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
39 #include <sys/cdefs.h>
40 __FBSDID("$FreeBSD$");
42 #include <sys/param.h>
43 #include <sys/systm.h>
46 #include <sys/kernel.h>
47 #include <sys/mount.h>
49 #include <sys/resourcevar.h>
50 #include <sys/signalvar.h>
51 #include <sys/vmmeter.h>
52 #include <sys/vnode.h>
55 #include <vm/vm_extern.h>
56 #include <vm/vm_page.h>
57 #include <vm/vm_object.h>
58 #include <vm/vm_pager.h>
59 #include <vm/vnode_pager.h>
61 #include <nfs/rpcv2.h>
62 #include <nfs/nfsproto.h>
63 #include <nfsclient/nfs.h>
64 #include <nfsclient/nfsmount.h>
65 #include <nfsclient/nfsnode.h>
68 * Just call nfs_writebp() with the force argument set to 1.
70 * NOTE: B_DONE may or may not be set in a_bp on call.
73 nfs_bwrite(struct buf *bp)
76 return (nfs_writebp(bp, 1, curthread));
79 struct buf_ops buf_ops_nfs = {
84 static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size,
88 * Vnode op for VM getpages.
91 nfs_getpages(struct vop_getpages_args *ap)
93 int i, error, nextoff, size, toff, count, npages;
101 struct nfsmount *nmp;
107 td = curthread; /* XXX */
108 cred = curthread->td_ucred; /* XXX */
109 nmp = VFSTONFS(vp->v_mount);
113 if (vp->v_object == NULL) {
114 printf("nfs_getpages: called with non-merged cache vnode??\n");
115 return VM_PAGER_ERROR;
118 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
119 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
120 (void)nfs_fsinfo(nmp, vp, cred, td);
123 npages = btoc(count);
126 * If the requested page is partially valid, just return it and
127 * allow the pager to zero-out the blanks. Partially valid pages
128 * can only occur at the file EOF.
132 vm_page_t m = pages[ap->a_reqpage];
135 /* handled by vm_fault now */
136 /* vm_page_zero_invalid(m, TRUE); */
137 for (i = 0; i < npages; ++i) {
138 if (i != ap->a_reqpage)
139 vm_page_free(pages[i]);
146 * We use only the kva address for the buffer, but this is extremely
147 * convienient and fast.
149 bp = getpbuf(&nfs_pbuf_freecnt);
151 kva = (vm_offset_t) bp->b_data;
152 pmap_qenter(kva, pages, npages);
154 cnt.v_vnodepgsin += npages;
156 iov.iov_base = (caddr_t) kva;
160 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex);
161 uio.uio_resid = count;
162 uio.uio_segflg = UIO_SYSSPACE;
163 uio.uio_rw = UIO_READ;
166 error = nfs_readrpc(vp, &uio, cred);
167 pmap_qremove(kva, npages);
169 relpbuf(bp, &nfs_pbuf_freecnt);
171 if (error && (uio.uio_resid == count)) {
172 printf("nfs_getpages: error %d\n", error);
173 for (i = 0; i < npages; ++i) {
174 if (i != ap->a_reqpage)
175 vm_page_free(pages[i]);
177 return VM_PAGER_ERROR;
181 * Calculate the number of bytes read and validate only that number
182 * of bytes. Note that due to pending writes, size may be 0. This
183 * does not mean that the remaining data is invalid!
186 size = count - uio.uio_resid;
188 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) {
190 nextoff = toff + PAGE_SIZE;
193 m->flags &= ~PG_ZERO;
195 if (nextoff <= size) {
197 * Read operation filled an entire page
199 m->valid = VM_PAGE_BITS_ALL;
201 } else if (size > toff) {
203 * Read operation filled a partial page.
206 vm_page_set_validclean(m, 0, size - toff);
207 /* handled by vm_fault now */
208 /* vm_page_zero_invalid(m, TRUE); */
211 * Read operation was short. If no error occured
212 * we may have hit a zero-fill section. We simply
213 * leave valid set to 0.
217 if (i != ap->a_reqpage) {
219 * Whether or not to leave the page activated is up in
220 * the air, but we should put the page on a page queue
221 * somewhere (it already is in the object). Result:
222 * It appears that emperical results show that
223 * deactivating pages is best.
227 * Just in case someone was asking for this page we
228 * now tell them that it is ok to use.
231 if (m->flags & PG_WANTED)
234 vm_page_deactivate(m);
245 * Vnode op for VM putpages.
