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;
453 if (uio->uio_rw != UIO_READ)
454 panic("ncl_read mode");
456 if (uio->uio_resid == 0)
458 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
462 mtx_lock(&nmp->nm_mtx);
463 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
464 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
465 mtx_unlock(&nmp->nm_mtx);
466 (void)ncl_fsinfo(nmp, vp, cred, td);
467 mtx_lock(&nmp->nm_mtx);
469 if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
470 (void) newnfs_iosize(nmp);
471 mtx_unlock(&nmp->nm_mtx);
473 if (vp->v_type != VDIR &&
474 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
477 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
478 /* No caching/ no readaheads. Just read data into the user buffer */
479 return ncl_readrpc(vp, uio, cred);
481 biosize = vp->v_mount->mnt_stat.f_iosize;
482 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
484 error = nfs_bioread_check_cons(vp, td, cred);
491 mtx_lock(&np->n_mtx);
493 mtx_unlock(&np->n_mtx);
495 switch (vp->v_type) {
497 NFSINCRGLOBAL(newnfsstats.biocache_reads);
498 lbn = uio->uio_offset / biosize;
499 on = uio->uio_offset & (biosize - 1);
502 * Start the read ahead(s), as required.
504 if (nmp->nm_readahead > 0) {
505 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
506 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
507 rabn = lbn + 1 + nra;
508 if (incore(&vp->v_bufobj, rabn) == NULL) {
509 rabp = nfs_getcacheblk(vp, rabn, biosize, td);
511 error = newnfs_sigintr(nmp, td);
517 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
518 rabp->b_flags |= B_ASYNC;
519 rabp->b_iocmd = BIO_READ;
520 vfs_busy_pages(rabp, 0);
521 if (ncl_asyncio(nmp, rabp, cred, td)) {
522 rabp->b_flags |= B_INVAL;
523 rabp->b_ioflags |= BIO_ERROR;
524 vfs_unbusy_pages(rabp);
535 /* Note that bcount is *not* DEV_BSIZE aligned. */
537 if ((off_t)lbn * biosize >= nsize) {
539 } else if ((off_t)(lbn + 1) * biosize > nsize) {
540 bcount = nsize - (off_t)lbn * biosize;
542 bp = nfs_getcacheblk(vp, lbn, bcount, td);
545 error = newnfs_sigintr(nmp, td);
546 return (error ? error : EINTR);
550 * If B_CACHE is not set, we must issue the read. If this
551 * fails, we return an error.
554 if ((bp->b_flags & B_CACHE) == 0) {
555 bp->b_iocmd = BIO_READ;
556 vfs_busy_pages(bp, 0);
557 error = ncl_doio(vp, bp, cred, td, 0);
565 * on is the offset into the current bp. Figure out how many
566 * bytes we can copy out of the bp. Note that bcount is
567 * NOT DEV_BSIZE aligned.
569 * Then figure out how many bytes we can copy into the uio.
574 n = min((unsigned)(bcount - on), uio->uio_resid);
577 NFSINCRGLOBAL(newnfsstats.biocache_readlinks);
578 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
580 error = newnfs_sigintr(nmp, td);
581 return (error ? error : EINTR);
583 if ((bp->b_flags & B_CACHE) == 0) {
584 bp->b_iocmd = BIO_READ;
585 vfs_busy_pages(bp, 0);
586 error = ncl_doio(vp, bp, cred, td, 0);
588 bp->b_ioflags |= BIO_ERROR;
593 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
597 NFSINCRGLOBAL(newnfsstats.biocache_readdirs);
598 if (np->n_direofoffset
599 && uio->uio_offset >= np->n_direofoffset) {
602 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
603 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
604 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
606 error = newnfs_sigintr(nmp, td);
607 return (error ? error : EINTR);
609 if ((bp->b_flags & B_CACHE) == 0) {
610 bp->b_iocmd = BIO_READ;
611 vfs_busy_pages(bp, 0);
612 error = ncl_doio(vp, bp, cred, td, 0);
616 while (error == NFSERR_BAD_COOKIE) {
618 error = ncl_vinvalbuf(vp, 0, td, 1);
620 * Yuck! The directory has been modified on the
621 * server. The only way to get the block is by
622 * reading from the beginning to get all the
625 * Leave the last bp intact unless there is an error.
626 * Loop back up to the while if the error is another
627 * NFSERR_BAD_COOKIE (double yuch!).
