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 * Since the caller has busied the requested page, that page's valid
133 * field will not be changed by other threads.
135 vm_page_assert_xbusied(pages[ap->a_reqpage]);
138 * If the requested page is partially valid, just return it and
139 * allow the pager to zero-out the blanks. Partially valid pages
140 * can only occur at the file EOF.
142 if (pages[ap->a_reqpage]->valid != 0) {
143 vm_pager_free_nonreq(object, pages, ap->a_reqpage, npages,
145 return (VM_PAGER_OK);
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_pager_free_nonreq(object, pages, ap->a_reqpage, npages,
178 return (VM_PAGER_ERROR);
182 * Calculate the number of bytes read and validate only that number
183 * of bytes. Note that due to pending writes, size may be 0. This
184 * does not mean that the remaining data is invalid!
187 size = count - uio.uio_resid;
188 VM_OBJECT_WLOCK(object);
189 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
191 nextoff = toff + PAGE_SIZE;
194 if (nextoff <= size) {
196 * Read operation filled an entire page
198 m->valid = VM_PAGE_BITS_ALL;
199 KASSERT(m->dirty == 0,
200 ("nfs_getpages: page %p is dirty", m));
201 } else if (size > toff) {
203 * Read operation filled a partial page.
206 vm_page_set_valid_range(m, 0, size - toff);
207 KASSERT(m->dirty == 0,
208 ("nfs_getpages: page %p is dirty", m));
211 * Read operation was short. If no error
212 * occured we may have hit a zero-fill
213 * section. We leave valid set to 0, and page
214 * is freed by vm_page_readahead_finish() if
215 * its index is not equal to requested, or
216 * page is zeroed and set valid by
217 * vm_pager_get_pages() for requested page.
221 if (i != ap->a_reqpage)
222 vm_page_readahead_finish(m);
224 VM_OBJECT_WUNLOCK(object);
229 * Vnode op for VM putpages.
232 ncl_putpages(struct vop_putpages_args *ap)
238 int iomode, must_commit, i, error, npages, count;
244 struct nfsmount *nmp;
250 td = curthread; /* XXX */
251 /* Set the cred to n_writecred for the write rpcs. */
252 if (np->n_writecred != NULL)
253 cred = crhold(np->n_writecred);
255 cred = crhold(curthread->td_ucred); /* XXX */
256 nmp = VFSTONFS(vp->v_mount);
259 rtvals = ap->a_rtvals;
260 npages = btoc(count);
261 offset = IDX_TO_OFF(pages[0]->pindex);
263 mtx_lock(&nmp->nm_mtx);
264 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
265 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
266 mtx_unlock(&nmp->nm_mtx);
267 (void)ncl_fsinfo(nmp, vp, cred, td);
269 mtx_unlock(&nmp->nm_mtx);
271 mtx_lock(&np->n_mtx);
272 if (newnfs_directio_enable && !newnfs_directio_allow_mmap &&
273 (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) {
274 mtx_unlock(&np->n_mtx);
275 ncl_printf("ncl_putpages: called on noncache-able vnode??\n");
276 mtx_lock(&np->n_mtx);
279 for (i = 0; i < npages; i++)
280 rtvals[i] = VM_PAGER_ERROR;
283 * When putting pages, do not extend file past EOF.
285 if (offset + count > np->n_size) {
286 count = np->n_size - offset;
290 mtx_unlock(&np->n_mtx);
293 * We use only the kva address for the buffer, but this is extremely
294 * convienient and fast.
296 bp = getpbuf(&ncl_pbuf_freecnt);
298 kva = (vm_offset_t) bp->b_data;
299 pmap_qenter(kva, pages, npages);
300 PCPU_INC(cnt.v_vnodeout);
301 PCPU_ADD(cnt.v_vnodepgsout, count);
303 iov.iov_base = (caddr_t) kva;
307 uio.uio_offset = offset;
308 uio.uio_resid = count;
309 uio.uio_segflg = UIO_SYSSPACE;
310 uio.uio_rw = UIO_WRITE;
313 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
314 iomode = NFSWRITE_UNSTABLE;
316 iomode = NFSWRITE_FILESYNC;
318 error = ncl_writerpc(vp, &uio, cred, &iomode, &must_commit, 0);
321 pmap_qremove(kva, npages);
322 relpbuf(bp, &ncl_pbuf_freecnt);
324 if (error == 0 || !nfs_keep_dirty_on_error) {
325 vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid);
327 ncl_clearcommit(vp->v_mount);
333 * For nfs, cache consistency can only be maintained approximately.
334 * Although RFC1094 does not specify the criteria, the following is
335 * believed to be compatible with the reference port.
337 * If the file's modify time on the server has changed since the
338 * last read rpc or you have written to the file,
339 * you may have lost data cache consistency with the
340 * server, so flush all of the file's data out of the cache.
341 * Then force a getattr rpc to ensure that you have up to date
343 * NB: This implies that cache data can be read when up to
344 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
345 * attributes this could be forced by setting n_attrstamp to 0 before
346 * the VOP_GETATTR() call.
349 nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred)
353 struct nfsnode *np = VTONFS(vp);
357 * Grab the exclusive lock before checking whether the cache is
359 * XXX - We can make this cheaper later (by acquiring cheaper locks).
360 * But for now, this suffices.
