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: src/sys/nfs/nfs_bio.c,v 1.83.2.4 2002/12/29 18:19:53 dillon Exp $
38 * $DragonFly: src/sys/vfs/nfs/nfs_bio.c,v 1.12 2004/02/13 18:52:35 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>
53 #include <vm/vm_extern.h>
54 #include <vm/vm_page.h>
55 #include <vm/vm_object.h>
56 #include <vm/vm_pager.h>
57 #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.
80 struct vop_getpages_args /* {
85 vm_ooffset_t a_offset;
88 struct thread *td = curthread; /* XXX */
89 int i, error, nextoff, size, toff, count, npages;
99 nmp = VFSTONFS(vp->v_mount);
103 if (vp->v_object == NULL) {
104 printf("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 * If the requested page is partially valid, just return it and
116 * allow the pager to zero-out the blanks. Partially valid pages
117 * can only occur at the file EOF.
121 vm_page_t m = pages[ap->a_reqpage];
124 /* handled by vm_fault now */
125 /* vm_page_zero_invalid(m, TRUE); */
126 for (i = 0; i < npages; ++i) {
127 if (i != ap->a_reqpage)
128 vnode_pager_freepage(pages[i]);
135 * We use only the kva address for the buffer, but this is extremely
136 * convienient and fast.
138 bp = getpbuf(&nfs_pbuf_freecnt);
140 kva = (vm_offset_t) bp->b_data;
141 pmap_qenter(kva, pages, npages);
143 iov.iov_base = (caddr_t) kva;
147 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
148 uio.uio_resid = count;
149 uio.uio_segflg = UIO_SYSSPACE;
150 uio.uio_rw = UIO_READ;
153 error = nfs_readrpc(vp, &uio);
154 pmap_qremove(kva, npages);
156 relpbuf(bp, &nfs_pbuf_freecnt);
158 if (error && (uio.uio_resid == count)) {
159 printf("nfs_getpages: error %d\n", error);
160 for (i = 0; i < npages; ++i) {
161 if (i != ap->a_reqpage)
162 vnode_pager_freepage(pages[i]);
164 return VM_PAGER_ERROR;
168 * Calculate the number of bytes read and validate only that number
169 * of bytes. Note that due to pending writes, size may be 0. This
170 * does not mean that the remaining data is invalid!
173 size = count - uio.uio_resid;
175 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
177 nextoff = toff + PAGE_SIZE;
180 m->flags &= ~PG_ZERO;
182 if (nextoff <= size) {
184 * Read operation filled an entire page
186 m->valid = VM_PAGE_BITS_ALL;
188 } else if (size > toff) {
190 * Read operation filled a partial page.
193 vm_page_set_validclean(m, 0, size - toff);
194 /* handled by vm_fault now */
195 /* vm_page_zero_invalid(m, TRUE); */
198 * Read operation was short. If no error occured
199 * we may have hit a zero-fill section. We simply
200 * leave valid set to 0.
204 if (i != ap->a_reqpage) {
206 * Whether or not to leave the page activated is up in
207 * the air, but we should put the page on a page queue
208 * somewhere (it already is in the object). Result:
209 * It appears that emperical results show that
210 * deactivating pages is best.
214 * Just in case someone was asking for this page we
215 * now tell them that it is ok to use.
218 if (m->flags & PG_WANTED)
221 vm_page_deactivate(m);
224 vnode_pager_freepage(m);
232 * Vnode op for VM putpages.
236 struct vop_putpages_args /* {
242 vm_ooffset_t a_offset;
245 struct thread *td = curthread;
250 int iomode, must_commit, i, error, npages, count;
254 struct nfsmount *nmp;
260 nmp = VFSTONFS(vp->v_mount);
263 rtvals = ap->a_rtvals;
264 npages = btoc(count);
265 offset = IDX_TO_OFF(pages[0]->pindex);
267 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
268 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
269 (void)nfs_fsinfo(nmp, vp, td);
271 for (i = 0; i < npages; i++) {
272 rtvals[i] = VM_PAGER_AGAIN;
276 * When putting pages, do not extend file past EOF.
279 if (offset + count > np->n_size) {
280 count = np->n_size - offset;
286 * We use only the kva address for the buffer, but this is extremely
287 * convienient and fast.
