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 1.2 2003/06/17 04:28:54 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>
59 #include <nfs/rpcv2.h>
60 #include <nfs/nfsproto.h>
62 #include <nfs/nfsmount.h>
63 #include <nfs/nqnfs.h>
64 #include <nfs/nfsnode.h>
66 static struct buf *nfs_getcacheblk __P((struct vnode *vp, daddr_t bn, int size,
69 extern int nfs_numasync;
70 extern int nfs_pbuf_freecnt;
71 extern struct nfsstats nfsstats;
74 * Vnode op for VM getpages.
78 struct vop_getpages_args /* {
83 vm_ooffset_t a_offset;
86 int i, error, nextoff, size, toff, count, npages;
98 p = curproc; /* XXX */
99 cred = curproc->p_ucred; /* XXX */
100 nmp = VFSTONFS(vp->v_mount);
104 if (vp->v_object == NULL) {
105 printf("nfs_getpages: called with non-merged cache vnode??\n");
106 return VM_PAGER_ERROR;
109 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
110 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
111 (void)nfs_fsinfo(nmp, vp, cred, p);
113 npages = btoc(count);
116 * If the requested page is partially valid, just return it and
117 * allow the pager to zero-out the blanks. Partially valid pages
118 * can only occur at the file EOF.
122 vm_page_t m = pages[ap->a_reqpage];
125 /* handled by vm_fault now */
126 /* vm_page_zero_invalid(m, TRUE); */
127 for (i = 0; i < npages; ++i) {
128 if (i != ap->a_reqpage)
129 vnode_pager_freepage(pages[i]);
136 * We use only the kva address for the buffer, but this is extremely
137 * convienient and fast.
139 bp = getpbuf(&nfs_pbuf_freecnt);
141 kva = (vm_offset_t) bp->b_data;
142 pmap_qenter(kva, pages, npages);
144 iov.iov_base = (caddr_t) kva;
148 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
149 uio.uio_resid = count;
150 uio.uio_segflg = UIO_SYSSPACE;
151 uio.uio_rw = UIO_READ;
154 error = nfs_readrpc(vp, &uio, cred);
155 pmap_qremove(kva, npages);
157 relpbuf(bp, &nfs_pbuf_freecnt);
159 if (error && (uio.uio_resid == count)) {
160 printf("nfs_getpages: error %d\n", error);
161 for (i = 0; i < npages; ++i) {
162 if (i != ap->a_reqpage)
163 vnode_pager_freepage(pages[i]);
165 return VM_PAGER_ERROR;
169 * Calculate the number of bytes read and validate only that number
170 * of bytes. Note that due to pending writes, size may be 0. This
171 * does not mean that the remaining data is invalid!
174 size = count - uio.uio_resid;
176 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
178 nextoff = toff + PAGE_SIZE;
181 m->flags &= ~PG_ZERO;
183 if (nextoff <= size) {
185 * Read operation filled an entire page
187 m->valid = VM_PAGE_BITS_ALL;
189 } else if (size > toff) {
191 * Read operation filled a partial page.
194 vm_page_set_validclean(m, 0, size - toff);
195 /* handled by vm_fault now */
196 /* vm_page_zero_invalid(m, TRUE); */
199 * Read operation was short. If no error occured
200 * we may have hit a zero-fill section. We simply
201 * leave valid set to 0.
205 if (i != ap->a_reqpage) {
207 * Whether or not to leave the page activated is up in
208 * the air, but we should put the page on a page queue
209 * somewhere (it already is in the object). Result:
210 * It appears that emperical results show that
211 * deactivating pages is best.
215 * Just in case someone was asking for this page we
216 * now tell them that it is ok to use.
219 if (m->flags & PG_WANTED)
222 vm_page_deactivate(m);
225 vnode_pager_freepage(m);
233 * Vnode op for VM putpages.
237 struct vop_putpages_args /* {
243 vm_ooffset_t a_offset;
250 int iomode, must_commit, i, error, npages, count;
256 struct nfsmount *nmp;
262 p = curproc; /* XXX */
263 cred = curproc->p_ucred; /* XXX */
264 nmp = VFSTONFS(vp->v_mount);
267 rtvals = ap->a_rtvals;
268 npages = btoc(count);
269 offset = IDX_TO_OFF(pages[0]->pindex);
271 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
272 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
273 (void)nfs_fsinfo(nmp, vp, cred, p);
275 for (i = 0; i < npages; i++) {
276 rtvals[i] = VM_PAGER_AGAIN;
280 * When putting pages, do not extend file past EOF.
283 if (offset + count > np->n_size) {
284 count = np->n_size - offset;
290 * We use only the kva address for the buffer, but this is extremely
291 * convienient and fast.