362 old_lock = ncl_upgrade_vnlock(vp);
363 if (vp->v_iflag & VI_DOOMED) {
364 ncl_downgrade_vnlock(vp, old_lock);
368 mtx_lock(&np->n_mtx);
369 if (np->n_flag & NMODIFIED) {
370 mtx_unlock(&np->n_mtx);
371 if (vp->v_type != VREG) {
372 if (vp->v_type != VDIR)
373 panic("nfs: bioread, not dir");
375 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
380 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
381 error = VOP_GETATTR(vp, &vattr, cred);
384 mtx_lock(&np->n_mtx);
385 np->n_mtime = vattr.va_mtime;
386 mtx_unlock(&np->n_mtx);
388 mtx_unlock(&np->n_mtx);
389 error = VOP_GETATTR(vp, &vattr, cred);
392 mtx_lock(&np->n_mtx);
393 if ((np->n_flag & NSIZECHANGED)
394 || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) {
395 mtx_unlock(&np->n_mtx);
396 if (vp->v_type == VDIR)
398 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
401 mtx_lock(&np->n_mtx);
402 np->n_mtime = vattr.va_mtime;
403 np->n_flag &= ~NSIZECHANGED;
405 mtx_unlock(&np->n_mtx);
408 ncl_downgrade_vnlock(vp, old_lock);
413 * Vnode op for read using bio
416 ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
418 struct nfsnode *np = VTONFS(vp);
420 struct buf *bp, *rabp;
422 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
426 int nra, error = 0, n = 0, on = 0;
429 KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode"));
430 if (uio->uio_resid == 0)
432 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
436 mtx_lock(&nmp->nm_mtx);
437 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
438 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
439 mtx_unlock(&nmp->nm_mtx);
440 (void)ncl_fsinfo(nmp, vp, cred, td);
441 mtx_lock(&nmp->nm_mtx);
443 if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0)
444 (void) newnfs_iosize(nmp);
446 tmp_off = uio->uio_offset + uio->uio_resid;
447 if (vp->v_type != VDIR &&
448 (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) {
449 mtx_unlock(&nmp->nm_mtx);
452 mtx_unlock(&nmp->nm_mtx);
454 if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG))
455 /* No caching/ no readaheads. Just read data into the user buffer */
456 return ncl_readrpc(vp, uio, cred);
458 biosize = vp->v_bufobj.bo_bsize;
459 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
461 error = nfs_bioread_check_cons(vp, td, cred);
468 mtx_lock(&np->n_mtx);
470 mtx_unlock(&np->n_mtx);
472 switch (vp->v_type) {
474 NFSINCRGLOBAL(newnfsstats.biocache_reads);
475 lbn = uio->uio_offset / biosize;
476 on = uio->uio_offset - (lbn * biosize);
479 * Start the read ahead(s), as required.
481 if (nmp->nm_readahead > 0) {
482 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
483 (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
484 rabn = lbn + 1 + nra;
485 if (incore(&vp->v_bufobj, rabn) == NULL) {
486 rabp = nfs_getcacheblk(vp, rabn, biosize, td);
488 error = newnfs_sigintr(nmp, td);
489 return (error ? error : EINTR);
491 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
492 rabp->b_flags |= B_ASYNC;
493 rabp->b_iocmd = BIO_READ;
494 vfs_busy_pages(rabp, 0);
495 if (ncl_asyncio(nmp, rabp, cred, td)) {
496 rabp->b_flags |= B_INVAL;
497 rabp->b_ioflags |= BIO_ERROR;
498 vfs_unbusy_pages(rabp);
509 /* Note that bcount is *not* DEV_BSIZE aligned. */
511 if ((off_t)lbn * biosize >= nsize) {
513 } else if ((off_t)(lbn + 1) * biosize > nsize) {
514 bcount = nsize - (off_t)lbn * biosize;
516 bp = nfs_getcacheblk(vp, lbn, bcount, td);
519 error = newnfs_sigintr(nmp, td);
520 return (error ? error : EINTR);
524 * If B_CACHE is not set, we must issue the read. If this
525 * fails, we return an error.
528 if ((bp->b_flags & B_CACHE) == 0) {
529 bp->b_iocmd = BIO_READ;
530 vfs_busy_pages(bp, 0);
531 error = ncl_doio(vp, bp, cred, td, 0);
539 * on is the offset into the current bp. Figure out how many
540 * bytes we can copy out of the bp. Note that bcount is
541 * NOT DEV_BSIZE aligned.
543 * Then figure out how many bytes we can copy into the uio.
548 n = MIN((unsigned)(bcount - on), uio->uio_resid);
551 NFSINCRGLOBAL(newnfsstats.biocache_readlinks);
552 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
554 error = newnfs_sigintr(nmp, td);
555 return (error ? error : EINTR);
557 if ((bp->b_flags & B_CACHE) == 0) {
558 bp->b_iocmd = BIO_READ;
559 vfs_busy_pages(bp, 0);
560 error = ncl_doio(vp, bp, cred, td, 0);
562 bp->b_ioflags |= BIO_ERROR;
567 n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
571 NFSINCRGLOBAL(newnfsstats.biocache_readdirs);
572 if (np->n_direofoffset
573 && uio->uio_offset >= np->n_direofoffset) {
576 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
577 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
578 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, 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);
590 while (error == NFSERR_BAD_COOKIE) {
592 error = ncl_vinvalbuf(vp, 0, td, 1);
594 * Yuck! The directory has been modified on the
595 * server. The only way to get the block is by
596 * reading from the beginning to get all the
599 * Leave the last bp intact unless there is an error.
600 * Loop back up to the while if the error is another
601 * NFSERR_BAD_COOKIE (double yuch!).
603 for (i = 0; i <= lbn && !error; i++) {
604 if (np->n_direofoffset
605 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
607 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
609 error = newnfs_sigintr(nmp, td);
610 return (error ? error : EINTR);
612 if ((bp->b_flags & B_CACHE) == 0) {
613 bp->b_iocmd = BIO_READ;
614 vfs_busy_pages(bp, 0);
615 error = ncl_doio(vp, bp, cred, td, 0);
617 * no error + B_INVAL == directory EOF,
620 if (error == 0 && (bp->b_flags & B_INVAL))
624 * An error will throw away the block and the
625 * for loop will break out. If no error and this
626 * is not the block we want, we throw away the
627 * block and go for the next one via the for loop.