289 bp = getpbuf(&nfs_pbuf_freecnt);
291 kva = (vm_offset_t) bp->b_data;
292 pmap_qenter(kva, pages, npages);
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);
311 pmap_qremove(kva, npages);
312 relpbuf(bp, &nfs_pbuf_freecnt);
315 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE;
316 for (i = 0; i < nwritten; i++) {
317 rtvals[i] = VM_PAGER_OK;
318 vm_page_undirty(pages[i]);
321 nfs_clearcommit(vp->v_mount);
327 * Vnode op for read using bio
330 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag)
332 struct nfsnode *np = VTONFS(vp);
334 struct buf *bp = 0, *rabp;
337 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
341 int nra, error = 0, n = 0, on = 0;
344 if (uio->uio_rw != UIO_READ)
345 panic("nfs_read mode");
347 if (uio->uio_resid == 0)
349 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
353 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
354 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
355 (void)nfs_fsinfo(nmp, vp, td);
356 if (vp->v_type != VDIR &&
357 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
359 biosize = vp->v_mount->mnt_stat.f_iosize;
360 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
362 * For nfs, cache consistency can only be maintained approximately.
363 * Although RFC1094 does not specify the criteria, the following is
364 * believed to be compatible with the reference port.
365 * For nqnfs, full cache consistency is maintained within the loop.
367 * If the file's modify time on the server has changed since the
368 * last read rpc or you have written to the file,
369 * you may have lost data cache consistency with the
370 * server, so flush all of the file's data out of the cache.
371 * Then force a getattr rpc to ensure that you have up to date
373 * NB: This implies that cache data can be read when up to
374 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
375 * attributes this could be forced by setting n_attrstamp to 0 before
376 * the VOP_GETATTR() call.
378 if ((nmp->nm_flag & NFSMNT_NQNFS) == 0) {
379 if (np->n_flag & NMODIFIED) {
380 if (vp->v_type != VREG) {
381 if (vp->v_type != VDIR)
382 panic("nfs: bioread, not dir");
384 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
389 error = VOP_GETATTR(vp, &vattr, td);
392 np->n_mtime = vattr.va_mtime.tv_sec;
394 error = VOP_GETATTR(vp, &vattr, td);
397 if (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;
410 * Get a valid lease. If cached data is stale, flush it.
412 if (nmp->nm_flag & NFSMNT_NQNFS) {
413 if (NQNFS_CKINVALID(vp, np, ND_READ)) {
415 error = nqnfs_getlease(vp, ND_READ, td);
416 } while (error == NQNFS_EXPIRED);
419 if (np->n_lrev != np->n_brev ||
420 (np->n_flag & NQNFSNONCACHE) ||
421 ((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) {
422 if (vp->v_type == VDIR)
424 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
427 np->n_brev = np->n_lrev;
429 } else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) {
431 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
436 if (np->n_flag & NQNFSNONCACHE) {
437 switch (vp->v_type) {
439 return (nfs_readrpc(vp, uio));
441 return (nfs_readlinkrpc(vp, uio));
445 printf(" NQNFSNONCACHE: type %x unexpected\n",
449 switch (vp->v_type) {
451 nfsstats.biocache_reads++;
452 lbn = uio->uio_offset / biosize;
453 on = uio->uio_offset & (biosize - 1);
456 * Start the read ahead(s), as required.
458 if (nfs_numasync > 0 && nmp->nm_readahead > 0) {
459 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
460 (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) {
461 rabn = lbn + 1 + nra;
462 if (!incore(vp, rabn)) {
463 rabp = nfs_getcacheblk(vp, rabn, biosize, td);
466 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
467 rabp->b_flags |= (B_READ | B_ASYNC);
468 vfs_busy_pages(rabp, 0);
469 if (nfs_asyncio(rabp, td)) {
470 rabp->b_flags |= B_INVAL|B_ERROR;
471 vfs_unbusy_pages(rabp);
483 * Obtain the buffer cache block. Figure out the buffer size
484 * when we are at EOF. If we are modifying the size of the
485 * buffer based on an EOF condition we need to hold
486 * nfs_rslock() through obtaining the buffer to prevent
487 * a potential writer-appender from messing with n_size.
488 * Otherwise we may accidently truncate the buffer and
491 * Note that bcount is *not* DEV_BSIZE aligned.
496 if ((off_t)lbn * biosize >= np->n_size) {
498 } else if ((off_t)(lbn + 1) * biosize > np->n_size) {
499 bcount = np->n_size - (off_t)lbn * biosize;
501 if (bcount != biosize) {
502 switch(nfs_rslock(np, td)) {
515 bp = nfs_getcacheblk(vp, lbn, bcount, td);
517 if (bcount != biosize)
518 nfs_rsunlock(np, td);
523 * If B_CACHE is not set, we must issue the read. If this
524 * fails, we return an error.