248 nfs_putpages(struct vop_putpages_args *ap)
254 int iomode, must_commit, i, error, npages, count;
260 struct nfsmount *nmp;
268 td = curthread; /* XXX */
269 cred = curthread->td_ucred; /* XXX */
270 nmp = VFSTONFS(vp->v_mount);
273 rtvals = ap->a_rtvals;
274 npages = btoc(count);
275 offset = IDX_TO_OFF(pages[0]->pindex);
277 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
278 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) {
279 (void)nfs_fsinfo(nmp, vp, cred, td);
282 for (i = 0; i < npages; i++)
283 rtvals[i] = VM_PAGER_AGAIN;
286 * When putting pages, do not extend file past EOF.
289 if (offset + count > np->n_size) {
290 count = np->n_size - offset;
296 * We use only the kva address for the buffer, but this is extremely
297 * convienient and fast.
299 bp = getpbuf(&nfs_pbuf_freecnt);
301 kva = (vm_offset_t) bp->b_data;
302 pmap_qenter(kva, pages, npages);
304 cnt.v_vnodepgsout += count;
306 iov.iov_base = (caddr_t) kva;
310 uio.uio_offset = offset;
311 uio.uio_resid = count;
312 uio.uio_segflg = UIO_SYSSPACE;
313 uio.uio_rw = UIO_WRITE;
316 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0)
317 iomode = NFSV3WRITE_UNSTABLE;
319 iomode = NFSV3WRITE_FILESYNC;
321 error = nfs_writerpc(vp, &uio, cred, &iomode, &must_commit);
323 pmap_qremove(kva, npages);
324 relpbuf(bp, &nfs_pbuf_freecnt);
327 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE;
328 for (i = 0; i < nwritten; i++) {
329 rtvals[i] = VM_PAGER_OK;
330 vm_page_undirty(pages[i]);
333 nfs_clearcommit(vp->v_mount);
340 * Vnode op for read using bio
343 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred)
345 struct nfsnode *np = VTONFS(vp);
347 struct buf *bp = 0, *rabp;
350 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
354 int nra, error = 0, n = 0, on = 0;
357 if (uio->uio_rw != UIO_READ)
358 panic("nfs_read mode");
360 if (uio->uio_resid == 0)
362 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */
366 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
367 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
368 (void)nfs_fsinfo(nmp, vp, cred, td);
369 if (vp->v_type != VDIR &&
370 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
372 biosize = vp->v_mount->mnt_stat.f_iosize;
373 seqcount = (int)((off_t)(ioflag >> 16) * biosize / BKVASIZE);
375 * For nfs, cache consistency can only be maintained approximately.
376 * Although RFC1094 does not specify the criteria, the following is
377 * believed to be compatible with the reference port.
379 * If the file's modify time on the server has changed since the
380 * last read rpc or you have written to the file,
381 * you may have lost data cache consistency with the
382 * server, so flush all of the file's data out of the cache.
383 * Then force a getattr rpc to ensure that you have up to date
385 * NB: This implies that cache data can be read when up to
386 * NFS_ATTRTIMEO seconds out of date. If you find that you need current
387 * attributes this could be forced by setting n_attrstamp to 0 before
388 * the VOP_GETATTR() call.
390 if (np->n_flag & NMODIFIED) {
391 if (vp->v_type != VREG) {
392 if (vp->v_type != VDIR)
393 panic("nfs: bioread, not dir");
395 error = nfs_vinvalbuf(vp, V_SAVE, cred, td, 1);
400 error = VOP_GETATTR(vp, &vattr, cred, td);
403 np->n_mtime = vattr.va_mtime.tv_sec;
405 error = VOP_GETATTR(vp, &vattr, cred, td);
408 if (np->n_mtime != vattr.va_mtime.tv_sec) {
409 if (vp->v_type == VDIR)
411 error = nfs_vinvalbuf(vp, V_SAVE, cred, td, 1);
414 np->n_mtime = vattr.va_mtime.tv_sec;
418 switch (vp->v_type) {
420 nfsstats.biocache_reads++;
421 lbn = uio->uio_offset / biosize;
422 on = uio->uio_offset & (biosize - 1);
425 * Start the read ahead(s), as required.
427 if (nmp->nm_readahead > 0) {
428 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
429 (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) {
430 rabn = lbn + 1 + nra;
431 if (!incore(vp, rabn)) {
432 rabp = nfs_getcacheblk(vp, rabn, biosize, td);
435 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
436 rabp->b_flags |= B_ASYNC;
437 rabp->b_iocmd = BIO_READ;
438 vfs_busy_pages(rabp, 0);
439 if (nfs_asyncio(rabp, cred, td)) {
440 rabp->b_flags |= B_INVAL;
441 rabp->b_ioflags |= BIO_ERROR;
442 vfs_unbusy_pages(rabp);
454 * Obtain the buffer cache block. Figure out the buffer size
455 * when we are at EOF. If we are modifying the size of the
456 * buffer based on an EOF condition we need to hold
457 * nfs_rslock() through obtaining the buffer to prevent
458 * a potential writer-appender from messing with n_size.
