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_pbuf_freecnt;
75 extern struct nfsstats nfsstats;
78 * Vnode op for VM getpages.
80 * nfs_getpages(struct vnode *a_vp, vm_page_t *a_m, int a_count,
81 * int a_reqpage, vm_ooffset_t a_offset)
84 nfs_getpages(struct vop_getpages_args *ap)
86 struct thread *td = curthread; /* XXX */
87 int i, error, nextoff, size, toff, count, npages;
98 nmp = VFSTONFS(vp->v_mount);
102 if (vp->v_object == NULL) {
103 kprintf("nfs_getpages: called with non-merged cache vnode??\n");
104 return VM_PAGER_ERROR;
107 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
108 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
109 (void)nfs_fsinfo(nmp, vp, td);
111 npages = btoc(count);
114 * NOTE that partially valid pages may occur in cases other
115 * then file EOF, such as when a file is partially written and
116 * ftruncate()-extended to a larger size. It is also possible
117 * for the valid bits to be set on garbage beyond the file EOF and
118 * clear in the area before EOF (e.g. m->valid == 0xfc), which can
119 * occur due to vtruncbuf() and the buffer cache's handling of
120 * pages which 'straddle' buffers or when b_bufsize is not a
121 * multiple of PAGE_SIZE.... the buffer cache cannot normally
122 * clear the extra bits. This kind of situation occurs when you
123 * make a small write() (m->valid == 0x03) and then mmap() and
124 * fault in the buffer(m->valid = 0xFF). When NFS flushes the
125 * buffer (vinvalbuf() m->valid = 0xFC) we are left with a mess.
127 * This is combined with the possibility that the pages are partially
128 * dirty or that there is a buffer backing the pages that is dirty
129 * (even if m->dirty is 0).
131 * To solve this problem several hacks have been made: (1) NFS
132 * guarentees that the IO block size is a multiple of PAGE_SIZE and
133 * (2) The buffer cache, when invalidating an NFS buffer, will
134 * disregard the buffer's fragmentory b_bufsize and invalidate
135 * the whole page rather then just the piece the buffer owns.
137 * This allows us to assume that a partially valid page found here
138 * is fully valid (vm_fault will zero'd out areas of the page not
141 m = pages[ap->a_reqpage];
143 for (i = 0; i < npages; ++i) {
144 if (i != ap->a_reqpage)
145 vnode_pager_freepage(pages[i]);
151 * Use an MSF_BUF as a medium to retrieve data from the pages.
153 msf_map_pagelist(&msf, pages, npages, 0);
155 kva = msf_buf_kva(msf);
161 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
162 uio.uio_resid = count;
163 uio.uio_segflg = UIO_SYSSPACE;
164 uio.uio_rw = UIO_READ;
167 error = nfs_readrpc(vp, &uio);
170 if (error && (uio.uio_resid == count)) {
171 kprintf("nfs_getpages: error %d\n", error);
172 for (i = 0; i < npages; ++i) {
173 if (i != ap->a_reqpage)
174 vnode_pager_freepage(pages[i]);
176 return VM_PAGER_ERROR;
180 * Calculate the number of bytes read and validate only that number
181 * of bytes. Note that due to pending writes, size may be 0. This
182 * does not mean that the remaining data is invalid!
185 size = count - uio.uio_resid;
187 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
188 nextoff = toff + PAGE_SIZE;
191 m->flags &= ~PG_ZERO;
193 if (nextoff <= size) {
195 * Read operation filled an entire page
197 m->valid = VM_PAGE_BITS_ALL;
199 } else if (size > toff) {
201 * Read operation filled a partial page.
204 vm_page_set_validclean(m, 0, size - toff);
205 /* handled by vm_fault now */
206 /* vm_page_zero_invalid(m, TRUE); */
209 * Read operation was short. If no error occured
210 * we may have hit a zero-fill section. We simply
211 * leave valid set to 0.
215 if (i != ap->a_reqpage) {
217 * Whether or not to leave the page activated is up in
218 * the air, but we should put the page on a page queue
219 * somewhere (it already is in the object). Result:
220 * It appears that emperical results show that
221 * deactivating pages is best.
225 * Just in case someone was asking for this page we
226 * now tell them that it is ok to use.
229 if (m->flags & PG_WANTED)
232 vm_page_deactivate(m);
235 vnode_pager_freepage(m);
243 * Vnode op for VM putpages.
