v_type == VLNK)
721 } while (error == 0 && uio->uio_resid > 0 && n > 0);
726 * The NFS write path cannot handle iovecs with len > 1. So we need to
727 * break up iovecs accordingly (restricting them to wsize).
728 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
729 * For the ASYNC case, 2 copies are needed. The first a copy from the
730 * user buffer to a staging buffer and then a second copy from the staging
731 * buffer to mbufs. This can be optimized by copying from the user buffer
732 * directly into mbufs and passing the chain down, but that requires a
733 * fair amount of re-working of the relevant codepaths (and can be done
737 nfs_directio_write(vp, uiop, cred, ioflag)
744 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
745 struct thread *td = uiop->uio_td;
749 mtx_lock(&nmp->nm_mtx);
750 wsize = nmp->nm_wsize;
751 mtx_unlock(&nmp->nm_mtx);
752 if (ioflag & IO_SYNC) {
753 int iomode, must_commit;
757 while (uiop->uio_resid > 0) {
758 size = min(uiop->uio_resid, wsize);
759 size = min(uiop->uio_iov->iov_len, size);
760 iov.iov_base = uiop->uio_iov->iov_base;
764 uio.uio_offset = uiop->uio_offset;
765 uio.uio_resid = size;
766 uio.uio_segflg = UIO_USERSPACE;
767 uio.uio_rw = UIO_WRITE;
769 iomode = NFSWRITE_FILESYNC;
770 error = ncl_writerpc(vp, &uio, cred, &iomode,
772 KASSERT((must_commit == 0),
773 ("ncl_directio_write: Did not commit write"));
776 uiop->uio_offset += size;
777 uiop->uio_resid -= size;
778 if (uiop->uio_iov->iov_len <= size) {
782 uiop->uio_iov->iov_base =
783 (char *)uiop->uio_iov->iov_base + size;
784 uiop->uio_iov->iov_len -= size;
793 * Break up the write into blocksize chunks and hand these
794 * over to nfsiod's for write back.
795 * Unfortunately, this incurs a copy of the data. Since
796 * the user could modify the buffer before the write is
799 * The obvious optimization here is that one of the 2 copies
800 * in the async write path can be eliminated by copying the
801 * data here directly into mbufs and passing the mbuf chain
802 * down. But that will require a fair amount of re-working
803 * of the code and can be done if there's enough interest
804 * in NFS directio access.
806 while (uiop->uio_resid > 0) {
807 size = min(uiop->uio_resid, wsize);
808 size = min(uiop->uio_iov->iov_len, size);
809 bp = getpbuf(&ncl_pbuf_freecnt);
810 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
811 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
812 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
813 t_iov->iov_len = size;
814 t_uio->uio_iov = t_iov;
815 t_uio->uio_iovcnt = 1;
816 t_uio->uio_offset = uiop->uio_offset;
817 t_uio->uio_resid = size;
818 t_uio->uio_segflg = UIO_SYSSPACE;
819 t_uio->uio_rw = UIO_WRITE;
821 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base, size);
822 bp->b_flags |= B_DIRECT;
823 bp->b_iocmd = BIO_WRITE;
824 if (cred != NOCRED) {
828 bp->b_wcred = NOCRED;
829 bp->b_caller1 = (void *)t_uio;
831 error = ncl_asyncio(nmp, bp, NOCRED, td);
833 free(t_iov->iov_base, M_NFSDIRECTIO);
834 free(t_iov, M_NFSDIRECTIO);
835 free(t_uio, M_NFSDIRECTIO);
837 relpbuf(bp, &ncl_pbuf_freecnt);
842 uiop->uio_offset += size;
843 uiop->uio_resid -= size;
844 if (uiop->uio_iov->iov_len <= size) {
848 uiop->uio_iov->iov_base =
849 (char *)uiop->uio_iov->iov_base + size;
850 uiop->uio_iov->iov_len -= size;
858 * Vnode op for write using bio
861 ncl_write(struct vop_write_args *ap)
864 struct uio *uio = ap->a_uio;
865 struct thread *td = uio->uio_td;
866 struct vnode *vp = ap->a_vp;
867 struct nfsnode *np = VTONFS(vp);
868 struct ucred *cred = ap->a_cred;
869 int ioflag = ap->a_ioflag;
872 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
875 int n, on, error = 0;
876 struct proc *p = td?td->td_proc:NULL;
878 KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
879 KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
881 if (vp->v_type != VREG)
883 mtx_lock(&np->n_mtx);
884 if (np->n_flag & NWRITEERR) {
885 np->n_flag &= ~NWRITEERR;
886 mtx_unlock(&np->n_mtx);
887 return (np->n_error);
889 mtx_unlock(&np->n_mtx);
890 mtx_lock(&nmp->nm_mtx);
891 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
892 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
893 mtx_unlock(&nmp->nm_mtx);
894 (void)ncl_fsinfo(nmp, vp, cred, td);
895 mtx_lock(&nmp->nm_mtx);
897 if (nmp->nm_wsize == 0)
898 (void) newnfs_iosize(nmp);
899 mtx_unlock(&nmp->nm_mtx);
902 * Synchronously flush pending buffers if we are in synchronous
903 * mode or if we are appending.
