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 $
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/resourcevar.h>
44 #include <sys/signalvar.h>
47 #include <sys/vnode.h>
48 #include <sys/mount.h>
49 #include <sys/kernel.h>
52 #include <vm/vm_extern.h>
53 #include <vm/vm_page.h>
54 #include <vm/vm_object.h>
55 #include <vm/vm_pager.h>
56 #include <vm/vnode_pager.h>
58 #include <nfs/rpcv2.h>
59 #include <nfs/nfsproto.h>
61 #include <nfs/nfsmount.h>
62 #include <nfs/nqnfs.h>
63 #include <nfs/nfsnode.h>
65 static struct buf *nfs_getcacheblk __P((struct vnode *vp, daddr_t bn, int size,
68 extern int nfs_numasync;
69 extern int nfs_pbuf_freecnt;
70 extern struct nfsstats nfsstats;
73 * Vnode op for VM getpages.
77 struct vop_getpages_args /* {
82 vm_ooffset_t a_offset;
85 int i, error, nextoff, size, toff, count, npages;
97 p = curproc; /* XXX */
98 cred = curproc->p_ucred; /* XXX */
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, cred, p);
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, cred);
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;
249 int iomode, must_commit, i, error, npages, count;
255 struct nfsmount *nmp;
261 p = curproc; /* XXX */
262 cred = curproc->p_ucred; /* XXX */
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, cred, p);
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 * We use only the kva address for the buffer, but this is extremely
290 * convienient and fast.
292 bp = getpbuf(&nfs_pbuf_freecnt);
294 kva = (vm_offset_t) bp->b_data;
295 pmap_qenter(kva, pages, npages);
297 iov.iov_base = (caddr_t) kva;
301 uio.uio_offset = offset;
302 uio.uio_resid = count;
303 uio.uio_segflg = UIO_SYSSPACE;
304 uio.uio_rw = UIO_WRITE;
307 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
308 iomode = NFSV3WRITE_UNSTABLE;
310 iomode = NFSV3WRITE_FILESYNC;
312 error = nfs_writerpc(vp, &uio, cred, &iomode, &must_commit);
314 pmap_qremove(kva, npages);
315 relpbuf(bp, &nfs_pbuf_freecnt);
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(vp, uio, ioflag, cred)
334 register struct vnode *vp;
335 register struct uio *uio;
339 register struct nfsnode *np = VTONFS(vp);
340 register int biosize, i;
341 struct buf *bp = 0, *rabp;
344 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
348 int nra, error = 0, n = 0, on = 0;
351 if (uio->uio_rw != UIO_READ)
352 panic("nfs_read mode");
354 if (uio->uio_resid == 0)
356 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
360 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
361 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
362 (void)nfs_fsinfo(nmp, vp, cred, p);
363 if (vp->v_type != VDIR &&
364 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
366 biosize = vp->v_mount->mnt_stat.f_iosize;
367 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
369 * For nfs, cache consistency can only be maintained approximately.
370 * Although RFC1094 does not specify the criteria, the following is
371 * believed to be compatible with the reference port.
372 * For nqnfs, full cache consistency is maintained within the loop.
374 * If the file's modify time on the server has changed since the
375 * last read rpc or you have written to the file,
376 * you may have lost data cache consistency with the
377 * server, so flush all of the file's data out of the cache.
378 * Then force a getattr rpc to ensure that you have up to date
380 * NB: This implies that cache data can be read when up to
381 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
382 * attributes this could be forced by setting n_attrstamp to 0 before
383 * the VOP_GETATTR() call.
385 if ((nmp->nm_flag & NFSMNT_NQNFS) == 0) {
386 if (np->n_flag & NMODIFIED) {
387 if (vp->v_type != VREG) {
388 if (vp->v_type != VDIR)
389 panic("nfs: bioread, not dir");
391 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
396 error = VOP_GETATTR(vp, &vattr, cred, p);
399 np->n_mtime = vattr.va_mtime.tv_sec;
401 error = VOP_GETATTR(vp, &vattr, cred, p);
404 if (np->n_mtime != vattr.va_mtime.tv_sec) {
405 if (vp->v_type == VDIR)
407 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
410 np->n_mtime = vattr.va_mtime.tv_sec;
417 * Get a valid lease. If cached data is stale, flush it.