245 * nfs_putpages(struct vnode *a_vp, vm_page_t *a_m, int a_count, int a_sync,
246 * int *a_rtvals, vm_ooffset_t a_offset)
249 nfs_putpages(struct vop_putpages_args *ap)
251 struct thread *td = curthread;
255 int iomode, must_commit, i, error, npages, count;
259 struct nfsmount *nmp;
266 nmp = VFSTONFS(vp->v_mount);
269 rtvals = ap->a_rtvals;
270 npages = btoc(count);
271 offset = IDX_TO_OFF(pages[0]->pindex);
273 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
274 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
275 (void)nfs_fsinfo(nmp, vp, td);
277 for (i = 0; i < npages; i++) {
278 rtvals[i] = VM_PAGER_AGAIN;
282 * When putting pages, do not extend file past EOF.
285 if (offset + count > np->n_size) {
286 count = np->n_size - offset;
292 * Use an MSF_BUF as a medium to retrieve data from the pages.
294 msf_map_pagelist(&msf, pages, npages, 0);
296 kva = msf_buf_kva(msf);
302 uio.uio_offset = offset;
303 uio.uio_resid = count;
304 uio.uio_segflg = UIO_SYSSPACE;
305 uio.uio_rw = UIO_WRITE;
308 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
309 iomode = NFSV3WRITE_UNSTABLE;
311 iomode = NFSV3WRITE_FILESYNC;
313 error = nfs_writerpc(vp, &uio, &iomode, &must_commit);
318 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE;
319 for (i = 0; i < nwritten; i++) {
320 rtvals[i] = VM_PAGER_OK;
321 vm_page_undirty(pages[i]);
324 nfs_clearcommit(vp->v_mount);
330 * Vnode op for read using bio
333 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag)
335 struct nfsnode *np = VTONFS(vp);
337 struct buf *bp = 0, *rabp;
340 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
346 int nra, error = 0, n = 0, on = 0;
349 if (uio->uio_rw != UIO_READ)
350 panic("nfs_read mode");
352 if (uio->uio_resid == 0)
354 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
358 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
359 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
360 (void)nfs_fsinfo(nmp, vp, td);
361 if (vp->v_type != VDIR &&
362 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
364 biosize = vp->v_mount->mnt_stat.f_iosize;
365 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
368 * For nfs, cache consistency can only be maintained approximately.
369 * Although RFC1094 does not specify the criteria, the following is
370 * believed to be compatible with the reference port.
372 * NFS: If local changes have been made and this is a
373 * directory, the directory must be invalidated and
374 * the attribute cache must be cleared.
376 * GETATTR is called to synchronize the file size.
378 * If remote changes are detected local data is flushed
379 * and the cache is invalidated.
381 * NOTE: In the normal case the attribute cache is not
382 * cleared which means GETATTR may use cached data and
383 * not immediately detect changes made on the server.
385 if ((np->n_flag & NLMODIFIED) && vp->v_type == VDIR) {
387 error = nfs_vinvalbuf(vp, V_SAVE, 1);
392 error = VOP_GETATTR(vp, &vattr);
395 if (np->n_flag & NRMODIFIED) {
396 if (vp->v_type == VDIR)
398 error = nfs_vinvalbuf(vp, V_SAVE, 1);
401 np->n_flag &= ~NRMODIFIED;
404 if (np->n_flag & NDONTCACHE) {
405 switch (vp->v_type) {
407 return (nfs_readrpc(vp, uio));
409 return (nfs_readlinkrpc(vp, uio));
413 kprintf(" NDONTCACHE: type %x unexpected\n", vp->v_type);
417 switch (vp->v_type) {
419 nfsstats.biocache_reads++;
420 lbn = uio->uio_offset / biosize;
421 on = uio->uio_offset & (biosize - 1);
422 loffset = (off_t)lbn * biosize;
425 * Start the read ahead(s), as required.
427 if (nmp->nm_readahead > 0) {
428 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
429 (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) {
430 rabn = lbn + 1 + nra;
431 raoffset = (off_t)rabn * biosize;
432 if (findblk(vp, raoffset, FINDBLK_TEST) == NULL) {
433 rabp = nfs_getcacheblk(vp, raoffset, biosize, td);
436 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
437 rabp->b_cmd = BUF_CMD_READ;
438 vfs_busy_pages(vp, rabp);
439 if (nfs_asyncio(vp, &rabp->b_bio2, td)) {
440 rabp->b_flags |= B_INVAL|B_ERROR;
441 vfs_unbusy_pages(rabp);
453 * Obtain the buffer cache block. Figure out the buffer size
454 * when we are at EOF. If we are modifying the size of the
455 * buffer based on an EOF condition we need to hold
456 * nfs_rslock() through obtaining the buffer to prevent
457 * a potential writer-appender from messing with n_size.
458 * Otherwise we may accidently truncate the buffer and
461 * Note that bcount is *not* DEV_BSIZE aligned.