629 for (i = 0; i <= lbn && !error; i++) {
630 if (np->n_direofoffset
631 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
633 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
635 error = newnfs_sigintr(nmp, td);
636 return (error ? error : EINTR);
638 if ((bp->b_flags & B_CACHE) == 0) {
639 bp->b_iocmd = BIO_READ;
640 vfs_busy_pages(bp, 0);
641 error = ncl_doio(vp, bp, cred, td, 0);
643 * no error + B_INVAL == directory EOF,
646 if (error == 0 && (bp->b_flags & B_INVAL))
650 * An error will throw away the block and the
651 * for loop will break out. If no error and this
652 * is not the block we want, we throw away the
653 * block and go for the next one via the for loop.
655 if (error || i < lbn)
660 * The above while is repeated if we hit another cookie
661 * error. If we hit an error and it wasn't a cookie error,
669 * If not eof and read aheads are enabled, start one.
670 * (You need the current block first, so that you have the
671 * directory offset cookie of the next block.)
673 if (nmp->nm_readahead > 0 &&
674 (bp->b_flags & B_INVAL) == 0 &&
675 (np->n_direofoffset == 0 ||
676 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
677 incore(&vp->v_bufobj, lbn + 1) == NULL) {
678 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
680 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
681 rabp->b_flags |= B_ASYNC;
682 rabp->b_iocmd = BIO_READ;
683 vfs_busy_pages(rabp, 0);
684 if (ncl_asyncio(nmp, rabp, cred, td)) {
685 rabp->b_flags |= B_INVAL;
686 rabp->b_ioflags |= BIO_ERROR;
687 vfs_unbusy_pages(rabp);
696 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
697 * chopped for the EOF condition, we cannot tell how large
698 * NFS directories are going to be until we hit EOF. So
699 * an NFS directory buffer is *not* chopped to its EOF. Now,
700 * it just so happens that b_resid will effectively chop it
701 * to EOF. *BUT* this information is lost if the buffer goes
702 * away and is reconstituted into a B_CACHE state ( due to
703 * being VMIO ) later. So we keep track of the directory eof
704 * in np->n_direofoffset and chop it off as an extra step
707 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
708 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
709 n = np->n_direofoffset - uio->uio_offset;
712 ncl_printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
718 error = uiomove(bp->b_data + on, (int)n, uio);
720 if (vp->v_type == VLNK)
724 } while (error == 0 && uio->uio_resid > 0 && n > 0);
729 * The NFS write path cannot handle iovecs with len > 1. So we need to
730 * break up iovecs accordingly (restricting them to wsize).
731 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
732 * For the ASYNC case, 2 copies are needed. The first a copy from the
733 * user buffer to a staging buffer and then a second copy from the staging
734 * buffer to mbufs. This can be optimized by copying from the user buffer
735 * directly into mbufs and passing the chain down, but that requires a
736 * fair amount of re-working of the relevant codepaths (and can be done
740 nfs_directio_write(vp, uiop, cred, ioflag)
747 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
748 struct thread *td = uiop->uio_td;
752 mtx_lock(&nmp->nm_mtx);
753 wsize = nmp->nm_wsize;
754 mtx_unlock(&nmp->nm_mtx);
755 if (ioflag & IO_SYNC) {
756 int iomode, must_commit;
760 while (uiop->uio_resid > 0) {
761 size = min(uiop->uio_resid, wsize);
762 size = min(uiop->uio_iov->iov_len, size);
763 iov.iov_base = uiop->uio_iov->iov_base;
767 uio.uio_offset = uiop->uio_offset;
768 uio.uio_resid = size;
769 uio.uio_segflg = UIO_USERSPACE;
770 uio.uio_rw = UIO_WRITE;
772 iomode = NFSWRITE_FILESYNC;
773 error = ncl_writerpc(vp, &uio, cred, &iomode,
775 KASSERT((must_commit == 0),
776 ("ncl_directio_write: Did not commit write"));
779 uiop->uio_offset += size;
780 uiop->uio_resid -= size;
781 if (uiop->uio_iov->iov_len <= size) {
785 uiop->uio_iov->iov_base =
786 (char *)uiop->uio_iov->iov_base + size;
787 uiop->uio_iov->iov_len -= size;
796 * Break up the write into blocksize chunks and hand these
797 * over to nfsiod's for write back.
798 * Unfortunately, this incurs a copy of the data. Since
799 * the user could modify the buffer before the write is
802 * The obvious optimization here is that one of the 2 copies
803 * in the async write path can be eliminated by copying the
804 * data here directly into mbufs and passing the mbuf chain
805 * down. But that will require a fair amount of re-working
806 * of the code and can be done if there's enough interest
807 * in NFS directio access.
