2 * Copyright (c) 1989, 1993
3 * The Regents of the University of California. All rights reserved.
5 * This code is derived from software contributed to Berkeley by
6 * Rick Macklem at The University of Guelph.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 3. All advertising materials mentioning features or use of this software
17 * must display the following acknowledgement:
18 * This product includes software developed by the University of
19 * California, Berkeley and its contributors.
20 * 4. Neither the name of the University nor the names of its contributors
21 * may be used to endorse or promote products derived from this software
22 * without specific prior written permission.
24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * @(#)nfs_bio.c 8.9 (Berkeley) 3/30/95
37 * $FreeBSD: /repoman/r/ncvs/src/sys/nfsclient/nfs_bio.c,v 1.130 2004/04/14 23:23:55 peadar Exp $
38 * $DragonFly: src/sys/vfs/nfs/nfs_bio.c,v 1.45 2008/07/18 00:09:39 dillon Exp $
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/resourcevar.h>
45 #include <sys/signalvar.h>
48 #include <sys/vnode.h>
49 #include <sys/mount.h>
50 #include <sys/kernel.h>
52 #include <sys/msfbuf.h>
55 #include <vm/vm_extern.h>
56 #include <vm/vm_page.h>
57 #include <vm/vm_object.h>
58 #include <vm/vm_pager.h>
59 #include <vm/vnode_pager.h>
61 #include <sys/thread2.h>
69 static struct buf *nfs_getcacheblk(struct vnode *vp, off_t loffset,
70 int size, struct thread *td);
71 static int nfs_check_dirent(struct nfs_dirent *dp, int maxlen);
72 static void nfsiodone_sync(struct bio *bio);
74 extern int nfs_numasync;
75 extern int nfs_pbuf_freecnt;
76 extern struct nfsstats nfsstats;
79 * Vnode op for VM getpages.
81 * nfs_getpages(struct vnode *a_vp, vm_page_t *a_m, int a_count,
82 * int a_reqpage, vm_ooffset_t a_offset)
85 nfs_getpages(struct vop_getpages_args *ap)
87 struct thread *td = curthread; /* XXX */
88 int i, error, nextoff, size, toff, count, npages;
99 nmp = VFSTONFS(vp->v_mount);
103 if (vp->v_object == NULL) {
104 kprintf("nfs_getpages: called with non-merged cache vnode??\n");
105 return VM_PAGER_ERROR;
108 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
109 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
110 (void)nfs_fsinfo(nmp, vp, td);
112 npages = btoc(count);
115 * NOTE that partially valid pages may occur in cases other
116 * then file EOF, such as when a file is partially written and
117 * ftruncate()-extended to a larger size. It is also possible
118 * for the valid bits to be set on garbage beyond the file EOF and
119 * clear in the area before EOF (e.g. m->valid == 0xfc), which can
120 * occur due to vtruncbuf() and the buffer cache's handling of
121 * pages which 'straddle' buffers or when b_bufsize is not a
122 * multiple of PAGE_SIZE.... the buffer cache cannot normally
123 * clear the extra bits. This kind of situation occurs when you
124 * make a small write() (m->valid == 0x03) and then mmap() and
125 * fault in the buffer(m->valid = 0xFF). When NFS flushes the
126 * buffer (vinvalbuf() m->valid = 0xFC) we are left with a mess.
128 * This is combined with the possibility that the pages are partially
129 * dirty or that there is a buffer backing the pages that is dirty
130 * (even if m->dirty is 0).
132 * To solve this problem several hacks have been made: (1) NFS
133 * guarentees that the IO block size is a multiple of PAGE_SIZE and
134 * (2) The buffer cache, when invalidating an NFS buffer, will
135 * disregard the buffer's fragmentory b_bufsize and invalidate
136 * the whole page rather then just the piece the buffer owns.
138 * This allows us to assume that a partially valid page found here
139 * is fully valid (vm_fault will zero'd out areas of the page not
142 m = pages[ap->a_reqpage];
144 for (i = 0; i < npages; ++i) {
145 if (i != ap->a_reqpage)
146 vnode_pager_freepage(pages[i]);
152 * Use an MSF_BUF as a medium to retrieve data from the pages.
154 msf_map_pagelist(&msf, pages, npages, 0);
156 kva = msf_buf_kva(msf);
162 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
163 uio.uio_resid = count;
164 uio.uio_segflg = UIO_SYSSPACE;
165 uio.uio_rw = UIO_READ;
168 error = nfs_readrpc(vp, &uio);
171 if (error && (uio.uio_resid == count)) {
172 kprintf("nfs_getpages: error %d\n", error);
173 for (i = 0; i < npages; ++i) {
174 if (i != ap->a_reqpage)
175 vnode_pager_freepage(pages[i]);
177 return VM_PAGER_ERROR;
181 * Calculate the number of bytes read and validate only that number
182 * of bytes. Note that due to pending writes, size may be 0. This
183 * does not mean that the remaining data is invalid!
186 size = count - uio.uio_resid;
188 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
189 nextoff = toff + PAGE_SIZE;
192 m->flags &= ~PG_ZERO;
194 if (nextoff <= size) {
196 * Read operation filled an entire page
198 m->valid = VM_PAGE_BITS_ALL;
200 } else if (size > toff) {
202 * Read operation filled a partial page.