293 bp = getpbuf(&nfs_pbuf_freecnt);
295 kva = (vm_offset_t) bp->b_data;
296 pmap_qenter(kva, pages, npages);
298 iov.iov_base = (caddr_t) kva;
302 uio.uio_offset = offset;
303 uio.uio_resid = count;
304 uio.uio_segflg = UIO_SYSSPACE;
305 uio.uio_rw = UIO_WRITE;
308 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
309 iomode = NFSV3WRITE_UNSTABLE;
311 iomode = NFSV3WRITE_FILESYNC;
313 error = nfs_writerpc(vp, &uio, cred, &iomode, &must_commit);
315 pmap_qremove(kva, npages);
316 relpbuf(bp, &nfs_pbuf_freecnt);
319 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE;
320 for (i = 0; i < nwritten; i++) {
321 rtvals[i] = VM_PAGER_OK;
322 vm_page_undirty(pages[i]);
325 nfs_clearcommit(vp->v_mount);
331 * Vnode op for read using bio
334 nfs_bioread(vp, uio, ioflag, cred)
335 register struct vnode *vp;
336 register struct uio *uio;
340 register struct nfsnode *np = VTONFS(vp);
341 register int biosize, i;
342 struct buf *bp = 0, *rabp;
345 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
349 int nra, error = 0, n = 0, on = 0;
352 if (uio->uio_rw != UIO_READ)
353 panic("nfs_read mode");
355 if (uio->uio_resid == 0)
357 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
361 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
362 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
363 (void)nfs_fsinfo(nmp, vp, cred, p);
364 if (vp->v_type != VDIR &&
365 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
367 biosize = vp->v_mount->mnt_stat.f_iosize;
368 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE);
370 * For nfs, cache consistency can only be maintained approximately.
371 * Although RFC1094 does not specify the criteria, the following is
372 * believed to be compatible with the reference port.
373 * For nqnfs, full cache consistency is maintained within the loop.
375 * If the file's modify time on the server has changed since the
376 * last read rpc or you have written to the file,
377 * you may have lost data cache consistency with the
378 * server, so flush all of the file's data out of the cache.
379 * Then force a getattr rpc to ensure that you have up to date
381 * NB: This implies that cache data can be read when up to
382 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
383 * attributes this could be forced by setting n_attrstamp to 0 before
384 * the VOP_GETATTR() call.
386 if ((nmp->nm_flag & NFSMNT_NQNFS) == 0) {
387 if (np->n_flag & NMODIFIED) {
388 if (vp->v_type != VREG) {
389 if (vp->v_type != VDIR)
390 panic("nfs: bioread, not dir");
392 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
397 error = VOP_GETATTR(vp, &vattr, cred, p);
400 np->n_mtime = vattr.va_mtime.tv_sec;
402 error = VOP_GETATTR(vp, &vattr, cred, p);
405 if (np->n_mtime != vattr.va_mtime.tv_sec) {
406 if (vp->v_type == VDIR)
408 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
411 np->n_mtime = vattr.va_mtime.tv_sec;
418 * Get a valid lease. If cached data is stale, flush it.
420 if (nmp->nm_flag & NFSMNT_NQNFS) {
421 if (NQNFS_CKINVALID(vp, np, ND_READ)) {
423 error = nqnfs_getlease(vp, ND_READ, cred, p);
424 } while (error == NQNFS_EXPIRED);
427 if (np->n_lrev != np->n_brev ||
428 (np->n_flag & NQNFSNONCACHE) ||
429 ((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) {
430 if (vp->v_type == VDIR)
432 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
435 np->n_brev = np->n_lrev;
437 } else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) {
439 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
444 if (np->n_flag & NQNFSNONCACHE) {
445 switch (vp->v_type) {
447 return (nfs_readrpc(vp, uio, cred));
449 return (nfs_readlinkrpc(vp, uio, cred));
453 printf(" NQNFSNONCACHE: type %x unexpected\n",
457 switch (vp->v_type) {
459 nfsstats.biocache_reads++;
460 lbn = uio->uio_offset / biosize;
461 on = uio->uio_offset & (biosize - 1);
464 * Start the read ahead(s), as required.
466 if (nfs_numasync > 0 && nmp->nm_readahead > 0) {
467 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
468 (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) {
469 rabn = lbn + 1 + nra;
470 if (!incore(vp, rabn)) {
471 rabp = nfs_getcacheblk(vp, rabn, biosize, p);
474 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
475 rabp->b_flags |= (B_READ | B_ASYNC);
476 vfs_busy_pages(rabp, 0);
477 if (nfs_asyncio(rabp, cred, p)) {
478 rabp->b_flags |= B_INVAL|B_ERROR;
479 vfs_unbusy_pages(rabp);
491 * Obtain the buffer cache block. Figure out the buffer size
492 * when we are at EOF. If we are modifying the size of the
493 * buffer based on an EOF condition we need to hold
494 * nfs_rslock() through obtaining the buffer to prevent
495 * a potential writer-appender from messing with n_size.
