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 1.18 2004/10/12 19:21:01 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>
68 static struct buf *nfs_getcacheblk (struct vnode *vp, daddr_t bn, int size,
71 extern int nfs_numasync;
72 extern int nfs_pbuf_freecnt;
73 extern struct nfsstats nfsstats;
76 * Vnode op for VM getpages.
78 * nfs_getpages(struct vnode *a_vp, vm_page_t *a_m, int a_count,
79 * int a_reqpage, vm_ooffset_t a_offset)
82 nfs_getpages(struct vop_getpages_args *ap)
84 struct thread *td = curthread; /* XXX */
85 int i, error, nextoff, size, toff, count, npages;
96 nmp = VFSTONFS(vp->v_mount);
100 if (vp->v_object == NULL) {
101 printf("nfs_getpages: called with non-merged cache vnode??\n");
102 return VM_PAGER_ERROR;
105 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
106 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
107 (void)nfs_fsinfo(nmp, vp, td);
109 npages = btoc(count);
112 * NOTE that partially valid pages may occur in cases other
113 * then file EOF, such as when a file is partially written and
114 * ftruncate()-extended to a larger size. It is also possible
115 * for the valid bits to be set on garbage beyond the file EOF and
116 * clear in the area before EOF (e.g. m->valid == 0xfc), which can
117 * occur due to vtruncbuf() and the buffer cache's handling of
118 * pages which 'straddle' buffers or when b_bufsize is not a
119 * multiple of PAGE_SIZE.... the buffer cache cannot normally
120 * clear the extra bits. This kind of situation occurs when you
121 * make a small write() (m->valid == 0x03) and then mmap() and
122 * fault in the buffer(m->valid = 0xFF). When NFS flushes the
123 * buffer (vinvalbuf() m->valid = 0xFC) we are left with a mess.
125 * This is combined with the possibility that the pages are partially
126 * dirty or that there is a buffer backing the pages that is dirty
127 * (even if m->dirty is 0).
129 * To solve this problem several hacks have been made: (1) NFS
130 * guarentees that the IO block size is a multiple of PAGE_SIZE and
131 * (2) The buffer cache, when invalidating an NFS buffer, will
132 * disregard the buffer's fragmentory b_bufsize and invalidate
133 * the whole page rather then just the piece the buffer owns.
135 * This allows us to assume that a partially valid page found here
136 * is fully valid (vm_fault will zero'd out areas of the page not
139 m = pages[ap->a_reqpage];
141 for (i = 0; i < npages; ++i) {
142 if (i != ap->a_reqpage)
143 vnode_pager_freepage(pages[i]);
149 * Use an MSF_BUF as a medium to retrieve data from the pages.
151 msf = msf_buf_alloc(pages, npages, 0);
152 kva = msf_buf_kva(msf);
154 iov.iov_base = (caddr_t) kva;
158 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
159 uio.uio_resid = count;
160 uio.uio_segflg = UIO_SYSSPACE;
161 uio.uio_rw = UIO_READ;
164 error = nfs_readrpc(vp, &uio);
167 if (error && (uio.uio_resid == count)) {
168 printf("nfs_getpages: error %d\n", error);
169 for (i = 0; i < npages; ++i) {
170 if (i != ap->a_reqpage)
171 vnode_pager_freepage(pages[i]);
173 return VM_PAGER_ERROR;
177 * Calculate the number of bytes read and validate only that number
178 * of bytes. Note that due to pending writes, size may be 0. This
179 * does not mean that the remaining data is invalid!
182 size = count - uio.uio_resid;
184 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
185 nextoff = toff + PAGE_SIZE;
188 m->flags &= ~PG_ZERO;
190 if (nextoff <= size) {
192 * Read operation filled an entire page
194 m->valid = VM_PAGE_BITS_ALL;
196 } else if (size > toff) {
198 * Read operation filled a partial page.
201 vm_page_set_validclean(m, 0, size - toff);
202 /* handled by vm_fault now */
203 /* vm_page_zero_invalid(m, TRUE); */
206 * Read operation was short. If no error occured
207 * we may have hit a zero-fill section. We simply
208 * leave valid set to 0.
212 if (i != ap->a_reqpage) {
214 * Whether or not to leave the page activated is up in
215 * the air, but we should put the page on a page queue
216 * somewhere (it already is in the object). Result:
217 * It appears that emperical results show that
218 * deactivating pages is best.
222 * Just in case someone was asking for this page we
223 * now tell them that it is ok to use.
226 if (m->flags & PG_WANTED)
229 vm_page_deactivate(m);
232 vnode_pager_freepage(m);
240 * Vnode op for VM putpages.
