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>
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 nfsstats newnfsstats;
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.
82 ncl_getpages(struct vop_getpages_args *ap)
84 int i, error, nextoff, size, toff, count, npages;
99 td = curthread; /* XXX */
100 cred = curthread->td_ucred; /* XXX */
101 nmp = VFSTONFS(vp->v_mount);
105 if ((object = vp->v_object) == NULL) {
106 ncl_printf("nfs_getpages: called with non-merged cache vnode??\n");
107 return (VM_PAGER_ERROR);
110 if (newnfs_directio_enable && !newnfs_directio_allow_mmap) {
111 mtx_lock(&np->n_mtx);
112 if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
113 mtx_unlock(&np->n_mtx);
114 ncl_printf("nfs_getpages: called on non-cacheable vnode??\n");
115 return (VM_PAGER_ERROR);
117 mtx_unlock(&np->n_mtx);
120 mtx_lock(&nmp->nm_mtx);
121 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
122 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
123 mtx_unlock(&nmp->nm_mtx);
124 /* We'll never get here for v4, because we always have fsinfo */
125 (void)ncl_fsinfo(nmp, vp, cred, td);
127 mtx_unlock(&nmp->nm_mtx);
129 npages = btoc(count);
132 * If the requested page is partially valid, just return it and
133 * allow the pager to zero-out the blanks. Partially valid pages
134 * can only occur at the file EOF.
136 VM_OBJECT_WLOCK(object);
137 if (pages[ap->a_reqpage]->valid != 0) {
138 for (i = 0; i < npages; ++i) {
139 if (i != ap->a_reqpage) {
140 vm_page_lock(pages[i]);
141 vm_page_free(pages[i]);
142 vm_page_unlock(pages[i]);
145 VM_OBJECT_WUNLOCK(object);
148 VM_OBJECT_WUNLOCK(object);
151 * We use only the kva address for the buffer, but this is extremely
152 * convienient and fast.
154 bp = getpbuf(&ncl_pbuf_freecnt);
156 kva = (vm_offset_t) bp->b_data;
157 pmap_qenter(kva, pages, npages);
158 PCPU_INC(cnt.v_vnodein);
159 PCPU_ADD(cnt.v_vnodepgsin, npages);
161 iov.iov_base = (caddr_t) kva;
165 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
166 uio.uio_resid = count;
167 uio.uio_segflg = UIO_SYSSPACE;
168 uio.uio_rw = UIO_READ;
171 error = ncl_readrpc(vp, &uio, cred);
172 pmap_qremove(kva, npages);
174 relpbuf(bp, &ncl_pbuf_freecnt);
176 if (error && (uio.uio_resid == count)) {
177 ncl_printf("nfs_getpages: error %d\n", error);
178 VM_OBJECT_WLOCK(object);
179 for (i = 0; i < npages; ++i) {
180 if (i != ap->a_reqpage) {
181 vm_page_lock(pages[i]);
182 vm_page_free(pages[i]);
183 vm_page_unlock(pages[i]);
186 VM_OBJECT_WUNLOCK(object);
187 return (VM_PAGER_ERROR);
191 * Calculate the number of bytes read and validate only that number
192 * of bytes. Note that due to pending writes, size may be 0. This
193 * does not mean that the remaining data is invalid!
196 size = count - uio.uio_resid;
197 VM_OBJECT_WLOCK(object);
198 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
200 nextoff = toff + PAGE_SIZE;
203 if (nextoff <= size) {
205 * Read operation filled an entire page
207 m->valid = VM_PAGE_BITS_ALL;
208 KASSERT(m->dirty == 0,
209 ("nfs_getpages: page %p is dirty", m));
210 } else if (size > toff) {
212 * Read operation filled a partial page.
215 vm_page_set_valid_range(m, 0, size - toff);
216 KASSERT(m->dirty == 0,
217 ("nfs_getpages: page %p is dirty", m));
220 * Read operation was short. If no error
221 * occured we may have hit a zero-fill
222 * section. We leave valid set to 0, and page
223 * is freed by vm_page_readahead_finish() if
224 * its index is not equal to requested, or
225 * page is zeroed and set valid by
226 * vm_pager_get_pages() for requested page.
230 if (i != ap->a_reqpage)
231 vm_page_readahead_finish(m);
233 VM_OBJECT_WUNLOCK(object);
238 * Vnode op for VM putpages.