459 * Otherwise we may accidently truncate the buffer and
462 * Note that bcount is *not* DEV_BSIZE aligned.
467 if ((off_t)lbn * biosize >= np->n_size) {
469 } else if ((off_t)(lbn + 1) * biosize > np->n_size) {
470 bcount = np->n_size - (off_t)lbn * biosize;
472 if (bcount != biosize) {
473 switch(nfs_rslock(np, td)) {
486 bp = nfs_getcacheblk(vp, lbn, bcount, td);
488 if (bcount != biosize)
489 nfs_rsunlock(np, td);
494 * If B_CACHE is not set, we must issue the read. If this
495 * fails, we return an error.
498 if ((bp->b_flags & B_CACHE) == 0) {
499 bp->b_iocmd = BIO_READ;
500 vfs_busy_pages(bp, 0);
501 error = nfs_doio(bp, cred, td);
509 * on is the offset into the current bp. Figure out how many
510 * bytes we can copy out of the bp. Note that bcount is
511 * NOT DEV_BSIZE aligned.
513 * Then figure out how many bytes we can copy into the uio.
518 n = min((unsigned)(bcount - on), uio->uio_resid);
521 nfsstats.biocache_readlinks++;
522 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td);
525 if ((bp->b_flags & B_CACHE) == 0) {
526 bp->b_iocmd = BIO_READ;
527 vfs_busy_pages(bp, 0);
528 error = nfs_doio(bp, cred, td);
530 bp->b_ioflags |= BIO_ERROR;
535 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid);
539 nfsstats.biocache_readdirs++;
540 if (np->n_direofoffset
541 && uio->uio_offset >= np->n_direofoffset) {
544 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ;
545 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1);
546 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td);
549 if ((bp->b_flags & B_CACHE) == 0) {
550 bp->b_iocmd = BIO_READ;
551 vfs_busy_pages(bp, 0);
552 error = nfs_doio(bp, cred, td);
556 while (error == NFSERR_BAD_COOKIE) {
557 printf("got bad cookie vp %p bp %p\n", vp, bp);
559 error = nfs_vinvalbuf(vp, 0, cred, td, 1);
561 * Yuck! The directory has been modified on the
562 * server. The only way to get the block is by
563 * reading from the beginning to get all the
566 * Leave the last bp intact unless there is an error.
567 * Loop back up to the while if the error is another
568 * NFSERR_BAD_COOKIE (double yuch!).
570 for (i = 0; i <= lbn && !error; i++) {
571 if (np->n_direofoffset
572 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset)
574 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td);
577 if ((bp->b_flags & B_CACHE) == 0) {
578 bp->b_iocmd = BIO_READ;
579 vfs_busy_pages(bp, 0);
580 error = nfs_doio(bp, cred, td);
582 * no error + B_INVAL == directory EOF,
585 if (error == 0 && (bp->b_flags & B_INVAL))
589 * An error will throw away the block and the
590 * for loop will break out. If no error and this
591 * is not the block we want, we throw away the
592 * block and go for the next one via the for loop.
594 if (error || i < lbn)
599 * The above while is repeated if we hit another cookie
600 * error. If we hit an error and it wasn't a cookie error,
608 * If not eof and read aheads are enabled, start one.
609 * (You need the current block first, so that you have the
610 * directory offset cookie of the next block.)