496 * Otherwise we may accidently truncate the buffer and
499 * Note that bcount is *not* DEV_BSIZE aligned.
504 if ((off_t)lbn * biosize >= np->n_size) {
506 } else if ((off_t)(lbn + 1) * biosize > np->n_size) {
507 bcount = np->n_size - (off_t)lbn * biosize;
509 if (bcount != biosize) {
510 switch(nfs_rslock(np, p)) {
523 bp = nfs_getcacheblk(vp, lbn, bcount, p);
525 if (bcount != biosize)
531 * If B_CACHE is not set, we must issue the read. If this
532 * fails, we return an error.
535 if ((bp->b_flags & B_CACHE) == 0) {
536 bp->b_flags |= B_READ;
537 vfs_busy_pages(bp, 0);
538 error = nfs_doio(bp, cred, p);
546 * on is the offset into the current bp. Figure out how many
547 * bytes we can copy out of the bp. Note that bcount is
548 * NOT DEV_BSIZE aligned.
550 * Then figure out how many bytes we can copy into the uio.
555 n = min((unsigned)(bcount - on), uio->uio_resid);
558 nfsstats.biocache_readlinks++;
559 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, p);
562 if ((bp->b_flags & B_CACHE) == 0) {
563 bp->b_flags |= B_READ;
564 vfs_busy_pages(bp, 0);
565 error = nfs_doio(bp, cred, p);
567 bp->b_flags |= B_ERROR;
572 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
576 nfsstats.biocache_readdirs++;
577 if (np->n_direofoffset
578 && uio->uio_offset >= np->n_direofoffset) {
581 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
582 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
583 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, p);
586 if ((bp->b_flags & B_CACHE) == 0) {
587 bp->b_flags |= B_READ;
588 vfs_busy_pages(bp, 0);
589 error = nfs_doio(bp, cred, p);
593 while (error == NFSERR_BAD_COOKIE) {
594 printf("got bad cookie vp %p bp %p\n", vp, bp);
596 error = nfs_vinvalbuf(vp, 0, cred, p, 1);
598 * Yuck! The directory has been modified on the
599 * server. The only way to get the block is by
600 * reading from the beginning to get all the
603 * Leave the last bp intact unless there is an error.
604 * Loop back up to the while if the error is another
605 * NFSERR_BAD_COOKIE (double yuch!).
607 for (i = 0; i <= lbn && !error; i++) {
608 if (np->n_direofoffset
609 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
611 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, p);
614 if ((bp->b_flags & B_CACHE) == 0) {
615 bp->b_flags |= B_READ;
616 vfs_busy_pages(bp, 0);
617 error = nfs_doio(bp, cred, p);
619 * no error + B_INVAL == directory EOF,
622 if (error == 0 && (bp->b_flags & B_INVAL))
626 * An error will throw away the block and the
627 * for loop will break out. If no error and this
628 * is not the block we want, we throw away the
629 * block and go for the next one via the for loop.
631 if (error || i < lbn)
636 * The above while is repeated if we hit another cookie
637 * error. If we hit an error and it wasn't a cookie error,
645 * If not eof and read aheads are enabled, start one.
646 * (You need the current block first, so that you have the
647 * directory offset cookie of the next block.)
649 if (nfs_numasync > 0 && nmp->nm_readahead > 0 &&
650 (bp->b_flags & B_INVAL) == 0 &&
651 (np->n_direofoffset == 0 ||
652 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
653 !(np->n_flag & NQNFSNONCACHE) &&
654 !incore(vp, lbn + 1)) {
655 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, p);
657 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
658 rabp->b_flags |= (B_READ | B_ASYNC);
659 vfs_busy_pages(rabp, 0);
660 if (nfs_asyncio(rabp, cred, p)) {
661 rabp->b_flags |= B_INVAL|B_ERROR;
662 vfs_unbusy_pages(rabp);
671 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
672 * chopped for the EOF condition, we cannot tell how large
673 * NFS directories are going to be until we hit EOF. So
674 * an NFS directory buffer is *not* chopped to its EOF. Now,
675 * it just so happens that b_resid will effectively chop it
676 * to EOF. *BUT* this information is lost if the buffer goes
677 * away and is reconstituted into a B_CACHE state ( due to
678 * being VMIO ) later. So we keep track of the directory eof
679 * in np->n_direofoffset and chop it off as an extra step
682 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
683 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
684 n = np->n_direofoffset - uio->uio_offset;
687 printf(" nfs_bioread: type %x unexpected\n",vp->v_type);
692 error = uiomove(bp->b_data + on, (int)n, uio);
694 switch (vp->v_type) {
702 * Invalidate buffer if caching is disabled, forcing a
703 * re-read from the remote later.