809 while (uiop->uio_resid > 0) {
810 size = min(uiop->uio_resid, wsize);
811 size = min(uiop->uio_iov->iov_len, size);
812 bp = getpbuf(&ncl_pbuf_freecnt);
813 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
814 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
815 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
816 t_iov->iov_len = size;
817 t_uio->uio_iov = t_iov;
818 t_uio->uio_iovcnt = 1;
819 t_uio->uio_offset = uiop->uio_offset;
820 t_uio->uio_resid = size;
821 t_uio->uio_segflg = UIO_SYSSPACE;
822 t_uio->uio_rw = UIO_WRITE;
824 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base, size);
825 bp->b_flags |= B_DIRECT;
826 bp->b_iocmd = BIO_WRITE;
827 if (cred != NOCRED) {
831 bp->b_wcred = NOCRED;
832 bp->b_caller1 = (void *)t_uio;
834 error = ncl_asyncio(nmp, bp, NOCRED, td);
836 free(t_iov->iov_base, M_NFSDIRECTIO);
837 free(t_iov, M_NFSDIRECTIO);
838 free(t_uio, M_NFSDIRECTIO);
840 relpbuf(bp, &ncl_pbuf_freecnt);
845 uiop->uio_offset += size;
846 uiop->uio_resid -= size;
847 if (uiop->uio_iov->iov_len <= size) {
851 uiop->uio_iov->iov_base =
852 (char *)uiop->uio_iov->iov_base + size;
853 uiop->uio_iov->iov_len -= size;
861 * Vnode op for write using bio
864 ncl_write(struct vop_write_args *ap)
867 struct uio *uio = ap->a_uio;
868 struct thread *td = uio->uio_td;
869 struct vnode *vp = ap->a_vp;
870 struct nfsnode *np = VTONFS(vp);
871 struct ucred *cred = ap->a_cred;
872 int ioflag = ap->a_ioflag;
875 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
878 int n, on, error = 0;
879 struct proc *p = td?td->td_proc:NULL;
882 if (uio->uio_rw != UIO_WRITE)
883 panic("ncl_write mode");
884 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread)
885 panic("ncl_write proc");
887 if (vp->v_type != VREG)
889 mtx_lock(&np->n_mtx);
890 if (np->n_flag & NWRITEERR) {
891 np->n_flag &= ~NWRITEERR;
892 mtx_unlock(&np->n_mtx);
893 return (np->n_error);
895 mtx_unlock(&np->n_mtx);
896 mtx_lock(&nmp->nm_mtx);
897 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
898 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
899 mtx_unlock(&nmp->nm_mtx);
900 (void)ncl_fsinfo(nmp, vp, cred, td);
901 mtx_lock(&nmp->nm_mtx);
903 if (nmp->nm_wsize == 0)
904 (void) newnfs_iosize(nmp);
905 mtx_unlock(&nmp->nm_mtx);
908 * Synchronously flush pending buffers if we are in synchronous
909 * mode or if we are appending.
911 if (ioflag & (IO_APPEND | IO_SYNC)) {
912 mtx_lock(&np->n_mtx);
913 if (np->n_flag & NMODIFIED) {
914 mtx_unlock(&np->n_mtx);
915 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
917 * Require non-blocking, synchronous writes to
918 * dirty files to inform the program it needs
919 * to fsync(2) explicitly.
921 if (ioflag & IO_NDELAY)
926 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
930 mtx_unlock(&np->n_mtx);
934 * If IO_APPEND then load uio_offset. We restart here if we cannot
935 * get the append lock.
937 if (ioflag & IO_APPEND) {
939 error = VOP_GETATTR(vp, &vattr, cred);
942 mtx_lock(&np->n_mtx);
943 uio->uio_offset = np->n_size;
944 mtx_unlock(&np->n_mtx);
947 if (uio->uio_offset < 0)
949 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
951 if (uio->uio_resid == 0)
954 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
955 return nfs_directio_write(vp, uio, cred, ioflag);
958 * Maybe this should be above the vnode op call, but so long as
959 * file servers have no limits, i don't think it matters
963 if (uio->uio_offset + uio->uio_resid >
964 lim_cur(p, RLIMIT_FSIZE)) {
972 biosize = vp->v_mount->mnt_stat.f_iosize;
974 * Find all of this file's B_NEEDCOMMIT buffers. If our writes
975 * would exceed the local maximum per-file write commit size when
976 * combined with those, we must decide whether to flush,
977 * go synchronous, or return error. We don't bother checking
978 * IO_UNIT -- we just make all writes atomic anyway, as there's
979 * no point optimizing for something that really won't ever happen.