612 if (nmp->nm_readahead > 0 &&
613 (bp->b_flags & B_INVAL) == 0 &&
614 (np->n_direofoffset == 0 ||
615 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&
616 !incore(vp, lbn + 1)) {
617 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td);
619 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) {
620 rabp->b_flags |= B_ASYNC;
621 rabp->b_iocmd = BIO_READ;
622 vfs_busy_pages(rabp, 0);
623 if (nfs_asyncio(rabp, cred, td)) {
624 rabp->b_flags |= B_INVAL;
625 rabp->b_ioflags |= BIO_ERROR;
626 vfs_unbusy_pages(rabp);
635 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is
636 * chopped for the EOF condition, we cannot tell how large
637 * NFS directories are going to be until we hit EOF. So
638 * an NFS directory buffer is *not* chopped to its EOF. Now,
639 * it just so happens that b_resid will effectively chop it
640 * to EOF. *BUT* this information is lost if the buffer goes
641 * away and is reconstituted into a B_CACHE state ( due to
642 * being VMIO ) later. So we keep track of the directory eof
643 * in np->n_direofoffset and chop it off as an extra step
646 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on);
647 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset)
648 n = np->n_direofoffset - uio->uio_offset;
651 printf(" nfs_bioread: type %x unexpected\n", vp->v_type);
656 error = uiomove(bp->b_data + on, (int)n, uio);
658 switch (vp->v_type) {
667 printf(" nfs_bioread: type %x unexpected\n", vp->v_type);
670 } while (error == 0 && uio->uio_resid > 0 && n > 0);
675 * Vnode op for write using bio
678 nfs_write(struct vop_write_args *ap)
681 struct uio *uio = ap->a_uio;
682 struct thread *td = uio->uio_td;
683 struct vnode *vp = ap->a_vp;
684 struct nfsnode *np = VTONFS(vp);
685 struct ucred *cred = ap->a_cred;
686 int ioflag = ap->a_ioflag;
689 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
692 int n, on, error = 0;
694 struct proc *p = td?td->td_proc:NULL;
699 if (uio->uio_rw != UIO_WRITE)
700 panic("nfs_write mode");
701 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread)
702 panic("nfs_write proc");
704 if (vp->v_type != VREG)
706 if (np->n_flag & NWRITEERR) {
707 np->n_flag &= ~NWRITEERR;
708 return (np->n_error);
710 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 &&
711 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0)
712 (void)nfs_fsinfo(nmp, vp, cred, td);
715 * Synchronously flush pending buffers if we are in synchronous
716 * mode or if we are appending.
718 if (ioflag & (IO_APPEND | IO_SYNC)) {
719 if (np->n_flag & NMODIFIED) {
721 error = nfs_vinvalbuf(vp, V_SAVE, cred, td, 1);
728 * If IO_APPEND then load uio_offset. We restart here if we cannot
729 * get the append lock.
732 if (ioflag & IO_APPEND) {
734 error = VOP_GETATTR(vp, &vattr, cred, td);
737 uio->uio_offset = np->n_size;
740 if (uio->uio_offset < 0)
742 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize)
744 if (uio->uio_resid == 0)
748 * We need to obtain the rslock if we intend to modify np->n_size
749 * in order to guarentee the append point with multiple contending
750 * writers, to guarentee that no other appenders modify n_size
751 * while we are trying to obtain a truncated buffer (i.e. to avoid
752 * accidently truncating data written by another appender due to
753 * the race), and to ensure that the buffer is populated prior to
754 * our extending of the file. We hold rslock through the entire
757 * Note that we do not synchronize the case where someone truncates
758 * the file while we are appending to it because attempting to lock
759 * this case may deadlock other parts of the system unexpectedly.
761 if ((ioflag & IO_APPEND) ||
762 uio->uio_offset + uio->uio_resid > np->n_size) {
763 switch(nfs_rslock(np, td)) {
778 * Maybe this should be above the vnode op call, but so long as
779 * file servers have no limits, i don't think it matters
781 if (p && uio->uio_offset + uio->uio_resid >
782 p->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
787 nfs_rsunlock(np, td);
791 biosize = vp->v_mount->mnt_stat.f_iosize;
794 nfsstats.biocache_writes++;
795 lbn = uio->uio_offset / biosize;
796 on = uio->uio_offset & (biosize-1);
797 n = min((unsigned)(biosize - on), uio->uio_resid);
800 * Handle direct append and file extension cases, calculate
801 * unaligned buffer size.
804 if (uio->uio_offset == np->n_size && n) {
806 * Get the buffer (in its pre-append state to maintain
807 * B_CACHE if it was previously set). Resize the
808 * nfsnode after we have locked the buffer to prevent
809 * readers from reading garbage.
812 bp = nfs_getcacheblk(vp, lbn, bcount, td);
817 np->n_size = uio->uio_offset + n;
818 np->n_flag |= NMODIFIED;
819 vnode_pager_setsize(vp, np->n_size);
821 save = bp->b_flags & B_CACHE;
823 allocbuf(bp, bcount);
825 bp->b_magic = B_MAGIC_NFS;
826 bp->b_op = &buf_ops_nfs;
830 * Obtain the locked cache block first, and then
831 * adjust the file's size as appropriate.
834 if ((off_t)lbn * biosize + bcount < np->n_size) {
835 if ((off_t)(lbn + 1) * biosize < np->n_size)
838 bcount = np->n_size - (off_t)lbn * biosize;
840 bp = nfs_getcacheblk(vp, lbn, bcount, td);
841 if (uio->uio_offset + n > np->n_size) {
842 np->n_size = uio->uio_offset + n;
843 np->n_flag |= NMODIFIED;
844 vnode_pager_setsize(vp, np->n_size);
854 * Issue a READ if B_CACHE is not set. In special-append
855 * mode, B_CACHE is based on the buffer prior to the write
856 * op and is typically set, avoiding the read. If a read
857 * is required in special append mode, the server will
858 * probably send us a short-read since we extended the file
859 * on our end, resulting in b_resid == 0 and, thusly,
860 * B_CACHE getting set.