905 if (ioflag & (IO_APPEND | IO_SYNC)) {
906 mtx_lock(&np->n_mtx);
907 if (np->n_flag & NMODIFIED) {
908 mtx_unlock(&np->n_mtx);
909 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
911 * Require non-blocking, synchronous writes to
912 * dirty files to inform the program it needs
913 * to fsync(2) explicitly.
915 if (ioflag & IO_NDELAY)
920 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
924 mtx_unlock(&np->n_mtx);
928 * If IO_APPEND then load uio_offset. We restart here if we cannot
929 * get the append lock.
931 if (ioflag & IO_APPEND) {
933 error = VOP_GETATTR(vp, &vattr, cred);
936 mtx_lock(&np->n_mtx);
937 uio->uio_offset = np->n_size;
938 mtx_unlock(&np->n_mtx);
941 if (uio->uio_offset < 0)
943 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
945 if (uio->uio_resid == 0)
948 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
949 return nfs_directio_write(vp, uio, cred, ioflag);
952 * Maybe this should be above the vnode op call, but so long as
953 * file servers have no limits, i don't think it matters
957 if (uio->uio_offset + uio->uio_resid >
958 lim_cur(p, RLIMIT_FSIZE)) {
966 biosize = vp->v_mount->mnt_stat.f_iosize;
968 * Find all of this file's B_NEEDCOMMIT buffers. If our writes
969 * would exceed the local maximum per-file write commit size when
970 * combined with those, we must decide whether to flush,
971 * go synchronous, or return error. We don't bother checking
972 * IO_UNIT -- we just make all writes atomic anyway, as there's
973 * no point optimizing for something that really won't ever happen.
975 if (!(ioflag & IO_SYNC)) {
978 mtx_lock(&np->n_mtx);
980 mtx_unlock(&np->n_mtx);
982 if (nmp->nm_wcommitsize < uio->uio_resid) {
984 * If this request could not possibly be completed
985 * without exceeding the maximum outstanding write
986 * commit size, see if we can convert it into a
987 * synchronous write operation.
989 if (ioflag & IO_NDELAY)
992 if (nflag & NMODIFIED)
994 } else if (nflag & NMODIFIED) {
996 BO_LOCK(&vp->v_bufobj);
997 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
998 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
1000 if (bp->b_flags & B_NEEDCOMMIT)
1001 wouldcommit += bp->b_bcount;
1004 BO_UNLOCK(&vp->v_bufobj);
1006 * Since we're not operating synchronously and
1007 * bypassing the buffer cache, we are in a commit
1008 * and holding all of these buffers whether
1009 * transmitted or not. If not limited, this
1010 * will lead to the buffer cache deadlocking,
1011 * as no one else can flush our uncommitted buffers.
1013 wouldcommit += uio->uio_resid;
1015 * If we would initially exceed the maximum
1016 * outstanding write commit size, flush and restart.
1018 if (wouldcommit > nmp->nm_wcommitsize)
1022 goto flush_and_restart;
1026 NFSINCRGLOBAL(newnfsstats.biocache_writes);
1027 lbn = uio->uio_offset / biosize;
1028 on = uio->uio_offset & (biosize-1);
1029 n = min((unsigned)(biosize - on), uio->uio_resid);
1032 * Handle direct append and file extension cases, calculate
1033 * unaligned buffer size.