419 if (nmp->nm_flag & NFSMNT_NQNFS) {
420 if (NQNFS_CKINVALID(vp, np, ND_READ)) {
422 error = nqnfs_getlease(vp, ND_READ, cred, p);
423 } while (error == NQNFS_EXPIRED);
426 if (np->n_lrev != np->n_brev ||
427 (np->n_flag & NQNFSNONCACHE) ||
428 ((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) {
429 if (vp->v_type == VDIR)
431 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
434 np->n_brev = np->n_lrev;
436 } else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) {
438 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
443 if (np->n_flag & NQNFSNONCACHE) {
444 switch (vp->v_type) {
446 return (nfs_readrpc(vp, uio, cred));
448 return (nfs_readlinkrpc(vp, uio, cred));
452 printf(" NQNFSNONCACHE: type %x unexpected\n",
456 switch (vp->v_type) {
458 nfsstats.biocache_reads++;
459 lbn = uio->uio_offset / biosize;
460 on = uio->uio_offset & (biosize - 1);
463 * Start the read ahead(s), as required.
465 if (nfs_numasync > 0 && nmp->nm_readahead > 0) {
466 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
467 (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) {
468 rabn = lbn + 1 + nra;
469 if (!incore(vp, rabn)) {
470 rabp = nfs_getcacheblk(vp, rabn, biosize, p);
473 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
474 rabp->b_flags |= (B_READ | B_ASYNC);
475 vfs_busy_pages(rabp, 0);
476 if (nfs_asyncio(rabp, cred, p)) {
477 rabp->b_flags |= B_INVAL|B_ERROR;
478 vfs_unbusy_pages(rabp);
490 * Obtain the buffer cache block. Figure out the buffer size
491 * when we are at EOF. If we are modifying the size of the
492 * buffer based on an EOF condition we need to hold
493 * nfs_rslock() through obtaining the buffer to prevent
494 * a potential writer-appender from messing with n_size.
495 * Otherwise we may accidently truncate the buffer and
498 * Note that bcount is *not* DEV_BSIZE aligned.
503 if ((off_t)lbn * biosize >= np->n_size) {
505 } else if ((off_t)(lbn + 1) * biosize > np->n_size) {
506 bcount = np->n_size - (off_t)lbn * biosize;
508 if (bcount != biosize) {
509 switch(nfs_rslock(np, p)) {
522 bp = nfs_getcacheblk(vp, lbn, bcount, p);
524 if (bcount != biosize)
530 * If B_CACHE is not set, we must issue the read. If this
531 * fails, we return an error.
534 if ((bp->b_flags & B_CACHE) == 0) {
535 bp->b_flags |= B_READ;
536 vfs_busy_pages(bp, 0);
537 error = nfs_doio(bp, cred, p);
545 * on is the offset into the current bp. Figure out how many
546 * bytes we can copy out of the bp. Note that bcount is
547 * NOT DEV_BSIZE aligned.
549 * Then figure out how many bytes we can copy into the uio.
554 n = min((unsigned)(bcount - on), uio->uio_resid);
557 nfsstats.biocache_readlinks++;
558 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, p);
561 if ((bp->b_flags & B_CACHE) == 0) {
562 bp->b_flags |= B_READ;
563 vfs_busy_pages(bp, 0);
564 error = nfs_doio(bp, cred, p);
566 bp->b_flags |= B_ERROR;
571 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
575 nfsstats.biocache_readdirs++;
576 if (np->n_direofoffset
577 && uio->uio_offset >= np->n_direofoffset) {
580 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
581 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
582 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, p);
585 if ((bp->b_flags & B_CACHE) == 0) {
586 bp->b_flags |= B_READ;
587 vfs_busy_pages(bp, 0);
588 error = nfs_doio(bp, cred, p);
592 while (error == NFSERR_BAD_COOKIE) {
593 printf("got bad cookie vp %p bp %p\n", vp, bp);
595 error = nfs_vinvalbuf(vp, 0, cred, p, 1);
597 * Yuck! The directory has been modified on the
598 * server. The only way to get the block is by
599 * reading from the beginning to get all the
602 * Leave the last bp intact unless there is an error.
603 * Loop back up to the while if the error is another
604 * NFSERR_BAD_COOKIE (double yuch!).