205 vm_page_set_validclean(m, 0, size - toff);
206 /* handled by vm_fault now */
207 /* vm_page_zero_invalid(m, TRUE); */
210 * Read operation was short. If no error occured
211 * we may have hit a zero-fill section. We simply
212 * leave valid set to 0.
216 if (i != ap->a_reqpage) {
218 * Whether or not to leave the page activated is up in
219 * the air, but we should put the page on a page queue
220 * somewhere (it already is in the object). Result:
221 * It appears that emperical results show that
222 * deactivating pages is best.
226 * Just in case someone was asking for this page we
227 * now tell them that it is ok to use.
230 if (m->flags & PG_WANTED)
233 vm_page_deactivate(m);
236 vnode_pager_freepage(m);
244 * Vnode op for VM putpages.
246 * nfs_putpages(struct vnode *a_vp, vm_page_t *a_m, int a_count, int a_sync,
247 * int *a_rtvals, vm_ooffset_t a_offset)
250 nfs_putpages(struct vop_putpages_args *ap)
252 struct thread *td = curthread;
256 int iomode, must_commit, i, error, npages, count;
260 struct nfsmount *nmp;
267 nmp = VFSTONFS(vp->v_mount);
270 rtvals = ap->a_rtvals;
271 npages = btoc(count);
272 offset = IDX_TO_OFF(pages[0]->pindex);
274 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
275 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
276 (void)nfs_fsinfo(nmp, vp, td);
278 for (i = 0; i < npages; i++) {
279 rtvals[i] = VM_PAGER_AGAIN;
283 * When putting pages, do not extend file past EOF.
286 if (offset + count > np->n_size) {
287 count = np->n_size - offset;
293 * Use an MSF_BUF as a medium to retrieve data from the pages.
295 msf_map_pagelist(&msf, pages, npages, 0);
297 kva = msf_buf_kva(msf);
303 uio.uio_offset = offset;
304 uio.uio_resid = count;
305 uio.uio_segflg = UIO_SYSSPACE;
306 uio.uio_rw = UIO_WRITE;
309 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
310 iomode = NFSV3WRITE_UNSTABLE;
312 iomode = NFSV3WRITE_FILESYNC;
314 error = nfs_writerpc(vp, &uio, &iomode, &must_commit);
319 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE;
320 for (i = 0; i < nwritten; i++) {
321 rtvals[i] = VM_PAGER_OK;
322 vm_page_undirty(pages[i]);
325 nfs_clearcommit(vp->v_mount);
331 * Vnode op for read using bio
334 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag)
336 struct nfsnode *np = VTONFS(vp);
338 struct buf *bp = 0, *rabp;
341 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
347 int nra, error = 0, n = 0, on = 0;
350 if (uio->uio_rw != UIO_READ)
351 panic("nfs_read mode");
353 if (uio->uio_resid == 0)
355 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
359 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
360 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
361 (void)nfs_fsinfo(nmp, vp, td);
362 if (vp->v_type != VDIR &&
363 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
365 biosize = vp->v_mount->mnt_stat.f_iosize;
366 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
369 * For nfs, cache consistency can only be maintained approximately.
370 * Although RFC1094 does not specify the criteria, the following is
371 * believed to be compatible with the reference port.
373 * NFS: If local changes have been made and this is a
374 * directory, the directory must be invalidated and
375 * the attribute cache must be cleared.
377 * GETATTR is called to synchronize the file size.
379 * If remote changes are detected local data is flushed
380 * and the cache is invalidated.
382 * NOTE: In the normal case the attribute cache is not
383 * cleared which means GETATTR may use cached data and
384 * not immediately detect changes made on the server.
386 if ((np->n_flag & NLMODIFIED) && vp->v_type == VDIR) {
388 error = nfs_vinvalbuf(vp, V_SAVE, 1);
393 error = VOP_GETATTR(vp, &vattr);
396 if (np->n_flag & NRMODIFIED) {
397 if (vp->v_type == VDIR)
399 error = nfs_vinvalbuf(vp, V_SAVE, 1);
402 np->n_flag &= ~NRMODIFIED;
405 if (np->n_flag & NDONTCACHE) {
406 switch (vp->v_type) {
408 return (nfs_readrpc(vp, uio));
410 return (nfs_readlinkrpc(vp, uio));
414 kprintf(" NDONTCACHE: type %x unexpected\n", vp->v_type);
418 switch (vp->v_type) {
420 nfsstats.biocache_reads++;
421 lbn = uio->uio_offset / biosize;
422 on = uio->uio_offset & (biosize - 1);
423 loffset = (off_t)lbn * biosize;
426 * Start the read ahead(s), as required.