1035 mtx_lock(&np->n_mtx);
1036 if (uio->uio_offset == np->n_size && n) {
1037 mtx_unlock(&np->n_mtx);
1039 * Get the buffer (in its pre-append state to maintain
1040 * B_CACHE if it was previously set). Resize the
1041 * nfsnode after we have locked the buffer to prevent
1042 * readers from reading garbage.
1045 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1050 mtx_lock(&np->n_mtx);
1051 np->n_size = uio->uio_offset + n;
1052 np->n_flag |= NMODIFIED;
1053 vnode_pager_setsize(vp, np->n_size);
1054 mtx_unlock(&np->n_mtx);
1056 save = bp->b_flags & B_CACHE;
1058 allocbuf(bp, bcount);
1059 bp->b_flags |= save;
1063 * Obtain the locked cache block first, and then
1064 * adjust the file's size as appropriate.
1067 if ((off_t)lbn * biosize + bcount < np->n_size) {
1068 if ((off_t)(lbn + 1) * biosize < np->n_size)
1071 bcount = np->n_size - (off_t)lbn * biosize;
1073 mtx_unlock(&np->n_mtx);
1074 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1075 mtx_lock(&np->n_mtx);
1076 if (uio->uio_offset + n > np->n_size) {
1077 np->n_size = uio->uio_offset + n;
1078 np->n_flag |= NMODIFIED;
1079 vnode_pager_setsize(vp, np->n_size);
1081 mtx_unlock(&np->n_mtx);
1085 error = newnfs_sigintr(nmp, td);
1092 * Issue a READ if B_CACHE is not set. In special-append
1093 * mode, B_CACHE is based on the buffer prior to the write
1094 * op and is typically set, avoiding the read. If a read
1095 * is required in special append mode, the server will
1096 * probably send us a short-read since we extended the file
1097 * on our end, resulting in b_resid == 0 and, thusly,
1098 * B_CACHE getting set.
1100 * We can also avoid issuing the read if the write covers
1101 * the entire buffer. We have to make sure the buffer state
1102 * is reasonable in this case since we will not be initiating
1103 * I/O. See the comments in kern/vfs_bio.c's getblk() for
1106 * B_CACHE may also be set due to the buffer being cached
1110 if (on == 0 && n == bcount) {
1111 bp->b_flags |= B_CACHE;
1112 bp->b_flags &= ~B_INVAL;
1113 bp->b_ioflags &= ~BIO_ERROR;
1116 if ((bp->b_flags & B_CACHE) == 0) {
1117 bp->b_iocmd = BIO_READ;
1118 vfs_busy_pages(bp, 0);
1119 error = ncl_doio(vp, bp, cred, td, 0);
1125 if (bp->b_wcred == NOCRED)
1126 bp->b_wcred = crhold(cred);
1127 mtx_lock(&np->n_mtx);
1128 np->n_flag |= NMODIFIED;
1129 mtx_unlock(&np->n_mtx);
1132 * If dirtyend exceeds file size, chop it down. This should
1133 * not normally occur but there is an append race where it
1134 * might occur XXX, so we log it.
1136 * If the chopping creates a reverse-indexed or degenerate
1137 * situation with dirtyoff/end, we 0 both of them.
1140 if (bp->b_dirtyend > bcount) {
1141 ncl_printf("NFS append race @%lx:%d\n",
1142 (long)bp->b_blkno * DEV_BSIZE,
1143 bp->b_dirtyend - bcount);
1144 bp->b_dirtyend = bcount;
1147 if (bp->b_dirtyoff >= bp->b_dirtyend)
1148 bp->b_dirtyoff = bp->b_dirtyend = 0;
1151 * If the new write will leave a contiguous dirty
1152 * area, just update the b_dirtyoff and b_dirtyend,
1153 * otherwise force a write rpc of the old dirty area.
1155 * While it is possible to merge discontiguous writes due to
1156 * our having a B_CACHE buffer ( and thus valid read data
1157 * for the hole), we don't because it could lead to
1158 * significant cache coherency problems with multiple clients,
1159 * especially if locking is implemented later on.