705 if (np->n_flag & NQNFSNONCACHE)
706 bp->b_flags |= B_INVAL;
709 printf(" nfs_bioread: type %x unexpected\n",vp->v_type);
712 } while (error == 0 && uio->uio_resid > 0 && n > 0);
717 * Vnode op for write using bio
721 struct vop_write_args /* {
725 struct ucred *a_cred;
729 struct uio *uio = ap->a_uio;
730 struct proc *p = uio->uio_procp;
731 struct vnode *vp = ap->a_vp;
732 struct nfsnode *np = VTONFS(vp);
733 struct ucred *cred = ap->a_cred;
734 int ioflag = ap->a_ioflag;
737 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
740 int n, on, error = 0, iomode, must_commit;
744 if (uio->uio_rw != UIO_WRITE)
745 panic("nfs_write mode");
746 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_procp != curproc)
747 panic("nfs_write proc");
749 if (vp->v_type != VREG)
751 if (np->n_flag & NWRITEERR) {
752 np->n_flag &= ~NWRITEERR;
753 return (np->n_error);
755 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
756 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
757 (void)nfs_fsinfo(nmp, vp, cred, p);
760 * Synchronously flush pending buffers if we are in synchronous
761 * mode or if we are appending.
763 if (ioflag & (IO_APPEND | IO_SYNC)) {
764 if (np->n_flag & NMODIFIED) {
766 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
773 * If IO_APPEND then load uio_offset. We restart here if we cannot
774 * get the append lock.
777 if (ioflag & IO_APPEND) {
779 error = VOP_GETATTR(vp, &vattr, cred, p);
782 uio->uio_offset = np->n_size;
785 if (uio->uio_offset < 0)
787 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
789 if (uio->uio_resid == 0)
793 * We need to obtain the rslock if we intend to modify np->n_size
794 * in order to guarentee the append point with multiple contending
795 * writers, to guarentee that no other appenders modify n_size
796 * while we are trying to obtain a truncated buffer (i.e. to avoid
797 * accidently truncating data written by another appender due to
798 * the race), and to ensure that the buffer is populated prior to
799 * our extending of the file. We hold rslock through the entire
802 * Note that we do not synchronize the case where someone truncates
803 * the file while we are appending to it because attempting to lock
804 * this case may deadlock other parts of the system unexpectedly.
806 if ((ioflag & IO_APPEND) ||
807 uio->uio_offset + uio->uio_resid > np->n_size) {
808 switch(nfs_rslock(np, p)) {
823 * Maybe this should be above the vnode op call, but so long as
824 * file servers have no limits, i don't think it matters
826 if (p && uio->uio_offset + uio->uio_resid >
827 p->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
834 biosize = vp->v_mount->mnt_stat.f_iosize;
838 * Check for a valid write lease.
840 if ((nmp->nm_flag & NFSMNT_NQNFS) &&
841 NQNFS_CKINVALID(vp, np, ND_WRITE)) {
843 error = nqnfs_getlease(vp, ND_WRITE, cred, p);
844 } while (error == NQNFS_EXPIRED);
847 if (np->n_lrev != np->n_brev ||
848 (np->n_flag & NQNFSNONCACHE)) {
849 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
852 np->n_brev = np->n_lrev;
855 if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) {
856 iomode = NFSV3WRITE_FILESYNC;
857 error = nfs_writerpc(vp, uio, cred, &iomode, &must_commit);
859 nfs_clearcommit(vp->v_mount);
862 nfsstats.biocache_writes++;
863 lbn = uio->uio_offset / biosize;
864 on = uio->uio_offset & (biosize-1);
865 n = min((unsigned)(biosize - on), uio->uio_resid);
868 * Handle direct append and file extension cases, calculate
869 * unaligned buffer size.
872 if (uio->uio_offset == np->n_size && n) {
874 * Get the buffer (in its pre-append state to maintain
875 * B_CACHE if it was previously set). Resize the
876 * nfsnode after we have locked the buffer to prevent
877 * readers from reading garbage.
880 bp = nfs_getcacheblk(vp, lbn, bcount, p);
885 np->n_size = uio->uio_offset + n;
886 np->n_flag |= NMODIFIED;
887 vnode_pager_setsize(vp, np->n_size);
889 save = bp->b_flags & B_CACHE;
891 allocbuf(bp, bcount);
896 * Obtain the locked cache block first, and then
897 * adjust the file's size as appropriate.