629 if (error || i < lbn)
634 * The above while is repeated if we hit another cookie
635 * error. If we hit an error and it wasn't a cookie error,
643 * If not eof and read aheads are enabled, start one.
644 * (You need the current block first, so that you have the
645 * directory offset cookie of the next block.)
647 if (nmp->nm_readahead > 0 &&
648 (bp->b_flags & B_INVAL) == 0 &&
649 (np->n_direofoffset == 0 ||
650 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
651 incore(&vp->v_bufobj, lbn + 1) == NULL) {
652 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
654 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
655 rabp->b_flags |= B_ASYNC;
656 rabp->b_iocmd = BIO_READ;
657 vfs_busy_pages(rabp, 0);
658 if (ncl_asyncio(nmp, rabp, cred, td)) {
659 rabp->b_flags |= B_INVAL;
660 rabp->b_ioflags |= BIO_ERROR;
661 vfs_unbusy_pages(rabp);
670 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
671 * chopped for the EOF condition, we cannot tell how large
672 * NFS directories are going to be until we hit EOF. So
673 * an NFS directory buffer is *not* chopped to its EOF. Now,
674 * it just so happens that b_resid will effectively chop it
675 * to EOF. *BUT* this information is lost if the buffer goes
676 * away and is reconstituted into a B_CACHE state ( due to
677 * being VMIO ) later. So we keep track of the directory eof
678 * in np->n_direofoffset and chop it off as an extra step
681 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
682 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
683 n = np->n_direofoffset - uio->uio_offset;
686 ncl_printf(" ncl_bioread: type %x unexpected\n", vp->v_type);
692 error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio);
694 if (vp->v_type == VLNK)
698 } while (error == 0 && uio->uio_resid > 0 && n > 0);
703 * The NFS write path cannot handle iovecs with len > 1. So we need to
704 * break up iovecs accordingly (restricting them to wsize).
705 * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf).
706 * For the ASYNC case, 2 copies are needed. The first a copy from the
707 * user buffer to a staging buffer and then a second copy from the staging
708 * buffer to mbufs. This can be optimized by copying from the user buffer
709 * directly into mbufs and passing the chain down, but that requires a
710 * fair amount of re-working of the relevant codepaths (and can be done
714 nfs_directio_write(vp, uiop, cred, ioflag)
721 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
722 struct thread *td = uiop->uio_td;
726 mtx_lock(&nmp->nm_mtx);
727 wsize = nmp->nm_wsize;
728 mtx_unlock(&nmp->nm_mtx);
729 if (ioflag & IO_SYNC) {
730 int iomode, must_commit;
734 while (uiop->uio_resid > 0) {
735 size = MIN(uiop->uio_resid, wsize);
736 size = MIN(uiop->uio_iov->iov_len, size);
737 iov.iov_base = uiop->uio_iov->iov_base;
741 uio.uio_offset = uiop->uio_offset;
742 uio.uio_resid = size;
743 uio.uio_segflg = UIO_USERSPACE;
744 uio.uio_rw = UIO_WRITE;
746 iomode = NFSWRITE_FILESYNC;
747 error = ncl_writerpc(vp, &uio, cred, &iomode,
749 KASSERT((must_commit == 0),
750 ("ncl_directio_write: Did not commit write"));
753 uiop->uio_offset += size;
754 uiop->uio_resid -= size;
755 if (uiop->uio_iov->iov_len <= size) {
759 uiop->uio_iov->iov_base =
760 (char *)uiop->uio_iov->iov_base + size;
761 uiop->uio_iov->iov_len -= size;
770 * Break up the write into blocksize chunks and hand these
771 * over to nfsiod's for write back.
772 * Unfortunately, this incurs a copy of the data. Since
773 * the user could modify the buffer before the write is
776 * The obvious optimization here is that one of the 2 copies
777 * in the async write path can be eliminated by copying the
778 * data here directly into mbufs and passing the mbuf chain
779 * down. But that will require a fair amount of re-working
780 * of the code and can be done if there's enough interest
781 * in NFS directio access.
783 while (uiop->uio_resid > 0) {
784 size = MIN(uiop->uio_resid, wsize);
785 size = MIN(uiop->uio_iov->iov_len, size);
786 bp = getpbuf(&ncl_pbuf_freecnt);
787 t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK);
788 t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK);
789 t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK);
790 t_iov->iov_len = size;
791 t_uio->uio_iov = t_iov;
792 t_uio->uio_iovcnt = 1;
793 t_uio->uio_offset = uiop->uio_offset;
794 t_uio->uio_resid = size;
795 t_uio->uio_segflg = UIO_SYSSPACE;
796 t_uio->uio_rw = UIO_WRITE;
798 KASSERT(uiop->uio_segflg == UIO_USERSPACE ||
799 uiop->uio_segflg == UIO_SYSSPACE,
800 ("nfs_directio_write: Bad uio_segflg"));
801 if (uiop->uio_segflg == UIO_USERSPACE) {
802 error = copyin(uiop->uio_iov->iov_base,
803 t_iov->iov_base, size);
808 * UIO_SYSSPACE may never happen, but handle
809 * it just in case it does.