1161 * as an optimization we could theoretically maintain
1162 * a linked list of discontinuous areas, but we would still
1163 * have to commit them separately so there isn't much
1164 * advantage to it except perhaps a bit of asynchronization.
1167 if (bp->b_dirtyend > 0 &&
1168 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1169 if (bwrite(bp) == EINTR) {
1176 error = uiomove((char *)bp->b_data + on, n, uio);
1179 * Since this block is being modified, it must be written
1180 * again and not just committed. Since write clustering does
1181 * not work for the stage 1 data write, only the stage 2
1182 * commit rpc, we have to clear B_CLUSTEROK as well.
1184 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1187 bp->b_ioflags |= BIO_ERROR;
1193 * Only update dirtyoff/dirtyend if not a degenerate
1197 if (bp->b_dirtyend > 0) {
1198 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1199 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1201 bp->b_dirtyoff = on;
1202 bp->b_dirtyend = on + n;
1204 vfs_bio_set_valid(bp, on, n);
1208 * If IO_SYNC do bwrite().
1210 * IO_INVAL appears to be unused. The idea appears to be
1211 * to turn off caching in this case. Very odd. XXX
1213 if ((ioflag & IO_SYNC)) {
1214 if (ioflag & IO_INVAL)
1215 bp->b_flags |= B_NOCACHE;
1219 } else if ((n + on) == biosize) {
1220 bp->b_flags |= B_ASYNC;
1221 (void) ncl_writebp(bp, 0, NULL);
1225 } while (uio->uio_resid > 0 && n > 0);
1231 * Get an nfs cache block.
1233 * Allocate a new one if the block isn't currently in the cache
1234 * and return the block marked busy. If the calling process is
1235 * interrupted by a signal for an interruptible mount point, return
1238 * The caller must carefully deal with the possible B_INVAL state of
1239 * the buffer. ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
1240 * indirectly), so synchronous reads can be issued without worrying about
1241 * the B_INVAL state. We have to be a little more careful when dealing
1242 * with writes (see comments in nfs_write()) when extending a file past
1246 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1250 struct nfsmount *nmp;
1255 if (nmp->nm_flag & NFSMNT_INT) {
1258 newnfs_set_sigmask(td, &oldset);
1259 bp = getblk(vp, bn, size, NFS_PCATCH, 0, 0);
1260 newnfs_restore_sigmask(td, &oldset);
1261 while (bp == NULL) {
1262 if (newnfs_sigintr(nmp, td))
1264 bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1267 bp = getblk(vp, bn, size, 0, 0, 0);
1270 if (vp->v_type == VREG) {
1273 biosize = mp->mnt_stat.f_iosize;
1274 bp->b_blkno = bn * (biosize / DEV_BSIZE);
1280 * Flush and invalidate all dirty buffers. If another process is already
1281 * doing the flush, just wait for completion.
1284 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1286 struct nfsnode *np = VTONFS(vp);
1287 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1288 int error = 0, slpflag, slptimeo;
1291 ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
1293 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1295 if ((nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF))
1298 slpflag = NFS_PCATCH;
1305 old_lock = ncl_upgrade_vnlock(vp);
1306 if (vp->v_iflag & VI_DOOMED) {
1308 * Since vgonel() uses the generic vinvalbuf() to flush
1309 * dirty buffers and it does not call this function, it
1310 * is safe to just return OK when VI_DOOMED is set.
1312 ncl_downgrade_vnlock(vp, old_lock);
1317 * Now, flush as required.
1319 if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) {
1320 VM_OBJECT_LOCK(vp->v_bufobj.bo_object);
1321 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1322 VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object);
1324 * If the page clean was interrupted, fail the invalidation.
1325 * Not doing so, we run the risk of losing dirty pages in the
1326 * vinvalbuf() call below.
1328 if (intrflg && (error = newnfs_sigintr(nmp, td)))
1332 error = vinvalbuf(vp, flags, slpflag, 0);
1334 if (intrflg && (error = newnfs_sigintr(nmp, td)))
1336 error = vinvalbuf(vp, flags, 0, slptimeo);
1338 mtx_lock(&np->n_mtx);
1339 if (np->n_directio_asyncwr == 0)
1340 np->n_flag &= ~NMODIFIED;
1341 mtx_unlock(&np->n_mtx);
1343 ncl_downgrade_vnlock(vp, old_lock);
1348 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1349 * This is mainly to avoid queueing async I/O requests when the nfsiods
1350 * are all hung on a dead server.