862 * We can also avoid issuing the read if the write covers
863 * the entire buffer. We have to make sure the buffer state
864 * is reasonable in this case since we will not be initiating
865 * I/O. See the comments in kern/vfs_bio.c's getblk() for
868 * B_CACHE may also be set due to the buffer being cached
872 if (on == 0 && n == bcount) {
873 bp->b_flags |= B_CACHE;
874 bp->b_flags &= ~B_INVAL;
875 bp->b_ioflags &= ~BIO_ERROR;
878 if ((bp->b_flags & B_CACHE) == 0) {
879 bp->b_iocmd = BIO_READ;
880 vfs_busy_pages(bp, 0);
881 error = nfs_doio(bp, cred, td);
891 if (bp->b_wcred == NOCRED)
892 bp->b_wcred = crhold(cred);
893 np->n_flag |= NMODIFIED;
896 * If dirtyend exceeds file size, chop it down. This should
897 * not normally occur but there is an append race where it
898 * might occur XXX, so we log it.
900 * If the chopping creates a reverse-indexed or degenerate
901 * situation with dirtyoff/end, we 0 both of them.
904 if (bp->b_dirtyend > bcount) {
905 printf("NFS append race @%lx:%d\n",
906 (long)bp->b_blkno * DEV_BSIZE,
907 bp->b_dirtyend - bcount);
908 bp->b_dirtyend = bcount;
911 if (bp->b_dirtyoff >= bp->b_dirtyend)
912 bp->b_dirtyoff = bp->b_dirtyend = 0;
915 * If the new write will leave a contiguous dirty
916 * area, just update the b_dirtyoff and b_dirtyend,
917 * otherwise force a write rpc of the old dirty area.
919 * While it is possible to merge discontiguous writes due to
920 * our having a B_CACHE buffer ( and thus valid read data
921 * for the hole), we don't because it could lead to
922 * significant cache coherency problems with multiple clients,
923 * especially if locking is implemented later on.
925 * as an optimization we could theoretically maintain
926 * a linked list of discontinuous areas, but we would still
927 * have to commit them separately so there isn't much
928 * advantage to it except perhaps a bit of asynchronization.
931 if (bp->b_dirtyend > 0 &&
932 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
933 if (BUF_WRITE(bp) == EINTR)
938 error = uiomove((char *)bp->b_data + on, n, uio);
941 * Since this block is being modified, it must be written
942 * again and not just committed. Since write clustering does
943 * not work for the stage 1 data write, only the stage 2
944 * commit rpc, we have to clear B_CLUSTEROK as well.
946 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
949 bp->b_ioflags |= BIO_ERROR;
955 * Only update dirtyoff/dirtyend if not a degenerate
959 if (bp->b_dirtyend > 0) {
960 bp->b_dirtyoff = min(on, bp->b_dirtyoff);
961 bp->b_dirtyend = max((on + n), bp->b_dirtyend);
964 bp->b_dirtyend = on + n;
966 vfs_bio_set_validclean(bp, on, n);
969 * If IO_NOWDRAIN then set B_NOWDRAIN (nfs-backed MD
972 if (ioflag & IO_NOWDRAIN)
973 bp->b_flags |= B_NOWDRAIN;
976 * If IO_SYNC do bwrite().
978 * IO_INVAL appears to be unused. The idea appears to be
979 * to turn off caching in this case. Very odd. XXX
981 if ((ioflag & IO_SYNC)) {
982 if (ioflag & IO_INVAL)
983 bp->b_flags |= B_NOCACHE;
984 error = BUF_WRITE(bp);
987 } else if ((n + on) == biosize) {
988 bp->b_flags |= B_ASYNC;
989 (void)nfs_writebp(bp, 0, 0);
993 } while (uio->uio_resid > 0 && n > 0);
996 nfs_rsunlock(np, td);
1002 * Get an nfs cache block.