606 for (i = 0; i <= lbn && !error; i++) {
607 if (np->n_direofoffset
608 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
610 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, p);
613 if ((bp->b_flags & B_CACHE) == 0) {
614 bp->b_flags |= B_READ;
615 vfs_busy_pages(bp, 0);
616 error = nfs_doio(bp, cred, p);
618 * no error + B_INVAL == directory EOF,
621 if (error == 0 && (bp->b_flags & B_INVAL))
625 * An error will throw away the block and the
626 * for loop will break out. If no error and this
627 * is not the block we want, we throw away the
628 * block and go for the next one via the for loop.
630 if (error || i < lbn)
635 * The above while is repeated if we hit another cookie
636 * error. If we hit an error and it wasn't a cookie error,
644 * If not eof and read aheads are enabled, start one.
645 * (You need the current block first, so that you have the
646 * directory offset cookie of the next block.)
648 if (nfs_numasync > 0 && nmp->nm_readahead > 0 &&
649 (bp->b_flags & B_INVAL) == 0 &&
650 (np->n_direofoffset == 0 ||
651 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
652 !(np->n_flag & NQNFSNONCACHE) &&
653 !incore(vp, lbn + 1)) {
654 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, p);
656 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
657 rabp->b_flags |= (B_READ | B_ASYNC);
658 vfs_busy_pages(rabp, 0);
659 if (nfs_asyncio(rabp, cred, p)) {
660 rabp->b_flags |= B_INVAL|B_ERROR;
661 vfs_unbusy_pages(rabp);
670 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
671 * chopped for the EOF condition, we cannot tell how large
672 * NFS directories are going to be until we hit EOF. So
673 * an NFS directory buffer is *not* chopped to its EOF. Now,
674 * it just so happens that b_resid will effectively chop it
675 * to EOF. *BUT* this information is lost if the buffer goes
676 * away and is reconstituted into a B_CACHE state ( due to
677 * being VMIO ) later. So we keep track of the directory eof
678 * in np->n_direofoffset and chop it off as an extra step
681 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
682 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
683 n = np->n_direofoffset - uio->uio_offset;
686 printf(" nfs_bioread: type %x unexpected\n",vp->v_type);
691 error = uiomove(bp->b_data + on, (int)n, uio);
693 switch (vp->v_type) {
701 * Invalidate buffer if caching is disabled, forcing a
702 * re-read from the remote later.
704 if (np->n_flag & NQNFSNONCACHE)
705 bp->b_flags |= B_INVAL;
708 printf(" nfs_bioread: type %x unexpected\n",vp->v_type);
711 } while (error == 0 && uio->uio_resid > 0 && n > 0);
716 * Vnode op for write using bio
720 struct vop_write_args /* {
724 struct ucred *a_cred;
728 struct uio *uio = ap->a_uio;
729 struct proc *p = uio->uio_procp;
730 struct vnode *vp = ap->a_vp;
731 struct nfsnode *np = VTONFS(vp);
732 struct ucred *cred = ap->a_cred;
733 int ioflag = ap->a_ioflag;
736 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
739 int n, on, error = 0, iomode, must_commit;
743 if (uio->uio_rw != UIO_WRITE)
744 panic("nfs_write mode");
745 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_procp != curproc)
746 panic("nfs_write proc");
748 if (vp->v_type != VREG)
750 if (np->n_flag & NWRITEERR) {
751 np->n_flag &= ~NWRITEERR;
752 return (np->n_error);
754 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
755 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
756 (void)nfs_fsinfo(nmp, vp, cred, p);
759 * Synchronously flush pending buffers if we are in synchronous
760 * mode or if we are appending.
762 if (ioflag & (IO_APPEND | IO_SYNC)) {
763 if (np->n_flag & NMODIFIED) {
765 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
772 * If IO_APPEND then load uio_offset. We restart here if we cannot
773 * get the append lock.
776 if (ioflag & IO_APPEND) {
778 error = VOP_GETATTR(vp, &vattr, cred, p);
781 uio->uio_offset = np->n_size;
784 if (uio->uio_offset < 0)
786 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
788 if (uio->uio_resid == 0)
792 * We need to obtain the rslock if we intend to modify np->n_size
793 * in order to guarentee the append point with multiple contending
794 * writers, to guarentee that no other appenders modify n_size
795 * while we are trying to obtain a truncated buffer (i.e. to avoid
796 * accidently truncating data written by another appender due to
797 * the race), and to ensure that the buffer is populated prior to
798 * our extending of the file. We hold rslock through the entire
801 * Note that we do not synchronize the case where someone truncates
802 * the file while we are appending to it because attempting to lock
803 * this case may deadlock other parts of the system unexpectedly.