527 if ((bp->b_flags & B_CACHE) == 0) {
528 bp->b_flags |= B_READ;
529 vfs_busy_pages(bp, 0);
530 error = nfs_doio(bp, td);
538 * on is the offset into the current bp. Figure out how many
539 * bytes we can copy out of the bp. Note that bcount is
540 * NOT DEV_BSIZE aligned.
542 * Then figure out how many bytes we can copy into the uio.
547 n = min((unsigned)(bcount - on), uio->uio_resid);
550 nfsstats.biocache_readlinks++;
551 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
554 if ((bp->b_flags & B_CACHE) == 0) {
555 bp->b_flags |= B_READ;
556 vfs_busy_pages(bp, 0);
557 error = nfs_doio(bp, td);
559 bp->b_flags |= B_ERROR;
564 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
568 nfsstats.biocache_readdirs++;
569 if (np->n_direofoffset
570 && uio->uio_offset >= np->n_direofoffset) {
573 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
574 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
575 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
578 if ((bp->b_flags & B_CACHE) == 0) {
579 bp->b_flags |= B_READ;
580 vfs_busy_pages(bp, 0);
581 error = nfs_doio(bp, td);
585 while (error == NFSERR_BAD_COOKIE) {
586 printf("got bad cookie vp %p bp %p\n", vp, bp);
588 error = nfs_vinvalbuf(vp, 0, td, 1);
590 * Yuck! The directory has been modified on the
591 * server. The only way to get the block is by
592 * reading from the beginning to get all the
595 * Leave the last bp intact unless there is an error.
596 * Loop back up to the while if the error is another
597 * NFSERR_BAD_COOKIE (double yuch!).
599 for (i = 0; i <= lbn && !error; i++) {
600 if (np->n_direofoffset
601 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
603 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
606 if ((bp->b_flags & B_CACHE) == 0) {
607 bp->b_flags |= B_READ;
608 vfs_busy_pages(bp, 0);
609 error = nfs_doio(bp, td);
611 * no error + B_INVAL == directory EOF,
614 if (error == 0 && (bp->b_flags & B_INVAL))
618 * An error will throw away the block and the
619 * for loop will break out. If no error and this
620 * is not the block we want, we throw away the
621 * block and go for the next one via the for loop.
623 if (error || i < lbn)
628 * The above while is repeated if we hit another cookie
629 * error. If we hit an error and it wasn't a cookie error,
637 * If not eof and read aheads are enabled, start one.
638 * (You need the current block first, so that you have the
639 * directory offset cookie of the next block.)
641 if (nfs_numasync > 0 && nmp->nm_readahead > 0 &&
642 (bp->b_flags & B_INVAL) == 0 &&
643 (np->n_direofoffset == 0 ||
644 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
645 !(np->n_flag & NQNFSNONCACHE) &&
646 !incore(vp, lbn + 1)) {
647 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
649 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
650 rabp->b_flags |= (B_READ | B_ASYNC);
651 vfs_busy_pages(rabp, 0);
652 if (nfs_asyncio(rabp, td)) {
653 rabp->b_flags |= B_INVAL|B_ERROR;
654 vfs_unbusy_pages(rabp);
663 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
664 * chopped for the EOF condition, we cannot tell how large
665 * NFS directories are going to be until we hit EOF. So
666 * an NFS directory buffer is *not* chopped to its EOF. Now,
667 * it just so happens that b_resid will effectively chop it
668 * to EOF. *BUT* this information is lost if the buffer goes
669 * away and is reconstituted into a B_CACHE state ( due to
670 * being VMIO ) later. So we keep track of the directory eof
671 * in np->n_direofoffset and chop it off as an extra step
674 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
675 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
676 n = np->n_direofoffset - uio->uio_offset;
679 printf(" nfs_bioread: type %x unexpected\n",vp->v_type);
684 error = uiomove(bp->b_data + on, (int)n, uio);
686 switch (vp->v_type) {
694 * Invalidate buffer if caching is disabled, forcing a
695 * re-read from the remote later.