242 * nfs_putpages(struct vnode *a_vp, vm_page_t *a_m, int a_count, int a_sync,
243 * int *a_rtvals, vm_ooffset_t a_offset)
246 nfs_putpages(struct vop_putpages_args *ap)
248 struct thread *td = curthread;
252 int iomode, must_commit, i, error, npages, count;
256 struct nfsmount *nmp;
263 nmp = VFSTONFS(vp->v_mount);
266 rtvals = ap->a_rtvals;
267 npages = btoc(count);
268 offset = IDX_TO_OFF(pages[0]->pindex);
270 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
271 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
272 (void)nfs_fsinfo(nmp, vp, td);
274 for (i = 0; i < npages; i++) {
275 rtvals[i] = VM_PAGER_AGAIN;
279 * When putting pages, do not extend file past EOF.
282 if (offset + count > np->n_size) {
283 count = np->n_size - offset;
289 * Use an MSF_BUF as a medium to retrieve data from the pages.
291 msf = msf_buf_alloc(pages, npages, 0);
292 kva = msf_buf_kva(msf);
294 iov.iov_base = (caddr_t) kva;
298 uio.uio_offset = offset;
299 uio.uio_resid = count;
300 uio.uio_segflg = UIO_SYSSPACE;
301 uio.uio_rw = UIO_WRITE;
304 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
305 iomode = NFSV3WRITE_UNSTABLE;
307 iomode = NFSV3WRITE_FILESYNC;
309 error = nfs_writerpc(vp, &uio, &iomode, &must_commit);
314 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE;
315 for (i = 0; i < nwritten; i++) {
316 rtvals[i] = VM_PAGER_OK;
317 vm_page_undirty(pages[i]);
320 nfs_clearcommit(vp->v_mount);
326 * Vnode op for read using bio
329 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag)
331 struct nfsnode *np = VTONFS(vp);
333 struct buf *bp = 0, *rabp;
336 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
340 int nra, error = 0, n = 0, on = 0;
343 if (uio->uio_rw != UIO_READ)
344 panic("nfs_read mode");
346 if (uio->uio_resid == 0)
348 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
352 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
353 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
354 (void)nfs_fsinfo(nmp, vp, td);
355 if (vp->v_type != VDIR &&
356 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
358 biosize = vp->v_mount->mnt_stat.f_iosize;
359 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
361 * For nfs, cache consistency can only be maintained approximately.
362 * Although RFC1094 does not specify the criteria, the following is
363 * believed to be compatible with the reference port.
364 * For nqnfs, full cache consistency is maintained within the loop.
366 * If the file's modify time on the server has changed since the
367 * last read rpc or you have written to the file,
368 * you may have lost data cache consistency with the
369 * server, so flush all of the file's data out of the cache.
370 * Then force a getattr rpc to ensure that you have up to date
372 * NB: This implies that cache data can be read when up to
373 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
374 * attributes this could be forced by setting n_attrstamp to 0 before
375 * the VOP_GETATTR() call.
377 if ((nmp->nm_flag & NFSMNT_NQNFS) == 0) {
378 if (np->n_flag & NMODIFIED) {
379 if (vp->v_type != VREG) {
380 if (vp->v_type != VDIR)
381 panic("nfs: bioread, not dir");
383 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
388 error = VOP_GETATTR(vp, &vattr, td);
391 np->n_mtime = vattr.va_mtime.tv_sec;
393 error = VOP_GETATTR(vp, &vattr, td);
396 if ((np->n_flag & NSIZECHANGED)
397 || np->n_mtime != vattr.va_mtime.tv_sec) {
398 if (vp->v_type == VDIR)
400 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
403 np->n_mtime = vattr.va_mtime.tv_sec;
404 np->n_flag &= ~NSIZECHANGED;
411 * Get a valid lease. If cached data is stale, flush it.
413 if (nmp->nm_flag & NFSMNT_NQNFS) {
414 if (NQNFS_CKINVALID(vp, np, ND_READ)) {
416 error = nqnfs_getlease(vp, ND_READ, td);
417 } while (error == NQNFS_EXPIRED);
420 if (np->n_lrev != np->n_brev ||
421 (np->n_flag & NQNFSNONCACHE) ||
422 ((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) {
423 if (vp->v_type == VDIR)
425 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
428 np->n_brev = np->n_lrev;
430 } else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) {
432 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
437 if (np->n_flag & NQNFSNONCACHE) {
438 switch (vp->v_type) {
440 return (nfs_readrpc(vp, uio));
442 return (nfs_readlinkrpc(vp, uio));
446 printf(" NQNFSNONCACHE: type %x unexpected\n",
450 switch (vp->v_type) {
452 nfsstats.biocache_reads++;
453 lbn = uio->uio_offset / biosize;
454 on = uio->uio_offset & (biosize - 1);
457 * Start the read ahead(s), as required.