241 ncl_putpages(struct vop_putpages_args *ap)
247 int iomode, must_commit, i, error, npages, count;
253 struct nfsmount *nmp;
259 td = curthread; /* XXX */
260 /* Set the cred to n_writecred for the write rpcs. */
261 if (np->n_writecred != NULL)
262 cred = crhold(np->n_writecred);
264 cred = crhold(curthread->td_ucred); /* XXX */
265 nmp = VFSTONFS(vp->v_mount);
268 rtvals = ap->a_rtvals;
269 npages = btoc(count);
270 offset = IDX_TO_OFF(pages[0]->pindex);
272 mtx_lock(&nmp->nm_mtx);
273 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
274 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
275 mtx_unlock(&nmp->nm_mtx);
276 (void)ncl_fsinfo(nmp, vp, cred, td);
278 mtx_unlock(&nmp->nm_mtx);
280 mtx_lock(&np->n_mtx);
281 if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
282 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
283 mtx_unlock(&np->n_mtx);
284 ncl_printf("ncl_putpages: called on noncache-able vnode??\n");
285 mtx_lock(&np->n_mtx);
288 for (i = 0; i < npages; i++)
289 rtvals[i] = VM_PAGER_ERROR;
292 * When putting pages, do not extend file past EOF.
294 if (offset + count > np->n_size) {
295 count = np->n_size - offset;
299 mtx_unlock(&np->n_mtx);
302 * We use only the kva address for the buffer, but this is extremely
303 * convienient and fast.
305 bp = getpbuf(&ncl_pbuf_freecnt);
307 kva = (vm_offset_t) bp->b_data;
308 pmap_qenter(kva, pages, npages);
309 PCPU_INC(cnt.v_vnodeout);
310 PCPU_ADD(cnt.v_vnodepgsout, count);
312 iov.iov_base = (caddr_t) kva;
316 uio.uio_offset = offset;
317 uio.uio_resid = count;
318 uio.uio_segflg = UIO_SYSSPACE;
319 uio.uio_rw = UIO_WRITE;
322 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
323 iomode = NFSWRITE_UNSTABLE;
325 iomode = NFSWRITE_FILESYNC;
327 error = ncl_writerpc(vp, &uio, cred, &iomode, &must_commit, 0);
330 pmap_qremove(kva, npages);
331 relpbuf(bp, &ncl_pbuf_freecnt);
333 if (error == 0 || !nfs_keep_dirty_on_error) {
334 vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid);
336 ncl_clearcommit(vp->v_mount);
342 * For nfs, cache consistency can only be maintained approximately.
343 * Although RFC1094 does not specify the criteria, the following is
344 * believed to be compatible with the reference port.
346 * If the file's modify time on the server has changed since the
347 * last read rpc or you have written to the file,
348 * you may have lost data cache consistency with the
349 * server, so flush all of the file's data out of the cache.
350 * Then force a getattr rpc to ensure that you have up to date
352 * NB: This implies that cache data can be read when up to
353 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
354 * attributes this could be forced by setting n_attrstamp to 0 before
355 * the VOP_GETATTR() call.
358 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
362 struct nfsnode *np = VTONFS(vp);
366 * Grab the exclusive lock before checking whether the cache is
368 * XXX - We can make this cheaper later (by acquiring cheaper locks).
369 * But for now, this suffices.
371 old_lock = ncl_upgrade_vnlock(vp);
372 if (vp->v_iflag & VI_DOOMED) {
373 ncl_downgrade_vnlock(vp, old_lock);
377 mtx_lock(&np->n_mtx);
378 if (np->n_flag & NMODIFIED) {
379 mtx_unlock(&np->n_mtx);
380 if (vp->v_type != VREG) {
381 if (vp->v_type != VDIR)
382 panic("nfs: bioread, not dir");
384 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
389 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
390 error = VOP_GETATTR(vp, &vattr, cred);
393 mtx_lock(&np->n_mtx);
394 np->n_mtime = vattr.va_mtime;
395 mtx_unlock(&np->n_mtx);
397 mtx_unlock(&np->n_mtx);
398 error = VOP_GETATTR(vp, &vattr, cred);
401 mtx_lock(&np->n_mtx);
402 if ((np->n_flag & NSIZECHANGED)
403 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
404 mtx_unlock(&np->n_mtx);
405 if (vp->v_type == VDIR)
407 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
410 mtx_lock(&np->n_mtx);
411 np->n_mtime = vattr.va_mtime;
412 np->n_flag &= ~NSIZECHANGED;
414 mtx_unlock(&np->n_mtx);
417 ncl_downgrade_vnlock(vp, old_lock);
422 * Vnode op for read using bio
425 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
427 struct nfsnode *np = VTONFS(vp);
429 struct buf *bp, *rabp;
431 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
435 int nra, error = 0, n = 0, on = 0;
438 KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
439 if (uio->uio_resid == 0)
441 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
445 mtx_lock(&nmp->nm_mtx);
446 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
447 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
448 mtx_unlock(&nmp->nm_mtx);
449 (void)ncl_fsinfo(nmp, vp, cred, td);
450 mtx_lock(&nmp->nm_mtx);
452 if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
453 (void) newnfs_iosize(nmp);
455 tmp_off = uio->uio_offset + uio->uio_resid;
456 if (vp->v_type != VDIR &&
457 (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) {
458 mtx_unlock(&nmp->nm_mtx);
461 mtx_unlock(&nmp->nm_mtx);
463 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
464 /* No caching/ no readaheads. Just read data into the user buffer */
465 return ncl_readrpc(vp, uio, cred);
467 biosize = vp->v_bufobj.bo_bsize;
468 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
470 error = nfs_bioread_check_cons(vp, td, cred);
477 mtx_lock(&np->n_mtx);
479 mtx_unlock(&np->n_mtx);
481 switch (vp->v_type) {
483 NFSINCRGLOBAL(newnfsstats.biocache_reads);
484 lbn = uio->uio_offset / biosize;
485 on = uio->uio_offset - (lbn * biosize);
488 * Start the read ahead(s), as required.