697 if (np->n_flag & NQNFSNONCACHE)
698 bp->b_flags |= B_INVAL;
701 printf(" nfs_bioread: type %x unexpected\n",vp->v_type);
704 } while (error == 0 && uio->uio_resid > 0 && n > 0);
709 * Vnode op for write using bio
713 struct vop_write_args /* {
717 struct ucred *a_cred;
721 struct uio *uio = ap->a_uio;
722 struct thread *td = uio->uio_td;
723 struct vnode *vp = ap->a_vp;
724 struct nfsnode *np = VTONFS(vp);
725 int ioflag = ap->a_ioflag;
728 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
731 int n, on, error = 0, iomode, must_commit;
735 if (uio->uio_rw != UIO_WRITE)
736 panic("nfs_write mode");
737 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread)
738 panic("nfs_write proc");
740 if (vp->v_type != VREG)
742 if (np->n_flag & NWRITEERR) {
743 np->n_flag &= ~NWRITEERR;
744 return (np->n_error);
746 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
747 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
748 (void)nfs_fsinfo(nmp, vp, td);
751 * Synchronously flush pending buffers if we are in synchronous
752 * mode or if we are appending.
754 if (ioflag & (IO_APPEND | IO_SYNC)) {
755 if (np->n_flag & NMODIFIED) {
757 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
764 * If IO_APPEND then load uio_offset. We restart here if we cannot
765 * get the append lock.
768 if (ioflag & IO_APPEND) {
770 error = VOP_GETATTR(vp, &vattr, td);
773 uio->uio_offset = np->n_size;
776 if (uio->uio_offset < 0)
778 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
780 if (uio->uio_resid == 0)
784 * We need to obtain the rslock if we intend to modify np->n_size
785 * in order to guarentee the append point with multiple contending
786 * writers, to guarentee that no other appenders modify n_size
787 * while we are trying to obtain a truncated buffer (i.e. to avoid
788 * accidently truncating data written by another appender due to
789 * the race), and to ensure that the buffer is populated prior to
790 * our extending of the file. We hold rslock through the entire
793 * Note that we do not synchronize the case where someone truncates
794 * the file while we are appending to it because attempting to lock
795 * this case may deadlock other parts of the system unexpectedly.
797 if ((ioflag & IO_APPEND) ||
798 uio->uio_offset + uio->uio_resid > np->n_size) {
799 switch(nfs_rslock(np, td)) {
814 * Maybe this should be above the vnode op call, but so long as
815 * file servers have no limits, i don't think it matters
817 if (td->td_proc && uio->uio_offset + uio->uio_resid >
818 td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
819 psignal(td->td_proc, SIGXFSZ);
821 nfs_rsunlock(np, td);
825 biosize = vp->v_mount->mnt_stat.f_iosize;
829 * Check for a valid write lease.
831 if ((nmp->nm_flag & NFSMNT_NQNFS) &&
832 NQNFS_CKINVALID(vp, np, ND_WRITE)) {
834 error = nqnfs_getlease(vp, ND_WRITE, td);
835 } while (error == NQNFS_EXPIRED);
838 if (np->n_lrev != np->n_brev ||
839 (np->n_flag & NQNFSNONCACHE)) {
840 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
843 np->n_brev = np->n_lrev;
846 if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) {
847 iomode = NFSV3WRITE_FILESYNC;
848 error = nfs_writerpc(vp, uio, &iomode, &must_commit);
850 nfs_clearcommit(vp->v_mount);
853 nfsstats.biocache_writes++;
854 lbn = uio->uio_offset / biosize;
855 on = uio->uio_offset & (biosize-1);
856 n = min((unsigned)(biosize - on), uio->uio_resid);
859 * Handle direct append and file extension cases, calculate
860 * unaligned buffer size.
863 if (uio->uio_offset == np->n_size && n) {
865 * Get the buffer (in its pre-append state to maintain
866 * B_CACHE if it was previously set). Resize the
867 * nfsnode after we have locked the buffer to prevent
868 * readers from reading garbage.
871 bp = nfs_getcacheblk(vp, lbn, bcount, td);
876 np->n_size = uio->uio_offset + n;
877 np->n_flag |= NMODIFIED;
878 vnode_pager_setsize(vp, np->n_size);
880 save = bp->b_flags & B_CACHE;
882 allocbuf(bp, bcount);
887 * Obtain the locked cache block first, and then
888 * adjust the file's size as appropriate.