900 if ((off_t)lbn * biosize + bcount < np->n_size) {
901 if ((off_t)(lbn + 1) * biosize < np->n_size)
904 bcount = np->n_size - (off_t)lbn * biosize;
906 bp = nfs_getcacheblk(vp, lbn, bcount, p);
907 if (uio->uio_offset + n > np->n_size) {
908 np->n_size = uio->uio_offset + n;
909 np->n_flag |= NMODIFIED;
910 vnode_pager_setsize(vp, np->n_size);
920 * Issue a READ if B_CACHE is not set. In special-append
921 * mode, B_CACHE is based on the buffer prior to the write
922 * op and is typically set, avoiding the read. If a read
923 * is required in special append mode, the server will
924 * probably send us a short-read since we extended the file
925 * on our end, resulting in b_resid == 0 and, thusly,
926 * B_CACHE getting set.
928 * We can also avoid issuing the read if the write covers
929 * the entire buffer. We have to make sure the buffer state
930 * is reasonable in this case since we will not be initiating
931 * I/O. See the comments in kern/vfs_bio.c's getblk() for
934 * B_CACHE may also be set due to the buffer being cached
938 if (on == 0 && n == bcount) {
939 bp->b_flags |= B_CACHE;
940 bp->b_flags &= ~(B_ERROR | B_INVAL);
943 if ((bp->b_flags & B_CACHE) == 0) {
944 bp->b_flags |= B_READ;
945 vfs_busy_pages(bp, 0);
946 error = nfs_doio(bp, cred, p);
956 if (bp->b_wcred == NOCRED) {
960 np->n_flag |= NMODIFIED;
963 * If dirtyend exceeds file size, chop it down. This should
964 * not normally occur but there is an append race where it
965 * might occur XXX, so we log it.
967 * If the chopping creates a reverse-indexed or degenerate
968 * situation with dirtyoff/end, we 0 both of them.
971 if (bp->b_dirtyend > bcount) {
972 printf("NFS append race @%lx:%d\n",
973 (long)bp->b_blkno * DEV_BSIZE,
974 bp->b_dirtyend - bcount);
975 bp->b_dirtyend = bcount;
978 if (bp->b_dirtyoff >= bp->b_dirtyend)
979 bp->b_dirtyoff = bp->b_dirtyend = 0;
982 * If the new write will leave a contiguous dirty
983 * area, just update the b_dirtyoff and b_dirtyend,
984 * otherwise force a write rpc of the old dirty area.
986 * While it is possible to merge discontiguous writes due to
987 * our having a B_CACHE buffer ( and thus valid read data
988 * for the hole), we don't because it could lead to
989 * significant cache coherency problems with multiple clients,
990 * especially if locking is implemented later on.
992 * as an optimization we could theoretically maintain
993 * a linked list of discontinuous areas, but we would still
994 * have to commit them separately so there isn't much
995 * advantage to it except perhaps a bit of asynchronization.
998 if (bp->b_dirtyend > 0 &&
999 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
1000 if (VOP_BWRITE(bp->b_vp, bp) == EINTR) {
1008 * Check for valid write lease and get one as required.
1009 * In case getblk() and/or bwrite() delayed us.
1011 if ((nmp->nm_flag & NFSMNT_NQNFS) &&
1012 NQNFS_CKINVALID(vp, np, ND_WRITE)) {
1014 error = nqnfs_getlease(vp, ND_WRITE, cred, p);
1015 } while (error == NQNFS_EXPIRED);
1020 if (np->n_lrev != np->n_brev ||
1021 (np->n_flag & NQNFSNONCACHE)) {
1023 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
1026 np->n_brev = np->n_lrev;
1031 error = uiomove((char *)bp->b_data + on, n, uio);
1034 * Since this block is being modified, it must be written
1035 * again and not just committed. Since write clustering does
1036 * not work for the stage 1 data write, only the stage 2
1037 * commit rpc, we have to clear B_CLUSTEROK as well.
1039 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1042 bp->b_flags |= B_ERROR;
1048 * Only update dirtyoff/dirtyend if not a degenerate
1052 if (bp->b_dirtyend > 0) {
1053 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
1054 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
1056 bp->b_dirtyoff = on;
1057 bp->b_dirtyend = on + n;
1059 vfs_bio_set_validclean(bp, on, n);
1062 * If IO_NOWDRAIN then set B_NOWDRAIN (e.g. nfs-backed VN
1063 * filesystem). XXX also use for loopback NFS mounts.
1065 if (ioflag & IO_NOWDRAIN)
1066 bp->b_flags |= B_NOWDRAIN;
1069 * If the lease is non-cachable or IO_SYNC do bwrite().