490 if (nmp->nm_readahead > 0) {
491 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
492 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
493 rabn = lbn + 1 + nra;
494 if (incore(&vp->v_bufobj, rabn) == NULL) {
495 rabp = nfs_getcacheblk(vp, rabn, biosize, td);
497 error = newnfs_sigintr(nmp, td);
498 return (error ? error : EINTR);
500 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
501 rabp->b_flags |= B_ASYNC;
502 rabp->b_iocmd = BIO_READ;
503 vfs_busy_pages(rabp, 0);
504 if (ncl_asyncio(nmp, rabp, cred, td)) {
505 rabp->b_flags |= B_INVAL;
506 rabp->b_ioflags |= BIO_ERROR;
507 vfs_unbusy_pages(rabp);
518 /* Note that bcount is *not* DEV_BSIZE aligned. */
520 if ((off_t)lbn * biosize >= nsize) {
522 } else if ((off_t)(lbn + 1) * biosize > nsize) {
523 bcount = nsize - (off_t)lbn * biosize;
525 bp = nfs_getcacheblk(vp, lbn, bcount, td);
528 error = newnfs_sigintr(nmp, td);
529 return (error ? error : EINTR);
533 * If B_CACHE is not set, we must issue the read. If this
534 * fails, we return an error.
537 if ((bp->b_flags & B_CACHE) == 0) {
538 bp->b_iocmd = BIO_READ;
539 vfs_busy_pages(bp, 0);
540 error = ncl_doio(vp, bp, cred, td, 0);
548 * on is the offset into the current bp. Figure out how many
549 * bytes we can copy out of the bp. Note that bcount is
550 * NOT DEV_BSIZE aligned.
552 * Then figure out how many bytes we can copy into the uio.
557 n = MIN((unsigned)(bcount - on), uio->uio_resid);
560 NFSINCRGLOBAL(newnfsstats.biocache_readlinks);
561 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
563 error = newnfs_sigintr(nmp, td);
564 return (error ? error : EINTR);
566 if ((bp->b_flags & B_CACHE) == 0) {
567 bp->b_iocmd = BIO_READ;
568 vfs_busy_pages(bp, 0);
569 error = ncl_doio(vp, bp, cred, td, 0);
571 bp->b_ioflags |= BIO_ERROR;
576 n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
580 NFSINCRGLOBAL(newnfsstats.biocache_readdirs);
581 if (np->n_direofoffset
582 && uio->uio_offset >= np->n_direofoffset) {
585 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
586 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
587 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
589 error = newnfs_sigintr(nmp, td);
590 return (error ? error : EINTR);
592 if ((bp->b_flags & B_CACHE) == 0) {
593 bp->b_iocmd = BIO_READ;
594 vfs_busy_pages(bp, 0);
595 error = ncl_doio(vp, bp, cred, td, 0);
599 while (error == NFSERR_BAD_COOKIE) {
601 error = ncl_vinvalbuf(vp, 0, td, 1);
603 * Yuck! The directory has been modified on the
604 * server. The only way to get the block is by
605 * reading from the beginning to get all the
608 * Leave the last bp intact unless there is an error.
609 * Loop back up to the while if the error is another
610 * NFSERR_BAD_COOKIE (double yuch!).
612 for (i = 0; i <= lbn && !error; i++) {
613 if (np->n_direofoffset
614 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
616 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
618 error = newnfs_sigintr(nmp, td);
619 return (error ? error : EINTR);
621 if ((bp->b_flags & B_CACHE) == 0) {
622 bp->b_iocmd = BIO_READ;
623 vfs_busy_pages(bp, 0);
624 error = ncl_doio(vp, bp, cred, td, 0);
626 * no error + B_INVAL == directory EOF,
629 if (error == 0 && (bp->b_flags & B_INVAL))
633 * An error will throw away the block and the
634 * for loop will break out. If no error and this
635 * is not the block we want, we throw away the
636 * block and go for the next one via the for loop.