428 if (nfs_numasync > 0 && nmp->nm_readahead > 0) {
429 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
430 (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) {
431 rabn = lbn + 1 + nra;
432 raoffset = (off_t)rabn * biosize;
433 if (findblk(vp, raoffset, FINDBLK_TEST) == NULL) {
434 rabp = nfs_getcacheblk(vp, raoffset, biosize, td);
437 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
438 rabp->b_cmd = BUF_CMD_READ;
439 vfs_busy_pages(vp, rabp);
440 if (nfs_asyncio(vp, &rabp->b_bio2, td)) {
441 rabp->b_flags |= B_INVAL|B_ERROR;
442 vfs_unbusy_pages(rabp);
454 * Obtain the buffer cache block. Figure out the buffer size
455 * when we are at EOF. If we are modifying the size of the
456 * buffer based on an EOF condition we need to hold
457 * nfs_rslock() through obtaining the buffer to prevent
458 * a potential writer-appender from messing with n_size.
459 * Otherwise we may accidently truncate the buffer and
462 * Note that bcount is *not* DEV_BSIZE aligned.
467 if (loffset >= np->n_size) {
469 } else if (loffset + biosize > np->n_size) {
470 bcount = np->n_size - loffset;
472 if (bcount != biosize) {
473 switch(nfs_rslock(np)) {
486 bp = nfs_getcacheblk(vp, loffset, bcount, td);
488 if (bcount != biosize)
494 * If B_CACHE is not set, we must issue the read. If this
495 * fails, we return an error.
498 if ((bp->b_flags & B_CACHE) == 0) {
499 bp->b_cmd = BUF_CMD_READ;
500 bp->b_bio2.bio_done = nfsiodone_sync;
501 bp->b_bio2.bio_flags |= BIO_SYNC;
502 vfs_busy_pages(vp, bp);
503 error = nfs_doio(vp, &bp->b_bio2, td);
511 * on is the offset into the current bp. Figure out how many
512 * bytes we can copy out of the bp. Note that bcount is
513 * NOT DEV_BSIZE aligned.
515 * Then figure out how many bytes we can copy into the uio.
520 n = min((unsigned)(bcount - on), uio->uio_resid);
523 biosize = min(NFS_MAXPATHLEN, np->n_size);
524 nfsstats.biocache_readlinks++;
525 bp = nfs_getcacheblk(vp, (off_t)0, biosize, td);
528 if ((bp->b_flags & B_CACHE) == 0) {
529 bp->b_cmd = BUF_CMD_READ;
530 bp->b_bio2.bio_done = nfsiodone_sync;
531 bp->b_bio2.bio_flags |= BIO_SYNC;
532 vfs_busy_pages(vp, bp);
533 error = nfs_doio(vp, &bp->b_bio2, td);
535 bp->b_flags |= B_ERROR | B_INVAL;
540 n = min(uio->uio_resid, bp->b_bcount - bp->b_resid);
544 nfsstats.biocache_readdirs++;
545 if (np->n_direofoffset
546 && uio->uio_offset >= np->n_direofoffset) {
549 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
550 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
551 loffset = uio->uio_offset - on;
552 bp = nfs_getcacheblk(vp, loffset, NFS_DIRBLKSIZ, td);
556 if ((bp->b_flags & B_CACHE) == 0) {
557 bp->b_cmd = BUF_CMD_READ;
558 bp->b_bio2.bio_done = nfsiodone_sync;
559 bp->b_bio2.bio_flags |= BIO_SYNC;
560 vfs_busy_pages(vp, bp);
561 error = nfs_doio(vp, &bp->b_bio2, td);
565 while (error == NFSERR_BAD_COOKIE) {
566 kprintf("got bad cookie vp %p bp %p\n", vp, bp);
568 error = nfs_vinvalbuf(vp, 0, 1);
570 * Yuck! The directory has been modified on the
571 * server. The only way to get the block is by
572 * reading from the beginning to get all the
575 * Leave the last bp intact unless there is an error.
576 * Loop back up to the while if the error is another
577 * NFSERR_BAD_COOKIE (double yuch!).
579 for (i = 0; i <= lbn && !error; i++) {
580 if (np->n_direofoffset
581 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
583 bp = nfs_getcacheblk(vp, (off_t)i * NFS_DIRBLKSIZ,
587 if ((bp->b_flags & B_CACHE) == 0) {
588 bp->b_cmd = BUF_CMD_READ;
589 bp->b_bio2.bio_done = nfsiodone_sync;
590 bp->b_bio2.bio_flags |= BIO_SYNC;
591 vfs_busy_pages(vp, bp);
592 error = nfs_doio(vp, &bp->b_bio2, td);
594 * no error + B_INVAL == directory EOF,
597 if (error == 0 && (bp->b_flags & B_INVAL))
601 * An error will throw away the block and the
602 * for loop will break out. If no error and this
603 * is not the block we want, we throw away the
604 * block and go for the next one via the for loop.
606 if (error || i < lbn)
611 * The above while is repeated if we hit another cookie
612 * error. If we hit an error and it wasn't a cookie error,
620 * If not eof and read aheads are enabled, start one.
621 * (You need the current block first, so that you have the
622 * directory offset cookie of the next block.)