459 if (nfs_numasync > 0 && nmp->nm_readahead > 0) {
460 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
461 (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) {
462 rabn = lbn + 1 + nra;
463 if (!incore(vp, rabn)) {
464 rabp = nfs_getcacheblk(vp, rabn, biosize, td);
467 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
468 rabp->b_flags |= (B_READ | B_ASYNC);
469 vfs_busy_pages(rabp, 0);
470 if (nfs_asyncio(rabp, td)) {
471 rabp->b_flags |= B_INVAL|B_ERROR;
472 vfs_unbusy_pages(rabp);
484 * Obtain the buffer cache block. Figure out the buffer size
485 * when we are at EOF. If we are modifying the size of the
486 * buffer based on an EOF condition we need to hold
487 * nfs_rslock() through obtaining the buffer to prevent
488 * a potential writer-appender from messing with n_size.
489 * Otherwise we may accidently truncate the buffer and
492 * Note that bcount is *not* DEV_BSIZE aligned.
497 if ((off_t)lbn * biosize >= np->n_size) {
499 } else if ((off_t)(lbn + 1) * biosize > np->n_size) {
500 bcount = np->n_size - (off_t)lbn * biosize;
502 if (bcount != biosize) {
503 switch(nfs_rslock(np, td)) {
516 bp = nfs_getcacheblk(vp, lbn, bcount, td);
518 if (bcount != biosize)
519 nfs_rsunlock(np, td);
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_flags |= B_READ;
530 vfs_busy_pages(bp, 0);
531 error = nfs_doio(bp, td);
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 nfsstats.biocache_readlinks++;
552 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
555 if ((bp->b_flags & B_CACHE) == 0) {
556 bp->b_flags |= B_READ;
557 vfs_busy_pages(bp, 0);
558 error = nfs_doio(bp, td);
560 bp->b_flags |= B_ERROR;
565 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
569 nfsstats.biocache_readdirs++;
570 if (np->n_direofoffset
571 && uio->uio_offset >= np->n_direofoffset) {
574 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
575 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
576 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
579 if ((bp->b_flags & B_CACHE) == 0) {
580 bp->b_flags |= B_READ;
581 vfs_busy_pages(bp, 0);
582 error = nfs_doio(bp, td);
586 while (error == NFSERR_BAD_COOKIE) {
587 printf("got bad cookie vp %p bp %p\n", vp, bp);
589 error = nfs_vinvalbuf(vp, 0, td, 1);
591 * Yuck! The directory has been modified on the
592 * server. The only way to get the block is by
593 * reading from the beginning to get all the
596 * Leave the last bp intact unless there is an error.
597 * Loop back up to the while if the error is another
598 * NFSERR_BAD_COOKIE (double yuch!).
600 for (i = 0; i <= lbn && !error; i++) {
601 if (np->n_direofoffset
602 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
604 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
607 if ((bp->b_flags & B_CACHE) == 0) {
608 bp->b_flags |= B_READ;
609 vfs_busy_pages(bp, 0);
610 error = nfs_doio(bp, td);
612 * no error + B_INVAL == directory EOF,
615 if (error == 0 && (bp->b_flags & B_INVAL))
619 * An error will throw away the block and the
620 * for loop will break out. If no error and this
621 * is not the block we want, we throw away the
622 * block and go for the next one via the for loop.
624 if (error || i < lbn)
629 * The above while is repeated if we hit another cookie
630 * error. If we hit an error and it wasn't a cookie error,
638 * If not eof and read aheads are enabled, start one.
639 * (You need the current block first, so that you have the
640 * directory offset cookie of the next block.)
642 if (nfs_numasync > 0 && nmp->nm_readahead > 0 &&
643 (bp->b_flags & B_INVAL) == 0 &&
644 (np->n_direofoffset == 0 ||
645 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
646 !(np->n_flag & NQNFSNONCACHE) &&
647 !incore(vp, lbn + 1)) {
648 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
650 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
651 rabp->b_flags |= (B_READ | B_ASYNC);
652 vfs_busy_pages(rabp, 0);
653 if (nfs_asyncio(rabp, td)) {
654 rabp->b_flags |= B_INVAL|B_ERROR;
655 vfs_unbusy_pages(rabp);
664 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
665 * chopped for the EOF condition, we cannot tell how large
666 * NFS directories are going to be until we hit EOF. So
667 * an NFS directory buffer is *not* chopped to its EOF. Now,
668 * it just so happens that b_resid will effectively chop it
669 * to EOF. *BUT* this information is lost if the buffer goes
670 * away and is reconstituted into a B_CACHE state ( due to
671 * being VMIO ) later. So we keep track of the directory eof
672 * in np->n_direofoffset and chop it off as an extra step
675 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
676 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
677 n = np->n_direofoffset - uio->uio_offset;
680 printf(" nfs_bioread: type %x unexpected\n",vp->v_type);
685 error = uiomove(bp->b_data + on, (int)n, uio);
687 switch (vp->v_type) {
695 * Invalidate buffer if caching is disabled, forcing a
696 * re-read from the remote later.