1004 * Allocate a new one if the block isn't currently in the cache
1005 * and return the block marked busy. If the calling process is
1006 * interrupted by a signal for an interruptible mount point, return
1009 * The caller must carefully deal with the possible B_INVAL state of
1010 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it
1011 * indirectly), so synchronous reads can be issued without worrying about
1012 * the B_INVAL state. We have to be a little more careful when dealing
1013 * with writes (see comments in nfs_write()) when extending a file past
1017 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td)
1021 struct nfsmount *nmp;
1026 if (nmp->nm_flag & NFSMNT_INT) {
1027 bp = getblk(vp, bn, size, PCATCH, 0);
1028 while (bp == (struct buf *)0) {
1029 if (nfs_sigintr(nmp, (struct nfsreq *)0, td->td_proc))
1030 return ((struct buf *)0);
1031 bp = getblk(vp, bn, size, 0, 2 * hz);
1034 bp = getblk(vp, bn, size, 0, 0);
1037 if (vp->v_type == VREG) {
1040 biosize = mp->mnt_stat.f_iosize;
1041 bp->b_blkno = bn * (biosize / DEV_BSIZE);
1047 * Flush and invalidate all dirty buffers. If another process is already
1048 * doing the flush, just wait for completion.
1051 nfs_vinvalbuf(struct vnode *vp, int flags, struct ucred *cred,
1052 struct thread *td, int intrflg)
1054 struct nfsnode *np = VTONFS(vp);
1055 struct nfsmount *nmp = VFSTONFS(vp->v_mount);
1056 int error = 0, slpflag, slptimeo;
1058 if (vp->v_flag & VXLOCK) {
1062 if ((nmp->nm_flag & NFSMNT_INT) == 0)
1072 * First wait for any other process doing a flush to complete.
1074 while (np->n_flag & NFLUSHINPROG) {
1075 np->n_flag |= NFLUSHWANT;
1076 error = tsleep((caddr_t)&np->n_flag, PRIBIO + 2, "nfsvinval",
1078 if (error && intrflg &&
1079 nfs_sigintr(nmp, (struct nfsreq *)0, td->td_proc))
1084 * Now, flush as required.
1086 np->n_flag |= NFLUSHINPROG;
1087 error = vinvalbuf(vp, flags, cred, td, slpflag, 0);
1090 nfs_sigintr(nmp, (struct nfsreq *)0, td->td_proc)) {
1091 np->n_flag &= ~NFLUSHINPROG;
1092 if (np->n_flag & NFLUSHWANT) {
1093 np->n_flag &= ~NFLUSHWANT;
1094 wakeup((caddr_t)&np->n_flag);
1098 error = vinvalbuf(vp, flags, cred, td, 0, slptimeo);
1100 np->n_flag &= ~(NMODIFIED | NFLUSHINPROG);
1101 if (np->n_flag & NFLUSHWANT) {
1102 np->n_flag &= ~NFLUSHWANT;
1103 wakeup((caddr_t)&np->n_flag);
1109 * Initiate asynchronous I/O. Return an error if no nfsiods are available.
1110 * This is mainly to avoid queueing async I/O requests when the nfsiods
1111 * are all hung on a dead server.
1113 * Note: nfs_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp
1114 * is eventually dequeued by the async daemon, nfs_doio() *will*.
1117 nfs_asyncio(struct buf *bp, struct ucred *cred, struct thread *td)
1119 struct nfsmount *nmp;
1126 nmp = VFSTONFS(bp->b_vp->v_mount);
1129 * Commits are usually short and sweet so lets save some cpu and
1130 * leave the async daemons for more important rpc's (such as reads
1133 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) &&
1134 (nmp->nm_bufqiods > nfs_numasync / 2)) {
1139 if (nmp->nm_flag & NFSMNT_INT)
1144 * Find a free iod to process this request.
1146 for (iod = 0; iod < nfs_numasync; iod++)
1147 if (nfs_iodwant[iod]) {
1153 * Try to create one if none are free.
1156 iod = nfs_nfsiodnew();
1163 * Found one, so wake it up and tell it which
1166 NFS_DPF(ASYNCIO, ("nfs_asyncio: waking iod %d for mount %p\n",
1168 nfs_iodwant[iod] = (struct proc *)0;
1169 nfs_iodmount[iod] = nmp;
1171 wakeup((caddr_t)&nfs_iodwant[iod]);
1175 * If none are free, we may already have an iod working on this mount
1176 * point. If so, it will process our request.
1179 if (nmp->nm_bufqiods > 0) {
1181 ("nfs_asyncio: %d iods are already processing mount %p\n",
1182 nmp->nm_bufqiods, nmp));
1188 * If we have an iod which can process the request, then queue
1193 * Ensure that the queue never grows too large. We still want
1194 * to asynchronize so we block rather then return EIO.
1196 while (nmp->nm_bufqlen >= 2*nfs_numasync) {
1198 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp));
1199 nmp->nm_bufqwant = TRUE;
1200 error = tsleep(&nmp->nm_bufq, slpflag | PRIBIO,
1201 "nfsaio", slptimeo);
1203 if (nfs_sigintr(nmp, NULL, td ? td->td_proc : NULL))
1205 if (slpflag == PCATCH) {
1211 * We might have lost our iod while sleeping,
1212 * so check and loop if nescessary.