624 if (nfs_numasync > 0 && nmp->nm_readahead > 0 &&
625 (bp->b_flags & B_INVAL) == 0 &&
626 (np->n_direofoffset == 0 ||
627 loffset + NFS_DIRBLKSIZ < np->n_direofoffset) &&
628 (np->n_flag & NDONTCACHE) == 0 &&
629 findblk(vp, loffset + NFS_DIRBLKSIZ, FINDBLK_TEST) == NULL
631 rabp = nfs_getcacheblk(vp, loffset + NFS_DIRBLKSIZ,
634 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
635 rabp->b_cmd = BUF_CMD_READ;
636 vfs_busy_pages(vp, rabp);
637 if (nfs_asyncio(vp, &rabp->b_bio2, td)) {
638 rabp->b_flags |= B_INVAL|B_ERROR;
639 vfs_unbusy_pages(rabp);
648 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
649 * chopped for the EOF condition, we cannot tell how large
650 * NFS directories are going to be until we hit EOF. So
651 * an NFS directory buffer is *not* chopped to its EOF. Now,
652 * it just so happens that b_resid will effectively chop it
653 * to EOF. *BUT* this information is lost if the buffer goes
654 * away and is reconstituted into a B_CACHE state ( due to
655 * being VMIO ) later. So we keep track of the directory eof
656 * in np->n_direofoffset and chop it off as an extra step
659 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
660 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
661 n = np->n_direofoffset - uio->uio_offset;
664 kprintf(" nfs_bioread: type %x unexpected\n",vp->v_type);
668 switch (vp->v_type) {
671 error = uiomove(bp->b_data + on, (int)n, uio);
675 error = uiomove(bp->b_data + on, (int)n, uio);
680 off_t old_off = uio->uio_offset;
682 struct nfs_dirent *dp;
685 * We are casting cpos to nfs_dirent, it must be
693 cpos = bp->b_data + on;
694 epos = bp->b_data + on + n;
695 while (cpos < epos && error == 0 && uio->uio_resid > 0) {
696 dp = (struct nfs_dirent *)cpos;
697 error = nfs_check_dirent(dp, (int)(epos - cpos));
700 if (vop_write_dirent(&error, uio, dp->nfs_ino,
701 dp->nfs_type, dp->nfs_namlen, dp->nfs_name)) {
704 cpos += dp->nfs_reclen;
708 uio->uio_offset = old_off + cpos - bp->b_data - on;
711 * Invalidate buffer if caching is disabled, forcing a
712 * re-read from the remote later.
714 if (np->n_flag & NDONTCACHE)
715 bp->b_flags |= B_INVAL;
718 kprintf(" nfs_bioread: type %x unexpected\n",vp->v_type);
721 } while (error == 0 && uio->uio_resid > 0 && n > 0);
726 * Userland can supply any 'seek' offset when reading a NFS directory.
727 * Validate the structure so we don't panic the kernel. Note that
728 * the element name is nul terminated and the nul is not included
733 nfs_check_dirent(struct nfs_dirent *dp, int maxlen)
735 int nfs_name_off = offsetof(struct nfs_dirent, nfs_name[0]);
737 if (nfs_name_off >= maxlen)
739 if (dp->nfs_reclen < nfs_name_off || dp->nfs_reclen > maxlen)
741 if (nfs_name_off + dp->nfs_namlen >= dp->nfs_reclen)
743 if (dp->nfs_reclen & 3)
749 * Vnode op for write using bio
751 * nfs_write(struct vnode *a_vp, struct uio *a_uio, int a_ioflag,
752 * struct ucred *a_cred)
755 nfs_write(struct vop_write_args *ap)
757 struct uio *uio = ap->a_uio;
758 struct thread *td = uio->uio_td;
759 struct vnode *vp = ap->a_vp;
760 struct nfsnode *np = VTONFS(vp);
761 int ioflag = ap->a_ioflag;
764 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
767 int n, on, error = 0, iomode, must_commit;
773 if (uio->uio_rw != UIO_WRITE)
774 panic("nfs_write mode");
775 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread)
776 panic("nfs_write proc");
778 if (vp->v_type != VREG)
780 if (np->n_flag & NWRITEERR) {
781 np->n_flag &= ~NWRITEERR;
782 return (np->n_error);
784 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
785 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
786 (void)nfs_fsinfo(nmp, vp, td);
789 * Synchronously flush pending buffers if we are in synchronous
790 * mode or if we are appending.
792 if (ioflag & (IO_APPEND | IO_SYNC)) {
793 if (np->n_flag & NLMODIFIED) {
795 error = nfs_flush(vp, MNT_WAIT, td, 0);
796 /* error = nfs_vinvalbuf(vp, V_SAVE, 1); */
803 * If IO_APPEND then load uio_offset. We restart here if we cannot
804 * get the append lock.
807 if (ioflag & IO_APPEND) {
809 error = VOP_GETATTR(vp, &vattr);
812 uio->uio_offset = np->n_size;
815 if (uio->uio_offset < 0)
817 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
819 if (uio->uio_resid == 0)
823 * We need to obtain the rslock if we intend to modify np->n_size
824 * in order to guarentee the append point with multiple contending
825 * writers, to guarentee that no other appenders modify n_size
826 * while we are trying to obtain a truncated buffer (i.e. to avoid
827 * accidently truncating data written by another appender due to
828 * the race), and to ensure that the buffer is populated prior to
829 * our extending of the file. We hold rslock through the entire
832 * Note that we do not synchronize the case where someone truncates
833 * the file while we are appending to it because attempting to lock
834 * this case may deadlock other parts of the system unexpectedly.