698 if (np->n_flag & NQNFSNONCACHE)
699 bp->b_flags |= B_INVAL;
702 printf(" nfs_bioread: type %x unexpected\n",vp->v_type);
705 } while (error == 0 && uio->uio_resid > 0 && n > 0);
710 * Vnode op for write using bio
712 * nfs_write(struct vnode *a_vp, struct uio *a_uio, int a_ioflag,
713 * struct ucred *a_cred)
716 nfs_write(struct vop_write_args *ap)
719 struct uio *uio = ap->a_uio;
720 struct thread *td = uio->uio_td;
721 struct vnode *vp = ap->a_vp;
722 struct nfsnode *np = VTONFS(vp);
723 int ioflag = ap->a_ioflag;
726 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
729 int n, on, error = 0, iomode, must_commit;
733 if (uio->uio_rw != UIO_WRITE)
734 panic("nfs_write mode");
735 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread)
736 panic("nfs_write proc");
738 if (vp->v_type != VREG)
740 if (np->n_flag & NWRITEERR) {
741 np->n_flag &= ~NWRITEERR;
742 return (np->n_error);
744 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
745 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
746 (void)nfs_fsinfo(nmp, vp, td);
749 * Synchronously flush pending buffers if we are in synchronous
750 * mode or if we are appending.
752 if (ioflag & (IO_APPEND | IO_SYNC)) {
753 if (np->n_flag & NMODIFIED) {
755 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
762 * If IO_APPEND then load uio_offset. We restart here if we cannot
763 * get the append lock.
766 if (ioflag & IO_APPEND) {
768 error = VOP_GETATTR(vp, &vattr, td);
771 uio->uio_offset = np->n_size;
774 if (uio->uio_offset < 0)
776 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
778 if (uio->uio_resid == 0)
782 * We need to obtain the rslock if we intend to modify np->n_size
783 * in order to guarentee the append point with multiple contending
784 * writers, to guarentee that no other appenders modify n_size
785 * while we are trying to obtain a truncated buffer (i.e. to avoid
786 * accidently truncating data written by another appender due to
787 * the race), and to ensure that the buffer is populated prior to
788 * our extending of the file. We hold rslock through the entire
791 * Note that we do not synchronize the case where someone truncates
792 * the file while we are appending to it because attempting to lock
793 * this case may deadlock other parts of the system unexpectedly.
795 if ((ioflag & IO_APPEND) ||
796 uio->uio_offset + uio->uio_resid > np->n_size) {
797 switch(nfs_rslock(np, td)) {
812 * Maybe this should be above the vnode op call, but so long as
813 * file servers have no limits, i don't think it matters
815 if (td->td_proc && uio->uio_offset + uio->uio_resid >
816 td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
817 psignal(td->td_proc, SIGXFSZ);
819 nfs_rsunlock(np, td);
823 biosize = vp->v_mount->mnt_stat.f_iosize;
827 * Check for a valid write lease.
829 if ((nmp->nm_flag & NFSMNT_NQNFS) &&
830 NQNFS_CKINVALID(vp, np, ND_WRITE)) {
832 error = nqnfs_getlease(vp, ND_WRITE, td);
833 } while (error == NQNFS_EXPIRED);
836 if (np->n_lrev != np->n_brev ||
837 (np->n_flag & NQNFSNONCACHE)) {
838 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
841 np->n_brev = np->n_lrev;
844 if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) {
845 iomode = NFSV3WRITE_FILESYNC;
846 error = nfs_writerpc(vp, uio, &iomode, &must_commit);
848 nfs_clearcommit(vp->v_mount);
851 nfsstats.biocache_writes++;
852 lbn = uio->uio_offset / biosize;
853 on = uio->uio_offset & (biosize-1);
854 n = min((unsigned)(biosize - on), uio->uio_resid);
857 * Handle direct append and file extension cases, calculate
858 * unaligned buffer size.
861 if (uio->uio_offset == np->n_size && n) {
863 * Get the buffer (in its pre-append state to maintain
864 * B_CACHE if it was previously set). Resize the
865 * nfsnode after we have locked the buffer to prevent
866 * readers from reading garbage.