981 if (!(ioflag & IO_SYNC)) {
984 mtx_lock(&np->n_mtx);
986 mtx_unlock(&np->n_mtx);
988 if (nmp->nm_wcommitsize < uio->uio_resid) {
990 * If this request could not possibly be completed
991 * without exceeding the maximum outstanding write
992 * commit size, see if we can convert it into a
993 * synchronous write operation.
995 if (ioflag & IO_NDELAY)
998 if (nflag & NMODIFIED)
1000 } else if (nflag & NMODIFIED) {
1001 int wouldcommit = 0;
1002 BO_LOCK(&vp->v_bufobj);
1003 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
1004 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
1006 if (bp->b_flags & B_NEEDCOMMIT)
1007 wouldcommit += bp->b_bcount;
1010 BO_UNLOCK(&vp->v_bufobj);
1012 * Since we're not operating synchronously and
1013 * bypassing the buffer cache, we are in a commit
1014 * and holding all of these buffers whether
1015 * transmitted or not. If not limited, this
1016 * will lead to the buffer cache deadlocking,
1017 * as no one else can flush our uncommitted buffers.
1019 wouldcommit += uio->uio_resid;
1021 * If we would initially exceed the maximum
1022 * outstanding write commit size, flush and restart.
1024 if (wouldcommit > nmp->nm_wcommitsize)
1028 goto flush_and_restart;
1032 NFSINCRGLOBAL(newnfsstats.biocache_writes);
1033 lbn = uio->uio_offset / biosize;
1034 on = uio->uio_offset & (biosize-1);
1035 n = min((unsigned)(biosize - on), uio->uio_resid);
1038 * Handle direct append and file extension cases, calculate
1039 * unaligned buffer size.
1041 mtx_lock(&np->n_mtx);
1042 if (uio->uio_offset == np->n_size && n) {
1043 mtx_unlock(&np->n_mtx);
1045 * Get the buffer (in its pre-append state to maintain
1046 * B_CACHE if it was previously set). Resize the
1047 * nfsnode after we have locked the buffer to prevent
1048 * readers from reading garbage.
1051 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1056 mtx_lock(&np->n_mtx);
1057 np->n_size = uio->uio_offset + n;
1058 np->n_flag |= NMODIFIED;
1059 vnode_pager_setsize(vp, np->n_size);
1060 mtx_unlock(&np->n_mtx);
1062 save = bp->b_flags & B_CACHE;
1064 allocbuf(bp, bcount);
1065 bp->b_flags |= save;
1069 * Obtain the locked cache block first, and then
1070 * adjust the file's size as appropriate.
1073 if ((off_t)lbn * biosize + bcount < np->n_size) {
1074 if ((off_t)(lbn + 1) * biosize < np->n_size)
1077 bcount = np->n_size - (off_t)lbn * biosize;
1079 mtx_unlock(&np->n_mtx);
1080 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1081 mtx_lock(&np->n_mtx);
1082 if (uio->uio_offset + n > np->n_size) {
1083 np->n_size = uio->uio_offset + n;
1084 np->n_flag |= NMODIFIED;
1085 vnode_pager_setsize(vp, np->n_size);
1087 mtx_unlock(&np->n_mtx);
1091 error = newnfs_sigintr(nmp, td);
1098 * Issue a READ if B_CACHE is not set. In special-append
1099 * mode, B_CACHE is based on the buffer prior to the write
1100 * op and is typically set, avoiding the read. If a read
1101 * is required in special append mode, the server will
1102 * probably send us a short-read since we extended the file
1103 * on our end, resulting in b_resid == 0 and, thusly,
1104 * B_CACHE getting set.
1106 * We can also avoid issuing the read if the write covers
1107 * the entire buffer. We have to make sure the buffer state
1108 * is reasonable in this case since we will not be initiating
1109 * I/O. See the comments in kern/vfs_bio.c's getblk() for
1112 * B_CACHE may also be set due to the buffer being cached
1116 if (on == 0 && n == bcount) {
1117 bp->b_flags |= B_CACHE;
1118 bp->b_flags &= ~B_INVAL;
1119 bp->b_ioflags &= ~BIO_ERROR;
1122 if ((bp->b_flags & B_CACHE) == 0) {
1123 bp->b_iocmd = BIO_READ;
1124 vfs_busy_pages(bp, 0);
1125 error = ncl_doio(vp, bp, cred, td, 0);
1131 if (bp->b_wcred == NOCRED)
1132 bp->b_wcred = crhold(cred);
1133 mtx_lock(&np->n_mtx);
1134 np->n_flag |= NMODIFIED;
1135 mtx_unlock(&np->n_mtx);
1138 * If dirtyend exceeds file size, chop it down. This should
1139 * not normally occur but there is an append race where it
1140 * might occur XXX, so we log it.