891 if ((off_t)lbn * biosize + bcount < np->n_size) {
892 if ((off_t)(lbn + 1) * biosize < np->n_size)
895 bcount = np->n_size - (off_t)lbn * biosize;
897 bp = nfs_getcacheblk(vp, lbn, bcount, td);
898 if (uio->uio_offset + n > np->n_size) {
899 np->n_size = uio->uio_offset + n;
900 np->n_flag |= NMODIFIED;
901 vnode_pager_setsize(vp, np->n_size);
911 * Issue a READ if B_CACHE is not set. In special-append
912 * mode, B_CACHE is based on the buffer prior to the write
913 * op and is typically set, avoiding the read. If a read
914 * is required in special append mode, the server will
915 * probably send us a short-read since we extended the file
916 * on our end, resulting in b_resid == 0 and, thusly,
917 * B_CACHE getting set.
919 * We can also avoid issuing the read if the write covers
920 * the entire buffer. We have to make sure the buffer state
921 * is reasonable in this case since we will not be initiating
922 * I/O. See the comments in kern/vfs_bio.c's getblk() for
925 * B_CACHE may also be set due to the buffer being cached
929 if (on == 0 && n == bcount) {
930 bp->b_flags |= B_CACHE;
931 bp->b_flags &= ~(B_ERROR | B_INVAL);
934 if ((bp->b_flags & B_CACHE) == 0) {
935 bp->b_flags |= B_READ;
936 vfs_busy_pages(bp, 0);
937 error = nfs_doio(bp, td);
947 np->n_flag |= NMODIFIED;
950 * If dirtyend exceeds file size, chop it down. This should
951 * not normally occur but there is an append race where it
952 * might occur XXX, so we log it.
954 * If the chopping creates a reverse-indexed or degenerate
955 * situation with dirtyoff/end, we 0 both of them.
958 if (bp->b_dirtyend > bcount) {
959 printf("NFS append race @%lx:%d\n",
960 (long)bp->b_blkno * DEV_BSIZE,
961 bp->b_dirtyend - bcount);
962 bp->b_dirtyend = bcount;
965 if (bp->b_dirtyoff >= bp->b_dirtyend)
966 bp->b_dirtyoff = bp->b_dirtyend = 0;
969 * If the new write will leave a contiguous dirty
970 * area, just update the b_dirtyoff and b_dirtyend,
971 * otherwise force a write rpc of the old dirty area.
973 * While it is possible to merge discontiguous writes due to
974 * our having a B_CACHE buffer ( and thus valid read data
975 * for the hole), we don't because it could lead to
976 * significant cache coherency problems with multiple clients,
977 * especially if locking is implemented later on.
979 * as an optimization we could theoretically maintain
980 * a linked list of discontinuous areas, but we would still
981 * have to commit them separately so there isn't much
982 * advantage to it except perhaps a bit of asynchronization.
985 if (bp->b_dirtyend > 0 &&
986 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
987 if (VOP_BWRITE(bp->b_vp, bp) == EINTR) {
995 * Check for valid write lease and get one as required.
996 * In case getblk() and/or bwrite() delayed us.
998 if ((nmp->nm_flag & NFSMNT_NQNFS) &&
999 NQNFS_CKINVALID(vp, np, ND_WRITE)) {
1001 error = nqnfs_getlease(vp, ND_WRITE, td);
1002 } while (error == NQNFS_EXPIRED);
1007 if (np->n_lrev != np->n_brev ||
1008 (np->n_flag & NQNFSNONCACHE)) {
1010 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
1013 np->n_brev = np->n_lrev;
1018 error = uiomove((char *)bp->b_data + on, n, uio);
1021 * Since this block is being modified, it must be written
1022 * again and not just committed. Since write clustering does
1023 * not work for the stage 1 data write, only the stage 2
1024 * commit rpc, we have to clear B_CLUSTEROK as well.
1026 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1029 bp->b_flags |= B_ERROR;
1035 * Only update dirtyoff/dirtyend if not a degenerate
1039 if (bp->b_dirtyend > 0) {
1040 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1041 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1043 bp->b_dirtyoff = on;
1044 bp->b_dirtyend = on + n;
1046 vfs_bio_set_validclean(bp, on, n);
1049 * If IO_NOWDRAIN then set B_NOWDRAIN (e.g. nfs-backed VN
1050 * filesystem). XXX also use for loopback NFS mounts.
1052 if (ioflag & IO_NOWDRAIN)
1053 bp->b_flags |= B_NOWDRAIN;
1056 * If the lease is non-cachable or IO_SYNC do bwrite().