869 bp = nfs_getcacheblk(vp, lbn, bcount, td);
874 np->n_size = uio->uio_offset + n;
875 np->n_flag |= NMODIFIED;
876 vnode_pager_setsize(vp, np->n_size);
878 save = bp->b_flags & B_CACHE;
880 allocbuf(bp, bcount);
885 * Obtain the locked cache block first, and then
886 * adjust the file's size as appropriate.
889 if ((off_t)lbn * biosize + bcount < np->n_size) {
890 if ((off_t)(lbn + 1) * biosize < np->n_size)
893 bcount = np->n_size - (off_t)lbn * biosize;
895 bp = nfs_getcacheblk(vp, lbn, bcount, td);
896 if (uio->uio_offset + n > np->n_size) {
897 np->n_size = uio->uio_offset + n;
898 np->n_flag |= NMODIFIED;
899 vnode_pager_setsize(vp, np->n_size);
909 * Issue a READ if B_CACHE is not set. In special-append
910 * mode, B_CACHE is based on the buffer prior to the write
911 * op and is typically set, avoiding the read. If a read
912 * is required in special append mode, the server will
913 * probably send us a short-read since we extended the file
914 * on our end, resulting in b_resid == 0 and, thusly,
915 * B_CACHE getting set.
917 * We can also avoid issuing the read if the write covers
918 * the entire buffer. We have to make sure the buffer state
919 * is reasonable in this case since we will not be initiating
920 * I/O. See the comments in kern/vfs_bio.c's getblk() for
923 * B_CACHE may also be set due to the buffer being cached
927 if (on == 0 && n == bcount) {
928 bp->b_flags |= B_CACHE;
929 bp->b_flags &= ~(B_ERROR | B_INVAL);
932 if ((bp->b_flags & B_CACHE) == 0) {
933 bp->b_flags |= B_READ;
934 vfs_busy_pages(bp, 0);
935 error = nfs_doio(bp, td);
945 np->n_flag |= NMODIFIED;
948 * If dirtyend exceeds file size, chop it down. This should
949 * not normally occur but there is an append race where it
950 * might occur XXX, so we log it.
952 * If the chopping creates a reverse-indexed or degenerate
953 * situation with dirtyoff/end, we 0 both of them.
956 if (bp->b_dirtyend > bcount) {
957 printf("NFS append race @%lx:%d\n",
958 (long)bp->b_blkno * DEV_BSIZE,
959 bp->b_dirtyend - bcount);
960 bp->b_dirtyend = bcount;
963 if (bp->b_dirtyoff >= bp->b_dirtyend)
964 bp->b_dirtyoff = bp->b_dirtyend = 0;
967 * If the new write will leave a contiguous dirty
968 * area, just update the b_dirtyoff and b_dirtyend,
969 * otherwise force a write rpc of the old dirty area.
971 * While it is possible to merge discontiguous writes due to
972 * our having a B_CACHE buffer ( and thus valid read data
973 * for the hole), we don't because it could lead to
974 * significant cache coherency problems with multiple clients,
975 * especially if locking is implemented later on.
977 * as an optimization we could theoretically maintain
978 * a linked list of discontinuous areas, but we would still
979 * have to commit them separately so there isn't much
980 * advantage to it except perhaps a bit of asynchronization.
983 if (bp->b_dirtyend > 0 &&
984 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
985 if (VOP_BWRITE(bp->b_vp, bp) == EINTR) {
993 * Check for valid write lease and get one as required.
994 * In case getblk() and/or bwrite() delayed us.
996 if ((nmp->nm_flag & NFSMNT_NQNFS) &&
997 NQNFS_CKINVALID(vp, np, ND_WRITE)) {
999 error = nqnfs_getlease(vp, ND_WRITE, td);
1000 } while (error == NQNFS_EXPIRED);
1005 if (np->n_lrev != np->n_brev ||
1006 (np->n_flag & NQNFSNONCACHE)) {
1008 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
1011 np->n_brev = np->n_lrev;
1016 error = uiomove((char *)bp->b_data + on, n, uio);
1019 * Since this block is being modified, it must be written
1020 * again and not just committed. Since write clustering does
1021 * not work for the stage 1 data write, only the stage 2
1022 * commit rpc, we have to clear B_CLUSTEROK as well.
1024 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1027 bp->b_flags |= B_ERROR;
1033 * Only update dirtyoff/dirtyend if not a degenerate
1037 if (bp->b_dirtyend > 0) {
1038 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1039 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1041 bp->b_dirtyoff = on;
1042 bp->b_dirtyend = on + n;
1044 vfs_bio_set_validclean(bp, on, n);
1047 * If IO_NOWDRAIN then set B_NOWDRAIN (e.g. nfs-backed VN
1048 * filesystem). XXX also use for loopback NFS mounts.