805 if ((ioflag & IO_APPEND) ||
806 uio->uio_offset + uio->uio_resid > np->n_size) {
807 switch(nfs_rslock(np, p)) {
822 * Maybe this should be above the vnode op call, but so long as
823 * file servers have no limits, i don't think it matters
825 if (p && uio->uio_offset + uio->uio_resid >
826 p->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
833 biosize = vp->v_mount->mnt_stat.f_iosize;
837 * Check for a valid write lease.
839 if ((nmp->nm_flag & NFSMNT_NQNFS) &&
840 NQNFS_CKINVALID(vp, np, ND_WRITE)) {
842 error = nqnfs_getlease(vp, ND_WRITE, cred, p);
843 } while (error == NQNFS_EXPIRED);
846 if (np->n_lrev != np->n_brev ||
847 (np->n_flag & NQNFSNONCACHE)) {
848 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
851 np->n_brev = np->n_lrev;
854 if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) {
855 iomode = NFSV3WRITE_FILESYNC;
856 error = nfs_writerpc(vp, uio, cred, &iomode, &must_commit);
858 nfs_clearcommit(vp->v_mount);
861 nfsstats.biocache_writes++;
862 lbn = uio->uio_offset / biosize;
863 on = uio->uio_offset & (biosize-1);
864 n = min((unsigned)(biosize - on), uio->uio_resid);
867 * Handle direct append and file extension cases, calculate
868 * unaligned buffer size.
871 if (uio->uio_offset == np->n_size && n) {
873 * Get the buffer (in its pre-append state to maintain
874 * B_CACHE if it was previously set). Resize the
875 * nfsnode after we have locked the buffer to prevent
876 * readers from reading garbage.
879 bp = nfs_getcacheblk(vp, lbn, bcount, p);
884 np->n_size = uio->uio_offset + n;
885 np->n_flag |= NMODIFIED;
886 vnode_pager_setsize(vp, np->n_size);
888 save = bp->b_flags & B_CACHE;
890 allocbuf(bp, bcount);
895 * Obtain the locked cache block first, and then
896 * adjust the file's size as appropriate.
899 if ((off_t)lbn * biosize + bcount < np->n_size) {
900 if ((off_t)(lbn + 1) * biosize < np->n_size)
903 bcount = np->n_size - (off_t)lbn * biosize;
905 bp = nfs_getcacheblk(vp, lbn, bcount, p);
906 if (uio->uio_offset + n > np->n_size) {
907 np->n_size = uio->uio_offset + n;
908 np->n_flag |= NMODIFIED;
909 vnode_pager_setsize(vp, np->n_size);
919 * Issue a READ if B_CACHE is not set. In special-append
920 * mode, B_CACHE is based on the buffer prior to the write
921 * op and is typically set, avoiding the read. If a read
922 * is required in special append mode, the server will
923 * probably send us a short-read since we extended the file
924 * on our end, resulting in b_resid == 0 and, thusly,
925 * B_CACHE getting set.
927 * We can also avoid issuing the read if the write covers
928 * the entire buffer. We have to make sure the buffer state
929 * is reasonable in this case since we will not be initiating
930 * I/O. See the comments in kern/vfs_bio.c's getblk() for
933 * B_CACHE may also be set due to the buffer being cached
937 if (on == 0 && n == bcount) {
938 bp->b_flags |= B_CACHE;
939 bp->b_flags &= ~(B_ERROR | B_INVAL);
942 if ((bp->b_flags & B_CACHE) == 0) {
943 bp->b_flags |= B_READ;
944 vfs_busy_pages(bp, 0);
945 error = nfs_doio(bp, cred, p);
955 if (bp->b_wcred == NOCRED) {
959 np->n_flag |= NMODIFIED;
962 * If dirtyend exceeds file size, chop it down. This should
963 * not normally occur but there is an append race where it
964 * might occur XXX, so we log it.
966 * If the chopping creates a reverse-indexed or degenerate
967 * situation with dirtyoff/end, we 0 both of them.