638 if (error || i < lbn)
643 * The above while is repeated if we hit another cookie
644 * error. If we hit an error and it wasn't a cookie error,
652 * If not eof and read aheads are enabled, start one.
653 * (You need the current block first, so that you have the
654 * directory offset cookie of the next block.)
656 if (nmp->nm_readahead > 0 &&
657 (bp->b_flags & B_INVAL) == 0 &&
658 (np->n_direofoffset == 0 ||
659 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
660 incore(&vp->v_bufobj, lbn + 1) == NULL) {
661 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
663 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
664 rabp->b_flags |= B_ASYNC;
665 rabp->b_iocmd = BIO_READ;
666 vfs_busy_pages(rabp, 0);
667 if (ncl_asyncio(nmp, rabp, cred, td)) {
668 rabp->b_flags |= B_INVAL;
669 rabp->b_ioflags |= BIO_ERROR;
670 vfs_unbusy_pages(rabp);
679 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
680 * chopped for the EOF condition, we cannot tell how large
681 * NFS directories are going to be until we hit EOF. So
682 * an NFS directory buffer is *not* chopped to its EOF. Now,
683 * it just so happens that b_resid will effectively chop it
684 * to EOF. *BUT* this information is lost if the buffer goes
685 * away and is reconstituted into a B_CACHE state ( due to
686 * being VMIO ) later. So we keep track of the directory eof
687 * in np->n_direofoffset and chop it off as an extra step
690 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
691 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
692 n = np->n_direofoffset - uio->uio_offset;
695 ncl_printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
701 error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio);
703 if (vp->v_type == VLNK)
707 } while (error == 0 && uio->uio_resid > 0 && n > 0);
712 * The NFS write path cannot handle iovecs with len > 1. So we need to
713 * break up iovecs accordingly (restricting them to wsize).
714 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
715 * For the ASYNC case, 2 copies are needed. The first a copy from the
716 * user buffer to a staging buffer and then a second copy from the staging
717 * buffer to mbufs. This can be optimized by copying from the user buffer
718 * directly into mbufs and passing the chain down, but that requires a
719 * fair amount of re-working of the relevant codepaths (and can be done
723 nfs_directio_write(vp, uiop, cred, ioflag)
730 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
731 struct thread *td = uiop->uio_td;
735 mtx_lock(&nmp->nm_mtx);
736 wsize = nmp->nm_wsize;
737 mtx_unlock(&nmp->nm_mtx);
738 if (ioflag & IO_SYNC) {
739 int iomode, must_commit;
743 while (uiop->uio_resid > 0) {
744 size = MIN(uiop->uio_resid, wsize);
745 size = MIN(uiop->uio_iov->iov_len, size);
746 iov.iov_base = uiop->uio_iov->iov_base;
750 uio.uio_offset = uiop->uio_offset;
751 uio.uio_resid = size;
752 uio.uio_segflg = UIO_USERSPACE;
753 uio.uio_rw = UIO_WRITE;
755 iomode = NFSWRITE_FILESYNC;
756 error = ncl_writerpc(vp, &uio, cred, &iomode,
758 KASSERT((must_commit == 0),
759 ("ncl_directio_write: Did not commit write"));
762 uiop->uio_offset += size;
763 uiop->uio_resid -= size;
764 if (uiop->uio_iov->iov_len <= size) {
768 uiop->uio_iov->iov_base =
769 (char *)uiop->uio_iov->iov_base + size;
770 uiop->uio_iov->iov_len -= size;
779 * Break up the write into blocksize chunks and hand these
780 * over to nfsiod's for write back.
781 * Unfortunately, this incurs a copy of the data. Since
782 * the user could modify the buffer before the write is
785 * The obvious optimization here is that one of the 2 copies
786 * in the async write path can be eliminated by copying the
787 * data here directly into mbufs and passing the mbuf chain
788 * down. But that will require a fair amount of re-working
789 * of the code and can be done if there's enough interest
790 * in NFS directio access.
792 while (uiop->uio_resid > 0) {
793 size = MIN(uiop->uio_resid, wsize);
794 size = MIN(uiop->uio_iov->iov_len, size);
795 bp = getpbuf(&ncl_pbuf_freecnt);
796 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
797 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
798 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
799 t_iov->iov_len = size;
800 t_uio->uio_iov = t_iov;
801 t_uio->uio_iovcnt = 1;
802 t_uio->uio_offset = uiop->uio_offset;
803 t_uio->uio_resid = size;
804 t_uio->uio_segflg = UIO_SYSSPACE;
805 t_uio->uio_rw = UIO_WRITE;
807 KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
808 uiop->uio_segflg == UIO_SYSSPACE,
809 ("nfs_directio_write: Bad uio_segflg"));
810 if (uiop->uio_segflg == UIO_USERSPACE) {
811 error = copyin(uiop->uio_iov->iov_base,
812 t_iov->iov_base, size);
817 * UIO_SYSSPACE may never happen, but handle
818 * it just in case it does.