1050 if (ioflag & IO_NOWDRAIN)
1051 bp->b_flags |= B_NOWDRAIN;
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 ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) {
1060 if (ioflag & IO_INVAL)
1061 bp->b_flags |= B_NOCACHE;
1062 error = VOP_BWRITE(bp->b_vp, bp);
1065 if (np->n_flag & NQNFSNONCACHE) {
1066 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
1070 } else if ((n + on) == biosize &&
1071 (nmp->nm_flag & NFSMNT_NQNFS) == 0) {
1072 bp->b_flags |= B_ASYNC;
1073 (void)nfs_writebp(bp, 0, 0);
1077 } while (uio->uio_resid > 0 && n > 0);
1080 nfs_rsunlock(np, td);
1086 * Get an nfs cache block.
1088 * Allocate a new one if the block isn't currently in the cache
1089 * and return the block marked busy. If the calling process is
1090 * interrupted by a signal for an interruptible mount point, return
1093 * The caller must carefully deal with the possible B_INVAL state of
1094 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it
1095 * indirectly), so synchronous reads can be issued without worrying about
1096 * the B_INVAL state. We have to be a little more careful when dealing
1097 * with writes (see comments in nfs_write()) when extending a file past
1101 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1105 struct nfsmount *nmp;
1110 if (nmp->nm_flag & NFSMNT_INT) {
1111 bp = getblk(vp, bn, size, PCATCH, 0);
1112 while (bp == (struct buf *)0) {
1113 if (nfs_sigintr(nmp, (struct nfsreq *)0, td))
1114 return ((struct buf *)0);
1115 bp = getblk(vp, bn, size, 0, 2 * hz);
1118 bp = getblk(vp, bn, size, 0, 0);
1121 if (vp->v_type == VREG) {
1124 biosize = mp->mnt_stat.f_iosize;
1125 bp->b_blkno = bn * (biosize / DEV_BSIZE);
1131 * Flush and invalidate all dirty buffers. If another process is already
1132 * doing the flush, just wait for completion.
1135 nfs_vinvalbuf(struct vnode *vp, int flags,
1136 struct thread *td, int intrflg)
1138 struct nfsnode *np = VTONFS(vp);
1139 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1140 int error = 0, slpflag, slptimeo;
1142 if (vp->v_flag & VRECLAIMED)
1145 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1155 * First wait for any other process doing a flush to complete.
1157 while (np->n_flag & NFLUSHINPROG) {
1158 np->n_flag |= NFLUSHWANT;
1159 error = tsleep((caddr_t)&np->n_flag, 0, "nfsvinval", slptimeo);
1160 if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, td))
1165 * Now, flush as required.
1167 np->n_flag |= NFLUSHINPROG;
1168 error = vinvalbuf(vp, flags, td, slpflag, 0);
1170 if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, td)) {
1171 np->n_flag &= ~NFLUSHINPROG;
1172 if (np->n_flag & NFLUSHWANT) {
1173 np->n_flag &= ~NFLUSHWANT;
1174 wakeup((caddr_t)&np->n_flag);
1178 error = vinvalbuf(vp, flags, td, 0, slptimeo);
1180 np->n_flag &= ~(NMODIFIED | NFLUSHINPROG);
1181 if (np->n_flag & NFLUSHWANT) {
1182 np->n_flag &= ~NFLUSHWANT;
1183 wakeup((caddr_t)&np->n_flag);
1189 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1190 * This is mainly to avoid queueing async I/O requests when the nfsiods
1191 * are all hung on a dead server.
1193 * Note: nfs_asyncio() does not clear (B_ERROR|B_INVAL) but when the bp
1194 * is eventually dequeued by the async daemon, nfs_doio() *will*.
1197 nfs_asyncio(struct buf *bp, struct thread *td)
1199 struct nfsmount *nmp;
1207 * If no async daemons then return EIO to force caller to run the rpc
1210 if (nfs_numasync == 0)
1213 nmp = VFSTONFS(bp->b_vp->v_mount);
1216 * Commits are usually short and sweet so lets save some cpu and
1217 * leave the async daemons for more important rpc's (such as reads
1220 if ((bp->b_flags & (B_READ|B_NEEDCOMMIT)) == B_NEEDCOMMIT &&
1221 (nmp->nm_bufqiods > nfs_numasync / 2)) {
1226 if (nmp->nm_flag & NFSMNT_INT)
1231 * Find a free iod to process this request.
1233 for (i = 0; i < NFS_MAXASYNCDAEMON; i++)
1234 if (nfs_iodwant[i]) {
1236 * Found one, so wake it up and tell it which
1240 ("nfs_asyncio: waking iod %d for mount %p\n",
1242 nfs_iodwant[i] = NULL;
1243 nfs_iodmount[i] = nmp;
1245 wakeup((caddr_t)&nfs_iodwant[i]);
1251 * If none are free, we may already have an iod working on this mount
1252 * point. If so, it will process our request.