466 if (loffset >= np->n_size) {
468 } else if (loffset + biosize > np->n_size) {
469 bcount = np->n_size - loffset;
471 if (bcount != biosize) {
472 switch(nfs_rslock(np)) {
485 bp = nfs_getcacheblk(vp, loffset, bcount, td);
487 if (bcount != biosize)
493 * If B_CACHE is not set, we must issue the read. If this
494 * fails, we return an error.
497 if ((bp->b_flags & B_CACHE) == 0) {
498 bp->b_cmd = BUF_CMD_READ;
499 bp->b_bio2.bio_done = nfsiodone_sync;
500 bp->b_bio2.bio_flags |= BIO_SYNC;
501 vfs_busy_pages(vp, bp);
502 error = nfs_doio(vp, &bp->b_bio2, td);
510 * on is the offset into the current bp. Figure out how many
511 * bytes we can copy out of the bp. Note that bcount is
512 * NOT DEV_BSIZE aligned.
514 * Then figure out how many bytes we can copy into the uio.
519 n = min((unsigned)(bcount - on), uio->uio_resid);
522 biosize = min(NFS_MAXPATHLEN, np->n_size);
523 nfsstats.biocache_readlinks++;
524 bp = nfs_getcacheblk(vp, (off_t)0, biosize, td);
527 if ((bp->b_flags & B_CACHE) == 0) {
528 bp->b_cmd = BUF_CMD_READ;
529 bp->b_bio2.bio_done = nfsiodone_sync;
530 bp->b_bio2.bio_flags |= BIO_SYNC;
531 vfs_busy_pages(vp, bp);
532 error = nfs_doio(vp, &bp->b_bio2, td);
534 bp->b_flags |= B_ERROR | B_INVAL;
539 n = min(uio->uio_resid, bp->b_bcount - bp->b_resid);
543 nfsstats.biocache_readdirs++;
544 if (np->n_direofoffset
545 && uio->uio_offset >= np->n_direofoffset) {
548 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
549 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
550 loffset = uio->uio_offset - on;
551 bp = nfs_getcacheblk(vp, loffset, NFS_DIRBLKSIZ, td);
555 if ((bp->b_flags & B_CACHE) == 0) {
556 bp->b_cmd = BUF_CMD_READ;
557 bp->b_bio2.bio_done = nfsiodone_sync;
558 bp->b_bio2.bio_flags |= BIO_SYNC;
559 vfs_busy_pages(vp, bp);
560 error = nfs_doio(vp, &bp->b_bio2, td);
564 while (error == NFSERR_BAD_COOKIE) {
565 kprintf("got bad cookie vp %p bp %p\n", vp, bp);
567 error = nfs_vinvalbuf(vp, 0, 1);
569 * Yuck! The directory has been modified on the
570 * server. The only way to get the block is by
571 * reading from the beginning to get all the
574 * Leave the last bp intact unless there is an error.
575 * Loop back up to the while if the error is another
576 * NFSERR_BAD_COOKIE (double yuch!).
578 for (i = 0; i <= lbn && !error; i++) {
579 if (np->n_direofoffset
580 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
582 bp = nfs_getcacheblk(vp, (off_t)i * NFS_DIRBLKSIZ,
586 if ((bp->b_flags & B_CACHE) == 0) {
587 bp->b_cmd = BUF_CMD_READ;
588 bp->b_bio2.bio_done = nfsiodone_sync;
589 bp->b_bio2.bio_flags |= BIO_SYNC;
590 vfs_busy_pages(vp, bp);
591 error = nfs_doio(vp, &bp->b_bio2, td);
593 * no error + B_INVAL == directory EOF,
596 if (error == 0 && (bp->b_flags & B_INVAL))
600 * An error will throw away the block and the
601 * for loop will break out. If no error and this
602 * is not the block we want, we throw away the
603 * block and go for the next one via the for loop.
605 if (error || i < lbn)
610 * The above while is repeated if we hit another cookie
611 * error. If we hit an error and it wasn't a cookie error,
619 * If not eof and read aheads are enabled, start one.
620 * (You need the current block first, so that you have the
621 * directory offset cookie of the next block.)