836 if ((ioflag & IO_APPEND) ||
837 uio->uio_offset + uio->uio_resid > np->n_size) {
838 switch(nfs_rslock(np)) {
853 * Maybe this should be above the vnode op call, but so long as
854 * file servers have no limits, i don't think it matters
856 if (td->td_proc && uio->uio_offset + uio->uio_resid >
857 td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
858 lwpsignal(td->td_proc, td->td_lwp, SIGXFSZ);
864 biosize = vp->v_mount->mnt_stat.f_iosize;
867 if ((np->n_flag & NDONTCACHE) && uio->uio_iovcnt == 1) {
868 iomode = NFSV3WRITE_FILESYNC;
869 error = nfs_writerpc(vp, uio, &iomode, &must_commit);
871 nfs_clearcommit(vp->v_mount);
874 nfsstats.biocache_writes++;
875 lbn = uio->uio_offset / biosize;
876 on = uio->uio_offset & (biosize-1);
877 loffset = uio->uio_offset - on;
878 n = min((unsigned)(biosize - on), uio->uio_resid);
881 * Handle direct append and file extension cases, calculate
882 * unaligned buffer size.
885 if (uio->uio_offset == np->n_size && n) {
887 * Get the buffer (in its pre-append state to maintain
888 * B_CACHE if it was previously set). Resize the
889 * nfsnode after we have locked the buffer to prevent
890 * readers from reading garbage.
893 bp = nfs_getcacheblk(vp, loffset, bcount, td);
898 np->n_size = uio->uio_offset + n;
899 np->n_flag |= NLMODIFIED;
900 vnode_pager_setsize(vp, np->n_size);
902 save = bp->b_flags & B_CACHE;
904 allocbuf(bp, bcount);
909 * Obtain the locked cache block first, and then
910 * adjust the file's size as appropriate.
913 if (loffset + bcount < np->n_size) {
914 if (loffset + biosize < np->n_size)
917 bcount = np->n_size - loffset;
919 bp = nfs_getcacheblk(vp, loffset, bcount, td);
920 if (uio->uio_offset + n > np->n_size) {
921 np->n_size = uio->uio_offset + n;
922 np->n_flag |= NLMODIFIED;
923 vnode_pager_setsize(vp, np->n_size);
933 * Issue a READ if B_CACHE is not set. In special-append
934 * mode, B_CACHE is based on the buffer prior to the write
935 * op and is typically set, avoiding the read. If a read
936 * is required in special append mode, the server will
937 * probably send us a short-read since we extended the file
938 * on our end, resulting in b_resid == 0 and, thusly,
939 * B_CACHE getting set.
941 * We can also avoid issuing the read if the write covers
942 * the entire buffer. We have to make sure the buffer state
943 * is reasonable in this case since we will not be initiating
944 * I/O. See the comments in kern/vfs_bio.c's getblk() for
947 * B_CACHE may also be set due to the buffer being cached
950 * When doing a UIO_NOCOPY write the buffer is not
951 * overwritten and we cannot just set B_CACHE unconditionally
952 * for full-block writes.
955 if (on == 0 && n == bcount && uio->uio_segflg != UIO_NOCOPY) {
956 bp->b_flags |= B_CACHE;
957 bp->b_flags &= ~(B_ERROR | B_INVAL);
960 if ((bp->b_flags & B_CACHE) == 0) {
961 bp->b_cmd = BUF_CMD_READ;
962 bp->b_bio2.bio_done = nfsiodone_sync;
963 bp->b_bio2.bio_flags |= BIO_SYNC;
964 vfs_busy_pages(vp, bp);
965 error = nfs_doio(vp, &bp->b_bio2, td);
975 np->n_flag |= NLMODIFIED;
978 * If dirtyend exceeds file size, chop it down. This should
979 * not normally occur but there is an append race where it
980 * might occur XXX, so we log it.
982 * If the chopping creates a reverse-indexed or degenerate
983 * situation with dirtyoff/end, we 0 both of them.
986 if (bp->b_dirtyend > bcount) {
987 kprintf("NFS append race @%08llx:%d\n",
988 (long long)bp->b_bio2.bio_offset,
989 bp->b_dirtyend - bcount);
990 bp->b_dirtyend = bcount;
993 if (bp->b_dirtyoff >= bp->b_dirtyend)
994 bp->b_dirtyoff = bp->b_dirtyend = 0;
997 * If the new write will leave a contiguous dirty
998 * area, just update the b_dirtyoff and b_dirtyend,
999 * otherwise force a write rpc of the old dirty area.
1001 * While it is possible to merge discontiguous writes due to
1002 * our having a B_CACHE buffer ( and thus valid read data
1003 * for the hole), we don't because it could lead to
1004 * significant cache coherency problems with multiple clients,
1005 * especially if locking is implemented later on.
1007 * as an optimization we could theoretically maintain
1008 * a linked list of discontinuous areas, but we would still
1009 * have to commit them separately so there isn't much
1010 * advantage to it except perhaps a bit of asynchronization.