1255 if (nmp->nm_bufqiods > 0) {
1257 ("nfs_asyncio: %d iods are already processing mount %p\n",
1258 nmp->nm_bufqiods, nmp));
1264 * If we have an iod which can process the request, then queue
1269 * Ensure that the queue never grows too large. We still want
1270 * to asynchronize so we block rather then return EIO.
1272 while (nmp->nm_bufqlen >= 2*nfs_numasync) {
1274 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp));
1275 nmp->nm_bufqwant = TRUE;
1276 error = tsleep(&nmp->nm_bufq, slpflag,
1277 "nfsaio", slptimeo);
1279 if (nfs_sigintr(nmp, NULL, td))
1281 if (slpflag == PCATCH) {
1287 * We might have lost our iod while sleeping,
1288 * so check and loop if nescessary.
1290 if (nmp->nm_bufqiods == 0) {
1292 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1297 if ((bp->b_flags & B_READ) == 0)
1298 bp->b_flags |= B_WRITEINPROG;
1301 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1307 * All the iods are busy on other mounts, so return EIO to
1308 * force the caller to process the i/o synchronously.
1310 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n"));
1315 * Do an I/O operation to/from a cache block. This may be called
1316 * synchronously or from an nfsiod.
1318 * NOTE! TD MIGHT BE NULL
1321 nfs_doio(struct buf *bp, struct thread *td)
1326 struct nfsmount *nmp;
1327 int error = 0, iomode, must_commit = 0;
1333 nmp = VFSTONFS(vp->v_mount);
1335 uiop->uio_iov = &io;
1336 uiop->uio_iovcnt = 1;
1337 uiop->uio_segflg = UIO_SYSSPACE;
1341 * clear B_ERROR and B_INVAL state prior to initiating the I/O. We
1342 * do this here so we do not have to do it in all the code that
1345 bp->b_flags &= ~(B_ERROR | B_INVAL);
1347 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp));
1350 * Historically, paging was done with physio, but no more.
1352 if (bp->b_flags & B_PHYS) {
1354 * ...though reading /dev/drum still gets us here.
1356 io.iov_len = uiop->uio_resid = bp->b_bcount;
1357 /* mapping was done by vmapbuf() */
1358 io.iov_base = bp->b_data;
1359 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1360 if (bp->b_flags & B_READ) {
1361 uiop->uio_rw = UIO_READ;
1362 nfsstats.read_physios++;
1363 error = nfs_readrpc(vp, uiop);
1367 iomode = NFSV3WRITE_DATASYNC;
1368 uiop->uio_rw = UIO_WRITE;
1369 nfsstats.write_physios++;
1370 error = nfs_writerpc(vp, uiop, &iomode, &com);
1373 bp->b_flags |= B_ERROR;
1374 bp->b_error = error;
1376 } else if (bp->b_flags & B_READ) {
1377 io.iov_len = uiop->uio_resid = bp->b_bcount;
1378 io.iov_base = bp->b_data;
1379 uiop->uio_rw = UIO_READ;
1381 switch (vp->v_type) {
1383 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1384 nfsstats.read_bios++;
1385 error = nfs_readrpc(vp, uiop);
1388 if (uiop->uio_resid) {
1390 * If we had a short read with no error, we must have
1391 * hit a file hole. We should zero-fill the remainder.
1392 * This can also occur if the server hits the file EOF.
1394 * Holes used to be able to occur due to pending
1395 * writes, but that is not possible any longer.