820 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
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);
834 free(t_iov->iov_base, M_NFSDIRECTIO);
835 free(t_iov, M_NFSDIRECTIO);
836 free(t_uio, M_NFSDIRECTIO);
838 relpbuf(bp, &ncl_pbuf_freecnt);
843 uiop->uio_offset += size;
844 uiop->uio_resid -= size;
845 if (uiop->uio_iov->iov_len <= size) {
849 uiop->uio_iov->iov_base =
850 (char *)uiop->uio_iov->iov_base + size;
851 uiop->uio_iov->iov_len -= size;
859 * Vnode op for write using bio
862 ncl_write(struct vop_write_args *ap)
865 struct uio *uio = ap->a_uio;
866 struct thread *td = uio->uio_td;
867 struct vnode *vp = ap->a_vp;
868 struct nfsnode *np = VTONFS(vp);
869 struct ucred *cred = ap->a_cred;
870 int ioflag = ap->a_ioflag;
873 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
875 int bcount, noncontig_write, obcount;
876 int bp_cached, n, on, error = 0, error1;
877 size_t orig_resid, local_resid;
878 off_t orig_size, tmp_off;
880 KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
881 KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
883 if (vp->v_type != VREG)
885 mtx_lock(&np->n_mtx);
886 if (np->n_flag & NWRITEERR) {
887 np->n_flag &= ~NWRITEERR;
888 mtx_unlock(&np->n_mtx);
889 return (np->n_error);
891 mtx_unlock(&np->n_mtx);
892 mtx_lock(&nmp->nm_mtx);
893 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
894 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
895 mtx_unlock(&nmp->nm_mtx);
896 (void)ncl_fsinfo(nmp, vp, cred, td);
897 mtx_lock(&nmp->nm_mtx);
899 if (nmp->nm_wsize == 0)
900 (void) newnfs_iosize(nmp);
901 mtx_unlock(&nmp->nm_mtx);
904 * Synchronously flush pending buffers if we are in synchronous
905 * mode or if we are appending.
907 if (ioflag & (IO_APPEND | IO_SYNC)) {
908 mtx_lock(&np->n_mtx);
909 if (np->n_flag & NMODIFIED) {
910 mtx_unlock(&np->n_mtx);
911 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
913 * Require non-blocking, synchronous writes to
914 * dirty files to inform the program it needs
915 * to fsync(2) explicitly.
917 if (ioflag & IO_NDELAY)
922 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
923 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
927 mtx_unlock(&np->n_mtx);
930 orig_resid = uio->uio_resid;
931 mtx_lock(&np->n_mtx);
932 orig_size = np->n_size;
933 mtx_unlock(&np->n_mtx);
936 * If IO_APPEND then load uio_offset. We restart here if we cannot
937 * get the append lock.
939 if (ioflag & IO_APPEND) {
941 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
942 error = VOP_GETATTR(vp, &vattr, cred);
945 mtx_lock(&np->n_mtx);
946 uio->uio_offset = np->n_size;
947 mtx_unlock(&np->n_mtx);
950 if (uio->uio_offset < 0)
952 tmp_off = uio->uio_offset + uio->uio_resid;
953 if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
955 if (uio->uio_resid == 0)
958 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG)
959 return nfs_directio_write(vp, uio, cred, ioflag);
962 * Maybe this should be above the vnode op call, but so long as
963 * file servers have no limits, i don't think it matters
965 if (vn_rlimit_fsize(vp, uio, td))
968 biosize = vp->v_bufobj.bo_bsize;
970 * Find all of this file's B_NEEDCOMMIT buffers. If our writes
971 * would exceed the local maximum per-file write commit size when
972 * combined with those, we must decide whether to flush,
973 * go synchronous, or return error. We don't bother checking
974 * IO_UNIT -- we just make all writes atomic anyway, as there's
975 * no point optimizing for something that really won't ever happen.
977 if (!(ioflag & IO_SYNC)) {
980 mtx_lock(&np->n_mtx);
982 mtx_unlock(&np->n_mtx);
984 if (nmp->nm_wcommitsize < uio->uio_resid) {
986 * If this request could not possibly be completed
987 * without exceeding the maximum outstanding write
988 * commit size, see if we can convert it into a
989 * synchronous write operation.
991 if (ioflag & IO_NDELAY)
994 if (nflag & NMODIFIED)
996 } else if (nflag & NMODIFIED) {
998 BO_LOCK(&vp->v_bufobj);
999 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
1000 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
1002 if (bp->b_flags & B_NEEDCOMMIT)
1003 wouldcommit += bp->b_bcount;
1006 BO_UNLOCK(&vp->v_bufobj);
1008 * Since we're not operating synchronously and
1009 * bypassing the buffer cache, we are in a commit
1010 * and holding all of these buffers whether
1011 * transmitted or not. If not limited, this
1012 * will lead to the buffer cache deadlocking,
1013 * as no one else can flush our uncommitted buffers.