623 if (nmp->nm_readahead > 0 &&
624 (bp->b_flags & B_INVAL) == 0 &&
625 (np->n_direofoffset == 0 ||
626 loffset + NFS_DIRBLKSIZ < np->n_direofoffset) &&
627 (np->n_flag & NDONTCACHE) == 0 &&
628 findblk(vp, loffset + NFS_DIRBLKSIZ, FINDBLK_TEST) == NULL
630 rabp = nfs_getcacheblk(vp, loffset + NFS_DIRBLKSIZ,
633 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
634 rabp->b_cmd = BUF_CMD_READ;
635 vfs_busy_pages(vp, rabp);
636 if (nfs_asyncio(vp, &rabp->b_bio2, td)) {
637 rabp->b_flags |= B_INVAL|B_ERROR;
638 vfs_unbusy_pages(rabp);
647 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
648 * chopped for the EOF condition, we cannot tell how large
649 * NFS directories are going to be until we hit EOF. So
650 * an NFS directory buffer is *not* chopped to its EOF. Now,
651 * it just so happens that b_resid will effectively chop it
652 * to EOF. *BUT* this information is lost if the buffer goes
653 * away and is reconstituted into a B_CACHE state ( due to
654 * being VMIO ) later. So we keep track of the directory eof
655 * in np->n_direofoffset and chop it off as an extra step
658 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
659 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
660 n = np->n_direofoffset - uio->uio_offset;
663 kprintf(" nfs_bioread: type %x unexpected\n",vp->v_type);
667 switch (vp->v_type) {
670 error = uiomove(bp->b_data + on, (int)n, uio);
674 error = uiomove(bp->b_data + on, (int)n, uio);
679 off_t old_off = uio->uio_offset;
681 struct nfs_dirent *dp;
684 * We are casting cpos to nfs_dirent, it must be
692 cpos = bp->b_data + on;
693 epos = bp->b_data + on + n;
694 while (cpos < epos && error == 0 && uio->uio_resid > 0) {
695 dp = (struct nfs_dirent *)cpos;
696 error = nfs_check_dirent(dp, (int)(epos - cpos));
699 if (vop_write_dirent(&error, uio, dp->nfs_ino,
700 dp->nfs_type, dp->nfs_namlen, dp->nfs_name)) {
703 cpos += dp->nfs_reclen;
707 uio->uio_offset = old_off + cpos - bp->b_data - on;
710 * Invalidate buffer if caching is disabled, forcing a
711 * re-read from the remote later.
713 if (np->n_flag & NDONTCACHE)
714 bp->b_flags |= B_INVAL;
717 kprintf(" nfs_bioread: type %x unexpected\n",vp->v_type);
720 } while (error == 0 && uio->uio_resid > 0 && n > 0);
725 * Userland can supply any 'seek' offset when reading a NFS directory.
726 * Validate the structure so we don't panic the kernel. Note that
727 * the element name is nul terminated and the nul is not included
732 nfs_check_dirent(struct nfs_dirent *dp, int maxlen)
734 int nfs_name_off = offsetof(struct nfs_dirent, nfs_name[0]);
736 if (nfs_name_off >= maxlen)
738 if (dp->nfs_reclen < nfs_name_off || dp->nfs_reclen > maxlen)
740 if (nfs_name_off + dp->nfs_namlen >= dp->nfs_reclen)
742 if (dp->nfs_reclen & 3)
748 * Vnode op for write using bio
750 * nfs_write(struct vnode *a_vp, struct uio *a_uio, int a_ioflag,
751 * struct ucred *a_cred)
754 nfs_write(struct vop_write_args *ap)
756 struct uio *uio = ap->a_uio;
757 struct thread *td = uio->uio_td;
758 struct vnode *vp = ap->a_vp;
759 struct nfsnode *np = VTONFS(vp);
760 int ioflag = ap->a_ioflag;
763 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
766 int n, on, error = 0, iomode, must_commit;
772 if (uio->uio_rw != UIO_WRITE)
773 panic("nfs_write mode");
774 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread)
775 panic("nfs_write proc");
777 if (vp->v_type != VREG)
779 if (np->n_flag & NWRITEERR) {
780 np->n_flag &= ~NWRITEERR;
781 return (np->n_error);
783 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
784 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
785 (void)nfs_fsinfo(nmp, vp, td);
788 * Synchronously flush pending buffers if we are in synchronous
789 * mode or if we are appending.
791 if (ioflag & (IO_APPEND | IO_SYNC)) {
792 if (np->n_flag & NLMODIFIED) {
794 error = nfs_flush(vp, MNT_WAIT, td, 0);
795 /* error = nfs_vinvalbuf(vp, V_SAVE, 1); */
802 * If IO_APPEND then load uio_offset. We restart here if we cannot
803 * get the append lock.
806 if (ioflag & IO_APPEND) {
808 error = VOP_GETATTR(vp, &vattr);
811 uio->uio_offset = np->n_size;
814 if (uio->uio_offset < 0)
816 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
818 if (uio->uio_resid == 0)
822 * We need to obtain the rslock if we intend to modify np->n_size
823 * in order to guarentee the append point with multiple contending
824 * writers, to guarentee that no other appenders modify n_size
825 * while we are trying to obtain a truncated buffer (i.e. to avoid
826 * accidently truncating data written by another appender due to
827 * the race), and to ensure that the buffer is populated prior to
828 * our extending of the file. We hold rslock through the entire
831 * Note that we do not synchronize the case where someone truncates
832 * the file while we are appending to it because attempting to lock
833 * this case may deadlock other parts of the system unexpectedly.