811 bcopy(uiop->uio_iov->iov_base, t_iov->iov_base,
813 bp->b_flags |= B_DIRECT;
814 bp->b_iocmd = BIO_WRITE;
815 if (cred != NOCRED) {
819 bp->b_wcred = NOCRED;
820 bp->b_caller1 = (void *)t_uio;
822 error = ncl_asyncio(nmp, bp, NOCRED, td);
825 free(t_iov->iov_base, M_NFSDIRECTIO);
826 free(t_iov, M_NFSDIRECTIO);
827 free(t_uio, M_NFSDIRECTIO);
829 relpbuf(bp, &ncl_pbuf_freecnt);
834 uiop->uio_offset += size;
835 uiop->uio_resid -= size;
836 if (uiop->uio_iov->iov_len <= size) {
840 uiop->uio_iov->iov_base =
841 (char *)uiop->uio_iov->iov_base + size;
842 uiop->uio_iov->iov_len -= size;
850 * Vnode op for write using bio
853 ncl_write(struct vop_write_args *ap)
856 struct uio *uio = ap->a_uio;
857 struct thread *td = uio->uio_td;
858 struct vnode *vp = ap->a_vp;
859 struct nfsnode *np = VTONFS(vp);
860 struct ucred *cred = ap->a_cred;
861 int ioflag = ap->a_ioflag;
864 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
866 int bcount, noncontig_write, obcount;
867 int bp_cached, n, on, error = 0, error1, wouldcommit;
868 size_t orig_resid, local_resid;
869 off_t orig_size, tmp_off;
871 KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
872 KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread,
874 if (vp->v_type != VREG)
876 mtx_lock(&np->n_mtx);
877 if (np->n_flag & NWRITEERR) {
878 np->n_flag &= ~NWRITEERR;
879 mtx_unlock(&np->n_mtx);
880 return (np->n_error);
882 mtx_unlock(&np->n_mtx);
883 mtx_lock(&nmp->nm_mtx);
884 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
885 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
886 mtx_unlock(&nmp->nm_mtx);
887 (void)ncl_fsinfo(nmp, vp, cred, td);
888 mtx_lock(&nmp->nm_mtx);
890 if (nmp->nm_wsize == 0)
891 (void) newnfs_iosize(nmp);
892 mtx_unlock(&nmp->nm_mtx);
895 * Synchronously flush pending buffers if we are in synchronous
896 * mode or if we are appending.
898 if (ioflag & (IO_APPEND | IO_SYNC)) {
899 mtx_lock(&np->n_mtx);
900 if (np->n_flag & NMODIFIED) {
901 mtx_unlock(&np->n_mtx);
902 #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */
904 * Require non-blocking, synchronous writes to
905 * dirty files to inform the program it needs
906 * to fsync(2) explicitly.
908 if (ioflag & IO_NDELAY)
912 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
913 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
917 mtx_unlock(&np->n_mtx);
920 orig_resid = uio->uio_resid;
921 mtx_lock(&np->n_mtx);
922 orig_size = np->n_size;
923 mtx_unlock(&np->n_mtx);
926 * If IO_APPEND then load uio_offset. We restart here if we cannot
927 * get the append lock.
929 if (ioflag & IO_APPEND) {
931 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
932 error = VOP_GETATTR(vp, &vattr, cred);
935 mtx_lock(&np->n_mtx);
936 uio->uio_offset = np->n_size;
937 mtx_unlock(&np->n_mtx);
940 if (uio->uio_offset < 0)
942 tmp_off = uio->uio_offset + uio->uio_resid;
943 if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)
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
955 if (vn_rlimit_fsize(vp, uio, td))
958 biosize = vp->v_bufobj.bo_bsize;
960 * Find all of this file's B_NEEDCOMMIT buffers. If our writes
961 * would exceed the local maximum per-file write commit size when
962 * combined with those, we must decide whether to flush,
963 * go synchronous, or return error. We don't bother checking
964 * IO_UNIT -- we just make all writes atomic anyway, as there's
965 * no point optimizing for something that really won't ever happen.
968 if (!(ioflag & IO_SYNC)) {
971 mtx_lock(&np->n_mtx);
973 mtx_unlock(&np->n_mtx);
974 if (nflag & NMODIFIED) {
975 BO_LOCK(&vp->v_bufobj);
976 if (vp->v_bufobj.bo_dirty.bv_cnt != 0) {
977 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd,
979 if (bp->b_flags & B_NEEDCOMMIT)
980 wouldcommit += bp->b_bcount;
983 BO_UNLOCK(&vp->v_bufobj);
988 if (!(ioflag & IO_SYNC)) {
989 wouldcommit += biosize;
990 if (wouldcommit > nmp->nm_wcommitsize) {
992 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
993 error = ncl_vinvalbuf(vp, V_SAVE, td, 1);
996 wouldcommit = biosize;
1000 NFSINCRGLOBAL(newnfsstats.biocache_writes);
1001 lbn = uio->uio_offset / biosize;
1002 on = uio->uio_offset - (lbn * biosize);
1003 n = MIN((unsigned)(biosize - on), uio->uio_resid);
1006 * Handle direct append and file extension cases, calculate
1007 * unaligned buffer size.
1009 mtx_lock(&np->n_mtx);
1010 if ((np->n_flag & NHASBEENLOCKED) == 0 &&
1011 (nmp->nm_flag & NFSMNT_NONCONTIGWR) != 0)
1012 noncontig_write = 1;
1014 noncontig_write = 0;
1015 if ((uio->uio_offset == np->n_size ||
1016 (noncontig_write != 0 &&
1017 lbn == (np->n_size / biosize) &&
1018 uio->uio_offset + n > np->n_size)) && n) {
1019 mtx_unlock(&np->n_mtx);
1021 * Get the buffer (in its pre-append state to maintain
1022 * B_CACHE if it was previously set). Resize the
1023 * nfsnode after we have locked the buffer to prevent
1024 * readers from reading garbage.