1142 * If the chopping creates a reverse-indexed or degenerate
1143 * situation with dirtyoff/end, we 0 both of them.
1146 if (bp->b_dirtyend > bcount) {
1147 ncl_printf("NFS append race @%lx:%d\n",
1148 (long)bp->b_blkno * DEV_BSIZE,
1149 bp->b_dirtyend - bcount);
1150 bp->b_dirtyend = bcount;
1153 if (bp->b_dirtyoff >= bp->b_dirtyend)
1154 bp->b_dirtyoff = bp->b_dirtyend = 0;
1157 * If the new write will leave a contiguous dirty
1158 * area, just update the b_dirtyoff and b_dirtyend,
1159 * otherwise force a write rpc of the old dirty area.
1161 * While it is possible to merge discontiguous writes due to
1162 * our having a B_CACHE buffer ( and thus valid read data
1163 * for the hole), we don't because it could lead to
1164 * significant cache coherency problems with multiple clients,
1165 * especially if locking is implemented later on.
1167 * as an optimization we could theoretically maintain
1168 * a linked list of discontinuous areas, but we would still
1169 * have to commit them separately so there isn't much
1170 * advantage to it except perhaps a bit of asynchronization.
1173 if (bp->b_dirtyend > 0 &&
1174 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1175 if (bwrite(bp) == EINTR) {
1182 error = uiomove((char *)bp->b_data + on, n, uio);
1185 * Since this block is being modified, it must be written
1186 * again and not just committed. Since write clustering does
1187 * not work for the stage 1 data write, only the stage 2
1188 * commit rpc, we have to clear B_CLUSTEROK as well.
1190 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1193 bp->b_ioflags |= BIO_ERROR;
1199 * Only update dirtyoff/dirtyend if not a degenerate
1203 if (bp->b_dirtyend > 0) {
1204 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1205 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1207 bp->b_dirtyoff = on;
1208 bp->b_dirtyend = on + n;
1210 vfs_bio_set_valid(bp, on, n);
1214 * If IO_SYNC do bwrite().
1216 * IO_INVAL appears to be unused. The idea appears to be
1217 * to turn off caching in this case. Very odd. XXX
1219 if ((ioflag & IO_SYNC)) {
1220 if (ioflag & IO_INVAL)
1221 bp->b_flags |= B_NOCACHE;
1225 } else if ((n + on) == biosize) {
1226 bp->b_flags |= B_ASYNC;
1227 (void) ncl_writebp(bp, 0, NULL);
1231 } while (uio->uio_resid > 0 && n > 0);
1237 * Get an nfs cache block.
1239 * Allocate a new one if the block isn't currently in the cache
1240 * and return the block marked busy. If the calling process is
1241 * interrupted by a signal for an interruptible mount point, return
1244 * The caller must carefully deal with the possible B_INVAL state of
1245 * the buffer. ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
1246 * indirectly), so synchronous reads can be issued without worrying about
1247 * the B_INVAL state. We have to be a little more careful when dealing
1248 * with writes (see comments in nfs_write()) when extending a file past
1252 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1256 struct nfsmount *nmp;
1261 if (nmp->nm_flag & NFSMNT_INT) {
1264 newnfs_set_sigmask(td, &oldset);
1265 bp = getblk(vp, bn, size, NFS_PCATCH, 0, 0);
1266 newnfs_restore_sigmask(td, &oldset);
1267 while (bp == NULL) {
1268 if (newnfs_sigintr(nmp, td))
1270 bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1273 bp = getblk(vp, bn, size, 0, 0, 0);
1276 if (vp->v_type == VREG) {
1279 biosize = mp->mnt_stat.f_iosize;
1280 bp->b_blkno = bn * (biosize / DEV_BSIZE);
1286 * Flush and invalidate all dirty buffers. If another process is already
1287 * doing the flush, just wait for completion.
1290 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1292 struct nfsnode *np = VTONFS(vp);
1293 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1294 int error = 0, slpflag, slptimeo;
1297 ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
1299 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1301 if ((nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF))
1304 slpflag = NFS_PCATCH;
1311 old_lock = ncl_upgrade_vnlock(vp);
1312 if (vp->v_iflag & VI_DOOMED) {
1314 * Since vgonel() uses the generic vinvalbuf() to flush
1315 * dirty buffers and it does not call this function, it
1316 * is safe to just return OK when VI_DOOMED is set.