1015 wouldcommit += uio->uio_resid;
1017 * If we would initially exceed the maximum
1018 * outstanding write commit size, flush and restart.
1020 if (wouldcommit > nmp->nm_wcommitsize)
1024 goto flush_and_restart;
1028 NFSINCRGLOBAL(newnfsstats.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 ncl_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 occured.
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) {
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 ((nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF))
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) && (vp->v_bufobj.bo_object != NULL)) {
1370 VM_OBJECT_WLOCK(vp->v_bufobj.bo_object);
1371 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1372 VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object);
1374 * If the page clean was interrupted, fail the invalidation.
1375 * Not doing so, we run the risk of losing dirty pages in the
1376 * vinvalbuf() call below.
1378 if (intrflg && (error = newnfs_sigintr(nmp, td)))
1382 error = vinvalbuf(vp, flags, slpflag, 0);
1384 if (intrflg && (error = newnfs_sigintr(nmp, td)))
1386 error = vinvalbuf(vp, flags, 0, slptimeo);
1388 if (NFSHASPNFS(nmp)) {
1389 nfscl_layoutcommit(vp, td);
1391 * Invalidate the attribute cache, since writes to a DS
1392 * won't update the size attribute.
1394 mtx_lock(&np->n_mtx);
1395 np->n_attrstamp = 0;
1397 mtx_lock(&np->n_mtx);
1398 if (np->n_directio_asyncwr == 0)
1399 np->n_flag &= ~NMODIFIED;
1400 mtx_unlock(&np->n_mtx);
1402 ncl_downgrade_vnlock(vp, old_lock);
1407 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1408 * This is mainly to avoid queueing async I/O requests when the nfsiods
1409 * are all hung on a dead server.
1411 * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1412 * is eventually dequeued by the async daemon, ncl_doio() *will*.
1415 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1424 * Commits are usually short and sweet so lets save some cpu and
1425 * leave the async daemons for more important rpc's (such as reads
1428 * Readdirplus RPCs do vget()s to acquire the vnodes for entries
1429 * in the directory in order to update attributes. This can deadlock
1430 * with another thread that is waiting for async I/O to be done by
1431 * an nfsiod thread while holding a lock on one of these vnodes.
1432 * To avoid this deadlock, don't allow the async nfsiod threads to
1433 * perform Readdirplus RPCs.
1435 mtx_lock(&ncl_iod_mutex);
1436 if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1437 (nmp->nm_bufqiods > ncl_numasync / 2)) ||
1438 (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
1439 mtx_unlock(&ncl_iod_mutex);
1443 if (nmp->nm_flag & NFSMNT_INT)
1448 * Find a free iod to process this request.
1450 for (iod = 0; iod < ncl_numasync; iod++)
1451 if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1457 * Try to create one if none are free.
1463 * Found one, so wake it up and tell it which
1466 NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1468 ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1469 ncl_iodmount[iod] = nmp;
1471 wakeup(&ncl_iodwant[iod]);
1475 * If none are free, we may already have an iod working on this mount
1476 * point. If so, it will process our request.
1479 if (nmp->nm_bufqiods > 0) {
1481 ("ncl_asyncio: %d iods are already processing mount %p\n",
1482 nmp->nm_bufqiods, nmp));
1488 * If we have an iod which can process the request, then queue
1493 * Ensure that the queue never grows too large. We still want
1494 * to asynchronize so we block rather then return EIO.
1496 while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1498 ("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1499 nmp->nm_bufqwant = TRUE;
1500 error = newnfs_msleep(td, &nmp->nm_bufq,
1501 &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1504 error2 = newnfs_sigintr(nmp, td);
1506 mtx_unlock(&ncl_iod_mutex);
1509 if (slpflag == PCATCH) {
1515 * We might have lost our iod while sleeping,
1516 * so check and loop if nescessary.
1521 /* We might have lost our nfsiod */
1522 if (nmp->nm_bufqiods == 0) {
1524 ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1528 if (bp->b_iocmd == BIO_READ) {
1529 if (bp->b_rcred == NOCRED && cred != NOCRED)
1530 bp->b_rcred = crhold(cred);
1532 if (bp->b_wcred == NOCRED && cred != NOCRED)
1533 bp->b_wcred = crhold(cred);
1536 if (bp->b_flags & B_REMFREE)
1539 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1541 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1542 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx);
1543 VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1544 VTONFS(bp->b_vp)->n_directio_asyncwr++;
1545 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx);
1547 mtx_unlock(&ncl_iod_mutex);
1551 mtx_unlock(&ncl_iod_mutex);
1554 * All the iods are busy on other mounts, so return EIO to
1555 * force the caller to process the i/o synchronously.