970 if (bp->b_dirtyend > bcount) {
971 printf("NFS append race @%lx:%d\n",
972 (long)bp->b_blkno * DEV_BSIZE,
973 bp->b_dirtyend - bcount);
974 bp->b_dirtyend = bcount;
977 if (bp->b_dirtyoff >= bp->b_dirtyend)
978 bp->b_dirtyoff = bp->b_dirtyend = 0;
981 * If the new write will leave a contiguous dirty
982 * area, just update the b_dirtyoff and b_dirtyend,
983 * otherwise force a write rpc of the old dirty area.
985 * While it is possible to merge discontiguous writes due to
986 * our having a B_CACHE buffer ( and thus valid read data
987 * for the hole), we don't because it could lead to
988 * significant cache coherency problems with multiple clients,
989 * especially if locking is implemented later on.
991 * as an optimization we could theoretically maintain
992 * a linked list of discontinuous areas, but we would still
993 * have to commit them separately so there isn't much
994 * advantage to it except perhaps a bit of asynchronization.
997 if (bp->b_dirtyend > 0 &&
998 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
999 if (VOP_BWRITE(bp->b_vp, bp) == EINTR) {
1007 * Check for valid write lease and get one as required.
1008 * In case getblk() and/or bwrite() delayed us.
1010 if ((nmp->nm_flag & NFSMNT_NQNFS) &&
1011 NQNFS_CKINVALID(vp, np, ND_WRITE)) {
1013 error = nqnfs_getlease(vp, ND_WRITE, cred, p);
1014 } while (error == NQNFS_EXPIRED);
1019 if (np->n_lrev != np->n_brev ||
1020 (np->n_flag & NQNFSNONCACHE)) {
1022 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
1025 np->n_brev = np->n_lrev;
1030 error = uiomove((char *)bp->b_data + on, n, uio);
1033 * Since this block is being modified, it must be written
1034 * again and not just committed. Since write clustering does
1035 * not work for the stage 1 data write, only the stage 2
1036 * commit rpc, we have to clear B_CLUSTEROK as well.
1038 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1041 bp->b_flags |= B_ERROR;
1047 * Only update dirtyoff/dirtyend if not a degenerate
1051 if (bp->b_dirtyend > 0) {
1052 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1053 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1055 bp->b_dirtyoff = on;
1056 bp->b_dirtyend = on + n;
1058 vfs_bio_set_validclean(bp, on, n);
1061 * If IO_NOWDRAIN then set B_NOWDRAIN (e.g. nfs-backed VN
1062 * filesystem). XXX also use for loopback NFS mounts.
1064 if (ioflag & IO_NOWDRAIN)
1065 bp->b_flags |= B_NOWDRAIN;
1068 * If the lease is non-cachable or IO_SYNC do bwrite().
1070 * IO_INVAL appears to be unused. The idea appears to be
1071 * to turn off caching in this case. Very odd. XXX
1073 if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) {
1074 if (ioflag & IO_INVAL)
1075 bp->b_flags |= B_NOCACHE;
1076 error = VOP_BWRITE(bp->b_vp, bp);
1079 if (np->n_flag & NQNFSNONCACHE) {
1080 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
1084 } else if ((n + on) == biosize &&
1085 (nmp->nm_flag & NFSMNT_NQNFS) == 0) {
1086 bp->b_flags |= B_ASYNC;
1087 (void)nfs_writebp(bp, 0, 0);
1091 } while (uio->uio_resid > 0 && n > 0);
1094 nfs_rsunlock(np, p);
1100 * Get an nfs cache block.
1102 * Allocate a new one if the block isn't currently in the cache
1103 * and return the block marked busy. If the calling process is
1104 * interrupted by a signal for an interruptible mount point, return
1107 * The caller must carefully deal with the possible B_INVAL state of
1108 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it
1109 * indirectly), so synchronous reads can be issued without worrying about
1110 * the B_INVAL state. We have to be a little more careful when dealing
1111 * with writes (see comments in nfs_write()) when extending a file past
1115 nfs_getcacheblk(vp, bn, size, p)
1121 register struct buf *bp;
1123 struct nfsmount *nmp;
1128 if (nmp->nm_flag & NFSMNT_INT) {
1129 bp = getblk(vp, bn, size, PCATCH, 0);
1130 while (bp == (struct buf *)0) {
1131 if (nfs_sigintr(nmp, (struct nfsreq *)0, p))
1132 return ((struct buf *)0);
1133 bp = getblk(vp, bn, size, 0, 2 * hz);
1136 bp = getblk(vp, bn, size, 0, 0);
1139 if (vp->v_type == VREG) {
1142 biosize = mp->mnt_stat.f_iosize;
1143 bp->b_blkno = bn * (biosize / DEV_BSIZE);
1149 * Flush and invalidate all dirty buffers. If another process is already
1150 * doing the flush, just wait for completion.