1214 if (nmp->nm_bufqiods == 0) {
1216 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp));
1221 if (bp->b_iocmd == BIO_READ) {
1222 if (bp->b_rcred == NOCRED && cred != NOCRED)
1223 bp->b_rcred = crhold(cred);
1225 bp->b_flags |= B_WRITEINPROG;
1226 if (bp->b_wcred == NOCRED && cred != NOCRED)
1227 bp->b_wcred = crhold(cred);
1231 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist);
1237 * All the iods are busy on other mounts, so return EIO to
1238 * force the caller to process the i/o synchronously.
1240 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n"));
1245 * Do an I/O operation to/from a cache block. This may be called
1246 * synchronously or from an nfsiod.
1249 nfs_doio(struct buf *bp, struct ucred *cr, struct thread *td)
1254 struct nfsmount *nmp;
1255 int error = 0, iomode, must_commit = 0;
1258 struct proc *p = td ? td->td_proc : NULL;
1262 nmp = VFSTONFS(vp->v_mount);
1264 uiop->uio_iov = &io;
1265 uiop->uio_iovcnt = 1;
1266 uiop->uio_segflg = UIO_SYSSPACE;
1270 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We
1271 * do this here so we do not have to do it in all the code that
1274 bp->b_flags &= ~B_INVAL;
1275 bp->b_ioflags &= ~BIO_ERROR;
1277 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp));
1280 * Historically, paging was done with physio, but no more.
1282 if (bp->b_flags & B_PHYS) {
1284 * ...though reading /dev/drum still gets us here.
1286 io.iov_len = uiop->uio_resid = bp->b_bcount;
1287 /* mapping was done by vmapbuf() */
1288 io.iov_base = bp->b_data;
1289 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1290 if (bp->b_iocmd == BIO_READ) {
1291 uiop->uio_rw = UIO_READ;
1292 nfsstats.read_physios++;
1293 error = nfs_readrpc(vp, uiop, cr);
1297 iomode = NFSV3WRITE_DATASYNC;
1298 uiop->uio_rw = UIO_WRITE;
1299 nfsstats.write_physios++;
1300 error = nfs_writerpc(vp, uiop, cr, &iomode, &com);
1303 bp->b_ioflags |= BIO_ERROR;
1304 bp->b_error = error;
1306 } else if (bp->b_iocmd == BIO_READ) {
1307 io.iov_len = uiop->uio_resid = bp->b_bcount;
1308 io.iov_base = bp->b_data;
1309 uiop->uio_rw = UIO_READ;
1311 switch (vp->v_type) {
1313 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
1314 nfsstats.read_bios++;
1315 error = nfs_readrpc(vp, uiop, cr);
1318 if (uiop->uio_resid) {
1320 * If we had a short read with no error, we must have
1321 * hit a file hole. We should zero-fill the remainder.
1322 * This can also occur if the server hits the file EOF.
1324 * Holes used to be able to occur due to pending
1325 * writes, but that is not possible any longer.
1327 int nread = bp->b_bcount - uiop->uio_resid;
1328 int left = uiop->uio_resid;
1331 bzero((char *)bp->b_data + nread, left);
1332 uiop->uio_resid = 0;
1335 if (p && (vp->v_flag & VTEXT) &&
1336 (np->n_mtime != np->n_vattr.va_mtime.tv_sec)) {
1337 uprintf("Process killed due to text file modification\n");
1339 psignal(p, SIGKILL);
1345 uiop->uio_offset = (off_t)0;
1346 nfsstats.readlink_bios++;
1347 error = nfs_readlinkrpc(vp, uiop, cr);
1350 nfsstats.readdir_bios++;
1351 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ;
1352 if (nmp->nm_flag & NFSMNT_RDIRPLUS) {
1353 error = nfs_readdirplusrpc(vp, uiop, cr);
1354 if (error == NFSERR_NOTSUPP)
1355 nmp->nm_flag &= ~NFSMNT_RDIRPLUS;
1357 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0)
1358 error = nfs_readdirrpc(vp, uiop, cr);
1360 * end-of-directory sets B_INVAL but does not generate an
1363 if (error == 0 && uiop->uio_resid == bp->b_bcount)
1364 bp->b_flags |= B_INVAL;
1367 printf("nfs_doio: type %x unexpected\n", vp->v_type);
1371 bp->b_ioflags |= BIO_ERROR;
1372 bp->b_error = error;
1376 * If we only need to commit, try to commit
1378 if (bp->b_flags & B_NEEDCOMMIT) {
1382 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff;
1383 bp->b_flags |= B_WRITEINPROG;
1385 bp->b_vp, off, bp->b_dirtyend-bp->b_dirtyoff,
1387 bp->b_flags &= ~B_WRITEINPROG;
1389 bp->b_dirtyoff = bp->b_dirtyend = 0;
1390 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1395 if (retv == NFSERR_STALEWRITEVERF) {
1396 nfs_clearcommit(bp->b_vp->v_mount);
1401 * Setup for actual write
1404 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size)
1405 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE;
1407 if (bp->b_dirtyend > bp->b_dirtyoff) {
1408 io.iov_len = uiop->uio_resid = bp->b_dirtyend
1410 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE
1412 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
1413 uiop->uio_rw = UIO_WRITE;
1414 nfsstats.write_bios++;
1416 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
1417 iomode = NFSV3WRITE_UNSTABLE;
1419 iomode = NFSV3WRITE_FILESYNC;
1421 bp->b_flags |= B_WRITEINPROG;
1422 error = nfs_writerpc(vp, uiop, cr, &iomode, &must_commit);
1425 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try
1426 * to cluster the buffers needing commit. This will allow
1427 * the system to submit a single commit rpc for the whole
1428 * cluster. We can do this even if the buffer is not 100%
1429 * dirty (relative to the NFS blocksize), so we optimize the
1430 * append-to-file-case.