1557 NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1562 ncl_doio_directwrite(struct buf *bp)
1564 int iomode, must_commit;
1565 struct uio *uiop = (struct uio *)bp->b_caller1;
1566 char *iov_base = uiop->uio_iov->iov_base;
1568 iomode = NFSWRITE_FILESYNC;
1569 uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1570 ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0);
1571 KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write"));
1572 free(iov_base, M_NFSDIRECTIO);
1573 free(uiop->uio_iov, M_NFSDIRECTIO);
1574 free(uiop, M_NFSDIRECTIO);
1575 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1576 struct nfsnode *np = VTONFS(bp->b_vp);
1577 mtx_lock(&np->n_mtx);
1578 if (NFSHASPNFS(VFSTONFS(vnode_mount(bp->b_vp)))) {
1580 * Invalidate the attribute cache, since writes to a DS
1581 * won't update the size attribute.
1583 np->n_attrstamp = 0;
1585 np->n_directio_asyncwr--;
1586 if (np->n_directio_asyncwr == 0) {
1587 np->n_flag &= ~NMODIFIED;
1588 if ((np->n_flag & NFSYNCWAIT)) {
1589 np->n_flag &= ~NFSYNCWAIT;
1590 wakeup((caddr_t)&np->n_directio_asyncwr);
1593 mtx_unlock(&np->n_mtx);
1596 relpbuf(bp, &ncl_pbuf_freecnt);
1600 * Do an I/O operation to/from a cache block. This may be called
1601 * synchronously or from an nfsiod.
1604 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1605 int called_from_strategy)
1609 struct nfsmount *nmp;
1610 int error = 0, iomode, must_commit = 0;
1613 struct proc *p = td ? td->td_proc : NULL;
1617 nmp = VFSTONFS(vp->v_mount);
1619 uiop->uio_iov = &io;
1620 uiop->uio_iovcnt = 1;
1621 uiop->uio_segflg = UIO_SYSSPACE;
1625 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We
1626 * do this here so we do not have to do it in all the code that
1629 bp->b_flags &= ~B_INVAL;
1630 bp->b_ioflags &= ~BIO_ERROR;
1632 KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1633 iocmd = bp->b_iocmd;
1634 if (iocmd == BIO_READ) {
1635 io.iov_len = uiop->uio_resid = bp->b_bcount;
1636 io.iov_base = bp->b_data;
1637 uiop->uio_rw = UIO_READ;
1639 switch (vp->v_type) {
1641 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1642 NFSINCRGLOBAL(newnfsstats.read_bios);
1643 error = ncl_readrpc(vp, uiop, cr);
1646 if (uiop->uio_resid) {
1648 * If we had a short read with no error, we must have
1649 * hit a file hole. We should zero-fill the remainder.
1650 * This can also occur if the server hits the file EOF.
1652 * Holes used to be able to occur due to pending
1653 * writes, but that is not possible any longer.
1655 int nread = bp->b_bcount - uiop->uio_resid;
1656 ssize_t left = uiop->uio_resid;
1659 bzero((char *)bp->b_data + nread, left);
1660 uiop->uio_resid = 0;
1663 /* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1664 if (p && (vp->v_vflag & VV_TEXT)) {
1665 mtx_lock(&np->n_mtx);
1666 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1667 mtx_unlock(&np->n_mtx);
1669 killproc(p, "text file modification");
1672 mtx_unlock(&np->n_mtx);
1676 uiop->uio_offset = (off_t)0;
1677 NFSINCRGLOBAL(newnfsstats.readlink_bios);
1678 error = ncl_readlinkrpc(vp, uiop, cr);
1681 NFSINCRGLOBAL(newnfsstats.readdir_bios);
1682 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1683 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1684 error = ncl_readdirplusrpc(vp, uiop, cr, td);
1685 if (error == NFSERR_NOTSUPP)
1686 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1688 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1689 error = ncl_readdirrpc(vp, uiop, cr, td);
1691 * end-of-directory sets B_INVAL but does not generate an
1694 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1695 bp->b_flags |= B_INVAL;
1698 ncl_printf("ncl_doio: type %x unexpected\n", vp->v_type);
1702 bp->b_ioflags |= BIO_ERROR;
1703 bp->b_error = error;
1707 * If we only need to commit, try to commit
1709 if (bp->b_flags & B_NEEDCOMMIT) {
1713 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1714 retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1717 bp->b_dirtyoff = bp->b_dirtyend = 0;
1718 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1723 if (retv == NFSERR_STALEWRITEVERF) {
1724 ncl_clearcommit(vp->v_mount);
1729 * Setup for actual write
1731 mtx_lock(&np->n_mtx);
1732 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1733 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1734 mtx_unlock(&np->n_mtx);
1736 if (bp->b_dirtyend > bp->b_dirtyoff) {
1737 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1739 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1741 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1742 uiop->uio_rw = UIO_WRITE;
1743 NFSINCRGLOBAL(newnfsstats.write_bios);
1745 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1746 iomode = NFSWRITE_UNSTABLE;
1748 iomode = NFSWRITE_FILESYNC;
1750 error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1751 called_from_strategy);
1754 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1755 * to cluster the buffers needing commit. This will allow
1756 * the system to submit a single commit rpc for the whole
1757 * cluster. We can do this even if the buffer is not 100%
1758 * dirty (relative to the NFS blocksize), so we optimize the
1759 * append-to-file-case.