1026 obcount = np->n_size - (lbn * biosize);
1027 bp = nfs_getcacheblk(vp, lbn, obcount, td);
1032 mtx_lock(&np->n_mtx);
1033 np->n_size = uio->uio_offset + n;
1034 np->n_flag |= NMODIFIED;
1035 vnode_pager_setsize(vp, np->n_size);
1036 mtx_unlock(&np->n_mtx);
1038 save = bp->b_flags & B_CACHE;
1040 allocbuf(bp, bcount);
1041 bp->b_flags |= save;
1042 if (noncontig_write != 0 && on > obcount)
1043 vfs_bio_bzero_buf(bp, obcount, on -
1048 * Obtain the locked cache block first, and then
1049 * adjust the file's size as appropriate.
1052 if ((off_t)lbn * biosize + bcount < np->n_size) {
1053 if ((off_t)(lbn + 1) * biosize < np->n_size)
1056 bcount = np->n_size - (off_t)lbn * biosize;
1058 mtx_unlock(&np->n_mtx);
1059 bp = nfs_getcacheblk(vp, lbn, bcount, td);
1060 mtx_lock(&np->n_mtx);
1061 if (uio->uio_offset + n > np->n_size) {
1062 np->n_size = uio->uio_offset + n;
1063 np->n_flag |= NMODIFIED;
1064 vnode_pager_setsize(vp, np->n_size);
1066 mtx_unlock(&np->n_mtx);
1070 error = newnfs_sigintr(nmp, td);
1077 * Issue a READ if B_CACHE is not set. In special-append
1078 * mode, B_CACHE is based on the buffer prior to the write
1079 * op and is typically set, avoiding the read. If a read
1080 * is required in special append mode, the server will
1081 * probably send us a short-read since we extended the file
1082 * on our end, resulting in b_resid == 0 and, thusly,
1083 * B_CACHE getting set.
1085 * We can also avoid issuing the read if the write covers
1086 * the entire buffer. We have to make sure the buffer state
1087 * is reasonable in this case since we will not be initiating
1088 * I/O. See the comments in kern/vfs_bio.c's getblk() for
1091 * B_CACHE may also be set due to the buffer being cached
1096 if (on == 0 && n == bcount) {
1097 if ((bp->b_flags & B_CACHE) == 0)
1099 bp->b_flags |= B_CACHE;
1100 bp->b_flags &= ~B_INVAL;
1101 bp->b_ioflags &= ~BIO_ERROR;
1104 if ((bp->b_flags & B_CACHE) == 0) {
1105 bp->b_iocmd = BIO_READ;
1106 vfs_busy_pages(bp, 0);
1107 error = ncl_doio(vp, bp, cred, td, 0);
1113 if (bp->b_wcred == NOCRED)
1114 bp->b_wcred = crhold(cred);
1115 mtx_lock(&np->n_mtx);
1116 np->n_flag |= NMODIFIED;
1117 mtx_unlock(&np->n_mtx);
1120 * If dirtyend exceeds file size, chop it down. This should
1121 * not normally occur but there is an append race where it
1122 * might occur XXX, so we log it.
1124 * If the chopping creates a reverse-indexed or degenerate
1125 * situation with dirtyoff/end, we 0 both of them.
1128 if (bp->b_dirtyend > bcount) {
1129 ncl_printf("NFS append race @%lx:%d\n",
1130 (long)bp->b_blkno * DEV_BSIZE,
1131 bp->b_dirtyend - bcount);
1132 bp->b_dirtyend = bcount;
1135 if (bp->b_dirtyoff >= bp->b_dirtyend)
1136 bp->b_dirtyoff = bp->b_dirtyend = 0;
1139 * If the new write will leave a contiguous dirty
1140 * area, just update the b_dirtyoff and b_dirtyend,
1141 * otherwise force a write rpc of the old dirty area.
1143 * If there has been a file lock applied to this file
1144 * or vfs.nfs.old_noncontig_writing is set, do the following:
1145 * While it is possible to merge discontiguous writes due to
1146 * our having a B_CACHE buffer ( and thus valid read data
1147 * for the hole), we don't because it could lead to
1148 * significant cache coherency problems with multiple clients,
1149 * especially if locking is implemented later on.
1151 * If vfs.nfs.old_noncontig_writing is not set and there has
1152 * not been file locking done on this file:
1153 * Relax coherency a bit for the sake of performance and
1154 * expand the current dirty region to contain the new
1155 * write even if it means we mark some non-dirty data as
1159 if (noncontig_write == 0 && bp->b_dirtyend > 0 &&
1160 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1161 if (bwrite(bp) == EINTR) {
1168 local_resid = uio->uio_resid;
1169 error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio);
1171 if (error != 0 && !bp_cached) {
1173 * This block has no other content then what
1174 * possibly was written by the faulty uiomove.
1175 * Release it, forgetting the data pages, to
1176 * prevent the leak of uninitialized data to
1179 bp->b_ioflags |= BIO_ERROR;
1181 uio->uio_offset -= local_resid - uio->uio_resid;
1182 uio->uio_resid = local_resid;
1187 * Since this block is being modified, it must be written
1188 * again and not just committed. Since write clustering does
1189 * not work for the stage 1 data write, only the stage 2
1190 * commit rpc, we have to clear B_CLUSTEROK as well.
1192 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1195 * Get the partial update on the progress made from
1196 * uiomove, if an error occured.
1199 n = local_resid - uio->uio_resid;
1202 * Only update dirtyoff/dirtyend if not a degenerate
1206 if (bp->b_dirtyend > 0) {
1207 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1208 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1210 bp->b_dirtyoff = on;
1211 bp->b_dirtyend = on + n;
1213 vfs_bio_set_valid(bp, on, n);
1217 * If IO_SYNC do bwrite().