1318 ncl_downgrade_vnlock(vp, old_lock);
1323 * Now, flush as required.
1325 if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) {
1326 VM_OBJECT_LOCK(vp->v_bufobj.bo_object);
1327 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1328 VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object);
1330 * If the page clean was interrupted, fail the invalidation.
1331 * Not doing so, we run the risk of losing dirty pages in the
1332 * vinvalbuf() call below.
1334 if (intrflg && (error = newnfs_sigintr(nmp, td)))
1338 error = vinvalbuf(vp, flags, slpflag, 0);
1340 if (intrflg && (error = newnfs_sigintr(nmp, td)))
1342 error = vinvalbuf(vp, flags, 0, slptimeo);
1344 mtx_lock(&np->n_mtx);
1345 if (np->n_directio_asyncwr == 0)
1346 np->n_flag &= ~NMODIFIED;
1347 mtx_unlock(&np->n_mtx);
1349 ncl_downgrade_vnlock(vp, old_lock);
1354 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1355 * This is mainly to avoid queueing async I/O requests when the nfsiods
1356 * are all hung on a dead server.
1358 * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1359 * is eventually dequeued by the async daemon, ncl_doio() *will*.
1362 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1371 * Unless iothreadcnt is set > 0, don't bother with async I/O
1372 * threads. For LAN environments, they don't buy any significant
1373 * performance improvement that you can't get with large block
1376 if (nmp->nm_readahead == 0)
1380 * Commits are usually short and sweet so lets save some cpu and
1381 * leave the async daemons for more important rpc's (such as reads
1384 mtx_lock(&ncl_iod_mutex);
1385 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1386 (nmp->nm_bufqiods > ncl_numasync / 2)) {
1387 mtx_unlock(&ncl_iod_mutex);
1391 if (nmp->nm_flag & NFSMNT_INT)
1392 slpflag = NFS_PCATCH;
1396 * Find a free iod to process this request.
1398 for (iod = 0; iod < ncl_numasync; iod++)
1399 if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1405 * Try to create one if none are free.
1408 iod = ncl_nfsiodnew(1);
1415 * Found one, so wake it up and tell it which
1418 NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1420 ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1421 ncl_iodmount[iod] = nmp;
1423 wakeup(&ncl_iodwant[iod]);
1427 * If none are free, we may already have an iod working on this mount
1428 * point. If so, it will process our request.
1431 if (nmp->nm_bufqiods > 0) {
1433 ("ncl_asyncio: %d iods are already processing mount %p\n",
1434 nmp->nm_bufqiods, nmp));
1440 * If we have an iod which can process the request, then queue
1445 * Ensure that the queue never grows too large. We still want
1446 * to asynchronize so we block rather then return EIO.
1448 while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1450 ("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1451 nmp->nm_bufqwant = TRUE;
1452 error = newnfs_msleep(td, &nmp->nm_bufq,
1453 &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1456 error2 = newnfs_sigintr(nmp, td);
1458 mtx_unlock(&ncl_iod_mutex);
1461 if (slpflag == NFS_PCATCH) {
1467 * We might have lost our iod while sleeping,
1468 * so check and loop if nescessary.
1470 if (nmp->nm_bufqiods == 0) {
1472 ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1477 /* We might have lost our nfsiod */
1478 if (nmp->nm_bufqiods == 0) {
1480 ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1484 if (bp->b_iocmd == BIO_READ) {
1485 if (bp->b_rcred == NOCRED && cred != NOCRED)
1486 bp->b_rcred = crhold(cred);
1488 if (bp->b_wcred == NOCRED && cred != NOCRED)
1489 bp->b_wcred = crhold(cred);
1492 if (bp->b_flags & B_REMFREE)
1495 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1497 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1498 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx);
1499 VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1500 VTONFS(bp->b_vp)->n_directio_asyncwr++;
1501 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx);
1503 mtx_unlock(&ncl_iod_mutex);
1507 mtx_unlock(&ncl_iod_mutex);
1510 * All the iods are busy on other mounts, so return EIO to
1511 * force the caller to process the i/o synchronously.