1352 * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1353 * is eventually dequeued by the async daemon, ncl_doio() *will*.
1356 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1365 * Unless iothreadcnt is set > 0, don't bother with async I/O
1366 * threads. For LAN environments, they don't buy any significant
1367 * performance improvement that you can't get with large block
1370 if (nmp->nm_readahead == 0)
1374 * Commits are usually short and sweet so lets save some cpu and
1375 * leave the async daemons for more important rpc's (such as reads
1378 mtx_lock(&ncl_iod_mutex);
1379 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1380 (nmp->nm_bufqiods > ncl_numasync / 2)) {
1381 mtx_unlock(&ncl_iod_mutex);
1385 if (nmp->nm_flag & NFSMNT_INT)
1386 slpflag = NFS_PCATCH;
1390 * Find a free iod to process this request.
1392 for (iod = 0; iod < ncl_numasync; iod++)
1393 if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1399 * Try to create one if none are free.
1402 iod = ncl_nfsiodnew(1);
1409 * Found one, so wake it up and tell it which
1412 NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1414 ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1415 ncl_iodmount[iod] = nmp;
1417 wakeup(&ncl_iodwant[iod]);
1421 * If none are free, we may already have an iod working on this mount
1422 * point. If so, it will process our request.
1425 if (nmp->nm_bufqiods > 0) {
1427 ("ncl_asyncio: %d iods are already processing mount %p\n",
1428 nmp->nm_bufqiods, nmp));
1434 * If we have an iod which can process the request, then queue
1439 * Ensure that the queue never grows too large. We still want
1440 * to asynchronize so we block rather then return EIO.
1442 while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1444 ("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1445 nmp->nm_bufqwant = TRUE;
1446 error = newnfs_msleep(td, &nmp->nm_bufq,
1447 &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1450 error2 = newnfs_sigintr(nmp, td);
1452 mtx_unlock(&ncl_iod_mutex);
1455 if (slpflag == NFS_PCATCH) {
1461 * We might have lost our iod while sleeping,
1462 * so check and loop if nescessary.
1464 if (nmp->nm_bufqiods == 0) {
1466 ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1471 /* We might have lost our nfsiod */
1472 if (nmp->nm_bufqiods == 0) {
1474 ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1478 if (bp->b_iocmd == BIO_READ) {
1479 if (bp->b_rcred == NOCRED && cred != NOCRED)
1480 bp->b_rcred = crhold(cred);
1482 if (bp->b_wcred == NOCRED && cred != NOCRED)
1483 bp->b_wcred = crhold(cred);
1486 if (bp->b_flags & B_REMFREE)
1489 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1491 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1492 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx);
1493 VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1494 VTONFS(bp->b_vp)->n_directio_asyncwr++;
1495 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx);
1497 mtx_unlock(&ncl_iod_mutex);
1501 mtx_unlock(&ncl_iod_mutex);
1504 * All the iods are busy on other mounts, so return EIO to
1505 * force the caller to process the i/o synchronously.
1507 NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1512 ncl_doio_directwrite(struct buf *bp)
1514 int iomode, must_commit;
1515 struct uio *uiop = (struct uio *)bp->b_caller1;
1516 char *iov_base = uiop->uio_iov->iov_base;
1518 iomode = NFSWRITE_FILESYNC;
1519 uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1520 ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0);
1521 KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write"));
1522 free(iov_base, M_NFSDIRECTIO);
1523 free(uiop->uio_iov, M_NFSDIRECTIO);
1524 free(uiop, M_NFSDIRECTIO);
1525 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1526 struct nfsnode *np = VTONFS(bp->b_vp);
1527 mtx_lock(&np->n_mtx);
1528 np->n_directio_asyncwr--;
1529 if (np->n_directio_asyncwr == 0) {
1530 np->n_flag &= ~NMODIFIED;
1531 if ((np->n_flag & NFSYNCWAIT)) {
1532 np->n_flag &= ~NFSYNCWAIT;
1533 wakeup((caddr_t)&np->n_directio_asyncwr);
1536 mtx_unlock(&np->n_mtx);
1539 relpbuf(bp, &ncl_pbuf_freecnt);
1543 * Do an I/O operation to/from a cache block. This may be called
1544 * synchronously or from an nfsiod.