1013 if (bp->b_dirtyend > 0 &&
1014 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1015 if (bwrite(bp) == EINTR) {
1022 error = uiomove((char *)bp->b_data + on, n, uio);
1025 * Since this block is being modified, it must be written
1026 * again and not just committed. Since write clustering does
1027 * not work for the stage 1 data write, only the stage 2
1028 * commit rpc, we have to clear B_CLUSTEROK as well.
1030 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1033 bp->b_flags |= B_ERROR;
1039 * Only update dirtyoff/dirtyend if not a degenerate
1043 if (bp->b_dirtyend > 0) {
1044 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1045 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1047 bp->b_dirtyoff = on;
1048 bp->b_dirtyend = on + n;
1050 vfs_bio_set_validclean(bp, on, n);
1054 * If the lease is non-cachable or IO_SYNC do bwrite().
1056 * IO_INVAL appears to be unused. The idea appears to be
1057 * to turn off caching in this case. Very odd. XXX
1059 * If nfs_async is set bawrite() will use an unstable write
1060 * (build dirty bufs on the server), so we might as well
1061 * push it out with bawrite(). If nfs_async is not set we
1062 * use bdwrite() to cache dirty bufs on the client.
1064 if ((np->n_flag & NDONTCACHE) || (ioflag & IO_SYNC)) {
1065 if (ioflag & IO_INVAL)
1066 bp->b_flags |= B_NOCACHE;
1070 if (np->n_flag & NDONTCACHE) {
1071 error = nfs_vinvalbuf(vp, V_SAVE, 1);
1075 } else if ((n + on) == biosize && nfs_async) {
1080 } while (uio->uio_resid > 0 && n > 0);
1089 * Get an nfs cache block.
1091 * Allocate a new one if the block isn't currently in the cache
1092 * and return the block marked busy. If the calling process is
1093 * interrupted by a signal for an interruptible mount point, return
1096 * The caller must carefully deal with the possible B_INVAL state of
1097 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it
1098 * indirectly), so synchronous reads can be issued without worrying about
1099 * the B_INVAL state. We have to be a little more careful when dealing
1100 * with writes (see comments in nfs_write()) when extending a file past
1104 nfs_getcacheblk(struct vnode *vp, off_t loffset, int size, struct thread *td)
1108 struct nfsmount *nmp;
1113 if (nmp->nm_flag & NFSMNT_INT) {
1114 bp = getblk(vp, loffset, size, GETBLK_PCATCH, 0);
1115 while (bp == NULL) {
1116 if (nfs_sigintr(nmp, NULL, td))
1118 bp = getblk(vp, loffset, size, 0, 2 * hz);
1121 bp = getblk(vp, loffset, size, 0, 0);
1125 * bio2, the 'device' layer. Since BIOs use 64 bit byte offsets
1126 * now, no translation is necessary.
1128 bp->b_bio2.bio_offset = loffset;
1133 * Flush and invalidate all dirty buffers. If another process is already
1134 * doing the flush, just wait for completion.
1137 nfs_vinvalbuf(struct vnode *vp, int flags, int intrflg)
1139 struct nfsnode *np = VTONFS(vp);
1140 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1141 int error = 0, slpflag, slptimeo;
1142 thread_t td = curthread;
1144 if (vp->v_flag & VRECLAIMED)
1147 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1157 * First wait for any other process doing a flush to complete.
1159 while (np->n_flag & NFLUSHINPROG) {
1160 np->n_flag |= NFLUSHWANT;
1161 error = tsleep((caddr_t)&np->n_flag, 0, "nfsvinval", slptimeo);
1162 if (error && intrflg && nfs_sigintr(nmp, NULL, td))
1167 * Now, flush as required.
1169 np->n_flag |= NFLUSHINPROG;
1170 error = vinvalbuf(vp, flags, slpflag, 0);
1172 if (intrflg && nfs_sigintr(nmp, NULL, td)) {
1173 np->n_flag &= ~NFLUSHINPROG;
1174 if (np->n_flag & NFLUSHWANT) {
1175 np->n_flag &= ~NFLUSHWANT;
1176 wakeup((caddr_t)&np->n_flag);
1180 error = vinvalbuf(vp, flags, 0, slptimeo);
1182 np->n_flag &= ~(NLMODIFIED | NFLUSHINPROG);
1183 if (np->n_flag & NFLUSHWANT) {
1184 np->n_flag &= ~NFLUSHWANT;
1185 wakeup((caddr_t)&np->n_flag);
1191 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1192 * This is mainly to avoid queueing async I/O requests when the nfsiods
1193 * are all hung on a dead server.
1195 * Note: nfs_asyncio() does not clear (B_ERROR|B_INVAL) but when the bp
1196 * is eventually dequeued by the async daemon, nfs_doio() *will*.