1071 * IO_INVAL appears to be unused. The idea appears to be
1072 * to turn off caching in this case. Very odd. XXX
1074 if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) {
1075 if (ioflag & IO_INVAL)
1076 bp->b_flags |= B_NOCACHE;
1077 error = VOP_BWRITE(bp->b_vp, bp);
1080 if (np->n_flag & NQNFSNONCACHE) {
1081 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1);
1085 } else if ((n + on) == biosize &&
1086 (nmp->nm_flag & NFSMNT_NQNFS) == 0) {
1087 bp->b_flags |= B_ASYNC;
1088 (void)nfs_writebp(bp, 0, 0);
1092 } while (uio->uio_resid > 0 && n > 0);
1095 nfs_rsunlock(np, p);
1101 * Get an nfs cache block.
1103 * Allocate a new one if the block isn't currently in the cache
1104 * and return the block marked busy. If the calling process is
1105 * interrupted by a signal for an interruptible mount point, return
1108 * The caller must carefully deal with the possible B_INVAL state of
1109 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it
1110 * indirectly), so synchronous reads can be issued without worrying about
1111 * the B_INVAL state. We have to be a little more careful when dealing
1112 * with writes (see comments in nfs_write()) when extending a file past
1116 nfs_getcacheblk(vp, bn, size, p)
1122 register struct buf *bp;
1124 struct nfsmount *nmp;
1129 if (nmp->nm_flag & NFSMNT_INT) {
1130 bp = getblk(vp, bn, size, PCATCH, 0);
1131 while (bp == (struct buf *)0) {
1132 if (nfs_sigintr(nmp, (struct nfsreq *)0, p))
1133 return ((struct buf *)0);
1134 bp = getblk(vp, bn, size, 0, 2 * hz);
1137 bp = getblk(vp, bn, size, 0, 0);
1140 if (vp->v_type == VREG) {
1143 biosize = mp->mnt_stat.f_iosize;
1144 bp->b_blkno = bn * (biosize / DEV_BSIZE);
1150 * Flush and invalidate all dirty buffers. If another process is already
1151 * doing the flush, just wait for completion.
1154 nfs_vinvalbuf(vp, flags, cred, p, intrflg)
1161 register struct nfsnode *np = VTONFS(vp);
1162 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1163 int error = 0, slpflag, slptimeo;
1165 if (vp->v_flag & VXLOCK) {
1169 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1179 * First wait for any other process doing a flush to complete.
1181 while (np->n_flag & NFLUSHINPROG) {
1182 np->n_flag |= NFLUSHWANT;
1183 error = tsleep((caddr_t)&np->n_flag, PRIBIO + 2, "nfsvinval",
1185 if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p))
1190 * Now, flush as required.
1192 np->n_flag |= NFLUSHINPROG;
1193 error = vinvalbuf(vp, flags, cred, p, slpflag, 0);
1195 if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p)) {
1196 np->n_flag &= ~NFLUSHINPROG;
1197 if (np->n_flag & NFLUSHWANT) {
1198 np->n_flag &= ~NFLUSHWANT;
1199 wakeup((caddr_t)&np->n_flag);
1203 error = vinvalbuf(vp, flags, cred, p, 0, slptimeo);
1205 np->n_flag &= ~(NMODIFIED | NFLUSHINPROG);
1206 if (np->n_flag & NFLUSHWANT) {
1207 np->n_flag &= ~NFLUSHWANT;
1208 wakeup((caddr_t)&np->n_flag);
1214 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1215 * This is mainly to avoid queueing async I/O requests when the nfsiods
1216 * are all hung on a dead server.
1218 * Note: nfs_asyncio() does not clear (B_ERROR|B_INVAL) but when the bp
1219 * is eventually dequeued by the async daemon, nfs_doio() *will*.
1222 nfs_asyncio(bp, cred, procp)
1223 register struct buf *bp;
1227 struct nfsmount *nmp;
1235 * If no async daemons then return EIO to force caller to run the rpc
1238 if (nfs_numasync == 0)
1241 nmp = VFSTONFS(bp->b_vp->v_mount);
1244 * Commits are usually short and sweet so lets save some cpu and
1245 * leave the async daemons for more important rpc's (such as reads
1248 if ((bp->b_flags & (B_READ|B_NEEDCOMMIT)) == B_NEEDCOMMIT &&
1249 (nmp->nm_bufqiods > nfs_numasync / 2)) {
1254 if (nmp->nm_flag & NFSMNT_INT)
1259 * Find a free iod to process this request.
1261 for (i = 0; i < NFS_MAXASYNCDAEMON; i++)
1262 if (nfs_iodwant[i]) {
1264 * Found one, so wake it up and tell it which
1268 ("nfs_asyncio: waking iod %d for mount %p\n",
1270 nfs_iodwant[i] = (struct proc *)0;
1271 nfs_iodmount[i] = nmp;
1273 wakeup((caddr_t)&nfs_iodwant[i]);
1279 * If none are free, we may already have an iod working on this mount
1280 * point. If so, it will process our request.