1058 * IO_INVAL appears to be unused. The idea appears to be
1059 * to turn off caching in this case. Very odd. XXX
1061 if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) {
1062 if (ioflag & IO_INVAL)
1063 bp->b_flags |= B_NOCACHE;
1064 error = VOP_BWRITE(bp->b_vp, bp);
1067 if (np->n_flag & NQNFSNONCACHE) {
1068 error = nfs_vinvalbuf(vp, V_SAVE, td, 1);
1072 } else if ((n + on) == biosize &&
1073 (nmp->nm_flag & NFSMNT_NQNFS) == 0) {
1074 bp->b_flags |= B_ASYNC;
1075 (void)nfs_writebp(bp, 0, 0);
1079 } while (uio->uio_resid > 0 && n > 0);
1082 nfs_rsunlock(np, td);
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, daddr_t bn, int size, struct thread *td)
1107 struct nfsmount *nmp;
1112 if (nmp->nm_flag & NFSMNT_INT) {
1113 bp = getblk(vp, bn, size, PCATCH, 0);
1114 while (bp == (struct buf *)0) {
1115 if (nfs_sigintr(nmp, (struct nfsreq *)0, td))
1116 return ((struct buf *)0);
1117 bp = getblk(vp, bn, size, 0, 2 * hz);
1120 bp = getblk(vp, bn, size, 0, 0);
1123 if (vp->v_type == VREG) {
1126 biosize = mp->mnt_stat.f_iosize;
1127 bp->b_blkno = bn * (biosize / DEV_BSIZE);
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,
1138 struct thread *td, int intrflg)
1140 struct nfsnode *np = VTONFS(vp);
1141 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1142 int error = 0, slpflag, slptimeo;
1144 if (vp->v_flag & VXLOCK) {
1148 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1158 * First wait for any other process doing a flush to complete.
1160 while (np->n_flag & NFLUSHINPROG) {
1161 np->n_flag |= NFLUSHWANT;
1162 error = tsleep((caddr_t)&np->n_flag, 0, "nfsvinval", slptimeo);
1163 if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, td))
1168 * Now, flush as required.
1170 np->n_flag |= NFLUSHINPROG;
1171 error = vinvalbuf(vp, flags, td, slpflag, 0);
1173 if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, td)) {
1174 np->n_flag &= ~NFLUSHINPROG;
1175 if (np->n_flag & NFLUSHWANT) {
1176 np->n_flag &= ~NFLUSHWANT;
1177 wakeup((caddr_t)&np->n_flag);
1181 error = vinvalbuf(vp, flags, td, 0, slptimeo);
1183 np->n_flag &= ~(NMODIFIED | NFLUSHINPROG);
1184 if (np->n_flag & NFLUSHWANT) {
1185 np->n_flag &= ~NFLUSHWANT;
1186 wakeup((caddr_t)&np->n_flag);
1192 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1193 * This is mainly to avoid queueing async I/O requests when the nfsiods
1194 * are all hung on a dead server.
1196 * Note: nfs_asyncio() does not clear (B_ERROR|B_INVAL) but when the bp
1197 * is eventually dequeued by the async daemon, nfs_doio() *will*.
1200 nfs_asyncio(struct buf *bp, struct thread *td)
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 nmp = VFSTONFS(bp->b_vp->v_mount);
1219 * Commits are usually short and sweet so lets save some cpu and
1220 * leave the async daemons for more important rpc's (such as reads
1223 if ((bp->b_flags & (B_READ|B_NEEDCOMMIT)) == B_NEEDCOMMIT &&
1224 (nmp->nm_bufqiods > nfs_numasync / 2)) {
1229 if (nmp->nm_flag & NFSMNT_INT)
1234 * Find a free iod to process this request.
1236 for (i = 0; i < NFS_MAXASYNCDAEMON; i++)
1237 if (nfs_iodwant[i]) {
1239 * Found one, so wake it up and tell it which
1243 ("nfs_asyncio: waking iod %d for mount %p\n",
1245 nfs_iodwant[i] = NULL;
1246 nfs_iodmount[i] = nmp;
1248 wakeup((caddr_t)&nfs_iodwant[i]);
1254 * If none are free, we may already have an iod working on this mount
1255 * point. If so, it will process our request.
1258 if (nmp->nm_bufqiods > 0) {
1260 ("nfs_asyncio: %d iods are already processing mount %p\n",
1261 nmp->nm_bufqiods, nmp));
1267 * If we have an iod which can process the request, then queue
1272 * Ensure that the queue never grows too large. We still want
1273 * to asynchronize so we block rather then return EIO.