1199 nfs_asyncio(struct vnode *vp, struct bio *bio, struct thread *td)
1201 struct buf *bp = bio->bio_buf;
1202 struct nfsmount *nmp;
1210 * If no async daemons then return EIO to force caller to run the rpc
1213 if (nfs_numasync == 0)
1216 KKASSERT(vp->v_tag == VT_NFS);
1217 nmp = VFSTONFS(vp->v_mount);
1220 * Commits are usually short and sweet so lets save some cpu and
1221 * leave the async daemons for more important rpc's (such as reads
1224 if (bp->b_cmd == BUF_CMD_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1225 (nmp->nm_bioqiods > nfs_numasync / 2)) {
1230 if (nmp->nm_flag & NFSMNT_INT)
1235 * Find a free iod to process this request.
1237 for (i = 0; i < NFS_MAXASYNCDAEMON; i++)
1238 if (nfs_iodwant[i]) {
1240 * Found one, so wake it up and tell it which
1244 ("nfs_asyncio: waking iod %d for mount %p\n",
1246 nfs_iodwant[i] = NULL;
1247 nfs_iodmount[i] = nmp;
1249 wakeup((caddr_t)&nfs_iodwant[i]);
1255 * If none are free, we may already have an iod working on this mount
1256 * point. If so, it will process our request.
1259 if (nmp->nm_bioqiods > 0) {
1261 ("nfs_asyncio: %d iods are already processing mount %p\n",
1262 nmp->nm_bioqiods, nmp));
1268 * If we have an iod which can process the request, then queue
1273 * Ensure that the queue never grows too large. We still want
1274 * to asynchronize so we block rather then return EIO.
1276 while (nmp->nm_bioqlen >= 2*nfs_numasync) {
1278 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp));
1279 nmp->nm_bioqwant = TRUE;
1280 error = tsleep(&nmp->nm_bioq, slpflag,
1281 "nfsaio", slptimeo);
1283 if (nfs_sigintr(nmp, NULL, td))
1285 if (slpflag == PCATCH) {
1291 * We might have lost our iod while sleeping,
1292 * so check and loop if nescessary.
1294 if (nmp->nm_bioqiods == 0) {
1296 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1303 * The passed bio's buffer is not necessary associated with
1304 * the NFS vnode it is being written to. Store the NFS vnode
1305 * in the BIO driver info.
1307 bio->bio_driver_info = vp;
1308 TAILQ_INSERT_TAIL(&nmp->nm_bioq, bio, bio_act);
1314 * All the iods are busy on other mounts, so return EIO to
1315 * force the caller to process the i/o synchronously.
1317 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n"));
1322 * Do an I/O operation to/from a cache block. This may be called
1323 * synchronously or from an nfsiod. The BIO is normalized for DEV_BSIZE.
1325 * A locked, completed I/O is returned and the caller is responsible for
1328 * NOTE! TD MIGHT BE NULL
1331 nfs_doio(struct vnode *vp, struct bio *bio, struct thread *td)
1333 struct buf *bp = bio->bio_buf;
1336 struct nfsmount *nmp;
1337 int error = 0, iomode, must_commit = 0;
1341 KKASSERT(vp->v_tag == VT_NFS);
1343 nmp = VFSTONFS(vp->v_mount);
1345 uiop->uio_iov = &io;
1346 uiop->uio_iovcnt = 1;
1347 uiop->uio_segflg = UIO_SYSSPACE;
1351 * clear B_ERROR and B_INVAL state prior to initiating the I/O. We
1352 * do this here so we do not have to do it in all the code that
1355 bp->b_flags &= ~(B_ERROR | B_INVAL);
1358 KASSERT(bp->b_cmd != BUF_CMD_DONE,
1359 ("nfs_doio: bp %p already marked done!", bp));
1361 if (bp->b_cmd == BUF_CMD_READ) {
1362 io.iov_len = uiop->uio_resid = bp->b_bcount;
1363 io.iov_base = bp->b_data;
1364 uiop->uio_rw = UIO_READ;
1366 switch (vp->v_type) {
1368 uiop->uio_offset = bio->bio_offset;
1369 nfsstats.read_bios++;
1370 error = nfs_readrpc(vp, uiop);
1373 if (uiop->uio_resid) {
1375 * If we had a short read with no error, we must have
1376 * hit a file hole. We should zero-fill the remainder.
1377 * This can also occur if the server hits the file EOF.
1379 * Holes used to be able to occur due to pending
1380 * writes, but that is not possible any longer.