835 if ((ioflag & IO_APPEND) ||
836 uio->uio_offset + uio->uio_resid > np->n_size) {
837 switch(nfs_rslock(np)) {
852 * Maybe this should be above the vnode op call, but so long as
853 * file servers have no limits, i don't think it matters
855 if (td->td_proc && uio->uio_offset + uio->uio_resid >
856 td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
857 lwpsignal(td->td_proc, td->td_lwp, SIGXFSZ);
863 biosize = vp->v_mount->mnt_stat.f_iosize;
866 if ((np->n_flag & NDONTCACHE) && uio->uio_iovcnt == 1) {
867 iomode = NFSV3WRITE_FILESYNC;
868 error = nfs_writerpc(vp, uio, &iomode, &must_commit);
870 nfs_clearcommit(vp->v_mount);
873 nfsstats.biocache_writes++;
874 lbn = uio->uio_offset / biosize;
875 on = uio->uio_offset & (biosize-1);
876 loffset = uio->uio_offset - on;
877 n = min((unsigned)(biosize - on), uio->uio_resid);
880 * Handle direct append and file extension cases, calculate
881 * unaligned buffer size.
884 if (uio->uio_offset == np->n_size && n) {
886 * Get the buffer (in its pre-append state to maintain
887 * B_CACHE if it was previously set). Resize the
888 * nfsnode after we have locked the buffer to prevent
889 * readers from reading garbage.
892 bp = nfs_getcacheblk(vp, loffset, bcount, td);
897 np->n_size = uio->uio_offset + n;
898 np->n_flag |= NLMODIFIED;
899 vnode_pager_setsize(vp, np->n_size);
901 save = bp->b_flags & B_CACHE;
903 allocbuf(bp, bcount);
908 * Obtain the locked cache block first, and then
909 * adjust the file's size as appropriate.
912 if (loffset + bcount < np->n_size) {
913 if (loffset + biosize < np->n_size)
916 bcount = np->n_size - loffset;
918 bp = nfs_getcacheblk(vp, loffset, bcount, td);
919 if (uio->uio_offset + n > np->n_size) {
920 np->n_size = uio->uio_offset + n;
921 np->n_flag |= NLMODIFIED;
922 vnode_pager_setsize(vp, np->n_size);
932 * Issue a READ if B_CACHE is not set. In special-append
933 * mode, B_CACHE is based on the buffer prior to the write
934 * op and is typically set, avoiding the read. If a read
935 * is required in special append mode, the server will
936 * probably send us a short-read since we extended the file
937 * on our end, resulting in b_resid == 0 and, thusly,
938 * B_CACHE getting set.
940 * We can also avoid issuing the read if the write covers
941 * the entire buffer. We have to make sure the buffer state
942 * is reasonable in this case since we will not be initiating
943 * I/O. See the comments in kern/vfs_bio.c's getblk() for
946 * B_CACHE may also be set due to the buffer being cached
949 * When doing a UIO_NOCOPY write the buffer is not
950 * overwritten and we cannot just set B_CACHE unconditionally
951 * for full-block writes.
954 if (on == 0 && n == bcount && uio->uio_segflg != UIO_NOCOPY) {
955 bp->b_flags |= B_CACHE;
956 bp->b_flags &= ~(B_ERROR | B_INVAL);
959 if ((bp->b_flags & B_CACHE) == 0) {
960 bp->b_cmd = BUF_CMD_READ;
961 bp->b_bio2.bio_done = nfsiodone_sync;
962 bp->b_bio2.bio_flags |= BIO_SYNC;
963 vfs_busy_pages(vp, bp);
964 error = nfs_doio(vp, &bp->b_bio2, td);
974 np->n_flag |= NLMODIFIED;
977 * If dirtyend exceeds file size, chop it down. This should
978 * not normally occur but there is an append race where it
979 * might occur XXX, so we log it.
981 * If the chopping creates a reverse-indexed or degenerate
982 * situation with dirtyoff/end, we 0 both of them.
985 if (bp->b_dirtyend > bcount) {
986 kprintf("NFS append race @%08llx:%d\n",
987 (long long)bp->b_bio2.bio_offset,
988 bp->b_dirtyend - bcount);
989 bp->b_dirtyend = bcount;
992 if (bp->b_dirtyoff >= bp->b_dirtyend)
993 bp->b_dirtyoff = bp->b_dirtyend = 0;
996 * If the new write will leave a contiguous dirty
997 * area, just update the b_dirtyoff and b_dirtyend,
998 * otherwise force a write rpc of the old dirty area.