1283 if (nmp->nm_bufqiods > 0) {
1285 ("nfs_asyncio: %d iods are already processing mount %p\n",
1286 nmp->nm_bufqiods, nmp));
1292 * If we have an iod which can process the request, then queue
1297 * Ensure that the queue never grows too large. We still want
1298 * to asynchronize so we block rather then return EIO.
1300 while (nmp->nm_bufqlen >= 2*nfs_numasync) {
1302 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp));
1303 nmp->nm_bufqwant = TRUE;
1304 error = tsleep(&nmp->nm_bufq, slpflag | PRIBIO,
1305 "nfsaio", slptimeo);
1307 if (nfs_sigintr(nmp, NULL, procp))
1309 if (slpflag == PCATCH) {
1315 * We might have lost our iod while sleeping,
1316 * so check and loop if nescessary.
1318 if (nmp->nm_bufqiods == 0) {
1320 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1325 if (bp->b_flags & B_READ) {
1326 if (bp->b_rcred == NOCRED && cred != NOCRED) {
1331 bp->b_flags |= B_WRITEINPROG;
1332 if (bp->b_wcred == NOCRED && cred != NOCRED) {
1339 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1345 * All the iods are busy on other mounts, so return EIO to
1346 * force the caller to process the i/o synchronously.
1348 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n"));
1353 * Do an I/O operation to/from a cache block. This may be called
1354 * synchronously or from an nfsiod.
1365 struct nfsmount *nmp;
1366 int error = 0, iomode, must_commit = 0;
1372 nmp = VFSTONFS(vp->v_mount);
1374 uiop->uio_iov = &io;
1375 uiop->uio_iovcnt = 1;
1376 uiop->uio_segflg = UIO_SYSSPACE;
1377 uiop->uio_procp = p;
1380 * clear B_ERROR and B_INVAL state prior to initiating the I/O. We
1381 * do this here so we do not have to do it in all the code that
1384 bp->b_flags &= ~(B_ERROR | B_INVAL);
1386 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp));
1389 * Historically, paging was done with physio, but no more.
1391 if (bp->b_flags & B_PHYS) {
1393 * ...though reading /dev/drum still gets us here.
1395 io.iov_len = uiop->uio_resid = bp->b_bcount;
1396 /* mapping was done by vmapbuf() */
1397 io.iov_base = bp->b_data;
1398 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1399 if (bp->b_flags & B_READ) {
1400 uiop->uio_rw = UIO_READ;
1401 nfsstats.read_physios++;
1402 error = nfs_readrpc(vp, uiop, cr);
1406 iomode = NFSV3WRITE_DATASYNC;
1407 uiop->uio_rw = UIO_WRITE;
1408 nfsstats.write_physios++;
1409 error = nfs_writerpc(vp, uiop, cr, &iomode, &com);
1412 bp->b_flags |= B_ERROR;
1413 bp->b_error = error;
1415 } else if (bp->b_flags & B_READ) {
1416 io.iov_len = uiop->uio_resid = bp->b_bcount;
1417 io.iov_base = bp->b_data;
1418 uiop->uio_rw = UIO_READ;
1420 switch (vp->v_type) {
1422 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1423 nfsstats.read_bios++;
1424 error = nfs_readrpc(vp, uiop, cr);
1427 if (uiop->uio_resid) {
1429 * If we had a short read with no error, we must have
1430 * hit a file hole. We should zero-fill the remainder.
1431 * This can also occur if the server hits the file EOF.
1433 * Holes used to be able to occur due to pending
1434 * writes, but that is not possible any longer.