1153 nfs_vinvalbuf(vp, flags, cred, p, intrflg)
1160 register struct nfsnode *np = VTONFS(vp);
1161 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1162 int error = 0, slpflag, slptimeo;
1164 if (vp->v_flag & VXLOCK) {
1168 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1178 * First wait for any other process doing a flush to complete.
1180 while (np->n_flag & NFLUSHINPROG) {
1181 np->n_flag |= NFLUSHWANT;
1182 error = tsleep((caddr_t)&np->n_flag, PRIBIO + 2, "nfsvinval",
1184 if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p))
1189 * Now, flush as required.
1191 np->n_flag |= NFLUSHINPROG;
1192 error = vinvalbuf(vp, flags, cred, p, slpflag, 0);
1194 if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p)) {
1195 np->n_flag &= ~NFLUSHINPROG;
1196 if (np->n_flag & NFLUSHWANT) {
1197 np->n_flag &= ~NFLUSHWANT;
1198 wakeup((caddr_t)&np->n_flag);
1202 error = vinvalbuf(vp, flags, cred, p, 0, slptimeo);
1204 np->n_flag &= ~(NMODIFIED | NFLUSHINPROG);
1205 if (np->n_flag & NFLUSHWANT) {
1206 np->n_flag &= ~NFLUSHWANT;
1207 wakeup((caddr_t)&np->n_flag);
1213 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1214 * This is mainly to avoid queueing async I/O requests when the nfsiods
1215 * are all hung on a dead server.
1217 * Note: nfs_asyncio() does not clear (B_ERROR|B_INVAL) but when the bp
1218 * is eventually dequeued by the async daemon, nfs_doio() *will*.
1221 nfs_asyncio(bp, cred, procp)
1222 register struct buf *bp;
1226 struct nfsmount *nmp;
1234 * If no async daemons then return EIO to force caller to run the rpc
1237 if (nfs_numasync == 0)
1240 nmp = VFSTONFS(bp->b_vp->v_mount);
1243 * Commits are usually short and sweet so lets save some cpu and
1244 * leave the async daemons for more important rpc's (such as reads
1247 if ((bp->b_flags & (B_READ|B_NEEDCOMMIT)) == B_NEEDCOMMIT &&
1248 (nmp->nm_bufqiods > nfs_numasync / 2)) {
1253 if (nmp->nm_flag & NFSMNT_INT)
1258 * Find a free iod to process this request.
1260 for (i = 0; i < NFS_MAXASYNCDAEMON; i++)
1261 if (nfs_iodwant[i]) {
1263 * Found one, so wake it up and tell it which
1267 ("nfs_asyncio: waking iod %d for mount %p\n",
1269 nfs_iodwant[i] = (struct proc *)0;
1270 nfs_iodmount[i] = nmp;
1272 wakeup((caddr_t)&nfs_iodwant[i]);
1278 * If none are free, we may already have an iod working on this mount
1279 * point. If so, it will process our request.
1282 if (nmp->nm_bufqiods > 0) {
1284 ("nfs_asyncio: %d iods are already processing mount %p\n",
1285 nmp->nm_bufqiods, nmp));
1291 * If we have an iod which can process the request, then queue
1296 * Ensure that the queue never grows too large. We still want
1297 * to asynchronize so we block rather then return EIO.
1299 while (nmp->nm_bufqlen >= 2*nfs_numasync) {
1301 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp));
1302 nmp->nm_bufqwant = TRUE;
1303 error = tsleep(&nmp->nm_bufq, slpflag | PRIBIO,
1304 "nfsaio", slptimeo);
1306 if (nfs_sigintr(nmp, NULL, procp))
1308 if (slpflag == PCATCH) {
1314 * We might have lost our iod while sleeping,
1315 * so check and loop if nescessary.