1513 NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1518 ncl_doio_directwrite(struct buf *bp)
1520 int iomode, must_commit;
1521 struct uio *uiop = (struct uio *)bp->b_caller1;
1522 char *iov_base = uiop->uio_iov->iov_base;
1524 iomode = NFSWRITE_FILESYNC;
1525 uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1526 ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0);
1527 KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write"));
1528 free(iov_base, M_NFSDIRECTIO);
1529 free(uiop->uio_iov, M_NFSDIRECTIO);
1530 free(uiop, M_NFSDIRECTIO);
1531 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1532 struct nfsnode *np = VTONFS(bp->b_vp);
1533 mtx_lock(&np->n_mtx);
1534 np->n_directio_asyncwr--;
1535 if (np->n_directio_asyncwr == 0) {
1536 np->n_flag &= ~NMODIFIED;
1537 if ((np->n_flag & NFSYNCWAIT)) {
1538 np->n_flag &= ~NFSYNCWAIT;
1539 wakeup((caddr_t)&np->n_directio_asyncwr);
1542 mtx_unlock(&np->n_mtx);
1545 relpbuf(bp, &ncl_pbuf_freecnt);
1549 * Do an I/O operation to/from a cache block. This may be called
1550 * synchronously or from an nfsiod.
1553 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1554 int called_from_strategy)
1558 struct nfsmount *nmp;
1559 int error = 0, iomode, must_commit = 0;
1562 struct proc *p = td ? td->td_proc : NULL;
1566 nmp = VFSTONFS(vp->v_mount);
1568 uiop->uio_iov = &io;
1569 uiop->uio_iovcnt = 1;
1570 uiop->uio_segflg = UIO_SYSSPACE;
1574 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We
1575 * do this here so we do not have to do it in all the code that
1578 bp->b_flags &= ~B_INVAL;
1579 bp->b_ioflags &= ~BIO_ERROR;
1581 KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1582 iocmd = bp->b_iocmd;
1583 if (iocmd == BIO_READ) {
1584 io.iov_len = uiop->uio_resid = bp->b_bcount;
1585 io.iov_base = bp->b_data;
1586 uiop->uio_rw = UIO_READ;
1588 switch (vp->v_type) {
1590 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1591 NFSINCRGLOBAL(newnfsstats.read_bios);
1592 error = ncl_readrpc(vp, uiop, cr);
1595 if (uiop->uio_resid) {
1597 * If we had a short read with no error, we must have
1598 * hit a file hole. We should zero-fill the remainder.
1599 * This can also occur if the server hits the file EOF.
1601 * Holes used to be able to occur due to pending
1602 * writes, but that is not possible any longer.
1604 int nread = bp->b_bcount - uiop->uio_resid;
1605 int left = uiop->uio_resid;
1608 bzero((char *)bp->b_data + nread, left);
1609 uiop->uio_resid = 0;
1612 /* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1613 if (p && (vp->v_vflag & VV_TEXT)) {
1614 mtx_lock(&np->n_mtx);
1615 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1616 mtx_unlock(&np->n_mtx);
1618 killproc(p, "text file modification");
1621 mtx_unlock(&np->n_mtx);
1625 uiop->uio_offset = (off_t)0;
1626 NFSINCRGLOBAL(newnfsstats.readlink_bios);
1627 error = ncl_readlinkrpc(vp, uiop, cr);
1630 NFSINCRGLOBAL(newnfsstats.readdir_bios);
1631 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1632 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1633 error = ncl_readdirplusrpc(vp, uiop, cr, td);
1634 if (error == NFSERR_NOTSUPP)
1635 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1637 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1638 error = ncl_readdirrpc(vp, uiop, cr, td);
1640 * end-of-directory sets B_INVAL but does not generate an
1643 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1644 bp->b_flags |= B_INVAL;
1647 ncl_printf("ncl_doio: type %x unexpected\n", vp->v_type);
1651 bp->b_ioflags |= BIO_ERROR;
1652 bp->b_error = error;
1656 * If we only need to commit, try to commit
1658 if (bp->b_flags & B_NEEDCOMMIT) {
1662 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1663 retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1666 bp->b_dirtyoff = bp->b_dirtyend = 0;
1667 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1672 if (retv == NFSERR_STALEWRITEVERF) {
1673 ncl_clearcommit(vp->v_mount);
1678 * Setup for actual write
1680 mtx_lock(&np->n_mtx);
1681 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1682 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1683 mtx_unlock(&np->n_mtx);
1685 if (bp->b_dirtyend > bp->b_dirtyoff) {
1686 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1688 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1690 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1691 uiop->uio_rw = UIO_WRITE;
1692 NFSINCRGLOBAL(newnfsstats.write_bios);
1694 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1695 iomode = NFSWRITE_UNSTABLE;
1697 iomode = NFSWRITE_FILESYNC;
1699 error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1700 called_from_strategy);
1703 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1704 * to cluster the buffers needing commit. This will allow
1705 * the system to submit a single commit rpc for the whole
1706 * cluster. We can do this even if the buffer is not 100%
1707 * dirty (relative to the NFS blocksize), so we optimize the
1708 * append-to-file-case.