1382 int nread = bp->b_bcount - uiop->uio_resid;
1383 int left = uiop->uio_resid;
1386 bzero((char *)bp->b_data + nread, left);
1387 uiop->uio_resid = 0;
1390 if (td && td->td_proc && (vp->v_flag & VTEXT) &&
1391 np->n_mtime != np->n_vattr.va_mtime.tv_sec) {
1392 uprintf("Process killed due to text file modification\n");
1393 ksignal(td->td_proc, SIGKILL);
1397 uiop->uio_offset = 0;
1398 nfsstats.readlink_bios++;
1399 error = nfs_readlinkrpc(vp, uiop);
1402 nfsstats.readdir_bios++;
1403 uiop->uio_offset = bio->bio_offset;
1404 if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
1405 error = nfs_readdirplusrpc(vp, uiop);
1406 if (error == NFSERR_NOTSUPP)
1407 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1409 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1410 error = nfs_readdirrpc(vp, uiop);
1412 * end-of-directory sets B_INVAL but does not generate an
1415 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1416 bp->b_flags |= B_INVAL;
1419 kprintf("nfs_doio: type %x unexpected\n",vp->v_type);
1423 bp->b_flags |= B_ERROR;
1424 bp->b_error = error;
1428 * If we only need to commit, try to commit
1430 KKASSERT(bp->b_cmd == BUF_CMD_WRITE);
1431 if (bp->b_flags & B_NEEDCOMMIT) {
1435 off = bio->bio_offset + bp->b_dirtyoff;
1436 retv = nfs_commit(vp, off,
1437 bp->b_dirtyend - bp->b_dirtyoff, td);
1439 bp->b_dirtyoff = bp->b_dirtyend = 0;
1440 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1445 if (retv == NFSERR_STALEWRITEVERF) {
1446 nfs_clearcommit(vp->v_mount);
1451 * Setup for actual write
1454 if (bio->bio_offset + bp->b_dirtyend > np->n_size)
1455 bp->b_dirtyend = np->n_size - bio->bio_offset;
1457 if (bp->b_dirtyend > bp->b_dirtyoff) {
1458 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1460 uiop->uio_offset = bio->bio_offset + bp->b_dirtyoff;
1461 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1462 uiop->uio_rw = UIO_WRITE;
1463 nfsstats.write_bios++;
1465 if ((bp->b_flags & (B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == 0)
1466 iomode = NFSV3WRITE_UNSTABLE;
1468 iomode = NFSV3WRITE_FILESYNC;
1470 error = nfs_writerpc(vp, uiop, &iomode, &must_commit);
1473 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1474 * to cluster the buffers needing commit. This will allow
1475 * the system to submit a single commit rpc for the whole
1476 * cluster. We can do this even if the buffer is not 100%
1477 * dirty (relative to the NFS blocksize), so we optimize the
1478 * append-to-file-case.
1480 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1481 * cleared because write clustering only works for commit
1482 * rpc's, not for the data portion of the write).
1485 if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1486 bp->b_flags |= B_NEEDCOMMIT;
1487 if (bp->b_dirtyoff == 0
1488 && bp->b_dirtyend == bp->b_bcount)
1489 bp->b_flags |= B_CLUSTEROK;
1491 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1495 * For an interrupted write, the buffer is still valid
1496 * and the write hasn't been pushed to the server yet,
1497 * so we can't set B_ERROR and report the interruption
1498 * by setting B_EINTR. For the async case, B_EINTR
1499 * is not relevant, so the rpc attempt is essentially
1500 * a noop. For the case of a V3 write rpc not being
1501 * committed to stable storage, the block is still
1502 * dirty and requires either a commit rpc or another
1503 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1504 * the block is reused. This is indicated by setting
1505 * the B_DELWRI and B_NEEDCOMMIT flags.
1507 * If the buffer is marked B_PAGING, it does not reside on
1508 * the vp's paging queues so we cannot call bdirty(). The
1509 * bp in this case is not an NFS cache block so we should
1513 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1515 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1516 if ((bp->b_flags & B_PAGING) == 0)
1519 bp->b_flags |= B_EINTR;
1523 bp->b_flags |= B_ERROR;
1524 bp->b_error = np->n_error = error;
1525 np->n_flag |= NWRITEERR;
1527 bp->b_dirtyoff = bp->b_dirtyend = 0;
1535 bp->b_resid = uiop->uio_resid;
1537 nfs_clearcommit(vp->v_mount);
1543 * Used to aid in handling ftruncate() operations on the NFS client side.
1544 * Truncation creates a number of special problems for NFS. We have to
1545 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1546 * we have to properly handle VM pages or (potentially dirty) buffers
1547 * that straddle the truncation point.
1551 nfs_meta_setsize(struct vnode *vp, struct thread *td, u_quad_t nsize)
1553 struct nfsnode *np = VTONFS(vp);
1554 u_quad_t tsize = np->n_size;
1555 int biosize = vp->v_mount->mnt_stat.f_iosize;
1560 if (np->n_size < tsize) {
1567 * vtruncbuf() doesn't get the buffer overlapping the
1568 * truncation point. We may have a B_DELWRI and/or B_CACHE
1569 * buffer that now needs to be truncated.
1571 error = vtruncbuf(vp, nsize, biosize);
1572 lbn = nsize / biosize;
1573 bufsize = nsize & (biosize - 1);
1574 loffset = nsize - bufsize;
1575 bp = nfs_getcacheblk(vp, loffset, bufsize, td);
1576 if (bp->b_dirtyoff > bp->b_bcount)
1577 bp->b_dirtyoff = bp->b_bcount;
1578 if (bp->b_dirtyend > bp->b_bcount)
1579 bp->b_dirtyend = bp->b_bcount;
1580 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1583 vnode_pager_setsize(vp, nsize);
1589 * Synchronous completion for nfs_doio. Call bpdone() with elseit=FALSE.
1590 * Caller is responsible for brelse()'ing the bp.
1593 nfsiodone_sync(struct bio *bio)
1596 bpdone(bio->bio_buf, 0);