1547 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1548 int called_from_strategy)
1552 struct nfsmount *nmp;
1553 int error = 0, iomode, must_commit = 0;
1556 struct proc *p = td ? td->td_proc : NULL;
1560 nmp = VFSTONFS(vp->v_mount);
1562 uiop->uio_iov = &io;
1563 uiop->uio_iovcnt = 1;
1564 uiop->uio_segflg = UIO_SYSSPACE;
1568 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We
1569 * do this here so we do not have to do it in all the code that
1572 bp->b_flags &= ~B_INVAL;
1573 bp->b_ioflags &= ~BIO_ERROR;
1575 KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1576 iocmd = bp->b_iocmd;
1577 if (iocmd == BIO_READ) {
1578 io.iov_len = uiop->uio_resid = bp->b_bcount;
1579 io.iov_base = bp->b_data;
1580 uiop->uio_rw = UIO_READ;
1582 switch (vp->v_type) {
1584 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1585 NFSINCRGLOBAL(newnfsstats.read_bios);
1586 error = ncl_readrpc(vp, uiop, cr);
1589 if (uiop->uio_resid) {
1591 * If we had a short read with no error, we must have
1592 * hit a file hole. We should zero-fill the remainder.
1593 * This can also occur if the server hits the file EOF.
1595 * Holes used to be able to occur due to pending
1596 * writes, but that is not possible any longer.
1598 int nread = bp->b_bcount - uiop->uio_resid;
1599 int left = uiop->uio_resid;
1602 bzero((char *)bp->b_data + nread, left);
1603 uiop->uio_resid = 0;
1606 /* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1607 if (p && (vp->v_vflag & VV_TEXT)) {
1608 mtx_lock(&np->n_mtx);
1609 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1610 mtx_unlock(&np->n_mtx);
1612 killproc(p, "text file modification");
1615 mtx_unlock(&np->n_mtx);
1619 uiop->uio_offset = (off_t)0;
1620 NFSINCRGLOBAL(newnfsstats.readlink_bios);
1621 error = ncl_readlinkrpc(vp, uiop, cr);
1624 NFSINCRGLOBAL(newnfsstats.readdir_bios);
1625 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1626 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1627 error = ncl_readdirplusrpc(vp, uiop, cr, td);
1628 if (error == NFSERR_NOTSUPP)
1629 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1631 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1632 error = ncl_readdirrpc(vp, uiop, cr, td);
1634 * end-of-directory sets B_INVAL but does not generate an
1637 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1638 bp->b_flags |= B_INVAL;
1641 ncl_printf("ncl_doio: type %x unexpected\n", vp->v_type);
1645 bp->b_ioflags |= BIO_ERROR;
1646 bp->b_error = error;
1650 * If we only need to commit, try to commit
1652 if (bp->b_flags & B_NEEDCOMMIT) {
1656 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1657 retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1660 bp->b_dirtyoff = bp->b_dirtyend = 0;
1661 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1666 if (retv == NFSERR_STALEWRITEVERF) {
1667 ncl_clearcommit(vp->v_mount);
1672 * Setup for actual write
1674 mtx_lock(&np->n_mtx);
1675 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1676 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1677 mtx_unlock(&np->n_mtx);
1679 if (bp->b_dirtyend > bp->b_dirtyoff) {
1680 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1682 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1684 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1685 uiop->uio_rw = UIO_WRITE;
1686 NFSINCRGLOBAL(newnfsstats.write_bios);
1688 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1689 iomode = NFSWRITE_UNSTABLE;
1691 iomode = NFSWRITE_FILESYNC;
1693 error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1694 called_from_strategy);
1697 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1698 * to cluster the buffers needing commit. This will allow
1699 * the system to submit a single commit rpc for the whole
1700 * cluster. We can do this even if the buffer is not 100%
1701 * dirty (relative to the NFS blocksize), so we optimize the
1702 * append-to-file-case.