1275 while (nmp->nm_bufqlen >= 2*nfs_numasync) {
1277 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp));
1278 nmp->nm_bufqwant = TRUE;
1279 error = tsleep(&nmp->nm_bufq, slpflag,
1280 "nfsaio", slptimeo);
1282 if (nfs_sigintr(nmp, NULL, td))
1284 if (slpflag == PCATCH) {
1290 * We might have lost our iod while sleeping,
1291 * so check and loop if nescessary.
1293 if (nmp->nm_bufqiods == 0) {
1295 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1300 if ((bp->b_flags & B_READ) == 0)
1301 bp->b_flags |= B_WRITEINPROG;
1304 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1310 * All the iods are busy on other mounts, so return EIO to
1311 * force the caller to process the i/o synchronously.
1313 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n"));
1318 * Do an I/O operation to/from a cache block. This may be called
1319 * synchronously or from an nfsiod.
1321 * NOTE! TD MIGHT BE NULL
1324 nfs_doio(struct buf *bp, struct thread *td)
1329 struct nfsmount *nmp;
1330 int error = 0, iomode, must_commit = 0;
1336 nmp = VFSTONFS(vp->v_mount);
1338 uiop->uio_iov = &io;
1339 uiop->uio_iovcnt = 1;
1340 uiop->uio_segflg = UIO_SYSSPACE;
1344 * clear B_ERROR and B_INVAL state prior to initiating the I/O. We
1345 * do this here so we do not have to do it in all the code that
1348 bp->b_flags &= ~(B_ERROR | B_INVAL);
1350 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp));
1353 * Historically, paging was done with physio, but no more.
1355 if (bp->b_flags & B_PHYS) {
1357 * ...though reading /dev/drum still gets us here.
1359 io.iov_len = uiop->uio_resid = bp->b_bcount;
1360 /* mapping was done by vmapbuf() */
1361 io.iov_base = bp->b_data;
1362 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1363 if (bp->b_flags & B_READ) {
1364 uiop->uio_rw = UIO_READ;
1365 nfsstats.read_physios++;
1366 error = nfs_readrpc(vp, uiop);
1370 iomode = NFSV3WRITE_DATASYNC;
1371 uiop->uio_rw = UIO_WRITE;
1372 nfsstats.write_physios++;
1373 error = nfs_writerpc(vp, uiop, &iomode, &com);
1376 bp->b_flags |= B_ERROR;
1377 bp->b_error = error;
1379 } else if (bp->b_flags & B_READ) {
1380 io.iov_len = uiop->uio_resid = bp->b_bcount;
1381 io.iov_base = bp->b_data;
1382 uiop->uio_rw = UIO_READ;
1384 switch (vp->v_type) {
1386 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1387 nfsstats.read_bios++;
1388 error = nfs_readrpc(vp, uiop);
1391 if (uiop->uio_resid) {
1393 * If we had a short read with no error, we must have
1394 * hit a file hole. We should zero-fill the remainder.
1395 * This can also occur if the server hits the file EOF.
1397 * Holes used to be able to occur due to pending
1398 * writes, but that is not possible any longer.
1400 int nread = bp->b_bcount - uiop->uio_resid;
1401 int left = uiop->uio_resid;
1404 bzero((char *)bp->b_data + nread, left);
1405 uiop->uio_resid = 0;
1408 if (td && td->td_proc && (vp->v_flag & VTEXT) &&
1409 (((nmp->nm_flag & NFSMNT_NQNFS) &&
1410 NQNFS_CKINVALID(vp, np, ND_READ) &&
1411 np->n_lrev != np->n_brev) ||
1412 (!(nmp->nm_flag & NFSMNT_NQNFS) &&
1413 np->n_mtime != np->n_vattr.va_mtime.tv_sec))) {
1414 uprintf("Process killed due to text file modification\n");
1415 psignal(td->td_proc, SIGKILL);
1420 uiop->uio_offset = (off_t)0;
1421 nfsstats.readlink_bios++;
1422 error = nfs_readlinkrpc(vp, uiop);
1425 nfsstats.readdir_bios++;
1426 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1427 if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
1428 error = nfs_readdirplusrpc(vp, uiop);
1429 if (error == NFSERR_NOTSUPP)
1430 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1432 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1433 error = nfs_readdirrpc(vp, uiop);
1435 * end-of-directory sets B_INVAL but does not generate an
1438 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1439 bp->b_flags |= B_INVAL;
1442 printf("nfs_doio: type %x unexpected\n",vp->v_type);
1446 bp->b_flags |= B_ERROR;
1447 bp->b_error = error;
1451 * If we only need to commit, try to commit
1453 if (bp->b_flags & B_NEEDCOMMIT) {
1457 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1458 bp->b_flags |= B_WRITEINPROG;
1459 retv = nfs_commit(bp->b_vp, off,