1710 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1711 * cleared because write clustering only works for commit
1712 * rpc's, not for the data portion of the write).
1715 if (!error && iomode == NFSWRITE_UNSTABLE) {
1716 bp->b_flags |= B_NEEDCOMMIT;
1717 if (bp->b_dirtyoff == 0
1718 && bp->b_dirtyend == bp->b_bcount)
1719 bp->b_flags |= B_CLUSTEROK;
1721 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1725 * For an interrupted write, the buffer is still valid
1726 * and the write hasn't been pushed to the server yet,
1727 * so we can't set BIO_ERROR and report the interruption
1728 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1729 * is not relevant, so the rpc attempt is essentially
1730 * a noop. For the case of a V3 write rpc not being
1731 * committed to stable storage, the block is still
1732 * dirty and requires either a commit rpc or another
1733 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1734 * the block is reused. This is indicated by setting
1735 * the B_DELWRI and B_NEEDCOMMIT flags.
1737 * EIO is returned by ncl_writerpc() to indicate a recoverable
1738 * write error and is handled as above, except that
1739 * B_EINTR isn't set. One cause of this is a stale stateid
1740 * error for the RPC that indicates recovery is required,
1741 * when called with called_from_strategy != 0.
1743 * If the buffer is marked B_PAGING, it does not reside on
1744 * the vp's paging queues so we cannot call bdirty(). The
1745 * bp in this case is not an NFS cache block so we should
1748 * The logic below breaks up errors into recoverable and
1749 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1750 * and keep the buffer around for potential write retries.
1751 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1752 * and save the error in the nfsnode. This is less than ideal
1753 * but necessary. Keeping such buffers around could potentially
1754 * cause buffer exhaustion eventually (they can never be written
1755 * out, so will get constantly be re-dirtied). It also causes
1756 * all sorts of vfs panics. For non-recoverable write errors,
1757 * also invalidate the attrcache, so we'll be forced to go over
1758 * the wire for this object, returning an error to user on next
1759 * call (most of the time).
1761 if (error == EINTR || error == EIO || error == ETIMEDOUT
1762 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1766 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1767 if ((bp->b_flags & B_PAGING) == 0) {
1769 bp->b_flags &= ~B_DONE;
1771 if ((error == EINTR || error == ETIMEDOUT) &&
1772 (bp->b_flags & B_ASYNC) == 0)
1773 bp->b_flags |= B_EINTR;
1777 bp->b_ioflags |= BIO_ERROR;
1778 bp->b_flags |= B_INVAL;
1779 bp->b_error = np->n_error = error;
1780 mtx_lock(&np->n_mtx);
1781 np->n_flag |= NWRITEERR;
1782 np->n_attrstamp = 0;
1783 mtx_unlock(&np->n_mtx);
1785 bp->b_dirtyoff = bp->b_dirtyend = 0;
1793 bp->b_resid = uiop->uio_resid;
1795 ncl_clearcommit(vp->v_mount);
1801 * Used to aid in handling ftruncate() operations on the NFS client side.
1802 * Truncation creates a number of special problems for NFS. We have to
1803 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1804 * we have to properly handle VM pages or (potentially dirty) buffers
1805 * that straddle the truncation point.
1809 ncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1811 struct nfsnode *np = VTONFS(vp);
1813 int biosize = vp->v_mount->mnt_stat.f_iosize;
1816 mtx_lock(&np->n_mtx);
1819 mtx_unlock(&np->n_mtx);
1821 if (nsize < tsize) {
1827 * vtruncbuf() doesn't get the buffer overlapping the
1828 * truncation point. We may have a B_DELWRI and/or B_CACHE
1829 * buffer that now needs to be truncated.
1831 error = vtruncbuf(vp, cred, td, nsize, biosize);
1832 lbn = nsize / biosize;
1833 bufsize = nsize & (biosize - 1);
1834 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1837 if (bp->b_dirtyoff > bp->b_bcount)
1838 bp->b_dirtyoff = bp->b_bcount;
1839 if (bp->b_dirtyend > bp->b_bcount)
1840 bp->b_dirtyend = bp->b_bcount;
1841 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1844 vnode_pager_setsize(vp, nsize);