1317 if (nmp->nm_bufqiods == 0) {
1319 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1324 if (bp->b_flags & B_READ) {
1325 if (bp->b_rcred == NOCRED && cred != NOCRED) {
1330 bp->b_flags |= B_WRITEINPROG;
1331 if (bp->b_wcred == NOCRED && cred != NOCRED) {
1338 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1344 * All the iods are busy on other mounts, so return EIO to
1345 * force the caller to process the i/o synchronously.
1347 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n"));
1352 * Do an I/O operation to/from a cache block. This may be called
1353 * synchronously or from an nfsiod.
1364 struct nfsmount *nmp;
1365 int error = 0, iomode, must_commit = 0;
1371 nmp = VFSTONFS(vp->v_mount);
1373 uiop->uio_iov = &io;
1374 uiop->uio_iovcnt = 1;
1375 uiop->uio_segflg = UIO_SYSSPACE;
1376 uiop->uio_procp = p;
1379 * clear B_ERROR and B_INVAL state prior to initiating the I/O. We
1380 * do this here so we do not have to do it in all the code that
1383 bp->b_flags &= ~(B_ERROR | B_INVAL);
1385 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp));
1388 * Historically, paging was done with physio, but no more.
1390 if (bp->b_flags & B_PHYS) {
1392 * ...though reading /dev/drum still gets us here.
1394 io.iov_len = uiop->uio_resid = bp->b_bcount;
1395 /* mapping was done by vmapbuf() */
1396 io.iov_base = bp->b_data;
1397 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1398 if (bp->b_flags & B_READ) {
1399 uiop->uio_rw = UIO_READ;
1400 nfsstats.read_physios++;
1401 error = nfs_readrpc(vp, uiop, cr);
1405 iomode = NFSV3WRITE_DATASYNC;
1406 uiop->uio_rw = UIO_WRITE;
1407 nfsstats.write_physios++;
1408 error = nfs_writerpc(vp, uiop, cr, &iomode, &com);
1411 bp->b_flags |= B_ERROR;
1412 bp->b_error = error;
1414 } else if (bp->b_flags & B_READ) {
1415 io.iov_len = uiop->uio_resid = bp->b_bcount;
1416 io.iov_base = bp->b_data;
1417 uiop->uio_rw = UIO_READ;
1419 switch (vp->v_type) {
1421 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1422 nfsstats.read_bios++;
1423 error = nfs_readrpc(vp, uiop, cr);
1426 if (uiop->uio_resid) {
1428 * If we had a short read with no error, we must have
1429 * hit a file hole. We should zero-fill the remainder.
1430 * This can also occur if the server hits the file EOF.
1432 * Holes used to be able to occur due to pending
1433 * writes, but that is not possible any longer.
1435 int nread = bp->b_bcount - uiop->uio_resid;
1436 int left = uiop->uio_resid;
1439 bzero((char *)bp->b_data + nread, left);
1440 uiop->uio_resid = 0;
1443 if (p && (vp->v_flag & VTEXT) &&
1444 (((nmp->nm_flag & NFSMNT_NQNFS) &&
1445 NQNFS_CKINVALID(vp, np, ND_READ) &&
1446 np->n_lrev != np->n_brev) ||
1447 (!(nmp->nm_flag & NFSMNT_NQNFS) &&
1448 np->n_mtime != np->n_vattr.va_mtime.tv_sec))) {
1449 uprintf("Process killed due to text file modification\n");
1450 psignal(p, SIGKILL);
1455 uiop->uio_offset = (off_t)0;
1456 nfsstats.readlink_bios++;
1457 error = nfs_readlinkrpc(vp, uiop, cr);
1460 nfsstats.readdir_bios++;
1461 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1462 if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
1463 error = nfs_readdirplusrpc(vp, uiop, cr);
1464 if (error == NFSERR_NOTSUPP)
1465 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1467 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1468 error = nfs_readdirrpc(vp, uiop, cr);
1470 * end-of-directory sets B_INVAL but does not generate an
1473 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1474 bp->b_flags |= B_INVAL;
1477 printf("nfs_doio: type %x unexpected\n",vp->v_type);
1481 bp->b_flags |= B_ERROR;
1482 bp->b_error = error;
1486 * If we only need to commit, try to commit
1488 if (bp->b_flags & B_NEEDCOMMIT) {
1492 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1493 bp->b_flags |= B_WRITEINPROG;
1495 bp->b_vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1497 bp->b_flags &= ~B_WRITEINPROG;
1499 bp->b_dirtyoff = bp->b_dirtyend = 0;
1500 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1505 if (retv == NFSERR_STALEWRITEVERF) {
1506 nfs_clearcommit(bp->b_vp->v_mount);
1511 * Setup for actual write
1514 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1515 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1517 if (bp->b_dirtyend > bp->b_dirtyoff) {
1518 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1520 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1522 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1523 uiop->uio_rw = UIO_WRITE;
1524 nfsstats.write_bios++;
1526 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1527 iomode = NFSV3WRITE_UNSTABLE;
1529 iomode = NFSV3WRITE_FILESYNC;
1531 bp->b_flags |= B_WRITEINPROG;
1532 error = nfs_writerpc(vp, uiop, cr, &iomode, &must_commit);
1535 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1536 * to cluster the buffers needing commit. This will allow
1537 * the system to submit a single commit rpc for the whole
1538 * cluster. We can do this even if the buffer is not 100%
1539 * dirty (relative to the NFS blocksize), so we optimize the
1540 * append-to-file-case.