1219 * IO_INVAL appears to be unused. The idea appears to be
1220 * to turn off caching in this case. Very odd. XXX
1222 if ((ioflag & IO_SYNC)) {
1223 if (ioflag & IO_INVAL)
1224 bp->b_flags |= B_NOCACHE;
1225 error1 = bwrite(bp);
1231 } else if ((n + on) == biosize) {
1232 bp->b_flags |= B_ASYNC;
1233 (void) ncl_writebp(bp, 0, NULL);
1240 } while (uio->uio_resid > 0 && n > 0);
1243 if (ioflag & IO_UNIT) {
1245 vattr.va_size = orig_size;
1246 /* IO_SYNC is handled implicitely */
1247 (void)VOP_SETATTR(vp, &vattr, cred);
1248 uio->uio_offset -= orig_resid - uio->uio_resid;
1249 uio->uio_resid = orig_resid;
1257 * Get an nfs cache block.
1259 * Allocate a new one if the block isn't currently in the cache
1260 * and return the block marked busy. If the calling process is
1261 * interrupted by a signal for an interruptible mount point, return
1264 * The caller must carefully deal with the possible B_INVAL state of
1265 * the buffer. ncl_doio() clears B_INVAL (and ncl_asyncio() clears it
1266 * indirectly), so synchronous reads can be issued without worrying about
1267 * the B_INVAL state. We have to be a little more careful when dealing
1268 * with writes (see comments in nfs_write()) when extending a file past
1272 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1276 struct nfsmount *nmp;
1281 if (nmp->nm_flag & NFSMNT_INT) {
1284 newnfs_set_sigmask(td, &oldset);
1285 bp = getblk(vp, bn, size, PCATCH, 0, 0);
1286 newnfs_restore_sigmask(td, &oldset);
1287 while (bp == NULL) {
1288 if (newnfs_sigintr(nmp, td))
1290 bp = getblk(vp, bn, size, 0, 2 * hz, 0);
1293 bp = getblk(vp, bn, size, 0, 0, 0);
1296 if (vp->v_type == VREG)
1297 bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE);
1302 * Flush and invalidate all dirty buffers. If another process is already
1303 * doing the flush, just wait for completion.
1306 ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg)
1308 struct nfsnode *np = VTONFS(vp);
1309 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1310 int error = 0, slpflag, slptimeo;
1313 ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf");
1315 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1317 if ((nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF))
1327 old_lock = ncl_upgrade_vnlock(vp);
1328 if (vp->v_iflag & VI_DOOMED) {
1330 * Since vgonel() uses the generic vinvalbuf() to flush
1331 * dirty buffers and it does not call this function, it
1332 * is safe to just return OK when VI_DOOMED is set.
1334 ncl_downgrade_vnlock(vp, old_lock);
1339 * Now, flush as required.
1341 if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) {
1342 VM_OBJECT_WLOCK(vp->v_bufobj.bo_object);
1343 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
1344 VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object);
1346 * If the page clean was interrupted, fail the invalidation.
1347 * Not doing so, we run the risk of losing dirty pages in the
1348 * vinvalbuf() call below.
1350 if (intrflg && (error = newnfs_sigintr(nmp, td)))
1354 error = vinvalbuf(vp, flags, slpflag, 0);
1356 if (intrflg && (error = newnfs_sigintr(nmp, td)))
1358 error = vinvalbuf(vp, flags, 0, slptimeo);
1360 if (NFSHASPNFS(nmp)) {
1361 nfscl_layoutcommit(vp, td);
1363 * Invalidate the attribute cache, since writes to a DS
1364 * won't update the size attribute.
1366 mtx_lock(&np->n_mtx);
1367 np->n_attrstamp = 0;
1369 mtx_lock(&np->n_mtx);
1370 if (np->n_directio_asyncwr == 0)
1371 np->n_flag &= ~NMODIFIED;
1372 mtx_unlock(&np->n_mtx);
1374 ncl_downgrade_vnlock(vp, old_lock);
1379 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1380 * This is mainly to avoid queueing async I/O requests when the nfsiods
1381 * are all hung on a dead server.
1383 * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1384 * is eventually dequeued by the async daemon, ncl_doio() *will*.
1387 ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td)
1396 * Commits are usually short and sweet so lets save some cpu and
1397 * leave the async daemons for more important rpc's (such as reads
1400 * Readdirplus RPCs do vget()s to acquire the vnodes for entries
1401 * in the directory in order to update attributes. This can deadlock
1402 * with another thread that is waiting for async I/O to be done by
1403 * an nfsiod thread while holding a lock on one of these vnodes.
1404 * To avoid this deadlock, don't allow the async nfsiod threads to
1405 * perform Readdirplus RPCs.
1407 mtx_lock(&ncl_iod_mutex);
1408 if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1409 (nmp->nm_bufqiods > ncl_numasync / 2)) ||
1410 (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) {
1411 mtx_unlock(&ncl_iod_mutex);
1415 if (nmp->nm_flag & NFSMNT_INT)
1420 * Find a free iod to process this request.
1422 for (iod = 0; iod < ncl_numasync; iod++)
1423 if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) {
1429 * Try to create one if none are free.
1435 * Found one, so wake it up and tell it which
1438 NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n",
1440 ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE;
1441 ncl_iodmount[iod] = nmp;
1443 wakeup(&ncl_iodwant[iod]);
1447 * If none are free, we may already have an iod working on this mount
1448 * point. If so, it will process our request.
1451 if (nmp->nm_bufqiods > 0) {
1453 ("ncl_asyncio: %d iods are already processing mount %p\n",
1454 nmp->nm_bufqiods, nmp));
1460 * If we have an iod which can process the request, then queue
1465 * Ensure that the queue never grows too large. We still want
1466 * to asynchronize so we block rather then return EIO.