1000 * While it is possible to merge discontiguous writes due to
1001 * our having a B_CACHE buffer ( and thus valid read data
1002 * for the hole), we don't because it could lead to
1003 * significant cache coherency problems with multiple clients,
1004 * especially if locking is implemented later on.
1006 * as an optimization we could theoretically maintain
1007 * a linked list of discontinuous areas, but we would still
1008 * have to commit them separately so there isn't much
1009 * advantage to it except perhaps a bit of asynchronization.
1012 if (bp->b_dirtyend > 0 &&
1013 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1014 if (bwrite(bp) == EINTR) {
1021 error = uiomove((char *)bp->b_data + on, n, uio);
1024 * Since this block is being modified, it must be written
1025 * again and not just committed. Since write clustering does
1026 * not work for the stage 1 data write, only the stage 2
1027 * commit rpc, we have to clear B_CLUSTEROK as well.
1029 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1032 bp->b_flags |= B_ERROR;
1038 * Only update dirtyoff/dirtyend if not a degenerate
1042 if (bp->b_dirtyend > 0) {
1043 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1044 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1046 bp->b_dirtyoff = on;
1047 bp->b_dirtyend = on + n;
1049 vfs_bio_set_validclean(bp, on, n);
1053 * If the lease is non-cachable or IO_SYNC do bwrite().
1055 * IO_INVAL appears to be unused. The idea appears to be
1056 * to turn off caching in this case. Very odd. XXX
1058 * If nfs_async is set bawrite() will use an unstable write
1059 * (build dirty bufs on the server), so we might as well
1060 * push it out with bawrite(). If nfs_async is not set we
1061 * use bdwrite() to cache dirty bufs on the client.
1063 if ((np->n_flag & NDONTCACHE) || (ioflag & IO_SYNC)) {
1064 if (ioflag & IO_INVAL)
1065 bp->b_flags |= B_NOCACHE;
1069 if (np->n_flag & NDONTCACHE) {
1070 error = nfs_vinvalbuf(vp, V_SAVE, 1);
1074 } else if ((n + on) == biosize && nfs_async) {
1079 } while (uio->uio_resid > 0 && n > 0);
1088 * Get an nfs cache block.
1090 * Allocate a new one if the block isn't currently in the cache
1091 * and return the block marked busy. If the calling process is
1092 * interrupted by a signal for an interruptible mount point, return
1095 * The caller must carefully deal with the possible B_INVAL state of
1096 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it
1097 * indirectly), so synchronous reads can be issued without worrying about
1098 * the B_INVAL state. We have to be a little more careful when dealing
1099 * with writes (see comments in nfs_write()) when extending a file past
1103 nfs_getcacheblk(struct vnode *vp, off_t loffset, int size, struct thread *td)
1107 struct nfsmount *nmp;
1112 if (nmp->nm_flag & NFSMNT_INT) {
1113 bp = getblk(vp, loffset, size, GETBLK_PCATCH, 0);
1114 while (bp == NULL) {
1115 if (nfs_sigintr(nmp, NULL, td))
1117 bp = getblk(vp, loffset, size, 0, 2 * hz);
1120 bp = getblk(vp, loffset, size, 0, 0);
1124 * bio2, the 'device' layer. Since BIOs use 64 bit byte offsets
1125 * now, no translation is necessary.
1127 bp->b_bio2.bio_offset = loffset;
1132 * Flush and invalidate all dirty buffers. If another process is already
1133 * doing the flush, just wait for completion.
1136 nfs_vinvalbuf(struct vnode *vp, int flags, int intrflg)
1138 struct nfsnode *np = VTONFS(vp);
1139 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1140 int error = 0, slpflag, slptimeo;
1141 thread_t td = curthread;
1143 if (vp->v_flag & VRECLAIMED)
1146 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1156 * First wait for any other process doing a flush to complete.
1158 while (np->n_flag & NFLUSHINPROG) {
1159 np->n_flag |= NFLUSHWANT;
1160 error = tsleep((caddr_t)&np->n_flag, 0, "nfsvinval", slptimeo);
1161 if (error && intrflg && nfs_sigintr(nmp, NULL, td))
1166 * Now, flush as required.
1168 np->n_flag |= NFLUSHINPROG;
1169 error = vinvalbuf(vp, flags, slpflag, 0);
1171 if (intrflg && nfs_sigintr(nmp, NULL, td)) {
1172 np->n_flag &= ~NFLUSHINPROG;
1173 if (np->n_flag & NFLUSHWANT) {
1174 np->n_flag &= ~NFLUSHWANT;
1175 wakeup((caddr_t)&np->n_flag);
1179 error = vinvalbuf(vp, flags, 0, slptimeo);
1181 np->n_flag &= ~(NLMODIFIED | NFLUSHINPROG);
1182 if (np->n_flag & NFLUSHWANT) {
1183 np->n_flag &= ~NFLUSHWANT;
1184 wakeup((caddr_t)&np->n_flag);
1190 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1191 * This is mainly to avoid queueing async I/O requests when the nfsiods
1192 * are all hung on a dead server.