1761 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1762 * cleared because write clustering only works for commit
1763 * rpc's, not for the data portion of the write).
1766 if (!error && iomode == NFSWRITE_UNSTABLE) {
1767 bp->b_flags |= B_NEEDCOMMIT;
1768 if (bp->b_dirtyoff == 0
1769 && bp->b_dirtyend == bp->b_bcount)
1770 bp->b_flags |= B_CLUSTEROK;
1772 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1776 * For an interrupted write, the buffer is still valid
1777 * and the write hasn't been pushed to the server yet,
1778 * so we can't set BIO_ERROR and report the interruption
1779 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1780 * is not relevant, so the rpc attempt is essentially
1781 * a noop. For the case of a V3 write rpc not being
1782 * committed to stable storage, the block is still
1783 * dirty and requires either a commit rpc or another
1784 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1785 * the block is reused. This is indicated by setting
1786 * the B_DELWRI and B_NEEDCOMMIT flags.
1788 * EIO is returned by ncl_writerpc() to indicate a recoverable
1789 * write error and is handled as above, except that
1790 * B_EINTR isn't set. One cause of this is a stale stateid
1791 * error for the RPC that indicates recovery is required,
1792 * when called with called_from_strategy != 0.
1794 * If the buffer is marked B_PAGING, it does not reside on
1795 * the vp's paging queues so we cannot call bdirty(). The
1796 * bp in this case is not an NFS cache block so we should
1799 * The logic below breaks up errors into recoverable and
1800 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1801 * and keep the buffer around for potential write retries.
1802 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1803 * and save the error in the nfsnode. This is less than ideal
1804 * but necessary. Keeping such buffers around could potentially
1805 * cause buffer exhaustion eventually (they can never be written
1806 * out, so will get constantly be re-dirtied). It also causes
1807 * all sorts of vfs panics. For non-recoverable write errors,
1808 * also invalidate the attrcache, so we'll be forced to go over
1809 * the wire for this object, returning an error to user on next
1810 * call (most of the time).
1812 if (error == EINTR || error == EIO || error == ETIMEDOUT
1813 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1817 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1818 if ((bp->b_flags & B_PAGING) == 0) {
1820 bp->b_flags &= ~B_DONE;
1822 if ((error == EINTR || error == ETIMEDOUT) &&
1823 (bp->b_flags & B_ASYNC) == 0)
1824 bp->b_flags |= B_EINTR;
1828 bp->b_ioflags |= BIO_ERROR;
1829 bp->b_flags |= B_INVAL;
1830 bp->b_error = np->n_error = error;
1831 mtx_lock(&np->n_mtx);
1832 np->n_flag |= NWRITEERR;
1833 np->n_attrstamp = 0;
1834 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1835 mtx_unlock(&np->n_mtx);
1837 bp->b_dirtyoff = bp->b_dirtyend = 0;
1845 bp->b_resid = uiop->uio_resid;
1847 ncl_clearcommit(vp->v_mount);
1853 * Used to aid in handling ftruncate() operations on the NFS client side.
1854 * Truncation creates a number of special problems for NFS. We have to
1855 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1856 * we have to properly handle VM pages or (potentially dirty) buffers
1857 * that straddle the truncation point.
1861 ncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1863 struct nfsnode *np = VTONFS(vp);
1865 int biosize = vp->v_bufobj.bo_bsize;
1868 mtx_lock(&np->n_mtx);
1871 mtx_unlock(&np->n_mtx);
1873 if (nsize < tsize) {
1879 * vtruncbuf() doesn't get the buffer overlapping the
1880 * truncation point. We may have a B_DELWRI and/or B_CACHE
1881 * buffer that now needs to be truncated.
1883 error = vtruncbuf(vp, cred, nsize, biosize);
1884 lbn = nsize / biosize;
1885 bufsize = nsize - (lbn * biosize);
1886 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1889 if (bp->b_dirtyoff > bp->b_bcount)
1890 bp->b_dirtyoff = bp->b_bcount;
1891 if (bp->b_dirtyend > bp->b_bcount)
1892 bp->b_dirtyend = bp->b_bcount;
1893 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1896 vnode_pager_setsize(vp, nsize);