1468 while (nmp->nm_bufqlen >= 2*ncl_numasync) {
1470 ("ncl_asyncio: waiting for mount %p queue to drain\n", nmp));
1471 nmp->nm_bufqwant = TRUE;
1472 error = newnfs_msleep(td, &nmp->nm_bufq,
1473 &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio",
1476 error2 = newnfs_sigintr(nmp, td);
1478 mtx_unlock(&ncl_iod_mutex);
1481 if (slpflag == PCATCH) {
1487 * We might have lost our iod while sleeping,
1488 * so check and loop if nescessary.
1493 /* We might have lost our nfsiod */
1494 if (nmp->nm_bufqiods == 0) {
1496 ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1500 if (bp->b_iocmd == BIO_READ) {
1501 if (bp->b_rcred == NOCRED && cred != NOCRED)
1502 bp->b_rcred = crhold(cred);
1504 if (bp->b_wcred == NOCRED && cred != NOCRED)
1505 bp->b_wcred = crhold(cred);
1508 if (bp->b_flags & B_REMFREE)
1511 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1513 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1514 mtx_lock(&(VTONFS(bp->b_vp))->n_mtx);
1515 VTONFS(bp->b_vp)->n_flag |= NMODIFIED;
1516 VTONFS(bp->b_vp)->n_directio_asyncwr++;
1517 mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx);
1519 mtx_unlock(&ncl_iod_mutex);
1523 mtx_unlock(&ncl_iod_mutex);
1526 * All the iods are busy on other mounts, so return EIO to
1527 * force the caller to process the i/o synchronously.
1529 NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n"));
1534 ncl_doio_directwrite(struct buf *bp)
1536 int iomode, must_commit;
1537 struct uio *uiop = (struct uio *)bp->b_caller1;
1538 char *iov_base = uiop->uio_iov->iov_base;
1540 iomode = NFSWRITE_FILESYNC;
1541 uiop->uio_td = NULL; /* NULL since we're in nfsiod */
1542 ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0);
1543 KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write"));
1544 free(iov_base, M_NFSDIRECTIO);
1545 free(uiop->uio_iov, M_NFSDIRECTIO);
1546 free(uiop, M_NFSDIRECTIO);
1547 if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) {
1548 struct nfsnode *np = VTONFS(bp->b_vp);
1549 mtx_lock(&np->n_mtx);
1550 if (NFSHASPNFS(VFSTONFS(vnode_mount(bp->b_vp)))) {
1552 * Invalidate the attribute cache, since writes to a DS
1553 * won't update the size attribute.
1555 np->n_attrstamp = 0;
1557 np->n_directio_asyncwr--;
1558 if (np->n_directio_asyncwr == 0) {
1559 np->n_flag &= ~NMODIFIED;
1560 if ((np->n_flag & NFSYNCWAIT)) {
1561 np->n_flag &= ~NFSYNCWAIT;
1562 wakeup((caddr_t)&np->n_directio_asyncwr);
1565 mtx_unlock(&np->n_mtx);
1568 relpbuf(bp, &ncl_pbuf_freecnt);
1572 * Do an I/O operation to/from a cache block. This may be called
1573 * synchronously or from an nfsiod.
1576 ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td,
1577 int called_from_strategy)
1581 struct nfsmount *nmp;
1582 int error = 0, iomode, must_commit = 0;
1585 struct proc *p = td ? td->td_proc : NULL;
1589 nmp = VFSTONFS(vp->v_mount);
1591 uiop->uio_iov = &io;
1592 uiop->uio_iovcnt = 1;
1593 uiop->uio_segflg = UIO_SYSSPACE;
1597 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We
1598 * do this here so we do not have to do it in all the code that
1601 bp->b_flags &= ~B_INVAL;
1602 bp->b_ioflags &= ~BIO_ERROR;
1604 KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp));
1605 iocmd = bp->b_iocmd;
1606 if (iocmd == BIO_READ) {
1607 io.iov_len = uiop->uio_resid = bp->b_bcount;
1608 io.iov_base = bp->b_data;
1609 uiop->uio_rw = UIO_READ;
1611 switch (vp->v_type) {
1613 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1614 NFSINCRGLOBAL(newnfsstats.read_bios);
1615 error = ncl_readrpc(vp, uiop, cr);
1618 if (uiop->uio_resid) {
1620 * If we had a short read with no error, we must have
1621 * hit a file hole. We should zero-fill the remainder.
1622 * This can also occur if the server hits the file EOF.
1624 * Holes used to be able to occur due to pending
1625 * writes, but that is not possible any longer.