1194 * Note: nfs_asyncio() does not clear (B_ERROR|B_INVAL) but when the bp
1195 * is eventually dequeued by the async daemon, nfs_doio() *will*.
1198 nfs_asyncio(struct vnode *vp, struct bio *bio, struct thread *td)
1200 struct buf *bp = bio->bio_buf;
1201 struct nfsmount *nmp;
1203 KKASSERT(vp->v_tag == VT_NFS);
1204 nmp = VFSTONFS(vp->v_mount);
1207 * If no async daemons then return EIO to force caller to run the rpc
1210 if (nmp->nm_rxstate > NFSSVC_PENDING)
1216 * The passed bio's buffer is not necessary associated with
1217 * the NFS vnode it is being written to. Store the NFS vnode
1218 * in the BIO driver info.
1220 bio->bio_driver_info = vp;
1221 TAILQ_INSERT_TAIL(&nmp->nm_bioq, bio, bio_act);
1223 nfssvc_iod_writer_wakeup(nmp);
1228 * Do an I/O operation to/from a cache block. This may be called
1229 * synchronously or from an nfsiod. The BIO is normalized for DEV_BSIZE.
1231 * A locked, completed I/O is returned and the caller is responsible for
1234 * NOTE! TD MIGHT BE NULL
1237 nfs_doio(struct vnode *vp, struct bio *bio, struct thread *td)
1239 struct buf *bp = bio->bio_buf;
1242 struct nfsmount *nmp;
1243 int error = 0, iomode, must_commit = 0;
1247 KKASSERT(vp->v_tag == VT_NFS);
1249 nmp = VFSTONFS(vp->v_mount);
1251 uiop->uio_iov = &io;
1252 uiop->uio_iovcnt = 1;
1253 uiop->uio_segflg = UIO_SYSSPACE;
1257 * clear B_ERROR and B_INVAL state prior to initiating the I/O. We
1258 * do this here so we do not have to do it in all the code that
1261 bp->b_flags &= ~(B_ERROR | B_INVAL);
1264 KASSERT(bp->b_cmd != BUF_CMD_DONE,
1265 ("nfs_doio: bp %p already marked done!", bp));
1267 if (bp->b_cmd == BUF_CMD_READ) {
1268 io.iov_len = uiop->uio_resid = bp->b_bcount;
1269 io.iov_base = bp->b_data;
1270 uiop->uio_rw = UIO_READ;
1272 switch (vp->v_type) {
1274 uiop->uio_offset = bio->bio_offset;
1275 nfsstats.read_bios++;
1276 error = nfs_readrpc(vp, uiop);
1279 if (uiop->uio_resid) {
1281 * If we had a short read with no error, we must have
1282 * hit a file hole. We should zero-fill the remainder.
1283 * This can also occur if the server hits the file EOF.
1285 * Holes used to be able to occur due to pending
1286 * writes, but that is not possible any longer.