1704 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1705 * cleared because write clustering only works for commit
1706 * rpc's, not for the data portion of the write).
1709 if (!error && iomode == NFSWRITE_UNSTABLE) {
1710 bp->b_flags |= B_NEEDCOMMIT;
1711 if (bp->b_dirtyoff == 0
1712 && bp->b_dirtyend == bp->b_bcount)
1713 bp->b_flags |= B_CLUSTEROK;
1715 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1719 * For an interrupted write, the buffer is still valid
1720 * and the write hasn't been pushed to the server yet,
1721 * so we can't set BIO_ERROR and report the interruption
1722 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1723 * is not relevant, so the rpc attempt is essentially
1724 * a noop. For the case of a V3 write rpc not being
1725 * committed to stable storage, the block is still
1726 * dirty and requires either a commit rpc or another
1727 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1728 * the block is reused. This is indicated by setting
1729 * the B_DELWRI and B_NEEDCOMMIT flags.
1731 * EIO is returned by ncl_writerpc() to indicate a recoverable
1732 * write error and is handled as above, except that
1733 * B_EINTR isn't set. One cause of this is a stale stateid
1734 * error for the RPC that indicates recovery is required,
1735 * when called with called_from_strategy != 0.
1737 * If the buffer is marked B_PAGING, it does not reside on
1738 * the vp's paging queues so we cannot call bdirty(). The
1739 * bp in this case is not an NFS cache block so we should
1742 * The logic below breaks up errors into recoverable and
1743 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1744 * and keep the buffer around for potential write retries.
1745 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1746 * and save the error in the nfsnode. This is less than ideal
1747 * but necessary. Keeping such buffers around could potentially
1748 * cause buffer exhaustion eventually (they can never be written
1749 * out, so will get constantly be re-dirtied). It also causes
1750 * all sorts of vfs panics. For non-recoverable write errors,
1751 * also invalidate the attrcache, so we'll be forced to go over
1752 * the wire for this object, returning an error to user on next
1753 * call (most of the time).
1755 if (error == EINTR || error == EIO || error == ETIMEDOUT
1756 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1760 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1761 if ((bp->b_flags & B_PAGING) == 0) {
1763 bp->b_flags &= ~B_DONE;
1765 if ((error == EINTR || error == ETIMEDOUT) &&
1766 (bp->b_flags & B_ASYNC) == 0)
1767 bp->b_flags |= B_EINTR;
1771 bp->b_ioflags |= BIO_ERROR;
1772 bp->b_flags |= B_INVAL;
1773 bp->b_error = np->n_error = error;
1774 mtx_lock(&np->n_mtx);
1775 np->n_flag |= NWRITEERR;
1776 np->n_attrstamp = 0;
1777 mtx_unlock(&np->n_mtx);
1779 bp->b_dirtyoff = bp->b_dirtyend = 0;
1787 bp->b_resid = uiop->uio_resid;
1789 ncl_clearcommit(vp->v_mount);
1795 * Used to aid in handling ftruncate() operations on the NFS client side.
1796 * Truncation creates a number of special problems for NFS. We have to
1797 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1798 * we have to properly handle VM pages or (potentially dirty) buffers
1799 * that straddle the truncation point.
1803 ncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1805 struct nfsnode *np = VTONFS(vp);
1807 int biosize = vp->v_mount->mnt_stat.f_iosize;
1810 mtx_lock(&np->n_mtx);
1813 mtx_unlock(&np->n_mtx);
1815 if (nsize < tsize) {
1821 * vtruncbuf() doesn't get the buffer overlapping the
1822 * truncation point. We may have a B_DELWRI and/or B_CACHE
1823 * buffer that now needs to be truncated.
1825 error = vtruncbuf(vp, cred, td, nsize, biosize);
1826 lbn = nsize / biosize;
1827 bufsize = nsize & (biosize - 1);
1828 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1831 if (bp->b_dirtyoff > bp->b_bcount)
1832 bp->b_dirtyoff = bp->b_bcount;
1833 if (bp->b_dirtyend > bp->b_bcount)
1834 bp->b_dirtyend = bp->b_bcount;
1835 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1838 vnode_pager_setsize(vp, nsize);