1627 int nread = bp->b_bcount - uiop->uio_resid;
1628 ssize_t left = uiop->uio_resid;
1631 bzero((char *)bp->b_data + nread, left);
1632 uiop->uio_resid = 0;
1635 /* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */
1636 if (p && (vp->v_vflag & VV_TEXT)) {
1637 mtx_lock(&np->n_mtx);
1638 if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) {
1639 mtx_unlock(&np->n_mtx);
1641 killproc(p, "text file modification");
1644 mtx_unlock(&np->n_mtx);
1648 uiop->uio_offset = (off_t)0;
1649 NFSINCRGLOBAL(newnfsstats.readlink_bios);
1650 error = ncl_readlinkrpc(vp, uiop, cr);
1653 NFSINCRGLOBAL(newnfsstats.readdir_bios);
1654 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1655 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) {
1656 error = ncl_readdirplusrpc(vp, uiop, cr, td);
1657 if (error == NFSERR_NOTSUPP)
1658 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1660 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1661 error = ncl_readdirrpc(vp, uiop, cr, td);
1663 * end-of-directory sets B_INVAL but does not generate an
1666 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1667 bp->b_flags |= B_INVAL;
1670 ncl_printf("ncl_doio: type %x unexpected\n", vp->v_type);
1674 bp->b_ioflags |= BIO_ERROR;
1675 bp->b_error = error;
1679 * If we only need to commit, try to commit
1681 if (bp->b_flags & B_NEEDCOMMIT) {
1685 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1686 retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1689 bp->b_dirtyoff = bp->b_dirtyend = 0;
1690 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1695 if (retv == NFSERR_STALEWRITEVERF) {
1696 ncl_clearcommit(vp->v_mount);
1701 * Setup for actual write
1703 mtx_lock(&np->n_mtx);
1704 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1705 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1706 mtx_unlock(&np->n_mtx);
1708 if (bp->b_dirtyend > bp->b_dirtyoff) {
1709 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1711 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1713 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1714 uiop->uio_rw = UIO_WRITE;
1715 NFSINCRGLOBAL(newnfsstats.write_bios);
1717 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1718 iomode = NFSWRITE_UNSTABLE;
1720 iomode = NFSWRITE_FILESYNC;
1722 error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit,
1723 called_from_strategy);
1726 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1727 * to cluster the buffers needing commit. This will allow
1728 * the system to submit a single commit rpc for the whole
1729 * cluster. We can do this even if the buffer is not 100%
1730 * dirty (relative to the NFS blocksize), so we optimize the
1731 * append-to-file-case.
1733 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1734 * cleared because write clustering only works for commit
1735 * rpc's, not for the data portion of the write).
1738 if (!error && iomode == NFSWRITE_UNSTABLE) {
1739 bp->b_flags |= B_NEEDCOMMIT;
1740 if (bp->b_dirtyoff == 0
1741 && bp->b_dirtyend == bp->b_bcount)
1742 bp->b_flags |= B_CLUSTEROK;
1744 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1748 * For an interrupted write, the buffer is still valid
1749 * and the write hasn't been pushed to the server yet,
1750 * so we can't set BIO_ERROR and report the interruption
1751 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1752 * is not relevant, so the rpc attempt is essentially
1753 * a noop. For the case of a V3 write rpc not being
1754 * committed to stable storage, the block is still
1755 * dirty and requires either a commit rpc or another
1756 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1757 * the block is reused. This is indicated by setting
1758 * the B_DELWRI and B_NEEDCOMMIT flags.
1760 * EIO is returned by ncl_writerpc() to indicate a recoverable
1761 * write error and is handled as above, except that
1762 * B_EINTR isn't set. One cause of this is a stale stateid
1763 * error for the RPC that indicates recovery is required,
1764 * when called with called_from_strategy != 0.
1766 * If the buffer is marked B_PAGING, it does not reside on
1767 * the vp's paging queues so we cannot call bdirty(). The
1768 * bp in this case is not an NFS cache block so we should
1771 * The logic below breaks up errors into recoverable and
1772 * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE
1773 * and keep the buffer around for potential write retries.
1774 * For the latter (eg ESTALE), we toss the buffer away (B_INVAL)
1775 * and save the error in the nfsnode. This is less than ideal
1776 * but necessary. Keeping such buffers around could potentially
1777 * cause buffer exhaustion eventually (they can never be written
1778 * out, so will get constantly be re-dirtied). It also causes
1779 * all sorts of vfs panics. For non-recoverable write errors,
1780 * also invalidate the attrcache, so we'll be forced to go over
1781 * the wire for this object, returning an error to user on next
1782 * call (most of the time).
1784 if (error == EINTR || error == EIO || error == ETIMEDOUT
1785 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1789 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1790 if ((bp->b_flags & B_PAGING) == 0) {
1792 bp->b_flags &= ~B_DONE;
1794 if ((error == EINTR || error == ETIMEDOUT) &&
1795 (bp->b_flags & B_ASYNC) == 0)
1796 bp->b_flags |= B_EINTR;
1800 bp->b_ioflags |= BIO_ERROR;
1801 bp->b_flags |= B_INVAL;
1802 bp->b_error = np->n_error = error;
1803 mtx_lock(&np->n_mtx);
1804 np->n_flag |= NWRITEERR;
1805 np->n_attrstamp = 0;
1806 KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp);
1807 mtx_unlock(&np->n_mtx);
1809 bp->b_dirtyoff = bp->b_dirtyend = 0;
1817 bp->b_resid = uiop->uio_resid;
1819 ncl_clearcommit(vp->v_mount);
1825 * Used to aid in handling ftruncate() operations on the NFS client side.
1826 * Truncation creates a number of special problems for NFS. We have to
1827 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1828 * we have to properly handle VM pages or (potentially dirty) buffers
1829 * that straddle the truncation point.
1833 ncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1835 struct nfsnode *np = VTONFS(vp);
1837 int biosize = vp->v_bufobj.bo_bsize;
1840 mtx_lock(&np->n_mtx);
1843 mtx_unlock(&np->n_mtx);
1845 if (nsize < tsize) {
1851 * vtruncbuf() doesn't get the buffer overlapping the
1852 * truncation point. We may have a B_DELWRI and/or B_CACHE
1853 * buffer that now needs to be truncated.
1855 error = vtruncbuf(vp, cred, nsize, biosize);
1856 lbn = nsize / biosize;
1857 bufsize = nsize - (lbn * biosize);
1858 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1861 if (bp->b_dirtyoff > bp->b_bcount)
1862 bp->b_dirtyoff = bp->b_bcount;
1863 if (bp->b_dirtyend > bp->b_bcount)
1864 bp->b_dirtyend = bp->b_bcount;
1865 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1868 vnode_pager_setsize(vp, nsize);