1432 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be
1433 * cleared because write clustering only works for commit
1434 * rpc's, not for the data portion of the write).
1437 if (!error && iomode == NFSV3WRITE_UNSTABLE) {
1438 bp->b_flags |= B_NEEDCOMMIT;
1439 if (bp->b_dirtyoff == 0
1440 && bp->b_dirtyend == bp->b_bcount)
1441 bp->b_flags |= B_CLUSTEROK;
1443 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
1445 bp->b_flags &= ~B_WRITEINPROG;
1448 * For an interrupted write, the buffer is still valid
1449 * and the write hasn't been pushed to the server yet,
1450 * so we can't set BIO_ERROR and report the interruption
1451 * by setting B_EINTR. For the B_ASYNC case, B_EINTR
1452 * is not relevant, so the rpc attempt is essentially
1453 * a noop. For the case of a V3 write rpc not being
1454 * committed to stable storage, the block is still
1455 * dirty and requires either a commit rpc or another
1456 * write rpc with iomode == NFSV3WRITE_FILESYNC before
1457 * the block is reused. This is indicated by setting
1458 * the B_DELWRI and B_NEEDCOMMIT flags.
1460 * If the buffer is marked B_PAGING, it does not reside on
1461 * the vp's paging queues so we cannot call bdirty(). The
1462 * bp in this case is not an NFS cache block so we should
1466 || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
1470 bp->b_flags &= ~(B_INVAL|B_NOCACHE);
1471 if ((bp->b_flags & B_PAGING) == 0) {
1473 bp->b_flags &= ~B_DONE;
1475 if (error && (bp->b_flags & B_ASYNC) == 0)
1476 bp->b_flags |= B_EINTR;
1480 bp->b_ioflags |= BIO_ERROR;
1481 bp->b_error = np->n_error = error;
1482 np->n_flag |= NWRITEERR;
1484 bp->b_dirtyoff = bp->b_dirtyend = 0;
1492 bp->b_resid = uiop->uio_resid;
1494 nfs_clearcommit(vp->v_mount);
1500 * Used to aid in handling ftruncate() operations on the NFS client side.
1501 * Truncation creates a number of special problems for NFS. We have to
1502 * throw away VM pages and buffer cache buffers that are beyond EOF, and
1503 * we have to properly handle VM pages or (potentially dirty) buffers
1504 * that straddle the truncation point.
1508 nfs_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize)
1510 struct nfsnode *np = VTONFS(vp);
1511 u_quad_t tsize = np->n_size;
1512 int biosize = vp->v_mount->mnt_stat.f_iosize;
1517 if (np->n_size < tsize) {
1523 * vtruncbuf() doesn't get the buffer overlapping the
1524 * truncation point. We may have a B_DELWRI and/or B_CACHE
1525 * buffer that now needs to be truncated.
1527 error = vtruncbuf(vp, cred, td, nsize, biosize);
1528 lbn = nsize / biosize;
1529 bufsize = nsize & (biosize - 1);
1530 bp = nfs_getcacheblk(vp, lbn, bufsize, td);
1531 if (bp->b_dirtyoff > bp->b_bcount)
1532 bp->b_dirtyoff = bp->b_bcount;
1533 if (bp->b_dirtyend > bp->b_bcount)
1534 bp->b_dirtyend = bp->b_bcount;
1535 bp->b_flags |= B_RELBUF; /* don't leave garbage around */
1538 vnode_pager_setsize(vp, nsize);