1436 int nread = bp->b_bcount - uiop->uio_resid;
1437 int left = uiop->uio_resid;
1440 bzero((char *)bp->b_data + nread, left);
1441 uiop->uio_resid = 0;
1444 if (p && (vp->v_flag & VTEXT) &&
1445 (((nmp->nm_flag & NFSMNT_NQNFS) &&
1446 NQNFS_CKINVALID(vp, np, ND_READ) &&
1447 np->n_lrev != np->n_brev) ||
1448 (!(nmp->nm_flag & NFSMNT_NQNFS) &&
1449 np->n_mtime != np->n_vattr.va_mtime.tv_sec))) {
1450 uprintf("Process killed due to text file modification\n");
1451 psignal(p, SIGKILL);
1456 uiop->uio_offset = (off_t)0;
1457 nfsstats.readlink_bios++;
1458 error = nfs_readlinkrpc(vp, uiop, cr);
1461 nfsstats.readdir_bios++;
1462 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1463 if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
1464 error = nfs_readdirplusrpc(vp, uiop, cr);
1465 if (error == NFSERR_NOTSUPP)
1466 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1468 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1469 error = nfs_readdirrpc(vp, uiop, cr);
1471 * end-of-directory sets B_INVAL but does not generate an
1474 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1475 bp->b_flags |= B_INVAL;
1478 printf("nfs_doio: type %x unexpected\n",vp->v_type);
1482 bp->b_flags |= B_ERROR;
1483 bp->b_error = error;
1487 * If we only need to commit, try to commit
1489 if (bp->b_flags & B_NEEDCOMMIT) {
1493 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1494 bp->b_flags |= B_WRITEINPROG;
1496 bp->b_vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1498 bp->b_flags &= ~B_WRITEINPROG;
1500 bp->b_dirtyoff = bp->b_dirtyend = 0;
1501 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1506 if (retv == NFSERR_STALEWRITEVERF) {
1507 nfs_clearcommit(bp->b_vp->v_mount);
1512 * Setup for actual write
1515 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1516 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1518 if (bp->b_dirtyend > bp->b_dirtyoff) {
1519 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1521 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1523 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1524 uiop->uio_rw = UIO_WRITE;
1525 nfsstats.write_bios++;
1527 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1528 iomode = NFSV3WRITE_UNSTABLE;
1530 iomode = NFSV3WRITE_FILESYNC;
1532 bp->b_flags |= B_WRITEINPROG;
1533 error = nfs_writerpc(vp, uiop, cr, &iomode, &must_commit);
1536 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1537 * to cluster the buffers needing commit. This will allow
1538 * the system to submit a single commit rpc for the whole
1539 * cluster. We can do this even if the buffer is not 100%
1540 * dirty (relative to the NFS blocksize), so we optimize the
1541 * append-to-file-case.
1543 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1544 * cleared because write clustering only works for commit
1545 * rpc's, not for the data portion of the write).
1548 if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1549 bp->b_flags |= B_NEEDCOMMIT;
1550 if (bp->b_dirtyoff == 0
1551 && bp->b_dirtyend == bp->b_bcount)
1552 bp->b_flags |= B_CLUSTEROK;
1554 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1556 bp->b_flags &= ~B_WRITEINPROG;
1559 * For an interrupted write, the buffer is still valid
1560 * and the write hasn't been pushed to the server yet,
1561 * so we can't set B_ERROR and report the interruption
1562 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1563 * is not relevant, so the rpc attempt is essentially
1564 * a noop. For the case of a V3 write rpc not being
1565 * committed to stable storage, the block is still
1566 * dirty and requires either a commit rpc or another
1567 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1568 * the block is reused. This is indicated by setting
1569 * the B_DELWRI and B_NEEDCOMMIT flags.
1571 * If the buffer is marked B_PAGING, it does not reside on
1572 * the vp's paging queues so we cannot call bdirty(). The
1573 * bp in this case is not an NFS cache block so we should
1577 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1581 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1582 if ((bp->b_flags & B_PAGING) == 0) {
1584 bp->b_flags &= ~B_DONE;
1586 if (error && (bp->b_flags & B_ASYNC) == 0)
1587 bp->b_flags |= B_EINTR;
1591 bp->b_flags |= B_ERROR;
1592 bp->b_error = np->n_error = error;
1593 np->n_flag |= NWRITEERR;
1595 bp->b_dirtyoff = bp->b_dirtyend = 0;
1603 bp->b_resid = uiop->uio_resid;
1605 nfs_clearcommit(vp->v_mount);
1611 * Used to aid in handling ftruncate() operations on the NFS client side.
1612 * Truncation creates a number of special problems for NFS. We have to
1613 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1614 * we have to properly handle VM pages or (potentially dirty) buffers
1615 * that straddle the truncation point.
1619 nfs_meta_setsize(struct vnode *vp, struct ucred *cred, struct proc *p, u_quad_t nsize)
1621 struct nfsnode *np = VTONFS(vp);
1622 u_quad_t tsize = np->n_size;
1623 int biosize = vp->v_mount->mnt_stat.f_iosize;
1628 if (np->n_size < tsize) {
1634 * vtruncbuf() doesn't get the buffer overlapping the
1635 * truncation point. We may have a B_DELWRI and/or B_CACHE
1636 * buffer that now needs to be truncated.
1638 error = vtruncbuf(vp, cred, p, nsize, biosize);
1639 lbn = nsize / biosize;
1640 bufsize = nsize & (biosize - 1);
1641 bp = nfs_getcacheblk(vp, lbn, bufsize, p);
1642 if (bp->b_dirtyoff > bp->b_bcount)
1643 bp->b_dirtyoff = bp->b_bcount;
1644 if (bp->b_dirtyend > bp->b_bcount)
1645 bp->b_dirtyend = bp->b_bcount;
1646 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1649 vnode_pager_setsize(vp, nsize);