1460 bp->b_dirtyend - bp->b_dirtyoff, td);
1461 bp->b_flags &= ~B_WRITEINPROG;
1463 bp->b_dirtyoff = bp->b_dirtyend = 0;
1464 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1469 if (retv == NFSERR_STALEWRITEVERF) {
1470 nfs_clearcommit(bp->b_vp->v_mount);
1475 * Setup for actual write
1478 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1479 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1481 if (bp->b_dirtyend > bp->b_dirtyoff) {
1482 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1484 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1486 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1487 uiop->uio_rw = UIO_WRITE;
1488 nfsstats.write_bios++;
1490 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1491 iomode = NFSV3WRITE_UNSTABLE;
1493 iomode = NFSV3WRITE_FILESYNC;
1495 bp->b_flags |= B_WRITEINPROG;
1496 error = nfs_writerpc(vp, uiop, &iomode, &must_commit);
1499 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1500 * to cluster the buffers needing commit. This will allow
1501 * the system to submit a single commit rpc for the whole
1502 * cluster. We can do this even if the buffer is not 100%
1503 * dirty (relative to the NFS blocksize), so we optimize the
1504 * append-to-file-case.
1506 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1507 * cleared because write clustering only works for commit
1508 * rpc's, not for the data portion of the write).
1511 if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1512 bp->b_flags |= B_NEEDCOMMIT;
1513 if (bp->b_dirtyoff == 0
1514 && bp->b_dirtyend == bp->b_bcount)
1515 bp->b_flags |= B_CLUSTEROK;
1517 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1519 bp->b_flags &= ~B_WRITEINPROG;
1522 * For an interrupted write, the buffer is still valid
1523 * and the write hasn't been pushed to the server yet,
1524 * so we can't set B_ERROR and report the interruption
1525 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1526 * is not relevant, so the rpc attempt is essentially
1527 * a noop. For the case of a V3 write rpc not being
1528 * committed to stable storage, the block is still
1529 * dirty and requires either a commit rpc or another
1530 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1531 * the block is reused. This is indicated by setting
1532 * the B_DELWRI and B_NEEDCOMMIT flags.
1534 * If the buffer is marked B_PAGING, it does not reside on
1535 * the vp's paging queues so we cannot call bdirty(). The
1536 * bp in this case is not an NFS cache block so we should
1540 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1544 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1545 if ((bp->b_flags & B_PAGING) == 0) {
1547 bp->b_flags &= ~B_DONE;
1549 if (error && (bp->b_flags & B_ASYNC) == 0)
1550 bp->b_flags |= B_EINTR;
1554 bp->b_flags |= B_ERROR;
1555 bp->b_error = np->n_error = error;
1556 np->n_flag |= NWRITEERR;
1558 bp->b_dirtyoff = bp->b_dirtyend = 0;
1566 bp->b_resid = uiop->uio_resid;
1568 nfs_clearcommit(vp->v_mount);
1574 * Used to aid in handling ftruncate() operations on the NFS client side.
1575 * Truncation creates a number of special problems for NFS. We have to
1576 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1577 * we have to properly handle VM pages or (potentially dirty) buffers
1578 * that straddle the truncation point.
1582 nfs_meta_setsize(struct vnode *vp, struct thread *td, u_quad_t nsize)
1584 struct nfsnode *np = VTONFS(vp);
1585 u_quad_t tsize = np->n_size;
1586 int biosize = vp->v_mount->mnt_stat.f_iosize;
1591 if (np->n_size < tsize) {
1597 * vtruncbuf() doesn't get the buffer overlapping the
1598 * truncation point. We may have a B_DELWRI and/or B_CACHE
1599 * buffer that now needs to be truncated.
1601 error = vtruncbuf(vp, td, nsize, biosize);
1602 lbn = nsize / biosize;
1603 bufsize = nsize & (biosize - 1);
1604 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1605 if (bp->b_dirtyoff > bp->b_bcount)
1606 bp->b_dirtyoff = bp->b_bcount;
1607 if (bp->b_dirtyend > bp->b_bcount)
1608 bp->b_dirtyend = bp->b_bcount;
1609 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1612 vnode_pager_setsize(vp, nsize);