1397 int nread = bp->b_bcount - uiop->uio_resid;
1398 int left = uiop->uio_resid;
1401 bzero((char *)bp->b_data + nread, left);
1402 uiop->uio_resid = 0;
1405 if (td && td->td_proc && (vp->v_flag & VTEXT) &&
1406 (((nmp->nm_flag & NFSMNT_NQNFS) &&
1407 NQNFS_CKINVALID(vp, np, ND_READ) &&
1408 np->n_lrev != np->n_brev) ||
1409 (!(nmp->nm_flag & NFSMNT_NQNFS) &&
1410 np->n_mtime != np->n_vattr.va_mtime.tv_sec))) {
1411 uprintf("Process killed due to text file modification\n");
1412 psignal(td->td_proc, SIGKILL);
1417 uiop->uio_offset = (off_t)0;
1418 nfsstats.readlink_bios++;
1419 error = nfs_readlinkrpc(vp, uiop);
1422 nfsstats.readdir_bios++;
1423 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1424 if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
1425 error = nfs_readdirplusrpc(vp, uiop);
1426 if (error == NFSERR_NOTSUPP)
1427 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1429 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1430 error = nfs_readdirrpc(vp, uiop);
1432 * end-of-directory sets B_INVAL but does not generate an
1435 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1436 bp->b_flags |= B_INVAL;
1439 printf("nfs_doio: type %x unexpected\n",vp->v_type);
1443 bp->b_flags |= B_ERROR;
1444 bp->b_error = error;
1448 * If we only need to commit, try to commit
1450 if (bp->b_flags & B_NEEDCOMMIT) {
1454 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1455 bp->b_flags |= B_WRITEINPROG;
1456 retv = nfs_commit(bp->b_vp, off,
1457 bp->b_dirtyend - bp->b_dirtyoff, td);
1458 bp->b_flags &= ~B_WRITEINPROG;
1460 bp->b_dirtyoff = bp->b_dirtyend = 0;
1461 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1466 if (retv == NFSERR_STALEWRITEVERF) {
1467 nfs_clearcommit(bp->b_vp->v_mount);
1472 * Setup for actual write
1475 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1476 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1478 if (bp->b_dirtyend > bp->b_dirtyoff) {
1479 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1481 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1483 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1484 uiop->uio_rw = UIO_WRITE;
1485 nfsstats.write_bios++;
1487 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1488 iomode = NFSV3WRITE_UNSTABLE;
1490 iomode = NFSV3WRITE_FILESYNC;
1492 bp->b_flags |= B_WRITEINPROG;
1493 error = nfs_writerpc(vp, uiop, &iomode, &must_commit);
1496 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1497 * to cluster the buffers needing commit. This will allow
1498 * the system to submit a single commit rpc for the whole
1499 * cluster. We can do this even if the buffer is not 100%
1500 * dirty (relative to the NFS blocksize), so we optimize the
1501 * append-to-file-case.
1503 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1504 * cleared because write clustering only works for commit
1505 * rpc's, not for the data portion of the write).
1508 if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1509 bp->b_flags |= B_NEEDCOMMIT;
1510 if (bp->b_dirtyoff == 0
1511 && bp->b_dirtyend == bp->b_bcount)
1512 bp->b_flags |= B_CLUSTEROK;
1514 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1516 bp->b_flags &= ~B_WRITEINPROG;
1519 * For an interrupted write, the buffer is still valid
1520 * and the write hasn't been pushed to the server yet,
1521 * so we can't set B_ERROR and report the interruption
1522 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1523 * is not relevant, so the rpc attempt is essentially
1524 * a noop. For the case of a V3 write rpc not being
1525 * committed to stable storage, the block is still
1526 * dirty and requires either a commit rpc or another
1527 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1528 * the block is reused. This is indicated by setting
1529 * the B_DELWRI and B_NEEDCOMMIT flags.
1531 * If the buffer is marked B_PAGING, it does not reside on
1532 * the vp's paging queues so we cannot call bdirty(). The
1533 * bp in this case is not an NFS cache block so we should
1537 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1541 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1542 if ((bp->b_flags & B_PAGING) == 0) {
1544 bp->b_flags &= ~B_DONE;
1546 if (error && (bp->b_flags & B_ASYNC) == 0)
1547 bp->b_flags |= B_EINTR;
1551 bp->b_flags |= B_ERROR;
1552 bp->b_error = np->n_error = error;
1553 np->n_flag |= NWRITEERR;
1555 bp->b_dirtyoff = bp->b_dirtyend = 0;
1563 bp->b_resid = uiop->uio_resid;
1565 nfs_clearcommit(vp->v_mount);
1571 * Used to aid in handling ftruncate() operations on the NFS client side.
1572 * Truncation creates a number of special problems for NFS. We have to
1573 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1574 * we have to properly handle VM pages or (potentially dirty) buffers
1575 * that straddle the truncation point.
1579 nfs_meta_setsize(struct vnode *vp, struct thread *td, u_quad_t nsize)
1581 struct nfsnode *np = VTONFS(vp);
1582 u_quad_t tsize = np->n_size;
1583 int biosize = vp->v_mount->mnt_stat.f_iosize;
1588 if (np->n_size < tsize) {
1594 * vtruncbuf() doesn't get the buffer overlapping the
1595 * truncation point. We may have a B_DELWRI and/or B_CACHE
1596 * buffer that now needs to be truncated.
1598 error = vtruncbuf(vp, td, nsize, biosize);
1599 lbn = nsize / biosize;
1600 bufsize = nsize & (biosize - 1);
1601 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1602 if (bp->b_dirtyoff > bp->b_bcount)
1603 bp->b_dirtyoff = bp->b_bcount;
1604 if (bp->b_dirtyend > bp->b_bcount)
1605 bp->b_dirtyend = bp->b_bcount;
1606 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1609 vnode_pager_setsize(vp, nsize);