1542 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1543 * cleared because write clustering only works for commit
1544 * rpc's, not for the data portion of the write).
1547 if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1548 bp->b_flags |= B_NEEDCOMMIT;
1549 if (bp->b_dirtyoff == 0
1550 && bp->b_dirtyend == bp->b_bcount)
1551 bp->b_flags |= B_CLUSTEROK;
1553 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1555 bp->b_flags &= ~B_WRITEINPROG;
1558 * For an interrupted write, the buffer is still valid
1559 * and the write hasn't been pushed to the server yet,
1560 * so we can't set B_ERROR and report the interruption
1561 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1562 * is not relevant, so the rpc attempt is essentially
1563 * a noop. For the case of a V3 write rpc not being
1564 * committed to stable storage, the block is still
1565 * dirty and requires either a commit rpc or another
1566 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1567 * the block is reused. This is indicated by setting
1568 * the B_DELWRI and B_NEEDCOMMIT flags.
1570 * If the buffer is marked B_PAGING, it does not reside on
1571 * the vp's paging queues so we cannot call bdirty(). The
1572 * bp in this case is not an NFS cache block so we should
1576 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1580 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1581 if ((bp->b_flags & B_PAGING) == 0) {
1583 bp->b_flags &= ~B_DONE;
1585 if (error && (bp->b_flags & B_ASYNC) == 0)
1586 bp->b_flags |= B_EINTR;
1590 bp->b_flags |= B_ERROR;
1591 bp->b_error = np->n_error = error;
1592 np->n_flag |= NWRITEERR;
1594 bp->b_dirtyoff = bp->b_dirtyend = 0;
1602 bp->b_resid = uiop->uio_resid;
1604 nfs_clearcommit(vp->v_mount);
1610 * Used to aid in handling ftruncate() operations on the NFS client side.
1611 * Truncation creates a number of special problems for NFS. We have to
1612 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1613 * we have to properly handle VM pages or (potentially dirty) buffers
1614 * that straddle the truncation point.
1618 nfs_meta_setsize(struct vnode *vp, struct ucred *cred, struct proc *p, u_quad_t nsize)
1620 struct nfsnode *np = VTONFS(vp);
1621 u_quad_t tsize = np->n_size;
1622 int biosize = vp->v_mount->mnt_stat.f_iosize;
1627 if (np->n_size < tsize) {
1633 * vtruncbuf() doesn't get the buffer overlapping the
1634 * truncation point. We may have a B_DELWRI and/or B_CACHE
1635 * buffer that now needs to be truncated.
1637 error = vtruncbuf(vp, cred, p, nsize, biosize);
1638 lbn = nsize / biosize;
1639 bufsize = nsize & (biosize - 1);
1640 bp = nfs_getcacheblk(vp, lbn, bufsize, p);
1641 if (bp->b_dirtyoff > bp->b_bcount)
1642 bp->b_dirtyoff = bp->b_bcount;
1643 if (bp->b_dirtyend > bp->b_bcount)
1644 bp->b_dirtyend = bp->b_bcount;
1645 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1648 vnode_pager_setsize(vp, nsize);