1288 int nread = bp->b_bcount - uiop->uio_resid;
1289 int left = uiop->uio_resid;
1292 bzero((char *)bp->b_data + nread, left);
1293 uiop->uio_resid = 0;
1296 if (td && td->td_proc && (vp->v_flag & VTEXT) &&
1297 np->n_mtime != np->n_vattr.va_mtime.tv_sec) {
1298 uprintf("Process killed due to text file modification\n");
1299 ksignal(td->td_proc, SIGKILL);
1303 uiop->uio_offset = 0;
1304 nfsstats.readlink_bios++;
1305 error = nfs_readlinkrpc(vp, uiop);
1308 nfsstats.readdir_bios++;
1309 uiop->uio_offset = bio->bio_offset;
1310 if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
1311 error = nfs_readdirplusrpc(vp, uiop);
1312 if (error == NFSERR_NOTSUPP)
1313 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1315 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1316 error = nfs_readdirrpc(vp, uiop);
1318 * end-of-directory sets B_INVAL but does not generate an
1321 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1322 bp->b_flags |= B_INVAL;
1325 kprintf("nfs_doio: type %x unexpected\n",vp->v_type);
1329 bp->b_flags |= B_ERROR;
1330 bp->b_error = error;
1334 * If we only need to commit, try to commit
1336 KKASSERT(bp->b_cmd == BUF_CMD_WRITE);
1337 if (bp->b_flags & B_NEEDCOMMIT) {
1341 off = bio->bio_offset + bp->b_dirtyoff;
1342 retv = nfs_commit(vp, off,
1343 bp->b_dirtyend - bp->b_dirtyoff, td);
1345 bp->b_dirtyoff = bp->b_dirtyend = 0;
1346 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1351 if (retv == NFSERR_STALEWRITEVERF) {
1352 nfs_clearcommit(vp->v_mount);
1357 * Setup for actual write
1360 if (bio->bio_offset + bp->b_dirtyend > np->n_size)
1361 bp->b_dirtyend = np->n_size - bio->bio_offset;
1363 if (bp->b_dirtyend > bp->b_dirtyoff) {
1364 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1366 uiop->uio_offset = bio->bio_offset + bp->b_dirtyoff;
1367 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1368 uiop->uio_rw = UIO_WRITE;
1369 nfsstats.write_bios++;
1371 if ((bp->b_flags & (B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == 0)
1372 iomode = NFSV3WRITE_UNSTABLE;
1374 iomode = NFSV3WRITE_FILESYNC;
1376 error = nfs_writerpc(vp, uiop, &iomode, &must_commit);
1379 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1380 * to cluster the buffers needing commit. This will allow
1381 * the system to submit a single commit rpc for the whole
1382 * cluster. We can do this even if the buffer is not 100%
1383 * dirty (relative to the NFS blocksize), so we optimize the
1384 * append-to-file-case.
1386 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1387 * cleared because write clustering only works for commit
1388 * rpc's, not for the data portion of the write).
1391 if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1392 bp->b_flags |= B_NEEDCOMMIT;
1393 if (bp->b_dirtyoff == 0
1394 && bp->b_dirtyend == bp->b_bcount)
1395 bp->b_flags |= B_CLUSTEROK;
1397 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1401 * For an interrupted write, the buffer is still valid
1402 * and the write hasn't been pushed to the server yet,
1403 * so we can't set B_ERROR and report the interruption
1404 * by setting B_EINTR. For the async case, B_EINTR
1405 * is not relevant, so the rpc attempt is essentially
1406 * a noop. For the case of a V3 write rpc not being
1407 * committed to stable storage, the block is still
1408 * dirty and requires either a commit rpc or another
1409 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1410 * the block is reused. This is indicated by setting
1411 * the B_DELWRI and B_NEEDCOMMIT flags.
1413 * If the buffer is marked B_PAGING, it does not reside on
1414 * the vp's paging queues so we cannot call bdirty(). The
1415 * bp in this case is not an NFS cache block so we should
1419 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1421 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1422 if ((bp->b_flags & B_PAGING) == 0)
1425 bp->b_flags |= B_EINTR;
1429 bp->b_flags |= B_ERROR;
1430 bp->b_error = np->n_error = error;
1431 np->n_flag |= NWRITEERR;
1433 bp->b_dirtyoff = bp->b_dirtyend = 0;
1441 bp->b_resid = uiop->uio_resid;
1443 nfs_clearcommit(vp->v_mount);
1449 * Used to aid in handling ftruncate() operations on the NFS client side.
1450 * Truncation creates a number of special problems for NFS. We have to
1451 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1452 * we have to properly handle VM pages or (potentially dirty) buffers
1453 * that straddle the truncation point.
1457 nfs_meta_setsize(struct vnode *vp, struct thread *td, u_quad_t nsize)
1459 struct nfsnode *np = VTONFS(vp);
1460 u_quad_t tsize = np->n_size;
1461 int biosize = vp->v_mount->mnt_stat.f_iosize;
1466 if (np->n_size < tsize) {
1473 * vtruncbuf() doesn't get the buffer overlapping the
1474 * truncation point. We may have a B_DELWRI and/or B_CACHE
1475 * buffer that now needs to be truncated.
1477 error = vtruncbuf(vp, nsize, biosize);
1478 lbn = nsize / biosize;
1479 bufsize = nsize & (biosize - 1);
1480 loffset = nsize - bufsize;
1481 bp = nfs_getcacheblk(vp, loffset, bufsize, td);
1482 if (bp->b_dirtyoff > bp->b_bcount)
1483 bp->b_dirtyoff = bp->b_bcount;
1484 if (bp->b_dirtyend > bp->b_bcount)
1485 bp->b_dirtyend = bp->b_bcount;
1486 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1489 vnode_pager_setsize(vp, nsize);
1495 * Synchronous completion for nfs_doio. Call bpdone() with elseit=FALSE.
1496 * Caller is responsible for brelse()'ing the bp.
1499 nfsiodone_sync(struct bio *bio)
1502 bpdone(bio->bio_buf, 0);